Split (1)

train · 148k rows

Search is not available for this dataset

text stringlengths 4.99k 445k
# Introduction Many researchers' results indicate that patients with adolescent idiopathic scoliosis, (AIS) suffer from back pain more frequently than healthy populations. It was found that the incidence of back pain in patients with scoliosis treated conservatively was lower than that of the AIS study group (11% vs. 32%). It was revealed that the stabilization of scoliosis with a brace may decrease the likelihood of back pain induced by the altered scoliotic spinal mechanics, and presumably a failure to stabilize the spine (curve progression) may increase the incidence of spinal pain. Additionally, some associations between types of orthoses and pain intensity in AIS patients have been discovered. Considering the incidence of spinal pain in the long-term follow-up to treatment for AIS, most results refer to patients treated surgically. For instance, Danielsson et al. indicated that, more than 20 years after the completion of treatment with a brace, AIS patients had significantly more degenerative lumbar disc changes than the controls. They indicated that, even if conservatively- treated patients admitted to having pain more often than the control group, their pain was mostly mild, analgesics were rarely used, and no patient had a major functional deficit. Weinstein et al. presented a 50-year follow-up of non- operatively treated late-onset persons with scoliosis. Overall, AIS patients were found to be “productive and functional at a high level” with some restraints due to back pain and cosmetic concerns. Meanwhile, Gabos et al., in a study of long-term radiographic and functional outcomes of females with AIS who had completed a Wilmington orthosis treatment, found that there was no significant overall difference between the group who had undergone orthotic treatment and the control group in terms of back pain or physical, functional or personal care activities. However, patients reported significantly greater difficulty with selected positional activities. However, it must be emphasized that these findings may relate either to the condition of scoliosis or the wearing of the brace and it is not possible to make such a distinction. At the same time, very little has been written about problems in the cervical spine and neck pain-related difficulties in performing everyday activities in AIS patients, especially in a long-term follow-up after the completion of brace treatment. The first signs of a problem were described by Moskowitz et al., who reviewed 61 patients who had had a posterior spine fusion for any type of scoliosis 25 years previously (range 20–30 yrs). An unexpectedly high incidence of cervical complaints (57%) was reported amongst those patients. Hilibrand et al. published a radiographic study of the sagittal alignment of the cervical spine in patients with AIS. As this was a study only of alignment and only in adolescents, no comments could be made regarding adult pain problems or the assessment of cervical degeneration. Edgar and Mehta, in the only study involving the long-term follow-up of fused and unfused persons with scoliosis, found an incidence of cervicodorsal pain in 7.8% of the unfused patients and in 17.6% of the surgically treated patients. Therefore, taking into account some inconsistencies between study results relating to back and neck pain in AIS patients after completing Milwaukee brace treatment and, in particular, the lack of comprehensive long-term assessments of cervical pain in adult persons with scoliosis, special emphasis should be placed on the following research questions: do adult persons with scoliosis treated with a Milwaukee brace during adolescence have more lower back pain (LBP) and neck pain compared with non-scoliotic subjects? How does spinal pain affect the daily life and the activities of these patients? Is there any correlation between the intensity of the pain, radiographic data and brace treatment-related variables? Thus, we aimed to provide a complex assessment of adult females with scoliosis treated with a Milwaukee brace, in a minimum 23-yrs follow-up. To sum up, the purpose of this report is to explore the long-term outcomes in terms of back and neck pain and functionality, in relation to radiological, clinical, pulmonary function and socio-demographic data. Our hypothesis is that no difference is expected between study and control groups in terms of socio-demographic characteristics, pain level and neck- or low back pain-disability. In addition, we assumed that there will be a significant relationship between pain intensity, patients’ everyday activities and radiographic, clinical, pulmonary and brace treatment-related variables. All study aims have been achieved. # Material and methods ## Study design Results concerning back- and neck pain and function in consecutively selected adult AIS females (scoliosis group-SG) treated with a Milwaukee brace were evaluated. Based on an extensive search of Pediatric Orthopedics and Traumatology Clinic charts, we retrospectively reviewed the clinical records and radiographs of all female patients who had successfully completed a course of treatment with the Milwaukee orthosis between 1974 and 1990. In those patients, the Risser sign 4 and minimum two years post-menarche was defined as a maturity, after that time the brace treatment was completed. Forty patients met the criteria for inclusion, but due to change some personal details (such as address or family name after getting married), not all of them were contacted. Finally, 30 women returned for a follow-up evaluation. To make the obtained data more reliable, an age- and sex-matched control group of 42 healthy individuals (healthy controls group–HG) was selected and asked to complete the same questionnaires as the patient group. ## Study protocol Study participants from both groups were examined using the same protocol, except for the radiological evaluation. The contents of interview in both study groups included age, work, marital status, number of children, incidence of caesarian sections and complications during delivery, educational level, place of residence and the pursuit of active hobbies. In addition, to evaluate the implications of the brace for back and neck pain and functionality, all study participants were asked to fill in the same series of questionnaires. ## Clinical and radiological examination Clinical and radiological examinations were performed three times, before and after completing the treatment and then in the follow-up, and were taken in an upright position with the iliac ala exposed in an anterior-posterior projection. Data concerning former treatment regimens and radiological findings were gathered from reviews of charts and radiographs. The physical examinations were performed by AO, the 3<sup>rd</sup> study author, and the radiographic measurements were conducted by JG and MG, the 2<sup>nd</sup> and the 5<sup>th</sup> study authors, respectively. The success rate at maturity was calculated according to Nachemson and Peterson, who defined a successful treatment as an increase in the curve of less than 6° from the start of bracing. The change in curvature from the end of treatment until the follow-up examination was also assessed. Pulmonary function, in terms of vital capacity (VC), was evaluated three times, before and after completing the treatment and then in the follow-up. The VC examinations at the present follow-up were performed by ML, the 4<sup>th</sup> study author. ## Healthy controls The exclusion criteria for the control group were as follows: previous back surgery or significant scoliosis, which was ruled out by clinical examination, including using Perdriolli’s scoliometer. None of the controls had a trunk rotation of more than 5°, according to Danielsson et al.. ## Patient sample Thirty AIS patients, with a minimum of 23-yrs of follow-up after completing Milwaukee brace treatment, were included in the study. All treatments were completed before the patient reached 19 yrs of age. In all cases, scoliosis had not been detected before the age of 10 and was not combined with any major spinal deformities at the time when the brace treatment was implemented. In addition, patients were excluded from the study if they, at the time of the follow-up examinations, suffered from any other disease leading to trunk deformity. The investigated sample sizes are equivalent (p = 0.157). During the whole treatment non-compliance monitoring such as temperature probe was not used. The average values were based on interviews with patients during the clinical examination. In addition, to make the obtained data more credible, daily duration of brace wearing was confirmed during the separate interviews carried out with parents. ## Ethical issues All study participants were informed in detail about the objective of the study. They understood that they would be anonymous and that their personal information would not be disclosed. All participants signed written informed consent forms in order to participate in the study. The study was approved by the Bioethics Commission of Poznan University of Medical Sciencesand was carried out following universal ethical principles. ## Evaluation of neck and back pain-related disability Polish versions of Revised Oswestry Low Back Pain Disability Index (RODI), Rolland-Morris Questionnaire (RMQ), Quebec Back Pain Disability Scale (QDS), Neck Disability Index (NDI) and Copenhagen Neck Functional Disability Scale (CNFDS) enabled us to assess the intensity of cervical and lumbar pain and the ability to perform everyday activities among both study samples. The RODI is a revised version of the original Oswestry Disability Index and focuses on, alongside the subjective evaluation of pain intensity, the degree to which everyday activities such as personal care, lifting, walking, sitting, rising/standing, sleep, social life, traveling and changes in pain intensity are affected. The answers are marked on a six-point scale (from 0 to 5), where 0 corresponds to no limitations on functional status, and 5 indicates maximum restrictions. The maximum sum of points from all the sections is 50. In order to present the general result in percentage values, reflecting the extent to which the ability to carry out everyday activities is restricted, the total is divided by 2. A result of 0–4 points is interpreted as no limitation to everyday activities, 5–14 points indicates mild limitation, 15–24 points is moderate limitation, 25–34 points indicates a serious disorder, and a score over 34 points is evidence of disability.. The RODI's psychometric properties have been well established. The RODI correlates with other outcome measures aiming at measuring disability due to LBP. The RODI shows good construct validity because it was used as the standard of comparison for other outcome measures assessing LBP-induced disability. Internal consistency has been shown to be at the acceptable level by different authors. Cronbach alpha ranges from 0.71 to 0.87. Test-retest reliability has been shown to be high. Values range from r = 0.83 to 0.99 and vary according to the time interval between measurements. The longer the wait between repeated measure, the lower the score becomes. Intraclass correlation coefficient values from 0.84 to 0.94 have been reported. Responsiveness has been reported to be high. The most common method for measuring responsiveness found in the current literature search was a receiver operating characteristic curve. It can also be used to estimate the minimum clinically important difference (MCID). Values for the area under the curve range from 0.723 to 0.94. The MCID has been reported to be between 4 and 10.5 points. The RMQ is widely used to evaluate the degree of disability due to lumbosacral back pain. It comprises 24 statements requiring a “yes”/”no” answer. The total score can range from 0 (no disability) to 24 (severe disability). The subjects were divided into the following four disability groups according to their scores: low degree of disability: 4–10 points, moderate: 11–17 and severe: 18–24. If the result was ≤ 3, the patient was considered to have no disability. Test-retest reliability of RMQ was 0.91, and Cronbach's alpha coefficients in outpatients with low back pain diagnosed with a musculoskeletal origin were 0.90, 0.84, 0.89, 0.92. Considering responsiveness in patients experiencing lower back pain, the Standard Response Mean was 0.55 (95% CI = -0.54 to 1.64). The QDS is comprised of 20 questions and 6 domains, reflecting everyday activities such as sleep/rest; sitting/rising; walking; moving; bending/squatting; lifting heavy objects. The responses are marked on a scale of 0–5, where 0 corresponds to no limitations, whereas 5 signifies maximum restrictions to everyday functional status. The overall result varies from 0 (no worsening of spine function) to 100 (maximum restrictions on functional status). Test-retest reliability of QDS was 0.92, and Cronbach's alpha coefficient was 0.96. The scale correlated as expected with other measures of disability, pain, medical history, and utilization variables, work-related variables, and socio- demographic characteristics. Significant changes in disability over time, and differences in change scores between patients that were expected to differ in the direction of change were also fund. The NDI questionnaire assesses pain intensity and related limitations to cervical spine function during everyday activities. The NDI is comprised of 10 questions regarding: pain intensity, personal care, lifting, reading, headaches, concentration, work, driving, sleeping, and recreation. Each item is scored from 0 (no disability) to 5 (total disability). The maximum possible score is 50. However, the total score obtained is often doubled to give a percentage score, out of 100. The interpretation is as follows: 0–20 normal, 21–40 mild disability, 41–60 moderate, 61–80 severe and \<80 complete or exaggerated disability. Test-retest reliability of NDI resulted in good statistical significance (Pearson's r = 0.89). The Cronbach alpha coefficients were calculated from a pool of questionnaires completed by 52 such subjects resulting in a total index alpha of 0.80, with all items having individual alpha scores above 0.75. Concerning concurrent validity, NDI scores were compared to scores on the McGill Pain Questionnaire, with similar moderately high correlations (0.69–0.70), whereas the responsiveness was excellent (effect size = 0.85). The CNFDS consists of 15 items that evaluate the impact of neck pain. Three items evaluate pain severity directly, including the patient’s perception of the future impact of neck pain, eight items evaluate disability during everyday activities and four items focus on social interaction and recreation. There are three possible answers to select from each item; ‘‘yes” (2 points), ‘‘occasionally” (1 point), and ‘‘no” (0 points). For items 1–5, however, the scoring is reversed and here “yes” carries a score of 0, “occasionally” 1 and “no” 2. The highest score attainable is 30, indicating the worst possible impact, and the lowest is 0, where no impact of neck pain can be identified. The Cronbach's alpha coefficient of the CNFDS for internal consistency was 0.9 for the entire scale, and the coefficients for individual items were all greater than 0.88. Disability scale scores correlated strongly to pain scores as well as to doctor and patient global assessments, indicating good construct validity. Relative changes in disability scores demonstrated a moderately strong correlation to changes in pain scores after treatment. Scale scores correlated weakly to all physical measurements. The disability scale demonstrated excellent practicality and reliability. The scale accurately reflects patient perceptions regarding functional status and pain as well as doctor's global assessment and is responsive to change over long periods of time. ## Statistical analysis For statistical quantitative (numerical) features, e.g. age, apical translation, Cobb angle, number of children or questionnaire results, we calculated the mean, a 95% confidence interval, the range and standard deviation (SD). With respect to qualitative features (information that has aspects that are impossible to be measured), e.g. curve type, educational level, marital status or place of residence, we gave the number of units that belong to the described categories of a given feature and respective percentages. To determine if the investigated sample sizes are equivalent, the chi-square test was used. The chi-square test was used to compare qualitative features between persons with scoliosis and healthy controls. In addition, a Mann-Whitney test was utilized to compare differences between both groups in regard to quantitative characteristics. To establish relations between quantitative data such as e.g. age, duration of brace application, apical translation, Cobb angle, and questionnaire results, we used Spearman's rank correlation (marked as rS). To determine dependency between quantitative and qualitative characteristics, e.g. between questionnaire numerical data and marital status, place of residence or curve type, ANOVA Kruskal-Wallis test was used. A p\<0.05 indicates statistical significance. Statistical calculations were performed by Statistica software. See the supplementary material file containing clinical, radiological, socio-demographic and questionnaire data. # Results ## Patient clinical and radiological data Patients’ follow-up period was mean 27.77 yrs ± SD 3.30 (range 23–35). The Milwaukee brace was worn for a mean of 22.9 hrs daily ± SD 0.31 (range 22–23). The duration of treatment was mean 45.47 months ± SD 20.00 (range 24–104). Following the criteria of the Scoliosis Research Society regarding the location of the apex, thoracic scoliosis was identified in 21 patients (70%), thoracolumbar in 2 patients (6.67%) and lumbar curves were identified in 7 AIS females (23.33%). Radiographic examinations at the beginning of brace the treatment resulted in Risser Grade 0 in 19 patients (63.33%), Risser Grade I in 2 patients (6.67%), and Risser Grade II in 9 patients (30%). Risser Grade IV was identified after completing treatment in all patients involved in the study (100%). The European Risser sign was used to define skeletal maturity status. The success rate at maturity, according to Nachemson and Peterson, was identified in 16 patients (53.33%). Five patients (16.67%) were eligible for scoliosis surgery after completing the brace treatment, but refused to undergo an operation. Change in spine curvature from the end of treatment until the follow-up was mean 9.1 angles ± SD 7.64 (range 0–27). Mean VC was 2526.67 cm<sup>3</sup> ± SD 561.98 in the pretreatment evaluation, 2983.33 cm<sup>3</sup> ± SD 617.05 after treatment and 2569.64 cm<sup>3</sup> ± SD 591.41 in the final evaluation. For additional clinical and radiological characteristics of the patient sample, see. ## Socio-demographic data The age of the patients during the follow-up was mean 41.13 yrs ± SD 3.87 (range 35–55), whereas age of the controls was mean 42.05 yrs ± SD 7.41 (range 22–61). Twenty-eight females with AIS (93.4%) and 29 healthy controls (69%) were married. Of 30 female patients, 29 (96.67%) had children, and the number of children was mean 2.0 ± SD 0.83 (range 0–4), whereas in HC group 33 females (78.57%) had children, and the number of children in this subgroup was mean 1.48 ± SD 0.99 (range 0–3). The rate of cesarean section was 30% (9 patients) in SG and 27.3% (9 controls) in HG. Ten patients (34.48%) and 8 controls (23.53%) had experienced problems during delivery (for additional data, see). ## Low back pain and function Regarding RODI, patients scored mean 13.67 ± SD 6.84 and healthy controls scored mean 5.26 ± SD 5.40. Regarding results expressed as percentage values, SG scored 27.34%, whereas HG scored 10.52%, which is interpreted as moderate and minimal disability respectively. In SG, 8 patients (26.67%) showed minimal disability, 16 subjects (53.33%) reported moderate disability and severe disability was indicated in 6 patients (20%), whereas 37 healthy controls (88.10%) reported minimal disability, and 4 and 1 participants, that is, 9.52% and 2.38% respectively, suffered moderate and severe restrictions in everyday activities. The value of the RMQ was mean 5.60 ± SD 4.68 in the SG, indicating a low degree of disability, and mean 1.79 ± SD 2.35 in HG, meaning no disability. In SG 12 patients (40%) reported no disability, 13 subjects (43.34%) reported low degree of disability, moderate disability was indicated in 4 patients (13.33%), and severe disability was identified in 1 patient (3.33%). 26 healthy controls (85.72%) reported no disability, and 5 and 1 participants, or 11.90% and 2.38% respectively, revealed low and moderate levels of restriction. The mean QDS-PL scores were 23.28 Mean ± SD 14.53 in SG and 8.23 Mean ± SD 7.79 in HG. Considering the results for particular criteria, the highest scores, relating to the most severe limitations, are related to moving and leaning/squatting in both study samples (mean 5.13 ± SD 2.92, mean 4.24 ± SD 3.19 in SG and mean 1.81 ± SD 1.69, mean 1.57 ± SD 1.84 in HG respectively). ## Cervical pain and functionality The value of the NDI general result equaled mean 11.66 ± SD 7.43 in SG and mean 4.38 ± SD 4.02 in HG. Considering results expressed as percentage values, SG scored 23.32%, whereas HG scored 8.76%, which is interpreted as mild and no disability respectively. In SG 5 patients (16.67%) reported no disability, 15 subjects (50%) reported mild disability, moderate disability was indicated in 8 patients (26.67%) and 2 patients (6.67%) demonstrated severe restrictions to their everyday activities, whereas 27 healthy controls (64.29%) reported no disability, and 14 and 1 participants, or 33.33% and 2.38% respectively, revealed mild and moderate levels of disability. Concerning the CNPDS total score, patients scored mean 12.63 ± SD 3.48, whereas the value in the healthy control sample was mean 2.14 ± SD 3.48. The highest values among both SG and HG concern the disability subscale (mean 8.17 ± SD 4.2.57 and mean 1.24 ± SD 1.75 respectively). ## Comparative analyses Among a variety of socio-demographic variables which might be affected by an individual’s spinal function, significant differences could be found between the groups such as marital status (p = 0.043; patients from SG got married more often), educational level (p = 0.008; patients from SG achieved lower levels of education), place of residence (p = 0.028, patients from SG lived more often in smaller town and cities), working time per week (p = 0.013, patients from SG spent less time each week on work-related activities), no. of children (p = 0.046, patients from SG had more children) and time spent on active hobbies per week (p = 0.038, patients from SG spend less hrs each week on active hobbies). For details, see. It was also revealed, regarding RODI, RMQ, QDS, NDI and CNFDS (both for total scores and individual domains), statistically significant differences (p \<0.001), between both samples, have been confirmed, indicating higher levels of pain, and neck and lower back pain-related disability among persons with scoliosis. For details see. In particular, regarding RODI, significant differences concerning personal care, lifting, walking, sitting, rising/standing, social life, traveling and changes in pain intensity (p \< 0.001), except for sleeping (p = 0.099) were confirmed. For details see. ## Associations between pain, functionality, clinical and radiological factors Considering lower back pain-related physical impairment, we identified significant associations between RMQ and BMI before treatment (rS = -0.44) and Cobb angle after completing the treatment (rS = 0.36). We also identified significant associations between particular subscales of QDS (moving and BMI before treatment (rS = -0.42)), VC before treatment (rS = -0.39), between lifting and Cobb angle during follow-up examinations (rS = 0.38) and between total score and apical translation after completing the treatment (rS = 0.38) and thoracic kyphosis during the follow-up (rS = 0.39). ## Correlational analysis by means of socio-demographic characteristics Considering neck pain, a significant correlation was discovered between the level of education and CNPDS total score (patients with secondary education declared lower neck pain-related disability than patients with a university education, p = 0.035). Moreover, the level of education was found to be associated with pain severity, social interaction and disability (patients with an occupational education reported lower neck pain-related disability than patients with a university education, p = 0.012, p = 0.028 and p = 0.039, respectively). For details see. ## Associations between neck and low back pain-related disability As seen in, all total scores of RODI, RMQ, QDS, NDI and CNFDS display a significant intercorrelation. The strongest associations, above rS = 0.50, regard relationships between RODI and RMQ (rS = 0.76) QDS (rS = 0.70), NDI (rS = 0.69) and CNFDS (rS = 0.60). RMQ was associated with QDS (rS = 0.71) and NDI (rS = 0.69), whereas QDS correlated with NDI (rS = 0.80) and CNFDS (rS = 0.60). NDI was also associated with CNFDS (rS = 0.81). For details results see. # Discussion The aim of this report was to elucidate the long-term outcomes of back and neck pain and functionality in a group of patients treated with a Milwaukee brace. The Milwaukee brace has been a standard of nonsurgical treatment for scoliosis since 1954. The Milwaukee brace, which should be worn for 23 hours every day, is most effective for the treatment of adolescent idiopathic scoliosis. Spinal deformity is corrected with the Milwaukee brace, in theory, by both passive and active forces. Passive correction is achieved by direct pressure from the pads or by traction from the brace design. Active correction is believed to occur through active movement of the body away from pressure points, as the patient’s muscles acted to pull the trunk away from contact with the lateral pads or chin support. The Milwaukee brace remains the orthosis with the longest clinical history and the highest reported success rate in halting the progression of AIS, with the caveat that, because of the prior lack of standardization in bracing studies, direct comparison with studies of alternative brace designs is problematic. Currently, the Milwaukee brace is primarily prescribed for patients with thoracic apices above Th7, for control of upper thoracic sagittal deformities, and for other spinal deformities not amenable to treatment with lower-profile designs. A number of long-term follow-up studies on AIS have been published. However, most of them refer to patients treated surgically, with no comparison group of healthy controls. Furthermore, to date, assessment of spinal pain in persons with scoliosis in long-term evaluations has been mainly limited to LBP- related disability. To our knowledge, this is the first study of patients at least 23 years after the completion of Milwaukee brace treatment, involving a detailed evaluation of cervical pain and disability influencing everyday activities. A control group of randomly-selected healthy females was also included for comparative purposes. Several authors have claimed that idiopathic scoliosis is not associated with a higher incidence of back pain. For example, as mentioned above, Gabos et al. indicated that there was no significant overall difference between the group who had undergone orthotic treatment and the control group in terms of back pain, physical, functional and self-care activities. Lange et al. evaluated the long-term outcome in AIS patients 12 years or more after treatment with the Boston brace. 12% of patients reported that they had consulted a physician for back complaints during the last year before follow-up, and 28% had had physiotherapy. Overall, back function was considered excellent, good or fair in 95% of the patients. Haefeli et al. in a retrospective study on patients 10 to 60 years of age after nonoperative treatment for AIS investigated long-term outcome regarding pain, disability, psychological disturbance, and health-related quality of life (HRQOL). Although pain, disability, HRQOL, and general psychological well-being were found to be quite satisfactory, curve size was found to be a significant predictor of pain in the long term. Danielsson and Hallerman investigated consecutively-selected patients with idiopathic scoliosis, who were invited to a clinical follow-up at least 10 years after treatment with a brace or surgery. Overall, 77% reported back pain, but analgesic use was rare and 88% had normal back function as measured by the ODI. They concluded that most braced and surgically treated patients had a normal quality of life and that their physical functionality was only mildly impaired. Despite frequent back pain, back function was not severely affected. In the examined group of AIS patients, back pain and pain-related disability was a significant problem when compared to healthy controls, which confirms our primary hypothesis. The difference is statistically significant at p \< 0.001 for RODI, RMQ and QDS, both in terms of total score and certain everyday activities such as walking, moving, lifting or leaning/squatting. Moreover, according to RODI, 73.33% of patients reported a moderate or severe degree of disability, whereas, in line with the RMQ criteria, moderate or severe disability was reported by just 16.63% of respondents. However, mean levels of disability, according to RMQ, are still higher among SG. In addition, according to QDS subscales, the highest levels of disability involved difficulties with activities such as moving and leaning/squatting. Those results are contrary to conclusions drawn by the Danielsson and Hallerman, Lange et al. or Danielsson and Nachemson studies. Considering the associations between LBP-related disability and clinical and radiological factors, Danielsson and Nachemson indicated that pain and reduced back function was not related to a variety of curve-related factors, including degenerative lumbar disc changes. This lack of correlation between curve and functional outcome scores have been previously reported. Ramirez et al. indicated that back pain after brace treatment for AIS correlated with a family history of scoliosis, sport activities, and with curve progression during treatment, with the latter being the most significant. Patients with painful scoliosis which progresses during the bracing should be fully evaluated for an underlying non-idiopathic cause, but curve progression is the most likely explanation for the development of pain. The incidence of pain is significantly increased in patients with a family history of scoliosis or in patients engaged in sport activities. According to Ramirez et al. pain incidence was found not to correlate with age at presentation, age at brace initiation, length of follow- up, gender, menarchal status, and limb-length discrepancy. Curve magnitude at presentation or at the time of brace prescription and skeletal maturity also did not show any correlation with symptoms. Interestingly, our study result confirmed significant associations between RMQ and BMI before treatment, Cobb angle after completing the treatment, and between particular subscales of QDS: moving and BMI before treatment, VC before treatment, between lifting and Cobb angle at follow-up examinations and between total score and apical translation after completing the treatment and thoracic kyphosis at follow-up examinations. This indicates that back pain is associated, among others, with curve progression in a long follow-up after Milwaukee brace treatment. However, those results may relate either to the condition of scoliosis or the wearing of the brace and it is not possible to make such a distinction. As previously stated, very little has been said on neck pain and cervical pain- related restrictions in performing everyday activities, such as social interaction and recreational activities, especially in long-term follow-up after the completion of brace treatment. As previously stated, Moskowitz et al. revealed an unexpectedly high incidence of cervical complaints (57%) amongst AIS patients treated surgically in a 25-years follow-up, whereas Edgar and Mehta found an incidence of cervicodorsal pain in 7.8% of the unfused persons with scoliosis and in 17.6% of the surgically treated patients. Interestingly, results derived from the current study confirm our primary hypothesis and support results derived from previous research conducted by e.g. Moskovitz et al or Mehta et al. Regarding NDI results, most patients (83.34%) report mild, moderate or severe disability, whereas most of the controls (64.29%) reported no disability. Referring to CNPDS scores, SG revealed higher levels of restrictions regarding the function of the cervical spine, regarding pain level, disability and social interactions. Moreover, neck pain-related disability is not, unlike LBP-related restrictions, associated with current or past deformity-related scoliosis parameters, such Cobb angle or apical translation. Furthermore, some interesting associations discovered between the level of education and CNPDS must be discussed. It was revealed that patients with university education declared higher levels of neck pain-related disability than patients with lower levels of education. Those results are in accordance with several cross-sectional studies which also have found a positive association between the duration of occupational sitting and occurrence of pain in the neck–shoulder region, while prospective studies found associations between sitting and neck–shoulder pain. It is well documented that white-collar workers spend a substantial proportion of their time at work sitting. Thus, investigations of associations between sitting and neck–shoulder disorders are often conducted on workers in what is usually considered “sedentary” occupations, such as office-based jobs. In addition, concerning associations between all study questionnaires, we identified intercorrelations between neck and lower back pain-related disability. Bearing in mind the advantages and disadvantages of the correlational study design, we cannot provide a conclusive reason as to why this relationship exists, but at the same time we are able to make some predictions e.g. about the incidence of neck pain, based on the LBP already reported by AIS patients. ## Study limitations Some limitations of the current study must be pointed out. Firstly, a long-term assessment of conservative treatment of AIS is necessarily retrospective, since the tools specific to LBP- and neck pain-related disability assessment, such as RODI or NDI, were not available at the time when the conservative treatment was implemented. Secondly, the response rate (n = 30) was relatively low. Thirdly, the current study specifically investigated female patients only, which may have had an impact on the distribution of scores and limit the scope of the findings. Fourthly, brace compliance monitoring was a speculation and lastly, as indicated above, a correlational study cannot be used to draw conclusions about the causal relationships between selected variables. To sum up, we cannot determine which is the cause and which is the effect in the relationships between e.g. neck- and LBP-related disability, which are significantly intercorrelated. ## Clinical and future research implications Despite the aforementioned limitations, we can provide clinicians with reliable data concerning the incidence of lumbar and cervical pain and disability in the long-term follow-up after completing a Milwaukee brace treatment. In addition, we believe long-term follow-up studies can provide reliable information to patients who will undergo conservative treatment for AIS. Further investigations, with a longer follow-up and in a group of patients treated surgically for comparative purposes, are needed to improve upon the results of our research. ## Conclusions This retrospective examination of the function of the lumbar and cervical spine, in a highly specific subset of patients, revealed that persons with scoliosis treated in adolescence with a Milwaukee brace display significant limitations in everyday activities, due to LBP- and neck-related impairment. In addition, back pain is associated with curve progression in a long follow-up after conservative treatment. Moreover, LBP-related disability coexists with restrictions experienced due to neck pain. # Supporting information [^1]: The authors have declared that no competing interests exist.
# Introduction Outsourcing is the transfer of the continuous management responsibility of a client party or service user to a third party, referred to as a vendor or provider, to perform an IT service under a contract of service level agreement. Outsourcing has since become a well-established academic subject because of its rapidly developing practice. Outsourcing determinations, outsourcing methods, challenges identified in outsourcing etc. are some of the hot topics in the literature. Some of the advantages of outsourcing include ensuring maximum profit, lower costs, increased productivity, flexibility in meeting service needs, higher quality, redirecting company resources to core activities, more customer satisfaction, continuity and risk management, faster time to market, access to skilled resources, flexibility to focus on key areas, and faster and better services. Proponents of IT outsourcing have underlined the importance of managing outsourced relationships. "A long-term commitment, shared risk and benefits, a sense of reciprocal cooperation, and other features congruent with concepts and theories of participatory decision making" is what a relationship is characterised as. While IT outsourcing interactions have gotten some attention in the literature, just a few academics have looked at the client-supplier IT outsourcing relationship. Collaboration across corporate boundaries is a necessary component of today’s company. This type of client-vendor relationship frequently goes beyond the typical contractual restrictions agreed upon at the start of collaboration. The risks and advantages of joint labours, as well as investments and work load, are evenly distributed among the collaborative partners. Inter- and intra- organizational collaboration helps businesses gain a competitive advantage. Bidirectional trust, reciprocal dependency, and a win-win mindset between partners are the foundations of long-term working relationships. Collaboration is typically developed by businesses to reduce the costs of collecting the necessary information/understanding, capabilities, and competencies for well- organized professional operations. All IT outsourcing contracts include components of collaboration, virtual cooperation, and the demands of increasingly sophisticated systems. But one of the most essential variables in deciding the success or failure of virtual collaboration is trust. Good relationships are tactical assets that necessitate the ongoing management effort and focus. An outsourcing project’s success or failure is determined by a number of elements, including project size, duration, and, most importantly, contract design and management among different stakeholders. Therefore, additional research is required to understand how various parts of outsourcing interact and what the further consequences of the dynamics of these factors are on various organizational outcomes. Outsourcing has progressed through several stages of client-vendor relationships. Such phases of the relationship are briefly illustrated via the framework in as 1\. Dyadic outsourcing relationship A client in a dyadic relationship relies on only one vendor to meet all of their demands, which might range from simple to complex. The majority of research in the literature considers such relationship as one-to-one, implying that one client seeks services independently of others, and suppliers do the same. 2\. Multi-vendor relationship In a one-to-many relationship, a client seeks out more than one vendor to supply their services in order to meet the client’s needs or achieve its goals, and information regarding the division of labor is shared and discussed with all parties to the agreement. 3\. Co-sourcing relationship A many-to-one or co-sourcing relationship is described as a collaboration between multiple clients for the delivery of a service through a single vendor contract. 4\. Complex outsourcing relationship The phrase complex relationship or complex outsourcing relationship (COR), which is the subject of this study, refers to a relationship between several clients and multiple vendors (i.e., many-to-many), in which multiple clients rely on multiple vendors to meet their needs under the same contract. OR Multiple vendor organizations may work for multiple client organizations through the same contractual agreement or connection in a "complex outsourcing relationship (COR). It is worth noting that this study’s focus is on such a phase or relationship, which is exceptionally difficult to manage owing to the complexity of the situation and the various types of stakeholders involved. The aforementioned phases and types of outsourcing relationships are shown in. Good relationships and carefully drafted contracts have a significant impact on outsourcing. However, there is a lack of comprehensive guidance in this area, specifically in the context of complex outsourcing relationships, due to the dearth of studies described in the literature. Outsourcing has become more complex and diversified as a result of the expansion of numerous actors connected via a contract, such as a client and one or more vendors, each with their own unique qualities. Furthermore, because of its complexity, complex outsourcing typically necessitates interactions among various stakeholders from diverse regions and cultures, making it significantly more challenging to manage than traditional outsourcing. Moreover, complex outsourcing when compared to other types of outsourcing is extremely difficult because it necessitates a variety of control and coordination mechanisms for project management, which proportionally increases the risk of project failure. Therefore, in light of these points, this study aims to identify the challenges and practices in the context of complex outsourcing relationships. Based on this hypothesis, the authors of this study formulated the following research question: 1. RQ. 1: What are the key challenges, as identified in the literature, faced by multiple stakeholders (clients and vendors) in the context of complex outsourcing relationships? 2. RQ. 2: What are the practices/solutions, as identified in the literature, faced by multiple stakeholders (clients and vendors) in the context of complex outsourcing relationships? The remainder of this paper is structured as follows. Background information is provided in Section II. Section III provides a detailed description of the research methodology. The conclusions of this investigation are presented in section IV. The study’s shortcomings are discussed in Section V, and conclusions and future work are discussed in Section VI. The remaining sections provide further information. # Background In support of a thorough software, the author proposed a robust architecture for security and compliance challenges. Furthermore, by establishing software requirements based on a study of work patterns of service providers acting in complex IT outsourcing agreements and proposing a software architecture that meets these requirements, this study contributes to filling the gap in software support. In COR setups, there are a number of solutions that focus on IT infrastructure compliance, but they are typically ineffective and inefficient and do not work well in general. I am not aware of any solutions that support all of the services we use, necessitating the use of multiple solutions. Complex IS outsourcing initiatives have distinct tasks and stakeholder connections compared to standard IS outsourcing projects. Because of the growing worldwide distribution of IT services, which involves complex interorganizational connections and the dispersion of decision rights and responsibilities across numerous organizations, a thorough understanding of governance procedures is required. The authors build and offer a theoretical framework for explaining various types of outsourcing partnerships, as well as an analysis of why client organizations seeking IT services might choose more collaborative or complicated outsourcing arrangements than dyadic interactions. Moreover, the authors believe that the COR relationships in the future will be more in use. The authors suggested a methodology and decision-making approach for examining incentive schemes and creating outsourcing contracts that benefit both the outsourcer and vendor in COR. Complex outsourcing contracts are difficult to manage and of themselves, but when time zone differences, language barriers, significant distance, rare face- to-face meetings, and diversity–all of which are common offshore deal companions–the task becomes exponentially more difficult. Organizations that deal with complex outsourcing arrangements on a regular basis have intelligently implemented the lifecycle model explained in this article. Many companies struggle to properly manage one or more complex outsourcing agreements, generate value, and mitigate risk. This research examines the significance of trust in complex contracts and relationships frameworks that underpin global outsourcing, as well as the difficult conflict-resolution procedures that can be employed to restore trust. In, the authors want to see what an Indian supplier considers to be significant in order to manage some big and complex outsourcing partnerships, and specifically, how reciprocally profitable and long-term partnerships are developed and maintained with European customers. By offering a descriptive framework incorporating essential governance variables, this article seeks to deliver a good understanding for managing complex-type IT-outsourcing agreements. The findings of the analysis, as well as the framework itself, demonstrate the numerous complex difficulties that arise when managing sophisticated IT outsourcing partnerships. # Research methodology We conducted a systematic literature review to collect relevant data from the literature, which is a more complete and rigorous method than the traditional literature review (OLR). The SLR was carried out in three stages: planning, execution, and reporting. ## Search strategy To conduct an SLR, we tracked Kitchenham’s guidelines and Mendes et al.. In addition, depending on the study objectives, we used the population, intervention, comparison, and outcomes (PICO) criteria to select keywords and build search strings. Population: Clients and Vendors in Complex Outsourcing Relationships Intervention: Challenges, Practices Comparison: For the sake of this investigation, no comparisons are made. Relevance Outcomes: To Assist Clients and Vendors in Complex Outsourcing Relationships Furthermore, to build the search strings, we employed Boolean connectors AND and OR to connect the features of PICO. ## Search strings To avoid a search break by different database constraints, we developed three different strings. String 1: ((“Complex outsourcing” OR “Complex software outsourcing” OR “Complex Information Systems outsourcing” OR “complex IT outsourcing” OR “complex outsourcing relationship”) AND (Challenges OR issues OR barriers OR risks OR Practices OR solutions). String 2: ((“Complex outsourcing relationship” OR “Complex software outsourcing relationship” OR “Complex Information Systems outsourcing relationship” OR “complex IT outsourcing relationship” OR “Complex outsourcing”) AND (Challenges OR barriers OR issues OR risks OR Practices OR solutions) AND (“Relationship management in complex outsourcing” OR “Co-ordination in complex outsourcing” OR “Communication in complex outsourcing” OR “Contract management in complex outsourcing” OR “Task allocation in complex outsourcing” OR “Trust building in Complex outsourcing”)) String 3: (("Complex software outsourcing" OR "Complex IS/IT outsourcing" OR "complex outsourcing") AND ("Challenges" OR "issues" OR "barriers" OR "risks" OR "Practices" OR "solutions" solutions)) ## Literature resources Using our search strings, we applied our search for famous libraries, such as IEEE Xplore, Google Scholar, ACM, Wiley Online Library, SpringerLink, and ScienceDirect. We also used the snowballing method to further support our research questions and not to miss any important data. It is worth noting that we started the search on 21<sup>th</sup> February, 2021, and systematically completed it on 8<sup>th</sup> April, 2021. ## Criteria for study selection The strings and authors’ recommendations were backed up by the researchers. Initially, we added a string to the library for metadata. The same procedure was followed to avoid interference with the title, abstract, or keyword restrictions. Each paper was properly documented by the first author, who kept a complete record. Based on this phase, other authors assessed the papers and assigned pertinent information for each paper, as well as its title and abstract. We established the following inclusion and exclusion criteria based on the aforementioned principles for SLR. ### Inclusion criteria Such criteria specify which portions of the literature will be considered for inclusion in the selection. Based on such criteria, we have only analyzed items that are relevant to complex outsourcing. The inclusion criteria were as follows: - Articles in full text of the English language. - Sources only relevant to COR. - Studies that describe the COR challenges. - Studies that describe the COR practices. - Studies published in journals and conferences. - White papers and standardized reports from trustworthy organizations. ### Exclusion criteria Such criteria specify which pieces of literature are not included for consideration. The exclusion criteria were as follows: - Studies, not relevant to our research questions. - Studies of other than English language. - Incomplete Studies. - Duplicate studies - Thesis or magazine and/or web articles. The search outcomes are presented in. We retrieved 281 publications from a total of 1372 using the inclusion criteria. Using the exclusion criteria, we narrowed it to 85 papers. During data extraction, the principal author finished each step, which was then examined by other authors. The primary goal of quality assessment was to identify and eliminate lower- quality studies and to determine the validity of a study’s conclusions. ## Criteria for quality assessment We created the assessment criteria using the guidelines and updates for the SLR from previous studies. In this way, we formulated the questions given in prior to the implementation of such criteria. Furthermore, we employed a three-tiered scale to rate each question in the reviewed papers. Yes, no, or partial. We assigned values of 1 to Yes, 0.5, partly, and 0 to No in order to produce quantitative results. Furthermore, the work had to be graded on an average of 0.5. The principal author was in charge of applying the assessment parameter for quality to the studies, whereas the remaining authors were responsible for confirming the same assessments on a minority group of previously nominated studies. A few papers were omitted from the study using the same mechanism. However, as seen in, after going through the entire process, the final number of studies was 85. Any discrepancies were resolved through additional discussion. Any discrepancies were resolved through additional discussion. ## Data extraction Counting the challenges and their practices is a difficult task because the majority of the challenges and practices from such areas are cascading in nature. Nonetheless, we discovered 11 primary challenges with 67 practices from a total of 85 papers. The identified challenges and their practices are presented in the next section via Tables, respectively. Furthermore, the papers extracted via the SLR are presented in with the respective IDs and Title. # Results and discussion The goal of this research is to uncover the key issues and practices experienced by multiple stakeholders (such as clients and vendors) when dealing with COR. The following are the challenges that we have identified and analyzed after thorough conduction of the SLR in such areas. ## Challenges in COR presents the 11 key challenges with the given frequencies as follows: 1. 1. Control and Coordination Challenges: ‘Control and Coordination’ are one amongst the major challenge in COR. Control refers to a company’s capability to command or manage scattered events so that they meet the company’s goals. The ability of a company to coordinate these disparate activities is referred to as coordination. 1. 2. Decision Problems: Information system outsourcing decisions are difficult to make because they entail a number of elements, including (a) establishing and managing a long-term partnership with an independent agent, and (b) revealing the critical organizational assets for controlling outside agents. When the same matter is evaluated from the perspective of COR, it becomes more challenging. 1. 3. High Cost: The overall cost, for example, the hidden cost, estimated cost, control cost, coordination cost, etc., for outsourcing is high in the case of COR. 1. 4. Security and Compliance Challenges: Security and Compliance is a composite challenge, including six sub-challenges (see above) in the context of COR. It is worth noting that this particular challenge focused on the cloud version of COR. 1. 5. Risky Contracts: Managing risk is also a major issue, specifically in the context of COR, where multiple stakeholders from different cultures and zones are involved. Furthermore, large or complex IT contracts, which are nearly always incomplete, enhance the chance of risk in a variety of ways. 1. 6. Poor Human Resource Management (HRM): People and staff management becomes extremely difficult in the situation of COR, where multiple personnel from various cultural and zone variances are coordinated, particularly in a security-related system. 1. 7. IT Transitions Difficulties: One reason for this failure is the complication of information system outsourcing transitions. In the case of COR, where several clients and vendors are involved, the transaction will undoubtedly grow more complicated. 1. 8. Poor Contract Management: Contract management, also known as the hard side, is one of the biggest issues and is discussed in more detail in the literature. For pre-contract, the supplier’s standard type contracts should not be utilized, even as a starting point, for complex outsourcing arrangements involving large sums of money, because they are always structured in favor of the outsourced vendor. 1. 9. Different Cultural Issues: Because of the increased level of uncertainty associated with the crossing of organisational, geographical and cultural boundaries, complex IS outsourcing strategies even become more complex. 1. 10. General Management Complexities: With several development centres in different time zones, geographies, and cultures, managing the complex outsourcing relationships is intrinsically tough. Furthermore, management becomes much more difficult when IT is mostly outsourced to one or more IT services providers i.e. COR. 1. 11. Modern Technological Challenges: Technology Challenges, Strategic Decision Challenges, Vendor Management Challenges, and Vendor Selection Challenges are the four key kinds of challenges often faced where more than one stackholders are involved. ## Analysis of these challenges based on research methodology used To analyze and well furnish the outcomes of the SLR, we further categorized the results of the identified challenges on 8 different commonly used research methodologies. The methodologies are Experimentation/Testing, Case Study/Case Studies, Empirical Survey, Interviews, Ordinary Literature Review (OLR), Systematic Literature Review (SLR), Others (i.e. Main Path Analysis and Content Analysis, Mixed Method (.i.e more than one method), as shown in. During the analysis, we concluded that Experimentation/Testing methodology has been used more in the literature in this domain. Because in these studies, models or methods were used to check or test the challenges quantitavely, for example, measuring the cost or taking decisions. Similarly, the second mostly used methodologies are Case Study/Case Studies or Empirical Survey. Once again the reason behind this analysis was to collect empirically sound data in the real context of a case. The Mixed Methodology shows progressive results in this particular area and that’s why it is reported at number three position in the list etc. ## Analysis of these challenges based on publication period Although, we did not put any date boundaries of data during the accomplishment of an SLR, but during the analysis, we pointed out a major contribution shift from a theoretical to practical implementation by the organizations in COR context. Therefore, based on this knowledge, we divided the publication era into two periods such as papers published before 2014 and papers published in 2014 and onwards, as shown in. We can further conclude from the analysis of that COR is one of the most prominent research areas to date. Because sufficient number of studies are published in the second period (i.e., in 2014 and onwards). ## Analysis of these challenges based on publication venue gives an analysis based on publication channel. We divided the publication channels into three categories, such as papers published in Conference/Proceeding, Journal Papers, Technical Reports/Research Working Papers. From the analysis of, it is clearly visible that the majority of papers were published in Journals. This analysis once again is an evidence that COR is always a strong research area in the literature. ## Practices for identified challenges However, it was quite difficult to accurately account for the concerned practices for the stated challenges due to the cascading nature of data. Nonetheless, we organized them to the best of our efforts and identified 67 practices for the the aforementioned challenges. Moreover, due to the time and space constraints, we present these practices for the concerned challenges in tabular form, as depicted in Tables to below. # Limitations and threats to validity We did our best to conduct a systematic literature review as a study approach, including ensuring the pertinency of appropriate string selection and a sufficient sample size, but it is still conceivable that we have overlooked some crucial information. In order to reduce the danger of construct validity in terms of the usage of digital libraries, we narrowed our focus to pertinent and famous computing libraries, but again it was confined to the very few libraries. We have listed the publications in for interested visitors to minimize the danger of the study’s internal validity. To prevent such dangers, each component of our SLR was confirmed via a rigorous approach and intermittent review processes by the participating researchers. The identification method has been used in the literature for similar investigations multiple times. It should be noted that the majority of the studies we gathered were from the COR database. However, only a few of them, or portions of them, do not properly identify the COR data, giving the context a hazy picture. Furthermore, only papers written in English were chosen. # Conclusion The benefits of a complex IT outsourcing arrangement include enhanced cost certainty and lower expenses, increased scalability, and flexibility on demand. Complex outsourcing, on the other hand, demands interactions between many stakeholders from various areas and cultures, making it substantially more difficult to manage than regular outsourcing. Complex outsourcing is also more challenging than other forms of outsourcing since it needs a variety of control and coordination mechanisms for project management, which increases the chance of project failure accordingly. In this regard, a more systematic research is needed to uncover the important difficulties and techniques in this area in order to overcome project failure in complex outsourcing relationships. As a result, this study uses systematic literature review as a research method and serves as a pioneering effort to achieve the aforementioned goals. The authors highlighted 11 key issues after receiving the SLR data, with 67 practises in hand from a total of 85 publications. The authors want to build a comprehensive framework based on their findings in the future by using some strong approaches like as AHP and fuzzy logic, etc. # Supporting information We are grateful to other members of the Software Engineering Research Group (SERG) for their valuable feedback and support during the research process. 10.1371/journal.pone.0262710.r001 Decision Letter 0 Haldorai Anandakumar Academic Editor 2022 Anandakumar Haldorai This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 15 Nov 2021 PONE-D-21-28732Challenges And Practices Identification In Complex Outsourcing Relationship: A Systematic Literature ReviewPLOS ONE Dear Dr. Khan, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Dec 30 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Anandakumar Haldorai, PhD Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please amend your Methods section to include the date range for the search. Please also state the latest date on which the search was performed. 3\. We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match. When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. 4\. Please note that in order to use the direct billing option the corresponding author must be affiliated with the chosen institute. Please either amend your manuscript to change the affiliation or corresponding author, or email us at <[email protected]> with a request to remove this option. 5\. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at <http://journals.plos.org/plosone/s/latex>. 6\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Additional Editor Comments: Please carefully address the issues raised in the comments and, up front in your revised paper. Your revised paper will be sent to the same reviewers, as well as possibly new reviewers, for evaluation. Make sure the Abstract briefly describes the paper as it is used in abstracting and citation services. Keep the Abstract between 200 words. Do not use any references in the Abstract. Spell out each acronym the first time used in the body of the paper. Spell out acronyms in the Abstract only if used there. Include a list of six to ten key words after the Abstract. You may ignore any suggestion of including self-references by reviewers if not applicable. Include a paragraph at the end of the Introduction describing the organization of the paper. Make sure that the Conclusion briefly summarizes the results of the paper it should not repeat phrases from the Introduction. Keep the Conclusion to about 300 words. Do not use any references or acronyms in the Conclusion. Make sure all figures and tables are referred to in the body of the paper. It is recommended to use a professional native English-speaking editor. Papers with less than excellent English will not be published even if technically perfect. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The structure of the paper is good, however, it needs the following modifications to be published in the journal. 1\. Abstract need to be revised to specify the demand for the review. 2\. In introduction part, authors should provide more priority about outsourcing relationship. 3\. The methodology of the proposed review is complex and readers find problems in understanding the procedure. 4\. There is a need for more research papers that present more detailed information about outsourcing relationships and its advantages. 5\. Finally, the conclusion part need a rigorous revision. Reviewer \#2: It’s a great work and great sense of methodology and great analyzing and I wish to see more relate study and more upgrades with the same concept of research in the future and keep it up and good work \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: Yes: Haitham Medhat Abdelaziz Elsayed Aboulilah \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0262710.r002 Author response to Decision Letter 0 21 Dec 2021 Rebuttal Letter Original Manuscript ID: PONE-D-21-28732 Original Article Title: “Challenges And Practices Identification In Complex Outsourcing Relationship: A Systematic Literature Review” To: PLOS ONE Re: Response to reviewers Dear Editor, Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments. We are uploading $1$ Our point-to-point response to the comments below (response to reviewer(s) and academic editor, respectively) $2$ A marked-up coy with track changes (Revised Manuscript with Track Changes) $3$ An unmarked version without track changes (Manuscript) Best regards, \<author name\> et al. Reviewer#1, Concern \# 1: Reviewer’s comment: Abstract need to be revised to specify the demand for the review Authors’ response: Many thanks for your kind suggestion. We have revised the abstract by adding such specific demand.\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ Reviewer#1, Concern \# 2: Reviewer’s comment: In introduction part, authors should provide more priority about outsourcing relationship. Authors’ response: We have added few more papers having specific focus on outsourcing relationships in the introduction section of the manuscript.\_\_\_\_ \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ Reviewer#1, Concern \# 3: Reviewer’s comment: The methodology of the proposed review is complex and readers find problems in understanding the procedure. Authors’ response: We have performed systematic literature review (SLR) as a research methodology. We have followed the SLR guidelines from \[25, 26\], respectively. SLR is widely used research methodology in software engineering and we have used this method in our previous work having Doi as follows: 10.1109/ACCESS.2021.3085707, 10.1016/[j.infsof.2010.08.003](http://j.infsof.2010.08.003), and 10.1109/ICGSE.2009.28 etc. \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ Reviewer#1, Concern \# 4: Reviewer’s comment: There is a need for more research papers that present more detailed information about outsourcing relationships and its advantages. Authors’ response: Our search is based on our pre-defined search string that was constructed during the planning phase of the SLR, i.e., SLR protocol. During the implementation phase of the SLR protocol, the search phase was completed in April, 2021. Now, according to this concern, we have added some more latest papers, through snowballing technique, relevant to outsourcing relationships. \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ Reviewer#1, Concern \# 5: Reviewer’s comment: Finally, the conclusion part need a rigorous revision. Authors’ response: We have revised the whole manuscript specifically the conclusion section. \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ NOTE: We have further addressed the Journal Requirements as follows: Concern \# 1: 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> Authors’ response: We have followed the PLOS ONE’S style and format in the revised manuscript. Concern \# 2: 2\. Please amend your Methods section to include the date range for the search. Please also state the latest date on which the search was performed. Authors’ response: We have amended the Methods section, accordingly. Please see the change tracked at page 8 under the Literature Resources section of Research Methodology. During the implementation phase of the SLR protocol, the search phase was completed in April, 2021. Concern \# 3: 3\. We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match. When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. Authors’ response: No fund has been issued to us by any organization/agency. It is solely accomplished by the authors. However, the publication charges are sponsored by the Qatar university. Concern \# 4: 4\. Please note that in order to use the direct billing option the corresponding author must be affiliated with the chosen institute. Please either amend your manuscript to change the affiliation or corresponding author, or email us at <[email protected]> with a request to remove this option. Authors’ response: ? Concern \# 5: 5\. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at <http://journals.plos.org/plosone/s/latex>. Authors’ response: We have updated the manuscript according to the PLOS LaTeX template, by following the given link. Concern \# 6: 6\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Authors’ response: We have reviewed the references list and revised it properly, and ensured its correctness and completeness. We have added the references of additional papers which we have included now, as according to reviewers’ suggestions. Furthermore, we have removed/retracted the references of excluded papers, because these references have been removed during merging and updating the table data for further improvements. Additional Editor Comments: Concerns \#: \- Please carefully address the issues raised in the comments and, up front in your revised paper. Your revised paper will be sent to the same reviewers, as well as possibly new reviewers, for evaluation. \- Make sure the Abstract briefly describes the paper as it is used in abstracting and citation services. Keep the Abstract between 200 words. Do not use any references in the Abstract. \- Spell out each acronym the first time used in the body of the paper. Spell out acronyms in the Abstract only if used there. \- Include a list of six to ten key words after the Abstract. \- You may ignore any suggestion of including self-references by reviewers if not applicable. \- Include a paragraph at the end of the Introduction describing the organization of the paper. \- Make sure that the Conclusion briefly summarizes the results of the paper it should not repeat phrases from the Introduction. Keep the Conclusion to about 300 words. \- Do not use any references or acronyms in the Conclusion. \- Make sure all figures and tables are referred to in the body of the paper. \- It is recommended to use a professional native English-speaking editor. Papers with less than excellent English will not be published even if technically perfect. Authors’ response: We have addressed all the mentioned comments/concerns, accordingly. NOTE: It is worth noting that we have revised or renamed few challenges, and similarly, merged together some of the previously identified practices for improvements. In this way, we finally collected 67 practices for 11 identified challenges. 10.1371/journal.pone.0262710.r003 Decision Letter 1 Haldorai Anandakumar Academic Editor 2022 Anandakumar Haldorai This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 3 Jan 2022 Challenges And Practices Identification In Complex Outsourcing Relationship: A Systematic Literature Review PONE-D-21-28732R1 Dear Dr. Khan, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Anandakumar Haldorai, PhD Academic Editor PLOS ONE Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The paper is acceptable in its current form. No need to modify. The author had addressed the comments Reviewer \#2: It’s nice research paper and really good analysis with a perfect methodology and based on good theoretical and looking forward to see more papers can doing with same content and research base as same as the current one cause really it’s open for opportunity to creat a new one, wish you all the best \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: Yes: Haitham Medhat Abdelaziz Elsayed Aboulilah 10.1371/journal.pone.0262710.r004 Acceptance letter Haldorai Anandakumar Academic Editor 2022 Anandakumar Haldorai This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 17 Jan 2022 PONE-D-21-28732R1 Challenges and practices identification in complex outsourcing relationships: A systematic literature review Dear Dr. Khan: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Anandakumar Haldorai Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction Boron neutron capture therapy (BNCT) requires the selective delivery of boron-10 (<sup>10</sup>B) to tumor cells. Following irradiation with neutrons, the nuclear capture and spontaneous fission reactions produce <sup>4</sup>He and <sup>7</sup>Li nuclei along with 2.4 MeV. These high-linear energy transfer (LET) particles travel less than ten micrometers from their sites of origin; therefore, they are only lethal to those cells that bind or internalize <sup>10</sup>B in sufficient concentrations. The effectiveness of BNCT is dependent upon the amount of <sup>10</sup>B delivered per cell. BNCT has been used in the experimental treatment of a number of different tumors, such as Glioblastoma, skin melanomas, head and neck cancer *, mesothelioma, and diffuse liver metastases and could provide a useful treatment option for tumors that are unaffected by conventional therapies or that are difficult to remove surgically. When metastases spread through an entire organ, the use of a selective BNCT agent might allow the selective destruction of each of the individual cells of the tumor nodules without requiring their selective irradiation (13). Selective <sup>10</sup>B incorporation into cancer cells requires boron carrier molecules that exhibit a particular affinity toward the targeted cells. A wide range of boron carriers has been designed, synthesized, and evaluated during the past several decades, including liposomes. Liposomes are slightly more advantageous for the selective delivery of <sup>10</sup>B to murine EMT-6 tumors due to increased tumor cell growth rates and the incorporation of liposome components into the cellular membrane. Additionally, the often tortuous and leaky tumor vasculature allows the accumulation of boron within the tumor interstitium. Liposome delivery to the tumors depends upon the blood supply to the tumors, with a higher blood supply resulting in a higher <sup>10</sup>B accumulation and lower blood supply culminating in lower boron accumulation in the tumors. Macrophage/monocytes are professional phagocytes and phagocytize cell debris, foreign particles such as bacteria, fungi and parasites or liposomes and other particles of similar size. The phagocytosis of liposomes results in boron accumulation in macrophages/monocytes. Neutron irradiation may result in BNCT reaction and the energy released could modify macrophages to either tumor promoting or tumor inhibiting phenotype. Macrophage polarization is the hallmark of the innate immune response against cancer and pathogenic invasion. The extent of macrophage modification is dependent on the microenvironmental factors and decides the fate of macrophage polarization. The polarization of macrophages is multifaceted due to the plasticity of macrophages which can accommodate signals from pathogens, injured tissues, and the basal tissue microenvironment. The polarization of macrophages is under the control of pathways which regulate the survival of the cell by either prolonging or reducing macrophage development and viability. The tissue microenvironment, microbial products, and cytokines decide the fate of macrophage polarization. The activation of macrophages influences other branches of the immune system due to these cells being the essential modulators and effectors of the immune response. A hypothesis put forward that subsets of T helper cells can be distinguished based on the cytokines secreted after their activation. These subsets mediate distinct regulatory and effector functions. In 1960s Mackaness introduced the term macrophage activation (classical activation) about infection to describe the antigen-dependent but non-specific anti-microbial response of macrophages to BCG (bacillus Calmette- Guerin), and Listeria upon subsequent exposure. Later the enhanced microbicidal activity was linked to T helper type 1 responses, along with IFNγ release by antigen-activated immune cells \[\] and these microbicidal effects regulated by TH1 and IFNγ also account for cytotoxic and antitumoral effects (). Stein, Doyle, and colleagues postulated that IL-4 and IL-13 induces an alternative activation phenotype due to the discovery of the mannose receptor which selectively enhances the TH2 response in murine macrophages. The alternative activation is an entirely different phase than classical activation, but different from deactivation. In other words, the lack of a classical activation does not entail loss of activation. Macrophages in mice with TH1 and TH2 backgrounds differed in the propensity to react to the classic stimuli (IFNγ or lipopolysaccharide or both). The ability to respond to stimuli carves a vital difference in metabolic pathways in macrophages. The macrophage nomenclature of M1 and M2 is similar and based on the activation of the T helper type cells, i.e., the macrophages which activate TH1 cells belong to the M1 phenotype, while M2 macrophages activate TH2 cells. For example, M1 macrophages following LPS or IFN stimuli release toxic nitric oxide (NO), while M2 macrophages release polyamines (). The official report for the alternative activation of macrophages in vivo similar to the observation of Mackness for pathogens came from the observation of Allen, de Baetselier, Brombacher, and colleagues in parasite infection. The parasite elicits a strong IgE and TH2 response \[\]. Montavani and colleagues grouped the macrophage activating factors in two functionally polarized states to integrate the phenotypic similarities and differences. The macrophages were grouped based on their effects on select markers as M1 (IFNγ and LPS or TNF-α), M2a (IL-4), M2c {IL-10 and glucocorticoids (GCs). # Results ## Optimizing the boron delivery to the EMT-6 tumors for effective boron neutron capture therapy (BNCT) We evaluated the boron accumulation in EMT-6 tumors of different sizes and observed that boron accumulation was extremely low in larger tumors (approximately 400 mg) but was significantly higher in smaller tumors (approximately 100 mg). These results suggest smaller EMT-6 tumors are more amenable to boron delivery. We further tested the boron distribution by incorporating fluorescein dye in liposome in order to track the boron distribution in tumors of varying size. distribution. We observed the uniform distribution of fluorescence in small tumors while fluorescence in large tumors was spotty and did not cover the large tumor area.. We evaluated the necrosis present in EMT-6 tumors ranging from 100–400 mg in size. Various 10-μm cryosections of the EMT-6 tumors revealed that necrosis progressively increases with the size of the tumor. These regions represent oxygen deprivation areas and hence cell death by mostly necrosis. Blood vessels are vital for the growth and development of tumors and the delivery of nutrients and drugs. As we have shown above, boron distribution is affected by tumor size. We examined tumor vasculature by staining tumor sections with an anti-CD31 antibody labeled with Fluorescein Isothiocyanate (FITC). CD31 staining of the EMT-6 tumor sections revealed that smaller EMT-6 tumors have a higher amount of CD31 staining than the larger EMT-6 tumors. ## Immunomodulatory effect of BNCT in EMT-6 tumors After optimizing the delivery of boron compounds in tumor tissue, we then moved to irradiation studies. One set of EMT-6 tumor laden mice received neutron irradiation, and the second set did not receive radiation but only the liposomes for biodistribution. Time kinetics of of boron distribution in tumor and blood revealed that tumors had approximately 60 ppm of boron and blood had 10 ppm of boron. The optimal tumor/blood ratio and optimal boron concentration in tumors was observed 54 hrs post systemic delivery of boron-rich liposomes. We used 54 hrs as optimal time point for optimal boron dose in all our irradiation experiments. We used this time point for all the irradiation studies. In the blood, peripheral blood mononuclear cells (PBMCs) are most likely to assimilate liposomes and hence boron, than other components of the blood. Analysis of blood following irradiation revealed that irradiation did not cause damage to the PBMCs despite these cells carrying boron which is why the total population of PBMCs remain unchanged. IL-12 is a known anti-tumor cytokine and also protect against infection caused by invading pathogens. In tumor microenvironment the intruding macrophages are modified to suppress IL-12 expression and enhance IL-10 expression. IL-10 suppresses anti-tumor immune response by modifying macrophages to release tumor promoting factors. Analysis of expression of IL-12 and IL-10 suggested that the PBMCs switched their phenotype to anti-tumor phenotype with increased levels of interleukin-12 (IL-12) and decreased IL-10 levels. We observed that irradiation of EMT-6 tumor- bearing mice significantly inhibited tumor growth in these mice. The point of significance in the tumor growth inhibition started at around 15 days (P = 0.04) and kept on increasing till the time study was ended at 22 days (P = 0.001) due to large tumor sizes and discomfort and restricted movement of neutron only irradiated mice. Time kinetics of boron distribution following direct tumor injection of boron- rich lipososmes suggested that the boron concetration did not decrease substantially. Direct injection of boron-rich liposomes into tumor did not have any effect on PBMCs cytokine profile following irradiation which remained as low IL-12 and high IL-10 levels. Caspases play important roles in the induction of cells. The caspase staining with FLICA (Fluorescent inhibitor of Caspase) to assess apoptosis following irradiation revealed that the tumors injected directly with boron-rich liposomes has significantly elevated caspase activity than the systemic delivery of boron-rich liposomes to the tumors. Irradiation of tumor-bearing mice following direct injection of boron-rich liposomes did not show increased inhibition of EMT-6 tumor growth despite the presence of boron at more than double the concentration compared to intravenous delivery. Infact the inhibition of tumor growth did not reach the point of significance. # Discussion Selective delivery of boron forms the basis of effective BNCT for the treatment of solid tumors. The capture of low-energy neutrons by <sup>10</sup>B results in fission reactions leading to the generation of high energy He2+ and Li3+ particles which traverse the path of 5–10 μm in diameter from initial point and hence are lethal to the cells carrying boron in sufficient concentrations. The results of our study indicate that boron distribution in the tumors is critical for effective BNCT therapy. We have shown that the size of the tumor is inversely proportional to the accumulation of boron compounds. We have further demonstrated that the reasons for low boron distribution include the high levels of necrosis and the fewer number of blood vessels observed in larger EMT-6 tumors. In other words, vascularization of EMT-6 tumors is inversely proportional to the necrosis in these tumors. Further vascularization of EMT-6 tumors directly proportional to the distribution of boron rich liposomes in these tumors and thus affect the efficiency of BNCT in treating the mammary tumors. Our results further reveal that boron-rich liposomes assimilated by PBMCs activated and modified these cells to an anti-tumor phenotype and aided in the inhibition of tumor growth. We observed that the tumor growth following systemic or direct tumor delivery of boron-rich liposomes revealed that the initial effects of irradiation in terms of caspase activity was higher in tumors directly injected with boron-rich liposomes than tumors receiving boron-rich liposomes systemically, however the long term effect on tumor growth were significantly higher in tumors receiving boron-rich liposomes systemically. These results suggest that vascular and liposomal delivery of boron compounds might have immunomodulatory effects and hence overtakes the overall impact on tumor growth. Eventhough the initial caspase activity is higher in tumor sections and cells obtained from the mice receiving direct tumoral injecrtion of boron rich liposomes, but overall tumor inhibition is lower than in the mice receiving systemic delivery. The plausible reason for the differential effect could be that intra-tumoral boron delivery produces acute effects while immune system modulation could produce long lasting effect. Similar, plausible immunomodulatory effects were seen in a study where rats implanted with colon tumors in both right and left legs followed by exposure to neutron beam in one leg resulted in significant reduction of tumors non-exposed leg. In this study boronophenylalanine (BPA) as a source of boron for BNCT reaction. Furthermore, cancer cells damaged by irradiation can also undergo autophagy to chew-up the damaged part to prevent cell death. In this study, we are showing initial results where boron-rich liposomes can modify the PBMCs to antitumor phenotype. Radiation therapy is the conventional treatment modality along with chemotherapy following surgical removal of tumor tissue to prevent relapse perpetrated by residual tumor tissue or cells. These therapies, however also result in severe side effects in patients thereby affecting their quality of life. There are two forms of radiation therapy, one is whole body irradiation therapy or radiation of exposed part making it somewhat taregeted, however the radiation does not differentiate between cancerous or non-cancerous tissue or cells, and the other is a specific therapy where radiation only affects the cells carrying the target molecule. In our study we observed little to no boron distribution in tissues such as heart, brain, kidney and lungs thereby preventing off target effects. We did not see the irradiation effects on noncancerous tissue probalby due to low boron distribution and hence proving our initial point of BNCT being a more targeted therapy than other forms of irradiation. The immunomodulatory effects occurring due to systemic delivery of boron molecules needs to be studied comprehensively. We are sharing our initial observation here for the scientific community so that it can ponder upon the therapeutic effects of BNCT before writing it off as another means of radiation therapy. BNCT has the potential to become an essential form of treatment for various cancers. However, most of the compounds developed for use in BNCT generally do not demonstrate significant inhibition of tumor growth, which could be the result of poor distribution of boron in tumors and also the mode of delivery. We suggest that for BNCT to be an effective therapy for inhibition of tumor growth or prevention of relapse in patients, the delivery of boron-rich compounds should be systemic and by nano drug delivery systems. The nano-delivery systems will deliver boron to the PBMCs which might activate these cells and help in the elimination of tumors. # Materials and methods ## Animals All work was performed by the general protocols of animal care, and experimental design committee and the experiments outlined in this study are on file and have been approved by the University of Missouri Committee on the Humane Care of Laboratory Animals (CHCLA). All animal caretakers and laboratory personnel have appropriate approvals based on specific American Association for Laboratory Animal Care (AALAC)-approved training programs. The facilities are regularly inspected by University Committee and by unannounced visits directed by the Federal Government. We obtained BALB/c mice from Harlan Sprague (Indianapolis, IN). Minimum of 10 mice were used in all the mice experiments listed in this study. All the animal experiments were done according the ACUC protocol \# 7993. ## Reagents, and cell lines We purchased Distearoyl-sn-glycero-3-phosphocholine (DSPC) from Avanti Polar Lipids (Alabaster, AL), Cholesterol from Sigma-Aldrich (St. Louis, MO) and syringe filters from Corning (Lowell, MA). The Sephadex G-25 gel used in the study was purchased from Sigma-Aldrich (St. Louis, MO). We bought the anti-CD31 FITC antibody from Thermo Fisher Scientific (Rockford, IL), the tissue-freezing medium from the Molecular Cytology Core Facility (University of Missouri, Columbia, MO), the EtBRIII (necrosis study) from Enzo Life Sciences (Farmingdale, NY). Lastly, the EMT-6 cells from ATCC (Manassas, VA). ## MAC/TAC liposome preparation First, 0.1675 g of 1,2-distearoyl-sn-glycero-3-phosphocholine (DSPC) was added to a 16 x 100 mm glass tube. Next, 0.0820 g of cholesterol (Sigma-Aldrich, St. Louis, MO) and 0.0505 g K\[nido-7- CH<sub>3</sub>(CH<sub>2</sub>)<sub>15</sub>−7,8- C<sub>2</sub>B<sub>9</sub>H<sub>ll</sub>\] (MAC) was added to achieve 0.3 g total of a 1:1:0.6 molar ratio of the respective components. The mixture was dissolved in chloroform and methanol before being vortexed, dried by a nitrogen stream until a film formed and then further dried for 12 hrs in a vacuum. Next, 6 ml of sterile, aqueous 250 mM (1000 mOsM) Na<sub>3</sub> \[1-(2'- B<sub>10</sub>H<sub>9</sub>)-2-NH<sub>3</sub>B<sub>10</sub>H<sub>8</sub>\] (TAC), which had been adjusted to pH seven via the addition of 1 M HCl, was added to the vacuum-dried lipid mixture. The mixture was vortexed then sonicated for 30 minutes at 65°C using a Sonics & Materials Vibracell and a 1/8” standard tapered microtip probe. Following sonication, the liposome mixture was purified from unincorporated materials via size-exclusion chromatography using Sephadex G-25 (medium) filtration with phosphate buffered lactose (9% w/v lactose, five mM phosphate, pH 7.4) as the eluting buffer solution. The collected and combined liposome suspension was then filtered and sterilized by passage through two 0.2 μm Corning syringe filters directly into an autoclave-sterilized, sealed serum bottle. Particle sizing was achieved utilizing a Zetatrac particle analyzer from Microtrac, Inc. (Montgomeryville, PA). ## Cell culture, tumor induction, and experimental design We obtained EMT6 cells from American Type Culture Collection (ATCC) and cultured in DMEM medium supplemented with 10% FBS as recommended by ATCC. TrypLE buffer (Life Technologies) was used to dissociate cells in log phase, DMEM + 10% FBS was added to stop TrypLE reaction. The cells were pelleted down using accuspin 3R centrifuge (Fisher Scientific) at 323 × g* for 8 min at room temperature, followed by resuspension of cell pellet in 1X phosphate buffered saline (PBS). Cells number was counted using an Automatic Cell Counter (Life Technologies). For tumor induction, one million EMT6 cells (1 × 10<sup>6</sup> cells/mouse) per mouse were inoculated into female BALB/c (n = 5/set) mice’s right flank. The mice were typically weighed around 20 ± 1 g for all the studies listed in this manuscript as described previously. ## Quantitative analysis of boron distribution in mice tissues using ICP-OES BALB/c mice (n = 5/time point) were implanted with EMT-6 tumors in the right flank of these mice. When the tumor size reached the desired size, the mice were injected with TAC/MAC liposomes either intravenous or directly into tumors for indicated time points. The injected dose of liposomes contained the boron at concentration of 350 μg or 17.5 μg 10B/gram of body mass. The mice were then euthanized by first anaesthetizing with cocktail of 10 mg/kg xylazine and 80 mg/kg ketamine followed by cervical dislocation. We collected tissue samples (blood and tumor) in clean, and dry tubes following intravenous or direct injections. Optima-grade nitric acid (1 ml) was added to the blood and tumor samples and allowed to sit followed by shaking. Also, 100 μl of scandium (200μg/ml) was added to each pressure vessel to serve as a standard for all samples. We used CEM Mar microwave for tissue digestion followed by measurement of tube weight after sample dilution. The samples were transferred to appropriately labeled 50-mL conical tubes and analyzed by ICP. Plasma was used as the source of light to excite the samples. The light intensities of calibrants were used to calculate the sample boron concentration. The calibrants (7 ml each) and a blank were placed in the auto-sampler and analyzed. We used a blank between each sample analysis, and a spiked solution was used approximately every four samples to ensure proper instrument function. ## Time kinetics for biodistribution studies The biodistribution studies BALB/c mice (n = 5/set) were implanted with EMT-6 tumor cells. The mice were injected with boron-rich liposomes for different time points either systemically or direct tumor injections. The mice euthanized and tumors and blood were harvested and analyzed for boron conc by ICP-OES. Boron levels in tumor tissue following systemic delivery of boron-rich liposomes peaked in 54 h. At 54 h, boron concentration in the tumors was 67.8 μg <sup>10</sup>B per gram tumor, and the tumor/blood boron ratio was 10:1. As clearance of boron from blood proceeded more rapidly than loss from tumors, the tumor/blood ratio continued to increase till 54 hrs. We used 54 hrs as an optimum time point for irradiation. ## Study of tumor necrosis The BALB/c mice (n = 5/set) were injected with EMT-6 cells at a density of 1X10<sup>6</sup> per mouse and allowed to grow to tumors of different sizes (80 mg-400 mg). Tumors harvested following euthanization as described above. The tumors were then incubated overnight in 4% paraformaldehyde and then washed with 1X PBS. The tumors were incubated in a 30% sucrose solution at 4°C overnight before being washed in a solution of 30% sucrose in distilled water. The tumors were then frozen in tissue-freezing medium (TFM) and sectioned in 10-μm-thick slices. The sections were then stained with EtBRIII and observed using a confocal microscope at a wavelength of 600 nm, and images taken of the sections. ## CD31 staining of the tumor section BALB/c mice (n = 5/set) were implanted with EMT-6 tumor cells in the right flank. Tumors were allowed to grow to indicated sizes and sectioned as described above and previously. Briefly, the tumor sections were stained with anti-mouse CD31-FITC overnight at 4°C. The sections were then washed with 1X PBS and observed under a confocal microscope at an excitation wavelength of 488 nm, and images taken of the sections. ## Neutron irradiation BALB/c mice (n = 5/set) were implanted with EMT-6 cells in the right flank. When tumors were 80- to the 150-mm<sup>3</sup> in volume, <sup>10</sup>B-enriched boronated liposome suspensions were administered to the mice either via tail- vein or direct tumor injections. The mice were administered intraperitoneally (i.p.) with cocktails of 10mg/kg xylazine and 80mg/ml of ketamine, and Cu/Au flux wires were implanted on the left and right thorax of each mouse before irradiation. <sup>6</sup>LiCO<sub>3</sub> was used to shield the head, thorax, and cranial abdomen of the mice during irradiation to avoid possible complications which also helped in improving the specificity of the treatment. Once the mice were placed in positioning gallantry, the gallantry was introduced in the irradiation chamber and irradiated for 30–45 mins. A camera was set to monitor the discomfort in the mice during irradiation. Mice were taken out of the irradiation chamber and allowed to recover from anesthesia followed by a collection of the Cu/Au wires from the irradiated mice. The dose of irradiation was calculated as described ealeir ## Therapeutic effect studies The effect of BNCT was determined based on tumor volume changes during the course of the study. The control group was not administered with the boron while receiving the same amount of irradiation as the boron group. Calipers were used to measure the tumor volumes during the course of the study. Mice were euthanized when tumor volume reached 2000 mm<sup>3</sup>, or the mice showed visible discomfort in moving and accessing the food and water . For effective BNCT mediated effects on tumors the total thermal (0–0.414 eV) neutron fluence must be at least 1 × 10<sup>12</sup> neutrons/cm<sup>2</sup>. A 30–45 min exposure to the University of Missouri Research Reactor neutron beam provides a total thermal neutron fluence of 1.6–2,4 × 10<sup>12**</sup> which was sufficient to produce desired therapeutic effects on EMT-6 tumors. Boron dose and irradiation dose are calculated in. ## PBMCs count The BALB/c mice (n = 5) injected with TAC/MAC liposomes for 48 hrs were either irradiated or left non-irradiated. Followed by a collection of blood through the heart puncture in a heparinized tube to prevent blood clotting. We isolated PBMCs by gradient centrifugation on Histopaque (Sigma, St Louis, MO) solution. The automatic cell counter was used to determine the PMBCs in blood collected from both irradiated and non-irradiated mice. ## Quantitation of cytokines in PBMCs following irradiation Total RNA was prepared from PBMCs isolated from above mice using the RNeasy mini kit (Qiagen). cDNA was made using the High Capacity cDNA Kit (Applied Biosystems), and PCR amplification of cDNA was performed using the Taqman probe- based gene expression assay (Applied Biosystems) as previously described. The probes for IL-12 (Mm01288989_m1) and IL-10 (Mm00439614_m1) were used to determine the PBMCs polarization status. ## Caspase activity in tumors following irradiation EMT-6 tumors were grown in four sets of BALB/c mice (n = 5/set) which either received direct tumor injection or tail vein injection for otpimum time point of 54 hrs when tumors reached the size of 100-150mg. Boron distribution in tumor and blood was done by ICP-OES in two sets of mice (direct tumor or tail vein injection). Other two sets (direct tumor tail vein injection) were irradiated as described previously. The tumors were then incubated overnight in 4% paraformaldehyde and then washed with 1X PBS, followed by incubation in a 30% sucrose solution at 4°C overnight before being washed in a solution of 30% sucrose in distilled water. The tumors were then frozen in tissue-freezing medium (TFM) and sectioned into 10-μm-thick slices. The sections were then stained with EtBRIII and observed using a confocal microscope at a wavelength of 600 nm, and images taken of the sections. We thank the staff of the Animal Facility at the Dalton Cardiovascular Research Center for allowing us to house our mice, which enabled us to run our experiments more efficiently. [^1]: The authors have declared that no competing interests exist.
# Introduction The hippocampus is a target of corticosterone (CORT) modulatory actions, with high levels of glucocorticoid receptor (GR) and mineralocorticoid receptor (MR) expression. Traditionally, CORT had been thought to be synthesized exclusively in the adrenal cortex, reaching the brain via blood circulation. De novo synthesis of CORT from PROG in the brain has been doubted partly because brain CORT disappears after adrenalectomy (ADX) in rats. However, recent evidence shows that the hippocampus expresses steroidogenic enzymes required for corticosteroid synthesis, including cytochromes P450scc, P450(11β1), P450(11β2), and 3β-hydroxysteroid dehydrogenase (3β-HSD). Some of these enzymes are also necessary for sex-steroid synthesis. Yet, the complete corticosteroid synthesis of ‘pregnenolone (PREG) → progesterone (PROG)→ deoxycorticosterone (DOC)→CORT or aldosterone (ALDO)’ in the hippocampus has not been proven. Although previous studies have shown parts of the corticosteroid synthesis pathway in brain, including, PREG→PROG and DOC→CORT and DOC→ALDO, the conversion of PROG→DOC to be demonstrated. Cytochrome P450(c21) (DOC synthase), a key enzyme catalyzing the conversion of PROG to DOC, has not been detected in the hippocampus, although mRNA expression has been demonstrated in other brain regions including the hypothalamus, cortex, cerebellum and striatum. In the current study, we demonstrated expression and activity of P450(c21) and the complete metabolism of PROG→DOC→CORT in rat hippocampus. In addition to P450(c21), expression of cytochrome P450(2D4), another candidate for DOC synthase, was also observed. Further, the net hippocampus-derived CORT content was determined in ADX rats by mass-spectrometric analysis. We further investigated the effect of low dose CORT on spinogenesis of hippocampal neurons in order to show a possible role of hippocampus-derived CORT. # Results ## Molecular biological analysis of corticosteroid synthesizing enzymes Cellular localization and expression of enzymes responsible for CORT synthesis including P450(c21), P450(2D4), P450(11β1), P450(11β2), and 11β-HSD (types 1 and 2) were examined. Typical RT-PCR patterns of mRNA transcripts are shown in. Because the expression level of P450(c21) is extremely low in the hippocampus, we performed careful primer design using minimization of Gibbs free energy (ΔG) upon hybridization of primers with target sequences. We also choose primers with ΔG for 5 bases in 3′-side of the primer to be larger than average ΔG for improved specificity, because the specificity of primer-target recognition is mainly governed by 3′-side sequence of the primer. In the current study, ΔG for 5 bases in 3′-side of the P450(c21) primer was –6 kcal/mol, which was larger than the average ΔG for the P450(c21) primer ( = \<1?show=\[to\]?\>–8 kcal/mol). As a result, we succeeded the detection of P450(c21) mRNA. Relative number of transcripts, expressed in the hippocampus of adult male rats, was in the order of 1/20,000 of that in the adrenal gland for P450(c21), almost same level of that in the liver for P450(2D4), in the order of 1/10,000 of that in the adrenal gland for P450(11β1), in the order of 1/5,000 of that in the adrenal gland for P450(11β2), approximately 1/50 of that in the liver for 11β-HSD (type 1), and approximately 1/500 of that in the kidney for 11β-HSD (type 2). The mRNA expression levels of P450(c21), P450(11β1), P450(11β2) and 11β-HSD (type 2) in the hippocampus were lower than those in the hypothalamus. The expression level of P450(2D4) in the hippocampus was almost equal of that in the hypothalamus. The expression level of 11β-HSD (type 2) was higher than that in the hypothalamus. The cellular localization of P450(2D4), P450(11β1) and P450(c21) was identified using in situ hybridization and immunohistochemical staining. Significant expressions of both P450(2D4) and P450(11β1) were observed in pyramidal neurons (CA1, CA3) and granule neurons (DG) and these enzymes were weakly expressed in glial cells. The weak expression of P450(c21) was observed in pyramidal and granule neurons. Because the expression level of mRNA for P450(c21) was very low, the Tyramide signal amplification system was used to obtain sufficient sensitivity. ## Cellular and Subcellular localization of enzymes for corticosteroid synthesis Immunoelectron microscopic analysis using postembedding immunogold was performed in order to determine the localization of enzymes for corticosteroid synthesis (P450(c21), P450(2D4), P450(11β1) and 3β-HSD) in the hippocampus. This method is particularly useful to detect enzymes with extremely low expression level such as P450(c21). All enzymes were mainly localized in principal neurons including pyramidal neurons of CA1 and CA3 regions as well as granule neurons of the DG. Weak expression was observed in some glial cells. P450(c21) and P450(2D4) were localized not only in the endoplasmic reticulum but also in both the axon terminals and dendritic spines of principal neurons. P450(11β1) was localized in both the mitochondria and synapses of principal neurons. We also observed 3β-HSD in the synapses in addition to the endoplasmic reticulum. The subcellular localization of these enzymes was confirmed by western blot with purified fractions of postsynaptic density, endoplasmic reticulum and mitochondria. ## Mass-spectrometric analysis of corticosteroids in the hippocampus The concentration of CORT and DOC was determined in steroid extracts from adult male rat hippocampus using a chromatogram analysis of the fragmented ions. After selection of mother ions, fragmentation and detection were performed via MS/MS procedures. Chromatographic profiles for the fragmented ions of CORT (m/z = 121) showed a clear peak with the retention time of 3.19 min which was the same as that of the fragmented ion of the standard CORT, indicating a good specificity of the analysis. In the chromatographic profiles of the fragmented ion of DOC (m/z = 97), a single peak was observed at 3.73 min. In order to determine the net corticosteroids synthesis in the hippocampus, we used ADX rats to eliminate adrenal-derived steroids (CORT and DOC) via the blood circulation. Results are summarized in. Two weeks after ADX, the concentrations of CORT in the hippocampus and plasma were 2.4 ng/g wet weight ( = 6.9 nM) and 0.8 ng/ml (2.3 nM), respectively. The DOC concentration was 1.9 ng/g wet weight (5.9 nM) in the hippocampus, and 0.5 ng/ml (1.4 nM) in plasma, respectively. Even at 4 weeks after the ADX, the level of CORT did not decrease further from the level at two weeks within experimental error (±2.1 nM, SEM). These results imply that the circulation-derived CORT in the hippocampus was cleaned up already at two weeks after ADX, therefore 6.9 nM is solely hippocampus- synthesized CORT. ## Validation of mass-spectrometric analysis To confirm the assay accuracy, the hippocampal homogenate spiked with known amounts of the steroids was prepared and its concentration of steroid was determined. Satisfactory accuracy was obtained, supporting the accuracy of determined hippocampal steroid content in. The limits of quantifications (LOQs) were defined in as the lowest value with an acceptable accuracy (91.6–107.1 %) and precision (i.e. RSD\<11 %). The results of intra- and inter-assay were shown in. The RSD for intra- and inter-assay was less than 10.2 % and 10.9 %, respectively. These results indicate that the present method is highly reproducible and accurate. In addition, we observed approximately 0.68 nM of 11-deoxycortisol (roughly 1/10 of DOC) in the hippocampus. Because P450(17α) is expressed in the hippocampus, DOC→11-deoxycortisol → cortisol pathway may also work. ## Hippocampal corticosteroid metabolism Analysis of the pathway of corticosteroid metabolism is necessary, because mass- spectrometric determination shows only the contents of individual steroids. The metabolism of radioactive steroids in hippocampal slices was investigated using normal phase HPLC. Hippocampal slices were incubated with 5×10<sup>6</sup> cpm of <sup>3</sup>H-labeled steroid substrate for 5 h. Typical results of HPLC analysis are illustrated in. A significant production of DOC from <sup>3</sup>H-PROG was observed. The incubation buffer contains 100 nM of finasteride, a specific inhibitor of 5α-reductase, in order to reduce the conversion to 5α-reduced metabolites such as allopregnanolone. Because of the strong activity of 5α-reductase and 3α-HSD, some 5α-dihydro PROG and TH-PROG (allopregnanolone) were produced even in the presence of finasteride. In the absence of finasteride, the production of DOC and CORT were very weak to observe. When <sup>3</sup>H-DOC was used as a substrate, CORT from DOC was produced. The conversion of CORT to other steroids was very weak, indicating that CORT is stably present once it produced. ## Low dose effect of CORT on spinogenesis of hippocampal neurons To determine the potential physiological significance of nanomolar concentrations of hippocampal CORT, we investigated CORT effects on dendritic spine density and morphology in hippocampal ‘acute’ slices. It should be noted that CORT levels were depleted in control ‘acute’ slices, containing only 1.9 nM CORT following 2 h recovery incubation in ACSF. Treatment with exogenous 10 nM CORT for 1 h significantly increased the total spine density (1.18 spines/µm) compared with control slices (0.98 spines/µm, with no exogenous CORT).On the other hand, treatment with 100 nM CORT only slightly increased the total spine density. Morphological changes in spine head diameter induced by 1 h CORT administration were also assessed. Spines were classified into three categories based on head diameters: small-head spines (0.2–0.4 µm): middle-head spines (0.4–0.5 µm): and large-head spines (0.5–1.0 µm). Morphological categorization of spines into three subclasses enabled complex responses in spine subpopulations upon CORT application to be distinguished. Treatment with 10 nM CORT significantly increased the density of small-head spines from 0.58 (control) to 0.81 spines/µm, while the density of middle-head spines (approx. 0.28 spines/µm) and large-head spines (approx. 0.12 spines/µm) was not significantly altered. Morphological changes induced by 100 nM CORT were smaller than those induced by 10 nM CORT. The majority of spines (\>95%) had a distinct head and neck, therefore these analysis cover major populations of spines. # Discussion In the current study, we clarified the complete pathway for corticosteroid synthesis ‘PREG→PROG→DOC→CORT’ in hippocampal neurons. We demonstrated for the first time the expression, neuronal localization and activity of P450(c21) in the hippocampus. In addition, the localization of P450(2D4), another enzyme participating in DOC synthesis, was demonstrated. ## Previous investigations of brain corticosteroids ### PROG→DOC conversion and P450(c21) expression To date, direct evidence of DOC synthesis in the adult brain had not been reported. No expression of P450(c21) in the rat/mouse hippocampus has been reported, although a few studies have reported weak expression of P450(c21) mRNA in some regions of the rat brain such as hypothalamus, striatum, cerebellum. As an exception, weak expression of P450(c21) mRNA was observed in the human hippocampus, with levels 1/20,000 of that of the adrenal gland. Thus far, it has been believed that P450(c21) is not physiologically active in the hippocampus, and all CORT in the hippocampus is derived from the adrenal glands via the blood circulation. Although P450(2D4) has been known as the drug metabolizing P450, recently, it has been demonstrated that P450(2D4) can also convert PROG to DOC in hippocampus. Therefore, both P450(c21) and P450(2D4) could contribute to DOC production. We observed the evidence of conversion of PROG to DOC, confirming de novo synthesis of DOC. ### DOC→CORT conversion and P450(11β1) expression The expression of P450(11β1) in the hippocampus has been described in several studies and neuronal localization of P450(11β1) has been demonstrated using immunohistochemistry. We observed the localization of P450(11β1) in cell bodies of pyramidal and granule neurons at mRNA and protein level. Gomez-Sanchez and co-workers observe a weak CORT production from <sup>3</sup>H-DOC, roughly 0.4 pmol/mg/3 h in the hippocampus. However, they doubted de novo synthesis of CORT from PROG since CORT levels in whole brain were nigligible after adrenalectomy (ADX) of rats. According to their report, CORT in the whole brain of five ADX rats was below the detection limit, but four ADX rats had measurable CORT in the brain (0.6 ng/g = 1.8 nM). Therefore, their results do not completely eliminate a possibility of the presence of brain-synthesized CORT. It should be noted that the expression level of P450(11β1) in the hippocampus does not change upon ADX. In the hippocampus of ADX rats we observed the direct conversion of DOC to CORT as well as low, but physiologically significant level of CORT. The expression of P450(11β1) (both mRNA and protein) in neurons supports endogenous synthesis. Interestingly, the expression level of P450(11β1) is roughly 10-fold more in the cortex than in the hippocampus. ## Functional significance of hippocampus-synthesized corticosteroids The low nanomolar level of CORT, synthesized in the hippocampal neurons, may play a essential role in enhancement of synaptic plasticity, in contrast to the deleterious effects (e.g., neuronal cell death or shrinkage of dendrites) elicited by micromolar plasma CORT secreted from the adrenal gland under stressful conditions. In the current study, low dose CORT (10 nM) enhanced spinogenesis, particularly increasing the density of small-head spines. It has been previously demonstrated that nanomolar doses of CORT (10 nM) drives the Erk MAP kinase pathway, increasing both the expression and phosphorylation of MAP kinase. Taken together, hippocampus-synthesized CORT (approx. 7 nM) might induce spinogenesis via activation of the Erk MAP kinase pathway. Earlier studies have also shown a range of beneficial effects of nanomolar doses of CORT including strengthened signaling, neuroprotection as well as a potential role in development. In vivo, low plasma levels of CORT increased the amplitude of population spike, as did application of low concentration (5–50 nM) of CORT to CORT-depleted slices. Low doses of CORT may also have neuroprotective effects. For example, low dose CORT attenuates dexamethasone – induced apoptosis in dentate gyrus via MR signaling pathway. Nanomolar low dose of CORT may also play a role in neuronal development, with 5–10 nM CORT found to enhance differentiation of neural stem cells by inducing the expression of astrocyte marker GFAP (J Kanno, K Igarashi, K Tanemura, H Asano, K Nakashima., 14<sup>th</sup> Int Cong of Endocrinol, Kyoto, 2010, P8-3-1). Importantly, all the 7 nM of hippocampal CORT could participate in modulation of neuronal functions such as spinogenesis, because corticosteroid binding globulin (CBG) and serum albumin are not present in the hippocampus, therefore all the hippocampal synthesized CORT should be biologically active. The persistence of a significant concentration of corticosterone in the plasma (2 nM) even after ADX may be derived from local synthesis in other peripheral tissue such as adipocytes. Although we have a significant concentration of CORT in the plasma, the majority (95–98%) of plasma CORT may be biologically inactive in basal conditions, with the majority of plasma CORT (90%) bound to CBG and a large proportion of the remaining CORT bound to serum albumin. Although basal plasma CORT levels range between 100 and 300 nM, biologically active free CORT comprises only 2–5% of the total plasma CORT levels. ## Moderate CORT level is kept by weak CORT production in the hippocampus A trace amount of endogenous DOC and CORT had been very difficult to detect by <sup>3</sup>H-steroid metabolism analysis, because many different pathways exist other than PROG→DOC→CORT. Therefore, we employed mass-spectrometric assay for detection of a trace amount of these steroids in the hippocampus of ADX rats. We determined the concentrations of DOC and CORT to be 5.8 nM and 6.9 nM, respectively. These values are comparable to hippocampus-synthesized sex steroids such as estradiol (roughly 7 nM), testosterone (3 nM) and dihydrotestosterone (0.6 nM). Because the volume of hippocampus is very small (nearly 0.1 mL for one whole hippocampus), the calculated concentrations were relatively high in nM range for DOC and CORT. The absolute contents of DOC and CORT were only around 0.27 ng and 0.34 ng in one hippocampus with a weight of 0.14 g, respectively. Although steroid production capacity is strong in the adrenal gland, circulation levels of steroids are diluted due to approx. 20 mL of blood (200-fold of the hippocampal volume). Though the hippocampal expression levels of enzymes (such as P450(c21) and P450(11β1)) are roughly 1/20,000 and 1/10,000 of those in the adrenal, they need to fill only small hippocampal volume (1/200 of the blood volume). It should be noted that the hippocampus-synthesized CORT is stably present for even 5 h once produced, resulting in keeping the CORT level sufficient for modulation of synaptic plasticity. It may be due to roughly 1000-fold higher expression level of 11β-HSD (type 1) in the hippocampus than that of 11β-HSD (type 2) as judged from the cycle number of PCR. Because 11β-HSD (type 1) converts 11-dehydrocorticosterone (inactive form) to CORT, hippocampal CORT is prevented from inactivating once produced. On the other hand, in peripheral tissues such as kidney, the expression level of 11β-HSD (type2) is higher than that of 11β-HSD (type1). ## Role of MR and GR In ADX rats, the stress due to application of ether anesthetics before decapitation did not increase hippocampal CORT level to a high stress-level (e.g., 0.5–1 µM), therefore hippocampus-synthesized CORT (roughly 7 nM) may not play a role in the negative feedback by occupying many GR. Under basal conditions, all MR may be occupied by ALDO or CORT. Negative feedback would be induced by adrenal CORT. In response to stress, 0.5–1 µM adrenal CORT, a part of which penetrates into the hippocampus, could further occupy many unoccupied GR in the hippocampus and could replace ALDO on MR. ## Conclusion The hippocampus is equipped with not only sex-steroid synthesis systems , but also corticosteroid synthesis systems. Synaptic localization of important enzymes including P450(c21), P450(2D4) and P450(11β1) suggests a potential synaptocrine function of corticosteroids in hippocampal neurons. These enzymes are also localized in the endoplasmic reticulum (microsome) and mitochondria of hippocampal neurons, suggesting potential synaptocrine function of corticosteroid synthesis in hippocampal neurons. The current study opens new field of investigations concerning possible physiological function of hippocampus-synthesized nanomolar CORT, such as modulation of synaptic plasticity. # Materials and Methods ## Animals Young adult male Wistar rats (12-week old) were purchased from Saitama Experimental Animals Supply (Japan). All animals were maintained under a 12 h light/12 dark exposure and free access to food and water. Adrenalectomy (ADX) and sham operations were performed two-weeks before the experiments. The experimental procedure of this research was approved by the Committee for Animal Research of the University of Tokyo (The permission number is 19–10.). ## Chemicals Corticosterone (CORT), deoxycorticosterone (DOC) and progesterone (PROG) were purchased from Sigma (USA). \[−<sup>13</sup>C<sub>3</sub>\]PROG was from Hayashi Pure Chemical (Japan). CORT-d<sub>8</sub> and DOC-d<sub>8</sub> were from CDN Isotope Inc. (Canada). Finasteride was from Aska Pharma Medical (Japan). \[<sup>3</sup>H\] or \[<sup>14</sup>C\] labeled steroids were purchased from Perkin Elmer (USA) and their specific activities were 76.5 Ci/mmol (\[1,2,6,7-<sup>3</sup>H\]-CORT) and 102 Ci/mmol (\[1,2,6,7-<sup>3</sup>H\]-PROG). \[1,2-<sup>3</sup>H\]-DOC (50 Ci/mmol) was purchased from Muromachi Yakuhin (Japan). ## RT-PCR and Southern hybridization The procedures were the same as described elsewhere. Rats were deeply anesthetized with ethyl ether and decapitated. The brains were then removed. Total RNAs including mRNA were isolated from adult rat tissues such as the hippocampus, hypothalamus, adrenal gland, kidney and liver, using a total RNA Purification Kit (Nippongene, Japan). The purified RNAs were quantified on the basis of the absorbance at 260/280 nm, and treated with RNase-free DNase to eliminate the possibility of genomic DNA contamination. The purified RNAs were reverse-transcribed, using a M-MLV Reverse Transcriptase (Promega, USA). The oligonucleotides for PCR amplification were designed as illustrated in. The PCR protocols comprised application of a 30 sec denaturation period at 95°C, a 2 sec annealing period at individual temperature for each enzyme, and a 30 sec extension at 72°C, for individual number of cycles for each enzyme. For semiquantitative analysis, the RT-PCR products were separated on 2% agarose gels, stained with ethidium bromide, and analyzed with a fluorescence gel scanner (Atto, Japan) and Image J software, in comparison with standard curves obtained from PCR of diluted RT products (between 1/100 and 1/10,000 in dilution), from adrenal gland, kidney and liver. To confirm the expression, Southern hybridization was performed. The amplified RT-PCR products of steroidogenic enzymes were directly cloned into TA-cloning vector (Promega, USA), and sequenced. The resulting sequence was identical to the reported cDNA sequences of these enzymes. These cloned products were used as the template of DNA probes for Southern hybridization. After transfer of the RT- PCR products from agarose gels to nylon membrane (Hybond N+, Amersham, USA), Southern hybridization was performed with <sup>32</sup>P-labeled cDNA probes for these enzymes. The Southern hybridization signals were then measured using a BAS-1000 Image analyzer (Fuji film, Japan). ## In situ hybridization of hippocampal slices Rats were deeply anesthetized with ethyl ether and decapitated. The brains were removed and the hippocampus was dissected. Preparation of adult hippocampal slices fixed with 4% paraformaldehyde, was performed essentially as described in elsewhere. The postfixed brain tissue was frozen-sliced coronally at 15 µm thickness with a cryostat (Leica, Germany) and slices were immediately mounted on the slide glass at −18°C. Digoxigenin (DIG)-labeled sense and antisense cRNA probes were in vitro transcribed from PCR products of P450(11β1), P450(2D4) and P450(c21) by using T7 or Sp6 promoters inherent in pGEM-T-Easy vector. The hippocampal slices were treated with 10 µg/mL of Proteinase K (Wako, Japan) for 10–20 minutes, and then postfixed with fixative solution for 10 minutes. After acetylation with acetic anhydride and dehydration, the mRNA in hippocampal slices were hybridized with 0.5 µg/mL of DIG-labeled sense or antisense cRNA probes. In order to digest and wash out the excess cRNA probes, the slices were treated with RNase A (Wako, Japan) and stringent washes after hybridization. For P450(11β1) and P450(2D4), the slices were incubated with alkaline phosphatase- conjugated anti-DIG antibody (Roche Diagnosis, USA) (1/1000) for 30 min. After washing the hippocampal slices twice, target mRNAs were visualized by color development with 0.45 mg/mL of nitro blue tetrazolium chloride (NBT) and 0.175 mg/mL of 5-bromo-4-chloro-3-indoryl phosphate (BCIP) (Roche Diagnosis, USA) for 12 h. For P450(c21), the sections were incubated with horseradish peroxidase- conjugated anti-DIG antibody (Roche, USA) (1/1000) for 30 min. To amplify signals of in situ hybridization, the slides were incubated with dinitrophenyl labeled amplification reagent using the TSA Plus DNP (AP) System (Perkin Elmer, USA) and stained with NBT/BCIP for 12 h. ## Post-embedding immunogold method for electron microscopy Rat hippocampal slices were prepared in essentially the same manner as described elsewhere, except that slicing was performed at 4°C using a vibratome instead of frozen slicing. Brains from five animals were used and from each brain, a single representative coronal section including the dorsal hippocampus was processed for ultrathin sectioning. Freeze substitution and low-temperature embedding of the specimens was performed as described previously. Briefly, slices were plunged into liquid propane. The samples were immersed in uranyl acetate in anhydrous methanol (−90<sup>°</sup>C), then infiltrated with Lowicryl HM20 resin (Electron Microscopy Sciences, USA); polymerization was performed with ultraviolet light. Ultrathin sections were cut using a Reichert-Jung ultramicrotome. For immunolabeling, sections were incubated with primary antibody for 3β-HSD(type1) (1∶500), P450(c21) (1∶2000), P450(2D4) (1∶500) or P450(11β) (1∶1000) in the above diluent overnight, then incubated with secondary gold-tagged (10 nm) Fab fragment in Tris-buffered saline (TBS). Sections were counter-stained with 1 % uranyl acetate, and viewed on a JEOL 1200EX electron microscope (Japan). Images were captured using the CCD camera (Advanced Microscopy Techniques, USA). Controls omitting the primary antibody were performed and no immunogold labeling was observed. Controls of preadsorption incubating with purified antigens were also performed and no immunogold labeling was observed. ## Mass-spectrometric assay of steroids Detailed procedures are described elsewhere. ### Step 1) Purification of steroids from hippocampi with normal phase HPLC The preparation of hippocampal homogenates from slices and the extraction of steroids were performed as described elsewhere. <sup>3</sup>H-steroids (20,000 cpm each) were added to homogenates as internal standards. The steroid extracts were applied to a C<sub>18</sub> Amprep solid phase column (Amersham Biosciences, USA) to remove contaminating fats. Then the steroids were separated into fractions of CORT, DOC and PROG using a normal phase HPLC system (Jasco, Japan) with an elution solvent of hexane: isopropylalcohol: acetic acid = 98∶2∶1. A silica gel column (Cosmosil 5SL, Nacalai Tesque, Japan) was used. By monitoring <sup>3</sup>H-steroids, the recovery of CORT, DOC and PROG were 35±1%, 34±1%, and 42±1%, respectively, after extraction, C<sub>18</sub> column treatment and normal phase HPLC separation. Plasma was prepared by centrifugation from trunk blood collected from decapitated rats. ### Step 2) Determination of the concentration for CORT, DOC and PROG using LC-MS/MS At first, 100 pg of isotope labeled steroids (<sup>13</sup>C<sub>3</sub>-PROG, DOC-d<sub>8</sub> and CORT-d<sub>8</sub>) were added to steroid extracts prepared via Step 1). The LC-MS/MS system, which consisted of a reverse phase LC (Agilent 1100, Agilent Technologies, USA) coupled with an API 5000 triple- stage quadrupole mass spectrometer (Applied Biosystems, USA), was operated with electrospray ionization in the positive-ion mode. The LC chromatographic separation was performed on a Cadenza CD-C<sub>18</sub> column (3×150 mm, 3 µm, Imtakt Japan). Detailed conditions for column purification were described in of Supporting Information. In the multiple reaction monitoring mode, the instrument monitored the m/z transition, from 347 to 121 for CORT, from 331 to 97 for DOC, and from 315 to 109 for PROG, respectively. Here, m and z represent the mass and charge of a steroid derivative, respectively. To examine specificity of LC-MS/MS analysis, samples were spiked with steroid isotopes as internal standards. Though the m/z transitions were different between CORT (from m/z = 347 to 121) and CORT-d<sub>8</sub> (from m/z = 355 to 125), their retention times were the same, because the affinity of CORT for LC- column is same as that for CORT-d<sub>8</sub>. In case of other steroids, there is also no difference in the retention time between steroids and their isotopes, though the m/z transitions were different. Isotope-labeled steroid derivatives were also used for internal standards in order to measure recovery of steroids. The recovery of CORT, DOC and PROG were determined as 89±8%, 75±4% and 71±6%, respectively, after purification and MS/MS detection. Total recovery during all the steps was determined via <sup>3</sup>H- and isotope-labeled steroids in Step 1) and Step 2). The limits of quantification for steroids were measured with blank samples, prepared alongside hippocampal samples through the whole extraction, fractionation and purification procedures. The limits of quantification for CORT, DOC and PROG were 2 pg, 1 pg, and 2 pg per 0.1 g of hippocampal tissue or 1 mL of plasma, respectively. From the calibration curve using standard steroids dissolved in blank samples, the linearity was observed between 2 pg and 4000 pg for CORT, between 1 pg and 1000 pg for DOC, and between 2 pg and 4000 pg for PROG, respectively. Male Wistar rats were anesthetized with pentobarbital and placed in a stereotaxic Because the LC-MS/MS enabled to determine the exact level of CORT in the brain, we applied LC-MS/MS to observe the circadian rhythm of CORT level in the cerebrospinal fluid (CSF) in combination with the transverse microdialysis. Microdialysate was collected from the cisterna magna in freely moving male Wistar rats every hour. Narishige apparatus (SR-5, Narishige, Japan) inserting a transverse dialysis tube placed at the cisterna magna under the guidance of a stainless steel wire attached in a horizontal position to a stereotaxic holder. The rats were allowed roughly 1 week to recover from the surgery. Thereafter, the animals were moved to an acrylic test box with the transverse probe being perfused with a Ringer's solution (138 mM NaCl, 2.4 mM KCl, 1.2 mM CaCl<sub>2</sub>, \[pH 7.0\]) at 1 µL/min. Daily samples of CSF from the cisterna magna were automatically collected every hour in a small vial. The rats had free access to food placed on the floor of the test box and to water given from the lid of the box. ## Metabolism analysis of radioactive steroids using normal phase HPLC Procedures were essentially the same as previously described in elsewhere. The hippocampus from one rat was sliced into 400 µm thickness with a vibratome and incubated with 5×10<sup>6</sup> cpm of \[<sup>3</sup>H\]-steroids at 30°C for 5 h in 4 ml of physiological saline containing 1.2 mM Mg<sup>2+</sup>. The incubation medium was gassed with 95% O<sub>2</sub> and 5% CO<sub>2</sub> during the incubation in order to keep the activity of hippocampal neurons. After termination of the reaction, the slices were homogenized. A portion of the purified radioactive metabolites (total of 10<sup>6</sup> cpm) was analyzed using an HPLC system. Procedures used for the analysis of steroid metabolites from hippocampal homogenates with normal phase HPLC were the same as described in Step 1) of Mass-spectrometric assay. Fraction radioactivity was measured using a liquid scintillation spectrometer LS6500 (Beckman, USA). The rate of steroid production was normalized as products (cpm) /g wet weight /5 h. ## Imaging and analysis of dendritic spine density and morphology ### Slice preparation Twelve weeks male rats were deeply anesthetized with ethyl ether and decapitated. Immediately after decapitation, the brain was removed from the skull and placed in ice-cold oxygenated (95% O<sub>2</sub>, 5% CO<sub>2</sub>) artificial cerebrospinal fluid (ACSF) containing (in mM): 124 NaCl, 5 KCl, 1.25 NaH<sub>2</sub>PO<sub>4</sub>, 2 MgSO<sub>4</sub>, 2 CaCl<sub>2</sub>, 22 NaHCO<sub>3</sub>, and 10 D-glucose (all from Wako); pH was set at 7.4. Hippocampal slices, 400 µm thick, were prepared with a vibratome (Dosaka, Japan). These slices were ‘freshly prepared’ slices without ACSF incubation. Slices were then incubated in oxygenated ACSF for 2 h (slice recovery processes) in order to obtain conventional ‘acute’ slices. These acute slices were then incubated at room temperature with CORT for 1 h. Then, slices were prefixed with 4% paraformaldehyde at 4°C for 2–4 h. Spine imaging and analysis with confocal microscopy was performed as described previously. ### Current injection of neurons by Lucifer Yellow Neurons within slices were visualized by an injection of Lucifer Yellow under a Nikon E600FN microscope (Japan) equipped with a C2400–79H infrared camera (Hamamatsu Photonics, Japan) and with a 40× water immersion lens (Nikon). Dye was injected with a glass electrode filled with 5% Lucifer Yellow for 15 min, using Axopatch 200B (Axon Instruments, USA). Approximately five neurons within a 100–200 µm depth from the surface of a slice were injected (Duan H et al., 2002). After labeling, slices were fixed again with 4% paraformaldehyde at 4°C overnight. ### Confocal laser microscopy and morphological analysis Imaging was performed from sequential z-series scans with LSM5 PASCAL confocal microscope (Zeiss, Germany). For analysis of spines, three-dimensional images were constructed from approximately 40 sequential z-series sections of neurons scanned every 0.45 µm with a 63× water immersion lens, NA 1.2 (Zeiss). The excitation and emission wavelengths were 488 nm and 515 nm, respectively. The applied zoom factor (3.0) yielded 23 pixels per 1 µm. The z-axis resolution was approximately 0.71 µm. The confocal lateral resolution was approximately 0.26 µm. Our resolution limits were regarded as sufficient to allow the determination of the density of thorns or spines. Confocal images were then deconvoluted using AUTODEBLUR software (AutoQuant, USA). The density of spines as well as the head diameter was analyzed with Spiso-3D (automated software calculating geometrical parameters of spines) developed by Kawato's group of Bioinformatics Project. Spiso-3D has an equivalent capacity with Neurolucida (MicroBrightField, USA), furthermore, Spiso-3D considerably reduces human errors and experimenter labor. We analyzed the secondary dendrites in the stratum radiatum, lying between 100 and 250 µm from the soma. The spine density was calculated from the number of spines having a total length of 50–80 µm. Spine shapes were classified into three categories as follows. (1) A small-head spine whose head diameter is 0.2–0.4 µm. (2) A middle-head spine whose head diameter is 0.4–0.5 µm. (3) A large-head spine whose head diameter is 0.5–1.0 µm. These three categories were useful to distinguish complex responses upon CORT application. Because the majority of spines (\>95%) had a distinct head and neck, and stubby spines and filopodium did not contribute much to overall changes, we analyzed spines having a distinct head. While counting the spines in the reconstructed images, the position and verification of spines was aided by rotation of three-dimensional reconstructions and by observation of the images in consecutive single planes. ## Statistical analysis Data were expressed as mean ± SEM. For, an unpaired, two-tailed t-test, under the assumption of unequal variances, was utilized to test the significance of observed differences between groups. Several numbers of independent experiments from different animals were used to determine the parameters of t-distribution for the test. For analysis of spinogenesis, we used Tukey–Kramer post hoc multiple comparisons test when one way ANOVA tests yielded P\<0.05. # Supporting Information Dr. Seijiro Honma (Aska Pharma Medical) is acknowledged for collaboration of mass-spectrometric analysis. Prof. John Morrison (Mount Sinai School of Medicine) is acknowledged for immunoelectron microscopic analysis. Prof. J. Ian Mason (University of Edinburgh), Dr. Fumiko Mitani (Keio University) and Prof. Yoshihiko Funae (Oosaka City University) are acknowledged for their kind gifts of 3β-HSD antibody, P450(11β1) antibody, and P450(2D4) antibody, respectively. [^1]: Conceived and designed the experiments: SH YH SK. Performed the experiments: SH YH HI YK YO GM HM TK. Analyzed the data: SH YH YK SK. Contributed reagents/materials/analysis tools: SK TY DN. Wrote the paper: SK SH YH AB. [^2]: The authors have declared that no competing interests exist.
The authors have discovered that an indexing error occurred in the code for the analysis of the article. Despite this error, the general conclusions of the article remain supported by the data: ecosystem functions across trophic levels are linked to functional and phylogenetic diversity. Furthermore, phylogenetic and functional diversity measures explain variation in ecosystem functions beyond that which is explained by traditional diversity measures such as species richness. After rerunning analysis, functional diversity is no longer the best predictor of zooplankton biomass. Rather, functional diversity, together with phylogenetic diversity, explains all the variation explained by species richness, along with additional unique variation. All together, this indexing error resulted in errors in Figs and and and and Tables and and and and, as well as in the Abstract, Results, Discussion, and Conclusion sections. Please see the corrected Figs and and Tables and here; please view the corrected and Figs and and Tables and below. Please see the corrected Abstract, Results, Discussion, and Conclusion text here. In the Abstract, there is an error in the ninth sentence. This sentence should be replaced with the following two sentences: “Measures of zooplankton phylogenetic diversity and trait-based functional diversity together explained all the variation in zooplankton community biomass that was explained by species richness, in addition to unique variation that was not explained by species richness alone. In contrast, phytoplankton abundance was best predicted by zooplankton phylogenetic diversity, which explained all variation explained by any of the three types of diversity indices.” In the Zooplankton community characteristics subsection of the Results section, there are a number of errors in the first and second paragraph. The correct first and second paragraphs are: “Average pond species richness was 4.65, ranging from 3 to 7, with a regional richness of 10. Across all ponds, zooplankton community biomass was 334.77 μg L<sup>-1</sup> on average, ranging from 2.66 to 1398.85 μg L<sup>-1</sup>. Species richness was positively, albeit weakly, related to the number of zooplankton present in our samples (R<sup>2</sup> = 0.29; p = 0.006). However, the rarefaction curves saturated in the majority of samples (87%), suggesting that our estimates of richness were not greatly biased by differences in the abundance of zooplankton amongst the ponds (Figure S5). Daphnia pulex comprised 44.9% of the zooplankton biomass over all ponds and was present in 14 of the 23 ponds. The next most abundant taxon, Harpactocoid copepods, comprised 26.4% of the zooplankton biomass in all ponds and was present in 21 of the 23 ponds. Acanthocyclops vernalis comprised 10.6% of the zooplankton biomass in all ponds and was present in 15 of the 23 ponds. This species is carnivorous as an adult but was retained in our analysis because it consumes phytoplankton in its juvenile stages \[44\]. There were seven rarer taxa, Alonella sp., Ceriodaphnia dubia, Chydorus sphaericus, Harpacticoida, Sida crystallina, Simocephalus sp., and Tropocyclops prasinus mexicanus, that each made up less than 5% of the average biomass. Phytoplankton abundance was 6.56 μg chl a L<sup>-1</sup> on average, ranging from undetectable to 33 μg L<sup>-1</sup>. Zooplankton species richness was not related to either chlorophyll a (R<sup>2</sup> = 0.10; p = 0.140) or total phosphorous (R<sup>2</sup> = 0.11; p = 0.131).” In the Zooplankton community biomass subsection of the Results section, there are a number of errors. The correct Zooplankton community biomass subsection is: “Three out of the 11 diversity measures tested explained a significant proportion of variance in zooplankton community biomass, and in all cases, there was a positive influence of diversity on biomass. These significant models included taxonomic, functional, and phylogenetic measures. Species richness (SR; ; R<sup>2</sup> = 0.38; p = 0.008) and abundance weighted standard effect size mean pairwise distance (sesMPD<sub>ab</sub>; ; R<sup>2</sup> = 0.34; p = 0.003), were jointly selected (equal AIC) as the best single diversity measure predictors of zooplankton community biomass. Species richness (SR) exhibited a unimodal relationship with zooplankton community biomass, where the highest biomass was found in ponds with intermediate species richness. This unimodal relationship between species richness and community zooplankton biomass outperformed a model that assumed a linear relationship (R<sup>2</sup> = 0.14, p = 0.077). Presence absence weighted functional evenness (Feve<sub>pa</sub>) explained the third highest proportion of variance of the single diversity measure models (; R<sup>2</sup> = 0.27; p = 0.010). Based on variation partitioning, SR, Feve<sub>pa</sub>, and sesMPD<sub>ab</sub> together explained 45% of the variation in zooplankton community biomass, and overlapped in explaining 10% of the variation. SR and Feve<sub>pa</sub> overlapped to explain 11% of the variation. SR and sesMPD<sub>ab</sub> overlapped to explain 12% of the variation. Feve<sub>pa</sub> and sesMPD<sub>ab</sub> overlapped to explain 1% of the variation. SR, Feve<sub>pa</sub>, and sesMPD<sub>ab</sub> uniquely explained 0%, 2%, and 8% of the variation respectively. Three of the five traits (trophic group, raptorial vs. filter feeding, and body length) individually explained a significant amount of variance in zooplankton biomass and body length outperformed the best functional diversity measures although ΔAIC was small. The best performing model of environmental variables for predicting zooplankton community biomass consisted of elevation and ln TP, and explained a significant amount of variation (R<sup>2</sup> = 0.39; p = 0.007). However, when either sesMPD<sub>ab</sub> or species richness was included in the model, none of these environmental variables remained as significant predictors, although the models outperformed all other models. Chlorophyll a did not explain a significant amount of variance in zooplankton community biomass (R<sup>2</sup> = 0.02; p = 0542). The pathway between sesMPD<sub>ab</sub> (chosen for use in the SEM models because it performed equally well to SR and was the best predictor when combined with the environmental variables) and zooplankton community biomass, was always significant, regardless of how we specified the effect of environment in our SEM. The most parsimonious model, based on AIC, only included the direct pathway from sesMPD<sub>ab</sub>. This suggests that our linear models adequately capture the relationship between diversity and zooplankton biomass.” In the Phytoplankton abundance subsection of the Results section, there are a number of errors. The correct Phytoplankton abundance subsection is: “Four out of the 11 zooplankton diversity measures explained a significant proportion of variance in chlorophyll a, and in all cases there was a negative influence of diversity on chlorophyll a. These significant models included functional and phylogenetic, but not taxonomic diversity measures. The best single diversity measure for predicting chlorophyll a was the phylogenetic diversity measure sesMPD<sub>pa</sub> (R<sup>2</sup> = 0.44; p = 0.001). There was one outlier in this relationship, which was found to have significant influence (Cook’s distance \> 0.5) on the analysis (—unfilled point). Chlorophyll a was not detectable in this pond, although the predicted concentration should have been relatively high based on the measured zooplankton phylogenetic diversity. We cannot be sure if this chlorophyll a concentration is a measurement error, so we compared model fit with and without including it. Removing this outlier from our analysis did not have a large effect on the slope of the relationship but greatly improved the model fit (dashed line; R<sup>2</sup> = 0.63, p \<0.001). The functional diversity measure with the lowest AIC was abundance weighted functional evenness (FEve<sub>pa</sub>; R<sub>2</sub> = 0.34, p = 0.004). However, this model slope was highly sensitive to the inclusion of the outlier, and the next best functional diversity measure FRic had a similar AIC but was not as influenced by the outlier and so we have elected to use FRic as our best functional diversity measure. FRic had a positive relationship with chlorophyll a (R<sub>2</sub> = 0.28, p = 0.010) and excluding the outlying pond did not change the slope of the relationship but improved the model fit (R<sup>2</sup> = 0.49, p \< 0.001). The best measure of taxonomic diversity was species richness (SR; ; R<sup>2</sup> = 0.10; p = 0.140), but no taxonomic diversity measure was able to explain a significant portion of variance in chlorophyll a. Again, excluding the outlying pond did not change the slope of the relationship but improved the model fit (dashed line; R<sup>2</sup> = 0.22, p = 0.02). Based on variation partitioning, SR, FRic, and sesMPD<sub>pa</sub> together explained 37% of the variation in chlorophyll a. However, all variation explained was captured by sesMPD<sub>pa</sub>, either alone (4%) or with SR (11%) or FRic (30%). SR and FDiv each uniquely did not contribute to explaining variation in chlorophyll a, nor did the overlap between all three indices, and this resulted in less variation explained by the three indices together than that explained by sesMPD<sub>pa</sub> on its own, because adjusted R<sup>2</sup> penalizes for the additional degrees of freedom used in the combined model. Two of the five traits (raptorial vs. filter feeding, and feeding type) individually explained a significant amount of variance in chlorophyll a. No single trait performed as well as the best phylogenetic diversity measure (sesMPD<sub>pa</sub>). The best performing model for chlorophyll a containing only environmental variables consisted of ln DIC, ln depth, ln pH, ln wet days and ln TP but was not significant (p = 0.052). The model combining these environmental variables plus sesMPD<sub>pa</sub> did not perform as well as the model with sesMPD<sub>pa</sub> alone. Neither zooplankton community biomass nor D. pulex biomass explained a significant amount of variation in chlorophyll a (Community Biomass R<sup>2</sup> = 0.02; p = 0.542; D. pulex–R<sup>2</sup> = 0.07; p = 0.221). The pathway between sesMPD<sub>pa</sub> and chlorophyll a, was always significant, regardless of how we specified the effect of environment in our SEM (Figure S8). Matching to the SEM with zooplankton community biomass, the most parsimonious model explaining variation in chlorophyll a did not include environment, but it included the direct pathway from sesMPD<sub>pa</sub>. This again suggests that our linear models adequately capture the relationship between diversity and chlorophyll a.” In the Discussion section, there are errors in the second, fourth, and fifth paragraph. The correct second paragraph is: “As predicted, we found that positive diversity ecosystem function relationships emerged most clearly when measures of functional and phylogenetic diversity were used, and that these measures explained variation in ecosystem function beyond that explained by taxonomic diversity measures, such as species richness. Previous studies relating the diversity of animals to ecosystem function have relied on taxonomic diversity measures \[47,48\], knowledge of the functional complementarity of species \[49,50\], single traits \[13\] or on taxonomic differences \[16\], but see \[14\]. Species richness, phylogenetic diversity (sesMPD<sub>ab</sub>), and functional diversity (Feve<sub>pa</sub>) all explained significant amounts of variation in zooplankton biomass. Furthermore, none of this variation was explained solely by species richness, whereas phylogenetic diversity and functional diversity each explained additional unique variation. Although we found a subset of single traits (e.g. trophic group, raptorial vs. filter feeding, and body length) performed almost as well in predicting zooplankton biomass, only body size explained as much variation as the best diversity measures. In contrast, phylogenetic diversity (sesMPD<sub>pa</sub>) was the best predictor of phytoplankton consumption, and no additional variation was uniquely explained by species richness and functional diversity (FRic). These findings suggest that metrics that quantitatively integrate trait or phylogenetic information have the potential to improve our understanding of variation in the functioning of complex multi- trophic ecosystems.” The fourth paragraph of the Discussion section is no longer relevant after the reanalysis and should be removed. The correct fifth paragraph of the Discussion is: “No single diversity index was consistently the best predictor of ecosystem function, and the degree to which they explained unique variation differed depending on the measured function. While taxonomic diversity (SR<sup>2</sup>) explained the most variation in zooplankton biomass, all of this variation was captured by either phylogenetic diversity (sesMPD<sub>ab</sub>) or functional diversity (Feve<sub>pa</sub>), which each captured additional unique variation. In contrast, phylogenetic diversity (sesMPD<sub>pa</sub>) explained the most variation in phytoplankton abundance and although both species richness and functional diversity (FRic) explained overlapping variation, these were subsets of that explained by phylogenetic diversity. These findings suggest that the three types of diversity indices capture some of the same functional differences in community composition. This is perhaps unsurprising because functional traits and niche differences are often phylogenetically conserved \[10,52\], as reflected by the high overall correlation between the traits and phylogeny. However, the diversity measures did not overlap completely in the variance in ecosystem function that they explained, and each function was best predicted by a different diversity measure. For example, body size showed little phylogenetic signal, but was predictive of zooplankton community biomass, and this correlation may explain why functional diversity explained unique variation not captured by phylogenetic diversity for this ecosystem function. In contrast, the fact that phylogenetic diversity explained additional variation in phytoplankton abundance to that explained by functional traits is suggestive of other important, but unmeasured, functional differences that covary with phylogeny. Each class of metric thus captured some unique aspect of the way that the communities use resources \[53\], highlighting the value of combining different diversity metrics in models explaining ecosystem function.“ In the Conclusion, there is an error in the third sentence of the second paragraph. The correct sentence is: “Zooplankton biomass production is best explained by a combination of functional and phylogenetic diversity, whereas suppression of phytoplankton production was best explained by phylogenetic diversity.” # Supporting information
# Introduction Plastid and mitochondria are essential organelles in plant cells. Chloroplasts conduct photosynthesis in the presence of sunlight and mitochondria indirectly supply energy within plant cells; together they form the powerhouses of the cell. Both chloroplasts and mitochondria possess their own genomes. The chloroplast (cp) genome and mitochondrial (mt) genomes are often used for the study of plant evolution. From the information of all sequenced cp genomes, most of them range from 120 to 160 kb in length and have GC contents of 30 to 40%. The quadripartite organization is shared by almost all cp genomes, consisting of a large-single-copy region (LSC; 80–90 kb) and a small-single-copy region (SSC; 16–27 kb), as well as two copies of inverted repeats (IRs) of ∼20 to 28 kb in size. The gene content and structure of angiosperm cp genome is highly conserved. Expansion and contraction of the IR as well as gene and intron losses have been documented in a wide range of angiosperms. The mt genome plays crucial roles in plant development and productivity. In comparison to other non-plant eukaryotes, plants have large and complex mt genomes. Mt genomes of seed plants are unusually large and vary in size at least in an order of magnitude. Much of these variations occur within a single family. Seed plant mitochondrial genomes are characteristic for their very low mutation rate, frequent uptake of foreign DNA by intracellular and horizontal gene transfer, and dynamic structure. The evolving land plants have gained new mechanisms to facilitate more frequent gene exchanges between mt and cp genomes as well as between mt and nuclear genomes, which make mt genomes increase their sizes.. In the past several years, we have witnessed a dramatic increase in the number of complete organellar genomes, especially those of plants. Until now, there are 206 complete cp genomes and 47 mt genomes having been deposited in GenBank Organelle Genome Resoures. With the emergence of next-generation sequencers, new approaches for genome sequencing have been gradually applied due to their high- throughput, time-saving, and low-cost advantages. With a new gene-based strategy and combining data from the two next-generation sequence platforms, pyrosequencing (Roche GS FLX) and ligation-based sequencing (Life Technologies SOLiD), we successfully assembled cp and mt genomes of resurrection plant B. hygrometrica (Bunge) R. Br. B. hygrometrica or Bh is a small dicotyledenous, homiochlorophyllous resurrection plant in the Gesneriaceae family, and it is widespread in China, inhabiting shallow rock crevices from humid tropical regions to arid temperate zones. In this study, we analyze genomic features and structures of both cp and mt genomes of B. hygrometrica. Through organellar genome comparison with other lower plants and angiosperms, we provide information for the better understanding of organellar genome evolution in land plants. # Results and Discussion ## Features of B. hygrometrica cp genome and mt genome The Bh cp genome is 153,493 bp in length and has a GC content of 37.59%. Similar to those of other angiosperms, the Bh cp genome maps as a circular molecule with the typical quadripartite structure: a pair of IRs (25,450 bp, covering 16.6%) separated by the LSC (84,692 bp, covering 55.1%) and SSC (17,901 bp, 11.7%) regions. It encodes 147 predicted functional genes and 19 of them are duplicated in the IR regions. Among these 147 genes, we identified 103, 36, and 8 protein-coding, tRNA, and rRNA genes, respectively ( and). 38% of the genome is non-coding, including introns, intergenic spacers, and pseudogenes. All the rRNA genes (rrn16, rrn23, rrn5 and rrn4.5) and 7 tRNA (trnI- CAU, trnL-CAA, trnV-GAC, trnI-GAU, trnA-UGC, trnR-ACG, and trnN- GUU) genes are located in IR regions. Similar to other dicot species, Bh has two genes (rps19 and trnH) located in the position of IR/LSC junctions. This is different in monocots, such as rice and maize, whose cpDNAs have a fully duplicated rps19 gene in the IR/LSC junctions. The average length of intergenic regions is 385 bp, varying from 1 to 2,221 bp. There are 4 cases of overlapping genes (psbD-psbC, ndhK-ndhC, atpE-atpB, and ycf1-ndhF), resulting in an average coding density (including conserved genes, unique ORFs and introns) of 1/1,058 bp. The cp genome has 19 intron-containing genes with 12 in protein-coding genes and 7 in tRNA genes. In terms of size, gene content, and intron composition, Bh cpDNA closely is mapped to Olea europaea cpDNA (155,872 bp, GC 37%) among all angiosperms. Sequence alignment shows that 93% (142,189 bp) of the Bh cp genome sequence are covered by that of O. europaea with 95.4% identity. 36 tRNAs are detected, enabling B. hygrometrica cp genome to decode all 61 codons. All of 3 stop codons are present with UAA being the most frequently used (UAA 40%, UAG 33% and UGA 27%). Phylogenetic analyses, which were constructed by 63 protein-coding sequences from 12 cp genomes (one green algae as outgroup, one land and 10 seed plants) , indicates that the phylogenetic position of B. hygrometrica is closer and older than V. vinifera among the analyzed dicots. The Bh master mt genome is assembled into a circular molecule of 510,519 bp in length and has an average GC content of 43.27%. This size is bigger than the mt genome of A. thaliana (366,924 bp), but smaller than Vitis vinifera (773,279 bp) among dicots. With only 12% of coding sequences, the largest part (88%) of this genome is non-coding, containing 1.45% repeat and 10.52% cp-derived sequences. The mt genome has 65 genes, including 33 protein-coding, 4 rRNA, and 28 tRNA genes, and 10 genes have exon-intron structure. Similar to other angiosperms, the Bh mitochondrion uses the canonical genetic code. All 61 codons are present in mt genome, and UAA (46%) is the most frequently-used stop codon. The known 33 protein-coding genes in mt genome are similar to other published angiosperm mt genomes, such as 9 subunits of NADH dehydrogenase (complex I), 5 subunits of ATP synthase (Complex V), and 3 subunits of cytochrome c oxidase (complex IV). Compared to other angiosperms, we observed that there is one sdh3 and no sdh4 in Bh mt genome. There are 3 copies of sdh3 in mt genome of Nicotiana tabacum, and both sdh3 and sdh4 are present in that of V. vinifera. The Bh cox1 has an intron/exon structure that is unlike other higher seed plants. There are two 5S rRNA (rrn5) copies and one copy of rrn5 from its cp genome. The best sequence alignment score belongs to V. vinifera mt genome, with 23% (119,377 bp) of Bh mt genome being alignable to that of V. vinifera with 94.2% identity. Plant cells often contain multiple clones or copies of cp and genomes, and thus the organellar genomes can be regarded as a population with genetic heterogeneity. Polymorphic sites can be detected by aligning thousands of high quality reads to assembling of the cp or mt genome. Our SNP analysis shows that there is no intervarietal SNPs (intraSNPs) found in Bh cp genome. However, we identified 729 SNPs in Bh mt genome and SNPs in mt genome occurred at a rate of 1 in 700 bp. We only detected 9 SNPs in gene regions with 2, 1, 1, and 5 in rrn5, rrn26, trnM-CAU, and rrn18, respectively. There are no intraSNPs detected in known protein-coding gene regions. The intraSNPs have been demonstrated to be present in both cp and mt genomes of rice,. As an indicator for the heterogeneous nature of cp and mt populations, the intravarietal polymorphisms provide us useful markers for the future genetic studies on B. hygrometrica. ## Structural dynamics of Bh mitochondrial genome in ontogeny All previous studies on complete sequencing of flowering plant mt genomes are based on the master circle (MC) hypothesis,. Arrieta-Montiel et al. reported on the structural dynamics of the common bean mt genome. The analysis of 10 recombinant clones supports existence of the MC molecule in wheat mt genome. However, the result of field-inversion electrophoresis suggests that Physcombitrella mt genome does not consist of a multipartite structure, as seen in angiosperms. As mitochondrial gene orders are significantly different between lower plants and higher flowering plants, the multipartite structures as seen among angiosperms may originate during the evolution of pteridophytes or seed plants. Repeat prediction by REPuter shows that there are 14 repeat pairs in Bh mt genome. Since not all the repeats are involved in recombination, from the mt assembly we detected 3 repeat-specific contigs that are candidates for the recombination among the MC and other isomeric (IO) and subgenomic molecules, which have been confirmed by SOLiD sequencing. Those 3 repeat contigs are the repeat pairs of two palindromic matches (1,474 bp and 843 bp) and one forward match (222 bp). By aligning all SOLiD long mate-pair reads to both ends of the repeats, we constructed the MC and 4 isomeric molecules. These subgenomic molecules are not discussed further because we have yet to find significant differences among the sequence reads. The length of recombinant repeats (222 bp, 843 bp, and 1,474 bp) of Bh is different from that of V. radiata, and demonstrates the recombination across short mt repeats (38–297 bp). The longer repeats are reminiscent of those found in other angiosperm mitochondrial genomes, which are involved in mt genome rearrangements and can result in stoichiometric shifting of subgenomic mt genome topologies, occasionally beyond detection level for one (or more) of alternative DNA topologies. The 3 recombinant repeats are located in gene-rich regions and split mt genome into 6 segments. Each segment has some essential conserved genes, such as nad1 e4 and rps13 in segment A, and atp6, nad2 e3-5 and coxb in segment B. From this point, it is possible that the mt genome of Bh do not have subgenomic molecules. There are 4 genes (nad1, nad2, nad5 and sdh3) with exon-intron structure separated by the recombinant repeats. The nad1 gene in MC molecule is cross-strand gene with exon 4 in positive strand and exon 1–3 in negative strand. Recombination involving introns might lead to rearranged molecules without loss of essential genes. In all rearranged 4 isomeric molecules, the Trans-splicing genes are different, and the IO3 molecule has all 4 Trans-splicing genes. Trans-splicing status of group II intron widely distributes in the mt genome of higher plant. We compared gene structures of 15 mt genomes from lower to higher plants, and found 3 conserved genes (nad1, nad2, and nad5) as well as other higher plants contain trans-splicing intron, while there is no intron in those genes of Chara vulgaris. The 3 genes structure supported the multipartite structures formed by multiple recombination may arise with the earliest tracheophytes, and can be a molecular signature of plant evolution. ## Comparative analysis of cp genome organization We compared 12 cp genomes ranging from green algae to angiosperm. The GC contents of cp genome are lowest in lower plants (Charophyta and Bryophyta) but highest in Cycads. Monocots seem to have slightly higher GC contents than dicots among their cp genomes. The genome size and structure of cp genomes are also different in those cp genomes. For example, C. vulgaris (184,933 bp; Charophyta) has the largest genome while the smallest genome is found in lower plant Marchantia polymorpha (121,024 bp). The genome size of angiosperms is more stable than lower plants with dicots larger than monocots. Compared to lower pants, the most variable portions of angiosperm cp genomes are percentages of IRs (34% in A. thaliana) and LSC (54.5% in A. thaliana) regions. This is the result of IR expansion into the LSC region from green algae to angiosperm. The cp genome contains genes that encode structural and functional components of the organelle. Although some genes and gene clusters are well conserved among all plants, the overall structure of cp genomes show remarkable differences. First, there are 63 core protein-coding genes, shared among all plants, whereas there are 3 additional core genes (chlB, chlL and ycf12) only found in the lower plant lineage (Charophyta, Bryophyta, and Cycads). The 63 core cp genes are involved in photosynthesis, energy metabolism, and other housekeeping functions. Second, there are 10 genes (psaM, rpl5, rpl12, rpl19, tufA, ycf20, ycf62, ycf66, odpB, and ftsH) are unique to green algae. All of them reside in the LSC region except ycf20 gene that is duplicated in IR regions. Compared to seed plants, there is only one gene, ribosomal protein L21 (rpl21) is conserved in both green algae and liverwort. Third, all four ribosomal RNA genes (rrn4.5, rrn5, rrn16, and rrn23) have 2 copies in IR regions except the 2 copies of rrn4.5 that is lost in Charophyta. Fourth, gene loss and transfer to the nucleus is a common feature of cp genomes. We detected 3 genes (petL, petN, and ycf3) that are lost at the base of the Bryophyta lineage and 2 genes (accD and ycf2) are lost in monocots as compared to dicots. There are also some species-specific gene lost events, such as psaJ in O. sativa and nadJ-ccsA lost in O. europaea. The unique loss of psbZ in LSC region testifies the convergent evolution of B. hygrometrica and O. europaea. The order of cp genes in plants is not constant, changing among different regions of the genome as large gene clustering become rare. Among 63 core protein-coding genes, 50 are always reside in LSC region, 5 (psaC, ndhD, ndhE, ndhG, and ndhI) in SSC region, and 8 (ndhA, ndhB, ndhF, ndhH, rpl2, rpl23, rpl32 and rps7) in variable positions among 12 examined cp genomes. No conserved protein-coding genes are found constant in IR regions. These mobile genes may serve as an indication of lineage markers, since 4 of them (ndhB, rpl2, rpl23, and rps7) locate on LSC region in lower plants and migrated to IR regions in higher plants. Genes residing in the boundary of LSC/IRA or IRB/LSC are usually ribosomal proteins S12 (rps12) in higher plants and the position-conserved ycf1 is more likely present in the boundary of IRA/SSC and SSC/IRB in dicots. ## Comparative analysis of plant mt genomes The plant mt genomes are exceptionally variable in size, structure, and sequence content and the accumulation of repetitive sequences contributes the most to such variation. From the feature comparison of 15 plant mt genomes, we noticed that their genome sizes vary from 67,737 bp in C. vulgaris to 773,279 bp in V. vinifera. Recently, the large mt genome have been reported in Cucurbita pepo with 982,833 bp. The GC contents of these genomes are also variable from 40 to 47%. There is a massive difference of coding sequences between lower and higher plants. The coding sequence in C. vulgaris is 90.7%, whereas it is 4.94% in Tripsacum dactyloides. Repeat content ranges from 1 to 41% among seed plants and are smallest in B. hygrometrica, composed of only 1.45% of the genome. Both large (\>1,000 nt) and small (\<50 nt) repeats affect recombination in seed plants. The protein-coding genes and tRNAs in mt genomes also vary largely because of the large number of function-unknown proteins or ORFs in mt genomes and frequent plastid DNA insertions including cp tRNA genes. We also carefully examined conserved genes in different plant lineages. First, there are 14 conserved core protein-coding genes shared among all lineages, including seven subunits of NADH dehydrogenase (Complex I), one subunit of ubichinol cytochrome c reductase (Complex III), three subunits of cytochrome c oxidase (Complex IV), and three subunits of ATP synthase (Complex V). All these genes play important roles either in proton movement across the inner membrance of the mitochondrion or electron transfer reactions in the respiratory chain. However, the gene structures are not conserved among them, and only two genes (nad4 and cox2) have exon-intron structures in all mt genomes. For comparison, there are 9 genes (nad3, nad4L, nad6, nad9, cob, cox3, atp1, atp9, and ccmFN) without exon-intron structure in both seed and early land plants and with exon-intron structure at least in one lower plant. Intron structure in mt genes is common as we only detected 6 genes (sdh3, sdh4, atp4, atp8, ccmB, and ccmFN) have no introns among all plants. Second, gene loss is more frequent in dicots than monocots, as genes in cytochrome c biogenesis are lost in both B. vulgaris and A. thaliana. Three species (Nicotiana tabacum, V. vinifera and B. hygrometrica) gained sdh3 as in this study. The number of ribosomal protein genes is different in various plant mt genomes. Most ribosomal proteins (23) are present in V. vinifera genome. Contrast to higher plants, there is no matR detected in liverwort and green algae in this study. However, it is reported that in mosses, Takakia and Sphagnum have part of matR. Most of mt genomes in plants have 3 ribosomal RNAs (rrn5, rrn18, and rrn26), but there are multiple copies found in angiosperms (such as T. aestivum and B. vulgaris). Copy number-variable mt genes are reported in wheat, rice, and maize. In summary, since the gene coding fraction is much less among mt genomes as compared to cp genomes, even conserved genes are also variable in gene content, structure, and intron positioning. ## Plastid DNA insertions in mt genome One of the important events in determining mt genome size in angiosperms is the frequent capture of sequences from the cp genome. A recently study demonstrates frequent DNA transfer from cp to mt genomes occur as far back as the common ancestor of the extant gymnosperms and angiosperms, about 300 MYA and the frequency of cp-derived sequence transfer is positively correlated with variations in mt genome size. For instance, B. hygrometrica mt genome contains fragments of cp origin, ranging from 50 to 5,146 bp in length. The total fraction of cp-derived sequences present in Bh mt genome is 53,440 bp, 10.5% of the whole mt genome. Most of these insertions are conserved, as evidenced from the observation that 45 out 80 insertions (over 50 bp) are identified in mt genomes of other plants. The average GC content between old insertions (at least found one homolog in the mt genomes of other plants) and new insertions (no homologs found in other mt genomes) is obviously different. The average GC content of old insertions (41.59%) is distinct from that of mt genomes (43.27%) as well as the average GC content of new insertions (35.32%) is close to the GC content of cp genomes (37.59%;). The result suggests positive correlations between the GC content of cp genomes and new insertions (coefficient value r<sup>2</sup> = 0.69) and between the GC content of mt genomes and old insertions (coefficient value r<sup>2</sup> = 0.64). There is a significant difference between the GC content of old and new insertions (T test: P\<0.01). All of cp-derived sequences in lower plants are old insertion, and it is strange to see there are no new insertions from cp genomes during the evolution of A. thaliana. Compared to other angiosperms, the most striking feature of Bh mt genome is thefrequent sequence transfer from the cp genome in its recent evolution. Cp-derived sequence analyses may provide clues for the understanding functional indications of DNA insertions from cp in mt genomes. We also analyzed the inserted cp genomic sequences to mt genomes. First, there are 85 cp-derived fragments in the Bh mt genome. Most of them, especially protein-coding sequences, have no intact gene structure or have frameshifts/indels and the fact suggests that theses cp-derived sequences are degenerated and lack functional constraints. However, protein-coding genes such as ribosomal proteins and tRNAs originated from cp genomes appear still functional in Bh mt genome. Similar genes are also seen in other angiosperms. Second, it is curious to investigate the locations of cp-derived sequences in cp DNAs to see if any particular regions in cp genomes are hotspots of DNA transfer. Among 80 transferred sequence fragments in Bh cp genome, 45, 7, and 28 are from LSC, SSC, and IR regions, respectively. The numbers of fragments appear correlating to the lengths of LSC (84,692 bp), SSC (17,901 bp) and IRs (50,900 bp), and such correlations are also seen in other angiosperms, including maize, rice, wheat, and tobacco. In conclusion, DNA transfer from cp genomes to mt genomes in angiosperms occurs randomly as it has been proposed earlier by Mastsuo et al in rice. ## tRNAs transfer between cp and mt genomes To investigate whether mt genomes encodes a full set of tRNAs species necessary for protein synthesis in the organelle, we identified 28 tRNA genes from the complete Bh assembly based on tRNA structures and realized that all 61 codons are used by Bh mitochondria. However, the tRNA genes encoded by the mt genome alone are not sufficient to decode all codons; for instance, trnA is missing in Bh, and it suggests that the role of the missing tRNA is supplied by either cp or nuclear genomes. tRNAs originated from plastids are called cp-derived tRNAs and their counterparts are native mt tRNAs. Half of the 28 mt tRNAs in B.hygrometrica are identified as cp-derived tRNAs and 19 amino acids are encoded by only one codon except for leucine (UAA and CAA) and serine (GCU, UGA, and GGA). In contrast to the protein coding genes, mt tRNA genes appear constantly being transferred from cp genomes during the evolution of angiosperms ( and), and the proportion of cp-derived tRNAs in mt genomes increases from 8% in Charophyta to 55% in dicotyledonous plants. There are 17 cp-derived tRNAs and 14 mt native tRNAs in mt genome of V. vinifera, which has the most cp-derived tRNAs among dicotyledonous plants. Seven mt-native tRNA genes (trnD-GUC, trnE-UUC, trnI-CAU, trnK-UUU, trnM-CAU, trnS-GCU, trnS-UGA and trnW-GUA) and one cp-derived tRNA gene (trnF-GAA) are common to all 15 species. Compared to the mt native tRNA genes in lower plants, there are three tRNA genes (trnH- GUG, trnN-GUU, and trnW-GUA) integrated as part of large cp genomic fragments into mt genomes among angiosperms. This indicates that frequent DNA transfer from cp to mt genomes occur as far back as the emergence of seed plants. We detected two different tRNAs transfer events in seed plants. One is trnC-GCA transfer in monocots and the other involves two (trnD-GUC and trnQ-UUG) gene transfer events in dicots. Cp-derived tRNA genes replace their mt counterparts were identified in all sequenced angiosperms, even in gymnosperm Cyas mtDNA. But these replacement not occurred in Marchantia, Reclinomonas, Cyanidioschyzon, Nephroselmis, Chara, and Physcomitrella.The mt-native tRNA gene (trnG-GCC) had all been lost in monocots and six mt-native genes (trnA-UGC, trnG-UCC, trnL-UAG, trnR- UCU, trnR-ACG, and trnT-GGU) are mostly lost in all angiosperms. ## Gene gain and loss in plant organelle genome Starting from Bh organellar genomes, we have analyzed in a systematic way representative cp and mt genomes of various lineages and our results provide information for a better understanding of organellar genome evolution and function. Sequence-based phylogenetic analysis clearly supports the conclusion that Bh is much close to V. vinifera. Structural dynamics of Bh mt genome suggest that the multipartite structures may have started during the evolution of seed plants. However, mechanisms for rapid mt genome rearrangement and expansion among plant lineages remain enigmatic. Based on eleven known cp and mt genomes of different lineages, we showed a strong relationship between the changing organellar genomes among angiosperms, and some of the lineage- associated gene gain and loss may provide excellent markers for phylogenetic studies. For instance, there are 9 cp and 4 mt genes lost during the evolution from green algae to lower land plants. It seems that monocots have a faster rate of evolution than dicots in organellar genomes in our study, because 3 cp and 9 mt genes are lost in monocots and only 2 mt genes are lost in dicots. In additional, gene structures and positioning of cp and mt genomes are also very informative for the understanding of land plant evolution. In agreement with the results of several previous studies, most of the transferred angiosperm sequences from cp to mt genomes become degenerated and are regarded as junk sequences, whereas some of the cp-derived tRNAs are still functional in mt genomes. As more plant organellar genome sequences become available, the evolution of plant organellar genomes will unveil its details and mechanisms. # Materials and Methods ## Genome sequencing and assembly We developed an efficient procedure for Bh organellar genome sequencing and assembly using whole genome data from 454 GS FLX sequencing platform. Briefly, we collected fresh leaves and extracted genomic DNA for 454 GS FLX sequencing (see manuals of GS FLX Titanium for detail). In order to validate genome assembly and to make sure for the assembly of the master circle or MC, we construed two mate-pair libraries (2×50 bp) for SOLiD 4.0 sequencing platform with insert sizes of 1–2 kb and 3–4 kb by following the SOLiD Library Preparation Guide. The method for assembling organellar genome was based on correlation between contig read depth and copy number in the genome. We first filtered cp reads from the raw data according known plant cp genome sequences and then assembled the “clean” read into cp genome into the major segments: large-single-copy (SSC), small-single-copy (SSC) and inverted repeats (IRs) regions. The mt genome assembly is more complicated than that of cp genomes. We filtered the contigs including mt conserve genes (such as NADH dehydrogenase and succinate dehydrogenase) and removed the contamination of cp sequences. The gene-based method for assembling mt genome has been reported earlier. Mapping all the SOLiD mate-pair reads to mt contigs with Bioscope, we obtained the major contig relationship map in the repeat regions to assemble the MC. ## Genome annotation The cp genome was annotated by using the program DOGMA (Dual Organellar GenoMe Annotator) coupled with manual corrections for start and stop codons. Protein- coding genes are identified by using the plastid/bacterial genetic code. Codon usage is predicted by using CodonW (<http://codonw.sourceforge.net/>). We construct a custom-designed amino acid database for protein-coding genes and nucleotide databases for rRNA and tRNA genes, compiled from all previously annotated plant mt genomes available at the organelle genomic biology website at NCBI (<http://www.ncbi.nlm.nih.gov/genomes/ORGANELLES/organelles.html>). NCBI BlastX and BlastN searches of the mt genome against the databases allow us to find protein and RNA genes, respectively. All BlastN and BlastX searches are carried out by using the default settings with e-value 1e-10. Putative RNA editing sites are inferred to create proper start and stop codons as well as to remove internal stop codons. We also used tRNAscan-SE to corroborate tRNA boundaries identified by BlastN. The annotated GenBank files of the cp and mt genomes of Bh are used to draw gene maps using OrganellarGenome DRAW tool (OGDRAW). The maps were then examined for further comparison of gene order and content. ## Analyses on SNPs, repeats, and cp-derived sequences We identified intra-specific SNPs in both cp and mt genomes. Using BioScope, we mapped two runs of SOLiD mate-pair reads to both cp and mt genomes (BioScope Software User Guide). We carried out repeat sequence analysis using the REPuter web-based interface (<http://bibiserv.techfak.uni-bielefeld.de/reputer/>), including forward, palindromic, reverse and complemented repeats with a minimal length of 50 bp. Transposable elements and other repeated elements were mapped with RepeatMasker Web Server (<http://www.repeatmasker.org/cgi- bin/WEBRepeatMasker>) running under the cross_match search engine. Cp-derived sequences are identified with BlastN search of mt genomes against Bh annotated cp genomes (Identity ≥80%, E-value≤1e-5, and Length ≥50 bp). The cp-derived sequences were then aligned to all known plant mt genomes by using BlastN (Identity ≥80%, E-value≤1e-5, and Coverage ≥50%). tRNAs transferred to the mt genome were identified by aligning to all tRNAs in the cp genome of the same species by using BlastN (Identity ≥80%, E-value≤1e-5, and Coverage ≥50%). ## Phylogenetic Analysis We compare the Bh cp genome with other plant organellar genomes, and use the homologous protein-coding sequences to construct phylogenetic tree. Sixty-three cp protein sequences (psaA, psaB, psaC, psaI, psbA, psbB, psbC, psbD, psbE, psbF, psbH, psbI, psbJ, psbK, psbL, psbM, psbN, psbT, petA, petB, petD, petG, atpA, atpB, atpE, atpF, atpH, atpI, rbcL, rpoA, rpoB, rpoC1, rpoC2, ndhA, ndhB, ndhC, ndhD, ndhE, ndhF, ndhG, ndhH, ndhI, ndhJ, rpl2, rpl14, rpl16, rpl20, rpl22, rpl23, rpl32, rpl33, rpl36, rps2, rps3, rps4, rps7, rps8, rps11, rps14, rps18, clpP, cemA, and ycf4) from 12 different organisms are aligned and concatenated into a dataset of 196,313 amino acids. We align amino acid sequences from individual genes using MUSCLE v3.8.31 , remove ambiguously aligned regions in each alignment using GBLOCKS 0.91b, and concatenate the aligned sequences. We use maximum likelihood method and PhyML v3.0 under Jones-Taylor-Thornton (JTT and gamma distribution of rates across sites with four categories) model of sequence evolution to construct phylogenetic trees. Confidence of branch points is estimated based on 100 bootstrap replications. We obtained the best tree after heuristic search with the help of Modelgenerator. ## Accession Numbers The GenBank accession numbers for the sequences mentioned in this article are as follows: Chara vulgaris, NC_008097 and NC_005255; Marchantia polymorpha, NC_001319 and NC_001660; Megaceros aenigmaticus, NC_012651; Cycas taitungensis, NC_009618 and NC_010303; Triticum aestivum, NC_002762 and NC_007579; Oryza sativa, NC_001320 and NC_011033; Sorghum bicolor, NC_008602 and NC_008360; Tripsacum dactyloides, NC_008362; Zea mays, NC_001666 and NC_007982; Beta vulgaris, NC_002511; Brassica napus, NC_008285; Arabidopsis thaliana, NC_000932 and NC_001284; Nicotiana tabacum, NC_001879 and NC_006581; Vitis vinifera, NC_007957 and NC_012119; Olea europaea, NC_013707; Boea hygrometrica, JN107811 and JN107812. # Supporting Information We thank Beijing Institute of Genomics, Chinese Academy of Sciences for technical support from. We also wish to thank Liancheng Huang for his preparation of the B. hygrometrica materials for this project. [^1]: Conceived and designed the experiments: JY XD SH. Performed the experiments: XW XZ. Analyzed the data: TZ YF. Contributed reagents/materials/analysis tools: XD. Wrote the paper: TZ YF. [^2]: The authors have declared that no competing interests exist.
# Introduction Pancreatic cancer is one of the most lethal cancers, and its 5-year survival rate is only 4%. It is characterized by excessive desmoplasia, which plays a crucial role in its aggressive behavior. Recently, research on cancer biology has focused on cancer–stroma interactions. Interactions between cancer cells and stromal tissues are essential for the development and progression of tumors. Therapies targeting cancer–stroma interactions may represent a new approach for the control of pancreatic cancer. Pancreatic stellate cells (PSCs) have been identified as the principal source of the excessive extracellular matrix observed in chronic pancreatitis and pancreatic adenocarcinoma. Similar to hepatic stellate cells, an important cell type for extracellular matrix production in hepatic fibrosis, PSCs store fat droplets containing vitamin A within their cytoplasm. PSCs become activated upon stimulation by various autocrine or paracrine factors. They express α-smooth muscle actin (α-SMA) and produce various extracellular matrix proteins. Soluble factors secreted by activated PSCs promote the proliferation, migration, invasion, and survival of pancreatic cancer cells against gemcitabine therapy. Thus, PSCs play an important role in cancer–stroma interactions in pancreatic cancer. Several reports suggest that stromal cells, such as myofibroblasts and mesenchymal cells, isolated from various human tissues exhibit different phenotypes. We previously reported that CD10<sup>+</sup> PSCs enhance the progression of pancreatic cancer cells. These observations indicate that PSCs have functional heterogeneity and further suggest that PSCs may contain several other cell subpopulations that individually or synergistically affect the progression of pancreatic cancer. Detailed characterization of the PSCs in pancreatic cancer would help to clarify the mechanism underlying the interactions between cancer cells and stromal cells, and may provide novel targets for stroma-directed therapies. CD271 (also known as nerve growth factor receptor, NGFR or p75NTR) is a neurotrophin receptor that has been implicated in the paracrine growth regulation of a number of neuronal and non-neuronal tumor types. Recent studies have focused on CD271 because it was identified as a marker of human mesenchymal stem cells and it has been evaluated as an important cancer stem cell marker in melanoma. CD271 might be a marker of a specific functional subpopulation, such as stemness. In the pancreas, CD271 expression in PSCs has been detected. However, CD271 expression rapidly decreased after cell isolation from pancreatic tissues. Therefore, the role of CD271 expression in PSCs remains unclear. The aim of this study was to identify the specific PSCs that affect the progression of cancer cells by focusing on the expression of CD271. We further assessed the significance of CD271 expression in PSCs. # Materials and Methods ## Ethics statement The study was approved by the Ethics Committee of Kyushu University (approval number, 23–64) and conducted according to the Ethical Guidelines for Human Genome/Gene Research enacted by the Japanese Government and the Helsinki Declaration. All the patients provided signed informed consent approving the use of their tissues for unspecified research purposes. For all experiments involving mice, the animals were housed in laminar-flow cabinets under specific pathogen-free conditions in facilities approved by Kyushu University (approval number, A24-112-0). ## Patients and pancreatic tissues Pancreatic cancer tissues were obtained from 105 patients who underwent pancreatic resection for pancreatic cancer at our institution. The clinicopathologic characteristics of the patients are described in. Survival was measured from the time of pancreatic resection until death. Prognosis was determined in September 2011. The median overall survival time was 23.5 months (range, 1–114 months). Sixty-two patients died during follow-up. All tissues adjacent to the specimens were evaluated histologically according to the criteria of the World Health Organization. The tumor stage was assessed according to the Union for International Cancer Control (UICC). We also obtained 31 normal pancreas (NP) samples from intact pancreas tissues resected for bile duct cancer, benign solid-pseudopapillary tumor or neuroendocrine tumor for use as control tissues. We further collected 10 pancreatic intraepithelial neoplasia (PanIN) and 39 intraductal papillary mucinous neoplasm (IPMN) samples. ## Immunohistochemical procedures and evaluation Immunohistochemical staining was performed using a Histofine SAB-PO Kit (Nichirei, Tokyo, Japan). Tissues were sectioned to a thickness of 4 µm and were incubated with mouse monoclonal anti-CD271 antibody (1∶100; Santa Cruz Biotechnology, Santa Cruz, CA) or mouse monoclonal anti-α-SMA antibody (1∶50; Dako, Glostrup, Denmark) overnight at 4°C. Cells were considered to be positively immunostained when the membrane or cytoplasm was stained. We identified activated PSCs based on the cell morphology (spindle-shaped cells) and their identities were confirmed by α-SMA staining. We counted the number of cells in at least 10 fields per section at 200× magnification. To account for the heterogeneity in CD271 expression, the distribution of CD271 immunostaining was evaluated as percentages of the stained cells and scored as follows: 0, 0%; 1, \<25%; 2, 26–50%; 3, 51–75%; or 4, 76–100%. Similarly, the staining intensity was scored as follows: 0, no staining; 1, weak staining; 2, moderate staining; or 3, strong staining. Finally, we calculated the staining score by multiplying the percentage score by the intensity score. In the pancreatic cancer cells, we divided the samples into high staining \>3 and low staining \<2 groups. All slides were evaluated independently by two investigators without any knowledge of the clinical features of each case. ## Cells and culture conditions Human PSCs were isolated from fresh pancreatic surgical specimens using the outgrowth method in our laboratory. The PSC cell type was confirmed by morphology (stellate-like or spindle-shaped cells) and by immunofluorescence staining for α-SMA and vimentin. PSC1 and PSC3-9 were isolated from pancreatic cancer. PSC2 and PSC10 were isolated from benign pancreatic cyst. PSC11 and PSC12 were isolated from normal pancreatic area of pancreatic cancer tissues. In addition, we evaluated two pancreatic cancer cell lines, SUIT-2 (Health Science Research Resources Bank, Osaka, Japan) and Capan-2 (American Type Culture Collection, Manassas, VA). SK-N-MC, a human neuroepithelioma cell line (American Type Culture Collection), was purchased as a positive control for CD271 expression. Cells at passages 3 to 8 were used for assays. The cells were maintained as previously described. ## Immunofluorescence staining and Laser-Scanning confocal microscopy PSCs (1×10<sup>5</sup>) were plated on glass-bottom dishes (Matsunami, Osaka, Japan) and incubated for 24 hours. In cultures with supernatant, cells were incubated in 10% fetal bovine serum (FBS)/DMEM with supernatant derived from pancreatic cancer cells for 72 hours. The cells were then fixed with methanol, blocked with 3% bovine serum albumin in phosphate-buffered saline solution (PBS), and incubated with a mouse monoclonal anti-α-SMA antibody (1∶50; Dako), a rabbit monoclonal anti-vimentin antibody (1∶50; Cell Signaling Technology, Beverly, MA) and a mouse monoclonal anti-CD271 antibody (1∶50; Santa Cruz Biotechnology) overnight at 4°C. Subsequently, the cells were incubated with Alexa 488-conjugated anti-mouse IgG or 546-conjugated anti-rabbit IgG (Molecular Probes, Eugene, OR) for 1 hour. Nuclear DNA was counterstained with 4′,6-diamidino-2-phenylindole (0.05 µg/mL). A laser-scanning confocal fluorescent microscope (A1R; Nicon, Tokyo, Japan) was used, and images were managed using NIS-Elements software (Nikon). ## Flow cytometry analysis Cultured cells were obtained from subconfluent monolayer cultures and incubated with a fluorescein isothiocyanate (FITC)-conjugated anti-CD271 antibody (Miltenyi Biotec, Auburn, CA). Mouse IgG1 K Isotype Control FITC (Miltenyi Biotec) was used as a negative control. The labeled cells were analyzed by flow cytometry using a FACS Calibur (Becton Dickson and Company, Franklin Lakes, NJ). ## Real-time quantitative RT-PCR (qRT-PCR) Total RNA was extracted from cultured cells using a High Pure RNA Isolation Kit (Roche Diagnostics, Mannheim, Germany) and DNase I (Roche Diagnostics). Real- time qRT-PCR was performed using a QuantiTect SYBR Green Reverse Transcription- PCR Kit (Qiagen, Tokyo, Japan) and a Chromo4 Real-Time PCR Detection System (Bio-Rad Laboratories, Hercules, CA). Primers for CD271 and β-actin were purchased from Takara Bio Inc. (Tokyo, Japan). The sequences of the oligonucleotide primers used in this study were as follows: CD271, 5′-TCAGTGGCATGGCTCCAGTC-3′ (forward) and 5′-GCAGTATCCAGTCTCAGCCCAAG-3′ (reverse); β-actin, 5′-TGGCACCCAGCACAATGAA-3′ (forward) and 5′-CTAAGTCATAGTCCGCCTAGAAGCA-3′ (reverse). Each reaction mixture was initially incubated at 50°C for 30 min to allow reverse transcription. PCR amplification was then initiated by incubation at 95°C for 15 min to activate the polymerase, followed by 40 cycles of 95°C for 5 s, 60°C for 20 s, and 72°C for 30 s. The expression levels of genes were calculated using a standard curve constructed using total RNA from SK-N-MC cells. The expression levels were normalized to the β-actin expression levels as an internal control, and expressed as the ratio of expression of the target gene to that of β-actin. All samples were run in triplicate. No detectable PCR products were amplified without prior reverse transcription. The accuracy and integrity of the PCR products were confirmed using an Agilent 2100 Bioanalyzer (Agilent Technologies Inc., Palo Alto, CA). ## In vitro coculture system In vitro cocultures were performed using 6-well cell culture insert companion plates and 3.0-µm cell culture inserts (Becton Dickinson Labware, Franklin Lakes, NJ) as previously described. SUIT-2 and Capan-2 cells (5×10<sup>5</sup>) were seeded separately into the upper chambers at 24 hours after seeding two types of PSCs (5×10<sup>5</sup>) into the lower chambers. ## Culture with a cancer cell-derived supernatant SUIT-2 and Capan-2 cells were cultured under FBS-free conditions for 48 hours. Thereafter, we collected the supernatants from the cancer cells and adjusted them to 10% FBS. The cancer cell-derived supernatants were then added to previously seeded PSCs for subsequent culture. ## Functional separation by Matrigel invasion assay Six-well cell culture insert companion plates and 8.0-µm cell culture inserts (Becton Dickinson Labware) were coated with Matrigel (150 µg/well; BD Biosciences, Bedford, MA). PSCs (5×10<sup>5</sup> cells/2 mL) were seeded into the upper chambers. Previously, cancer cells (5×10<sup>5</sup>) were seeded into the lower chambers. Thereafter, the cells were cultured for 72 hours. Subsequently, we removed the cells from both sides of the upper chambers using trypsin. After two centrifugations, we extracted mRNA from the cells. Each experiment was carried out in triplicate wells, and independent experiments were repeated three times. ## In vivo experiments SUIT-2 pancreatic cancer cells and two types of PSCs were used for in vivo experiments. Cells were divided into three groups, comprising SUIT-2 cells alone, SUIT-2 cells with PSC1 cells, and SUIT-2 cells with PSC2 cells. SUIT-2 cells (5×10<sup>5</sup>) and PSCs (5×10<sup>5</sup>) suspended in 100 µL of PBS were cotransplanted into the pancreatic tail of 6-week-old female SCID mice (FOX CHASE SCID®, C.B-17/lcr-scid/scidJcl; Clea Japan Inc., Tokyo, Japan). Six mice were used in each group. The tumors were resected on days 8, 15, and 22 after implantation. The tissues were fixed in 10% neutral-buffered formalin and embedded in paraffin. Four µm tissue sections were stained with a mouse monoclonal anti-α-SMA antibody (Dako) and a rabbit monoclonal anti-CD271 antibody (Millipore, Billerica, MA). ## Statistical analysis Values are expressed as means ± SD. Comparisons between two groups were performed using Student's t-test. Statistical significance was defined as values of p\<0.05. A χ<sup>2</sup> test was used to analyze the correlations between CD271 expression and the clinicopathologic characteristics observed in the immunohistochemical analyses. Survival analyses were undertaken using the Kaplan–Meier method, and the curves were compared using the log-rank test. To evaluate independent prognostic factors associated with survival, a multivariate Cox proportional-hazards regression analysis was used, with CD271 expression, pT category, lymph node metastasis, UICC stage, perilymphatic invasion, perivascular invasion, and pathological margin as covariates. All statistical analyses were performed using JMP 8.0 software (SAS Institute, Cary, NC). # Results ## Stromal cells express CD271 in pancreatic ductal adenocarcinoma (PDAC) To evaluate CD271 expression in pancreatic tissues, we performed immunohistochemistry for CD271. Consistent with a previous report, we observed strong staining of nerves as a positive control. The epithelial cells of normal pancreatic ducts showed no expression of CD271, and PDAC cells were unstained. In normal pancreas (NP), stromal cells around the pancreatic duct were rarely stained. However, we observed that stromal cells were strongly stained around pancreatic intraepithelial neoplasia (PanIN) and intraductal papillary mucinous neoplasm (IPMN). In PDAC tumors, stromal cells around pancreatic cancer cells were partly stained. However, CD271<sup>+</sup> stromal cells were not adjacent to cancer cells in PDAC, and were strongly stained at the edge rather than the center of the tumors. CD271 high staining in stromal cells was present in 6.5% (2/31) of NP, 30.0% (3/10) of PanIN, 27.6% (29/105) of PDAC and 84.6% (33/39) of IPMN. The CD271 high staining rates were significantly higher in stromal cells of IPMN (p\<0.0001) and PDAC (p = 0.0069) than those in stromal cells of NP. ## Stromal CD271 expression independently correlated with good prognosis We evaluated correlations between stromal CD271 expression and the clinicopathologic factors of PDAC. Although differences did not reach statistical significance, pT1/pT2, no lymph node metastasis, UICC stage I, no perilymphatic invasion, no perivascular invasion, and pathological margin/negative were observed more frequently in the stromal CD271high staining group than in the stromal CD271 low staining group. Interestingly, stromal CD271 expression was associated with a good prognosis in pancreatic cancer (p = 0.0040). The median survival times for CD271 high staining and CD271 low staining patients were 62 and 18.5 months, respectively. Next, we performed a multivariate survival analysis based on the Cox proportional hazard model for all parameters found to be significant by univariate analyses, including stromal CD271 expression, pT category, lymph node metastasis, UICC stage, perilymphatic invasion, perivascular invasion, and pathologic margin positivity. Stromal CD271 expression was found to be an independent prognostic marker in pancreatic cancer patients, and the relative risk of stromal CD271 expression was 0.495. ## CD271 stromal cells are a subpopulation of activated PSCs To characterize the CD271<sup>+</sup> stromal cells, we performed immunohistochemistry for α-SMA, a marker for activated PSCs, on serial PDAC sections. We found that α-SMA was expressed in most stromal cells around cancer cells and neoplastic tubules. CD271<sup>+</sup> stromal cells were restricted to areas with strong α-SMA expression. These findings indicate that CD271<sup>+</sup> stromal cells are a subpopulation of activated PSCs. ## CD271 expression in human PSCs We isolated five primary cultures of PSCs from fresh human pancreatic tissues, and their identity was confirmed by immunofluorescence staining. PSCs were stellate-like or spindle-shaped and expressed α-SMA and vimentin. In these immunofluorescence analyses, we did not detect CD271 expression in PSCs (data not shown). Next, we evaluated CD271 expression in 12 primary cultures of human PSCs using flow cytometry, and found CD271-positive rates of 0.0–2.1% in PSCs. ## CD271<sup>+</sup> PSCs are increased through cancer–stroma interactions We investigated the effects of PSCs on the phenotype of cancer cells using a coculture system. We used two pancreatic cancer cell lines, SUIT-2 and Capan-2, and two primary cultures of human PSCs isolated from PDAC and benign pancreatic cysts. After monoculture or coculture with pancreatic cancer cells for 72 hours, we evaluated CD271 expression in PSCs using flow cytometry and real-time qRT- PCR. By flow cytometry analyses, we observed the percentage of CD271<sup>+</sup> cells was higher in coculture than in monoculture (0.7% vs. 0.1%). By real-time qRT-PCR analyses, the level of CD271 mRNA expression in two cultures of PSCs was significantly higher in coculture with pancreatic cancer cells than in the monoculture (p\<0.001). Next, we collected supernatant from SUIT-2 pancreatic cancer cells, and added it to two primary cultures of PSCs. We compared the levels of CD271 mRNA expression among monocultures, cultures with the cancer cell-derived supernatant, and cocultures with pancreatic cancer cells for 72 hours. The level of CD271 mRNA expression was highest in cocultures with pancreatic cancer cells. In addition, the level of CD271 mRNA expression was higher in cultures with the cancer cell-derived supernatant than in monocultures (p = 0.0053). By immunofluorescence analyses, we detected CD271 expression in PSCs was slightly increased in culture with the cancer cell-derived supernatant in contrast to monocultures. ## CD271 expression decreases after the transient increase in expression when cocultured with pancreatic cancer cells We used a coculture system consisting of two pancreatic cancer cell lines, SUIT-2 and Capan-2, and two primary cultures of human PSCs to evaluate the time- dependent changes in CD271 mRNA expression. After monoculture or coculture of PSCs with pancreatic cancer cells for 1 to 5 days, we evaluated the level of CD271 mRNA expression in PSCs. In monocultures, the expression of CD271 mRNA in PSCs did not change. However, in cocultures with pancreatic cancer cells, the level of CD271 mRNA expression increased on day 3 (p = 0.0249) and peaked at day 4 (p\<0.0001). Interestingly, the levels of CD271 mRNA expression decreased the day after peak expression (p\<0.0001). Next, we collected supernatant from Capan-2 pancreatic cancer cells, added it to two primary cultures of PSCs and evaluated the time-dependent changes in CD271 mRNA expression. In cultures with supernatant, the level of CD271 mRNA expression increased on day 2 (p\<0.0001) and then decreased on day 4 (p\<0.0001). Similar to coculture with pancreatic cancer cells, CD271 expression decreased after the transient increase in expression when cocultured with supernatant. ## CD271 expression is decreased in PSCs migrating through Matrigel toward pancreatic cancer cells Next, we evaluated CD271 expression in PSCs after functional separation using Matrigel invasion assay. In this culture system involving upper chambers coated with Matrigel, we evaluated SUIT-2 pancreatic cancer cells and two primary cultures of human PSCs. PSCs in the upper chambers were incubated for 72 hours in culture with pancreatic cancer cells in the lower chambers, or as controls without pancreatic cancer cells in the lower chambers. PSCs were collected from the top and bottom sides of the upper chambers separately. The top side-derived PSCs were cells that did not migrate toward the cancer cells through the Matrigel and remained in the upper chambers. The bottom side-derived PSCs were cells that migrated toward the pancreatic cancer cells through Matrigel and collected on the bottom side of the upper chambers. We then analyzed CD271 mRNA expression levels in the top or bottom side-derived PSCs to evaluate the functional differences between CD271<sup>+</sup> and CD271<sup>−</sup> PSCs. The level of CD271 mRNA expression was highest in top side-derived PSCs after culture with pancreatic cancer cells (p\<0.001). Interestingly, the level of CD271 mRNA expression in the bottom side-derived PSCs migrating toward the pancreatic cancer cells was lower than that in the top side-derived PSCs that did not move toward cancer cells (p\<0.01). In controls, there were no differences in CD271 mRNA expression levels between top and bottom side-derived PSCs, and these levels were low compared with the levels of CD271 mRNA expression in cultures with pancreatic cancer cells. ## CD271<sup>+</sup> PSCs are present in tumor margins/periphery and are absent in the tumor core To evaluate the localization of CD271<sup>+</sup> PSCs in vivo, we transplanted SUIT-2 pancreatic cancer cells into the pancreatic tail of SCID mice, or cotransplanted them with two primary cultures of human PSCs. We euthanized the mice on days 8, 15, or 22, and evaluated the localization of CD271<sup>+</sup> PSCs in the xenograft tumors. Stromal CD271 expression was detected at all days after transplantation, regardless of transplantation or cotransplantation. Stromal CD271 high staining rates were 50% (3/6), 60% (3/5), and 67% (4/6) on days 8, 15, and 22, respectively. In every case, we detected CD271 expression on nerve bundles as a positive control, and found that pancreatic cancer cells never expressed CD271. We confirmed the presence of activated PSCs by a spindle-shaped morphology and by α-SMA staining. Activated PSCs existed in the xenograft tumors and normal pancreatic areas around the tumors. However, CD271<sup>+</sup> PSCs only existed in the normal pancreatic areas around the xenograft tumors, and were absent in the tumor core. # Discussion Previously, Trim et al., reported that CD271 was expressed in PSCs. Haas et al., also reported that CD271 was slightly expressed in PSCs and that the level of CD271 mRNA expression rapidly decreased in primary rat PSCs during culture. We found that CD271<sup>+</sup> PSCs existed in surgical pancreatic tissues using immunohistochemistry. However, qRT-PCR and flow cytometry analyses revealed that CD271 expression in PSCs isolated from pancreatic tissues was very weak. These findings suggest that CD271 expression in primary human PSCs also rapidly decreased during culture. We found that CD271 mRNA expression was transiently increased in PSCs cocultured with pancreatic cancer cells, suggesting that pancreatic cancer cells enhance CD271 expression in PSCs. However, the present immunohistochemical analyses revealed that CD271<sup>+</sup> PSCs also existed around PanIN and IPMN, precancerous lesions. Furthermore, Trim et al., reported that CD271<sup>+</sup> PSCs were found in tissues from chronic pancreatitis patients. These observations suggest the appearance of CD271<sup>+</sup> PSCs is not specific to malignant diseases, and is related to desmoplasia. Haas et al., reported that CD271 mRNA expression was enhanced in PSCs by transforming growth factor beta (TGFβ), a strong profibrogenic activator of PSCs. Addition of TGFβ to PSCs slightly increased the level of CD271 mRNA expression in PSCs, but without a significant difference (data not shown). Although TGFβ secreted from pancreatic cancer cells may be related to CD271 expression in PSCs, it is possible that other soluble factors also affect CD271 expression. Interestingly, CD271<sup>+</sup> stromal cells were not adjacent to pancreatic cancer cells in PDAC as observed by immunohistochemical analyses. In vivo, CD271 expression decreased after the transient increase in expression in PSCs cocultured with pancreatic cancer cells. These findings suggest that CD271 expression is decreased in PSCs following long-term interactions with pancreatic cancer cells. In pancreatic cancer tissues, CD271 high staining cases have a better prognosis than CD271 low staining cases. Progression of PDAC may cause the decrease of CD271 positive rate of PSCs due to the long-term influences of pancreatic cancer cells. The functional role of CD271<sup>+</sup> PSCs remains unclear. Previous reports identified CD271 as a marker of activated PSCs. However, CD271 expression in PSCs rapidly decreased during primary culture, even though α-SMA expression in PSCs remained after several passages of primary cultures. Recently, several articles have reported that normal fibroblasts, which prevented tumor growth and invasiveness, became reprogrammed by unknown mechanisms to co-evolve with epithelial tumor cells and provide an environment conducive to tumor initiation and progression. CD271 expression may be a marker for such a reprogramming stage, suggesting that CD271 expression is a temporary marker of PSCs in the early stages of interaction with pancreatic cancer cells. Previously, we reported that CD10<sup>+</sup> PSCs enhanced the malignant progression of pancreatic cancer. CD10<sup>+</sup> PSCs existed near pancreatic cancer cells, and the existence of CD10<sup>+</sup> PSCs suggested a poor prognosis. In the present study, we found that CD10 mRNA expression continuously increased in cocultures with pancreatic cancer cells but did not decrease in PSCs migrating through Matrigel toward pancreatic cancer cells (data not shown). In contrast, CD271<sup>+</sup> PSCs existed near pancreatic tumors, but were separate from pancreatic cancer cells, rather than adjacent, and the existence of CD271<sup>+</sup> PSCs was an independent factor for good prognosis. These findings suggest that CD271<sup>+</sup> PSCs may play a defensive role in pancreatic carcinogenesis and/or progression of pancreatic cancer, and that long-term interactions with cancer cells may reduce the number of CD271<sup>+</sup> PSCs. However, CD271<sup>+</sup> PSCs were too difficult to isolate after primary cultures. Therefore, we were unable to evaluate directly the function of CD271<sup>+</sup> PSCs, and further investigations are required. In conclusion, CD271<sup>+</sup> PSCs existed near pancreatic cancer tumors, and CD271 expression correlated with a better prognosis of human pancreatic cancer. A short interaction with cancer cells transiently increased the expression of CD271 in PSCs, while a long interaction with cancer cells decreased CD271 expression in PSCs. CD271<sup>+</sup> PSCs appeared during the early stages of cancer–stroma interactions. Taken together, the present findings suggest that the CD271<sup>+</sup> subpopulation of PSCs may play a role in the resistance to pancreatic carcinogenesis and progression of pancreatic cancer. # Supporting Information The authors thank the Research Support Center, Graduate School of Medical Sciences, Kyushu University, for technical support. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: KF KO KM NI MT. Performed the experiments: KF DE SK. Analyzed the data: KF KS NI SA. Contributed reagents/materials/analysis tools: KF DE SK NI TO ST MT. Wrote the paper: KF KO KM NI MT.
# Introduction The genus Cryptosporidium is a multispecies complex with extensive genetic variation. Approximately 27 species and more than 60 genotypes have been identified. A wide diversity of Cryptosporidium spp. and subtypes infect humans, and each may have a range of transmission routes of public-health significance. DNA analysis of Cryptosporidium parasites from humans has shown the anthroponotic C. hominis and the zoonotic C. parvum to be the most common cause of human cryptosporidial infections, with 90% of reported cases attributed to them. Recently an additional 8 species have been identified as causes of cryptosporidiosis in humans, including C. meleagridis, C. felis, C. canis, C. suis, C. muris, C. fayeri, C.ubiquitum, and C. cuniculus. The contribution of these species to human cryptosporidiosis varies globally and has been documented to be associated with seasonality, demographics, immune status, and contact with reservoir hosts. Further intra-species variation has been observed in Cryptosporidium isolates which further classifies species to the subtype family and subtype levels. To date six C. hominis subtype families (Roman numeral I) and 11 C. parvum subtype families (Roman numeral II) have been identified. Six subtype families have also been identified in C. meleagridis (Roman numeral III) and C. fayeri (Roman numeral IV). Advances in molecular biology and full genome sequencing has contributed to knowledge on molecular epidemiology and better understanding of the biology and transmission of cryptosporidiosis in humans, as well as associations of different species or subtypes to clinical manifestations and infection risk factors, such as age and HIV status. In addition, information gained from such studies provides support for treatment, prevention and control strategies. Amplification and sequencing of one or more genetic loci (markers) have been used for the differentiation of Cryptosporidium species, genotypes and subtypes. In particular, a PCR-RFLP tool that targets \~830-bp fragment of the small subunit (SSU) rRNA gene and uses SspI and VspI restriction enzymes for genotyping is commonly used, due to the multi-copy nature of the gene and presence of semi-conserved and hyper-variable regions. Other genotyping tools based on the oocyst wall protein (COWP) gene have narrow specificity and only amplify DNA of C. parvum, C. hominis, C. meleagridis, and species/genotypes closely related to C. parvum hence this technique is rarely used for diverse samples. Mini- and micro-satellites, or simple sequence repeats, constitute a rich source of polymorphism and have been extensively used for high-resolution genotyping and subtyping. In particular, the GP60 gene is useful for such studies as it contains multiple regions with high mutation rates, including a ‘hyper-variable’ microsatellite region. The GP60 gene is the most polymorphic marker identified so far in the Cryptosporidium genome and exhibits extensive sequence differences in the non-repeat regions, which categorize C. parvum and C. hominis each to several subtype families. Within each subtype family, differences are attributed to the number of trinucleotide repeats (TCA, TCG or TCT microsatellite). Molecular analysis of the GP60 gene has facilitated the identification of transmission pathways and zoonotic disease contamination sources and highlighted the importance of certain genetic variants to human health, and the public health risk posed by particular Cryptosporidium subtypes. There are also significant differences in clinical presentations and virulence among some common C. hominis subtype families in cryptosporidiosis- endemic areas. Although cryptosporidiosis is prevalent in Sub Saharan Africa, information on the molecular characteristics of Cryptosporidium spp. is scant. In Kenya particularly, Cryptosporidium has been found to be a major cause of diarrhoea in children, yet information on genetic diversity of circulating Cryptosporidium spp. and subtypes in different populations is limited. The aim of this study was to determine the prevalence, epidemiology and the genetic diversity of Cryptosporidium spp. isolated from children with diarrhoea, presenting at outpatient clinics in Mukuru informal settlement, or admitted at the Mbagathi District Hospital. The understanding of this relationship may represent the starting point for further extended studies on the epidemiology and genetic diversity of Cryptosporidium in different populations, and establish routes of transmission and intervention measures. # Materials and Methods ## DNA isolation Genomic DNA was extracted from faecal specimens that were positive for Cryptosporidium from an earlier published study (Mbae et al., 2013)using QiAmp<sup>®</sup> DNA stool Mini kit (Qiagen, Crawley, West Sussex, United Kingdom) with slight modifications. Briefly 200 μl of fecal suspension was washed five times with triple-distilled water by centrifugation. To this suspension 1.4 ml of ASL buffer was added and subjected to five times thawing (80°C) and freezing (-80°C) to rupture the rigid oocysts. The genomic DNA was eluted in 50 μl of Nuclease free water and stored at -20°C until use. ## Species and genotype identification PCR Amplification targeting the Cryptosporidium 18S rRNA and restriction fragment length polymorphism (RFLP) of the PCR amplicon were used to genotype Cryptosporidium isolates essentially as described. Briefly, a two stage nested PCR of the 18S rRNA gene was done; the primary PCR amplified a 1,325 bp, using forward primer `5’-TTCTAGAGCTAATACATGCG-3’` and reverse primer `5’-CCCATTTCCTTCGAAACAGGA-3’`. The illustra<sup>™</sup> PuReTaq Ready-To-Go PCR Beads (GE Healthcare, UK) was used for amplification in a 25 μl final volume, to the beads (2.5 units of puRe Taq DNA polymerase, 200 μM of each dNTP, 10 mM Tris-HCl pH 9.0, 50 mM KCl, 1.5 mM MgCl<sub>2</sub>), 0.25 μM of each primer, an extra 1.5 mM MgCl<sub>2</sub>, 21.25 μl of nuclease free water and 1.0 μl of DNA were added. The secondary amplification of an internal fragment of between 826–864 bp was done using forward primer`5’-GGAAGGGTTGTATTTATTAGATAAAG-3’`and reverse primer `5’-CTCATAAGGTGCTGAAGGAGTA-3’`. The same conditions described for primary PCR were used for the secondary PCR reaction except 0.5 μM of each primer, 20 μl of nuclease free water and 1 μl of the primary PCR product were added to the beads. Two step restriction digestion of the secondary PCR productswas carried in a total reaction volume of 40 μl. The reaction mix composed of; 15 μl of the secondary PCR amplicon as the template, 2 μl (20 units) of SspI (Promega) and 4 μl of 1 X restriction buffer. The second set of restriction digestion using VspI (Promega) enzyme was also performed in a total reaction volume of 40 μl, consisting of 1 μl (10–12 units) of the enzyme, with similar volumes of buffer and PCR amplicons as the SspI reaction. Digestion was carried out in a water bath at 37°C for a minimum of 4 hours. The digestion products were resolved on 2% agarose gel stained with ethidium bromide and visualized under UV. Fragment sizes were then compared with the expected banding patterns from known Cryptosporidium species and genotypes as previously published for species/genotype identification. DNA sequencing was carried out only on isolates that did not give distinct bands for confirmation. ## C. parvum and C. hominis subtyping Subtyping of C. parvum andC. hominis was carried out by a GP60-based tool, which amplified an\~850 bp fragment of the GP60 gene, and the sequences obtained analysed for identification of subtype families and subtypes, as described previously. In the primary PCR forward primer (AL3531 `5’-ATAGTCTCCGCTGTATTC-3’`) and reverse primer (AL3535 `5’-GAGATATATCTTGGTGCG-3’`) were used, while the forward primer (AL3532 `5’-TCCGCTGTATTCTCAGCC-3’`) and reverse primer (AL3534 `5’-GCAGAGGAACCAGCATC-3’`) were used for the nested PCR. The same conditions described for 18rRNA PCR were used for GP60 semi-nested PCR. Reaction mixtures containing the correct size fragment were purified using QIAquick PCR purification kit (Qiagen, Crawley, West Sussex, United Kingdom) according to the manufacturer's protocol. ## Sequencing of GP60 PCR products The purified nested PCR amplicons of Cryptosporidium isolates of GP60 gene, were sequenced in both directions using forward primer AL3532 and reverse primer AL3534. Bidirectional sequencing of DNA samples using ABI3730 and Big Dye terminator v3.1 kit was carried out at the International Livestock Research Institute, BeCA laboratories, Nairobi, Kenya. Base calling for each sequence run was done using Sequence Analysis v5.2 software Sequences received were first edited and consensus sequence generated from the forward and reverse sequence using CLC DNA workbench 6.1. Each consensus sequence from individual isolates was used for the identification of Cryptosporidium GP60 subtypes. Basic local alignment search tool (BLAST) ([www.ncbi.nlm.nih.gov/blast](http://www.ncbi.nlm.nih.gov/blast)) was used to assess identity and degrees of similarities with CryptosporidiumGP60 subtypes in the GenBank as well as compared with other reference sequences of C. parvum and C. hominis subtypes from different geographical regions. Multiple sequence alignment of Cryptosporidium isolates with GenBank reference sequences of C. parvum and C. hominis subtypes were aligned using ClustalX 2.1. The phylogenetic analysis of the GP60 gene nucleotide sequence data were conducted using the software package MEGA vs.5.05. The evolutionary relationship was inferred using the Neighbor-Joining method. Branches that had less than 50% bootstrap value were collapsed. The tree was rooted using C. meleagridis (Accession No. AF401499) as the outgroup. The previously established nomenclature system was used to differentiate subtype families and subtypes within each species. All the GP60Cryptosporidium sequences were analysed for ‘‘TCA” microsatellite region. GP60 sub genotype results analysis display high mutation rates, in particular, a ‘‘hyper-variable” microsatellite region. The GP60 sub-genotype ‘‘TCA” micro satellite region, showed triplet cordons were categorized according to the number of trinucleotide repeats coding for the amino acid serine. Cryptosporidium GP60subtypes consist of a variable number of ‘‘A” (TCA), ‘‘G” (TCG), ‘‘T” (TCT) and ‘‘R” (ACATCA). The PCR-RFLP of the 18S rRNA gene and sequence analysis of the GP60gene locus provided information on the genetic inter-relationships of the Cryptosporidium spp. in the study population. ## Statistical analysis Data from the questionnaires was entered into a Microsoft access database using EpiInfo<sup>™</sup> version 3.3 (CDC, 2004). Data cleaning procedures were performed before importing data for analysis into Stata 9.2 (Stata Corporation, Texas USA) for analysis. Frequencies and proportions for patients’ characteristics categorized by cryptosporidium species, families and subtypes were calculated and reported. Univariate analyses were used to identify potential patients’ characteristics, clinical symptoms and seasonality correlates of infection with particular cryptosporidium species and assessed for significance to determine suitability for multivariate logistic regression analyses. Odds ratios were used to describe associations and a p-value of \<0.05 was considered significant. ## Ethical approval The study was approved by the Kenya National Ethical Review Committee. All guardians of participating children were informed of the study objectives and voluntary consent was sought before inclusion. # Results ## Genotyping Cryptosporidium from faecal samples A total of 151 out of 187 samples positive by microscopy from a previously published study. were genotyped. Thirty six samples that were positive by microscopy did not give an amplication by 18S rRNA PCR. Results from PCR-RFLP analysis of 18S rRNA PCR revealed banding patterns distinctive of four different species. Majority of the isolates;125/151(82.78%),were identified as C. hominis, making it the most predominant species identified in the study population. Among these, 71(56.8%) were outpatients and 54 (43.2%) inpatients. C. parvum was identified in 18/151(11.92%), 6 (33.3%) of which were from outpatients and 12 (66.67%) from inpatients. Among the other samples, 4/151(2.64%) were identified as C. felis. Three (75%) of C. felis positive samples were isolated from inpatients and 1(25%) from outpatients. Lastly, C. meleagridis was seen in 2/151(1.32%) of the samples analysed, and these came from children recruited at the outpatient clinics. Mixed infection with C. parvum and C. hominis was seen from an outpatient. Occurrence of these species in relation to patient type (inpatients and outpatients) is shown in. ## Sub-typing C. parvum and C. hominis using GP60 based tool A total of 101 isolates were successfully amplified using the GP60 gene target, and the nested PCR amplicons sequenced. The remaining samples did not give amplification with the GP60 PCR, while one sample that had not amplified with 18S rRNA PCR amplified with GP60 PCR. Of these, 82/101 (81.2%) were identified as C. hominis, while 19/101 (18.8%) were C. parvum. Age-specific distribution of cryptosporidiosis due to the two species showed the highest prevalence in the 0–24 months -old age group, with all children except one (18/19, 94.8%), infected with C. parvum belonging to this age-group. The odd one was isolated in an older child in the 49–60 months-old age group. With C. hominis, highest prevalence was observed in 0–12 months age group (45.7%), followed by 13–24 months age-group (33.3%), while 13.6% were isolated in 25–36 month age-group. Only 3 isolates were from those over 36 months old. There were more patients infected with C. parvum among the inpatients (63.2%), while equal numbers were infected with C. parvum among HIV infected and HIV uninfected children, with 9 cases in each group. Prevalence of C. hominis was 53.7% and 46.3% among outpatients and inpatients, respectively, while more HIV negative children were infected with C. hominis as compared with HIV positive children (68.4% vs. 31.6%) respectively. However the difference in these groups was not significant. Associated clinical manifestations varied between the different Cryptosporidium spp. Generally, C. hominis infections were associated with more diverse and severe clinical manifestations of fever, vomiting, abdominal pains and acute diarrhoea. Vomiting was observed in 84% of those infected with C. hominis and 68.4% of C. parvum infections, while fever was reported among the 72% of C. hominis and 53% of C. parvum infections. More children infected with C. hominis presented with chronic diarrhoea than those infected with C. parvum: 35% versus 26%, respectively. More cases of cryptosporidiosis due to C. hominis were identified during the wet season compared to dry season (80% vs.81.6%), and the same pattern was observed with C. parvum infections. These differences were not however significant (73.4% vs. 26%). ## Microsatellite analysis of CryptosporidiumGP60 subtypes and subtype families Phylogenetic relationships based onGP60 nucleotides of the C. hominis sequences clustered isolates from this study into five distinct subtype families and one subtype family of C. parvum. The C. hominis subtype families identified among 82 isolates were; Ia, (n = 5, 6.1%), Ib, (n = 20, 24.4%), Id, (n = 31, 37.8%), 1e, (n = 23, 28%) and If (n = 3, 3.7%). Among subtype family Ia, four subtypes were found: IaA7R1 (in two cases), IaA25R5 (in one case), IaA27R3 (in one case) and IaA30R3 (in one case). Within subtype family Ib, only two subtypes were identified with subtype IbA9G3 being the most common, present in 17 cases and 1bA9G3R2 which was present in 3 cases. Subtype family Id showed the highest diversity of C. hominis with 10 subtypes. These included IdA22, being the most common in 14 cases, IdA25 in 5 cases, IdA24 and IdA15G1 in 3 cases each while subtypes IdA19, IdA21, IdA20, IdA18, IdA17G1 and IdA23G1 were identified in one case each. Subtype family Ie was the least diverse with subtype Ie11G3T3R1 in 19 samples, while 4 samples belonged to subtype IeA11G3T3. Subtype family If had 3 subtypes, comprising IfA19G1, IfA14G1, IfA12G1and these were identified in one case each. Nucleotide sequence of GP60 in C. parvum recognized the presence of a single subtype family, IIc from positive patients and all belonged to subtype IIcA5G3R2. C. hominis distribution across age groups was indiscriminate, with subtype families Ia, Ib, Id and Ie more frequently isolated in younger children than older ones. Isolation of the three cases of subtype If were confined to children within age group 0–12 months- old category. Subtype family Ia was isolated from HIV positive outpatients with a single isolate from HIV negative patient. Subtypes IA25R5, IaA27R3 and IaA27R1 were all from HIV negative outpatients, while IaA30R3 was identified in an outpatient HIV infected child. The most common subtype identified, IbA9G3, was present in four HIV infected outpatients and two HIV infected inpatients, with the rest being among HIV negative children. Two of the subtypes IbA9G3R2 were detected in HIV positive inpatients and 1 in HIV negative outpatient. The most common subtype within subtype family Id, i.e. IdA22 was identified in 6 HIV negative inpatients, 4 HIV negative outpatients, 3 HIV positive inpatients and 1 HIV positive outpatient. Almost equal numbers were infected with subtype family Ie at the outpatient and inpatient settings (11cases and 12 cases respectively). Two out of the three subtype family If cases were in inpatients who were HIV negative and one was from an HIV positive outpatient Detailed distribution of subtype families isolated based on HIV status, patient type, age and other factors is shown in. Patterns of clinical manifestations also varied among C. hominis subtype families. All the 5 subtype Ia infections were associated with chronic diarrhoea with none associated with acute diarrhoea, while all the other subtypes were associated with acute diarrhoea. Notably 3 subtype If cases were associated acute diarrhoea. All the subtype families except Ie, showed similar patterns of distribution with vomiting, abdominal pains and fever. ## Phylogenetic relationships of C. hominis and C. parvumbased on GP60 subtyping tool Phylogenetic analysis of the GP60 gene, from selected Cryptosporidium isolates showed two distinct clades; clade 1 and 2. Clade 1, was composed of, four main sub-clades i –iv. Sub-clade i was composed of an admixture of subtype (Ib, If, Ie, Ia, Id), from Kenyan isolates. Sub-clade ii, contained the subtype Ib, with Kenya isolates clustering very closely with isolates from Australia. The sub-clade III, was composed of subtype Ie with the Kenyan isolates comparable with Nigeria isolates from retrieved from GenBank. Sub-clade IV was composed of isolates from the subtype If, that clustered with South Africa isolates retrieved from the GenBank. The second major clade 2, was composed of two sub-clades, i and ii, that were contained subtype families Id and Ia respectively. Multiple sequence alignment of Cryptosporidium subtype Id against references Genbank sequences, indicated single point mutations/substitution within the sequences. Representative phylogenetic relationship within the C. parvum isolates had two distinct clades. Clade I was composed mainly of Kenyan isolates, that clustered with bootstrap values of \>50% with reference isolates from Australia and India and Uganda, with Clade II composed of distinct C. parvum isolates from Kenya, that did not cluster with any of the reference samples from the Genbank. The accession numbers of the reference sequences are indicated on the dendrograms. # Discussion The present study produced data on the importance of Cryptosporidium spp., and the predominant species and subtypes circulating in this patient population of children presenting with diarrhoea in an informal settlement of Nairobi. The results also show the epidemiologic heterogeneity of Cryptosporidium spp. in the population. PCR-RFLP analysis demonstrated the existence of at least four species of Cryptosporidium in the study population; C. hominis, C. parvum, C. felis and C. meleagridis. Of these four species, C. hominis which is almost exclusively a human parasitewas the most common with a prevalence of 82.7%. The other three are zoonotic species but are commonly associated with human cryptosporidiosis. Previous studies on the prevalence of Cryptosporidium species and genotypes infecting children in Kenya also found that C. hominis was the dominant species, and C. parvum, C. meleagridis, and C. muris were identified in HIV-infected persons. More outpatients (54%) than inpatients (46%) were infected with C. hominis, while more inpatients (63%) than outpatients (37%) were infected with C. parvum. This may imply that C. parvum caused more severe disease than C. hominis. This agrees with findings by Cama et al., (2007), where C. parvum was seen to be more pathogenic. There was equal number of HIV infected and uninfected children found to have C. parvum, while more HIV negative children were infected with C. hominis. This study involved children living in the urban informal settlement of Nairobi, where they live in overcrowded rooms and belong to low socioeconomic classes with poor hygiene and sanitation, but may have minimal direct contact with animals, hence the few zoonotic species. However, the presence of these few zoonotic species including C. felis, and C. meleagridis indicates that animal reservoirs are still important. The distribution of Cryptosporidium species and genotypes in a population is an indication of the potential infection source, thus, person-to-person transmission probably played an important role in cryptosporidiosis epidemiology in the children in this study. On the other hand, it is difficult to confirm with certainty that the zoonotic species of Cryptosporidium were transmitted directly from animal to child, rather than via contamination of water, food, or hands with animal or human faeces. These findings are consistent with results from previous studies in paediatric populations in Africa and other developing countries, where 79–90% of infections are caused by C. hominis. In Malaysia and Kuwait, the same four species were identified, however in their study, C. parvum was the most frequently detected species followed by C. hominis. Cryptosporidium felis is one of the five most common Cryptosporidium spp. that are responsible for human cryptosporidiosis. C.felis was first described in humans in Kenya by Gatei et al., (2006b) in two children, and this is the second study to report C. felis in children in Kenya. C. meleagridis, one of the zoonotic species identified in this study has previously been recognized as an important human pathogen in Africa (Kenya), Peru, India and Thailand. C. muris and C. canis identified in an earlier study by were not among the species isolated in this study. However, an earlier report of possible asymptomatic C. muris infection in healthy persons and in an immunocompromised patient by suggest that this may be yet another Cryptosporidium species with a zoonotic potential. Given that C. parvum and C. hominis constituted 88% of the Cryptosporidium infections in the paediatric population studied here, a further evaluation of the genetic variation within each of these two species was carried out using partial gp40/15 or GP60subtyping tool, which is currently, the most common locus for identifying Cryptosporidium subtype families and subtypes. Understanding the subtypes of C. hominis and C. parvum may provide clues into the mechanisms of transmission and infection of these organisms and lay foundations for effective prevention and treatment strategies. The GP60 analysis of the analysed isolates in this study revealed the high genetic variation of GP60 subtype families and subtypes in the study population, in which 5 different subtype families and 21 different subtypes within C. hominis were identified. These included subtype families Ia, Ib, Id, Ie and If, and each had different subtypes. Similar subtype families were reported in children in Uganda, except that they did not detect any subtype family If. In S. Africa and Sweden, the same range of C.hominis subtype families have been recently reported. Four different subtypes within subtype family Ia were reported. A study in Indian urban children identified subtype family Ia to be the most common. Although less frequently reported than C. hominis Ib, subtype C. hominis Ia is more genetically diverse at the sub-genotypic level. However, only four subtypes were identified in the study population, which included IaA25R5, IaA27R3, IaA30R3 and IaA7R1. While 3 of the subtypes (IaA27R3, IaA30R3, IaA7R1) have been reported in various parts of the world, subtype IaA25R5 has not been published previously. Worldwide, IbA9G3 and IbA10G2 are the two common subtypes within the Ib subtype family. IbA9G3, which is one of the subtypes in this study is commonly seen in humans in Malawi, India, and has been isolated in children and in baboons (Li et al., 2011). Subtype IbA10G2, commonly seen in South Africa, Botswana, and European countries, and isolated in HIV infected in Jamaica, and known to be responsible for more than half of the waterborne outbreaks of gastroenteritis was not isolated in this study. However subtype IbA9G3R2 which was present in 3 cases has not been reported in other areas before. One of the striking findings in our study is the predominance of, and high genetic diversity of subtype family Id in the study population; 31(31%) C. hominis isolates were identified as subtype Id. This is much higher than prevalence reported earlier, as reviewed by Jex and Gasser, (2010). Subtype IdA22 was the most common as also observed previously by Gatei (2006a or b). In the present study, a relatively rare subtype IdA21 that has only been reported from South Africa, Jordan and China was detected. Subtype IdA24, isolated in 3 children has previously been reported in Kenyan children, in the US, in a waterborne outbreak of gastroenteritis and in Jordan. Within subtype family Id, subtype IdA15G1, also reported in this study, is usually the most commonly reported. On the other hand, to the best of our knowledge, subtype IdA23G1, which was isolated from an inpatient has not been reported previously. Therefore, the diversity of Id subtype populations might indicate the presence of a unique C. hominis genotype transmission in Kenya. Subtype family Ie is presently of low genetic diversity with only 3 subtypes reported. The C. hominis Ie subtype identified, IeA11G3T3, has been reported in human infections from other developing countries, such as Nigeria, and the predominance of subtype IeA11G3T3 in our study agrees with findings by Hira et al., (2011). This subtype was also previously isolated from HIV infected persons in Jamaica and Kenya. Likewise, most infections with subtype Ie in humans are caused by IeA11G3T3, but our study identified a less common subtype, Ie11G3T3R1. The 3 subtypes within the If subtype family reported in this study were; IfA19G1, IfA14G1 and IfA12G1. To the best of our knowledge, this is the first report of subtype If in East Africa and second in humans in Africa after it was detected in South Africa by Samra et al., 2012. However one of the subtypes IfA12G2, although not found in this study, has recently been reported in baboons in Kenya. The genetic diversity within the subtype family If differs from the subtypes isolated in Bangladesh where all the 11 isolates were identified as IfA13G1, but in agreement with the study in South African children where two of the subtypes in our study; IfA14G1 and IfA12G1were reported. However subtype IfA19G1, reported in this study, has not previously been reported in Africa. All 19 C. parvum samples in our study were of the IIc subtype family. The C. parvum allelic family IIc has been frequently recorded and described almost exclusively in humans. Our findings are similar to those observed in Nigeria by Molloy, and Cama, where IIc was the most common subtype family of C. parvum. Subtype families IIa and IId which are commonly reported in different parts of the world and are zoonotic were not identified in the present study. Variation within clades of the same subtype family could be attributed to repeat motif difference; however there was lack of distinct clustering based on either HIV status or patient type (inpatient and outpatient). The most dominant C. hominis subtype isolated from HIV positive patients was Id (n = 7) and Ib (n = 7) with least genotype Ia and If (n = 1 each). We observed differences in clinical manifestations among subtype families of C. hominis: morepersons infected with subtype families Ib, Id and Ie were observed to present with fever, abdominal pains, vomiting and acute diarrhoea, whereas infections with subtype family Ia none had acute diarrhoea. This finding is supported by an earlier study that showed that subtype Id was more virulent than other C. hominis subtype families in Peruvian HIV-positive people, but differs with another study that reported that subtype family Ib may be more pathogenic than Ia, Id, and Ie due to its significant association with diarrhoea, nausea, vomiting and general malaise. On the other hand, the Peru study indicated that infections with the Ia and Ie subtype families were more likely asymptomatic. These results demonstrated that different Cryptosporidium subtypes and subtype families may be linked to different clinical manifestations. Indeed, further detailed investigations are warranted in order to improve understanding of these associations. In conclusion, we demonstrate high genetic diversity of C. hominisGP60 subtype families, and children in the study area may have different clinical responses to infections with different C. hominis subtype families. Considering that C. hominis was the predominant species in the study population confirms that transmission in children in this area is predominantly anthroponotic. # Supporting Information We would like to acknowledge financial support from Wellcome Trust programme, Kilifi (strategic award, 084538) and the National Council of Science, Technology and Innovation (NACOSTI). The children of Mukuru/Mbagathi and their parents/guardians who participated in the study. Special thanks to the entire field, clinical and laboratory staff of the Kenya Medical Research Institute, Mukuru clinics and Mbagathi District hospital, involved in collection of all the data used in this project. This article is published with permission from the Director, KEMRI. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CM SK AW. Performed the experiments: CM EM. Analyzed the data: AW CM BN JW. Contributed reagents/materials/analysis tools: SK CM AW JW BN. Wrote the paper: CM. Planning, coordination and supervision of data collection in the field; data entry and cleaning: CM. Revised the manuscript: SK EM JW.
# Introduction B cell Chronic Lymphocytic Leukemia (CLL) is an adult leukemia characterized by the accumulation of B cells in the blood, bone marrow and secondary lymphoid tissues due to apparent survival advantages and/or apoptosis resistance of these cells in vivo. There is significant heterogeneity in the disease progression between CLL patients. A more aggressive form of the disease, which results in lower patient survival time, correlates with markers including unmutated immunoglobulin heavy chain variable region (IgHV) status and high expression of the tyrosine kinase ZAP-70 (ZAP-70+). Although the accumulation of a monoclonal population of CD5+/CD19+ B cells is characteristic of both prognostic groups, aggressive CLL appears to have some distinct characteristics and signaling properties compared to indolent CLL. Despite their enhanced survival in vivo, when CLL cells from patients are cultured in vitro, they rapidly undergo apoptosis under conditions that support the survival of normal B cells, underscoring the dependence of these cells on survival cues from the microenvironment. In the microenvironment, marrow stromal cells are believed to secrete factors that promote CLL cell survival in patients; correspondingly, when monocytes isolated from peripheral blood of CLL patients are cultured, they differentiate into “Nurse-like cells” (NLCs) that promote CLL survival in vitro. One of the factors known to be secreted by these NLCs and to support CLL survival, is the chemokine, CXCL12 (SDF-1). Additionally, CXCR4, the receptor for CXCL12, is overexpressed on CLL cells compared to normal B cells, and thus has the potential for enhanced responsiveness to CXCL12 signaling. Although another chemokine receptor, CXCR7, can also bind CXCL12 and was previously shown to be expressed on B cells, surface expression of CXCR7 was not observed on CLL B cells. Therefore the CXCL12 signaling effects are likely mediated exclusively by CXCR4 in these cells. While chemokines and their G-protein coupled receptors are best known for their role in directing the migration of immune cells, it is clear that these proteins are involved in many other biological functions. The CXCL12/CXCR4 axis is critical for developmental processes including lymphopoiesis, and central nervous system and cardiac development, and knockout of either the ligand or receptor in mice results in embryonic lethality. Due to the involvement of CXCL12/CXCR4 in migration, angiogenesis, and development, it is not surprising that this axis is often exploited by cancer cells for metastasis as well as survival and proliferation. However, the specific molecular mechanisms by which these various functions are effectuated and how these signaling pathways target different downstream signaling molecules in cancer cells compared to non- malignant counterpart cells is largely unknown. Similarly, while it is known that Akt and ERK1/2 are activated by CXCL12 in CLL, the downstream targets of these pathways and activation of other pathways have not been elucidated. Despite the upregulation of CXCR4 and strong Akt and ERK signaling demonstrated by CLL cells in response to CXCL12, the CLL cells actually migrate less efficiently to CXCL12 than B cells from healthy donors in a transwell migration assay. Thus, in CLL cells, it appears that signaling downstream of CXCL12/CXCR4 may be redirected towards survival signaling in lieu of cell migration. To better characterize the signaling responses to CXCL12 stimulation, primary CLL cells isolated from 5 patients were subjected to phosphoproteomic analysis by liquid chromatography and tandem mass spectrometry (LC-MS/MS). Rather than attempting to characterize the complete phosphoproteome of CLL cells, this approach was designed to generate new hypotheses about the CXCL12/CXCR4 signaling network in CLL survival, and to identify downstream proteins that might be good therapeutic targets. While many phosphoproteins were identified in the CLL cells, comparison of spectral counts between CXCL12 stimulated and unstimulated cells allowed identification of proteins phosphorylated as a consequence of CXCL12/CXCR4 signaling. With follow-up experiments, the tumor suppressor PDCD4 was validated as a downstream phosphorylation target of CXCL12 signaling in all CLL patient cells examined (n = 10) and HSP27 was similarly validated in a subset of CLL patients (∼25%). Although these proteins have been previously linked to cancer cell survival, they have not been previously associated with CLL nor has PDCD4 been established as a downstream phosphorylation target of CXCL12 signaling. Furthermore, a number of other proteins (many of which do not have commercially available phospho-specific antibodies available) have been proposed as potential downstream phosphorylation targets of CXCL12 stimulation based on spectral count analysis of the CLL phosphoproteomics data. # Methods ## Cells and reagents Peripheral blood mononuclear cells (PBMCs) were obtained from leukopheresis samples of CLL patients following written consent at the Rebecca and John Moores Cancer Center at the University of California San Diego (UCSD), in compliance with the Declaration of Helsinki. These studies were approved by the Institutional Review Board of UCSD. PBMCs were isolated by Ficoll-Paque (GE Healthcare) density gradient centrifugation as previously described. The isolated PBMCs were used fresh and cultured for phosphoproteomics analysis or frozen as liquid nitrogen stocks in 90% heat inactivated fetal bovine serum (FBS)/10% DMSO for follow-up analysis by western blot. PBMCs used in the proteomics experiments were determined to be \>90% CLL B cells as assessed by CD5+/CD19+ staining and flow cytometry analysis. For western blot validation, CLL B cells were purified from the PBMCs by negative selection using the magnetic associated cell sorting (MACS) (Miltenyi Biotec, Auburn, CA) by depletion of CD14+ (monocytes) and CD2+ (T cells) cells, leading to \>99% CLL B cell purity. Normal B cells were purified from PBMCs from healthy donors (San Diego Blood Bank) using the MACS B cell Isolation Kit II (Miltenyi Biotec, Auburn, CA) according to the manufacturer's protocol and were determined to be \>90% pure by flow analysis staining for CD19+/CD3-/CD14- cells. RPMI-1640 glutamax media and FBS were obtained from Gibco (Invitrogen, Carlsbad, CA). ## Recombinant CXCL12 preparation CXCL12 was expressed recombinantly in BL21 E. coli as previously described. In brief, CXCL12 was expressed as a His-tag fusion protein and purified from inclusion bodies. Bacterial cell pellets were sonicated and washed with deoxycholate following resuspension in 10 mM Tris pH 8.0 with 1 mM MgCl<sub>2</sub>, 200 µg DNAse, and Complete Protease Inhibitor Cocktail (EDTA- free) (Roche, Indianapolis, IN). Protein was then solubilized in 6 M Guanadine- HCl, 100 mM sodium phosphate, 10 mM Tris-Cl, pH 8.0, using a dounce homogenizer. CXCL12 was purified over a Ni-NTA column and refolded with Hampton Fold-It Buffer \#8 (Hampton Research, Aliso Viejo, CA), then dialyzed and concentrated using Amicon Ultra centrifugal concentrators (MWCO = 5000). The His-tag was removed by cleaving with enterokinase (NEB, Ipswich, MA) at a 1∶100,000 molar ratio overnight at room temperature. CXCL12 was then purified by HPLC and the identity and purity was validated by ESI mass spectrometry. Transwell migration assays on Jurkat cells were used to validate functionality of the purified CXCL12. ## Migration assays Transwell migration assays (Corning, Corning, NY) were performed on purified CLL B cells and B cells from healthy donors using inserts with a 6.5 mm diameter, 5.0 µm pore size. Cells were resuspended at 2.5×10<sup>6</sup> cells/mL in RPMI+10%FBS and 100 µL of cell suspension was added to the inserts. CXCL12 was diluted over a concentration range of 0 nM to 500 nM in a 600 µL total volume of RPMI+10%FBS in the bottom wells. As a positive control and cell count reference, cells were added directly to the wells without inserts. Transwell migration was conducted for 2 h at 37°C/5%CO<sub>2</sub>. Cells that had migrated into the bottom wells were then collected and counted by flow cytometry on a FACSCalibur (BD Biosciences, San Jose, CA). Data was normalized to no chemokine control and percent migration was calculated from the positive reference control. ## Preparation of CLL lysates for proteomics CLL cell lysates for phosphoproteomic analysis were prepared as previously described. Briefly, 3×10<sup>9</sup> total CLL PBMCs were washed with sterile PBS and resuspended at 1×10<sup>7</sup> cells/mL in serum-free RPMI-1640 media. The CLL cell suspension was distributed evenly into five 15 cm plates (6×10<sup>8</sup> cells/plate) (Corning Inc, Corning, NY) and cultured for 2 h at 37°C/5% CO<sub>2</sub> prior to stimulation with CXCL12. CLL cells were then either unstimulated or stimulated for 3 min, 10 min, 30 min, or 60 min with 30 nM CXCL12. All plates were harvested at the same time with 3 mL ice cold cytoplasmic lysis buffer containing 10 mM HEPES, pH 7.9, 1.5 mM MgCl<sub>2</sub>, 10 mM KCl, 0.5 mM dithiothreitol (DTT) (Sigma, St. Louis, MO), Complete Protease Inhibitor Cocktail (Roche Diagnostics, Indianapolis, IN), and Halt Phosphatase Inhibitor Cocktail (Pierce, Rockford, IL) for 30 min on ice. Lysates were clarified by centrifugation at 20,000 rcf for 20 min at 4°C. The supernatants were distributed into protein LoBind Eppendorf tubes (Eppendorf, Westbury, NY) and stored at –80°C. The total protein concentration of the CLL lysates was determined using a BCA protein assay (Pierce, Rockford, IL). ## IMAC phosphopeptide enrichment IMAC enrichment was performed as previously described. Briefly, 2 mg of CLL lysates were denatured with 1% sodium dodecyl sulfate (SDS) (Fisher Scientific, Pittsburgh, PA), reduced with 10 mM DTT, and alkylated with iodoacetamide (Sigma, St. Louis, MO). Proteins were then precipitated with 50% ethanol/50% acetone/0.1% acetic acid (HAC). The pellets were resuspended in 6 M urea/0.1 M Tris, pH 8.0, and vortexed to solubilize the protein. The urea concentration was then diluted five-fold by addition of 50 mM Tris, pH 8.0 and protein was digested overnight at 37°C using sequencing-grade modified trypsin (Promega, Madison, WI) at a ratio of 1∶50 (trypsin:protein). Trypsin was inactivated by acidification of the digests with trifluoroacetic acid to 0.3 to 0.5% (v/v). Prior to IMAC enrichment, peptide mixtures were desalted with 50 mg Sep-pak C18 cartridges (Waters Corp, Milford, MA). IMAC beads were prepared by stripping Ni- NTA spin column resin (Qiagen, Valencia, CA) and recharging the beads with 100 mM FeCl<sub>3</sub> (Fluka reagent, Sigma, St. Louis, MO). IMAC beads were then packed into gel loading tips with glass wool and conditioned with 25% Acetonitrile (ACN)/0.1% HAC. Nonspecific peptides were removed by washing twice with 30 µL of 25% ACN/0.1% HAC/0.1 M NaCl, twice with 0.1% HAC, and twice with 30 µL of Milli-Q H<sub>2</sub>O. Phosphopeptides were eluted with a total volume of 50 µL over three elutions with 1% phosphoric acid. All fractions were collected in protein LoBind Eppendorf tubes, speed-vac dried, and stored at –20°C until MS analysis. Analysis of IMAC washes and flow through by LC-MS/MS confirmed successful binding of phosphopeptides to the IMAC columns as less than 0.1% of these peptides were found to be phosphorylated, and the few that were detected were redundant with phosphopeptides identified in the IMAC enriched samples. ## Mass spectrometry and data processing IMAC-enriched CLL peptides were resuspended in Milli-Q H<sub>2</sub>O +0.1%HAC and analyzed by reversed-phase, C18 capillary liquid chromatography and tandem mass spectrometry (LC-MS/MS) on a Thermo-Finnigan LTQ ion trap mass spectrometer. The capillary LC columns (∼17 cm) were pulled and packed in-house using deactivated fused silica (100 µm) (Agilent, Santa Clara, CA) as previously described. Angiotensin II (Sigma-Aldrich, St. Louis, MO) was run after every two CLL sample runs as a control for column performance. The standard method used for all samples was as follows: 95% A/5% B (buffer A = 0.1% HAC in HPLC-grade Milli-Q H<sub>2</sub>O, buffer B = 0.1% HAC in HPLC-grade ACN) for 20 min, 60% A/40% B for 30 min, 20% A/80% B for 6 min, followed by a final washing step of 95% A/5% B for 30 min at 250 µl/min. A flow rate of 200 to 500 nL/min through the capillary column was achieved by splitting the flow of solvent before it reached the column. Samples were run in data-dependent mode in which the spectrometer performed one full MS scan followed by six MS/MS scans of the top six most intense ions in the parent spectrum with an m/z ranging from 400 to 2000. The dynamic exclusion list was varied in order to get a range of coverage and spectral counts. The standard dynamic exclusion list applied had a repeat count of 1, a repeat duration of 30 s, an exclusion size of 100, an exclusion duration of 180 s, and an exclusion mass width of 1.50. Other variations included a list of the top 25 peptides with a repeat duration of 60 s, an exclusion of the top 5 most abundant peptides, and a dynamic exclusion list turned off. The spray voltage was 1.8 kV. On average, the scan rate in this experiment ranged from four to eight scans per second. A comprehensive phosphoproteomics data set from cells of one particular patient (CLL A) with unmutated IgHV and ZAP-70+ status (indicative of more aggressive disease) was collected from two separate triplicate runs and one duplicate run in order to obtain good coverage and number of spectra for comparison. Phosphoproteomics analyses of 4 additional CLL patients' cells (CLL B – E) (each run in single triplicate experiments) were also performed to ensure reproducibility between different patient cells and stimulations. CLL B, D and E also had more aggressive (high ZAP-70) characteristics (high ZAP-70) while CLL C was of the indolent (low ZAP-70) subgroup. Following data collection, RAW data files were converted to mzXML data files using the program ReAdW (<http://tools.proteomecenter.org/ReAdW.php>). Data analysis of the MS/MS spectra was performed using the open-access database search tool, InsPecT, with the UniProt human database, the UniProt shuffled human decoy database as well as common contaminants databases (e.g. keratin). Peptide sequencing searches were defined for tryptic cleavage restraints and to allow for modification of up to two phosphorylation sites (Ser, Thr, or Tyr) on a peptide. Spectra were sorted by p-values according to the InsPecT scoring function and the target decoy database was used as a measure of the overall quality of MS/MS data. Peptides smaller than 7 amino acids and peptides with more than one missed cleavage were excluded from analysis. Peptides with a false discovery rate (FDR) of less than 1–2% were manually validated for positive identification. ## Western blots and antibody reagents For western blot analysis, patient CLL cells used for phosphoproteomics as well as additional patient's cells were used. Purified CLL B cells were cultured in serum-free RPMI at 1×10<sup>7</sup> cells/mL for 2 h at 37°C/5%CO<sub>2</sub> and then stimulated with 30 nM CXCL12 over an hour time course (unstimulated, 3, 10, 30 and 60 min) or for 4 h, 10 h, and 24 h for PDCD4 degradation analysis. For inhibitor studies, CLL cells were pre-treated with 40 µM AMD3100 (Sigma) or 200 ng/ml Pertussis toxin (List Biological Laboratories) for 1 h prior to stimulation with CXCL12. Coculture of CLL cells with NLCs was performed as previously described, and then the CLL cells were collected and centrifuged for harvest. Cells were lysed on ice for 30 min in Ripa buffer (10 mM Tris pH 7.4, 150 mM NaCl, 1% Triton X, 0.1% Na-Deoxycholate, 0.1%SDS and 5 mM EDTA) containing Complete Protease Inhibitor Cocktail (Roche, Palo Alto, CA) and Halt Phosphatase Inhibitors (Pierce). Lysates were clarified by centrifugation at 20,000 rcf for 10 min at 4°C. A BCA protein assay (Pierce) was performed to determine total protein concentration and 20 µg of total protein was loaded into each well of a Criterion 4–12% Bis-Tris gel and run with the XT MES buffer system (Bio-Rad, Hercules, CA). Gels were transferred onto PVDF membranes (Bio- Rad), blocked with 5% milk-TBST, and incubated overnight at 4°C with primary antibodies. Blots were washed 3 times for 10 min with Tris Buffered Saline +0.1% Tween (TBST) and then incubated for 1 h at room temperature with secondary antibodies conjugated to HRP, washed again 3 times with TBST and then developed with Amersham ECL-plus (GE-healthcare) or SuperSignal West femto-sensitivity reagent (Pierce). Blots were stripped with Restore western blot stripping solution (Pierce) for 10 min at room temperature and then re-probed with other antibodies and/or β-actin as a loading control. Primary antibodies were diluted into 5% BSA-TBST at recommended concentrations. Phospho-PDCD4 and PDCD4 antibodies were obtained from Rockland Immunochemicals. Phospho-p38, phospho- HSP27, HSP27, phospho-S6K, and β-actin were obtained from Cell Signaling Technology. Densitometry analysis was performed using ImageJ software (NIH) and normalized to β-actin loading controls. ## Flow cytometry CLL B cells or normal B cells were purified for flow cytometry analysis. Cells were washed and resuspended in a 0.5% bovine serum albumin (BSA) (Sigma) in Phosphate Buffered Saline (PBS) solution and stained for CXCR7 expression using an APC-conjugated antibody (clone 11G8) (R&D systems) or an APC-conjugated IgG1 isotype control according to manufacturer's protocol (R&D systems). Flow cytometry data was collected on a FACSCalibur cytometer (BD Biosciences) and analyzed using FlowJo software. # Results ## Normal B cells migrate with higher efficacy and potency to CXCL12 than CLL B cells, despite having lower levels of CXCR4 Chemokines, including CXCL12, are best known for their role in directing cell migration, but it is well established that they can also induce cell survival and proliferation. This observation can be rationalized by the fact that some of the major pathways involved in cell migration (e.g. PI3K/Akt and Raf/MEK/ERK), are also important for survival and proliferation signaling. However, little is known regarding the extent to which there is overlap or divergence of the upstream and downstream effectors of these pathways in the context of migration versus survival/proliferation. Since it has been established that CLL cells have up-regulated expression of CXCR4 compared to normal B cells, and that CXCL12 stimulation of CLL cells activates Akt and ERK1/2 pathways, transwell migration assays were performed on purified CLL B cells and normal B cells to compare their ability to migrate towards CXCL12. Surprisingly, the normal B cells showed a significantly greater (p\<0.0001) ability to migrate to CXCL12, with respect to both efficacy (4.5+/−1.2% migration in normal vs. 0.85+/−0.48% migration in CLL cells) and potency (∼10 nM in normal vs ∼50 nM in CLL cells). Although it was expected that the CLL cells would have the stronger migratory response due to higher CXCR4 expression, these results are consistent with previously published observations that showed weak migration of CLL cells to CXCL12 compared to a much more robust response to the CCR7 ligands, CCL19 and CCL21. These data suggest that the downstream effects of CXCR4 may be redirected for survival rather than migration, and led us to consider what other pathways or downstream targets of Akt or ERK1/2 might be activated to bias the CXCL12 signaling response towards survival. Taking a global approach to this question, we performed mass spectrometry-based phosphoproteomics analysis of unstimulated and CXCL12-stimulated CLL cells. ## Characterization of phosphopeptides/phosphoproteins in CXCL12-stimulated CLL cells via mass spectrometry Fresh PBMCs from 5 CLL patients were stimulated over an hour time course with CXCL12, and lysates were generated for IMAC enrichment and LC-MS/MS analysis. Multiple (3 separate duplicate or triplicate experiments with variable acquisition methods) phosphoproteomics data sets were acquired on lysates from the cells of a patient with ZAP-70+ aggressive CLL, referred to as “CLL A”, in order to ensure good coverage of the proteomic space. Smaller phosphoproteomics data sets (single triplicate experiments) were collected from cells of 4 additional patients (CLL B to E) for comparison. Protein lysates from the CLL cells were trypsin digested and enriched for phosphopeptides via immobilized metal affinity chromatography (IMAC) to yield a highly enriched population of phosphorylated peptides. Phosphopeptides were analyzed by LC-MS/MS using a linear iontrap (LTQ) mass spectrometer. Data was processed using the InsPecT database search algorithm and the spectra were manually inspected for validation (see flow diagram). Over 10,000 spectra were collected from the combined phosphoproteomics analysis of unstimulated as well as stimulated CLL cells, which was comprised of 1470 phosphopeptides (\>1200 unique phosphosites) from 696 phosphoproteins. In general, a \>30% enrichment of annotated phosphopeptides was achieved from our IMAC procedure, which is on par with other phosphopeptide enrichment studies of mammalian cells. To ensure that good coverage of the CLL A data set was obtained, a comparison was made of the overlapping phosphoproteins identified in the CLL A data sets and the smaller CLL B - E data sets. 538 of the 696 total phosphoproteins (\>77%) identified in the CLL A data sets were also identified in data sets B - E, suggesting detection of the majority of phosphopeptides detectable by these methods. Also, many of the non-overlapping phosphoproteins identified in CLL B to E but not in CLL A were isoform variants of phosphoproteins detected in CLL A. depicts the overlap of CLL B - E with CLL A and the matrix shows overlap between all of the CLL samples. ## Identification of phosphoproteins with prior correlations to CLL and other leukemias Adding confidence to our phosphoproteomics results, a number of hematopoietic- specific phosphoproteins as well as phosphoproteins with prior implications in CLL and other leukemias were identified. highlights 9 of the phosphoproteins detected with previous links to CLL/leukemia, including Hematopoietic cell- specific Lyn substrate (Hcls1) and SH2-containing inositol phosphatase-1 (SHIP-1). The phosphorylation status of both Hcls1 and SHIP-1 have been shown to correlate with disease aggressiveness and shorter mean survival, consistent with the aggressive characteristics of CLL A. Also included in are phosphoproteins which have been linked to CLL but not necessarily phosphorylation status, including HSP90, B cell novel protein 1, promyelocitic leukemia protein, and formin-like 1 (FMNL1),. These proteins have general implications in CLL, but whether differences in phosphorylation status affect the activity or function in disease is unclear. A few additional proteins including minichromosome maintenance protein 2 and stathmin 1 have been linked to disease progression of other leukemias, but not directly to CLL and thus warrant further investigation in CLL. Many of these phosphoproteins did not appear to exhibit changes in phosphorylation in response to CXCL12 or spectral numbers were too low to make an assessment of stimulation response, but HSP-90 and Mcm2 could be potential phosphorylation targets and are thus also highlighted as proteins of interest in. ## Identification of novel downstream targets of CXCL12/CXCR4 signaling in CLL To semi-quantitatively assess whether phosphorylation of some of the proteins is a consequence of CXCL12 stimulation, spectral counts from the mass spectrometry runs on the stimulated samples off CLL A were compared to those from unstimulated samples. Select candidate targets of CXCL12-induced phosphorylation are reported in along with their associated spectral counts. A number of known targets of survival signaling pathways including programmed cell death factor 4 (PDCD4) and heat shock protein 27 (HSP27) were identified as having more spectral counts in the stimulated versus unstimulated samples, and were selected for validation by western blot and further analysis, discussed in detail later. Validation was performed on lysates from CLL cells used in phosphoproteomics analysis (lettered CLL A to E) as well as additional patient cells not examined by phosphoproteomics (CLL1, CLL2, etc) in order to determine consistency of these responses across different patients, since CLL is a heterogeneous disease. Some of the phosphoproteins identified in this study have been previously implicated in cancer malignancy such as Mcm2 and adenylyl cyclase associated protein (CAP1), while little information is available on some of the other potential targets including small acidic protein. Although two of the proteins (PDCD4 and HSP27) are validated herein as targets of CXCL12-signaling in CLL, the remaining phosphoproteins, while beyond the scope of this work, pose interesting targets for future investigations. ## CXCL12 induces the phosphorylation and degradation of PDCD4 PDCD4 is one of the phosphoproteins that appeared to be induced by CXCL12 stimulation based on spectral counts. It is a known tumor suppressor, and downstream phosphorylation target of Akt, which is known to be activated by CXCL12 in CLL cells. A phospho-specific antibody is also commercially available, making it attractive for follow-up studies. As a tumor suppressor protein, PDCD4 has been implicated in a number of cancers where it is often inhibited and/or downregulated, disrupting its ability to inhibit eIF4A translational and AP-1 transcriptional activity, processes that are important for cell growth and survival. Phosphorylation of PDCD4 is known to occur by Akt and p70 S6Kinase (p70S6K) which inhibits its activity and leads to its ubiquitination and proteosomal degradation. Three separate phosphopeptides from PDCD4 were detected from our analysis: R.FVSpEGDGGR.V (Ser457), R.SGLTVPTSpPK.G (Ser94) and R.DSGRGDSpVSDSGSDALR.S (Ser76). The R.FVSpEGDGGR.V phosphopeptide corresponds to Ser457 phosphorylation, a site known to be phosphorylated by Akt. Therefore, multiple CLL patient samples were examined for PDCD4 phosphorylation in response to CXCL12 by western blot. An increase in phosphorylation of PDCD4 at Ser457 was observed upon CXCL12 stimulation in CLL A cells, as well as all 9 additional CLL patient cells examined (representative western blot). Increases in PDCD4 phosphorylation levels was variable between patients and ranged from 1.7-fold to 7.4-fold and averaged to approximately 3.4-fold (n = 10), as quantified by densitometry analysis of western blots. Although variability was noted, this variation did not cluster according to disease aggressiveness. As a control to ensure that the phosphorylation was dependent on CXCL12/CXCR4 signaling and to determine if the effects were dependent on signaling through the G-protein, Gi, CLL cells were pretreated with the small molecule CXCR4 antagonist, AMD3100, or the Gi-inhibitor, pertussis toxin (PTx) prior to a 3 min stimulation with CXCL12. As shown in, both AMD3100 and PTx completely abrogated phosphorylation suggesting it is CXCR4 and G-protein signaling dependent. Furthermore, to ensure that phosphorylation of PDCD4 has relevance in a more physiological context, the levels of PDCD4 phosphorylation were examined in CLL cells that had been cultured with NLCs (+) compared to those without NLCs (−). As with CXCL12 stimulation, the coculture of CLL cells with NLCs led to an increase in the phosphorylated levels of PDCD4. Since the phosphorylation of PDCD4 is known to lead to its ubiquitination and degradation, we examined levels of total PDCD4 over a 24 h time period (0, 4, 10 and 24 h) following CXCL12 stimulation. As shown in, 24 h post-stimulation resulted in PDCD4 degradation to ∼45% of starting levels. Additionally, although it is well established that Akt is phosphorylated downstream of CXCL12 signaling in CLL cells, it has not been established whether p70S6K, another kinase known to phosphorylate PDCD4 leading to its ubiquitination and degradation, is activated in CLL cells by CXCL12. Western blot analysis revealed that phosphorylation and thus activation of p70S6K (Thr389) was induced by CXCL12-stimulation in the CLL cells. ## HSP27 expression and phosphorylation is variable in CLL cells The other phosphoprotein investigated further was HSP27 (phosphopeptide: R.QLSphosSGVEIR.H, Ser82) (mass spectrum shown). HSP27 was also selected due to spectral counts indicative of phosphorylation upon CXCL12 activation of CXCR4, its implications in cancer and protection from apoptosis, and the availability of a phospho-specific antibody at the Ser82 phosphorylation site. Additionally, HSP27 is an interesting target since it is downstream of p38-MAPK signaling which has not received much attention in association with CLL and therefore represents a pathway with potentially novel implications in CLL survival. Interestingly, while PDCD4 was found to be a common target of CXCL12 signaling in all CLL samples examined, phospho-HSP27 and total HSP27 protein were only detectable by western blot in a subset of ∼25% (3 out of 12) of CLL patients examined (representative western blot). Nevertheless, HSP27 was indeed present and did exhibit an increase in phosphorylation upon CXCL12 stimulation in the CLL A patient samples from which the phosphoproteomics data was collected. Since p38-MAPK is known to be upstream of HSP27 phosphorylation, we also examined p38 phosphorylation among the CLL patient samples; correspondingly, we observed detectable p38 phosphorylation only in the samples that also exhibited the HSP27 phosphorylation (CLL A). Although no common factor could be determined among the patients examined with detectable HSP27, a larger sample size might identify common features of these cells and determine whether HSP27 is influencing the survival of this subset of CLL patients. # Discussion CLL is the most common leukemia in the Western world. The accumulation of CLL B cells is believed to result from low rates of precursor cell proliferation and via recruitment of accessory cells that create a supportive microenvironment by producing factors that foster CLL survival. The chemokine, CXCL12, is one of the cytokines produced by cells in the microenvironment that enhances CLL survival in vitro and likely in vivo. Although chemokines are best known for their role as chemoattractants, we show here that CLL cells are much less capable of migrating to CXCL12 compared to CLL B cells, despite an upregulation of CXCR4 on the CLL cells. While the low levels of migration may still play a role in vivo, it is evident that the CXCL12/CXCR4 axis is also networked into pathways involved in survival. In contrast, CXCR7, the other receptor of CXCL12, is not expressed on the surface of CLL cells although it is expressed on normal B cells. Thus, while there is overlap in signaling pathways activated by CXCL12 in CLL cells and normal B cells, the differences in migration and CXCR7 expression, and the potential bias towards survival in CLL cells, suggest significant differences in the role that the CXCL12/CXCR4 axis plays in the context of the normal and pathological cells. Phosphoproteomics analysis of CXCL12-stimulated CLL cells was performed in an effort to determine potential downstream signaling targets that could contribute to the survival and malignancy of CLL cells. As these are precious non-renewable primary patient cells, the intent of our phosphoproteomics approach was to generate hypotheses rather than an exhaustive analysis of the CLL phosphoproteome. Therefore, while it would be ideal to use a number of phospho- enrichment strategies in addition to IMAC (e.g. TiO<sub>2</sub>) and to employ additional liquid chromatography separation steps besides C18 (e.g. hydrophilic interaction liquid chromatography (HILIC)) to expand the number of phosphoproteins identified, we focused our efforts on well established methods. Along these lines, the use of quantitative phosphoproteomics strategies is limited since these cells do not replicate and cannot be cultured long term. Thus, stable isotope labeling of amino acids in cell culture (SILAC) is not possible. Post-digest labeling with iTRAQ or ICAT isotopic labels is also difficult due to limited sample availability, limitations in instrumentation (one-third rule with the LTQ spectrometer restricting detection of the labels in the low molecular weight range), and the labile nature of phosphates and labels which causes reduced fragmentation and detection in MS/MS spectra. While understanding its limitations, spectral counting was employed as a semi- quantitative assessment of the CXCL12-stimulation responses and several candidates were followed up by western blot validation. In addition to the above examples with PDCD4 and HSP27, which showed that the spectral counting provides a relatively good approximation of stimulation response, spectral counts reflecting fairly even levels of phosphorylated p21-activated kinase (PAK2), another target of the PI3K pathway, was also confirmed by western blot in all six patient cells probed for phospho-PAK2 (Ser141). Through this phosphoproteomics approach, we were able to confidently identify close to 700 phosphoproteins in the CLL samples, including numerous proteins previously implicated in CLL disease. Additionally, we identified many proteins that appear to exhibit changes in phosphorylation levels in response to CXCL12. This data led to the identification and validation of several previously unknown phosphorylation targets of CXCL12 signaling. Typical approaches (e.g. western blot) for investigating signaling in response to stimuli require a priori knowledge of specific targets and the availability of phospho-specific antibodies, which limits the ability to globally assess cellular signaling events. Furthermore, validation studies of CLL cells are difficult since their viability in culture is limited. They are also difficult to manipulate through transfection and transduction since they do not divide in culture or infect well (e.g. they require high MOI and/or pre-activation of cells with CD40L and IL-4 which could complicate interpretation of signaling analysis). Therefore, this mass-spectrometry-based approach seemed the most effective method for gaining new insight into the function of CXCL12 on CLL cell survival and possibly disease aggressiveness. Although not done in this study, comprehensive MS analysis of many patients may help to distinguish variations between patients, and with disease stratification and the identification of judicious therapeutic targets. To our knowledge, this is the first report demonstrating that PDCD4 is a phosphorylation target downstream of CXCL12 signaling in CLL or other cell types. This finding is exciting due to the established role of PDCD4 as a tumor suppressor and as a substrate of Akt. Although little is known regarding the function of two of the phosphorylation sites of PDCD4 (Ser94 and Ser76) identified from the LC-MS/MS analysis, phosphorylation at Ser457 near the C-terminus of the protein is a well-established site with known functional implications. Phosphorylation at Ser457 by Akt has been shown to result in nuclear translocation of PDCD4 and a decrease in its ability to inhibit AP-1-mediated transcription and eIF4A-mediated translation, ; although Ser67 phosphorylation was not directly identified in our LC-MS/MS analysis, it is likely that this sight is also phosphorylated in response to CXCL12 stimulation since PDCD4 degradation was observed following stimulation. In combination, these effects may reduce the growth regulating/tumor-suppressor capacity of PDCD4, thereby contributing to CLL cell survival and the malignancy phenotype. Validation of PDCD4 phosphorylation also led to the identification of p70S6K phosphorylation and activation downstream of CXCL12 signaling in CLL cells. Based on the phosphoproteomics analysis, HSP27 appeared to be another promising phosphorylation target of CXCL12-signaling. HSP27 and several other HSPs have received attention in the context of cancer due to their cytoprotective/anti- apoptotic functions. Specifically, HSP27 indirectly inhibits cytochrome c release and caspase activation and it sequesters cytosolic cytochrome c. It also promotes degradation of the inhibitor of NF-kB (IkB) and p27kip, and interacts with and supports the activity of Akt under stressful conditions, all leading to protection from apoptosis. We identified the presence of phosphorylated and total HSP27 protein and its upstream activator p38-MAPK in a subset (∼25%) of cell samples from different CLL patients. Of note, the observed variability in signaling between different CLL patient cells highlights the underlying heterogeneity of the disease. Furthermore, it serves as a reminder of how different parameters including patient differences (age and gender), variations in clinical course such as aggressiveness, stage and prognosis, and different treatments (e.g. chemotherapy, gene therapy, etc) may alter how the cells respond to different stimuli. Such variability is not unprecedented as Messmer et al (submitted manuscript) have demonstrated differences in CXCL12-mediated MEK and ERK activation in different ZAP-70 subgroups of CLL, and Montresor et al demonstrated differences in CXCL12-mediated lymphocyte function-associated antigen-1 (LFA-1) activation in normal B cells compared to CLL B cells and amongst the cells of different CLL patients. Thus, it is reasonable to expect that different patients will exhibit different responses to stimuli, whether it is a survival stimulus from the microenvironment or a therapeutic agent used to treat the disease. These results showing patient variability in HSP27 expression emphasize the strength of utilizing primary cells for understanding disease pathogenesis as opposed to cell lines, which are much more homogenous, but can be less insightful and sometimes misleading. While there was considerable overlap in the phosphoproteins identified from LC-MS/MS analysis between different patients (CLL A – E), more comprehensive MS analysis of multiple patients may help to distinguish variations in signaling responses between patients. Along these lines, since HSP27 is often induced following stressful cellular events such as treatment with chemotherapeutics, its induction in certain patients could reflect a response to treatment. For example, lymphoma cells which did not express Hsp27 were sensitive to apoptosis while those expressing Hsp27 were resistant to apoptosis induced by Bortezomib (PS-341), a proteasome inhibitor. Silencing of Hsp27 in the resistant lymphoma cells then rendered them susceptible to Bortezomib-induced death, demonstrating its link in resistance to this chemotherapeutic treatment. Thus, a larger patient sample size may reveal if expression of HSP27 is induced by certain chemotherapeutics or in particular subsets of patients and whether there is any correlation to refractory disease, as resistance to chemotherapy is one of the major hurdles in treating CLL. Herein we present follow-up data to PDCD4 and HSP27, although there are numerous other candidate phosphoprotein targets of CXCL12 signaling in CLL cells that have been proposed. A summary of our findings from phosphoproteomics analysis combined with some previously established pathways of CXCL12 signaling in CLL are summarized in a signaling diagram. Overall, our data suggests that CXCL12 may preferentially activate survival signaling pathways rather than those involved in cell migration in CLL cells, although some of the pathway components (Gi, Erk, Akt) are common nodes. We have demonstrated that the use of phosphoproteomics is a feasible and informative means of evaluating signaling responses to CXCL12 in CLL, which could be employed for investigating a variety of other stimuli in these or other primary cells. Through phosphoproteomics detection and western blot validation, PDCD4 was found to be a common phosphorylation target of CXCL12-signaling in CLL while HSP27 was present in only a subset of CLL patients. Although our focus was on CXCL12 as a survival factor, it is likely that other growth and survival stimuli may synergistically activate these pathways and downstream targets. Therefore, PDCD4 and HSP27, which have previous implications in regulation of apoptosis and carcinogenesis, may represent potential therapeutic targets for treatment of CLL. In fact, small molecule stabilizers of PDCD4, that enhance its function as a tumor suppressor by inhibiting its degradation, are currently being developed due to its potential as a therapeutic target for numerous cancers. Such agents could prove to be useful agents in combination with other therapeutic modalities for the treatment of CLL. # Supporting Information We gratefully acknowledge the contributions of Dr. Huilin Zhou and Marie Reichart for guidance with the IMAC technique, Larry Gross, Dario Meluzzi and Mike Meehan for assistance with LC-MS/MS and Dr. Steve Bark and Dr. Elizabeth Komives for many useful discussions and critical reading of this work. [^1]: Conceived and designed the experiments: MO CLS DM PCD TMH. Performed the experiments: MO CLS. Analyzed the data: MO CLS. Contributed reagents/materials/analysis tools: TJK DM PCD. Wrote the paper: MO. [^2]: The authors have declared that no competing interests exist.
# Introduction The foetal and developmental origins of adult disease may also be relevant to the aetiology of dementia. Late-life leg length and skull circumference provide useful information about the nutritional environment in early life, and brain development, respectively. Adult leg length is sensitive to diet in infancy, specifically breast feeding and energy intake. Increasing population height is mainly accounted for by increasing leg length, linked to improvements in childhood nutrition. Skull dimensions are a long-term and stable marker of early-life brain size. Inverse associations between skull circumference and prevalent Alzheimer’s disease (AD) were reported in five cross-sectional studies: three of which were community based, from the USA, Brazil, and Korea, and two of communities of Catholic nuns from the US and Germany. There are two negative reports, both small case-control studies with cases recruited from clinical services. In the only cohort study an inverse association was noted with incident AD among Japanese-Americans. Leg length was inversely cross-sectionally related to dementia prevalence in Brazil, and among women in a population-based study in Korea, and with cognitive impairment among Caribbean migrants to the UK. In the only cohort study, from the USA, knee height was inversely associated with incident dementia, but among women only. This was a post hoc finding, and the interaction with gender was not statistically significant. In the baseline phase of the 10/66 Dementia Research Group studies in Latin America, China and India, longer legs and larger skulls were independently associated with a lower prevalence of dementia. There was little heterogeneity of estimates among sites. Earlier reports of effect modification were not replicated. The aims of the current analysis based upon the incidence phase of the 10/66 Dementia Research Group surveys are similar, although the outcome is now incident dementia, and we explore the role of reverse causality. Specifically we aim to 1. test hypotheses that longer leg lengths and larger skull circumferences are associated with a lower incidence of dementia in large, representative population-based dementia-free cohorts, checking for consistency across diverse urban and rural settings in Latin America, and China 2. test for effect modification, with a priori hypotheses based upon previously observed associations any protective effect of longer leg length on dementia incidence is modified by gender (stronger in women) any protective effect of greater skull circumference on dementia incidence is modified by education (stronger among the least educated), or gender (stronger in women) 3. assess the possibility that reverse causality may have accounted for previously observed cross-sectional associations by testing whether cognition at baseline is associated with changes in leg and skull measurements from baseline to follow-up reduction in leg lengths and skull circumferences are associated with the incidence of dementia. # Materials and methods The 10/66 population-based baseline and incidence wave study protocols are published in an open access journal, and the profile of the resultant cohort has also been described. Relevant details are provided here. One-phase population- based surveys were carried out of all residents aged 65 years and over in geographically defined catchment areas (urban sites in Cuba, Dominican Republic, Puerto Rico and Venezuela, and urban and rural sites in Mexico, Peru, and China). Baseline surveys were completed between 2003 and 2007, other than in Puerto Rico (2007–2009). The target sample was 2000 for each country, and 3000 for Cuba. The baseline survey included clinical and informant interviews, and physical examination. Incidence waves were subsequently completed, with a mortality screen, between 2007 and 2011 (2011–2013 in Puerto Rico) aiming for 3–4 years follow-up in each site. Assessments were identical to baseline protocols for dementia ascertainment, and similar in other respects. We revisited participants’ residences on up to five occasions. If no longer resident, we sought information on their vital status and current residence from contacts recorded at baseline. If moved away, we sought to re-interview them, even outside the catchment area. If deceased, we recorded the date, and completed an informant verbal autopsy, including evidence of cognitive and functional decline suggestive of dementia onset between baseline assessment and death. ## Measures The 10/66 survey interview generates information regarding dementia diagnosis, mental disorders, physical health, anthropometry, demographics, an extensive dementia and chronic diseases risk factor questionnaire, disability, and health service utilisation. Only assessments relevant to the analyses of associations between leg length, skull circumference and dementia incidence are described in detail here. ### Anthropometric measures Skull circumference was measured using a cloth tape measure encircling the nuchal tuberosity and the brow. Leg length was measured, standing (or lying down if not feasible), from the highest point of the iliac crest to the lateral malleolus. In the event of deformity the non-deformed leg was measured. Both dimensions were measured, to the nearest centimeter, at both time points. ### Sociodemographic variables Age was determined at baseline interview from stated age, documentation and informant report, and, if discrepant, according to an event calendar. We also recorded sex, and education level (none/ did not complete primary/ completed primary/ secondary/ tertiary). ### Frailty At baseline, we assessed four of the five indicators of the Fried physical frailty phenotype; exhaustion, weight loss, slow walking speed and low energy expenditure. Since hand grip strength was not measured we considered participants as frail if they fulfilled two or more of the four frailty indicators. ### Dementia Dementia diagnosis according to the 10/66 Dementia Research Group methodology (referred to throughout as ‘incident dementia’) is allocated to those scoring above a cutpoint of predicted probability for dementia, calculated using coefficients from a logistic regression equation developed, calibrated and validated cross-culturally in the 25 centre 10/66 pilot study, applied to outputs from a) a structured clinical interview, the Geriatric Mental State, b) two cognitive tests; the Community Screening Instrument for Dementia (CSI-D) COGSCORE and the modified CERAD 10 word list learning task delayed recall, and c) informant reports of cognitive and functional decline from the CSI-D RELSCORE. For those who died between baseline and follow-up we diagnosed ‘probable incident dementia’ by applying three criteria: 1. A score of more than two points on the RELSCORE, from the post-mortem informant interview, with endorsement of either ‘deterioration in memory’ or ‘a general deterioration in mental functioning’ or both, and 2. an increase in RELSCORE of more than two points from baseline, and 3. the onset of these signs noted more than six months prior to death. In the baseline survey, the first criterion would have detected those with either DSM-IV or 10/66 dementia with 73% sensitivity and 92% specificity. Prevalence and incidence of dementia in the current cohorts has been reported. ## Analyses We used STATA version 11 for all analyses. For each site 1. we describe participants’ characteristics; age, sex, educational level, mean leg length, and skull circumference 2. we model the effect of covariates on dementia incidence using a competing-risks regression derived from Fine and Gray’s proportional subhazards model. This is based on a cumulative incidence function, indicating the probability of failure (dementia onset) before a given time, acknowledging the possibility of a competing event (case dementia-free death). In our analysis of baseline data, men had larger skulls and longer legs than women in all sites, while in most sites older people had smaller skulls and shorter legs, consistent with birth cohort effects. Better education was generally associated with larger skulls and longer legs. Age, gender and education were also the major determinants of incident dementia in this cohort. We therefore controlled for age, gender, and education when modelling the effects of leg length (quarters) and skull circumference (quarters) on dementia incidence, comparing the longest or largest quarters with the shortest or smallest. We report fully adjusted sub-hazard ratios (ASHR) with robust 95% confidence intervals adjusted for household clustering. As a sensitivity analysis, we also estimate the linear effect of leg length and skull circumference (per centimetre) on dementia incidence, controlling for the same covariates. We fit the models separately for each site and then use a fixed effects meta-analysis to combine them. Higgins I<sup>2</sup> quantifies the proportion of between-site variability accounted for by heterogeneity, as opposed to sampling error; up to 40% heterogeneity is conventionally considered negligible, while up to 60% may reflect moderate heterogeneity. 3. we test for effect modification of linear effects of leg length on dementia by gender, and of skull circumference upon dementia by gender and education, by extending the adjusted models by appropriate interaction terms. 4. we estimate correlations between baseline and follow-up measures of skull circumference and leg length, and explore the possibility of reverse causation by a) comparing change in leg length and skull circumference by baseline cognitive status (cognitively normal, versus ‘cognitive impairment no dementia (CIND), versus dementia), and b) testing whether change in leg length or skull circumference is associated with incident dementia. Participation in the study was by signed informed consent, or by signed independently witnessed verbal consent for participants who could not read or write. The study protocol and the consent procedures were approved by the King's College London research ethics committee and in all countries where the research was carried out: 1- Medical Ethics Committee of Peking University the Sixth Hospital (Institute of Mental Health, China); 2- the Memory Institute and Related Disorders (IMEDER) Ethics Committee (Peru); 3- Finlay Albarran Medical Faculty of Havana Medical University Ethical Committee (Cuba); 4- Hospital Universitario de Caracas Ethics Committee (Venezuela); 5- Consejo Nacional de Bioética y Salud (CONABIOS, Dominican Republic); 6- Instituto Nacional de Neurología y Neurocirugía Ethics Committee (Mexico); 7- University of Puerto Rico, Medical Sciences Campus Institutional Review Board (IRB) # Results ## Sample characteristics Across the 10 sites, 15,027 interviews were completed at baseline. Response proportions varied between 72% and 98%. The ‘at risk’ cohort comprised 13,587 dementia-free persons aged 65 years and over. Mean age at baseline varied from 72.0 to 75.4 years, lower in rural than urban sites and in China than in Latin America. Women predominated over men in all sites, accounting for between 53% and 67% of participants, by site. Educational levels were lowest in rural Mexico (83% not completing primary education), Dominican Republic (70%), rural China (69%), and urban Mexico (54%) and highest in urban Peru (9%), Puerto Rico (20%), and Cuba (23%). In other sites, between a third and a half of participants had not completed primary education. There was significant between site variation in leg length (shorter legs in rural compared with urban sites; F = 64.7, p\<0.04), and skull circumference (no obvious pattern of variation; F = 82.3, p\<0.001). Site accounted for 5.5% of the variance in skull circumference, and 4.4% of the variance in leg length. From the ‘at risk’ cohort, 2,443 participants (18.0%) were lost to follow-up; 10,540 with baseline skull circumference data were followed up for 40,466 person years, and 10,400 with baseline leg length data were followed up for 39,954 person years. In the skull circumference cohort, there were 1,766 deaths, of which 1,605 were judged to be dementia-free, and 1,009 cases of incident dementia were observed (of which, 161 were ‘probable’ cases diagnosed retrospectively from post-mortem informant interviews). ## Leg length and incident dementia In all sites other than Cuba, there was a non-significant trend towards inverse associations between leg length and incident dementia. The pooled meta-analysed adjusted sub-hazard ratio (ASHR) for leg length (longest vs. shortest quarter) was 0.80 (95% CI, 0.66–0.97) with no heterogeneity between sites (I<sup>2</sup> = 0%). The linear effect, per centimetre of leg length, was also statistically significant when pooled across sites (SHR 0.986, 95% CI 0.977–0.995, I<sup>2</sup> = 0%). The interaction between leg length (per quarter) and gender (male versus female) was not statistically significant (pooled ASHR 1.03, 95% CI 0.90–1.18, I<sup>2</sup> = 38.3%). ## Skull circumference and incident dementia There was no evidence for an association between skull circumference and incident dementia in any site, or after meta-analysis, whether comparing quarters with the largest and smallest skulls (ASHR 1.02, 95% CI, 0.84–1.25, I<sup>2</sup> = 0%), or the linear effect per centimetre of skull circumference (ASHR 0.987, 95% CI 0.960–1.015, I<sup>2</sup> = 0%). The interaction between skull circumference (per quarter) and gender (male versus female) was statistically significant (pooled ASHR 0.86, 95% CI 0.75–0.98, I<sup>2</sup> = 0.0%), but in the reverse direction to that hypothesized with a greater protective effect of larger skulls in men than in women. In a further exploratory analysis stratified by gender the effect of skull circumference did not attain statistical significance either in men (pooled ASHR 0.91, 95% CI 0.81–1.02, I<sup>2</sup> = 0.0%), or in women (pooled ASHR 1.07, 95% CI 0.99–1.15, I<sup>2</sup> = 0.0%). There was a non-significant trend towards an interaction between skull circumference and education (per level) (pooled ASHR 0.986, 95% CI 0.971–1.001, I<sup>2</sup> = 32.1), also in the opposite direction to that hypothesized, with a greater protective effect at higher educational levels. ## Changes in leg length and skull circumference measures Correlations between baseline and follow-up skull circumferences (0.53) and leg lengths (0.43) were generally modest. They also varied among sites; Cuba (skull circumference 0.62, leg length 0.50), Dominican Republic (0.48, 0.34), Puerto Rico (0.47, 0. 51), Peru urban (0.83, 0.70), Peru rural (0.44, 0.36), Venezuela (0.56, 0.37), Mexico urban (0.64, 0.46), Mexico rural (0.56, 0.46), China urban (0.51, 0.50), China rural (0.20, 0.21). Inspection of the distribution of change scores revealed some biologically implausible differences at the extremes of the distributions. Even after exclusion of the outliers (top and bottom 5% of the change score distributions), there was a tendency for both skull circumferences (mean difference -0.20, 95% CI -0.23 to -0.17 cms) and leg lengths (mean difference -0.99, 95% CI -1.12 to -0.85 cms) to be smaller on the second measurement occasion. Apparent shortening of leg length was more marked among those with dementia at baseline, than in those with CIND, and those who were cognitively normal. Changes in skull circumference did not differ between those with dementia at baseline and those who were cognitively normal. ## Post hoc sensitivity analyses Given modest correlations between baseline and follow-up measures of skull circumference and leg length, we repeated the regressions reported in, limited to those individuals with baseline and follow-up anthropometric assessments, and excluding the 10% with the most discrepant baseline and follow-up measures. This restriction did not affect the main findings (skull circumference pooled AHR 1.05, 95% confidence intervals 0.80–1.36, I<sup>2</sup> = 37.9%; leg length AHR 0.76, 95% CI 0.60–0.97, I<sup>2</sup> = 0.0%). Given the association between baseline dementia status and subsequent change in measured leg length, we assessed the association of change in leg length (per centimetre) with incident dementia, controlling for age, gender and education. The pooled effect size (AHR 1.006, 95% CI 0.992–1.020, I<sup>2</sup> = 19.3%) did not suggest any association. Neither was change in skull circumference associated with incidence of dementia (pooled AHR 1.016, 95% CI 0.961–1.074, I<sup>2</sup> = 7.7%). Since anthropometric measures were not available at follow-up on those who were deceased, these analyses could not take account of the competing risk of dementia free death, and were simple Cox’s Proportional Hazard regressions. Finally, other than baseline cognitive status, physical frailty seemed to be the most prominent independent determinant of change in leg length, (adjusted mean difference, controlling for age, gender, education and CIND, -1.19, 95% CI -1.89 to -0.70 cms). We therefore repeated the regression for the effect of baseline leg length on incident dementia (reported) controlling for frailty in addition to age, education and gender, and there was only a very small attenuation of the effect (ASHR 0.82, 95% CI 0.67–0.99, I<sup>2</sup> = 0%). # Discussion ## Main findings This study provides additional support for the salience of early life development to late-life dementia risk. However, while the hypothesis that longer legs are associated with a reduced risk of incident dementia was supported, the hypothesis that larger skulls would also be associated with a reduced risk was not. The association between longer legs and reduced dementia risk was independent of age, gender, education and physical frailty, and there was no heterogeneity of effect across sites. There was no evidence that the effect of leg length on incident dementia was modified by gender. The effect of skull circumference on incident dementia was modified by gender, but in the contrary direction to the only previous report, from a cross-sectional study from Korea. ## Strengths and weaknesses of the study The associations were assessed longitudinally, in large population-based dementia-free cohorts, comprising rural and urban catchment area sites in the Caribbean, Latin America, and China. Fixed effect meta-analysis was appropriately applied to increase the precision of our estimates, given the negligible heterogeneity for all of the associations studied. There have been only two previous longitudinal studies of these associations, neither of which approached the power and precision conferred by our 10,000 participants and 40,000 person years of follow-up. A unique aspect of our approach was the repeated anthropometric measures at baseline and follow-up. The moderate agreement for the repeated measures, particularly for skull circumference suggested random measurement error, with potential for underestimation of the true effect of these exposures upon dementia risk. In a post-hoc analysis, the inverse association between leg length and incident dementia was more pronounced when excluding those with more discrepant baseline and follow-up leg length measurements, suggesting regression dilution bias. Random measurement error might have been reduced by more appropriate measurement methods (for example knee height or sitting minus total height), and more training and quality control. We also found a consistent tendency for both leg and skull dimensions to be smaller on the second measurement occasion. For legs, but not skulls, this was more pronounced for those with dementia at baseline than those who were cognitively normal. Although mortality was high among those with dementia, and those who died did not have a second anthropometric assessment, selective mortality seems an unlikely explanation. Leg lengths, per se, are unlikely to shorten other than after hip fracture. However, our measurement approach, measuring standing leg length, may have led to apparent shortening (and, in some cases, lengthening) due to difficulty in keeping the knee joint fully extended, particularly for those with arthritis, sarcopaenia, general fatigue, and, possibly, cognitive impairment. Physical frailty (capturing several of these elements) was associated with leg length shortening. Given these findings, the possibility of reverse causation, or residual confounding cannot be positively excluded. However, the association between baseline leg length and incident dementia persisted after controlling for physical frailty, and there was no association between change in leg length (or skull circumference) and incident dementia. Finally, losses to follow up may have led to attrition bias, if for example those with longer legs who developed dementia were selectively more likely to refuse, or not be traced. However, losses to follow-up for reasons other than death were negligible to modest, and were not generally associated with the important determinants of dementia risk. The possible biasing effects of attrition through death were addressed, in part, through post mortem ascertainment of probable incident dementia, and the competing risk regression. ## Contextualisation with other research Our finding of an inverse association between leg length and incident dementia is consistent with findings from the baseline phase of our study, and with three other cross-sectional studies and one longitudinal study. We did not confirm an inverse association between skull circumference and incident dementia previously reported in one cohort study of Japanese-Americans, and also observed cross- sectionally in the baseline phase of our study, and in five other cross- sectional studies. Any suggestion that the association of skull circumference with incident dementia might be modified by gender must remain tentative, given the absence of a main effect, and that the observed interaction is in the opposite direction to that reported in the only other study to test for this interaction, a cross-sectional study from Korea. Our finding is therefore contrary to our a priori hypothesis, and may well be accounted for by Type 1 error. Further research may clarify this issue. The relationships between birthweight, subsequent growth, and cognitive development are well established, with possible ‘sensitive periods’ in infancy, childhood and adolescence. In a recent analysis of historical cohort data from the UK MRC National Survey of Health and Development, trunk and leg length were inversely associated with cognitive function at age 53 years. These associations attenuated after controlling for adverse early life circumstances, and were entirely accounted for controlling for cognition at age 15 years. Hence early- life skeletal growth may be a marker for cognitive development, and associations of adult leg length with late-life cognition and dementia risk may be accounted for by tracking of cognitive ability across the life course, and the cognitive reserve hypothesis, respectively. This would not, however, adequately account for the discrepancy, in our analyses, between leg length and skull circumference as predictors of incident dementia. Although adult leg length and height is mainly influenced by early growth and its determinants, legs continue to grow for up to a decade after skull growth is complete; skull circumference and leg length may therefore be markers of different phases, and, possibly, different aspects of the developmental process. Shorter limb length is associated with cardiovascular risk factors in later life, particularly those linked to insulin resistance. This may be another mediating pathway, best explored in birth cohorts with mid-life assessment of cardiovascular risk profiles. Leg length, in contrast to trunk height, and skull circumference are thought to be stable across the adult life course. We found few previous studies of changes in measured leg length and skull dimensions in older populations. As we highlighted in our previous cross-sectional analysis, in a report from our 10/66 DRG site in Chennai, India where follow-up was limited to those with dementia or cognitive impairment, reassessment of skull circumferences and leg lengths three years after baseline indicated that leg lengths did diminish, but at a similar rate in those with and without dementia, while skull circumferences did not. Furthermore, in the previous cross-sectional analysis, we found no associations between dementia severity and either leg length or skull circumference. This concurs with findings from a longitudinal computed tomography study of skull dimensions of people with dementia. # Conclusions Our cohort study clarifies the association between markers of early life development; leg length (an inverse association) and skull circumference (no association) and the incidence of dementia in late life. It addresses previous unresolved concerns regarding direction of causality through repeated measures of leg and skull dimensions. Main effects and interactions are estimated with much more power and precision than had been previously possible. Findings from previous studies in high income countries, can now be generalised to diverse urban and rural settings in Latin America and China. Consistent findings across settings and previous studies provide quite strong support for an association between adult leg length and dementia incidence in late-life, with early life nutrition most plausibly implicated as the antecedent cause. While attention has focused upon education as a modifiable risk factor for dementia, much less has been accorded to other developmental factors. Mean male adult height, after a long period of relative stability, has increased by 5-10cms since 1920 in high income countries. This secular trend may, in part, explain recent trends towards a declining incidence of dementia in high income countries, which are not adequately accounted for by improvements in education and cardiovascular risk profile. Our findings would predict a 7% reduction in the incidence of dementia with each five centimetre increment in height. Further clarification of these associations could inform predictive models of future incidence, prevalence and numbers affected, and invigorate global efforts to improve childhood nutrition, growth and development. [^1]: Michael E Dewey is a statistical advisor for PLOS Medicine. All other authors have declared that no competing interests exist.
# Introduction Many studies have examined the association between tobacco smoking and breast cancer risk. However, the findings have been controversial. Some studies have reported no increased risk, while others have reported increased risk for passive smoking exposure. The review by Canadian expert panel showed that the evidence for a relationship between passive smoking and breast cancer remained tenuous, although they suggested that the relationship between passive smoking and breast cancer in younger, primarily premenopausal women was consistent with causality. The most recent review suggested that the role of passive smoking was less clear. In China, traditionally, few women are smokers, but the rate of passive smoking has known to be high. A survey conducted between 2005 and 2007 showed a high rate of 43.6%. Even higher level of passive smoking was reported in younger women and in rural areas. Some studies have evaluated the association between passive smoking and breast cancer risk among Chinese women. Although most studies reported the positive association of passive smoking with breast cancer risk, these studies had relatively smaller sample size with 108 to 704 study subjects, and few studies have used quantitative measures to evaluate the exposures of passive smoking both at home and in the workplace. Inadequate evaluations of exposure may result in an under-estimation of the risks, if they do exist. We conducted this case-control study in Guangdong Province, China to investigate the association of passive smoking at home and in the workplace with breast cancer risk. # Materials and Methods ## Ethics Statement The procedures and protocols of the study were approved by The Ethical Committee of School of Public Health, Sun Yat-sen University. All participants signed informed consent forms before the interview. ## Study Subjects Breast cancer cases and controls were recruited through two stages. The first stage was conducted from June 2007 to August 2008 and the second stage was from September 2011 to September 2013. Potential case patients were recruited from inpatients admitted to the surgical units of two affiliated hospitals of Sun Yat-sen University. Eligible cases were female subjects with histologically confirmed breast cancer diagnosed no more than 3 months before the interview, aged 25–70 years and natives of Guangdong province or having lived in Guangdong for at least 5 years. Women were excluded if they had a history of breast cancer or other cancers. Totally, 925 eligible cases were identified and 896 were interviewed, with a response rate of 96.9%. Control subjects were patients admitted to the same hospitals during the same time period as the case subjects. Eligibility criteria for controls were the same as described for the cases except that they had no history of any cancers. They were frequency matched with cases by age (5-year interval) and residence (rural/urban). These patients presented with a wide spectrum of non-neoplastic conditions including eye disorders (glaucoma, uveitis, keratitis, pterygium, dacryocystitis, and optic neuritis), ear-nose-throat diseases (sudden deafness, acute bacterial/viral otitis media, sinusitis, deviation of nasal septum, tonsillitis), trifacial neuralgia, varicose veins, osteoarthritis, degenerate joint disease, orthopedics diseases, facial paralysis and acute appendicitis. In total, 939 controls were identified and 912 were interviewed, with 2.9% patients refused to participate. ## Data Collection A structured questionnaire was used to collect information through face-to-face interview by trained interviewers. The collected information included socio- demographic and anthropometric parameters, dietary habits, menstrual and reproductive factors, use of hormone and contraceptive drugs, family history of cancer, alcohol drinking, active smoking, passive smoking history, disease history, and physical activity. The interview time was limited to exposures that occurred before diagnosis date for cases and the interview date for controls. Relevant medical information, medical diagnosis, and histological findings were abstracted from the hospital medical records. Women were classified as non-smokers if they reported never smoking or smoking less than 100 cigarettes over their lifetime. Passive smoking history was collected for two exposure sources. First, the subject was asked whether her husband or other family members ever smoked in her house, then she was asked the average number of cigarettes they smoked per day and the number of years she had been exposed at home. Second, the subject was asked whether someone ever smoked within three meters around her in her workplace, then she was asked the number of people and the number of years she had been exposed in the workplace. Women were categorized as having been exposed to passive smoking if they reported ever being exposed to tobacco smoke at home or in the workplace. Among women who were passive smokers, the duration or intensity of exposure with smoker-years, cigarettes/day or pack-years of exposure were calculated. Smoker-years were defined as sum of the number of years of exposure to each smoker. Total smoker- years were calculated as the sum of smoker-years at home and in the workplace. Pack-years were defined as the number of years of exposure multiplied by the pack of cigarettes (1 pack = 20 cigarettes) smoked per day for a given smoker. Pack-years were summed across smokers to generate a total pack-years measure. Body mass index (BMI) was calculated by dividing weight (kg) by height (m<sup>2</sup>). Menopausal status was defined as at least 12 months since the last menstrual cycle. Women were considered to be premenopausal if they were currently menstruating, or if they were not menstruating because of a hysterectomy and younger than 50 years old. Women were defined as postmenopausal if they had either undergone a natural menopause, or surgery to remove both ovaries, or if their ovarian function was unknown but they were older than 50 years. ## Statistical analysis Since the socio-demographic and established breast cancer risk factors of the two-stage study subjects are comparable, we pooled the two stage data for these analyses. The analysis excluded 41 subjects (19 cases and 22 controls) who reported past or current history of personal tobacco smoking. Analyses were based on the remaining 1767 non-smokers (877 cases and 890 controls). Differences in characteristics were assessed by using either χ<sup>2</sup> tests for categorical variables or t tests for continuous variables. Unconditional logistic regression models were used to estimate odds ratios (OR) and 95% confidence intervals (CI) for the association between passive smoking and breast cancer risk. Based on the comparison of baseline characteristics between cases and controls, the following variables, BMI, physical activity, age at menarche, age at first live birth, age at menopause, history of benign breast disease, and mother/sister/ daughter with breast cancer, were selected to be adjusted for as potential confounding factors. Age, residence and study stage were also controlled for in all logistic models. Tests for trend were performed by entering categorical variables as continuous parameters in the models. Stratified analyses by menopausal status were conducted. As breast cancer is a heterogeneous disease with estrogen receptor (ER) and progesterone receptor (PR) subtypes, stratified analyses by ER/PR status were also conducted. All analyses were performed using SPSS 13.0 (SPSS Inc., Chicago, Illinois, USA). All tests were two-sided, with P values \< 0.05 indicating statistical significance. # Results Compared to controls, cases were more likely to have higher BMI, an earlier age at menarche, a later age at first live birth, a later age at menopause, a history of benign breast disease and a family history of breast cancer, and were less likely to be physically active. All of the above variables were considered potential confounders and adjusted for in subsequent analyses. No significant differences were found between cases and controls in socio-demographic factors, including marital status, educational level, occupation, and household income, or in reproductive factors, including nulliparous, number of live births, months of breast feeding, menopausal status, and use of an oral contraceptive. As shown in, of all subjects, 495 (56.4%) cases and 442 (49.7%) controls reported ever having been exposed to passive smoking at home or in the workplace. Compared with women who were never exposed to passive smoking, women who were ever exposed had a higher risk of breast cancer, with the adjusted OR (95% CI) of 1.35 (1.11–1.65). When subjects were categorized according to sources of exposure, 39.7% of cases and 37.4% of controls were exposed only at home, 7.1% of cases and 7.2% of controls were exposed only in the workplace, and 9.7% of cases and 5.1% of controls reported both exposures. The adjusted ORs (95% CIs) of breast cancer were 1.30 (1.05–1.61) for passive smoking exposure only at home, 1.05 (0.71–1.56) for passive smoking exposure only in the workplace and 2.17 (1.45–3.23) for both exposures, compared with women unexposed to passive smoking. Passive smoking exposure at home was examined in detail. Women who were ever exposed to passive smoking at home (53.1% cases and 45.8% controls) had a higher risk of breast cancer compared with women who were never exposed to passive smoking, with the adjusted OR (95% CI) of 1.30 (1.05–1.61). Dose-response relationships between breast cancer risk and smoker-years, cigarettes per day and pack-years of exposure at home were observed (P<sub>trend</sub> = 0.003, 0.006 and 0.009, respectively). Compared with women with no passive smoking exposure, the adjusted ORs (95% CIs) of more than 26 smoker-years, more than 16 cigarettes per day and more than 16 pack-years of exposure at home were 1.66 (1.21–2.26), 1.56 (1.17–2.09) and 1.61 (1.17–2.19), respectively. Analysis on the association between passive smoking exposure at workplace and breast cancer risk (418 cases and 465 controls) showed no association, with the adjusted OR (95% CI) of 1.19 (0.80–1.78) comparing women who were ever exposed to tobacco smoke in the workplace with never exposed women. No significant association was found between smoker-years of exposure in the workplace and breast cancer risk (P<sub>trend</sub> = 0.313). A strong dose-response relationship and a positive association were observed between total smoker-years of passive smoking exposure at home and in the workplace and breast cancer risk. Compared with women who were never exposed to passive smoking, the adjusted ORs (95% CIs) was 0.95 (0.70–1.29) for 1–16 smoker-years, 1.34 (1.04–1.74) for 17–30 smoker-years, and 1.95 (1.43–2.66) for more than 31 smoker-years, respectively (P<sub>trend</sub> \< 0.001). Analyses stratified by menopausal status showed a positive association between passive smoking and breast cancer risk, primarily among postmenopausal women. Compared with non-exposed women, the adjusted OR (95% CI) was 1.83 (1.29–2.60) for women with any passive smoking exposure and 1.80 (1.24–2.61) for women ever exposure to passive smoking at home. The significant dose-response relationships between breast cancer risk and smoker-years, cigarettes per day and total pack- years of exposure at home were also found only among postmenopausal women (P<sub>trend</sub> = 0.001, \< 0.001 and 0.001, respectively). But the significant dose-response relationship between total smoker-years of any exposure and breast cancer risk was found in both pre/post-menopausal women. No significant association was found between passive smoking in the workplace and breast cancer risk in both the pre/post-menopausal women. We also evaluated the relationship between passive smoking and breast cancer risk stratified by ER and PR status, excluding 92 (10.5%) cases with no ER/PR information. These analyses included 567 ER+ cases, 218 ER- cases, 596 PR+ cases and 189 PR- cases. A positive association was observed in all subtypes of ER and PR status, although the associations among women with ER+/PR+, and ER+/PR- or ER-/PR+ breast cancer tumors were not statistically significant. # Discussion This study found that passive smoking exposure was associated with an increased risk of breast cancer. The significant dose-response relationship between total smoker-years and breast cancer risk was found in both the pre and post- menopausal women, and in all ER/PR subtypes of breast cancer. This study also showed that passive smoking exposure at home was associated with increased risk of breast cancer, and there were significant dose-response relationships in smoker-years, cigarettes/day and total pack-years. However, no evidence of a relationship between passive smoking exposure in the workplace and breast cancer risk was found. Results from previous studies on passive smoking and breast cancer risk were inconclusive. Some studies suggested an increased risk with passive smoking exposure and others showed no effect. A meta-analysis containing seven cohort studies and twelve case-control studies supported the notion that passive smoking exposure was positively associated with breast cancer risk (relative risk = 1.27, 95% CI = 1.11–1.45). A later meta-analysis reported the combined relative risk of breast cancer in the 17 studies with retrospective reporting of exposure was 1.21 (1.11–1.32), based on a total of 5696 breast cancer women. The result of the present study is consistent with that reported from the meta- analyses, showing a positive association between passive smoking and an increased risk of breast cancer. The observed association between passive exposure to tobacco smoke and breast cancer risk is biologically plausible. Tobacco smoke contains over a dozen fat- soluble compounds that are known to induce mammary tumors in rodents. Studies have strongly suggested that breast tissue is a target for the carcinogenic effects of tobacco smoke. DNA adducts with derivatives of tobacco smoke are more common in the breast tissue of smokers than that of non-smokers. It has been demonstrated that most of the tobacco smoke is not inhaled by the smokers and the highest amounts of many components, such as carbon monoxide, nicotine, benzene, formaldehyde, N-nitrosamines, nickel and tar, are found in side-stream smoke. Moreover, the vapor-phase constituents from side-stream smoke are also more quickly absorbed into blood and lymph systems than the particulate-phase particulates found in main stream smoke. This study found a significant dose-response relationship between total smoker- years of any exposure and breast cancer risk. Some previous studies have reported increased breast cancer risks for various duration of exposure, or regardless of duration of exposure. However, many studies have not observed any linear dose–response relationship between passive smoking and breast cancer risk . In this study, every smoker around passive smokers were considered and smoker-years was used. This may reflect a better measure to evaluate the duration and intensity of passive smoke exposure than years. Inconsistencies in exposure assessment methods may contribute to the inconsistent findings across studies. Ideally, an exhaustive assessment of exposure to passive smoking should include the duration and intensity of childhood home exposure, adult home exposure, and workplace exposure. However, some passive smoking studies relied on husband's smoking history as the index of exposure and did not quantify additional sources of exposure. This may lead to possibly non-differential misclassification of the exposure status and may dilute the risk estimates. In the present study, we had no information on childhood home exposure. Although more recent studies considered childhood exposure, almost universally, these studies tend to report null results for breast cancer risk. This may in part be due to the fact that self-report of parental smoking is subject to even greater error. Analyses according to sources of passive smoking exposure were also conducted. A positive association and significant dose-response relationships in smoker- years, cigarettes/day and total pack-years were found. Our findings are consistent with that observed in some home exposure studies. However, a number of home exposure studies found null association. Some of these studies solely relied on husband's smoking history as the index of home exposure, and most studies had no information on workplace exposure. As such, inadequate passive smoke exposure assessment (for example, ignoring occupational exposure) could result in classifying those with only workplace exposure as “unexposed”, thus leading to underestimates of risks, if they should exist. In the present study, we collected detailed information on home and workplace passive smoke exposure, and used no any exposure as the referent group. We observed no significant association between workplace exposure to passive smoking and risk of breast cancer. Our results are in accordance with most other previous results. However, a case-control study conducted in Shanghai, China, found some evidence of a slightly elevated breast cancer risk associated with workplace exposure of 5 hr. or more per day (OR = 1.6, 95% CI = 1.0–2.4; P <sub>trend</sub> = 0.02) among women who worked during the 5 years after excluding the influence of home exposure. A cohort study conducted among California teachers observed an increased risk in the most highly exposed subgroup of postmenopausal women exposed in adulthood (age ≥20 years) (hazard ratio = 1.25, 95%CI = 1.01–1.56). Another cohort study observed a 32% excess risk of breast cancer associated with the most extensive exposure to passive smoking among postmenopausal women who had never been active smokers. Other epidemiological studies also showed a statistically significant positive association between passive smoking and breast cancer risk in postmenopausal women. Consistent with these results, our study also provided strong supporting evidence that passive smoking was associated with an increased risk of breast cancer in postmenopausal women. One possible explanation for the positive association of passive smoking with breast cancer in the postmenopausal only might be related to the anti-estrogenic effects of passive smoking. Smoking women have an earlier menopause and thus fewer years of menstruation. And cigarette smoking alters estrogen metabolism, which may contribute to the absence of a positive association of passive smoking with premenopausal breast cancer. However, some reports suggested that passive smoking was associated with an increased risk of breast cancer among premenopausal women or both pre/post- menopausal women. Since this was a stratified analysis, chance findings might arise. More studies with a larger sample size might be needed to confirm this association. Some studies have examined the association between passive smoking and breast cancer risk by ER/PR status, and yielded inconsistent results \[, , –\]. Tong et al. reported that passive smoking exposure from partners was associated with increased risk of ER+/PR+ breast cancer among non-smoking Chinese urban women. Morabia et al found that passive smoking increased the risk of both ER+ and ER- breast cancer. However, other studies found no significant association between passive smoking and breast cancer risk stratified by ER/PR status. Our study found a strong positive association between passive smoking and all subtypes of ER/PR status of breast cancer, although the association was statistically non-significant for some subtypes. ## Strengths The present study has some strengths. We conducted detailed comprehensive measurements of passive smoking exposure at home and in the workplace, including duration and intensity of exposure (e.g., smoker-years, cigarettes/day and pack- years).The data were collected using face-to-face interviews by trained interviewers and the response rates of cases and controls were relatively high. Furthermore, some major potential confounding factors were adjusted in all logistic regression models. ## Limitations This study had some limitations. First, selection bias was inevitable in hospital-based case-control studies. To minimize this bias, great attempt was made to recruit controls from patients with a wide spectrum of non-neoplastic conditions. Moreover, the high participation rate (96.9% and 97.1% for cases and controls, respectively) and high comparability in socio-demographic factors between the two groups indicated that selection bias should not be a serious problem. Second, recall bias was also of concern in case-control studies. To reduce recall bias, we tried to interview patients as soon as diagnosis was made and take great effort to interview cases before their surgery. Third, limited sample size in some subgroups might lead to limited power to detect the associations. Fourth, there were some missing data on the duration or intensity of exposure, which may lead to an under estimation of the association. But the percentage of missing data was less than 5%. Fifth, this study had no information on genetic polymorphisms, which had been reported to modify the association of passive smoking with breast cancer risk. # Conclusions In summary, this study suggested that passive smoking was associated with an increased risk of breast cancer among non-smoking Chinese women. A stronger positive association with breast cancer was seen among postmenopausal women and all subtypes of ER/PR status of breast cancer. Future studies are needed to confirm these results. # Supporting Information We greatly appreciate the participation of the study subjects and the contributions of students in the data collection. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: BL SCH CXZ. Performed the experiments: BL LW MSL. Analyzed the data: BL LW. Contributed reagents/materials/analysis tools: XFM FYL. Wrote the paper: BL SCH CXZ.
# Introduction Legume root and stem nodules fix atmospheric dinitrogen (N<sub>2</sub>) yielding anabolic-N, which augments growth and reproduction of host plants. In these nodules, the biochemical conversion of N<sub>2</sub> to ammonium is owed to endosymbiotic rhizobia (bacteroids) who carry the N<sub>2</sub> fixation genes encoding the dinitrogenase complex. Whether N<sub>2</sub> fixation occurs in legume nodules or in pure cultures of diazotrophic (able to use N<sub>2</sub> as sole N-source) bacteria, hydrogen gas (H<sub>2</sub>) is then co-produced. From subsequent mechanistic studies of dinitrogenase activity, H<sub>2</sub> co- production is both stoichiometric and requires 2 ATP per H<sub>2</sub> formed. Yet in agronomic surveys, many legume nodules typically evolve H<sub>2</sub> at high levels, and such H<sub>2</sub> evolution rates correlate with N<sub>2</sub> fixation rates. However, in certain symbiotic legume nodules, bacteroids avidly fix N<sub>2</sub> yet reproducibly evolve little or no H<sub>2</sub>. As this endogenous H<sub>2</sub> production consumes metabolic energy, H<sub>2</sub> recycling, which recoups that energy, allows increased efficiency of N<sub>2</sub> fixation and, in principle, increased plant biomass yields –. This symbiotic nodule H<sub>2</sub> recycling capability correlates with specific bacteroid strains, although host legume cultivars also contribute to H<sub>2</sub> recycling and yield. Indeed, in biochemical assays, bacteroids isolated from H<sub>2</sub> recycling (non-evolving) nodules show high levels of respiratory uptake hydrogenase activity. H<sub>2</sub> recycling during N<sub>2</sub> fixation was first observed with the aerobe Azotobacter vinelandii, a diazotroph but not a legume symbiont. In pure culture, A. vinelandii induces a particulate (respiratory) hydrogenase activity which oxidizes H<sub>2</sub> at the expense of and tolerant of O<sub>2</sub>. In following studies with legume nodule bacteroids, such uptake hydrogenase activity was also affirmed. In the ensuing forty years, hydrogenases, extensively studied, have proven both biochemically diverse and broadly distributed across bacteria and archaea. Among aerobes and microaerophiles able to use H<sub>2</sub> as energy source, uptake hydrogenase activities are typically classified as group I: heterodimeric, globular, hydrophilic proteins carrying a heteronuclear Ni,Fe-catalytic center; group I hydrogenases are generally O<sub>2</sub> tolerant. In cellular terms, the group I, Ni,Fe uptake hydrogenases are tightly associated with respiratory membranes via integral diheme b-type cytochromes, required for physiological activity, \[Bernhard\]. As the group I cell membrane-peripheral complexes face the periplasm, or cell exterior, they may be termed exo-hydrogenases. Azorhizobium caulinodans, a microaerophilic α-proteobacterium originally isolated as nodule endosymbiont of the host legume Sesbania rostrata, is capable of N<sub>2</sub> fixation both in planta and in pure diazotrophic culture. Recently, we discovered in A. caulinodans a second, novel respiratory hydrogenase encoded by the seven-gene hyq operon. The inferred Hyq hydrogenase includes six different structural proteins, including a heterodimeric Ni,Fe- catalytic center hydrogenase conserved with group I enzymes. From bioinformatic analyses, the remaining four Hyq proteins are all membrane-integral. Because all six Hyq hydrogenase subunits are NADH:quinone oxidoreductase (respiratory complex I) homologs, the Hyq complex is classified with the reversible group IV hydrogenases. Given structural and functional homology to respiratory complex I, the Ni,Fe-catalytic center heterodimers of group IV complexes associated with respiratory membranes presumably face the cell-interior and thus may be termed endo-hydrogenases. # Results ## In symbiotic legume nodules both exo- and endo-hydrogenases recycle H<sub>2</sub> produced by N<sub>2</sub> fixation To assess physiological roles for both bacteroid Hup exo- and Hyq endo-hydrogenases in symbiotic legume nodules fixing N<sub>2</sub> and recycling H<sub>2</sub>, A. caulinodans haploid derivatives carrying precise (to the nucleotide pair) in-frame deletions of hup and hyq structural genes encoding the conserved catalytic subunits of, respectively, exo- and endo-hydrogenases were constructed and verified by nucleotide sequencing of mutant loci. Specifically, A. caulinodans exo-hydrogenase null mutants carried in-frame, precise, complete hupSL deletions; endo-hydrogenase null mutants comprised precise, complete hyqRBCEFGI operon deletions. As well, haploid recombinant double-mutants carrying both exo- and endo-hydrogenase null alleles were also constructed. Pure A. caulinodans cultures were used to inoculate both stems and roots of S. rostrata seedlings aseptically germinated and individually cultivated under N-limitation. In S. rostrata, symbiotic nodules are developmentally determinate, not meristematic. While both stem- and root-nodules subsequently developed on inoculated plants only, as they are invariably absent on uninoculated plants, stem nodules were chosen for further study. Three week-old and five week-old determinate stem nodules were excised from inoculated plants and individually tested for N<sub>2</sub> fixation activity, assaying acetylene-dependent ethylene production by gas chromatography with flame ionization detection. Excised stem nodules all showed similar (±15%) high levels of acetylene reduction activity when normalized per fresh nodule biomass. Accordingly, all A. caulinodans strains tested were assigned both nodulation-competent (Nod<sup>+</sup>) and N<sub>2</sub> fixation-competent (Fix<sup>+</sup>) phenotypes. Additional excised nodules from these S. rostrata plants were simultaneously tested under air for H<sub>2</sub> evolution activity using gas chromatography coupled to a reducing-compound photometric detector. In kinetic studies with excised nodules elicited by A. caulinodans strains 61305R (parental), 66081 (exo-hydrogenase mutant) and 66132 (endo-hydrogenase mutant), H<sub>2</sub> evolution was nonexistent. Whereas, nodules elicited by double (exo- and endo-hydrogenase) mutant 66204 evolved H<sub>2</sub> at very high rates comparable to those measured for acetylene reduction. Thus, H<sub>2</sub> evolution by double-mutant 66204-elicited nodules was quantitatively owed to N<sub>2</sub> fixation (dinitrogenase) activity. Results with five week-old determinate nodules from additional S. rostrata plants entirely corroborated results with three week-old nodules (data not presented). Pure bacterial cultures were reestablished from aseptically crushed nodules and strain identities verified by nucleotide sequencing of hup and hyq loci. In conclusion, A. caulinodans bacteroids in S. rostrata nodules employ both exo- and endo-hydrogenases to recycle endogenous H<sub>2</sub> produced by N<sub>2</sub> fixation. Moreover, H<sub>2</sub> recycling is quantitative, entirely accounting for N<sub>2</sub> fixation activities. Yet as measured by H<sub>2</sub> evolution rates, bacteroid exo- and endo-hydrogenase are interchangeable and individually are fully competent to handle endogenous H<sub>2</sub> recycling in symbiotic S. rostrata nodules. ## N<sub>2</sub> fixing, microaerophilic α-proteobacteria able to recycle H<sub>2</sub> carry exo- and endo-hydrogenase gene-sets From bioinformatic analyses, orthologous hyq<sup>+</sup> operons encoding endo-hydrogenase are generally present in N<sub>2</sub> fixing microaerophiles able to recycle endogenous H<sub>2</sub>. These strains include both free-living diazotrophs as well as certain rhizobia, such as B. japonicum, the endosymbiont of Glycine max (soy). In Rhizobium leguminosarum, a metastable species with several descendant biovars each with genomes comprised of variable multipartite replicons, H<sub>2</sub> recycling capability in symbiotic legume nodules varies among strains. As well, both the hup<sup>+</sup>/hyp<sup>+</sup> (exo-hydrogenase) and the hyq<sup>+</sup> (endo-hydrogenase) gene-sets are also variables,. Yet other diverse rhizobia (e.g. Sinorhizobium meliloti 1021; Mesorhizobium loti MAFF303099; Rhizobium etli CFN42; Rhizobium sp. NGR234) all incapable of H<sub>2</sub> recycling in symbiotic legume nodules, completely lack both hup<sup>+</sup>/hyp<sup>+</sup> and hyq<sup>+</sup> gene-sets. As the hyq<sup>+</sup> operon is also absent from anaerobic (fermentative) diazotrophs, fully aerobic diazotrophs (e.g. Azotobacter spp.), and non-diazotrophs generally, Hyq endo-hydrogenase seems co-selected with N<sub>2</sub> fixation in microaerophilic (non-fermentative) α-proteobacteria. Nevertheless, in every N<sub>2</sub> fixing microaerophile with both exo- and endo-hydrogenases, these gene-sets, as well the nif genes encoding N<sub>2</sub> fixation activities are all unlinked. Moreover, A. caulinodans haploid strains carrying complete (20-gene) hup<sup>+</sup>/hyp<sup>+</sup> (including hupSL<sup>+</sup>) operon deletions entirely lacking exo-hydrogenase and ancillary activities, nevertheless retain full H<sub>2</sub> recycling activity both in pure cultures and in S. rostrata stem nodules. As well, Rhodocista centenaria (aka Rhodospirillum centenum) SW, which possesses the hyq<sup>+</sup> operon but not the hup<sup>+</sup>/hyp<sup>+</sup> operon, completely recycles H<sub>2</sub> in diazotrophic culture (data not presented). Accordingly, these exo- and endo-hydrogenase gene-sets seem fully autonomous. # Discussion Among legume-Rhizobium symbioses, H<sub>2</sub> recycling was first reported in Pisum sativum (garden pea) nodules elicited by specific Rhizobium leguminosarum bv. viciae strains. Genetic studies were subsequently undertaken with \[Brady\]Rhizobium japonicum strains able to recycle H<sub>2</sub> in Glycine max (soy) nodules. Many subsequent studies with H<sub>2</sub> recycling legume nodules all presumed uptake hydrogenase activity a single entity. These studies include combined genetic and physiological analyses which might have challenged this assertion. For the case of A. caulinodans, single mutants W58, U58 as well as hupSL impaired strain ORS571.2 all were reported to suffer substantial to complete loss of uptake hydrogenase activity. Such conclusions are incompatible with the present finding: A. caulinodans employs two structurally and functionally distinct, genetically-independent, respiratory hydrogenases to recycle endogenous H<sub>2</sub> produced by N<sub>2</sub> fixation. Whereas, early on the investigative timeline, B. japonicum single mutants unable to be cultured autotrophically on exogenous H<sub>2</sub> yet still able to recycle endogenous H<sub>2</sub> in soy nodules were identified. As these strains showed induction of uptake hydrogenase activity in cultures shifted to O<sub>2</sub> limitation (≤11 µM DOT), they were perhaps understandably considered transcriptional control mutants hypersensitive to O<sub>2</sub>. With the benefit of hindsight, this phenotype is precisely that expected of true loss-of-function point mutants affecting hup operon structural genes encoding Hup exo-hydrogenase activity, were the observed limiting-DOT uptake hydrogenase activity in fact owed to Hyq endo-hydrogenase. In A. caulinodans, hyq operon expression requires NifA as transactivator, and the pnifA<sup>+</sup> promoter is in turn strongly transactivated by Fnr, which process requires physiological O<sub>2</sub> limitation in diazotrophic culture. In principle, both exo- and endo-hydrogenase gene-sets, despite being encoded at disparate loci in all organisms identified, might nevertheless share a common genetic predisposition, allowing strategic single mutations to convey dual loss-of-function. However, as strains carrying complete hyq operon deletions still possess wild-type Hup exo-hydrogenase activity, and vice versa, evidence for any genetic, post-transcriptional interaction or interdependence between the two gene-sets is entirely lacking. As shown previously, in pure diazotrophic (N<sub>2</sub> as sole N-source) cultures, A. caulinodans exo-hydrogenase knockout mutants grow as wild-type, whereas endo-hydrogenase knockout mutants exhibit slow growth. Are exo- and endo-hydrogenase H<sub>2</sub> recycling efficiencies in pure culture and in legume nodules then demonstrably different? Or, do diazotrophic phenotypes imply additional endo-hydrogenase function(s), e.g. chemiosmotic work associated with membrane ion translocation not undertaken by exo-hydrogenase? Obviously, effective exo- and endo-hydrogenase cellular concentrations and/or distributions might be dissimilar in legume nodules and in pure diazotrophic cultures, even though both hup<sup>+</sup>/hyp<sup>+</sup> (exo-hydrogenase) and hyq<sup>+</sup> (endo-hydrogenase) gene-sets are then strongly transcribed. Because hup mutants suffer loss of chemoautotrophy with exogenous H<sub>2</sub> as energy substrate, exo-hydrogenase kinetic behavior may constitute simple diffusion control. Because hyq mutants do not adversely impact chemoautotrophy with exogenous H<sub>2</sub>, endo-hydrogenase kinetic behavior might not constitute simple diffusion control. A critical test of this hypothesis is still lacking. Diazotrophic liquid batch cultures typically employ constant sparging with relatively high gas-phase exhaust rates (0.5 min<sup>−1</sup>), complicating kinetic behavior and analysis of cellular processes with gaseous substrate(s) subject to simple diffusion control. In such pure liquid diazotrophic batch cultures bacterial densities typically reach 10<sup>8</sup> cc<sup>−1</sup>, whereas in determinate S. rostrata nodules, bacteroid densities approach 10<sup>11</sup> cc<sup>−1</sup>, the latter obviously more conducive to endogenous H<sub>2</sub> recycling under simple diffusion control. Notwithstanding, given their apparent co-selection in N<sub>2</sub> fixing micoaerophilic α-proteobacteria capable of H<sub>2</sub> recycling, exo- and endo-hydrogenases likely possess additional, distinctive functionalities yet to be elucidated. # Methods ## Bacterial strains and culture conditions Azorhizobium caulinodans ORS571 wild-type (strain 57100), originally isolated from Sesbania rostrata stem-nodules, was cultured as previously described. As 57100 wild-type is a pyridine nucleotide auxotroph, to serve as ‘virtual’ wild- type, all experiments reported here employ A. caulinodans 61305R, a 57100 derivative carrying an IS50R insertion in the (catabolic) nicotinate hydroxylase structural gene. Precise, in-frame deletion mutants were constructed by a ‘crossover PCR’ method. Haploid exo-hydrogenase knockout mutants each carry a hupΔSL2 allele in which the (upstream) hupS translation initiation codon is fused in-frame to a synthetic 21np linker sequence fused in-frame to the (downstream) hupL termination codon. Similarly, haploid endo-hydrogenase mutants each carry a hyqΔRI7 allele, in which the hyqRBCEFGI operon has been replaced by a deletion allele comprising the hyqR initiation codon fused in-frame to the 21np linker sequence fused in-frame to the hyqI termination codon. After gene replacement, haploid strains carrying deletion alleles were verified by PCR and DNA sequencing analyses. ## Sesbania rostrata nodulation tests S. rostrata plants were germinated, cultivated aseptically, and stem inoculated with pure A. caulinodans strain cultures as described. Either three or five weeks post-inoculation, stem nodules were detached, weighed, individually placed in septated vials. Dinitrogenase activity was assayed kinetically by acetylene reduction and product ethylene was measured by gas chromatography with flame-ionization detection. H<sub>2</sub> evolution was assayed kinetically and measured by gas chromatography with reducing compound photometer detection (RCP1; Peak Laboratories LLC, Mountain View, CA.), both at atmospheric pressure and 29°C. Enzymatic activities are expressed per gram nodule fresh-biomass at 29°C. The authors thank Henk Vreman (Stanford University), Steve Hartman and Alex Lowe (Peak Laboratories LLC) for technical assistance with gas chromatographic H<sub>2</sub> analyses. [^1]: Conceived and designed the experiments: RL. Performed the experiments: CC NR JM. Analyzed the data: CC NR JM DC RL. Contributed reagents/materials/analysis tools: RL. Wrote the paper: RL. [^2]: The authors have declared that no competing interests exist.
# Introduction The ability of cancer cells to invade or metastasize to distant organs is linked to their ability to traverse the extracellular matrix. The extracellular matrix (ECM) is a heterogeneous protein matrix that consists of a variety of collagens, laminins, fibronectin and proteoglycans. Matrix metalloproteinases (MMP) cleave and remodel ECM components, and increased activity of some of the MMPs, including MMP2, 9, and 14 can promote cancer cell invasion in vivo. MMP13 (collagenase-3) was originally identified in human breast cancer tissue. It is secreted from cells as an inactive zymogen that can be activated by the MMP2/MMP14/tissue inhibitor of metalloproteinase (TIMP2) complex or by plasmin. MMP13 acts in the extracellular environment as a potent collagenase capable of degrading a variety of collagens. Mmp13 mRNA is expressed in a subpopulation of myofibroblasts in invasive ductal breast carcinomas, but rarely in normal breast, benign breast lesions and ductal carcinoma in situ (DCIS). Interestingly, the presence of microinvasion in DCIS is associated with focal expression of Mmp13 mRNA in stromal fibroblasts. Direct comparison of the Mmp13 mRNA expression pattern with that of the Mmp2, Mmp11 and Mmp14 mRNAs indicates that Mmp13 is unique in this respect since these other MMPs are also present in DCIS in the absence of invasion. These observations raise the question as to whether MMP13 is a rate-limiting proteinase that mediates the initial steps in breast cancer invasion. The role of MMPs and other ECM-degrading proteinases during breast cancer progression has been studied using a variety of murine mammary tumor models. In one model, tumors are induced by expressing the polyoma middle T oncogene under the mouse mammary tumor virus LTR (MMTV-PyMT) leading to early onset hyperplasia in the mammary glands. The PyMT-oncogene induced mammary tumors go through typical pre-invasive, invasive and metastatic phases similarly to human breast cancer. The onset of tumorigenesis in the MMTV-PyMT mouse model is observed at 2–3 weeks after birth, and the non-invasive early stage tumors, mammary intraepithelial neoplasia (MIN), progress into invasive carcinomas at 7–9 weeks after birth. Transition to invasive carcinoma takes place in the central core of the tumors, where epithelial atypia and nuclear pleomorphism first appear. Tumors isolated from 13-week-old mice, late stage tumors, show evidence of dedifferentiation and squamous metaplasia and have spread to the lungs and lymph nodes. The tumors share both morphological and molecular characteristics with human invasive ductal carcinomas,. Furthermore, we have reported that Mmp2, 3, 11, 13 and 14 mRNAs are expressed in the stromal compartment in a series of late stage MMTV-PyMT tumors, similar to the expression in human ductal breast cancers. In late stage tumors, Mmp13 mRNA shows high, focal expression in stromal cell populations. Since this expression pattern is similar to what we have observed in human breast cancers, we chose the MMTV-PyMT mouse breast tumor model to test the functional role of MMP13 in breast cancer progression and metastasis. # Materials and Methods ## Mice MMTV-PyMT and Mmp13<sup>−/−</sup> mice have previously been described. Both strains were backcrossed more than 8 generations to FVB/n. The two strains were intercrossed and the resulting MMTV-PyMT;Mmp13<sup>+/−</sup> male offspring were mated with female Mmp13<sup>+/−</sup> mice to establish sibling cohorts of MMTV-PyMT;Mmp13<sup>+/+</sup> and MMTV-PyMT;Mmp13<sup>−/−</sup> females. All animal experiments were conducted according to institutional guidelines and approved by the Danish Animal Experiments Inspectorate. A concurrent health report compliant with the guidelines of the Federation of European Laboratory Animal Science Associations revealed no infections. ## Genotyping Genotyping was performed on chromosomal DNA purified from tail tips in a single PCR reaction using MMTV-PyMT primers, control wt plasminogen primers and Mmp13 primers (MMP13_in5 anti: GGT GGT ATG AAC AAG TTT TCT GAG C, MMP13_in2: CAG ACC CTA CAG TGC CAG ATT TTA G, MMP13_ex5: TGA TGA CGT TCA AGG AAT TCA GTT T). Bands representing MMTV-PyMT transgene (159 bp), plasminogen (268 bp), wt Mmp13 (572 bp) and Mmp13 knock-out allele (672 bp) were identified by agarose gel electrophoresis. ## Quantitative real time PCR Mice were anesthetized by intraperitoneal administration of a 1∶1 mixture of Dormicum (Roche A/S, Basel, Switzerland) and Hypnorm (Janssen-Cilag Ltd, High Wycombe, UK) and sacrificed by intracardial perfusion with 10 ml ice-cold phosphate-buffered saline (PBS). For analysis of Mmp13 mRNA expression during normal mammary gland development and MMTV-PyMT tumorigenesis, the \#4 mammary glands were isolated from MMTV-PyMT mice and wild type FVB/n mice at 3, 5, 7, 9, 11 and 13 weeks (n = 5 for each age and genotype). For analyses of the expression of other MMPs and TIMPs, tumors from 11-week old MMTV- PyMT;Mmp13<sup>+/+</sup> (n = 6) and MMTV-PyMT;Mmp13<sup>−/−</sup> mice were used (n = 6). Total mRNA was isolated from 100 mg tissue using the Qiagen RNeasy midi Kit (Qiagen, West Sussex, UK). The purified RNA samples were treated with DNase I for 15 minutes at 37°C to remove residual DNA. Reverse transcription was performed essentially as described followed by TaqMan qPCR performed as described using an ABI7900 (Applied Biosystems, Foster City, CA) and primers specified in except that Mmp13 primers were chosen from the universal probe library (number 105, Roche Applied Science, Burgess Hill, UK), left: CTT TTC CTC CTG GAC CAA ACT, right: TCA TGG GCA GCA ACA ATA AA. Gene expression levels relative to 18S were calculated. ## In situ hybridization 13-week-old mice were anesthetized as above and perfused intracardially with 10 mL PBS followed by 10 mL of 4% paraformaldehyde (PFA). Mammary tumors were dissected from glands \#1–3 and \#4, fixed in 4% formalin for 5–7 days at 4°C and paraffin embedded. In situ hybridization was performed on samples from MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 10) with <sup>35</sup>S labeled RNA probes for MMP13 as described previously. No Mmp13 mRNA signal was detected in tumors from MMTV-PyMT;Mmp13<sup>−/−</sup> mice (n = 4). ## Immunohistochemistry For immunohistochemical staining of α-smooth muscle-actin (mouse mAb, Dako, Glostrup, Denmark) and CD204 (rat mAb clone 2F8, Serotec, Oxford, UK), sections were heat-treated at 98°C in TEG buffer (10 mmol/L Tris, 0.5 mmol/L EGTA, pH 9.0) for 5 minutes in a T/T Micromed microwave processor (Milestone, Sorisol, Italy). The α-smooth muscle-actin antibody was mixed with biotinylated anti- mouse Fab fragments using the Animal Research Kit (Dako) and used for staining following the recommendations of the manufacturer. The rat mAb against CD204 was detected with rabbit anti-rat antibodies followed by Envision-Rabbit reagent (Dako). All antibody incubation steps were performed in an OptiMax automated immunostainer (BioGenex, San Ramon, CA). Immunohistochemical staining combined with in situ hybridization was performed on mouse glands \#1–3 tumor samples from MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 6) as described previously. The sections were developed with DAB chromogene, fixed for 40 minutes in 4% PFA, and washed three times with sterile water. After dehydration, the sections were incubated with <sup>35</sup>S-labeled Mmp13 antisense probes and further treated according to the in situ hybridization as described previously. ## Morphometric analysis of mammary gland whole mounts The \#4 mammary glands were isolated from 4- and 6-week-old MMTV- PyMT;Mmp13<sup>+/+</sup> mice (n = 4 and n = 4, respectively) and MMTV- PyMT;Mmp13<sup>−/−</sup> mice (n = 3 and n = 7, respectively) and prepared for whole mount as described. The whole mounts were photographed using a DFC320 CCD camera (Leica Microsystems, Wetzlar, Germany) and the ductal epithelial invasion determined by measuring the distance from the branch point closest to the nipple to the three most far-reaching ducts and the edge of the fat pad. The penetration is presented as the mean of the three measures of ductal penetration related to the length to the edge of the fat pad. All measurements were done using the Leica IM500 software. ## Tumor growth and lung metastasis Tumor growth in MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 38) and MMTV- PyMT;Mmp13<sup>−/−</sup> mice (n = 25) was followed by weekly palpations of all 10 mammary glands. The length and width of all palpable tumors were measured by caliper. The tumor volume (assuming that tumors took the shape of an ellipsoid) was calculated using the formula: V = (π/6)×W<sup>2</sup>×L, where L = length and W = width. At 13 weeks, MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 35) and MMTV-PyMT;Mmp13<sup>−/−</sup> mice (n = 25) were anesthetized and perfusion fixed (three mice had been terminated before due to extensive tumor burden). The lungs were removed for determination of the metastasis burden by stereological analysis using Cavalieri's principle as described previously. ## Tumor grading Nuclear grading of the MMTV-PyMT tumors was performed on haematoxylin and eosin (H&E) stained sections from the \#1–3 tumors of MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 34) and MMTV-PyMT;Mmp13<sup>−/−</sup> mice (n = 24). We defined grade I tumors as tumors with adenoma and atypia in Mammary Intraepithelial Neoplasia (MIN)-like foci and containing up to 5% of MIN with grade II nuclear pleiomorphy. Grade II tumors (“Early carcinoma”) were tumors with widespread invasive carcinoma with grade II nuclear pleiomorphy and contained areas with MIN Grade I/II. Grade III tumors (“Late carcinoma”) were tumors with focal or widespread invasive carcinoma with grade III nuclear pleiomorphy that also contained areas with Grade I and II carcinoma. Most tumors with grade III carcinoma showed areas of squamous metaplasia. ## Quantification of angiogenesis Mice were anesthetized as above and 100 µL of 1 mg/mL FITC-conjugated Lycopersicon esculentum (tomato) lectin (Vector labs) injected i.v. in the tail vein. After 5 min, the mice were cardiac perfused with 4% PFA in PBS at 120–140 mm Hg until the flow-through fluid was clear. Tumors were dissected and post-fixed in 4% PFA overnight, incubated in increasing sucrose concentrations (12%, 15%, 18% in PBS), embedded in Tissue-Tek O.C.T. compound and thick sections (30–40 µm) cut. The tissue sections were stained with 1∶500 solution of Propidium Iodide in PBS (Molecular probes, P-3566) for 1 hour and the slides were mounted with Gel/Mount (Biomeda, \#M01). Confocal image stacks with 1-µm spacing were collected from the most vascularized area in the middle of the tumors using a 20x lens with a Solamere micro-lensed spinning disk confocal microscope (Solamere Technologies, Salt Lake City UT) equipped with an intensified charge-coupled device (ICCD) camera (XRMega-10EX S-30, Stanford Photonics, Palo Alto, CA). The acquired images were analyzed with Bitplane Imaris version 5.5 for Windows software using ‘filament tracer’ with an approximate filament diameter set at 1.5. Minor manual corrections of the computerized filaments were uniformly done. Vessel (segment) average diameter, length and volume, were exported to a Microsoft Excel file and analyzed using GraphPad 4 statistical software. Eight tumors from MMTV- PyMT;Mmp13<sup>+/+</sup> and nine tumor from MMTV-PyMT;Mmp13<sup>−/−</sup> were analyzed. All image acquisition and analysis was done blindly. ## Cell proliferation analysis MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 19) and MMTV- PyMT;Mmp13<sup>−/−</sup> mice (n = 19) were injected i.p. with 2 mg bromodeoxyuridine (BrdU, Sigma-Aldrich) 2 hours before euthanasia. The mice were anesthetized, perfusion fixed, the tumors isolated and fixed in 4% PFA and finally paraffin embedded as described above. For immunohistochemistry, we used a rat mAb against BrdU at 2 µg/mL (clone BU1/75-ICR1, Abcam, Cambridge, UK). The BrdU labeled cells were counted as a percentage of the total number of epithelial tumor cells using CastGRID (Visiopharm, Hørsholm, Denmark) essentially as previously described. In brief, frames of 9391 µm<sup>2</sup> were systematically but randomly placed over the BrdU stained sections with a final magnification of 658 on the monitor. All BrdU positive neoplastic cells within the frames were counted to give total numbers varying from 250–400. The total number of neoplastic cells within the frames was estimated based on the usage of systematic and randomly positioned reference points. ## Collagen evaluation Serial sections (3-µm thick) from mammary tumor glands \#1–3 obtained from 13-week-old MMTV-PyMT;Mmp13<sup>+/+</sup> mice (n = 26) and MMTV- PyMT;Mmp13<sup>−/−</sup> mice (n = 21) were stained with haematoxylin and eosin (H&E) or picrosirius red (PSR). Areas with invasive tumor were identified in H&E stained sections. The same areas were then analyzed in the PSR stained sections by obtaining 2.3×1.8 mm images with a charge-coupled device (CCD) camera (Leica) under linearly polarized light using a Leica DMRBE microscope. In linearly polarized light, thick collagen bundles appear orange-red and thin collagen fibers green. Images were analyzed with MetaMorph software for collagen content. Total collagen was identified with Hue-Saturation-Intensity (HSI) intervals 0–256, 1–256 and 70–256, respectively. For thin fibrillar collagen, HSI settings of 26–83, 70–135, and 57–155 were applied to all the images. All image acquisition and analysis was done blindly. The area covered by total collagen, thin fibrillar collagen, and the fraction of thin fibrillar collagen was calculated and used for statistical analysis (t-test). ## Statistics For statistical analyses of the tumor burdens obtained by caliber measurements, the total tumor volumes were logarithmically transformed and analyzed using a general linear model with repeated measures. Only measurements obtained from week 7 to 13 were included in this analysis. Estimates were obtained using generalized estimating equations (SAS v9.1). Statistical interaction between genotype and time were evaluated for the entire observation period considering repeated measures (SAS). For statistical analyses of the metastasis burdens, the values were logarithmically transformed and analyzed using a general linear model. # Results ## Mmp13 is expressed by myofibroblasts in invasive stage MMTV-PyMT tumors Mmp13 mRNA is expressed by a subpopulation of myofibroblasts in human invasive ductal breast carcinomas, but is rarely expressed in normal breast, benign breast lesions and ductal carcinoma in situ (DCIS). To determine if the expression level of MMP13 changes during tumor progression in the MMTV-PyMT mouse model, Mmp13 mRNA levels were assessed in tumor tissue from mice at 3-13 weeks of age by qPCR. The level of Mmp13 mRNA was highly increased in MMTV- PyMT tumors after 7-weeks of age, with the highest levels found in older mice, corresponding to late stage tumors. Mmp13 mRNA levels were low in normal mammary glands at all ages. The increase in Mmp13 mRNA thus took place during the time of transition to invasive stages in this tumor model. To determine where Mmp13 was expressed, we performed in situ hybridization in tumors from 13-week-old mice, containing different stages of nuclear grading. The Mmp13 mRNA signal was seen in stromal fibroblast-like cells located in the tumor core in areas with invasive grade II and III carcinoma (“late carcinoma”), but was generally absent in the tumor periphery that was dominated by normal- looking mammary tissue, hyperplasia or MIN grade I and II. In particular, Mmp13 mRNA positive cells were often seen in stromal branches located close to the tumor cores with focal necrosis. Sporadic Mmp13 mRNA positive single cells and small cell clusters were also found in areas with strongly keratinized, squamous metaplastic changes. This focal expression of Mmp13 in the tumor core was very different from the expression of Mmp2 and Mmp14, which were found throughout the tumor stroma, including at the tumor periphery ( e–h), and in areas with hyperplasia, grade I and grade II MIN (not shown). The localization of the cells that expressed Mmp13 suggested that they were myofibroblasts. Using combined in situ hybridization and immunohistochemical staining, we verified that most of the Mmp13 mRNA positive cells were positive for the myofibroblast marker α-smooth muscle-actin, paralleling the expression we found in human breast cancers. Myofibroblasts are prevalent in the invasive MMTV-PyMT tumors, but only a small subpopulation was positive for Mmp13 mRNA. We noted that some of the Mmp13 mRNA positive myofibroblasts were located in narrow bands between tumor cell clusters. CD204 positive macrophages did not express Mmp13 mRNA and the location of the Mmp13 mRNA positive cells was not consistent with myoepithelial cells (data not shown), supporting our conclusion that the stromal Mmp13 mRNA expressing cells were myofibroblasts. ## Early tumor formation does not require MMP13 in MMTV-PyMT The expression levels and patterns of Mmp13 mRNA suggested that the proteinase would play a role in the transition to the invasive stages. To determine the role of MMP13 in mouse mammary tumor progression, we took a genetic approach using Mmp13<sup>−/−</sup> mice. We verified that the absence of MMP13 did not affect the development of the mammary epithelium and therefore would not affect tumor progression indirectly by influencing the number of mammary epithelial cells. We next crossed the MMTV-PyMT mice with Mmp13<sup>−/−</sup> mice and measured the total hyperplastic area on whole-mounts of mammary glands from mice at age 4 or 6 weeks, before the transition to invasive carcinoma. No difference was found between MMTV-PyMT;Mmp13<sup>+/+</sup> and MMTV- PyMT;Mmp13<sup>−/−</sup> mice , consistent with the lack of Mmp13 mRNA expression at the early cancer stage. ## MMP13 is not required for tumor growth, tumor vascularization or metastasis in MMTV-PyMT mammary carcinomas Mmp13 mRNA was induced in tumors from mice of 7–9 weeks-of-age, and we therefore expected that MMP13 would be required in the later stages of tumorigenesis, possibly affecting tumor growth and metastasis. We followed tumor growth in a cohort of littermate MMTV-PyMT;Mmp13<sup>+/+</sup> (n = 38) and MMTV-PyMT;Mmp13<sup>−/−</sup> females (n = 25) by weekly palpations and caliper measurements until 13 weeks of age, when tumors were isolated for histological examination and the lungs isolated to determine metastasis burden. A minor reduction in the growth of the tumors in MMTV-PyMT;Mmp13<sup>−/−</sup> mice were seen at 8, 9, 11, 12 weeks, but this was not reproduced in a different cohort. Furthermore, there were no differences in cancer cell proliferation between tumors from MMTV-PyMT;Mmp13<sup>+/+</sup> and MMTV- PyMT;Mmp13<sup>−/−</sup> mice at 8 or 11 weeks of age as determined by BrdU- labeling. In addition, histological examination and nuclear grading of the tumors revealed no significant differences in the percentages of tumors that had progressed to grade III, late-stage carcinoma (8% of the MMTV- PyMT;Mmp13<sup>−/−</sup> vs. 18% of MMTV-PyMT;Mmp13<sup>+/+</sup> tumors). Furthermore, analyses of H&E stained sections revealed no differences in the presence of necrosis or stroma between tumors from MMTV- PyMT;Mmp13<sup>−/−</sup> and MMTV-PyMT;Mmp13<sup>+/+</sup> mice (not shown). Finally, no effects were seen on tumor vascularization, when analyzing vessel diameter, length, or volume in 3D reconstructions of vessels in tumors of lectin-perfused mice. Thus, the absence of MMP13 did not influence progression in the primary tumors. Since MMP13 is upregulated concomitantly with the onset of cancer cell dissemination in the MMTV-PyMT model, we determined the metastatic burden in the lungs from end-stage MMTV-PyMT;Mmp13<sup>+/+</sup> and MMTV- PyMT;Mmp13<sup>−/−</sup> mice using a morphometric approach. Although there was a trend for a reduced lung metastatic volume in the absence of MMP13, this was not significant (.2 mm<sup>3</sup> for MMTV-PyMT;Mmp13<sup>+/+</sup>, n = 35; 0.7 mm<sup>3</sup> for MMTV-PyMT;Mmp13<sup>−/−</sup>, n = 25, p = 0.34). We therefore conclude that the absence of MMP13 does not affect lung metastasis in the MMTV-PyMT breast cancer model. ## There are no compensatory changes in expression of MMPs or TIMPs in MMTV-PyMT;Mmp13<sup>−/−</sup> tumors Since we found no changes in tumor progression in the absence of MMP13, we explored the possibility that the lack of MMP13 activity was compensated for by changes in expression of other MMPs or in their inhibitors, the TIMPs. However, there were no significant differences in the mRNA levels for MMP2, -3, -8, -9, -10, -11, -12, -13, -14 or TIMP1, -2, -3, or -4 in tumors with and without MMP13 from 11-week-old mice by qPCR. ## Increased amounts of thin collagen fibers is found in the absence of MMP13 in invasive areas of MMTV-PyMT tumors To determine whether any effects of MMP13 could be seen on its major substrate, fibrillar collagen, we evaluated the fibrillar collagen content in invasive areas of the tumors by picrosirius red staining of tumor sections. Although total collagen was not affected by the absence of MMP13, the fraction of thin collagen fibers relative to total collagen was increased significantly by ∼30% in the invasive areas of tumors from MMTV-PyMT;Mmp13<sup>−/−</sup> mice (t-test, p = 0.007). Thus, MMP13 played a role in collagen metabolism in invasive areas of MMTV-PyMT tumors, but this did not affect tumor progression. # Discussion In this study, we analyzed the effects of absence of MMP13 activity on the progression of MMTV-PyMT mammary carcinoma. Our previous studies of human DCIS with early invasion led us to hypothesize that MMP13 would be a fibroblast- derived, potential rate-limiting proteinase in the transition from non-invasive to invasive breast carcinoma. Interestingly, a recent study reported upregulation of several MMPs, including MMP13, at the transition from non- invasive to invasive breast cancer cells in an in vitro model. However, although Mmp13 mRNA also was strongly induced in a subpopulation of myofibroblasts in MMTV-PyMT mammary carcinomas, concurrent with early transition to invasive carcinoma, there were no effects on tumor cell proliferation, tumor growth, lung metastasis, vascularization or differentiation of the primary tumors by the absence of MMP13. We excluded that the lack of effects was due to compensatory transcriptional changes of other MMPs or their inhibitors expressed in the MMTV-PyMT tumors. MMP13 has a broad substrate profile that includes several fibrillar collagens, tenascin, fibronectin and proteoglycans. Mmp13 mRNA was found in carcinoma- associated fibroblasts located in the central areas of the MMTV-PyMT tumors, where tumor dedifferentiation and invasion is prominent. In these areas, we found increased thin collagen fibers relative to total collagen in MMTV- PyMT;Mmp13<sup>−/−</sup> compared to MMTV-PyMT;Mmp13<sup>+/+</sup> tumors, indicating that MMP13 influences collagen fiber formation and/or metabolism of fibrillar collagen in the MMTV-PyMT stroma. Mice lacking Mmp13 display no major phenotypic abnormalities, but a significant contribution of MMP13 in collagen metabolism has been reported during bone development, skeletal regeneration and in atherosclerotic plaques. We found that most of the Mmp13 mRNA expressing cells in the MMTV-PyMT tumors are carcinoma-associated myofibroblasts. The myofibroblast is an abundant cell type in the MMTV-PyMT tumors (this study) as it is in invasive human breast cancers. There are several reports suggesting that carcinoma-associated myofibroblasts promote cancer progression. In addition, proteins expressed by myofibroblasts, such as PAI-1 and uPA, are associated with poor prognosis for breast cancer patients. It is unknown what induces Mmp13 in the breast carcinoma-associated myofibroblasts, but the restricted expression of Mmp13 in dense fibrous stromal septae within the central parts of the invasive MMTV-PyMT tumors suggests a locally expressed tumor-cell derived factor. In vitro studies have shown that IL-1α, IL-1β, and transforming growth factor-β1, in particular, can induce Mmp13 mRNA expression in fibroblasts. Interestingly, TGFβ signaling through fibroblast-expressed TGFβ receptor type II has a considerable impact on tumor growth. Mmp13 mRNA expressing myofibroblasts are often found in the tumor core close to areas with necrosis suggesting that MMP13 may be upregulated in response to hypoxia for stimulation of angiogenesis. Indeed, Mmp13 mRNA is induced by hypoxia in various cell lines – including fibroblasts , and vasculature of ossification sites and migration of endothelial cells into cartilage is impaired in Mmp13<sup>−/−</sup> mice. However, our measurements of vascular diameter, length and volumes in MMTV-PyMT tumors did not reveal any effect of MMP13 on tumor vascularization. It has been reported that Mmp8 and Mmp9 mRNAs are upregulated during bone development in the ossification sites in Mmp13<sup>−/−</sup> mice and that Mmp8 mRNA is upregulated during skin wound healing in Mmp13<sup>−/−</sup> mice. We tested whether the mRNA expression levels of several MMPs were changed in MMTV-PyMT;Mmp13<sup>−/−</sup> tumors, but found no significant changes. We found that tumor progression in the MMTV-PyMT model was associated with strong upregulation of the Mmp13 mRNA level as compared to the normal mammary gland. Preliminary results suggest that Mmp10 and Mmp12 mRNA also are upregulated, while Mmp2, -3, -8, -9, -11, and -14 are not (TXP, CP, DE, unpublished data). This suggests that the three MMPs are upregulated by a common mechanism and potentially implicates MMP10 and MMP12 in tumorigenesis in the MMTV-PyMT model even though MMP10 and MMP12 mRNAs were not significantly upregulated in the MMTV-PyMT;Mmp13<sup>−/−</sup> tumors compared to MMTV-PyMT wild type tumors. However, we cannot exclude that differences exist at the level of enzyme activity or that other classes of proteolytic enzymes compensate for the absence of MMP13. The expression of Mmp13 mRNA is strongly restricted compared to the expression of Mmp2 and Mmp14 mRNAs that are expressed in a broader group of (myo)fibroblasts. We have previously reported that mRNAs for several MMPs, including Mmp2, Mmp3, Mmp11 and Mmp14 are expressed in the MMTV-PyMT tumor stroma. Of these MMPs, MMP2 and MMP14 may have substrates in common with MMP13, for example, both are also collagenases in vivo. Therefore MMP2, MMP14 or the activity of other MMPs may compensate for the lack of MMP13 in the MMTV- PyMT;Mmp13<sup>−/−</sup> tumors. It is possible that MMP13 plays both promoting and inhibiting roles during cancer progression, resulting in a neutral net effect in its absence in the MMTV-PyMT model. Tumor promoting and inhibiting functions have been found for MMP9 in the human papilloma virus (HPV) 16 skin cancer model where the absence of Mmp9 reduced the number of tumors, but the tumors that formed were more aggressive. A tumor protective role of a number of matrix degrading proteases, including MMP8 and MMP12, has been described. These studies emphasize the need for better understanding the role not only of the individual proteases but also of their separate activities in cancer progression. Breast cancer progression in humans may have a larger contribution from the stromal environment than murine tumors, since human tissue have a more fibroblast-rich stroma than murine. Nevertheless, the MMTV-PyMT breast cancer model shares many histological and molecular characteristics with human luminal breast cancer, but it differs from human ductal carcinoma in one interesting aspect: the myoepithelial cells, which represent a cellular barrier for tumor cells and are an important source for basement membrane, are partly lost early during MMTV-PyMT oncogene-induced tumor progression. The absence of an intact myoepithelial barrier may lead to pre-invasive precursor lesions different from human DCIS and render MMP13 a dispensable proteinase in the critical transition phase from non-invasive to invasive carcinoma. Thus, despite the presence of a heterogeneous stromal environment and the appearance of a considerable myofibroblast population in the MMTV-PyMT tumors, myofibroblasts may play a minor role for the overall progression of the MMTV-PyMT tumor compared to human breast cancer. Therefore, models in which transition from DCIS to invasive carcinoma occurs similar to in human tumors are needed to definitely rule out a role of MMP13 in breast cancer. In conclusion, we found that the absence of MMP13 in the MMTV-PyMT mice did not result in any differences in tumor progression to invasive and metastatic breast carcinoma. Carcinoma-associated fibroblasts have been shown to facilitate tumor progression through stimulation of angiogenesis, growth factor secretion and invasion. However, there are no good markers allowing systematic sub- classification of these cells. Thus, even though MMP13 itself apparently is dispensable for tumor progression in the MMTV-PyMT model, its restricted expression in myofibroblasts in invasive regions suggests that it marks a subpopulation of carcinoma-associated fibroblasts that are involved in the transition to invasive carcinoma. Charlotte Lønborg, Lotte Frederiksen, Öznur Turan and Birthe Larsen for excellent technical assistance and John Post for photographic assistance. [^1]: Conceived and designed the experiments: BSN ME ZW LRL. Performed the experiments: BSN ME CJP TXP LRL. Analyzed the data: BSN ME FR HAA CJP TXP IJC LRL. Contributed reagents/materials/analysis tools: DRE ZW. Wrote the paper: BSN ME. Critically revised the manuscript: DRE ZW LL. [^2]: Current address: Exiqon A/S, Diagnostic Product Development, Vedbæk, Denmark [^3]: Current address: Department of Clinical Biochemistry, Rigshospitalet, Copenhagen, Denmark [^4]: The authors have declared that no competing interests exist.
# Introduction Long-term cryopreservation of female genetic material could be very useful for assisted reproductive technologies including breeding programs. However, obtaining viable embryos from cryopreserved oocytes has proved to be a difficult and inefficient labor in most mammals. Vitrification is a cryopreservation procedure characterized by a high concentration of cryoprotectants and high cooling rate that mostly prevents the formation of ice crystals and replaces, at least for oocytes and embryos, the standard method of slow freezing. Unlike slow freezing, in which extracellular water crystallizes generating an osmotic gradient that dehydrate cells, in vitrification, extra and intracellular compartments vitrify after cellular dehydration. Different vitrification devices have been designed in order to decrease the vitrified volume, thus increasing the cooling rate and reducing the exposure to cryoprotectants to minimize its toxic and osmotic hazardous effects (for review, see). Still, in spite of the several advances in the field, the development of oocyte vitrification in cattle seems to have reached a plateau. In line with this, although similar rates of survival and development have been achieved by different methods, the rates of development of bovine oocytes after vitrification are still low. A major site of injury during mammalian oocyte cryopreservation is the plasma membrane. Part of the damage caused at the membrane level is induced by the change from a fluid state to an ordered state as temperature is reduced below the transition temperature of the membrane. Graham and co-workers prevented this damage by the addition of cholesterol to sperm membrane of different domestic animals in order to increase membrane fluidity at low temperatures and make it more tolerant to cryopreservation. To add cholesterol to the membrane and to increase its cholesterol level, they used methyl-β-cyclodextrin (MβCD), a cyclic oligosaccharide which has a high affinity for inclusion of cholesterol and other sterols in its hydrophobic cavity. Modulation of plasma membrane cholesterol to increase post-cryopreservation survival is currently a booming topic for the male gamete of different species (for review, see) but has been little addressed in oocytes. Previously, Horvath and Seidel and, more recently, Spricigo et al. reported the use of this method to add cholesterol to bovine oocytes prior to vitrification. Horvarth and Seidel claimed that cholesterol could enter mature oocytes through the cumulus cells and the zona pellucida, improving its viability after vitrification in a chemically defined system. Spricigo et al. improved nuclear maturation after treating immature bovine oocytes with cholesterol loaded MβCD prior to vitrification. However, these effects were not reflected in the production of embryos in vitro. Whereas Horvath and Seidel analyzed the presence and absence of serum in their assays, Spricigo et al. used fetal calf serum in order to load MβCD with cholesterol from serum, and also during the recovery of cumulus-oocyte complexes (COCs), in vitro maturation and vitrification. More recently, rabbit oocytes with increased cholesterol content by exposure to MβCD/cholesterol complexes were found to fail to improve their developmental competence after either vitrification or slow freezing. In none of these previous studies was the possibility of removing the added cholesterol analyzed after warming to recover membrane conditions prior to vitrification. Previous research has shown that whereas an excess of cholesterol does not affect meiotic progression, it induces mouse egg activation and it may cause female infertility. Conversely, a sub-physiological level of cholesterol in mouse oocytes causes a delay in second polar body extrusion and low fertilization rates. Cholesterol mediates membrane curvature during fusion events and recent advances start to reveal common mechanisms also in gamete interaction. It is clear that abnormal levels of membrane cholesterol could adversely affect fertilization and subsequent embryo development. Moreover, membrane cholesterol segregates laterally in small domains (10–200 nm) rich in sterols (cholesterol) and sphingolipids (sphingomyelin, gangliosides) known as membrane rafts. Gangliosides are glycosphingolipids that contain sialic acid in their structure and, particularly, ganglioside GM1 has been found in the mouse oocyte and in the cleavage furrow of embryos. These specialized microdomains permit plasma membrane sub-compartmentalization and formation of signaling platforms that mediate physiological responses. It has been shown that membrane raft integrity is necessary to efficiently accomplish fertilization in the mouse oocyte. However, the importance of preserving membrane rafts in postcryopreserved bovine oocytes as well as in mammalian oocytes in general still remains unexplored. Exploring the effect of vitrification on membrane rafts may be particularly relevant considering the important role they play in cell signaling and gamete interaction. In order to increase the survival of bovine oocytes after cryopreservation, we proposed not only to increase cholesterol level of oocyte membranes before vitrification but also to remove the added cholesterol after warming to recover its original level. In this respect, MβCD is a unique tool to modulate membrane cholesterol in living cells. The high affinity of MβCD for cholesterol can be used not only to generate cholesterol inclusion complexes that donate cholesterol to the membrane but also to remove cholesterol from biological membranes when used free of cholesterol. The objectives of this study were to: (1) modulate cholesterol from membranes of bovine oocytes during vitrification, (2) determine the time of exposure to MβCD at which oocyte viability is not compromised, (3) quantify cholesterol incorporation by oocytes and cumulus cells, and (4) evaluate possible disturbance of membrane organization caused by cryopreservation through the analysis of the raft marker GM1 in living oocytes. Further understanding how bovine oocytes modulate cholesterol under these experimental conditions is a starting point to increase cryotolerance and to improve developmental competence of vitrified-warmed oocytes. # Materials and methods ## 1- Chemicals and reagents All chemicals and reagents were purchased from Sigma-Aldrich (St. Louis, MO, USA) unless otherwise stated. ## 2- Oocyte collection and in vitro maturation ### 2.1 Oocyte collection Bovine ovaries from cycling beef heifers (Bos taurus) were collected from local slaughterhouses and transported within 60–90 minutes in a thermic container to the laboratory. COCs were aspirated from follicles ranging from 2 to 8 mm in diameter by a vacuum system. COCs with homogeneous ooplasm and more than four complete layers of cumulus cells, corresponding to grades 1 and 2 according to de Loos et al., were selected under a stereomicroscope and washed 3 times in phenol red-free HEPES buffered Synthetic Oviductal Fluid (H-SOF) supplemented with 1% polyvinyl alcohol (PVA) (w:v). ### 2.2 In vitro maturation Selected COCs were incubated in four-well culture plates (NUNC, Thermo Fisher Scientific, Loughborough, Leicestershire, UK) in groups of 60 per well, with 400 μl of serum- and gonadotropin-free maturation medium: M199 with 0.1 mg/ml L-glutamine and 2.2 mg/ml NaHCO<sub>3</sub> supplemented with 10 ng/ml epidermal growth factor (EGF), 30 μg/ml hyaluronic acid, and 100 μM cysteamine as anti- oxidant. COCs were incubated for 20 hours at 38.5°C under 5% CO<sub>2</sub> in humidified air. When necessary, cumulus cells were removed either mechanically or by a brief exposure to 1 mg/ml hyaluronidase (Type IV-S from bovine testes), and zona pellucida (ZP) was dissolved with 1 mg/ml pronase (protease from Streptomyces griseus) under visual monitoring. ZP-free oocytes were rapidly washed 5 times and kept at 38.5°C under 5% CO<sub>2</sub> in humidified air for 1-hour recovery. Experimental groups were defined as COCs matured in chemically defined medium, namely (1) without treatment (control), (2) loaded with cholesterol the last 45 minutes or 2 hours of maturation, (3) loaded with cholesterol the last 45 minutes or 2 hours of maturation and subsequently depleted of cholesterol for the same periods of time. ## 3- Cholesterol enrichment and depletion with methyl-β-cyclodextrin MβCD complexed with cholesterol or alone was used to enrich and remove cholesterol from cell membranes, respectively. A stock solution of MβCD (0.5 M) in M199 was stored at 4°C in a glass tube. MβCD/cholesterol complexes (molar ratio 8:1) were prepared in M199 according to Christian et al.. Briefly, cholesterol in chloroform:methanol 1:1 (v:v) was completely dried under a stream of nitrogen. A 15 mM MβCD aqueous solution was subsequently added to the dried material. The mixture was clarified by vigorous mixing, sonicated in bath sonication for 1–3 minutes and incubated in a rotating water bath at 37°C overnight. Before using the solution, it was filtered through a 0.45 μm syringe filter and equilibrated at the incubator. COCs with partially removed cumulus cells (2–3 cumulus cell layers) and the bulk of cumulus cells were incubated separately with 15 mM MβCD/cholesterol for 45 minutes or 2 hours at 38.5°C under 5% CO<sub>2</sub> in humidified air during the last period of in vitro maturation. In addition, in order to remove the added cholesterol from membranes, COCs were incubated with different concentrations of “empty” MβCD for 45 minutes or 2 hours. Before incubation with fluorescent probe Amplex® Red, oocytes were completely denuded. ## 4- Vitrification and warming COCs with partially removed cumulus cells (2–3 cumulus cell layers) were vitrified with the surface device Cryotech® (ex Cryotop®) using vitrification and warming solutions of known composition. These protein free solutions were adapted from the literature available replacing serum by the synthetic polymer PVA. Phenol red- and calcium-free H-SOF supplemented with 1% PVA was used to handle COCs and as base medium to prepare vitrification and warming solutions. Vitrification was performed following the protocol of Zhou et al. with some modifications. Briefly, COCs with 2–3 cumulus cell layers were vitrified with heated stage at 39°C (except for cholesterol loaded COCs that were vitrified at room temperature) and vitrification solutions at room temperature (25–27°C). COCs were equilibrated for 10 minutes in Equilibration Solution (ES) with 7.5% ethylene glycol and 7.5% DMSO and vitrified in Vitrification Solution (VS) with 15% ethylene glycol, 15% DMSO and 0.5M sucrose for 45–60 seconds including mounting onto Cryotech® (3 COCs/Cryotech®) and plunging into liquid nitrogen. Almost all VS was removed to leave only a thin layer covering the COCs. Warming was performed stepwise into decreasing sucrose solutions (1 M for 1 minute and 0.5 M for 3 minutes) and subsequently washed twice for 5 minutes in H-SOF/PVA. The heated stage was used at 39°C and warming solutions were maintained at 37°C during the whole procedure. Recovery of COCs was achieved at the incubator for 2 hours. ## 5- In situ detection of activated caspases Apoptosis was analyzed in living cells through detection of activated caspases with a specific fluorescent inhibitor (VAD-FMK-FITC, Calbiochem®). Caspase inhibitor (VAD-FMK) conjugated to FITC is cell permeable, nontoxic and irreversibly binds to activated caspases in apoptotic cells. Denuded oocytes were incubated in 1:300 VAD-FMK-FITC in M199 for 45 minutes at 38.5°C under 5% CO<sub>2</sub> in humidified air. Oocytes were washed in H-SOF/PVA with 1:500 propidium iodide (Calbiochem®) to assess membrane integrity. They were finally mounted for detection by fluorescence microscopy using a B-2A filter (microscope Nikon TE-300; Nikon, Tokyo, Japan). Oocytes showing brilliant green fluorescence were considered caspase positive. ## 6- Quantification of free cholesterol and cholesterol esters by Amplex® Red Based on an enzyme-coupled reaction that detects both free cholesterol and cholesteryl esters, cholesterol levels were measured through a fluorometric method (Amplex® Red Cholesterol Assay Kit, Molecular Probes®) using an SLM model 4800 spectrofluorimeter (SLM Instruments, Urbana, IL) with a vertically polarized light beam from a Hannovia 200-W Hg/Xe arc lamp, obtained by a Glan- Thompson polarizer (4-nm excitation and emission slits), and 3- x 3-mm quartz cuvettes. Temperature for the assay (37°C) was set with a thermostat-controlled circulating bath (Haake, Darmstadt, Germany). Free cholesterol was measured performing incubation for 30 minutes in the dark. The fraction of cholesterol that is in the form of cholesterol esters was determined adding cholesterol esterase after free cholesterol measurement (time zero) and recording measurements after 30, 60 and 90 minutes. The latter incubation-time was the maximum at which all measurements reached a plateau. Measurements were performed in samples of 45–50 matured oocytes and the corresponding cumulus cells derived from MβCD-treated or untreated COCs. Fluorescence was measured with an excitation of 560 nm and emission of 590 nm. Background fluorescence determined for the non-cholesterol control sample was subtracted from each value. Cholesterol content was expressed as pmol of cholesterol per oocyte and nmol of cholesterol per mg of protein for cumulus cells. Protein content of cumulus cells was quantified by a direct method using a spectrophotometer (Picodrop, P100, UK). ## 7- Live cell imaging of lipids ### 7.1 Staining of lipid droplets with Nile Red Denuded oocytes were incubated with 1 μg/ml Nile Red (Molecular Probes®) in H-SOF/PVA for 5 minutes at room temperature. The fluorescent lipophilic dye was excited by a 450 to 500 nm line and yellow emission was detected with a B-2A filter. Digital photographs of the equatorial part of the oocyte were taken with a 20X objective in an epifluorescence inverted microscope (Eclipse TE-300; Nikon, Tokyo, Japan) connected to a DS-Fi1c camera (Nikon, Tokyo, Japan). Fluorescence intensity from oocytes was measured using ImageJ (Fiji) software. Barceló-Fimbres and Seidel validated the use of Nile Red to quantify the content of cytoplasmic neutral lipids in bovine oocytes and embryos. ### 7.2 Fluorescent cholesterol in living oocytes A stock solution (5 mM) of the fluorescent cholesterol analog, BODIPY- cholesterol (BPY-chol; Avanti Polar Lipids), was prepared in ethanol and stored in a dark glass tube under nitrogen at -20°C. Working solution (1 μM) was obtained diluting the stock in M199 medium (amount of ethanol less than 1%). Pulse-labeling was performed incubating COCs, cumulus-free oocytes and ZP-free oocytes with BPY-chol for 5 minutes at 38.5°C. Oocytes were immediately washed and subsequently imaged to avoid internalization of the lipid probe. Digital photographs were taken with a 20X objective in an epifluorescence inverted microscope (Eclipse TE-300; Nikon, Tokyo, Japan) connected to a DS-Fi1c camera (Nikon, Tokyo, Japan). Quantification of fluorescence intensity at the plasma membrane was measured in ZP-free oocytes by outlining regions of interest (ROI) in background corrected images using ImageJ (Fiji). ### 7.3 Localization of the raft marker lipid GM1 Glycosphingolipid GM1 was detected in living cumulus-free oocytes by using the fluorescent-labeled cholera toxin B subunit (CTB-AF<sup>488</sup>, Molecular Probes®), which binds specifically to the ganglioside. Oocytes were incubated at 38.5°C for 10 minutes in M199 supplemented with CTB-AF<sup>488</sup> (20 μg/ml), washed with H-SOF/PVA, mounted and immediately imaged in an epifluorescence inverted microscope (Eclipse TE-300; DS-Fi1c camera, Nikon, Tokyo, Japan) to avoid internalization of toxin-GM1. Quantification of fluorescence intensity at the plasma membrane was measured by outlining regions of interest (ROI) in background corrected images using ImageJ (Fiji). ## 8- Statistical analysis Statistical analysis was carried out using InfoStat software. Fluorescence intensity was analyzed through Student’s t test when two mean values were compared and analysis of variance (ANOVA) when more than two mean values were compared, followed by post hoc test analysis of multiple comparisons Bonferroni or Fisher’s Least Significant Difference (LSD Fisher). Caspase variable (binomial distribution) was compared using Generalized Linear Mixed Models (GLMM) with Binomial family, logit link function and LSD Fisher contrast. Cholesterol content was analyzed through ANOVA and post hoc tests of Bonferroni for autumn-spring samples and LSD Fisher for winter-summer samples. Differences were considered significant at P\<0.05. # Results ## Effect of cholesterol incorporation and depletion on oocyte viability Cholesterol-binding drug MβCD was used in order to increase membrane cholesterol content before vitrification and to remove the added cholesterol after warming. COCs with partially removed cumulus cells were incubated with 15 mM MβCD/cholesterol for either 45 minutes or 2 hours during the last period of in vitro maturation. Matured COCs were subsequently vitrified-warmed and depleted of cholesterol within the period of recovery at the incubator. COCs loaded with cholesterol for 45 minutes were depleted of cholesterol for 45 minutes with 4.25 mM MβCD after warming and maintained in M199 at the incubator until completing 2 hours. COCs loaded with cholesterol for 2 hours were depleted of cholesterol with 8.5 mM MβCD for the same period of time. To ensure that the treatment with MβCD produced no alterations in the membrane with negative implications in viability, programmed cell death (apoptosis) indicator was analyzed. Detection of activated caspases was analyzed in live cells by a fluorescent specific inhibitor (FITC-VAD-FMK). This permeable fluorescent probe binds irreversibly to activated caspases in apoptotic cells allowing direct detection by fluorescence microscopy. Moreover, these assays were carried out in the presence of propidium iodide to assess possible alterations in membrane integrity. This method allowed the analysis of the possible effect of modulating cholesterol (enrichment/removal) in vitrification as well as the effect of vitrification itself and the toxicity caused by exposure to vitrification and warming solutions. Results revealed that vitrification generated 12% of oocytes with activated caspases while fresh oocytes presented only \~1% of oocytes caspase positive. In addition, when cholesterol was incorporated into oocytes for 2 hours prior to vitrification and removed after warming, the percentage of caspase activation was 26% higher than that in the vitrified control (12%). No statistical difference in caspase activation was found when cholesterol was incorporated into oocytes for 45 minutes before vitrification and removed after warming, with respect to the vitrified control. Moreover, no statistical difference was found in caspase activation compared to the fresh control. Interestingly, cholesterol modulation of oocytes for 45 minutes showed a similar level of caspase positive oocytes with respect to toxicity control (5% vs 7%, respectively) in which untreated oocytes were exposed to vitrification and warming solutions with no vitrification. In addition, oocytes were not positive for propidium iodide (except for some isolated case) in any of the experimental groups analyzed. On the contrary, when cumulus cells were left during incubation with fluorescent probes, they showed caspase/propidium iodide positive staining at both incubation times. ## Estimation of relative cholesterol incorporation into bovine oocytes Taking into account that modulation of cholesterol for 45 minutes did not affect the apoptotic status of oocyte post-vitrification, fluorescent cholesterol analog BPY-chol was used to estimate the extent of cholesterol incorporation under these experimental conditions. COCs, cumulus-free oocytes and ZP-free oocytes were incubated with 1 μM BPY-chol after MβCD/cholesterol treatment of COCs for 45 minutes. Incubation was performed until finding a time-point (5 minutes) at which the fluorescent cholesterol was highly located at the plasma membrane of the oocyte. That cholesterol incorporation mediated by MβCD/cholesterol complexes occurred effectively as well as its extent were both rapidly verified. An increase in fluorescence intensity was observed under all conditions analyzed (lower panels) but quantification of fluorescence was performed in ZP-free oocytes. Cholesterol-specific fluorescence in the plasma membrane increased \~43% after MβCD/cholesterol treatment compared to BPY-chol- control oocytes. ## Response of bovine oocytes to cholesterol removal Once incubation times with MβCD/cholesterol (45 minutes) and BPY-chol (5 minutes) were determined, we analyzed the response of oocytes to cholesterol depletion with no previous cholesterol loading or vitrification. Interestingly, ZP-intact oocytes incubated with 15 mM “empty” MβCD for 45 minutes showed effects only at lipid droplet level. In order to simplify the contribution of any other unknown variable and to effectively assess cholesterol removal, ZP- free oocytes were directly exposed to 15 mM MβCD. Even under this direct exposure, treated oocytes showed no difference in membrane BPY chol fluorescence level compared to controls (upper panels). Conversely, MβCD-treated oocytes showed marked differences in the distribution of lipid droplets (LD) observed in transmission images (middle panels). The cortex of MβCD-treated oocytes was devoid of LD compared to untreated oocytes. Fluorescent probe Nile Red was in fact used to visualize these differences in the distribution of LD in MβCD- treated oocytes. Nile Red is a lipophilic probe that fluoresces yellow with neutral lipids stored in LD, such as triacylglycerols and cholesterol esters. Absence of LD in the oocyte cortex was confirmed by specific staining with Nile Red in live oocytes (lower panels). Quantification of Nile Red fluorescence intensity showed a decrease of the emitted fluorescence in ZP-free oocytes of the MβCD-treated group. When ZP-intact oocytes were exposed to MβCD under the same conditions, this distribution pattern of LD was observed in 35.5 ± 4.2% of the oocytes, with the cortex showing a lesser extent of absence of LD. ## Quantification of free cholesterol and cholesterol esters in bovine oocytes and Cumulus cells To quantify cholesterol incorporation by oocytes and cumulus cells, we used a highly sensitive fluorometric method that detects both free cholesterol (membrane cholesterol) and cholesterol esters (stored in LD) using an enzyme coupled reaction. After free cholesterol is measured (time zero), the cholesterol fraction in the form of cholesterol esters is determined by adding a cholesterol esterase within the samples. Measurements were performed in fresh oocytes and cumulus cells from untreated COCs, from COCs incubated with MβCD/cholesterol complexes for 45 minutes, and from COCs firstly loaded with cholesterol for 45 minutes and subsequently depleted of cholesterol with empty MβCD for another 45 minutes. An interesting finding revealed that the content of membrane cholesterol in bovine oocytes varies among seasonal periods. Cholesterol was therefore quantified all throughout the seasons. Membrane cholesterol level of oocytes from autumn was observed to be the same as that of spring oocytes, whereas winter oocytes showed a membrane cholesterol level similar to that of summer oocytes. Likewise, a difference of 3 pmol of membrane cholesterol per oocyte was found between autumn-spring and winter-summer oocytes, the latter showing the lowest cholesterol level. No differences were found in basal membrane cholesterol of cumulus cells among seasons. However, when analyzed grouped in both seasonal periods, autumn-spring cumulus cells showed a higher membrane cholesterol level than winter-summer cells (p = 0.04). Therefore, cholesterol incorporation and recovery to its original level was analyzed with respect to controls in autumn-spring oocytes versus winter-summer ones and the same criterion was used for the corresponding cumulus cells. After incubation with MβCD/cholesterol, oocytes showed an increase of membrane cholesterol of \~2.5 pmol/oocyte, regardless of the season. Cumulus cells showed \~2-fold more membrane cholesterol than controls in both seasonal groups. Recovery of original cholesterol levels after incubation of cholesterol-loaded COCs with 4.25 mM MβCD was achieved for both oocytes and cumulus cells regardless of the season. Cholesterol esterase hydrolyzed cholesterol esters stored in cytoplasmic LD of autumn-spring oocytes until reaching a plateau after 60 minutes of incubation with the enzyme. This behavior, which was observed under all conditions analyzed, represents -in terms of cholesterol content- total oocyte cholesterol (free cholesterol + esterified cholesterol). As to winter-summer oocytes, hydrolysis of cholesterol esters reached the maximum level after 90 minutes of enzymatic incubation. Under control conditions, cholesterol esters represented 38% of the total cholesterol content of autumn-spring oocytes while they represented \~60% of the total cholesterol content of winter-summer oocytes. In absolute terms, oocyte cholesterol esters remained unchanged among seasonal periods. In cumulus cells, \~30% of total cholesterol corresponded to the ester fraction in autumn-spring but accounted only for 19% in winter-summer. Hydrolysis of cholesterol esters in cumulus cells was more evident during the first 30 minutes of incubation with the enzyme for both seasonal groups. ## Effect of vitrification on raft marker lipid GM1 The relevance of functional membrane rafts in postcryopreserved bovine oocytes has not been fully explored to date. The degree of compromise of raft marker lipid GM1 in vitrification was analyzed using a fluorescent-labeled cholera toxin B subunit which binds specifically to the ganglioside and permits its identification in living cells. Bovine oocytes showed enrichment in membrane rafts evidenced by the presence of the glycolipid GM1 all along the membrane of living oocytes. As the size of membrane rafts is smaller than the resolution of light microscopy, this uniform distribution of the GM1-associated fluorescence was expected. (upper and middle panel) shows how vitrification affected both localization and level of GM1 at the plasma membrane. The decrease of GM1-associated fluorescence caused by vitrification either did not occur or, if it did, it was at least restored when cholesterol was incorporated into oocytes before vitrification and removed after warming, thus indicating that GM1 related-raft integrity is being preserved by cholesterol modulation. On the other hand, under these experimental conditions, membrane cholesterol showed no differences in terms of BPY-chol fluorescence levels between fresh oocytes and postvitrified treated oocytes. Postvitrified control oocytes showed a tendency to decrease BPY-chol fluorescence at the plasma membrane. However, the measurement of BPY-chol fluorescence after enzyme treatment to remove ZP after warming revealed high variability and therefore no statistical differences could be observed among the experimental groups. Membrane cholesterol was nonetheless maintained after warming by previously removing the added cholesterol with a concentration of empty MβCD but quite lower (4.25 mM) than that used to enrich membrane cholesterol (15 mM). # Discussion Biological membranes undergo a change from a fluid state to an ordered state as temperature is reduced below the transition temperature of the membrane (\~10°C in the mature bovine oocyte). This phenomenon is dependent on their lipid composition. Membranes with high cholesterol:phospholipid ratio or either with polyunsaturated fatty acids or short chain fatty acids are less sensitive to sub-physiological temperatures. Bovine oocytes are particularly susceptible to chilling which occurs at different cellular levels, such as the zona pellucida, plasma membrane, meiotic spindles and cytoskeleton. Cholesterol is the major non-polar lipid of mammalian cell membranes, where it typically accounts for 20–25% of total membrane lipid content. The planar structure of cholesterol confers special biophysical properties that contribute to generate a semipermeable barrier and to regulate membrane fluidity. Exposing cells to MβCD/cholesterol complexes containing saturating amounts of cholesterol leads to cholesterol enrichment depending on the concentration, exposure duration and cell type. Exposure times to “empty” MβCD required to remove cellular cholesterol are generally much shorter than those needed to enrich a membrane that already has its original cholesterol level. To our knowledge, two studies have been conducted in bovine oocytes using MβCD as a strategy to load cholesterol in their membranes prior to vitrification. In the first study, the authors showed that cholesterol diffused through cumulus cells and reached the oocyte but failed to reflect these effects in the production of embryos. The same methodology was applied later in immature bovine oocytes. In this case, both the presence of serum and the compact structure of the cumulus of immature oocytes could have limited the availability of MβCD/cholesterol complexes, affecting the results collected. Overall, the low success reported in these studies could be due to two main reasons which, to our knowledge, have not been taken into account to date. (1) Incubations with MβCD/cholesterol complexes were performed for 25–60 minutes at the concentrations typically used for sperm (0.5–2 mM MβCD). The lipid mass representing the plasma membrane of an oocyte (diameter \~80–100 μm), which also has numerous microvilli, compared to a cell as a sperm, is significantly higher. Therefore, the concentrations and/or incubation times assayed could not have been sufficient to effectively increase the level of cholesterol in the membranes of oocytes. (2) If before vitrification, there was a genuine incorporation of cholesterol into the oocyte plasma membrane, after warming it would be necessary to remove the added cholesterol to recover the physiological level of cholesterol at the membrane. In the cholesterol-loaded sperm, this scenario seems not to affect them to a great extent since cryopreservation itself produces an efflux of cholesterol that may recover -either totally or partially- cholesterol levels in the membrane after thawing (this being the reason why it is said that sperm undergo "cryocapacitation"). In the present work, we have increased MβCD/cholesterol concentration and we have analyzed two incubation times (45 minutes and 2 hours). We have also hypothesized that removal of cholesterol after warming better preserves the organization of the plasma membrane by recovering its original cholesterol level. In situ detection of activated caspases revealed that modulation of cholesterol for 45 minutes does not affect oocyte apoptotic status. On the contrary, incubation for 2 hours increases apoptosis after vitrification. Based on this, it can by hypothesized that the level of apoptosis of oocytes loaded with cholesterol for 45 minutes, vitrified and recovered (cholesterol removed), could be due at least by the cytotoxicity caused by cryoprotectants since no differences were found with respect to the toxicity control group. However, no significant differences were found either between the apoptotic level of these oocytes and the fresh controls. It was, in fact, the only experimental group that showed this condition compared to fresh oocytes, thus indicating a level of cell death closer to that of fresh oocytes. Interestingly, it has been shown that cholesterol protects the phospholipid bilayer from oxidative damage. Also, most part of cumulus cells was found to be positive for caspases and propidium iodide, a marker of cell viability that penetrates when the membrane loses integrity. Vitrification protocols are not particularly designed to preserve cumulus cells but the oocyte, with the large volume that this cell type presents. This raises a controversy at the moment of vitrifying bovine oocytes with the corona radiata and some layers of cumulus cells to proceed to the subsequent in vitro fertilization (IVF), as is usually done. In mammals, fertilization rates in oocytes without cumulus are markedly low. Furthermore, intracytoplasmic sperm injection (ICSI), which is performed with success in human and mouse denuded oocytes (without cumulus), still remains inefficient in the bovine. Because vitrified bovine oocytes lack great part of viable cumulus cells capable of responding to the signaling cascades generated by sperm, further studies should converge in the adaptation of an IVF protocol for denuded (and vitrified) oocytes. Recent research has demonstrated that the presence of several layers of cumulus cells during vitrification reduces the survival of bovine matured oocytes. This research also showed that vitrifying oocytes either with a few layers of cumulus cells or without cumulus cells, but co-incubated with fresh COCs during fertilization, may lead to higher survival and embryo development. Our data showed that cholesterol was incorporated into the oocyte plasma membrane as evidenced by comparative labeling of the fluorescent probe BPY-chol. However, direct exposure of ZP-free oocytes to empty MβCD was not sufficient to achieve cholesterol removal in oocytes with physiological levels of cholesterol. Instead, it affected the distribution of oocyte LD which may reflect that cholesterol repletion occurs at the plasma membrane. It seems possible that cholesterol removal occurs concomitantly to cholesterol replenishment driven by cortical LD through specific hydrolysis of cholesterol esters stored within these organelles, thus maintaining BPY-chol labeling at the plasma membrane. Macro-autophagy of LD, termed lipophagy, is a physiological mechanism to degrade LD in some circumstances. It is important to note that this effect was only observed in oocytes with physiological levels of cholesterol exposed to MβCD but not in cholesterol-loaded oocytes in which excess cholesterol was fairly simple to remove. Nevertheless, under this experimental condition (ZP-free oocytes exposed to MβCD), restructuring or even disturbance of the underlying cytoskeleton cannot be discarded. At present, LD have gain notoriety due to increasing evidence related to their roles in many aspects of health and disease. The prevalence of metabolic syndromes, obesity, steatosis and atherosclerosis has prompted further biomedical research on LD since early embryo stages or even at the oocyte level. On the other hand, earlier studies have addressed the analysis of the fatty acid composition of total phospholipids, major phospholipid sub-classes and triacylglycerols from bovine oocytes and embryos. However, none of these studies have quantified oocyte cholesterol levels, except for the work of Kim et al. who measured total cholesterol content. More recently, mass spectrometry has provided new insights into the composition of different molecular species of phospholipids, free fatty acids, and triacylglycerols present in single oocytes and preimplantation embryos. Relative abundance of squalene, a key intermediate in cholesterol biosynthesis, showed higher levels in bovine immature oocytes compared to in vitro matured oocytes analyzed through this approach. In our work, an analytical strategy has been conducted to quantify free cholesterol and esterified cholesterol in bovine oocytes and cumulus cells. An interesting finding revealed that the level of membrane cholesterol in bovine oocytes varies among seasonal periods. In contrast, cholesterol esters remained stable among seasons. These variations at the membrane level were enough to account for the differences also found in total cholesterol. Even when no design was performed to detect a direct effect of the season, in particular, these results could be quite close to show this effect. For subsequent analysis, oocytes were grouped into autumn-spring and winter-summer oocytes taking into account similar levels found in membrane cholesterol among seasons. Autumn and spring oocytes evidenced the highest content of membrane cholesterol as well as of total cholesterol. Equal amount of total cholesterol was found by Kim and co-workers in bovine oocytes matured in a serum-free medium (9.2 pmol per oocyte), irrespective of the season. However, when oocytes were matured in the presence of fetal bovine serum, cholesterol content was significantly higher (15.1 pmol per oocyte) which can be explained by the fact that serum provides a variety of lipids including cholesterol. On the other hand, our results showed that the differences found in total cholesterol of cumulus cells were explained by the differences found in the content of membrane cholesterol and of cholesterol esters. Furthermore, membrane cholesterol of cumulus cells showed the same pattern as that of oocytes when grouped in these seasonal categories. Our results demonstrate that the whole COC is dynamic in terms of cholesterol homeostasis. In addition, both oocytes and cumulus cells increased membrane cholesterol after incubation with MβCD/cholesterol and recovered the original level, regardless of the season. Also, total cholesterol was restored to physiological levels. Interestingly, cholesterol-loaded oocytes from winter and summer showed that membrane cholesterol levels were the same to those of autumn-spring oocytes at basal conditions. Similar profiles were also observed in the cholesterol ester fraction and total cholesterol. Altogether, these results suggest a distinctive cholesterol metabolic status of COCs among seasons. In dairy cows, a seasonal study showed a higher percentage of saturated fatty acids in phospholipids of germinal vesicle oocytes and granulosa cells during summer whereas mainly mono- and polyunsaturated fatty acids (PUFA) predominated during winter. Biophysical analysis of oocyte membranes revealed that the lipid phase transition decreased 6°C between summer and winter. These alterations in membrane composition were associated with decreased developmental competence during high ambient temperature months. Conditions of heat stress may aggravate the negative energy balance of dairy cows contributing to the metabolic imbalance responsible for reduced fertility. Heat stress seems, in fact, to be relevant for reproductive performance of cows in tropical and subtropical climates. In this respect, in vivo produced embryos of Bos indicus during the rainy season, compared to those produced in the dry season, had a lower number of apoptotic cells in fresh embryos as well as after freezing. A connection between nutrition and quality of oocytes and embryos has been suggested. Nevertheless, even when there is some coincidence in that PUFA rich diets have beneficial effects on bovine oocyte and embryo quality correlation to changes in gamete lipid composition still remains to be established. It has also been proposed that PUFA uptake by the oocyte is selective and highly regulated to avoid the risk of cellular damage. A recent study has shown that dietary PUFA supplementation has no effect on the relative abundance of phosphatidylcholine or sphingomyelin molecular species in bovine blastocysts produced in vitro. As to cholesterol, it is known that plants only have minor amounts of cholesterol and that the diet is therefore not a cholesterol source for herbivorous, however, a particular nutritional condition may influence the metabolic status of oocytes and embryos. Cholesterol may account for up to 45 mol% of the membrane´s total lipid content in animal cells. Measuring only choline- phospholipids and total cholesterol, Kim et al. showed that cholesterol to phospholipid ratio is important in bovine oocytes. Our work provides new data on unesterified cholesterol and cholesterol esters in bovine oocytes and cumulus cells at different seasonal periods. Taking into account that cryopreservation may disturb plasma membrane organization required for fertilization, we analyzed putative raft molecule GM1 in living oocytes. Gangliosides have been found to be involved in multiple physiological functions, and it is important to understand how their distribution is regulated in the cell membrane. It has been shown that the clustering of both GM1 and GM3 at the plasma membrane depends primarily on membrane cholesterol levels and also on intact actin cytoskeleton. To our knowledge, this is the first report in which ganglioside GM1 is described in the bovine oocyte. Vitrification was observed to clearly affect the localization and level of GM1 at the plasma membrane. In line with this, dispersion of GM1 clusters by chilling was reported in mouse fibroblast analyzed by immunoelectron microscopy after quick-freezing and freeze-fracture. However, when cholesterol was incorporated into oocytes before vitrification and removed after warming, GM1 levels at the plasma membrane were not altered or at least restored. Likewise, membrane cholesterol was maintained after warming. Cholesterol modulation during vitrification and warming could therefore be a useful tool to preserve membrane raft integrity. A glycosylphosphatidylinositol (GPI)-anchored protein named Juno has been identified as the sperm receptor at the plasma membrane of mammalian oocytes. GPI-anchored proteins are enriched in non- invaginated membrane rafts highlighting how important plasma membrane organization is for fertilization. Further analyses will be needed to address membrane lipid compromise during vitrification of bovine oocytes. Emerging studies in embryos have found differences between fresh and vitrified bovine blastocysts in certain molecular species of choline-phospholipids and triacylglycerols. Cholesterol is a key contributory factor to membrane properties, particularly membrane fluidity. This is the first study conducted to date on free cholesterol and cholesterol esters in bovine oocytes and cumulus cells. Interestingly, cholesterol levels differed among seasonal periods revealing a distinctive metabolic status of COCs. Also, the whole COC showed a dynamic organizational structure in terms of cholesterol homeostasis. Modulation of membrane cholesterol by MβCD improved oocyte survival after vitrification yielding levels of cell death closer to those of fresh oocytes. Oocytes effectively incorporated membrane cholesterol prior to vitrification and recovered their original level after cholesterol removal. Cholesterol modulation also preserved membrane localization of the raft lipid GM1 after vitrification, thus suggesting its possible role as a cryotolerance marker. Future studies on the biophysical properties of oocyte and cumulus cell membranes will provide new insights to understand the role of lipid behavior in the developmental competence of cryopreserved oocytes. The authors thank translator Viviana Soler for controlling the use of English language and Adriana Lauro for her technical assistance in oocyte collection and handling. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: JB SSA RHA. Formal analysis: JB JRC SSA RHA. Funding acquisition: JB SSA RHA. Investigation: JB GLR JRC SSA. Methodology: JB GLR SSA RHA. Project administration: JB SSA RHA. Resources: JB JRC SSA RHA. Visualization: JB SSA RHA. Writing – original draft: JB SSA RHA. Writing – review & editing: SSA RHA.
# Worrying me softly with your tweets: Anxiety contagion between leaders and followers in computer-mediated communication during COVID-19 Emotional contagion, “a process in which a person or group influences the emotions or behavior of another person or group through the conscious or unconscious induction of emotion states and behavioral attitudes” \[, p. 50\], plays an important role in several domains, including social interactions and leadership. Previous work has identified that in face-to-face leader-follower interactions, leaders do not only transfer emotions to their followers via tacit, subconscious processes, but also can transfer emotions deliberately “with the intention of attaining certain interactions or task outcomes” \[, p. 659\]. Leaders are likely to influence follower affect, due to leaders’ status role, which corresponds to access to resources and authority to shape the work environment as leaders see fit. However, less is known about the transfer of emotions between organizational leaders and followers in computer-mediated communication (CMC) on public social media platforms such as Twitter. The presented research addresses this gap. Based on emotional contagion, we explore the relationship between leader and follower anxiety. We focus on anxiety due to its state <u>and</u> trait properties. And while previous studies have focused on the dispositional differences and situational characteristics that trigger anxiety, a longitudinal examination of the transfer of leader anxiety to followers, in particular using CMC, has not been conducted. In this respect, our study responds to the need of understanding further emotional contagion in organizational life by examining a widely used and powerful means of communication between leaders and followers. Finally, we explore possible changes in the transfer of anxiety from leaders to followers before and during a threatening situation, namely the COVID-19 pandemic. Anxiety is considered the default crisis emotion, given that individuals in a crisis (i.e., a particularly threatening) situation experience anxiety more intensely and frequently than any other crisis emotion \[incl. anger, fear or sadness; \]. Hence, we examine whether during a crisis the occurrence and impact of anxiety contagion from leaders to followers differ compared to normal circumstances. We examine this particular phenomenon in an exploratory manner, using an innovative ML approach to detect both state and trait anxiety in a large sample of 197 leaders and 958 followers and derive 43,283 daily indications of anxiety from posts and interactions on the Twitter platform by leaders and followers over 316 days. ## Emotional contagion in face-to-face and computer-mediated communication Emotional contagion constitutes the transfer and sharing of emotions from one person to another. As stated by Barsade, Coutifaris, emotional contagion is 1) comprised of distinct emotions (e.g., anger, anxiety), 2) “occurs via subconscious and conscious processes that transpire when people are both elicitors and targets of emotional contagion” \[(p. 138), \] occurs on an interpersonal level and finally 4) not only influences “how people feel but also what they subsequently think and do” \[p. 138\]. Based on this definition, we examine emotional contagion as a social influence, in which “\[person\] A has power over \[person\] B to the extent that \[A\] can get B to do something that B would otherwise not do” \[, pp. 202–203\], while the power of person A over person B is affective in nature. If person A is successful in transferring their emotions to person B, person B non-consciously imitates the communicated affect, which leads to convergence in both interaction partners’ emotions. Importantly, previous research oftentimes equates emotions with affect–a term that comprises all aspects of subjective feelings. To stay consistent with past research and avoid confusion, throughout this manuscript, we use the term emotional contagion when describing the transfer of anxiety, as part of follower affect. Similar to previous studies (Barsade, 2018), we do distinguish between emotions, short-term affective reactions to stimuli, and dispositional affect, a trait-like predisposition to experience certain feelings at any given time. The majority of scholars have studied the occurrence and underpinnings of emotional contagion in face-to-face interactions \[for a review see \]. However, there is some indication that emotional contagion is not only possible in CMC but also that emotions effectively can spread both directly and indirectly in social networks such as social media. For example, Cheshin, Rafaeli confirmed that the transfer of emotion also occurs in virtual teams and that these “text- based communications of emotion were detected and ‘caught’ by partners interacting via text-based instant messaging” (p. 3)\]. And in dyadic CMC interactions partners exchanged messages slower and used shorter messages when experiencing negative emotions, compared to participants experiencing neutral emotions. Interaction partners in CMC could also distinguish discrete emotions in messages, namely anger or happiness. Hence, it seems clear that when communicating electronically individuals may influence others affectively and even impact behavioral outcomes. Yet, the majority of existing studies have examined emotional contagion in CMC using text-based communication on Web 1.0 platforms, i.e. communication between two individuals via text, largely ignoring another mainstream means of communication–social media. Barsade, Coutifaris point out, social media platforms are much more open, interactive, and dynamic, allowing communication between an individual and entire groups or communities at once. Hence, social media platforms allow researchers to study the effects of emotion transfer from leader to several followers (i.e., from A→B and A→C) simultaneously. Indeed, previous studies have found that emotions <u>can</u> spread between users via CMC on social media platforms, and can do so from one user to several others. Yet, less is known about the spread of discrete emotions via CMC, in particular, whether some emotions are more likely to be spread and transferred to others. And as the use of public social media platforms is becoming more prevalent and accepted by both organizations and employees, we argue that more attention needs to be paid to the study of emotional contagion from leaders to followers, and its repercussions on follower affect, via CMC. ## Anxiety contagion between leaders and followers Anxiety is characterized by negative valence and high arousal, as well as cognitive appraisals of uncertainty and low control and is a transient emotional response to an event- or situation-specific context. Throughout this paper anxiety is based on the presented sub-clinical definition, as opposed to the clinical diagnosis of anxiety. This transient episode of anxiety, triggered by a threatening or uncertain situation, is referred to as state anxiety. State anxiety is short in duration, unlike mood or other forms of dispositional affect, and is particularly important in the transfer of emotion in communication because anxiety is functional and “can facilitate constructive behavior”. For example, in response to situation-specific threats, anxious individuals are more likely to detect and recognize potential threats, protect themselves or others, and in general, become more vigilant. Hence, anxiety can help individuals adapt to potentially threatening and harmful environmental demands, especially when facing overwhelming threats such as a crisis. However, most previous research has not focused on the transfer of anxiety but rather negative and positive affect in general. In addition, to the best of our knowledge, no previous studies have examined anxiety contagion between organizational leaders and followers using social media. Yet, based on previous evidence of emotional contagion between users of social networks and the signaling power that organizational leaders possess, we would expect leader state anxiety via CMC to influence follower state anxiety over time. 1. Hypothesis 1: In CMC, leader state anxiety positively predicts follower state anxiety. ## The role of trait anxiety It is important to note that the predisposition to anxiety can vary between individuals in regards to both intensity <u>and</u> frequency. Hence, some individuals have a greater tendency to experience anxiety, known as trait anxiety. Trait anxiety, in combination with state anxiety, influences an individual’s magnitude of experienced anxiety in any given situation. Individuals who are predisposed to experience anxiety (i.e., trait anxiety) tend to perceive situations as more threatening, pay more attention to presented negative information, and tend to experience stronger physiological and psychological sensations. Therefore, trait anxiety plays a significant role in determining the intensity and frequency of experienced state anxiety. Concerning the leader-follower relationship, we know that leader trait affectivity can influence contagion susceptibility but also can “influence leaders’ ability to influence others through emotion” \[, p. 658\]. Hence, we would expect followers of high trait anxiety leaders to be more used to their leader displaying heightened anxiety to situations in general. Hence, it is likely that when leaders express heightened state anxiety, followers of trait anxious leaders are likely to express less state anxiety compared to followers of leaders who rarely exhibit anxiety. We hypothesize the following: 1. Hypothesis 2: In CMC, leader trait anxiety moderates the relationship between leader state anxiety and follower state anxiety, in that followers of highly trait anxious leaders experience less increased state anxiety than followers of low trait anxious leaders. ## The role of context While more anxious leadership could be associated with more anxious followers in general, this might change due to context, such as a crisis as “individuals differ in their adaptation to events, with some individuals changing their set point and others not changing in reaction to some external event” \[, p. 306\]. Hence, we argue that the transfer of anxiety from organizational leaders (i.e., individuals with resources and authority) to their followers would be particularly useful to examine considering the critical role anxiety plays in crises. Accordingly, we examine the relationship between leader and follower anxiety in pre-and during crisis contexts, namely the COVID-19 pandemic. ### Emotional contagion during the COVID-19 pandemic With terms such as “social distancing” and “flattening the curve” becoming part of everyday vocabulary, the COVID-19 pandemic is the most challenging health crisis of the 21<sup>st</sup> century. Coronavirus officially was declared a worldwide pandemic by 11<sup>th</sup> March 2020 by the World Health Organization, with governmental work-from-home orders following shortly thereafter. Employees soon found themselves in a new and, for most, quite unfamiliar social territory, with homes all over the world converting into offices, schools, daycares, gyms, etc. while communication to the outside world mainly moved online. The changing working conditions and the increased economic pressures have led to a wider socioeconomic crisis, typically followed by adverse working conditions for employees. Given the increased need for remote working, managers and employees alike have turned to CMC as a primary means of communication. Yet, research on information communication and sharing in a crisis (or crisis- like situation) is limited. One study examined how communication on Twitter occurs during the Great East Japan Earthquake in 2011 and concurs with findings of related studies that negative feelings, such as worry and anxiety, were more likely to be shared online than neutral or positive feelings. As information grows explosively in a crisis, people flock online, especially to public communication platforms such as Twitter, to communicate and share information with others. Individuals also rely on these platforms to discover and evaluate how others have been affected by a crisis and in turn can use these platforms to emotionally influence others. In sum, via social media individuals are both influenced by and “influence each other more instantly and frequently” \[, p. 2033\]. Therefore, communication on social media platforms provides a feasible alternative to study emotional contagion between leaders and followers, especially in the case of a crisis marked by social distancing and a strong reliance on CMC between organizational members. Nevertheless, it is important to remember that not all crises are the same, nor do all crises impact all individuals to the same degree. For example, Barsade, Coutifaris suggest that in the last financial crisis, individuals who were in financial distress were more likely influenced by the anxiety described in the press and social media, and making the transferred anxiety their own, which in turn affected behavioral outcomes such as restricted spending. However, in the case of the COVID-19 pandemic, the affective impact of the crisis itself is not yet clearly defined. On the one hand, previous studies have found that state anxiety levels have increased for entire populations, on the other hand, recent studies have also found that exogenous individual differences such as age, risk behavior, and resilience also influence state anxiety. In short, the COVID-19 crisis provides a unique setting, in which interactions between leaders and followers may be studied in extreme circumstances, and compared to CMC in normal circumstances (i.e., before the onset of the COVID-19 pandemic). Yet, due to the uniqueness of the COVID-19 crisis concerning its (universal vs. individualized) influence on employees’ affect, we hesitate to form specific hypotheses regarding pre-and-during-crisis differences in the relationship between leader and follower anxiety. Hence, we decided to examine the COVID-19 pandemic context as a possible moderating factor in an exploratory manner. # Methodology Due to the state- and trait-like properties of anxiety, it is crucial to study follower anxiety over time. The presented study overcomes typical diary study limitations (e.g., high dropout rates of respondents) by applying anxiety and personality detection algorithms on a large sample of leaders and followers. The final sample comprised 197 leaders and 958 followers in 79 companies engaging in CMC on the public social media platform Twitter. This resulted in a total of 43,283 matched daily leader-follower observations. The respective minimal anonymized data can be found on Open Science Framework (<https://osf.io/3r8z7/>). The collection method complied with the terms and conditions for the websites from which the data was collected. ## Sample and procedure Our methodology comprises five key steps. The first step involves the preparation and pre-processing of our dataset. Using an initial database of organizational leaders and employees ([https://crunchbase.com](https://crunchbase.com/)), and their social media information, we selected organizations and respective employee social media data based on the number of listed social media handles. An organization was included as long as it listed at least ten employees and their public social media information in our dataset. We define leaders as C-suite executives, i.e., individuals with a job title of e.g., Chief Executive Officer (or CEO), Chief Financial Officer (or CFO), etc. The remaining individuals were classified as followers. Notably, we only included employees from the United States of America, as the anxiety detection algorithm was trained exclusively on U.S. data (see section “Predicting state and trait anxiety”). We recognize that this categorization approach is limited, primarily for two reasons. Firstly, this distinction between leaders and followers does not allow us to match followers with their direct supervisors. However, C-suite executives are largely recognized as senior organizational leaders, and therefore wield greater signaling power, which in turn influences all employees. Secondly, one could argue that high-level leaders might be inclined to censor their social media communications to ensure that a positive message is conveyed. Such communications also might be crafted with assistance from the organization’s communications team or even HR since they enhance employee voice. While we agree with this argument, we would add that because emotional contagion can occur both purposefully or intendedly, leaders’ communications still affect their followers even if leaders are not themselves tweeting. In other words, emotional contagion still occurs independently whether emotions are communicated on purpose or unintentionally. In addition, to further ensure the validity and accuracy of provided information of leaders and followers and increase the robustness of the provided data, we manually searched all included leaders and followers on the LinkedIn platform ([https://linkedin.com](https://linkedin.com/)). Doing so allowed us to confirm that leaders and their respective followers were indeed employed in their respective organizations throughout the examined timeframe, including pre-and post-onset of the COVID-19 pandemic. Finally, we chose to include gender as an additional control in our model, because gender is an exogenous variable that cannot be influenced by the outcome variable. Due to the limitations of the Twitter API (twitteR Package documentation - <https://rpubs.com/Kyleen1991/594933>), a maximum of 3,200 tweets per profile were extracted. ### Predicting state and trait anxiety In the second step, we annotate the dataset with an anxiety prediction algorithm. Extracted tweets of all available leaders and followers curated in the previous step were annotated using the anxiety detection algorithm as described in Gruda and Hasan. This algorithm was trained on a dataset of 600 randomly selected tweets from 10,386 users, scored by 604 zero-acquaintance human raters from the United States based on a six-item short-form of the Spielberger State-Trait Anxiety Inventory \[STAI; \]. On average, each tweet was rated five times. Each tweet was assigned an anxiety score of between 1 (“Not at all”; low anxiety) and 4 (“Very much”; high anxiety) by each rater. Two types of features were extracted from the texts of the tweets. The first type of feature is based on the pre-trained Global Vectors for Word Representation (Glove) embedding. The second type of feature comprises the unigram and bigram terms (including emojis) and the corresponding term frequency (TF). The ML algorithm was implemented as two Linear Ridge Regression models corresponding to the sets of features described earlier. For the prediction of anxiety scores of non- labeled tweets, the average of the two predicted scores from the two models is taken as the final score. The training procedure involved the use of a 6-fold cross-validation resampling plan and resulted in a model with R<sup>2</sup> = 0.49 and a Root-Mean-Square Error (RMSE) of 0.52. The trained model was validated using a set of 3.33 million tweets. The model predicted the anxiety scores of the tweets to be between 1 and 4 for 99.7% of the tweets (M = 2.34, SD = 0.36). Trait anxiety was accounted for using the same algorithm as described above. Given that trait anxiety is traditionally measured as the frequency of anxiety experiences, in this study trait anxiety constituted a 30 day average of anxiety scores per user, derived from available tweets before the examined period (before 5<sup>th</sup> October 2019). This provided us with approximately one month of anxiety ratings per individual (i.e., leaders and followers) in our dataset. Manual evaluation of the predicted anxiety scores for our dataset showed that the scores were highly correlated with the emotion expressed in the tweets. For instance, the tweet “Thank you everyone—it’s been a busy week! We love our wonderful \[…\] community!” was predicted to have an anxiety score of 1.7217, while the tweet “How it feels to be a web team fighting DDoS \[…\]” was assigned an anxiety score of 2.5432. Finally, the tweet “\[…\] no no no. Not a happy place.” was scored 3.099. ### Predicting the Big-Five personality traits Emotional contagion susceptibility may be influenced by individual differences. Hence, we measure and control for all Big Five leader and follower personality traits. Previous work \[e.g., –\] has shown that personality traits can be measured accurately and successfully in online contexts using social media data. Therefore, in our third step, each Twitter profile was fed into the IBM Watson Personality Insights API, which extracts and analyzes social media textual data to identify personality traits based on linguistic analysis. IBM Watson relies on an open-vocabulary machine-learning approach and is used to compute raw trait scores and subsequently compare raw scores to a reference sample of 1,000,000 individuals. A pre-condition for using this service is that the examined Twitter profiles need to be public and have a minimum of 100 words across tweets. The provided mean absolute error indicates the difference between estimated or predicted scores and actual scores. The IBM Watson Personality Insights algorithm provides a Watson estimates error rate of ca. 12%. In addition, the IBM Watson algorithm also provides six individual facets for each Big Five personality dimension. All facets were combined into higher-order Big Five personality dimensions (all Cronbach α ≥.70, except for Openness to Experience: α =.68, see). Although a higher Cronbach alpha could have been achieved in the case of Openness to Experience by excluding the facet “adventureness” (α =.73), we decided to include all facets to ensure a full picture of the examined data. ### Post-annotation dataset consolidation Our fourth step comprised the consolidation of state and trait anxiety scores (Step 2) with predicted Big-five personality traits (Step 3). From the resulting dataset, we formed pairwise combinations between leaders and followers, within companies, based on the time of the tweet. For example, tweets by followers from 1<sup>st</sup> March 2020 were paired with leader tweets of the same day. Due to the maximum number of (3,200) tweets limitations by the Twitter API, and since we also examine the role of context, specifically the COVID-19 crisis, as a secondary research question, we focused our main analyses on matched leader- follower observations between 5<sup>th</sup> October 2019 and 13<sup>th</sup> August 2020. Doing so provided us with a somewhat balanced dataset of 158 days before and 158 days after the onset of COVID-19. We classified the onset of COVID-19 to be the 11<sup>th</sup> March 2020, during which COVID-19 was officially declared a worldwide pandemic by the World Health Organization. Finally, since leaders and followers are nested within companies, we excluded companies with less than four daily leader-follower observations. Put differently, we disregarded cases in which there were less than four posted dyads between leaders and their respective followers on the same day. This threshold was identified as the bottom 10% of our dataset and was implemented to ensure data reliability on a company level. This final step resulted in a total of 43,283 matched daily leader-follower observations between 197 leaders and 958 followers across 79 companies. ### Analytical strategy In our dataset, followers are nested within leader-follower dyads, which in turn are nested within companies; observations in our dataset are not independent. Hence, we use a multi-level cross-classified mixed-effects model for repeated measures to test our hypotheses. We define a three-level model with random intercepts at the company level. To compare the overall goodness of fit across models, we used the Aikake Information Criterion (AIC) and the Bayesian Information Criterion (BIC), which facilitate a comparison of mixed models with different numbers of levels and predictors. All analyses were conducted using Stata 16.0. # Results Summary statistics and pairwise correlations of variables on the leader and follower level are shown in. Results of our multi-level cross-classified mixed- effects model for repeated measures are shown in. All outlined models in include various control variables, as noted below in the respective table. To minimize the potential of an unaccounted “third variable” causing a shift in both leader and follower state anxiety on the same day, we apply a multi-day lagged regression design. Doing so allows us to test the following: if a leader posts a highly (or less) anxious tweet today, are respective followers more likely to post (less) anxious tweets on subsequent days? The proposed lag analysis also allowed us to account for potential sleeper effects of the leader-follower anxiety influence. In we present the results of this multi-day lagged regression design, namely by examining whether leader state anxiety on Day 1 predicts follower state anxiety on Day 2, controlling for both leader and follower state anxiety (Day 0 and Day 1), respectively. We also controlled for possible spillover effects of follower state anxiety on preceding days (i.e., Day 0 and Day 1) and of leader state anxiety on preceding days (i.e., Day 0). The main two-way interaction was significant (M2: b = 0.10, SE = 0.05, z = -2.08, p = 0.037). ## Context as a moderating factor We further examine our findings in the context of the COVID-19 pandemic. To do so, we created a dummy variable, which specifies the pre-and-post onset of COVID-19 (11<sup>th</sup> March 2020, the WHO declares COVID-19 a worldwide pandemic). Model 3 includes the examined three-way interaction and all considered controls (identical to M1 and M2). We find a significant three-way interaction, between leader state- and trait-anxiety on follower anxiety and the pre-and-post onset of the COVID-19 pandemic(M3: b = -0.19, SE =.09, z = -2.10, p = 0.036). To better understand this interaction, we plotted the results of the complete model (M3) in. Graphing this three-way interaction (5<sup>th</sup> and 95<sup>th</sup> percentile) showed that in the case of less trait anxious leaders, leader state anxiety on Day 1 was associated positively with high follower state anxiety on the next day during the pandemic (simple slope = 0.06, SE = 0.02, z = 3.53, p \<.001). Hence, it seems that in a crisis context, followers of less trait anxious leaders seemed to be most impacted by increased leader state anxiety than followers of more trait anxious leaders. In the case of trait anxious leaders, this trend was not significant (simple slope = -0.02, SE = 0.01, z = -1.26, p \> 0.10). Results before the onset of the COVID-19 pandemic were even less pronounced and proved to be not significant both in the case of less trait anxious (simple slope = 0.01, SE = 0.01, z = 1.03, p \> 0.10) as well as trait anxious leaders (simple slope = 0.01, SE = 0.01, z = 0.60, p \> 0.10). ### Robustness checks We also tested the same three-way interaction using a continuous-time variable instead of the aforementioned dummy variable. Results remained unchanged with the examined three-way interaction between leader state anxiety (Day 1) predicting follower state anxiety (Day 2) over time (b =.00, SE = 0.00, z = -2.29, p = 0.02). # Discussion Previous research has found that leaders effectively transfer emotions to their followers in face-to-face communications. However, less clarity exists concerning the transfer of emotion via CMC in a naturally occurring environment. In this study, we find that leader state anxiety predicts follower state anxiety, even when accounting for a series of leader and follower personality traits and demographics (e.g., leader and follower trait anxiety, gender, etc.). We also find that follower state anxiety is a function of both leader state- (i.e., leaders’ experience of event-specific anxiety over time) and leader trait anxiety (i.e., leaders’ tendency to experience anxiety in general). Hence, followers of highly trait anxious leaders and who are experiencing increased state anxiety (e.g., due to COVID-19 pandemic and its repercussions), experience less state anxiety than followers of less trait anxious leaders. Put differently, followers of leaders who do not tend to be anxious in general (unrelated to a specific event), seem to be less used to their leaders’ affect and might be more susceptible to their leaders’ increased state anxiety. This could be because followers of trait anxious leaders might discount their leaders’ expressed anxiety and instead look for emotional cues from other important people in their lives (e.g., peers, colleagues, etc.), while the same is not the case for followers of less trait anxious leaders. Relational individual differences might play a role here as well \[e.g., \]. # Implications We argue that the study of anxiety as a follower outcome is particularly important because the experience of anxiety can have destructive consequences, including depleted self-regulation, and increased and prolonged emotional exhaustion. In addition, anxious individuals, in their aim to lower their anxiety oftentimes reach out to others for advice, but also are more likely to accept un-advantageous or even inappropriate and harmful advice from others. Finally, the transfer of anxiety from leader to follower is important because it can serve as a precursor of the subsequent cascading emotional contagion between team members, which can determine team affect over time. Hence, we suggest that even low rates of “anxiety spread” may be meaningful in influencing team members’ affect over time. We found that the proposed effects seem to be dependent on the contextual effects of the COVID-19 pandemic. The COVID-19 crisis emerged very fast with an all-encompassing impact on individuals, organizations, and communities. Although there are similarities between the COVID-19 pandemic and previous crises regarding economic consequences, the COVID-19 crisis brings to the forefront critical health and safety issues as well. For example, social distancing created serious knock-on effects on the way people communicate, collaborate, and work. This in turn means that the use of social platforms of communication increased dramatically with various consequences including mental health. In our study, we found that the transfer of anxiety via social media from leaders to followers is stronger throughout the pandemic compared to the examined pre- pandemic period. In that sense, we suggest that the role of crisis context is of particular importance in transferring anxiety between leaders and followers and should not be ignored. The crisis context makes people more vulnerable and stressed resulting in decreasing their commitment to organizational leaders and their organization overall. In this respect, leaders need to understand that during a crisis everything counts, even the words they are publicly expressing via social media. There is a specific trend for leaders to try to take more responsibility and respond to employees’ demands during a crisis period. However, this seems not to be always the ideal action, since some evidence suggests that leaders’ effort to take responsibility during a crisis and communicate their thoughts can harm their own as well as their followers’ well- being. This also seems to be the case when communicating online based on the results of this study. Since social media is becoming a dominant way of triggering collective behavior in organizations; leaders that once used to influence their follower’s anxiety levels in face-to-face communication, seem to equally affect their emotional status by simply tweeting. In this respect, our study demonstrates the potential negative implications of leadership communication via social media by highlighting the impact of being a communicative leader under adverse situations on follower well-being. Based on our findings, one clear implication is related to the way that leaders can use social media communication. We suggest that leaders should be more aware of the impact of their communication to others using CMC and learn to behave more strategically, specifically avoiding comments on issues that might be be perceived as negative by their employees. In addition, mentoring programs or even CMC training could potentially help leaders develop their social media communication skills. # Limitations and future research Our study is not without limitations. Firstly, although our applied ML methodology allows us to score millions of tweets across a long period of time, it could be argued that social media profiles might not reflect actual personality but rather an idealized form of user representation. Yet, previous research has found strong support for assessing personality traits in social media posts. Indeed, we would argue that the anxiety detection algorithm specifically is superior to self-report ratings of anxiety because the onset of anxiety is not always immediately evident to the experiencing person and lead to an increased likelihood of burnout and exhaustion over time if not recognized. However, due to the behavioral residue in CMC, more accurate changes in anxiety can be detected. Secondly, one could argue that the leader-follower dyads in the present paper are not sufficient, since we do not guarantee that followers actually read their leaders’ posts. Based on the work of Kramer, we argue that it would not be computationally feasible to assess whether followers read all of their leaders’ posts. For example, a follower viewing their leaders’ newsfeed likely will be presented with the most recent or important (i.e., pinned) posts, with more posts appearing on each page-down, which go unread. It would be unlikely to expect that social media users’ affect is influenced by unread posts. Hence, our decision to not systematically include or exclude followers based on which leader posts followers had viewed constitutes an error, which would only make it more difficult to find statistically significant results. Yet, we do find that leaders’ tweets are significantly associated with follower state anxiety as expressed in followers’ subsequent posts. Thirdly, we recognize that we could have restricted our data collection approach to followers who actively commented on their leaders’ posts. While this might have increased our observed effect size, we chose not to restrict our sample in this way, due to possible existing disclosure norms. We argue that it is likely that even if the transferred leader emotion is not felt by their followers, followers might feel required to respond, most likely in agreement with and in support of their leaders’ respective post content and tone. This is likely because a) organizational leaders hold positions of status, authority, and power over lower-ranked employees and b) interactions on Twitter constitute public statements. Moreover, leaders who communicate using social media usually do not direct their posts to specific followers but rather use these platforms to communicate with entire communities. Hence, by not restricting our sample to only responding followers, we can be confident that any observed effects are not due to disclosure norms or proximal communication. Finally, although the presented regression models are based on multi-day lagged analyses, in which a variable on a given day predicts another variable on subsequent days, this study is not based on an experimental research design and therefore causal claims cannot be made to the same extent. However, we would argue that studying wide- reaching phenomena such as the effect of a real (not simulated) crisis context in such a large sample would be difficult if not impossible to do using a truly experimental research design, as randomizing participants into various experimental conditions is likely not feasible and brings with ethical considerations as well. Instead, the presented approach allows the study of interactions in a naturally occurring non-obtrusive manner. [^1]: The authors have declared that no competing interests exist.
# Introduction Adipose tissue is a specialized, loose connective tissue laden with adipocytes and sometimes stored in ectopic locations, such as near muscle, liver, the heart, and within the abdominal cavity. Adipose tissue usually infiltrates within or around the muscle fibers, with adipose tissue infiltration “within” the muscle referred to as intramuscular adipose tissue or “IntraMAT”. IntraMAT is an important measure in research, and in older individuals and patients with Duchenne muscular dystrophy, IntraMAT is inversely associated with insulin resistance, force generation, and a range of functional ability tests (e.g., 6-minute walk distance, gait speed, sit-to-stand, stair descent time, and timed up and go tests). Furthermore, studies show that IntraMAT values of the lower limb muscles, such as the thigh muscles, increase with age. IntraMAT values therefore aid in our understanding of insulin resistance, muscle functional ability, and muscle quality and could be used to create thresholds that identify patients who may be at risk of developing type 2 diabetes in the future. The values could also help in identifying patients with early myopathy as a window into muscular dystrophy, for example. Magnetic resonance imaging (MRI) is a powerful tool to assess the quality and quantity of human skeletal muscles, and to investigate the IntraMAT of target muscles, T1-weighted (T1W) imaging and the Dixon method have been frequently used in previous studies. IntraMAT has been calculated based on T1W images combined with segmentation analysis in previous studies. The Dixon method is an alternate, relatively inventive technique for the evaluation of fat fractions based on the different precessional frequencies of water and lipid protons. Previous studies confirmed that IntraMAT content by both the T1W and Dixon methods were consistent or correlated with muscle biopsy, which is the gold standard to measure IntraMAT. The IntraMAT content in the thigh measured by the three-point Dixon (3PD) method was approximately 3% in younger men and 6% in older men. Others have reported specific muscle values: 3.0% for younger and 5.4% for older men in the vastus lateralis (VL) and 3.9% for younger and 7.2% for older men in the biceps femoris-long head (BF-L). Recently, a study reported no significant difference (P = 0.83) between the fat fractions determined by two-point Dixon (2PD) and lipocytes assessed by histological analysis taken from muscle biopsies, indicating that the Dixon technique is reliable for the quantification of IntraMAT. However, T1W imaging has been widely used for the diagnosis of diseases as well as in medical imaging research. Previous studies with T1W imaging have used histograms to separate skeletal muscle tissue and adipose tissue based on differences in pixel intensity, which mainly comes from differences in the spin-lattice relaxation times between skeletal muscle and adipose tissues, whereas others have used different muscle and adipose tissue threshold settings for discrimination \[, –\]. We previously reported that IntraMAT content using T1W images is 4.6% in the VL and 16.2% in the BF-L in older subjects (70.7 ± 3.8 years). These values are slightly higher than those taken by 3PD images (0.8% for VL and 9% for BF-L). Several factors affect IntraMAT content between two imaging techniques, including subject characteristics, differences in spatial resolution, and differences in the sensitivity of the imaging apparatus to adipose tissues located inside and outside of the muscle cells. The proportion of IntraMAT by 2PD is consistent with the results of histological evaluations from biopsy specimens, and a correlation between the proportion of IntraMAT by T1W and biopsy results has been reported. It is therefore plausible that IntraMAT values of T1W and 2PD may match; yet, to the best of our knowledge, no study has tested the agreement between T1W imaging and Dixon imaging for IntraMAT content on the same subjects using the same locations. Thus, the purpose of this study was to compare IntraMAT content determined by 2PD imaging and T1W imaging, calculated using thresholding techniques. This study is useful for judging whether there is inconsistency in the IntraMAT contents by the two methods reported in previous studies. Those results are instrumental for selecting the imaging method used for diagnosis of IntraMAT in diseased patients. We hypothesized that IntraMAT values obtained with T1W would be slightly higher than those taken using 2PD. # Methods ## Subjects Nineteen younger men (26.2 ± 4.9 years, 22.2 ± 2.2 kg/m<sup>2</sup>) and 13 older men (72.2 ± 6.0 years, 23.9 ± 1.6 kg/m<sup>2</sup>) volunteered to participate in this study. Subjects were widely recruited healthy and non-obese (BMI \< 25.0 kg/m<sup>2</sup>) volunteers using a poster in the library. This experiment was conducted as a prospective study in 2016. Before the experiment, the procedure, purposes, risks, use data for research and benefits associated with the study were explained and written consent was obtained with all subjects. This study was approved by the Research Ethics Committee (Nagoya University; 2016–0254) and the investigation was performed according to the principles outlined in the Declaration of Helsinki. ## MRI acquisition Subjects were assessed with a 3.0 T whole-body MRI scanner (MAGNETOM Verio, Siemens Healthcare Diagnostics K.K., Tokyo, Japan). Subjects were placed in a supine position and images of the thigh were acquired using a body coil. We took consecutive images of the entire thigh using T1W and 2PD imaging. We defined the mid-thigh according to markers attached at the middle point between the greater trochanter and the lateral condyle of the femur. The durations for image acquisition were 4 min for the T1W and 8 min for the 2PD. T1W spin-echo transaxial images of the right thigh were collected with the following sequences: three-dimensional, TR = 604 ms; TE = 12 ms; flip angle = 120°; optimized field of view = 256 × 256 mm; slice thickness = 10 mm; and interslice gap = 0 mm. All subjects were instructed to remain as still as possible. In- phase and out-of-phase water and fat transaxial images were obtained to create water and fat images for analysis. 2PD images of the right thigh were acquired with the following sequences: three-dimensional, TR = 20 ms; TE1 = 2.450 ms; TE2 = 3.675 ms; flip angle = 9°; optimized field of view = 288 × 288 mm; slice thickness = 5 mm; and interslice gap = 0 mm. We calculated the cross-sectional area (CSA) of the thigh in younger and older subjects from T1W and 2PD images to ensure that the same locations were being compared. As a result, CSAs of the thigh were not significantly different between T1W and 2PD values (T1W, 196.2 ± 30.7 cm<sup>2</sup>; 2PD, 195.3 ± 31.0 cm<sup>2</sup>). Therefore, we were able to measure the same area of the thigh using T1W and 2PD imaging. ## Analysis of thigh composition MR images were read in random order for the analysis. We measured CSA of skeletal muscle and IntraMAT of the VL, adductor magnus (AM), and BF-L at the mid-thigh. Serial axial images were used to help identify muscle boundaries. ### 1) T1-weighted images We used the Medical Image Processing, Analysis and Visualization software (version 4.4.0; National Institutes of Health, Bethesda, MD) to analyze images on a personal computer. The MRI data analysis procedure was essentially the same as described in previously studies. Briefly, we first corrected for image heterogeneity caused by suboptimal radiofrequency coil uniformity, or gradient- driven eddy currents, using a well-established nonparametric nonuniform intensity normalization (N3) algorithm. This step was essential for subsequent analyses that assume homogenous signal intensities across images. Optimized image correction parameters (N3) were determined (end tolerance, 0.0001; maximum iterations, 100; signal threshold, 1; field distance, 25 mm; subsampling factor, 4; Kernel full width, half maximum of 0.15; Wiener filter noise, 0.01), and the same parameters were applied to all images. Second, we calculated the CSA of IntraMAT content at the mid-thigh in T1W images using the threshold method, as described previously. We then drew six regions of interest (ROIs) of 25 mm<sup>2</sup> each, with three ROIs on the vastus intermedius and three ROIs on the subcutaneous adipose tissue. The vastus intermedius, which is 99% skeletal muscle, was chosen to obtain a pure skeletal muscle peak in the pixel number–signal intensity histogram. The total number of pixels within the six ROIs was used to produce a frequency distribution and histogram of all pixels and signal intensities. To separate muscle and adipose tissues in the pixel number–signal intensity histogram with minimal investigator bias, we implemented the Otsu threshold method, a reliable histogram-shape–based thresholding technique used in medical imaging analysis. To minimize manual tracing-induced errors on thresholding values, the mean of three trials was used, and the values were applied to the VL, AM, and BF-L. After carefully tracing the edge of each muscle, the following parameters were calculated: 1) the total number of pixels within the ROI; 2) the number of pixels with a signal intensity lower than the threshold value (skeletal muscle); and 3) the number of pixels with a value higher than the threshold value (IntraMAT). The IntraMAT content for each muscle was then calculated using the following equation: <img src="info:doi/10.1371/journal.pone.0231156.e001" id="pone.0231156.e001g" /> IntraMAT content ( % ) = ( IntraMAT pixel numbers ) / \[ ( skeletal muscle pixel numbers ) \+ ( IntraMAT pixel numbers ) \] × 100 ### 2). Water and fat images by 2-point Dixon Dixon images were analyzed using ImageJ (version 1.44; National Institutes of Health, Bethesda, MD, USA). ROIs were drawn to match the corresponding T1W image voxel locations for fat and water, calculated using the 3.5 ppm chemical shift between water and lipid at 3.0 T. From these VOIs, the mean signal intensities from water and fat images by 2PD were measured to create a Dixon-based fat–water ratio using the following equation: <img src="info:doi/10.1371/journal.pone.0231156.e002" id="pone.0231156.e002g" /> IntraMAT content ( % ) = 100 × Fat mean intensity / ( Water mean intensity \+ Fat mean intensity ) ### Reproducibility analysis Manual segmentation of the thigh muscle compartments was repeated twice in 10 randomly selected subjects by one researcher with 5 years of experience in muscle image analysis to assess intra-observer reproducibility of the segmentation process. Intraclass correlation coefficient \[(ICC 2.1)\] in individual muscles indicated values ranging between 0.92 to 0.97 in T1W and 0.89 to 0.99 in 2PD for IntraMAT content (all P \< 0.001). ## Statistical analysis IntraMAT content calculated by the two MRI methods were compared using Student’s unpaired t-test. We compared IntraMAT content by the two MRI methods between the younger and the older men using Student’s unpaired t-test. We estimated agreement using the 95% limits of agreement method developed by Bland and Altman, where the difference between the adipose tissue obtained with water and fat images by T1W and 2PD imaging is plotted against their means. The limits of agreement were between 1.96 SD and -1.96 SD. All continuous variables are expressed as the mean ± SD; a two-tailed P \< 0.05 was considered to indicate statistical significance. All statistical analyses were performed using IBM SPSS statistics (version 22.0; IBM, Tokyo, Japan). # Results shows the IntraMAT content and range for the VL, AM, and BF-L. IntraMAT content in the VL determined by T1W imaging was not significantly different from that determined by 2PD; however, IntraMAT contents in the AM and BF-L were significantly higher with T1W than with 2PD imaging. In the older group, IntraMAT content was significantly higher than in the younger group in all muscles and with both the T1W and 2PD methods. The relationship between IntraMAT content determined using T1W and 2PD is shown in 2. There was a high positive correlation between these two values using T1W and 2PD for the different muscles (VL; r = 0.735, P \< 0.01, AM; r = -0.717, P \< 0.01, BF-L; r = 0.790, P \< 0.01, VL + AM + BF-L; r = 0.686, P \< 0.01;). shows the Bland-Altman plot of IntraMAT content for each of the three muscles determined using the T1W and 2PD methods. The limits of agreement were between 1.96 SD and -1.96 SD. However, Bland-Altman plots showed that the Pearson r (difference versus mean) was -0.698 to -0.914 (P \< 0.01;); therefore, a proportional bias was observed in all muscles. This suggests that subjects with higher IntraMAT contents showed larger differences between the two imaging modalities. The limits of agreement in the differences between T1W and 2PD vs. the mean of the two measurements were between 1.96 SD and −1.96 SD. Mean value, ------ 95% limits of agreement. Y = 0 is a line of perfect average agreement. Thus, we next investigated why there were such differences in IntraMAT content between the two methods for different muscles. compares a representative T1W image, its respective binary image, and the fat image from 2PD for three subjects with relatively higher, middle, and lower IntraMAT contents. Comparing the arrow portions of the T1W image with the binary image and the 2PD fat image, it is clear that the portions of adipose tissue do not coincide. This tendency was observed particularly in subjects with high IntraMAT in T1W images (Figs). Moreover, in subjects with a large difference in IntraMAT content between T1W and 2PD, we found a difference in the thresholds in T1W and boundary values of the muscle and adipose tissue when applying the 2PD IntraMAT content to the T1W image. Finally, we selected 10 subjects at random (five young subjects and five older subjects) and calculated the difference in the threshold in T1W and boundary values for muscle and adipose tissue and the values after applying the IntraMAT content of the 2PD to the T1W image for those subjects. Interestingly, there was a high negative correlation between these two values for the different muscles (VL; r = -0.874, P \< 0.01, BF-L; r = -0.944, P \< 0.01, VL + BF-L; r = -0.955, P \< 0.01;). This result indicates that the difference in IntraMAT content between T1W and 2PD methods depends on the threshold setting. Difference in IntraMAT content (%) = 2PD’s IntraMAT content − T1W’s IntraMAT content. Difference in signal intensity (AU): Threshold of muscle and adipose tissue on T1W images − boundary value of muscle and adipose tissue when applying 2PD IntraMAT content to T1W images. # Discussion The main findings of this study were that: 1) IntraMAT content was significantly higher in T1W images than in 2PD images for the AM and the BF-L, but not the VL; and 2) systematic errors were found between T1W and 2PD imaging techniques that can be explained by variations in the threshold setting. Many attempts have been made to calculate adipose tissue content within soft tissues, such as skeletal muscle, heart, and liver, using the Dixon method in healthy subjects and those with diseases. Previous studies have reported that IntraMAT content of the VL by the Dixon method ranged from 3.0 to 3.2% in younger subjects. However, in the present study, the IntraMAT content by 2PD for young subjects was 10.9 ± 1.2% in the VL, which was three times higher than that found in previous studies. Imaging methods (2PD or 3PD), differences in volume and cross-sectional area, subject characteristics or physical activity levels could all explain these differences in IntraMAT values. In the present study, the IntraMAT content by T1W was two times higher than that found in previous studies. A previous study reported that IntraMAT contents of the VL were 3.6% and 7.7% in younger and older subjects, respectively, who were men and women. Subject characteristics, physical activity levels, and sex differences could all explain these differences in IntraMAT values. We showed that IntraMAT content in the VL and BF-L were 7% and 19%, respectively, in the older group, and 3% and 13%, respectively, in the younger group. The IntraMAT in the BF-L was thus 3- to 4-fold higher than that in the VL. Overend and colleagues, who used computed tomography (CT) imaging for relatively non- contractile tissue (i.e., IntraMAT and connective tissues) within the quadriceps and hamstring muscle groups, found that IntraMAT was 3.6% and 5.4%, respectively, in younger subjects, and 7.7% and 13.6%, respectively, in older subjects. This difference in content aligns with our findings, and thus, it is a reasonable finding that the levels of adipose tissue within the VL were less than that in the BF-L. We also previously measured intramyocellular lipids (IMCL) and extramyocellular lipids (EMCL) using <sup>1</sup>H magnetic resonance spectroscopy (<sup>1</sup>H-MRS) in male and female subjects across a broad age range, and found collective lipid levels ranging from 9.4 to 40.2 mmol/kg wet weight, respectively, in the VL and from 10.5 to 94.0 mmol/kg wet weight, respectively, in the BF-L. Accordingly, it is plausible that IntraMAT in the VL is likely to be lower than that in the BF-L on various imaging modalities, including MRI, CT, and <sup>1</sup>H-MRS. Although IntraMAT content in the AM and BF-L by T1W were significantly higher (by 3.0% and 8.7%, respectively) than that found by 2PD, there was no difference in IntraMAT content for VL between the two methods, suggesting that the absolute conformity of the IntraMAT content was muscle-specific. Thus, particular attention must be paid to the differences noted for AM and BF-L between the two methods. As shown in, T1W and threshold-adjusted T1W images gave the same IntraMAT values as that found with 2PD; i.e., 2PD boundary value, located at the right edge of the muscle peak in the VL, indicated that almost all of the ROI was skeletal muscle using this thresholding technique. The 2PD threshold was higher than that for T1W, and this may have led to false results in terms of the amount of adipose tissue in the AM and BF-L, even though the area was categorized as skeletal muscle tissue. As a result, IntraMAT content determined by T1W was significantly higher than that determined by 2PD in the AM and BF-L. The further assessments conducted in show that the difference in IntraMAT content determined by T1W and 2PD was almost perfectly associated with differences in threshold values in randomly selected subjects; i.e. differences were dependent on signal intensity. This result suggests that the difference in IntraMAT by T1W and 2PD was closely related but dependent on threshold settings. This would be the main reason for the systematic errors in the Bland-Altman analysis of VL, BF-L and AM shown in. Inhomogeneity of the magnetic field in the MR system may also affect IntraMAT content values for AM and BF-L. Indeed, a previous study reported that the magnetic field on T1W images was not constant, with correction necessary to rectify for such inhomogeneity effects. We used the well-established N3 algorithm to correct for shading in the images due to heterogeneity linked to suboptimal radiofrequency coil uniformity or gradient-driven eddy currents. Although we confirm that shading was much decreased compared with the original MR image, there is still some shading evident (e.g., from anterior to posterior or medial to lateral). We expect that these persistent shading defects may have influenced our IntraMAT values. In addition, we considered that the location of muscles and imaging methods could be related changes in the signal-to-noise ratio (SNR). We measured SNR with the subtraction method, which is the most commonly used method. In this method, a difference image was obtained by subtracting two identical images. We measured the SNR between T1W and water images, and T1W and fat images for 10 subjects, in the same size ROI of four corners (ROI 1 to ROI 4) to compare the SNR between the ROIs. The SNR in ROI 4 was a significantly higher than in ROI 1, but there was no significant difference between the other ROIs. In this study, VL was on the ROI 3 side, AM was on the ROI 2 side, and BF-L was on the ROI 4 side. These results suggest that the effect of SNR was not observed independent of the location of muscles in this study. Unfortunately, we could not use the same voxel size between the T1W and 2PD imaging methods because of the MR device; therefore, the effect of the SNR on the results of IntraMAT could not be excluded. Furthermore, the effect of local distortion in the MR by distance of the center of the gantry to the body coil and the measurement site could also be related to the difference in IntraMAT content between T1W and 2PD. Thus, this study had some important limitations. Further human and phantom experiments under stable conditions of voxel size, SNR, and specific absorption rate would be required to reveal what caused the difference in IntraMAT content between T1W and 2PD. Importantly, T1W has been used to measure IntraMAT in previous studies, with a close relationship between IntraMAT content by T1W and histochemical analysis of muscle biopsies. Furthermore, T1W can be used in any MRI systems from lower to higher magnetic fields. However, the Dixon method is recommended for MRI with a 1.5T or higher because of its technical limitations in being able to clearly distinguish between water and fat signals. IntraMAT content by both T1W and 2PD methods was consistent or correlated with muscle biopsy in previous studies, However, when T1W imaging is used, extra care must be taken of the IntraMAT content discrepancy with 2PD found in this study. The discrepancy in IntraMAT contents of the T1W and 2PD methods is unknown in this study. We previously showed that IntraMAT content by T1W primarily represents EMCL, not IMCL, in the VL and BF-L muscles. On one hand, this could be because a relatively 5 to 10 times larger amount of lipids accumulates in the interstitium compared with the cytoplasm of the muscle cells. On the other hand, Fischer et al. (2014) reported that IntraMAT content by 2PD was closely related with the sum of IMCL and EMCL contents detected by <sup>1</sup>H-MRS (r = 0.918 to 0.990, P \< 0.001). However, it has not been established that IntraMAT contents by both T1W and 2PD methods reflected EMCL only, as shown by Akima et al., or IMCL and EMCL, as shown in Fischer et al. using <sup>1</sup>H-MRS. The major limitation in the present study was that we did not directly compare biopsy or <sup>1</sup>H-MRS results with MR images (T1W and 2PD), and we did not detect the exact changes in fat content (IMCL and EMCL) or total adipose tissue content. Therefore, it is important that additional research is carried out to confirm the validity of segmentation analyses for T1W and 2PD images by combining IMCL and EMCL data with <sup>1</sup>H-MRS or histochemical analysis. # Conclusions Although there was no significant difference between T1W and 2PD in IntraMAT content for the VL, the IntraMAT content using T1W was significantly higher in the AM and BF-L as compared with the 2PD method. We suggest that this is primarily because of differences in threshold settings when using T1W, particularly for measurements of the BF-L. Thus, these results suggest that care should be taken when selecting an imaging modality for IntraMAT, especially for patients who would be suspected of having higher IntraMAT values. The authors gratefully thank the volunteers for participation as well as Dr. Haruo Isoda, radiological technologist, Mr. Akira Ishizuka (Graduate School of Medicine, Nagoya University), and nurse, Yoko Onoda. This study was supported in part by a Grant-in-Aid for JSPS Research Fellow (to MO) and a Suzuken Memorial Foundation and Editage Grant (to NT). 10.1371/journal.pone.0231156.r001 Decision Letter 0 Sanada Kiyoshi Academic Editor 2020 Kiyoshi Sanada This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Nov 2019 PONE-D-19-29508 Comparing intramuscular adipose tissue on T1-weighted and two-point Dixon images PLOS ONE Dear Dr. Ogawa, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. We would appreciate receiving your revised manuscript by Dec 29 2019 11:59PM. When you are ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission- guidelines#loc-laboratory-protocols> Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out. We look forward to receiving your revised manuscript. Kind regards, Kiyoshi Sanada, PhD Academic Editor PLOS ONE Journal Requirements: 1\. When submitting your revision, we need you to address these additional requirements. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf> and <http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_f ormatting_sample_title_authors_affiliations.pdf> 2\. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified who gave consent and whether it was informed. 3\. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. 4\. Please amend either the abstract on the online submission form (via Edit Submission) or the abstract in the manuscript so that they are identical. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: In this manuscript, the author aimed to compare intramuscular adipose tissue content determined by two-point Dixon imaging and T1-weighted imaging, calculated using thresholding techniques. This study showed that there was no significant difference between T1W and 2PD in IntraMAT content for the VL, whereas the IntraMAT content using T1W was significantly higher in the AM and BF-L as compared with the 2PD method. Additionally, this is primarily because of differences in threshold settings when using T1W. This manuscript may be potentially interesting and clinical significance, however, there are several key concerns that need to be addressed. Major comments: 1\. The authors concluded that T1W imaging is likely to be the better option for IntraMAT measurements. It is ambiguous about the process of this conclusion. Therefore, the authors need to discuss the more clearly. IntraMAT content determined by T1W was significantly higher than that determined by 2PD in the AM and BF-L. Which value is appropriate? 2\. What is the gold standard to assess IntraMAT? T1W? Dixon? biopsies? Please describe in the introduction. 3\. Is the order of measurement randomized? Can the authors clarify the study protocols? 4\. In the Table 1, the authors should present the comparison between young and older. It will be interesting to know whether the effects of aging differ between T1W and 2PD methods. 5\. Instead of showing only the relationship between difference in IntraMAT contents and difference in signal intensity, Figure should present the relationship between IntraMAT content determined by T1W and 2PD 6\. In the Figure 3, is there the representative image of Lower IntraMAT? 7\. the authors described “the IntraMAT content of this study by 2PD which was three times higher than that found in previous studies”. How is T1W compared to previous studies? Please discuss. Minor comments: 1\. Please change “positive correlation” to “negative correlation” (page 14, line 7). Reviewer \#2: The paper brings some original data and I have read it with a great interest. Nevertheless I have few comments and suggestions: The MR image shows a change in contrast when there is a distance change from the magnetic field center in the gantry. Therefore, in addition to this experimental data, 1) Change the subject's imaging position, 2) After imaging at a site away from the center of the bore, recalculate to see if similar results are obtained. In addition, the calculated MR signal intensity depends on the voxel size. Therefore, change the Voxel size and verify that similar results are obtained. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0231156.r002 Author response to Decision Letter 0 5 Jan 2020 Journal Requirements: Q1. When submitting your revision, we need you to address these additional requirements. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf> and <http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_f ormatting_sample_title_authors_affiliations.pdf> A1. I confirmed PLOS ONE style templates and changed our maniscript. Q2. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified who gave consent and whether it was informed. A2.I added Ethics Statement in online and manuscript. However, if you need to more information, please tell me. Q3. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. A3.Our data contain potentially identifying information and ethics committee didn't allow the connecting internet those data. Data are available from Ethics Committee for researchers who meet the criteria for access to confidential data. Ethics committee of Nagoya University Graduate School of Medicine Address: Nagoya University 65 Tsurumai, Showa-ku, Nagoya, Aichi 466-8550, Japan E-mail: <[email protected]> Q4. Please amend either the abstract on the online submission form (via Edit Submission) or the abstract in the manuscript so that they are identical. A4. We apologize for this oversight. We confirmed and changed our abstract in manuscript and online. Q5. Your ethics statement must appear in the Methods section of your manuscript. If your ethics statement is written in any section besides the Methods, please move it to the Methods section and delete it from any other section. Please also ensure that your ethics statement is included in your manuscript, as the ethics section of your online submission will not be published alongside your manuscript. A5. We deleted ethics statement in end of our manuscript. We written ethics statement in Methods. 10.1371/journal.pone.0231156.r003 Decision Letter 1 Sanada Kiyoshi Academic Editor 2020 Kiyoshi Sanada This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 29 Jan 2020 PONE-D-19-29508R1 Comparing intramuscular adipose tissue on T1-weighted and two-point Dixon images PLOS ONE Dear Dr. Ogawa, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. We would appreciate receiving your revised manuscript by Mar 14 2020 11:59PM. When you are ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission- guidelines#loc-laboratory-protocols> Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out. We look forward to receiving your revised manuscript. Kind regards, Kiyoshi Sanada, PhD Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors have addressed my major concerns, and the manuscript has much improved. I have no further comments for the authors. Reviewer \#2: I understand your response. It can be inferred from the text that T1w and the 2PD method was performed continuously. Since it is described as "consective," it is not necessary to describe it as "continuously." As you know and use it, MR devices have different image data and contrast (power from changes in SN and SAR) depending on continuous imaging and shimming conditions. These problems have been corrected in the image conversion process. If it is far from the center of the gantry, the effect of local distortion, due to MR, also appears. It is not a matter of continuity, but the MR system. In addition, the distance between the body-coil and the measurement site is also an issue. This paper is a valuable paper that utilizes an uncomplicated evaluation and hopes for future development. Therefore, it is essential that the reliability of the published data is established and whether the reader can obtain comparative data by the same method. I understand your response is giving up its verification. The previous advice does not consider continuous imaging as a problem but points out the reproducibility of data. I highly recommend you to describe problems with MR equipment properly. The advice for voxel size is the same. These points should also be considered in the limitations. If you use a method that can be evaluated easily, it is recommended that you consider it properly. \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0231156.r004 Author response to Decision Letter 1 11 Mar 2020 Dear Dr. Sanada and Reviewer \#2 We wish to express our deep appreciation to the reviewer for the insightful comment on our manuscript. We apologize for not answering your previous question perfectly. We have worked hard to incorporate your feedback and hope you will agree that that these revisions have made our manuscript suitable for publication in PLOS ONE. Sincerely, Madoka 10.1371/journal.pone.0231156.r005 Decision Letter 2 Sanada Kiyoshi Academic Editor 2020 Kiyoshi Sanada This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 18 Mar 2020 Comparing intramuscular adipose tissue on T1-weighted and two-point Dixon images PONE-D-19-29508R2 Dear Dr. Ogawa, We are pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it complies with all outstanding technical requirements. Within one week, you will receive an e-mail containing information on the amendments required prior to publication. When all required modifications have been addressed, you will receive a formal acceptance letter and your manuscript will proceed to our production department and be scheduled for publication. Shortly after the formal acceptance letter is sent, an invoice for payment will follow. To ensure an efficient production and billing process, please log into Editorial Manager at <https://www.editorialmanager.com/pone/>, click the "Update My Information" link at the top of the page, and update your user information. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, you must inform our press team as soon as possible and no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. With kind regards, Kiyoshi Sanada, PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: (No Response) Reviewer \#2: I have no further comments for the authors. I read the revised version. Although the experimental data can be easily measured by using the MR device, the verification of the data becomes uncertain if the experiment is performed without understanding its essence. I think the paper is more polite than the previous one. I am looking forward to reading your future paper. \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No 10.1371/journal.pone.0231156.r006 Acceptance letter Sanada Kiyoshi Academic Editor 2020 Kiyoshi Sanada This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 23 Mar 2020 PONE-D-19-29508R2 Comparing intramuscular adipose tissue on T1-weighted and two-point Dixon images Dear Dr. Ogawa: I am pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. For any other questions or concerns, please email <[email protected]>. Thank you for submitting your work to PLOS ONE. With kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Kiyoshi Sanada Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Nippon Sport Science University, Setagaya-ku, Tokyo, Japan
# Introduction Mouse parvovirus (MPV) is a commonly recognized infectious agent relatively prevalent in mouse colonies and found in cell culture which can impact research results, such as potentiate rejection of tumor allograft. Elimination of MPV from infected colonies relies principally on rederivation; however, there are aspects of MPV infection that suggest mice may become immune tolerant, for example MPV has been detected in the gametes of mice. Since parvoviruses replicate readily in rapidly dividing tissues such as a growing embryo, and gametes may be infected with MPV, vertical transmission of the virus could lead to early in utero exposure. There have been recent studies demonstrating rederivation by embryo transfer may not be effective at eliminating minute virus of mice (MVM), a parvovirus similar to MPV. In these studies, fertilized oocytes and morulae were exposed to varying concentrations of MVM. Recipient dams and their progeny receiving embryos infected with as little as 100 ID<sub>50</sub> seroconverted to MVM. The antibody titers of the progeny dissipated after 20 weeks suggesting they were maternal antibodies, and the authors mentioned unpublished observation that MVM was shed in the experimental colonies for up to one year. One plausible cause for this phenomenon is immune tolerance due to in utero exposure. Within the last several years, we have confirmed MPV seropositivity in three separate immunocompetent mouse breeding rooms, resulting in attempts to eliminate the virus from the colony. On each occasion, at least 100 mice in the colony were tested, and none were identified as seropositive. Although a large representative number of mice were tested, not all mice were tested, so the use of microisolator caging may have sequestered low prevalence breaks within the room, and even within a cage, variation among cage mates could occur. We have also had experiences in which sentinel mice would seroconvert to MPV and subsequent sentinel testing was seronegative for several months, only to be followed by another seropositive sentinel. The cost of confirming or culling colonies due to erratic serological responses can be significant. This erratic serologic detection in sentinels is not uncommon based on conversations with colleagues and presentations at the national meetings. The sporadic nature of the serologic detection of MPV suggested the possibility that the immune response to this pathogen is not consistent with the typical exposure- seroconversion responses noted for most pathogens. The primary rational for pursuing this study was to determine if immune tolerance was responsible for the sporadic serologic responses seen in colony mice during an MPV serologic screen. If parvovirus persists following in utero exposure, it could explain the seroconversion of naïve sentinels, but failure to confirm MPV in subsequent serologic screening of colony animals. For a host to develop an appropriate immune response to pathogens, T cell development occurs in the thymus to make CD4+ or CD8+ T cells. In addition, these cells must undergo positive and negative selection. Positive selection enables T cells to recognize self major histocompatibility complex, while negative selection removes T cells that react to self antigens. The end result is CD4+ and CD8+ T cells capable of recognizing foreign antigens. The mouse thymus begins development at embryonic day 10–12 and is not completely developed until 3–4 weeks after birth. Early experiments with mice have demonstrated one can induce tolerance to foreign antigens if they are introduced in the neonatal period prior to thymic maturation. More recently administration of adeno- associated viral vectors in utero and at 2 days post-partum failed to develop detectable antibodies to the viral vector and subsequent challenge with the adeno-associated virus demonstrated tolerance as viral vector expression continued up to 11 months. Immune tolerance due to in utero exposure to viral pathogens has been demonstrated with lymphocytic choriomeningitis virus of mice, and pestiviruses such as bovine viral diarrhea virus and hog cholera. Similarly, we hypothesized that mice exposed in utero develop tolerance to MPV infection resulting in the inability to mount a serologic response following post natal challenge with persistent infection. However, we found that offspring mice exposed in utero by oral inoculation of the dam do not develop tolerance to MPV infection. # Materials and Methods ## Mice Eighteen female and six male, 10–12 week old outbred Crl:CD1(ICR) mice were purchased from Charles River Laboratories (Wilmington, MA) which were pathogen free for Sendai virus, pneumonia virus of mice, mouse hepatitis virus, minute virus of mice, mouse parvovirus, mouse norovirus, Theiler’s murine encephalitis virus, reovirus, rotavirus, lymphocytic choriomeningitis virus, ectromelia virus, mouse adenovirus, mouse cytomegalovirus, K virus, polyoma virus, Hantavirus, lactate dehydrogenase elevating virus, mouse thymic virus, Bordetella bronchiseptica, CAR bacillus, Citrobacter rodentium, Corynebacterium kutcherii, Helicobacter species, Klebsiella oxytoca, Mycoplasma pulmonis, Pasteurella pneumotropica, Salmonella species, Streptobacillus moniliformis, Streptococcus pneumonia, Clostridium piliforme, pinworms, and ectoparasites. Mice were housed in individually ventilated caging (Thoren Caging System, Inc, Hazelton, PA). Mice were provided Teklad Irradiated Diet 2918 (Harlan Laboratories, Madison, WI) and filter sterilized water and allowed to acclimate for 7 days prior to initiating the studies. All mice were maintained under a 12:12-h light:dark cycle at temperatures of 21 to 24°C. All animal experiments were approved by the Colorado State University Institutional Animal Care and Use Committee (09-143A-01) and carried out with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. ## Experimental Infections Mice were maintained in harems of three females to one male and allowed to breed naturally. Females were checked for vaginal plugs daily to confirm mating. Following successful plugs, female mice were separated. Thirteen time-pregnant female mice were inoculated with 50 ID50 of MPV1e, a mouse passaged strain of MPV, via oral gavage at day 5 and 12 gestation as previously described. Briefly, MPV1e was isolated from a naturally occurring infection and was maintained by oral inoculation of naïve mice with filter-sterilized tissue homogenate. The ID50 was determined in in weaning mice by oral inoculation of serial dilutions. Four control time-mated female mice received a control inoculation of 100 microliters of phosphate buffered saline at day 5 and 12 gestation. The pups were fostered to time-mated, MPV negative dams prior to the presence of a milk spot. Once fostered, the original dams were euthanized by asphyxiation with carbon dioxide, blood was collected by cardiocentesis, and feces was collected for PCR analysis. At 21 days, mice were weaned from the foster dams. Foster dams were euthanized, and blood and tissues collected as described above. Blood from the tail vein and feces were collected from weanling mice to confirm MPV negative status. After weanling mice were confirmed MPV seronegative and PCR negative, 10 male and 10 female mice were orally challenged by gastric gavage with MPV1e, and 5 of each sex received mock inoculum as control, as described above and tested weekly by fecal PCR. Four weeks post-inoculation, mice were euthanized with carbon dioxide. Blood and feces were collected to evaluate viral presence by serology and PCR, respectively. ## Mouse parvovirus Assays Mice were monitored for fecal shedding by PCR every other day post-inoculation for the first 7 days to confirm MPV infection. Feces was collected and homogenized in PBS as previously described, and DNA isolated from fecal pellets using a Qiagen DNA extraction kit according to the manufacturer’s instructions. Quantitative (real time) PCR reactions were performed in a BioRad iCycler using well established primers and cycle times. Serum was collected by centrifugation of blood at 3,000 rpm for 10 minutes and diluted 1:5 in saline. Serologic assays using the Multiplex Fluorometric ImmunoAssay® were performed by Charles River Laboratories using their established protocols to evaluate MPV specific antibodies. Capture antigens used on the Multiplesx Fluorometric ImmunoAssay® platform included the non-structural protein NS-1 and two virus capsid proteins (VP-2) for MPV-1 and MPV-2. ## Statistical Analysis Analysis was done using SAS 9.4 (SAS Institute, Inc., Cary, NC). Fisher’s exact test was used to compare positive rates (PCR or serology) across inoculation groups (sham/sham, MPV/sham, sham/MPV, MPV/MPV). A separate test was conducted for each testing method and day (PCR- 0, 2, 4, 6, and 28; serology 0 and 28). Similarly, Fisher’s exact test was used to compare conception rates across inoculation groups (sham and MPV). A homoscedastic two-tailed t-test was used to compare litter size between MPV and sham infected dams. P-values less than 0.05 were considered statistically significant. # Results ## Mouse parvovirus infectivity Of the thirteen pregnant mice inoculated with MPV at day 5 and 12 gestation, eleven of them were confirmed infected by fecal PCR. The foster dams were all fecal PCR negative for MPV. The weanling mice were all PCR and serologically negative to MPV prior to infection. After confirming they were MPV negative, 21 day old weanling mice from MPV inoculated or sham inoculated dams were assigned to two groups for subsequent MPV challenge. Group 1 had 20 mice (10 females, 10 males) from sham inoculated dams that received a sham inoculation; group 2 had 9 mice (5 females, 4 males) from MPV inoculated dams that received a sham inoculation; group 3 had 20 mice (10 females, 10 males) from MPV inoculated dams that received a sham MPV challenge inoculation; and group 4 had 23 mice (11 females, 12 males) from MPV inoculated dams that received an MPV challenge inoculation. Serum and feces were collected from weanling mice at day 2, 4, 6 and 28 days post inoculation for serology and PCR evaluation. Sham inoculated mice from sham inoculated (Group 1) and MPV inoculated dams (Group 2) were consistently negative at all time points. Fecal PCR results for MPV inoculated mice from sham inoculated dams (Group 3) demonstrated 19 of 20 mice were MPV positive for at least one time point, and fecal PCR assays for MPV inoculated mice from MPV inoculated dams (Group 4) demonstrated 18 of 23 mice were MPV positive for at least one time point. All foster dams were serologically negative to MPV after mice were weaned. Antibodies to MPV were not detected in the weanling mice originating from either sham or MPV inoculated dams prior to MPV inoculation. At the conclusion of the study, there were no serum antibodies detected in the sham inoculated mice from sham (Group 1) or MPV infected dams (Group 2). Serum antibodies were detected in all MPV inoculated mice from both sham (Group 3) and MPV infected dams (Group 4). demonstrates the frequency of positive fecal PCR assays over time for each group and pre- and post-inoculation serology results. ## Conception rates and litter size Seventeen of the 18 mice bred successfully as indicated by the presence of a vaginal plug. Following oral inoculation with MPV1e, five of the 13 MPV infected dams (38%) had pups. All four of the sham inoculated dams (100%) had viable pups. The litter sizes varied greatly between the two groups, although it was not statistically significant (p = 0.08). The litter size of the MPV infected mice ranged from 1–13 with an average litter size of 6.5, while the sham inoculated mice had a litter size ranging from 8–13 with an average of 10. While this is notable, it was not statistically significant (p = 0.24). # Discussion Clearly, our hypothesis that mice exposed in utero develop tolerance to MPV infection resulting in the inability to mount a serologic response following post natal challenge with persistent infection was not supported in this study. The mice exposed in utero to MPV were able to mount an antibody response to MPV challenge post natal suggesting that oral inoculation of MPV at gestational day 5 and 12 does not result in an immune tolerance in their offspring. Pregnant mice were inoculated with mouse parvovirus at day 5 and 12 gestation. The initial inoculation of MPV would result in a viremia at approximately 12 days gestation, coinciding with thymic development at embryonic day 10–12. The subsequent inoculation at day 12 gestation corresponded with thymic development in an effort to maximize thymic exposure to MPV. At this point in thymic development, the T cells would recognize the parvoviral antigens as self, resulting in immune tolerance. This would manifest itself with the inability to mount an antibody response to MPV as an adult, yet continue to have virus persist within the mice. This was not the case in this experiment as all MPV challenged mice seroconverted after exposure in utero; therefore, minimizing the chances that immune tolerance is responsible for the sporadic serology responses typically seen in endemic MPV mouse colonies. There are several studies that indicate that immune tolerance to specific antigens can be induced if the antigen is present during thymic development, including post natal thymic development. In an early experiment demonstrating this phenomenon, CBA mice were inoculated in utero with adult tissue cells from an A strain mouse. CBA feti were exposed by a surgical laparotomy and given an intra-embryonic injection of adult tissues from an A strain mouse. The offspring of the CBA mice were subsequently capable of accepting skin grafts from A strain mice, whereas, those that were not exposed in utero rejected the skin grafts. The cause for the acceptance of the skin graft from a different mouse strain resulted from immune tolerance induced by early fetal exposure to the antigen. This technique is still used experimentally to overcome the immune response. For example, adeno-associated viral vector used in gene therapy often result in an immune response to the vector and the transgene. To overcome this, immune tolerance was induced in mice via the administration of the protein in utero and at day 2 post natal. In utero exposure was done by surgically exposing the feti and injecting the antigen intramuscularly. Neonatal exposure was generated via an intramuscular injection at 2 days post natal. In both sets of experiments and routes, challenge with antigen resulted in immune tolerance as demonstrated by a failure to induce an antibody response to the protein. These provide examples of inducing immune tolerance from in utero exposure by very invasive means. It is unlikely that this invasive type of exposure would occur in conventional mouse colonies. Because the primary route of infection of MPV is fecal-oral, the oral route of inoculation was chosen to best mimic exposure in a traditional mouse colony. These mice failed to develop immune tolerance with post natal inoculation. Similarly, tolerance was not induced to a viral vector when fetal mice were exposed to the antigen at day 13 to 15 gestation by intramuscular injection. It is certainly possible that the route of inoculation or the concentration of the inoculum was insufficient to induce tolerance; however, the route is typical of what may occur in a conventional mouse facility at a dose that likely exceeds what may be found by the fecal-oral route of exposure. Although every parameter was not investigated, this study does not support that immune tolerance could be an explanation for perceived hidden infections in mouse colonies. The most common method to eliminate pathogens from mouse colonies is rederivation or embryo transfer, and previous studies have demonstrated successful rederivation of MPV-infected mice by embryo transfer. Since parvoviruses have a predilection for rapidly dividing cells, such as the growing embryo or fetus, one potential complication with this procedure is vertical transmission of pathogens which could result in embryo infection, and consequently immune tolerance. When pregnant mice were challenged with MPV orally in this study, the offspring seroconverted to subsequent MPV challenge after maternal antibodies were found not to be present. Thus the common oral route of transmission in a conventional mouse facility does not result in in utero exposure sufficient to result in immune tolerance of the progeny, suggesting vertical transmission is not an important mode of transmission. Therefore rederivation or embryo transfer procedures would be effective in eliminating mouse parvovirus from mouse colonies. There have been several reports using cross fostering of neonatal mice to eliminate murine pathogens. This includes mouse hepatitis virus, Theiler’s murine encephalomyelitis virus, mouse rotavirus and Helicobacter hepaticus. These studies removed neonates from their dams 24–48 hours after birth and placed them with pathogen free foster dams in a new cage. One study performed an iodine dip of the neonates prior to placing with the foster dam to treat any pathogens on the neonates prior to transfer. When neonates from MPV infected dams were removed prior to developing a milk spot and placed on foster dams in new cages, the mice remained MPV negative by fecal PCR and serology up to 7 weeks of age. Furthermore, their foster dams did not develop antibody titers to MPV suggesting that the pups were not infected while in their original home cage prior to transferring them to the foster dam. While a more extensive study would need to be performed to confirm these observations, it appears that cross fostering may be useful to eliminate MPV from mouse colonies since our study failed to demonstrate in utero transmission. One final observation from this study was the reduced litter size in the MPV infected dams. While there was not a statistically significant difference, the MPV infected dams had reduced conception rates and reduced litter sizes. This may be due to the effects of MPV infection on the developing embryos, similar to the effects human parvovirus may have during pregnancy in people. These results demonstrate that pregnant mice infected with MPV via the oral route do not induce immune tolerance in the offspring. This information bodes well for those using rederivation, embryo transfer or cross fostering as a means to eliminate MPV from their mouse colonies as vertical transmission does not appear to be significant. # Supporting Information We would like to thank Elisa French for technical assistance and animal care. This work was funded by a grant from the American College of Laboratory Animal Medicine Foundation and Merck & Co., Inc. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors have declared no financial or non-financial competing interest exist. MPV mouse parvovirus MVM minute virus of mice [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LVK KSH. Performed the experiments: LVK CA KSH. Analyzed the data: LVK CA KSH. Contributed reagents/materials/analysis tools: LVK KSH. Wrote the paper: LVK CA KSH.
# Introduction The fact that many insect species are difficult to be discriminated at the morphological level, as well as the huge number of cryptic species, makes the global species count uncertain. The adoption of DNA-based molecular markers represents a satisfactory alternative. Since the proposal of DNA barcoding in 2003, subunit I (658 bp) of the mitochondrial cytochrome C oxidase (COX) gene (namely COI) became the most universal marker for species identification in the animal kingdom. The recently developed Barcode Index Number (BIN) system (Ratnasingham and Hebert, 2013) can act as a powerful alternative to morphological species that easily distinguishes the occurrence of diversity or possible speciation. Technically speaking, this system complements the molecular-based approaches of species identification to strengthen and support the evolutionary analysis in insects. Documentation of insect diversity in the Sahara-Arabian region including Saudi Arabia has recently taken place as the biodiversity data available for this region is insufficient, compared with those in Canada or the US. Documentation based on the criteria of the International Union for Conservation of Nature allows recognizing invasive alien species or list threatened species. Malaise traps were successfully used in scoring richness of insect species and biodiversity surveillance in regions that are difficult-to-access and/or in the absence of formal taxonomic assignments. COX enzyme functionally participates in the electron transport chain by reducing oxygen and pumping protons across the inner mitochondrial membrane. The encoded protein of COI comprising 219 amino acids (AAs) consists of six polypeptide chains and a few metallic ligands. The latter includes two iron atoms bound in two heme groups, three coppers, one zinc and one magnesium. Changes in the AA sequence within the COI region likely reduce cellular energy metabolism, especially when changes occur close to the active sites of the enzyme. Purifying selection predominantly occurs for animal mitochondrial protein-coding genes, thus, results in the scarce of AA substitutions, especially in the COX genes. However, evidence of nonsynonymous AA substitution supports the notion of positive selection generally in animal and specifically in class Insecta. Due to the functional constraints criterion, changes should be biased to the nonfunctioning regions of the gene supporting the claim that evolution is not neutral. Positive selection for changes in mitochondrial proteins during evolution could be caused by lifestyle changes, and this has been observed in snakes changing their metabolic rate to complement lifestyle. In this work, we introduce a model study on the COI subunit of specimens of six insect orders collected from Saudi Arabia to generate DNA barcodes and detect variation among species and orders at the AA level. Based on the latter, changes in COI protein structure at the three-dimensional model are projected. Through this model study, we gained new insights into the possible mechanisms by which COI protein evolves in insects living in Saudi Arabia. # Materials and methods For sample collection, a Malaise trap (<u>Hutcheson and Jones, 1999; Hill et al., 2005</u>) was installed at Hada Al-Sham station, King Abdulaziz University (KAU) located in the western region (near Makkah) of Saudi Arabia (21.795<sup>o</sup>N, 39.711<sup>o</sup>E). Samples were collected on a weekly basis for four weeks (May 1–28, 2017) into 95% ethanol and stored at -20°C for further analysis. Specimens were morphologically classified down to the species level, then total DNAs were recovered from individual tissue samples according to Evans and Paulay. PCR was performed to recover the COI gene fragment (650 bp) as described by Ashfaq et al. and amplicons were shipped to Macrogen (South Korea) for Sanger sequencing. Recovered sequences were checked for quality and those meeting the standard criteria were further assigned to subsequent analysis. Good quality sequences were trimmed and algorithmically aligned with Clustal Omega with strict parameters and the resulting alignments were manually refined before translation. Deduced AA sequences were recovered consulting the invertebrate mitochondrial code. Samples with nucleotide deletions were excluded upon DNA multi-sequence alignment, and only in-frame AA sequences were retained. A consensus sequence of Odonata (accession no. JN294479) was used as a reference for comparison. This order is the most ancient in the hexapoda insect phenogram. Binary data matrices were entered into TFPGA (version 1.3) and analyzed using qualitative routine to generate a similarity coefficient. Dissimilarity coefficients were used to construct dendrograms using unweighted pair group method with arithmetic average (UPGMA) and sequential hierarchical and nested clustering (Neighbor-Joining or NJ) routine using NTSYSpc (version 2.10, Exeter software). Phylogeny tree generated by Bybee et al., was taken as a reference ancestral order for testing insect lineage and evolution. The AA sequence of cattle (Bos taurus) was used as a reference for detecting changes of COI protein of selected species at the 3D structure level. We recruited the bovine protein X-ray structure (Protein Data Bank ID 1OCC) as a homology model of the COI barcode region to build 3D structures of selected insect species using the I-TASSER Suite. Distance measurements between AAs, on one side, and the two heme ligands referenced as 515 and 516, on the other side, were estimated using UCSF Chimera 1.13.1. Electrostatic potential of the COI protein was estimated using DeepView—Swiss-PdbViewer (v4.1) (<http://www.expasy.org/spdbv>) Nucleotide substitution patterns, e.g., synonymous (S) and nonsynonymous (N), were generated using the Estimate Substitution Matrix feature in MEGA v. 6 and estimates of the numbers of S and N substitutions per site were made using the joint Maximum Likelihood reconstructions of ancestral states of codon substitution and Felsenstein model of nucleotide substitution. Changes in AA sequences were scored referring to the consensus sequence of Odonata. # Results and discussion COI barcode was analyzed at the DNA and protein levels for insects living in Saudi Arabia in order to gain new insights as to how this gene and its encoded protein evolve among different insect orders. A number of 560 samples were collected by the Malaise trap within a period of four weeks. Photographs were taken for all specimens, while the number was eventually narrowed to one photograph per species as shown in. DNAs were purified from these samples and the recovered mitochondrial COI gene fragments were sequenced. We projected to include only sequences with ≥ 634 nucleotides encoding all loops and helices of the COI protein. The number of deduced AAs is 211 starting at codon position 12 referring to the numbering made by Pentinsaari et al.. Based on the rigid quality control criterion, the number of samples was narrowed to 175 for further analysis. The consensus DNA sequence of Odonata was included for comparison as this order has an ancient phylogenetic position being the representative of the first ancient winged insects in the hexapoda phylogeny. This order has extensively been used as a reference in contemporary evolutionary genomic studies. BLAST analysis indicated that a number of 30 species representing six insect orders were collected during the course of this study. Two of which belong to Hemimetabolous, e.g., Blattodea (one species) and Hemiptera (five species), while four belong to Holometabolous, e.g., Hymenoptera (six species), Coleoptera (six species), Lepidoptera (five species) and Diptera (seven species). One of the Hymenopteran species was not precisely identified at both the morphological and molecular levels, while showed the closest relationship (98%) with Apoidea sp. Multiple DNA sequence alignment indicated that a number of 265 nucleotides out of 634 are common (conserved) in the recovered COI barcode, while the rest varied among orders and species as shown in. A phylogenetic tree constructed from the sequences indicated the close relationship between the species of Odonata and Blattodea as both orders are hemimetabolous. Recent research indicated that Blattodea is the closest among the six orders of the present study to Odonata. Interestingly, a close relationship was shown between species of Hemiptera and Hymenoptera although the first is hemimetabolous, while the second is holometabolous. Species of Coleoptera were shown in close relationship with those of Diptera. As it is the case with animal mitochondrial DNA, the DNA COI barcode sequences analyzed for the six orders is AT-biased (e.g., AT content = \~68%) as shown in. The AT content was lowest (\~57%) for the nucleotides in the first position of all codons, while \~58% in the second position and as high as \~90% in the third position. The G nucleotide has proven to be the least degenerate among the four nucleotides in the third position. Interestingly, A nucleotide in the second position was unexpectedly low (\<15%), while C was high (\~25%). No G nucleotide was scored in the third position for four species namely Hymenoptera sp., Belenois aurota, Asyndetus sp. and Carpomya vesuviana. The first two species are Hymenoperan and Lepidopteran, respectively, while the other two are Dipteran. The lowest C nucleotide in the third position was scored for the two Hymenoptran species Tachytes crassus (\<2%) and Nomioides facilis (\<1%). Amino acid sequences of COI barcode were sorted out in this study based on their chemical properties into standard groups: nonpolar aliphatic (G, A, V, L, M and I), polar uncharged (S, T, C, P, N and Q), aromatic (F, Y and W), positively charged (K, R and H) and negatively charged (D and E). The barcode sequences of different species largely encode nonpolar AAs. There are differences in the AA composition of the barcode sequences among orders. Alanine (A) and valine (V) are lower in barcodes of Hymenopteran species as well as of the Hemipteran species Batracomorphus angustatus compared with those of the other orders. On the other hand, lysine (K) is exclusively encoded by barcodes of the latter species although there are very few. Four species of the latter two orders) Hymenoptera and Hemiptera) also exclusively encode cysteine (C). The two AAs K and C do not exist in the sequence of cattle COI barcode. We speculate that they appeared due to AA substitutions in several positions of the animal AA sequence of the barcode of Batracomorphus angustatus. Three other AAs seem to be encoded in very small amounts in few orders such as glutamine (Q), tyrosine (Y) and glutamic acid (E). Similar results were previously reached by Pentinsaari et al. when studying Coleopteran and Lepidopteran species. The scarcity of these AAs in barcodes can possibly be compensated by AAs of the same chemical groups at the same position in the polypeptide chain. For example, tyrosine (Y) exists in three positions, e.g., 38, 113 and 215. The position in the middle is conserved among the six orders, while substituted to phenylalanine (F) in the other two positions. The two AAs exist in the same chemical group (aromatic). Interestingly, the five rare AAs (C, K, Q, Y, and E) are encoded by twofold degenerate codons. Cysteine in the COI region exists in four species of Hymenoptera and one species in Hemiptera although in very small amounts, while completely absent in other species of the six orders. Changes in cysteine in the barcode sequence can severely affect the secondary structure as well as the stabilization of tertiary and quaternary structures of the COI protein. Multiple AA sequence alignment of the COI barcode of the 30 species is shown in. The encoded AA sequences of the DNA barcode fragment across the six orders cover 211 AAs starting from position 12 to 222 based on the numbering made by Pentinsaari et al.. This portion of the barcode includes the enzymatically active part of COX mediating electron transfer from Cu to heme. Amino acid sequence variation in each position was scored in which we referred to conserved AAs by asterisks while increasing variability by reducing the number from 9 to 1. Based on the criterion followed by Pentinsaari et al. (2016), the substitution of AAs with the same chemical property at any given position is not considered to significantly influence enzyme function. This also depicts the substitution of AAs from one chemical group to the other in a position far from the enzyme ligands. However, substitutions that change the AA chemical group occur close to the enzyme ligands, may likely influence enzyme function. As indicated earlier, the secondary structure of the barcode region includes six α-helices connected by five loops. The loops encompass 62 AAs of which all the eight-plus 11 AAs of loops 1–2 and 3–4, respectively, vary among the six orders. AAs in loop 2–3 vary in three out of the eight AAs, while 10 out of 23 and five of 12 for loops 3–4 and 5–6, respectively. The loop 3–4 is important for pointing towards the heme group at the active site of the COI protein. The higher number of conserved AAs in this study was expected due to the narrow genetic distances among the six insect orders compared with those shown by Pentinsaari et al. across the Metazoan barcodes. The number of the conserved AAs of the barcode sequence among the six insect order is 102 including the known 23 conserved AAs among Metazoan. Overall, AA variations occurring at positions in the helices 3, 4 and 5 likely have less influence on enzyme function compared with that in the helices 1 or 2. Besides, AA variations are more pronounced in loops than in helices. Functional constraints during the evolution of the insect COI barcode was checked in the predicted three-dimensional structure in terms of the distances between AAs and the two heme ligands referred to as hemes 515 and 516. The analysis indicated a number of 16 AAs mostly in close proximity with heme 515. These AAs exist in helices 2 (10) and 1 (4) and loop 3–4 (2). They are Thr (T), Ser (S), Ile (I), Arg (R), Thy (Y), Ile (I), Val (V), His (H), Ala (A), Met (M), Ile (I), Met (M), Val (V), Ile (I), Gly (G) and Trp (W). They exist at codon positions 15, 18, 21, 22, 38, 41, 42, 45, 46, 49, 50, 53, 54, 57, 109 and 110, respectively. Six of them at positions 15, 21, 38, 41, 42 and 57 vary among the six insect orders, while the rest are conserved. We selected Batrachedra amydraula as a model barcode of the insect conserved sequences for the 16 AAs in the heme 515 vicinity in which the predicted 3D structure was generated. We mainly focused on four AAs that directly share six bonds with hemes 515 and 516. These AAs are R (position 22), Y (position 38), H (position 45) and W (position 110). Two bonds exist between H at position 45 and heme 515 and two other bonds exist between W and the two heme ligands 515 and 516. Chemical bonds occur between NH1 molecule of R (position 22) and OMA molecule of heme 515, while OH molecule of Y (position 38) and O1A molecule of heme 515. Two chemical bonds occur between NE2 molecule of H (position 45) side chain and both NA and NC molecules of heme 515. W (position 110) shares N atom of its backbone with O1D molecule of heme 515 and NE1 molecule of its side chain with O2D molecule of heme 516. Distances between the four selected AAs and heme ligands of COI barcode are shown in. All the four AAs are \< 4 Å apart from the respective heme ligand. for the AA multiple sequence alignment indicated that the four AAs are conserved among insect orders except for tyrosine (position 38) that unexpectedly showed substitution to phenylalanine in two Hymenopteran species, e.g, Camponotus maculatus and Monomorium junodi. These two AAs differ in the occurrence of an OH molecule at the side chain of tyrosine, which is lacking in phenylalanine. The predicted 3D structure of Camponotus maculatus was investigated at this position in order to detect the possible severity accompanying this substitution. Tyrosine and phenylalanine share the same chemical group, however, the structure indicates no chemical bonding between the AA and heme 515 albeit the possible occurrence of Van der Waals interaction. This is the result of polarity change due to substitution that increases the chance of splitting heme from the protein. We speculate that this change in the 3D structure of COI and its consequence can impair the process of electron transfer from Cu to heme, thus, potentially affect energy metabolism in the mitochondria. depicts the alignment of the predicted three-dimensional structures of COI proteins of Batrachedra amydraula and Camponotus maculatus. The figure indicates the occurrence of a hydrogen bond between tyrosine of the first insect at position 38 and Heme 515 while lacking bond between phenylalanine of the second insect at the same position. The figure also underpins poor alignment of loop 1–2 or 4–5 of the two insect species, which reflects the high rate of varied AA in these two regions in line with the results of Pentinsaari et al. and those of the present study. The three-dimensional structure of the protein also indicates that the change from tyrosine to phenylalanine resulted in the loss of bonding with heme, thus likely affect energy metabolism. The phenylalanine appears to form a proper geometric complementarity; however, lacks electrostatic complementarity due to the absence of the hydroxyl group that contributes to the electrostatic interaction with the heme, resulting in an increase in hydrophobicity. This suggests that the protein retains its structure, but not all of its functions. Moreover, the absence of a single hydrogen bond, computing the electrostatic potential showed no clear difference when tyrosine is substituted with phenylalanine. Phylogenetic tree based on multiple AA sequence alignment of COI barcode indicated that Hymenoptera is the most genetically distant of all the six insect orders under study. The tree structure complements that of the tree generated based on multiple DNA sequence alignment in which Blattodea was shown to be closest to Odonata. The same criterion applies to the relatedness between the two orders Hemiptera and Hymenoptera and the two other orders Coleoptera and Diptera. We further analyzed the six varied AAs at positions 15, 21, 38, 41, 42 and 57, existing within the distance of \< 4Å from heme 515, in terms of transitions from one biochemical group to the other. At position 15 of the COI protein, threonine is changed to either serine with the same chemical group (polar uncharged) or methionine (two-codon AA) with different chemical groups (nonpolar) in all species of Hymenoptera and Hemiptera, respectively. At position 21, the substitution of isoleucine to valine in the same chemical group (nonpolar) has taken place in only Melanotus villosus species. At position 38, the substitution of tyrosine to phenylalanine in the same chemical group has taken place in the two Hymenopteran species Monomorium junodi and Camponotus maculatus. At position 41, the substitution of isoleucine (two-codon AA) to either leucine (six-codon AA) or methionine (two-codon AA) in the same chemical group (nonpolar) has taken place in Hymenopteran species except for Camponotus maculatus. An extra substitution at the same position has taken place to valine also in the same chemical group (nonpolar). At position 42, the substitution of valine to isoleucine in the same chemical group has taken place in Hymenopteran species, which was not fully detected at the species level. Interestingly, substitution at position 57 from isoleucine (nonpolar) to phenylalanine in a different chemical group (aromatic) has taken place in all Hymenopteran species. The AA transitions from one chemical group to the other might result in impaired energy metabolism. The study indicated a small distance between Hymenoptera and Hemiptera in the phylogenetic tree than between Coleoptera and Hemiptera. The data of AA substitution support our speculation that Hymenoptera is the most genetically distant of all the other insect orders under study. As Hymenopteran species are shown closely related to Hemipteran species and the fact that Hymenopteran species showed AA substitutions in five out of the six varied AAs in close proximity with the heme ligand indicate that Hymenoptera might be among the oldest insect orders. The estimated nucleotide substitution matrices in the COI barcode of the six orders generally showed a slightly higher probability of transversion compared with transition. The probability of transversion ranged between \~40% for the Coleopteran species Melanotus villosus and \~65% for the Hymenopteran species Tachytes crassus. Substitutions of A followed by T to any other nucleotide represents the lowest probability compared with the rest. Changes from C to T followed by G to A showed the highest transition frequency compared with the others. These two high transition frequencies justify the high TA ratio in insect orders and the bias against G or C during evolution. Both Coleoptera and Diptera showed the most notable bias against G. Maximum Likelihood computations of codon-by-codon synonymous (s) and nonsynonymous (n) substitutions in COI protein were conducted using the HyPhy software package in which Odonata was used as the reference ancestral order. The total number of nonsynonymous codons is 109, out of 211, of which 102 of them showed positive values at different rates across the six orders. The other seven codons showed negative dN-dS values, hence, listed under purifying selection. Hymenoperan followed by Hemipteran species are severely under positive selection as the values of dN-dS were positive in 90 and 61 out of the 102 codons, respectively. Twenty-two extremely high positive dN-dS values (\> 1.0) in 13 codons were scored for species only of these two orders; the most frequent are codons 98 (A, a four-codon AA) for Hymenopteran species and 126 (S, an eight- codon AA) for Hemipteran species. On the other hand, the only species (e.g., Balta vilis) of Blattodea showed positive dN-dS values for as low as 17 codons. This low number reflects the close relatedness of this order with Odonata in line with the results of the phylogenetic analysis. The other three orders, e.g., Lepidoptera, Coleoptera, and Diptera showed a total of 30, 38 and 34 positive dN-dS values across species, respectively. This indicates that most codons of COI barcode in these three orders seem to be mostly under purifying selection. There are 16 unique codons to a given species of either Hymenopteran (13 codons) or Hemipteran (3 codons) that are under positive selection. These Hymenopteran species are Camponotus maculatus (1 codon), Hymenoptera sp. (4 codons), Nomioides facilis (3 codons) and Tachytes crassus (4 codons), while Hemipteran species is Batracomorphus angustatus (4 codons). On the other hand, 10 codons frequently showed positive selection across most species of the six orders; five of them are located in the helix regions and five are located in the loop regions. Positions of these codons are 16, 23, 27, 48, 81, 118, 119, 123, 162 and 199. As for the dN-dS values in terms of domain structure, the results indicated that percentages of codons under positive selection are \~ 80, 31, 61, 52, 40 and 9% in helices 1 to 6, while 100, 38, 52, 100 and 33% in loops 1–2 to 5–6, respectively. These results indicate that codons in helices 2, 5 and 6 and in loops 2–3 and 5–6 are mostly conserved and are under strong purifying selection. The study indicates that COI barcodes can provide insights by which the encoded protein evolve and function in insect at different taxonomic scales. The COI or Folmer region is basically used for species identification based on variations at the DNA level. The region is sufficiently conserved within species, while variable among species. This region is located in the enzymatically active part of COI, thus involved in the respiratory chain by transferring electrons from Cu to heme. Accordingly, this region is under functional constraints and mutations in close proximity with heme ought to be lethal. There is less evolutionary pressure against variation in positions within the extra-membranous loop structures (e.g., loops 1–2 at AA positions 27–34, 3–4 at AA positions 102–124 and 4–5 at AA positions 156–166). This conclusion was previously reached by Panchenko et al.. Amino acids in the two highly conserved loops (e.g., loops 2–3 and 5–6) seem important in holding the two sides of the enzyme and avoid severe conformational changes. However, tryptophan within the variable loop 3–4 at position 99 rigidly evolves as it is the only AA of COI protein that bonds with the two hemes. Another AA in positions close to the active site is rigid membrane-embedded α-helices undergo functional constraints and restricted evolution (negative evolution). A similar pattern of variation was observed in the protein encoded by mitochondrial ribosomal RNA gene, where the loops freely evolve. The study generally indicated that almost half the positions of the AA of the barcode region are variable, however, they are mostly beyond the atomic interaction distance of the heme groups as previously reported. Six of these AAs exist within a distance of \>4Å with four of them changed to AAs of the same chemical groups. However, the three-dimensional structure of the protein indicates that the change from tyrosine to phenylalanine resulted in the loss of bonding with heme, thus likely affect energy metabolism. Estimates of dN-dS were used for detecting codons under either positive or purifying selection. dS is the number of synonymous substitutions per site (s/S) and dN is the number of nonsynonymous substitutions per site (n/N). Positive values indicate the overabundance of nonsynonymous substitutions in these positions. During evolution, positive (Darwinian) selection promotes the sweeps of new beneficial alleles, and negative (or purifying) selection impedes the spread of harmful alleles. The results indicated that 102 codons are under positive selection and tend to evolve more rapidly. Phenotypic plasticity or polyphenism is another potential player of evolutionary changes and shaping of ecosystems that allows species to phenotypically adapt to different environmental conditions during evolution. This type of evolutionary force is not genome structure-based, rather, it relies on the differential gene expression (gene expression-biased) and is particularly prominent in insects. Gene expression bias is predominant in molecular evolution when selection becomes less effective at removing deleterious alleles. Studying polyphenism provides insight into the molecular basis of phenotypic differentiation during evolution. This approach is ideal especially for cases where low genetic differences exist among species. Hymenopteran species have the highest AA variation in line with the presumed age of the order that is likely among the oldest, thus had a long time to molecularly evolve by changing AA sequence. Previous reports speculated that differences in patterns of variation, substitution, and selection among insect orders are due to the number of metabolism-related factors that are high in actively flying insect species (like Hymenoptera) than in non-flying species. Selection is biased towards the higher active metabolic rate and the increase of the resting metabolic rate. Pentinsaari et al. indicated a higher relative rate of CT/GA transitions in the insect is the result of higher oxidative damage to DNA. This high variation is the result of a weak purifying selection in Hymenopteran species. We strongly recommend making wet laboratory analysis to support whether the changes in AAs consequently affect metabolism in the target species or not. The overall results argue the possible evolutionary position of Hymenopteran species among those of other insect orders living in Saudi Arabia, especially Hemiptera and Coleoptera. # Supporting information [^1]: The authors have declared that no competing interests exist.
# Introduction According to John Gray's best-seller, Men Are from Mars, Women Are from Venus, males and females not only communicate differently but also think, feel, perceive, respond, love, and appreciate differently. Such differences can be reflected in the domain of emotion regulation. Emotion regulation refers to a person's ability to regulate one's own and others' emotional states and is regarded as a crucial component of emotional intelligence. Numerous behavioral studies have suggested that males score higher than females with regard to self- report measurement of emotion regulation ability and other similar constructs. For example, using the Trait-Meta Mood Scale (TMMS), Amelang and Steinmayr (2006) and Extremera, Durán and Rey (2007) found that males had higher mood repair scores. Using the Wong and Law Emotional Intelligence Scale (WLEIS), Kong, Zhao and You (2012) found that males scored higher in emotion regulation ability. Using the Trait Emotional Intelligence Questionnaire (TEIQue), Mikolajczak, Luminet, Leroy and Roy (2007) found that males had higher self- control scores, and finally, using the Emotional Quotient Inventory (EQI), Bar- On, Brown, Kirkcaldy and Thomé (2000) found that males scored higher in stress tolerance and impulse control. However, some studies found inconsistent results in the sex differences in emotional intelligence. For example, using Brain Resource Inventory for Emotional Intelligence Factors (BRIEF), Craig et al. (2009) found that females scored higher on Empathy than males, whereas males scored higher on Self-concept, so females scored higher on the total scale than males. Using the Trait Emotional Intelligence Questionnaire, Mikolajczak et al. (2007) found that females scored higher on Emotionality, whereas males scored higher on Self-Control and Sociability, as a result, males scored higher on the total scale than females. The discrepancy may be due to the choice of measurement instrument, but these inconsistent findings are not accidental, which may reflect the gender differences in many emotional aspects. These sex differences may be explained by the “extreme male brain theory of autism” proposed by Baron-Cohen (2002). According to this theory, the masculine brain predominantly seeks to understand and construct systems (i.e., “systemize”), whereas the feminine brain is predominantly structured to feel empathy (i.e., “empathize”). Here we focused on emotional regulation ability and used magnetic resonance imaging (MRI) to investigate the sex-specific neural basis of this ability. Previous functional neuroimaging studies have mainly focused on emotion regulation strategies (e.g., cognitive reappraisal); such studies have demonstrated that both subcortical regions and cortical regions are involved in emotion regulation strategies. Subcortical regions include the amygdala, hippocampus and cerebellum. Cortical regions include the prefrontal cortex (dorsolateral prefrontal cortex (DLPFC), ventromedial prefrontal cortex (VMPFC), orbitofrontal cortex (OFC), and the anterior cingulate cortex (ACC)) and the insular cortex. However, emotion regulation ability is differentiated from emotion regulation strategies because emotion regulation ability reflects effective choices and flexible application of different strategies for managing emotionally charged situations. Recently, several structural neuroimaging studies have explored the neuroanatomical basis of emotion regulation ability, but the results are less clear. For instance, Killgore et al. (2012) found that emotion regulation ability correlated positively with regional gray matter volume (rGMV) in the bilateral VMPFC, while Takeuchi et al. (2011) found that emotion regulation ability correlated with regional gray matter density (rGMD) in the right anterior insula, the right cerebellum, the precuneus, and the medial prefrontal cortex. Koven, Roth, Garlinghouse, Flashman and Saykin (2011) found that better mood repair was related to larger rGMV in three clusters in frontal and inferior parietal areas. Despite the limited statistical power caused by the relatively small sample size used in these studies, we speculate that sex differences may be another significant factor in these inconsistent findings. To our knowledge, there are three functional neuroimaging studies that have explored sex-related differences on cognitive reappraisal strategies (for a review see Ref.). Using an effortful reappraisal task, in which participants consciously maintained or down-regulated their negative emotions, two of the studies reported stronger prefrontal activity in males, while the third study found stronger prefrontal activity in females. In addition, the third study reported that the amygdala was more activated in females, while the two other studies failed to find an effect of gender on amygdala activity,. One possible reason for this inconsistency is that the regulation instructions for participants varied across the studies, and thus different types and number of reappraisal strategies may have been used. Although the results are mixed, these findings indicate distinct roles for prefrontal and subcortical regions in emotion regulation between males and females. However, to the best of our knowledge, no study has directly explored the sex-linked neuroanatomical correlates of emotion regulation ability. To investigate this issue, here we extended the previous studies in three ways. First, we assessed each participant's general ability to regulate emotions in daily life using a well-established self-report questionnaire (WLEIS). Second, we used voxel-based morphometry (VBM) to examine structural differences underlying emotion regulation ability in a large sample size of males and females (N = 299; 159 females). Functional and structural MRI studies are believed to complement each other, but structural MRI studies are particularly suitable for describing the neural correlates of emotion regulation ability because the short reappraisal tasks in functional MRI studies may hardly tap the cognitive-affective processes involving the whole range of emotion regulation ability. Finally, a large sample size of young adults was used in this study, which will provide a higher statistical power to test and identify the sex- specific neural correlates of emotion regulation ability. Given that males have better emotion regulation skills than females, we hypothesized that males and females would show differences with regard to the brain regions involved in emotion regulation ability. Specifically, we hypothesized that males' emotion regulation ability would be more strongly associated with rGMV in cortical regions, particularly in the prefrontal cortex, which is associated with cognitive processes. Similarly, females would mainly have a stronger association between emotion regulation ability and rGMV in regions (e.g., the amygdala) that are implicated in emotional processes. # Methods ## Participants College students (N = 299; 159 females; mean age = 21.55 years, SD = 1.01) from Beijing Normal University, Beijing, China, participated in this study. Both behavioral and MRI protocols were approved by the Institutional Review Board of Beijing Normal University. Written informed consent was obtained from all participants prior to the experiment. Two participants were excluded due to missing items or erroneous reports in the questionnaire. Another five participants were removed from further analyses due to extraordinary scanner artifacts or abnormal brain structures (e.g., unusually large ventricles). Thus, 292 participants contributed to our study findings (133 males, mean age = 21.57 years, SD = 1.00; 159 females. mean age = 21.54 years, SD = 1.02). ## Psychological measurement Participants' emotion regulation ability was assessed using the 4-item Regulation of Emotion (ROE) scale of the WLEIS. The ROE scale measures individuals' ability to regulate their emotions and their ability to quickly recover from psychological stress; sample items include, “I am quite capable of controlling my own emotions” and “I can always calm down quickly when I am very angry.” Participants were then instructed to indicate the extent to which they agree or disagree with each statement using a 7-point Likert-type scale. Higher scores in the ROE scale indicate better emotion regulation ability. The scale has been demonstrated to have high internal consistency, convergent/discriminant validity with related constructs of loneliness, positive affect, negative affect, depression, and empathy, and good concurrent validity with other emotion regulation measures, including the Optimism/Mood Regulation subscale of the Schutte' Emotional Intelligence Scale, the Repair of Emotion subscale of Trait Meta-Mood Scale, and the Emotional Control Questionnaire. ## MRI acquisition Participants were scanned using a Siemens 3T scanner (MAGENTOM Trio, a Tim system) with a 12-channel phased-array head coil at BNU Imaging Center for Brain Research, Beijing, China. MRI structural images were acquired using a 3D magnetization prepared rapid gradient echo (MP-RAGE) T1-weighted sequence (TR/TE/TI = 2530/3.39/1100 ms, flip angle = 7 degrees, FOV = 256×256 mm). One hundred and twenty-eight contiguous sagittal slices were acquired with 1×1-mm in-plane resolution and 1.33-mm slab thickness for whole brain coverage. ## Image Processing for VBM VBM was employed to characterize the neuroanatomical differences in gray matter volume (GMV) and the neuroanatomical correlates of behavioral performance across participants. In this study, VBM was performed using SPM8 (Statistical Parametric Mapping, Wellcome Department of Imaging Neuroscience, London, UK), with an optimized VBM protocol on T1-weighted structural MRI images. First, image quality was assessed by manual visual inspection. Five participants whose images had excessive scanner artifacts or showed gross anatomical abnormalities were excluded. Second, the origin of the brain was manually set to the anterior commissure for each participant. Third, images were segmented into four distinct tissue classes: gray matter, white matter, cerebrospinal fluid, and everything else (e.g., skull and scalp) using a unified segmentation approach. Forth, the MNI152 template was used to spatially normalize the gray matter images for each participant using the Diffeomorphic Anatomical Registration through Exponential Lie algebra (DARTEL) registration method. DARTEL registration involves repetitively computing the study-specific template based on the average tissue probability maps of all participants and then warping all participants' tissue maps into a generated template to improve the alignment. Fifth, gray matter voxel values were modulated by multiplying the Jacobian determinants derived from the normalization procedure to preserve the volume of tissue from each structure after warping. The modulated gray matter images were then smoothed using an 8-mm full width at half maximum (FWHM) isotropic Gaussian kernel. Finally, to exclude noisy voxels, the modulated images were masked using absolute masking with a threshold of 0.2. The masked-modulated gray matter images were used for further statistical analyses. ## Statistical Analysis of VBM Sex difference in the correlation between emotion regulation ability and rGMV was tested using the condition by covariate interaction analysis. The interaction analysis treated sex as a condition, the score of emotion regulation ability as a covariate of interest, and the total GMV and age as covariates of no interest. Statistical analysis was performed using a general linear model (GLM). Multiple comparison correction was performed by setting the voxel-wise intensity threshold at p\<0.05 and a cluster-level threshold determined by Monte Carlo simulations (10,000 iterations) conducted in the AlphaSim program within AFNI. Accordingly, significant effects were reported when the volume of a cluster was greater than the Monte Carlo simulation determined minimum cluster size on whole brain GMV (i.e., 943 voxels), above which the probability of type I error was below 0.01. To test the specificity of correlation, we also detected the neuroanatomical correlates of individual differences in emotion regulation ability in the whole sample. Statistical analysis treated total GMV, age and sex as confounding covariates and the score of emotion regulation ability as a covariate of interest. The threshold for statistical significance was also set at MC-cluster- corrected p\<0.05 and significant effects were reported when the volume of a cluster was greater than 943 voxels. # Results The summed score for the ROE scale was used as an index of emotion regulation ability, whereby higher scores indicated a better ability. Demographic characteristics for the participants are presented in. As indicated in, measures of emotion regulation ability, use of emotion ability, self-emotion appraisal ability, and others-emotion appraisal ability had good internal consistency. Moreover, emotion regulation ability was moderately and positively related to use of emotion ability (r = 0.36, p\<0.001), self-emotion appraisal ability (r = 0.44, p\<0.001), and others-emotion appraisal ability (r = 0.28, p\<0.001), thus indicating that the ROE scale has good discriminant validity. The independent sample t-test analyses revealed no significant sex differences in age (t = 0.29; p = 0.77) and use of emotion ability (t = 1.15; p = 0.25), self-emotion appraisal ability (t = 0.83; p = 0.41), and others-emotion appraisal ability (t = −1.28; p = 0.20). Importantly, we found a significant difference in emotion regulation ability between male and female groups (t(290) = 2.92, p = 0.004, Cohen's d = 0.34), which is consistent with previous findings that males have higher levels of emotion regulation ability than females. Next, we examined whether the sex-specific differences in emotion regulation, as observed in the behavioral measurements, had distinct neural substrates. To do this, we conducted a condition by covariate interaction analysis with sex as a condition and the score of emotion regulation ability as a covariate of interest. We revealed a significant interaction of sex by emotion regulation ability in two anatomical clusters: one in the right DLPFC (MNI coordinate: 20, 28, 60; Cluster-corrected p\<0.01) and the other that extends from the left brainstem to the left hippocampus, the left amygdala and the insular cortex (MNI coordinate: 0, −14, −10; Cluster-corrected p\<0.01). Brain structures involved in the second cluster are known as central parts of the limbic system. Specifically, males showed a stronger positive relation between emotion regulation ability and rGMV in the right DLPFC (r = 0.30, p\<0.001), relative to females (r = −0.08, p\>0.05). In contrast, females showed a stronger positive relation between emotion regulation ability and rGMV in the other anatomical cluster (r = 0.20, p\<0.05), relative to females (r = −0.11, p\>0.05). The anatomical location of these clusters identified by the VBM analysis in this study was close to regions identified by fMRI using emotion regulation tasks. In addition, we tested the gender differences of rGMV in these two clusters. We found that males had larger rGMV in these two clusters than females (ps\<0.001). To test the specificity of these correlations, we also detected the neuroanatomical correlates of individual differences in emotion regulation ability in the whole sample. A significant positive correlation between emotion regulation ability and rGMV was found in a cluster that included the left precuneus (MNI coordinate: −2, −72, 32; Cluster-corrected p\<0.01). Taken together, these results suggest that the left hippocampus and the left amygdala, as well as the right DLPFC, have a sex-specific correlation with emotion regulation ability. # Discussion The aim of the present study was to investigate the sex-linked neuroanatomical basis of emotion regulation ability in a large sample of young healthy adults. As expected, males reported higher levels of emotion regulation ability relative to females, which is consistent with previous behavioral findings that males have a greater ability to regulate emotions than females. More importantly, VBM analysis revealed males relying more on the right DLPFC and females relying more on limbic regions including the left hippocampus, the left amygdala and insular cortex. Thus, the present study provides the first empirical evidence for sex- related neuroanatomical basis of emotion regulation ability. Males demonstrated a stronger positive relation between emotion regulation ability and rGMV in the right DLPFC, which is in line with previous VBM and lesion studies that report involvement of the DLPFC in emotion regulation ability using a self-report emotion regulation ability assessment. Moreover, numerous fMRI studies have consistently shown increased neural activity in the DLPFC when participants are instructed to deploy a variety of emotion regulation strategies, including cognitive reappraisal and expressive suppression,. The DLPFC is known for its critical role in cognitive control, working memory, and response selection. Taken together, recruitment of the DLPFC may facilitate emotion regulation through a range of effortful cognitive processes, such as selecting, shifting, and maintaining regulation strategies. For example, the DLPFC may be used to direct attention to reappraisal-relevant stimulus features and hold in mind reappraisal goals as well as the content of one's reappraisal. Our finding that the DLPFC has a male-specific correlation with emotion regulation ability seems to support behavioral observations that report males use more rational and detachment coping strategies in a stressful environment than females. In contrast, we found that females' emotion regulation ability was associated with rGMV in the anatomical cluster that extends from the left brainstem to the left hippocampus, the left amygdala and the insular cortex, which are implicated in emotional processes. Numerous fMRI studies have demonstrated that the hippocampus, the amygdala and the insular cortex are involved in successful emotion regulation,. In particular, the amygdala plays a critical role in detecting, as well as attending to and encoding affectively arousing and threatening stimuli into memory, and its hyperactivation is associated with a variety of affective disorders. The hippocampus is involved in establishing declarative or episodic memory for autobiographic events. Therefore, recruitment of the hippocampus and amygdala may help females to accurately perceive emotional events and detect cues that signal potential threats, thus facilitating their regulation of emotion. The insular cortex is a fundamental multimodal sensory integration region, and has highly reciprocal connections with cortical (e.g., the anterior cingulate and prefrontal cortex) and subcortical regions (e.g., amygdala and basal ganglia). It is involved in the experience of bodily self-awareness, emotional processing, sexual memory, and regulating autonomic functions, and possibly plays a pathophysiological function in anxiety disorders and emotion dysregulation. The insular cortex, as well as the hippocampus and the left amygdala are central components of the limbic system that has been extensively implicated in many emotional processes such as emotional memory and emotion recognition. This may be consistent with the findings of previous behavioral studies that report females use emotion-focused coping strategies more frequently when under stress. These results may contribute to our understanding of sex differences in disorders and maladaptive behaviors that are associated with emotion regulation ability. For example, it has been reported that hippocampal and amygdala volume are smaller in depressed females than that in normal participants, and, more specifically, than in depressed males,. This is in accordance with our findings that rGMV in the hippocampus and the amygdala is more strongly associated with emotion regulation ability in females than in males, which in part supports the finding of a higher prevalence of affective disorders (e.g., depression) in females. On the other hand, maladaptive behaviors (e.g., aggressive behavior and alcohol-related problems) are more prevalent in males. The finding that rGMV in the DLPFC is more strongly associated with emotion regulation ability in males than in females, similarly suggests that maladaptive behaviors might result from difficulties in the cognitive control of emotions. In conclusion, our study provides the first evidence for the sex-specific neuroanatomical basis of emotion regulation ability. Specifically, we found that females' emotion regulation ability is associated with rGMV in emotion-focused brain regions, whereas males' emotion regulation ability tends to be associated with rGMV in brain regions implicated in cognitive processes. However, several limitations need to be addressed for future research. First, we assessed each participant's emotion regulation ability using a self-report questionnaire. Although this questionnaire predicts real-world outcomes, such as well-being and depressive symptomatology, it would perhaps be useful to employ objective measures of emotion regulation ability when examining sex-specific neuroanatomy involved in the process in order to examine convergence across diverse measures. Second, only healthy participants were recruited, and therefore, it is unknown whether our results relate to clinical samples. Given the relationship between emotion dysregulation and affective disorders, studying emotion regulation ability in a clinical population is necessary to improve our understanding of emotion regulation deficits and to provide insight into treatment implications. Third, this study is a cross-sectional design, which does not draw causal inferences. Future investigations should ideally utilize longitudinal designs to examine changes in emotion regulation ability over the course of development. [^1]: The authors have declared that no competing interests exist. [^2]: Analyzed the data: FK ZZ XW LH. Wrote the paper: FK ZZ JL. Designed research: FK JL. Performed research: FK ZZ JL. Contributed unpublished reagents/analytic tools: ZZ LH XW YS.
# Introduction There are more than 40 species in the genus Borrelia, currently divided into the Lyme borreliosis (LB) and Relapsing fever (RF) groups with human pathogens confirmed in both of these. Whilst the major focus has been on Lyme disease borreliae the health burden of Relapsing fever (RF) species is now being recognized. B. miyamotoi classified in the RF group was first identified in Japan in 1995 and subsequently found in many countries including the USA, Europe and England. A prior study of Borrelia infections in ticks in Ireland identified three species, B burgdorferi, B garinii, B valaisiana (VS116) plus un-typeable Borrelia species. In order to ensure detection of a wide range of Borrelia species that might be present in Ireland experiments were carried out to identify PCR primers for species that have been reported in Europe. In the current study, we used a pair of genus-specific PCR primers to amplify a highly conserved segment with single-nucleotide polymorphisms of the borrelial 16S rRNA gene shared by all known pathogenic borrelial strains to survey the borrelial infections among the I. ricinus ticks collected in Ireland. Since this PCR amplicon is 357/358 bp long, the PCR products can be used as the template for direct Sanger sequencing for amplicon validation and for speciation. This metagenomic 16S rRNA gene sequencing assay is most suitable for molecular diagnosis of borrelial infections in human-biting ticks and in clinical specimens in Europe because of the great diversity of causative agents in European Lyme borreliosis which needs a broad-spectrum tool to detect the target DNA from various borrelial strains and to prepare the template for Sanger sequencing to ensure diagnostic accuracy. In the present study, we developed a protocol for using a single pair of genus- specific PCR primers to amplify a highly conserved segment with hypervariable regions of the borrelial 16S rRNA gene for detection of all species of Borrelia infecting the I. ricinus ticks collected in Ireland and to use the positive crude nested PCR products as the templates for direct Sanger sequencing to determine the species of the Borrelia detected. # Materials and methods ## Ixodes tick collection and extraction Unfed, questing I. ricinus nymphs were collected by “flagging”, which involves brushing the vegetation with a white towel from which the ticks can then be removed. Nymphs were chosen for analysis since they occur in greater numbers than adult ticks, and pose the greatest risk to humans. In late May and early June 2018, samples were taken at six localities within Ireland, designed to provide a representative view of tick borrelial infection across the country. The following areas were sampled (county in parenthesis): Killarney (Kerry), Kilmacthomas (Waterford), Clifden (Galway-West), Portumna (Galway-South), Glendalough (Wicklow), and Glenveagh (Donegal) and their locations are shown on the accompanying map of Ireland. No permission to access tick collection sites were required. All ticks were collected along the verges of public roads without requiring access to private land. There were no known endangered or protected species at any of the collection sites which all had public right of way access Each individual sample was made up of at least 5 sub-samples taken at different points within a locality to minimize any sampling bias. Distance between sub- sampling points was never less than 100 m. The collected ticks were disinfected in 70% ethanol, air-dried on filter paper and sent to Milford Molecular Diagnostics Laboratory in Milford, Connecticut, U.S.A. to be tested. The general procedure for extraction of the crude DNA from archived ticks and for Sanger sequencing detection of the borrelial 16S rRNA gene previously published was followed. Initially, 300 ticks were analyzed, 50 from each of the six sampled locations listed above. On the day of testing, each dried tick was placed in a 1.5 mL plastic tube and immersed in 300 μL of 0.7 mol/L NH<sub>4</sub>OH overnight at room temperature. On the following day, the test tubes were heated at 95°C to 98°C for 20 minutes with closed caps, followed by 10 minutes with open caps. After the test tubes were cooled to room temperature and the carcass of the tick discarded, 700 μL of 95% ethanol and 30 μL of 3 mol/L sodium acetate (Sigma) were added to each NH<sub>4</sub>OH digestate. The precipitated crude DNA was spun down in the pellet after centrifugation at \~ 16,000 x g for 5 minutes, washed in 1 mL of 70% ethanol, air dried, and re-dissolved in 100 μL of tris(hydroxymethyl)aminomethane hydrochloride–EDTA (TE) buffer, pH 7.4 (Sigma) by heating the DNA extract at 95°C to 98°C for 5 minutes. After a final centrifugation at \~ 16,000 x g for 5 minutes, the supernatant was used for molecular detection of borrelial 16S rRNA genes by nested PCR amplification. ## Molecular detection of borrelial 16S rRNA genes A 3μL aliquot of the crude DNA extract was used to initiate a primary PCR, followed by a same-nested PCR using a pair of M1/M2 borrelial genus-specific primers in a total 25 μL low-temperature PCR mixture for a 30-cycle amplification at primary PCR, followed by another 30-cycle amplification at the same-nested PCR. In carrying out the same-nested PCR, a single pair of M1 (`5'-ACGATGCACACTTGGTGTTAA-3'`) and M2 (`5' TCCGACTTATCACCGGCAGTC-3'`) primers were used for both primary PCR and nested PCR so that a small amount of the primary PCR products was re-amplified with the same pair of PCR primers in a new PCR mixture. An original target DNA segment in the PCR mixture might have been amplified for 60 cycles exponentially by the same pair of primers to increase the sensitivity of detection. The PCR amplicons of the 357-bp 16S rRNA gene segment of the B. burgdorferi sensu lato complex and the corresponding 358-bp 16S rRNA gene segment of the relapsing fever borreliae, both defined by the M1 and M2 PCR primer pair, were visualized by agarose gel electrophoresis of the nested PCR products. The same-nested PCR products were used as the template for Sanger reaction without purification. A positive Borrelia coriaceae DNA and a negative (no extract) reagent control were routinely included in each run of our experiments, as detailed in a previously published reference ## Sanger sequencing confirmation of Borrelia species For DNA sequencing, the positive nested PCR products were transferred by a micro-glass rod into a Sanger reaction tube containing 1 μL of 10 μmolar sequencing primer, 1 μL of the BigDye Terminator (v 1.1/Sequencing Standard Kit), 3.5 μL 5× buffer, and 14.5 μL water in a total volume of 20 μL for 20 enzymatic primer extension/termination reaction cycles according to the protocol supplied by the manufacturer (Applied Biosystems, Foster City, CA, USA). After a dye-terminator cleanup with a Centri-Sep column (Princeton Separations, Adelphia, NJ, USA), the reaction mixture was loaded in an automated ABI 3130 four-capillary Genetic Analyzer for sequence analysis. Sanger sequencing of positive 357/358 bp M1/M2 same-nested PCR products is capable of accurate identification of many species including, B. valaisiana, B. afzelii, B. mayonii, B. spielmanii, B. lusitaniae, B. recurrentis, B. miyamotoi, B. hermsii, B. lonestari, B. coriaceae and several other members in the relapsing fever group based on known species-specific single-nucleotide polymorphisms in the gene segment. The “genus-specific” M1/M2 PCR primer pair can amplify a “core genome” of all pathogenic borreliae for the purpose of detection. However, to design a pair of reliable PCR primers to amplify a segment of borrelial 16S rRNA gene with single-nucleotide polymorphism among various borrelial strains turned out to be challenging. It took several weeks of experimental work before we found 3 PCR primers to generate a 282-bp heminested PCR amplicon useful as the template for Sanger sequencing to distinguish B. garinii from B. burgdorferi and to discriminate among the various B. garinii strains. The sequences of these 3 heminested PCR primers are listed as follows. Primary PCR Forward Primer Bg1: `5’- GACGTTAATTTATGAATAAGC -3’` Primary PCR Reverse Primer Bg 6: `5’- TTAACACCAAGTGTGCATCGT– 3’` Heminested PCR Forward Primer Bg5: `5’- CGGGATTATTGGGCGTAAAGGGTGAG-3’` Heminested PCR Reverse Primer Bg 6: `5’- TTAACACCAAGTGTGCATCGT– 3’` Reference sequences were retrieved from the GenBank, the Bg5/Bg6 heminested PCR primer pair defines a 282 bp segment of the borrelial 16S rRNA gene with single- nucleotide polymorphisms. These were used to discriminate Borrelia burgdorferi strain B31 (ID# [CP019767](https://www.ncbi.nlm.nih.gov/nucleotide/CP019767.1?re port=genbank&log$=nuclalign&blast_rank=3&RID=NPJ52KA801R)), Borrelia garinii BgVir (ID# [CP003151)](https://www.ncbi.nlm.nih.gov/nucleotide/CP003151.1?report =genbank&log$=nuclalign&blast_rank=17&RID=NPF83DHW014), Borrelia garinii strain Bernie (ID# [D89900)](https://www.ncbi.nlm.nih.gov/nucleotide/D89900.1?re port=genbank&log$=nuclalign&blast_rank=3&RID=NPGTV1D5014), Borrelia garinii strain T25 (ID# [AB035388) and](https://www.ncbi.nlm.nih.gov/nucleotide/AB035388 .1?report=genbank&log$=nuclalign&blast_rank=1&RID=NGYV838S014) Borrelia garinii strain L20 (ID# [X85198).](https://www.ncbi.nlm.nih.gov/nucleotide/X851 98.1?report=genbank&log$=nuclalign&blast_rank=50&RID=N8ZTRD0H01R) However, after the first 300 ticks were tested, it was realized that the extracellular DNA of the Borrelia 16S rRNA gene in the crude DNA extract from the ticks was not stable on storage even at -20° C. By the time when the Bg5 and Bg6 heminested PCR primers were readied to be put into routine practice, the borrelial 16S rRNA gene DNA in 9 of the 12 samples initially found to be positive for B. burgdorferi sensu lato already degraded and were no longer amplifiable with any PCR primers. Therefore, an additional series of 50 ticks from each of the Portumna and Kilmacthomas samples were analyzed for the specific purpose of species confirmation and strain determination of the B. garinii isolates. # Results ## Multiple Borrelia species found in I. ricinus ticks in Ireland The same-nested PCR amplification of a 357/358 bp segment of borrelial 16S rRNA gene by the M1/M2 genus-specific PCR primers followed by Sanger sequencing of the nested PCR products provided metagenomic evidence of B. burgdorferi sensu lato (B.b.s.l.), B. valaisiana and B. miyamotoi infection in the ticks collected in Ireland. Samples of the 16S rRNA gene sequencing with the M2 primer in support of these molecular diagnoses are illustrated by the 3 selected electropherograms presented in. ## Strain diversity of B. garinii Sanger sequencing of the PCR amplicon with the reverse M1 primer proved that none of the B. burgdorferi sensu lato isolates were B. afzelii. A 177-base sequence of the 282-bp amplicon defined by the Bg5 and Bg6 heminested PCR primers distinguished the heterogeneous strains of B. garinii from B. burgdorferi and B. bavariensis due to the presence of single-nucleotide polymorphisms among strain BgVir {1}, strain 25 {2} and strain Bernie {3} of the B. garinii species and distinguished these strains from B. burgdorferi sensu stricto {4}. Alignment of these four 177-base reference sequences retrieved from the GenBank showing single-nucleotide polymorphisms is presented below with the ending 26-base Bg5 primer site underlined. {1}CCTTCGCCTCCGGTATTCTTTCTGATATCAACAGATTCCACCCTTACACCAGAAATTCTAACTTCCTCTATCAGACT CTAGAC {2}············································································· ······ {3}······································································ ············T {4}············································································· ······ {1}ATATAGTTTCCAACATAGTTCCACAGTTGAGCTGTGGTATTTTATGCATAGACTTATATATCCGCCTA<u>CTCACC CTTTACGCC</u> {2}································C····C······························<u>······ ·········</u> {3}····································································<u>······ ·········</u> {4}··················G·················································<u>······ ·········</u> {1}<u>CAATAATCCCG</u> B. garinii Strain BgVir Sequence ID: CP003151 (Range:447801–447977) {2}<u>···········</u> B. garinii Strain T25 Sequence ID: AB035388 (Range: 684–508) {3}<u>···········</u> B. garinii Strain Bernie Sequence ID: D89900 (Range:692–516) (4}<u>···········</u> B. burgdorferi strain B31 Sequence ID: CP019767 (Range: 444582–444758) The corresponding electropherograms of these four 177-base sequences (1–4), using the Bg6 sequencing primer, are illustrated in, arranged from top to bottom in numerical order of sequences 1–4 as Strain BgVir, Strain T25, Strain Bernie and B. burgdorferi strain B31. A reverse sequencing with Bg5 primer of the 282-bp PCR amplicon confirmed that the isolates identified as strains BgVir were not strain SL20 (Sequence ID: X85198) because there is no single-nucleotide polymorphism located near the Bg6 primer site of the amplicon which is unique for strain SL20. ## Tick infection rates varied with locations We conducted two surveys. Survey 1 consisted of 300 ticks collected in six locations showing 15 of the 300 ticks (5%) infected with B. burgdorferi sensu lato of which 3 were confirmed to be B. garinii (2 matching strain Bernie and 1 matching strain BgVir), 2 were B. valaisiana and 10 were not speciated. Survey 2 consisted of 100 ticks collected in two of the six locations previously surveyed, showing 5 of the 100 ticks (5%) infected by borrelial species, including 4 isolates of B. garinii (3 matching strain BgVir and 1 matching strain T25) and 1 isolate of B. miyamotoi. The Borrelia infection rates varied from 2% to 12% depending on the locations of the tick collection. It is noteworthy that all B. burgdorferi sensu lato isolates in Survey 2 were confirmed to be B. garinii as the Bg5/BG6 PCR followed by Sanger sequencing for speciation was carried out without delay after initial detection by M1/M2 PCR. The final species distributions in different locations were summarized in. ## Selection of PCR primers for 16S rRNA gene PCR and sequencing We used the “genus-specific” M1/M2 primer pair to generate a PCR amplicon for Sanger reaction to detect all species of the B.b.s.l. complex and the relapsing fever borreliae, in particular B. miyamotoi. And a heminested PCR system to amplify an adjacent 282 bp segment defined by the Bg5 and Bg6 primers. When the Bg5/Bg6 nested PCR products were used as the DNA sequencing template for confirmation of B. miyamotoi in tick samples, the electropherograms showed numerous ambiguous base calling peaks as a result of co-amplification of unwanted DNAs in the sample extract. When the sequencing electropherogram generated by the M1/M2 primer PCR products from the same tick extract showed no ambiguous base calling labels. ## Degradation of 16S rRNA gene DNA in samples We also observed that free borrelial 16S rRNA gene DNA in crude extracts from the ticks was unstable even in TE buffer stored at -20°C for an extended period of time. For example, when the NH<sub>4</sub>OH extracts of the 50 ticks from the Portumna area were processed for M1/M2 same-nested PCR screening on July 20, 2018, 3 samples were positive for borrelial 16S rRNA gene DNA, with a robust 357/358 bp band on lanes 11, 15 and 25 which eventually proved to be B. garinii strain BgVir, B. miyamotoi and B. garinii BgVir, respectively. When the NH<sub>4</sub>OH extracts of the 50 ticks from the Kilmacthomas area were processed for M1/M2 same-nested PCR screening on July 24, 2018, 2 samples were positive for borrelial 16S rRNA gene DNA, with a robust 357/358 bp band on lanes 372 and 382 which eventually proved to be B. garinii with sequence matching strain BgVir and B. garinii strain T25, respectively. However, when the same-nested PCR was repeated on these 5 NH<sub>4</sub>OH extracts after 7 days and 3 days storage in a -20°C freezer, respectively, the 16S rRNA gene DNA in sample 11 was no longer detectable and the intensities of the nested PCR bands using the same extracts of samples 15, 25, 372 and 382 as primary PCR templates for amplification under identical experimental conditions decreased markedly over a period of 3–7 days, as demonstrated on the agarose gel dated July 27, 2018. The image of the gel electrophoresis dated July 27, 2018 also showed that nested PCR is generally required for detection of borrelial infections by 16S rRNA gene analysis. Primary PCR products are usually invisible after gel electrophoresis. # Discussion This is the first time that B. miyamotoi has been detected in a tick collected in Ireland. Our study shows that I. ricinus ticks in Ireland are infected by a diversity of pathogenic borreliae. The number of species and strain diversity may prove to be greater if surveys with larger samples are carried out. For example, in our two surveys B. miyamotoi and B. garinii were demonstrated only during Survey 2, but not in Survey 1. Our results demonstrate that Borrelia-infected tick populations exist in the south-east of Ireland, an area hitherto not considered to be significantly tick-infested and hence not considered an area of risk to humans for contraction of Lyme borreliosis. In the prior work carried out by Kirstein et al tick infection rates were reported between 3.5% and 26.7% depending on the locations of tick collection using reverse line blot or PCR. In the current study, we used metagenomic 16S rRNA gene sequencing to survey the I. ricinus ticks, and found the overall rate of borrelial infection to be 5%, with a range from 2% to 12% depending on the locations of tick collection. In Glenveagh (Donegal) there were two different strains of B. garinii with sequences that matched BgVir, and Bernie). This compares to infected nymphs collected in Glenveagh National Park 20 years earlier which were reported to be positive for B. valasiana (labeled as strain VS116). We identified B. valaisiana in Gelendalough and Killarney, and B. miyamoti plus B. garinii in Portumna. For the purpose of patient care, it is probably not necessary to determine the species or strain of the Borrelia infection in order to initiate timely antibiotic treatment. But, for serological diagnosis a knowledge of the borrelial species and strains carried by the human-biting ticks collected in the endemic areas is crucial since polymorphism of ospC, and variation of the VlsE antigens among species and strains of B. burgdorferi sensu lato have been well documented in the literature. There is limited information regarding the sensitivity of commercial tests for different species of Borrelia. The manufacturers of Western Blot antibody tests specify the target species. In Europe these are typically B. afzelii, B. burgdorferi and B. garinii. A study of the most frequently used C6 synthetic peptide ELISA test for initial screening does appear to have variable sensitivity for some species but with no data available for many of the European species including B. valaisiana identified as prevalent in Ireland in this study. In summary, our study confirms that the genus-specific M1/M2 PCR primers can amplify a highly conserved segment of the borrelial 16S rRNA gene for Sanger sequencing-based molecular diagnosis of tick-borne borreliae. Three species of Borrelia were identified with B. garinii the most common (70% of those speciated), followed by B. valaisiana (20%), and for the first time in Ireland B. miyamotoi (10%). This study and future expanded surveillance of the Borrelia species and prevalence will contribute to optimized testing for patients and to help quantify the risks and potential burden of human Borrelia infections in Ireland. # Supporting information [^1]: SHL is Director of Milford Molecular Diagnostics Laboratory. There are no patents, products in development or marketed products associated with this research to declare. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
# Introduction The Mare medi terraneum (in Latin) describes the Mediterranean as a “sea in the middle of the land.” This basin is the largest (2,969,000 km<sup>2</sup>) and deepest (average 1,460 m, maximum 5,267 m) enclosed sea on Earth. Situated at the crossroads of Africa, Europe, and Asia, the Mediterranean coasts have witnessed the flourishing and decline of many civilizations. The region was an important route for merchants and travelers of ancient times, allowing for trade and cultural exchange, and today it is notable for contributions to global economy and trade. Its coasts support a high density of inhabitants, distributed in 21 modern states, and it is one of the top tourist destinations in the world, with 200 million tourists per year. The Mediterranean Sea connects through the Strait of Gibraltar to the Atlantic Ocean in the west and through the Dardanelles to the Sea of Marmara and the Black Sea in the northeast. In the southeast, the Suez Canal links the Mediterranean to the Red Sea and the Indian Ocean. In the Strait of Sicily, a shallow ridge at 400 m depth separates the island of Sicily from the coast of Tunisia and divides the sea into two main subregions: the western (area = 0.85 million km<sup>2</sup>) and the eastern (area = 1.65 million km<sup>2</sup>). General oceanographic conditions in the Mediterranean have been previously described in detail \[e.g., –\]. It is a concentration basin: evaporation is higher in its eastern half, causing the water level to decrease and salinity to increase from west to east. The resulting pressure gradient pushes relatively cool, low-salinity water from the Atlantic across the Mediterranean basin. This water warms up to the east, where it becomes saltier and then sinks in the Levantine Sea before circulating west and exiting through the Strait of Gibraltar. The climate in the region is characterized by hot, dry summers and cool, humid winters. The annual mean sea surface temperature shows a high seasonality and important gradients from west to east and north to south. The basin is generally oligotrophic, but regional features enrich coastal areas through changing wind conditions, temporal thermoclines, currents and river discharges, and municipal sewage. The basin is characterized by strong environmental gradients, in which the eastern end is more oligotrophic than the western. The biological production decreases from north to south and west to east and is inversely related to the increase in temperature and salinity. The Mediterranean has narrow continental shelves and a large area of open sea. Therefore, a large part of the Mediterranean basin can be classified as deep sea and includes some unusual features: (1) high homothermy from 300–500 m to the bottom, where temperatures vary from 12.8°C–13.5°C in the western basin to 13.5°C–15.5°C in the eastern, and (2) high salinity of 37.5–39.5 psu. Unlike in the Atlantic Ocean, where temperature decreases with depth, there are no thermal boundaries in the deep sea of the Mediterranean. Shelf waters represent 20% of the total Mediterranean waters, compared with the 7.6% of the world oceans, and therefore play a proportionally greater role here than in the world's oceans. Shelves in the south are mainly narrow and steep (e.g., Moroccan, Algerian, and Libyan coasts, with the exception of the Gulf of Gabés), while those in the north are wider (e.g., the north and central Adriatic Sea, the Aegean Sea, and the Gulf of Lions). These features influence the morphology and constrain the connections to the Atlantic, the Red Sea, and the Indian Ocean. The enclosed Mediterranean had a varied geological history, including isolation from the world ocean, that led to its near drying out during the Messinian crisis (5.96 million years ago) and to drastic changes in climate, sea level, and salinity. The geological history, biogeography, ecology, and human history have contributed to the Mediterranean's high cultural and biological diversity. The recent marine biota in the Mediterranean Sea is primarily derived from the Atlantic Ocean, but the wide range of climate and hydrology have contributed to the co-occurrence and survival of both temperate and subtropical organisms. High percentages of Mediterranean marine species are endemic. This sea has as well its own set of emblematic species of conservation concern, such as sea turtles, several cetaceans, and the critically endangered Mediterranean monk seal (Monachus monachus). It is the main spawning grounds of the eastern Atlantic bluefin tuna (Thunnus thynnus) \[e.g., –\]. There are several unique and endangered habitats, including the seagrass meadows of the endemic Posidonia oceanica, vermetid reefs built by the endemic gastropod Dendropoma petraeum, coralligenous assemblages \[e.g., –\], and deep-sea and pelagic habitats that support unique species and ecosystems \[e.g., –\]. Many sensitive habitats exist within the coastal ecosystems. There are 150 wetlands of international importance for marine and migrating birds, and some 5,000 islands and islets. The region has numerous laboratories, universities, and research institutes dedicated to exploring the sea around them \[e.g., \]. In addition to the unique geologic, biogeographic, physical, and ecological features, our current understanding of the high biodiversity of the Mediterranean Sea is built on the long tradition of study dating from the times of the Greeks and Romans. Historical documentation began with Aristotle, who contributed to the classification and description of marine biodiversity, and was followed by the work of Plinius (Historia naturalis, liber IX) in the first century B.C., Carl von Linné in the eighteenth century, and many others to the middle of the nineteenth century \[e.g., –\]. The first deep-sea investigations began at the end of the nineteenth century \[e.g., –\]. The expeditions of the R.V. “Calypso” by Jacques-Yves Cousteau in the Mediterranean during the 1950s and 1960s provided as well valuable material that supported many important publications on the Mediterranean diversity. The history of ecological research and species discovery in the region has been thoroughly reviewed by Riedl, Margalef, and Hofrichter, though mostly confined to the western Mediterranean. Numerous detailed taxonomic inventories now exist, most of which are specific to sub-regions or to a range of organisms \[e.g.,, among many others\]. Efforts continue to provide complete datasets of taxonomic groups for the entire basin \[e.g., –\], although they need periodic updates. Freely available databases for macroorganism inventory include the Medifaune database, the Food and Agriculture Organization Species Identification Field Guide for Fishery Purposes, the FNAM (Fishes of the North-Eastern Atlantic and the Mediterranean) atlas, and the ICTIMED database. However, Web-based datasets often lack updates because of limitations in funding or expertise, and in general, the marine biodiversity of the Mediterranean is less known than its terrestrial counterpart. There are still important gaps at population, community, habitat, and sub-region levels, as well as in basic information about taxonomy distribution, abundance, and temporal trends of several groups. In some areas biodiversity data exist, but it is not easily accessible, because the inventories are not publicly available. Data are also lacking to evaluate the conservation status of many species. The Mediterranean region has been inhabited for millennia, and ecosystems have been altered in many ways \[e.g.,\]. Therefore, impacts of human activities are proportionally stronger in the Mediterranean than in any other sea of the world. Therefore, combined natural and anthropogenic events shaped the biodiversity of the Mediterranean Sea in the past and are likely to continue to do so. Within this complex framework, our aims were threefold: 1. Review available estimates of Mediterranean marine biodiversity, including new estimates of less conspicuous organisms, updating previous checklists, and incorporating living organisms from microbes to top predators. 2. Describe the main spatial and temporal patterns of biodiversity, including innovative ways of describing these patterns. 3. Summarize the main drivers of change and threats to marine biodiversity. We have collated available information, generated coherent patterns, and identified the current state of knowledge and information gaps, challenges, and prospects for future research. We embrace the concept of biodiversity in its broader definition as the variation of life at all levels of biological organization, but we have focused our efforts on documenting species-level diversity. # Methods ## Diversity estimates ### Total estimates of biodiversity We used our updated taxonomic estimates of species diversity to revise the total estimate of Mediterranean marine biodiversity and to compare it with previous studies. We assessed online data availability by comparing these estimates with global and regional datasets that store an important portion of Mediterranean information, including the World Register of Marine Species database (WoRMS), Marbef Data System (European Register of Marine Species, ERMS) and the Ocean Biogeographic Information System (OBIS), FishBase and SeaLifeBase, AquaMaps, and ICTIMED,. We also calculated the percentage that Mediterranean species of macrophytes and metazoans make up of their global counterpart, by comparing our estimates with global number of marine species according to Bouchet and Green and Short for flowering plants, and Groombridge and Jenkins for other [Vertebrata](http://www.eol.org/pages/2774383) species. ### Estimates by taxonomic group We combined an extensive literature analysis with expert opinions to update publicly available estimates of major taxa and to revise and update several species lists. While most of this information has been incorporated into the supporting materials, here we present detailed summaries of the diversity of some specific groups inhabiting either the extreme ends of the food web (microbes and predators) or the deep-sea environment that represents the most prevalent habitat type in the Mediterranean Sea. In addition, we provide an overview of the newly introduced species. We also identified information gaps by taxonomic group and assessed species discoveries over time for several taxa to visualize the rates of diversity description. and summarize specific information for each taxonomic group for which such analysis is possible, and lists the experts contributing to this synthesis. also lists several experts and taxonomic guides by taxa, although it is not an exhaustive list of experts by taxonomic group in the Mediterranean Sea. provides methodological specifications and the detailed taxonomic review of several groups too, as well as revised checklists, detailed references, and additional information. To classify the estimates of organisms, we followed the taxonomic classification by WoRMS. This classification is followed in the other regional syntheses of marine diversity of the Census of Marine Life (Census) and enables comparison between regions. We therefore used a practical division of the [Eukarya](http://www.eol.org/pages/2908256) into Plantae, Animalia, Protists, and Chromists even though the current kingdom division in the eukaryotes ranges between 6 and 12 and few coincide with these traditional divisions. However, we placed together Archaea and Bacteria because little information exists for either of these divisions. Our review included only generic information on prokaryotic (Bacteria and Archaea) and eukaryotic (Protists) marine microbes and detailed quantification of diversity of a few groups, such as seaweeds and seagrasses (a phylogenetically heterogeneous group of eukaryotic photosynthetic organisms) and metazoans (invertebrates and vertebrates). Within Animalia, we especially focused on the phyla [Porifera](http://www.eol.org/pages/3142), [Cnidaria](http://www.eol.org/pages/1745) (with emphasis on benthic forms), Mollusca, Annelida (with emphasis on Polychaeta), [Arthropoda](http://www.eol.org/pages/164) (with emphasis on Decapoda, Cumacea, and [Mysidacea](http://www.eol.org/pages/4291563)), [Bryozoa](http://www.eol.org/pages/2060), [Echinodermata](http://www.eol.org/pages/1926), [Sipuncula](http://www.eol.org/pages/8871), some other invertebrates forming part of the meiobenthos (Nematoda, benthic Harpacticoida \[[Crustacea](http://www.eol.org/pages/2598871): Copepoda\], benthic Foraminifera, and [Gastrotricha](http://www.eol.org/pages/8728)), Tunicata (with emphasis on Ascidiacea), and the subphylum [Vertebrata](http://www.eol.org/pages/2774383). We did not include the Fungi occurring in the Mediterranean Sea (which are reported to be approximately 140 species). ## Depiction of patterns ### Spatial and bathymetric patterns To describe spatial patterns, we used published available information by region or subregions and by taxonomic group regarding sighting locations, home ranges, or general information on distribution of species in the Mediterranean Sea. We also included information on biodiversity patterns by depth, reviewing data of several taxa available in the literature. Spatial patterns of benthic primary producers and invertebrate species were explored at the scale of large regions or basins. When available, we used detailed spatial data, mostly available in the form of expert-drawn maps or sighting locations, to map spatial patterns of vertebrate species using GIS (geographical information system) software (ArcView by ESRI). For each 0.1×0.1 degree grid cell within the Mediterranean, we estimated the species richness of different taxonomic groups as the sum of the species co-occurring by overlapping expert-drawn distribution maps. We compiled data about exotic fish species from the CIESM (The Mediterannean Science Commission) atlas, and the paper by Quignard and Tomasini. Data for other fish species were available from the FNAM atlas and data compiled by Ben Rais Lasram et al.. We used maps of species occurrence and sighting locations as point data to draw the distributional ranges of resident marine mammals and turtles, but we excluded nonresident or visiting species from the species richness maps. We represented the latter information as point data showing their sighting locations,. The current distribution of Mediterranean monk seal was drawn by integrating information in recent literature,. Information on the distribution of seabird colonies around the Mediterranean, and of Audouin's gull Larus audouinii in particular, was collected from different observations. In our analysis, we considered those regions with uncertain or insufficient data (mainly identified by a question mark in distribution maps) as “no occurrence.” However, we recognize that the absence of data may well reflect a lack of study effort in a given area rather than actual absence of a species, and thus we used the missing data to identify regions that are insufficiently studied. Moreover, available data have been collected mainly from the 1980s to 2000s. Therefore, species richness maps generated in this study should be considered as cumulative distribution maps rather than current distributions. We also used the global species distribution model AquaMaps to generate standardized range maps of species occurrence. AquaMaps is a modified version of the relative environmental suitability (RES) model developed by Kaschner et al.. This is an environmental envelope model that generates standardized range maps, within which the relative probability of occurrence for marine species is based on the environmental conditions in each 0.5×0.5 degree cell of a global grid (see specifications of Mediterranean AquaMaps). We produced AquaMaps of predicted patterns of biodiversity for different taxa in the Mediterranean by overlaying the respective subsets of the 685 available distribution maps for Mediterranean species and counting all species predicted to occur in a given cell. We assumed a species to be present in each cell for which the species- specific predicted relative probability of occurrence was greater than zero. For the prediction of marine mammal biodiversity, we used a probability threshold of species occurrence of at least 0.4 to define presence in a given area, since there is some evidence that lower probabilities for species in this taxa often describe a species' potential rather than its occupied niche. We then used these predictions to visualize species richness patterns by selected latitudinal and longitudinal transects. These results were compared with the maps generated using regional distributions and sighting locations. ### Temporal patterns The analysis of temporal changes in Mediterranean marine biodiversity requires the integration of diverse data from paleontological, archaeological, historical, and fisheries data, as well as ecological surveys and monitoring data \[e.g., –\]. We summarized temporal changes of diversity using studies that dealt with this challenge using available data that informed on changes over past centuries and millennia. We integrated historical records of Mediterranean monk seals and sea turtles around the Mediterranean to explore examples of historical spatial changes,. For the north Adriatic Sea, we analyzed data from Lotze et al., who used a multidisciplinary approach to assess the ecological changes and overall shift in diversity over historical time scales in 12 estuaries and coastal seas worldwide, including the north Adriatic Sea. They assessed the number of species that became depleted (\>50% decline), rare (\>90% decline), or extirpated (locally extinct) in the north Adriatic Sea over past centuries and millennia, based on records for 64 species or species groups that used to be of ecological or economic importance in the Adriatic Sea. These records included marine mammals, birds, reptiles, fish, invertebrates, and plants and were grouped into ten distinct cultural periods. ## Threats to biodiversity Changes in diversity are partially driven by anthropogenic factors, in addition to natural forces. Therefore, our last aim was to identify and quantify the importance of historical and current human-induced drivers and threats to marine biodiversity. We used the aggregated results presented by Lotze et al. and explicitly separated the data available for the north Adriatic Sea as an example to explore historical threats in the Mediterranean. Those authors evaluated human impacts that caused or contributed to the depletion or extirpation of species in the north Adriatic Sea over historical time scales. We also identified current human threats to diversity using published data on specific taxa and areas of the Mediterranean and the opinion of experts. Each expert was asked to (1) list main threats to diversity for their taxonomic expertise group using data available and experience, and (2) rank those threats from 1 to 5, taking into account the relative importance of each threat to the biodiversity (0: no importance, 1: lowest in importance, 5: highest in importance). The experts repeated the ranking exercise considering data available and projecting their results 10 years into the future. In addition and to visualize the impacts of climate warming on species diversity, we documented the mean location of February (the coldest month of the year in the Mediterranean) sea surface isotherms (°C) for the period 1985 to 2006, integrating several data sources. We also generated current and projected future temperature maps, which we compared with sea surface temperature (SST) data from the 1980s. First, we compiled weekly SST data from the National Climatic Data Center (National Operational Model Archive and Distribution System Meteorological Data Server, NOMADS, NOAA Satellite and Information Service), and interpolated maps at 0.1° resolution. Next, we averaged weekly SST values from 1981 to 1984 for each 0.1° grid cell. Last, we used the Mediterranean model OPAMED8 based on the A2 IPCC scenario to visualize the future climate. This model considers main forcing parameters (river runoffs, exchanges with connected seas, and wind regimes) and was used to generate climate data for the middle (2041–2060) and the end of the twenty-first century (2070–2099). Finally, we visualized potential hot spots for conservation efforts by linking predicted species distributions from the AquaMaps model to status information reported by the International Union for Conservation of Nature. From the available AquaMaps, a total of 110 maps belonged to vertebrate species that had been classified as critically endangered, endangered, vulnerable, or near threatened in the Mediterranean Sea. This represented the 16% of all species included in the Mediterranean AquaMaps. We subsequently mapped the richness of these species using a probability threshold of more than 0.4, which usually corresponds to the most frequently used and ecologically most important habitats. # Results ## Diversity estimates in the Mediterranean Our analysis revealed approximately 17,000 species occurring in the Mediterranean Sea. Of these, at least 26% were prokaryotic (Bacteria and Archaea) and eukaryotic (Protists) marine microbes. However, the data available for Bacteria, Archaea, and Protists were very limited, so these estimates have to be treated with caution (see next section), as well as data for several invertebrate groups (such as [Chelicerata](http://www.eol.org/pages/2579982), [Myriapoda](http://www.eol.org/pages/2631567), and Insecta). Within the Animalia, the greater proportion of species records were from subphylum [Crustacea](http://www.eol.org/pages/2598871) (13.2%) and phyla Mollusca (12.4%), Annelida (6.6%), Platyhelminthes (5.9%), [Cnidaria](http://www.eol.org/pages/1745) (4.5%), the subphylum [Vertebrata](http://www.eol.org/pages/2774383) (4.1%), [Porifera](http://www.eol.org/pages/3142) (4.0%), [Bryozoa](http://www.eol.org/pages/2060) (2.3%), the subphylum Tunicata (1.3%), and [Echinodermata](http://www.eol.org/pages/1926) (0.9%). Other invertebrate groups encompassed 14% of the species, and Plantae included 5%. Detailed biodiversity estimates of main taxonomic groups of benthic macroscopic primary producers and invertebrates are summarized in and documented in in detail. Available information showed that the highest percentage of endemic species was in [Porifera](http://www.eol.org/pages/3142) (48%), followed by [Mysidacea](http://www.eol.org/pages/4291563) (36%), Ascidiacea (35%), Cumacea (32%), [Echinodermata](http://www.eol.org/pages/1926) (24%), [Bryozoa](http://www.eol.org/pages/2060) (23%), seaweeds and seagrasses (22%), Aves (20%), Polychaeta (19%), Pisces (12%), Cephalopoda (10%), and Decapoda (10%). The average of the total endemics was 20.2%. In some groups the percentage of endemics was now lower than in the past, partly due to new finding of Mediterranean species in adjacent Atlantic waters. ## The biodiversity of the “smallest” An important bulk of species diversity was attributed to the prokaryotic (Bacteria and Archaea) and eukaryotic (Protists) marine microbes. However, the differences in the methodologies and types of studies and the continuously changing state of our knowledge of marine microbial diversity make it difficult to provide species estimates for the Mediterranean (or from anywhere else) and establish comparisons. Current methods cannot yet provide reliable estimates of the microbial richness of a system \[e.g., \] because of (i) our limited capacity to describe morphological variability in these organisms, (ii) the limited development and the biases associated with molecular techniques used to identify them, even with the use of the most powerful of these techniques, and (iii) the uncertainty in determining a “microbial species” and where to draw the line that differentiates one species from another. Morphological variability is used to describe diversity of some groups of microbes, such as ciliates and microphytoplankton, but this is not useful for most nano- and almost all picoplanktonic organisms, including all Archaea and most Bacteria. Therefore, until recently, surveys of microbial diversity were mainly limited to those taxa with enough features to be described under an optical microscope. Among phytoplankton, the best-studied groups included thecate dinoflagellates, diatoms, coccolithophores, and silicoflagellates. Among microzooplankton, groups like tintinnids, foraminifers, or radiolarians attracted most attention. Much less information is available on “naked” auto- or heterotrophic flagellates and on small picoplankton species. However, researchers have made efforts to obtain estimates of the dominant microbial species in Mediterranean waters. The expansion of electron microscopy in the last decades of the twentieth century helped to untangle inconsistencies in the distribution of some described species and to consolidate the establishment of a biogeography of many protist taxa. More recently, molecular techniques (metagenomics) have been used to enumerate the microorganisms present in a given sample and have completely transformed the field by changing ideas and concepts. These advances have highlighted the problems with the species concept when applied to microbial communities, which may be based on morphology, biology, or phylogeny. Furthermore, different methodologies have biases that give different views of microbial diversity \[e.g.,\], and now we know that microdiversity is a general characteristic of microbial communities, making the delimitation of “diversity” units difficult. To avoid some of the problems with the “species” delimitation, some authors prefer to use “functional diversity”: the amount and types of microbial proteins (e.g., functions) in the sample \[e.g., \], rather than “species” diversity. According to the compilation published in Hofrichter, the number of described protist species in the Mediterranean is approximately 4,400. However, this estimate requires cautious interpretation and it is likely that many morphospecies, more or less well described, will include a number of cryptic or pseudocryptic variants \[e.g., \]. Molecular methods have recently uncovered new sequences that are being associated with the organisms they represent. Fingerprinting techniques have been used to compare microbial communities and establish the scale of variability of these communities. For example, Schauer et al. determined that, along the coastal northwestern Mediterranean, the time of the year was more important than exact location in determining bacterial community structure. Acinas et al. and Ghiglione et al. showed that microbial communities tend to be similar in the horizontal scale and much more variable on the vertical scale, but these techniques are not appropriate to determine the number of species present and usually refer only to the dominant organisms. Recent application of new methodologies (such as metagenomics and 454-tag sequencing) will in the near future provide more accurate estimates. All studies to date concur in identifying members of the SAR11 group as some of the most abundant Mediterranean bacteria, comprising 25–45% of the reported sequences \[e.g.,\]. These are followed by other Alphaproteobacteria, which tend to be more common in coastal regions and during algal blooms (such as [Roseobacter](http://www.eol.org/pages/11716174)-like). Cyanobacteria (Prochlorococcus and Synechococcus), diverse culturable (Alteromonadales) and unculturable Gammaproteobacteria and Bacteroidetes form the rest of the diversity with some differences with depth and with distance from land. Several studies have concentrated in the diversity of subgroups of these abundant bacteria in the Mediterranean \[e.g., ,\]. Additionally, the diversity of deep samples and the communities from which they are taken have received considerable attention in the Mediterranean. Specific and likely unique ecotypes of some bacteria appear at certain depths, \[e.g., \], free-living communities appear to be as complex as epipelagic communities, and appear to vary seasonally, as do surface communities. The deep-sea Mediterranean maintains several extremely peculiar and interesting ecosystems, such as the deep hypersaline anoxic “lakes” in the Ionian Sea that are reported to include several new and little-known microbial lineages \[e.g., \]. Some studies have shown that bacterial richness peaks in tropical latitudes \[e.g., \] and concluded that at Mediterranean latitudes the number of detectable “operational taxonomic units” (OTUs) is between 100 and 150. Zaballos et al. arrived at a similar value that, once extrapolated, indicated a value of approximately 360 OTUs for surface waters. A slightly lower value was estimated for the coastal Blanes Bay Microbial Observatory \[e.g., \] based on a different approach. Archaeal richness is known to be lower than bacterial richness \[e.g., \], and this has been seen in the Mediterranean and in other oceans. Results of these new sequencing techniques suggest that microbial richness in the sea is much higher because of the presence of a “rare biosphere” composed of very few individuals of many distinct organism types. Application of this technique to data from the northwestern Mediterranean indicates that the numbers should be raised to about 1,000 “bacterial species” per sample. Again, the real magnitude of bacterial richness in the Mediterranean cannot be appreciated with the techniques available. A similar situation to that with prokaryotes occurs with small eukaryotes, which are photosynthetic, heterotrophic, or mixotrophic organisms. These small eukaryotes are found in abundances of 10<sup>3</sup>–10<sup>4</sup> ml<sup>−1</sup> and have low morphological variability. Thus we must rely on molecular techniques to grasp their diversity. Molecular work has allowed the discovery of new groups of eukaryotes present in this smallest size class. The study of Mediterranean protists has benefited from the early establishment of marine laboratories and a number of illustrated books and checklists \[e.g., –\]. More recent inventories can be found in Velasquez and Cruzado and Velasquez for diatoms, Gómez for dinoflagellates and Cros for coccolithophorids. The compilation of northwestern Mediterranean diatom taxa of Velasquez records 736 species and 96 genera. The checklist of Gómez contains 673 dinoflagellate species in 104 genera. Cros lists 166 species of coccolithophorids of the northwestern Mediterranean and revised the classification of several important taxa \[see also 160\]. Recently, the discovery of a number of combination coccospheres bearing holo- and heterococcoliths fostered the recognition that holococcolithophores do not belong to a separate family, as previously accepted, but are part of a life cycle that includes holo- and heterococcolithophore stages. The biodiversity of photosynthetic nano- and picoflagellates other than coccolithophores is poorly known for most groups, as may be expected from the difficulties involved in their identification. However, in the last decade, work using optical and electron microscopy, often in combination with molecular and culturing techniques, has considerably increased the taxonomic knowledge of many of these groups and has highlighted the potential existence of much cryptic or unknown diversity \[e.g.,\]. There are few taxonomic surveys of heterotrophic flagellates \[e.g., \], although many phytoplankton studies based on microscopy also included taxa from these groups. Massana et al. describes a high diversity of picoeukaryotic sequences, belonging to two groups of novel alveolates (I with 36% and II with 5% of clones), dinoflagellates (17%), novel stramenopiles (10%), prasinophytes (5%), and cryptophytes (4%). Later work has shown that these novel stramenopiles are free-living bacterivorous heterotrophic flagellates. Most of the biodiversity work on ciliates has focused on tintinnids or loricate ciliates, while studies involving naked ciliates tend to use groupings based on ecological morphotypes and only rarely include detailed taxonomical work \[e.g.,–\]. Numbers of species ranging from 40 to 68 were recorded in one to several-year surveys of various Mediterranean sites \[among others 154\]. Other groups, such as the Foraminifera, which have calcium carbonate tests, and the Radiolaria, which produce siliceous or strontium sulfate skeletons, have been the subject of many stratigraphical and paleoceanographical studies. However, biodiversity work on living Foraminifera and Radiolaria in the Mediterranean is scarce \[e.g.,\]. Hofrichter provided a systematic summary of the main groups and species of both autotrophic and heterotrophic protists found in the Mediterranean. ## The biodiversity at high trophic levels Species that occupy the upper trophic levels, normally beyond the level of secondary consumers, are classified as predators. They have lower diversity than other taxonomic groups, but information available is usually more detailed. We reviewed data available for fish, seabirds, marine mammals, and turtles in the Mediterranean Sea. Ground-breeding species such as seabirds (gulls and terns) are counted using census bands and monitored by satellite tracking. However, procellariiforms reproduce in caves and burrows in cliffs on remote, inaccessible islets, and census methods to estimate population densities are not totally reliable. Population models, based on demographic parameters, allow researchers to estimate extinction probabilities. A census of marine mammals or turtles normally uses transect data collected from aerial or boat-based sighting surveys developed to assess abundance, while movement patterns are tracked with transmitters and monitored by satellite tracking as well. Fish species are mainly studied using scuba diving or fishing techniques. There is still some discussion about diversity estimates for these taxonomic groups. For fish species, for example, several estimates of Mediterranean diversity exist: Quignard lists a total of 562 fish species occurring in the Mediterranean Sea; Whitehead et al. mention 589; Fredj and Maurin list a total of 612 species (and identified 30 species as uncertain); and Quignard and Tomasini register 664 species. Hofrichter summarizes 648 species, and Golani et al. report a total of 650 fishes. Fish diversity estimates also change as new species are described or reclassified. The updated list of exotic fish species reveals that the Mediterranean currently contains 116 exotic species, although more species are likely to be cited. There is also a long-standing controversy regarding genetic differentiation among a few fish populations and sub-basins, especially of commercial species due to management implications (for example for the European anchovy Engraulis encrasicolus), although results are still under debate \[e.g., \]. Approximately 80 fish species are elasmobranchs, although the status of some is uncertain because of infrequency or uncertain reporting \[e.g., ,\]. According to Cavanagh and Gibson, nine of these elasmobranch species may not breed in the Mediterranean, while some are rare because the Mediterranean represents the edge of their distribution ranges. Only four batoid species are Mediterranean endemics: the Maltese skate (Leucoraja melitensis), the speckled skate (Raja polystigma), the rough ray (R. radula), and the giant devilray (Mobula mobular). Nine species of marine mammals are encountered regularly in the Mediterranean. Of these species, five belong to the Delphinidae, and one each to the Ziphiidae, Physeteridae, Balaenopteridae, and Phocidae. Other 14 species are sporadically sighted throughout the basin and are considered “visitors” or “non-residents.” Of the seven living species of sea turtles, two (the green and the loggerhead Chelonia mydas and Caretta caretta - Cheloniidae) commonly occur and nest in the Mediterranean, and one (leatherback turtle Dermochelys coriacea - Dermochelyidae) is regularly sighted but there is no evidence of nesting sites. The other two (hawksbill and Kemp's riddle turtles Eretmochelys imbricata and Lepidochelys kempi - Cheloniidae) are extremely rare and considered to be vagrants in the Mediterranean. Seabirds from the Mediterranean have a low diversity (15 species) and their population densities are small, consistent with a relatively low-productivity ecosystem compared with open oceans, and particularly with upwelling regions. Ten of the Mediterranean species are gulls and terns (Charadriiformes), four are shearwaters and storm petrels (Procellariiformes), and one is a shag (Pelecaniformes). Three of the ten species are endemics. ## What is hidden in the deep? Because of the large size of the Mediterranean deep-sea ecosystems, our knowledge of the benthic deep-sea diversity is incomplete. In the past 20 years, several studies on deep-sea sediment diversity have been undertaken in various oceans \[e.g.,\] but have been limited to a few taxonomic groups. However, due to technological improvements that render the deep waters more accessible, the deep-sea benthos of the Mediterranean has received increased attention and there is progress toward a more comprehensive view of the levels, patterns, and drivers of deep-sea biodiversity in this semienclosed basin. Its paleoecological, topographic, and environmental characteristics suggest that the Mediterranean Sea is a suitable model for investigating deep-sea biodiversity patterns along longitudinal, bathymetric and energetic gradients across its different regions. There are few areas with depths greater than 3,000 m, and typically bathyal or abyssal taxonomic groups are limited. Cold-water stenothermal species that elsewhere represent the major part of the deep-sea fauna are also unknown in the Mediterranean Sea. The Mediterranean abyssal macrobenthos comprises a large number of eurybathic species and only 20–30 true abyssal species. In the western basin, where the depth does not exceed 3,000 m, the abyssal fauna is less abundant than in the deeper eastern basin, where abyssal species are dominant in the Matapan trench, which is more than 5,050 m deep. The close affinity between Mediterranean and Atlantic congeneric deep- water species suggests that the ancestors of the Mediterranean bathyal endemic species moved from the Atlantic when conditions were favorable (i.e. when larvae of deep Atlantic fauna was able to enter in the Western Mediterranean due to hydrodynamic and physico-chemical conditions allowed it). According to Pérès, the deep-water fauna of the Mediterranean has a lower degree of endemism than that of the Atlantic at similar depths. So while the Mediterranean basin is recognized as one of the most diverse regions on the planet, the deep sea in the Mediterranean may contain a much lower diversity than deep-sea regions of the Atlantic and Pacific oceans. The reasons for such a low diversity may be related to (a) the complex paleoecological history characterized by the Messinian salinity crisis and the almost complete desiccation of the basin, and (b) the Gibraltar sill that is, potentially, a physical barrier to the colonization of larvae and deep-sea benthic organisms from the richer Atlantic fauna. These factors may explain the composition of the benthos in the deep sea of the Mediterranean. It may also be that the high deep- sea temperatures (about 10°C higher than in the Atlantic Ocean at the same depth) have led to a Mediterranean deep-sea fauna that consists of reproductively sterile pseudopopulations that are constantly derived through larval inflow. These postulates were based on the analysis of the macrobenthos, characterized by life cycles with meroplanktonic larvae that are spread by currents. However, the populations of the most common benthic mollusks in depths greater than 1,000 m off the Israeli coast are composed of both adult and juvenile specimens, and one species, Yoldia micrometrica, the most common and abundant species in the eastern Mediterranean, is unrecorded from the westernmost part of the sea. In addition, and though much reduced in diversity and richness compared with the deep-sea fauna of the western and central basins of the Mediterranean, the Levantine bathybenthos is composed of autochthonous, self-sustaining populations of opportunistic, eurybathic species that have settled there following the last sapropelic event. Macpherson and Briggs have suggested that within the Atlantic-Mediterranean region, the fauna (including invertebrates and fishes) of the Mediterranean is more diverse than that of the Atlantic and displays considerable endemism. For strictly deep-dwelling species (e.g., the deep-water decapod crustacean family Polychelidae), the Gibraltar sill is not an impenetrable barrier for some deep- waters macrobenthic species. Moreover, available hypotheses did not consider meiofauna diversity, which is characterized by direct development but also by a small size, which allows organisms' resuspension and drifting over wide regions. This is consistent with information on the most abundant deep-sea phylum, the Nematoda, which often accounts for more than 90% of total meiofauna abundance. Nematode diversity has been investigated only in a few areas of the deep sea in the Mediterranean: slopes of the Gulf of Lions, Catalan margin and Corsica, Tyrrhenian basin, and Eastern Mediterranean \[e.g., –\]. Recent collections from a limited number of sites throughout the Mediterranean basin (at approximately 1,000 m, 3,000 m, and 4,000 m depth), suggest that, conversely to what was expected, the deep-sea nematode fauna of the Mediterranean basin is rather diverse. At bathyal and abyssal depths, levels of nematode genera and species richness are similar to those reported from other deep-sea areas of the world oceans. In the deep sea of the Mediterranean, small-bodied taxa (e.g., meiofauna) can reach a high diversity, and with the presence of a high prokaryotic diversity in the sediments of the deep-sea Mediterranean, this may change the view that the Mediterranean deep-sea biota is impoverished in comparison with its Atlantic counterpart. Endemic macrobenthic species account for approximately 13–15% of total species number at depths from 200 m to 1,000 m, and approximately 20% at 2,000 m. These estimates are similar for each taxon and are further supported by the continuous discovery of new species (both within the highly diverse Nematoda and in rare phyla such as the [Loricifera](http://www.eol.org/pages/1537)) in different sectors of the deep Mediterranean. Therefore, the general conclusion that the biodiversity is high in coastal systems and low in the deep sea of the Mediterranean might not hold true. Detailed references about the deep Mediterranean can be found in. ## New biodiversity The biodiversity of the Mediterranean is definitively influenced by the introduction of new species \[e.g.,–\]. Since the first review of exotic species in the Mediterranean, the studies in this topic have intensified. Now more than 600 metazoan species have been recorded as alien, these representing 3.3% of the total estimates (and for detailed information by taxonomic group). However, this estimate is continuously increasing and may be as high as 1,000 species if unicellular aliens and foraminiferans are included \[e.g.,\]. Most of these introductions are littoral and sublittoral benthic or demersal species (or their symbionts). Because the shallow coastal zone, and especially the benthos, has been extensively studied and is more accessible than deeper waters, new arrivals probably will be encountered and identified in shallow waters. The species most likely to be introduced by the predominant pathways (the Suez Canal, vessels, and mariculture) are shallow-water species. A taxonomic classification of the alien species showed that the alien phyla most frequently recorded are Mollusca (33%), [Arthropoda](http://www.eol.org/pages/164) (18%), Chordata (17%), Rhodophyta (11%), and Annelida (8%). The data are presumably most accurate for large and conspicuous species that are easily distinguished from the native biota and for species that occur along a frequently sampled or fished coast and for which taxonomic expertise is readily available. Data are entirely absent for many of the small members of invertebrate phyla. Thus, the true numbers of alien species are certainly downward biased. The native range of the alien species in the Mediterranean was most commonly the Indo-Pacific Ocean (41%), followed by the Indian Ocean (16%), and the Red Sea (12%), while some species have a pantropical or circumtropical distribution (19%). The actual origins of the Mediterranean populations of a species widely distributed in the Indo-Pacific Ocean may be their populations in the Red Sea, both from the Indian or Pacific oceans, or a secondary introduction from already established populations in the Mediterranean itself \[e.g., \]. However, and with few notable exceptions \[e.g.,\], the source populations of alien species in the Mediterranean have not been assessed by molecular means. Even so, it is clear that most alien species in the Mediterranean are thermophilic and therefore originated in tropical seas. The exceptions are exotic algae, of which the largest numbers are in the Gulf of Lions and the northern Adriatic, and a few other examples \[e.g., \]. As far as can be deduced, the majority of aliens in the Mediterranean entered through the Suez Canal (Erythrean aliens) (53%), and an additional 11% were introduced primarily through the Canal and then dispersed by vessels. Introductions from vessels from other parts of the world account for 22% of introduced species, and aquaculture accounts for 10%. A further 2% arrived with the introduction of aquaculture and were secondarily spread by vessels. The means of introduction differ greatly among the phyla: whereas of the alien macrophytes, 41% and 25% were introduced through mariculture and vessels, respectively, the majority of alien crustaceans, mollusks, and fish are Erythrean aliens (59%, 64%, and 86%, respectively), and mariculture introductions are few (4%, 5%, and 4%, respectively) \[, B.S. Galil, personal observation\]. The numbers of alien species that have been recorded over the past century have increased in recent decades. The increasing role of the Mediterranean as a hub of international commercial shipping, a surge in the development of marine shellfish farming over the last 25 years, and the continued enlargement of the Suez Canal have contributed to the resurgence of introductions since the 1950s. Many introduced species have established permanent populations and extended their range: 214 alien species have been recorded from three or more peri- Mediterranean countries, and 132 have been recorded from four or more countries \[, B.S. Galil, personal observation\]. A comparison of the alien species recorded along the Mediterranean coasts of Spain and France and an equivalent length of coast in the Levantine Sea (from Port Said, Egypt, to Marmaris, Turkey) showed marked differences in their numbers, origin, and means of introduction. There are nearly four times as many alien species along the Levantine coast (456 species) as along the western coast of the Mediterranean (111 species). The majority of aliens in the Eastern Mediterranean entered through the Suez Canal (68% of the total, 14% vessel- transported, 2% mariculture), whereas mariculture (42%), vessels (38%), or both (5%) are the main means of introduction in the Western Mediterranean \[, B.S. Galil, in preparation\]. Climate change favors the introduction of Red Sea species in the southeastern Mediterranean and their rapid spreading northwards and westwards (see section 4.2c and d). It similarly favors species coming from the African Atlantic coasts to enter the western basin. ## Spatial patterns of Mediterranean biodiversity ### Longitudinal and latitudinal patterns Describing the distribution of marine diversity is as important as quantifying it. In the Mediterranean, a northwestern-to-southeastern gradient of species richness was observed in most groups of invertebrate species analyzed here, with a highly heterogeneous distribution of species in the different regions (and for detailed information). We noticed only a few exceptions. For example, while there was the same number of Euphausia species in the western and central basins, estimates for several other invertebrate groups were higher in the Aegean Sea than in central areas of the Mediterranean. These exceptions may be due to different species tolerance to environmental factors (such as temperature and salinity), connectivity between regions, and to the lack of data in some regions. We found similar results for vertebrate species. There was a decreasing gradient from northwest to the southeast, while the sea around Sicily had the highest richness (375 species per 0.1×0.1 degree cell), followed by other northwestern coastal and shelf areas. The distribution of elasmobranch species was not homogenous either, showing a higher concentration of species in the west. The endemic richness gradient of fish species was more pronounced with latitude, the north side exhibiting a greater richness, and the Adriatic appearing as a hot spot of endemism with 45 species per cell. Spatial patterns also showed how most of Mediterranean coastal waters have been colonized by exotic species. The highest richness of exotic species occurred along the Israeli coast. Marine mammals were concentrated in the Western Mediterranean and Aegean seas. Of the nine resident marine mammals, eight were found in the western part of the basin. This distribution pattern was also observed for the visiting marine mammals. Two of the three resident sea turtles (loggerhead, green, and leatherback turtles) occurred in the central Mediterranean and Aegean seas, while the two visiting turtles were absent from the eastern side. There were fewer seabird colonies and seabird density was lower in the southeast than the northwest. Spatial patterns of benthic biodiversity in the deep sea are poorly known in comparison with other ecosystems. Available information is scarce and our maps and estimates include only approximations for the deep sea. In this context, metazoan meiofauna and, in particular, nematodes can be used to describe the biodiversity patterns in the deep sea. Deep-sea nematode diversity appears to be related to that of other benthic components such as foraminifers, macrofauna, and the richness of higher meiofauna taxa in the deep sea. Results for the deep sea of the Mediterranean show a clear longitudinal biodiversity gradient that also occurs along the open slopes, where values decrease eastward, from Catalonia to the margins of southern Crete. The analysis of the Nematoda indicates that at equally deep sites, nematode diversity decreases from the western to the eastern basin and longitudinal gradients are evident when comparing sites at 3,000 m or 1,000 m depth. Complementary information on spatial patterns of the deep Mediterranean fauna can be found in. Additional information from the literature on spatial patterns of Mediterranean marine diversity suggests that the measurement of local α-diversity is not sufficient to draw a clear picture for the whole Mediterranean basin. Whittaker defined α-diversity as the number of species found in a sample (or within a habitat), β-diversity as the extent of species replacement along environmental gradients (termed “turnover diversity” by Gray), and γ-diversity as the diversity of the whole region. The analysis of β-diversity of Nematoda among different sites in the deep sea of the Mediterranean and across bathymetric and longitudinal gradients reveals an extremely high species turnover. By comparing nematode assemblages at (a) different depths, (b) similar depths in two different basins, and (c) similar depths within the same basin, the dissimilarity of biodiversity among deep-sea samples is always greater than 70%. On average, the dissimilarity of nematode diversity between western and eastern Mediterranean at about 3,000 m depth is greater than 80% and at similar depths the dissimilarity between Atlantic and Western Mediterranean exceeds 90%. These findings indicate that each region is characterized by the presence of a specific assemblage and species composition. This has important implications for estimating the overall regional diversity (γ-diversity) but also suggests the presence of high biogeographic complexity in the Mediterranean. However, these patterns may not hold for all the taxonomic groups, and a broader comparison is needed. ### Spatial patterns predicted with AquaMaps Predicted patterns of overall species richness based on AquaMaps showed a concentration of species in coastal and continental waters most pronounced in the Western Mediterranean, Adriatic, and Aegean seas. Less than half of the species were predicted to occur in the deeper waters of the central Mediterranean, and biodiversity was particularly low in offshore waters at the eastern end. Given the overall proportion of ray-finned fishes in AquaMaps dataset, overall biodiversity patterns from these figures were largely dominated by Actinopterygii. The concentration in coastal waters was more pronounced in the map focusing on these taxa. Predicted species richness of elasmobranchs was similar to that for Actinopterygii, but rays and sharks occurred farther offshore, especially in the waters of Tunisia and Libya. The Aegean Sea, especially its northern sector, also showed high invertebrate species richness, which was otherwise low in most of the remaining central and eastern basin. Biodiversity patterns for the marine mammals contrasted with patterns for fishes and invertebrates in that many species were also predicted to occur in the offshore western and central basin waters, and particularly in slope waters. The biodiversity patterns of sea turtles broadly mimic those of the other more species-rich taxa in that there was a concentration in coastal areas and a decline in species richness from the northwest to the southeast. Therefore, there were similarities and differences between expert-drawn maps and modeling results. The pattern describing species richness of ray-finned fish was similar overall, but for the elasmobranchs there were some noticeable differences. While both methods identified areas around Sicily, the coast of Tunisia, and the Western Mediterranean as high diversity hot spots, the Adriatic and Aegean seas showed up as high in species richness only in the predicted maps. Both types of analyses arrived at similar patterns for marine mammals, although the lack of distinction between resident and visitor species in the AquaMaps analysis hampered the direct comparison of diversity patterns for these taxa. Nevertheless, differences could be seen around the Aegean and Alboran seas. Maps of sea turtle diversity showed peaks in the western region based on both types of analysis, but there were a few discrepancies regarding the eastern Mediterranean. AquaMaps analysis of predicted species richness of invertebrates also showed a geographical gradient. Latitudinal transects corresponding to cross sections through the species richness map highlighted the importance of coastal habitats for fishes and invertebrates. These habitats were represented by peaks in species numbers in areas corresponding to shelf waters. Cross-section gradients followed a similar pattern for fishes and invertebrates; large variations were mostly determined by depth changes along the respective transects. There was also an overarching trend of decreasing species richness from western to eastern waters, a trend that became particularly pronounced in the southern transects. Marine mammal transects diverged from the general trend in that species richness was less directly linked to depth variation. Changes in fish and invertebrate species richness along three different longitudinal cross sections again followed similar depth contours. Marine mammal longitudinal biodiversity patterns in the Western Mediterranean followed a different trend with highest numbers predicted to occur in deeper waters, such as the southern Tyrrhenian Sea. There appeared to be a general decrease of diversity from northern to southern regions. ### Bathymetric patterns Because seaweeds and seagrasses are photosynthetic organisms, their development is limited to shallow areas where there is enough light for growth. They are distributed between the mediolittoral zone and the deepest limit of the circalittoral zone, situated at 110 m in the clearest waters of the western Mediterranean and a bit deeper in the even more oligotrophic waters of the eastern part. Their growth occurs only on the continental shelves and the uppermost parts of seamounts above 150 m depth. Seaweeds, which have a limited distribution across the whole bathymetric gradient, show an increase in species richness from the highest levels of the mediolittoral rocks down to the lower infralittoral and upper circalittoral communities. There they display the highest species richness, as many as 150 species reported in a surface of 1,600 cm<sup>2</sup> at 18 m depth. Species richness then decreases along the circalittoral zone from the shallowest down to the deepest parts, becoming nil at the beginning of the bathyal zone. The pattern of a generally decreasing diversity with increasing depth was also documented here for invertebrate and fish species and is consistent with previous studies \[e.g.,\]. Diversity was concentrated in coastal areas and continental shelves, mainly above 200 m depth. However, patterns did not necessarily show a monotonic decrease with depth. For example, more polychaete species inhabited shallow waters than deep waters, particularly below 1,000 m deep, but this pattern was less clear when looking at maximum ranges of depth. It is not clear whether this is a real pattern of lower deep-sea diversity or a result of the lack of proper faunistic studies in the Mediterranean at those depths. Larger numbers of cumacean species were found in shallow waters of 0–99 m depth (48 species) and between 200 m and 1,400 m depth, but species richness decreased below this depth (references in). The highest endemism (43.8%) was found between 0 and 99 m depth. The largest number of mysidaceans (54 species) was also found in shallow waters less than100 m deep. At depths between 100 m and 1,000 m, 27 species were found, and below 1,000 m, 21 species. The level of endemism was also higher in the 0–100 m depth interval (29 species, 78.4% of total endemism) than in the 100–1,000 m interval (3 species, 8.1%) or below 1,000 m (5 species, 13.5%), in line with results obtained for cumaceans. The circalittoral zone was the region with highest anthozoan species richness (61.8% by numbers of species) followed by the infralittoral (57.6%) and bathyal (40%) zones. Half of the total number of species were restricted to one of the infra-, circa-, or bathyal zones, and 9.7% were eurybathic, while the remaining species (40%) were intermediate in depth distribution. We also found exceptions to the pattern of decreasing diversity with depth. The bathymetric range of Mediterranean sipunculans was generally quite wide. Most of the Mediterranean records were bathyal, whereas there were few sublittoral records. Other studies carried out on depth-related distribution of marine biodiversity in the deep sea of the Mediterranean available form the literature suggest a generally unimodal pattern of species richness, the highest values of which are observed at intermediate depths (about 2,000 m) and lower values at upper bathyal (\<2,000 m) and abyssal (\>2,000 m) plains. More recent studies, however, have demonstrated that such patterns are not always recognizable \[e.g., –\]. In open slope systems, bathymetric gradients of species diversity have been widely documented \[e.g., –\]. In the Mediterranean, nematode diversity also decreases with depth, but the degree of species decrease is limited and ample ranges of biodiversity are observed at the same depth. These results suggest that the eurybathy of the Mediterranean fauna (3,613 species) could be lower than previously reported. For example, analysis of all the existing nematode diversity data from the Aegean Sea showed that there is a gradual increase of diversity with depth from the littoral zone down to the bathyal areas (2,000 m) (N. Lampadariou, personal observation). Complementary information on bathymetric patterns of the deep Mediterranean fauna are explored with detail in. ## Temporal trends Available data from the literature show that environmental factors have led to profound changes in the abundance, distribution, and composition of Mediterranean marine species in the distant past \[e.g.,\]. For example, during the Cretaceous, the Mediterranean Sea (called Tethys) was connected to the Atlantic on its western side and the Indo-Pacific on its eastern side. The two oceans contributed very different faunas to the Tethys. During the Miocene, the Tethys was isolated from the Indo-Pacific Ocean and at the Messinian stage, the connection with the Atlantic Ocean was also closed. During this Messinian salinity crisis, the Mediterranean underwent severe desiccation that drove most species to extinction. Although some shallow areas remained on the two sides of the Siculo-Tunisian Strait, and there were many allopatric speciations, , the reopening of the Strait of Gibraltar 5 million years ago led to restocking of the Mediterranean with fauna and flora from the Atlantic. Up to the nineteenth century, the Mediterranean had been connected with the eastern Atlantic Ocean only. In this section, however, we summarized main changes since the end of the last ice age (approximately 12,000 years ago). During this time there were notable climate-driven fluctuations but also human-induced changes due to the long periods of exploration and exploitation, and more recently the reopening to the Red Sea through the Suez Canal, the globalization of commerce and trade, increasing pollution and eutrophication of coastal areas, habitat modification and loss, and finally the looming climate change. Early evidence of human interaction with marine fauna in the Mediterranean Sea comes from the Paleolithic period and continues through the Mesolithic and Neolithic periods (approximately 20,000–4000 B.C.). Zooarchaeological remains are found in Franchthi Cave in the southern Argolid, Greece, Las Cuevas de Nerja in southern Spain, Athlit Yam, a submerged site south of Haifa Bay in Israel, Cape Andreas Kastros in Cyprus, and the Strait of Gibraltar. In Greece, fish bones of large tuna, Sparidae and Mugillidae, were found. Zooarchaeological remains in Spain include 20 taxa and show changes in mean fish size and range over time that have been considered as indication of overfishing. At Cape Andreas Kastros in Cyprus and in Athilit Yam, 90% of the remains are grey trigger fish (Balistes capriscus), which points to intensive fishing regardless of size. In Gibraltar, remains of Mediterranean monk seals and mollusks consumed by humans were found. However, stable isotope analyses of human bones show that between 10,000 and 8000 B.C., the main Mediterranean coastal populations did not rely significantly on marine food. Since the fifth century B.C., humans have exploited marine resources. Aristotle, in his zoological works dating to the fourth century B.C., focuses his scientific interest on fish and invertebrates exploited by humans in various ways. Fisheries in the Aegean communities by that period are characterized by variability both in the nature and abundance of the exploited fish and in the manner of their exploitation. Mollusks and other invertebrates are part of the diet of ancient Greeks, and their consumption is connected with the treatment or prevention of various health problems and diseases. Bath sponges of the genera Spongia and Hippospongia, collected by skillful divers, are widely exploited for household and personal hygiene purposes, and play a principal role in medical practice. Commercial fishing and fish processing activities play an important role in the Pontic economy. The export of fish and fish products, including salt-fish (tarichos) and fish sauce (garum) mainly from European anchovy to the Aegean Sea, continue into the Roman period. These products are exported from the western Mediterranean, but garum is forgotten in the west by the tenth century, although it is still prepared in Constantinople in the fifteenth and sixteenth centuries. Naval trade traffic becomes intense, and invasions of islands from the mainland are already common, and they result in the beginning of the introduction of alien species in those ecosystems. Some of these introductions (rats, carnivores) trigger the extirpation of many seabird colonies, and they have shaped the current distribution of several seabird species. Seafood becomes increasingly popular toward the end of Roman domination, probably because of the proximity of, and access to, marine resources. There is historical evidence of overfishing in some parts of the Western Mediterranean in the early Imperial period. Even then, certain fishing techniques are prohibited to manage or counteract the decline in fish stocks (such as fishing by torch lights at night), and efforts are made to boost natural availability with introduced fish and shellfish stocks. For example, the parrot fish (Sparisoma cretense) is captured in the Aegean Sea and released in the Tyrrhenian Sea. There are also pictorial remains that show fishing gear and a large variety of targeted species during Roman times. Gastropods, the red coral Corallium rubrum, and several species of sponges were exploited on an industrial scale. Fishing, fish processing, industrial exploitation of several marine species, and development of improved fishing gear continue during the Byzantine period. Various literary sources point out that targeted species, among them the currently overfished tuna, are conspicuous. There is a 200-year gap between the Moslem conquest of the Near East and northern Africa and the appearance in the ninth century of the first Arabic written sources. In northern Africa, the first written evidence dates from the tenth century and refers to fishing gear used to catch mullets, Atlantic bluefin tuna (with large spears), and fish in shallow waters. Zooarchaeological material from the Israeli coastline dating from the Byzantine through the Moslem Crusader and Mamluk periods (fourteenth century) points to a high consumption of marine and freshwater fish that are still fished in Israel today, such as the thin-lipped grey mullet (Liza ramada), Sparidae, and the parrot fish. There is noticeable fishing activity dating from the Byzantine, Moslem (tenth century), and later Norman periods (eleventh to thirteenth centuries) in southern Italy and in Sicily, where Atlantic bluefin tuna is the main target species exploited by traps (tonnara). Harvesting of the gastropods Hexaplex trunculus and Bolinus brandaris is an example of the successive exploitation of marine resources from the Iron Age until the thirteenth century in the Eastern Mediterranean. These species are specifically harvested for the purple pigments extracted from their shells and used to dye clothes. This harvest disappear from the Levantine area in the late twelfth century, and from Greece a century later, although both species are still abundant to this day. Another example of human exploitation of marine resources from historical times is the hunting of seabirds on islands, particularly of shearwaters, which probably constituted the only source of protein in periods of scarcity especially on small islands. In places such as Formentera (Balearic Islands), humans contribute to the depletion, and partial extinction, of Balearic shearwaters (Puffinus mauretanicus), with consequences at the level of the marine trophic web. Human impacts on marine biodiversity grow increasingly stronger as the Mediterranean cities and ports continue to grow and more recent centuries witnessed substantial advances in technology. It is assumed that since the fourteenth century, the adoption of new fishing methods (such as the tonnara, a sort of drift net mainly used for tuna fishing) in the Western Mediterranean, their spread to southern Italy, , and their introduction to the Adriatic in the seventeenth century, increase fishing catches. Fishing catches increase to an extent that even the early fishermen organizations (sixteenth century), such as Cofradias in Catalonia and the Prud'homies in Provence, are concerned about possible negative effects on exploited stocks. Such effects are further intensified by the increasing industrialization in the nineteenth century, with an increase in the efficiency of existing fishing gear (e.g., otter trawl) and the introduction of new ones (such as midwater pelagic trawls, hydraulic dredges, and iron-toothed dredges). Industrialized fishing had severe impacts on species, habitats, and ecosystems. Several studies also show historical changes in fish communities of different regions of the basin \[e.g.,–\]. These findings point to a general severe depletion of top predators in the basin, including Atlantic bluefin tuna, which is considered critically endangered according to the declining trend observed in the Atlantic and the Mediterranean in the last 50 years. Historical fluctuations in the abundance of this species have been described on the basis of a centuries-long time-series of tuna trap catches, starting in the seventeenth century, and suggested to be linked to climate fluctuations. Despite this comparative wealth of historic information about temporal trends mainly linked to the history of human exploitation of Mediterranean marine biodiversity, many unknowns remain in spatial and chronological gaps from prehistoric periods to the present. Ancient, medieval, and early modern records contain qualitative rather than quantitative data, and it is difficult to depict general diversity trends at either a species or ecosystem level at the scale of the whole Mediterranean. Interesting results do emerge from analyses of specific regions. The overall trends reported by Lotze et al. for the north Adriatic Sea indicated that prehistoric people had no measurable effect on marine resources around this basin (see for species included in the analysis). This changed during the Classical period (500 B.C. to A.D. 600), and especially during Roman times, when reports of species depletion and overexploitation in coastal waters increased. It is possible that marine species recovered from heavy exploitation after the collapse of the Roman Empire, as has been documented for terrestrial resources. However, human population increased during the Medieval period (approximately A.D. 600 to 1500), increasing the pressure on marine resources. With the onset of the industrialization in Europe in the nineteenth century, signs of species depletions and rareness increased and accelerated throughout the twentieth century, when the first extirpations of species were also recorded. Biodiversity did not decrease, however, because some species were newly introduced into the Adriatic Sea. No temporal trend is known for alien species in the Adriatic Sea, so we showed a timeline of mollusk invasions in the Mediterranean as a whole, which started in the late nineteenth century and accelerated during the twentieth century. The depletion of formerly abundant species and the invasion of new species caused a shift in species composition and diversity in the north Adriatic Sea. Local species depletions and extirpations mostly occurred among large species, including marine mammals, birds, reptiles, and commercial fish and invertebrates, while species invasions were mainly by smaller species at lower trophic levels, such as invertebrates and algae. Such fundamental changes in species composition had effects on the structure and functioning of food webs and ecosystems. Population declines have also been noted among marine mammals throughout the Mediterranean. These species include sperm whales, which have been declining since the end of the 1980s ; short-beaked common dolphins, which began to decline around the 1970s, ; common bottlenose dolphins, which have decreased by at least 30% over the past 60 years, ; and striped dolphins, which have been in decline since the early 1990s. The Mediterranean monk seal, in particular, was deliberately hunted during the Roman period, and it disappeared in the greatest part of the Mediterranean basin during the early 1900s. Currently, it mainly occurs in small, isolated areas of the Greek and Turkish coasts, and northwest African coastal waters, but the presence of Mediterranean monk seal in some of these areas is uncertain. There are fewer loggerhead and green turtles throughout the Mediterranean, although historical records were available to determine the severity of their population decline. Known nesting sites especially for the loggerhead turtle disappeared in several areas of the basin. Although the population trends for most seabird species are not well known, all reliable long-term information suggests that most seabird species have recovered on the European coasts during the last three decades. This recovery is due to more restrictive conservation policies at national and international levels. With the exception of shearwaters, seabird species show relatively stable population trends. Gulls and terns, after two decades (1980s and 1990s) of sharp increase in their densities (up to an average 13% annual growth rate in Audouin's gull), are now in dynamic equilibrium. Sparse data on shags suggest a slow recovery in the last two decades. Storm petrel populations are stable at the few long-term monitored sites, but many suitable breeding sites have been destroyed since historical times along coastlines. Paleontological records confirm that the distribution of many species was much larger, even occupying habitats in the interior of large islands relatively far from the sea, where recolonization is now impossible. Population recoveries of Mediterranean seabirds must be considered only partial, and only occurring where protection is effective. ## Threats to diversity and hot spots As shown above, anthropogenic factors have influenced the general patterns and temporal trends of Mediterranean marine diversity with varying degrees of intensity. Quantifying the importance of each threat is essential for future analysis. Lotze et al. provided data to evaluate the human impacts that caused or contributed to the depletion or extirpation of species in the north Adriatic Sea over historical time scales. Exploitation stood out as the most important factor causing or contributing to 93% of depletions and 100% of local extinctions or extirpations. Habitat loss or destruction was the second-most-important human impact, followed by eutrophication, introduced predators, disease, and general disturbance. While 64% of depletions and 88% of local extinctions were caused by a single human impact, in all other cases the combination of two or several human causes was responsible for the decline or loss. This highlights the importance of cumulative human impacts, especially in coastal ecosystems, with emphasis on species with commercial interest. Recently, anthropogenic drivers and threats to diversity increased and further diversified in the Mediterranean, as observed elsewhere. Published information and the opinion by experts identified and ranked current threats to diversity in the Mediterranean. The sum of the ranking (0–5 for each threat) showed that for 13 large taxonomic groups, habitat loss and degradation are considered the primary impact on diversity, followed by exploitation, pollution, climate change, eutrophication and species invasions. These were the most conspicuous threats and also affect the greatest number of taxonomic groups. Other threats to diversity were maritime traffic (collisions with vessels) and aquaculture. Within 10 years from now, habitat degradation and exploitation were predicted to retain the predominant roles, while pollution and climate change will likely increase in importance, followed by eutrophication. Of all current threats to biodiversity in the Mediterranean, climate change was predicted to show the largest growth in importance within the next 10 years (10.8%), followed by habitat degradation (9.2%), exploitation (6.2%), and pollution, eutrophication, and invasion of species (4.6% each). shows past changes and projected future increases in sea surface temperature (SST) in the Mediterranean Sea. The 15°C isotherm, whose one-century climatological mean crosses the Straits of Sicily, may have moved northward in recent times. This can imply that a number of tropical Atlantic species that entered the Mediterranean during the last interglacial (125,000 to 110,000 years ago) will reenter the Western Mediterranean in the near future. In the meantime, in the Western Mediterranean, the “14°C divide”, the one-century climatological mean of the surface isotherm for February that coincides with a frontal system created by mesoscale eddies in the Algerian Basin and that may act as a barrier to dispersal, has apparently moved northward in recent times. The southern sectors of the Mediterranean harbor many native warm-water species that do not occur or get much rarer in the northern sectors. These “southerners” are apparently confined by the 14°C divide. Perhaps not coincidentally, many of these native but “meridional” warm-water species have colonized the northern sectors, which are thus facing a process of “meridionalization” \[e.g.,\]. In addition, the mean SST made in early 1980s revealed that the warmest area of the Mediterranean was the Levantine Basin, with a mean SST of 21.8°C, and the coolest areas were the Gulf of Lions and the Ligurian Sea, with a mean SST of 16.9°C. Climate models predicted that by 2041–2060, the major part of the Mediterranean will become warmer except the northern Adriatic, which is expected to become cooler (OPAMED8 model based on the A2 IPCC scenario). By 2070–2099, the Mediterranean is projected to warm by 3.1°C, the last cool enclaves being the Gulf of Lions and the northern Adriatic, with a mean SST of 18°C. Taking into account data regarding marine biodiversity and threats, we mapped vertebrate endangered species and have tried to locate potential hot spot areas of special concern for conservation in the Mediterranean. The first attempt included fish, marine mammals, and sea turtles, which are considered important sentinels for ocean health. The identified hot spots highlighted the ecological importance of most of the western Mediterranean shelves. The Strait of Gibraltar and adjacent Alboran Sea and African coast were identified as representing important habitat for many threatened or endangered vertebrate species. The most threatened invertebrate species in the Mediterranean, the limpet Patella ferruginea, is also distributed along this area. Both the northern Adriatic and Aegean seas also showed concentrations of endangered, threatened, or vulnerable species. Other equally species-rich waters along the northeast African coast, and the southern Adriatic Sea, were of lesser concern for the protection of endangered species. # Discussion ## Estimates and patterns of marine diversity in the Mediterranean Sea Our estimate of 17,000 species for marine biodiversity in the Mediterranean updated and exceeded previous values, which were on the order of 8,000–12,000 species. In comparison with the 1992 estimate, the total number of recorded species has increased substantially. As a result of recent efforts and improvements in analytical methods and instruments, our estimates of invertebrates and protists, in particular, have undergone an upward revision in recent years. Current estimates of sponges, cnidarians, polychaetes, mollusks, arthropods, echinoderms, ascidians, and other invertebrates all exceed those dating back to the early 1990s. However, since most microbial diversity is basically unknown, global numbers and their evolution are uncertain. Estimates from global databases that include Mediterranean information up to September 2009 range from 4% and 25% of the total species diversity estimated in our study. They covered vertebrate taxa fairly comprehensively, but other taxonomic groups were underrepresented. WoRMS included 8,562 records of Mediterranean marine species, which represented 50% of species registered in this study. Mediterranean databases such as ICTIMED (specialized in fish diversity) included about 70% of fish diversity reported in our study. Total estimates of Mediterranean species of macrophytes and metazoans represented 6.4% of their global counterpart. Macrophytes showed the highest percentage of shared species with global estimates, and Heterokontophyta and Magnoliophyta scored the highest (17.2% and 11.7%, respectively). Among metazoans, Mediterranean sponges showed the highest percentage (12.4%), followed by polychaetes (9.4%) and cnidarians (7.7%). Other groups represented much lower percentages of the total, such as echinoderms (2.2%), fish species (4%), and mollusks (4%). Previous studies claim the existence of a gradient of species richness from the northwest to the southeast Mediterranean \[e.g.,–\], in agreement with differences in key environmental variables, such as latitude, salinity, temperature, and water circulation, in addition to the distance from the Strait of Gibraltar. Our results confirmed this general decreasing trend and showed that the distribution of marine diversity in the Mediterranean is highly heterogeneous. The Western Mediterranean displays the highest values of species richness, likely owing to the influx of Atlantic species and the wide range of physicochemical conditions. The central Mediterranean, Adriatic, and Aegean seas are areas of second-highest species richness, although with exceptions. The Adriatic Sea sometimes displays lower species numbers because of restricted exchange with the western basin, decreasing depth toward the north, the presence of fresh water, and the larger amplitude of temperature variations. However, this basin shows a large number of endemics possibly owing to its higher isolation. The Aegean Sea normally follows the western areas, mainly because of its more direct exchange with the western basin and its higher habitat diversity. The Levantine Basin and southeastern side have in general the lowest species richness, which is due to the unfavorable conditions prevailing in the area (such as high salinity) as well as the less intensive sampling effort. In fact, a lack of data is evident in several eastern and southern regions of the Mediterranean basin. This may have strongly influenced some of our results regarding spatial patterns, so generalizations have to be made carefully. Marine research in the Mediterranean has been regionally biased, reflecting sparse efforts along the southern and easternmost rim. It has even been suggested that the relative species richness of different taxa by sector of the Mediterranean is a better indicator of the level of research effort than of true species richness. Therefore, as new species are assessed in the eastern and southern areas, patterns may be modified. Moreover, the diversity in the eastern end is more influenced by species introductions. The Suez Canal, opened in 1869, has restored the connection between the Mediterranean and the Indian Ocean, and in recent years we have witnessed an exponential increment in the number of Indo- Pacific species recorded in the Eastern Mediterranean \[e.g.,\]. This trend will continue to influence the biodiversity of the Mediterranean Sea. In addition, the data used to draw spatial patterns were collected from the 1980s to 2000s, so results may differ from the current situation and may represent potential ranges and values rather than current ones. However, similarities exist between results achieved with distribution maps drawn with expert data and predicted results using AquaMaps models. These similarities indicated that the species richness maps resulting from this study are a useful first attempt to represent comprehensive species richness patterns at the Mediterranean scale. Differences encountered using both methods may be due to limitations of the data. By their nature, expert-drawn maps or sightings often represent underestimates of total species distributions because of the absence or lack of effort in certain areas (in our case the southern shorelines of Mediterranean along the coasts of northern Africa and the eastern sites) and the inability to detect rarer species without sufficient efforts. On the other side, AquaMaps model predictions do not currently factor human impacts or ecological interactions and may be closer to fundamental or historical niche rather than realized niche. Therefore some AquaMaps predictions may represent overestimates (a good example is the Mediterranean monk seal; see [www.aquamaps.org](http://www.aquamaps.org)). Besides, the relative probability of occurrence calculated from AquaMaps does not distinguish between a rare species that might only have been sighted once in a given cell, and a more abundant species that might be sighted every day. AquaMaps rely exclusively on data accessible through OBIS/GBIF, which currently contains few Mediterranean records. Therefore, for many species, occurrence was inferred from habitat use outside of the Mediterranean. Because the Mediterranean environment represents some environmental extremes (such as salinity and temperature records), occurrences in the eastern part may not have been captured adequately by AquaMaps, and this could partially explain the low values in this region. These limitations are extended to our first attempt to depict hot spot areas in the Mediterranean. The eastern region hosts important populations of elasmobranchs and marine mammals that are currently threatened, but their probability of occurrence estimated by AquaMaps model is lower than 0.4. Further studies should be able to reconcile both mapping sources and confirm or correct patterns. Explanations for the observed heterogeneity of species richness in the Mediterranean Sea include the threshold of the Siculo-Tunisian Strait that divides the Mediterranean into two basins, and the paleo-biogeographical history of the Mediterranean Sea. The western basin shows more biological similarity with the Atlantic Ocean, hosting a higher number of cold-temperate species, while the eastern basin shows more biological similarities with the Indo- Pacific, and hosts a larger number of subtropical species. The Siculo-Tunisian Strait still partially acts as a barrier to the dispersal of many species between the two basins and constitutes their meeting point. Diversity differences between areas may also reflect changes in water masses and circulation, as well as changes in temperature and salinity. The diversity of some groups is definitively influenced by this temperature gradient. For the sipunculans, richness may be linked to the temperature of the water masses during the year, which reflects a physiological barrier between cold and warm water for cold- and warm-water species. For example, Golfingia margaritacea is mainly a temperate and boreal species, and its presence in the Mediterranean may indicate the prevalence of colder water masses. In contrast, other thermophilic species, such as Phascolion convestitum and Aspidosiphon elegans, have been proposed as Lessepsian migrants. Diversity distribution in the Mediterranean is also associated with a productivity gradient. Higher productivity areas show higher diversity partially because they are important feeding and reproductive sites for several taxa. Most of these areas occur in the Western Mediterranean and the northern Adriatic that, for example, host many species of fish, seabirds, marine mammals, and turtles \[e.g.,\]. Their distribution is associated with feeding habits \[e.g.,\]. Moreover, some fish, seabirds, sea turtles, and mammals show opportunistic feeding behavior, exploiting discards from trawling and purse seines, and to a lesser extent from artisanal long-lining \[e.g., –\]. In developed Mediterranean countries, discards from trawl fishing can be up to 400% of the commercially valuable catches, and such amounts of food, which may be predictable in space and time, are scavenged by many species. Most Mediterranean marine mammals are predominantly offshore and prefer deep-water habitats, but a few species can venture to inshore waters and scavenge fishery discards. The three main categories explaining the drivers of biodiversity in the deep Mediterranean are (i) bathymetric gradients, which are associated with increasing pressure and decreasing food availability in deeper sediments; (ii) geographical and physicochemical features, which are responsible for the north- northwest–south-southeast gradient in trophic conditions; and (iii) environmental heterogeneity (e.g., grain size distribution, habitat complexity, distribution of food inputs). Our understanding of the mechanisms driving deep- sea biodiversity patterns is still limited, but some of the factors frequently invoked are (a) sediment grain size and substrate heterogeneity ; (b) productivity, organic content, or microbial activity ; (c) food resources ; (d) oxygen availability ; (e) water currents ; and (f) occasional catastrophic disturbances. Thus, the spatial distribution of available energy may influence the distribution of benthic abundance, biomass, and biodiversity,. Food availability depends almost entirely on the supply of energy from the water column and decreases with depth, which may explain most of the variability between the observed spatial patterns of the benthic biodiversity in the deep Mediterranean Sea. ## Threats to diversity In the past, geological and physical changes lie at the root of the most dramatic changes in biodiversity in the Mediterranean Sea. Today, human activities are essential elements to consider as well, and several of them threaten marine diversity. The most important threats in this region are habitat loss, degradation and pollution, overexploitation of marine resources, invasion of species, and climate change. ### Habitat degradation, pollution, and eutrophication Our results show that habitat degradation and loss is currently the most widespread threat and was also important in the past. Human interventions, such as coastal modification, that can be traced back to before the Roman period, have important consequences for diversity. Coastal development, sediment loading, and pollution reduced the extent of important habitats for marine diversity, such as seagrass meadows, oyster reefs, maërl, and macroalgal beds, and affected Mediterranean ecosystem functioning well before the 1900s. Most species depend strongly on their habitats (such as bryozoans, sponges, echinoderms, benthic decapods, and organisms of the suprabenthos and meiobenthos); hence, its loss and degradation have major effects on marine diversity. Cultural eutrophication, in particular in semienclosed basins such as the Adriatic Sea, can also be traced back for centuries. This phenomenon reached its peak in the late 1980s and, in addition to fishing, may be the cause of the sequence of jellyfish outbreaks, red tides, bottom anoxia events leading to benthic mass mortalities, and mucilage events that have occurred in recent ecological history of the Adriatic Sea. Direct and indirect pollution is generated directly from the coast, or through fluvial contributions, and ends up in the sea. Pollution affects a wide range of marine species \[e.g.,–\] and is of primary concern for the conservation of the deep-sea ecosystems. The main threats for most seabirds and marine turtles in the Mediterranean arise from habitat degradation and loss. The breeding habitat for seabirds is relatively well protected along the northern Mediterranean shore, but the protection of many seabird colonies and hot spots is less effective along the southern shore because of limited resources. Marine wind farms, which are expected to increase in some countries, may represent a new conservation concern for seabird populations. Marine turtles are also affected primarily by degradation of habitats but also by marine pollution, driftnets, gillnet and longline by-catches, and boat strikes. The continuing increase of coastal settlements is important for the region's economic activity, but it is also causing intense environmental degradation through excessive coastal development, further pollution, and consumption of natural resources, all of which add pressure to coastal areas and the marine environment. ### Exploitation of marine species This study also illustrates that the oldest and one of the most important maritime activities that has become a threat to diversity is human exploitation of marine resources. People around the Mediterranean have exploited marine resources since earliest times. Maybe not surprisingly, negative effects of the exploitation of the Mediterranean marine biodiversity were first reported in the fourth century B.C. by Aristotle. He mentioned that scallops had vanished from their main fishing ground (Gulf of Kalloni, in Lesvos Island) since fishermen began using an instrument that scratched the bottom of the sea. Early records of overfishing and depletion of coastal resources become evident during Roman and medieval times and are driven by human population growth and increasing demand and the increasing commercialization and trade of food and products. The current high demand for marine resources continues and has resulted in high levels of fishing or harvesting intensity. Several fish resources are highly exploited or overexploited \[e.g.,–\]. Other organisms that are exploited or affected by exploitation in the Mediterranean include macrophytes, sponges, cnidarians, echinoderms, mollusks, arthropods, polychaetes, ascidians, and other invertebrates \[e.g.,–\]. The threats to currently endangered marine mammals and sea turtles include unwanted by-catch, as well as historical exploitation. For sea turtles, the overall mortality rate caused by entanglement in fishing gear and by habitat degradation is poorly known, but for marine mammals the major threats clearly derive from human activities: direct or indirect effects of exploitation, such as prey depletion, direct killing, and fishery by-catch,. At sea, threats to seabirds mainly come from fisheries, particularly by-catch in longlining. Fishing is being expanded toward deeper areas and is threatening several ecosystems \[e.g.,\], while management effectiveness in the Mediterranean is low. Fishing activity may also be the cause of ecosystem structural and functional changes and ecosystem degradation \[e.g.,–\]. ### Bioinvasions A few Mediterranean invasive aliens have drawn the attention of scientists, managers, and media for the conspicuous impacts on the native biota attributed to them. A pair of coenocytic chlorophytes, Caulerpa taxifolia and Caulerpa racemosa var. cylindracea, are the most notorious invaders due to their high impact on marine benthic ecosystems, thus the best-studied invasive species in the Mediterranean. Other work has traced the impacts of invasive aliens that entered the Mediterranean from the Red Sea through the Suez Canal and displaced native species. Tropical species have been entering the Mediterranean through either the Suez Canal (Lessepsian migration) or the Strait of Gibraltar for decades, and mainly by ship transportation. The Mediterranean is highly susceptible to ship- transported bioinvasions: one-fifth of the alien species recorded in the Mediterranean were first introduced by vessels. In 2006, 13,000 merchant vessels made 252,000 calls at Mediterranean ports, and an additional 10,000 vessels passed through the sea (REMPEC/WG.29/INF.9). The increase in shipping-related invasions may be attributed to the increase in shipping volume throughout the region, changing trade patterns that result in new shipping routes, improved water quality in port environments, augmented opportunities for overlap with other introduction vectors, and increasing awareness and research effort. The swarms of the vessel-transported American comb jelly (Mnemiopsis leidyi) that spread across the Mediterranean from Israel to Spain in 2009 raise great concern because of their notorious impacts on the ecosystem and fisheries \[ansamed.info and 360\]. Moreover, with the development of large-scale marine aquaculture (mariculture) in the late twentieth century, the commercially important alien shellfish Crassostrea gigas and Ruditapes philippinarum were intentionally introduced to the Mediterranean. The high permeability of aquaculture facilities, transport, and transplantation of these species have resulted in many unintentional introductions: oyster farms have become veritable gateways into Mediterranean coastal waters for alien macrophytes. The massive “official” and “unofficial” importation of foreign spat (young bivalves both before and after they become adherent) in the 1970s and 1980s coincided with a marked increase of alien species around oyster farms, and the aliens were considered to have arrived with the oysters. Segments of the industry may still resort to illegal importation: neither the Turkish authorities nor the UN Food and Agricultural Organization were aware of the importation of the bilaterally ablated female banana prawn (Fenneropenaeus merguiensis) that was found in the Bay of Iskenderun, Turkey. Although some aliens are responsible for reducing the population of some native species, others have become locally valuable fishery resources. Some Erythrean aliens were exploited commercially almost as soon as they entered the Levantine Sea, and their economic importance was quickly acknowledged. Levantine fisheries statistics record the growing prominence of the Erythrean aliens: the Erythrean prawns are highly prized and, beginning in the 1970s, a shrimp fishery developed in the Levantine Sea. Nearly half of the trawl catches along the Levantine coast consist of Erythrean fish, but the commercially exploitable species were accompanied each summer by swarms of the scyphozoan jellyfish Rhopilema nomadica, washed ashore along the Levantine coast. The shoals of jellyfish adversely affect tourism, fisheries, and coastal installations, and severe jellyfish envenomations require hospitalization. The recent spread of the silver stripe blaasop (Lagocephalus sceleratus) and the striped catfish (Plotosus lineatus) pose severe health hazards. Other work has traced the impacts of invasive aliens that entered the Mediterranean from the Red Sea through the Suez Canal and displaced native species. Pronounced thermal fluctuations and a significant increase in the average temperature of the waters in the Mediterranean during the past two decades have coincided with an enlarged pool of warm-water alien species that have become established and expanded their distributions (see next section). These thermophilic aliens have a distinct advantage over the native Mediterranean biota. Though no extinction of a native species is yet attributable to invasion of new species, sudden declines in abundance, concurrent with proliferation of aliens, have been recorded. Examination of the profound ecological impacts of some of the most conspicuous invasive alien species underscores their role, among many anthropogenic stressors, in altering the infralittoral benthic communities. Local population losses and niche contraction of native species may not induce immediate extirpation, but they may trigger reduction of genetic diversity and loss of ecosystem functions and processes, and habitat structure. ### Impacts of climate change Climate change is exerting a major effect on Mediterranean marine biodiversity through seawater warming \[e.g., –\]. The increase in seawater temperature has affected the distribution and abundance of native and alien species, and has had both direct and indirect effects on invertebrates and fish \[e.g.,, see \]. The increase in water temperature in the Mediterranean also alters jellyfish population dynamics \[e.g., \] and may act in addition to indirect fishing impacts \[e.g., \]. Seawater of the Mediterranean Sea has been warming since at least the 1970s. Rising temperature enlarges the pool of alien species that could establish themselves, enables the warm-water species (native and alien) present in the sea to expand beyond their present distributions, and provides the thermophilic aliens with a distinct advantage over the native Mediterranean biota. The appearance of numerous allochthonous species of tropical origin is leading to what is called the “tropicalization” of the Mediterranean Sea. Although tropical invaders have been recorded in the northernmost sectors of the Mediterranean \[e.g.,\], tropicalization is especially obvious in the southern sectors, where species of tropical origin now form a significant portion of the biota. Tropical species have been entering the Mediterranean through either the Suez Canal (Lessepsian migration) or the Strait of Gibraltar for decades , but they used to remain in the eastern or western basin, respectively. Thus it conformed to the traditional physiographic and biogeographic subdivision of the Mediterranean. However, in the last two decades, the number of tropical species that have also spread through the entire basin is growing. Examples of Erythrean aliens that crossed the Strait of Sicily include algae, a seagrass, many invertebrates and fish \[e.g.,–\]. Species coming from the tropical Atlantic have traveled the opposite way to reach the Levantine Sea \[e.g.,\]. The Strait of Sicily is today a crossroad for species of distinct tropical origins (Atlantic and Indo-Pacific), expanding their range longitudinally within the Mediterranean. If the southern sectors of the Mediterranean are being “tropicalized” (higher occurrence of tropical aliens) and the northern sectors “meridionalized” (increased proportion of indigenous thermophilic species), it is uncertain what will happen to those species of boreo-Atlantic origin, which entered the Mediterranean during glacial periods and have been established in the northern and colder areas of the basin. Because they cannot move farther northward, they may dramatically decrease or even be at risk of extirpation. Although the total extinction of flora and fauna from a basin as wide as the Mediterranean may be unrealistic, the signs of increased rarity or even disappearance of cold-water species deserve further investigation,. An example is the deep-water white coral, Lophelia pertusa, reefs of which have become rare in the Mediterranean. These coldest parts of the Mediterranean (Gulf of Lions, northern Adriatic) could act as a sanctuary for cold-temperate species, but if warming intensifies, those areas may act as traps without any cooler water for escape. Global warming may cause thermophilic species of the southern Mediterranean to appear more frequently in the northern and colder parts \[e.g.,–\], and an increasing colonization by southern exotic species may be seen. But there may also be habitat fragmentation and local extinction of species unable to undertake migrations. Lack of (evidence of) species extinctions, coupled with establishment of alien species, is apparently leading to an increased species richness of the Mediterranean, a much debated issue. Richness is increasing at the whole-basin scale (γ-diversity), but it is difficult to establish what is happening at local scales (α-diversity) in coastal areas. Instances of species replacement \[e.g.,–,\], and mass mortalities due to high temperature or pathogens \[e.g.,–\] and perhaps aliens have been observed. Climate warming, moving physiological barriers and inducing the spatial overlap between alien and indigenous species, causes biotic homogenization and hence a depression in β-diversity. Thus, the relationship between tropicalization, meridionalization, and biodiversity is not straightforward. In general, the establishment of tropical invasive aliens may cause Mediterranean communities to lose their particular character and to become similar to their tropical analogs, especially in the southern portions of the basin. Cladocora caespitosa, the most important shallow-water zooxanthellate species living in the Mediterranean, was more abundant and built more conspicuous formations during periods of the Quaternary, when the Mediterranean climate was subtropical. However, warming episodes in recent summers coincided with mass-mortality events of this coral \[e.g., \]. Hence, it is unlikely that the Mediterranean in the future will contain significant coral constructions. The overwhelming number of Lessepsian immigrants will move the composition of the biota more and more like that of the Red Sea, but Mediterranean communities will probably look like those that today characterize southern Macaronesia and the Cape Verde region, with scanty coral and abundant algae \[e.g., \], rather than those of the Red Sea and the Indo-Pacific. Seawater acidification may also be a threat to Mediterranean marine biodiversity. The most obvious consequence of the increased concentration of CO<sub>2</sub> in seawater is a reduced rate of biogenic calcification in marine organisms. This could affect both planktonic and benthic communities. Calcifying phytoplankton (coccolithophores) play a significant role in the primary productivity of the oligotrophic Mediterranean Sea, whereas many benthic habitats are engineered by sessile organisms that lay down carbonate crusts. Calcareous red algae are the builders of coralligenous reefs, one of the most important Mediterranean ecosystems, and seawater acidification will probably impair their role. However, noncalcifying photosynthetic plants, such as frondose algae and seagrasses, may take advantage of a greater availability of CO<sub>2</sub>. But large, erect species of brown algae as well as Mediterranean seagrass are now in decline because of the environmental degradation, induced primarily by human activities. ## The unknowns and limitations The study of Mediterranean marine diversity over many years has produced a significant amount of information. Yet this information remains incomplete with the discovery and description of new species, especially of smaller, less conspicuous and cryptic biota. The biodiversity in the Mediterranean Sea may be in fact much higher than is currently known. We do not have credible measures of microbial richness, but development of new technologies will allow us to decide whether this is knowable or not. The description of microbial diversity is probably better approached through the continued study at selected sites, such as the Microbial Observatories, for which data exist on both identification methodologies and the functioning of the ecosystem. Current Mediterranean observatories are at Blanes Bay, Gulf of Naples, Villefranche's Point B, Dyfamed station, and the Mola and SOLA stations in Banyuls. Sites in the southern and eastern Mediterranean are still to be added. Further exploration and taxonomic work on seaweeds and seagrasses is needed in all the African countries (mainly in Libya and Egypt), the Levantine Sea (Israel, Lebanon, Cyprus, Syria), and the Aegean Sea (Greece and Turkey). Phycological surveys are also required in Croatia, because several species (and even genera) described from the Adriatic have never been found again and require taxonomic reevaluation. We do not expect a significant increase in the rate of description of new species, but the description of new macroalgal species continues \[e.g., ,\]. A large number of species are poorly known, and our checklist includes several taxa inquirenda. Accurate morphological studies, and new molecular tools, are required to decipher the taxonomy of several genera, including Ectocarpus, Cystoseira, Acrochaetium, [Polysiphonia](http://www.eol.org/pages/9653), and Ulva. A similar situation exists for the invertebrates. Most of the small fauna of the Mediterranean are typical of current scientific knowledge: in one of the best- known geographic areas of the world, there are many regions and habitats that remain insufficiently studied, and several taxonomic groups in deep-sea areas and portions of the southern region are still poorly known. The description of new species is still a high priority. As illustrative examples, the accumulation curves for cumaceans, mysids, polychaetes, and ascidians discovered (described or first recorded) show that no asymptote has been reached, and there has been no slowing in the rate of discovery for less conspicuous species in the Mediterranean, as it is observed when analyzing accumulation curves in other parts of the world. The shortage of taxonomists for many groups is a particularly serious problem worldwide, and it also applies to the Mediterranean Sea. Several of the main invertebrate specialists have retired or are close to retirement and few are being replaced. Many samples are not being properly identified, which leads to a corresponding underestimation of biodiversity. The current spread of invasive species requires serious taxonomic attention. Many, if not most, taxonomic groups are subject to anthropogenic threats in one way or another, and researchers must work against time to avoid losing valuable biological information. Undescribed invertebrate species may become extinct before we even know of their existence. In addition, and paradoxically, some of the commonest and most accessible ecosystems such as beaches, among other habitats in the Mediterranean, have been poorly studied,. Sampling biases are another source of uncertainty in the estimation of marine biodiversity. In particular, the three-dimensional character of marine ecosystems requires much more study at depths where light penetration is perceived as important but is poorly understood. Light intensity decreases with increasing depth and species perform extensive migrations within the water column or along the seabed. Endobenthic species display rhythms of emergence, including burying or burrowing within the substrate and sheltering in natural holes. Marine species react to light intensity cycles, which may include movements in and out of our sampling windows. Information gathered without attention to such rhythmicity will affect perceived population distribution, biomass, and estimated biodiversity. These issues have been integral to land ecology since the early twentieth century but have been rarely considered in the marine environment. In the Mediterranean, Sardà et al. considered this problem during day-night sampling at and below the end of the twilight zone (1,000 m depth) and observed day-night fluctuations in their catches. Midday and midnight trawl catches at different depths during October showed great differences in fish, cephalopod, and crustacean species composition and relative abundance in the deeper areas. Waveform analysis of crustacean catches showed behavioral rhythms that affected presence or absence from catches made at different times during a 24-hour cycle. Because trawl surveying is one of the commonest methods of sampling in marine waters, and is one of the most used in the Mediterranean Sea, future biodiversity studies should correct for the practice of sampling only during daytime. In addition, observations of important diel variation in the fauna associated with seagrasses include a notable increase of species richness and abundance in nighttime samples. This issue brings together the problem of biodiversity and climate change due to expected changes in species migrations and rhythmicity. While Mediterranean vertebrate species are better known than the invertebrates, our understanding is still incomplete and often outdated. The FNAM atlas, which contains data collected and edited during the 1980s and 1990s, is based on regional data and expert knowledge and is the only record of geographic ranges for all Mediterranean fish species. Several areas of the southern Mediterranean have never been surveyed scientifically. Long-term monitoring programs are absent or unavailable for many countries. Since vertebrate species may be useful indicators of changes in ocean food webs, a major challenge that remains is to achieve time-series sampling of species diversity, abundance, and habitat data. These time series should have large spatial and temporal scales to develop useful indicators of changes in Mediterranean marine ecosystems and provide measures of ecological connections and ecosystem services. A clearer identification of hot spot areas will require the inclusion of new data on macroalgae and seagrasses, invertebrates, and seabirds. Most of the Mediterranean seabird species (with the exception of some large gulls) are protected by European laws because of their small or declining populations or the small number of breeding sites. Nine species are included in Annex II of the EU list of endangered or threatened species. The Balearic shearwater is critically endangered, and the monitored colonies of Cory's and Mediterranean shearwaters are slowly declining. Although information is incomplete for macroalgae and invertebrates, a total of 11 species of macroalgae, 3 of flowering plants, 9 of sponges, 3 of cnidarians, 17 of mollusks, 2 of crustacean decapods, and 3 of echinoderms are now listed as endangered or threatened in the Annex II of the Barcelona Convention for the Protection of the Marine Environment and the Coastal Region of the Mediterranean (1995). A recent proposal (2009) for amendments in the annex II increased to four the number of flowering plants and to 16 plus all the species of the genus Cystoseira (with the exception of C. compressa) the number of endangered species of macroalgae. ## Conclusions The Mediterranean Sea is a region of high biodiversity that ranks among the best known in the world, although much work remains to be done. The description of new species, especially of invertebrates and protists, undergoes upward revision, and new discoveries continually modify previous estimates. Increased efforts are required in taxonomy and sampling of poorly known ecosystems and on long-term monitoring programs of species and habitats. The invasion of alien species will continue to change the biodiversity of the Mediterranean Sea and requires continuous monitoring. The first attempt to integrate the spatial data and temporal trends presented here enables one to visualize macroecological patterns at the Mediterranean scale. These results depict a region of high diversity and heterogeneity, but they also evidence the need for further study of geographical areas that are largely unexplored, mainly the African coasts and certain zones of the southeastern basin and the deep sea. Our study illustrates that the Mediterranean is a complex region where ecological and human influences meet and strongly interact, posing a large and growing potential impact to marine biodiversity. Although much is known about individual threats, knowledge is very limited about how multiple impacts will interact. Therefore, there is the need to develop comprehensive analysis of conservation and management initiatives to preserve Mediterranean biodiversity. Owing to the Mediterranean physically, ecologically, and socioeconomically steep gradients, this region may be seen as a model of the world's oceans and a suitable laboratory to study marine ecosystems and decipher future trends. In addition to further sampling and taxonomic efforts, much of what remains to be done requires free distribution of publicly available data from national and regional research initiatives. This will facilitate database updates and enable scientific discussion. Marine surveys are not always accessible at the regional level and, when available, data coverage is often incomplete. Regional initiatives (such as MedObis) provide promising platforms for the integration of efforts devoted to marine biodiversity within the Mediterranean region, but they must be kept up to date. Individual and collaborative research efforts must continue to advance our knowledge of marine biodiversity in the Mediterranean Sea and narrow down the unknowns. # Supporting Information The authors gratefully acknowledge the support given by the European Census of Marine Life, the Census of Marine Life, and the Fisheries Centre (University of British Columbia, Canada). They thank Charles Griffiths, Michele DuRand, Dale Langford, and Iain Taylor for edition and correction of the paper. [^1]: The authors have declared that no competing interests exist.
# Introduction The organisms in the euryarchaeal order Halobacteriales are generally extreme halophiles requiring at least 1.5 M salt and growing optimally at 3.5–4.5 M salt, although some have recently been found to grow at lower salt concentrations. The haloarchaea are found in the water and sediment of salt lakes and salterns, and also in saline soils. Their mechanism of adaptation to high salinity involves the accumulation of molar concentrations of KCl in the cytoplasm and the production of proteins with a higher number of negative charges than in other organisms. Bacteria of the order Halanaerobiales and the genus Salinibacter also accumulate KCl internally. Other halophilic organisms accumulate compatible solutes such as glycerol or glycine betaine to counter high external salt concentrations, but this is energetically more costly than accumulation of KCl. The haloarchaea are heterotrophs, growing with amino acids and/or carbohydrates as carbon and energy sources. They are either aerobic or facultatively anaerobic using various electron acceptors. Glycerol is a particularly important nutrient for haloarchaea as it is produced by eukaryotic algae in high-salt environments. Dihydroxyacetone, produced by Salinibacter ruber, can also be present at high concentrations. While some haloarchaea are capable of growth on a wide range of compounds, others are very limited in their metabolism. The most extreme example of this is Haloquadratum walsbyi, which has only been found to grow well on pyruvate or dihydroxyacetone. Six haloarchaeal genomes have been sequenced previously, and analyses of the genomes have been published for Halobacterium salinarum NRC-1, Halobacterium salinarum R1, Haloarcula marismortui, Natronomonas pharaonis, Haloquadratum walsbyi, and Haloferax volcanii. Since the two H. salinarum genomes are very similar, we included only strain NRC-1 in the analysis. Here we present an analysis of these five genomes along with five new genomes, four of which (Halogeometricum borinquense, Halomicrobium mukohataei, Halorhabdus utahensis and Haloterrigena turkmenica) were sequenced as part of the Genomic Encyclopedia of Bacteria and Archaea (GEBA) project. The remaining genome, Halorubrum lacusprofundi, was sequenced as part of a Joint Genome Institute Community Sequencing Program project to sequence diverse archaeal genomes. With this work we are doubling the amount of genomic information from these extremophilic organisms and derive novel information and conclusions about the breadth of their metabolic capabilities. # Methods ## Genome sequencing Sequencing and annotation of H. borinquense, H. mukohataei, H. utahensis, and H. turkmenica have been previously described. A publication describing the sequencing and annotation of H. lacusprofundi is in preparation. Genome sequences are available from GenBank. Genome analysis was carried out within the Integrated Microbial Genomes Expert Review (IMG-ER) system. ## Sequence data used in phylogenetic analysis Protein sequences from 19 Halobacteria and outgroup (Methanomicrobia) genomes were retrieved from the IMG website (<http://img.jgi.doe.gov/>) or from NCBI (Methanocella paludicola SANAE; NC_013665). Methanomicrobia appear as sister group of Halobacteriaceae in a recent comprehensive 16S rRNA tree. Accession numbers for the genomes used in this study are listed in. ## Orthologs and alignments All-against-all protein BLAST was performed using mpiBLAST version 1.5 (<http://www.mpiblast.org/>), a parallel implementation of NCBI BLAST, using soft masking instead of complexity filtering. To determine orthologs, BLAST e-values were transformed using our own re-implementation of the OrthoMCL algorithm in conjunction with MCL version 08-312 (<http://micans.org/mcl/>) using an inflation parameter of 3.0 (slightly deviating from the default, 2.0, thus yielding slightly more clusters). The transformation of the e-values corrects for genome-specific biases as, e.g., caused by a GC bias ; after transformation, the BLAST results are reduced to the reciprocal best hits for each pair of genomes, which are then clustered using the MCL algorithm. OrthoMCL clusters containing inparalogs were reduced by selecting the most ‘central’ of several sequences from the same genome, that is, the sequence with the highest sum of within-cluster BLAST scores. The reduced OrthoMCL clusters were aligned using MUSCLE version 3.7 under default settings. The program scan_orphanerrs from the RASCAL package version 1.3.4 was applied to detect orphan sequences (overall poorly aligned genes) within the alignments. After removal of orphan sequences (if present), poorly aligned columns and divergent regions were eliminated with GBLOCKS version 0.91b using a minimum block length of two amino acids and allowing gap positions in all sequences. Each OrthoMCL cluster was also assigned to a COG category using a majority-rule approach. Above-mentioned parameter settings for ortholog determination and alignment filtering had previously been optimized using a genome-scale dataset for Actinobacteria type strains (unpublished data). During this optimization, it also turned out that modifying the program parameters had comparatively little influence on the phylogenetic outcome. ## Phylogenetic inference Filtered OrthoMCL cluster alignments comprising at least four sequences were concatenated to form a supermatrix for phylogenetic analysis. Maximum likelihood (ML) phylogenetic trees were inferred from the supermatrix with the Pthreads- parallelized RAxML package version 7.2.5, applying fast bootstrapping with subsequent search for the best tree, the autoMRE bootstopping criterion and the LG model of amino acid evolution in conjunction with gamma-distributed substitution rates and empirical amino acid frequencies (F). Among all amino acid models implemented in RAxML (except GTR, which was rejected for performance reasons), LGF was the empirically preferred one, as it produced the highest likelihood if optimized on a RAxML parsimony starting tree. Tree searches under the maximum parsimony (MP) criterion were conducted with PAUP\* version 4b10 using 100 rounds of random sequence addition and subsequent TBR branch swapping, saving no more than 10 best trees per round and collapsing potential zero-length branches during tree search. MP bootstrap support was calculated with PAUP\* using 1,000 replicates with 10 rounds of heuristic search per replicate. ## Assessing incongruence between gene trees and species tree After reducing the OrthoMCL cluster alignments to ingroup sequences, congruence between gene trees and the species tree was assessed by calculating partitioned Bremer support (PBS), for each OrthoMCL cluster using the newick.tcl script in conjunction with PAUP\. The Bremer support value for a certain bipartition (split) in the tree topology is the difference between the score (number of steps) of the best-known MP tree in which this split does not occur (forced using a converse constraint) and the score of the unconstrained best-known MP tree. While the total Bremer support must be positive, PBS values may be positive, zero, or negative, indicating that the OrthoMCL cluster supports the split, contains no information, or prefers another topology, respectively. As the total PBS (summed over all branches) for each cluster positively correlated with both its number of genes and its number of informative characters, the residuals from a linear regression with these two independent variables were determined. The residuals were used to determine their correlation with the COG categories and classes and the clusters most in conflict with the species tree (i.e., those with the most negative total PBS). ## Spectral clustering Protein sequences for the ten halophiles were downloaded from IMG-ER. We applied a spectral clustering procedure, for discriminating between groups of homologous proteins. The proteins are represented as nodes in a connected undirected graph with edges that carry weights based on node-to-node similarity according to the protein identity. The clustering procedure is analogous to a random walk of a particle moving on this graph from one node to another. In each node the particle moves to another node based on the probabilities corresponding to the weights of the edges. The amount of time the particle spends in a given subgraph will determine whether this is indeed a cluster of its own or not. After infinite time, the distinction between these subgraphs will become more clear, and to model this we calculated the normalized transition matrix at equilibrium. The eigenvalues of the transition matrix provide a measure of how distinct (or entwined) subgraphs are in relation to each other. An eigenvalue (or strength of partition) of 1 suggests a complete distinction between two subgraphs, while an eigenvalue of 0 suggests that no further partitions can be made. The eigenvectors are ordered by their eigenvalues in descending order, and the graph is successively partitioned while the second eigenvalue is greater than 0.8. As a result, a graph is partitioned into one or more subgraphs until the distinction between subgraphs becomes less clear. In this work, the choice of 0.8 as a cutoff suggests a balance between partitioning only fully non- homologous proteins and allowing proteins with weak similarities to be separated into subgraphs. ## Proteases and glycosyl hydrolases The heat maps and clustering of secreted proteases and glycosyl hydrolases were generated using hierarchical clustering with the program Cluster and visualized with TreeView (<http://rana.lbl.gov/EisenSoftware.htm>). Glycosyl hydrolases were identified based on Pfam domains. Proteases were obtained from MEROPS. Non- peptidase homologues were eliminated from the analysis. Signal peptides were identified with SignalP. # Results ## Orthologs and alignments The number of aligned and filtered OrthoMCL clusters containing at least four entries (i.e., genes from distinct genomes) was 3,853, their length ranging from 22 to 1650 amino acids (262.542 on average). The concatenated supermatrix thus comprised 1,011,575 columns, including 708,296 variable and 353,936 parsimony- informative characters. ## Phylogenetic inference The ML phylogeny inferred from the concatenated gene alignments is shown in (species tree) together with ML and MP bootstrap support values. The final highest log likelihood obtained was −12,342,390.79, whereas the single best MP tree (excluding uninformative sites) had a score of 1,409,187. ML and MP topologies were identical. Support was maximum (100%) for all branches under ML, and maximum for all but four branches under MP; only a single branch entirely lacked support under MP. As only one genome per genus was included in the sample, there is no taxonomic subdivision of Halobacteriaceae* to compare the tree with. However, outgroup taxonomy was well recovered, the tree showing the monophyly of Methanomicrobiales, Methanosarcinales, and Methanosarcinaceae, each of which were represented with at least two genomes. ## Incongruence between gene trees and species tree After reducing the dataset to the ingroup taxa and to the OrthoMCL clusters present in at least four ingroup genomes, total PBS per OrthoMCL cluster ranged between 142 and −219 (average: 4.941, standard deviation: 19.588, median: 1, MAD: 8.896). These data are plotted against the number of parsimony-informative characters in supplementary. Within a total of 2,891 OrthoMCL clusters, 1,506 genes showed overall positive support and 764 showed overall negative support. Trees inferred from the five clusters least congruent with the species tree are depicted in. They are uniformly characterized by high bootstrap support for groupings in conflict with the species tree estimate. Total PBS values per cluster vary between the COG categories (which could be assigned to 2,213 clusters; see); on average, COGs related to information storage and processing display higher PBS than those associated with metabolism or cellular processes and signaling; but individual COG categories may differ from this general trend. ## Core clusters We used a spectral clustering method to generate gene clusters from the haloarchaeal genomes. There were 887 core clusters, those found in all of the haloarchaeal genomes, and these accounted for 40% to 50% of the genes in each genome. As expected, the core clusters contain genes involved in basic cellular processes such as transcription, translation, DNA replication, DNA repair, RNA modification, protein modification, and protein secretion. The core clusters also include many genes involved in biosynthesis of essential metabolites – amino acids, purines and pyrimidines, lipids, and cofactors. This is somewhat unexpected as the haloarchaea are heterotrophs, but they appear to be relatively self-sufficient in being able to make most essential metabolites. Biosynthetic pathways in the haloarchaea have recently been reviewed, so we will not go into more detail here. The number of genes in each genome belonging to all clusters ranges from 78% to 88%, showing that 12% to 22% of the genes in each genome have no hits to genes in the other halophile genomes. ## Signature clusters We also identified signature gene clusters, those that are shared by all haloarchaea but are not found in any other archaea. There are 112 of these clusters, 89 of which contain proteins with completely unknown function. Of the clusters with a predicted function, two are protein kinases related to the Bacillus subtilis PrkA protein. These two kinase genes are adjacent to each other on the chromosome in all haloarchaeal genomes and are always found with two other genes: one with a domain of unknown function DUF444, and one related to B. subtilis SpoVR, the function of which is unknown. Proteins of these three families are found together on the chromosome in many other microbial genomes suggesting functional linkage. Halophiles are known to accumulate gamma-glutamyl-cysteine, and two of the signature clusters may be involved in gamma-glutamyl-cysteine metabolism. Cluster 491.1× contains proteins related to glutamate-cysteine ligase, and the H. volcanii member of cluster 491.1× was recently shown to have glutamate- cysteine ligase activity. Cluster 1151× includes genes related to glutathione S-transferase, which inactivates toxic compounds by linking them to glutathione. In the halophiles, these may function as gamma-glutamylcysteine S-transferases. ## Habitat-specific clusters Of the ten haloarchaea with sequenced genomes, four were isolated from water and four were isolated from soil or sediment. The ones isolated from water are H. walsbyi, N. pharaonis, H. marismortui, and H. borinquense. H. mukohataei and H. turkmenica were isolated from saline soils, while H. volcanii and H. utahensis were isolated from lake sediments. The water halophiles and soil/sediment halophiles do not form separate clades in the phylogenetic tree. We looked for clusters present in all water-isolated halophiles that are not present in soil/sediment halophiles and vice versa. There were no clusters specific to water halophiles and only three specific to soil/sediment haloarchaea. Proteins belonging to two of the soil/sediment- specific clusters (326.1.0× and 2168×) are often found in the vicinity of nucleotide-sugar metabolic enzymes and glycosyl transferases, suggesting they are involved in cell wall biosynthesis. Since there were few protein clusters completely specific to the water or soil/sediment halophiles, we looked for clusters present in three out of four organisms from one group and absent from the other group. There were 16 clusters present in three out of four water halophiles, of which 11 contain hypothetical proteins and four have only general functional annotations. The only cluster with a specific annotation is cluster 1816×, formate-tetrahydrofolate ligase. Of the 26 clusters found in three out of four soil/sediment halophiles and not present in water halophiles, 11 are hypothetical proteins. Several of the clusters are involved in polysaccharide degradation. These include a glycosyl hydrolase of family GH4, an alpha-L-arabinofuranosidase of family GH51, a polysaccharide deacetylase and a trehalose utilization protein. Two additional clusters present in three out of four soil/sediment halophiles encode a monooxygenase and an acyltransferase, which are found adjacent to each other on the chromosome and flanked by two IucA/IucC family proteins. These proteins are likely to be involved in siderophore synthesis. H. borinquense has three IucA/IucC family proteins but lacks the monooxygenase and acyltransferase, thus it is unclear if it has a complete siderophore biosynthesis pathway. ## All-but-one clusters We also looked for clusters conserved in all but one genome. These probably indicate recent gene losses in each species. The genomes fell into two groups – those that had 20 or less such clusters and those that had greater than 60. The three genomes that had greater than 60 clusters lost were H. salinarum, H. walsbyi, and H. utahensis. Many of the clusters lost from H. salinarum are involved in amino acid synthesis, including genes for the synthesis of glutamate, lysine, ornithine, methionine, and branched chain amino acids. To make up for this, H. salinarum does not have more amino acid transporters or secreted proteases than the other haloarchaea, but it is one of only two of the haloarchaea to have a putative peptide symporter of the OPT family (TC 2.A.67). Symporters have low affinity but high capacity, suggesting that H. salinarum may prefer to live where there is an ample supply of peptides. Of the clusters not present in H. walsbyi, many are involved in flagellum biosynthesis and chemotaxis. However, H. walsbyi has a set of gas vesicle proteins to enable motility in the absence of flagella. The clusters in all except H. utahensis include several enzymes involved in cobalamin synthesis and several enzymes of biotin utilization and propionate metabolism. H. utahensis appears to lack the enzymes for the early steps of cobalamin biosynthesis up to the incorporation of cobalt, but all of the halophiles, including H. utahensis, contain the enzymes for the later steps of cobalamin biosynthesis. Biotin and propionate metabolism are discussed further below. ## Central metabolism Some haloarchaea are known to use the semi-phosphorylated Entner-Doudoroff (ED) pathway for glucose degradation, and genes encoding enzymes of this pathway have been identified in several haloarchaea. With the addition of the new genomes, we find that the semi-phosphorylated Entner-Doudoroff pathway is likely to be present in all sequenced haloarchaea except N. pharaonis, which does not utilize carbohydrates. Aldolases belonging to two different protein families may be involved. All of the halophiles except H. salinarum and N. pharaonis have one or more bacterial-type 2-keto-3-deoxy-6-phosphogluconate (KDPGlc) aldolase (COG0800). Also all except N. pharaonis have at least one potential aldolase related to the characterized Sulfolobus aldolase (COG0329), which is active on KDPGlc and unphosphorylated 2-keto-3-deoxygluconate (KDGlc). The enzymes of the semi-phosphorylated Entner-Doudoroff pathway are highly conserved in sequence among the haloarchaea, suggesting descent from a common ancestor. The standard Embden-Meyerhof pathway of glycolysis appears to be incomplete in the halophiles as no 6-phosphofructokinase could be identified. This agrees with previous experimental studies and analysis. Gluconeogenesis is likely to be present in all of the halophile genomes with the possible exception of H. utahensis. All except H. utahensis have phosphoenolpyruvate (PEP) synthase and/or pyruvate, phosphate dikinase (COG0574). In addition, H. lacusprofundi and H. turkmenica have ATP-utilizing PEP carboxykinase (COG1866). In H. utahensis the only enzyme that potentially can generate PEP for gluconeogenesis is pyruvate kinase. All except H. utahensis have a fructose 1,6-bisphosphatase belonging to the same family as the E. coli Fbp enzyme (COG0158). All of the halophiles including H. utahensis have at least one gene belonging to COG0483, which includes inositol phosphatases and some archaeal fructose 1,6-bisphosphatases. H. utahensis has two genes belonging to this family, but they are weakly related to characterized fructose 1,6-bisphosphatases. These findings suggest that H. utahensis may lack the gluconeogenesis pathway or have an unusual gluconeogenesis pathway. Unlike the rest of the archaea, halophiles are thought to use the oxidative pentose phosphate pathway for generation of pentoses. This pathway also generates NADPH for anabolic pathways. All except H. utahensis have a probable 6-phosphogluconate dehydrogenase (COG1023), the key enzyme of this pathway. In contrast, H. utahensis is the only sequenced haloarchaeon to have transaldolase (Huta_0859) and transketolase (Huta_0860 and Huta_0861), the enzymes of the non-oxidative pentose phosphate pathway. For NADPH generation, H. utahensis possesses genes encoding a NAD/NADP transhydrogenase (Huta_2005–2007). None of the other haloarchaea has the genes for this enzyme. The presence of these enzymes only in H. utahensis suggests that they may have been acquired through lateral transfer, but phylogenetic analysis was unable to identify the donor (data not shown). ## Nutrient transport presents an overview of nutrient transport in the haloarchaea. All of the haloarchaea have at least five symporters for amino acids and at least two ABC transporters for peptides. Since all except H. salinarum can synthesize most or all amino acids, this suggests that amino acids are an important carbon and energy source even in the species that can grow on carbohydrates. All of the haloarchaea also have at least one symporter for nucleosides or nucleobases. Carbohydrate transport is variable. Only half of the halophiles have symporters for sugars, and either none or one ABC transporter for sugars is found in the non-carbohydrate-utilizing organisms. Surprisingly no transporters for sugars could be identified in the H. utahensis genome, suggesting that it uses uncharacterized families of sugar transporters. There appears to be a connection between some transporters and universal stress protein A (UspA) family proteins. Most amino acid transporters of the amino acid-polyamine-organocation (APC) family are either fused to a UspA domain or adjacent on the chromosome to a UspA protein, and some are both fused and adjacent to UspA family proteins (e.g. Htur_0566). This appears to be specific for the APC family as other potential amino acid symporters of the neurotransmitter∶sodium symporter (NSS) family and the dicarboxylate/amino acid∶cation symporter (DAACS) family are not associated with UspA family proteins. Several transporters of the formate-nitrite transporter (FNT) family are also fused or adjacent to UspA domains (Hmuk_1674, HQ1451A, NP6264A, Hlac_2299, and Htur_2705). The FNT family proteins with associated UspA domains are closely related to each other and to a transporter from H. marismortui that lacks a UspA domain (rrnAC0187). They are likely to be formate transporters as the H. marismortui and H. mukohataei proteins are adjacent to enzymes of folate metabolism. Another transporter with adjacent UspA proteins is a putative acetate transporter of the solute∶sodium symporter (SSS) family. This transporter is found in seven of the ten halophile genomes (e.g. NP5136A), and in all cases is followed by a UspA domain protein. In six of the seven genomes with this transporter, it is close on the chromosome to two acetyl-CoA synthetase genes (e.g. NP5128A and NP5132A). The transporter has highest similarity to subfamily 7 of the SSS family, which includes acetate, propionate, and phenylacetate transporters (see the Transporter Classification Database at [www.tcdb.org](http://www.tcdb.org)). UspA family proteins are expressed during many stressful conditions, and they are known to bind ATP, but their exact molecular function is unknown. The UspA domains associated with transporters may play a regulatory role, or may be involved in maintaining transporter function during stressful conditions. A recent report shows that a UspA domain protein is involved in regulation of a transporter. ## Secreted proteases Since the halophiles have numerous transporters for amino acids and peptides, we analyzed the distribution of secreted proteases within their genomes. Only secreted proteases were considered because these are likely to be involved in the utilization of proteins as a nutrient source, while intracellular and integral membrane proteases are involved in a variety of cellular processes. We included proteases that have signal peptides as well as proteins that are likely to be attached to the membrane with the protease domain outside the cell. Signal peptidases (family S26) were excluded from the analysis since they have a specific cellular function. The numbers of secreted proteases in the genomes ranged from 3 to 11. Hierarchical clustering shows that the halophiles fall into two groups with respect to protease distribution. The main feature separating these groups appears to be the presence or absence of secreted members of protease family S8, which includes subtilisin as well as halolysins from halophilic archaea. The organisms having secreted S8 proteases do not correspond to a habitat-specific or phylogenetic group. The presence of at least three secreted proteases in each genome suggests that all of the halophiles may be capable of degradation of extracellular proteins. ## Amino acid utilization Since all of the halophiles have at least five amino acid symporters and two peptide ABC transporters, we investigated pathways of amino acid utilization to see if all of the halophiles are capable of using many amino acids. A summary is provided in. Three degradation pathways were found in all of the ten genomes. All of them had an alanine dehydrogenase similar to the enzyme characterized in Archaeoglobus fulgidus. This enzyme could potentially be involved in synthesis of alanine as well as its degradation. All had at least one asparaginase from COG0252 or COG1446. Finally all of them had a pyruvoyl-dependent arginine decarboxylase similar to the enzyme characterized in Methanocaldococcus jannaschii and agmatinase. This combination of enzymes produces putrescine and urea from arginine. Additional enzymes of arginine utilization were present in some genomes. Six of the genomes have arginase, which produces ornithine and urea. None of the genomes had ornithine decarboxylase, but all of the ones that have arginase also have ornithine cyclodeaminase which produces proline. Several other amino acid degradation pathways are found in a subset of the genomes. The glycine cleavage system is found in all genomes except those of H. walsbyi and H. utahensis. This enzyme complex produces CO<sub>2</sub>, NH<sub>3</sub>, methylene-tetrahydrofolate (THF), and NADH. Methylene-THF has a variety of possible uses within the cell. Another pathway found in all but H. walsbyi and H. utahensis is isoleucine degradation. A 2-oxoacid dehydrogenase complex involved in isoleucine degradation was recently identified in H. volcanii, and seven of the other halophiles have genes with at least 68% similarity to the H. volcanii genes, indicating that they probably have the same function. Five of the genomes have tryptophanase, which produces pyruvate from tryptophan. A histidine degradation pathway with formiminoglutamate as an intermediate is also found in five of the genomes. Seven have proline dehydrogenase, but only H. lacusprofundi has pyrroline-5-carboxylate dehydrogenase to complete proline conversion to glutamate. All of the genomes have threonine dehydratase, but this may be used only for biosynthesis. Five of the genomes in addition have threonine aldolase, which produces glycine and acetaldehyde. Finally, all of the genomes have glutamate dehydrogenase, which may have a biosynthetic role. Four of the genomes have glutamate mutase and methylaspartate ammonia-lyase. These are the first two enzymes of a four-step pathway that produces acetate and pyruvate from glutamate with mesaconate and citramalate as intermediates. However, these two enzymes are likely to be involved in a new pathway for acetate assimilation. Many of the pathways for amino acid degradation are found in a subset of the genomes. They could have been acquired independently by lateral gene transfer or lost in some species. The genes for amino acid degradation in the halophiles are closely related in sequence, suggesting that the common ancestor of haloarchaea was able to degrade many amino acids and that some organisms have lost these pathways. ## Polysaccharide degradation According to the distribution of glycosyl hydrolase domains and carbohydrate- binding modules, halophilic archaea can be divided into 3 groups: those that may be capable of degrading plant biomass (H. utahensis, H. turkmenica and to a lesser extent H. marismortui and H. volcanii), those harboring family 18 glycosyl hydrolases with possible chitinase activity (H. salinarum, H. mukohataei and H. borinquense) and organisms that are unlikely to degrade any externally provided polysaccharides (H. lacusprofundi, H. walsbyi and N. pharaonis). H. utahensis and H. turkmenica have the two largest sets of proteins with glycosyl hydrolase domains among halophilic archaea (43 and 44, respectively). However, their glycosyl hydrolase complements are markedly different. While H. utahensis has five proteins of GH10 family and two proteins of GH11 family, which probably have xylanase activity, H. turkmenica has only one GH10 protein and no GH11 members. The abundance of predicted xylanases in the H. utahensis genome is in agreement with experimental data that showed xylan-degrading activity of this archaeon. The H. utahensis genome contains seven GH5 family proteins and one GH9 family protein (as compared to three and zero in H. turkmenica genome). These proteins may have endo-beta-glucanase activity, thus enabling H. utahensis to degrade components of the plant cell wall. One of the GH5 proteins in H. utahensis (Huta_2387) has been shown experimentally to have cellulolytic activity (T. Zhang et al., in press). H. utahensis also has two GH94 proteins that may have cellobiose or cellodextrin phosphorylase activity. On the other hand, the H. turkmenica genome encodes four GH32 family proteins predicted to have beta-fructosidase (levanase or invertase) activity that are absent from the H. utahensis genome, while both genomes have several GH2 family proteins that may have beta-galactosidase activity. Three of the genomes from haloarchaea isolated from soil or sediment encode enzymes involved in degradation of pectin. The backbone chains of pectin are made up of either homogalacturonan or rhamnogalacturonan with various side chains, and the main chains are linked together by α-1,5-arabinan chains. H. turkmenica has four family 1 polysaccharide lyases (PL) which likely have pectate lyase activity, three of which are close together on the chromosome (Htur_4783, Htur_4785, Htur_4789). Also in the vicinity of these three genes is a family 2 polysaccharide lyase related to pectate lyases (Htur_4786) and two glycosyl hydrolases, one of which may have polygalacturonase activity (Htur_4790). The other H. turkmenica PL1 family protein (Htur_4440) is close to one of two pectin methylesterases (Htur_4438) and a rhamnogalacturonan acetylesterase (Htur_4445). H. turkmenica also has two family 11 polysaccharide lyases (Htur_3890, Htur_3891) that are highly similar to a B. subtilis rhamnogalacturonan lyase and a GH105 protein similar to the RhiN protein of Dickeya (formerly Erwinia) chrysanthemi, which is involved in degradation of rhamnogalacturonate oligosaccharides. H. utahensis and H. volcanii have much lower capacity for pectin degradation: H. utahensis has two probable pectate lyases from family 1, while H. volcanii has one pectate lyase and one pectin methylesterase. In addition to enzymes capable of degrading the main chains of pectin, H. turkmenica, H. utahensis, and H. volcanii have GH43, GH51, and GH93 glycosyl hydrolases with similarity to endo- and exo- arabinases that may be capable of degrading the arabinan linking chains of pectin. ## Galactose utilization H. lacusprofundi and H. mukohataei have been shown to grow on galactose, but the genome sequences suggest that other haloarchaea can also utilize galactose. Also the genome sequences show that two different pathways for galactose metabolism may exist in haloarchaea: the Leloir pathway in H. utahensis, and the De Ley-Doudoroff pathway in H. lacusprofundi, H. marismortui, H. volcanii, H. borinquense, H. mukohataei, and H. turkmenica. H. utahensis is the only haloarchaeon with genes encoding the three enzymes of the Leloir pathway. No other archaeon possesses a gene for hexose 1-phosphate uridylyltransferase (COG1085, Huta_2170), and H. volcanii is the only other haloarchaeon to have a probable galactokinase (HVO_1487). Six haloarchaea (listed above) have genes with high similarity (65–75%) to E. coli galactonate dehydratase, one of the enzymes of the De Ley-Doudoroff pathway. Phylogenetic analysis shows that the genes for galactonate dehydratase in the haloarchaea cluster together (data not shown). In the De Ley-Doudoroff pathway, after 2-dehydro-3-deoxygalactonate (KDGal) is formed by galactonate dehydratase, it is phosphorylated by KDGal kinase to form 2-dehydro-3-deoxy-6-phosphogalactonate (KDPGal). KDPGal is then split by KDPGal aldolase to form pyruvate and glyceraldehyde 3-phosphate. None of the halophiles has a gene related to known KDGal kinases (COG3734). KDPGal aldolases belong to the same family as bacterial-type KDPGlc aldolases of the Entner-Doudoroff pathway (COG0800). H. lacusprofundi has two proteins related to KDGlc kinase and two proteins related to bacterial-type KDPGlc aldolase. One kinase (Hlac_2870) and aldolase (Hlac_2860) are close on the chromosome to each other and to galactonate dehydratase (Hlac_2866), beta-galactosidase (Hlac_2868), and a probable alpha- galactosidase (Hlac_2869), suggesting that the kinase and aldolase may be involved in the utilization of galactose via the De Ley-Doudoroff pathway. Similarly, H. volcanii has three COG0800 proteins, one of which (HVO_A0329) is close on the chromosome to a KDPGlc kinase-related protein (HVO_A0328), galactonate dehydratase (HVO_A0331) and beta-galactosidase (HVO_A0326). Some of the halophiles that have galactonate dehydratase only have one protein related to KDGlc kinase and KDPGlc aldolase. It is possible that in these organisms the proteins are bifunctional, working with both KDGlc and KDGal, similar to the proteins of the Sulfolobus solfataricus ED pathway. ## Fructose utilization Fructose can be utilized by some haloarchaea: H. marismortui, H. borinquense, H. utahensis, and H. turkmenica have been shown to grow on fructose, while H. mukohataei has been shown to grow on sucrose and thus will likely also metabolize fructose. The enzyme ketohexokinase was characterized in Haloarcula vallismortis but the gene was not identified. H. marismortui, H. volcanii, and H. turkmenica each have one transporter of the phosphotransferase system (PTS), and operon evidence suggests that these are fructose transporters that produce fructose 1-phosphate. The PTS transporters from H. volcanii and H. turkmenica are close on the chromosome to putative fructose 1-phosphate kinases (COG1105), while in H. volcanii the PTS proteins are also close to fructose bisphosphate aldolase. In addition to the three haloarchaea with PTS transporters, H. utahensis and H. mukohataei also have putative fructose 1-phosphate kinases. In H. mukohataei the fructose 1-phosphate kinase (Hmuk_2661) is close on the chromosome to fructose bisphosphate aldolase (Hmuk_2663) and another putative sugar kinase (Hmuk_2662) which may be a ketohexokinase. Surprisingly H. borinquense does not have a member of COG1105, despite its known ability to grow on fructose. The protein sequences of all components of the PTS system transporter and fructose 1-phosphate kinase are strongly conserved among the halophiles. ## Xylose utilization The pathway by which H. volcanii utilizes xylose has recently been characterized. The pathway involves formation of xylonate, followed by two dehydratase steps to generate 2-oxoglutarate semialdehyde. This pathway also appears to be present in H. turkmenica and H. lacusprofundi, both of which were not previously known to utilize xylose. H. marismortui is known to produce acid from xylose, and it appears to have 2-dehydro-3-deoxyxylonate dehydratase and xylose dehydrogenase, but it does not have a gene with high similarity to the H. volcanii xylonate dehydratase. H. utahensis is known to degrade xylan and to be able to grow on xylose, and it uses a different pathway for xylose degradation. It has a xylose isomerase (COG2115, Huta_2443) and xylulokinase (TIGR01312, Huta_2446). The resulting D-xylulose 5-phosphate then feeds into the non-oxidative pentose phosphate pathway. H. utahensis is the only one of the sequenced haloarchaea to have transaldolase and transketolase of the non-oxidative PPP, which allows it to use this pathway of xylose utilization. H. borinquense is known to utilize xylose, but it does not have identifiable genes for either of the pathways found in the other haloarchaea. Phylogenetic analysis of xylose isomerase using both neighbor joining (Clustal W) and Bayesian (MrBayes) methods show that H. utahensis xylose isomerase branches deeply within Firmicutes with high bootstrap support (not shown), but xylulokinase did not associate closely to any group of organisms. ## Glucuronate utilization Both H. utahensis and H. turkmenica have putative glucuronate isomerase (COG1904) and mannonate dehydratase (COG1312), suggesting that they may utilize glucuronate by the same pathway as found in E. coli. This pathway produces KDGlc which feeds into the Entner-Doudoroff pathway. H. utahensis also has a probable alpha-glucuronidase (Huta_0871) belonging to glycosyl hydrolase family 67, that is adjacent on the chromosome to glucuronate isomerase (Huta_0870) and mannonate dehydratase (Huta_0869). H. lacusprofundi has a putative mannonate dehydratase but no glucuronate isomerase, therefore it is unclear whether it has the capacity to break down glucuronate. Since H. lacusprofundi is known to grow on mannose, it is possible that mannonate dehydratase is used in a pathway for mannose degradation. ## L-arabinose utilization None of the haloarchaea have been shown to grow on L-arabinose, but the genomes suggest that it may be utilized by some haloarchaea. H. utahensis, H. volcanii, and H. turkmenica all have putative alpha-L-arabinofuranosidases (COG3534). In H. utahensis the arabinofuranosidase (Huta_1152) is close on the chromosome to L-arabinose isomerase (Huta_1154) and ribulose 5-phosphate 4-epimerase (Huta_1149), suggesting that H. utahensis uses the known bacterial pathway of L-arabinose degradation. A gene similar to ribulokinase was not found in H. utahensis, but there is a gene with similarity to xylulokinases (Huta_1150) close to the arabinose degradation genes. Huta_2446 is likely to be a xylulokinase in H. utahensis (see above), and Huta_1150 may be a ribulokinase, completing the pathway. This pathway produces D-xylulose 5-phosphate which enters the non-oxidative pentose phosphate pathway. H. utahensis is the only haloarchaeon to have the non-oxidative pentose phosphate pathway, which allows it to use this pathway, and it is also the only haloarchaeon to have L-arabinose isomerase. ## N-acetylglucosamine utilization Since the presence of family 18 glycosyl hydrolases in H. salinarum, H. mukohataei and H. borinquense indicates that they may possess chitinase activity and use chitin as a growth substrate, we attempted to identify enzymes for subsequent degradation of chitooligosaccharides, N-acetyl-glucosamine or glucosamine. We found that H. mukohataei likely possesses a beta-N- acetylhexosaminidase (Hmuk_3174) that has 51% similarity to characterized enzymes from Streptomyces thermoviolaceus and Bacillus subtilis. A chitobiose deacetylase has been identified in Thermococcus kodakaraensis belonging to COG2120, which includes other carbohydrate deacetylases. Both H. mukohataei and H. borinquense have genes belonging to this family, although they are distantly related to the T. kodakaraensis enzyme. None of the organisms with family 18 glycosyl hydrolases has been tested for growth on N-acetylglucosamine or glucosamine, so the presence of chitinase activity and chitinolytic pathway in haloarchaea needs further experimental elucidation. ## Glycerol metabolism and transport The haloarchaea encode genes for two different glycerol utilization pathways. All except N. pharaonis have a glycerol kinase and glycerol 3-phosphate dehydrogenase, and the genes for both enzymes are found close together on the chromosome. Another pathway involving glycerol dehydrogenase and dihydroxyacetone kinase is present only in H. lacusprofundi. H. volcanii and H. walsbyi have a dihydroxyacetone kinase without glycerol dehydrogenase, and this may be used for metabolism of dihydroxyacetone from the environment. H. salinarum encodes a glycerol dehydrogenase but no dihydroxyacetone kinase. Only H. mukohataei has an identifiable glycerol transporter within the genome. It encodes a member of the Major Intrinsic Protein (MIP) family adjacent to glycerol kinase, providing strong evidence for a glycerol transport function. All other haloarchaea that have a glycerol kinase have an uncharacterized membrane protein adjacent (e.g. rrnAC0550), and we predict that these genes encode a new family of glycerol transporters. H. borinquense has two glycerol kinases and both have this uncharacterized membrane protein family adjacent to the kinase gene. There are also bacterial homologs of this membrane protein family, and many of them are adjacent to genes involved in glycerol or propanediol metabolism. ## Propionate metabolism H. lacusprofundi is the only halophile that has been shown to grow on propionate, but all of the haloarchaeal genomes, with the exception of H. utahensis, contain genes that may encode the methylmalonate pathway for conversion of propionate to succinyl-CoA. All except H. utahensis have methylmalonyl-CoA epimerase (TIGR03081) and methylmalonyl-CoA mutase (COG1884 and COG2185). Also they have a biotin carboxylase protein (COG4770) and a carboxyltransferase protein (pfam01039), subunits of a biotin-dependent carboxylase. All except H. utahensis also contain a biotin-protein ligase (COG0340) and a BioY family biotin transporter (pfam02632). Propionate or propionyl-CoA may be produced intracellularly from the breakdown of fatty acids, amino acids, or other compounds, or these organisms may be able to use propionate from the environment produced as a result of fermentation. ## Glycine betaine metabolism and transport Glycine betaine is a compatible solute which is likely to be present in high- salt environments. All of the haloarchaeal genomes except that of H. mukohataei encode members of the betaine/carnitine/choline transporter (BCCT) family, which transport glycine betaine and related compounds. Most have one or two members of this family, but H. turkmenica has seven. In addition, H. turkmenica has an ABC transporter for compatible solutes. Four of the haloarchaeal genomes – H. marismortui, H. walsbyi, H. volcanii, and H. turkmenica – encode one or two genes with high similarity to dimethylglycine oxidase from Arthrobacter globiformis. In all four genomes these oxidase genes are close on the chromosome to BCCT family transporters. The presence of these enzymes and transporters raises the possibility that some of the halophiles may be able to utilize glycine betaine, dimethylglycine, and/or sarcosine. However, in H. walsbyi, betaine was not found to enhance growth and H. utahensis could not grow on betaine. # Discussion ## Phylogenetic inference Our analysis of the ten haloarchaeal genomes provided a phylogenetic tree in which all nodes were well-supported. We here followed the ‘total evidence’ approach for inferring the species tree, which dictates that the best phylogenetic hypothesis is the one based on all available data , and that conflict between the species tree and gene trees can at least as well as via inferring the gene trees be quantified after combined analysis. Total evidence has been criticized as a ‘verificationist’ approach, particularly by researchers who question the concept of a microbial tree of life (TOL) in general. Other authors counter that concatenated analysis is well rooted in the ‘Popperian principles of background knowledge and corroboration’, and that whether a TOL exists can be assessed, among other means, by checking for high branch support, which would indicate a strong treelike signal, and by comparing previous phylogenetic or taxonomic hypotheses. These issues are of considerable importance for the future of Prokaryote classification. The mostly maximal bootstrap support observed here indicates that Halobacteriaceae have evolved in manner consistent with a TOL, which is supported by the majority of the genes. Moreover, our supermatrix tree is fully in agreement with the current taxonomy of ingroup and outgroup, whereas considerable conflict was observed in a recent total evidence analysis of Pasteurellaceae. The single exception is the branch connecting H. borinquense and H. walsbyi, which is supported by 700 genes and contradicted by 565 ones, yielding a total positive support of 2,422 steps and a total negative support of −2,366 steps. Considerable incongruence between gene trees may thus be responsible for the low support under MP, but even this branch is maximally supported under ML. ## Incongruence between gene trees and species tree While our results are in agreement with the optimistic view regarding TOL inference from supermatrices, they nevertheless also indicate that disagreement with the species tree is considerable for some genes ( and). It is possible to distinguish the effects of artifacts of the phylogenetic inference method, caused e.g. by a low signal-to-noise-ratio, from those of horizontal gene transfer (HGT) because in the former case a correlation between gene length and the extent of disagreement between gene tree and species tree occurs. This correlation is significant in our analysis, but a number of genes appear as outliers of the regression, indicating distinct causes of incongruence such as HGT. The distribution of the median total PBS values over the COG categories reveals that genes related to information processing and storage are more in accordance with the species tree, corresponding to the view that such genes are less frequently horizontally transferred between organisms. However, many exceptions from this rule exist, as also observed in a recent analysis of a comprehensive genomic dataset of Prokaryotes. In any case, total evidence analysis in conjunction with partitioned Bremer support apparently worked well in identifying those genes most in disagreement with the species trees. Among the five ‘worst’ gene trees, those for L-lactate permease, Tyrosyl-tRNA synthetase and Cobalamin biosynthesis protein CobN did not comprise any ingroup paralogs in their OrthoMCL clusters; thus, horizontal gene transfer is the most likely explanation for their considerable incongruence with the species tree. HGT between Candidatus Desulforudis audaxviator and Archaea has been suggested in the literature for CobN and between Opisthokonta and Archaea for Tyrosyl-tRNA synthetase. In contrast, the OrthoMCL clusters containing either Glucosamine 6-phosphate synthetase or Acyl-coenzyme A synthetase comprised a combination set of inparalogs and possible outparalogs; that is, their disagreement with our species tree estimate is most likely due to a complicated gene duplication/gene loss pattern which may not optimally have been resolved by the OrthoMCL algorithm. ## Soil/sediment vs. water halophiles While there were few habitat-specific clusters, soil/sediment halophiles were found to have more glycosyl hydrolases as well as genes involved in siderophore synthesis and cell wall metabolism. Some of the halophiles analyzed here were isolated from water (H. walsbyi, N. pharaonis, H. marismortui, H. borinquense), while others were isolated from soil (H. mukohataei, H. turkmenica) or lake sediment (H. volcanii, H. utahensis). We looked for gene clusters found exclusively in the water halophiles or in the soil/sediment halophiles, but very few clusters were found. There may be several explanations for this finding. One possibility is that the change from water to sediment or vice versa has happened several times independently, and different changes in the genome occurred during each round of adaptation. This explanation is supported by the fact that the water and soil/sediment halophiles do not form separate clades in the species phylogenetic tree. Another potential explanation is that the division between water and sediment is not clear-cut. Even though some of these organisms were isolated from water or sediment, they may live part of the time in both environments. Nevertheless, there were some clusters with known functions found in the soil/sediment group and not in the water group. The soil/sediment halophiles tend to have greater numbers of glycosyl hydrolases. Each of the four soil/sediment halophiles has between 15 and 44 glycosyl hydrolases, while the haloarchaea isolated from water have between zero and twelve glycosyl hydrolases. The soil/sediment halophiles also have clusters involved in siderophore synthesis that are missing from the water group. This may be explained by the fact that siderophores are less likely to diffuse away from the producing organism in soil or sediment. Finally the soil/sediment group has two clusters likely to be involved in cell wall metabolism. This may indicate a reaction to the more complex environments that they inhabit. ## Uniqueness of H. utahensis Differences in central metabolism in H. utahensis may make it easier for this organism to grow on pentoses. One of the findings from this study is that H. utahensis is substantially different from the other halophiles in its central metabolism. It may lack gluconeogenesis, and it appears to use the non-oxidative pentose phosphate pathway and a transhydrogenase instead of the oxidative pentose phosphate pathway used by other haloarchaea. It has the fewest amino acid degradation pathways of any of the sequenced haloarchaea, and thus appears to be the only one of the ten haloarchaea studied here that is specialized only for carbohydrate utilization. The use of the non-oxidative pentose phosphate pathway may allow for more flexibility in the ability to utilize pentoses as carbon and energy sources. H. utahensis is known to degrade xylan and can grow on xylose. The other halophiles have a different xylose degradation pathway that produces 2-oxoglutarate. The xylose and L-arabinose degradation pathways of H. utahensis are likely to feed into the pentose phosphate pathway, thus producing fructose 6-phosphate and glyceraldehyde 3-phosphate from phosphorylated pentoses. However, the fate of fructose 6-phosphate is unclear. Halophiles are thought to lack phosphofructokinase, but several including H. utahensis have putative fructose 1-phosphate kinase. A possible pathway for fructose utilization would involve the conversion of fructose 6-phosphate to fructose 1-phosphate, followed by phosphorylation to form fructose 1,6-bisphosphate, which would then enter the Embden-Meyerhof pathway. However, further experimental work will be needed to determine how pentoses are metabolized in H. utahensis. ## Distribution of catabolic pathways The distribution of catabolic pathways and the protein sequence conservation of the associated enzymes suggest that the ancestor of haloarchaea could degrade many amino acids but few carbohydrates. The haloarchaea possess pathways for the utilization of amino acids, carbohydrates, and other compounds. The amino acid degradation pathways are found in four to ten of the genomes, and the sequences of the enzymes are closely related, suggesting that they were present in the ancestor of haloarchaea. Three of the genomes have five or fewer amino acid catabolic pathways (H. walsbyi, H. utahensis, and N. pharaonis), while the remaining genomes have at least seven pathways. H. walsbyi and N. pharaonis also have few glycosyl hydrolases and carbohydrate utilization pathways, and they use very few compounds as energy and carbon sources. H. utahensis on the other hand has a large number of glycosyl hydrolases and sugar utilization pathways, showing that it has become specialized for growth on carbohydrates. In contrast to amino acid utilization pathways, alternative pathways are used for sugar catabolism. For example, H. utahensis uses a different xylose degradation pathway than the others, and fructose phosphorylation can occur via the phosphotransferase system or ketohexokinase. In addition, H. borinquense is known to grow on fructose and xylose, but it lacks the catabolic enzymes found in other haloarchaea. However, some genes and pathways are highly conserved. The enzymes of the semi-phosphorylated Entner-Doudoroff pathway are closely related in sequence, suggesting that glucose utilization by this pathway was present in the haloarchaeal ancestor. Fructose 1-phosphate kinase is found in five haloarchaeal genomes and is highly conserved. Similarly galactonate dehydratase is found in six of the ten genomes and forms a single phylogenetic group. Overall it appears that the ancestor of haloarchaea was able to use up to ten amino acids as energy and carbon sources, while its ability to use sugars was limited. ## Biotechnological applications Halophilic glycosyl hydrolases have potential uses in polysaccharide degradation related to biofuel production. An unexpected result from this analysis was the high number of glycosyl hydrolases in two of the organisms. H. utahensis has numerous genes related to cellulases and xylanases, while H. turkmenica has a large number of predicted pectin-degrading enzymes. Salt-adapted glycosyl hydrolases from haloarchaea may have applications in the depolymerization of plant material for biofuel production. Plant material is pretreated in order to remove lignin and hemicellulose and to reduce the crystallinity of cellulose. Several different methods are used, some of which involve incubation with acid or base. After treatment with these chemicals, a neutralization step is required before further processing, and this produces a salty solution. Salt-adapted glycosyl hydrolases may be useful at this point in the process. Also, treatment of plant material with ionic liquids has been shown to reduce the crystallinity of cellulose and enhance its hydrolysis. One of the glycosyl hydrolase family 5 enzymes of H. utahensis has been shown to be a cellulase and to be active in an ionic liquid (T. Zhang et al., in press). Carrying out hydrolysis of lignocellulose under high-salt conditions has the advantage that the possibility of contamination is low. # Supporting Information [^1]: Conceived and designed the experiments: H-PK NK. Performed the experiments: IA CS MG KM SDH IP NI. Analyzed the data: IA CS MG IP NI. Contributed reagents/materials/analysis tools: KM SDH. Wrote the paper: IA MG NI. [^2]: The authors have declared that no competing interests exist.
# Introduction In recent decades, studies have been addressing a possible contribution of traffic related air pollution to allergic diseases. Interestingly, tyrosine residues of pollen allergens are efficiently nitrated by the air pollutants nitrogen dioxide and ozone at levels reached in urban air. In sera of birch pollen-allergic patients, the levels of IgE recognizing nitrated major birch pollen allergen Bet v 1.0101 (referred to as Bet v 1 nitro) were significantly higher compared to IgE specific for unmodified Bet v 1.0101 (Bet v 1) and in mouse models, nitrated Bet v 1 and nitrated Ovalbumin are more potent allergens when compared to their unmodified forms. These findings suggest that post- translational modifications (PTMs), such as nitration, can increase the potential of pollen allergens to trigger immune responses and might play a role in the emergence of allergies. PTMs within the human body have been observed and characterized in numerous studies. Although the majority of PTMs are required for the biological function of the proteins, several modifications were also identified in the context of autoimmune diseases. Nitrated proteins were found to be present in multiple sclerosis, Alzheimer's disease, M. Parkinson, and atherosclerosis and are a hallmark of inflammation,. Some modified self proteins induce immune responses leading to the generation of antibodies which recognize the modified and/or the unmodified protein. These findings suggest that PTMs might alter processing and presentation of proteins by professional antigen presenting cells, leading to the generation of new antigenic epitopes and potential induction of a T cell response. The presentation of protein fragments via HLA-DR molecules by antigen presenting cells, such as mature dendritic cells (DCs), is a key event in the induction of a T cell response. After internalization by dendritic cells, proteins are enzymatically cleaved within endolysosomal compartments. Some of the resulting peptides, which are of considerably variable length, bind to HLA-DR molecules in a sequence dependent and HLA-DM-edited manner. It has been established that PTMs can increase the peptide binding affinity to MHC class II molecules, or interfere with the proteolysis of proteins. This may, in addition to the alterations introduced by the modified amino acid residue itself, result in the generation of new, naturally processed HLA-DR associated peptides, potentially giving rise to T cell epitopes. For some PTMs, such as maleylation – and nitration, there is evidence that protein uptake by antigen presenting cells can be altered. We have studied whether there is a difference between the peptides derived from the allergen Bet v 1 presented via HLA-DR and those derived from post- translationally chemically modified Bet v 1 nitro. For this purpose, immature DCs were loaded with unmodified Bet v 1 or Bet v 1 nitro. After affinity purification of the HLA-DR peptide complexes, the HLA-DR associated peptides were isolated by acidic elution and identified by liquid chromatography-mass spectrometry and the identified Bet v 1 or Bet v 1 nitro derived peptides were compared with respect to peptide clusters, peptide length variants and copy number of peptides. Since changes in the pattern of presented HLA-DR associated peptides on DCs can also change the recognition by T lymphocytes, and since the conversion of tyrosine to nitrotyrosine has already been shown to affect the reactivity of T cells for other proteins, we also addressed the question whether peripheral blood mononuclear cells (PBMCs) loaded with Bet v 1 nitro can activate T lymphocytes more efficiently than PBMCs loaded with unmodified Bet v 1. For this purpose Bet v 1-specific T cell lines were generated from birch pollen allergic patients and T cell proliferation towards unmodified Bet v 1 or Bet v 1 nitro was analyzed. # Results ## Structural analysis and nitration of the allergen Bet v 1 and human serum albumin The UCSF-Chimera was used for the structural analysis of the seven tyrosine residues of Bet v 1 regarding accessibility and electrostatics on the protein databank entry 1BV1. With the exception of Y120, all tyrosine residues have exposed 3-carbon atoms, indicating that they are available for posttranslational modifications. Residues Y5, Y66, Y120, Y150 and Y158 are on the outside, whereas Y81 and Y83 are exposed in the large cavity of Bet v 1. In terms of accessibility, all tyrosine residues except Y120 are candidate targets for nitration. In the X-ray structure, several water molecules are found in the cavity, which indicates that there is a reasonable space and polarity in the inside of the molecule. It has been reported that one factor determining the selectivity of nitration is the charge environment of the tyrosine resides. The electrostatic potential was therefore calculated using DelPhi and mapped with the Chimera program to 1BV1. Most of the targetable groups, with the exception of Y66, are in a negative environment, which favours nitration. We subsequently applied GPS-YNO2, a recently developed predictor for potential nitration sites based on sequence information. The predictor ranks Y150 and Y158 with the highest scores, but also the inaccessible Y120 with a high score. This data does not completely correspond to the above mentioned structural analysis and suggests that the prediction might not be fully optimized for Bet v 1. Overall, the data suggest that Y5, Y81, Y83, 150 and 158 are preferably nitrated, the latter two with a remarkable high score in the GPS-YNO2 prediction. In contrast, Y66 is located in a positive electrostatic environment and Y120 is not accessible and those two residues are probably not nitrated to the same degree. Treatment of Bet v 1 and human serum albumin (HSA) with TNM led to nitration of tyrosine residues. We determined that 58.1% of all tyrosine groups within one protein molecule were nitrated in both batches of Bet v 1 (data not shown). Donors B01 and B02 were loaded with nitro Bet v 1 of batch 1 and donors B03–B10 were loaded with nitro Bet v 1 of batch 2. Two batches of HSA nitro were generated with nitration grades of 21.4% and 26.3%. ## Loading of DCs with unmodified Bet v 1 or Bet v 1 nitro led to an alteration of the Bet v 1 derived peptide repertoire presented via HLA-DR Results of the peptides identified from DCs loaded with unmodified Bet v 1 and Bet v 1 nitro are given in. A total of 1076 to 13367 peptides containing 381 to 1347 different peptides were identified per sample generated from cells loaded with Bet v 1, while Bet v 1 nitro derived samples gave rise to a total of 758 to 14821 peptides containing 379 to 1380 different peptides per sample. Despite this huge number of identified naturally processed HLA-DR associated peptides some less frequent peptides may not have been detected due to sensitivity limits of the LC-ESI-MS/MS setup. The majority of identified peptides were derived from a broad panel of proteins present in the endolysosome. From this total number of different identified peptides, an average of 0–0.8% of the peptides were derived from Bet v 1 in samples generated from DCs loaded with unmodified Bet v 1 and 0.3–7.2% of the total peptides were derived from Bet v 1 nitro in samples generated from DCs loaded with Bet v 1 nitro. Thus, the percentage of different allergen-derived peptides related to the total number of identified peptides in each sample was considerably increased in samples generated from Bet v 1 nitro loaded DCs. All Bet v 1 derived peptides identified from different donors are listed in supplementary. In none of the negative control samples, false positive Bet v 1-derived peptides were identified, ruling out false positive identification of allergen-derived peptides. Allergen-derived peptides could be identified in samples generated from DCs loaded with Bet v 1 or Bet v 1 nitro in eight out of ten donors, or in all ten donors, respectively. Peptides were derived from all regions of the Bet v 1 protein sequence, except for region aa66–72. It is notable that the average length of Bet v 1 derived peptides differed only marginally in samples generated from DCs loaded with unmodified Bet v 1 (peptide length: 22.34 on average) compared to Bet v 1 nitro (peptide length: 22.05 on average). The sequences of naturally processed peptides found for Bet v 1 matched those identified by another study using DCs from 4 allergic patients. ## Nitration increases the number of clusters as well as the number of different length variants of Bet v 1-derived naturally processed peptides Identified peptides “clustered” in several regions along the amino acid sequence. In peptide samples derived from DCs loaded with unmodified Bet v 1, an average of 0.9 peptide clusters per donor could be observed, while in samples derived from Bet v 1 nitro loaded DCs, an average number of 2.6 peptide clusters could be identified. In region aa66–72, containing a tyrosine residue, no allergen derived peptides could be identified. In all other sequence regions containing tyrosine residues, both, peptides with or without tyrosine modification could be identified in samples generated from cells loaded with Bet v 1 nitro. These peptides are listed in Supplementary and the data displayed in this table indicate that all tyrosine residues were nitrated to some extent. For all ten donors studied, peptide samples generated from Bet v 1 nitro loaded DCs showed a higher number of Bet v 1 derived peptide length variants (with an average of 21.2 per donor) compared to peptide samples isolated from DCs loaded with unmodified Bet v 1 (with an average of 3 peptide length variants per donor). Nitration of Bet v 1 resulted in a 2.9-fold increase in the number of identified Bet v 1 a-derived peptide clusters and in a 7.2-fold increase in the number of identified peptide length variants (i.e. 2.5-fold increase of peptide length variant per cluster), resulting in a broader range of different peptides being presented to T-lymphocytes. ## Nitration of Bet v 1 increases the copy number of identified Bet v 1-derived peptides According to their amino acid composition, peptides with different sequences exhibit different flight characteristics within electrostatic and electrodynamic fields. Therefore, the detection efficiency for different peptides can vary, which complicates a quantitative comparison of the identified peptides in different samples. Overall, 94.3%+/−22.4 of the total peptides in samples treated with Bet v 1 nitro was similar to the total number of peptides in samples treated with unmodified Bet v 1. This implies a similar peptide loading efficiency, recovery, and peptide distribution in samples generated from the same donor. Thus, within the same donor it was eligible to compare quantities of peptides with identical amino acid sequence and consequently identical flight characteristics in the mass spectrometer within samples loaded with unmodified Bet v 1 or Bet v 1 nitro. Therefore, peptides in samples generated from DCs loaded with unmodified Bet v 1 were compared with the identical peptides from samples generated from DCs loaded with Bet v 1 nitro for each donor. Peptides with the same sequence but containing nitrated tyrosine residues were not taken into account, since the introduction of NO<sub>2</sub> groups might alter the flight characteristics of a peptide. Exemplarily for two different peptide stretches aa133–157 in two different donors B05 and B07 and aa102–132 in two different donors B06 and B08, the top matched sequences are shown in the supplementary. Among 17 comparable sequences, we observed in all cases an increase in the number of allergen derived peptides in samples generated from DCs loaded with Bet v 1 nitro ranging from 1.1-fold up to 106-fold in comparison to peptides in samples generated from DCs loaded with unmodified Bet v 1. The mean increase in peptides with identical sequences in DCs loaded with Bet v 1 nitro compared to DCs loaded with unmodified Bet v 1 was 12.2 fold. This indicates that nitration had a remarkable impact on the presentation of allergen-derived peptides, and that DCs loaded with Bet v 1 nitro accumulated a higher copy number of Bet v 1 derived HLA-DR peptide complexes Of note an increase in allergen-derived peptides was not only observed in sequence stretches containing tyrosine residues but also regions devoid of tyrosine. ## Nitration of Bet v 1 leads to an enhanced proliferation of Bet v 1 specific T cell lines In four out of five tested birch pollen allergic patient samples, the proliferation of Bet v 1 specific T cell lines gave a higher response to Bet v 1 nitro when compared to unmodified Bet v 1 at the lowest tested concentration (1.25 µg/ml). Statistical analysis was performed for the lowest concentration for which the difference in proliferation between Bet v 1 and Bet v 1 nitro was the highest (1.25 µg/ml for donors 1, 2, 4, 5 and 2.5 µg/ml for donor 3). With p = 0.046, proliferation of Bet v 1 specific T cell lines in response to Bet v 1 nitro was significantly higher than proliferation induced by unmodified Bet v 1. Proliferation of Bet v 1 specific T cell lines was very low upon control treatment with HSA or HSA nitro, showing that the presence of nitrotyrosine as such does not induce an increased T cell response. Interestingly, increasing the Bet v 1 concentrations had little effect on the proliferation and the difference between native and nitrated Bet v 1 was smaller and no longer significant. ## HLA typing of donors For donors B01–B10 used for isolation of HLA-DR associated peptides from DCs, DRB1, DRB3, DRB4 and DRB5 alleles were determined. In general, peptide patterns of donors sharing one allele also shared peptide clusters. Three out of four donors with allele DRB1\0301 shared HLA-DR associated peptides at aa1–48 and three out of three donors with allele DRB1\1501 shared peptides at aa130–159. DRB1\0701 (present in donors B02, B03 and B10) appears to be correlated with low presentation of Bet v 1 peptides. Thus, in general, shared HLA-DRB1 alleles correlated with the presentation of peptides derived from the same region within Bet v 1. Interestingly, the effect of nitration on the presentation of Bet v 1 peptides seems to be independent of HLA alleles. # Discussion Our study demonstrates that presentation of HLA-DR associated peptides was altered upon nitration of Bet v 1. Nitration resulted in a 2.9-fold increased number of identified peptide clusters, a 7.2-fold increase in the overall number of peptide length variants and a 12.2-fold increase in the copy number of identified peptides derived from major birch pollen allergen. An increase in allergen-derived peptide presentation was observed not only for sequence stretches containing tyrosine residues but also in regions devoid of tyrosine. This indicates a general change in uptake and/or processing of the Bet v 1 nitro protein. There could be several explanations for the increase in the presentation of Bet v 1 nitro derived peptides; 1) an enhanced uptake of the allergen by the DCs, which could be due to multimerization of the allergen. Consistently, for Bet v 1 nitro an elevated tendency to form dimers and trimers was observed using high performance size exclusion chromatography and Western blot analysis (data not shown). 2) there are membrane bound receptors recognizing nitrotyrosine; this would facilitate the uptake of the nitrated allergen, and/or alter the processing of the allergen within endolysosomal compartments of DCs. 3) nitration of tyrosine residues has the potential to alter the sterical properties and interaction with neighboring amino acids, resulting in changes of the protein conformation and ultimately altered antigen processing. 4) nitrotyrosine residues could render the allergen either more or less susceptible to protease activity of degrading enzymes without affecting the processing and presentation of unmodified proteins. A decrease in the stability of nitrated allergens was confirmed in a recent study on food allergy. In this study, the nitrated allergens administered orally in a mouse model had a reduced allergenicity and were more easily digested. In contrast, intravascular injection of nitrated food proteins did increase their allergenicity. Our data show that the average length of allergen derived peptides differed only marginally between samples from DCs loaded with unmodified allergen (22.34 amino acids) or samples from DCs loaded with nitrated allergen (22.98 amino acids). However, nitration could still affect the kinetics of nitrated Bet v 1 degradation, consequently leading to a change in the presentation kinetics of Bet v 1-derived peptides and eventually the peptide profile which is presented to the T-cells. Which of the delineated processes finally contributes to the increase of Bet v 1 nitro derived peptide presentation remains unknown and the interplay seems to be complex. Noteworthy, the nitration of human serum albumin did not lead to an enhanced peptide presentation (data not shown); suggesting that nitration per se does not result in enhanced peptide presentation of the nitrated protein. Thus, the impact of nitration on antigen presentation also seems to depend on the properties of the protein itself. However, formally we cannot rule out that the grade of nitration impacts on antigen presentation as well. The identification of Bet v 1 derived HLA-DR associated peptides led to the question, whether these peptides would be recognized by T cell receptors of T lymphocytes and if this would result in activation and proliferation of the T lymphocytes. For this purpose PBMCs loaded with Bet v 1 nitro were assessed for their capacity to stimulate Bet v 1 specific T cell lines. In response to Bet v 1 nitro, proliferation was significantly higher as compared to unmodified Bet v 1 using a concentration of 1.25 µg/ml protein. Increasing the protein concentration did not in result in higher proliferation and the impact of nitration was lower. Altogether, our data showed that nitration not only enhanced presentation of Bet v 1 derived HLA-DR associated peptides on DCs, but also had an impact on T cell activation. It has been hypothesized that nitration of autologous proteins may contribute to autoimmunity ; additionally, the fact that nitration occurs in inflamed tissue should be taken into account. Nitration of tyrosine residues may have evolved as a strategy to intensify immune responses against foreign proteins derived from viruses or bacteria, while a possible contribution to autoimmunity had to be accepted as an unfortunate side effect. Improved presentation of pathogen derived nitrated peptides may in contrast be beneficial to the host. Tyrosine nitration could also be seen as a danger signal – a type of stimulus which is thought to play an important role in the regulation of immune responses. Airborne allergens bearing nitro-tyrosine mimic nitrated foreign proteins present in inflamed tissue, which may explain our findings that nitration of allergens intensifies the presentation of allergen derived HLA-DR associated peptides. Previous studies have shown increased immunogenicity of Bet v 1 nitro compared to Bet v 1 : Sera from patients with birch pollen allergy contain higher titers for IgE against Bet v 1 nitro compared to Bet v1; the reactivity against Bet v 1 nitro cannot be fully removed by absorption with normal Bet v 1, indicating a specific recognition of the nitrated allergen. The same study showed that nitrated Bet v 1 and nitrated Ovalbumin were more potent allergens compared to their unmodified forms when tested in mouse models. Regarding the issue of HLA haplotypes and the predisposition to allergies published studies show diverging results. Several studies have shown associations between IgE reactivity and the presence of distinct HLA-DRB1 alleles; most notably in patients allergic to ragweed Amb a 5, Alternaria Alt a 1, Parietaria Par o 1, birch Bet v 1, cat Fel d 1, as well as cockroach and house dust mite allergens. In these cases HLA-DRB1 haplotypes could favor susceptibility to allergy. However, Jahn-Schmid et al. have recently shown that the dominant T cell epitopes of the major ragweed allergen Amb a 1 were presented by different HLA- DR, DP and DQ molecules. These findings suggest that, alternatively, a broad HLA class II restriction profile might contribute to the high allergenic properties of Amb a 1. Several questions remain to be addressed, e.g. if and/or how nitrated proteins may interfere with uptake and/or processing pathways of DCs or if potential alternative uptake mechanisms for nitro-proteins e.g. via specialized receptors expressed on DCs might exist. Furthermore, the questions whether chemical nitration of the protein compared to nitration by NO<sub>2</sub> and ozone in polluted air have different characteristics (e.g. act on different tyrosine residues) and whether they contribute to nitration to a similar extent could not be investigated in the scope of the present study. Environmental pollutants might nitrate tyrosine residues less eagerly and more selectively than the chemical agent used here. These aspects will have to be addressed in follow-up investigations. In summary, our data show that nitration has an enhancing effect on processing and presentation of Bet v 1 derived HLA-DR associated peptides, by enhancing both the quality and the quantity of the Bet v 1 specific peptide repertoire. # Materials and Methods ## Allergen Recombinant Bet v 1 isoform a (Bet v 1a/Bet v 1.0101) was purchased from Biomay, Vienna. ## HLA-DR-specific antibody The hybridoma cell line L243 was used for the production of monoclonal antibodies specific for HLA-DRαβ dimers. HLA-DR specific antibodies were purified from hybridoma supernatants by Protein A chromatography, immobilized on CNBr-activated sepharose beads (Pharmacia) according to the manufacturer's protocol and stored containing 0.02% sodium azide. ## Nitration of major birch pollen allergen Recombinant Bet v 1 was dissolved in H<sub>2</sub>O at a concentration of 1 mg/ml. Protein concentration was determined using a Bio-Rad Protein assay (Bio- Rad, Munich, Germany). Nitration of the allergen was performed with an excess of 30 tetranitromethane (TNM) molecules per tyrosine residue. To generate Bet v 1 nitro, half of the allergen solution was mixed with 0.5 M TNM/methanol solution (Chemos, Regenstauf, Germany) to a final concentration of 12.03 mM TNM. The other half of the allergen solution was mixed with the corresponding amount of methanol without TNM, to generate unmodified Bet v 1. After incubation for 60 min at room temperature, the protein solutions were purified by size exclusion chromatography using PD-10 Sephadex G-25M columns (GE Healthcare, Uppsala, Sweden), in order to remove potential smaller protein fragments generated during the chemical modification. The protein concentrations were determined using the Bio-Rad Protein assay and the nitration grade was determined based on a standard curve using nitrotyrosine by measurement of the samples at 420 nm. ## Generation of monocyte derived DCs DCs were generated from peripheral blood mononuclear cells (PBMCs). PBMCs were purified from buffy coats obtained from healthy donors from the blood bank in Basel using Ficoll gradient centrifugation. Monocytes were purified from peripheral blood mononuclear cells by positive selection using CD14-specific antibody coated MicroBeads (Miltenyi Biotech, Auburn, CA) and differentiated to immature DCs in complete RPMI supplemented with granulocyte macrophage-colony stimulating factor (GM-CSF, 33 ng/ml) and interleukin-4 (IL-4, 3 ng/ml) for 5 days at 37°C and 5% CO<sub>2</sub>. DCs were shown to be in the immature state, as characterized by very low expression of CD86 and lack of expression of CD83. ## Loading of DCs with allergens Maturation of the immature DCs was induced after five days with LPS (1 µg/ml, Sigma, St. Louis, MO). The DCs were either loaded with 10 µg/ml unmodified Bet v 1 or Bet v 1 nitro. Unloaded DCs served as background control. After 24 hours, DCs were harvested after re-suspension and washing in PBS. The cell pellets were immediately frozen at −70°C. ## Isolation of HLA-DR restricted peptides DC pellets (obtained from 1×10<sup>7</sup> cells) were lyzed in hypotonic buffer containing 1% Triton X-100 and protease inhibitors for 1 hour on a horizontal shaker. The lysate was cleared from cell debris and immunoprecipitated with mAb L243 conjugated sepharose beads. After washing with double-distilled water, elution of peptides from HLA-DR molecules was achieved with 0.1% trifluoracetic acid (Fluka, Buchs, Switzerland) at 37°C. The peptides were lyophilized in an Eppendorf Concentrator 5301 (Eppendorf AG, Hamburg, Germany). Cell lysates, before and after immunoprecipitation, were analyzed by Western Blotting using the HLA-DRα-specific mAb 1B5 to determine the HLA-DR depletion efficacy. ## Mass spectrometry Peptide identification was performed using multidimensional protein identification technology (MudPIT) combining a two dimensional liquid chromatography with a mass spectrometric analysis (LC-ESI-MS/MS). Lyophilized peptides were resuspended in hydrophilic buffer containing 5% acetonitrile, 0.5% acetic acid, 0.012% heptafluorbutyric acid and 1% formic acid. Peptides were fractionated using a MudPIT column packed with C18 reversed phase material and SCX material. Elution of the peptides from the column was performed in 10 cycles on an UltiMate 3000 nanoflow HPLC (Dionex Corporation, Sunnyvale, CA). The first six cycles started with a salt step with increasing concentrations of ammonium acetate (0–225 mM) followed by a nonlinear acetonitrile gradient. The seventh cycle consisted of a salt step with 250 mM ammonium acetate and a nonlinear acetonitrile gradient. The last three cycles started with a salt step with 1500 mM ammonium acetate followed by a nonlinear acetonitrile gradient. The MudPIT column was coupled to a LTQ/Orbitrap ion trap mass spectrometer (Finnigan, San Jose, CA). Peptide identification was performed using the SEQUEST Algorithm. For database searches, the sequence of the allergen Bet v 1 was included in a Swiss- Prot deduced database. Peptides were evaluated according to their quality requirements based on the sequence variables cross correlation (X<sub>corr</sub>) and delta cross-correlation (dCn). Only peptides with a dCn \>0.1 and a cross correlation of X<sub>corr</sub> \>1.8 for singly charged ions, X<sub>corr</sub> \>2.3 for doubly charged ions and X*<sub>corr</sub> \>2.8 for triply charged ions were considered. Due to limited amount of available cells and thus limited amounts of sample, multiple injections were not feasible. Therefore each sample could only be measured once. Nevertheless, most of the Bet v 1-derived peptides were detected several times during one run, each resulting in a separate mass spectrum. In addition, within one sequence clusters usually several length variants were identified and most of the presented peptide clusters were present in several donors with the same HLA-types. This confirmed the identity of the presented peptide stretches. As a negative control and to rule out false positive identification of Bet v 1-derived peptides, cell- derived peptides from mature unloaded DCs were used. ## Determination of HLA-DRB alleles Low resolution determination of the HLA-DRB1 tissue type was performed for each blood donor using a commercial SSO typing kit (DynalRELI SSO HLA-DRB Typing kit, Invitrogen, Bromborrough, UK). Low resolution analysis of the HLA-DRB3/4/5 tissue type was performed using a ligation based typing approach. High resolution analysis of HLA-DRB1 and HLA-DRB3/4/5 tissue types was performed by nucleotide sequencing of exon 2 (BigDye Terminator Cycle Sequencing Kit, ABI, Foster City, CA). In addition, for high resolution typing of DRB4 alleles, a commercial SSP kit was used (Olerup SSPTM DRB4, Qiagen, Hilden, Germany). ## Patient selection All birch pollen-allergic patients had a typical case history, specific IgE RAST/CAP class \>3 to Bet v 1 (Pharmacia Diagnostics, Uppsala, Sweden), and positive skin prick reactions (wheal diameter \>5 mm) to Bet v 1. All patients gave written consent before enrolment in the study, which was approved by the Ethics Committee of the Medical University of Vienna. ## Allergen-specific T cell lines (TCL) PBMC were isolated from the blood of birch pollen allergic donors by Ficoll- Hypaque density gradient centrifugation. For the generation of allergen-specific T cells, 1.5×10<sup>6</sup> PBMCs were stimulated with 10 µg/ml unmodified Bet v 1 or Bet v 1 nitro in 24-well flat-bottom culture plates (Costar, Cambridge, MA). On day 5, 10 U/ml human IL-2 (Boehringer Mannheim, Mannheim, Germany) were added. On day 7, T cell blasts were enriched by density gradient centrifugation and cultures were expanded at weekly intervals with irradiated PBMCs and IL-2. Before the experiment, T cells were rested for 10 to 14 days. To test allergen specificity, the T-cell lines (TCL) were stimulated in the presence of 5×10<sup>4</sup> irradiated autologous PBMCs for 48 h in duplicate with varying concentrations of unmodified Bet v 1 (1.25 µg/ml, 2.5 µg/ml, 5 µg/ml) and equimolar amounts of Bet v 1 nitro, respectively. After a 16 h pulse with 0.5 µCi of \[<sup>3</sup>H\]TdR (GE Healthcare, Munich, Germany), cultures were harvested and radionuclide uptake was measured by scintillation counting. The stimulation index (SI) was calculated as ratio between counts per minute (cpm) of TCL, PBMCs and antigen, and cpm of TCL and PBMCs only. Delta cpm (dpm) were calculated as cpm of TCL plus PBMCs plus protein minus cpm of TCL and PBMC only. TCL were considered as specific when the SI was \>2.5. ## Statistical analysis T-cell proliferation data were analyzed by Wilcoxon Signed Ranks Test using the software package SPSS (SPSS Inc., Chicago, IL, USA). # Supporting Information The authors thank Dr. Martin Himly for providing data concerning the aggregation status of the allergens and Dr. Gabriele Gadermaier for performing the amino acid analysis. We also thank Ingrid Faé for HLA-typing and Nikolaos Berntenis for generating the fragment spectra figures. [^1]: Conceived and designed the experiments: AK AV AD. Performed the experiments: AK GO SM. Analyzed the data: AK SM GF PL. Contributed reagents/materials/analysis tools: GO FF BB PL. Wrote the paper: AK GO AV. [^2]: Current address: Novartis Pharma AG, Basel, Switzerland [^3]: The authors have read the journal's policy and have the following conflicts: Barbara Bohle and Fatima Ferreira are supported by Biomay AG, Vienna, Austria. Anne Vogt is an employee of F. Hoffmann-La Roche Ltd. Anette Karle was an employee of F. Hoffmann-La Roche Ltd when she performed the experiments outlined in the present manuscript. The other authors declare no conflict of interest. This does not alter the authors' adherence to all the PLoS ONE policies on sharing data and materials.
# Introduction Acinetobacter baumannii (A. baumannii) is mainly implicated in hospital infections and is responsible for 80% of the Acinetobacter infections. A. baumannii can also be found on normal human skin, but it generally does not pose a threat to a healthy person, besides the not-so-frequent skin and soft tissue infections, infections in the surgical site, urinary tract infection, etc. In the past 30 years, A. baumannii has evolved into a multidrug resistant (MDR) opportunistic pathogen that selectively infects seriously ill patients in intensive care unit (ICU), trauma or burn patients. The presence of intrinsic efflux pump and high rates of genetic adaptation, contributes to adaptation against the antibiotics. Besides, it also possesses several beta-lactamase genes which offer resistance against beta-lactam antibiotics. A. baumannii has also been developing resistance against carbapenem which had been one of the last line of drugs against it. Combination therapies such as of colistin, polymixin B, and tigecycline are used to treat MDR strains, but these are complex compared to a single drug when it comes to quantification of the effect and the validation of their safety. Due to the growing concern about MDR, new types of antimicrobial agents are needed. Antimicrobial peptides (AMP) are a fundamental part of the innate defense system and are reportedly present in organisms from bacteria and fungus to humans. Although several modes of AMP activity, including DNA damage, RNA damage and targeting ribosomes regulatory enzymes or other proteins have been proposed, it is generally believed that the positively charged AMPs act by disrupting the bacterial membrane and the membrane disruption is one of the key factor for the AMP activity. Because of this fundamental difference in the mechanism compared to the traditional drugs, it is believed that the bacteria do not develop resistance easily against AMPs. The low toxicity of AMPs towards human cells and their tendency not to result in resistant strains makes them an ideal rational choice as the next generation antimicrobial agents, possibly eventually becoming effective drugs for A. baumannii. Quantitative Structure and Activity Relationship (QSAR) is an approach in computer aided rational drug design, which uses biophysical or biochemical parameters of the molecules to develop a quantitative relation with the measured activities. Once validated, the computational model can be used for predicting the activities of the possible drug candidates and for pre-screening them. Recent studies have developed a QSAR relation using 29 small molecule drug candidates which act on the oxphos metabolic path of A. baumannii. As noted above, since bacteria are less likely to develop resistance against AMP based drugs, we focus on QSAR for AMPs against A. baumannii. The present work has three major objectives. Several experimental groups have independently evaluated the activity of AMP against A. baumannii. We curated these experimental results against a single, well studied target, ATCC 19606 strain, whose activity is quantified using Clinical and Laboratory Standards Institute (CLSI) or related protocols. We developed a computational model using neural networks to rationally predict the activity from the biochemical attributes of the AMP. Since A. baumannii is a growing threat, while realizing the potential limitations of training on 75 peptides, we also predict the activity of all the naturally occurring AMPs in the AMP database to enable a rational screening of AMPs against A. baumannii. # Methods ## Curation of data Training QSAR models with data from multiple sources, obtained with different protocols and on different strains can lead to poor predictive capabilities. In order to standardize the data used in the analysis, we used three criteria for inclusion- the tests should be on ATCC 19606 strain, with cationic antimicrobial peptides and studied according to the CLSI or equivalent guidelines. With these inclusion criteria, we believed that the mechanism of antibiotic action will be similar and the data curated from different sources can be compared. Since data availability was limited, we had to include data from different groups. AMP sequence and activity data against A. Baumannii was curated from different sources and is presented in Table A in. The curated AMP data set had the activity of 75 AMPs with their length ranging from 10 to 43 amino acids and charges in the range +1 to +12. Of these, for 63 AMPs the MIC was available (referred to as quantitative data), and for the remaining 12, only the lower bound of minimum inhibitory concentration (MIC) (refered to as the qualitative data). ## Parameter computation in vivo aggregation propensity is calculated by using a web-based software AGGRESCAN. Where the aggregation propensity is calculated on the basis of aggregation- propensity scale of amino acids. in vitro aggregation propensity is calculated by using TANGO software (with ionic strength 0.02M, pH 7.0 and T = 298K), where we only consider the β-sheet aggregation term. Aliphatic index of the peptides is calculated as described by Ikai. Grand average hydropathy is calculated on the scale given by Kyte-Doolittle and the hydrophobic moment is calculated by using HELIQUEST software. The toxicity of the AMPs was predicted using ToxinPred (<http://crdd.osdd.net/raghava/toxinpred/>). The method allows for the prediction of toxicity of peptides shorter than 50 amino acids. However, this was not a limitation as peptides longer than that are anyways complicated to synthesize and may not be ideal drug candidates. ## Artificial neural network Since the available data is limited, we used used both the quantitative and the qualitative data, albeit with different proportions, to train and test the models. We used 63 of the MIC values from the quantitative data and 3 from the qualitative data for which the cited lower bound was treated as the MIC for the purpose of this analysis. We performed a 10-fold cross validation to check the robustness of our models. To do the 10-fold cross validation, we divide the data set into 10 different test sets, each contains 7 data points. We performed the artificial neural network (ANN) calculation for each test set by taking 53 data points for training and 6 data points for validation. Rest of the 9 points from the qualitative data are used for an independent qualitative test. The activity of the AMPs was predicted by ANN model with an open module for machine learning called Scikit-learn in Python. For the activation function, logistic function was used and low memory BFGS optimization algorithm was used a solver. Three independent neural network calculations have been performed to do the 10-fold cross validation, by using a hidden layer of 6 neurons, 8 neurons and 10 neurons. 2500 trial runs in each case were made by taking 50 different random initializations for the input biases and 50 random choices for the training and validation sets. We screened the results of these 2500 trials with $R_{training}^{2} > 0.7$ and $R_{validation}^{2} > 0.6$. Two best models were selected based on the result obtained from the 10-fold cross validation. The models were expected to perform with $R_{test}^{2} > 0.8$ for the quantitative data and at least 5 predictions for the qualitative data set. These models were then used to predict the MIC values of a complete AMP database (<https://aps.unmc.edu>). # Results ## Curated data for AMPs and their effectiveness The data on the activity of AMPs on A. baumannii is scattered in literature. We curated the data mainly with the goal of developing a quantitative model, and hence restricted the focus to the most commonly studied ATCC 19606 strain. To maintain uniformity of standards, we included studies which were performed according to CLSI or equivalent guidelines. The sequence data and the antimicrobial activity of these peptides measured as the MIC was gathered (Table A in). Overall, the comprehensive collection of the data on AMP activity allowed a classification based on the various biophysical parameters which are commonly used for developing a quantitative relation with activity: (1) charge, which draws the AMPs selectively to anionic membrane, (2) length, reflecting how it has to be commensurate with the membrane thickness for an improved activity (3) molecular weight, which gives an idea of the bulkiness and membrane penetration efficiency (4) hydrophobic moment (μ<sub>H</sub>), which quantifies the amphipathic characters required to form pores in the membrane, (5) aliphatic index, which indicates the volume of aliphatic content (A, V, I and L) of the peptide, (6) grand average of hydropathy (GRAVY) based on Kyte-Doolittle hydropathicity scale, (7) in vivo aggregation propensity, calculated by using a web-based software AGGRESCAN and (8) in vitro (β-sheet) aggregation propensity, calculated by using TANGO software (with ionic strength 0.02 M, pH 7.0 and temperature 298 K). The in vitro aggregation, before interaction with the membrane can at times stop proteolytic degradation by the bacteria but in many other cases reduce the drug potency \[, \]. Further, the aggregation propensity affects the barrel-stave and carpet mechanisms of action differently. Toxicity of peptides obtained from ToxinPred was categorical, and it was used only to classify the AMPs from the database as potential drug candidates, and not for the activity prediction. The distribution of the eight parameters for all the curated AMPs are given in Fig A in and their individual relation with MIC in, which shows that each of the parameters individually is not sufficient to describe the activity. ## Quantitative models for AMP activity ANN model was used to obtain the relationship between the various above- mentioned parameters and MIC values (Methods). A schematic of how we developed the model is shown in Fig B in. The first step was to create a model with the activity data from 75 AMPs, of which some were used for an internal assessment of the quality of predictions. The second step was to use the test set in the 75 AMP data analysis as a secondary validation for refining the choice of model that can be used for making the predictions for the AMP database. The details are as follows. Out of the 75 AMPs curated, for 12 of them a lower bound of MIC, as being greater than a certain value (Table B in), rather than a precise number was cited. To include them in the analysis, and not to reduce the data size which is already small (75 AMPs), we created two independent test sets, one in which a quantitative MIC comparison was made (referred to as quantitative data) and another qualitative one in which the calculated MIC was checked if it was more than the experimental lower bound (referred to as qualitative data). The combined data set with quantitative and qualitative data was used to construct training, validation and test sets (Methods). We performed a 10-fold cross validation with three different architectures with 6, 8 and 10 hidden neurons respectively. The overall error in the architecture with 8 neurons was optimal, thus justifying a small sampling around it with 6 and 10 neurons (Table C in). However, all three architectures were satisfactory in their predictions (Figs C, D and E respectively in), resulting in many models, which qualify for the criteria ($R_{training}^{2} > \mspace{360mu} 0.7$ & $R_{validation}^{2} > \mspace{180mu} 0.6$). Several of these models also had good predictions for the test sets, which are about 10% of the data. ## Selecting the best model In a traditional QSAR analysis, the choice of the best model would be guided by the combination of the best $R_{training}^{2}$ and $R_{validation}^{2}$, following which $R_{test}^{2}$ on a small fraction of the data, in our case 7 data points, comes as a consequence. Since the goal of screening through the large set of potential AMPs whose activities against an extremely important pathogen are not yet available is more ambitious than performing well on these 7 points, we performed a secondary validation check to select the best models. We used two additional criteria: $R_{test}^{2} > 0.6$ for the quantitative and that at least 5 predictions in qualitative data set were correct to within a factor of 2 (Table B in). Two models satisfied these conditions, with $R_{test}^{2} > 0.8$ and they were selected. The best among these models (referred to as Model-1) obtained from the calculation with 8 hidden neurons, had good predictions ($R_{training}^{2} = 0.975$, $R_{validation}^{2} = 0.866$ and $R_{test}^{2} = 0.827$). The experimental MIC for the quantitative data set versus MIC values predicted from Model-1 is shown in. Results obtained from another model (Model-2) are given in Fig F in. ## Predicting the results for naturally occurring AMPs Considering the health threat A. baumannii is posing, and the potential of AMPs for antibiotic-resistance-free activity, we propose a rational basis for an in silico screening of AMPs active against A. baumannii. Our models were used to predict the MIC values of the 2338 AMPs obtained from database (<https://aps.unmc.edu>) of naturally occurring AMPs. We made the predictions from Model-1 and Model-2. In order to reduce the risk of a poorly trained ANN model with limited data, we filtered these results for a consistent prediction that is within ΔMIC ≤ 5 μg/ml for both the models. Despite the potential statistical limitations of training and validating on 75 AMPs, a pre-screening to rationally sort multiple AMPs with their predicted activity, in vitro and in vivo aggregation potential, toxicity and length (a surrogate for synthetic complexity), all are provided in and in the. The computational scripts and the predictions are made accessible, to provide an immediate access to a pool of rational choices that can help progress towards large scale experimental testing, considering the extreme urgency of developing effective strategies to combat the superbug, A. baumannii ## Parameter importance in model It is important to know which are the parameters (P<sub>i</sub>) that are most responsible for the activity on A. baumannii. In the combined training and validation set used for accepting the models, we replaced (P<sub>i</sub>) with its average \<P<sub>i</sub>\> and measure the difference $\Delta R_{P_{i}}^{2} = R_{training + validation}^{2} - R_{training + validation, < P_{i} >}^{2}$. $\Delta R_{P_{i}}^{2}$ is treated as reflecting the importance of the parameter. The results obtained from Model-1 are given in and the result obtained from another model is given in Fig G in. From our calculations, we found out that the aliphatic index is the most important parameter in both the models. ## Relevance of predictions for MDR strains In order to reduce the uncertainties, our computational model was trained on data standardized in three ways, A. baumannii strain used, choice of cationic AMPs and measurements by CLSI method. However, considering the threat that A. baumannii MDR strains are posing, it is important to ask whether our calculations have any relevance to these clinical variants. The two limitations of this work are the smaller data size used for training, and it was based on ATCC 19606 strain. Interestingly, in the limited studies that we found the activity of cationic AMPs against ATCC 19606 and other MDR strains of A. baumannii are comparable, thus potentially removing the latter strain specific data limitation for A. baumannii, although for other bacteria, such as S. aureus the activity changes quite significantly with the strain. Drawing confidence from this fact, we used our models to predict the activity for a few MDR strains. The results reported in Table D in are encouraging at this stage, although more such validations will be helpful in establishing the utility of the screening models we proposed. ## Conclusions To our knowledge, the present work is the only QSAR study for predicting AMP activity against A. baumannii. The present work is different from the only other QSAR in two different ways, using AMPs instead of small molecules for a better tolerance to antibiotic resistance and a slightly larger set (75 AMPs compared to 29 small molecules). Using the ANN models we developed, we could make quantitative predictions for the entire database of naturally occuring AMPs. We hope that our work will inspire the further studies quantifying the activity of AMPs on A. baumannii, some of which may follow the activity predictions and others that differ offer an opportunity to retrain the ANN models. # Supporting information A.M. wants to thank Swagatam Barman for scientific discussions. M.K.P. wants to thank JNCASR for stipend. [^1]: The authors have declared that no competing interests exist.
# Introduction Renewable energy has been studied for a long time due to high electricity generation costs and global warming. Sustainable energy production plays a vital role in reducing global warming and the impact of climate change. Therefore, the energy sector is now more focused on seeking the appropriate tools for the widespread use of renewable energy resources. Global demand for renewable energy rises to 30 percent relative to 435 installed capacities, with an annual growth of 17 percent in recent years. In 2020, renewable energy provided an estimated 12% of total energy output worldwide. Wind power consumption has increased nearly twice, while biogas consumption has increased by 2.8% over the past five years. As a result, overall greenhouse gas emissions are decreased by 3% and environmental damage by 23%, respectively. Estonia has taken steps to expand the use of renewable energy in the form of the Short Assessment of Alternative Energy Sources (AES) by 30%, 45%, and 80% between 2020, 2030, and 2050. Renewable wind power is further categorized into short-term and long-term wind power production. Short-term wind power varies from few minutes to a day, while long-term wind power is based on weeks, months, or even years. Both short-term and long-term wind data can be utilized to forecast wind power generation. However, short-term wind power forecasting is deemed preferable compared to long-term power forecasting. Power sectors of different countries use various wind power forecasting methods to generate sufficient wind energy. These methods largely focused on either mathematical or physical models, whereas the hybrid models rely on the integration of both models is also popular in the energy sector. Physical models can be affected by particular characteristics such as wind turbine position, barriers, surface coarseness, and blade turbulence, etc. Physical models are specifically used for long-term wind power forecasting. Contrary, the mathematical models use historical wind data to estimate potential power generation and may also take advantage of various hybrid frameworks such as machine learning and deep learning-based power forecasting. Common forecasting methods include Support Vector Regressor (SVR), Multi-layer perceptron, deep neural network, as well as a combination of various related methods. Wind power has uncertain characteristics and weak controllability, which also raises the problem of inconstancy and fluctuations in power systems. Apart from that, air velocity may also be influenced by the possibility of hazardous modes and postures. Therefore, an effective power monitoring system is needed for efficient transmission and to resolve the power generation demands. Various estimation strategies have been introduced and structured to examine and estimate sustainable wind power. Wind power forecasting strategies are generally classified into three theoretical classifications, including Forecast Analysis (NWP), Survey Data Analysis (SDA), and a mixture of the two approaches. Probabilistic units are typically known to estimate the relevant NWP since these approaches have become more and more significant at periodic intervals with better accuracy. However, it would be very difficult to develop an efficient numerical framework without any in-depth analysis of the systems engineering as well as the wind area atmosphere. These types of models need to construct certain variables with the aid of illustrative variables, census data algorithms, to overcome the association between empirical evaluation and extracted wind properties. These methods may only need observational data for estimates and are also fairly convincing to widespread forms of research implementation. Despite their significance, the accuracy of these models has been declining over time. Nowadays, most of the statistical techniques are used in collaboration with the Artificial Neural Network (ANN), Convolution Neural Network (CNN), Support Vector Regression (SVR), and Back Propagation Neural Network (BPNN) respectively. Several data science algorithms have been developed for complex database systems to efficiently use and train data to solve forecasting problems in multiple domains. Such algorithms use specific behavior of data and afterward generates predictive outcomes. CNN is a supervised learning algorithm that uses a perceptron to extract estimations from adaptive learning. Alternatively, BPNN is also a supervised learning algorithm that determines the cumulative marginal distribution of feed-forward backpropagation in a fully connected neural network. The wind turbine data is based on a time series with a real-time wind speed reading for a particular area. Wind power forecasting models mainly use these readings along with other characteristics to predict potential power output for that area. For short-term forecasting, minutes and hourly time series data are more accurate as they can be used for stochastic wind signals and have been used in many models such as Kalman filters, Box-Jenkins, as well as ANN. The Box- Jenkins model uses probabilistic approximation and cannot be used to predict future wind power. Therefore, their model involves a full assessment of the distinct forecasts. The Kalman filter is considered to be a direct symmetric estimator, as it uses a limited direct estimation feature relative to Box- Jenkins and thus requires less computational costs. As a result, Kalman filters are steady and needless computation, but Box-Jenkins provides more effective results with a significant number of simulations. However, considering the risk of diverse knowledge and extremely nonlinear wind behavior, the ANN has gained more momentum in handling complex and dynamic information. ANN estimates forecasting based on neural structures such as multilayer perception with formal specifications that may not be fully suitable for wind power forecasting. The predictive potential of these approaches falls beyond the longer forecast horizons. As mentioned, our proposed deep learning model based on combined approach of CNN and LSTM which has a significant advantage over certain state- of-the-art approaches. The main contribution of this paper is as follows: - First, we use the NREL toolkit to analyze and extract offshore wind power data. The collected data is further divided into three regions that span large-scale offshore wind turbines in the United States. - Second, wind power data is primarily based on time series, and LSTM is well- suited to handle time series forecasting regardless of duration lags. Thus, a combined approach of CNN and LSTM is used to analyze hidden features of offshore winds for reliable wind power forecasting. - Next, data is preprocessed using deep auto-encoders prior to wind power forecasting to minimize the error ratio. Offshore winds are diverse and continuously changing, whereas deep auto-encoders can learn low-dimensional data efficiently and with minimal reconstruction error. - Furthermore, the proposed CNN-LSTM forecasting model is optimized using fine-tuned parameters for each offshore region, and the performance is compared to the current state-of-the-art methods. - As per empirical analysis, the proposed method demonstrated excellent forecasting accuracy with a low error ratio at various intervals, making it more suited approach for offshore wind power forecasting. # Literature review Current wind power forecasting technologies have limitations and needed to be enhanced to estimate actual wind power generation. It is essential to investigate the different perspectives of wind power to design an effective model. Most recent researches have been focused on hybrid techniques to take advantage of integrated methods. Individual approaches usually produce low efficiency than hybrid approaches, while the hybrid approaches can be further divided into key parts. A hybrid approach combined with computational intelligence can be more effective and widely studied in energy and power forecasting. For instance, Kaluri et al used predictive power of rough sets to forecast battery power life and, Maddikunta el al used hybrid algorithmic strategy for efficient power communication between networks. A hybrid computational intelligence technique can learn effective model features from wind direction and wind speed adjacent to wind turbines before actual wind power forecasting. The outcomes of hybrid deep learning approaches recommended that the multiple neural networks must efficiently enhance short-term wind speed analysis for effective forecasting. Long Short-Term Memory (LSTM) is particularly used in different adapted distinctions with high basic criteria. For instance, Alazab et al used multi-directional LSTM model to forecast stability of power in smart grids. However, if the interactive mechanism is not needed, it still needs a lot of resources to converge neural networks quickly as the existing improved version does not match hardware power acceleration. Lu et al proposed an encoder-decoder LSTM model for wind power estimation by mapping wind power time series data into fixed-length representations. The results showed that the LSTM auto-encoder based preprocessing can perform better compared to simple LSTM wind power forecasting. The Fourier representation of linear and steady-state analysis approaches focus on optimal filtering using Variation Mode Decomposition (VMD) and Single or Multi-Kernel Regularized Pseudo Inverse Neural Network (MKRPINN). An Ensemble Empirical Mode Decomposition (EEMD) further decreases the impact of projected modeling of pseudo wind power data which is also scalable to data reduction without any need for further specifications. Different parameter fine-tuning has the additional advantage to acquire useful data characteristics. The one-hour forecasting of seven locations of ground radiation showed that the wind power forecasting performance can be improved by 30% compared to traditional benchmark estimates. Fu et al presented a multi-step ahead technique based on RNN combined with LSTM or GRU also known as Gated Recurrent Unit to improve wind power forecasting accuracy. Shao et al further combined Infinite Feature Selection (Inf-FS) with RNN to overcome short-term wind power forecasting problems. Analysis of data from the US National Renewable Energy Laboratory (NREL) reveals that the efficiency of short-term wind energy forecasts in spring, summer, autumn, and winter has been greatly improved by 11 percent, 29 percent, 33 percent, and 19 percent respectively. However, the RNN has one weakness associated with the high power of the matrix caused by its vanishing gradient. As a result, it is difficult to determine the long-term dependence of the time series in wind power forecasting. However, it has been found that the average error rate and the maximum error rate of LSTM-RNN are lesser than other methods. Lai et al proposed Long and Short Term Times Series Network (LST-Net) to resolve the accuracy problem of time-series forecasting. The empirical evaluation showed that the earlier evaluation of repeated trends in time series data can improve the overall forecasting accuracy. A conceptual hybrid method associated with Wavelet Packet Decomposition (WPD) is proposed by Liu et al, which is a combination of CNN and LSTM to minimize the effect of non-stationary raw data for short-term wind power forecasting. WPD decomposes initial wind speed data based on time series into multiple sub-layer levels where CNN and CNN-LSTM were used for high-frequency and low-frequency evaluation. The empirical evaluation of various test samples has shown that hybrid modeling of various wind speeds can improve the precision of short-term wind power forecasting. Chen et al presented another hybrid approach based on LSTM and an evolutionary algorithm named ELM for wind speed and power forecasting. The empirical evaluation of the four evaluation parameters found that the preferred hybrid approach had achieved the predicted outcomes from the estimation criteria, while the projected results had an actual benefit over the forecasting accuracy. Adopting several models as the mainstream model can only improve the aggregate average forecasting performance. Khodayar et al proposed a graph-based deep neural network to analyze unguided high wind positions consisting of LSTM and graph based CNN model to overcome short-term wind speed forecasting errors. CNN-based deep learning models are adaptable and have been widely used in forecasting studies. For instance, Vasan et al used an ensemble CNN model to improve forecasting accuracies. The deep learning architecture (GCDLA) captures the temporal characteristics of the wind forecasting process. A study of GCDLA findings has shown that hybrid approaches paired with separate neural networks can improve the predictive accuracy of wind speeds. Chang et al developed an improved deep learning network by selecting input features with suitable interfaces and established two types of probabilistic forecasting models. In previous studies, it was reported that the absence of appropriate input features and lack of analysis and selection causes a negative impact when forecasting is applied to high-dimensional wind power data from multi-input. This requires a huge number of computational resources that can also influence usability. Wang et al suggested an ensemble approach to resolving such probabilistic wind power forecasting issues. Time-frequency classification including distinguishing feature extraction for various wavelet transformations along with CNN is processed to achieve desirable estimation. The empirical evaluation showed that some uncertainties in wind power data can be identified using ensemble approaches. Furthermore, Xiao et al proposed hybrid modified architecture centered upon Bat Algorithm (BAT) in combination with Conjugate Gradient (CG) process in order to forecast wind speeds. In recent studies, Lin and Liu used high frequency SCADA data to forecast wind power of Levenmouth offshore wind turbine. The outliers were detected and removed from data using isolation forest filtering while a deep learning model is finetuned to forecast offshore wind power. Although the isolation forest filtering needs limited memory, it does not employ any distance or density measures to eliminate outliers. As a result, the deep learning model may be affected when applied to multiple offshore wind turbines. Zameer et al used an ensemble ANN model along with genetic programming to overcome instability problem of wind power. The proposed approach was tested on five windfarms and produced reasonable outcomes. Devi et al also used ensemble strategy mainly focused on improving the forecasting performance using LSTM-EFG model combined with cuckoo search optimization and ensemble empirical mode decomposition. In general, the ensemble methods are quite helpful in improving forecasting results by combining multiple models but their improved performance is due to the reduction in the variance component of forecasting errors generated by the participating models. Yildiz et al used variational mode decomposition (VMD) to convert wind power features into RGB images. The image data was then used as input for CNN model to perform short term power forecasting. The accuracy of image-based forecasting is effective but image pixels are not true geographical objects due to limited pixel topology. Therefore, spectral and spatial effects of images on CNN may further needed to be elaborated for effective wind power forecasting. Acikgoz et al used extreme machine learning (EML) strategy to forecast terrain-based wind power on one year data in turkey. The empirical evaluation of seasonal performance showed improvement compared to tradition ANN with minimal forecasting errors. However, the training and testing were conducted on k-folds, which means that the training algorithms had to be ran many times, which may be computationally expensive for large datasets. Niu et al used sequence-to-sequence modeling and attention-based GRU network to improve accuracy and stability of traditional wind power forecasting. GRU is an alternative and improved version of LSTM but computational time of deep learning process increases as attention mechanism applied to wind power forecasting. The summarized comparison of recent studies on wind power forecasting is shown in. As discussed in the literature study, classifier training is necessary from the beginning for offshore wind power forecasting, which is also a time-consuming practice. Secondly, ensemble-based deep learning strategies for wind power forecasting utilize two or more DNN models, with one predictor propagating forecasting outcomes while others improving the forecasting error. Additionally, the wind speed is more diverse at offshore regions compared to land-based installations, thus observations may contain a significant quantity of outlier values. As a result, eliminating all outliers may have a negative effect on the reliability of wind power forecasting strategies. Instead, our proposed model employs deep auto-encoders to efficiently learn low-dimensional data ensuring minimal reconstruction error. Furthermore, training several layers with a deep autoencoder is more efficient than training single huge transformation based on dimension reduction schemes. As a result, the neural network’s time-consumption problem can be handled more precisely. In particular, a deep auto-encoding based hybrid CNN-LSTM model can facilitate rapid feature analysis and estimation of wind power forecasting. # Overall methodology for wind power forecasting Due to the turbulent and diverse real-time behavior of wind signals, predicting accurate wind power is a difficult task. In order to evaluate the predictability of wind power, the raw wind data is obtained from three separate offshore Windfarms. Next, the raw wind data is analyzed and the valuable variables are chosen that can hold worthy information from the vast range of offshore wind turbines. In all three data sets, a sequence of five Deep Auto-Encoders (DAC) is used to exploit hidden information and focus on meaningful features that can facilitate the predictability of deep learning models. Our proposed deep learning wind power forecasting methodology can efficiently compress and encode wind data, then realizes data in minimized encoded form towards a classification model. Furthermore, auto-encoders has been applied for dimensional reduction that can effectively improve forecasting accuracy. The trained model is then transferred to the deep learning model based on Convolutional Neural Network (CNN) combined with LSTM to predict wind power on all three offshore Windfarms. Finally, the efficiency and effectiveness of the proposed model over different offshore Windfarms are evaluated in terms of prediction accuracy and forecasting errors as shown in. ## Deep Auto Encoder (DAC) DAC is a simple deep auto-encoder and feedforward artificial neural network consisting of three basic layers, i.e., the entry layer also known as the input layer the hidden layer responsible for applying weights using the activation function, and the output layer to produce final results. Each node in the neural network except input nodes is a neuron that utilizes a non-linear activation function. Several levels, including non-linear stimulus, discriminate against the Hidden layer by a standard DAC classifier to differentiate non-linearly separable data. DAC auto-encoders each with one hidden layer and hidden weights minimize the squared error with a quadratic penalty (weight decay). Contractive auto-encoder (CAE) is also an alternative to the weight decay DAC strategy. We used a specific normalization that has the advantage of robustness with slight deviations around the test points to favor interfaces that converge more precisely on the training set. Like the Contractive auto-encoder, the DAC auto- encoder can obtain a comparable measure of regularization on objective function distribution as shown in. <img src="info:doi/10.1371/journal.pone.0256381.e001" id="pone.0256381.e001g" /> J C A E = ∑ x ∈ D n ( L ( x , g ( f ( x ) ) ) \+ λ ∥ J f ( x ) ∥ F 2 ) Where the approximation error is L, λ hyper-parameter controls the regularization power, $J_{f}(x) \parallel_{F}^{2}$ promotes the estimation with training data across suburb as a constructive for feature space. ### Correlation between DAC and weight decay Since it becomes obvious to assume Frobenius standard of Jacobian squared corresponds towards an L2 weight decay (for instance, an activation function s<sub>f</sub>). Having lower weights through a static scenario is also the best solution towards reduction. However, in the case of the sigmoid non-linear activation function, reduction and robustness can also be achieved by placing hidden units into their saturated state. If an encoder represented as f function transforms the source $x \in R^{d_{x}}$ to a hidden $h(x) \in R^{d_{h}}$ representation, then the resulting form can be expressed as. <img src="info:doi/10.1371/journal.pone.0256381.e005" id="pone.0256381.e005g" /> h = f ( x ) = s f ( W x \+ b h ) Here s<sub>f</sub> is a type of sigmoid non-linear activation function, while the d<sub>h</sub>×d<sub>x</sub> weight matrix W, and a bias vector $b_{h} \in R^{d_{h}}$ are parametrized towards the encoder. The key features of the proposed auto-encoding strategy are as follows: 1. It renders the encoding less vulnerable from its training dataset to slight alterations. 2. Encoding is achieved by utilizing a regularization or penalty scheme to an objective function. 3. The overall outcome is to minimize the sensitivity of the learning representation against the training feedback. 4. Encoder activation sequences are regularized and must comply with the Frobenius norm of the Jacobian matrix concerning input data. 5. The DAC auto-encoder is generally used similar to other auto-encoders by activating only when the data point is not labeled by other encoding schemes. The unpredictable and unstable nature of the wind makes it very difficult to extract certain wind patterns in order to precisely forecast reliable wind energy. The instability of wind contributes to the wide range of training samples that have a significant effect on the precision of power prediction. First, a group of five DAC deep auto-encoders is used in all three Offshore Windfarm datasets to extract hidden features and meaningful data patterns in a low-dimensional space. DAC auto-encoder is an unsupervised neural network that learns how to compress and encode information efficiently, and then learns how to reconstruct the features of a reduced encoded representation roughly similar to the actual input. The dataflow diagram of the proposed DAC based Deep Auto- Encoding framework is shown in. In the proposed auto-encoder framework, the dimensionality reduction strategy is also used for data to further improve the efficiency of wind power forecasting during pre-processing wind data. ## Offshore Windfarm (NREL) dataset Windfarms data is pre-processed based on offshore wind turbine data from U.S. NREL repository consisting of multiple offshore regions. Datasets provide wind potential and real-time measurements of wind turbines for certain intervals located in different offshore regions. The first dataset is processed based on a scale of 30 meters to 90 meters with a measurement duration of 1 hour, 7 hours, and 12 hours, including the climatic data of 164,000 Windfarms. The second Windfarm data is based on the Hawaii region of the United States, with an average grid of 2 kilometers recorded during the month of January. Additionally, the third Windfarm data is based on an offshore wind analytics database compiled via different wind speed design parameters. Moreover, the real-time wind data are evaluated for 17 years only by the MERRA time series, and the various meteorological specifications. shows the annual wind speed measured by the NREL on different offshore regions of the United States. Higher wind speeds indicate the potential of power generation by offshore wind turbines in the United States. A significant number of offshore windfarms are located in these regions, that are also utilized in wind power forecasting analysis. ## CNN-LSTM neural network architecture The pre-trained model is then transferred towards a CNN-LSTM neural network for offshore wind power forecasting. The CNN-LSTM framework is used as the primary indicator throughout this research. The proposed Convolution Neural Network strategy has succeeded in reducing the effect of computational complexity and has also achieved significant improvement in extracting and generalizing features. LSTM is capable of processing 1-decision statistical analysis as well as in making assumptions by generating an outcome for each timespan. In order to isolate the features of precompiled results, a 1-D convolutional layer along with the LSTM framework is implemented in this paper as a major indicator. The designated CNN-LSTM configuration with the addition of CNN layers in the LSTM model is shown in. The LSTM layer along with two dense layers and three fully connected layers is configured, while max-pooling is also used within the hybrid CNN-LSTM model. Through multiple tests, the activation function of each convolutional layer is calculated against Parametric ReLU (PReLU), then the sigmoid function is estimated to activate that LSTM layer. Moreover, zero paddings have been used only between different components of convolution layers to preserve a certain proportion. Indeed, the convolution layer is the main focus of the convolution neural network. Within this layer, there are also two main processes, local independent correlation and temporal classification of outputs. To further simplify the basic calculations, the former calculation is also adopted for the calculation of each filtered relevant information. Although described, the performance of the convolution layer may be enforced: $$h^{k} = f((W^{k}x) + b_{k})$$ Where \ signify convolutional process, f (·) is the activation function, W<sup>k</sup> and b<sub>k</sub> are the weights and biases of the k<sup>th</sup> function. ## LSTM (long-short term memory) network LSTM is more resourceful relative to conventional Recurrent Neural Network (RNN). However, based on the input gate, the forget gate, and even the corresponding output, the same weight of perception appears to be malicious. The defined specific tasks can be observed as follows: 1. How data are excluded again from the convolutional layer is decided by the forget gate. 2. It will change the cell state, but how much additional data throughout the input gate will be applied to the main contribution. 3. A sigmoid function is executed as an output gate, and the cell state is analyzed by an activation function that pairs with the sigmoid output layer to produce the desired total output. Although there four key LSTM parameters such as internal memory state c<sub>t</sub>, forgot gate f<sub>t</sub>, input gate i<sub>t</sub>, and output gate o<sub>t</sub>. The computation phase at each component is represented using Eqs – at period t time: $$f_{t} = \sigma(W_{f}.\lbrack x_{t},h_{t - 1}\rbrack + V_{f}.c_{t - 1} + b_{f})$$ $$i_{t} = \sigma(W_{i}\lbrack x_{t},h_{t - 1}\rbrack + V_{i}.c_{t - 1} + b_{i})$$ $$c_{t} = c_{t - 1}f_{t} + i_{t}\tanh(W_{c}\lbrack x_{t},h_{t - 1} + b_{c})$$ $$o_{t} = \sigma(W_{o}.\lbrack x_{t} + h_{t - 1}\rbrack + V_{o}.c_{t} + b_{o})$$ $$h_{t} = o_{t}\tanh\left( c_{t} \right)$$ For offshore Windfarm datasets, optimum fine-tuned specifications are shown in Tables, accordingly. In the hidden layers, the number of neurons ranges from 50 to 250, whereas the number of epochs varies from 200 to 400. All offshore Windfarm datasets are pre-trained and saved as train models. The training model is then used in the proposed LSTM-CNN deep learning model to predict wind power for all three offshore Windfarms. For the input nodes, the Rectified Linear Unit (ReLU) is used. In deep learning models, ReLU is the most widely selected activation function. This activation function returns 0 if any negative response is obtained, and retains input for any positive value of x. It is numerically represented using. <img src="info:doi/10.1371/journal.pone.0256381.e013" id="pone.0256381.e013g" /> f ( x ) = x \+ = max ( 0 , x ) Where the input of the neural network is x also termed as a ramp function identical to split-wave rectification. The Softmax activation function is used with the output layer to evaluate forecast errors which is also a multi-dimensional generalization of the logistic equation. It can be used for multiple regression analysis and can also be used as the final activation function of the neural network to stabilize the probability distribution performance over the expected output groups. The neural network is generally optimized for output N values for each class of classification model, while the Softmax function can normalize these outputs by transforming the weights to the sum of probabilities. Each value in the output of the Softmax function is interpreted as the probabilities of being a member of each class. It is statistically represented using. <img src="info:doi/10.1371/journal.pone.0256381.e014" id="pone.0256381.e014g" /> σ ( z → ) i = e z i ∑ j = 1 K e z j Where σ* signifies the softmax function, $\overset{\rightarrow}{z}$ is the input vector, $e^{z_{i}}$ is the standard exponential function for input vector, K shows the variety of sections throughout the algorithm for a multi-class, while $e^{z_{j}}$ is a standard exponential function for output vector. In order to build a deep learning model, the entropy function is often used to detect accuracy loss. It considers each tensor as input and targets the same shape of a tensor as an output. The Adam optimizer is used to assemble and configure a deep learning model that is often known as a stochastic descending gradient. Adam Optimizer modifies network weights and determines unique active learning rates for each element of the deep learning architecture. The decaying mean (DM) of the pas squared gradient is expressed using Eqs and respectively. <img src="info:doi/10.1371/journal.pone.0256381.e018" id="pone.0256381.e018g" /> m t = β 1 m t − 1 \+ ( 1 − β 1 ) g t <img src="info:doi/10.1371/journal.pone.0256381.e019" id="pone.0256381.e019g" /> v t = β 2 v t − 1 \+ ( 1 − β 2 ) g t 2 Where m<sub>t</sub> and v<sub>t</sub> are the approximate measures during the first and second moment gradients. As a result, the respective gradient is indicated for each moment. Adam optimizer further counteracts these biases by estimating the bias-corrected first and second moments as shown in Eqs and respectively. <img src="info:doi/10.1371/journal.pone.0256381.e020" id="pone.0256381.e020g" /> m ^ t = m t 1 − β 1 t <img src="info:doi/10.1371/journal.pone.0256381.e021" id="pone.0256381.e021g" /> v ^ t = v t 1 − β 2 t Sparse Categorical Cross-entropy is used to compare the predicted label and true label to calculate the accuracy loss and validation errors. Sparse Categorical Cross-entropy is an arithmetic variant of the categorical cross-entropy loss function which does not require the transformation of target labels into category formats. Mathematically, the relationship between different parameters of the loss function can be expressed using. <img src="info:doi/10.1371/journal.pone.0256381.e022" id="pone.0256381.e022g" /> J ( w ) = − 1 N ∑ i = 1 N \[ y i log ( y ^ i ) \+ ( 1 − y i ) log ( 1 − y ^ i ) \] Whereas w indicates the parameters of the deep learning model such as weights of the neural network, while y<sub>i</sub> represents the true label and ${\hat{y}}_{i}$ becomes the predicted label respectively. Following the formation of the CNN-LSTM model, the number of hidden layers, neurons, and operational parameters are fine-tuned for each offshore Windfarm. Tables – shows the precise fine-tuned settings used in power forecasting of Windfarms. # Results and discussion The Keras Interface integrated with the TensorFlow platform makes it easy to transform layers, activation and loss functions, etc. into a variety of prototype deep learning models. It also offers a compilation strategy for customizing the training process in each layer, including losses, optimization, and other built-in learning configurations to train the constructed model. The training is effectively conducted corresponding to the input design, but also dense layers. The training process is automatically performed by TensorFlow corresponding to the input shape of data and dense layers. Deep learning further drives the integration of user-defined enhanced functionality to train feature representation in a short period to solve complex issues. For all three Offshore Windfarms, the actual and predicted wind power is displayed in Figs – respectively. The orange curve indicates the predicted power by DAC-CNN-LSTM strategy while the grey curve shows the actual wind power generated from US offshore turbines collected by NREL. In, the offshore wind farms predicted values were generated using fine-tuned setting. There is a small error margin between actual and predicted wind power which clearly shows effectiveness of proposed methods for offshore wind power forecasting. Furthermore, the R-squared correlation values of predicted and actual wind power are 0.01 and 0.03 also shows a close association between predicted and actual results. In the case of Figs and, the predicted and actual R-squared correlation of Offshore Windfarm 2 is 0.07 and 0.06, while the predicted and actual R-square correlation of Offshore Windfarm 3 is 0.01 and 0.02 respectively. The observations in all figures reveal that predicted power is quite comparable to actual power, with minor variations at some intervals. However, the overall consistency between the predicted and actual power shows the flexibility and reliability of proposed model in forecasting wind energy across multiple offshore wind turbines. Offshore Windfarms forecasting performance was further assessed in terms of MAE and RMSE errors. The MAE and RMSE errors are widely used to evaluate performance of time-series and other non-label data. In, we used boxplots to show the distribution of errors in the form of minimum, maximum and Q1-Q3 percentile ratios based on outcomes generated by deep learning models. The minimum MAE error for all three Offshore Windfarms is between 0.01 to 0.04, while the minimum RMSE for all three Offshore Windfarms is between 0.07 to 0.15 respectively. The RMSE error of DAE-CNN-LSTM is little higher than its MAE error but the overall effectiveness of proposed model is the same. In the case of Q1-Q3 percentile ratios, all Offshore Windfarms MAE and RMSE error is less than 0.20. The lower error percentile ratios show higher confidence level in the evaluation and outcomes of the proposed DAE-CNN-LSTM deep learning model. The low MAE and RMSE error also indicate that the proposed model is quite effective in forecasting offshore wind power regardless of offshore regions. The proposed model is also fine-tuned to attain the appropriate parameters for individual mean absolute error (MAE), along with root mean square error (RMSE). For fine-tuning the deep learning model, optimization function, the training error rate, activation, and loss specifications were used for training and testing. The (MAE, RMSE) values of Offshore Windfarm 1 are (0.02, 0.0747) and (0.0103, 0.0786) for Offshore Windfarm 2 and (0.0324, 0.1485) for Offshore Windfarm 3 respectively. Numerically, Eqs and were used to measure MAE and RMSE for three Windfarms. <img src="info:doi/10.1371/journal.pone.0256381.e024" id="pone.0256381.e024g" /> M A E = ∑ i = 1 n \\| y i − x i \\| n Where the variable y is predicted, x is the actual value, n is the number of observed data points. Aggregate variations among different components and samples are rooted in terms of RMSE. The predicted value is the first, while the actual value is the second value. <img src="info:doi/10.1371/journal.pone.0256381.e025" id="pone.0256381.e025g" /> R M S E = ∑ t = 1 T ( x 1 , t − x 2 , t ) 2 T Where expected values are x<sub>1,t</sub>, x<sub>2,t</sub> in observation and T is the cumulative number of measurements or observations. The predicted wind power error values further compared with the common state-of- the-art algorithms, i.e., REP Tree, SVM, Random Forest, J48, Back and Forward Procedure Ensemble Selection, and BPNN respectively. demonstrates the MAE and RMSE errors for all three Offshore Windfarm datasets for different state-of-the- art wind power forecasting models. In offshore windfarm 1, the MAE and RMSE are improved relative to existing state-of-the-art methods. Comparably, the MAE and RMSE for offshore windfarm 2 achieved better results compared to other windfarms. Additionally, for the offshore windfarm 3, the MAE and RMSE gave better results for the proposed model in comparison with other classification models. The Rep tree and BPNN have improved MAE and RMSE scores (0.0229, 0.1204) for the offshore windfarm 1 dataset, concerning our proposed model. Compared to other approaches, the random forest performs the best MAE and RMSE prediction scores (0.0154, 0.1022) for the offshore windfarm 2 dataset. Other than the proposed DAE-CNN-LSTM model, the classification models such as J48 and ensemble selection showed a strong MAE and RMSE scores (0.0446, 0.1653) for the offshore windfarm 3 dataset. Fortunately, the outcomes of showed that our proposed model is more flexible and reliable compared to other state-of-the-art wind power forecasting approaches. Furthermore, the state-of-the-art approaches such as Random Forest, J48, REP Tree, SVM, Back and Forward Method Ensemble Selection, BPNN are also compared to proposed DAE-CNN-LSTM strategy in terms of Normalized Mean Absolute Error (NMAE) and the Normalized Root Mean Square Error (NRMSE) respectively. NMAE and NRMSE are also used to evaluate the effectiveness of times-series or non-label data under normalized scales. Thus, lowering the NRMSE value maximizes the reliability of wind power forecasting models. shows the comparison of the proposed model to other state-of-the-art forecasting approaches on-premise of NMAE and NRMSE. It is certain that across all three Offshore Windfarms, our proposed DAE-CNN-LSTM model outperforms existing state-of-the-art methods in terms of NMAE and NRMSE. Offshore Windfarms 1–3 showed (0.0020, 0.0116), (0.0011,0.0102) and (0.0032, 0.0105) normalized errors respectively. The NMAE and NRMSE analyses further demonstrate that the feature engineering and forecasting strategies are precise and quite useful in accurate forecasting regardless of diverse behavior of offshore winds. The dynamic forecasting plot of all three Offshore Windfarms (1, 2 and 3) in terms of loss configuration is given in Figs. The epoch values are provided on the x-axis and also the accuracy loss amongst all three Offshore Windfarm datasets is specified on the y-axis. In offshore windfarm 1, the loss curve starts at a high range on the y-axis, and loss becomes lowest and stable approximately after the 25 epochs. Similarly, the loss curve for offshore windfarm 2 starts at a high range on the y-axis, and approximately after the 50 epochs, it becomes stable and behaves in the same direction. Lastly, for the offshore windfarm 3, the loss curve starts from a high range on the y-axis and after 50 epochs, the curve is in a linear direction with the lowest and stable loss. The accuracy loss figures are generated by fitting the trained data against test data. The linear curve with fewer error losses indicates the effectiveness of the proposed DAE-CNN-LSTM model on the data used for both training and testing sets. In general, the validation loss shows how much a model is affected by error generated by forecasting models. Low validation error indicates the higher stability of forecasting model for different epoch intervals. Lastly, the R-Squared correlation is used to illustrate the association of wind data characteristics and actual wind power generated by offshore regions. High association between predictor (wind power) and response variable (wind speed) indicates the probability of accurate forecasting on Offshore Windfarms by forecasting models. R-squared describes the measure by which the uncertainty with one component influences the fluctuation for the other component. It is a statistical measure of how the regression line is correlated with wind power forecasting results. The R-Squared curves are estimated for all three Offshore Windfarms, as shown in Figs, respectively. The vertical line displays the predicted values in the model within each curve, whereas the horizontal line indicates the attributes of the wind data as observed values. The green dots comprise the specification and the dynamics of the wind data as well as the linear regression curve depicts a model strength dependent on R-Squared. We also derived R-Squared curves to determine the performance of the proposed DAE-CNN- LSTM model. The closer wind points along linear regression axis shows the better variance and accuracy of the model. All three Offshore Windfarm datasets have obtained an R-Squared correlation of 91.77%, 85.42%, 90.77% respectively. Figs – illustrate that the data points in Offshore Windfarm 1 are more proportionate to the linear axis compared to Offshore Windfarm 2 and Offshore Windfarm 3. As a result, the R-squared correlation of Offshore Windfarm 1 is higher than other Offshore Windfarm datasets. In general, Figs – demonstrates the applicability of forecasting measures on offshore wind turbines data collected by NREL. # Conclusion A new hybrid approach based on deep auto-encoding and CNN-LSTM neural network is designed to estimate wind power for offshore wind turbine datasets. The DAE model is used as the forecasting engine for the initial outcomes. First, the collected data was screened and those variables were selected that primarily lead to precise predictions. The pre-trained model is then used by CNN-LSTM neural network to forecast actual wind power. Finally, the MAE and RMSE were estimated and their corresponding error ratios were evaluated by comparing them with the widely studied state-of-the-art wind power forecasting models. Experiment findings also proved that the proposed model outperforms alternative approaches in terms effectiveness and sustainability of offshore wind power forecasting. To enhance forecasting performance in time series datasets, the optimizer function and capabilities of the LSTM neural network can be further improved. For instance, the Gated Neural Network (GRU) is a new and enhanced version of LSTM-based neural networks that can perform faster and train better with minimal data to train forecasting models. In future research, we intend to improve our proposed approaches for short-term wind power forecasting by integrating GRU and other optimization algorithms where limited data is accessible particularly for short-term power forecasting. [^1]: The authors have declared that no competing interests exists.
# Background 12 million new cases of syphilis occur each year, the majority in developing countries. Probable active syphilis occurs in 1.5 million pregnancies and contributes to 305,000 neonatal deaths and stillbirths each year. Half of pregnant women with untreated syphilis will experience adverse pregnancy outcomes, such as miscarriage, stillbirth, premature delivery, low birth weight and neonatal infection. Congenital syphilis has been a neglected public health problem in Sub-Saharan Africa (SSA), where syphilis prevalence among pregnant women ranges from 1.4 to 17%. Universal syphilis screening of pregnant women is recommended as part of the basic antenatal care (ANC) package promoted by the World Health Organization (WHO), since the symptoms of early syphilis too often go unnoticed and late stages of the disease may be completely asymptomatic. Despite the availability of screening tests and ample evidence that antenatal testing and treatment with single-dose Benzathine Penicillin improves pregnancy outcomes and is highly cost-effective, screening rates range from 1.7 to 79.9% of women attending antenatal care in SSA. A recent analysis of antenatal surveillance data estimated 66% of syphilis-associated adverse pregnancy outcomes occurred in ANC attendees who were either not appropriately tested or treated for syphilis. A number of key barriers to implementation of universal antenatal syphilis screening have been identified: inconsistent supply chain, patient cost and need to return for results, health worker absence or insufficient training, low prioritisation by health policy implementers and use of Rapid Plasma Reagin (RPR) as the standard diagnostic tool, which requires laboratory capacity, cold storage and electricity. To address these barriers, new health technologies: rapid, specific and validated point-of-care (POC) syphilis tests, have been developed and successfully implemented in a variety of clinical settings. However, experience gained in human immunodeficiency virus (HIV) and malaria treatment programmes illustrates that ensuring POC test use and reliability of results, particularly on scale-up, presents its own set of challenges. Reliability is aided by adopting high quality test kits that are easy-to-use; providing adequate training to all healthcare workers (HCW); and integrating Quality Assurance/Quality Control (QA/QC) and supervision systems into programmes from the outset. POC test implementation involves shifting testing to non-laboratory settings and non-laboratory HCWs, often unaccustomed to performing tests or routine QA/QC. Such task-shifting has important planning implications for workload of HCWs in already burdened health systems, and for testing strategy, diagnostic algorithms, QA/QC continuity and supply chain management. Furthermore, HCW acceptance of POC tests and trust in their accuracy are key to ensuring that tests are performed and results are acted upon. In this paper, we explore the end-user experience of new health technology introduction. We compare Zambian HCWs’ experience of RST usability, training and quality systems during a highly-supported pilot project versus a pared down national implementation programme. Through HCW and key informant interviews, we examine how health system planning and infrastructure influenced implementation of both pilot and first phase of Zambia’s national RST programme. # Methodology ## Study design ### Study context Zambia is a lower-middle income country with a population of 14.1 million, spread over a large geographical area. 94% of pregnant women attend at least one antenatal visit and ANC is free-of-charge. The health care system is arranged in six tiers: outreach services, health posts, urban and rural health centres, district hospitals, secondary referral and tertiary referral hospitals. Rural facilities are remote, located several hours’ drive from district hubs, lacking or with limited electricity and limited transport to convey patient samples, results or supplies back and forth to district laboratories. In 2010–11, a six-country pilot study, including Zambia, evaluated the feasibility of introducing RSTs into existing maternal and child health (MCH) programmes. The study was the first of its kind to incorporate QA systems to monitor training and accuracy of POC test results. The non-governmental organisation (NGO)-led Zambian study arm introduced RST into existing prevention of mother-to-child transmission (PMTCT) of HIV services in 15 sites in two districts: Mongu, a rural district, with a syphilis prevalence of 7%, and Lusaka, an urban district, with a prevalence of 2.5%. It showed that integrating RST into PMTCT programmes increased testing and treatment for syphilis in HIV positive pregnant women without compromising HIV service delivery. Following the pilot’s success, the Zambian Ministry of Health (MoH) rapidly adopted RST into national policy and led the rollout of a national RST programme, incorporating QA/QC into programme design. The rollout, supported by the Elizabeth Glaser Pediatric AIDS Foundation (EGPAF), commenced in 2012, initially in four underserved districts with high rates of maternal mortality: Kalomo, Lundazi, Mansa and Nyimba. Key implementation changes were made between pilot and rollout: supervision was reduced from monthly to quarterly and QA/QC activities were devolved from central to district laboratory level. Prior to the pilot, antenatal syphilis screening was carried out using the laboratory-based RPR test when available. MCH HCWs sent blood samples to a laboratory for batched testing, with patients usually returning for results at a later date. In some rural facilities, non-laboratory workers performed RPR tests without a systematic QA approach. During the pilot, RST was performed by MCH HCWs who started same-day treatment based on a reactive RST test. National rollout guidelines described several diagnostic algorithms, depending on RST, RPR and Treponema pallidum haemoagglutination (TPHA) test availability at each site. None of the sites included in this evaluation performed TPHA. ### Data collection This qualitative evaluation used a HCW questionnaire and key informant interviews to determine the feasibility and acceptability of introducing RST into PMTCT services, comparing pilot and national rollout implementation phases. The concept of “feasibility”, drawn from the health technology literature, was defined as the process of RST programme implementation leading to end-user acceptance and utilisation, discussed further below. Acceptability was defined as health workers’ positive satisfaction levels and their correct and consistent use of RST. A health economic analysis of pilot and rollout phases is presented in our accompanying paper, Shelley et al.. The questionnaire utilised both closed questions (including five-point Likert scales) and open-ended questions eliciting qualitative responses. The following domains were included: RST advantages and disadvantages, patient experience, organisational environment and workflow, training, skill and confidence in test performance and RST acceptability. For the rollout evaluation, the pilot questionnaire was adapted using current literature on technology acceptance; domains on QA/QC and supervision were added; and a topic guide was designed, covering the same domains as the questionnaire. Further adjustment, following a pilot interview and feedback from completed MoH/EGPAF supervisory visits, strengthened the internal validity of the study. During the pilot assessment, SAS software (version 9.1) was used to randomly select four sites, two in each study district. They included a district hospital and both low and high volume health centres. Data on service and patient volumes, staffing numbers and location, time/motion studies and cost data were collected to document changes in syphilis and HIV testing and treatment and are reported elsewhere. For the rollout evaluation health facilities were selected by convenience sampling (based on MoH guidance and practical limitations such as distance, staffing and vehicle availability) to reflect a range of urban and rural clinical settings with varying laboratory capacity. Review of records and key informant interviews took place at District Health Offices (DHO). Face-to-face interviews were conducted in English with consenting health workers, with the exception of one interview, which used a translator. In November and December 2010, two EGPAF study staff administered the questionnaire to pilot HCWs, including seven midwives, two nurses and seven lay counsellors. Twenty four rollout HCWs (including four midwives, four nurses, five lay counsellors, one clinical officer, six laboratory technicians, three environmental health technologists and one psychosocial counsellor) were interviewed in August 2012 by one of the authors (É.A.), an independent researcher who accompanied MoH/EGPAF on supervisory visits. Responses were hand- written and rollout interviews were also audio-recorded and subsequently transcribed by the interviewer. Interviewees were selected by convenience sampling, aiming to interview four HCWs at each pilot site and any HCW who had ever performed RST at each rollout site. Using the MoH guideline, any errors identified in test performance were immediately remediated post-interview and discussed with the implementing partner. Informal interviews were held with key informants during MoH supervisory visits to gain understanding of the planning, implementation and costs of the national RST programme. Key informants included EGPAF senior and pilot study staff, MoH HIV/AIDS STI Programme staff, and DHO staff from Mansa and Kalomo districts. Data were collected in the form of field notes recorded after each informal interview. ### Conceptual framework and analysis A conceptual framework was created to guide the analysis, based on a model used by Asiimwe et al. which explored the feasibility of new health technology introduction. The framework divides the concept of feasibility into two inter- related domains, acceptance and usability. Technology acceptance and usability originate in the study of human-computer interaction. Usability has been further broken down into various attributes. Here, acceptance and usability were divided into six sub-domains: learnability, willingness, suitability, satisfaction, efficacy and effectiveness, attributes which have been described in other settings. The framework recognises that acceptance and usability may be influenced by factors related to the end-user (both HCW and client), the diagnostic tool and the health system. Health system factors include guidelines and training, quality monitoring and evaluation, supply chain, and policy, budget and planning. Data from the two evaluation phases were collected, entered and analysed by separate pilot and rollout teams. Quantitative data were double-entered, verified and cleaned using MS Access (2007–10) and Microsoft Excel 2003 respectively. Qualitative data were entered into MAXqda (V10) and NVivo 10, respectively, which are computer-assisted qualitative data analysis software packages. Two authors (M.G. and É.A.) used template analysis to independently analyse qualitative data from each phase. The initial coding template was based on the questionnaire and included codes for: HCW acceptance and satisfaction, patient and partner experience of RST, training, workflow and integration with existing services and quality assurance activities and supervision. For this paper, the authors compared data from each phase under these theme headings and explored and refined emerging sub-themes through iterative discussion. ## Ethical considerations The pilot study protocol and consent procedures were approved by the University of Zambia Biomedical Research Ethics Committee (UNZAREC), the University of Alabama at Birmingham’s Institutional Review Board, the World Health Organization Research Ethics Review Committee and the Zambian MoH. The rollout study protocol and consent procedures were approved by UNZAREC, the London School of Hygiene and Tropical Medicine Research Ethics Committee and the Zambian MoH. To minimise use of personal identifiers, verbal informed consent was obtained from pilot HCWs and recorded on an approved consent form by interviewers. During rollout evaluation, written informed consent was obtained from each relevant District Health Officer and from each interviewed HCW using an approved consent form; verbal informed consent to participate was obtained from key stakeholders and recorded by the interviewer. # Results In evaluating the pilot, 16 HCW interviews were conducted at four sites, located in two study districts (Mongu and Lusaka). In the rollout evaluation, 24 HCW interviews were conducted in two facilities in Kalomo and five facilities in Mansa. Informal interviews were conducted with four senior administrative and management staff in each DHO. Using the conceptual framework, we compare and contrast results from pilot study and rollout evaluation according to the following themes: RST acceptance and satisfaction; RST usability; training and guidelines; quality assurance activities and supervision; health system and supply chain. ## RST acceptance and satisfaction Overall, both pilot and rollout HCWs accepted RST as a suitable addition to existing services. One rollout clinical officer stated: “ “I would encourage the MoH to embrace it and to roll it out in the whole country.” ” Most HCWs were strongly motivated by the opportunity to deliver a better service and offer effective, same-day testing and treatment (STAT) to patients to ensure safer pregnancies. They reported perceived increases in testing rates and treatment access for pregnant women and their partners as well as perceived reductions in loss to follow-up and adverse pregnancy outcomes. They were satisfied with RST reliability and accuracy. HCWs in both phases reported patient satisfaction with the RST but this was notably less in the rollout: 93.8% of pilot HCWs (15/16) and 62.5% of rollout HCWs (15/24) thought patients were somewhat or very accepting of the RST. A rollout lay counsellor commented that patients: “received (it) with both their hands.” Patient benefits reported by HCWs in both phases included: STAT, reduced clinic waiting time, reduced travel time and increased case detection and treatment. Reduced patient waiting time was more evident in the pilot; 68.8% (11/16) of pilot HCWs reported a reduction compared to 39.1% (9/23) of rollout HCWs. By contrast, several (26%, 6/23) rollout HCWs reported an increase in patient waiting times following RST introduction. POC technology allows patients to directly observe their test result. Rollout HCWs reported that this, along with patient counselling, assisted in alleviating patients’ initial distrust of RST: “ “Every change is quite hard to comprehend…We have to try hard so…people really adapt to new situations,” rollout lay counsellor. ” HCWs encountered some difficulties communicating RST test results to patients, in cases where tests were positive due to previously treated infection (and, therefore, requiring only a single dose of BP) or in cases of partner sero- discordance. Counselling was referred to repeatedly by all rollout interviewees as key to client and partner acceptance of testing and treatment. Many drew on their HIV counselling training. ## RST usability ### Health technology usability Both pilot and rollout HCWs found the RST highly learnable; prior experience with POC tests added to their confidence. They rated the individual steps in test performance as easy or very easy. > “The \[RST\] procedure is simple, good and very short,” > > pilot lay counsellor. Compared to RPR testing, it was less “cumbersome” and time-consuming to perform, and easier to interpret. The mobility and convenience of RST allowed each stage of syphilis testing and treatment to be performed in one facility department, or out in the community: “ “It’s like a supermarket, we do everything here…HIV test, results…Hb, the syphilis… the same day,” rollout midwife. ” However, some aspects of RST were less user-friendly. HCWs in both phases described difficulties reporting results within the appropriate time-frame and with interpreting results. These difficulties were more pronounced during rollout: only 30% (7/23) of rollout HCWs rated reporting results within the correct time-frame as “very easy” compared to 87.5% (14/16) of pilot HCWs; half of rollout HCWs were unsure how to correctly interpret weak positive results, which should be reported as positive. Several described repeating the RST or referring to the laboratory for confirmation of weak positive results; neither practice is in accordance with guidelines. ### Usability—integration within existing services Before RSTs were introduced at study sites, HCWs reported that, despite inclusion in national ANC guidelines, antenatal syphilis testing using RPR was inconsistently performed. RPR was performed at only 3 of 4 pilot sites and 6 of 9 rollout sites during the interviews. Most HCWs agreed RSTs were successfully integrated into facility PMTCT services, (16/16 pilot and 23/24 rollout HCWs) but there was high inter-site variability in algorithm followed and effect on workflow. The pilot integration pattern was standard: a MCH HCW (midwife, nurse or lay counsellor) ran both RST and HIV tests and returned results; a midwife or nurse then initiated treatment thus efficiently and effectively providing STAT. RPR confirmation was not included in the pilot protocol. By contrast, integration at rollout sites depended on the facility’s health system level, available HCW cadres and laboratory capacity. Fewer rollout HCWs reported consistent delivery of STAT (83%, 20/23) compared with pilot HCWs (100%, 16/16). This may in part be due to variable patterns of task shifting. At larger facilities HCWs reported splitting testing steps amongst different HCWs and different departments, which led to delays in returning results and initiating treatment. In some cases confirmatory RPR was performed by the same HCW; in others, clients were required to return for RPR testing or results. Notably, midwives in one rollout facility reported a high reactivity rate of 56.3% on initial RST introduction. To verify results, they repeated RST or referred to the laboratory, where the lab technician reported repeating RST followed by RPR on all positive ANC samples, out of keeping with guideline algorithms. It also proved initially confusing to some rollout HCWs that the treatment algorithm differed in the case of negative RPR. Pilot HCWs regarded RST as a user-friendly addition to PMTCT services which resulted in overall time-savings, despite the additional time required to counsel patients on both syphilis and HIV: “ “…results are out in a short time…thus, cuts on time we spend on syphilis testing giving us more time to do other routine activities in MCH,” pilot lay counsellor. ” In the rollout, the effect on workload depended on both the role the HCW had in syphilis testing prior to RST introduction and the algorithm employed. Those previously involved in RPR testing reported overall time-savings irrespective of whether they then went on to perform RPR for some clients; those not previously involved felt RST introduction increased their workload and staffing levels were insufficient to manage: “ “I have a lot of clients …I have to see them… counsel them and do the RST to the woman and the man. So, it becomes difficult at times when I’m alone. That is in terms of staffing,” rollout midwife. ” As a result, they deferred syphilis testing to a later ANC visit or delayed reading test results. ## Training and guidelines Training experiences varied between pilot and rollout, with pilot HCWs more likely to have received training from EGPAF staff. 25% (4/16) of pilot interviewees reported attending an EGPAF training workshop; the remainder received on-the-job training from colleagues or EGPAF study staff; half received remedial booster training from EGPAF during monthly supervisory visits. 29% (7/24) of rollout interviewees attended the MoH/EGPAF-led training workshop. The remainder (17/24) received on-the-job training from colleagues. According to MoH staff, the rollout training workshop emphasised on-the-job training but did not include specific “train-the-trainer” techniques. A training manual was provided to workshop attendees but few reported using it to train colleagues. Following training, only three (3/23) rollout HCWs used SOPs to aid accurate testing; all (16/16) pilot HCWs used them. Most rollout HCWs found both workshop and on-the-job training acceptable. Several HCWs believed on-the-job training was too short and contained inaccurate or incomplete “second-hand information”. Some HCWs described colleagues’ distrust of them as trainers, their dissatisfaction at missing “official” training and their resultant unwillingness to perform the test: “ “...they think, since you’re the one who went to the workshop you will be the one to do (the RST),” rollout counsellor. ” ## Quality assurance activities and supervision Acceptance and usability of QA/QC activities were investigated as part of the rollout evaluation only. 100% of rollout respondents accepted QA/QC activities and recognised the potential benefits: “ “I think it will help us know if the test kits are working properly and you would be confident that the result you would give would be the correct result,” rollout midwife. ” However, few rollout HCWs accurately recalled the QA/QC programme components described in RST training manuals. A small number reported performing internal QC to check test kit quality (e.g. ensuring the test control window was positive before reporting results). With the exception of one PMTCT counsellor who displayed self-efficacy in using RPR control samples to check kits, no HCW reported performing systematic external QC or completing proficiency panels. Guidelines stated that district laboratories should prepare and provide positive and negative control samples for external QC and dried-tube specimens for proficiency panel testing. Laboratory facilities reported performing their own weekly QC; one district laboratory sent proficiency panels to local facilities but no results were returned. Reported barriers to the provision of district-led QA/QC were lack of trained staff, dedicated time, transport and reporting infrastructure. Supervision differed from pilot to rollout. EGPAF made monthly support visits during the pilot and supported the MoH on quarterly supervisory visits during rollout. Rollout HCWs expressed willingness to undergo supervisory visits, describing them as an opportunity to learn and to correct mistakes. Most believed supervision was a suitable and effective tool, which boosted morale and encouraged HCWs to work more diligently and more accurately. Some dissatisfaction with supervision was expressed; a minority felt demoralized, intimidated or uncomfortable when observed performing tests. ## Health system and supply chain Supply of penicillin was consistent during the pilot but was less reliable during rollout: 12.5% (3/24) of rollout interviewees said stock-outs occurred once a month. RST kit supply was also less reliable during rollout; half of pilot sites (2/4) reported a stock-out on one occasion whereas almost a third of rollout sites (7/24) reported a complete RST stock-out during the month preceding interviews. EGPAF key informants reported that RST kits were not supplied through the usual national supply chain but rather through the DHO Pharmacy. This was to avoid incorrectly delivering to districts not yet involved in rollout but it led to confusion and ruptures in RST supply. # Discussion This two-phased evaluation provides a detailed account of new POC test implementation, examining the transition from an NGO-led pilot project to the first phase of a national RST programme in Zambia. Novel, inexpensive and easy- to-use POC tests can address inequality in access to diagnostics in lower and middle income countries such as Zambia. However, implementation can be complex and requires a usable test, acceptance by HCWs and patients, good quality training, ongoing HCW support, quality monitoring systems and robust supply chains, which are more difficult to ensure for large-scale rollout with budget constraints. Jafari et al described a lack of quality reporting of implementation research outcomes, other than accuracy, for syphilis POC tests, which our study helps to address. This early programme evaluation takes place at multiple health system levels in four Zambian districts and may inform the Zambian MoH as they plan further phases of RST programme rollout. Using HCW surveys and key stakeholder interviews, we examined the success of RST implementation from the perspective of “feasibility”, a concept borrowed from the health technology literature, which was further broken down into two components: “acceptance” and “usability”. This study was not designed to report impact of RST introduction into PMTCT services in Zambia, which is reported elsewhere. Overall, the study findings were positive with both pilot and rollout HCWs reporting that RST largely met our criteria for acceptability and usability. However, feasibility, as defined here, is contingent not only on end-user and technology attributes but also on health system and programmatic factors. We identified that several aspects of the rollout programme differed from the pilot, particularly training, testing algorithm, QA/QC, supervision, and supply chain; these may have negatively influenced the end-user experience and therefore the feasibility of the RST rollout programme. Both groups found some aspects of RST non-user-friendly but rollout HCWs demonstrated significantly less correct and consistent use of RST, likely influenced by less exposure to MoH/NGO-provided training and regular supervision. Rollout on-the-job training quality appeared highly variable; the effectiveness of the cascaded training model utilised in the rollout was clearly limited. The inconsistent implementation of rollout testing and treatment guidelines may reflect the nature of the training and scaled-down supervision received by rollout HCWs compared to pilot HCWs and may have influenced the quality of care delivered. The POC literature emphasises the importance of high quality training to ensure test accuracy and reliability. This is increasingly critical as HCWs are now expected to perform multiple rapid tests in parallel, each with varying specifications and manufacturers’ instructions. The WHO and global health workforce alliance advocate the cascaded model of training, or “train-the-trainer” (TTT) approach, as one way to meet the global health workforce crisis. The effectiveness of TTT is supported by a recent systematic review. However, there is dearth of knowledge on TTT programme content, the training techniques that should be incorporated, or on programme sustainability. Rollout HCWs were willing to take part in cascaded training, although some expressed resentment at missing the opportunity to attend an off-site training workshop. This may be linked to the widespread use of per diem payments for health workers in some low and middle-income countries, such that they have become an expected supplement, sometimes even exceeding health sector salaries. While these payments may encourage productivity and motivation, they have been shown to create conflict between HCWs and even to demotivate those who don’t receive them. Supervision and monitoring of HCWs’ performance has been shown to increase their confidence in test performance and accuracy of test results, a finding echoed in this study. During the pilot, rates of both HIV and syphilis testing and treatment improved, which was attributed in part to regular supervision. Overall, rollout HCWs were less satisfied with the increased workload resulting from the addition of another POC test to MCH services, but this was dependent on their previous role in syphilis testing and the algorithm employed. Our findings support a New Zealand study which showed that POC test introduction can have important, unforeseen implications for workforce planning. Most rollout HCWs inaccurately described the steps involved in RST performance, confusing them with other POC tests. The assumption that integrating multiple testing programmes serves to strengthen health systems may not always hold true. The new dual RST and HIV POC tests, which are currently under evaluation by the WHO and partners, may eliminate some of the problems engendered by running multiple POC tests in parallel. Less frequent supervision may be one reason rollout HCWs reported greater difficulty with test performance and employed a variety of integration patterns, resulting in lower rates of STAT. Rollout HCWs perceived lower levels of patient satisfaction than their pilot colleagues, which may result from the longer patient waiting times reported by some. HCWs also described difficulties around negotiating patient expectations and understanding of test results. Similar issues have been described in the literature on malaria rapid diagnostic test (RDT) introduction, particularly challenging patient expectations of malaria treatment in the face of a negative malaria test. Specific training in counselling patients and partners around RST results may prove useful to include in further editions of RST training manuals. The importance of QA/QC systems has been repeatedly emphasised in recent literature on HIV and malaria POC tests. Despite this, most HIV testing programmes in developing countries operate without them. Ensuring reliable POC test results is of paramount importance as inaccuracies lead to incorrect diagnosis and treatment and negatively impact patient care and cost- effectiveness. Assuming the same issues regarding POC accuracy might beset RST implementation, QA/QC were incorporated into both pilot study and national rollout. Anticipating that QA/QC could be an area of weakness during rollout, several modules of the MoH training guide were devoted to QA, QC and monitoring procedures. All rollout HCWs interviewed understood the rationale for QA and demonstrated willingness to take part. However, the first round of rollout supervisory visits revealed limited implementation of these procedures. Several factors may have contributed to this: ineffective planning and communication (in that no specific person at facility or district level was nominated to coordinate QA/QC), lack of dedicated budget and logistics, lack of local expertise in this type of activity which is often undertaken by NGOs, or lack of efficacy on the part of local laboratory staff. This suggests that HCW training on QA/QC was not learnable or effective in its current form. Further research is needed to establish the ideal mix of RST QA/QC activities which is usable, acceptable, effective and cost-effective. Supply chain issues may have impacted rollout success. For POC tests to maximise their potential, supplies of test kits, treatment and other consumables must be consistently available. Weaknesses in both MoH district pharmacy and central medical supply chains were identified, which should be addressed before commencing further phases of the rollout. Limitations of this study include the fact that it was a single-country case study whose findings may not be generalizable to other countries. Interviews and data analysis were performed by two separate research teams using slightly different collection tools which may reduce comparability of data from pilot and rollout phases. The pilot surveys were conducted by the implementing organisation and surveys in both phases were carried out during supervisory visits, increasing the risk of social desirability bias. The majority of study findings are based on a HCW survey; data are therefore response-based and susceptible to social desirability bias and measurement error. Convenience sampling of interviewees may have introduced selection bias. Quantitative data are also subject to measurement error. In addition this study design was not able to ascertain programme impact. The Zambian RST scale-up model incorporated many success factors defined by Yamey in a recently proposed framework for successful scale-up of global health interventions. However, we suggest that there are a number of aspects which ought to be refined before further rollout of the Zambian RST programme; our findings have been shared with the MoH and their implementing partner. The ideal training model, supervision type and frequency and mix of QA/QC activities for RST programmes has not been established and further implementation studies, which also examine impact, are required. To strengthen the cascade model of training, the MoH could adopt a comprehensive “training of trainers” model incorporating adult education techniques and certifying workshop attendees as “local trainers”. Integrating training with HIV, malaria and Haemoglobin POC test training would facilitate emphasis on differing test characteristics and reduce HCW confusion. QA programme aspects would be strengthened by a distinct laboratory QA rollout strategy identifying specific responsible local personnel, and facilitating them with dedicated financing and logistics. Collaboration with those organisations already performing monitoring and evaluation of HIV testing may prove effective. Advocacy with the RST kit manufacturer to include positive and negative control samples could facilitate QC testing and address some of the difficulties experienced in communication and transport between laboratories and the surrounding health facilities. In addition, the supply chain must be strengthened. # Conclusion This study provides insight into the challenges of integrating new health technology into existing health systems and explores the transition from pilot study to national rollout. It illustrates programmatic experience of implementing cascaded training and QA, QC and supervision activities. Use of a conceptual framework adopted from a paper on health technology “usability” facilitated the exploration of programme implementation and may prove useful in other settings. We demonstrate that it is feasible to implement a MoH-led rapid syphilis testing programme in Zambia, following a successful pilot, but that issues around training, QA/QC and supply must be addressed in planning for future phases of this and other programmes incorporating POC tests. The Zambian MoH has already demonstrated commitment to monitoring and evaluation practices and addressing these programme aspects should serve to strengthen the health system as a whole. # Supporting Information The authors would like to acknowledge the support of the Zambian Ministry of Health in carrying out this research. We would like to thank the health workers who participated in the study and the staff of the District Health Offices and health facilities who participated in Mongu, Lusaka, Kalomo and Mansa Districts. The pilot study was made possible by the support of the Sexually Transmitted Diseases Diagnostics Initiative of the UNICEF/UNDP/World Bank/WHO Special Programme for Research and Training in Tropical Diseases (TDR) and the support and coordination of The London School of Hygiene and Tropical Medicine (LSHTM). During pilot study conception and implementation, technical support was provided by Catherine Wilfert; Helen Kelly (LSHTM); Laura Guay (EGPAF, Washington, DC); the Centre for Infectious Disease Research (CIDRZ); and TDR. LSHTM supported the rollout research and Colette Fleischer contributed to rollout data collection and analysis. The staff of the EGPAF Lusaka office provided invaluable assistance and support. The views expressed in this article are the opinions of the authors and do not necessarily reflect the official policies of USAID or the US government. The authors also thank the journal reviewers for helpful comments and suggestions for revising the original version of the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: EMA MMG JR SS KDS TS FTP RWP DM. Performed the experiments: EMA ATN GTM. Analyzed the data: EMA MMG JR. Contributed reagents/materials/analysis tools: EMA MMG KDS. Wrote the paper: EMA MMG KDS FTP JR. [^3]: Current address: Department of International Health, Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland, United States of America [^4]: Current Address: ICAP program, Mailman School of Public Health, Columbia University, New York, United States of America
# Introduction Coastal and ocean systems have played a crucial role in human evolution, providing food, health, well-being and livelihoods since prehistoric times. Marine biodiversity, in particular, provides enhanced nutritional benefits to human societies, while ensuring fundamental ecosystem functions, structures and services. Since the second half of the 18th century (Industrial Era), and especially in recent decades, pervasive and compounded anthropogenic impacts on ocean and coastal systems have accelerated marine biodiversity loss and population decline, and increased the vulnerability of marine ecosystem services. Understanding the scale of anthropogenic stressors, however, is extremely complex and exacerbated by the paucity of long-term records on species composition, abundance and distribution pre-dating the most recent decades of anthropogenic impacts. This form of historical amnesia creates misconceptions that have far-reaching consequences for sustainability and conservation actions. For example, it may conceivably increase tolerance for progressive environmental degradation and contribute to setting inappropriate sustainability targets and responses by ocean stakeholders; and hence underpin their expectations as to what is a desirable and achievable state of social-ecological systems. These are typical symptoms of the “shifting baseline syndrome” that require locally- impactful and globally-relevant new knowledge to engage and support stakeholders in designing conservation pathways to ocean sustainability. Moreover, the Intergovernmental Oceanographic Commission of the United Nations Educational, Scientific, and Cultural Organisation (IOC-UNESCO) has recently launched important global efforts to educate people on the oceans (Ocean Literacy), including fundamental concepts about how oceans function and human- ocean interdependencies. Understanding humanity’s past, present and future relationships with the oceans are therefore of utmost importance to both the on- going implementation of the UN Decade on Ecosystem Restoration (2021–2030;) and the UN Decade on Ocean Science for Sustainable Development. The southern Atlantic Forest coast of Brazil is a world biodiversity hotspot with a long history of human-ocean interaction, and a priority region for socially equitable sustainable development in cooperation with biological conservation and restoration. It is also, however, a data poor region in relation to fisheries statistics despite being the largest fish producing region (Santa Catarina state) in the country. Stock monitoring data are regionally and chronologically dispersed, and mostly limited to the second half of the 20th and the first decade of the 21st centuries. In Brazil, there has been an increased effort to expand stock reconstructions with data collected from local fisheries and seafood wholesale market data, but these approaches are inherently limited to only the most recent decades affected by large national fisheries subsidies, specifically from the 1960’s to 1970’s. Overall, little is known about the diversity of targeted species and their socio-cultural, economic and market values before more recent political and financial support were introduced to the fishing industry. Historical sources can provide valuable insights into past species composition, distribution and abundance. Among written public media documents, newspapers offer some advantages for qualitative and quantitative historical analyses. Newspapers generally have large distribution ranges and relatively high frequency of issues (released on a daily to semestral basis), while digitization programmes through public funds and advances in optical character recognition (OCR) technology are making historical newspapers increasingly accessible. Moreover, public media offers insights into societal (e.g. cultural, economic and political) drivers of change, including stakeholder perception of environmental conditions through time. Public media also shapes public opinion, influencing peoples’ literacy of ocean, human and ecological affairs, and thus may be a major driver of social-ecological baselines in all societies. There is a rich body of literature conceptualising the use of newspapers for reconstructing historical events that can help marine historical ecologists assess data quality while framing their research in the evolving editorial and political agenda of popular media through time. In Brazil, newspapers have been issued since the beginning of the 19th century, and are one of the few sources of information documenting the development of political and economic institutions, and commercial activities over the last centuries. Gallardo et al. have recently shown that historical newspapers retained insightful information on the drivers of anthropogenic impacts on coastal areas of Brazil over the last 150 years. Here we expand their approach to derive long-term data on marine organisms targeted by Brazil’s fishing sector between 1840 and 2019, and notably during the first half of the 20th century, preceding “blue acceleration” of the last decades, where a major gap in understanding still exists. We investigated historical newspapers published in Santa Catarina state and scrutinised more than twenty thousand digitised editions in the Brazilian Digital Newspapers and Periodicals Library (Hemeroteca Digital Brasileira). Our aim is to uncover the diversity of species commercially exploited, traded and consumed in southern Brazil since the late 19th century, prior to the start of the global-level blue acceleration social-ecological phenomena and the first large fisheries subsidies in the country. Our paper offers new insights into the origin and changing nature of significant human impacts on marine systems in the southwestern Atlantic Ocean; as well as implications for ocean sustainability initiatives aiming to bring society’s social-ecological baselines closer to historical reality. # Material and methods ## Data collection The sampling method follows Gallardo et al. and, in brief, it can be summarised in the following steps: A) Historical newspapers were sourced from the Brazilian Digital Newspapers and Periodicals Library (Hemeroteca Digital Brasileira, Fundação Biblioteca Nacional, BNDigital), made available by the National Library Foundation (Fundação Biblioteca Nacional, <https://bndigital.bn.gov.br/hemeroteca-digital/>). B) News items (e.g. articles, opinion letters, regulations and sanctions) were sourced using the keywords “pesca” (fishing) and "peixe" (fish), selecting the “Locality” as the region of Santa Catarina using the State abbreviation “SC”, opting for single decades as the “Period”, and including all newspapers for the selected decades. Details of the searching facilities including OCR available through the BNDigital can be found at <https://bndigital.bn.gov.br/hemeroteca-digital/>. The keywords “pesca” (fishing) and "peixe" (fish) were chosen for the following reasons: i) The keyword “pesca” captured several types of marine resources, and also captured the word "pescado" which categorically addresses different types of resources exploited by fisheries; ii) within Brazilian folk taxonomy (ethnotaxonomy), various marine animals ("pescado") are classified as “fish”; iii) the popular association of fish such as shrimp, lobster, oysters, octopus and squid with crustaceans and molluscs is something very recent and thus this nomenclature is infrequently used in everyday life; iv) given the large diversity of organisms captured by coastal fisheries in Brazil and the general scope of the paper (not targeting a specific species), it would be unfeasible to extend the keywords to all potential captured organisms. C) Single items were then assessed for duplicate results and non-relevant information (e.g. commercial advertisements) which were then excluded. Once the search was completed, single items were fully read in order to assess the content and quality of information based on two inclusion criteria: i) the items focused on coastal and ocean areas, including estuaries and coastal lagoons, or coastal rivers; and ii) the items reported marine animals (e.g. fish, crustaceans, molluscs, mammals, reptiles), including non-native taxa. D) An analytical framework was created to collect and categorise content and contextual information, which consisted of 1) a catalogue of reported taxa and their overall trophic ecology, and 2) the type of fishing gear, when reported (SI1). Aquatic animals were mostly reported using vernacular names, with conversion to scientific names (using the minimum identified taxonomic level), whenever possible, based on local literature. The catalogued animals and gear were quantified for their absolute and relative abundances (the number of times a species or type of gear occurred in a given number of newspaper items) and richness (the number of different species and gear types in a given number of newspaper items) aggregated per decade. If species X was reported more than once in a single item, its abundance in the given item was counted as one. ## Data analysis For each species, the trophic level (TL) was assigned according to data in FishBase, using the main diet of adults. For taxa where only genus and/or family were identified we used the average values of the main species in the region according to SiBBr ([https://sibbr.gov.br](https://sibbr.gov.br/?lang=pt_BR)). With this information we computed the weighted average trophic level (WATL) for each decade. In addition, we computed fish-based piscivores/planktivores (Pis/Pla) and invertebrates/fish (Inv/Fis) to explore changes in overall catch composition through time. For this we used the absolute abundance of species reported in newspapers per decade. These are simple ecological indicators that are used here to infer changes in fishing pressure, trophic level (fishing down the food web), biological productivity, market preferences and technological changes. Non-native and freshwater organisms (e.g. fish, shrimp), reptiles, birds and sea mammals were excluded from WATL, Pis/Pla, Inv/Fis analyses. The results from Inv/Fis were compared with reconstructed and officially reported industrial and artisanal catch data (volume in tons) for Santa Catarina from 1950 to 2015. Piscivores included fish species consuming finfish and pelagic cephalopods, while for planktivores their diet consisted of zooplankton and early stages of fish and jellyfish. For these two trophic categories we used references from Quimbayo et al., complemented with other literature. When combined, multiple indicators are expected to provide a more comprehensive picture of historical changes in the system. Fishing gear were standardised following the Fishing Gear Type of the ICMBio (Centro Nacional de Pesquisa e Conservação da Biodiversidade Marinha do Sudeste e Sul (<https://www.icmbio.gov.br/cepsul/artes-de-pesca.html>)). The information collected from the newspapers were complemented with additional sources, such as the reports of the Ministry of Agriculture (available at <http://ddsnext.crl.edu/>). Data collection was carried out from March 2021 to March 2022. # Results ## Species composition and fishing technology The keywords “pesca” (fishing) and “peixe” (fish) generated 9,190 and 14,168 matches respectively, spanning a period of 180 years (1840 to 2019). Over 23,400 newspaper reports (items) were individually assessed and read. Of these, 598 items were selected following the inclusion criteria reported above. The selected items were reported in 57 newspapers, representing 4.7% of the total newspapers available in the HBD for Santa Catarina over the studied time period (n = 1206, 18th March 2022). The selected items provided information for 37 municipalities, of which 49% (n = 298) were related to Florianópolis (formerly Desterro), the capital of the state. From 1950 onwards there was an increase in news items from municipalities located by the Itajaí River and the northern coastal regions of the state (e.g. Barra Velha and São Francisco do Sul), where Brazil’s largest commercial fishing industry developed over the last decades. The keywords “pesca” (fishing) and "peixe" (fish) captured animals from different taxonomic groups (fish, crustaceans, molluscs, mammals, reptiles), which is significant since in folk taxonomy several taxa are also considered as “fish”. Native and non-native animals were mostly reported as vernacular (v) and, to a lesser extent, as scientific (s) names, the latter sporadically appearing in more recent decades. We found 1655 mentions of marine animals, totaling 277 and 203 vernacular and scientific taxa respectively (SI1). As reported by Feire and Pauly, we also found evidence that a single folk nomenclature was used for several species (e.g. Hypanus sp., Dasyatis sp., Fontitrygon sp., Bathytoshia sp. = raia prego), and several folk nomenclatures were also used for a single species (e.g. Mugil liza = tainha, tainhota, tanhota, do corso, facão, tainha-assú). As a consequence, attempts to correlate scientific and vernacular names were complex and not always possible with an equal level of confidence. Native marine fish dominated in terms of number of species, both vernacular (v = 192) and scientific names (s = 143), as well as in abundance (n = 1,041). This was followed by crustaceans (v = 17, s = 9, n = 181), freshwater animals (v = 22, s = 15, n = 111), molluscs (v = 11, s = 9, n = 99), sea mammals (v = 9, s = 7, n = 81), echinoderms (v = 2, s = 2, n = 2), reptiles (v = 1, s = 1, n = 7) and birds (v = 1, s = 1, n = 1). Non- native animals (including fish and crustaceans), introduced for commercial aquaculture (e.g. tilapia, carp, channel catfish) or imported as processed food (e.g. cod, conger, hake), also had relatively high species richness (v = 22, s = 16) and abundance (n = 132). The most frequently reported species was the demersal and estuarine-dependent tainha (mullet, Mugil liza, Mugillidae, 11.3%, n = 187). Tainha appeared in the initial period of 1880–1889 and reached maximum relative frequency in 2010–2019. Tainha was followed by camarão (shrimp, Penaeidae, 9.3%, n = 154) and by pelagic and estuarine-demersal taxa including anchova (bluefish- enchova, Pomatomus saltatrix, Pomatomidae, 3.8%, n = 63), baleia (whale, several Mysticeti, 3.8%, n = 63), cação/tubarão (sharks, several Elasmobranchii, 3.2%, n = 53), bagre (catfish, Ariidae, 2,8%, n = 47), corvina (whitemouth croaker, Micropogonias furnieri, Sciaenidae, 2.8%, n = 47), cascudo (armored catfish, Loricariidae, 2.2%, n = 37), tilapia (tilapia, Oreochromis niloticus, 2.1%, n = 35), pescada (drums/croakers, Sciaenidae, 1.9%, n = 31), robalo (fat snook, Centropomus parallelus, Centropomidae, 1.9%, n = 31), ostra (oyster, Crassostrea rhizophorae, Ostreidae, 1.7%, n = 28), sardinha (sardine, Clupeidae, 1.6%, n = 27), garoupa (dusky gruper, Epinephelus marginatus, Serranidae, 1.6%, n = 26), among others. Some generic taxa were reported with distinct vernacular names, such as cação/tubarão (e.g. mangona, anequim, tintureira, baleeiro, cação gralha branca, c. salteador, c. azul) and camarão (e.g. camarão branco/legítimo/do corso, c. perereca, c. sete barbas, c. rosa/pistola, c. cabeça azul). When combined, the relative abundance of sharks and shrimps increased to 4.5% and 9.3%, respectively. The number of items reporting marine animals using the keywords “pesca” (fishing) (n = 269) and "peixe" (fish) (n = 329) were significantly positively correlated with the number of newspapers for Santa Catarina available in the HDB archives (pesca r = 0.61; R<sup>2</sup> = 0.37; p \< 0.005; peixe r = 0.52; R<sup>2</sup> = 0.27; p \< 0.02), a pattern also observed by Gallardo et al.. Similarly, species richness and their abundances were also significantly positively correlated with the number of newspapers (species richness, r = 0.73; R<sup>2</sup> = 0.53; p \< 0.001, frequency, r = 0.68; R<sup>2</sup> = 0.46; p \< 0.001) in Santa Catarina, indicating that, in absolute terms, our data is affected by the amount of digitised newspapers available in the HDB database. When considering the relative proportion of distinct taxonomic categories and species, the results point to variations possibly expressing changes in socio- cultural, economic and market importance of aquatic animals through time. Native marine fish and sea mammals had higher relative abundance and species richness between the decades of 1900–1909 to 1930–1939, followed by a decreasing trend toward the most recent decades (2000–2009 and 2010–2019). By contrast, the abundance and richness of invertebrates (molluscs, crustaceans) increased in the decades between 1900–1909 and 1930–1939, and again in 2000–2009 and 2010–2019. Freshwater fish and non-native species (both freshwater and marine), instead, showed the opposite trend, with higher relative abundances and species richness in the decades of 2000–2009 and 2010–2019 compared to previous decades. The WATL shows the prevalence of high trophic level species over the last 180 years, including medium to large top predators such as cherne (groupers—cherne-pintado (snowy grouper), garoupa (dusky grouper), mero (goliath grouper)) and tubarão/cação (sharks—cação mangona (sand tiger shark), c. baleeiro (copper shark), c. gralha branca (whitetip shark)), among others. However, a significant decline in WATL was observed between the late 19th and early 21st centuries (r = -0.56; R<sup>2</sup> = 0.31; p \< 0.05). The ratio between piscivores and planktivores (Pis/Pla) was quite variable from 1880–1889 to 1960–1969 (ranging from 0.7 to 5), but maintained consistently low values (0.6 to 1) from 1980–1989 onwards. These low trophic level species are mostly represented by carapeba (possibly caratinga (mojarras, Eucinostomus sp.), gordinho (American harvestfish, Prepilus paru), parati (white mullet, Mugil curema), and several Scianidae (cangulo, pescadinhas, cangoá, guete). Fluctuations in Pis/Pla did not correlate with time (r = -0.01; R<sup>2</sup> = 0.0002; p \> 0.05), nevertheless, the results from combining these indicators point to a greater relative importance of high trophic level species (notably piscivores) from the late 19th to the mid 20th century, and a later decrease from the 1970’s and 1980’s. The high diversity and the relative frequency of camarão (shrimp), ostra (oyster), marisco (mussels), siri (crab), and lula (squid) since the late 19th century indicates that historical fisheries targeted organisms from all trophic levels. Interestingly, a significant increase of short-lived invertebrates (r = 0.58; R<sup>2</sup> = 0.35; p \< 0.05) is observed from 1970–1979 onwards, which is partially consistent with their increased captures by both artisanal and industrial fleets in the 1970’s and 1980’s in Santa Catarina. The reported organisms were associated with a range of fishing gear (n = 43), some of which only appeared in the late 19th and early 20th centuries (e.g. chae-chae, caçoal) (SI1). The diversity and abundance of fishing gear did not change as a function of time (n gear, r = 0.157; R<sup>2</sup> = 0.024; p \< 0.532; frequency, r = 0.078; R<sup>2</sup> = 0.006; p \< 0.757), but were significantly correlated with the number of newspapers (n gear, r = 0.641; R<sup>2</sup> = 0.411; p \< 0.005; frequency, r = 0.806; R<sup>2</sup> = 0.649; p \< 0.001). The broad category “rede” (net) dominated both in frequency (n = 132) and diversity of gear (n = 18), including variations such as anchored, surface, bottom, drift, small mesh, large mesh, and trammel. According to ICMBio (Fishing Gear Type), these could be tentatively assigned to drift gillnets (rede de emalhe), which are some of most common fishing gear used to catch demersal species in south Brazil, and with documented use since the 1870’s (A Gazeta de Joinville, 11 february 1879). Other relatively frequently used gear included “aparelhos com anzol” (use of a hook), “tarrafa” (cast nets), and “arrasto” (trawl, beach seine), to name a few. # Discussion ## Species composition and the impact of early fisheries subsidies Our study revealed that historical newspapers retain insightful information on the perceived value, diversity and abundance of marine animals in southern Brazil over the last 180 years. We identified nearly 300 aquatic animals, the majority of which were captured and traded by coastal communities and commercial fisheries in Santa Catarina since the 1840’s, and others that were imported for direct consumption or aquaculture throughout the 20th and the early 21st centuries. The compiled information (e.g. frequency and richness), however, was unevenly distributed across the studied time interval, with fluctuations positively correlated with the number of digitised newspapers in the HDB database. This highlights the challenges of extracting absolute values to reconstruct ecological baselines from historical sources. Therefore, in order to minimise sampling biases, we focus our discussion on the decades with greater amounts of information in the newspapers. Significantly, 51% of our data (n = 307) was compiled from newspapers published between the decades of 1880–1889 and 1930–1939, which was a time interval of unprecedented political efforts to boost the commercial fishing sector in Brazil. For example, in 1856, the Decree 876 of 10 September authorised the incorporation of companies for catching, salting and drying fish along the coast and rivers of the Brazilian Empire. Later in 1912, Decree 9672 of 17 July created the Fisheries Agency (Inspetoria da Pesca), replaced by the Biological Marine Station in 1915 (Estação de Biologia Marinha, Decree 11507 of 4 March), which promoted institutional actions for “studying and disseminating the natural resources of Brazilian waters, to develop them as much as possible and regulate their use". In 1933, Decree 23348 of 14 November established fish landing, processing and commercialization sites (entreposto de pesca), in order to increase efficiency of commercial fishing. Surprisingly, little is known about the targeted species and the ecological implications of these public policies. From the late 19th century, several fishing concessions were released to private investors in Santa Catarina (A Regeneração, 12 August 1884, 12 October 1888), while in the early 20th century, landing and seafood markets (Entrepostos de Pesca) were implemented through public investments in several cities along the Brazilian coast and increased state control of economic activities. Urban population growth increased demand for fish, enhancing market opportunities for private investors that greatly benefited from fiscal incentives in key urban centres such as Rio de Janeiro and Santos. The extent of commercial fisheries development in Santa Catarina since the early 19th century is alluded to by export records from the city of Laguna for the years 1819 and 1820, which documented, among other fish species, the sale of more than 1 million catfish in 2 years to several Brazilian states (Bahia, Pernambuco, Rio de Janeiro) (Revista Catarinense, 1911). However, consolidated markets and long-trade networks are better documented for later periods. Between 1910 and 1939, several locally caught fish (e.g. mullet, bluefish-enchova, shark, catfish, black and other types of drums) and invertebrates (e.g. shrimps, oysters, mussels) were reported in newspapers in relation to markets (22%, n = 133), landing reports (18%, n = 110), industrial fishing (14%, n = 83) and local food security (13%, n = 81). They were captured for direct human consumption (seafood) as well as for other byproducts including fishmeal, fish skin (as "sandpaper"), fish stomach and gut (as fertiliser) and oils, which were regularly traded in local and regional markets (O Despertador, 23 October 1880; Republica, 3 July 1903; O Estado, 22 December 1933). These local catches supplied the growing demands in urban centres (A Gazeta, 2 October 1935, but see also), as well as the nutritional needs of small coastal communities (O Estado, 22 July 1915; O Dia, 3 August 1916; A Noticia, 18 September 1936; Republica, 22 October 1937). We suspect that commercial fisheries equally benefited from the specialised labour force, infrastructure and technology released with the decline of whaling in the late 19th century. In the case of mullet, the most cited species with various distinct vernacular names (“tainha”, “tainhota, “tanhota”, “tainha do corso”, “tainha facão”, “tainha-assu”), intensive exploitations were attested since the late 19th century, when the species was primarily captured during its reproductive migration in winter (A Regeneração, 27 May and 7 June 1885; O Estado, 3 June 1915). Its fishing and processing (escalada, involving salting and desiccating) mobilised entire communities (A Regeneração,18 June 1885; O Despertador, 14 May 1881), and the catches were sold both locally (O Estado, 2 February and 1 March 1917) and in markets located hundreds of kilometres away, such as São Paulo (O Estado, 12 June 1923). Single catch events could consist of 30–60 thousand mullets (O Estado, 4 June 1941), even up to 100 thousand specimens (O Estado, 16 June 1933), which were captured using relatively simple fishing technology (gillnets and beach seine made of plant fibres). According to the captain Vieira da Rosa, on request of the Ministry of Agriculture in 1916, “The most abundant fish in the state is tainha do corso \[winter mullet schools\]…It can be estimated that more than one million mullets are caught every winter…” (O Dia, 8 August 1916). Considering the average weight of modern mullet during their winter reproductive migration (2 kg, but see also O Estado, 12 June 1923), this would correspond to 2000 t catch/winter, which is higher or comparable with artisanal landings/winter (May- July) in 2017 (1423 t), 2018 (1122 t) and 2019 (2522 t) for mullets in Santa Catarina state (data from PMAP-SC, <http://pmap-sc.acad.univali.br/index.html>). Although the productivity of winter migratory mullet schools can be affected by weather conditions, the reduced or equivalent landings in relation to the significantly increased fishing efforts of the more recent years supports the growing evidence that this species is currently overexploited in Santa Catarina. Other species nowadays considered overexploited, such as miraguaia (black drummer, Pogonias courbina) and corvina (whitemouth croaker, Micropogonias furnieri), were perceived as abundant and thus intensively targeted by coastal fisheries (Sul Americano, 12 August 1900, A Noticia, 18 February 1939, O Estado, 16 August 1930). High trophic level marine fish such as anchova (bluefish-enchova) and cação-mangona (sand tiger shark) were the main source of livelihood and financial capital to some coastal groups, as attested to in 1901: “…the most important trade that Ribeirão makes at the head of the municipality, in the city of Florianópolis, is that of fish, which annually yields a hundred thousand reis, according to a general estimate. Two species of fish mainly occupy the activity of this honourable people, anchova \[bluefish-enchova\] and a variety of cação \[sharks\], known locally by the name of mangona \[sand tiger shark\] or xarque by the populations of the southernmost part of the island” (Sul Americano, 22 July 1902). Newspaper evidence of fishing intensification in the early 20th century is largely corroborated by the official reports of the Ministry of Agriculture, which enable linking the study region discussed here (Santa Catarina) with the widespread ecological effects of historical fishing subsidies. In 1945, reports documented an increase of national fish production (fish byproduct) of 431% from 1934 (2112 t) to 1941 (11208 t). Brazil also shifted from importing 500 t of canned fish to exporting around 800 t of the product in 1943, mostly to South Africa and Guyana. This suggests that commercial fisheries in Brazil benefited from reduced competition with foreign producers due to significant disruptions in fishing trade from the North Atlantic, North Sea and the Baltic during World War II. The impact of these early subsides was also evident in the first official landing reports from the Entreposto Federal de Pesca in Rio de Janeiro, which indicated a 43% increase in landings (fish, invertebrates) from 1934 (13879 t) to 1940 (19913 t). As observed in most recent statistics, fisheries subsidies also caused a noticeable increase in overfishing and bycatching (prevalently juvenile specimens and small fish) in the first half of the 20th century in Brazil. According to the Ministry of Agriculture, the confiscation of marine animals due to being under the minimum established size (bycatching) and deterioration increased 707% between 1934 (10580 t) and 1940 (85445 t) in the same Entreposto Federal de Pesca. Strikingly, confiscated fish and shellfish due to bycatching only, performed a dramatic increment of \~6000% in just five years, from 945 t in 1934 to 57585 t in 1939. Overfishing and bycatching, however, were not limited to marine species as official reports expressed growing concerns about the declining catches of freshwater animals. For example, decreasing landings of pirarucu (Arapaima gigas) in Belém (north Brazil) between 1919 (1.8 t) and 1938 (1.45 t) was largely attributed to overfishing. Other species such as manatees (Trichechus sp.) and arrau turtle (Podocnemis expansa) also saw a decrease in abundance in the Amazon by the 1930’s, again attributed to intensive exploitation. ## Long-term change in species composition When considering native marine fish only, the WATL shows consistently high values between the decades of 1900–1909 (WATL = 3.6) and 1940–1949 (WATL = 3.4), suggesting that top predators were highly desirable market items for quite some time, even though several species were already under noticeable fishing pressure and declining in wholesale markets. Comparison with values for 2010–2019 reveals a sharp decline with the lowest values of the time series (WATL = 3.0). Interestingly, other independent indicators expressed the importance of high trophic level species in early decades, and their decreasing popularity in the most recent news. For example, planktivorous fish (Pis/Pla) and short-lived invertebrates (Inv/Fis) increased in their relative frequency of occurrence in newspapers from 1970 onwards at the expense of piscivorous fish. Significantly, long-term catch reconstructions for the coast of Santa Catarina show a similar trend in Inv/Fis, with catches of invertebrates increasing in the 1970’s and 1980’s, although this is followed by a downtrend in more recent decades. This possibly reflects fiscal incentives provided to the fisheries by SUDEPE from 1967 onwards, along with fishing agreements between Brazil, Uruguay and Argentina, and the increased investment in shrimp and mollusc aquaculture. For example, between 1966 and 1972 the number of fishing vessels employed in the catch of pink shrimp (Farfantepenaeus sp.) in Rio Grande (Rio Grande do Sul) increased from 80 to 396. In Santa Catarina shrimps were intensively exploited between 1970 and 1990 to the extent that coastal stocks declined considerably (A Ponta, July 1993,). Moreover, increased market opportunities and public investment for farming of molluscs (e.g. mussels, oysters) and crustaceans (e.g. prawn and shrimps) accelerated production from the second half of the 20th century, and notably from the late 1980s onward, which may explain their increased popularity in public media. Overall, the newspapers appear to capture the effect of market changes due to historical overexploitation and increased fishing efforts in Santa Catarina over the last 180 years. From 2000–2009 the newspapers documented a substantial increase in freshwater and non-native species. Reports of freshwater species (mostly non-native) were primarily associated with aquaculture and imports for direct human consumption (Correio do Povo, 23 February 2000; O Município, 15 August 2008; O Município, 13–14 April 2017), along with recreational fishing (Correio do Povo, 3 May 1969) and re-population of rivers (Correio do Povo, 9–15 August 1980; Correio do Povo, 7 December 2002). Such a trend conceivably reflects a combination of factors, including the reduced contribution of local marine stocks to domestic consumption, changes in population demography, and increased investments in aquaculture in Brazil, notably between the 1990 and 2020. Significantly, the growing global importance of aquaculture in recent years, known as the “blue revolution”, currently supplies half of all seafood for human consumption worldwide. ## Implications for marine conservation and Ocean Literacy principles Our study reinforces the value of historical newspapers in providing insightful information on long-term species composition, perceived abundance and the drivers of change in marine biodiversity in Brazil. However, contrary to other records such as official landing statistics, species composition in newspapers reflect distinct socio-ecological contexts, spanning from production to consumption systems, as well as considerations on the cultural and ecological values of some species, which are subject to evolving political, ideological and economic agendas, notably in social contexts with high levels of illiteracy as was the case in the late 19th and early 20th centuries in Brazil. Moreover, we acknowledge that information from newspapers is often fragmented, with substantial temporal and spatial gaps. Nevertheless, their use in historical ecology has proved fruitful and here we show their potential to uncover species composition (caught or observed) and related social, economic and market values for decades prior to monitoring programmes and scientific observations. Specifically, the late 19th and early 20th centuries witnessed the increased commoditization of marine resources in south Brazil. Profit- and efficiency- seeking fisheries policies accelerated the “economy of things” at the expense of the “economy of relationships” between communities and environments, exacerbating unsustainable exploitation and weak governance in the following decades. The early 20th century in particular can be considered the “incubation chamber” of fisheries industrialization in the country, but knowledge gaps on species composition and the ecological impacts of these early subsides may lead to the misconception that the commercial sector was limited in extent and volume, and that most fisheries in the late 19th and early 20th centuries were for local subsistence or incipient local markets. Such views impinge upon our understanding of the extent to which marine and ocean systems have been altered by anthropogenic activities. Newspapers provide valuable insights into the changing nature of long-term human interaction with marine biodiversity, helping to track the origin and evolution of Traditional Ecological Knowledge. This knowledge plays a fundamental role in fisheries management, biological conservation and coastal livelihoods, and newspapers reveal that it developed in a historical context of conflicts with institutions and political and economic actors, who often associated local fishers with backwardness, poverty, and destitution, and attributed the richness of their fishing practices with detrimental impacts and economic underdevelopment. The establishment of the Fishery Inspection Agency in 1912 is an example of how political institutions attempted to override local fisheries knowledge with “modern” knowledge in order to increase efficiency and unlock the market potential of aquatic resources in the country: “Until now, unfortunately, this industry \[fishing\], among us, has not passed from the backward and old primitive processes, the hook, the trawler and, lately, the dynamite, ruthlessly damaging the creations of the coasts and almost extinguishing several lacustrine, fluvial and maritime species. The improved fishing gear that is currently being used, with great results, without impoverishing the waters in which they are exercised, and mechanical fishing by means of steamboats have not yet been used…”. This market logic has for decades contributed to dismantling traditional forms of socio-ecological interactions. Their legacy can still be perceived among modern small-scale fisheries and the social and political context around them. Given their role in shaping public opinion through time, historical newspapers offer a rich repository of shifting ecological baselines. Therefore they are valuable resources to inform the ambitious Mission of the UN Decade of Ocean Science for Sustainable Development to promote “Transformative ocean science solutions for sustainable development, connecting people and our ocean”. A core driving concept advanced by the UN Ocean Decade is to push countries to initiate Ocean Literacy programs across the globe, in order to promote better public understanding, communication and informed policy decisions for ocean sustainability. By resurrecting previously hidden social-ecological baselines, historical records can help Ocean Literacy programs advance on essential principles and fundamental concepts about the functioning of the ocean, including: clarity on the diversity of life and ecosystems oceans once supported (principle 5); acknowledgment of deep, historical interconnections humans have held with the ocean (principle 6); and more realistic baselines about the level of human exploration of coastal marine environments (principle 7). # Conclusion Humans have depended on marine ecosystems as a source of food and livelihood for thousands of years along the Brazilian coast, a process that has laid down the foundation of a diversity of fishing cultures in the region. Over the last few decades, increased fishing demand, overoptimistic cycles of profit-driven subsidy programmes and weak governance models intensified commercial exploitations, leading to unprecedented levels of catches and the decline of a range of stocks. This work advances our understanding of this convoluted historical process by expanding the current knowledge of captured and consumed marine animals in southern Brazil by nearly two centuries, covering decades predating national official landing reports and market information. Our results highlight the utility of newspaper records in resurrecting previously lost marine social-ecological baselines, with a high potential to support Ocean Literacy programs to rightfully contextualise historical realities when shaping sustainability pathways with ocean stakeholders at all levels. # Supporting information We would like to thank Leopoldo Cavaleri Gerhardinger and Krista McGrath for their insightful comments on the first draft of the manuscript. The authors are also grateful to Natalia Hanazaki and the Laboratory of Human Ecology and Ethnobotany (Universidade Federal de Santa Catarina, Brazil). 10.1371/journal.pone.0284024.r001 Decision Letter 0 MacKenzie Brian R. Academic Editor 2023 Brian R. MacKenzie This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 18 Jan 2023 PONE-D-22-32178180 years of marine animal diversity as perceived by public media in southern BrazilPLOS ONE Dear Dr. Herbst, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please consider the comments of reviewer 1 regarding potential use of official fishery statistics to support and complement your findings. This should be possible for some of the more recent decades though probably not for the earliest decades where official records may be more limited or non-existent. Also please make changes according to comments of both reviewers regarding several points for clarification and readability. Please submit your revised manuscript by Mar 04 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Brian R. MacKenzie, Ph. D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatt ing_sample_title_authors_affiliations.pdf> 2\. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. 3\. We note that Figure 1 in your submission contain map image which may be copyrighted. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For these reasons, we cannot publish previously copyrighted maps or satellite images created using proprietary data, such as Google software (Google Maps, Street View, and Earth). For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figure 1 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. The following resources for replacing copyrighted map figures may be helpful: USGS National Map Viewer (public domain): <http://viewer.nationalmap.gov/viewer/> The Gateway to Astronaut Photography of Earth (public domain): <http://eol.jsc.nasa.gov/sseop/clickmap/> Maps at the CIA (public domain): <https://www.cia.gov/library/publications/the- world-factbook/index.html> and <https://www.cia.gov/library/publications/cia- maps-publications/index.html> NASA Earth Observatory (public domain): <http://earthobservatory.nasa.gov/> Landsat: <http://landsat.visibleearth.nasa.gov/> USGS EROS (Earth Resources Observatory and Science (EROS) Center) (public domain): <http://eros.usgs.gov/#> Natural Earth (public domain): <http://www.naturalearthdata.com/> 4\. We note that Figure 2 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figure 2 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. 5\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: [http://journals.plos.org/plosone/s/supporting- information. ](http://journals.plos.org/plosone/s/supporting-information. ) \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Overview: The work brings a historical rescue of fishing data for one of the largest production regions at national level (Santa Catarina). Although the approach does not bring methodological novelties, the work is relevant considering the scarcity/fragmentation of information at the local/national level. The work is well structured, however, below are some suggestions to mitigate the bias related to the search for information in newspapers. I believe that authors should look for ways to compare these historical records with official fishing statistics existing at the regional level. In this way, it will be possible to assess whether, in this specific study, newspaper records reflect (or not) official statistics, at least for the most recent periods (20th and 21st centuries). Methods: Page 5, Lines 131-132: It is necessary to indicate which were the keywords used for the search of all the resources (fish, crustaceans, molluscs, mammals, reptiles), since the results show several resources and not just 'fish'. P 6, L 141-143: Please include how the search was performed on this platform (keywords, etc.). P 6, L 165-167: Did you use the ratio between abundance, catch weight (landings)? Make it more clear to the reader. Results: P 11, L 272-273: This result may be related to the projects that promoted the implementation of the industrial productive segment of shrimp in Santa Catarina in the 1970s (e.g., SUDEPE). In addition to projects aimed at growing mangrove oysters in association with shrimp in the 1980s. In this way, there is a need for a greater deepening in the discussion about public policies directed to the productive sector that promoted the cultivation of these aquatic organisms, bearing in mind that this may influence the publicity of journalistic articles on the subject. Discussion: P 15, L 397-421: This paragraph is outside the scope of the study, as it provides an overview of fishing at the national level, while the study focuses on the regional level (Santa Catarina). Is the pattern described at the national level also observed at the regional level? I suggest that this paragraph be revised according to the purpose of the study. P 16, L 433-435: this may be related to what was pointed out above (see comment in Results section). P 17, L 451-452: Is this trend also observed in Santa Catarina? P 18, L 505-506: As the authors themselves recognize above, the records obtained from newspapers can suffer from several biases. One possibility of mitigating this bias and verifying whether newspaper records actually reflect catches is through cross-analysis with historical sources of fisheries statistics. I agree that statistical records are incomplete, fragmented and scattered, but it is possible to use some official data sources that can serve to assess general trends. For example, it is possible to use ICMBio/CEPSUL records (<https://www.icmbio.gov.br/cepsul/biblioteca/acervo- digital/38-download/artigos-cientificos/112-artigos-cientificos.html>), IBGE (<https://biblioteca.ibge.gov.br/biblioteca- catalogo.html?id=7132&view=detalhes>) and IBAMA (<https://www.gov.br/ibama/pt- br/assuntos/biodiversidade/biodiversidade-aquatica/gestao-da-biodiversidade- aquatica/estatistica-pesqueira>). I believe it might be possible to check the general trend between invertebrates/fishes with this database. Thus, we can assess whether the methodology used in the work reflects (at least partially) the official statistics (albeit fragmented). Figure 3A,B: include in the legend the meaning of the colors present in the figure (A, B). There are two axes but three colors. Reviewer \#2: This is an interesting paper that is scientifically sound. My comments are minor and mostly relate to clarifying parts of the text to ensure accurate interpretation of methods and results by readers. L43: ‘…while ensuring fundamental ecosystem functions, structures and services’ Please provide a reference for this. L50: ‘This sort of historical amnesia creates misconceptions…’ This sentence is very long. Suggest splitting up to improve readability. L107: ‘…scrutinised more than twenty thousand digitised editions in the Brazilian Digital Newspapers and Periodicals Library (Hemeroteca Digital Brasileira)’. What do you mean by ‘scrutinised’? If you did not examine at all of these (rather, you conducted key-word searches via OCR across this number of newspapers) this phrasing is rather misinformative. Fig 1 Legend: The coloured dots should be labelled as \>1 – 10 reports, \>10 – 20 reports… \>30 – 200 reports etc for greater clarity. Also, ‘The colour gradient of the localities represents the number of reports IDENTIFIED per municipality’. L133: what dates did you undertake data collection? You mention dates in the results section but it is not clear if these were dates of collection – this is important to know given that more newspaper articles/editions are being uploaded to digital collections all the time. Given you identified animals from articles initially located by keywords fishing and fish, how likely is it that you’ve missed accounts of species that were gathered or cultured, or otherwise exploited/identified using alternative words to ‘fishing’ or ‘fish’? L159: Suggest ‘Species richness REPORTED by decade’ for clarity. L184: Suggest restructuring of this paragraph as it initially reads that you individually screened the \>20,000 articles, rather than identify the ca. 1200 from SC which were then individually screened. Regarding the 4.7% of the total newspapers you mention, were the 95.3% of papers available from the SC region but that did not get identified from the OCR searches (i.e. did not mention fishing or fish)? In general, this paragraph is unclear. L198: I.e. ‘(v = 203, s = 107)’: Do you mean that a total of 310 species were identified, or are these two numbers not additive? Same question for the other numbers that follow. L199: ‘frequency (n = 1,041)’: Do you mean frequency of mentions? L234: ‘The number of items reporting marine animals using the keywords fishing (n = 269) and fish (n = 329) were significantly positively correlated with the number of newspapers (fishing r = 0.61; R2 = 0.37; p \< 0.005; fish r = 0.52; R2 = 0.27; p \< 0.02)’ Can you explain this more clearly? Are you referring to the total number of newspapers available (i.e. published online) in the region, or those identified through OCR searches as containing the words ‘fishing’ or ‘fish’? This is clarified somewhat by the legend in Fig 3 but it needs to be clear in the text, too. Fig. 3A and 3B – there are three colours in this panel figure, the black I assume correlates to the black text on the y axis descriptor (fig 3a), but there are also two red block colours in this figure and I am now sure how these align with B and the red text on the secondary y axis (#items). If these are aggregated per the legend description, why two red colours? Also, what does ‘items’ mean? Is this referring to individual articles within a newspaper edition? Fig. 3C-E – there seems to be very little difference between absolute species counts and species richness. It may be worth briefly repeating in the text (or legend description) the difference between these two measures, as it gets a bit confusing when examining the figure. In general, the way your results are written it reads as absolute species richness or species numbers, when actually it is richness or numbers reported in newspapers. I think this needs to be stated more clearly throughout the results. I.e., L256: By contrast, the richness and frequency of invertebrates REPORTED increased… L263: The WATL shows the prevalence of high trophic level species REPORTED over the last 180 years… L287: The diversity and frequency of fishing gear REPORTED did not change… L466: you state the potential for newspapers to uncover species composition, better to clarify here that it is composition of fishery catch and/or landings? This of course reflects to some extent underlying ecological composition, but acknowledges the influences of social (in terms of what is reported) and market forces and technological trends (in terms of what is caught) on newspaper reporting. L473: misspelling of ‘unsustainable’. L482: The relevance of this paragraph on LEK to the wider paper is unclear and it doesn’t tie in overly well. The historical context of conflicts you mention regarding LEK will undoubtedly also play out in popular media, it being a window into past cultures. But I’m not sure if you’re mentioning this to say that this is indeed the case with newspapers or to say that historical newspaper sources can mitigate this issue? L485. Your definition of LEK seems more akin to traditional ecological knowledge in terms of intergenerational transmission. While I am aware that TEK can be defined as a form of LEK, I think this needs to be clarified. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: Yes: Cleverson Zapelini Reviewer \#2: No \\\\\\\\\\ \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0284024.r002 Author response to Decision Letter 0 26 Feb 2023 PONE-D-22-32178 180 years of marine animal diversity as perceived by public media in southern Brazil PLOS ONE Dear Academic Editor Thank you for considering our manuscript entitled 180 years of marine animal diversity as perceived by public media in southern Brazil, by Herbst et al. for publication in PLOS ONE. We have now completed the revision of our manuscript according to the comments made by Reviewer \#1 and \#2 and the Academic Editor. We would like to thank you in the quality of Academic Editor and both reviewers for your constructive comments which helped to improve the quality of the manuscript and to clarify some of our statements. Please find below our rebuttal. All changes to the manuscript are marked in red to facilitate their identification. We would like to update our financial statement with the statement below: This work was funded by the ERC Consolidator project TRADITION, which has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme under Grant Agreement No 817911. This work was also funded by EarlyFoods (Evolution and impact of early food production systems), 2021 SGR 00527. This work contributes to the ICTA-UAB “María de Maeztu'' Programme for Units of Excellence of the Spanish Ministry of Science and Innovation (CEX2019-000940-M). We hope the manuscript can now be accepted for publication. Kind regards, Herbst et al. ------------------------------- Dear Dr. Herbst, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please consider the comments of reviewer 1 regarding potential use of official fishery statistics to support and complement your findings. This should be possible for some of the more recent decades though probably not for the earliest decades where official records may be more limited or non-existent. Also please make changes according to comments of both reviewers regarding several points for clarification and readability. Please submit your revised manuscript by Mar 04 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: ● A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. ● A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. ● An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Brian R. MacKenzie, Ph. D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatt ing_sample_title_authors_affiliations.pdf> 2\. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. 3\. We note that Figure 1 in your submission contain map image which may be copyrighted. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For these reasons, we cannot publish previously copyrighted maps or satellite images created using proprietary data, such as Google software (Google Maps, Street View, and Earth). For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figure 1 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. The following resources for replacing copyrighted map figures may be helpful: USGS National Map Viewer (public domain): <http://viewer.nationalmap.gov/viewer/> The Gateway to Astronaut Photography of Earth (public domain): <http://eol.jsc.nasa.gov/sseop/clickmap/> Maps at the CIA (public domain): <https://www.cia.gov/library/publications/the- world-factbook/index.html> and <https://www.cia.gov/library/publications/cia- maps-publications/index.html> NASA Earth Observatory (public domain): <http://earthobservatory.nasa.gov/> Landsat: <http://landsat.visibleearth.nasa.gov/> USGS EROS (Earth Resources Observatory and Science (EROS) Center) (public domain): <http://eros.usgs.gov/#> Natural Earth (public domain): <http://www.naturalearthdata.com/> R: The figures were generated using free access images (Public Domain or CC BY 4.0 licence) and references to their sources are included in the revised figure captions. We provide more information about the individual licences below: ● Natural Earth (<https://www.naturalearthdata.com/about/terms-of-use/>) ● National Institute for Space Research - INPE (<http://terrabrasilis.dpi.inpe.br/citacoes-e-licenca-de-uso/>) ● CGIAR Consortium for Spatial Information (<https://cgiarcsi.community/data/srtm-90m-digital-elevation-database-v4-1/>) 4\. We note that Figure 2 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figure 2 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. R: The figures were generated using free access images (Public Domain or CC BY 4.0 licence) and references to their sources are included in the revised figure captions. We provide more information about the individual licences below: ● PhyloPic (<http://phylopic.org/>) 5\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: <http://journals.plos.org/plosone/s/supporting-information>. R: Done, captions for Supporting Information files now reported \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes 3\. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Overview: The work brings a historical rescue of fishing data for one of the largest production regions at national level (Santa Catarina). Although the approach does not bring methodological novelties, the work is relevant considering the scarcity/fragmentation of information at the local/national level. The work is well structured, however, below are some suggestions to mitigate the bias related to the search for information in newspapers. I believe that authors should look for ways to compare these historical records with official fishing statistics existing at the regional level. In this way, it will be possible to assess whether, in this specific study, newspaper records reflect (or not) official statistics, at least for the most recent periods (20th and 21st centuries). R: We thank the reviewer \#1 for his/her/their suggestion. We want to highlight that the emphasis of our paper is on the perception of mass media on marine biodiversity changes, for this reason, we believe that attempts to validate data from newspapers (and derivatives) with independent landing records may risk to open more questions than answers, and the nature of these questions may remain largely contentious. Moreover, official fishing statistics could also carry other analytical and statistical biases complex to resolve, as exemplified by Freire et at \[1,2\]. However, following suggestions by the Reviewer \#1 we have now compared our Inv/Fis index with the estimated index from reconstructed and officially reported industrial and artisanal catch data (volume in tons) for Santa Catarina from 1950 to 2015 \[1\] (see lines 184 to 186, page 6). We have revised the text which now reads (lines 474 to 488, page 15): “Significantly, long-term catch reconstructions for the coast of Santa Catarina show a similar trend in Inv/Fis, with catches of invertebrates increasing in the 1970’s and 1980’s, although this is followed by a downtrend in more recent decades \[1\]. This possibly reflects fiscal incentives provided to the fisheries by SUDEPE from 1967 onwards, along with fishing agreements between Brazil, Uruguay and Argentina, and the increased investment in shrimp and mollusc aquaculture \[3–6\]. For example, between 1966 and 1972 the number of fishing vessels employed in the catch of pink shrimp (Farfantepenaeus sp.) in Rio Grande (Rio Grande do Sul) increased from 80 to 396\[4\]. In Santa Catarina shrimps were intensively exploited between 1970 and 1990 to the extent that coastal stocks declined considerably (A Ponta, July 1993,\[5\]). Moreover, increased market opportunities and public investment for farming of molluscs (e.g. mussels, oysters) and crustaceans (e.g. prawn and shrimps) accelerated production from the second half of the 20th century, and notably from the late 1980s\[6,7\] onward, which may explain their increased popularity in public media. Methods: Page 5, Lines 131-132: It is necessary to indicate which were the keywords used for the search of all the resources (fish, crustaceans, molluscs, mammals, reptiles), since the results show several resources and not just 'fish'. R: The keywords used are those presented in the original text (lines 131 and 132, page 5 of the original manuscript): “pesca” (fishing) and "peixe" (fish). We have added the following texts to clarify the importance of these two keywords: Material and Methods (Lines 141 to 151, page 5) “The keywords “pesca” (fishing) and "peixe" (fish) were chosen for the following reasons: i) The keyword “pesca” captured several types of marine resources, and also captured the word "pescado" which categorically addresses different types of resources exploited by fisheries; ii) within Brazilian folk taxonomy (ethnotaxonomy), various marine animals ("pescado") are classified as “fish”\[8,9\]; iii) the popular association of fish such as shrimp, lobster, oysters, octopus and squid with crustaceans and molluscs is something very recent and thus this nomenclature is infrequently used in everyday life; iv) given the large diversity of organisms captured by coastal fisheries in Brazil and the general scope of the paper (not targeting a specific species), it would be unfeasible to extend the keywords to all potential captured organisms.” Results (Lines 215 to 217, page 7) “The keywords “pesca” (fishing) and "peixe" (fish) captured animals from different taxonomic groups (fish, crustaceans, molluscs, mammals, reptiles), which is significant since in folk taxonomy several taxa are also considered as “fish”.” P 6, L 141-143: Please include how the search was performed on this platform (keywords, etc.). R: The method is reported in detail in Galardo et al. \[10\], and succinctly described in the Material and Methods. We have now expanded it to clarify the steps and process, please see lines 131 to 169 (Pages 5 and 6). P 6, L 165-167: Did you use the ratio between abundance, catch weight (landings)? Make it more clear to the reader. R: We used the absolute abundance of species per decade. The sentence has been revised for clarification. Now it's reads (lines 176 to 179, page 6): “In addition, we computed fish-based piscivores/planktivores (Pis/Pla)\[11\] and invertebrates/fish (Inv/Fis) to explore changes in overall catch composition through time. For this we used the absolute abundance of species reported in newspapers per decade.” Results: P 11, L 272-273: This result may be related to the projects that promoted the implementation of the industrial productive segment of shrimp in Santa Catarina in the 1970s (e.g., SUDEPE). In addition to projects aimed at growing mangrove oysters in association with shrimp in the 1980s. In this way, there is a need for a greater deepening in the discussion about public policies directed to the productive sector that promoted the cultivation of these aquatic organisms, bearing in mind that this may influence the publicity of journalistic articles on the subject. R: Yes, we agree and have now expanded the argument in the discussion, see please lines 477 to 488, page 15. The impact of public policies, as well as of evolving ideological and economic agendas on the quality and number of news is acknowledged in lines 459 and 464 (page 17 original version). We have expanded the argument (lines 507 to 515, page 16): “Our study reinforces the value of historical newspapers in providing insightful information on long-term species composition, perceived abundance and the drivers of change in marine biodiversity in Brazil. However, contrary to other records such as official landing statistics, species composition in newspapers reflect distinct socio-ecological contexts, spanning from production to consumption systems, as well as considerations on the cultural and ecological values of some species, which are subject to evolving political, ideological and economic agendas, notably in social contexts with high levels of illiteracy as was the case in the late 19th and early 20th centuries in Brazil\[12–14\].” Discussion: P 15, L 397-421: This paragraph is outside the scope of the study, as it provides an overview of fishing at the national level, while the study focuses on the regional level (Santa Catarina). Is the pattern described at the national level also observed at the regional level? I suggest that this paragraph be revised according to the purpose of the study. R: We disagree with the Reviewer \#1. The evidence discussed in this paragraph is highly relevant and within the scope of the paper. It enabled us to scale up our showcase (Santa Catarina) with the widespread ecological effects of increased fishing effort in Brazil. In the discussion, by considering the results of the reports from the Ministry of Agriculture (responsible for fisheries management in Brazil during the analysed timeframe), we enlarge the context of fishing to national and international levels, which is also crucial for tracking the development of fishing in Santa Catarina. We have slightly modified the text to emphasise its relevance and now it reads (lines 534 to 537, page 14): “Newspaper evidence of fishing intensification in the early 20th century is largely corroborated by the official reports of the Ministry of Agriculture, which enable linking the study region discussed here (Santa Catarina) with the widespread ecological effects of historical fishing subsidies.” P 16, L 433-435: this may be related to what was pointed out above (see comment in Results section). R: Thanks, the comment has been addressed above. P 17, L 451-452: Is this trend also observed in Santa Catarina? R: Yes, this references mentioned analysis and trend of Santa Catarina. P 18, L 505-506: As the authors themselves recognize above, the records obtained from newspapers can suffer from several biases. One possibility of mitigating this bias and verifying whether newspaper records actually reflect catches is through cross-analysis with historical sources of fisheries statistics. I agree that statistical records are incomplete, fragmented and scattered, but it is possible to use some official data sources that can serve to assess general trends. For example, it is possible to use ICMBio/CEPSUL records (<https://www.icmbio.gov.br/cepsul/biblioteca/acervo- digital/38-download/artigos-cientificos/112-artigos-cientificos.html>), IBGE (<https://biblioteca.ibge.gov.br/biblioteca- catalogo.html?id=7132&view=detalhes>) and IBAMA (<https://www.gov.br/ibama/pt- br/assuntos/biodiversidade/biodiversidade-aquatica/gestao-da-biodiversidade- aquatica/estatistica-pesqueira>). I believe it might be possible to check the general trend between invertebrates/fishes with this database. Thus, we can assess whether the methodology used in the work reflects (at least partially) the official statistics (albeit fragmented). R: Done, comparison of Inv/Fis was performed with data published in Freire et al. \[1\] that includes data of different sources. See previous comments Figure 3A,B: include in the legend the meaning of the colors present in the figure (A, B). There are two axes but three colors. R: Done, colours made clear now. Reviewer \#2: This is an interesting paper that is scientifically sound. My comments are minor and mostly relate to clarifying parts of the text to ensure accurate interpretation of methods and results by readers. L43: ‘…while ensuring fundamental ecosystem functions, structures and services’ Please provide a reference for this. R: Done! L50: ‘This sort of historical amnesia creates misconceptions…’ This sentence is very long. Suggest splitting up to improve readability. R: Done, it now reads (lines 49 to 55 page 2): “This form of historical amnesia creates misconceptions that have far-reaching consequences for sustainability and conservation actions. For example, it may conceivably increase tolerance for progressive environmental degradation and contribute to setting inappropriate sustainability targets and responses by ocean stakeholders; and hence underpin their expectations as to what is a desirable and achievable state of social-ecological systems\[15–17\].” L107: ‘…scrutinised more than twenty thousand digitised editions in the Brazilian Digital Newspapers and Periodicals Library (Hemeroteca Digital Brasileira)’. What do you mean by ‘scrutinised’? If you did not examine all of these (rather, you conducted key-word searches via OCR across this number of newspapers) this phrasing is rather misinformative. R: The search via OCR with the keywords "pesca" and "fish" found 9,190 and 14,168 matches respectively (see in the results), in other words, more than twenty thousand digitised editions (newspaper reports) were accessed with these keywords. Scrutinised means that we examined closely and thoroughly all the news items. Fig 1 Legend: The coloured dots should be labelled as \>1 – 10 reports, \>10 – 20 reports… \>30 – 200 reports etc for greater clarity. Also, ‘The colour gradient of the localities represents the number of reports IDENTIFIED per municipality’. R: Thank you for your suggestion! Done! L133: what dates did you undertake data collection? You mention dates in the results section but it is not clear if these were dates of collection – this is important to know given that more newspaper articles/editions are being uploaded to digital collections all the time. R: From March 2021 to March 2022. This information has now been added to the text (Lines 196 to 197, page 6) Given you identified animals from articles initially located by keywords fishing and fish, how likely is it that you’ve missed accounts of species that were gathered or cultured, or otherwise exploited/identified using alternative words to ‘fishing’ or ‘fish’? R: It is possible that we could have had a more extensive database. However, within a long list of complementary and/or alternative keywords, the use of two generic keywords (fishing and fish) that can be applicable elsewhere covered a broad range of marine organisms. L159: Suggest ‘Species richness REPORTED by decade’ for clarity R: Done! L184: Suggest restructuring of this paragraph as it initially reads that you individually screened the \>20,000 articles, rather than identify the ca. 1200 from SC which were then individually screened. Regarding the 4.7% of the total newspapers you mention, were the 95.3% of papers available from the SC region but that did not get identified from the OCR searches (i.e. did not mention fishing or fish)? In general, this paragraph is unclear. R: The text has been revised for clarity. About the other 95.3% of newspapers available from SC did not meet the inclusion criteria, namely i) the items focused on coastal and ocean areas, including estuaries and coastal lagoons, or coastal rivers; and ii) the items reported marine animals (e.g. fish, crustaceans, molluscs, mammals, reptiles), including non-native taxa (Supplementary Information 1, SI1). L198: I.e. ‘(v = 203, s = 107)’: Do you mean that a total of 310 species were identified, or are these two numbers not additive? Same question for the other numbers that follow. R: These are two independent classifications, therefore they are not additive. We have amended the text for clarification, and now it reads (lines 215 to 227, pages 7 and 8): “The keywords “pesca” (fishing) and "peixe" (fish) captured animals from different taxonomic groups (fish, crustaceans, molluscs, mammals, reptiles), which is significant since in folk taxonomy several taxa are also considered as “fish”. Native and non-native animals were mostly reported as vernacular (v) and, to a lesser extent, as scientific (s) names, the latter sporadically appearing in more recent decades. We found 1655 mentions of marine animals, totaling 277 and 203 vernacular and scientific taxa respectively (SI2). As reported by Feire and Pauly\[18\], we also found evidence that a single folk nomenclature was used for several species (e.g. Hypanus sp., Dasyatis sp., Fontitrygon sp., Bathytoshia sp. = raia prego), and several folk nomenclatures were also used for a single species (e.g. Mugil liza = tainha, tainhota, tanhota, do corso, facão, tainha-assú). As a consequence, attempts to correlate scientific and vernacular names were complex and not always possible with an equal level of confidence.” L199: ‘frequency (n = 1,041)’: Do you mean frequency of mentions? R: We have amended it for “abundance”. In Material and Methods it now reads (lines 164 to 169, page 5): “The catalogued animals and gear were quantified for their absolute and relative abundances (the number of times a species or type of gear occurred in a given number of newspaper items) and richness (the number of different species and gear types in a given number of newspaper items) aggregated per decade. If species X was reported more than once in a single item, its abundance in the given item was counted as one.” L234: ‘The number of items reporting marine animals using the keywords fishing (n = 269) and fish (n = 329) were significantly positively correlated with the number of newspapers (fishing r = 0.61; R2 = 0.37; p \< 0.005; fish r = 0.52; R2 = 0.27; p \< 0.02)’ Can you explain this more clearly? Are you referring to the total number of newspapers available (i.e. published online) in the region, or those identified through OCR searches as containing the words ‘fishing’ or ‘fish’? This is clarified somewhat by the legend in Fig 3 but it needs to be clear in the text, too. R: Thanks, the sentence has been amended. Now it reads (lines 265 to 274, page 9): “The number of items reporting marine animals using the keywords “pesca” (fishing) (n = 269) and "peixe" (fish) (n = 329) were significantly positively correlated with the number of newspapers for Santa Catarina available in the HDB archives (pesca r = 0.61; R2 = 0.37; p \< 0.005; peixe r = 0.52; R2 = 0.27; p \< 0.02), a pattern also observed by Gallardo et al.\[10\] (Figure 3A-B). Similarly, species richness and their abundances were also significantly positively correlated with the number of newspapers (species richness, r = 0.73; R2 = 0.53; p \< 0.001, frequency, r = 0.68; R2 = 0.46; p \< 0.001) in Santa Catarina (Figure 3C-D), indicating that, in absolute terms, our data is affected by the amount of digitised newspapers available in the HDB database.” Fig. 3A and 3B – there are three colours in this panel figure, the black I assume correlates to the black text on the y axis descriptor (fig 3a), but there are also two red block colours in this figure and I am now sure how these align with B and the red text on the secondary y axis (#items). If these are aggregated per the legend description, why two red colours? Also, what does ‘items’ mean? Is this referring to individual articles within a newspaper edition? R: There were only two variables, but the transparency in the plot created a third colour. The figure has been amended. “Items” refer to “articles, opinion letters, regulations and sanctions”, see lines 136 (page 5). We revised the figure caption for clarity. Fig. 3C-E – there seems to be very little difference between absolute species counts and species richness. It may be worth briefly repeating in the text (or legend description) the difference between these two measures, as it gets a bit confusing when examining the figure. R: Their trends are similar but the numbers are different (please see their respective y-axes). However we noticed that we had switched the labels, which have now been corrected. In general, the way your results are written it reads as absolute species richness or species numbers, when actually it is richness or numbers reported in newspapers. I think this needs to be stated more clearly throughout the results. I.e., L256: By contrast, the richness and frequency of invertebrates REPORTED increased… L263: The WATL shows the prevalence of high trophic level species REPORTED over the last 180 years… L287: The diversity and frequency of fishing gear REPORTED did not change… R: Yes, but it is implicit that these refer to reported species and gear in newspapers, not in absolute terms. However, for clarity we have revised the terminology through the text and in the Figure 3D-E. Absolute (relative) frequency has been replaced with absolute (relative) abundance (the amount of a species/gear in newspaper items aggregated per decade) L466: you state the potential for newspapers to uncover species composition, better to clarify here that it is composition of fishery catch and/or landings? This of course reflects to some extent underlying ecological composition, but acknowledges the influences of social (in terms of what is reported) and market forces and technological trends (in terms of what is caught) on newspaper reporting. R: The text has been revised to address this particular comment and now it reads (lines 509 to 515, page 16): “However, contrary to other records such as official landing statistics, species composition in newspapers reflect distinct socio-ecological contexts, spanning from production to consumption systems, as well as considerations on the cultural and ecological values of some species, which are subject to evolving political, ideological and economic agendas, notably in social contexts with high levels of illiteracy as was the case in the late 19th and early 20th centuries in Brazil\[12–14\].” L473: misspelling of ‘unsustainable’. R: Thanks! L482: The relevance of this paragraph on LEK to the wider paper is unclear and it doesn’t tie in overly well. The historical context of conflicts you mention regarding LEK will undoubtedly also play out in popular media, it being a window into past cultures. But I’m not sure if you’re mentioning this to say that this is indeed the case with newspapers or to say that historical newspaper sources can mitigate this issue? R: Thanks. The sentence has been revised for clarification. Our point here is about understanding how LEK are formed through time. The sentence now reads (lines 535 to 542, pages 17): “Newspapers provide valuable insights into the changing nature of long-term human interaction with marine biodiversity, helping to track the origin and evolution of Traditional Ecological Knowledge. This knowledge plays a fundamental role in fisheries management\[19\], biological conservation\[20,21\] and coastal livelihoods\[22–24\], and newspapers reveal that it developed in a historical context of conflicts with institutions and political and economic actors, who often associated local fishers with backwardness, poverty, and destitution, and attributed the richness of their fishing practices with detrimental impacts and economic underdevelopment\[10\].” L485. Your definition of LEK seems more akin to traditional ecological knowledge in terms of intergenerational transmission. While I am aware that TEK can be defined as a form of LEK, I think this needs to be clarified. R: Yes, it is on ecological knowledge that is transmitted through generations. We have replaced LEK for TEK. References 1\. Freire KMF, Almeida Z da S de, Amador JRET, Aragão JA, Araújo AR da R, Ávila-da-Silva AO, et al. Reconstruction of Marine Commercial Landings for the Brazilian Industrial and Artisanal Fisheries From 1950 to 2015. Frontiers in Marine Science. 2021;8: 946. 2\. Freire KMF, Aragão JAN, Araújo ARR, Ávila-da-Silva AO, Bispo MCS, Velasco G, et al. Reconstruction of catch statistics for Brazilian marine waters (1950-2010). University of British Columbia. Fisheries Centre; 2015. Report No.: 23 (4). doi:10.14288/1.0354313 3\. Abdallah PR, Sumaila UR. An historical account of Brazilian public policy on fisheries subsidies. Mar Policy. 2007;31: 444–450. 4\. Reis EG, D’Incao F. The present status of artisanal fisheries of extreme Southern Brazil: an effort towards community-based management. Ocean Coast Manag. 2000;43: 585–595. 5\. D’Incao F, Valentini H, Rodrigues LF. Avaliação da pesca de camarões nas regiões Sudeste e Sul do Brasil (1965-1999). 2002. Available: <http://repositorio.furg.br/handle/1/5716> 6\. Lopes PFM. Extracted and farmed shrimp fisheries in Brazil: economic, environmental and social consequences of exploitation. Environ Dev Sustainability. 2008;10: 639. 7\. Valenti WC, Barros HP, Moraes-Valenti P, Bueno GW, Cavalli RO. Aquaculture in Brazil: past, present and future. Aquaculture Reports. 2021;19: 100611. 8\. Souza SP, Begossi A. Whales, dolphins or fishes? The ethnotaxonomy of cetaceans in São Sebastião, Brazil. J Ethnobiol Ethnomed. 2007;3: 9. 9\. Previero M, Minte-Vera CV, Moura RL de. Fisheries monitoring in Babel: fish ethnotaxonomy in a hotspot of common names. Neotrop Ichthyol. 2013;11: 467–476. 10\. Sandoval Gallardo S, Fossile T, Herbst DF, Begossi A, Silva LG, Colonese AC. 150 years of anthropogenic impact on coastal and ocean ecosystems in Brazil revealed by historical newspapers. Ocean Coast Manag. 2021;209: 105662. 11\. Caddy JF, Garibaldi L. Apparent changes in the trophic composition of world marine harvests: the perspective from the FAO capture database. Ocean Coast Manag. 2000;43: 615–655. 12\. Pallares-Burke MLG. A imprensa periódica como uma empresa educativa no século XIX. Cadernos De Pesquisa. 2013;104: 144–161. 13\. Oliveira RS. A relação entre a história e a imprensa, breve história da imprensa e as origens da imprensa no Brasil (1808-1930). Historiæ. 2011;2: 125–142. 14\. Marchelli PS. As minorias alfabetizadas no final do período colonial e sua transição para o império: um estudo sobre a história social e educação no Brasil. EU. 2006;10: 187–200. 15\. Soga M, Gaston KJ. Shifting baseline syndrome: causes, consequences, and implications. Front Ecol Environ. 2018;16: 222–230. 16\. Fulton EA, Sainsbury K, Noranarttragoon P, Leadbitter D, Staples DJ, Porobic J, et al. Shifting baselines and deciding on the desirable form of multispecies maximum sustainable yield. ICES J Mar Sci. 2022 \[cited 27 Sep 2022\]. doi:10.1093/icesjms/fsac150 17\. Campbell L, Gray N, Hazen E, Shackeroff J. Beyond Baselines: Rethinking Priorities for Ocean Conservation. Ecol Soc. 2009;14. doi:10.5751/ES-02774-140114 18\. Freire KMF, Pauly D. RICHNESS OF COMMON NAMES OF BRAZILIAN MARINE FISHES AND ITS EFFECT ON CATCH STATISTICS. Journal of Ethnobiology. 2005;25: 279–296. 19\. Tietze U. Technical and socio-economic characteristics of small-scale coastal fishing communities and opportunities for poverty alleviation and empowerment. FAO Fisheries and Aquaculture Circular; Rome N.o C. search.[proquest.com](http://proquest.com); 2016. pp. 1–88. Available: <https://search.proquest.com/openview/72e02c4d532ae5760e1481415b405b1b/1?pq- origsite=gscholar&cbl=237324> 20\. Reyes-García V, Fernández-Llamazares Á, Aumeeruddy-Thomas Y, Benyei P, Bussmann RW, Diamond SK, et al. Recognizing Indigenous peoples’ and local communities' rights and agency in the post-2020 Biodiversity Agenda. Ambio. 2022;51: 84–92. 21\. Begossi A. Temporal Stability in Fishing Spots: Conservation and Co- Management in Brazilian Artisanal Coastal Fisheries. Ecol Soc. 2006;11. doi:10.5751/ES-01380-110105 22\. Begossi A. Cultural and ecological resilience among caiçaras of the Atlantic Forest coast and caboclos of the Amazon. Linking social and ecological systems for resilience and sustainability The Beijer International Institute of Ecological Economics, Stockholm. 1998; 129–157. 23\. Diegues ACS. Pescadores, camponeses e trabalhadores do mar. Editora Ática; 1983. 24\. Diegues AC. A pesca construindo sociedades. NUPAUB-USP; 2004. 10.1371/journal.pone.0284024.r003 Decision Letter 1 MacKenzie Brian R. Academic Editor 2023 Brian R. MacKenzie This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 23 Mar 2023 180 years of marine animal diversity as perceived by public media in southern Brazil PONE-D-22-32178R1 Dear Dr. Herbst, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Brian R. MacKenzie, Ph. D. Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0284024.r004 Acceptance letter MacKenzie Brian R. Academic Editor 2023 Brian R. MacKenzie This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 21 Jun 2023 PONE-D-22-32178R1 180 years of marine animal diversity as perceived by public media in southern Brazil Dear Dr. Herbst: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Prof. Dr. Brian R. MacKenzie Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction Since their identification eukaryotic Transposable Elements (TEs) developed into an important field of genetic and genomics investigation. Nevertheless, recent advances in sequencing technologies offer a unique and somewhat unappreciated opportunity to increase our understanding of several aspects of the TEs biology, e.g. structure, evolution and regulation (see for a review). Besides their detrimental role as an endogenous source of mutations, TEs transposition and accumulation serve as an evolutionary substrate for genes and genomes evolution. Indeed, inactive TEs play a significant role in macroevolution, contributing in chromosomal rearrangements or being recruited to evolve novel functions. In addition, defective elements and ancient relics of autonomous copies are quite informative to trace the evolution of single TEs families. The Tc1-mariner constitutes one out of 17 super-families of the class II transposons. Tc1-mariner elements are widespread among all life kingdoms and their diffusion is mainly due to their simple transposition mechanism, and their proposed ability for cross-species diffusion by horizontal transfer mechanisms. The Bari family of transposons belongs to the Tc1-mariner superfamily and Bari-like elements have been identified in several species of the Drosophila genus. Based on their structural and evolutionary features Bari-like elements fall into three distinct subfamilies, Bari1, Bari2 and Bari3 after the founder elements discovered in D. melanogaster, D. erecta and D. mojavensis respectively. The Bari1 and Bari3 subfamilies contain autonomous elements able to perform transposition due to the presence of Terminal Inverted Repeats (TIRs) sequences surrounding a central sequence encoding a functional transposase. On the other hand, Bari2-type elements are non-autonomous due to the accumulation of inactivating mutations. Besides the functional criterion, Bari elements can be classified standing to the structural differences of the terminal sequences. Bari1-type elements harbor short TIRs, usually 28 nucleotides long, called Short Inverted Repeats (SIR) while Bari2 and Bari3 possess Long terminal Inverted Repeats (LIR), roughly 250 nucleotides long. Interestingly, non-autonomous Bari1 elements possessing LIRs have been described in D. ananassae, a species of the melanogaster group, suggesting that the ancestor of the Bari1 subfamily had LIRs that were lost during Bari1 evolution. Regardless their length, both SIR- and LIR- containing elements share three highly similar 18 nucleotides long domains, called Direct Repeats (DRs). DRs are found within the 250 terminal nucleotides at both ends and are responsible for the transposon-transposase interaction, a crucial step in the transposition event. Such interaction has been previously demonstrated for Bari1 and Bari3. An interesting genomic feature of the Bari1 subfamily is the presence of a heterochromatic array in the D. melanogaster species. It was known from previous studies that at least two distinct clusters exist in the reference genome of D. melanogaster. The first cluster maps to the h39 region, a cytological band adjacent to the centromere of the second chromosome of D. melanogaster, the same cytological band where the Responder satellite maps. It contains several tens of Bari1 copies and is interrupted by a MAX insertion. The second cluster contains few copies and had uncertain heterochromatic map location. Heterochromatic Bari1-Bari1 junctions (Right TIR-Left TIR) are characterized by the deletion of the first two nucleotides (CA) in the left terminus of each element. The terminal copies of both clusters have been also previously characterized thus helping in the reconstruction of their origin. Genome-scale comparison studies are an important tool for both understanding the forces that shaped modern forms of transposable elements, and highlight non- mendelian modes of transposons’ transmission. Early investigations on Bari-like transposable elements at the genomic level were essentially performed in 12 Drosophila sequenced genomes available at that time. The availability of 11 additional Drosophila genome draft sequences together with the availability of new assembly releases and data from re-sequencing projects prompted us to investigate the genomic distribution of the Bari transposon family in a wider pool of studied and unexplored genomes. In this study, we performed an extensive annotation of the Bari-like elements in 23 Drosophila genomes analyzed, and uncovered additional structural variability as compared to previous analyses. In addition, we disclosed the presence of MITE-like forms of the Bari transposons in 7 Drosophila species including, interestingly, species apparently devoid of full-length Bari elements. Analyses of the integration site of Bari elements revealed a preference for AT-rich sequences in which the TA target is duplicated upon integration. Furthermore, annotation of unrelated TEs insertions in the proximity of Bari elements revealed significant co-occurrence of other Tc1-mariner elements while class I TEs avoid these regions. Finally, we propose that incongruences revealed by our phylogenetic analyses could be explained by horizontal transfer events. Taken together our results significantly increase our understanding of the evolution of Bari elements. # Materials and Methods ## Bari transposon search strategy and sequence analyses Searches for Bari homologous elements were carried out in Drosophila species listed in. A BLAST strategy was applied to identify of Bari-like elements at the NCBI WGS (Whole Genome Shotgun) database (<http://www.ncbi.nlm.nih.gov/genbank/wgs/>) or at the FlyBase database. Query sequences for tBLASTn searches were either the Bari1 or Bari3 transposases (GenBank CAA47913 and conceptual translation of GenBank accession CH933806 position 6274049–6275068 respectively). Queries for BLASTn searches were performed using either the whole DNA sequence or the 250 terminal nucleotides containing the three DR sequences of Bari1 or Bari3. All BLAST analyses were performed using the default parameters. Subject sequences from tBLASTn searches with E \< 10<sup>−120</sup> and similarity greater than 75% over the whole transposase length (339 amino acids), were further analyzed. The threshold E-value was set to higher values in BLASTn searches aimed at MITEs identification (E \>10<sup>−20</sup>). Two criteria were used to identify full-length Bari-like elements; 1) the detection of a high-scoring subject sequence by tBLASTn search, using either the Bari1 or the Bari3 transposase protein as query sequence; 2) the presence of homologous DRs in the terminal sequences, surrounding the coding region of the elements. Terminal inverted repeats and homologous DRs in the transposon’s termini were identified by a combined analysis using the Dot Plot matrix analyses implemented in the DNA Strider package software, the Einverted software (<http://embossbioinformaticsnl/cgi-bin/emboss/einverted>) and by multiple sequence alignment of terminal ends of Bari-like elements as previously described. Elements’ names were assigned according to the binary nomenclature used in Repbase, consisting of the subfamily identifier (either Bari1, Bari2 or Bari3) followed by the species identifier (i.e. the first letter of the genus and three letters of the species name to avoid ambiguity for some species). Similarly, MITE’s elements names look like Bari_Dxyz_MITE-# (where “xyz” is a three-letters species identifier) followed by a number (#) to distinguish different MITEs subfamilies, where necessary. Once a novel full-length element was identified, it was used to identify other full-length copies as well as truncated elements, by BLASTn analyses against species-specific WGS database. Each copy was then carefully annotated. Additional BLASTn searches were performed using as query the mel-ER (consensus), the sim-ER and the sec-ER elements (here named Bari2_Dmel, Bari2_Dsim and Bari2_Dsec in the text) described in, three Bari2-type elements in the genomes of D. melanogaster, D. simulans, and D. sechellia respectively. Redundancy can be excluded for all elements except for split elements, i.e. sequences overlapping either the beginning or the end of a genomic scaffold in which they have been identified. These sequences have been annotated as “partial” elements in and were counted as elements. No further analysis was performed using partial elements. Redundant elements were removed on the basis of the flanking sequences comparison. Global alignments were performed using LALIGN (<http://wwwchembnetorg/software/LALIGN_formhtml>). The Open Reading Frames (ORFs) were inferred using ORF Finder (<http://wwwncbinlmnihgov/gorf/gorfhtml>) or using a custom script (available upon request) written in Biopython (version 1.63). The similarity with previously reported sequences was established using CENSOR at the RepBase database. dS analysis was performed using the Nei and Gojobori method implemented in MEGA5. Fisher’s exact test (one-tailed) was conducted using 2 x 2 tables to verify if transposon dS values were statistically lower than those presented by the host genes. The number of base substitutions per site between sequence pairs was calculated using MEGA 5.0. Analyses were performed using the Kimura 2-parameter model. Rate variation among sites was modeled with a gamma distribution (shape parameter = 1). All ambiguous positions were removed for each sequence pair. The whole transposon sequence alignment was used to perform Plotcon analyses (<http://emboss.bioinformatics.nl/cgi-bin/emboss/help/plotcon>) by moving a window of 100 nucleotides along the aligned sequences. Sequences marked as “partial sequence” in were excluded from p-distance and Plotcon analyses. Data analysis was carried out with the R System (R version 3.1.0) ([https://www.r-project.org](https://www.r-project.org/)). The non-parametric Kruskal-Wallis test and Tukey test were used to assess significance for the observed differences, considered significant at p values\<0.05. Hartigan dip test was carried out to assess unimodality/multimodality of the observed similarity values distributions. ## Insertion sites analyses WebLogo analyses were performed in those species where it was possible to retrieve at least four flanking sequences of equally oriented Bari elements. A larger scale insertion site analysis was also performed. Regions of 2500 bp flanking downstream and upstream the insertion sites of Bari elements, when available, were analyzed with RepeatMasker-open-4.0.5 ([http://www.repeatmasker.org](http://www.repeatmasker.org/)) using the RepBase20.02 dataset. The same analysis was performed, as control, on 1000, non- overlapping, random sequences, 2,5 kb in length, irrespective of their gene content selected from species containing at least four Bari insertions, taking care to avoid regions with internal sequencing gaps. The distance of each masked transposon from the sequence origin (fixed at the first base flanking Bari elements or to the first nucleotide of the randomly-selected scaffold) was calculated and used to create a positions dataset. The detected distances were then classified in 5 distance ranges from the sequence origin, considering a 500 bp cumulative increment (0–500; 0–1000; 0–1500; 0–2000 and 0–2500). Poisson distribution has been applied to the data, given that the repeat frequency in 1000 random regions was considered as the expected value, and the frequency in the Bari flanking regions as the observed value to test. Sequence accessions and coordinates of the control dataset are reported in. Relative probability was calculated to identify a significant association between Bari and the selected repeat families. ## Multiple alignments and phylogenetic analysis Multiple sequence alignments were performed using ClustalW2 with default parameters. Nucleotide sequences relative either to the Coding Sequence (CDS) or to the terminal sequences were used in the analyses of Bari elements. The Bari1_Dsim and Bari1_Dsec elements were not included in the multiple alignment of terminal sequences and transposase sequences because they are identical in sequence to the Bari1_Dmel element. The D. yakuba element was excluded from the alignment of TIR sequences because it has been re-classified as MITE, and thus included in the MITE elements’ analysis. The BioEdit software was used for alignment editing and visualization. Alignment slices were obtained using the Alignment Slicer tool (<http://www.hiv.lanl.gov/content/sequence/Slice_Align/>). Species phylogeny was based on the alignment of nine concatenated orthologous CDSs encoding subunits of the V-ATPase complex, for which orthologous genes could be identified in the 23 species analyzed. FlyBase annotation symbols, GenBank accession of mRNA and length of D. melanogaster CDSs used in this analysis are the following: CG3762 NM_143747 (1845 bp), CG2934 NM_130724 (1054 bp), CG7071 NM_142770 (909 bp), CG34131 NM_001043279 (483 bp), CG12403 NM_001273497 (1845 bp), CG1088 NM_169073(681 bp), CG7007 NM_143753 (639 bp), CG3161 NM_057453 (480 bp), CG6213 NM_058089 (354 bp). The jModelTest program v 2.1.7 was used to select the simplest evolutionary model that fitted adequately the sequence data. A Bayesian Markov chain Monte Carlo (MCMC) method implemented in BEAST package 1.8.0 was used for Bayesian analysis. The MCMC chains were run for at least 50 million generations and, sampled every 5.000 steps. Convergence was assessed on the basis of the effective sampling size after a10% burn-in. Phylogenetic trees were visualized and edited using FigTree 1.4.2 software (<http://treebioedacuk/software/figtree/>). Paris and S elements were used as outgroups. # Results ## Bari elements identification and their structural diversity in 23 sequenced genomes We applied a BLAST-based search to identify insertions of Bari-like elements in 23 Drosophila genomes. The evolutionary time of the species analyzed in this work spans roughly 40 million year. We used the full-length Bari1 and Bari3 transposase as query in tBLASTn analyses, and the whole transposon DNA sequences in BLASTn analyses against the genome shotgun database or reference genomes. Using this approach, we identified 401 Bari-like elements comprising full- length copies, truncated elements and MITEs. The actual number of Bari insertions annotated in the species analyzed could be slightly inflated because we have also annotated partial sequences (53 out of 249 sequences), i.e. split sequences overlapping either the beginning or the end of a genomic scaffold in which they have been identified, for which we cannot exclude redundancy. Redundancy can be instead excluded in the remaining sequences for which we compared the flanking sequences to establish uniqueness in the dataset. Inflation is particularly relevant in D. pseudoobscura (22/33 total insertions) and in D. miranda (10/16 total insertions). However, inflation could be compensated by the possible underestimation of the Bari copy number due to the draft status of the genome assembly in some species. New Bari-related sequences were annotated in 11 species (namely D. biarmipes, D. bipectinata, D. rhopaloa, D. takahashii, D. kikkawai, D. eugracilis, D. ficusphila, D elegans, D. suzukii, D. miranda, and D. albomicans). For clarity purposes, data concerning full-length and truncated elements are presented separately from data concerning MITEs. The features of representative elements (i.e. the top scoring and the most complete subjects after BLAST search) are summarized in, while information concerning all the annotated insertions can be found in and Tables. A snapshot of the distribution of Bari subfamilies and the presence of potentially active elements in the species analyzed is reported in. Details relative to individual insertions identified in this study are reported in and Tables. The presence of Bari-like elements in every genome analyzed, with the exception of D. grimshawi, confirms previous evidence on the widespread diffusion of this transposon family in the Drosophila genus. As can be observed in, distinct Bari subfamilies have colonized specific genomes, with little subfamilies overlaps, as can be observed in the species of the melanogaster complex (D. melanogaster, D. simulans, D. sechellia) for the Bari1 and Bari2 subfamilies. Drosophila melanogaster contains the highest number of Bari-like element copies (84 elements annotated) mainly clustered in a specific heterochromatic region, (best described in the next paragraph), followed by D. pseudoobscura (33 insertions detected), while D. albomicans (1 element) virtually lacks Bari elements and D. grimshawii is devoid of Bari elements. A preliminary analysis of all the elements identified suggests that Bari-like elements have variable length ranging from 17 bp (D. melanogaster sheet, element \#25) to 4353 bp (D. ananassae sheet, element \#12), due to the existence of canonical copies, truncated copies (i.e. transposon chunks and elements overlapping the extremities of the sequence contigs) or other mobile elements nested within Bari elements. The median length of all annotated Bari elements is 696 bp (IQR = 1635 bp; max. length = 4353; min. length = 17). In more details, the median length of full-length and truncated elements is 1704 bp (IQR = 593; max. length = 4353; min. length = 17), while the median length of MITEs is 82 bp (IQR = 394; max. length = 939; min. length = 69). Considering only those elements containing two terminal sequences (each one containing three DRs) framing an intervening sequence, irrespective of its coding potential (134 elements) Bari-like elements look more homogeneous (median length = 1723 bp; IQR =; min. length = 746 bp in D. pseudoobscura; max length = 4353 bp in D. ananassae). The TA dinucleotide TSD, a feature of the Tc1-mariner superfamily, was identified in 60 out of these 134 sequences. Twenty-four copies of Bari, distributed in seven species (D. melanogaster, D. simulans, D. rhopaloa, D. persimilis, D. pseudoobscura, D. miranda and D. mojavensis) can be considered as autonomous elements due to the presence of two terminal sequences (either with LIR or SIR structure) bracketing a CDS encoding homologous Bari transposase. The presence of three conserved DR sequences within the 250 terminal nucleotides bracketing the transposase gene, typical of the Bari family, suggest that all the new elements identified are bona fide Bari-like elements. We performed a comparative analysis to estimate the sequence variability of Bari elements and the deterioration profile both at the inter-species and intra-species level. We calculated the pairwise distance (p-distance) relative to the left and right terminal sequences, comprising the DRs, and to the transposase coding sequence. At the inter-species level, no significant differences can be observed among the CDS and the terminal sequences in the three Bari subfamilies, suggesting the uniform distribution of point mutations in the three regions analyzed. However, some differences were observed by analyzing the same elements at the intra- species level, where the p-distance analysis was also coupled with the determination of the degradation level of the Bari elements in each species. As an example, three representative results obtained are shown in (see for complete results). In two species (D. willistoni and D. simulans) the p-distances at the terminal sequences are significantly different if compared to the central region, suggesting an increased mutational load at the terminal sequence level. In four species one of the TIRs has a significantly higher distance if compared to the CDS (D. ananassae D. erecta, D. pseudoobscura, D. miranda). In the remaining three species (D. melanogaster, D. persimilis, D. mojavensis) no significant difference was observed, suggesting uniform mutation pattern. The respective degradation profiles show that truncated elements are generated in each of the analyzed species by deletion and/or unrelated TEs nested insertions, occurring without any obvious positional preference pattern. A description of Bari elements in those genomes containing few copies, and thus not analyzed as described above, is given below. The genome of D. takahashii contains two inactive elements due to insertions and deletions. The Bari1_Dtak element was reconstructed removing an unreported MITE insertion (not shown). Bari1_Dtak includes a 912 bp-long ORF encoding a protein that shares 81% and 72% similarity with Bari1 and Bari3 transposase respectively framed by two LIRs containing 3 DRs. The Bari1\_D.rho element coding sequence encodes a protein 84% similar to the Bari1 transposase, framed by LIR-type terminal sequences. Bari1-like elements with a similar structure have been identified in the genomes of D. bipectinata, D. biarmipes, D. suzukii, D. takahashii and previously in D. ananassae but all them are inactive, while the Bari1\_D.rho can be considered as a putatively active element (although the presence of the very last two nucleotides in the right TIR has not been confirmed, see also). The D. miranda Bari element encodes a Bari3-like transposase (73% similarity, see) and shows divergent terminal sequences (71.8% similarity). A single putatively autonomous element can be found in the current genome assembly of D. miranda along with several defective elements carrying internal deletions. In D. albomicans a 109 bp long fragment can be only identified on the basis of its nucleotide similarity (75%) with the Bari2_Dere element (Bari2). This sequence, arbitrarily assumed to be a left terminus, contains the left middle (Lm) and left inner (Li) DRs, that can be easily recognized with multiple alignment analysis and allow its classification in the Bari family, and specifically its assignment to the Bari2 subfamily (see last paragraph of the section). The putative ancestral sequence of Bari elements of D. suzukii and D. bipectinata element were reconstructed from truncated elements (Bari1_Dsuz element reconstructed from AWUT01004624 and AWUT01009036; Bari1_Dbip element reconstructed from contigs AFFE02004473 and AFFE02005005. See for sequence coordinates). Both the Bari1_Dbip and the Bari1_Dsuz elements are members of the Bari1 subfamily and possess LIR-type terminal sequences. In D. eugracilis a low-scoring sequence contained a truncated copy of a Bari element with single terminal sequences with three DRs ( and). While in D. grimshawi we were not able to detected elements related to the Bari family, in D. ficusphila and D. elegans we have identified Bari-related MITEs, described in a specific paragraph. In D. grimshawi the absence of Bari-like elements is possibly due to the actual absence of the transposon or early draft status of the genome sequencing. ## New insights into the Bari1 heterochromatic cluster in D. melanogaster We identified 61 independent heterochromatic Bari1 copies in heterochromatic tandem repeat configuration. The contig JSAE01000772 contains the LTR- retrotransposon MAX embedded in 20 Bari1 copies repeated in tandem. This observation matches with previous findings concerning the discontinuity in the Bari1 repeat. Two additional contigs identified in this study (ac. nos. JSAE01000400 and JSAE01000412) contain respectively 18 and 17 Bari1 copies in tandem repeat configuration. These contigs map to the borders of the Bari1 cluster and contain terminal copies matching also identified in previous studies. We can orient of these contigs with respect to the centromere of the second chromosome considering that: 1) the Responder repeats, representing the major satellite DNA in this heterochromatic region are abundant in the JSAE01000412 contig (388 copies); 2) Responder maps proximally to the centromere with respect to the Bari repeat as demonstrated by previous studies. It can be concluded that contigs JSAE01000400 and JSAE01000412 map, respectively, distally and proximally to the centromere of the second chromosome. Scaffold JSAE01000412 contains two adjacent truncated Bari1 copies (position 169962–173070, corresponding to elements \#31 and \#32,). We hypothesized that either inter-monomer recombination or unequal crossing over events involving adjacent or non-adjacent copies of the cluster and occurring at short homologous sequences, as shown in (panel A). Comparison of these two adjacent elements to two canonical adjacent Bari1 heterochromatic copies (i.e. containing two full- length Bari1 elements carrying the deletion of the fist two nucleotides) revealed a 340 bp deletion flanked by short stretches of identical sequences that could mediate recombination leading to the observed deletion (panel B). The last Bari1 copy of the cluster toward the centromere of chromosome 2 represents an additional defective Bari1 element. This truncated copy retains only 16 terminal nucleotides of the right TIR, previously described in. A possible origin of the terminal copies of the h39 cluster, linked to an aberrant activity of the transposase has been previously proposed. The last worth-noting scaffold (JSAE01000184) contains six Bari1 copies arranged in a tandem repeat configuration, embedded in more than 200 kb of repetitive DNA. RepeatMasker analyses revealed the presence of transposable elements belonging to the R1 and the R2 class, typically inserted into rDNA sequences as well as few unmasked sequences related to rDNA genes (at least 4 copies). As previously suggested, this short Bari1 cluster could either map in a distinct h39 sub-region or could be located on the X chromosome, in the region of the nucleolar organizer region. Based on the recent findings that no rDNA repeat nor type 1 rDNA insertions exist outside the nucleolar organizer regions D. melanogaster, it could be concluded that the small Bari1 cluster may be located in the heterochromatin of the X or Y chromosomes, where the nucleolar organizers map. ## MITE elements related to Bari transposons The tBLASTn strategy described above fails in detecting Bari copies carrying extensive deletions of the transposase gene, as can be observed in MITEs. Dias and Carareto recently described two MITE elements related to Bari1 (msechBari1 and msechBari2, hereafter Bari_Dsec_MITE1 and Bari_Dsec_MITE2) present in multiple copies in the genome of D. sechellia, therefore, the existence of Bari-related MITEs in other species cannot be excluded. To test this possibility we performed a BLASTn-based search of Bari-like sequences in the species in which Bari-like elements were not found on the first attempt (namely D. ficusphila, D. elegans and D. grimshawi). The genomes of D. ficusphila and D. elegans harbor 36 and 4 Bari-related MITEs respectively. We, therefore, searched for Bari-related MITEs in all the available sequenced Drosophila genomes. We identified and annotated 152 copies of Bari-related MITEs in nine genomes, including D. sechellia. Names were assigned according to the rule given in the Materials and Methods section. In the genomes of D. simulans, we identified a Bari-related MITE identical in sequence to Bari_Dsec_MITE1 of D. sechellia. However, while Bari_Dsim_MITE is present as single copy element in the genome of D. simulans, more than twenty Bari_Dsec_MITE1 copies are found in D. sechellia ( and this study). In D. melanogaster, we identified two Bari-related MITE families based on their length. The Bari_Dmel_MITE1 family is an 116 bp long element with nearly identical, 46 bp long, terminal inverted repeats. The Bari_Dmel_MITE2 family contains longer elements (692 bp) and has 50 bp long terminal inverted repeats. D. suzukii harbors a single family of Bari-related MITEs, nearly 500 bp long, sharing high sequence similarity if compared to each other. A retrospective analysis of the single copy element previously identified in D. yakuba led us to conclude that it can be classified as a Bari-related MITE. A peculiar feature of this MITE consists in the presence of two DR sequences in the left terminus while all the other Bari-related MITEs display a single DR in each terminus. An overall comparison of the terminal sequences of the Bari-related MITEs with the terminal sequences of Bari1_Dmel, Bari2_Dere and Bari3_Dmoj, representative of the Bari1, Bari2 and Bari3 subfamilies respectively, is shown in. All the terminal sequences of Bari-related MITEs are very similar to each other and share similarity with the corresponding full-length elements. Bari-MITE sequences contain an intervening sequence between the two terminal that can be either short or long but with poor sequence similarity if compared to reference elements, suggesting that complex mechanisms might cause their origin from master elements. Although the sequence similarity at the terminal sequences level suggests that these MITEs belong to the Bari family, no obvious ancestor can be inferred using this approach. MITEs with identical sequence have been identified in different species (i.e. Bari_Dsec_MITE1 and Bari_Dsim_MITE) while in other cases MITEs from different species have a very similar sequence (Bari_Drho_MITE and Bari_Dtak_MITE share 75 out of 78 nucleotides). To investigate the mechanisms of MITEs expansion, we postulated that the analyzed MITEs derived from the same ancestor and performed pairwise nucleotide diversities analysis of these elements. Rogers and Harpending reported that episodes of population growth leave a characteristic signature in the distribution of nucleotide differences between pairs of individuals, and this concept was used to describe the mode of amplification of MITEs in Oryza sativa. We adapted this type of analysis to the Bari-derived MITEs and statistically evaluated whether the wave-like form of the histograms could fit unimodal distribution. Our data do not allow rejection of unimodality in three species (D. sechellia, D. ficusphila, D. rhopaloa; Hartigan dip test for unimodality: Drho p = 0,598; Dfic, p = 0,2135; Dsec, p = 0,4119) suggesting single burst of transposition in these genomes. In the genomes of D. suzukii and D. takahashii, a multimodal distribution of the pairwise differences (Hartigan dip test for unimodality: Dsuz, p = 0,01805; Dtak, p = 0, 01374) suggests that multiple rounds of MITEs transposition occurred. ## Bari-like elements and their target site selection preferences To gain insight into the choice of integration target sites for Bari-like elements, we performed flanking sequences analyses. It is well established that Tc1-mariner elements integrate themselves into the TA target, which is duplicated upon element insertion. Multiple alignment of 15 bp long sequences encompassing the Bari elements insertion sites we identified, showed that this is also true for Bari elements, as also highlighted by the WebLogo analyses. Furthermore, the Bari transposase targets AT-rich sequences irrespective of the Bari subfamily and of the host species, as can be observed in. We further investigated the target site preferences of Bari-like elements on a larger sequence scale by analyzing the sequence context in which Bari transposons insert themselves. The presence of transposable elements belonging to three LTR-retrotransposons superfamilies (Copia, Gypsy and PAO), three non-LTR retrotransposons superfamilies (LOA, I and Jockey), the Tc1-mariner superfamily and to the Helitron superfamily, was annotated within a 2,5 kb long sequence flanking upstream and/or downstream the Bari elements in each species analyzed. To assess whether the observed distribution of transposable elements within the neighborhoods of Bari, significantly differ from their genomic distribution in the respective genomes, we collected 1000, non-overlapping, random sequences, of the same length (2,5 kb), irrespective of their gene content, and recorded the occurrence of tested TEs. This analysis was performed in genomes containing a minimum of 4 Bari insertions (an arbitrarily chosen threshold), resulting in 15 investigated genomes. In these genomes, the Bari transposon insertions occur in genomic loci depleted in Copia, LOA and I elements and enriched in Tc1-mariner elements. In 5 genomes we also found a significant tight association (within 500 bp from the Bari elements) with elements of the Jockey and PAO families; however in 7 genomes, taking into account the same sequence range, Bari elements insertions occur in genomic regions significantly depleted in Jockey and PAO elements. Gypsy-like retrotransposons do not display significant enrichment or depletion within the 2,5 kb range considered. Helitron-like elements also significantly co-occur within 2,5 kb from Bari elements. ## Evolutionary analyses of Bari transposons We multi-aligned the transposase-coding sequences from 16 representative Bari copies and built a phylogenetic tree using Bayesian inference methods. The phylogenetic tree created recovered three clades with significant statistical support (left tree) and corresponding to the previously described Bari1, Bari2 and Bari3 clades. This analysis placed the newly identified Bari elements of D. suzukii, D. takahashii, D. rhopaloa, D. bipectinata, D. biarmipes, D. kikkawai and of D. eugracilis in the Bari1 clade while the element of D. miranda cluster together with other Bari3 elements. We also performed phylogenetic analyses based on the terminal sequences of Bari-like elements. Roughly 250 nucleotides containing the three DRs from the left terminal sequence of Bari-like elements were analyzed in a Bayesian tree (right tree). The phylogenetic tree obtained displays a topology similar to that obtained for the transposase gene-based tree, with three distinct and well- supported clades consisting of the three Bari subfamilies. A careful inspection of the transposase-based phylogenetic tree revealed an inconsistency in a well-supported Bari1 sub-clade if compared to the phylogeny of species. The Bari1_Dbia and Bari1_Dbip elements are indeed very close to each other in the transposase tree, while their respective host species, D. biarmipes and D. bipectinata, are more distantly related, being members of the suzukii and ananassae subgroups respectively. These elements share high sequence similarity and are part of a highly supported clade in the tree. In order to discern if the incongruences observed in the phylogeny of Bari-like elements and host species were due to vertical transmission or horizontal transfer, the divergence at synonymous sites (dS), taken as a measure of neutral evolution in the absence of a strong codon usage bias, was compared between TEs and Adh or Gapdh1 genes of the two species. Fisher exact test-based comparisons were performed to verify whether the dS of the Bari-like elements was significantly lower than the dS of nuclear genes. The dS of Bari elements (0,0603) was significantly different (p\<0,005) if compared to dS of Adh (0,3281) or Gapdh1 (0,4006) genes in these species, supporting the hypothesis of horizontal transfer of Bari-like elements between these species. # Discussion This work aimed to identify Bari transposon insertions in the sequenced Drosophila genomes, a step forward in the annotation of the complete set of TEs insertions in these genomes. We have annotated 401 copies ( and Tables) of Bari-like transposon in the genomes of 22 Drosophila species out of 23 available sequenced genomes. Although this data may seem as an overestimation of the actual number of Bari elements in the analyzed species, we expect that elements being missed due to the early draft status of some genome assembly (especially in the heterochromatin) would somewhat compensate any number discrepancies. Importantly, Southern blot experiments testing the genomic distribution of Bari elements in some of the genomes analyzed in this work are in substantial agreement with our sequence analyses, suggesting that Bari is present in low copy number in these genomes and that D. melanogaster is the only species (among those studied) containing heterochromatic clusters. This work provided a number of novel insights into the biology of the Bari transposon’s family: 1. high sequence variability of the terminal sequences coupled with high conservation of the DRs and wide diffusion in the Drosophila species’ genomes. 2. MITE-like forms of this transposon, arising occasionally during the species evolution. 3. strong target site preference for AT-rich DNA regions, which is also the case for other Tc1-mariner elements, that are instead avoided by class I transposable elements. 4. occasional horizontal transfer events involving Bari-like elements. ## Bari structural diversity Transposons structural diversity is an obvious evolutionary phenomenon if large families or super-families of mobile elements are taken into account. However, the emergence of structural variants can also be detected in small transposons’ families, such as the Bari family. The dynamic processes leading to TEs replication, amplification, degradation and elimination from a given genome are difficult to be elucidated. However the graphic representation of the deterioration pattern of TEs proposed by Fernandez Medina et al., can be used to shed light into this task. Our results suggest that Bari elements have undergone deterioration in different ways depending on the host genome. Point mutations and indels probably weighted differently in producing non-functional Bari copies during evolution, because mutations in the terminal sequences are more tolerated if compared to the coding sequence. Despite some degree of diversification observed at the terminal sequence level, a strong conservation of the DR sequences in all elements’ subfamilies from all species, including those containing only MITEs suggests that these sequences may have acquired a potential role during the evolution. The observed differences in the p-values of the terminal sequences can be ascribed either to the transposition activity of Bari elements or to the presence of functional elements in the terminal sequences. For example D. mojavensis contains 8 potentially active and 7 inactive Bari3-like elements and a median p-distance value of 0 for both terminal sequences, while D. miranda contains 1 potentially active and 4 inactive Bari3 elements and a median p-distance values of 0,171 and 0,06 for left and right terminal sequences respectively. Similarly D. melanogaster contains 4 potentially active and 1 inactive Bari1-type elements with a median p-distance value of 0 for both terminal sequences while D. simulans contains two potentially active elements with a median p-distance value of 0,01 for both terminal sequences. In conclusion, it can be speculated that species displaying lower p-distance values host active Bari elements, which produce identical copies upon transposition thus lowering the p-distance. By contrast, species with higher p-distance values contain many dead or almost dead elements, which rapidly accumulate more mutations and become divergent in sequence. However, we have evidence of Bari transposition only in D. melanogaster and in D. mojavensis, needing additional studies in order to establish if Bari is active in non-model species. The presence of functional elements could also explain the differences observed in p-distance values between the terminal sequences if compared to the CDS. While both the left and right termini contain DRs able to bind the transposase, the left terminus might carry the promoter or part of it, and the right terminus might contain important signals for the termination of transcription. Interestingly in the context of the terminal sequences the DRs appear strongly conserved in sequence even in species lacking active Bari elements. Based on this it could be speculated that DRs have acquired a new function (e.g. production of siRNA, binding of different protein partner etc) in the genomes where Bari elements have been completely inactivated. There is evidence from previous studies suggesting a well-defined structural vs phylogenetic relationship among Bari-like elements. While elements of the inactive Bari2 subfamily contain LIR-type terminal sequences, elements of the Bari1 and Bari3 subfamilies usually harbor SIR and LIR respectively. The non-autonomous Bari1 elements identified in D. ananassae represent an exception as they harbor LIR. In this paper, we show the existence in D. rhopaloa of potentially active Bari1 elements containing identical LIR, which is a previously unreported feature in the Bari1 subfamily that increases the diversity among Bari elements. In addition we identified Bari1-type elements with LIRs in D. bipectinata, D. biarmipes, D. suzukii and D. takahashii, but in these species all the Bari elements are inactive. To date, the Bari1_Drho represents the only potentially active Bari-1 type element with LIRs. It was previously suggested that the identical LIRs might represent the first stage of the evolution of the terminal repeats of Bari-like elements. The LIR structure may subsequently evolve into SIR structures as a consequence of the intrinsic instability associated with the long terminal repeats structure. This hypothesis was best fitting with the Bari3 subfamily, in which the Bari3_Dmoj represents a “young” Bari element having perfect LIR and the Bari3_Dper and Bari3_Dpse elements, representing older elements that are going to lose the similarity between their TIRs. The same hypothesis can be now formulated for the Bari1 subfamily, where SIR- and LIR-containing elements were previously identified. The two TIRs of the Bari_Dana element indeed share 94% similarity and the two TIRs of the Bari1_Drho element reported here are identical to each other. In this view the Bari1_Drho element represents a young Bari1-type element, still awaiting the divergence of its terminal sequences. This conclusion raises further questions concerning the mechanisms that generate Bari1- and Bari3-type elements with perfectly matching terminal sequences, which are still to be identified. The powerful sequencing technologies developed in the last years facilitate the molecular determination of repeat-rich genomic such as the h39 heterochromatic region of D. melanogaster which hosts roughly 80 clustered copies of Bari1 and the Responder satellite. Our detailed analysis allowed orienting the Bari1 cluster with respect to the centromere of the second chromosome, thanks to the presence of the Responder repeats in one of the sequences flanking the Bari cluster. It was previously observed that the Rsp satellite displayed an extreme quantitative and structural variability while Bari1 cluster showed remarkable homogeneity. The occurrence of recombination events between clustered copies of Bari1, as inferred from our data, suggests that this transposon cluster could be subjected to expansions or contractions as observed for other complex DNA repeats over evolutionary time. Our sequence analyses of heterochromatic copies of Bari1 in D. melanogaster support the hypothesis that an additional Bari1 cluster might exist outside the h39 region of the mitotic chromosomes, specifically on the X or Y chromosome, since it is associated with DNA sequences specific of the Nucleolar Organizer Region (rDNA and R1 element insertions). The presence of such kind of clusters in a single Drosophila species (i.e. D. melanogaster) is peculiar and it could be speculated that the reiterated formation of clusters, might depend on species-specific host factors contributing to an error-prone activity of the transposase. In addition resolving the structure of heterochromatic genomic blocks, rich in transposable elements would help understanding important regulatory loci, as reported for the flamenco locus in D. melanogaster. ## MITEs related to Bari elements Miniature Inverted-repeat Transposable Elements (MITEs) are non-autonomous, short repeats that mobilize within the host genome even without the potential to encode the key proteins responsible for their mobilization (i.e. the transposase). MITEs are generally smaller than 600 bp with few exceptions, have conserved TIRs, a target site preference, do not display coding potential, are AT-rich and amplified within the host genome. In general, they are supposed to originate by deletions internal to autonomous elements, which leave untouched the TIRs and, sometimes, just portions of the transposase. This origin supports the hypothesis of their mobilization in trans by a transposase encoded by a full-length element. In addition to the previously described Bari-related MITEs in the genome of D. sechellia, we identified this element type in eight additional Drosophila species. These novel Bari-related MITE sequences match the definition of MITEs; however some of them (Bari_Dmel_MITE2, Bari_Dsim_MITE, Bari_Dyak_MITE) lack genomic amplification while others show only modest amplification (Bari_Dmel_MITE1, Bari_Dtak_MITE, Bari_Dele_MITE). By contrast Bari_Dsec_MITE1, Bari_Dsec_MITE2, Bari_Dfic_MITE and Bari_Drho_MITE are quite abundant in the respective genomes. It is possible that MITEs in D. melanogaster, D. simulans and D. takahashi could be in a very initial stage of their amplification. Alternatively, they might represent the product of abortive amplification. Wallau et al., recently described 27 independent sub-lineages of mariner-derived MITEs in 20 Drosophila species, which have internal sequences and TIRs similar to the sequences of the full-length copies and a typical size of 900–1000 bp with few exceptions. Bari-related MITEs are shorter than the mariner-related and possess internal regions not easily comparable with autonomous elements, reflecting a more complex rearrangement differing from abortive gap repair. Furthermore, Bari-derived MITEs can be short (less than 120 bp) or long (greater than 120 bp and less than 700 bp) in sequence, and the intervening sequence between the TIRs, where present, is apparently unrelated to Bari elements. Contrarily to the mariner MITEs, Bari-related MITEs apparently originated by deletions in the same point, with the possible exception of the Bari_Dyak_MITE. It can be hypothesized that Bari-related MITEs originated through internal deletion of a master element, with the deletion breakpoints occurring between the Lo-Lm and between Rm-Ro direct repeats, which has been possibly followed either by direct junction of the broken extremities or by the addition of unrelated sequences. Finally, as possible explanations for the presence of MITEs in those species lacking full- length elements (namely D. yakuba, D. ficusphila and D. elegans), it is plausible that these MITEs derived from the ancestral transposition of full- length elements, followed by generation of MITEs and elimination of the original founders. Thus, the single elements identified are just the relic of this elimination. Expansion of MITEs has probably occurred one or multiple times within the host genomes as suggested by sequence similarity distribution. Notably, the D. sechellia, D. suzukii D. ficusphila and D. takahashii genomes lack autonomous Bari elements, so it would be of particular interest to know if transposition events involving MITEs occurred after the elimination of autonomous elements, resulting in transposition mediated by unrelated transposases. ## Bari target site preferences and the physical relationships with other mobile elements It is widely accepted that mobile elements do not integrate themselves randomly. Primary and secondary DNA structure, as well as chromatin status, could affect the target site selection. Our results point out a strong preference of Bari-like elements for AT-rich sequences. Similar results were obtained for other Tc1-like transposons, like SB which has a preference for a palindromic AT-repeat (ATATATAT) in which the central AT is the cleaved and duplicated target. In addition to the above-described analysis we have performed a larger-scale analysis aiming to the identification of TEs that preferentially insert, or preferentially avoid, the DNA neighborhood of Bari elements. We found that loci in which Bari transposons are inserted are also populated by other Tc1-mariner elements (p\<0,05). The significant association between Bari elements and other transposons of the same class lead us to hypothesize a possible correlation between the DNA (or chromatin) structure and the insertion preferences of transposons belonging to the same superfamily. On the other hand, the members of the copia, Jockey and I families are significantly under- represented in the range of 2,5 kb around Bari elements in the analyzed species. Helitron elements also display a preference for sites near Bari insertions, while gypsy–like elements do not show significant preference/avoidance respect to Bari elements insertion sites suggesting that gypsy-like elements have a random genomic distribution. Besides its importance in identifying the structural organization of the single loci in which Bari elements lie, this kind of analyses would hopefully help understanding how a DNA domain is built during evolution. This is especially interesting for heterochromatic DNA blocks, which are largely composed of transposons’ arrays and can develop to master loci producing small regulatory RNAs. ## Bari phylogeny inconsistencies and possible horizontal transfer events The evolution of Bari elements has been extensively studied in previous works. The non-uniform distribution of the three subfamilies across the species of the Drosophila and Sophophora subgenera was observed by Moschetti et al., and explained with a distinct evolutionary history in different genomes. The overlap observed here, between the Bari1-Bari2 subfamilies in some species (D. melanogaster, D. simulans, D. sechellia), might be explained hypothesizing two independent waves of genomic invasion: the first, most ancient, by Bari2 elements which probably occurred in the ancestor of the melanogaster subgroup (comprising the melanogaster, simulans, sechellia and erecta species analyzed in this work), the second, more recent, by Bari1 which have invaded the ancestor of the melanogaster complex (including the melanogaster simulans and sechellia species). Assuming only vertical transmission, Bari2 elements have been inactivated while Bari1 elements are still active in D. melanogaster and, potentially, in D. simulans. After these invasions, Bari2 elements were probably inactivated resulting in the absence of active copies in these genomes. It would be interesting to know if Bari2 elements are also present in non-sequenced genomes of the Drosophila subgenus, or if Bari2 subfamily is restricted to species of the Sophophora subgenus. The emergence of transposable elements in a genome can occur in three ways: de novo emergence, horizontal transfer and introgression. Patchy TEs distribution, incongruent TE vs host species phylogeny and the presence of highly similar sequences in distantly related species are used as proofs in support of TEs horizontal transfer. Several horizontal transfer events involving Tc1-mariner elements have been described so far. In a recent paper, Dupeyron and collaborators described horizontal transfer events between terrestrial isopod crustaceans and hexapods involving Tc1-mariner elements. Horizontal transfer events were also suggested for Bari-like elements between the sibling species D. melanogaster and D. simulans as reported by former studies. Here we present evidence of horizontal transfer events involving Bari-like elements between D. bipectinata and D. biarmipes. Although these two species diverged at least 27 million years ago, Bari elements within the respective genomes are nearly identical. Analysis of synonymous substitutions differences between transposases and host genes suggest that horizontal transfer occurred, which explain the incongruences observed in the phylogenetic trees and the high similarity between the Bari elements in these species. The geographic distribution of the Drosophila species involved in transposon horizontal transfer events supports this possibility, as these two species coexist in the Indian subcontinent. Very recently Wallau et al., reported the same horizontal transfer event described here by applying a sophisticated and statistically supported method, called VHICA, which used 50 orthologous, vertically transmitted genes, as a reference set to infer horizontal transposon transfer of Bari between D. biarmipes and D. bipectinata. Other horizontal transposon transfer events involving different species have been also detected using this method suggesting that complex evolutionary mechanisms have originated the actual distribution of Bari elements, complicating the inference of their evolutionary history. # Conclusions Our annotation and analyses of 401 insertions unveiled sequence and structure variability of Bari-like elements in the sequenced genomes of Drosophila species. Besides the dynamic structure of the TIRs and the presence of active and inactive elements in the three Bari subfamilies, we detected the presence of MITEs derived from Bari elements in 9 Drosophila species, suggesting that the generation of such inactive form can be considered as a common event in this family. The analysis of genomic sites targeted by Bari transposition showed that the same sites are also preferred or avoided by other mobile elements, and this may be important to understand how transposons model genomic domains. Finally our phylogenetic analysis showed that three subfamilies (Bari1, Bari2 and Bari3) can be recognized both in TIR- and transposase-based phylogenetic trees and that a previously unreported horizontal transfer event has probably occurred between D. biarmipes and D. bipectinata. # Supporting Information We gratefully acknowledge Dr. Kostantinos Lefkimmiatis for critical discussion and language editing of the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: RMM. Performed the experiments: AP DL PDA RC RMM. Analyzed the data: AP DL PDA RC RMM. Wrote the paper: RC RMM.
# Introduction G protein-coupled receptor (GPCR) kinases (GRKs) are a family of serine/threonine kinases that phosphorylate GPCRs and desensitize GPCR-mediated signaling. GRK-catalyzed receptor phosphorylation leads to the recruitment of beta-arrestins to phosphorylated receptors, induces receptor internalization, and thus down-regulates cellular responses to extracellular signal. Many studies indicate that GRKs are able to phosphorylate a variety of non-GPCR substrates such as synuclein, p38, NF-κB1 p105, ezrin, arrestin-2, and p53. It has also been shown that GRKs can regulate signaling pathways via direct interaction with other proteins in a phosphorylation-independent manner. GRK2 is able to interact with Gαq to regulate GPCR signaling. Binding of GRK5 with IκB inhibits NF-κB- mediated transcription. Our earlier research showed that the kinase activity- independent regulation of the cyclin pathway by GRK2 is essential for zebrafish early development and GRK5 acts as a scaffold to promote F-actin bundling and targets bundles to membrane structures to control neuronal morphogenesis. These studies implicate that GRKs, especially GRK5, may exert multiple physiological functions via various mechanisms including those independent of their kinase activities. Change in GRK protein level has been detected in a variety of human disorders including heart failure, acute myocardial infarction, hypertension, brain ischemia, rheumatoid arthritis, Parkinson’s disease, Alzheimer’s disease and depression, suggesting that protein turnover plays a key role in GRK regulation. The regulation of GRK2 turnover has been studied. Mdm2 plays a key role in regulation of GRK2 ubiquitination and degradation. Hsp90 interacts with and stabilizes GRK2. However, little is known about regulation of other GRK subtypes. Damaged DNA-binding protein 1 (DDB1) is part of an E3 ligase complex that includes the cullin proteins CUL4A and CUL4B. The DDB1–CUL4 complex is a conserved cullin-RING ubiquitin ligase, that regulates DNA repair, replication, and transcription. The DDB1–CUL4 complex can also be subverted by pathogenic viruses to benefit viral infection. CUL4 assembles ubiquitin ligase by binding to ROC1, a RING protein, and to DDB1, a triple β propeller adapter protein, which functions as a linker to recruit substrates or substrate receptors to CUL4 E3 ubiquitin ligase. In an effort to identify proteins that interact with GRK5, we used a proteomic approach to isolate GRK5 interacting proteins and identified several proteins in the GRK5 immunocomplex including DDB1. We further demonstrated that DDB1 acts as an adapter to link GRK5 to CUL4-ROC1 E3 ligase complex and regulates GRK5 ubiquitination and degradation. # Results ## Identification of GRK5-interacting Proteins by Mass Spectrometry A proteomic approach was used to identify interacting proteins of GRK5. Flag- tagged GRK5 was affinity purified from MDA-MB-231 cells stably expressing GRK5-Flag. As a control, cells stably expressing GFP were taken through the same procedure. The resulting protein complexes were electrophoresed on a 4–20% gradient polyacrylamide gel and then stained with Coomassie Blue. Protein bands of ∼300 kDa, ∼230 kDa, ∼130 kDa, ∼85 kDa and ∼55 kDa were selectively copurified with Flag-tagged GRK5 but not control beads. The gel bands were excised and analyzed by mass spectrometry as indicated in. The corresponding bands in the control lane were also excised and analyzed by mass spectrometry, and proteins detected in both the control lane and GRK5-immuocomplex lane were excluded from further analysis. This work led to the identification serine/arginine repetitive matrix 2 (SRRM2, 300 kDa), Myosin-9 (MYH9, 227 kDa), AP-3 complex subunit delta 1 (AP3D1, 130 kDa), damage-specific DNA binding protein 1 (DDB1, 127 kDa), heat shock protein Hsp90-alpha (Hsp90AA1, 83 kDa), heat shock protein Hsp90-beta (Hsp90AB1, 85 kDa), upstream binding transcription factor, RNA polymerase 1 (UBTF4, 89 kDa), nucleolin (NCL, 77 kDa) and serine/threonine kinase 38 (STK38, 55 kDa) as the major interacting proteins of GRK5 in MDA-MB-231 cells. DDB1 functions as a linker to recruit substrates or substrate receptors to CUL4-ROC1 E3 ubiquitin ligase. Interestingly, peptides derived from other components of DDB1-CUL4 complex were also identified in the GRK5 immunoprecipitates by mass spectrometry including CUL4B, which interacts with DDB1 and assembles ubiquitin ligase by binding to ROC1; WD40-repeat protein 22 (WDR22) and glutamate-rich WD40-repeat protein 1 (GRWD1), which contain WD40 repeat domain and constitute the substrate-specific adaptor that recruits substrates to the DDB1-CUL4 complexes through interaction with DDB1; COP9 signalosome complex subunit 7a (COPS7A), component of the COP9 signalosome complex, which cleaves NEDD8 from CUL4 and regulates the CUL4 ubiquitin ligase activity. Similar proteomic approach was carried out in human umbilical vein endothelial Cells (HUVEC) cells transiently transfected with Flag-GRK5. SRRM2, AP3D1, Hsp90, UBTF4 and NCL were also detected in HUVEC cells by mass spectrometry (supplemental,). The DDB1-CLU4 complex proteins including DDB1, CUL4B, WDR22, GRWD1 and COPS7A were all detected in HUVEC cells (supplemental), and two additional WD40-repeat domain containing proteins were identified in HUVEC cells including WDR12 and WDR82. However, STK38, which was readily detected in the GRK5 immunocomplex in MDA- MB-231 cells by mass spectrometry, was not detected in the HUVEC cells ( and,S2,S3,S4). As we identified several components of DDB1-CUL4 complex in both MDA-MB-231 cells and HUVEC cells by mass spectrometry, we further examined the association of GRK5 with DDB1-CUL4 ubiquitin ligase complex. ## GRK5 Associates with DDB1-CUL4 Ubiquitin Ligase Complex The association of GRK5 with DDB1-CUL4 complex was confirmed by immunoprecipitation and Western analysis in MDA-MB-231 cells stably expressing GRK5-Flag and in 293 T cells transiently transfected with GRK5-Flag. DDB1, CUL4A, and CUL4B were detected in the immunocomplex of the Flag-tagged GRK5. The RING finger protein ROC1 was also detected in GRK5 precipitates in 293 T cells and MDA-MB-231 cells (data not shown). Damage DNA binding protein 2 (DDB2), a known binding protein of DDB1, functioning in the global genome nucleotide excision repair, was also detected in the GRK5 precipitates. Furthermore, endogenous GRK5 was detected in DDB1 immunocomplexes in 293 T cells, demonstrating association between endogenous GRK5 and endogenous DDB1 in vivo. ## DDB1 Targets GRK5 to CUL4 to Form a Complex To test the idea that DDB1 acts as an adapter to recruit GRK5 to CUL4 to form a complex, epitope-tagged CUL4A, DDB1, and GRK5 were expressed in 293 T cells and protein associations were analyzed by immunoprecipitation. As shown in, the association of DDB1 with GRK5 could be detected in the absence of CUL4A-HA expression, and similarly, the association of DDB1 with CUL4A-HA could occur in the absence of GRK5 expression, confirming the interaction of GRK5 and CUL4A with DDB1. When CUL4A-HA, DDB1-Flag, and GRK5 were co-expressed, all three proteins were present in a complex. However, in the absence of DDB1-Flag, the association between CUL4A-HA and GRK5 was significantly attenuated. These data suggest that DDB1 is required for the association of GRK5 with CUL4A. The basal interaction of GRK5 with CUL4A-HA in the absence of DDB1-Flag expression was likely a result from co-immunoprecipitated endogenous DDB1 by CUL4A-HA, since significant amount of endogenous DDB1 could be detected in the CUL4A immunoprecipitates (panel 3). Similar results were obtained when using GRK5-HA to immunoprecipitates DDB1 and CUL4A. Furthermore, knockdown of DDB1 significantly reduced the amount of CUL4A in GRK5 immunoprecipites. However, knockdown of CUL4A had no significant effect on the association between GRK5 and endogenous DDB1. These results further support the notion that DDB1 acts as an adaptor to link CUL4A with GRK5. Based on the overall structural organization and homology of GRKs, GRK family proteins can be divided into three subfamilies: GRK1 and GRK7; GRK2 and GRK3; and GRK4, GRK5, and GRK6. We examined the interaction of DDB1 with other members of GRK family proteins. As shown in, DDB1 could be detected in the immunoprecipitates of GRK5-Flag as well as GRK6-Flag, but not in GRK2-Flag immunoprecipitates, suggesting the specific complex formation of DDB1 with GRK4 subfamily proteins. ## DDB1 Regulates GRK5 Ubiquitination Since DDB1 targets GRK5 to the CUL4 ubiquitin ligase complex, we tested whether DDB1 regulates GRK5 ubiquitination and degradation. 293 T cells were co- transfected with plasmids encoding Flag-tagged GRK5 and HA-tagged ubiquitin, in combination with control plasmid, DDB1, or myc-CUL4A. As shown in, expression of DDB1 alone or co-expression of DDB1 with CUL4A significantly increased GRK5 ubiquitination, while expressing CUL4A alone did not affect the ubiquitination of GRK5. To examine the requirement of DDB1-CUL4 complex for GRK5 ubiquitination, 293 T cells were co-transfected with GRK5-Flag, HA-ubiquitin, and shRNA targeting DDB1 or CUL4. As shown in, GRK5 ubiquitination was severely decreased by the depletion of either DDB1 or CUL4A/B. ## DDB1 Regulates GRK5 Degradation We next examined the effect of DDB1 on GRK5 degradation. 293 T cells were infected by lentivirus expressing control-shRNA or DDB1-shRNA, and the degradation of endogenous GRK5 was determined in the presence of CHX in chasing experiment. As shown in, in the control cells, 50% of GRK5 was degraded in less than 3 h, whereas knockdown of DDB1 markedly attenuated the degradation of GRK5 and extended the half-life of GRK5 (∼25% reduction of GRK5 in 8 h). Hsp90 has been shown to interact with GRK2 and regulate the stability of GRK2 by preventing it from proteasome-dependent degradation. A significant amount of Hsp90 peptides were also detected in the GRK5 immunoprecipitates by mass spectrometry. We next examined the role of Hsp90 in regulating GRK5 stability. As shown in, treatment of 293 T cells with geldanamycin, an Hsp90-specific inhibitor, significantly down-regulated the level of endogenous GRK5 in a dose- dependent manner, suggesting that Hsp90 regulates GRK5 stability. This is consistent with the previous report that treatment with geldanamycin resulted in about 80% down-regulation of GRK5 transiently expressed in COS-7 cells. Furthermore, pretreatment of cells with MG132, a proteasome inhibitor, significantly inhibited geldanamycin-induced down-regulation of GRK5, suggesting that GRK5 degradation induced by geldanamycin was predominantly through the proteasome pathway. We further examined the role of DDB1 in Hsp90 inhibition- induced degradation of GRK5. As shown in, knockdown of DDB1 significantly inhibited geldanamycin-induced down-regulation of GRK5 in 293 T cells, suggesting that DDB1 mediates Hsp90 inhibition-induced proteasome-dependent degradation of GRK5. ## DDB1 Mediates UV Irradiation-induced GRK5 Degradation We further explored the potential role of the DDB1-CUL4 ubiquitin ligase in GRK5 degradation in 293 T cells. It has been shown that DDB1-CUL4 ubiquitin ligase promotes degradation of many proteins including p21, CDT1, and DDB2 following ultraviolet light (UV) irradiation. The response of GRK5 following DNA damage with UV irradiation was examined. The cellular GRK5 level was significantly down-regulated following UV treatment with as little as 20 J/m<sup>2</sup> UV, whereas GRK2 levels remained largely unchanged under the same condition. Remarkably, knock down of DDB1 effectively prevented UV-induced GRK5 degradation, suggesting that UV irradiation-induced degradation of GRK5 is mediated by DDB1. The effect of GPCR activation on the stability of GRK5 was also examined. As shown in, treatment of cells with isoproterenol to activate endogenous β2-adrenergic receptors in 293 T cells, had no significant effect on GRK5 protein level (data not shown) or UV irradiation-induced GRK5 degradation. # Discussion In the current study, a proteomic approach was used to screen GRK5 interacting proteins in MDA-MB-231 cells and HUVEC cells. Several proteins were detected in the GRK5 immunocomplex including SRRM2, MYH9, AP3D1, DDB1, Hsp90, UBTF4, NCL and STK38. Interestingly, other components of DDB1-CUL4 ubiquitin ligase complex including CUL4B, WDR22, GRWD1 and COPS7A, were also detected in the GRK5 immunocomplex in both MDA-MB-231 cells and HUVEC cells. We further provided evidence that GRK5 forms a complex with DDB1-CUL4-ROC1 E3 ubiquitin ligase and DDB1 acts as an adaptor to link GRK5 to CUL4 to form the complex. DDB1 regulates the ubiquitination and degradation of GRK5. Furthermore, depletion of DDB1 inhibited Hsp90 inhibitor-induced GRK5 destabilization and UV irradiation- induced GRK5 degradation. Thus, our immunoprecipitation-mass spectrometry data provide useful information about GRK5 interacting proteins in cells, and our results reveal DDB1 as a key regulator of GRK5 stability. Altered GRK protein expression and/or activity have profound effects on cell signaling and physiological functions, and changed GRK expression has been observed in a variety of human disorders. The mechanisms that govern GRK2 cellular levels have recently been addressed. GRK2 is rapidly degraded via the proteasome pathway. The ubiquitination and turnover of GRK2 are stimulated by β2-adrenergic receptor activation, through a mechanism involving GRK2 phosphorylation by c-Src or MAPK in a beta-arrestin-dependent manner. Mdm2 is subsequently identified as the key E3 ubiquitin ligase involved in GRK2 ubiquitination and degradation. However, how the stability of other GRK subtypes is controlled remains largely unknown. In the current study, we identified DDB1-CUL4 complex as the key ubiquitin ligase responsible for GRK5 ubiquitination and degradation. Several lines of evidence support the notion that DDB1 serves as a linker to target GRK5 to DDB1-CUL4 E3 ligase for GRK5 ubiquitination and degradation. First, DDB1 was detected in GRK5 immunoprecipitates of lysates from different cell lines. Second, a pool of endogenous GRK5 and DDB1 can be found in the same molecular complex, as indicated by co-immunoprecipitation. Moreover, overexpression or knockdown of the protein reveals DDB1 is an adapter linking GRK5 to DDB1-CUL4 complex. Third, GRK5 ubiquitination and degradation is significantly impaired in DDB1/CUL4-knockdown cells. Finally, we found that degradation of GRK5 induced by both Hsp90 inhibitor and UV-irradiation could be both inhibited in DDB1 deficient cells. Thus this may serve as a new regulation mechanism for GRK5 stability. It will be interesting to further explore if the DDB1-mediated GRK5 degradation is DDB1/GRK5 interaction-dependent, or if there is a direct interaction between GRK5 and DDB1. We also show that DDB1 preferentially associates with GRK4 family proteins. DDB1 could be observed in both GRK5 and GRK6, but not GRK2 immunoprecipitates. GRK2 has been shown to interact with Mdm2. These results suggest that different GRK subtypes may form complexes with diverse E3 ubiquitin ligases. In addition, Penela et al. reported that GRK2 has a half-life of about 1 h in C6 glioma and Jurkat cells. In this study, we observed that in 293 T cells, endogenous GRK5, but not GRK2, was rapidly degraded after CHX treatment. Our results are consistent with previous reports showing that the half-life of GRK2 in HL60 cells is 20–24 h. This may reflect difference in degradation mechanisms of GRK2 among these cell types. Hsp90 has been shown to interact with GRK2 and GRK3, and regulates the stability of GRK2 and GRK3. Inhibition of Hsp90 results in rapid, proteasome-dependent, down-regulation of GRK2 and GRK3. In this study, we found that inhibition of Hsp90 resulted in proteasome-dependent degradation of GRK5. These studies and our data implicate that Hsp90-mediated stabilization and proteasome-dependent degradation may play a general role in regulation of the stability of GRK family proteins. Hsp90 has also been reported to interact with and participate in the regulation of other kinases, such as Erb2, Akt/PKB, and Raf-1. Inhibition of Hsp90 by geldanamycin results in enhanced degradation of these proteins. Therefore, stabilization of client protein seems to be a general function of Hsp90. We further provide the evidence that the degradation of endogenous GRK5 induced by Hsp90 inhibition was mediated by DDB1. Depletion of DDB1 inhibited geldanamycin-induced degradation of GRK5. These results suggest that disrupting the interaction between GRK5 and Hsp90 induces DDB1-CUL4 complex-mediated ubiquitination-proteasome dependent degradation of GRK5, and put forward the notion that the stability of GRK5 was counter-regulated through interaction with DDB1-CUL4 complex and Hsp90. DDB1-CUL4 complex controls ubiquitination and stability of many cellular substrates including p21, DDB2, Cdt1, p27, TSC2, Merlin, Chk1, c-Jun, and histones, through DDB1 bridged interaction of substrates with CUL4 E3 ubiquitin ligase. p21 and Cdt1 are the most established substrates of DDB1-CUL4 ubiquitin ligase complex after UV irradiation. The degradation of p21 and Cdt1 following UV irradiation, which promotes DNA repair and suppresses replication licensing, is critical for the cell to respond to DNA damage. Interestingly, in this study, we found that GRK5, but not GRK2, is degraded after UV irradiation, and DDB1 mediates UV irradiation-induced GRK5 degradation. The identification of GRK5 as a DDB1-CUL4 complex target following UV irradiation adds it to the growing list of cellular substrates targeted to CUL4 by DDB1. The physiological consequence of UV irradiation-induced GRK5 degradation remains to be explored. The current study is aimed to explore the possible regulation of GRK5 by the DDB1-CUL4 ubiquitin ligase complex. However, the association of GRK5 with DDB1-CUL4 complex may also result in the phosphorylation of DDB1 and/or Cul4 by GRK5, and thus regulates the function of DDB1-CUL4 ubiquitin ligase complex itself. In fact, cells deficient of GRK5 or DDB1 exhibit similar phenotypes, such as apoptosis and cell growth retardation. It would be interesting to determine whether GRK5 regulates the function of DDB1-CUL4 ubiquitin ligase complex and the ubiquitination of other substrates of DDB1-CUL4 complex, and investigate whether DDB1-CUL4 complex is involved in GRK5 deficiency-induced cell growth retardation in future studies. We also identified other new interacting proteins of GRK5, including SRRM2 and STK38, by immunoprecipitation-mass spectrometry in MDA-MB-231 cells. SRRM2, which has been shown to be involved in pre-mRNA splicing, , is the most abundant among proteins identified by mass spectrometry in GRK5-immunocomplex in both MDA-MB-231 cells and HUVEC cells (supplemental). These results implicate possible function of GRK5 in mRNA splicing. Another interesting interacting protein of GRK5 is STK38 which is identified in GRK5 immunocomplex in cancer cell lines (MDA-MB-231 cells), but not in normal cells (HUVEC). STK38, also known as NDR1, is a serine-threonine protein kinase of AGC kinase subfamily, it has been shown to regulate many essential processes including cell division, proliferation, centrosome duplication and apoptosis. Interestingly, GRK5 has also been shown to be involved in proliferation and apoptosis. It will be of interest to examine the possible role of interaction between GRK5 and STK38 in regulating these processes. # Materials and Methods ## Reagents and Antibodies Dulbecco’s modified Eagle’s and Leibovitz’s L-15 media were purchased from Invitrogen. Fetal bovine serum (FBS) was obtained from Hyclone. Mouse anti-DDB1 antibody, rabbit anti-ROC1 antibody, and mouse antibody against Myc epitope were from Invitrogen. Rabbit anti-DDB2 antibody, rabbit anti-GRK2 antibody, and protein A/G agarose were from Santa Cruz Biotechnology. Rabbit anti-CUL4B antibody, rabbit anti-β-actin antibody, mouse antibody against FLAG epitope, rabbit antibody against HA epitope, mouse IgG, anti-Flag M2 affinity gel, and anti-HA agarose were from Sigma. Goat anti-GRK5 antibody was from R&D. Geldanamycin and cycloheximide (CHX) were from Sigma. ## Plasmid Construction Plasmids encoding tagged bovine GRK2 and GRK5, and human GRK6 were constructed by PCR mutagenesis. Construction of shRNA plasmids for human DDB1, human CUL4A, and human CUL4B was performed as described. LacZ shRNA was used as a control. The shRNA lentivirus system, which uses the FG12 and package vector to enable simultaneous expression of the GFP protein and shRNA were obtained from Dr. Gang Pei (Chinese Academy of Sciences, Shanghai, China). Sequences for DDB1 RNAi are: GCG AGA GCA TTG ACA TCA TTA (sense) and TAA TGA TGT CAA TGC TCT CGC (antisense). CUL4A: GCA GGA CCA CTG CAG ACA AAT (sense) and ATT TGT CTG CAG TGG TCC TGC (antisense). CUL4B: GCC ACG TAC CGA TAC AGA AGA (sense) and TCT TCT GTA TCG GTA CGT GGC (antisense). The plasmids DDB1-Flag (Addgene plasmid 19918), CUL4A-HA (Addgene plasmid 19907), and Myc-CUL4A (Addgene plasmid 19951) were gifts of Dr. Yue Xiong. The plasmid DDB1 was constructed into pcDNA3.0. ## Cell Culture and Plasmid Transfection 293 T cells were cultured in Dulbecco’s modified Eagle’s medium containing 10% FBS. Cells were seeded in 60 or 100 mm tissue culture dishes at 0.6–2×10<sup>6</sup>/dish 20 h before transfection and transfected plasmid using calcium phosphate/DNA co-precipitation method or infected by lentivirus. Assays were performed 48 h (for expressing experiments) or 72 h (for RNAi experiments) after transfection. MDA-MB-231 cells stably expressing GFP or GRK5-Flag were cultured in Leibovitz’s L-15 plus 10% FBS. For protein decay analysis, 293 T cells were treated with CHX at 100 µg/ml. ## Generation and Titer of Lentivirus The sequences predicted to target the human DDB1 or CUL4A were used for the construction of shRNA lentivirus. FG12-hU6-shRNA and packaging vectors were co- transfected into 293 T cells and the resulting supernatant were collected after 48 h. Virus was recovered after ultracentrifugation resuspension in phosphate- buffered saline. Titers were determined by infecting 293 T cells with serial dilutions of concentrated lentivirus. GFP expression in infected cells was determined 48 h after infection, and for a typical preparation, the titer of lentivirus was approximately 10<sup>8</sup> infectious units (IFU)/ml. ## Isolation and Identification of GRK5-associated Proteins MDA-MB-231 cells were stably transfected with GRK5-Flag or GFP lentivirus. HUVEC cells were transiently transfected with GRK5-Flag or GFP by lentivirus. About 10<sup>8</sup> cells were harvested and lysed in IP buffer (50 mM Tris.HCl, pH 7.5, 150 mM NaCl, 0.5% NP-40, 10% glycerol, 1 mM EDTA, 10 mM NaF, plus 10 µg/ml aprotinin, 10 µg/ml benzamidine, and 0.2 mM PMSF) for 2 h. The lysate was centrifuged and the supernatant was incubated with anti-FLAG M2 affinity gel at 4°C for 8 h. The beads were subsequently washed. Bound proteins were eluted by addition of SDS sample buffer. The resulting protein complexes were electrophoresed on a 4–20% gradient polyacrylamide gel and revealed by staining with Coomassie Blue. Six selected bands were excised from the gel and analyzed by mass spectrometry. The corresponding GFP control lane were also excised and analyzed by mass spectrometry. The mass spectrometry analysis was performed at the Shanghai Applied Protein Technology Co.Ltd. ## LC-MS/MS Gel pieces were destained with 30% ACN/100 mM NH<sub>4</sub>HCO<sub>3</sub> and dried in a vacuum centrifuge. The in-gel proteins were reduced with dithiothreitol (10 mM DTT/100 mM NH<sub>4</sub>HCO<sub>3</sub>) for 30 minutes at 56°C, then alkylated with iodoacetamide (50 mM IAA/100 mM NH<sub>4</sub>HCO<sub>3</sub>) in the dark at room temperature for 30 minutes. Gel pieces were briefly rinsed with 100 mM NH<sub>4</sub>HCO<sub>3</sub> and ACN, respectively. Gel pieces were digested overnight in 12.5 ng/mL trypsin in 25 mM NH<sub>4</sub>HCO<sub>3</sub>. The peptides were extracted three times with 60% ACN/0.1% TFA. The extracts were pooled and dried. EttanTM MDLC system (GE Healthcare) was applied for desalting and separation of tryptic peptides mixtures. In this system, samples were desalted on RP trap columns (Agilent Technologies), and then separated on a RP column (Column technology Inc.). Mobile phase A (0.1% formic acid in HPLC-grade water) and the mobile phase B (0.1% formic acid in acetonitrile) were selected. 20 µg of tryptic peptide mixtures was loaded onto the columns, and separation was done at a flow rate of 2 µL/min by using a linear gradient of 4–50% B for 30 min. A LTQ VELOS (Thermo Electron) equipped with an electrospray interface was connected to the LC setup for eluted peptides detection. Data-dependent MS/MS spectra were obtained simultaneously. Each scan cycle consisted of one full MS scan in profile mode followed by five MS/MS scans in centroid mode with the following Dynamic ExclusionTM settings: repeat count 2, repeat duration 30 s, exclusion duration 90 s. Each sample was analyzed in triplicate. MS/MS spectra were automatically searched against the non-redundant International Protein Index (IPI) human protein database (version 3.53) using the BioworksBrowser rev. 3.1(Thermo Electron). ## Immunoprecipitation and Western Blotting Cells were washed with ice-cold PBS and lysed in IP buffer for 1.5 h as described. The lysate was centrifuged and the supernatant was incubated with anti-FLAG M2 affinity gel or anti-HA agarose at 4°C for 8 h. The beads were subsequently washed, and the proteins bound to the beads were eluted. The samples were detected in the subsequent Western procedures with the corresponding antibody. Blots were incubated with IRDye 800CW-conjugated or 700CW-conjugated antibody (Rockland Biosciences) and infrared fluorescence images were obtained with the Odyssey infrared imaging system (Li-Cor Bioscience). ## Ubiquitination Assay 293 T cells seeded in 60 mm dishes were transfected with different combinations of expression vectors for HA-ubiquitin, GRK5-Flag, and other indicated plasmids. Cells were lysed in ubiqutination buffer (50 mM HEPES, pH 8.0, 250 mM NaCl, 0.5% NP-40, 10% glycerol, 2 mM EDTA, 20 mM NaF, plus 10 µg/ml aprotinin, 10 µg/ml benzamidine, and 0.2 mM PMSF). Cell lysate was immunoprecipitated with anti-Flag M2 affinity gel and analyzed by immunobloting with anti-HA antibody. ## CHX Treatment For protein decay analysis, 293 T cells were infected by LacZ RNAi or DDB1 RNAi lentivirus for 48 h and then reseeded onto 60 mm culture dish at 7×10<sup>5</sup> cells per dish. After 40 h, cells were exchanged with fresh culture medium and CHX (100 µg/ml) was added at the proper time point for CHX treatment for 0 h, 2 h, 4 h, or 8 h. After CHX treatment, cells were washed with PBS, and total cell lysates were prepared for Western blotting. ## UV-C Irradiation UV-C (254 nm) irradiation was performed with a UV-C transilluminator (Hoefer UVC 500). The culture medium was replaced with complete medium two hours before the UV-C treatment. Cells were irradiated at 80% confluency with a 20 J/m<sup>2</sup> dose or indicated dose after removal of the medium in the 60 mm petri dish without the lid. After irradiation, cells were post-incubated in their culture medium for the indicated time and were lysed directly in RIPA<sup>+</sup> buffer (50 mM Tris.HCl, pH 8.0, 150 mM NaCl, 1% NP-40, 0.5% Deoxycholic Acid, 0.1% SDS, 5 mM EDTA, 10 mM NaF, 10 mM Disodium pyrophosphate, plus 10 µg/ml aprotinin, 10 µg/ml benzamidine, and 0.2 mM PMSF), and equal amount of protein was loaded on gels. ## Statistical Analysis Data were analyzed using two-way ANOVA for comparison of independent means with pooled estimates of common variances. # Supporting Information We thank Junjun Wang and Wenqing Yao for technical assistance. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YC LM. Performed the experiments: ZW YC TY QG MY. Analyzed the data: ZW YC TY LM. Contributed reagents/materials/analysis tools: QG MY. Wrote the paper: YC ZW TY.
# Introduction An important issue in health geography and health policy is the evaluation of accessibility to healthcare services, with hundreds of research papers published on the topic since the 2000s. However, the concept of accessibility is multi- dimensional, which often presents challenges to its operationalization in empirical research. According to Joseph and Bantock, accessibility can be defined by both aspatial and spatial dimensions. The first dimension considers factors such as the quality of the services and their cost, as well as the income, social class, ethnicity, and mobility profile of potential users of services. From a geographical perspective, the spatial dimension is key, and considers the distribution of available healthcare services across the landscape, in addition to the cost or friction that potential users incur when trying to reach these services. By taking these geographical factors into account, estimates of accessibility can help researchers, planners, and policy makers identify areas with high or low accessibility to healthcare services. This, in turn, can provide valuable information related to social and spatial inequalities and guidance for health policy and resource allocation. Spatial accessibility can be estimated in various ways. At a high level, provider-to-population ratios (PPR) offer an indication of the level of service within a community. These measures conceptualize a region as a container of population and services, and therefore are sometimes called container approaches. PPRs are straightforward to interpret as the supply of a service (say number of doctors, beds, etc.) divided by demand (say, number of people who require the service). Despite this convenient and intuitive interpretation, container approaches are limited in the amount of spatial information that they provide, especially if applied to large regions. When applied to smaller regions these approaches present other shortcomings, such as the assumption that the population in the container is captive and does not cross the boundaries of the container in search of services—and that users do not come into the container from other regions to avail themselves of local services. An alternative to container approaches is provided by gravity measures. Gravity measures offer a more sophisticated approach to measuring spatial accessibility to healthcare that moreover addresses some of the limitations of the container approach. Instead of defining rigid container boundaries, gravity measures consider the mobility characteristics of the public to produce flexible (and often overlapping) catchment areas for both services and population. Accordingly, one of the most popular approaches for estimating healthcare accessibility in the literature is the Two-Step Floating Catchment Area (2SFCA) method proposed by Luo and Wang after research by Radke and Mu. The 2SFCA method is an ensemble of two gravity models with a simplified binary distance function to account for crowding of facilities and allocation of levels of service. Numerous applications of this methods are found in the international literature, including work from Germany, South Korea, Japan, China, Australia, and Canada. Accessibility to healthcare is estimated in two stages in the 2SFCA: in the first step, a level of service at a given healthcare provider is determined based on the supply (e.g., number of physicians in a clinic) and the estimated demand from the surrounding population within some catchment area. This level of service resembles a local provider-to-population ratio (PPR). In the second step, the level of service of different healthcare providers is aggregated for each population center. By operationalizing accessibility in terms of demand and level of service, the 2SFCA method is appealing for health policy analysis. Still, several improvements have been proposed that seek to address the method’s most important perceived shortcomings. The result is a family of Floating Catchment Areas (FCA) methods that include more realistic conceptualizations of the friction of distance by specifying variable catchment area sizes and/or the use of stepped, continuous, and adaptive distance-decay functions. Other authors have added multi-modal transportation, age-adjusted healthcare demand profiles, as well as ways to counteract the modifiable areal unit problem. A major focus of FCA research, in addition to the improvements mentioned above, has been the introduction of competition for available opportunities or the allocation of services to the population. More concretely, the original 2SFCA approach has been criticized for over-estimating the levels of demand and/or level of service in the system. This is a consequence of the way catchment areas for facilities and population centers typically overlap in any realistic spatial system—an artifact of FCA methods that can lead to misleading estimates of accessibility. In effect, when aggregating the population within the overlapping catchment areas of multiple facilities, the original 2SFCA framework leads to double- counting of the population that tends to inflate the level of demand at supply points in the healthcare system. We call this effect demand inflation. Inflated demand, in turn, tends to deflate the level of service for populations serviced by the facilities so affected. A similar effect, which we call level of service inflation, happens when the levels of service of various service points are aggregated for population centers. Ultimately, accessibility estimates are affected in potentially complex ways, depending on the geography of the problem, and their interpretation as PPRs becomes suspect. Various solutions to the issues of demand and level of service inflation have been proposed, including the addition of selection weights based on a travel impedance function in the Three-Step Floating Catchment Area (3SFCA) method; the use of a Huff model to generate probability-based estimate of the selection weights in the 3SFCA method; and, on the supply side, a modified 2SFCA (M2SFCA) method to address suboptimal spatial configuration of services. In this paper we are interested in the way demand and level of service are calculated in FCA methods. We review how different approaches deal with the issue of inflation, and then propose a simple and intuitive approach to proportionally allocate supply and demand. Our solution consists on adjusting the impedance weights used in the estimation of FCA methods. More concretely, by incorporating methods drawn from the field of spatial statistics and econometrics, proportional allocation has the feature that it preserves the levels of demand and service in the system. To illustrate the key aspects of our proposal, we conduct a case study of access to family physicians in Hamilton, Canada. Our results indicate that the proposed adjustments produce more intuitive measures of accessibility to healthcare measured in terms of local PPRs. Moreover, these outputs can be used to provide estimates of access disparity across a region that are both easily understood and robust to demand and level of service inflation. # Background: Floating Catchment Area methods To motivate the discussion to follow we begin by reviewing some popular FCA methods. In general terms, FCA approaches are implemented as ensembles of two gravity models in two steps, using an impedance function to represent the cost required to overcome distance. Impedance functions implement a distance-decay effect that mimics a commonly observed cost-minimization behavior, namely that people in general prefer to spend less time/money/effort travelling to destinations. In this way, the impedance function defines a catchment area for the points of service and population centers alike. In the first step of FCA methods, the impedance function defines catchment areas for facilities j, which could be clinics, parks, libraries, etc. A weighted sum of the population within a catchment area is allocated to the corresponding facility or service point to represent demand. In the second step of the algorithm, the catchment areas are “floated” to population centers i. Accessibility at location i is calculated as the weighted sum of the level of service at every location j that includes i within its catchment area. The following methods are popular in the literature. ## Two-Stage Floating Catchment Areas (2SFCA) The original 2SFCA implements a binary impedance function W with a threshold cost d<sub>0</sub> as follows: $$W\left( d_{ij} \leq d_{0} \right) = \left\{ \begin{array}{ll} 1 & { d_{ij} \leq d_{0}} \\ 0 & { d_{ij} > d_{0}} \\ \end{array}\operatorname{} \right.$$ This function assumes equal potential within a catchment area (i.e., d<sub>ij</sub> ≤ d<sub>0</sub>), and zero beyond (d<sub>ij</sub> \> d<sub>0</sub>). This implies that 1) travellers are equally likely users of a service point within the catchment area, irrespective of how proximate or distant they are from it; and 2) no users travel to the service point from beyond the threshold cost. Given the impedance function, the level of demand D<sub>j</sub> is calculated as the weighted sum of the population at i: $$D_{j} = \sum\limits_{i}D_{ij} = \sum\limits_{i}{P_{i}W\left( d_{ij} \leq d_{0} \right)}$$ The supply S of the service offered at location j (say, number of beds/doctors in a clinic) is then divided by the demand to obtain a measure of level of service (e.g., beds/person, sq.m of park space/person, library floor space/person). This gives a level of service L<sub>j</sub> at the service point: $$L_{j} = \frac{S_{j}}{D_{j}} = \frac{S_{j}}{\sum_{i}D_{ij}} = \sum\limits_{i}\frac{S_{j}}{D_{ij}} = \sum\limits_{i}L_{ij}$$ The level of service resembles a PPR. Aggregation of demand creates a congestion effect that depends on the number of potential users from different origins i that converge at service point j: at a fixed level of supply, greater demand results in lower levels of service. The different decompositions of L<sub>j</sub> help to understand how different population centers contribute to the level of demand at facility j. In the second step of the algorithm, catchment areas are “floated” to population centers i. A second gravity model is used to calculate the accessibility at i: $$A_{i} = \sum\limits_{j}{L_{j}W\left( d_{ij} \leq d_{0} \right)}$$ Since accessibility is calculated as the weighted sum of the level of service at facilities, it is conventionally interpreted as a PPR. ## Enhanced Two-Stage Floating Catchment Areas (E2SFCA) A criticism of the binary impedance function of the 2SFCA is that it does not account for the declining probability of using a facility as distance grows. As a result of this criticism, other impedance functions have since been proposed, including the stepwise formulation of the Enhanced Two-Stage Floating Catchment Area method: $$W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right) = \left\{ \begin{array}{ll} k_{1} & { d_{ij} \leq d_{1}} \\ k_{2} & { d_{1} < d_{ij} \leq d_{2}} \\ \ldots & \\ k_{R - 1} & { d_{R - 1} < d_{ij} \leq d_{R}} \\ 0 & { d_{ij} > d_{R}} \\ \end{array}\operatorname{} \right.$$ A stepwise function does not assume identical potential within the catchment area (i.e., the space contained within d<sub>ij</sub> ≤ d<sub>R</sub>), but rather declining potential with increasing cost of travel. It is worthwhile noting that impedance functions have long been studied in geographical analysis in general, and accessibility research in particular. However, it is only relatively recently that alternative impedance functions have been incorporated in FCA approaches, including continuous functions and mixtures of continuous and step functions. Besides the use of a non-binary impedance function, the method remains the same. In the first step, demand is calculated as a weighted sum of the population within the catchment area: $$D_{j} = \sum\limits_{i}D_{ij} = \sum\limits_{i}{P_{i}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)}$$ Note that non-binary impedance functions discount the level of demand as a function of cost more rapidly than binary functions. How rapidly this happens depends on the definition of the cutoff values d<sub>1</sub>, d<sub>2</sub>, ⋯, d<sub>R</sub> and weights k<sub>1</sub>, k<sub>2</sub>, ⋯, k<sub>r−1</sub> of the function. In the second step of the algorithm, accessibility at i is calculated as the weighted sum of the level of service of service points j: $$A_{i} = \sum\limits_{j}{\frac{S_{j}}{D_{j}}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)} = \sum\limits_{j}{L_{j}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)}$$ Again, the use of a non-binary impedance function discounts the level of service more rapidly compared to binary functions. ## Three-Stage Floating Catchment Areas (3STCA) Wan et al. proposed a Three-Stage Floating Catchment Area method (3SFCA) that aims at refining the estimates of level of demand and accessibility by means of the use of selection weights. This approach operates by introducing an aditional step where selection weights are calculated as follows: $$G_{ij} = \frac{T\left( d_{ij} \right)}{\sum_{j\forall d_{ij} \leq d_{0}}T\left( d_{ij} \right)}$$ where T(d<sub>ij</sub>) are Gaussian weights (essentially an impedance function), and the summation in the denominator is for all sites j that are within a critical threshold d<sub>0</sub>. Notice that a property of the selection weights is that their sum over j equals one: $$\sum\limits_{j}G_{ij} = 1$$ Given a set of selection weights, the level of demand is caclulated by this algorithm in the following manner: $$D_{j}^{} = \sum\limits_{i}G_{ij}P_{i}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right) = \sum\limits_{i}G_{ij}D_{ij}$$ Notice how demand in this method takes what is essentially the demand in the E2SFCA, and allocates it proportionally to service points j. Accessibility, in the final step, becomes (with the subindices of the selection weights reversed, to reflect the displacement of the catchment area to population centers) is calculated in the following manner: $$A_{i}^{} = \sum\limits_{j}G_{ji}\frac{S_{j}}{D_{ij}^{}}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right) = \sum\limits_{j}G_{ji}L_{j}^{}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)$$ ## Modified Two-Stage Floating Catchment Areas (M2SFCA) Delamater discusses the application of FCA methods for systems that are not optimally configured to service the whole population. To address this issue, he proposes a modification to the second step of the 2SFCA algorithm that increases the friction of distance. Demand in this modification is the same as in 2SFCA. However, accessibility is calculated in the following manner: $$A_{i} = \sum\limits_{j}{L_{j}W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right)} = \sum\limits_{j}{L_{j}\left( W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right) \right)^{2}}$$ In other words, the level of service is discounted by the square of the impedance function, thus increasing the rate of decay. This is done to reflect the possibility that some population centers may experience increased friction to reach destinations in suboptimally configured systems. # Inflation effects in FCA methods Having reviewed a selection of FCA approaches, we now proceed to discuss the issue of inflation. Inflation has been identified, among others, by Wan et al. and Delamater. As discussed by these authors, inflation happens when demand or level of service are overestimated. Inflation is a consequence of the way in which Dj and A<sub>i</sub> are calculated, with some population centers contributing to the level of demand at more than one facility and then the level of service of facilities allocated to multiple population centers. Calculating demand, in particular, generally fails to preserve the population, and therefore lacks the pycnophilactic property discussed by Tobler. In practical terms, this implies that the population used to calculate the demand component of level of service will often exceed (but sometimes fall short of) the actual population in a region, depending on the weighting scheme. We term the consequent effect demand inflation. Let us illustrate this inflation effect by means of a simple example using the conventional 2SFCA approach with a binary impedance function. In this case, the population value at i is multiplied by zero or one, meaning that the contribution of i to demand at j whenever d<sub>ij</sub> does not exceed the threshold is: $$D_{ij} = P_{i}$$ If we concentrate for a moment on a single population center that enters the catchment areas of several service points (see, left panel), we can see that when the demand at each of the service points is calculated, the population in question is added two times, and the levels of service are L<sub>1</sub> = L<sub>2</sub> = 1/s00. More generally, when calculating the level of service at L<sub>j</sub>, the population at i contributes to demand every time that d<sub>ij</sub> ≤ d<sub>0</sub> for any j. And, since since D<sub>i</sub> j = P<sub>i</sub>, it follows that the sum of the population to be serviced over all clinics is: $$\sum\limits_{j}D_{ij} = K_{i}P_{i}$$ where K<sub>i</sub> is the number of service points j that include i as part of their catchment areas. Therefore, the system-wide contribution of the population at i to the level of demand implied by these calculations, vastly exceeds the actual population at i, since: $$\sum\limits_{j}D_{ij} = K_{i}P_{i} > P_{i}$$ Let us consider next what happens when enhanced (i.e., non-binary) impedance weights are used. These functions aim to capture more realistically the rule that most members of the population prefer to travel shorter distances to reach a destination. For the example, assume a set of weights with decay as follows (see, right panel): $$W\left( d_{ij} \middle\| d_{1},d_{2},\ldots,d_{R} \right) = \left\{ \begin{array}{ll} {0.9} & { d_{ij} \leq d_{1}} \\ {0.8} & { d_{1} < d_{ij} \leq d_{2}} \\ {0.4} & { d_{R - 1} < d_{ij} \leq d_{R}} \\ 0 & { d_{ij} > d_{R}} \\ \end{array}\operatorname{} \right.$$ The population center in the example is relatively distant from the service points. Accordingly, its potential demand is reduced by assuming that some people do not travel at all. In this example, the contribution of the population center to demand is only 0.8P to each clinic, and therefore the system-wide demand of this center is 1.6P—less than the all-or-nothing allocation of the binary impedance weights, but still in excess of the actual population. More generally, when calculating the level of service at j locations, the population at i contributes to demand every time that d<sub>ij</sub> is within the service area for any j. The precise contribution depends on the weights in the distance-decay function and the position of the population center with relative to all service points. In a function with faster decay, the total demand attributed to i (i.e., ∑<sub>i</sub> D<sub>ij</sub>) can be less than the population of i. In other words, depending on the steepness of decay, the total demand can be greater than, equal to, or less than the population at i: $$\sum\limits_{j}D_{ij} \lessgtr P_{i}$$ Clearly, only when the full population at i is allocated exclusively to one service point (i.e., when K<sub>i</sub> = 1) the implied demand equals the population—something that seldom happens in practical situations. It is important to acknowledge that demand in accessibility analysis represents the potential for spatial interaction, not realized interaction. That said, the expectation that facilities need to serve multiple times the size of the population in a region can easily lead to misleading conclusions about the need for resources. A logical question, however, is whether the inflation of demand (with the consequence deflation of level of service) is not offset in the second step of the method, when the population at i has potential access to multiple service points? Let us consider what happens in the second step of the algorithm in the example, when catchment areas are floated to the population center (see). When a binary impedance function is used, the aggregation of the level of service means that, despite the inflation of demand due to double-counting, accessibility matches the level of service as well as the regional PPR of 2/100 (left panel). In the case of the stepwise function, the level of implied demand is less than the population, but the population is also assumed to receive less of the available level of service. In this case, again, the accessibility matches the level of service despite the fact that segments of the population were assumed to not contribute to demand. Clearly, the example is too simplistic (in fact just a variation of the container approach), and it is unclear what the implications would be for a system with even just a slightly more complex geography. To explore this, consider the addition of two population centers to the landscape. Notice how the three population centers are in the catchment areas of the two clinics. When the binary impedance function is used, demand at each clinic is calculated as 300, and demand over all clinics is therefore 600, or twice the population of the region. When the stepwise impedance function is used, the demand by each center is: $$\begin{array}{r} \begin{matrix} {D_{1j} = 0.8 \times 100 + 0.8 \times 100 = 160} \\ {D_{2j} = 0.8 \times 100 + 0.4 \times 100 = 120} \\ {D_{3j} = 0.4 \times 100 + 0.8 \times 100 = 120} \\ \end{matrix} \\ \end{array}$$ and the total load on the system is therefore 400, still well in excess of the total population of the region. When demand is used to calculate the level of service, and then accessibility in the second step of the algorithm, the following occurs. When the binary impedance function is used (left panel), the level of service at each clinic is: $$\begin{array}{r} \begin{matrix} {L_{1} = \frac{10}{300} = 0.033} \\ {L_{2} = \frac{10}{300} = 0.033} \\ \end{matrix} \\ \end{array}$$ The level of service at the clinics is only half of the regional PPR, since each clinic is assumed to serve the entire population of the region. Unfortunately, since demand has been inflated for each clinic, these levels of service cannot be meaningfully interpreted as local PPRs. The sum over the clinics, on the other hand, is 20/300—which is consistent with the regional PPR. Interestingly, as seen in the figure, the accessibility of each population center matches the regional PPR—but the sum of accessibility over all population centers exceeds the sum of the level of service over all the clinics as a consequence of allocating the same level of service to several population centers. Continuing with the stepwise impedance function, we can see (right panel) that the levels of service are calculated as: $$\begin{array}{r} \begin{matrix} {L_{1} = \frac{10}{0.8 \times 100 + 0.8 \times 100 + 0.4 \times 100} = \frac{10}{200} = 0.05} \\ {L_{2} = \frac{10}{0.8 \times 100 + 0.4 \times 100 + 0.8 \times 100} = \frac{10}{200} = 0.05} \\ \end{matrix} \\ \end{array}$$ Notice how the level of service is higher in this case: this is a consequence of assuming (as the stepwise impedance function does) that some of the population does not demand service. Demand, however, is still inflated, and interpretation of the levels of service as local PPRs is still inappropriate. Accessibility is higher for population center 1 but lower for the two peripheral centers. Furthermore, the sum of accessibility over all population centers exceeds the sum of the level of service of all clinics in the region. At issue is the interpretability of the levels of service, which as the example illustrates do not accurately represent PPRs, and how accessibility, which is a weighted sum of levels of service, cannot be interpreted as the PPR for a population center either. Two methods reviewed above, namely the Three-Stage Floating Catchment Area method and the Modified Two-Stage Floating Catchment Area method aim to address the overestimation of demand and/or levels of service when calculating accessibility. As discussed previously, they do this by compounding the effect of the impedance function. In the case of 3SFCA, demand is deflated by assuming that demand declines more rapidly with distance. Then, when calculating accessibility, the levels of service are allocated more locally, again, as a consequence of steeper distance-decay. In the case of M2SFCA, demand is not deflated, however, the levels of service are allocated more locally as a consequence of steeper distance-decay. In other words, these methods correct for inflation by assuming that fewer people demand helath care services, and that the levels of service are allocated to fewer people too. For comparison, the levels of service and accessibility for the example according to these two methods are shown in. Notice how the levels of service in the 3FSCA are considerably higher as a consequence of excluding potential users with a steeper rate of decay. On the other hand, the levels of accessibility are also lower, as a consequence of allocating service more locally. The levels of service in the M2SFCA are identical to the E2SFCA, however, accessibility is lower, again as a result of allocating service more locally. # A simulated example The examples in the preceding section illustrate the way demand and level of service can be overestimaged (and in some cases underestimated) in FCA algorithms. However, they are too simplistic to indicate what would happen in a realistic situation. In particular, it is possible that the consequences depend on the geography of the problem as the examples in Delamater suggest. Based on the way demand and level of service are allocated, we conjecture that the effects are likely more pronounced in areas with higher density of population and service, since inflation is a consequence of overlapping catchment areas. Furthermore, we conjecture that demand inflation will be reduced when stepwise/continuous distance-decay functions are used, since their effect is to reduce the overlap by reducing the contribution of population at different distances, and to allocate levels of service more locally as well. We explore these issues further by means of a simple but realistic simulated example. The setup for the simulated example is shown in. There are three clinics and nine population centers. Assume that the supply at the three clinics is one physician at clinic 1, three physicians at clinic 2, and two physicians at clinic 3. Further, assume that the population at 1, 2, 8, and 9 is 250; population at 3, 4, and 6 is 250; and population at 5 and 7 is 1000. The total population in the region therefore is 4, 500. Under this setup, the level of service across the whole system is 1.33 physicians per thousand people, which we will refer to as the Regional PPR. For this experiment, we consider binary and stepwise impedance functions. The former is simply the traditional 2SFCA method, whereas the latter is the Enhanced 2SFCA approach. The catchment areas for the first step of the algorithm (demand allocation) are shown in (binary impedance) and (stepwise impedance). Notice that some population centers are inside the catchment areas of more than one clinic. For instance, Population Center 5 is in the catchment areas of Clinics 2 and 3, whereas Population Center 4 is in the catchment areas of all three clinics. To see how the overlap of catchment areas impacts the calculations in the first step of the algorithm, we define impedance matrices using the same criteria as for the buffers seen in Figs and. These matrices are shown in. The demand for each clinic is calculated as the population of the centers multiplied by the values of the corresponding impedance weight with respect to that clinic, and then aggregated for all population centers. The level of service is the supply divided by the demand, multiplied by 1, 000. The last row of the table shows the total population as well as the total demand at each clinic. First we discuss the results according to the binary impedance function. As seen in, the population of Center 3 (which is in the catchment area of three clinics) is assumed to contribute 1, 500 patients to the demand across the system, whereas Center 1 (which is in the catchment area of only one clinic) contributes exactly its population of 250. Since the population of several centers is counted multiple times, the apparent demand exceeds the actual population. In effect, when we calculate the total demand (the sum of the demand across clinics), we find that this is 9,750 according to the binary impedance function, which far exceeds the actual population. Turning now to the stepwise function, we see that Center 3 contributes 500 × 0.242 + 500 × 0.600 + 500 × 0.242 = 542 to the demand across the system, but Center 1 contributes only 250 × 0.242 = 60.5. The total demand now is 4,316, which is less than the total population. This example illustrates a vexing effect in how FCA methods operate: when multiple service points are within the threshold travel cost of a population center, it is assumed that some (and possibly all) of the same persons crowd more than one service point, resulting in inflated demand and deflated levels of service. On the other hand, when stepwise or continuous functions (e.g., E2SFCA) are used to weigh down the population of distant population centers, the apparent effect is that some segments of the population do not demand service, even when clinics are within their threshold travel cost. This effect is even more marked in the case of 3SFCA, which produces considerably higher levels of service, as a consequence of stacking the effects of two impedance functions. In effect, demand is deflated and the level of service is inflated. While the assumption that some members of the population drop out from the total demand pool may be acceptable for discretionary services, it is suspect when it comes to essential services such as many health care services, and particularly primary health care. Recall as well that the Regional Average PPR in this example is 1.33 physicians per thousand. If the total implied demand according to the binary impedance function is 9,750 the corresponding PPR is 0.615 physicians per thousand, or about half of the regional ratio. The corresponding PPR for the stepwise impedance function (implied demand = 4315.5) is 1.39 physicians per thousand, much closer to the Regional Average PPR. However, this PPR is misleading in that it assumes that some segments of the population are served multiple times, and some are not served at all. Clearly, the first step of the algorithm can lead to inflation or deflation of the levels of demand. But do these matter? Or do they somehow average out when the levels of service are aggregated in the second step of the algorithm? Again, the situation is not clear-cut when multiple population centers and/or service clinics interact through overlapping catchment areas. To illustrate this, we proceed to estimate the accessibility for the example using the binary and the stepwise impedance matrices. The results appear in. Accessibility in the table is calculated as the level of service of the clinics multiplied by the the values of the impedance function with respect to a population center, and then aggregated for all clinics. As seen in the table, the levels of accessibility vary considerably depending on the method. As anticipated, use of non-binary impedance functions reduces the inflation effect, and can even lead to deflation. Consider for instance the case of the binary impedance matrix: the total level of service in the system is the sum of the level of service at the three clinics, or 1.87. The level of service allocated to population centers, on the other hand, is the sum of the accessibility in the system, or 11.8. When using the stepwise impedance function, the total level of service in the system is 4.3, and the level of service allocated to population centers is 11.2. Compare this to the case of 3SFCA, where the total level of service in the system is 8.71, but the level of service allocated to population centers is only 2.21; or the case of M2SFCA, which estimates the total level of service in the system as 4.3 (same as E2SFCA) but allocates 5.74 to population centers. Clearly, all the methods give qualitatively similar results, with peripheral centers displaying lower accessibility and more central places higher. But there are important differences in how demand and level of service are allocated throughout the system to calculate accessibilty. shows how the different methods penalize peripheral centers at different rates. And, since the demand is not consistent with the population and the accessibility is not consistent with the level of service of the clinics, it is difficult to interpret the results in terms PPRs. For instance, when we inspect the results for the binary impedance matrix (2SFCA), we can see in the table that the accessibility of Population Center 1 is simply the level of service of Clinic 1. But, as we saw before, this level of service was deflated by double counting the population of Centers 2, 3, and 4, which contribute to the calculation of demand at multiple clinics. Things become more complex as the number of overlapping catchment areas grows. For example, Population Center 2 contributed to the congestion effect of two clinics. However, demand at one of those clinics was calculated using the population of eight out of nine population centers. What this suggests is that, at the very least, some population centers (likely those in the periphery of regions) will have artificially low accessibility levels as a consequence of demand inflation. # A method for proportional allocation of demand and supply As the examples in the preceding subsection illustrate, FCA methods can induce quite substantial inflation (or deflation) of demand and level of service. This, in turn, can affect the estimates of accessibility in potentially complex ways. The results, furthermore lack a clear interpretation. In this section, we propose a simple and intuitive adjustment to avoid the inflation artifacts inherent in current implementations of FCA methods. Refer again to. Demand inflation occurs because of the overlap in catchment areas—with the underlying assumption that a service location services the population within its catchment area. More realistically, only a fraction of that population will demand service at the location if other service points are within reach (i.e., inside its “floated” catchment area). For instance, assuming (as the binary impedance function does), that individuals at Population Center 1 are indifferent between Clinics 1 and 2, then it is reasonable to think that the population will sort itself proportionally to these two clinics—in this example, this means that half of the population will attend one of two different clinics (importantly, this assumes that the services on offer are undifferentiated; one would not generally consider cancer screening and hair removal clinics competitors). This suggests the following adjustment to the way the level of demand is calculated. Given an impedance function, a set of adjusted weights, say $W_{ij}^{i}$, are precalculated by dividing the original impedance weights by the sum of the weights for population center i* over all service points j: $$W_{ij}^{i} = \frac{W_{ij}}{\sum_{j}W_{ij}}$$ Please notice that these weights are identical to the selection weights of the 3SFCA method. A key property of the adjusted weights is the following: $$\sum\limits_{j}W_{ij}^{i} = 1$$ This adjustment procedure has the effect that, when the level of demand of i is summed over all service points j, the aggregated level of demand due to i is identical to its population: $$\sum\limits_{j}P_{i}W_{ij}^{i} = P_{i}$$ As a result of standardizing the impedance weights, population is allocated proportionally to clinics. On the supply side, inflation happens because the level of service available at location j is assumed to be available to every population center i within its catchment area. To adjust this, another set of weights, say $W_{ij}^{j}$, is pre-calculated by dividing the original impedance weights W<sub>ij</sub> by the sum of the weights for service point j* over all population centers i: $$W_{ij}^{j} = \frac{W_{ij}}{\sum_{i}W_{ij}}$$ Again, the resulting weights have the property that: $$\sum\limits_{i}W_{ij}^{j} = 1$$ As before, the result of this procedure is that, when the level of service of j is aggregated by population centers, the total level of service for that service point is preserved: $$\sum\limits_{i}L_{j}W_{ij}^{j} = L_{j}$$ Note that, since the weights add up to one, they can be interpreted as a probability or frequency of contact, similar to the Huff model of. In reference to (left panel), we can see that the original binary (unadjusted) weights for Population Centre 1 are W<sub>11</sub> = 1, W<sub>12</sub> = 1, the weights of population center 2 are W<sub>21</sub> = 1, W<sub>22</sub> = 1, and the weights of population center 3 are W<sub>31</sub> = 1, W<sub>32</sub> = 1. On the demand side, the adjusted weights become for Population Center 1, $W_{11}^{i} = 1/2$, $W_{12}^{i} = 1/2$, for Population Center 2 $W_{21}^{i} = 1/2$, $W_{22}^{i} = 1/2$, and for Population Center 3 $W_{31}^{i} = 1/2$, $W_{32}^{i} = 1/2$. Using the adjusted weights, it can be seen that the level of demand due to each population center equals its respective population: $$\begin{array}{r} \begin{array}{l} {\sum\limits_{j}D_{1}j = 1/2P_{1} + 1/2P_{1} = P_{1}} \\ {\sum\limits_{j}D_{2}j = 1/2P_{2} + 1/2P_{2} = P_{2}} \\ {\sum\limits_{j}D_{3}j = 1/2P_{3} + 1/2P_{3} = P_{3}} \\ \end{array} \\ \end{array}$$ Coming next to the supply side, the adjusted weights for Clinic 1 are $W_{11}^{j} = 1/3$, $W_{21}^{i} = 1/3$, and $W_{23}^{i} = 1/3$; for Clinic 2 the adjusted weights are $W_{12}^{j} = 1/3$, $W_{22}^{i} = 1/3$, and $W_{32}^{i} = 1/3$. It can be seen that the level of service is preserved across clinics, and therefore across the system: $$\begin{array}{r} \begin{array}{l} {\sum\limits_{i}L_{i1} = L_{1}/3 + L_{1}/3 + L_{1}/3 = L_{1}} \\ {\sum\limits_{i}L_{i2} = L_{2}/3 + L_{2}/3 + L_{2}/3 = L_{2}} \\ \end{array} \\ \end{array}$$ The method to adjust the weights used above is identical to a procedure that will be familiar to readers acquainted with the literature in the fields of spatial statistics and econometrics. The same adjustment is widely used there under the names of row- and column-standardization of a weights matrix. The proposed adjustment can be easily implemented. We will present next the implementation using a compact matrix notation. Begin by defining the following impedance matrix: $$\begin{array}{r} {\mathbf{W} = \begin{pmatrix} W_{11} & \cdots & W_{1J} \\ \vdots & \ddots & \vdots \\ W_{N1} & \cdots & W_{NJ} \\ \end{pmatrix}} \\ \end{array}$$ where W<sub>ij</sub> is an impedance function evaluated at d<sub>ij</sub>. Subindex i is for population centers (i = 1, …, N) and subindex j is for service points (j = 1, …, J). Note that the matrix does not need to be square. A row-standardized set of weights is obtained as follows: $$\begin{array}{r} {\mathbf{W}^{i} = \begin{pmatrix} \frac{W_{11}}{\sum_{j}W_{1j}} & \cdots & \frac{W_{1J}}{\sum_{j}W_{1j}} \\ \vdots & \ddots & \vdots \\ \frac{W_{N1}}{\sum_{j}W_{Nj}} & \cdots & \frac{W_{NJ}}{\sum_{j}W_{Nj}} \\ \end{pmatrix}} \\ \end{array}$$ Next, a column-standardized set of weights is calculated as: $$\begin{array}{r} {\mathbf{W}^{j} = \begin{pmatrix} \frac{W_{11}}{\sum_{i}W_{i1}} & \cdots & \frac{W_{1J}}{\sum_{i}W_{iJ}} \\ \vdots & \ddots & \vdots \\ \frac{W_{N1}}{\sum_{i}W_{i1}} & \cdots & \frac{W_{NJ}}{\sum_{i}W_{iJ}} \\ \end{pmatrix}} \\ \end{array}$$ In the first example above, the binary impedance matrix is: $$\begin{array}{r} {\mathbf{W}_{binary} = \begin{pmatrix} 1 & 1 \\ 1 & 1 \\ 1 & 1 \\ \end{pmatrix}} \\ \end{array}$$ The row-standardized weights that correspond to this matrix are: $$\begin{array}{r} {\mathbf{W}_{binary}^{i} = \begin{pmatrix} {1/2} & {1/2} \\ {1/2} & {1/2} \\ {1/2} & {1/2} \\ \end{pmatrix}} \\ \end{array}$$ and the column-standardized weights are: $$\begin{array}{r} {\mathbf{W}_{binary}^{j} = \begin{pmatrix} {1/3} & {1/3} \\ {1/3} & {1/3} \\ {1/3} & {1/3} \\ \end{pmatrix}} \\ \end{array}$$ The stepwise impedance weights in the example are: $$\begin{array}{r} {\mathbf{W}_{stepwise} = \begin{pmatrix} {0.8} & {0.8} \\ {0.8} & {0.4} \\ {0.4} & {0.8} \\ \end{pmatrix}} \\ \end{array}$$ The row-standardized weights in turn are: $$\begin{array}{r} {\mathbf{W}_{stepwise}^{i} = \begin{pmatrix} {1/2} & {1/2} \\ {2/3} & {1/3} \\ {1/3} & {2/3} \\ \end{pmatrix}} \\ \end{array}$$ whereas the column-standardized weights are: $$\begin{array}{r} {\mathbf{W}_{stepwise}^{j} = \begin{pmatrix} {4/10} & {4/10} \\ {4/10} & {2/10} \\ {2/10} & {4/10} \\ \end{pmatrix}} \\ \end{array}$$ Once that the impedance weights have been adjusted, a vector of adjusted level of demand D\* can be obtained by multiplying the transposed impedance matrix by a vector of population values as follows: $$\mathbf{D}^{} = \left\lbrack \mathbf{W}^{i} \right\rbrack^{T}\mathbf{P}$$ where the <sup>T</sup> operator is for “transpose”, and P* is: $$\begin{array}{r} {\mathbf{P} = \begin{pmatrix} P_{1} \\ \vdots \\ P_{N} \\ \end{pmatrix}} \\ \end{array}$$ The level of demand for the service points in the binary impedance function example is (in vector form): $$\begin{array}{r} {\mathbf{D}_{binary}^{} = \begin{pmatrix} {1/2} & {1/2} & {1/2} \\ {1/2} & {1/2} & {1/2} \\ \end{pmatrix}\begin{pmatrix} 100 \\ 100 \\ 100 \\ \end{pmatrix} = \begin{pmatrix} {300/2} \\ {300/2} \\ \end{pmatrix} = \begin{pmatrix} 150 \\ 150 \\ \end{pmatrix}} \\ \end{array}$$ Notice how each clinic is expected to service only 150, and the level of demand over the system is identical to the total population. The level of demand for the service points in the stepwise impedance function example is (in vector form): $$\begin{array}{r} {\mathbf{D}_{sw}^{} = \begin{pmatrix} {1/2} & {2/3} & {1/3} \\ {1/2} & {1/3} & {2/3} \\ \end{pmatrix}\begin{pmatrix} 100 \\ 100 \\ 100 \\ \end{pmatrix} = \begin{pmatrix} {50 + 200/3 + 100/3} \\ {50 + 100/3 + 200/3} \\ \end{pmatrix} = \begin{pmatrix} 150 \\ 150 \\ \end{pmatrix}} \\ \end{array}$$ As can be seen, the aggregated level of demand, after the adjustment, equals (as desired) the actual population of the region. In the case of the stepwise function, total demand has been adjusted to the population of the region without the restrictive assumption that some people are excluded from the system. This is achieved by assuming an assortative process that leads to proportional allocation of the demand. The levels of demand can then be used to calculate the level of service at the individual clinic locations by performing Hadamard division (⊘) of the vector of supply by the vector of adjusted demand. This is the first step of the 2SFCA (aggregating demand over catchment areas for service points): $$\mathbf{L}^{} = \mathbf{S} \oslash \mathbf{D}^{}$$ Since Hadamard division is an element-by-element operation, the adjusted levels of service in the first example (using the binary impedance function) are: $$\begin{array}{r} {\mathbf{L}_{b}^{} = \begin{pmatrix} 10 \\ 10 \\ \end{pmatrix} \oslash \begin{pmatrix} 150 \\ 150 \\ \end{pmatrix} = \begin{pmatrix} {10/150} \\ {10/150} \\ \end{pmatrix} = \begin{pmatrix} {0.067} \\ {0.067} \\ \end{pmatrix}} \\ \end{array}$$ The levels of service in the second example, when using the stepwise impedance function, are also: $$\begin{array}{r} {\mathbf{L}_{sw}^{} = \begin{pmatrix} 10 \\ 10 \\ \end{pmatrix} \oslash \begin{pmatrix} 150 \\ 150 \\ \end{pmatrix} = \begin{pmatrix} {10/150} \\ {10/150} \\ \end{pmatrix} = \begin{pmatrix} {0.067} \\ {0.067} \\ \end{pmatrix}} \\ \end{array}$$ Unlike the 2SFCA, E2SFCA, and 3SFCA methods that produce levels of service that resemble PPRs but with values that are inconsistent with total demand given the population, this operation returns values that are genuinely local PPRs that are consistent with the population of the region. As we saw above, the demand equals the population. Here, the supply also equals the number of physicians in the region. Because both demand and supply are not inflated or deflated in this rectified method, these values are easily interpretable relative to the Regional Average PPR of 20/300 or 0.067 physicians per person. In the case of the example, it is clear that both clinics have PPRs that are identical to the Regional Average PPR. Accessibility, finally, is calculated as the matrix product of the column- standardized weights and the adjusted level of service: $$\mathbf{A}^{} = \mathbf{W}^{j}\mathbf{L}^{}$$ which, continuing with the example, gives the following for the binary impedance function: $$\begin{array}{r} {\mathbf{A}_{b}^{} = \begin{pmatrix} {1/3} & {1/3} \\ {1/3} & {1/3} \\ {1/3} & {1/3} \\ \end{pmatrix}\begin{pmatrix} {10/150} \\ {10/150} \\ \end{pmatrix} = \begin{pmatrix} {10/450 + 10/450} \\ {10/450 + 10/450} \\ {10/450 + 10/450} \\ \end{pmatrix} = \begin{pmatrix} {0.044} \\ {0.044} \\ {0.044} \\ \end{pmatrix}} \\ \end{array}$$ Notice how the sum of accessibility over the region is consistent with the total level of service over all clinics (i.e., 0.133). The level of service has been allocated in its totality. When using the stepwise impedance function, accessibility is calculated as: $$\begin{array}{r} {\mathbf{A}_{sw}^{} = \begin{pmatrix} {4/10} & {4/10} \\ {4/10} & {2/10} \\ {2/10} & {4/10} \\ \end{pmatrix}\begin{pmatrix} {10/150} \\ {10/150} \\ \end{pmatrix} = \begin{pmatrix} {4/150 + 4/150} \\ {4/150 + 2/150} \\ {2/150 + 4/150} \\ \end{pmatrix} = \begin{pmatrix} {0.053} \\ {0.040} \\ {0.040} \\ \end{pmatrix}} \\ \end{array}$$ Again, the sum of accessibility is consistent with the level of service available from all clinics in the region. As with the Local PPRs, accessibility is interpreted as population-to-provider ratios for each population center in such a way that all calculations are with total demand and total level of service. In particular, accessibility can be interpreted as the share of level of service that a population center receives from all the clinics that service it. For the sake of comparison, levels of service and accessibility are reported for the simulated example in Tables and. An important point to remark is the following. The use of row- and column- standardized impedance weights assumes that the full population of every population center within the catchment of a clinic will receive service. However, the allocation, although proportional, is different when binary or stepwise impedance weights are standardized. When binary weights are employed, the underlying idea is that potential for use is identical within the catchment area irrespective of distance. When stepwise weights are used, proportionally more of the population is allocated to closer clinics. Depending on the definition of cost of travel, this allows a research to accommodate directional effects as well. For example, use of network travel time would tend to favor movement away from congested locations. ## Suboptimal systems The research of Delamater illustrates how accessibility estimates can be misleading when systems are not optimally configured. We understand this to mean that some population centers are located too far away from service points to actually benefit from them. In the modified 2SFCA method (M2SFCA), Delamater addresses this issue by increasing the friction of distance. A slight inconsistency in this approach is that some of the centers that contribute to demand fail to benefit from the service due to the increased friction to which the allocation of the level of service is subjected. Our suggestion in the case of suboptimal systems is to use an impedance function that reflects limiting conditions. For instance, in urban settings a travel time longer than 2 hours might be considered too long to be serviced by any clinic. ## System efficiency The approach proposed in this paper allocates population and level of service proportionally and exactly. This assumes that the population sorts itself into clinics in the most efficient way. But what if some members of the population lack full information about the spatial distribution of clinics? Or have some bias towards centric locations? The vagaries of human behavior could create excess demand in some locations, and as a consequence supply surpluses in others. Situations like this can be accommodated in a relatively straightforward way using our approach. Here, we describe the use of slack factors. Demand and level of service are allocated proportionally and exhaustively (i.e., 100%). But the standardization could allow for some slack, by inflating demand and/or supply in a controlled way. Our proposal to standardize the weights was as follows, for the case of rows and columns respectively: $$\begin{array}{r} {\mathbf{W}^{i} = \begin{pmatrix} \frac{W_{11}}{\sum_{j}W_{1j}} & \cdots & \frac{W_{1J}}{\sum_{j}W_{1j}} \\ \vdots & \ddots & \vdots \\ \frac{W_{N1}}{\sum_{j}W_{Nj}} & \cdots & \frac{W_{NJ}}{\sum_{j}W_{Nj}} \\ \end{pmatrix}\text{and}\mathbf{W}^{j} = \begin{pmatrix} \frac{W_{11}}{\sum_{i}W_{i1}} & \cdots & \frac{W_{1J}}{\sum_{i}W_{iJ}} \\ \vdots & \ddots & \vdots \\ \frac{W_{N1}}{\sum_{i}W_{i1}} & \cdots & \frac{W_{NJ}}{\sum_{i}W_{iJ}} \\ \end{pmatrix}} \\ \end{array}$$ A set of slack factors, say $k_{i}^{i}$, could be introduced in the following manner: $$\begin{array}{r} {\mathbf{W}^{i} = \begin{pmatrix} \frac{k_{1}^{i}W_{11}}{\sum_{j}W_{1j}} & \cdots & \frac{k_{1}^{i}W_{1J}}{\sum_{j}W_{1j}} \\ \vdots & \ddots & \vdots \\ \frac{k_{N}^{i}W_{N1}}{\sum_{j}W_{Nj}} & \cdots & \frac{k_{N}^{i}W_{NJ}}{\sum_{j}W_{Nj}} \\ \end{pmatrix}} \\ \end{array}$$ A value of $k_{1}^{i} = 1.10$ would inflate the demand of population center i = 1 by 10%, whereas a value of $k_{1}^{i} = 1.20$ would inflate the demand by 20%. In a similar way, a set of slack factors $k_{i}^{j}$ could be introduced to modulate the allocation of supply: $$\begin{array}{r} {\mathbf{W}^{j} = \begin{pmatrix} \frac{k_{1}^{j}W_{11}}{\sum_{i}W_{i1}} & \cdots & \frac{k_{J}^{j}W_{1N_{J}}}{\sum_{i}W_{iJ}} \\ \vdots & \ddots & \vdots \\ \frac{k_{1}^{j}W_{N1}}{\sum_{i}W_{i1}} & \cdots & \frac{k_{J}^{j}W_{NJ}}{\sum_{i}W_{iJ}} \\ \end{pmatrix}} \\ \end{array}$$ A value of $k_{1}^{j} = 0.9$, for example, would deflate the supply of clinic j = 1 by 10%. The use of slack factors provides an interesting way of modulating demand and level of service allocation in a very precise and controlled way, and presents interesting opportunities as well to introduce expert opinion or other empirical approaches to callibrate slack factors. # Empirical example In the reminder of the paper we present an empirical example to illustrate the application of the methods presented above. Based on the preceding discussion, the adjusted 2SFCA method employed can be summarized as: $$L_{j} = \sum\limits_{i}\frac{S_{j}}{P_{i}W_{ij}^{i}}$$ with row-standardized impedance weights $W_{ij}^{i}$ in the first step, and: $$A_{i} = \sum\limits_{j}{L_{j}W_{ij}^{j}}$$ with colum-standardized impedance weights $W_{ij}^{j}$ in the second step. The same approach is used to re-weight the impedance function for the stepwise approach (i.e., E2SFCA). The case study is based on accessibility to family physicians in the Hamilton Census Metropolitan Area (CMA), in Ontario, Canada. For this, we use data collected about the distribution of the population and primary health care clinics (i.e., family physicians) in the region. Time use data from Canada’s General Social Survey (GSS) was also used to inform the selection of thresholds for the impedance functions. The data collection and preprocessing protocols are described next. ## Family physicians and clinic locations In regards to the supply of clinics, the locations of family physicians were obtained using the College of Physicians and Surgeons of Ontario (CPSO) database for the Province of Ontario. We chose this organization beacuse all physicians practicing in Ontario are required to register with the CPSO, as set out in the Ontario Regulation 865/93: Registration. Our search of CPSO’s database was conducted attending to the following criteria. 1. Only physicians who are registered as family physicians were selected (this excluded specialists such as pediatric physicians). 2. The spatial extent of the search was determined using forward sortation areas (FSAs), which are the first three initial characters of a postal code. Using a GIS, the regions of interest were selected by choosing FSAs within a 10 kilometer buffer distance from the Hamilton CMA boundary. This involved 72 different FSA regions. Each FSA region code was then searched in the CPSO database in addition to the family physician specification. The parameters of the search were deliberately conservative, and the search identified a total of 2,224 family physicians practicing in the region, of which, 864 are located in the Hamilton CMA. The resulting dataset was manually verified by the third author to ensure that the information was consistent and suitable for geocoding. Prior to geocoding, locational information was organized in three columns, containing street address, city name, and province name. After family physicians were geocoded, locations were further examined. When family physicians overlapped or were within a 50 meter distance of each other we merged the records to identify 535 unique locations that we term “clinics”. Many of these clinics are not in the Hamilton CMA proper, but provide a buffer to minimize edge effects in the analysis. The distribution of clinics and family physicians is shown in for the Hamilton CMA. ## Population Population information was obtained from the 2011 Canadian Census. To maximize the spatial resolution, population data were acquired at the Dissemination Area (DA) level of geography for all DAs within the selected FSAs. As a result, this includes DAs not in the Hamilton CMA proper, but that provide a buffer against edge effects. From this, the region contains a population of 2,959,090, of which 720,725 are in the Hamilton CMA. The distribution of population in the Hamilton CMA is shown in. ## Travel time matrix Calculation of impedance weights requires that we evaluate an impedance function at values of d<sub>i</sub> j, that is, the cost of travel between DA i and clinic j. In this research we used travel time as our cost variable. To this end, we computed a matrix of travel times measured over the road network. To calculate the travel time between population centers and clinics we used the DA centroids and the geocoded locations of clinics. Shortest paths on the network and subsequently travel times were computed using a Geographic Information System. ## Impedance functions For the experiments we use two different impedance functions, corresponding to the 2SFCA and E2SFCA algorithms. We do not implement the 3SFCA or the M2SFCA methods because, as noted above, they are equivalent to using steeper impedances. For the 2SFCA apprach, impedance is given by a binary function, whereas for E2SFCA it is given by a stepwise function. The impedance functions require that we define cost (i.e., travel time) thresholds to implement them. To select the thresholds, we retrieved time use data from Canada’s General Social Survey Cycle 24 (see <http://odesi2.scholarsportal.info/webview/>). From the time use files, we filtered all activity episodes corresponding to respondents living in CMAs/CAs (metropolitan regions) in Ontario. Next, we filtered all episodes taking place in a car (as driver) while traveling for personal care activities for household adults (which includes traveling to see a doctor) and traveling for shopping or obtaining services (which includes traveling to go to health clinic or doctor’s office). It is worthwhile noting that travel by car accounts for over 95% of trips for the selected purposes in Ontario CMAs/CAs. Once episodes were filtered by mode of travel and purpose of the trip, their durations (in minutes) were examined by means of quantile analysis, using episode weights to ensure the representativeness of the analysis. From the results, we learned that 50% of all trips by car for the aforementioned purposes are less than 15 minutes long, and we selected this value as the threshold d<sub>0</sub> for the binary function. In other words, this part of the analysis assumes that any person who has to travel longer than 15 minutes to reach a clinic is outside its catchment area. We deem this value appropriate for the scale, density, and level of congestion of Hamilton CMA. Quantile analysis of trip durations was also used to calibrate a Gaussian function with standard deviation set at 15 minutes, to match the value selected for the binary impedance above. This produced the following stepwise function, with any trips longer than 45 minutes assumed to be outside of catchment: $$W\left( d_{ij} \right) = \left\{ \begin{array}{ll} {0.946} & { d_{ij} \leq 5} \\ {0.801} & { 5 < d_{ij} \leq 10} \\ {0.607} & { 10 < d_{ij} \leq 15} \\ {0.411} & { 15 < d_{ij} \leq 20} \\ {0.135} & { 20 < d_{ij} \leq 30} \\ {0.011} & { 30 < d_{ij} \leq 45} \\ {0.00} & { 45 < d_{ij}} \\ \end{array}\operatorname{} \right.$$ Notice how the stepwise function has weights greater than 0.5 for d<sub>ij</sub> ≤ 15min and less than 0.5 for d<sub>ij</sub> \> 15min. This means that it will count fewer people than the binary function when d<sub>ij</sub> ≤ 15min, but more when d<sub>ij</sub> \> 15min. ## Results We begin our discussion of the results by noting that with a total population of the region of 2,959,090 and 2,222 family physicians, the Regional Average PPR ratio is 0.751 family physicians per 1,000 people. This value is somewhat lower than the value of 1.16 for Ontario reported by CIHI and lower than the 1.20 estimated based on the population and physician data for the Hamilton CMA, which we attribute to our conservative search criteria of family physicians in the rest of the region. The nominal levels of demand, service, and accessibility are calculated for the 2SFCA and E2SFCA using both the unadjusted and adjusted impedance matrices. summarizes the results by each impedance matrix. As seen there, when no adjustment is made, the nominal demand explodes to several times the actual population in the region. However, when the impedance weights are standardized, demand is now only slightly less than the total population for the region, since the system is not optimal in the sense discussed by Delamater, and a small proportion of the population turns out to be outside of catchment. The nominal demand under binary impedance is lower due to the stricter catchment area condition (i.e., less than 15 minutes), compared to the stepwise function (i.e., less than 45 minutes). This, in turn, is somewhat lower than the total demand in the Regional Average PPR, which does not impose catchment area constraints within the region. It is clear that the rectified demand leads to results that are considerably more realistic than the conventional approaches. In addition to the nominal system-wide demand, this is seen as well when calculating the regional provider- to-population ratios for each case (i.e., Family Physicians per 1,000 people). As seen in the table, the mean levels of service for clinics in the region in the case of the adjusted binary and stepwise weights are in line with their corresponding Regional Average PPRs. Since the levels of service in the case of the adjusted weights can be interpreted as local PPRs, this indicates that the average clinic offers approximately the same level of service as the regional system does for the whole population. Furthermore, the mean accessibility of a DA according to the adjusted weights is identical to the mean LOS: this is because the LOS is allocated completely to DAs. The total LOS and accessibility in the region match when the adjusted weights are used. This is not the case when the unadjusted weights are used. Clearly, the use of the unadjusted weights can lead to a substantial amount of accessibility inflation, by factors as high as five or six times the estimates of the proposed proportional allocation approach. These results demonstrate how inflation of the supply (i.e., the level of service) leads to much higher values of accessibility in the case of the conventional 2SFCA and E2SFCA methods. The procedure to rectify the population and level of service, on the other hand, leads to accessibility outputs that are consistent with the regional population and overall supply of health care services. This, in turn, makes interpretation of the output more robust and intuitive. Another important issue is that spatial distribution of inflation of demand and level of service. If inflation happened in a uniform way, the upward bias in the estimates could to some extent be ignored, as long as relative differences by location remained relatively constant. Unfortunately, as seen in Figs and, demand inflation is far from uniform. In fact, inflation of demand tends to happen, as per our earlier conjecture, in areas with higher population density. Inflation factors are also substantially higher when the binary impedance function is used. Since this function lacks a gradual distance-decay mechanism, it is more generous in terms of counting population serviced. Notice the magnitude of the inflation factors: since the inflation of demand depends on the number of overlapping catchment areas, a factor of 160, for instance, would suggest that a clinic is expected to simultaneously serve approximately that number of DAs in the conventional 2SFCA method, and a proportionally similar number in the conventional E2SFCA method. The map of accessibility for the implementation of 2SFCA is shown in and with the adjusted weights for proportional allocation in. The general patterns observed in the figures are as expected, with higher accessibility in denser, better connected parts of the region. Relatively high accessibility in the north and west of the CMA is due to proximity to other major population centers such as Oakville, Kitchener, and Waterloo. A question, however, is the degree of inflation of accessibility in the original 2SFCA? plots the ratio of the binary and adjusted binary accessibility measures. Here it can be seen that the unadjusted accessibility values are at least three times greater than their adjusted counterparts within the study area. This inflation, moreover, is not uniform across space, with inflation of the binary accessibility values up to 8 times greater than those from the adjusted model at the edges of the city where the 15-minute catchment areas begin to overlap with neighboring municipalites. Why is this important? As noted by various authors, in traditional FCA methods, the sum of the population-weighted average of accessibility across all population centers is equal to the regional average provider-to-population ratio. In the present case, the weighted sum of accessibility in the unadjusted binary and stepwise measures is 0.751. However, while this value is indeed identical to the regional average provider-to-population ratio, it is problematic because the share of the population correlates poorly with the pattern of inflation observed. The key issue here is that accessibility is deflated by the share of the population in a DA i; however, the degree of inflation of demand and supply depends not only of the population DA i, but on the population of every DA j with which DA i interacts via overlapping catchment areas. As a consequence, deflating accessibility using population shares in previous FCA methods does not accurately offset demand and supply inflation. Figs and present the results for the stepwise E2SFCA with and without the rectification. The results are qualitatively similar to the 2FSCA, with the expected differences. The inflation factors are even more substantial, given the larger catchment areas used. ## Disparity analysis An advantage of the use of adjusted weights for proportional allocation of demand and level of service it that, after rectifying the inflation artifact, they make it is possible to conduct accessibility disparity analysis in a very intuitive way. For instance, an analyst interested in equity analysis could allocate the total level of service uniformly to every DA. In other words, the total level of service (which equals the sum of accessibility over the system) can be divided by the number of population centers in the system to return the Average Local Population Center PPR. The resulting mean value, call it $L_{i}^{e}$ then would be assigned to the population centers as their “equitable” share of the total level of service in the system. Next, the equitative distribution of the level of service in each population center is substracted from the estimated mean accessibility to arrive at a disparity index. When the difference between these two quantities is positive, this would indicate that a DA’s accessibility exceeds its equitable share of level of service. On the other hand, when the difference is negative, the DA’s accessibility is below its equitable share of the level of service. This approach is reminiscent of the Spatial Access Ratio (SPAR) proposed by Wan et al., which is calculated as the ratio between a population center’s accessibility and the mean accessibility across all population centers. Wan et al. calculate SPAR based on the results of their 3SFCA method, by rescaling the accessibility measures to reflect the percentage difference in each population center’s accessibility relative to the mean. This measure is designed to overcome the sensitivity of existing FCA metrics to the impedance function. In contrast, the approach proposed here, enables more intuitive and interpretable results by preserving the system-wide population and level of service. In this way, a disparity index is useful to highlight the absolute difference in accessible provider-to-population ratios across population centers. Disparity maps for the adjusted binary and stepwise impedance functions are shown in Figs and. These figures reveal the spatial distribution in disparity, with levels of access that are lower than the mean in more rural parts of the city (where travel times are longer and the distribution of physicians is more spatially disperse) compared to levels of access that are greater than the mean in the higher-density and more connected urban center. # Conclusion Accessibility to healthcare is an issue of continued interest in health geography. One of the most popular approaches to estimating accessibility is the 2SFCA method and its associated family of FCA models due to their simplification of more complex gravity models and their interpretation as proxies for provider- to-population ratios. These properties make FCA approaches particularly appealing for health policy. In this paper, we have argued that the overestimation of demand and level of service in FCA approaches poses a challenge to the interpretation of accessibility and the identification of spatial disparities in access, with potentially deleterious consequences for policy analysis. The issue of overestimation of demand and level of service has been recognized before, notably by Wan et al. and Delamater, and alternative approaches have been proposed that seek to offset or reduce the problem. Nevertheless, the present paper has shown that the inflation of demand is present in all existing FCA methods. Moreover, we also show that in some cases, demand is deflated, and detail the potential for inflation/deflation on the supply side. To overcome these issues, we draw from the fields of spatial statistics and econometrics, to incorporate row-standardized impedance weights in the calculation of demand, and column-standardized impedance weights to adjust the level of service. These adjustments ensure that allocation of demand and level of service are proportional. As a result, both the system-wide population and level of service are preserved in the estimation of accessibility. The case study in Hamilton CMA reveals the extent of inflation in accessibility inherent in the unadjusted approaches compared to the adjusted binary and stepwise FCA methods. Furthermore, the adjustments result in local provider-to- population ratios which can be easily understood relative to the system-wide equitable level of service through the calculation of a disparity index. The applicability of these values is particularly enhanced by the use of a travel survey to inform the estimated impedance functions. Taken together, these innovations provide estimates of spatial accessibility and disparity that are robust to the regional distribution of supply and demand, as well as observed travel behaviour. By extension, these properties mean that the adjusted approach employed here can offer more rigorous recommendations for health policy. Finally, 1) we proposed a set of slack factors to modulate the estimates of demand and/or level of supply to account for system inefficiencies; and 2) demonstrated the use of a disparity index to conduct equity analysis. In conclussion, the research presented in this paper demonstrates how a relatively simple adjustment of the impedance weights can help to overcome the inflation/deflation issue inherent in previous FCA approaches. By incorporating these methods into the estimation of accessibility to healthcare services, future research can help to ensure that the FCA approach continues to live up to its promise as an intuitive and policy-relevant method for investigating access and disparity. # Replicable research This research was conducted and the paper prepared using `R` and RStudio, along with the following packages: `tidyverse`, `knitr`, `rgdal`, `sf`, `gridExtra`, `raster`, `readr`, `kableExtra`, `ggthemes`, `ggrepel`. Data and source code used in this research are available at <https://github.com/paezha/Demand-and- Supply-Inflation-in-Floating-Catchment-Area-FCA-Methods->. [^1]: The authors have declared that no competing interests exist.
# Introduction Vocal nodules (VN) belong to the most common benign vocal cord disease, and is characterized by bilateral nodular protrusions in the anterior 1/3 of the vocal cords and voice disorder. Researches have suggested potential associations between VN and phonetic trauma/abuse, while their exact etiopathogeneses remain unknown. Due to hoarseness, vocal fatigue and pharyngeal discomfort, VN decreases patients’ quality of life, physiologically and socially. According to a Korean epidemiological report, the incidence of VN was 0.99%-1.72%, while it was more frequent in male children and adult female, which was same as a Turkish survey (diagnosed in 30.3%, 187 cases, of school age children). Formally, therapies for VN are divided into surgical and non-surgical approaches, and specific superiority or inferiority concerning any of them were still not evaluated in a Cochrane systematic review. However, clinically, non- surgical therapies are more selected and better preferred by medical staff and patients, which may owe to its non-invasive nature. Among them, evidence showed that voice training (VT) is beneficial in improving voice quality, while it was not popular in most of the world due to insufficient (or without) speech pathology specialty and limited medical resources. In addition to VT and other non-surgical therapies, some VN patients have been turning to complementary and alternative therapies, acupuncture and Chinese herbal medicine (CHM) especially, for better efficacy, longer duration and fewer costs. In traditional Chinese medicine (TCM), VN lies in Qi stagnation, phlegm, fluid retention, and blood stasis, which are resulted from unbalanced visceral function, and gather in vocal cord. Originating about three thousand years ago, acupuncture could regulate Yin, Yang, and visceral function based on the principles of meridians and acupoints of TCM, as well as circulate qi and blood. As a frequently used adjuvant, trials suggested potential benefits of acupuncture for benign nodular/hyperplasia diseases, such as benign prostatic hyperplasia, hyperplasia of mammary glands, and thyroid nodule. However, no systematic review has been published to explore efficacy and safety of acupuncture for VN. The aim of this systematic review and meta-analysis is to fill the vacancy above with rigorous design and comprehensive analyses, and incorporate new evidence about acupuncture for VN. In addition, potential variations between studies, quality of outcomes and strength of evidence recommendations were also measured for better clinical application. # Material and methods ## Protocol and registration This systematic review was registered in PROSPERO with the registration number CRD42022350916 (available from <https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=350916>). ## Search strategy Four English databases, including PubMed, Embase, Cochrane Library, Web of Science, and four Chinese databases, including Chinese Biomedical Literature Database, VIP Database for Chinese Technical Periodicals, China National Knowledge Infrastructure, and Wanfang Database, were searched from inception until December 10<sup>th</sup>, 2022. The PRISMA agreement was followed in decision of search strategy and inclusion criteria. Two subsets of terms were searched with the term ‘AND’, including the experimental intervention (‘acupuncture’, ‘needle’, ‘electro-acupuncture’, ‘electropuncture’, ‘electroacupuncture’, ‘acusector’, and ‘acupoint’) and terms of the disease (‘vocal nodule’, ‘behavioral voice disorder’, ‘voice disorder’, ‘voice’, ‘vocal cord’, ‘hoarseness’, ‘speech disorder’). Two authors processed the searches independently, and we also searched the references of the original and review studies manually for trials. ## Inclusion criteria 1\) Trials in which participants were diagnosed with vocal nodules by laryngoscopy, and there was no limitation on type of laryngoscopy and vocal nodules; 2\) Prospective randomized controlled trials (RCTs). 3\) Trials in which acupuncture (manual or electro type) was applied as the only therapy in experimental groups. The participants in control groups received sham acupuncture, blank (wait-list) controls, western medicine (WM), TCM, VT, etc.. There was no restriction on duration, frequency, acupoint, or stimulation of acupuncture, while trials with other acupoint-based therapy (e.g., acupressure, moxibustion, acupoint injection, catgut embedding) designed as the control were excluded. 4\) Primary outcomes included clinical effective rate (CER). Secondary outcomes included scales on objective sound quality and scores of subjective symptoms. 5\) Trials published in Chinese or English. ## Study selection and data extraction Two reviewers (RZ and QF) searched the online databases listed above and recorded the titles and abstracts of all the articles. Two evaluators (XL and YL) assessed the eligibility of these articles and made decisions on every research (inclusion or exclusion) independently. If they did not reach the same decision, the concerned articles were discussed with a fourth reviewer (ZX). Two reviewers (QP, and XX) extracted data independently from each study. Differences of extracted data were solved after discussion with a third reviewer (LZ). ## Quality assessment Quality assessment of all the trials included in this review was independently evaluated by three reviewers (QF, XL and RZ) using the Cochrane Collaboration risk of bias tool by RevMan 5.4 software. Any disagreement was resolved by discussions with a fourth reviewer (ZX or LZ). ## Statistical analysis Statistical analyses were measured with RevMan V.5.4, Stata V.14 and TSA 0.9 software. Effect sizes were determined as weighted mean difference (WMD) or standard mean difference for continuous outcomes, and risk ratio (RR) for binary outcomes with their 95% confidence intervals (CI). The Q and I<sup>2</sup> test statistics were conducted to examine heterogeneity, with I<sup>2</sup>\>50% indicating significant heterogeneity, with fixed or random-effects model applied, with P \<.05 indicating significant differences for effect sizes. If the heterogeneity was still obvious (I<sup>2</sup>\>50%) and more than two trials were included, then meta-influence analysis (for sensitivity analysis) was conducted. Further, meta-regression and regression-based sub-group analyses for the primary outcome were performed to identify potential variables leading to high heterogeneity with interpretation. Exploration of publication bias by Egger’s tests were planned for the primary outcome, together with trim and fill test for further identification of the stability. P \<.05 indicates significant differences for meta-regression, and p\<0.1 for Egger’s test. For the primary outcome, trial sequential analyses were performed by TSA 0.9 with type I error α = 0.05 and type II error β = 0.1, aiming at examining and minimizing the impact of type 1 errors due to sparse data and repeated significance testing following updates with new trials. We also conducted penalized test for further verification. Strength of evidence recommendations for effect estimation in each outcome was evaluated by The Grading of Recommendations Assessment, Development and Evaluation system (GRADE) pro GDT. In addition, cluster analyses and correlation coefficient of the variables were performed concerning acupoints applied among the included trials by R 4.1.3. # Results ## Study inclusion Originally, 4161 studies were retrieved from the seven databases, of which 2363 were removed due to duplication, and 1352 studies were eliminated according to titles and abstracts. The remaining 446 studies were downloaded for further consideration, and among which 311 studies were excluded with reasons. Finally, 15 trials from 12 studies (three three-arm studies were recombined to six trials for comparison) were included. ## Study characteristics All of the 15 trials were conducted in China and published in Chinese, with the range of publish years from 1999 to 2022. Specifically, 1118 participants aging from 12 to 65 with VN from 20 days to 6 years were involved. As for experimental intervention, acupuncture alone was performed in all of the groups, with WM alone applied in six trials, CHM alone applied in eight trials, and VT alone in one trial. Though some of the specific prescriptions of the acupoints and TCM decoction applied were different among the included trials, their principles and theories showed similarity according to the theory of TCM. detailed characteristics of the included trials are listed in. In addition, two specific acupoints, Kai yin yi hao and Sang yin point, were applied among the trials. Located on neck, the two acupoints are widely selected for improving quality of voice during acupuncture concerning VN. Location of Kai yin yi hao and Sang yin point, and summary of acupoints applied and frequency in the included trials are listed in. ## Assessment of quality and bias According to the Cochrane Handbook and risk of bias tool, the randomization methods were described clearly and appropriately in 10 trials, but were not detailed in the other trials. The manners of allocation concealment, blinding of outcome assessment, and selective reporting were assessed as unclear bias in all the trials due to insufficient reporting in full-text. The situation was same concerning blinding of participants and personnel, but we evaluated it as high bias in all the trials because it was impossible to perform the blinding between acupuncture and the controls (WM, CHM, and VT), obviously. In addition, all included trials were better reported in terms of completeness of outcome data, which were rated as low risk ( and Figs). ## Pooled results of acupuncture for patients with vocal nodules ### Effects of acupuncture vs. WM Pooled results favored acupuncture groups with significantly (p \< 0.05) higher CER (RR = 1.27, P for RR = 0.002, 95%CI:1.12–1.43, I<sup>2</sup> = 60%). Heterogeneity was still high (I<sup>2</sup> \> 50%) with random-effects model applied, and without extreme trial evaluated in meta-influence analysis. Furtherly, meta-regression and sub-group analysis according to variations of age, gender, duration of disease, frequency of treatment, course of treatment, selection of acupoints, and report of Deqi were designed. Results showed that the heterogeneity was reduced to acceptable level (I<sup>2</sup> \< 50%) with statistical significance for meta-regression (tau<sup>2</sup> = 0, I<sup>2resid</sup> = 0%, Adjusted R<sup>2</sup> = 100%, P = 0.04) when were sub-grouped based on selection of acupoints (local acupoints only, or local and remote acupoints). ### Effects of acupuncture vs. CHM Compared with CHM, results of systematic reviews favored acupuncture groups with significantly (p \< 0.05) higher CER (RR = 1.16, P for RR = 0.001, 95%CI:1.06–1.26, I<sup>2</sup> = 0%). Findings of meta-analyses also showed that acupuncture could reduce symptom scores (WMD = 3.92, P for WMD \< 0.001, 95%CI:2.65–5.19, I<sup>2</sup> = 0%), and improve scores of voice analyses (WMD = 3.70, P for WMD \<0.001, 95%CI:2.26–5.15, I<sup>2</sup> = 0%) more, significantly (p \< 0.05). ### Effects of acupuncture vs. VT However, results of one trial suggested that acupuncture was inferior to VT concerning lower CER (RR = 0.78, P for RR = 0.09, 95%CI:0.59–1.04, I<sup>2</sup> = NA). Similarly, during follow-ups, results also suggested fewer reduction on scores of grade, roughness, and breathiness (GRB) in one month after treatment (WMD = -0.75, P for WMD \< 0.001, 95%CI:-1.13 - -0.37, I<sup>2</sup> = NA), scores of GRB in two months after treatment (WMD = -0.65, P for WMD = 0.09, 95%CI:-1.41–0.11, I<sup>2</sup> = NA), scores of voice handicap index (VHI) in one month after treatment (WMD = -5.67, P for WMD \< 0.001, 95%CI:-9.00 - -2.34, I<sup>2</sup> = NA), and scores of VHI in two months after treatment (WMD = -3.75, P for WMD = 0.17, 95%CI:-9.14–1.64, I<sup>2</sup> = 0%) for the experimental compared with VT ( and – Figs). ### Meta-regression based sub-group analyses of the effects For the outcomes of CER (acupuncture vs. WM and acupuncture vs. CHM), sub-group analyses based on meta-regression were performed to explore characteristics which might influence sizes of effect and heterogeneities, potentially. Sub- group analyses were also designed based on variations of age, gender, duration of disease, frequency of treatment, course of treatment, selection of acupoints, and report of Deqi. Results of sub-group analyses revealed some common findings concerning the two comparisons. Firstly, groups with older ages (40 ± 2.5, years) yielded comparatively higher RR than younger ages (30 ± 2.5, years and 35 ± 2.5, years). Secondly, results favored the experimental with comparatively higher RR when more proportions of male patients were included. Thirdly, as for duration of disease, results favored the experimental with comparatively higher RR when the time was shorter than one year, compared with those were not shorter than one year. Fourthly, as for course of treatment, results favored the experimental with comparatively higher RR when it was shorter than one month, compared with those were one month. Finally, groups with local and remote acupoints applied yielded comparatively higher RR than those with local acupoints applied only. ## Trial sequential analysis and penalized test Trial sequential analysis and penalized tests were conducted for CER of acupuncture vs. WM and acupuncture vs. CHM. For the two outcomes, results favored exclusion of the possibility of false positive before and after penalized test. However, the results need to be interpreted with caution because of the lack of inclusion (≤10 trials), more inclusion was needed to reach enough superiority and to meet the required information size for them. ## Adverse event reported in trials Among the included trials, six of them reported that there was no acupuncture specific adverse event, and it was not mentioned in other nine trials. ## Publication bias and trim and fill test Egger’s test was performed in two comparisons, and publication bias was not detected in both of them ( and Figs). Trim and fill test showed that statistical significances of the two comparisons were not reversed after certain number of missing studies filled, indicating the stability of them ( and Figs). ## Levels of evidence As for acupuncture vs. CHM, levels of evidence were suggested as moderate (⨁⨁⨁◯, GRADE B) for all of the three comparisons. However, levels of evidence for all comparisons concerning acupuncture vs. WM and acupuncture vs. VT were low (⨁⨁◯◯, GRADE C). ## Cluster analyses and correlation coefficient of the variables As for cluster analyses, results of heat map showed some clustering effects between acupoints selected for acupuncture in the trials. In addition to several one-time application of acupoints \[including Shui tu (ST-10), Fu tu (LI-18), Sang yin point, Jia ji (EX-B2), Xue hai (SP-10), Feng long (ST-40), Lie que (LU-7), Zhao hai (KI-6), Yu ji (LU-10)\], Kai yin yi hao and He gu (LI-4) were the most frequently applied matching-acupoints in trials (as component or alone). Other frequently selected acupoints including Zu san li (ST-36) and Ren yin (ST-9). Results of correlation coefficient between the acupoints were also showed in. # Discussion VN, which is also known as singer’s nodules, teacher’s nodules, or “shouting (caused) nodules” when occurred in children, is a special type of chronic laryngitis induced by inflammatory lesions. The main clinical symptom of VN is hoarseness, and in research with patients were suffering from voice disorders, vocal nodules was diagnosed among 40% of them. However, Shah, et al. revealed that objective and subjective voice measurements were not different in various vocal nodule sizes statistically, other than pitch reduction. Meanwhile, in a study with 79 VN patients showed that there was little evidence to suggest that the nodules themselves were “driving” the severity of the dysphonia. In this review, results of meta-regression based sub-group analyses suggested that sizes of effect were varying in different ranges of ages, genders, duration of disease. As a result, we also believe that the variables above are also associated with sizes of VN, and further researches concerning relationships between different sizes of VN and severities of voice disorder are needed to improve this study and provide more objective evidence of subgroups. A diagnosis of VN can be made through laryngoscopy clinically, and modern artificial intelligence technologies, such as deep-learning-based computer-aided diagnosis, could provide valuable references for diagnosis of benign, precancerous, and cancer lesions during laryngoscopy examination. Treatments for VN worldwide are divided into two categories, mainly, as surgery and non-surgery therapy, while surgery therapy and drug therapy (such as antibiotic, steroid, or anti-reflux medications) are not routinely recommended in guidelines. Different from considerable one-off expenditure and physical quality requirements of VN surgery, VT has a high status in non-surgical therapies for VN. However, in China and many other developing countries, there are no specific speech pathology specialty, and VT is poorly developed or absent. Results of this review indicate that compared with oral WM alone and oral CHM alone, pooled results of our study covering 15 trials with 1118 participants favored acupuncture alone concerning higher CER (acupuncture vs. WM, and acupuncture vs. CHM), more reduction on symptom scores (acupuncture vs. CHM), and more improvement on scores of voice analyses (acupuncture vs. CHM). However, VT exhibited favorable improvements compared with acupuncture for VN, including higher CER, and more reduction on scores of GRB and VHI during follow-ups. Among all of the trials included in this study, acupuncture was performed manually. After pooling all the acupoints applied in the trials, two types of acupoint protocols were noticed, including local acupoints (located at face and neck) and remote acupoints (located at upper and lower limbs). In TCM, primarily, acupoints are classified into 1) acupoints of fourteen meridians, 2) extra acupoints (in addition to the fourteen meridians), and 3) ashi acupoints. For the first and second types, there are names, locations and indications for each of the acupoints, while there is no fixed name, location or indication for ashi acupoints, the third type. Based on these, in terms of general efficacy of certain disorder or disease, the classifications above can be recombined and divided into 1) local (effects) acupoints, 2) remote (effects) acupoints, and 3) specific (effects) acupoints. Local effects are equipped by all acupoints (including acupoints of fourteen meridians, extra acupoints, and ashi acupoints), such as head acupuncture for post-stroke aphasia, acupuncture at Jingming (BL1) for dry eye disease, and local acupressure for diabetic peripheral neuropathy of lower limbs. Remote effects are equipped by acupoints of fourteen meridians and extra acupoints, especially those located at distal end of elbow joints of upper limbs and distal end of knee joints of lower limbs, with effects of their belonging meridians, such as acupuncture at Qu chi (LI-11) and Zu san li (ST-36) for left hemiplegia after ischemic stroke, and wrist-ankle acupuncture for hypertension after intubation during induction of general anesthesia. In addition, from clinical experience to medical evidence, in thousands of years, many specific acupoints were discovered, tested, and summarized for acupuncture with good efficacy, long duration, fewer needles, such as acupuncturing at sphenopalatine ganglion acupoint for allergic rhinitis. In this review, cluster analyses and correlation coefficient of the acupoints revealed that Kai yin yi hao and He gu (LI-4) were the most frequently applied matching-acupoints in the included trials, followed by Zu san li (ST-36) and Ren yin (ST-9). Commonly, local acupoints are selected and matched with remote and (or) specific acupoints in acupuncture, which is beneficial for achieving enough acupuncture stimulation, longer efficacy duration, and balanced meridian effects per treatment. Our results of meta-regression for CER (acupuncture vs. WM and acupuncture vs. CHM) also suggested superiorities of local acupoints plus remote acupoints, compared with local acupoints applied alone. In our study, only six of the included trials reported that there was no acupuncture specific adverse event, while it was not mentioned in other nine trials. Sometimes, patients may feel local and mild sensations of sour, numb, distension, painful during or after acupuncture. The sensations are associated with De Qi (obtaining Qi) in TCM, and will clear up several days later without specific medical care required. Such mild and self-limited events include local bruising and radiating pain. In addition, severer events may appear on very few first-time acupuncture patients due to fasting state, uncomfortable position, overstrain, or movement during acupuncture, including fainting, sticking of needle, broken of needle, vascular injury, nervous system injury, or even visceral injury. As a result, detailed communication between doctors or acupuncture practitioners and patients before acupuncture, required preparations, and complete post treatment orders are necessary. As for study quality and risk of bias, all the 15 trials are RCTs, but none of them implied placebo control. Randomization method was clear and appropriate in 10 trials, while it was in unclear risk of bias for the other 5 trials. Allocation concealment and blinding (for outcome assessment) method were of unclear risk of bias in all trials. No study reported drop-out, and a protocol or registration ahead of experiment was not mentioned. Publication bias (by Egger’s tests) or instability (by trim and fill tests) was not suspected in results, and more inclusions were required to meet the required information size for all comparisons according to trial sequential analyses and penalized tests. According to the GRADE, levels of evidence were moderate (⨁⨁⨁◯, GRADE B) concerning acupuncture vs. CHM, but were low (⨁⨁◯◯, GRADE C) concerning acupuncture vs. WM and acupuncture vs. VT. As a result, more placebo or blank- controlled, double-blind, prospective, randomized trials of acupuncture for VN are urgently needed. # Conclusion Our study, the first one with RCT evidence from 15 trials involving 1,118 participants, proved that applying acupuncture yielded better improvement for patients with VN compared with WM and CHM. However, negative results were discovered in acupuncture vs. VT groups in this study. There is also need for RCTs with improvements on designing and interventions in experimental and controls. ## Limitations This study had several limitations. Firstly, most of the trials included were of moderate to high risk of bias, with reasons such as without mentioning details of random sequence generation method, allocation concealment, and blinding of participants, personnel and outcome assessment. This is the main reason for low quality of the included trials. Secondly, interventions and follow-up periods were short among most of the trials, while longer treatment duration and follow- up periods for VN, a chronic and recurrent disorder, is essential and required. Finally, inclusion of trials and participants were limited due to the few numbers of published trials and small sample sizes, relatively. # Supporting information VN Vocal nodules VT Voice training CHM Chinese herbal medicine TCM Traditional Chinese medicine WM Western medicine CER Clinical effective rate GRB Grade, roughness, breathiness VHI Voice handicap index RR Risk ratio WMD Weighted mean difference CI Confidence interval GRADE The Grading of Recommendations Assessment, Development and Evaluation system [^1]: The authors have declared that no competing interests exist. [^2]: ‡ QF share Co first authorship on this work.
# Introduction Tuberculosis (TB) is a serious global health problem and is one of the leading causes of death worldwide. In 2013, an estimated 9.0 million people were infected with TB disease and 1.5 million (approximately 17%) died from this disease. Importantly, these recent epidemiological estimates are higher than what were previously estimated. Notably, of the 1.5 million deceased cases, 360,000 recorded among people with HIV infection. Despite of its life-threating pathogenesis, TB is a curable disease when it is correctly diagnosed and effectively treated. However, rapid and accurate diagnosis of TB can be difficult due to the paucibacillary characteristics of the disease (especially for cases with smear-negative, co-infection with HIV and drug-resistance) and the challenge of sample collection from deep-seated tissues. In fact, approximately 35% of all the worldwide TB infections are undiagnosed. Furthermore, the ratio of patients with undiagnosed multi-drug resistant TB remains much staggering (\~75%). Less than 3% of patients who are diagnosed with TB infection are proved to have certain pattern of drug resistance. Solid and/or liquid culture is generally considered as the standard reference for TB diagnosis. However, limited laboratory facilities in resource-limited settings and prolonged culturing period restrict the utility of culture-based diagnosis in clinical practice. Histology is widely applied to the diagnosis of TB where the technical expertise exists, but this is technically demanding, time-consuming and it lacks specificity. In early 2011, a novel, automated, rapid, cartridge-based nucleic acid amplification test, named the Xpert® MTB/RIF assay (Cepheid, Sunnyvale, USA) was authorized by the World Health Organization (WHO) to be used for TB diagnosis. Xpert® MTB/RIF can simultaneously test both TB and rifampicin resistance through examination of the DNA of Mycobacterium tuberculosis and detection of major mutations which confer rifampicin resistance. This assay was first endorsed by WHO as an initial diagnostic test in patients with human immunodeficiency virus (HIV)-associated pulmonary TB or suspected pulmonary MDR-TB. Xpert® MTB/RIF showed a substantial accuracy for detection of pulmonary TB in adults with 89% sensitivity and 99% specificity. In late 2013, WHO expanded its recommendations to include the diagnosis of TB in some special subjects such as children and patients with certain forms of extrapulmonary TB. A systematic review by Detjen et al. revealed that Xpert offers a better sensitivity (62%) and specificity (98%) for the diagnosis of pulmonary tuberculosis in children. However, the information concerned the accuracy of Xpert MTB/RIF in different TB endemic areas is lacking. In fact, the prevalence of TB is clearly varying among different regions. Based on global TB epidemiology in 2013, 56% of TB cases worldwide were in the Western Pacific Regions and South-East Asia while 25% of the cases were in the African Region, which also had the highest rates of cases and deaths relative to population. Notably, India and China alone had 24% and 11% of total cases, respectively. Therefore, in this systemic review, we aimed to determine the diagnostic accuracy of Xpert MTB/RIF assay in different regions with different TB prevalence regardless of sample type, subject’s age, HIV co-infection or smear-positivity. # Methods Following the standard guidelines, we designed a protocol before commencing the study. ## Literature search strategy MEDLINE, EMBASE, the Cochrane Library, and Web of Knowledge were used to retrieve published work without language or date restrictions. The last search was done on June 20, 2015. The keywords used for searching were: ‘‘Xpert”, ‘‘Gene Xpert”, “Xpert MTB/RIF”, ‘‘Cepheid”, ‘‘tuberculosis” and ‘‘Mycobacterium tuberculosis”. ## Study selection and data extraction Two researchers (Shiying Li and Bin Liu) carried out the process of study- retrieval and data extraction independently. Any disagreements in the process were solved by discussing with a third researcher (Peng Hu). ### Inclusion criteria Inclusion criteria used in this meta-analysis were: (i) peer-reviewed, full- text, randomized controlled trials, cohort studies and cross-sectional studies, which used Xpert MTB/RIF for TB detection; (ii) specimens were tissues or body fluid collected from suspected TB patients; (iii) the number of cases ≥30; (iv) original data were sufficient to calculate the true positive (TP), false positive (FP), true negative (TN) and false negative (FN); (v) culture and/or a composite reference standard (CRS) was used as the reference standard in each individual study; and (vi) nationalities of individuals were clearly described. ### Exclusion criteria The initially selected articles were further screened based on the following exclusion criteria: (i) non-clinical research; (ii) abstract of any conference; (iii) case report; and (iv) review. ### Data extraction The basic characteristics of selected studies such as the year of publication, the number of the study population, number and type of samples, patients’ epidemiological and laboratory results, were collected. Additionally, the diagnostic characteristics of Xpert MTB/RIF such as TP, FP, TN, and FN were extracted. If data were insufficient in any study to perform a meta-analysis, we contacted the authors by e-mail for further information. If we were unable to obtain target data for certain studies, these studies were excluded. ## Imperfect reference standard Culture is a gold standard for many infectious diseases except TB due to its paucibacillary characteristics, which may lead to a misdiagnosis of tested sample. Assuming that Xpert MTB/RIF typically identifies TB in specimens with negative culture, this result would be considered as FP causing an underestimation of Xpert MTB/RIF’s true specificity. However, a CRS, which diagnoses the TB based on comprehensive results of clinical manifestations, signs and laboratory tests, might sometimes confirm the positivity of Xpert MTB/RIF for a sample with negative-TB culture and hence overestimate Xpert MTB/RIF’s accuracy. On the other hand, a CRS itself might reduce the accuracy of Xpert MTB/RIF by considering the result as FN. Thus, to provide a more credible range of accuracy, we compared the accuracy of Xpert MTB/RIF to both the culture and CRS. ## Statistical analysis MIDAS, a professional module of diagnostic test in STATA statistical software (version 12.0; STATA Corporation, College Station, TX, USA), was used to carry out the meta-analysis. The summary receiver operating characteristic model and bivariate random-effects model were carried out in this study to estimate the diagnostic accuracy of Xpert MTB/RIF. We calculated the sensitivity and specificity of Xpert MTB/RIF to diagnose TB for each individual study, then a pooled sensitivity, specificity, and area under the curve (AUC) were obtained, comparing with culture or CRS, along with 95% confidence intervals. ## Assessment of methodological qualities The Review Manager software (version 5.3, The Nordic Cochrane Centre, Copenhagen, Denmark), which contains a Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool, was used to evaluate the quality of selected studies. In QUADAS-2 tool, each question has three choices: yes, unclear or no. If1338 there was at least one ‘no’ or ‘unclear’ answer to a given question of a given domain, the risk for bias was considered as high or unclear, respectively. ## Publication bias In a systematic review, publication bias should be assessed to estimate whether the relevant studies with positive results are more likely to be published than the ones with negative results. However, there is no validated method for publication bias in test accuracy reviews yet. So in this research, we did not estimate the publication bias. ## Heterogeneity analysis A bivariate boxplot was used to assess the heterogeneity between included studies. It describes the degree of interdependence including the central location and identification of any outliers with an inner oval representing the median distribution of the data points and an outer oval representing the 95% confidence bound (by visually examining the position of each individual study, within the range of boxplot suggesting more heterogeneity). We predicted pre- existing heterogeneity in terms of sample types, patient age, status of HIV and smear-positivity. Therefore, studies were pre-specified into four subgroups: pulmonary versus extrapulmonary, adults versus children, HIV positive versus negative, and smear positive versus negative. Meta-analysis in each subgroup was only performed when at least four studies were available. # Results ## Description of studies In total, we identified 106 studies (PRISMA flow chart and Supplementary reference) that included 52,410 samples for TB detection. All studies were in English except for two which were in Turkish. Among the 106 studies, 54 studies (47.8%) were carried out in 22 countries with high TB burden. The median number of specimens was 494 for TB detection. The proportions of HIV- positive patients ranged from 0% to 100%. In particular, two studies only included HIV patients, one study had no HIV patients and HIV status was unknown in 39 studies. Children were included in 21 studies, while patient’s age was unknown in another 30 studies. For sample type, pulmonary samples (such as sputum and bronchoalveolar lavage) were included in 71 studies, extrapulmonary samples (such as body fluid, fine needle aspiration, stool, and blood) were included in 25 studies and a mixture of pulmonary and extrapulmonary samples were used in 9 studies. The information of smear-positivity was reported in 44 included studies. The details of diagnostic accuracy of each individual study were shown in. ## Methodological quality of selected studies The methodological qualities were estimated based on the culture and CRS. The overall methodological quality of included studies is summarized in (details of the quality assessment for each study are individually shown in and Figs). In the majority of studies, data was collected consecutively or randomly (n = 72; 67.9%). All included studies were carried out either at reference laboratories or at tertiary care centers. Based on index tests, 11 studies (9.7%) were considered to carry an unclear risk of bias. Accordingly, the results of index test were interpreted blindly regardless of the reference standard results. The result of heterogeneity analysis was shown in. A significant heterogeneity was found based on the bivariate box plot (culture, I<sup>2</sup> = 99.90; CRS, I<sup>2</sup> = 99.88). ## Sensitivities, specificities and AUCs of Xpert MTB/RIF for TB detection For the overall diagnostic accuracy of Xpert MTB/RIF, culture was used as the gold reference standard in 95 studies, while CRS was used in 20 studies. Pooled sensitivity, specificity and AUC were 0.85 (95% confidence interval \[CI\] 0.82–0.88), 0.98 (95% CI 0.96–0.98) and 0.97 (95% CI 0.95–0.98) compared to culture reference standard, respectively, while they were 0.59 (95% CI 0.44–0.77), 0.99 (95% CI 0.92–0.96) and 0.95 (95% CI 0.92–0.96) compared to CRS, respectively. Since included studies were significantly heterogeneous, thus we further performed a separate meta-analysis for each of the four pre-identified subgroups based on the type of sample (pulmonary and extrapulmonary), subject’s age, HIV co-infection and smear-positivity. In the pulmonary and extrapulmonary subgroup, 74 studies included pulmonary samples while 26 studies included extrapulmonary samples. Pooled sensitivity, specificity and AUC of Xpert MTB/RIF for detecting pulmonary TB (PTB) were 0.87 (95% CI 0.83–0.90), 0.97 (95% CI 0.96–0.98) and 0.97 (95% CI 0.95–0.98) compared to culture reference standard, respectively; while they were 0.76 (95% CI 0.53–0.90), 0.97 (95% CI 0.87–1.00) and 0.95 (95% CI 0.92–0.96) compared to CRS, respectively. Likewise, pooled sensitivity, specificity and AUC of Xpert MTB/RIF for extrapulmonary TB (EPTB) detection were 0.80 (95% CI 0.69–0.88), 0.97 (95% CI 0.94–0.98) and 0.97 (95% CI 0.95–0.98) compared to culture reference standard, respectively; while they were 0.49 (95% CI 0.32–0.67), 0.99 (95% CI 0.97–1.00) and 0.94 (95% CI 0.92–0.96) compared to CRS, respectively. For different HIV co-infection status, diagnostic accuracy parameters could be collected on 26 studies and 18 studies that included HIV positive and negative cases, respectively. Pooled sensitivity, specificity and AUC of Xpert MTB/RIF for detecting TB in HIV positive patients were 0.81 (95% CI 0.73–0.87), 0.98 (95% CI 0.96–0.99) and 0.96 (95% CI 0.94–0.97) comparted to culture reference standard, respectively; while they were 0.63 (95% CI 0.45–0.87), 0.94 (95% CI 0.87–0.97) and 0.94 (95% CI 0.92–0.96) compared to CRS, respectively. Likewise, pooled sensitivity, specificity and AUC of Xpert MTB/RIF for detecting TB in HIV negative patients were 0.77 (95% CI 0.67–0.85), 0.99 (95% CI 0.97–0.99) and 0.98 (95% CI 0.96–0.99) comparted to culture reference standard, respectively; while they were 0.44 (95% CI 0.08–0.87), 0.99 (95% CI 0.93–1.00) and 0.98 (95% CI 0.97–0.99) compared to CRS, respectively. For smear-positivity status, samples of 38 studies were positive while they were negative in 43 studies. Compared to culture reference standard, pooled sensitivity, specificity and AUC of Xpert MTB/RIF for detecting TB in patients with positive smears were 0.99 (95% CI 0.97–0.99), 0.88 (95% CI 0.76–0.94) and 0.99 (95% CI 0.98–1.00), respectively; while they were 0.70 (95% CI 0.64–0.75), 0.98 (95% CI 0.96–0.99) and 0.89 (95% CI 0.86–0.929) in patients with negative smears, respectively. Similarly, compared to the CRS, Xpert MTB/RIF’s pooled sensitivity, specificity and AUC of for TB detection in patients with negative smears were 0.52 (95% CI 0.41–0.63), 0.99 (95% CI 0.95–1.00) and 0.68 (95% CI 0.63–0.72), respectively. We were unable to determine diagnostic accuracy parameters of Xpert MTB/RIF compared to CRS in patients with positive smears due to the lack of this information. Diagnosis of TB in children is rather difficult due to its typical paucibacillary characteristics and the difficulty of collecting sputum sample. Microscopic examination has a little value in diagnosis TB in children. Culture methods have a greater benefit, yet have a highly variable sensitivity. Clinical diagnosis of children TB relies mainly on a combination of symptoms, radiological findings, and identification of a tuberculosis contact. Thus, the assessment of diagnostic accuracy of Xpert MTB/RIF for detecting TB in children is importantly needed. In the subgroup of subject’s age, there were 48 studies that included adults while 18 studies included children. Compared to culture reference standard, pooled sensitivity, specificity and AUC of Xpert MTB/RIF for detecting TB in adults were 0.82 (95% CI 0.76–0.86), 0.98 (95% CI 0.96–0.99) and 0.97 (95% CI 0.95–0.98) respectively; while they were 0.76 (95% CI 0.70–0.81), 0.98 (95% CI 0.96–0.99) and 0.90 (95% CI 0.87–0.93) in children, respectively. Likewise, compared to CRS, pooled sensitivity, specificity and AUC of Xpert MTB/RIF’s for TB detection in adults were 0.52 (95% CI 0.35–0.69), 0.99 (95% CI 0.97–1.00) and 0.96 (95% CI 0.94–0.97), respectively; while they were 0.55 (95% CI 0.41–0.65), 0.99 (95% CI 0.97–1.00) and 0.92 (95% CI 0.89–0.94) in children, respectively. ## Sensitivities, specificities and AUCs of Xpert MTB/RIF in TB detection at different endemic degree regions To evaluate the diagnostic accuracy of Xpert MTB/RIF assay in different regions, multiple meta-analyses were further carried out based on different TB prevalence. Since the data on CRS were limited within studies of different endemic regions, therefore all analyses were only relied on culture reference standard. As shown in, 54 studies were carried out in 22 high TB burden countries while 56 studies were carried out in middle and low prevalence countries. Pooled sensitivity, specificity and AUC in highly endemic countries were 0.84 (95% CI 0.80–0.88), 0.97 (95% CI 0.95–0.98) and 0.96 (95% CI 0.94–0.98), respectively; while they were 0.89 (95% CI 0.84–0.93), 0.98 (95% CI 0.97–0.99) and 0.99 (95% CI 0.97–0.99) in countries with middle/low endemics, respectively. Additionally, we sought to re-evaluate the diagnostic accuracy of Xpert MTB/RIF’s for the four subgroups among regions with different TB endemic level. However, data are quite limited for children; therefore, analyses were carried out considering the other 3 subgroups (sample’s type, HIV co-infection, and smear-positivity). For studies including PTB samples, pooled sensitivity, specificity and AUC were 0.84 (95% CI 0.80–0.88), 0.97 (95% CI 0.95–0.98) and 0.96 (95% CI 0.94–0.98) in high TB endemic countries, respectively; while they were 0.92 (95% CI 0.88–0.95), 0.98 (95% CI 0.96–0.99) and 0.99 (95% CI 0.97–0.99) in middle/low endemic countries, respectively. For studies including EPTB samples, pooled sensitivity, specificity and AUC were 0.82 (95% CI 0.67–0.91), 0.92 (95% CI 0.84–0.96) and 0.94 (95% CI 0.92–0.96) in high TB endemic countries, respectively; while they were 0.81 (95% CI 0.63–0.91), 0.98 (95% CI 0.96–0.99) and 0.99 (95% CI 0.97–0.99) in middle/low endemic countries, respectively. For different HIV co-infection status, pooled sensitivity, specificity and AUC in high TB endemic countries were 0.80 (95% CI 0.75–0.84), 0.97 (95% CI 0.90–0.99) and 0.86 (95% CI 0.83–0.89) in HIV positive patients, respectively; while they were 0.75 (95% CI 0.66–0.82), 0.98 (95% CI 0.96–0.99) and 0.96 (95% CI 0.94–0.97) in HIV negative patients, respectively. Among middle/low TB endemic countries, pooled sensitivity, specificity and AUC were 0.81 (95% CI 0.61–0.92), 0.99 (95% CI 0.97–1.00) and 0.99 (95% CI 0.98–1.00) in HIV positive patient, respectively; while they were 0.76 (95% CI 0.47–0.92), 0.99 (95% CI 0.96–0.99) and 0.99 (95% CI 0.97–0.99) in HIV negative patients, respectively. Based on smear-positivity, pooled sensitivity, specificity and AUC in studies including positive smears were 0.97 (95% CI 0.95–0.98), 0.89 (95% CI 0.60–0.98) and 0.97 (95% CI 0.96–0.99) for countries with high TB endemic, respectively; while they were 0.99 (95% CI 0.97–1.00), 0.88 (95% CI 0.75–0.94) and 0.99 (95% CI 0.98–1.00) for countries with middle/low TB endemic, respectively. For studies including negative smears, pooled sensitivity, specificity and AUC were 0.68 (95% CI 0.60–0.75), 0.96 (95% CI 0.90–0.98) and 0.84 (95% CI 0.81–0.87) for regions with high TB endemic, respectively; while they were 0.73 (95% CI 0.63–0.81), 0.99 (95% CI 0.97–1.00) and 0.93 (95% CI 0.91–0.95) for countries with middle/low TB endemic, respectively. # Discussion Major advantages of this systemic review were the use of a pre-designed protocol, a comprehensive search, independent researchers, an effective model for meta-analysis (a bivariate random-effects model) and analyzing of four pre- identified subgroups to investigate the heterogeneity. In addition, effective communication with authors of studies with insufficient information resulted in a more efficient data analysis. The comprehensive meta-analyses carried out in this systemic review revealed that Xpert MTB/RIF is a highly sensitive diagnostic tool for TB detection regardless sample’s type, status of HIV co-infection, subject’s age and smear- positivity. In addition, our analyses revealed that the sensitivity of Xpert MTB/RIF for TB detection compared to culture reference standard was higher than that compared to CRS, which was consistent with previous studies. Culture is an imperfect reference standard for TB detection, however, it requires more strict skills and long time. On the other hand, CRS is a composite of clinical investigations and laboratory examinations which offer more information to make a decision. However, there is a possibility that CRS could lead to a false positive diagnosis. Therefore, a combined application of the two reference standards (culture and CRS) may help to make more accurate clinical decisions. Heterogeneity of diagnostic accuracy of Xpert MTB/RIF among all included studies was significantly higher when either compared to culture reference standard or CRS indicating the variability of Xpert MTB/RIF’s diagnostic accuracy among different populations. Notably, the impact of sample type, age of patients, and status of HIV co-infection and smear-positivity are the main factors studied for their impact on Xpert MTB/RIF’s diagnostic accuracy. In this context, our results showed that Xpert MTB/RIF has higher sensitivity in patients with positive smears (0.99, 95% CI 0.97–0.99), in patients with PTB samples (0.87, 95% CI 0.83–0.90), in adults (0.82, 95% CI 0.76–0.86) and in HIV- positive patients (0.81, 95% CI 0.73–0.87). Overall, the sensitivities, specificities, and AUCs were ≥0.70, ≥0.97 and ≥89%, respectively, indicating a high diagnostic accuracy of Xpert MTB/RIF for TB detection which is consistent with previous reports. Given the extraordinary prevalence of TB worldwide since it is estimated that every year, nearly 10 million people fell ill with TB, Xpert MTB/RIF would be a perfect quick and accurate diagnostic assay for TB diagnosis. Notably, we found that Xpert MTB/RIF was more sensitive in HIV- positive than in HIV-negative individuals (81% versus 77%, compared to culture), which disagrees with the results of a previous review by Steingart et al., (HIV- positive versus HIV-negative: 79% versus 86%, compared to culture). This difference could be attributed to the fact that Steingart et al addressed the accuracy of Xpert MTB/RIF only in the pulmonary samples of adults, while in our study, we addressed the overall diagnostic accuracy of Xpert MTB/RID regardless the sample type and subject’s age. TB remains one of the world’s deadliest communicable diseases. There are some reviews that have addressed the accuracy of Xpert MTB/RIF either among specific population, such as adults or in children, or by testing specific samples, such as pulmonary, or extrapulmonary samples. Therefore, it is important to evaluate the overall diagnostic accuracy of Xpert MTB/RIF among countries with different levels of TB prevalence as well as different ages, sample sites, and HIV and smear statuses. To the best of our knowledge, our study was the first to investigate this issue. Based on our study, the overall pooled sensitivity of Xpert MTB/RIF for TB detection was lower in high TB prevalence countries than that of middle/low ones. Also, the same trend was found when different four subgroups were considered. Taken together, our results suggested that the diagnostic accuracy of Xpert MTB/RIF has a higher efficiency in countries with middle/low TB endemic burden than in countries with high endemic. There are still some limitations in the current analysis such as there was no protocol available on how to handle non-respiratory specimens, specimen processing was extremely variable across studies and the CRS standard differed between studies. Therefore, the overall outcome should be interpreted with caution. In conclusion, based on our meta-analyses using a bivariate model and the sufficient number of specimens, the diagnostic accuracy of Xpert MTB/RIF for TB detection was quite high. The overall sensitivity of Xpert MTB/RIF was lower in high TB burden countries than that of the middle/low burden. The results obtained in the current study will significantly help the physicians especially those in high-risk regions to make their clinical decision. # Ethics approval and consent to participate The experimental protocol was established, according to the ethical guidelines of the Helsinki Declaration and was approved by the Human Ethics Committee of Department of Infectious Diseases, The Second Affiliated Hospital, Chongqing Medical University. Written informed consent was obtained from individual participants. # Supporting information [^1]: The authors have declared that no competing interests exist.
# Introduction For the past decade, the removal of the internal limiting membrane (ILM) has been an important step for anatomical and functional success in macular hole, macular pucker, and even retinal detachment surgeries. Because of its anatomical characteristics, the ILM is challenging to identify during surgical procedures. With the assistance of a vital dye such as indocyanine green (ICG) or brilliant blue G (BBG), the technique is much easier. Therefore, the use of dyes to identify structures during vitreoretinal surgery, “chromovitrectomy,” has become a popular technique in recent years. Although the dyes are used temporarily during the operation, some of the dyes may remain on the unpeeled part of the ILM. Several groups have reported that ICG may persist in the ocular cavity up to 6 weeks after its application during surgery. Several groups reported toxicity to retinal pigment epithelial cells and the neurosensory retina, as well as cases of optic nerve atrophy, after the use of ICG. Therefore, several alternative dyes have been introduced for use in vitreoretinal surgery, including infracyanine green (IfCG), trypan blue (TB), bromophenol bue (BPH), patent blue, and BBG. Even so, all of the previously mentioned dyes were reported to exhibit toxicity on RPE cells following acute exposure during surgical doses. IfCG, BBG, and BPH have been shown to be less toxic on retinal ganglion cells and RPE cells compared with ICG. BBG, it was claimed, provided a good staining to the ILM and was not toxic in experimental studies and a case series in humans. However, recent reports showed a selective toxicity to photoreceptors related to BBG after intravitreal injection in rabbit eyes and RPE changes on fluorescein angiography, as well as macular damage following accidental subretinal dye injection in humans. Another report of the intraocular safety of ICG, TB, Evans blue (EB), and BBG on ARPE-19 cell lines and murine retinal ganglion/Müller glial (RGC) primary cell cultures showed that all dyes demonstrated relatively safe viability profiles in both cell lines at surgically relevant concentrations and times. BBG was the only dye that caused toxicity in ARPE-19 cell lines after short exposure times, and ICG had a favorable viability profile at almost all of the concentrations and times tested. Mitochondria have been implicated in the cytotoxicity caused by the dyes. Mitochondrial membrane potential (ΔΨ<sub>m</sub>) was altered after exposure to surgical does of ICG, TB, PB, or a four-fold surgical dose of BrB. An in vitro RPE cell study by Penha et al. revealed that expression of Bax, cytochrome c, and caspase-9 was upregulated at the mRNA and protein level after ICG exposure, but Bcl-2, an anti-apoptotic protein, was downregulated; however, brilliant blue (BriB), resulted in upregulation of Bcl-2. During real-life procedures, not only the RPE, but also the photoreceptors may be exposed to the dyes cause cytotoxicity in the these cells. Until now, no study had focused on the effect of the intraoperative dyes on the in vitro safety of photoreceptor cells. Cultures of 661W cells, an in vitro model that mimics photoreceptor cells, have been widely used in the study of retinal degeneration, retinal neuroprotection, and retinal regeneration. Macroautophagy is typically referred to as a degradation process that proceeds in a lysosome-dependent manner by which microtubule-associated proteins 1A/1B light chain 3B (LC3) facilitates elongatation of autophagosome and fuses with lysosomes for degradation and recycling. Sequestome 1 (SQSTM1) contains LC3 and ubiquitin-binding motifs to recruit ubiquitinated proteins to the autophagosome, which serves as an autophagy receptor, for selective bulk degradation. Autophagy plays a beneficial role in several ocular cell types to maintain the eye’s normal physiological function. Autophagy is involved in maintaining inner segment turnover in photoreceptors, and it protects cells from stress and melanin degradation in RPE cells. However, autophagy is activated to promote autophagic cell death in retinal ganglion cells during chronic intraocular pressure elevation, suggesting the role of autophagy might be varied depending on types of ocular cells or stress. The role of autophagy in RPE and photoreceptor cells in response to vital dyes remains unknown. Here, we examine the autophagic effects of vital dyes in RPE and photoreceptor cells and showed that autophagy was inhibited and induced in ARPE-19 and 661W cells, respectively, when exposed to ICG and BBG. Nevertheless, ablation of autophagy enhanced ICG and BBG-induced cytotoxicity in both ARPE-19 and 661W cells. Administration of dietary supplements, including resveratrol, lutein, and CoQ10, induced autophagy and diminished the cytotoxic effects of ICG and BBG, suggesting autophagy may play a cytoprotective role in RPE and photoreceptor cells during exposure to ICG or BBG. # Material and methods ## Compounds The dyes ICG (I2633) and BBG (B0770) and the dietary supplements resveratrol (R5010), coenzyme Q10 (CoQ10, C9538), and lutein (xanthophyll from marigold, X6205) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Reagents and materials for cell culture were obtained from GIBCO (Life Technologies; City, Country). CellTiter-Glo Luminescent Cell Viability Assay (G7572) was purchased from Promega Corporation (Madison, WI, USA). ## Dye preparation ICG and BBG (5mg each) was dissolved in 100 μL dimethyl sulfoxide (DMSO) and phosphate-buffered saline (PBS) was added to obtain a stock of 5 mg/mL. The ICG solution should be made fresh before use. ICG and BBG were mainly used at concentrations of 0.05 mg/mL in this study. The surgical concentrations of ICG and BBG solutions are 0.5 mg/mL and 0.25 mg/mL, respectively. ## Cell culture of human RPE cells Adult human RPE cell cultures (ARPE-19) were obtained from the American Type Culture Collection (CRL-2302; ATCC, Manassas, VA, USA). These cells were maintained in a 1:1 mixture of Dulbecco's modified Eagle's medium (DMEM) and Ham's F12 medium supplemented with 10% fetal bovine serum (Life Technologies), sodium bicarbonate (1.2 g/L), L-glutamine (2.5 mM), HEPES (15 mM), and sodium pyruvate (0.5 mM). The cells were cultured at 37°C in a humidified atmosphere of 95% air and 5% CO<sub>2</sub>. The mouse photoreceptor–derived 661W cell line was generously provided by Dr. Muayyad Al-Ubaidi (University of Oklahoma, Norman, OK, USA). Cells were grown in DMEM supplemented with 10% fetal bovine serum in 100-mm tissue culture dishes in which cells typically were seeded at a concentration between 1 × 10<sup>5</sup> and 1 × 10<sup>6</sup> in 10 mL of growth medium. They were allowed to grow to 80%–90% confluence before harvesting for use in experiments. ## CellTiter-Glo luminescent cell viability assay ARPE19 and 661W cells were plated at 3000–5000 cells/well in 96-well plates and incubated at 37°C in an incubator containing 5% CO<sub>2</sub>. Cell viability was assessed using the CellTiter-Glo luminescent cell viability kit (G7572) from Promega Corporation according to the manufacturer’s instructions. The method is based on quantitation of the adenosine triphosphate present in the cells, detected by the generation of A luminescent signal. All experiments were repeated at least three times. ## Autophagic flux measurement and immunoblotting For more precise monitoring of autophagic activity, cells were treated with or without 20 μM chloroquine (CQ; Sigma-Aldrich, C6628), an inhibitor of autophagy, for 2 h prior to harvesting. The cell lysates were used to detect LC3-II accumulation by immunoblotting to verify autophagic flux. For immunoblotting, the cells were briefly rinsed in PBS (02-023-1, Biological Industries, City, Country) and lysed with RIPA buffer (1% NP40 \[MDBio, City, Country, 101-9016-45-9\], 50 mM Tris HCl, pH 7.5, 150 mM NaCl, 0.25% sodium deoxycholate \[Sigma-Aldrich, D6750\], 0.1% sodium dodecyl sulfate \[SDS; Calbiochem, City, Country, 428015\], and protease inhibitor cocktail \[Roche, City, Country, 11873580001\]). The cell lysates were resolved by sodium dodecyl sulfate–polyacrylamide gel electrophoresis and transferred electrophoretically onto nitrocellulose membranes. The membrane was blocked with 5% skim milk and then incubated with primary antibodies against LC3 (L7543), and ACTB (β-actin, A5441) (all from Sigma-Aldrich), SQSTM1 (BD Pharmingen, 610832), ATG5 (8540) and ATG7 (8558) (all purchased from Cell Signaling Technology) overnight at 4°C. The proteins were probed with an HRP-labeled secondary antibody (Santa Cruz, sc-2004 or sc-2005) and detected with an ECL reagent. The membrane was scanned and analyzed for the protein expression level with the ChemiDoc XRS Imaging System (Bio-Rad, Hercules, CA, USA). ## GFP-LC3 punctuation GFP-LC3 plasmids (1500 cells/40 μL) were purchased from Addgene and were transfected into human ARPE-19 cells or mouse 661W cells for stable selection with G418 (600 μg/mL). The stable cells were seeded with ICG or BBG in plates for 30 min and recovered for 6 h. The cells were then fixed with 3.7% paraformaldehyde and cell images were collected with a 40× objective for GFP-LC3 puncta. Next, GFP-LC3–labeled autophagosomes were detected with fluorescence spot numbers and areas in each cell after ICG or BBG treatment. The percentages of changes of GFP-LC3 were calculated. Alternatively, the cells expressing GFP- LC3 were exposed to vital dyes and trypsinized for cell harvest. The cells were washed with PBS once and then resuspended with 200 μL 0.05% saponin to remove nonlipidated GFP-LC3 from cells. One mL PBS was added to the cells to quantificate membrane-bound GFP-LC3-II with flow cytometry as reported previously. ## Transfection and shRNA infection For siRNA transfection, the cells were seeded at 20% to 30% confluence and reversely transfected with RNAiMAX (Life Technologies; 13778–150) in the presence of 10 nM scrambled siRNA (Life Technologies, 12935–112), siRNA against ATG5 (GE Healthcare Dharmacon, City, Country, 9474) or ATG7 (Life Technologies, s20652) or BECN1 (Life Technologies, 4392420) for 72 h. For shRNA infection, shRNAs against ATG5 (TRCN0000151963, The RNAi Consortium, Taiwan), ATG7 (TRCN0000007584, The RNAi Consortium, Taiwan), and pLKO.1 Scrambled shRNA (Addgene, 1864). The plasmids were transfected into HEK293FT cells with Lipofectamine 2000 (Life Technologies, 11668–027) for 2 days, and the supernatant were used to infect ARPE-19 cells. The infected ARPE-19 cells were selected with puromycin (1 μg/ml) for 10 days to obtain stable knockdowned cells. The cells were harvested for knockdown efficiency by immunoblotting. ## Statistical analysis All the data are expressed as the mean ± standard error of the mean from at least 3 individual experiments. The statistical analysis was performed using a nonparametric 2-tailed Student’s t test with Prism 5.0 (Graph-Pad, La Jolla, CA, USA). P values \<0.05 were considered significant (\* indicates P\<0.05, \\ indicates P\<0.01, and \\\* indicates P\<0.001). # Results ## ICG and BBG diminish cell viability and modulate autophagy in ARPE-19 and 661W cells ARPE-19 and 661W cells were exposed to various concentrations (0.05, 0.5, or 1.0 mg/mL) of ICG or BBG for 5 or 30 min. The treated cells were then recovered for 24 h to determine the cytotoxic effects of ICG and BBG with CellTiter-Glo. Both ICG and BBG reduced cell viability in ARPE-19 cells. In contrast with ICG, BBG treatment is relatively more cytotoxic to ocular cells. Because the cells treated with ICG and BBG at the dose of 0.05 mg/mL for 30 min reached the minimal cytotoxicity, these conditions were used in the subsequent experiments. Moreover, to corroborate whether ICG and BBG modulate autophagy in ocular cells, ARPE-19 and 661W cells expressing GFP-LC3 were exposed to ICG and BBG to examine the membrane-bound GFP-LC3 puncta (GFP-LC3-II) with confocal microscope. ICG and BBG increased the GFP-LC3 puncta in both ARPE-19 and 661W cells. The GFP- LC3–expressing cells treated as above were also rinsed with the detergent saponin to remove nonlipidated GFP-LC3-I, and retained GFP-LC3-II, likely autophagosome associated form, were analyzed by flow cytometry. As with the results of fluorescent microscopy, the flow cytometry results showed that ICG and BBG increased lipidated GFP-LC3-II in ocular cells. Further, the cells treated with vital dyes were harvested to detect the level of the autophagy marker LC3 and adaptor SQSTM1 with immunoblotting. ICG increased the ratio of LC3-II/actin and the SQSTM1 level in ARPE-19, suggesting ICG and BBG may inhibit autophagy in ARPE-19 according to the guideline of autophagy assays reported previously. Although SQSTM1 was undetectable in 661W cells, the effects of vital dyes on lipidated LC3-II in 661W cells were similar to those in ARPE-19, suggesting that ICG and BBG modulate autophagy in ocular cells. ## ICG and BBG inhibited autophagic flux in ARPE-19 cells but induced autophagic flux in 661W cells Autophagy induction transiently increases the GFP-LC3-II puncta and LC3-II/actin ratio in cells, whereas a block in autolysosomes causes an accumulation of GFP- LC3-II puncta and increases the LC3-II/actin ratio. To precisely evaluate the autophagic effects of the vital dyes ICG and BBG in ARPE-19 cells and 661W cells, these cells were exposed to ICG and BBG in the presence or absence of the autophagy inhibitor CQ. The net difference in LC3-II levels between cells treated with and those without CQ was used to determine autophagic flux in the treated cells. Like the autophagy inhibitor CQ, ICG and BBG significantly increased the SQSTM1 protein level in ARPE-19 cells. LC3-II flux was significantly decreased in ARPE-19 cells treated with ICG or BBG. Conversely, ICG and BBG increased the LC3-II flux in photoreceptor 661W cells compared with the control cells, indicating that vital dyes may suppress autophagy in retinal pigment epithelial ARPE-19 cells but induce autophagy in photoreceptor 661W cells. These results suggest vital dyes differentially modulate autophagy, which may depend on cell types. ## Autophagy protected ocular cells from ICG or BBG-induced damage Autophagy can be protective or detrimental for cells in response to stresses. We examined the role of autophagy in ocular cells after treatment with ICG and BBG. Autophagy ablation with CQ notably enhanced the cytotoxicity of ICG and BBG in ARPE-19 and 661W cells. Knockdown of ATG5 or ATG7 or BECN1, which are essential genes for autophagy signaling, with shRNA or siRNA enhanced cytotoxic effects of ICG and BBG in ARPE-19 cells. Additionally, we examined the effect of three dietary supplements on autophagy in ocular cells, including resveratrol, lutein, and CoQ10. Among these dietary supplements, resveratrol has been reported to be an inhibitor of mechanistic target of rapamycin (mTOR) that induces autophagy. As shown in, resveratrol, lutein, and CoQ10 induced autophagic activity in both retinal pigment epithelial ARPE-19 cells and in photoreceptor 661W cells. Resveratrol and CoQ10 are reportedly to facilitate cell proliferation in neural progenitor cells and fibroblast, respectively. Indeed, pretreatment with the dietary supplements increased cell viability, likely due to promotion of cell proliferation as mentioned above. Furthermore, resveratrol, lutein, and CoQ10 diminished the cytotoxic effects of ICG and BBG in ocular cells compared to the untreated cells. These results suggest that autophagy may play a cytoprotective role in ocular cells in response to the damage of ICG and BBG. # Discussion Vital dyes have been reported to cause cytotoxicity in RPE cells, likely through mitochondrial damage and ROS production. The role of autophagy in survival and death in ocular cells during stresses has been documented. However, the involvement of autophagy in vital dyes–induced cytotoxicity in ocular cells remains unclear. Our study is the first to show that ICG and BBG induced cytotoxicity and modulated autophagy in both ARPE-19 and 661W cells. Interestingly, ICG and BBG inhibited autophagic flux in ARPE-19 cells, whereas the vital dyes induced autophagic flux in 661W cells. In addition, genetic and pharmacological ablation of autophagy diminished cell viability in both ARPE-19 and 661W cells. Dietary supplements, including resveratrol, lutein, and CoQ10, induced autophagy and diminished the cytotoxic effects of vital dyes in the ocular cells. Autophagic response in the same stimuli might be varied in different cells. For example: doxorubicin induces autophagy in rat cardiac myocytes, whereas it reduces autophagy in mouse cardiac myocytes. However, the mechanisms for differential autophagy regulation in difference cells remain unknown. Similarly, one of the interesting findings in our study is that ICG and BBG inhibited autophagic flux in ARPE-19 cells, whereas the vital dyes induced autophagic flux in 661W cells. Thus, the discrepancy in the responses to dyes between ARPE-19 and 661W cells may be through the different signaling pathways. In addition, ROS play a dual role on autophagy in various cells. Autophagy is induced in cells in response to oxidative stress via several factors, such as Protein kinase RNA-like endoplasmic reticulum kinase (PERK), Hypoxia-inducible factor 1 (HIF1), p53, nuclear factor erythroid-2-related factor (NRF2), ATG4, and Forkhead box O3 (FOXO3). Conversely, ROS also down- regulates ULK1, a homologue of yeast ATG1, to inhibit autophagy via phosphorylation of p53 in NB4 cells. We also noticed that ROS was induced by ICG and BBG in both ARPE-19 and 661W cells. Nevertheless, elucidating these mechanisms requires further work to identify if ROS is involved in differential effects of vital dyes on autophagy in different ocular cells. On the other hand, autophagy can reduce oxidative stress through the nuclear factor erythroid 2–related factor 2 (NRF2)/kelch-like ECH-associated protein 1 (KEAP1) and the SQSTM1 pathway. NRF2 is a transcription factor of the leucine zipper family and can activate antioxidant-defense genes such as glutathione peroxidase, superoxide dismutase, and thioredoxin. Under normal conditions, NRF2 is sequestered by KEAP1 ubiquitination and degradation. Following oxidative stress, either the ubiquitin activity of KEAP1 is inhibited or SQSTM1 interacts with KEAP1 for its degradation, leading to the liberation and activation of NRF2, which in turn limits ROS production and cell damage. These results raise the possibility that the vital dyes may initially inhibit autophagy to elevate ROS and cause oxidative damage in ARPE-19 cells. The RPE is a single layer of epithelial cells lining the posterior segment of the eye. It is located between the light-sensing photoreceptors and the choriocapillaris and serves essential function for vision. A failure of any one of these functions can lead to degeneration of the retina, loss of visual function, and blindness. Our results suggest that both ICG and BBG at the routine dose used in chromovitrectomy can affect autophagy not only RPE but also photoreceptor cells. Autophagy ablation with CQ or silencing ATG genes notably enhanced the cytotoxicity of ICG and BBG in ocular cells, indicating autophagy might be a survival mechanisms in ocular cells exposed to vital dyes. Moreover, Autophagy is a stress responsive pathway to limit damage-induced death, whereas autophagy may not be sufficient to block death in cells during excessive stresses. For example: autophagy is enhanced in cardiac cells during ischemia/reperfusion. Pretreatment with autophagy inducer rapamycin protects cells from ischemia/reperfusion injury. Besides, many antioxidants are reportedly to promote cell proliferation and resistant to oxidative stress- induced cell death, Three dietary supplements, which are reported to be antioxidants, induced autophagic activity and increased cell growth in both ARPE-19 cells and photoreceptor 661W cells. Pretreatment with the dietary supplements diminished the cytotoxic effects of ICG and BBG in ocular cells compared to the cells that did not receive the dietary supplements. Several mechanisms may be involved in cytoprotective effects of the dietary supplements in cells exposed to ICG or BBG, such as reduction of oxidative stress and promotion of mitochondrial biogenesis. Our current results imply that dietary supplements-induced autophagy could be one of the mechanisms for protecting ocular cells from stress caused by ICG and BBG. In conclusion, autophagy plays a protective role in ARPE-19 and 661W cells treated with vital dyes. The dietary supplements studied may protect the ocular cells against the vital dye–induced cytotoxicity via induction of autophagy. # Supporting information ICG Indocyanine green BBG brilliant blue G ILM internal limiting membrane LC3 microtubule associated protein 1 light chain 3 SQSTM1 sequestosome 1 ATG autophagy-related BECN1 beclin 1 CQ chloroquine RPE retinal pigment epithelial cells ROS reactive oxygen species [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: SJS CWS. Data curation: SHL YAC. Formal analysis: CWS YSB. Funding acquisition: SJS CWS. Investigation: JLC YSB. Methodology: SJS CWS. Project administration: SJS CWS. Resources: JLC YSB. Software: CWS. Supervision: SJS CWS. Validation: JLC YSB. Visualization: SJS CWS. Writing – original draft: SJS CWS. Writing – review & editing: SJS CWS.
# Introduction Ophiobolins (OPs) are a family of sesterpenoid phytotoxic metabolites produced mainly by fungi of the Bipolaris genus, pathogen of rice, maize and sorghum. In plants, OPs induce a wide array of toxic effects, such as inhibition of coleoptiles and root growth and seed germination inhibition. At the cellular level, OPs affect membrane permeability, causing electrolytes and nutrients leakage, proton extrusion inhibition, decrease of CO<sub>2</sub> photosynthetic assimilation and inhibition of protein and nucleic acid synthesis. In the last years, accumulating evidence demonstrates that OPs are able to affect cellular functions also in mammalian cells. Ophiobolin A (OP-A) results strongly cytotoxic to mouse leukaemia cells, where it induces shrinkage of cell soma, chromatin condensation and DNA laddering, typical features of apoptotic cell death. Moreover, ophiobolin O (OP-O) from Aspergillus ustus induces apoptosis in multidrug-resistant MCF-7 breast cancer cells.. By contrast, OP-A displays the same cytostatic effect on both apoptosis- sensitive and apoptosis-resistant cancer cells, whereas in human glioblastoma cells it is able to induce cell death, through a paraptosis-like mechanism. Melanoma is a highly malignant tumour induced by transformation of melanocytes, whose incidence rate is rapidly increasing in the world. Due to its high resistance to cytotoxic agents, metastatic melanoma has a very poor prognosis. Therefore, finding new anti-cancer molecules able to integrate or enhance chemical treatments of drug-resistant tumours such as melanoma is a relevant research issue. In the present study we characterized the OP-A effects on A375 (BRAF V600E) and CHL-1 (BRAF wt) melanoma derived cell lines, as compared to the HaCaT (immortalised keratinocytes) cell line. To this purpose, we analysed cell viability, nuclear and mitochondria morphology and functionality, cell death induction, as well as cell cycle progression. Finally, we performed a comparative proteomic analysis on A375 cell line treated with OP-A. # Materials and Methods ## Cell culture and treatments A375 human melanoma cell line was grown in RPMI 1640 medium (Lonza, Switzerland) supplemented with 2 mM L-glutamine (Thermo Fisher Scientific, MA, USA), CHL-1 human melanoma and HaCaT immortalised human keratinocytes cell lines were grown in DMEM medium (Lonza), both supplemented with 10% Foetal Bovine Serum (FBS, Thermo Fisher Scientific), and penicillin/streptomycin (Sigma Aldrich, MO, USA) in an humidified 5% CO<sub>2</sub> atmosphere at 37°C. Cell treatments: 1x10<sup>5</sup> or 2x10<sup>6</sup> cells were seeded in 12 wells plates or 100 mm dishes and the next day treated with the indicated amount of OP-A, diluted in fresh culture medium, for the indicated times. For the necrostatin treatments cells were incubated for 2 h with 20 μM necrostatin-1 (Santa Cruz Biotechnology, TX, USA) in complete medium, before the addition of OP-A. ## MTS viability assay Cell viability was assessed by Acqueous One Solution Proliferation Assay (MTS assay, Promega, WI, USA), following the manufacturer indications. ## Western blotting Whole cell extracts were prepared by lysis in RIPA buffer (50 mM Tris-HCl, pH 7.4, 150 mM NaCl, 0.5% Na-deoxycolate, 0.1% SDS, 1% NP-40, 2 mM Na<sub>2-</sub>EDTA), supplemented with protease inhibitors (Roche, Germany). Protein concentration was determined by Bio-Rad protein assay (Bio-Rad, CA, USA) and 10–25 μg of proteins were separated on 4–12% Nu-PAGE pre-cast gels (Thermo Fisher Scientific). After blotting on PVDF and 1 h saturation in PBS containing 0.05% Tween-20 and 5% skim milk, membranes were incubated for 1 h or overnight with primary antibody, diluted in PBS containing 0.05% Tween-20 and 0.5% skim milk, washed three times for 10 min in PBS containing 0.05% Tween-20, incubated for 1 h with the appropriate horseradish peroxidase-conjugated secondary antibody (Bio-Rad) and the signals detected with Chemiglow by means of a FluorChem SP system (Alphainnotech, Germany). Primary antibodies were against: PARP, (BioMol, Germany, 1 μg/ml), Caspase 3 (9662, Cell Signaling, MA, USA, 1 μg/ml), Caspase 9 (9502, Cell Signaling, 1 μg/ml), LC3 (2775, Cell Signaling, 1 μg/ml), LC3B (D11 XP, Cell Signaling, 1 μg/ml), PINK1 (D8G3, Cell Signaling, 1 μg/ml), BAX (2D2 and N-20, Santa Cruz, 0.5 μg/ml), BAK (N-20, Santa Cruz, 0.5 μg/ml), cytochrome c (556432, Becton Dickinson, NJ, USA, 1 μg/ml). β-Tubulin (Sigma Aldrich, 1 μg/ml) was used as a loading control for cell extracts. ## Mitochondrial imaging, mitochondrial membrane potential, mitochondrial mass, lysosome contents, and mitochondrial reactive oxygen species (ROS) measurement Mitochondrial network imaging was performed by incubating untreated and treated cells for 20 min at 37°C with 1 μM MitoTracker Red CMXRos regent (Thermo Fisher Scientific) in RPMI medium and nuclei counterstained with 1 μM HOECHST 33342 (Thermo Fisher Scientific). Images were captured by means of a Floid Instrument (Thermo Fisher Scientific). For mitochondrial membrane potential, mitochondrial mass, lysosome content and mitochondrial ROS measurement, cells were detached with trypsin and, after washing with PBS, 1x10<sup>6</sup> cells were incubated for 30 min at 37°C with 2 μM JC-1 (5,5’,6,6’-tetrachloro-1,1’,3,3’tetraethylbenzimidazolylcarbocyanine iodide), or 100 nM MitoTracker Green FM (Thermo Fisher Scientific), or 100 nM LysoTracker Red DND 99 (Thermo Fisher Scientific), or 5 μM MitoSOX Red (Thermo Fisher Scientific), diluted in pre-warmed complete medium. At the end of the incubation, samples were acquired by means of a FACSCalibur (Becton Dickinson). ## Flow cytometry analysis of cell death Cells, untreated or treated with OP-A for the indicated times, were detached with trypsin, washed twice with PBS and after suspension in propidium iodide solution (50 μg/ml propidium iodide, 0,1% Triton X-100, 0,1% Na-citrate in PBS), incubated 8 h or overnight at 4°C in the dark. Samples were acquired by means of a FACSCalibur and flow cytometry data analysed with the FlowJo software (TreeStar, [www.flowjo.com](http://www.flowjo.com)). ## 2D electrophoresis Cells were treated with 0.6 μM OP-A for 24 h. After incubation, cells were detached with trypsin, washed twice with ice-cold PBS and centrifuged. Cell pellet was resuspended in Lysis Buffer (50 mM Tris-HCl pH 7.4, 1% Triton-X-100, 250 mM NaCl, 5 mM EDTA) and incubated overnight at 4°C. Cell lysates were centrifuged at 14,000 x g for 20 min and protein concentration in the supernatant was determined by Bradford assay. Equivalent protein amounts (300 μg) of control and treated cell samples were desalted by precipitation with cold ethanol (overnight at -20°C). Precipitates were centrifuged at 15,000 x g for 15 min. Protein pellets were dissolved in IEF buffer (9 M urea, 4% w/v CHAPS, 0.5% v/v Triton X-100, 20 mM DTT, 1% w/v Bio-Rad carrier ampholytes pH 3–10 NL). Protein concentration was estimated by using the Bradford assay, modified according to Ramagli and Rodriguez. IPG strips (17 cm pH 3–10 NL, Bio-Rad ReadyStrip) were rehydrated overnight with IEF buffer containing 350 μg of total proteins. Proteins were focused using a Protean IEF Cell (Bio-Rad) at 12°C, by applying the following voltages: 250 V (90 min), 500 V (90 min), 1000 V (180 min) and 8000 V for a total of 52 KVh. After focusing, proteins were reduced by incubating the IPG strips with 1% w/v DTT in 10 ml of equilibration buffer (50 mM Tris-HCl pH 8.8, 6 M urea, 30% w/v glycerol, 2% w/v SDS and a dash of bromophenol blue), for 20 min, and then alkylated with 2.5% w/v iodoacetamide in 10 ml of equilibration buffer, for 20 min. Electrophoresis in the second dimension was carried out on 12% T polyacrylamide gels (180 x 240 x 1 mm) running on a Protean apparatus (Bio-Rad) in 25 mM Tris-HCl pH 8.3, 1.92 M glycine and 1% w/v SDS, with 120 V (for 12 h), until the dye front reached the bottom of the gel. 2-DE gels were then stained with colloidal Coomassie G250; resulting images were acquired by using a GS-800 imaging systems (Bio-Rad). For quantitative analysis, each biological sample was analyzed in technical triplicates. ## Gel image analysis Digitalized images of Coomassie-stained gels were analyzed by using the PD Quest (vers. 7.3.1) 2-D analysis software (Bio-Rad), which allowed spot detection, landmarks identification, aligning/matching of spots within gels, quantification of matched spots and their analysis, according to manufacturer's instructions. Manual inspection of the spots was performed to verify the accuracy of automatic gel matching; any error in the automatic procedure was manually corrected prior to the final data analysis. The spot volume was used as the analysis parameter for quantifying protein expression. The protein spot volume was normalized to the spot volume of the entire gel (i.e., of all the protein spots). Fold-changes in protein spot levels were calculated between spot volumes in the treated group, relative to that in the control gels. Statistically significant changes in protein expression were determined by using two sequential data analysis criteria. First, a protein spot had to be present in all gels for each sample to be included in the analysis. Next, statistically significant changes in protein expression were determined by using the distribution of fold-change values in the data. Spots were determined to be statistically significant if the difference between the average intensity of a specific protein spot in the treated and control cells (three technical replicates of three biological samples) was greater than one standard deviation of the spot intensities for both groups. An absolute two-fold change in normalized spot densities was considered indicative of a differentially represented component; values 2 or 0.5 were associated with increased or decreased protein amounts after treatment, respectively. ## Protein digestion and mass spectrometry analysis Spots from two-dimensional electrophoresis were manually excised from gels, triturated and washed with water. Proteins were in-gel reduced, S-alkylated and digested with trypsin, as previously reported. Protein digests were subjected to a desalting/concentration step on microZipTipC18 devices (Millipore, Bedford, MA, USA). Peptide mixtures were then analyzed by nano-liquid chromatography coupled to electrospray-linear ion trap-tandem mass spectrometry (nanoLC-ESI- LIT-MS/MS) using a LTQ XL mass spectrometer (Thermo Fisher Scientific) equipped with Proxeon nanospray source connected to an Easy-nanoLC (Proxeon, Denmark). Peptide mixtures were separated on an Easy C18 column (100 x 0.075 mm, 3 μm) (Proxeon) using a gradient of acetonitrile containing 0.1% formic acid in aqueous 0.1% formic acid; acetonitrile ramped from 5% to 35% over 15 min, and from 35% to 95% over 2 min, at a flow rate of 300 nl/min. Spectra were acquired in the range m/z 400–2000. Acquisition was controlled by a data-dependent product ion scanning procedure over the three most abundant ions, enabling dynamic exclusion (repeat count 2 and exclusion duration 1 min). The mass isolation window and the collision energy were set to m/z 3 and 35%, respectively. ## Protein identification MASCOT software package version 2.2.06 (Matrix Science, [www.matrixscience.com](http://www.matrixscience.com)) was used to identify proteins within spots from an updated human non-redundant sequence database (NCBI 2014/12). NanoLC-ESI-LIT-MS/MS data were searched by using a mass tolerance value of 2 Da for precursor ion and 0.8 Da for MS/MS fragments, trypsin as proteolytic enzyme, a missed cleavages maximum value of 2, and Cys carboxamidomethylation and Met oxidation as fixed and variable modification, respectively. Protein candidates with at least 2 assigned unique peptides with an individual MASCOT score \>25, both corresponding to p \< 0.05 for a significant identification, were further evaluated by the comparison with their calculated Mr and pI values, using the experimental ones obtained from two- dimensional electrophoresis. ## Data analysis All the data reported were verified in at least six different replicates and are reported as mean ± SEM. Statistical analysis on flow cytometry data was performed by means of ANOVA and Bonferroni post-test. Analysis of gel spot quantitative differences was carried out using the Student's t test. # Results ## OP-A induced cell death in human melanoma cells by activating the mitochondrial pathway of apoptosis We first assessed the effect of OP-A on A375 human melanoma cell line, by analyzing cell viability by means of MTS assay in a dose-response treatment (0.3, 0.6, 1.2 μM OP-A) for 24 and 48 h. shows that administration with 0.3 μM OP-A for 24 h was effective in reducing cell metabolic activity at about 60% of control cells; metabolism was quite completely inhibited by 0.6 and 1.2 μM. The effect was slightly increased after 48 h administration. These results suggested that OP-A treatment resulted in an impairment of mitochondrial functionality, an effect that could lead to the induction of the intrinsic pathway of apoptosis. To verify this hypothesis, we then analyzed mitochondrial and nuclear morphology of cells treated as above using MitotrackerRed and Hoechst staining. shows that OP-A did not lead to the appearance of the cytoplasmic vacuolization observed during paraptosis induction, but resulted in a fragmentation of the mitochondrial network, even at the lowest dose used, and in the clustering of mitochondria. Moreover, we detected the appearance of picnotic and apoptotic nuclei (black and white arrowheads) when cells were treated with 0.6 and 1.2 μM OP-A. Taken together, these results suggest that OP-A treatment of A375 cell line might lead to apoptosis induction through the activation of the mitochondrial pathway, and prompted us to expand our analysis to other melanomas and normal cell lines. To this purpose we compared OP-A-dependent cell death induction of A375 (BRAF V600E) with CHL-1 (BRAF wt) melanoma, and HaCaT (immortalised keratinocytes) cell lines., shows that OP-A was able to induce cell death in the three cell lines. Nevertheless, treatment with 0.3 and 0.6 μM OP-A did not induce significant cell death levels in the HaCaT cell line, whereas these doses were effective in the two melanoma cell lines. At the highest dose tested (1.2 μM), OP-A caused massive cell death in all the three cell lines. Western blot analysis of proteins involved in apoptosis, revealed the activation of the mitochondrial apoptotic pathway. In fact, we could detect PARP and caspase-9 and -3 cleavage (quantified), respectively. In addition, we checked for the induction of autophagy using the specific marker LC3II, whose levels increase upon LC3I cleavage. shows that OP-A induced autophagy in all the three cell lines, as demonstrated by the increase of LC3II/LC3I ratio (panel B). Densitometric analysis revealed that autophagy induction was particularly efficient in the A375 cell line. In order to exclude the possibility that other types of death could be responsible for the effects we observed, we performed the same OP-A treatments in presence of 20 μM necrostatin-1., right panel, shows that necrostatin-1 did not inhibit OP-A-induced cell death, thus excluding the possible induction of necrosis or necroptosis. The OP-A-dependent mitochondrial perturbation we observed in the A375 cell line should be coupled not only to mitochondrial network rearrangements, but also to other alterations such as loss of mitochondrial membrane potential, which could result in mitophagy induction, and ROS generation. To verify these hypothesis, we first analysed mitochondrial membrane potential, mitochondrial mass, lysosome content and mitochondrial ROS production, by means of flow cytometry after staining with JC-1, MitoTracker Green, LysoTracker Red or MitoSOX Red, respectively. shows that OP-A treatment resulted in a reduction of mitochondrial membrane potential in all the three cell lines. As far as the HaCaT cell line, significant reduction of membrane potential occurred already at 0.6 μM OP-A, with a complete loss of the potential at 1.2 μM. On the other hand, the two melanoma cell lines showed a significant reduction of membrane potential only at 1.2 μM. Mitochondrial damages resulting in loss of membrane potential are frequently coupled to mitophagy induction, in order to remove the damaged organelles. For this reason, we measured mitochondrial mass and lysosome content. Surprisingly, even if we observed significant induction of autophagy we were not able to detect any significant decrease of the mitochondrial mass, and for the HaCaT cell line we revealed a significant increase of the mass in samples treated with 0.6 and 1.2 μM OP-A. Nevertheless, we observed a significant reduction of the lysosome content in cells treated with 1.2 μM OP-A, thus suggesting autophagy (mitophagy) induction. We measured also mitochondrial-derived ROS, by means of MitoSOX Red. shows that OP-A induced ROS production in the HaCaT and A375 cell lines, while no significant ROS levels were detected for the CHL-1 cell line. Taken together these results suggest that OP-A damaged mitochondria, causing loss of mitochondrial membrane potential and ROS production, even though at different extents in the three cell lines. As far as the observed increase of mitochondrial mass, we could speculate that the reported effect of OP-A on the Endoplasmic Reticulum (ER) dilation coupled with alteration of the cell’s membranes permeability, could determine MitoTracker Green accumulation also in the ER, thus leading to an apparent increase in the mitochondrial mass. In order to confirm the occurrence of mitochondrial damage and induction of the mitochondrial pathway of apoptosis, we performed a western blot analysis on different cellular fractions of HaCaT, CHL-1 and A375 cell lines treated with OP-A for 24 h. shows the results of the analysis of PINK1, a marker of mitochondrial depolarization, Bax, Bak and cytochrome c, markers of the mitochondrial pathway of apoptosis. OP-A induced significant PINK1 accumulation on the mitochondria in all the three cell lines, thus confirming the mitochondrial damage and induction of mitophagy. Furthermore OP-A caused relocalisation to mitochondria of Bax and Bak proteins, as well as, the release of cytochrome c in the cytosol. The activation of the mitochondrial pathway of apoptosis in melanoma cell lines was also corroborated by the analyses of the autophagy marker LC3II, which was increased in all the three cell lines upon OP-A treatment. ## OP-A induced alteration of the cell cycle In order to test the effect of OP-A on cell cycle progression, we analysed the distribution of cells in each phase of the cycle after 24 h of OP-A administration, by means of flow cytometry after propidium iodide staining. As shown in, OP-A induced alterations of the cell cycle in all cell line tested. As far as the HaCaT cell line, the only significant alteration was an increase of the percentage of cells in the G1 phase, from 47% (ctrl) to 57% (0.6 μM, p\<0.01), coupled to a decrease of the percentage of cells in the S phase, from 33% (ctrl) to 22% (0.6 μM, p\<0.01). A similar behaviour was observed for the CHL-1 cell lines, but at a lower dose. In fact, OP-A induced an increase of the percentage of cells in the G1 phase, from 61% (ctrl) to 75% (0.3 μM, p\<0.05), coupled to a decrease of the percentage of cells in the S phase, from 28% (ctrl) to 15% (0.3 μM, p\<0.05). Differently, in the A375 cell line, OP-A induced alteration of the cell cycle at the highest doses of the treatment. The percentage of cells in the G2/M phase, increased from 8% (ctrl) to 12% (0.6 μM, p\<0.05) and to 12% (1.2 μM, p\<0.001), coupled to a decrease of the percentage of cells in the G1 phase, from 57% (ctrl) to 50% (1.2 μM, p\<0.001). Taken together, our results suggest that prolonged exposure to OP-A leads to an impairment of mitochondrial functions, which are reflected in the alteration of cell cycle progression, induction of autophagy and ultimately in cell death by apoptosis, even if with different timings and modalities in the three cell lines. ## Effect of OP-A treatment on the protein repertoire of melanoma cells In order to obtain information about the molecular mechanisms underlying OP-A effect, a proteomic analysis on the A375 cell line was carried out. Total proteins were extracted from control or 0.6 μM OP-A treated cells and resolved by two-dimensional gel electrophoresis (2-DE). To detect quantitative changes in relative protein spot volume, colloidal Coomassie-stained gels were subjected to software-assisted image analysis. Statistical evaluation of the relative volumes allowed to detect spots whose representation varied significantly (p \< 0.05). The 2-D Master gel of the A375 human melanoma cell proteome is shown in , upper panel, where encircled spots indicate differentially represented proteins. The overall 2-DE profiles of control and treated cells were very similar; however, 24 protein spots whose abundance varied at least two-fold in response to OP-A challenge were detected. These differentially represented spots were excised from the gel, proteolyzed and subjected to MS analysis. Database searching with results from nanoLC-ESI-LIT-MS/MS experiments allowed the identification of the proteins migrating within these spots. The list of the identified polypeptides is reported in. Globally, spots assayed were associated with 16 non-redundant protein entries. Analysis of spot 1, 11, 17, 18, 21 and 24 resulted in a multiple identification, and they were not further discussed. One protein, peptidyl prolyl isomerase was present in multiple spots (4–6) whose structural differences were not further characterized; probably, they resulted from post-translational modifications or sequence-related isozymes. Fifteen proteins resulted down-regulated, whereas 1 was up-regulated. Functional categorization according to Gene Ontology annotation and literature data (data not shown), showed that differentially- represented proteins grouped into different functional categories, including components involved in glucose metabolism, protein folding, mitochondrial transport, cytoskeleton organization and cell proliferation. Reprogramming of metabolism, involving enhanced glycolysis is a hallmark of cancer. Hence, down-regulation of glycolytic enzymes can contribute to reduce energy fuelling to cancer cells. Two enzymes of the glycolytic metabolism have been identified, namely fructose 1,6 bisphosphate aldolase A (ALDOA, spot 3) and triose phosphate isomerase (TPI, spots 13 and 15). Notably, ALDOA is highly expressed in a variety of malignant cancers, including renal cancer, human lung squamous cell carcinoma, and hepatocellular carcinomas, suggesting it could promote cancer growth by enhancing glycolysis. It is worth to note that OP-A has been described as a calmodulin inhibitor and that calmodulin antagonists induce a reduction of ALDOA levels, coupled to cell death in melanoma cell lines. TPI catalyzes the conversion of dihydroxyacetone phosphate into D-glyceraldehyde 3-phosphate, a key substrate of the glycolytic pathway. This enzyme is up- regulated in different types of lung and urinary cancers and in ovarian carcinoma. In etoposide-treated HeLa cells, inhibition of TPI by cyclin A/Cdk2 phosphorylation hampers energy production, thereby inducing apoptosis. Interestingly, a recent body of evidence supports the idea that glycolytic enzymes play also non-glycolytic roles, which are essential for promoting cancer cell proliferation and chemoresistance. Six proteins involved in protein folding were down-regulated by OP-A, namely peptidyl-prolyl cis-trans isomerase A (spots 4–6), prefoldin (spot 7), heat shock protein (HSPA8) (spot 16), mitochondrial 60 kDa and 10 kDa heat shock proteins (spot 19, 23), and tapasin (spot 22). Peptidyl-prolyl cis-trans isomerases A (PPIase A) belongs to the cyclophilins class of PPIases. Recently, evidence of the involvement of members of this family in cancer development has been reported. In fact, knockdown of cyclophylin A reverses chemoresistance in endometrial cancer cells, while the over-expression of peptidyl-prolyl isomerase-like 1 is associated with growth of colon cancer cells. Prefoldins (PFDs) are hetero-oligomer chaperones involved in cancer development, since their main targets are actin and tubulin. Up-regulation of members of the PFD protein family has been reported in different tumours, such as glioblastoma, breast, pancreatic, colon, bladder and cervical carcinomas. The heat shock proteins (HSPs) have been reported to be significantly elevated in many kind of human cancers and their over-expression has been correlated to therapeutic resistance and poor survival. HSP71 (HSPA8, spot 16) is a constitutively expressed isoform of the HSP70 protein family, essential for normal protein homeostasis in unstressed cells. In addition to facilitating folding, HSC70 is involved in the degradation of misfolded proteins. Although HSC70 is over- expressed in cancer cells, little is known about how it contributes to their survival. A recent proteomic investigation of human colon cancer cells showed that HSC70 interacts with and prevents the degradation of key proteins involved in cancer survival. As far as mitochondrial HSPs isoforms (spot 19, 23), they can enhance tumourigenicity by promoting the retention into the mitochondria of pro-apoptotic factors. Mitochondrial voltage-dependent anion channel 1 (VDAC-1, spot 2) is the most abundant and physiologically relevant isoform of the VDAC family of pore-forming proteins, located at the outer membrane of mitochondria (OMM). VDACs are voltage-sensitive ion channels that regulate the energy flux across the OMM by facilitating the diffusion of key molecules such as nucleotides, pyruvate and malate. Altering VDAC-1 levels may lead to opposite effects in tumour cells. In fact, VDAC-1 favours the release of pro-apoptotic proteins from the mitochondrial inter-membrane space to the cytosol. On the other hand, VDAC-1 sustains the high glycolytic rates of tumour cells by associating with hexokinase, and inhibits the formation of the mitochondrial permeability transition pore (MPTP) in the OMM, which is necessary for the release of pro- apoptotic factors. Recently, a systematic analysis demonstrated that VDAC-1 is up regulated in breast, colon, liver, lung, pancreatic, and thyroid tumour tissues. VDAC-1 is considered a marker of mitochondrial mass, and its decrease in OP-A treated A375 cells fairly correlates with the mitochondrial damage and reduction of mitochondrial mass due to mitophagy induction. Three proteins involved in cytoskeleton dynamics regulation have been found as differentially represented after treatment with OP-A, namely Pr22 protein (spot 8) and calumenin (spot 20) (both down-represented), and P1725 protein (spot 9) (over-represented). Microtubule destabilization is a crucial process during cancer progression. Pr22 protein is over-expressed in different highly malignant tumours, such as breast or ovarian cancers and leukaemias, whereas its reduction can reverse the malignant phenotype. Calumenin is involved in the regulation of cytoskeleton and localization of intracellular proteins. In a recent proteomic study concerning the effect of the anti-tumoural natural drug gambogic acid on breast carcinoma cells, levels of calumenin were consistently decreased. Other proteins with different functions were observed as down-represented after treatment with OP-A. Among that, translationally-controlled tumour protein (TCTP, spot 12) that is an ubiquitous eukaryotic component involved in pivotal cell functions like growth, cell cycle, apoptosis and stem cells pluripotency. TCPT is over-represented in diverse tumours, and its down-regulation hampers tumour viability. Recent data indicate that TCTP has a chaperone-like structure that makes it able to interact with anti-apoptotic proteins like Bcl-xL, thereby functioning as a pro-survival factor in cancer cells. # Discussion In this study, we analysed the effect of OP-A on A375 (BRAF V600E) and CHL-1 (BRAF wt) human melanoma cell lines, as compared to the HaCaT immortalised keratinocytes cell line. In fact, despite a growing body of evidence indicates that OPs possesses anti-proliferative activity towards different cancer-derived cell lines, the specific mechanism underlying this action remains unclear. Our results demonstrated that OP-A induced mitochondrial network fragmentation, membrane potential dissipation and mitochondrial ROS production, leading to induction of autophagy and ultimately resulting in activation of the mitochondrial pathway of apoptosis. Although flow cytometry analysis did not allow to show any significant reduction of mitochondrial mass, probably due to dye accumulation also in swelled ER, evidence of mithocondrial damage was obtained by western blot analysis. In fact, the strong increase in the mitochondrial levels of PINK1 protein, a marker of mitochondrial membrane potential depolarization, and of the autophagy marker LC3II, strongly put forward the occurrence of mitochondrial damage and induction of autophagy. A further piece of evidence resulted from proteomic investigation of A375 cell line, in which a dramatic reduction of the mitochondrial mass marker VDAC-1 was observed. Mitochondrial damage and consequent mitophagy ultimately resulted in cell death by the activation of mitochondrial pathway of apoptosis, as indicated by the analysis of specific markers. In fact, western blot analysis demonstrated caspase-9 and -3 activation, PARP cleavage, as well as relocalisation of the pro-apoptotic members of the Bcl-2 family Bax and Bak. The activation of the mitochondrial pathway of apoptosis was further supported by cytochrome c release in the cytosol. Prolonged exposure to OP-A leads to an impairment of mitochondrial functions, induction of autophagy and ultimately in cell death by apoptosis in both melanoma cell lines, even though with different timings and modalities in different cell lines. In particular, CHL-1 cells appeared more sensitive to OP-A. This difference could be related to survival signalling pathways activated by the BRAF V600E mutation present in the A375 cell line. Furthermore, the proteomic approach on A375 cell line allowed us to identify down-regulation of different proteins, essential to maintain cell integrity and viability, whose levels normally increase in different types of cancer. Taken together, our results demonstrate that OP-A possess a strong cytotoxic activity on melanoma cells, even on recalcitrant A375 cell line, where it is able to induce mitochondrial damage and cell death by apoptosis. Under this respect, it is noteworthy that OP-A is effective at nanomolar/micromolar concentrations after 24 h of treatment, whereas the BRAF inhibitor vemurafenib, the elective drug for treatment of recalcitrant melanomas is not (data not shown). However, the mechanism of action of the toxin deserves further investigation. In fact, even though mitochondria appear as the primary site of action, the apparent lack of mitochondrial mass reduction and the proteomic identification of many down-regulated proteins regulating fundamental cellular processes suggest that other cell death mechanisms may be involved, such as for instance the ER stress. In conclusion, even further investigation is needed, OP-A appears as a promising molecule to be studied for its potential use in the treatment of melanoma, possibly in association to other type of anti-cancer agents. We thank prof. Maurizio Mattei, University of Rome Tor Vergata, for providing the A375 human melanoma cell line, and dr. Marco Corazzari, University of Rome Tor Vergata, for providing the CHL-1 human melanoma and HaCaT immortalised human keratinocytes cell lines. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: CR LC MM. Data curation: AS MS MR. Formal analysis: CR MR AS MS LCat. Funding acquisition: AS. Investigation: CR MS MT LCat MR. Methodology: CR MM LC. Project administration: MM AS LC PA. Resources: PA AS MM MR. Software: AS MS MT CR. Supervision: MM AS PA. Validation: AS MR LC MM. Visualization: CR MR LC. Writing – original draft: LC CR MM. Writing – review & editing: LC CR MM AS.
# Introduction Nowadays obesity has become epidemic in developed and even in developing countries worldwide.This disease is characterized by low-grade systemic inflammation. Regarding this affirmation, it is fully accepted that the adipose tissue is a metabolically active endocrine organ secreting a variety of bioactive substances known as adipokines, that could act as functional links between energy balance and insulin resistance. Visfatin is a recently described adipocytokine, also known as pre-B cell colony-enhancing factor (PBEF) or nicotinamide phosphorybosil transferase (NAMPT).NAMPT was cloned by Samal and his colleagues in 1994 from activated human peripheral blood lymphocytes during their attempt to discover new factors for the earliest events in B-cell development. This protein of 52–55 KDa, was described as a growth factor for early B cells called pre-B cell colony–enhancing factor 1 (PBEF1). Recently, it has been identified as an intracellular enzyme and called nicotinamide 5-phosphoribosyl-1-pyrophosphate transferase (NAMPT) that catalyzes the rate- limiting step in nicotinamide adenine dinucleotide (NAD) biosynthesis and mediates the conversion of nicotinamide to nicotinamide mononucleotide.It has been proposed that this adipose derived hormone exerts insulin mimicking effects and play a positive role in attenuating insulin resistance. However, the precise mechanisms underlying the beneficial effects of visfatin on insulin sensitivity remain largely unknown. It has been suggested that visfatin might have both endocrine and paracrine effects, mostly related to obesity and insulin sensitivity although there are important discrepancies in the literature. As regards to obesity, Fukuhara et al. reported that, in adipose tissue, visfatin is predominantly expressed by visceral fat as estimated by abdominal computed tomography; and also its circulating levels correlate with the amount of visceral fat in both humans and mice. However, these observations were not confirmed by other authors. Furthermore, the significance of the correlation between visfatin levels and body mass index (BMI) is still unclear probably due to deficiencies in the specificity of the applied immunoassays. As it relates to the impact of visfatin on insulin sensitivity, Fukuhara et al reported that visfatin binds to and activates the insulin receptor explaining the reported metabolic effect of this citokyne on peripheral organs. Recently, these authors retracted from the original paper, being this hypothesis currently under debate. Moreover, other groups have reported on the presumed insulin mimetic effects of visfatin in osteoblasts and cultured mesangial cells. Visfatin is a proinflammatory mediator and might participate in a variety of inflammatory conditions, such as autoimmune diseases and adipose tissue inflammation-induced insulin resistance. The human visfatin gene is located on chromosome 7q22.3 and includes 11 exons encompassing 34.7 kb. Interestingly, this region appears to be linked to metabolic syndrome–related phenotypes in several populations including non diabetic Mexican-Americans, Mexicans–American families and hypertensive Hispanic families according to multipoint variance component analysis. Our aim was to investigate if genetic variations in the visfatin gene could be associated with obesity, glucose tolerance (Diabetes Mellitus Type 2) and cardiovascular (CV)risk–related alterations on a large sample of a population–based study in Spain. # Methods ## Participants The Segovia Study was designed as a cross-sectional, population–based survey of the prevalence of anthropometric and physiological parameters related to obesity and other components of the metabolic syndrome. It was conducted in rural and urban areas of the province of Segovia, in Central Spain. A detailed report of this study has been previously published. In brief, a random sample of 2,992 men and non—pregnant women aged 35–74 years were selected from a target population of 63,417 inhabitants. A total of 587 individuals gave written consent after receiving detailed information on the purposes and the objectives of the study. The protocol was approved by the Ethics Committee of the Hospital Clínico San Carlos in Madrid. ## Anthropometrical and biochemical measurements The BMI and waist circumference (WC) were used as estimations of total body fat mass and visceral obesity, respectively. Those individuals with a BMI higher than 30 kg/m<sup>2</sup> were classified as obese. Systolic and diastolic blood pressures were measured three times in a seated position after 10 min of rest to the nearest even digit using a random-zero sphygmomanometer. 20 ml of blood were obtained from an antecubital vein without compression after about a ten-hour overnight period. Plasma glucose was determined twice by a glucose-oxidase method adapted to autoanalyze (Hitachi 704, Boehringer Mannheim, Germany). An oral glucose tolerance test (OGTT) was performed following the criteria of the American Diabetes Association. The categories of glucose values were as follows: 1) normoglycaemia (NG): fasting plasma glucose \< 100 mg/dl or 2-h postload glucose lower than 140 mg/dl; 2) impaired fasting glucose (IFG): fasting plasma glucose levels ≥ 100 mg/dl but \<126 mg/dl and 2h postload glucose lower than 140 mg/dl; 3) impaired glucose tolerance (IGT): 2-h postload glucose between 140 and 200 mg/dl; 4) diabetes: fasting plasma glucose levels ≥ 126 mg/dl or 2-h postload glucose ≥ 200 mg/dl. Total serum cholesterol, triglycerides and high- density lipoprotein cholesterol (HDL-C) were determined by enzymatic methods using commercial kits (Boehringer Mannheim, Germany). Low-density lipoprotein cholesterol (LDL-C) was calculated by the Friedewald formula. Plasma visfatin concentrations were measured with a human visfatin (COOH-terminal) enzyme immunometric assay (Phoenix, Belmont, CA). The minimum detectable concentration was 2.3 ng/ml and the intra / inter-assay variation´s coefficient \<5% and \<14% respectively. Insulin, proinsulin, leptin and adiponectin serum concentrations were determined by highly specific / sensitive RIAs (Linco Research Inc., St LouisMO, USA). Insulin resistance (IR) was estimated by the homeostasis model assessment (HOMA-IR) method according to the formula: insulin (μU/ml) x glucose (mmol/l) / 22.5, using a cut-off point for HOMA-IR ≥ 3.8 as described in the Spanish population. To estimate high CV risk we chose: a) Framingham chart by D´Agostino et al in which the top sex-specific quintiles of predicted CVD risk identifies \~ 48% of men and 58% of women who experienced a first CVD event on follow-up (sensitivity) and proportions of men and women without events who were not in the top quintile of risk were 85% and 83% respectively (specificity); b) SCORE Cholesterol risk chart for low risk regions: the 5% threshold has a 35% sensitivity and 88% specificity and the ROC area is 0.74 (95%CI 0.72–0.76). A CVR ≥20% according to the Framingham chart is associated with high risk of cardiovascular morbidity and a CVR ≥5% with the SCORE chart indicates high risk of cardiovascular mortality. ## Genotyping Three previously described (<http://www.ncbi.nlm.nih.gov/SNP>) visfatin gene variants were genotyped. In order to select the polymorphic sites in the NAMPT gene that would be genotyped in our sample, we used information from the HapMap database. Previous studies have reported association of NAMPT variants with homeostasis glucose, fasting insulin and metabolic disorders. Among these, the SNPs rs7789066 and rs11977021 reside within the promoter region and the SNP rs4730153 is located in an intronic region. DNA was extracted from venous blood using a kit by Qiagen. Genotyping was performed using the ABI PRISM 7900 Sequence Detection System by TaqMan for allelic discrimination assay (Applied Biosystems). Genotyping was done using Assays–on Demand, and SNP Genotyping Products (Applied Biosystems). The reaction was amplified on a GeneAmp polymerase chain reaction system: cycling conditions were as follows: an enzyme heat activation step at 95°C for 10 min, 40 two-step amplification cycles at 95°C for 15 second for denaturation and a last heating step at 60°C for 1 min for annealing and extension. To assess genotyping reproducibility, a random \~20% selection of the samples was genotyped, again in three SNPs with 100% concordance. ## Statistical Analysis Genotypic and allelic frequencies were analysed as categorical variables. Hardy- Weinberg equilibrium was computed to the expected genotype distribution. Triglycerides, fasting insulin, HOMA-IR, and leptin values were log-transformed because of their skewed distributions. The one-way analysis of variance (ANOVA) was used to compare continuous variables expressed as the mean ± standard deviation (S.D.), while categorical variables were compared using the Chi- squared test. Successive multiple linear regression analyses, adjusted for potential confusion factors, were carried out to asses the effect of clinical variables and genetic variants on glucose tolerance and obesity–related alterations. The null hypothesis was rejected in each statistical test when the p value was less than 0.05. The statistical analysis was performed using the Windows SPSS version 15.0 software (SPSS, Inc, Chicago, IL, USA). # Results ## Segovia study sample characteristics The clinical characteristics of the studied population are shown in Tables and. One hundred and forty-nine participants were obese, and the remaining ones (438) were non–obese. Statistical significant differences were found for most of the clinical parameters when comparing obese against normal weight subjects, although only 2 hours glucose, fasting insulin and proinsulin, 2 hours insulin and proinsulin, HOMA-IR and fasting leptin could be considered of different significant clinical relevance between groups. Fasting visfatin and adiponectin were similar in both categories subjects. Women showed lower mean values of WC, triglycerides, fasting glucose, HOMA-IR, fasting proinsulin and proinsulin 2 hours than men. However, fasting HDL-C, leptin and adiponectin levels were higher in women than men. ## Circulating visfatin levels and clinical parameters Plasma visfatin levels were significantly higher in obese subjects with type 2 diabetes than in other categories of glucose tolerance: Normoglycaemia (NG): 9.69±3.77 ng/ml, prediabetes: impaired fasting glucose/impaired glucose tolerance (IFG/IGT): 9.61± 2.72 ng/ml, and Diabetes Mellitus (DM): 13.48± 9.66ng/ml; global p = 0.018, NG vs DM p = 0.007; IFG/IGT vs DM p = 0.016. This association between elevated visfatin levels and type 2 diabetes mellitus was confirmed by a lineal regression analysis adjusted for gender, age, BMI, and WC (β = 0.064, 95%CI: 1.002–1.135, p = 0.021). No evidence of that association was found in non obese subjects (β = 1.03, 95% CI 0.89–1.19, p = 0.067). ## Genotype distributions and association study Genotype distributions of the analyzed SNPs are shown in. None of theSNPs deviated significantly from the Hardy-Weinberg, and our genotype distributions are similar to those reported in the NCBI (National Center for Biotechnology Information) for European populations (<http://www.ncbi.nlm.nih.gov/entrez/>). The analysed SNPs were in Hardy–Weinberg´s equilibrium (data not shown). In order to detect some possible associations between these genetic variants of the visfatin gene, obesity and other glucose tolerance status–related disturbances, the association study was performed in two groups of obese and non–obese individuals. Evidence of some associations with insulin sensitivity–related parameters was detected in the subgroup of obese participants. The significant associations found by ANOVA were confirmed by linear regression analysis showing that, in obese participants, the genotype AA of the rs4730153 SNP was significantly associated with fasting glucose (β = 10.84, p = 0.040), fasting insulin (β = 9.004, p = 0.003) and HOMA-IR (β = 0.173, p = 0.031) after adjustment for gender, age, BMI and WC. We performed the same analysis excluding diabetic subjects and the results were similar. On the other hand, the obese individuals carrying the CC genotype of the rs11977021 SNP showed higher circulating levels of fasting proinsulin (β = 4.231, p = 0.029) after adjustment for the same variables. No evidence of these associations was found in the non- obese subgroup or between the analysed genotypes and the other metabolic parameters (data not shown). ## Cardiovascular risk The genotype AA of the rs4730153 SNP offers protection against CV risk either estimated by Framingham or SCORE charts in general population as well as in obese and non-obese.No associations with CV risk were observed for the other studied SNPs (rs11977021,rs7789066). # Discussion Most recently it has been demonstrated that the systemic visfatin-mediated nicotinamide adenine dinucleotide synthesis is necessary for β cell function. In this study, we performed a genetic association analysis to investigate whether the visfatin gene is a candidate gene influencing glucose tolerance, obesity and CV risk. The results here reported show for the first time, that in obese adult population, the SNP rs4730153 is associated with insulin resistance as estimated by HOMA-IR regardless of BMI, WC, age and gender. Moreover, another SNP, namely rs11977021 is independently associated with serum proinsulin levels. On the other hand the genotype AA of the rs4730153 SNP seemed to be protective for CV risk in either Framingham or SCORE in general population as well as in obese and non-obese. As for the significance of circulating visfatin levels, several authors have reported higher circulating visfatin levels in type 2 diabetic subjects than in non diabetic individuals from Caucasian and non–Caucasian populations, though it was unclear if this association was primarily related to obesity. Moreover, no associations between insulin sensitivity related parameters and circulating visfatin levels have been found in most of the reported studies or, when they found significance, it was lostafter adjusting for BMI. Chen et al, in a case–control study on 120 non–obese individuals, reported higher plasma levels of circulating visfatin in patients with type 2 diabetes mellitus than in non diabetic subjects after adjustment for gender, age and BMI. Similarly, García- Fuentes et al showed that plasma visfatin levels are increased in patients with severe obesity but only when it is accompanied by some degree of impaired glucose tolerance. Using the same EIA commercial kit (Phoenix Pharmaceuticals), we found that elevated visfatin levels were associated with a higher prevalence of type 2 diabetes mellitus just in obese subjects. However, the conclusions derived from our study and the above referred reports must have some limitations because of the specificity of this particular assay, as well as others currently available, has been recently questioned as the different immunoassays might detect distinct visfatin compounds. On the other hand, the analysis of genetic variants in the visfatin gene might help knowing if visfatin plays an important physiological role in glucose homeostasis–related alterations. We found that, in our obese subpopulation, the SNPs rs11977021 and rs4730153 are associated with glucose metabolism–related parameters regardless of BMI, WC, age and gender. As for the rs4730153 variant, the allele frequency was in agreement with that reported in other Caucasian populations. The homozygous subjects for the A allele had higher mean values of fasting glucose, fasting insulin and HOMA index. These associations have been previously explored in 626 subjects without type 2 diabetes mellitus and in 167 obese and 508 non-obese children from Germany, but no statistically significant associations were found. The physiological mechanism by which an intronic polymorphism such as the rs4730153 affects changes in glucose metabolism remains unclear. So far, there are only 2 SNPs located in the promoter region of the visfatin gene (-948C/A, rs9770242) which have been found to be associated with type 2 diabetes–related parameters, but no proof of its functionality is yet available. Of further interest is to consider that polymorphisms in the phosphoinositide 3 –kinase gene is a potential candidate gene to develop type 2 diabetes also located in the chromosome region 7q22.3. The rs11977021 had been previously associated with total and LDL cholesterol but its possible association with proinsulin has not been previously analysed. We report here a significant association between the CC genotype of the rs11977021 and higher levels of fasting proinsulin. This finding might be of relevance since there is data indicating that the rs11977021 variant might be a functional polymorphism key in type 2 diabetes mellitus. On the other hand, in the last years, several studies have established positive associations between enhanced circulating visfatin/NAMPT levels and atherogenic inflammatory diseases, therefore supporting a role for visfatin as a potential biomarker of cardiovascular complications associated to metabolic disorders. In another study, Belo et al explored different SNPs (rs1319501 and rs3801266) in obese children and adolescence with and without ≥ 3 traditional metabolic risk factors (MRFs). They found that the SNPs rs1319501 had no effect on plasma visfatin levels in obese and obese children with ≥ 3 MRFs as compared to lean children. In contrast, rs3801266 was associated with higher plasma visfatin concentrations in lean and obese children but not in obese children with ≥ 3 MRFs, suggesting that obesity and MRFs are more influential than genetic polymorphisms in the determination of visfatin levels in obese children. Other study in obese Chinese children and adolescents found that rs4730153 GG genotype polymorphism was related to an increase in insulin sensitivity and a decrease in triglycerides in response to exercise. So previous described results focused on obesity in earlier stages of life, need to be corroborated in larger studies and the comparison to our results here reported is difficult. Finally, Dou et al. suggested that NAMPT may play an important role in the development of dilated cardiomyopathy. Our study shows that visfatin genotype AA SNP rs4730153 is associated with a lower estimated CV-risk, although the precise mechanism by which the SNP influence CV-risk remains to be established. In summary, this is the first study which concludes that the genotype AA of the rs4730153 SNP appear to protect against CV events in obese and non-obese individuals, estimated by Framingham and SCORE charts. Our results confirm that the different polymorphism in the visfatin gene might be influencing the glucose homeostasis in obese individuals. Additional genetic studies in large obese population samples and functional characterization of the genetic variants are, however, warranted to corroborate our results. The authors also wish to thank María Peiró-Camaró Adán who helped revise the English. We gratefully acknowledge Members of the Segovia Insulin Resistance study Group (SIRSG) lead author: Manuel Serrano Ríos.e-mail: <[email protected]>: Ángel Agudo-Benito, Juana Alonso-Barbolla, Eva M Álvarez-de Castro, Mónica Álvarez-González, María José Andrés-Francés, Paz de Andrés-Luis, Juan Manuel de Andrés-Rubio, Enrique Arrieta-Antón, Yolanda Artiñano-del Pozo, Maria de los Ángeles Benito-Benito, Pinal Benito-Zamarriego, Miguel Ángel Betés-Ruíz, Teresa Calvo-Navajo, Juan José Cañas-Sanz, Ovidio Campo-Hernández, María Jesús Cardiel-Mañas, Rosario Cayuela-Caravaca, Tomás Conde-Macías, Lucia Corral-Cuevas, Rafael Cuervo-Pinto, Máximo Durán-Ramos, Raúl Fernández-Lambea, Eva M García-Arahuetes, Eugenio P García-de Santos, Francisco J García-de Santos, José García-Velásquez, Luís García de Yébenes-Prous, Magdalena Garrido-Mesa, Esther González-Aispiri, Esther González-Garrido, Pedro González-Sastre, Luis Gonzálvez-López, Enrique Guilabert-Pérez, Argimiro Gutiérrez-Mata, Benito de la Hoz-García, Mariana Egido-Martín, Belén Estampa- Santiago, Lucia Fuentes-Fuentes, José María García Arres, Belén García-Márquez, Julia García Múgica, Juan F Gil-García, Carmen González-Ferreiro, Mercedes Herranz-Rosa, María Carmen Herrero-de la Cruz, María de la Infanta-Pérez, Mariano Illana-Sanz, Jesús Izquierdo-Sánchez, Carlos de la Lama López-Areal, Ana M de Lucas-Herrero, José Carlos Martín-Cuesta, Maria M Martín-García, Elena Martín-Muñoz, Elvira Martín-Tomero, Manuel Monsalve-Torrón, Agustín Moreno- Aragoneses, Vicente Negro-Dimas, Luis Ortega-Suárez, Felipe de la Osa-Plaza, José de Pablo-Álamo, José Luis Palacio, María Luz Pardo-Duque, Jesús Pérez- Tarrero, José M.ª Pinilla-Sánchez, Dolores Piñuela-de la Calle, Gloria Poza- Martín, Rosa Amparo Ramos-Herranz, María de los Ángeles Requejo-Grado, Laura Rincón-Heras, Javier Roca-Bernal, María Rodríguez-Herrera, José Rojas-Mula, Jesús Ruiz-Barrio, Noelia Sánchez-Jiménez, Belén Sánchez-Martín, Luis A Santos- López, César Sanz-Herrero, Juan José Sanz-Vicente, Marina Silva-Guisasola, Virginia Silva-Guisasola, Carlos Solís-Camba, Carmen Tapia-Valero, Antonia Valverde-Martín, María del Pilar de Vega-Codes, Saturio Vega-Quiroga, Cristina Velarde-Mayol, Julio Zamarrón-Moreno and Laura Zamarrón-Sanz. Collaborating centers (allin the Province of Segovia): Villacastín, El Espinar, Segovia I, Segovia II, Segovia III, Segovia Rural, Cantalejo, Carbonero el Mayor, Sacramenia, Sepúlveda, Cuéllar, Nava de la Asunción, Riaza and San Ildefonso Primary Care Centers. CV risk Cardiovascular-risk PBEF1 Pre-B cell colony–enhancing factor 1 NAMPT nicotinamide 5-phosphoribosyl-1-pyrophosphate transferase NAD nicotinamide adenine dinucleotide HOMA-IR Homeostasis model assessment- insulin resistance BMI Body Mass Index CV cardiovascular WC waist circumference OGTT oral glucose tolerance test NG normoglycaemia IFG impaired fasting glucose IGT impaired glucose tolerance HDL-C high-density lipoprotein cholesterol LDL-C Low density lipoprotein cholesterol IR Insulin resistance SCORE Systematic Coronary Risk Evaluation NCBI National Center for Biotechnology Information [^1]: The authors received funding from a commercial source (Eli Lilly Lab, & Bayer Pharmaceutical Co.) but this does not alter their adherence to PLOS ONE policies on sharing data and materials. No conflict of interest related to this work is declared. [^2]: Conceived and designed the experiments: MSR. Performed the experiments: ACA MTML MPB YLR. Contributed reagents/materials/analysis tools: MTML MPB YLR. Wrote the paper: MTML MSR ACA. Study concept and design: MSR. Acquisition of data: ACA MTML MPB YLR. Analysis and interpretation of data: CFP MTML MSR. Drafting of the manuscript: MTML MSR ACA. Analysis and interpretation of data: CFP MTML MSR. [^3]: ¶ Membership of the SIRSG is provided in the Acknowledgments.
# Introduction Fecal hormone analysis has become a widely used technique for measuring an animal’s endocrine status and can provide valuable information to conservation and monitoring programs. Fecal samples are often easily found and identified to the species level and can be collected without disturbing wildlife. Analysis of hormones in these samples can provide a variety of stress, reproductive, and metabolic status measures that can be correlated with environmental pressures over time. Noninvasive sample collection, however, often includes samples that have been exposed to variable environmental conditions for varying and unknown time periods. Understanding how time and exposure in the natural environment affect hormone degradation is a prerequisite to reliable interpretation of fecal hormone levels, particularly if the same natural conditions causing variation in hormone levels (e.g., ambient temperature) also promote hormone degradation. Physiological measures of climate-related thermoregulatory demands provide a case in point. Monitoring effects of climate change can be difficult because cumulative effects take place on large geographic scales over long time frames. Noninvasive physiological measures of thermoregulatory demands and associated impacts from habitat shifts in wildlife over large landscapes could provide a sensitive tool for early detection and monitoring of such impacts. Fecal glucocorticoids (fGC: cortisol, corticosterone, and their metabolites) could provide one such measure, having been shown to reflect physiological responses to ambient temperature and thermal exposure in mammals. Monitoring these impacts in nature can be difficult, however, because warmer ambient temperatures and environmental exposure may also hasten fGC degradation, and both temperature and exposure effects are likely to vary with time of year. Nutritional status, which is likely to vary with time of year, can also be detected through fGC levels. Yet, diet composition itself can impact steroid excretion rates. Here, we ask: can we detect a biological relationship between thermoregulatory demands or nutrition and fGC concentrations in bears? How does fGC degradation change with time, temperature, and exposure? Are these degradation effects large enough to mask fGC biological indices of thermoregulation and nutrition? We separate thermoregulatory from degradation-related effects on fGC levels in a study of grizzly bears (Ursus arctos horribilis) and American black bears (U. americanus) in Glacier National Park, MT. Julian date (estimating ambient temperature) plus aspect, slope, and elevation (collectively estimating thermal exposure), and precipitation at sample collection locations were used to predict thermoregulatory impacts by measuring fGC levels in fresh fecal samples at time 0 (i.e., samples \<24 hours old) compared to fGC in those same samples left under the same natural field conditions for an additional, randomly assigned period of 1–28 days. Scat contents were also used in those samples to examine how fGC varies with diet/nutrition, as well as to separate diet impacts from effects of thermoregulation and degradation on fGC. # Methods ## Study Area We collected bear scats from 2 sections of trail located near the southern boundary of Glacier National Park (GNP), MT. Generally, these trails paralleled the park boundary with a mix of National Forest and private lands along the opposite side of the Middle Fork of the Flathead River. This area has very little human development; however, U.S. Highway 2 and the Burlington Northern Santa Fe railroad experience intermittent heavy human use. Our study occurred at the end of the tourist season when highway traffic volumes were well below the peaks experienced in mid-summer. The study area was located just west of the Continental Divide, in a relatively moist climate reflecting a maritime influence. Elevation ranged from approximately 950 m to 1250 m. Surveyed trails ran mostly on westerly or southerly facing slopes through mature forest consisting largely of Douglas fir (Pseudotsuga menziesii), Engelmann spruce (Picea engelmannii), lodgepole pine (Pinus contorta), and western larch (Larix occidentalis). Although GNP provides a high level of security for wildlife within its borders, bears are known to readily move outside the park and across the U.S. Highway 2/railroad corridor. Outside GNP bears are exposed to a variety of stressors related to human activities such as train and highway traffic, private residences and small public lodgings, and various forms of outdoor recreational and industrial activities such as logging. ## Sample Collection and Processing The study period was from Julian date 250 (early September), when day time temperatures were 27° to 35°C, until Julian date 300 (end of October), when temperatures were near freezing, ca. 0° to −5°C. Two trails, approximately 19 km in length each, were surveyed daily to ensure that each new scat found was \<24 hours old. When a fresh scat (\<24 hrs) was encountered, the sample was divided in half. One half was mixed thoroughly to equalize the distribution of hormone throughout the sample and stored frozen immediately upon returning from the field (within 4 hrs of collection on average with an 8 hr maximum). This was referred to as the “time 0” subsample. The second half of the sample was left intact to retain its original characteristics until collection, and in its original position in the environment (or moved slightly \[\<1 m\] if at risk of being trampled). The second half, referred to as the “time 1” subsample, was randomly assigned to one of five time exposure groups: 1, 2, 7, 14, or 28 days. Field crews returned to the second half of each sample at the designated exposure time, thoroughly mixed the remaining half sample and stored it frozen, as described above. Although hormones are not evenly distributed in scat, we assumed that each sample was sufficiently large that hormone concentrations would be relatively comparable between the sample halves. At the least, hormone concentrations would have varied randomly between the two halves. We collected the following information about each sample: GPS location, elevation, date, primary and secondary food contents, moisture, odor, odor strength, presence of mold, substrate, and initial exposure conditions. We used date and location information to derive proxies for thermal stress based on average temperatures and exposure (i.e., aspect, slope, and elevation). We transformed aspect, which reflects microclimate, to a continuous variable ranging from +1 to −1 using the following formula:. Thus, +1 is the hottest aspect (west-facing) and −1 (NE facing) the coolest. We used a geographic information system to average aspect across every 30 m pixel that intersected a 100 m buffer around each sample’s location. Aspect would be important if bears seek out and compete for these locations to better regulate thermal demands. Our method assumes that fecal sample location is a reasonable proxy of the bear’s average location. Any random error associated with this assumption should decrease rather than increase the significance of the covariates in our model. All fecal samples were freeze-dried within 30 days post-collection and stored at −80°C until extracted and analyzed for glucocorticoid concentration. In other omnivores, freeze-drying samples and expressing hormone content per g of dry feces has been shown to minimize dietary impacts on fecal hormone excretion rates, independent of nutrition. All samples were extracted and assayed for fGC using the method described by Wasser et al.. Briefly, a 0.2 g subsample of dried, pulverized, well-mixed sample was weighed to the nearest 0.0001 g and transferred to a 15 ml vial. After adding 4 ml of 90% methanol to the sample, the vial was capped and then shaken for 30 min in a pulsing vortexer. Samples were subsequently centrifuged (20 min) and 1 ml of supernatant removed per sample and stored at −20°C until assayed. Extracts were then diluted four-fold in assay buffer and quantified for fGC with a double antibody <sup>125</sup>I radioimmunoassay kit (MP Biomedicals \[previously ICN\], Solon, OH, catalog \#07-120103). We used the manufacturer’s protocol except with half-volumes throughout. Low and high controls were included in every assay. Non-specific binding tubes and blanks were assayed in quadruplicate, and standards, controls, and samples in duplicate; any sample with a CV\>10% between duplicates was re-assayed to confirm results. This assay has previously been validated for black and grizzly bear fGC. For other assay details, see Wasser et al. and Hunt and Wasser. ## Statistical Analyses Fecal glucocorticoid measures were normalized by log transformation (log10(X+1)) in all analyses. We used generalized linear models (GLM) to first predict fGC concentrations in fresh (time 0) samples based on proxies of nutrition (berry and meat\>vegetation), ambient temperature (Julian date;), and thermal exposure (e.g., aspect, slope, elevation, and heat load index). Degradation effects were assumed to be negligible in the time 0 samples (≤24 hrs old) because previous work showed GCs in bear scat to be highly stable. We found support for this assumption by showing that precipitation (from time of defecation to sample collection) had no effect on fGC concentrations of the time 0 samples after discovering precipitation to be the principle cause of degradation in the time 1 samples (see results). The covariates from the time 0 GLM were next used to predict fGC concentrations in the exposed sample halves to determine whether the same environmental covariates still reliably predict fGC concentrations, with and without including effects of exposure time and precipitation. All variables included in our final models were derived from two separate GLMs, support for which was based on AIC<sub>c</sub> values. All significant variables in these two GLMs were then included in a repeated measures multivariate analysis of variance (MANOVA) to identify the environmental variables that affected both between-sample variation in fGC (independent of exposure time), and the within-sample degradation effects of exposure time. This enabled us to determine whether degradation impacts were severe enough in the time 1 subsamples (i.e., the sample halves that were left in the field for 1–28 days) to mask differences in fGC levels due to environmental stressors detected in time 0 (fresh) subsamples. Effect size was also examined in the MANOVA analyses to compare the whole models as well as all variables included in each model, using Cohen’s statistic, where: By convention, effect sizes of 0.02, 0.15, and 0.35 are termed small, medium, and large, respectively. # Results ## Analysis 1: Factors Influencing fGC Levels at Time 0 The main effects explaining variation in fGC concentrations at time 0 (i.e., prior to degradation effects), based on AIC<sub>c</sub> model support, were food contents (vegetation vs berries or meat), Julian date, aspect, as well as an interaction between Julian date and aspect. Time 0 fGCs increased with Julian date, being lowest at earlier dates (warmer temperatures) and highest at later dates (colder temperatures;). The interaction with aspect, however, indicated that fGCs were lowest in cooler aspects at earlier dates (higher temperatures), and in warmer aspects at later dates (colder tempteratures). Essentially, fGCs were lower in scats collected in cool locations (aspects) during hot weather and in warm locations during cold weather. Time 0 fGCs also appeared to reflect nutritional quality of the diet; fGCs were highest for scats containing primarily vegetation (grass, leaves, stems, roots, tubers, and corms, presumably indicating relatively poor nutrition and/or high stress) and lowest for scats containing berries and meat (implying relatively good nutrition and/or low stress;). As expected, precipitation had no effect on fGC concentrations in the time 0 samples, supporting our assumption that degradation effects were undetectable in the time 0 samples. ## Analysis 2: Factors Influencing fGC Levels at Time 1 We examined fGC stability over time by assessing whether the same variables predicting fGC concentrations in the time 0 samples remained significant in the time 1 samples, with or without the addition of all degradation measures (e.g., exposure time, precipitation, and their interactions with aspect). All predictors of fGC in the time 0 samples remained significant predictors in the time 1 samples. Adding precipitation and its interaction with aspect, however, significantly improved the model predicting fGC concentrations in the time 1 samples. Increased cumulative precipitation between time 0 and time 1 decreased fGC concentrations of the time 1 samples and the cooler the aspect, the more this effect was exacerbated. ## Analysis 3: Covariates Influencing GC Degradation The repeated measures MANOVA was consistent with the above findings. Diet, Julian date, and the interaction between Julian date and aspect remained significant predictors of between-sample variation in fGC. Cumulative precipitation, which significantly affected fGC concentrations in the time 1 but not time 0 samples, was also a significant predictor of between-sample variation in fGC in the MANOVA. However, only diet and the interaction between cumulative precipitation and aspect were significant predictors of within-sample variation over time. Cumulative precipitation in samples exposed in cooler aspects showed increased degradation over time. # Discussion The biological effects of temperature (Julian date) and thermal exposure (aspect) on bear fecal hormones at time 0 suggested that fGC may be a good indicator of thermoregulatory demands in ursids. In particular, fGC increased linearly with Julian date, indicating increased fGC in the fall as ambient temperatures declined. Similar effects of elevated cortisol levels related to thermoregulation have been described for other mammals. The interaction between Julian date and aspect also suggests that fGC during the hottest temperatures were lowest in bears that presumably spent more time in the coolest aspect. Similarly, fGC in the coldest temperatures were lowest in bears that spent more time in the hottest aspect. The colder the temperature, the greater the effect that aspect had on fGC levels. This reversal in effect of aspect on fGC would not be expected if Julian date were simply reflecting fGC degradation occurring up until the time 0 sample was first located (i.e., ≤24 hours post defecation). If degradation resulted from hotter ambient temperature, fGCs should be lower, not higher, in samples collected in the hotter aspect during the hottest time of year. If degradation resulted from cooler temperatures and/or greater precipitation, fGCs should be lower, not higher, at later Julian dates and particularly in cooler aspects at that time. Degradation effects from precipitation were also ruled out in these time 0 samples. Overall, our data suggest that bear fGC concentrations may be affected by thermoregulatory demands. These patterns suggest that fGC measures should be investigated as a potential technique for landscape-wide assessments of thermoregulatory load in free-ranging animals, for example, in studies of climate change and related effects. We also found a significant effect of diet on fGC concentrations at time 0, indicating that fGC concentrations also corresponded to diet quality in ursids, as has been found in other vertebrates. We found that fGC concentrations were highest for the least nutritious diet (coarse vegetation) and lowest for the most nutritious diets (berries and meat). An alternative possibility is that dietary fiber content might affect steroid excretion rates directly via changes in gut transit time or fecal mass. We controlled for excretion rate effects by freeze-drying samples prior to extraction, as this has proved effective in other omnivores. Moreover, berries and vegetation both contain high amounts of dietary fiber compared to meat, but only vegetation had elevated fGC. Gut transit time is also faster in grizzly bears fed vegetation diets compared to meat diets, which would tend to decrease fGC in vegetation-based scats, yet we observed an increase. We therefore conclude that the patterns we observed reflect nutritional impacts on circulating GC as opposed to effects of gut transit time or fiber content on fGC excretion rates per se. Vegetation diets are known to be a less preferred diet for grizzly bears, and vegetation diets have substantially lower protein content and digestibility. Vegetation-based diets appear to represent a nutritional stressor for grizzly bears that is accurately reflected in fGC content of fecal samples. It is possible that the part of the between-sample variance in fGC explained by Julian date was also due to seasonal changes in diet, as nutritionally important berries were most abundant earlier in the sampling season when fGC was lowest. Such nutritional effects would not, however, explain the interaction between Julian date and aspect on fGC; those can only be reconciled by the joint effects of temperature and exposure. Further, scats had a mixture of high and low quality food items throughout the sampling season. For example, highly nutritious berries were the primary contents in some scats as late as mid- October. Some fGC degradation did occur over the 28 day exposure period, largely due to cumulative precipitation, particularly in cooler aspects where evaporation would be reduced. Overall, fGC declined in the time 1 samples as a function of aspect interacting with precipitation. Yet in time 0 samples, this relationship was reversed; fGC increased in time 0 samples when aspect interacted with Julian date (cold temperatures), despite greater precipitation at later Julian dates. These differing relationships in time 0 vs. time 1 samples strongly support our conclusions that (a) fGC reflects thermoregulatory demands and (b) that some fGC degradation does occur in samples exposed to cumulative precipitation for periods of weeks prior to collection. These latter effects, however, were still insufficient to mask the effects of the biological covariates predicting fGC concentrations in the time 1 samples. Finally, autocorrelation of fGC measurements is unlikely to contribute to study results because multiple samples were collected in close proximity to each other on only one occasion. Future field studies could genotype scat as an added safeguard against autocorrelation, as done by Wasser et al.. ## Conclusions Scat can provide a wide variety of physiological and genetic information and is the most accessible biological product from wildlife in nature. Collectively, this provides the opportunity to partition impacts from a number of multiple environmental pressures, given the right study design. Our results suggest that fGC analyses can provide insight about thermoregulation and nutrition in grizzly bears and black bears, and may serve as an index of physiological response to climate change. We further showed that some sample degradation impacts do occur in the wild. These effects were, however, insufficient to mask detection of the biological environmental impacts on ursids in samples exposed for up to one month. Regardless, controlling for cumulative precipitation over the sample exposure period can be used as an added precaution against such degradation impacts masking biological effects. Similar studies examining other hormones, environments, diets, and species will be important to evaluate how far our findings can be generalized. We thank our field technicians L. Erickson, J. Hebert, J. Mascia, D. Reding, B. Wollenzien, and K. Yale for hiking the same trail every day for 6 weeks regardless of weather or boredom. We thank J. Reimers at the University of Montana for access to and assistance with lyophilizing samples. We thank Liz Addis for assistance in laboratory and data analyses. We thank Tyler Coleman, Clayton Miller, and two anonymous reviewers for providing helpful comments on early drafts of this manuscript. This research was conducted under Glacier National Park collections permit GLAC-2001-SCI-0024. [^1]: In accordance with PLOS ONE‚s policy on documenting conflicts of interest, the authors herewith declare that there are no competing interests with any authors. Specifically, the author affiliated with a private entity (JBS) has no vested interests in any of the methods or results included in this manuscript. As such, author affiliations do not alter the authors‚ adherence to all of the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: JS SW. Performed the experiments: JS KH SW. Analyzed the data: JS KH SW. Contributed reagents/materials/analysis tools: JS KH KK SW. Wrote the paper: JS KH KK SW. [^3]: Current address: New England Aquarium, Boston, Massachusetts, United States of America
# Introduction Smoking is a major public health problem worldwide. There have been thousands of studies investigating the impact of active smoking on health, and the overall toxic effects of active smoking are generally recognized. In comparison, the effects of passive smoking on health are not fully understood. Existing studies suggest that passive smoking and active smoking might equally increase the risk of certain diseases, such as female breast cancer, allergic rhinitis, allergic dermatitis, and food allergy. As early as 1928, Schonherr suspected that inhalation of husbands’ smoke could cause lung cancer among non-smoking wives. Since then a substantial body of research about environmental tobacco smoke and health has appeared. But the impact of passive smoking on health remains largely inconclusive and has not been systematically summarized. Due to the relative small health risks associated with exposure to passive smoking, investigation of this issue requires large study sizes. Difficulties in measuring passive smoking and controlling various confounding factors further add to the uncertainty in any investigation of the effects of passive smoking. Consequently, a meta-analysis, pooling together individual original studies quantitatively, has played an important part in establishing the evidence about the health effects of passive smoking. Since Zmirou evaluated the respiratory risk of passive smoking by a meta-analysis in the early 1990s, many meta- analyses of observational epidemiological studies have been published to identify the impact of passive smoking on health. Recognizing that the evidence is accumulating constantly worldwide, we conducted an overview of systematic reviews that have summarized the evidence from observational epidemiological studies on the health effects of passive smoking. # Methods No protocol exists for this overview of systematic reviews. ## Ethics Data for this research was acquired from previously published papers. Written consent and ethical approval were not required. ## Literature search strategy We attempted to conduct this overview of systematic reviews in accordance with the rationale and guideline recommended by Cochrane handbook 5.1.0. A systematic literature search of PubMed, Embase, Web of Science, and Scopus was conducted in January 2015 using the following search terms with no restrictions: passive smoking, secondhand smoking, environmental tobacco smoke, involuntary smoking, and tobacco smoke pollution. The reference lists of the retrieved articles were also reviewed. We did not contact authors of the primary studies for additional information. ## Selection of relevant systematic reviews Systematic reviews meeting the following criteria were regarded as eligible: (1) the design was meta-analysis, (2) passive smoking was an exposure variable and the outcome was the incidence of certain diseases or health problems, (3) the included original studies were cross-sectional, case-control, or/and cohort study design, (4) the literature search was international or worldwide, and (5) the pooled relative risk (RR) or odds ratio (OR) and the corresponding 95% confidence interval (CI) of specific diseases relating to exposure to passive smoking were reported or could be calculated from the data provided. Systematic reviews in which all included original studies were conducted in one country or region were excluded. We also excluded the meta-analyses that investigated the association between maternal smoking in pregnancy and the health risk of offspring. All potential meta-analyses were independently screened by two authors (SC and CY), who reviewed the titles or/abstracts first and then conducted a full-text assessment. Disagreements between the two reviewers were resolved through discussion with the third investigator (ZL). ## Data extraction The following information was extracted from the studies by two investigators (SC and CY): first author, publication year, country, number and design of the included original studies, and main quantitative estimates of the association of interest. ## Quality appraisal We appraised all the included meta-analyses using the Assessment of Multiple Systematic Reviews (AMSTAR) standard, an 11-item assessment tool designed to appraise the methodological quality of systematic reviews. The maximum score is 11, and 0–4, 5–8, and 9–11 respectively indicates low, moderate, and high quality. Disagreements on assessment scores were resolved by discussion among the authors. ## Synthesis of the evidence There may be more than one meta-analysis published regarding the association between passive smoking and risk of a specific disease. We only included the latest meta-analysis and excluded all the previous ones. For each included meta- analysis, we summarized the number and design of the included original studies, the main quantitative estimates of association of interest, heterogeneity between original studies, and so on. In any included meta-analyses, when estimates of association between passive smoking and certain diseases were reported separately for subgroups, we combined the results of the subgroups and calculated common estimates using a fixed-effects model if appropriate. # Results ## Literature search shows the process of study identification and inclusion. Initially, we retrieved 2,079 articles from Pubmed, Emabse, Web of Science, and Scopus. After 1,105 duplicates were excluded, 974 articles were screened through titles and abstracts, of which 858 were excluded mainly because they were original studies or irrelevant reviews. After full-text review of the remaining 116 articles, 100 were further excluded because they did not report the outcomes of interest or their findings were already updated by newer systematic reviews. Finally, 16 meta-analyses were included. ## Characteristics and quality of the included systematic reviews The main characteristics of the sixteen meta-analyses were summarized in. These meta-analyses covered a total of 130 cohort studies, 159 case-control studies, and 161 cross-sectional studies. They were published between 1998 and 2014. The quality scores of these meta-analyses appraised using AMSTAR ranged from 3 to 10. The numbers of meta-analyses with high quality, middle quality, and low quality were 5, 9, and 2 respectively. ## The Main Health Consequences of Passive Smoking shows the integrated results on the impact of passive smoking on specific diseases. The included 16 meta-analyses covered 25 diseases or health problems. There was statistically significant positive relationship between exposure environmental tobacco smoke and the risk of eleven diseases, especially invasive meningococcal disease in children (OR 2.18; 95% CI 1.63–2.92) and other three diseases or health problems with a 1.5 to 2.0-fold increase in the risk: cervical cancer (OR 1.73; 95% CI 1.35–2.21), Neisseria meningitidis carriage (OR 1.68; 95% CI 1.19–2.36), and Streptococcus pneumoniae carriage (OR 1.66; 95% CI 1.33–2.07). The increase in the risk of other seven diseases associated with exposure to passive smoking was statistically significant but small in impact size (OR was less than 1.5): lower respiratory infections in infancy (OR 1.42; 95% CI 1.33–1.51), food allergy (OR 1.43; 95% CI 1.12–1.83), childhood asthma (OR 1.32; 95% CI 1.23–1.42), lung cancer (OR 1.27; 95% CI 1.17–1.37), stroke (OR 1.25; 95% CI 1.12–1.38), allergic rhinitis (OR 1.09; 95% CI 1.04–1.14), and allergic dermatitis (OR 1.07; 95% CI 1.03–1.12). Of these 25 diseases or health problems, eight diseases were not found to be significantly associated with passive smoking. They were invasive Haemophilus influenzae type B (Hib) disease, invasive pneumococcal disease, Crohn's disease, pancreatic cancer, ulcerative colitis, breast cancer, bladder cancer, and pharyngeal carriage for Hib. In addition, the effects of passive smoking on increased risk of coronary heart disease, tuberculosis, diabetes, and middle ear disease in children (recurrent otitis media, middle ear effusion, and glue ear) were not conclusive, because the number of included studies was small or the quality of the corresponding meta-analysis was low. ## Passive smoking and cancer risk We investigated the association of passive smoking with the risk of lung cancer, cervical cancer, pancreatic cancer, breast cancer, and bladder cancer. Based on 55 observational studies (7 cohort studies, 25 population-based case-control studies and 23 non-population-based case-control studies), passive smoking were found to be associated with the increased risk of lung cancer (OR 1.27; 95% CI 1.17 to 1.37). The ORs for lung cancer in North America, Asia, and Europe were similar. 11 case-control studies, involving 3,230 cases and 2,982 controls, suggested a positive relationship between passive smoking and cervical cancer (OR 1.73; 95% CI 1.35–2.21). Pancreatic cancer, breast cancer, and bladder cancer were not found to be associated with passive smoking. ## Passive smoking and allergic diseases A meta-analysis of observational studies published in PLOS Medicine systematically reviewed the effects of exposure to environmental smoke on allergic diseases. The pooled ORs of 63 studies for allergic rhinitis, 58 studies for allergic dermatitis, and 6 studies for food allergies were 1.07 (95% CI 1.03–1.12), 1.09 (95% CI 1.04–1.14), and 1.43 (95% CI 1.12–1.83) respectively. Another meta-analysis investigated the association between passive smoking and the risk of physician-diagnosed childhood asthma, and suggested that there was consistent evidence of a modest positive association between them (OR 1.32; 95% CI: 1.23–1.42). ## Passive smoking and pediatric invasive bacterial disease and bacterial carriage Passive smoking was also thought to be associated with pediatric invasive bacterial disease and bacterial carriage. A meta-analysis involving 30 case- control studies for invasive bacterial disease and 12 cross-sectional studies for bacterial carriage indicated that the risk of invasive meningococcal disease, pharyngeal carriage for Neisseria, meningitidies and Streptococcus pneumoniae were significantly associated with passive smoking, and the ORs were 2.18, 95% CI 1.63 to 2.92), 1.68 (95% CI, 1.19–2.36), and 1.66 (95% CI 1.33–2.07), respectively. The risk of invasive pneumococcal disease, invasive Hib disease, and pharyngeal carriage for Hib were not found to be related to exposure to environmental smoke. # Discussion The health effects of environmental tobacco smoke are attracting more and more attention worldwide. Increasing numbers of original studies and meta-analyses are being published focusing on this important issue. In the present overview of systematic reviews based on sixteen systematic reviews involving 450 original observational studies, we found that passive smoking could significantly increase the risk of eleven diseases, especially invasive meningococcal disease in children, cervical cancer, Neisseria meningitidis carriage, and Streptococcus. pneumoniae carriage, but not associated with other eight diseases. Cancers were one of the most common investigated health outcomes associated with passive smoking. We found that exposure to environmental tobacco smoke could increase the risk of lung cancer and cervical cancer, but was not the risk of pancreatic cancer, breast cancer, or bladder cancer. It appears that passive smoking could increase the risk of some diseases among children, especially bacterial infections (e.g., lower respiratory infections in infancy, middle ear disease in children, invasive meningococcal disease in children, allergic diseases in children, and childhood asthma). Previously, there were some reviews focusing on the health effects of exposure to environmental tobacco smoke. But they were qualitative or only involved children or limited to several diseases. We used a systematic overview to summarize the quantitative estimates of the associations between passive smoking and various diseases based on all latest available meta-analyses. It should be noted that, in the present overview, we excluded meta-analyses evaluating the effects of smoking during pregnancy on fetus or offspring health, because the effects was obviously different from the health effects of active smoking or conventional passive smoking in the general population. The quality of included original studies influences the reliability of the results and conclusions of the corresponding meta-analysis; similarly, the validity of the results of an overview of systematic reviews depends on the quality of the included systematic reviews. We used AMSTAR protocol, an internationally recognized assessment tool, to appraise the methodological quality of all included meta-analyses, and found that there were two meta- analyses with low quality. Accordingly, the conclusions drawn based on these two meta-analyses involving middle ear disease in children and coronary heart disease need to be interpreted with caution. The evidence level of meta-analyses partly depends on the number and the design type of included original studies. Although there was no consensus about the minimum number of original studies included in meta-analysis, but more caution is needed when an association is assessed based on a small number of original studies. In our overview, we found a significant positive association between passive smoking and tuberculosis (OR 4.01; 95% CI 2.54–6.34), but it was only based on 4 case-control studies. More studies should be conducted to further assess the relationship between them. Similarly, the effect of passive smoking on diabetes was based on 6 cohort studies (OR 1.21; 95% CI 1.07–1.38), and more original studies are also needed. There were several strengths in our research. Firstly, we followed the primary rationale and method of Cochrane overviews of reviews to summarize the health consequences of certain exposure. Overview of systematic reviews is primarily intended to summarize multiple reviews addressing the effects of two or more potential interventions for a single condition or health problem. Up to now, most of overviews have been conducted to evaluate the effects of several interventions, and very few overviews have addressed the effects of a single exposure factor on multiple diseases or health problems based on observational studies. Our present overview expands the application of overviews of systematic reviews. Additionally, our study provides robust and comprehensive scientific information for smoking ban in public places and for educational pamphlets about passive smoking. Some limitations in our overview should be noted. Firstly, we only included systematic reviews but not original studies. The associations of passive smoking with some diseases might have been investigated by original studies but not synthesized by meta-analyses and, therefore, were not summarized in this overview. Secondly, the mechanism on the health effects of passive smoking was not be examined since our study only intended to summarize relevant observational epidemiological evidence. In summary, our overview of systematic reviews of up-to-date epidemiological evidence suggests that passive smoking is significantly associated with an increasing risk of many diseases and health problems, especially diseases in children and cancers. This study provides comprehensive population-based evidence about toxic effect of exposure to environmental tobacco smoke and should benefit developing health promotion strategies of smoking control. Stricter regulations against cigarette smoking should be formulated and implemented, because smoking harms not only own health but also the health of neighboring people. # Supporting Information # Access to data All the data in this review are from publicly published papers, and we take responsibility for the integrity of the data and the accuracy of the data analysis. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ZL. Performed the experiments: SC CY. Analyzed the data: SC YG. Contributed reagents/materials/analysis tools: CY. Wrote the paper: SC.
# Introduction Myxoviruses are enveloped, negative-strand RNA viruses that are transmitted through the respiratory route. The orthomyxovirus family comprises five different genera of which the influenza viruses are clinically most relevant. Of the paramyxoviridae, respiratory syncytial virus (RSV), measles virus (MeV), mumps virus (MuV), human parainfluenzaviruses (HPIV) and the recently emerged, highly pathogenic zoonotic henipaviruses constitute major human pathogens. Although clinical complications associated with some myxoviruses involve persistent infections, the viruses predominantly induce acute respiratory or systemic disease. Collectively, myxoviruses are responsible for the majority of human morbidity and mortality due to viral respiratory illness globally. In particular, influenza virus is the leading cause of morbidity and mortality from respiratory disease in North America despite the existence of vaccine prophylaxis. This is due to the fact that the vaccines currently in use reduce illness in approximately 70% of healthy adults when homologous to the prevalent circulating virus, but protection in the elderly reaches only approximately 40%. Vaccine efficacy is reduced substantially when the circulating strains differ from those constituting the vaccine. Despite extensive research and in contrast to, for instance, MeV and MuV, no vaccines are currently available against several major pathogens of the paramyxovirus family such as RSV or different HPIVs. Infection with RSV is the leading cause of pneumonia and bronchiolitis in infants, both associated with significant mortality, while HPIV types 1 and 2 are the primary cause of croup syndrome and can likewise result in serious lower respiratory diseases such as pneumonia and bronchiolitis. The availability of effective antiviral therapy for most clinically significant myxovirus infections is limited. Licensed neuraminidase inhibitors for influenza therapy, Zanamivir and Oseltamivir, show efficacy when administered within a 48-hour window after the onset of symptoms, but are increasingly compromised by pre-existing or emerging viral resistance. Ribavirin, although approved for RSV treatment, shows limited utility due to efficacy and toxicity issues. The polyclonal immunoglobulin RSV-IVIG and the humanized monoclonal antibody Synagis provide RSV prophylaxis, but use is limited to high-risk pediatric patients. Considering the high mutation rates seen in particular with RNA viruses, the development of novel types of myxovirus inhibitors that circumvent the rapid development of resistance is highly desirable. Of the strategies conceivable towards this goal, targeting host factors required for completion of the viral life cycle rather than pathogen-encoded factors directly has received heightened interest in recent years. This approach is expected to establish a significant barrier against spontaneous viral escape from inhibition, since individual viral mutations are less likely to compensate for the loss of an essential host cofactor than to prevent high-affinity binding of a conventional, pathogen-directed antiviral. Given some degree of overlap of host cell pathways required for successful replication of related viral pathogens, host-directed antiviral approaches also have the potential to move beyond the one-bug one-drug paradigm by broadening the pathogen target range of a chemical scaffold. Naturally, targeting host factors for antiviral therapy bears an inherently higher potential for undesirable drug-induced side effects than conventional pathogen-directed strategies. While the approach is nevertheless under investigation for the treatment of chronic viral infections such as HSV-1 and HIV-1, an application to the inhibition of infections by pathogens predominantly associated with severe acute disease, such as most members of the myxovirus families, is anticipated to render drug-related side effects tolerable to some extent, since the necessary treatment time and concomitant host exposure to the drug remain limited. In the case of influenza infections, for instance, typical neuraminidase inhibitor regimens consist of twice daily administration for a five-day period for treatment, or a 10-day period for prophylaxis. Relying on a broadened anti-myxovirus target spectrum as the main selection criterion in secondary screening assays, we have mined results of a recently completed high throughput chemical library screen to identify hit candidates with a possible host-directed mechanism of action. This has yielded a compound class with broad anti-viral activity, which was subjected to synthetic scaffold optimization, quantification of active concentrations for a select group of clinically relevant ortho- and paramyxovirus family members, testing against a panel of exposed host cells of different species origin, and characterization of the compound-induced point-of-arrest in viral life cycle progression. Viral adaptation to growth in the presence of inhibitor has been employed to compare escape rates from inhibition by this new compound class with those from a well- characterized, pathogen-directed antiviral. # Results To identify small-molecule hit candidates that block the myxovirus life cycle through a host-directed mechanism, we analyzed the results of a high-throughput cell-based anti-MeV screen of a 137,500-entry library of the NIH diversity set that we recently reported. The primary screening agent, serving as the myxovirus representative, was the wild type MeV isolate MVi/Alaska.USA/16.00 (MeV-Alaska). It was chosen based on its ease of growth and readily quantifiable cytopathic effect in the automated system. In search of candidates with a host-directed antiviral profile, we anticipated three distinct features of desirable compounds: a) potent inhibition of virus replication at the screening concentration (3.3 µM); b) a primary screening score, representative of the selectivity index (CC<sub>50</sub>/EC<sub>50</sub>), close to the cut-off value for hit candidates due to some anticipated host-cell interference ( = 1.9); and c) a broadened viral target spectrum in counterscreening assays that extends to other pathogens of the myxovirus families. ## Identification of a chemical scaffold with broad anti-viral activity When inhibition of paramyxovirus family members was assessed, six compounds efficiently blocked the closely related canine distemper virus (CDV) and the more distantly related human parainfluenzavirus type 3 (HPIV3) in addition to MeV-Alaska, while leaving cell metabolic activity essentially unaffected. Of these independent hits, three share a common molecular scaffold. Since HTS scores of these analogs best matched the target criteria and antiviral activity was highest in this group, we subjected them to further characterization and developmental efforts. Synthetic optimization and structural confirmation of the scaffold returned a lead analog JMN3-003, which showed potent activity against MeV, a selection of clinically significant members of the para- and orthomyxovirus families, and, albeit to a lesser degree, representatives of positive strand RNA virus (sindbis virus of the Alphaviridae) and DNA virus (vaccinia virus of the Poxviridae) families (inhibitory concentrations for a larger panel of myxovirus family members are summarized). As observed for the primary hit compound, metabolic activity of different established cell lines exposed to JMN3-003 was unchanged at 75 µM, the highest assessable concentration based on solubility of the substance in growth media. Of different primary human cells examined, metabolic activity was unaffected (PBMCs, smooth muscle cells) or only slightly affected (bronchial epithelial cells) by the compound. These data support potent anti-myxovirus activity of the compound with active concentrations ranging from 10 to 80 nM depending on the target virus. ## Antiviral activity of lead compound JMN3-003 is host cell-specific To further explore whether JMN3-003 meets the profile of a host-directed antiviral, we examined whether the extent of inhibition is determined by the species origin of the host cell used for virus propagation. Based on its broad host cell range, inhibition of influenza A/WSN replication was monitored. In addition to higher mammalian (HT1080 (ATCC CCL-121), HeLa (ATCC CCL-2), MDCK (ATCC CCL-34)) cell lines, cells of rodent (NIH-3T3 (ATCC CRL-1658), MEL B16 (ATCC CRL-6322), BHK-21 (ATCC CCL-10), CHO (ATCC CCL-61)) and avian (DF-1 (ATCC CRL-12203)) origin were tested, which are all permissive for influenza A/WSN infection. While inhibitory concentrations obtained for all higher mammalian cell lines examined were similar, A/WSN inhibition by JMN3-003 was found inactive on some rodent cell lines and when virus was propagated on murine or avian cells. However, inhibitory activity extended fully to primary human PBMCs. For the latter, inhibition of MeV-Alaska was monitored due to efficient growth of MeV isolates on PBMCs. The host cell species effect of antiviral activity of JMN3-003 is consistent with specific targeting of cellular factors by the compound, while arguing against docking to conserved viral factors or an undesirable promiscuous, unspecific mode of activity. ## JMN3-003 shows high metabolic stability in vitro The central 2-thio-connector found in the chemical scaffold of JMN3-003 may render the compound susceptible to rapid phase I oxidation in vivo , thus possibly compromising its developmental potential. To test metabolic stability of the substance early in development, we exposed JMN3-003 to human S-9 hepatocyte subcellular fractions as an in vitro indicator for phase I metabolism. After a 60-minute exposure, approximately 80% of the input material remained intact, corresponding to an extrapolated half-life of approximately 200 minutes. Unstable analogs of JMN3-003, JMN5-165 and JMN5-166, returned half lives of 38 and 5 minutes in this assay, respectively, confirming metabolic competency of the S9 fractions used. Assessment of JMN3-003 stability in human plasma in comparison with unstable Procaine and stable Procainamide corroborated these results, since JMN3-003 integrity was virtually unaffected after a 120-minute incubation period. Taken together, these findings suggest desirable metabolic stability for the JMN3-003 scaffold, recommending it for further mechanistic characterization. The data are corroborated by the good metabolic stability reported for the structurally similar compound RDEA-806, a non-nucleoside inhibitor of HIV reverse transcriptase and clinical precedent, which shares the 2-thio-connector of JMN3-003 but lacks MeV inhibitory activity in our assays (data not shown). ## Temporary arrest in cell cycle progression Since direct cytotoxicity of JMN3-003 was low for all cell lines examined, we next tested the effect of the substance on cell cycle progression. Analysis of the DNA content of cells continuously treated with JMN3-003 for 36 hours by flow cytometry revealed accumulation of cells in a single population with 2N DNA content, which closely resembled the profile of a reference cell population exposed to hydroxyurea but markedly differed from the 4N DNA content of nocodazole-treated cells. Nocodazole interferes with microtubule polymerization, resulting in a G<sub>2</sub>/M arrest, whereas hydroxyurea is thought to lead to an arrest in the G<sub>1</sub>/S-phase through depletion of cellular dNTP pools. To further explore the effect of JMN3-003 on cell cycle progression, we monitored the phosphorylation status of the cdc2-cyclin B kinase after exposure of cells to either the compound, hydroxyurea, nocodazole, or alsterpaullone, a nanomolar small molecule inhibitor of cyclin-dependent kinases that reportedly induces a potent G<sub>1</sub>/S-phase cell cycle arrest. Pivotal in regulating the G<sub>2</sub>/M transition, cdc2-cyclin B kinase is inactivated through phosphorylation during the G<sub>2</sub>-phase. Accumulation in its phosphorylated form thus indicates a G<sub>1</sub> arrest. As in hydroxyurea- and alsterpaullone-treated controls, exposure of cells to JMN3-003 resulted in increased steady state levels of phosphorylated cdc2-cyclin B kinase, supporting a G<sub>1</sub>-phase arrest. To test whether this JMN3-003-induced arrest is permanent or temporary, we next incubated cells in the presence of compound or vehicle alone for 30 hours, followed by removal of the substance and reseeding of cells at identical densities. Monitoring cell growth over an additional 72-hour incubation period in the absence of JMN3-003 revealed that proliferation rates resumed those of untreated control cells after removal of the compound, indicating reversibility of the growth arrest. In contrast to members of the orthomyxovirus family, paramyxovirus replication takes place in the cytosol and, thus, is considered not to be immediately dependent on active cell proliferation. In fact, MeV itself has been shown to induce a G<sub>1</sub>/S arrest in infected T lymphoyctes, confirming that cell cycle progression is not required for virus replication. To directly test whether the JMN3-003-mediated growth arrest per se is causal for the antiviral effect of the compound, we generated MeV-Alaska inhibition curves of JMN3-003 in comparison with the cyclin-dependent kinase inhibitor alsterpaullone. Even at the highest concentration assessed (50 µM), alsterpaullone caused only a marginal reduction in MeV yields. These findings indicate that the antiviral effect of JMN3-003 is based on an upstream effect of the compound rather than being a consequence of the cell cycle arrest itself. ## Cellular mRNA production and protein biosynthesis are unperturbed by JMN3-003 To explore whether growth arrest of treated cells coincides with reduced host cell RNA synthesis or overall cell protein biosynthesis, we next assessed the effect of JMN3-003 on host mRNA and protein production. Relative levels of three signature host mRNAs with short half lives, MCL1, ASB7 and MKP1, were determined by real time PCR after incubation of cells in the presence of different JMN3-003 concentrations ranging from 0.01 to 10 µM. In all cases, mRNA levels of JMN3-003-exposed cells were similar to those of the vehicle-treated references, while exposure to Actinomycin D, which blocks RNA synthesis through arrest of the transcription initiation complex, resulted in a major reduction in relative mRNA levels. Immunodetection of cellular GAPDH and plasmid-encoded MeV F protein under the control of the CMV promoter demonstrated that productive transcription in the presence of the compound furthermore coincides with uninterrupted translation and, in the case of F, co-translational insertion into the host secretory system. Furthermore, equivalent levels of proteolytically processed F<sub>1</sub> material in JMN3-003 and vehicle-exposed cells indicated that intracellular vesicular transport remains intact in the presence of JMN3-003, since cleavage is mediated by the cellular protease furin in a late-Golgi compartment. In contrast to host-encoded or transiently expressed proteins, expression of virus-encoded proteins in the context of paramyxovirus or orthomyxovirus infection was fully blocked by 100 nM JMN3-003. Thus, these observations demonstrate that the compound efficiently suppresses the expression of virus-encoded proteins, but that this is not due to general interference of the inhibitor with cellular mRNA synthesis or translation. This phenotype suggests possible interference of JMN3-003 with early steps of the viral life cycle, such as entry or viral RdRp activity, as the basis for antiviral activity. ## Inhibition of a post-entry step of the viral life cycle To differentiate between those alternatives and identify the point of arrest in the viral life cycle induced by JMN3-003, we first examined whether the compound blocks membrane fusion and thus viral entry. Expression of plasmid-encoded paramyxovirus envelope glycoproteins in receptor-positive cells typically results in extensive cell-to-cell fusion, the hallmark cytopathic effect associated with most paramyxovirus infections in vitro. Transient membrane fusion assays allow a quantitative assessment of whether an inhibitor blocks viral entry or post-entry steps of the viral life cycle. When we examined MeV glycoprotein-mediated cell-to-cell fusion microscopically and in a luciferase reporter-based quantitative cell-to-cell fusion assay in the presence of JMN3-003, we observed extensive membrane fusion indistinguishable from that seen in vehicle-treated controls, indicating that the compound does not act as an entry inhibitor. To determine whether JMN3-003 predisposes host cells against viral infection by inducing an antiviral state, we pre-treated cells with the compound, followed by wash-out of the substance and virus infection after different time periods. Independent of incubation time after removal of the compound, we could not detect any substantial inhibitory effect in this set-up, arguing against priming of the innate antiviral response by JMN3-003. Likewise, preincubation of viral particles with JMN3-003 prior to removal of the article and infection lacked any appreciable antiviral effect, excluding direct virucidal activity of the substance. When added in a time-of-addition experiment at distinct time points post- infection in comparison with two previously characterized, pathogen-targeted antivirals, the inhibition profile of JMN3-003 was distinct from that of the entry inhibitor AS-48 but very closely resembled the profile of the AS-136A RdRp blocker class. Thus, these data point towards inhibition of the viral RdRp activity by JMN3-003 as one possible underlying mechanism for antiviral activity of the compound. ## Host-directed inhibitor of viral RdRp activity For myxovirus infection, the viral RdRp complex mediates both genome transcription and replication to express viral proteins and generate progeny genomes, respectively. Replication occurs through generation of an antigenome of positive polarity, which then serves as template for negative strand genome synthesis. To directly test whether JMN3-003 affects viral RdRp activity in the context of virus infection, we determined the copy numbers of MeV-Alaska mRNA and antigenome in infected, compound-treated cells relative to vehicle-treated controls by quantitative RT-PCR. Presence of JMN3-003 caused a dose-dependent reduction in viral RNA levels. At a concentration of 100 nM, for instance, we observed a \>100-fold reduction of viral mRNA and antigenome copy numbers relative to vehicle-treated samples, indicating potent inhibition of viral replication. For comparison, a concentration of 25 µM of the RdRp inhibitor AS-136A, a nanomolar blocker of MeV replication, was required to achieve comparable mRNA and antigenome reduction levels. When this assay was applied to orthomyxovirus infection, we likewise observed a dose-dependent inhibition of influenza A/WSN antigenome levels relative to vehicle treated controls. Parallel quantification of genome copy numbers of released progeny virus demonstrated that an approximate \>100-fold drop in relative viral antigenome levels correlates to a \>10,000-fold reduction in genome copies of released progeny virions. Assessment of viral RdRp activity in a plasmid-based minireplicon reporter system confirmed dose-dependent inhibition of RdRp by JMN3-003 also in a sub- infection setting, since luciferase reporter expression was fully blocked at compound concentrations of approximately 100 nM. Taken together, these data suggest indirect inhibition of the viral polymerase complex through interaction of the compound with a cellular cofactor required for RdRp activity as the basis for the antiviral effect of JMN3-003. ## JMN3-003 does not induce rapid emergence of viral resistance It has been suggested for different viral pathogens that a host-directed antiviral approach has the potential to reduce the frequency of viral escape from inhibition compared to direct targeting of pathogen components. To explore whether resistance to JMN3-003 could be induced experimentally, we attempted stepwise viral adaptation to growth in the presence of the compound in comparison with the pathogen-specific MeV RdRp inhibitor AS-136A. Following an escalating dose scheme, inhibitor concentrations were doubled when virus-induced cytopathicity became detectable microscopically. While robust resistance to the pathogen-targeted AS-136A control emerged rapidly in an approximate 15 to 20-day time window (tolerated dose at the end of adaptation was ≥30 µM, equivalent to ≥100-fold resistance), only marginal increases in the tolerated dose could be detected for JMN3-003 after 90 days of continued viral incubation in the presence of the substance. These results are consistent with a host-directed mechanism of action of JMN3-003 and suggest the existence of a systemic barrier that prevents rapid viral escape from inhibition by the article. # Discussion In recent years, host cell-directed antivirals have experienced growing recognition as a new concept for the development of advanced generation antivirals with the potential to counteract the challenge of preexisting or rapidly emerging viral resistance. Novel automated genomics and proteomics analyses have greatly advanced our insight into host-pathogen interactions. These studies have underscored the notion that several cellular pathways are exploited for virus replication, supporting the hypothesis that a host-directed antiviral may enjoy an expanded viral target range, rendering it effective for the treatment of several related viral diseases. Technologies applied for host-directed drug discovery include cDNA and siRNA- based microarray analyses combined with pathway-guided data mining , loss-of- function screens using aptamers or small oligonucleotides, protein array analyses and chemical library screening. By combining automated library screening with counter screens against a variety of related viral pathogens of the myxovirus families, we have identified a candidate scaffold that, after moderate hit-to-lead chemistry, adheres to the profile of a host-directed antiviral based on several lines of evidence: I) antiviral activity is host cell species-dependent, indicating specific interaction with a distinct host factor rather than a viral component. Host cell-specific activity is incompatible with compound docking to conserved viral factors. For example, carbohydrate structures exposed on viral envelope glycoproteins that are targeted by antiviral lectins such as pradimicin A. Furthermore, it is incompatible with an undesirable unspecific, promiscuous mode of action of the compound ; II) affinities against a panel of human pathogens of the paramyxovirus family as well as laboratory adapted and wild type influenza virus isolates were very similar throughout (average EC<sub>50</sub> concentrations are ∼40 nM). Equivalent active concentrations argue against compound docking to distinct viral components and suggest that inhibition of distinct myxovirus families follows the same mechanism of action; III) in vitro adaptation attempts to induce viral resistance were unsuccessful even after extended exposure times to the drug. A full assessment of the frequency of viral escape from inhibition by JMN3-003 will certainly need to include in vivo virus adaptation attempts in suitable animal models, since the rate of resistance build-up may vary between tissue culture and in vivo settings. We nevertheless reliably induced resistance in less than 30 days to a pathogen-directed MeV RdRp inhibitor that was analyzed in parallel, which is fully consistent with our previous experience and provides confidence for the validity of our overall experimental design for viral adaptation. Mechanistic analysis of the bioactivity of the JMN3-003 compound class through characterization of exposed cells and time-of-addition experiments revealed two distinct phenotypes, a temporary cell cycle arrest in the G<sub>1</sub>/S phase and an arrest in the myxovirus life cycle at a post-entry step. Current libraries of chemical analogs of JMN3-003 do not yet permit a definitive conclusion as to whether both activities adhere to discrete structure-activity relationships or are causally linked, but a bulk of experimental data demonstrate that host cell cycle arrest per se has no inhibitory effect on replication of paramyxoviruses such as MeV. Not only does the virus itself induce a G<sub>1</sub>/S-phase arrest in infected T lymphocytes, we also found that exposure of infected cells to alsterpaullone, a potent blocker of G<sub>1</sub>/S-phase cell cycle progression through nanomolar inhibition of cellular cyclin-dependent kinases, did not affect the extent of virus replication even at concentrations exceeding reported alsterpaullone EC<sub>50</sub> values by more than 1,000-fold. Likewise consistent with the notion that the antiviral activity of JMN3-003 is not based on cell cycle arrest itself, virus inhibition was not restricted to the context of immortalized, rapidly dividing tissue culture cell lines but extended with equal potency to primary human PBMCs. Reversible cell cycle arrest and block of virus replication indicate non- covalent docking of JMN3-003 to its target structures, which is corroborated by the compound's stability, low chemical reactivity profile and the complete absence of virucidal activity in pre-incubation settings. An inhibition profile of JMN3-003 closely mimicking that of AS-136A, the pathogen-directed blocker of MeV RdRp targeting the viral L polymerase protein, and the block in viral RdRp activity in the context of viral infection and minireplicon reporter assays by JMN3-003 consistently point towards interaction of the compound with a host cofactor essential for RdRp function as the basis for its antiviral activity. While viral RdRp depends on a variety of host cell components , unperturbed cellular mRNA synthesis and, thus, uninterrupted host RNA polymerase function in the presence of compound exclude interference of JMN3-003 with essential transcription initiation factors. Recently, accumulating evidence has implicated host cell kinases as regulators of the activity of RdRp complexes of different negative-strand RNA viruses : host cell kinases of the PI3K-Akt pathway manipulate paramyxovirus RdRp activity through Akt-mediated phosphorylation of the viral phosphoprotein, an essential component of the RdRp complex. Furthermore, Akt activity itself is upregulated through activation of PI3K during influenza A infection via direct interaction of the viral NS1 protein with PI3K. In the case of MeV, however, published data, and our own observations (Krumm and Plemper, unpublished) demonstrate that Akt inhibition causes a moderate reduction in virus release, whereas titers of cell- associated progeny particles remain unchanged. While this rules out the PI3K-Akt pathway as a direct target for JMN3-003, it illuminates the intricate regulatory interactions between pathogen and host, which provide a wealth of possible points of entry for antiviral intervention. Future identification of the molecular target of JMN3-003 carries high potential to further our understanding of these interactions and may conceivably provide a basis for pharmacophore extraction and structure-driven scaffold optimization. We note that the central sulfur in the JMN3-003 chemical scaffold could potentially render the molecule vulnerable to rapid phase I oxidation and thus compromise both metabolic stability and bioavailability. For instance, it has been reported that flavin-containing monooxygenases, dioxygenases and cytochrome P-450 enzymes catalyze oxidation of alkylaryl sulfides to sulfoxides (R<sub>2</sub>S = O). However, the high stability of JMN3-003 in the presence of human hepatocyte subcellular fractions and human plasma argues against an undesirable short in vivo half-life of the substance. This is corroborated by good metabolic stability of the structurally similar HIV reverse transcriptase inhibitor RDEA-806, which shares the central 2-thio-acetamide connector with JMN3-003 and has achieved success in clinical trials: the compound was well tolerated in both Phase 1 and 2a studies after single or multiple oral doses and showed no drug-related CNS toxicity, creating a clinical precedence for the applicability of the broader scaffold. Although RDEA-806 follows a different mechanism of action than JMN3-003 and lacks any anti-paramyxovirus activity, the structural similarities provide sufficient confidence for the overall developmental potential of the JMN3-003 class to recommend it as a promising candidate for advanced synthetic optimization towards preclinical validation and development. In toto, we have identified a novel chemical class of viral inhibitors that block viral RdRp activity with a host factor-mediated profile. A complete activity workup after synthetic identification of a clinical lead analog will be required to fully appreciate the range of the different viral families inhibited by the substance. However, we consider human pathogens of the myxovirus families that are primarily associated with acute disease among the most suitable for host-directed antiviral efforts due to anticipated short treatment regimens. While we cannot exclude that resistance to JMN3-003 may eventually emerge in in vivo settings, our in vitro adaptation efforts support the hypothesis that the mechanism of action of this compound class establishes a strong barrier against rapid viral escape from inhibition. # Materials and Methods ## Cells and viruses All cell lines were maintained at 37°C and 5% CO<sub>2</sub> in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum. Vero (African green monkey kidney epithelial) cells (ATCC CCL-81) stably expressing human signaling lymphocytic activation molecule (CD150w/SLAM), called in this study Vero-SLAM cells, and baby hamster kidney (BHK-21) cells stably expressing T7 polymerase (BSR-T7/5 (BHK-T7) cells) were incubated at every third passage in the presence of G-418 (Geneticin) at a concentration of 100 µg/ml. Lipofectamine 2000 (Invitrogen) was used for cell transfections. Peripheral blood mononuclear cells (PBMCs) were prepared through overlay of whole blood samples from mixed, healthy human donors (Emory University Institutional Review Board approval IRB00045690, Phlebotomy of Healthy Adults for Research in Infectious Diseases and Immunology) on Ficoll Hypaque solution, followed by centrifugation at 240×g for 30 minutes at room temperature and removal of the interphase material. Red blood cells were lysed with RBC lysis solution (Sigma), followed by repeated washing of extracted PBMCs with phosphate buffered saline and transfer to tissue culture plates pre-coated with poly-L-lysine (Sigma). Other primary human cell lines were obtained from PromoCell, Germany. Virus strains used in this study were MeV isolate MVi/Alaska.USA/16.00, genotype H2 (MeV-Alaska), HPIV3, MuV strain South Africa, RSV strain Long, laboratory adapted influenza A strains WSN (H1N1) and PR8/34 (H1N1), swine-origin influenza virus isolates S-OIV Texas and Mexico, vaccinia virus and sindbis virus. To prepare virus stocks, cells permissive for the virus to be amplified (Vero-Slam, Vero, HepG2 (ATCC HB-8065), and Madin-Darby canine kidney (MDCK)) were infected and incubated at 37°C. Cell- associated paramyxovirus and vaccinia virus particles were harvested by scraping cells in OPTIMEM (Invitrogen), followed by release of virus through two consecutive freeze-thaw cycles. Influenza virus and sindbis virus particles were harvested from cell culture supernatants. Titers of MeV and MuV were determined through 50% tissue culture infective dose (TCID<sub>50</sub>) titration according to the Spearman-Karber method as described, titer of all other viruses were determined by plaque assay on permissive cells. ## Influenza A titration by TaqMan RT-PCR To determine genome copy numbers of released progeny influenza A particles (strains WSN, PR8/34, S-OIV Texas and Mexico), culture supernatants of infected MDCK cells (4×10<sup>5</sup> cells/well in a 12-well plate format) were harvested and total RNA prepared using a QIAcube automated extractor and the QIAamp viral RNA mini kit reagent. Purified RNA was then subjected to quantitative real time (qRT) PCR analysis using an Applied Biosystems 7500 Fast real-time PCR system and the qRT-PCR TaqMan Fast Virus 1-Step Master Mix (Applied Biosystems). Primers and probe are based on recent reports and universally reactive with all influenza A strains including the recent S-OIV (H1N1) isolates. To generate a qRT-PCR standard, genome segment seven of influenza A/WSN was subcloned into pCR2.1-TOPO vector (Invitrogen) and copy numbers of the resulting standard calculated using Promega's BioMath Calculator tools (<http://www.promega.com/biomath/>). For each TaqMan reaction, 10-fold serial dilutions of the linearized plasmid ranging from 10<sup>7</sup> to 10<sup>1</sup> were amplified in parallel. ## Compound synthesis Chemical synthesis of compounds AS-48, AS-136A and RDEA-806 was achieved as previously described. Synthesis of JMN3-003, N-(4-methoxyphenyl)-2-nitroaniline (substance (3) in), and analogs JMN5-165 and JMN5-166 was achieved as outlined schematically in. To prepare inhibitor stocks, compounds were dissolved at 75 mM in DMSO. ## Viral CPE-reduction assay Vero-SLAM cells were infected with MeV-Alaska at an MOI of 0.4 pfu/cell in the presence of the inhibitor analyzed ranging from 75 µM to 293 nM in two-fold dilutions. At 96 hours post-infection, cell monolayers were subjected to crystal violet staining (0.1% crystal violet in 20% ethanol), and the absorbance of dried plates at 560 nm determined. Virus-induced cytopathicity was then calculated according to the formula \[% rel. CPE = 100−(experimental- minimum)/(maximum-minimum)\100\], with minimum referring to infected, vehicle- treated wells and maximum to mock-infected wells. ## Virus yield reduction assay Cells were infected with the specified myxovirus at an MOI = 0.1 pfu/cell (all paramyxoviruses assessed), 0.05 pfu/cells (influenza viruses), 1.0 (vaccinia virus), or 10 sindbis virus) in the presence of a range of compound concentrations or equivalent volumes of solvent (DMSO) only, and incubated in the presence of compound at 37°C. When vehicle treated controls approached the end of the logarithmical growth phase, progeny viral particles were harvested and titered by TCID<sub>50</sub> titration, plaque assay or TaqMan real-time PCR, respectively, as described above. Plotting virus titers as a function of compound concentration allowed quantitative assessment of resistance. Where applicable, 50% inhibitory concentrations were calculated using the variable slope (four parameters) non-linear regression-fitting algorithm embedded in the Prism 5 software package (GraphPad Software). ## Quantification of compound cytotoxicity A non-radioactive cytotoxicity assay (CytoTox 96 Non-Radioactive Cytotoxicity Assay, Promega) was employed to determine the metabolic activity of cell after exposure to the compound. In a 96-well plate format, 10,000 cells per well were incubated at 37°C for 24 hours in four replicates per concentration tested in the presence of compound in two-fold dilutions starting at 75 µM. Substrate was then added and color development measured at 490 nm using a BioRad plate reader. Values were calculated according to the formula \[% toxicity = 100−((experimental-background)/(maximum(vehicle treated)-background)\100)\]. Values were plotted in dose-response curves and, if applicable, CC<sub>50</sub> concentrations calculated. ## In vitro assessment of metabolic and plasma stability JMN3-003 was mixed with liver S9 fractions (protein concentration 2.5 mg/ml) from pooled mixed gender humans (XenoTech) at a final concentration of 1 µM and reactions initiated by the addition of cofactors (1.14 mM NADPH, 1.43 mM glucose-6-phosphate, 1.43 mM uridine 5′-diphosphoglucuronic acid, 9.42 mM potassium chloride, 2.28 mM magnesium chloride) in 100 mM potassium phosphate buffer, pH 7.4. Samples were incubated at 37°C with mixing, aliquots removed after 0, 15, 30 and 60 minutes and subjected to reversed-phase LC-MS/MS (Applied Biosystems API 4000 QTRAP with heated nebulizer; Turbo IonSpray for JMN5-166) analysis. Peak areas were measured to calculate half life and percent of input compound remaining according to the formulas t<sub>1/2</sub> = (−0.693/slope of linear regression analysis of log transformed peak area versus) and % input remaining = (peak area of test compound at t<sub>x</sub>/peak area of test compound at t<sub>0</sub>)\100. Positive controls to assess the metabolic competency of the liver S9 fractions were 7-Ethoxycoumarin, Propranolol, and Verapamil (Sigma), which were analyzed in parallel to the article. To determine compound plasma stability, articles were mixed with freshly prepared human plasma at a final concentration of 0.5 mM and incubated at 37°C for up to 120 minutes. Aliquots were removed at distinct time points as indicated and analyzed by LC-MS/MS with detection of the compound at 254 nm. Values are expressed as percent of compound remaining at each time relative to the amount of that compound present at the starting time point. ## Flow-cytometric analysis of cell cycle progression Actively proliferating HeLa cells were exposed to JMN3-003 (10 µM), hydroxyurea (4 mM), or nocodazole (200 ng/ml) for 36 hours, followed by resuspension in buffer I (20 mM citrate/PO, pH 3.0, 0.1 mM EDTA, 0.2 M Sucrose, 0.1% Triton X-100) and staining in buffer II (10 mM Citrate/PO, pH 3.8, 0.1 M sodium chloride, 20 µg/ml acridine orange) as described. Green fluorescence at 525 nm resulting from DNA intercalating acridine orange was then measured using a BD LSRII flow cytometer and FlowJo software (Tree Star) for data analysis. For comparison, unstained and stained, solvent-only exposed cells were examined in parallel. ## SDS-PAGE and immunoblotting Cells were lysed with RIPA buffer (50 mM Tris/CL, pH 7.2, 1% deoxycholate, 0.15% sodium dodecylsulfate, 150 mM sodium chloride, 50 mM sodium fluoride, 10 mM EDTA, 1% NP-40, 1 mM PMSF, protease inhibitors). Aliquots with equal total concentrations of cleared lysates (20,000×g; 10 min; 4°C) were mixed with 2x-urea buffer (200 mM Tris, pH 6.8; 8 M urea; 5% sodium dodecyl sulfate (SDS); 0.1 mM EDTA; 0.03% bromphenol blue; 1.5% dithiothreitol) and denatured for 25 min at 50°C. Samples were then fractionated on 10% SDS-polyacrylamide gels, blotted to polyvinylidene difluoride (PVDF) membranes (Millipore) and subjected to enhanced chemiluminescence detection (Pierce) using specific antisera directed against phosphorylated or non-phosphorylated cdc2-cyclin B kinase (Cell Signaling Technology), GAPDH (Abcam), the cytosolic tail of the MeV F protein, or influenza A/WSN virus M2 (Thermo Scientific). Immunostained PVDF membranes were developed using a ChemiDoc XRS digital imaging system (Bio-Rad) and horseradish peroxidase conjugated anti-species IgG (mouse or rabbit) antibodies. For densitometry, signals were quantified using the QuantityOne software package (Bio-Rad). ## Assessment of cell growth rates Vero cells were seeded at a density of 6×10<sup>5</sup> cells and incubated in the presence of 10 µM JMN3-003 or vehicle only for 30 hours at 37°C. Cells were then washed extensively and reseeded at a density of 1×10<sup>5</sup> cells per well, followed by continued incubation at 37°C and assessment of life/dead cell numbers every 24 hours using a Countess automated cell counter (Invitrogen). Cells were reseeded as before when fastest growing cultures approached confluency. Growth rates were calculated for each 24-hour time interval using the Prism software package (GraphPad Software Inc.) based on the formula Y = Y<sub>0</sub>\exp(K\X) with Y equaling life cell numbers, Y<sub>0</sub> the Y value at the starting time (t<sub>0</sub>), and K the growth constant equaling ln(2)/doubling-time. ## Quantification of cellular and viral mRNA levels Cells were infected with either recombinant MeV Edmonston (recMeV-Edm) (Vero cells, MOI = 1.0) or influenza A/WSN (MDCK cells, MOI = 0.05), followed by removal of inocula one hour post-infection and addition of JMN3-003 in growth media at 0.1 µM or 1 µM. All MeV infected wells received in addition fusion inhibitory peptide (FIP, Bachem) at 100 µM to prevent premature breakdown of the monolayer through viral CPE in the vehicle control wells prior to RNA extraction. Twenty-four (influenza A/WSN) or forty (recMeV-Edm) hours post- infection, total RNA was prepared from all wells using the QIAcube automated extractor and the RNeasy Mini Kit (Qiagen), and subjected to reverse transcription using Superscript II Reverse Transcriptase (Invitrogen). For RNA samples originating from recMeV-Edm infected cells, antigenome-specific primer 5-GGCTCCCTCTGGTTGT or oligo-dT primer (viral mRNA and GAPDH quantification) were used for cDNA priming. In the case of samples originating from influenza A/WSN infected cells, primers for cDNA synthesis were 5-AGTAGAAACAAGGTAGTTT (antigenome) or oligo-dT (mRNA and canine GAPDH). Real-time reactions were carried out using an Applied Biosystems 7500 Fast real-time PCR system and iQ* Fast SYBR Green Supermix with ROX (Bio-Rad). Probes were a fragment at the N/P junction (MeV antigenomic RNA, 5-AACCAGGTCCACACAG and 5-GTTG TCTGATATTTCTGAC), a fragment of MeV F mRNA (5-GTCCACCATGGGTCTCAAGGTGAACGTCTC and 5-CAGTTATTGAGGAGAGTT), a fragment of human GAPDH (SABiosciences proprietary primers), a fragment of influenza A/WSN segment seven (influenza A/WSN antigenomic RNA, 5-tagctccagtgctggtct and 5-AAGGCCCTCCTTTCAGTCC), and a fragment of canine GAPDH (Qiagen proprietary primer). Melting curves were generated at the end of each reaction to verify amplification of a single product. To calculate ΔΔC<sub>T</sub> values, CT values obtained for each sample were normalized for GAPDH as reference and then ΔC<sub>T</sub> values of JMN3-003-treated samples normalized for the FIP-treated controls. Final quantification was based on three independent experiments in which each treatment condition and RT primer setting were assessed in triplicate. To assess the relative quantities of cellular mRNA, 9×10<sup>5</sup> HeLa cells were incubated in the presence of JMN3-003 (0.01, 0.1, 1.0, 10.0 µM final concentration), AS-136A (25 µM), Actinomycine D (5 µg/µl), or vehicle only for six hours at 37°C, followed by preparation of total RNA as described above. Quantitative TaqMan RT-PCR was again achieved using the TaqMan Fast Master Mix (Applied Biosystems) combined with proprietary primer and probe sets specific for Induced myeloid leukemia cell differentiation protein 1- (MCL1), MAPK phosphatase 1 (MKP1), and ankyrin repeat and SOCS box-containing protein 7- (ASB7) encoding mRNAs (Applied Biosystems). Samples were standardized for GAPDH as before and normalized values expressed relative to the equally analyzed vehicle-treated controls. ## Quantitative cell-to-cell fusion assays An effector cell population (3×10<sup>5</sup> cells/well) was cotransfected with 2 µg each of MeV H and F expression plasmids. To inhibit fusion until the cell overlay, the effector cells are incubated in the presence of 100 µM fusion inhibitory peptide (Bachem). Single transfections of plasmids encoding MeV F served as controls. Target cells (6×10<sup>5</sup> cells/well) were transfected with 4 µg of the reporter plasmid encoding firefly luciferase under the control of the T7 promoter. Two hours post-transfection, modified vaccinia virus Ankara expressing T7 polymerase at an MOI of 1.0 PFU/cell was added to the effector cells. Following incubation for 16 h at 37°C, target cells were detached and overlaid on washed effector cells at a 1∶1 ratio and incubated at 37°C in the presence of different JMN3-003 concentrations as indicated. Four hours post- overlay, cells were lysed using Bright Glo lysis buffer (Promega), and the luciferase activity determined in a luminescence counter (PerkinElmer) after addition of Britelite substrate (PerkinElmer). The instrument's arbitrary values were analyzed by subtracting the relative background provided by values of the controls, and these values were normalized against the reference constructs indicated in the figure legends. On average, background values were \<1% of the values obtained for reference constructs. For qualitative assessment, transfected Vero-SLAM cells were photographed 18 hours post-transfection at a magnification of ×200. ## Time of compound addition analysis For virus pre-incubation assays, 10<sup>7</sup> infectious MeV-Alaska particles were incubated for 60 minutes at 37°C in the presence of JMN3-003 (1.0 µM final concentration) or vehicle only, followed by 1,000-fold dilution in growth media and transferred to 3×10<sup>5</sup> Vero-Slam cells/well (corresponding to final compound concentrations after pre-incubation of 1 nM and an MOI = 0.033). Reference wells were kept at 1.0 µM JMN3-003 for the duration of the experiment. Cell-associated viral particles were harvested 24 hours post-infection and infectious titers determined by TCID<sub>50</sub> titration. To assess cell priming, Vero-Slam cells (3×10<sup>5</sup>/well) were incubated in the presence of JMN3-003 at 1.0 µM for one hour at 37°C at the indicated time points pre- infection, followed by washing and further incubation in growth media. Immediately before infection, cells were reseeded at a density of 2.5×10<sup>5</sup> per well and infected with MeV-Alaska at an MOI = 0.2 pfu/cell. Inocula were replaced with growth media four hours post-infection and cells incubated for approximately 20 hours. Cell-associated viral particles were then harvested and infectious titers determined by TCID<sub>50</sub> titration. For post-infection time-of-addition studies, 3×10<sup>5</sup> Vero-Slam cells/well were infected with MeV-Alaska as before, followed by addition of JMN3-003 (1.0 µM final concentration), entry inhibitor AS-48 (75 µM), or RdRp inhibitor AS-136A (25 µM). Controls received vehicle only. All wells were harvested 19 hours post-infection and titers of cell-associated progeny virus determined by TCID<sub>50</sub> titration. ## Minireplicon assays BSR T7/5 cells (5×10<sup>5</sup>/well) were transfected with plasmid DNAs encoding MeV-L (0.24 µg), MeV-N (0.94 µg) or MeV-P (0.29 µg) and 2 µg of the MeV luciferase minigenome reporter plasmid. Control wells included identical amounts of reporter and helper plasmids but lacked the L-encoding plasmid. At the time of transfection, JMN3-003 was added as specified, while control wells received vehicle only for comparison. Thirty-six hours post-transfection, cells were lysed with Bright GLO lysis buffer and relative luciferase activities determined using the Britelite substrate and a luminescence counter as outlined above. ## In vitro virus adaptation Adaptations were carried out essentially as we have previously described. Briefly, Vero-SLAM cells were infected with MeV-Alaska at an MOI of 0.1 pfu/ml and incubated in the presence of gradually increasing JMN3-003 concentrations starting at 0.5 µM. Equally infected cells treated with the virus polymerase targeted RdRp inhibitor AS-136A were examined in parallel. When cultures became over confluent, cells were reseeded for continued incubation in the presence of the same compound concentration as before. At detection of extensive cell-to- cell fusion, cell-associated viral particles were harvested, diluted 10-fold and used for parallel infections of fresh cell monolayers in the presence of compound at unchanged and doubled concentrations. Cultures treated with the highest compound concentrations in which virus-induced cytopathicity became detectable were used for further adaptation. The approach was terminated after 90 days of continued incubation or when virus-induced cytopathicity was readily detectable in the presence of 30 µM compound in accordance with previous results. # Supporting Information We are grateful to D. C. Liotta and R. W. Arrendale (Emory University and Emory Institute for Drug Development) for support and LC-MS/MS analysis of compound samples, and A. L. Hammond for critical reading of the manuscript. MuV strain South Africa, sindbis virus, and SO-influenza isolates Texas and Mexico were kind gifts of P. A. Rota (Centers for Disease Control and Prevention), W. J. Bellini (Centers for Disease Control and Prevention) and D. A. Steinhauer (Emory University), respectively. We also thank Scynexis Inc. for assistance with experimentation involving human S9 hepatocytes. [^1]: Conceived and designed the experiments: SAK MN JPS RKP. Performed the experiments: SAK JMN J-JY MD AS RKP. Analyzed the data: SAK JMN JPS RKP. Contributed reagents/materials/analysis tools: AS MN. Wrote the paper: SAK JMN RKP. [^2]: The authors have declared that no competing interests exist.
# Introduction HIV misdiagnosis can lead to severe individual and public health consequences. A false-positive HIV diagnosis may lead to stigma and discrimination, strains on family relationships and reproductive choices, and unnecessary lifelong use of medication. From a public health perspective, HIV misdiagnosis can undermine the public’s trust in HIV test results and in testing programs, and, can lead to inefficient spending associated with costly antiretroviral treatment (ART) and large settlements from lawsuits brought on by cases of misdiagnosis. The use of recommended HIV rapid testing algorithms, comprised of at least two consecutive reactive assays in a high prevalence settings, or at least three consecutive reactive assays in a low prevalence settings, minimizes false- positive results. However, a variety of factors, including but not limited to: user error, poor recordkeeping, inadequate management and supervision, use of incorrect testing algorithms, cross-reactivity and over-interpretation of weak reactive results, contribute to HIV misdiagnoses. In global settings, studies have reported HIV misdiagnosis rates ranging from less than 1% to more than 10%. The World Health Organization (WHO) recommends retesting for verification of all HIV-positive cases prior to ART initiation, in order to reduce the frequency of HIV misdiagnosis and prevent unnecessary initiation of ART. The recommendation calls for repeating the same nationally validated rapid testing algorithm on all HIV cases initially diagnosed, using a second blood sample, and a different tester. Retesting for verification is additional and subsequent to the two or three assays used as part of a rapid HIV testing algorithm leading to an initial HIV-positive diagnosis. Retesting for verification applies only to persons not on ART because HIV diagnostic tests validated for use on persons taking ART are not available. And, once persons are on ART, rapid tests may give false negative results due to waning of antibodies. WHO’s 2015 “Treat All” guidelines call for the immediate initiation of ART for all HIV-diagnosed persons, irrespective of their CD4 count. The benefits of “Treat All” include improved health outcomes for persons living with HIV, a more efficient linkage to treatment process and population benefit of reduction in transmission. ART initiation upon testing positive for HIV, does not allow a health care provider to observe the natural history of CD4 levels and question the diagnosis in the absence of a CD4 decline. Therefore, the adoption of “Treat All” and the associated surge in ART initiations creates an imminent window of opportunity and impetus for adopting retesting for verification as a manageable component to strengthening the quality of HIV care services. While WHO first recommended retesting for verification of new diagnoses in 1997, the recommendation has been re-emphasized with the introduction and adoption of the “Treat All” guidelines. Retesting for verification has not been widely adopted nor implemented. This slow adoption may be related to: lack of knowledge about the retesting recommendation; providers’ reliance on other clinical assessments as required under previous ART guidelines that were indicative of HIV infection; lack of data about the magnitude of misdiagnosis; concerns of the additional costs and resources needed to implement retesting for verification or suspicions that the operational requirements of re-testing could be an impediment to rapid ART initiation. An analysis among pregnant women suggests that retesting is less expensive than treatment of those with false HIV-positive status. A study of misdiagnosis also among pregnant women suggests that laboratory-based confirmation of HIV among those with an undetectable pre-ART viral load, is cost-saving when compared to lifetime cost of ART program enrolment for those misdiagnosed. Another study of HIV misdiagnosis modeled the effect of retesting on a low- and high-prevalence cohort of 10,000 persons and suggests that the cost of retesting is recouped quickly when compared to the cost of ART for those misdiagnosed over 30 years. Our study estimates the costs of implementing retesting across countries in Africa, and, to our knowledge, is the first cost analysis to base data inputs on reviews of misdiagnosis rates and testing costs, and the gap in the number of persons needing to be initiated on ART to reach 90% coverage by 2020. We conducted a cost analysis to estimate and compare the total costs of implementing the WHO HIV retesting for verification recommendation in Africa, to the costs of not retesting. We define the positive result of an initial testing algorithm, as an initial diagnosis, and the positive result following retesting for verification as a verified diagnosis. We do not use the terms confirmatory testing or confirmed diagnosis to avoid confusion. We compared the cost of retesting individuals for verification of an initial diagnosis, to the cost of ART associated with the cases who would be HIV misdiagnosed, without retesting for verification. With this model-based evaluation, we aim to provide policy- makers in governments of low- and middle-income countries, and other resource- constrained settings, with evidence to support their decision-making process around HIV retesting for verification strategies. # Methods We conducted a cost analysis of HIV retesting for verification, from the perspective of the health care provider. We compared two scenarios: retesting, implementing retesting for verification of all initially diagnosed HIV-positive cases prior to ART initiation; and no retesting, which is the current practice in most developing countries. In terms of ART costs, we considered only the costs incurred for those misdiagnosed as HIV-positive because the cost of ART for true positives would be the same in both scenarios. We considered all countries in Africa listed as member states of the United Nations African Group, except Equatorial Guinea, and the Seychelles, Cape Verde and São Tomé & Príncipe islands. We presented both the total costs and per person cost of these scenarios for these 50 African countries, grouped by income level: 26 low-income countries (LICs), 15 lower-middle income countries (LMICs) and 9 upper middle- income countries (UMICs). Regions outside of Africa were not evaluated due to paucity of data. We assumed retesting for verification would largely be conducted in facility- based settings, even when the initial diagnosis is provided in community-based settings because the recommendations for retesting call for a different person to administer the verification test. We used facility-based testing costs, which do not typically include recruitment and mobilization efforts, to approximate retesting costs. We reviewed the literature for studies, set in African countries, published in the last ten years and reporting facility-based HIV testing costs per person. We searched PubMed using the terms: cost, HIV, testing, and facility, and limited the publication year from 2007 to present. To identify further literature, we searched reference lists. We identified five relevant studies, from two systematic review of HIV testing costs, and an additional three relevant studies published subsequent to those systematic reviews. We present the facility-based cost per person tested resulting from these eight studies in. HIV counseling and testing services in these studies are largely provided by trained counselors, occasionally supported by nurses or laboratory assistants, and all are using rapid diagnosis testing in the initial algorithm. All eight studies considered costs from the perspective of the heath care provider, and, all but one study adopted an economic costing approach. For each of the three income level country groups, we used the average cost per person tested to estimate the cost of retesting for verification, in our base- case analysis. In sensitivity analysis, we varied the retesting costs, between the lowest and the highest reported facility-based cost per person tested, by country income level. To estimate the resources needed to fund the 2016 UNAIDS Fast-Track Approach, Stover et al. conducted a review of published ART cost studies. We used this review, limited to Africa, to estimate the average annual cost of ART for each of the three income level country groups. The average annual cost of ART per client included antiretroviral drugs, other drugs, laboratory services and other service delivery costs. All cost data were adjusted to 2017 United States Dollars. In sensitivity analysis, we varied these costs between the lowest and the highest reported average annual cost of ART, by country income level. A recent systematic review found 30 studies reporting a false positive HIV diagnostic error rate, with a median rate of 3.1%. We used a subset of those studies to determine the false positive misdiagnosis rate following an initial testing algorithm. From those 30 studies, we excluded seven not set in Africa; ten were excluded for reporting on misdiagnosis following discordant rapid test results, including use of a tie-breaker algorithm which are known to have lower specificity. Further reasons for exclusion were reporting on acute infection among negative samples, and reporting on oral fluid testing, or insufficient data. The remaining 11 studies were deemed relevant and allowed for pooling of the individual study estimates. For each country income level, we estimated a weighted mean false positive misdiagnosis rate by summing the number of false positives reported across studies and dividing by the number of positives, also summed across studies. These data are presented in. For LICs, the misdiagnosis rate ranged from 0.7% to 10% and the weighted mean was 2.7%. For LMICs, the misdiagnosis rate ranged from 0.3% to 5% and the weighted mean was 1.1%. For UMICs, the misdiagnosis rate ranged from 0.3% to 2% and the weighted mean was 0.85%. In sensitivity analysis, we evaluated the results across these ranges. Few studies indicate the reduction in misdiagnoses that could be obtained from retesting. Reports from Malawi show that misdiagnosis has decreased from 7% to 1% following the implementation of retesting for verification and related quality assurance measures. Another study reported that re-testing for verification prevented six out of eight false positive misdiagnoses. In our model, we assumed a 75% reduction in misdiagnoses would result from retesting for verification, and varied this parameter in sensitivity analysis and threshold analysis. Varying the reduction in misdiagnosis following retesting can also serve to capture any loss to follow-up attributable to the extra step of retesting for verification, and to correct any overestimate resulting from the concurrent effects associated with the implementation of other quality improvement measures. To close the HIV treatment gap and reach 90% ART coverage by 2020, UNAIDS reports the number of persons who would need to be initiated on ART, by 2020 and by country. We used these data to estimate the number of initially diagnosed HIV-positive persons who are not on ART, and the total ART costs of the projected misdiagnosed cases, for each of the three country income level groups. Our analysis is presented in the context of “Treat All” and assumes that all those diagnosed as HIV-positive, would be initiated and incur ART costs. We define costs per person initially diagnosed as the average per person costs across the total number of initially diagnosed, ART-naïve HIV cases, prior to retesting for verification. We estimate the per person costs and total costs of the retesting and no retesting scenarios over a ten year time horizon, as the base case. In sensitivity analysis, we explored a 5-year and 20-year time horizon. We assumed that all those initially-diagnosed HIV-positive persons who are not on ART would incur costs from the start of the time horizon. Future costs were discounted at rate of 3%. To gauge the robustness of the findings, we conducted univariate sensitivity analysis and threshold analysis on the model input parameters. All input values for the base-case analysis and ranges examined in sensitivity analysis are presented in. # Results and discussion Using the base-case values, total costs under the no retesting scenario exceed the total costs under the retesting scenario in all three country groups, suggesting that cost savings are associated with the adoption of retesting for verification. Over the ten-year horizon, the estimated number of misdiagnoses in the absence of retesting for verification was 156,117, 52,720 and 29,884, for LICs, LMICs and UMICs, respectively, totaling 238,721 for Africa. Using a 75% reduction in misdiagnosis from retesting for verification, the estimated number of misdiagnosed cases with retesting for verification was 39,029, 13,180 and 7,471 for LICs, LMICs and UMICs, respectively, totaling 59,680 for Africa, over the ten-year horizon. The total cost of treatment for those misdiagnosed cases under the no retesting scenario was \$727 million, \$199 million, and \$266 million (2017 United States Dollars (USD)) for LICs, LMICs and UMICs, respectively, over the ten-year horizon. And, costs per person initially diagnosed under the no retesting scenario, defined as the total costs divided by the estimated of number of positives to be initiated on ART, were: \$125, \$43, and \$75 for (2017 USD) for LICs, LMICs, and UMICs, respectively. Total costs under the retesting scenario include the cost of retesting all positives and the costs of ART for those who would be misdiagnosed in spite of the retesting efforts; these are: \$231 million, \$98 million, and \$146 million (2017 USD) for LICs, LMICs and UMICs, respectively. Costs per person initially diagnosed under the retesting scenario, defined as the total costs divided by the estimated of number of positives retested, were: \$40, \$21, and \$42 for (2017 USD) for LICs, LMICs, and UMICs, respectively. When comparing, the two scenarios, the estimated savings from retesting totals \$717 million over the ten-year horizon, across all country income levels when compared to the no retesting scenario. And, the savings per person retested are \$85, \$21, and \$33 for LICs, LMICs, and UMICs, respectively, when retesting for verification is implemented. Results of the base-case analysis are presented in. In one-way sensitivity analysis, we varied the base-case input parameters, across the ranges indicated in, and the time horizon to 5 and 20 years, for all three country groups. In all sensitivity analysis scenarios explored, the total cost of no retesting exceeded the total cost of retesting, with three exceptions. First, at a per person cost of treatment as low as \$257, the estimated cost of retesting in UMICs would exceed the cost no retesting by \$29 million. Also, in both LMICs and UMICs, the lower bound on the rate of positive misdiagnosis causes the cost of retesting to exceed the cost no retesting by \$6 million and \$5 million, respectively. We conducted threshold analysis on the input parameters to establish the parameter values for which for the total cost of both scenarios would be the same. For the total cost of retesting to equal the total cost of no retesting, with other parameters held constant, the per person cost of HIV retesting would have to increase from \$9 to \$98 in LICs, from \$11 to \$33 in LMICs, and from \$23 to \$58 in UMICs. Also, for the total costs to be the same in both scenarios, the mean annual cost of ART would have to decrease from \$532 to \$48 in LICs, from \$432 to \$139 in LMICs, and from \$1,016 to \$405 in UMICs. Moreover, for the total cost of retesting to equal the total cost of no retesting, with other parameters held constant, the misdiagnosis rate at initial diagnosis would have to be reduced to less than 0.2% in LICs, less than 0.4% in LMICs, and less than 0.3% in UMICs. Alternatively, the reduction in misdiagnosis rate following retesting would have to be reduced from 75% to less than 30%, 24% and 7%, in UMICs, LMICs, and LICs, respectively. Lastly, the cost of HIV treatment for misdiagnosed cases would have to be cumulated over a time horizon of less than ten months for the total cost of retesting to equal the total cost of no retesting in LICs, less than 3 years in LMICs, and, less than four years for UMICs. Results of the sensitivity analyses and threshold analyses are presented in. Our results indicate that adopting retesting for verification leads to significant savings in medical costs averted, even under conservative cost and outcome assumptions. While there is uncertainty in the value of the model’s input parameters, sensitivity analyses and threshold analyses demonstrate that the conclusion holds under a majority of scenarios. For each scenario, reviewing where threshold lies relative to the range defined by the lower and upper bound in sensitivity analysis, provides an indication of the stability of the analysis. For LICs, all threshold values lie considerably outside of the range explored in sensitivity analysis indicating that reasonable variations in parameter input values, would not alter the findings; this is critical because of the three country groups considered, LICs have both the highest rate of positive misdiagnosis and the most number of positives to be initiated, and thus the greatest potential for misdiagnosis in the absence of retesting. Among LMICs, the threshold value for misdiagnosis rate is 0.36%, and a lower bound of 0.31% misdiagnosis was set based on a study, in Nigeria which found one misdiagnosis among 318 tested. All other ranges of input values explored for LMICs excluded the threshold. Lastly, for UMICs, the threshold for misdiagnosis rate was 0.34% which lies faintly beyond the lower bound of 0.32%, and the threshold for the treatment costs was \$405 while the lower bound in was \$257. These trends in country income levels suggests that, while retesting is cost- saving, the case for retesting is greater where resources are more constrained. It is possible, but not known, whether any correlation exists between country income level and misdiagnosis rates. Better resourced countries may be able to afford more quality assurance activities such as training and proficiency testing. Our findings are also consistent with those of a recent analysis by Eaton, in spite of the different approach taken. Our analysis considers discounted ART costs over ten years, rather than a 30-year time horizon, by which antibody and viral load testing technologies are likely to have changed the landscape of HIV diagnosis and treatment. Our study also benefited from a systematic review to establish a baseline rate and range for false positive diagnosis, while the aforementioned study relied on assumptions based on rapid test specificity data. Lastly, our study provides the total costs of retesting using the estimated numbers of persons requiring ART initiation across 50 countries in Africa, grouped by income level, while the Eaton study evaluates the costs of retesting assuming a 1% and 10% HIV prevalence among a cohort of 10,000. In spite of these differences both analyses, along with another, suggest that, at present, retesting for verification is cost-saving relative to the treatment of false positive HIV cases. Our analysis has limitations. First, the costs considered are those incurred by the health care provider only, and broader societal costs were not considered. At the individual level, the consequences of an HIV diagnosis may include stigma and discrimination, violence, psychological trauma, and productivity losses; and it can influence partner, reproductive, and professional choices. Published case reports and media accounts of HIV misdiagnosis suggest that significant jury awards have been made to victims of HIV misdiagnosis and this negative publicity can lead to credibility loss for the public health system and its providers. We did not attempt to quantify or cost the effects of these repercussions. However, considering these costs would increase the total cost of no retesting, and would serve to encourage the implementation of HIV retesting for verification prior to ART initiation. Second, the model considered the cost of the retesting events but did not consider the costs associated with the adoption and rollout of a retesting for verification strategy such training and dissemination. Training is an integral component of introducing new operating procedures in HIV testing, and, all HIV testing services must be implemented along with supervision and quality assurance programs. Third, retesting would call for a two-test algorithm, in high prevalence settings, or a three-test algorithm, in low prevalence settings. To save time and labor costs when retesting, we assumed that all tests would be run in parallel, and therefore the impact on cost of a two- vs. three-test algorithm is the cost of the rapid test kit itself, which is typically a small fraction of testing costs. Also, in our analysis, retesting for verification was defined as repeating the same testing algorithm. And, retesting can rule out possible technical or clerical errors, including specimen mix-up through mislabeling and transcription errors, as well as random error by either the provider or the test device, though retesting for verification will not exclude misdiagnosis related to poor choice of a testing algorithm or cross- reactivity. However, this risk should be reduced assuming the testing algorithm used is validated. The WHO recommends following the approved national HIV testing algorithm and ensuring adherence to a quality-testing program with continuous quality assurance. Retesting may provide significant correction from multiple sources of errors including, lack of training, quality assurance and over interpretation of weakly reactively test lines, as a different provider should be reading the results using a new blood sample. Retesting is common for diseases, including HIV, by reference laboratories in Western countries, but was not practiced routinely in resource-limited countries. Given the wide adoption of “Treat All” guidelines, it is critical to incorporate retesting for verification as part of quality assurance programs, to avoid misdiagnosis and unnecessary costs. Other approaches to prevent misclassification include viral load testing, and supplemental HIV testing methods, that are more expensive and more specific than a rapid testing algorithm. We did not explicitly model these approaches as they are used to address different points in the diagnostic error cascade from those addressed by retesting. Viral load testing, may further confirm a positive diagnosis when provided to ART-naïve individuals, is not a substitute for HIV retesting for verification because it addressed a different portion of errors. Also, an undetectable viral load at baseline cannot be used to rule out infection because up to 10% of HIV-positive persons may have an undetectable viral load as their initial baseline measurement, without having been on ART. Therefore, an undetectable viral load at baseline can include both true HIV- positive and false positives. And, the turnaround time to obtain viral load testing results is, at present, much greater than that of obtaining results from retesting using a rapid HIV testing algorithm, and any delay introduced in getting patients diagnosed and initiated on treatment increases the risk of losing patients to follow-up and puts the achievement of 90% ART coverage by 2020 in peril. Lastly, in low- to middle-income country settings, there is little capacity to provide viral load testing prior to initiation of ART, and it will take several years to ramp up this capacity sufficiently to cover all newly diagnosed HIV cases. In the future, viral load testing may have a role in diagnosing HIV, but at present, given current technology and capacity, detection of HIV antibody testing is the most appropriate method to diagnose and confirm infection, especially in developing countries. # Conclusion In sum, this analysis suggests significant savings in HIV treatment costs averted from the adoption of WHO’s retesting for verification guidelines and the consequent reduction in misdiagnosis. Our study ought to drive attention to the incorporation of retesting for verification as part of quality assurance for “Treat All” services, and to the policy and operational needs for routinizing retesting. Beyond adoption of the guidelines, participating in continuous quality improvement and adhering to the nationally validated HIV testing algorithm also plays a key role in reducing HIV misdiagnosis rates, and a comprehensive approach to quality assurance of HIV testing is critical whether for initial diagnosis or for verification purposes. Implementation and operational aspects of retesting for verification may present complexities, for example, where HIV testing is conducted in community settings, or where health care facilities do not have an another person qualified to administer retesting for verification. We suggest that program managers proactively design the implementation of retesting for verification to overcome any hurdles to operationalizing the strategy. For countries adopting a phased approach, retesting for verification may be prioritized to testing sites that appear to indicate poor performance on external quality assessments and areas where higher level of discordant results are observed. We also recommend that countries review their data, consider the long term financial, personal, and societal costs associated with not retesting for verification, and include retesting for verification as an integral quality assurance component in the development or revision of their “Treat All” plans. This work has been supported by the President’s Emergency Plan for AIDS Relief (PEPFAR) through the U.S. Centers for Disease Control and Prevention (CDC). # Disclaimer The findings and conclusions in this report are those of the author(s) and do not necessarily represent the official position of the Centers for Disease Control and Prevention. [^1]: The authors have declared that no competing interests exist.
# Introduction Fusarium head blight (FHB) is a notorious wheat disease prevailing in warm and humid environments, exerting global impact on food and feed safety due to the presence of mycotoxins produced by Fusarium species, the causal agents of FHB. Deoxynivalenol (DON) has been considered the most important FHB-related mycotoxin and legislation has been set up in many countries/organizations for controlling DON content in food and feed. Host resistance to FHB is of quantitative inheritance and influenced significantly by environment, making breeding for this trait a difficult task. Multiple mechanisms of host resistance to FHB has been recognized, including Type I for resistance to initial infection, Type II for spread of pathogen in spike tissues, Type III for DON accumulation, Type IV for kernel infection, and Type V for yield reduction. In relation to food safety, Type III resistance is the most important; but so far no validated QTL specific for this resistance mechanism has been identified, and some researchers still regard it as a consequence of FHB infection and not an independent trait. Of the first two resistance mechanisms, Type I resistance exhibited more frequent association with phenological, morphological, and flower biology traits, such as days to heading (DH), plant height (PH) and anther extrusion (AE). The negative association between PH and FHB susceptibility in wheat has long been observed, and it happened also in barley and oat. Three possible mechanisms have been proposed for the association, i.e. disease escape, pleiotropy of reduced height (Rht) genes, and tight linkage. In the last decade researches provided molecular evidence for this relationship and several QTL responsible for both FHB and PH were identified, including Rht-B1, Rht-D1 and Rht8. Dwarfing genes Rht-B1b and Rht-D1b (formally known as Rht1 and Rht2, respectively) were derived from the Japanese cultivar ‘Norin 10’ and contributed greatly to the Green Revolution. Strong evidences are available for the association between Rht-D1b and Type I FHB susceptibility in European varieties. For example, Rht-D1b increased FHB severity by 52% in a ‘Mercia’ background and 38% in a ‘Maris Huntsman’ background. Lu et al. demonstrated in a mapping population that two major resistant QTL may be required to counteract the negative effect of Rht-D1b. In an association mapping study on European winter wheat materials, Miedaner et al. also reported the significant association of Rht-D1b with increased FHB susceptibility, but to a lesser degree than reported previously; the authors concluded that the negative effects of Rht-D1b in bi-parental populations may have been overestimated. In the case of Rht-B1b, Srinivasachary et al. found that it showed little or no negative impact on Type I FHB resistance under moderate FHB pressure, but exerted negative effects similar to Rht-D1b under severe infection. Miedaner and Voss also reported the different performance of Rht-B1b under different genetic backgrounds. In the three mapping populations tested by Buerstmayr et al., Rht-B1b was associated with increased FHB susceptibility, with phenotypic effects ranging from 3–18%. The negative effects of the two dwarfing genes on field FHB resistance have also been reported in Chinese and US wheat materials. In point inoculated experiments, Rht-B1b exhibited significant effects on Type II resistance, whereas Rht-D1b showed little or no effects on this type of FHB resistance. Many researchers ascribed the association to the pleiotropic effects of dwarfing genes; but Yan et al. claimed that it was the micro-environmental condition around spikes that contributed to the relationship, since the negative effects on Type I resistance disappeared when Rht-B1b and Rht-D1b near- isogenic lines were physically raised to the same heights as their tall counterparts. The importance of anthers in FHB infection has long been observed. Pugh et al. observed that retained anthers were the first tissues to be colonized as a base for further infection. Strange and Smith also found this phenomenon and reported that the presence of anthers favoured greatly the FHB infection, whereas emasculation significantly reduced the disease severity. They ascribed this to the fungal growth stimulants in anthers, of which choline and betaine were the two major components. Based on these findings, Strange et al. suggested the selection of wheat lines with low anther retention (or high AE) to facilitate FHB resistance breeding. Three decades later, Skinnes et al. and Graham and Browne reported the association of FHB with AE in European wheat varieties, where those having high AE tended to had low FHB severity. The positive correlation between AE and FHB resistance has also been reported in Chinese, Japanese and CIMMYT germplasm. QTL mapping studies revealed the underlying mechanisms for this relationship by identifying linked or coincided QTL for the two traits. Like PH, AE was also found to be associated with Type I FHB resistance. Considering the associations of Type I FHB resistance with both PH and AE, it is tempting to investigate the association between the latter two traits, and clues do exist in literature. In the Shanghai-3/Catbird x Naxos population, Rht-B1 explained 10% of the phenotypic variation of AE, and in the Hermann × Skalmeje population, lines with Rht-B1b or Rht-D1b showed reduced AE and double dwarfs (Rht-B1b/Rht-D1b) had a high degree of anther retention (+99%) compared to tall lines with Rht-B1a/Rht-D1a. It was also observed by hybrid wheat breeders that PH and AE are positively correlated. The objectives of the current study were to map QTL for FHB and its related traits and to evaluate the impacts of dwarfing genes Rht-B1b and Rht-D1b on field FHB resistance and AE in two mapping populations. # Materials and Methods ## Plant material Two doubled haploid populations were used in this study. The first one was developed from a cross between ‘TRAP#1/BOW//Taigu derivative’ and ‘Ocoroni F86’ with 135 progenies (referred to as the TO population hereafter), while the second was from ‘Ivan/Soru#2’ and ‘Ocoroni F86’ with 92 progenies (the IO population). Both the two female parents were FHB resistant lines bred at CIMMYT, while ‘Ocoroni F86’ (pedigree JUPATECO-73/(SIB)EMU//(SIB)GRAJO) is a CIMMYT breeding line moderately susceptible to FHB. Both of the two resistant parents carried Rht-B1b/Rht-D1a whereas the susceptible parent had the Rht-B1a/Rht-D1b genotype, resulting in both dwarfing genes segregating in the two populations. ## Field trials and phenotyping The field FHB experiments were conducted at the El Batán experimental station (altitude of 2,240 meters above sea level, coordinate 19.5°N, 98.8°W, with an average annual precipitation of 625 mm) of CIMMYT, Mexico, during the summer season (May to September) when rainfall is concentrated. The two populations were evaluated from 2010 to 2012, sown in 1 m double rows with randomized complete block design with three replications. Each year, a mixture of 5 aggressive F. graminearum isolates were collected, characterized, and used for field inoculation, following the protocols described by He et al.. Spray inoculation was targeted to each line’s anthesis stage with an inoculum of 50,000 spores/ml and was repeated two days later. From anthesis to early dough stages, the nursery was misted from 9am to 8pm with 10 minutes of spraying each hour, to create a humid environment favourable for FHB development. A wheat/maize rotation and conservation agricultural practices were followed in the nursery to enhance natural inoculum. FHB symptoms were evaluated at 25 days post inoculation (dpi) on the 10 spikes that had been tagged at anthesis. Numbers of infected spikes and symptomatic spikelets of each spike were counted for calculating FHB index with the formula: FHB index = Severity x Incidence, where Severity stands for the averaged percentage of diseased spikelets, and Incidence for the percentage of symptomatic spikes. Plots were sickle harvested and threshed with a belt thresher set at low wind speed to retain scabby kernels. Fusarium damaged kernels (FDK) was estimated only in 2012 for the two populations through visually evaluating a random sample in a petri dish, where both scabby and shrivelled kernels were regarded as FDK. DON content was quantified in 2010 and 2012 for the TO population and in 2011 and 2012 for the IO population, based on 2 g flour sampled from 20 g ground grain of each accession, using the Ridascreen Fast DON ELISA kit (RBiopharm GmbH, Darmstadt, Germany) following the manufacturer’s instructions. AE and PH were scored in 2011 and 2012 for the IO population and in 2012 for the TO population. In 2015, the TO population was planted in 40x15 cm hill plots with two replications for an additional evaluation of AE and PH. AE was rated with a linear scale from 0 (no extrusion) to 9 (full extrusion) according to Skinnes et al., and PH was measured before harvest from ground to the average spike tips excluding awns in each plot. Days to heading (DH) was scored for the two populations in all the experiments. ## Statistical analyses The phenotypic data was analysed by the SAS program ver. 9.2. Analysis of variance (ANOVA) was carried out with the PROC GLM module, and Pearson correlation coefficients were calculated using the PROC CORR function. The results of ANOVA were used for calculating the heritability estimates, using the formula $h^{2} = \sigma_{g}^{2}/\left( \sigma_{g}^{2} + \frac{\sigma_{e}^{2}}{r} \right)$ for single years and $h^{2} = \sigma_{g}^{2}/\left( \sigma_{g}^{2} + \frac{\sigma_{gy}^{2}}{y} + \frac{\sigma_{e}^{2}}{ry} \right)$ for multiple years; in which $\sigma_{g}^{2}$ stands for genetic variance, $\sigma_{gy}^{2}$ for genotype-by-year interaction, $\sigma_{e}^{2}$ for error variance, y for the number of years, and r for the number of replications. ## Genotyping The two populations were genotyped with the DArTseq genotyping-by-sequencing (GBS) platform at the Genetic Analysis Service for Agriculture (SAGA) in Guadalajara, Mexico. This genotyping method is a combination of complexity reduction methods developed for array-based DArT and sequencing of resulting representations on next-generation sequencing platforms, for detailed information please check Li et al.. Additionally, two dwarfing genes Rht-B1 and Rht-D1 were also genotyped, using the KASPar technology (KBioscience) based SNP markers developed at CIMMYT. A few SSR markers linked to known FHB resistance QTL were also applied. Markers with missing data points greater than 20% and segregation ratio beyond the range 0.5–2.0 were discarded from further analysis. ## Linkage and QTL analysis Linkage groups (LGs) were constructed using the JoinMap v.4 software, where groupings were based on LOD values from 5 to 10, and ordering within each LG was done with the Maximum Likelihood algorithm. LGs were assigned to chromosomes according to the consensus GBS map by Li et al.. QTL mapping was carried out with MapQTL v6.0, in which interval mapping (IM) was first performed to detect potential QTL for each trait, followed by multiple QTL mapping (MQM) for each QTL, using the closest linked markers to each QTL detected in IM as cofactors. QTL were taken as significant and were reported if they were over the LOD threshold of 3 in at least one environment or over the threshold of 2 in multiple environments. LGs and LOD curves were drawn by the software MapChart ver. 2.3. # Results FHB development of the two populations was satisfactory, ranging from slight infection to around 50% of FHB index in all the three years. The resistant parents ‘TRAP#1/BOW//Taigu derivative’ and ‘Ivan/Soru#2’ showed always significantly higher resistance than the susceptible parent ‘Ocoroni F86’, in terms of all three FHB parameters. In both populations, ‘year’ effect contributed the most variation of FHB and DON, followed by ‘genotype’ and ‘genotype x year’ effects which were also significant except for DON in the TO population. Usually high heritability estimates were obtained for the FHB parameters, but DON in the TO population had a value of merely 0.22. Significantly positive correlations were found among all the FHB traits, although in several cases the r values were low. AE and PH showed wide segregation in both populations. High heritability estimates of 0.79 and 0.87 were obtained for AE in the TO and IO populations, respectively, and in the case of PH the values were even higher. The two traits showed significantly negative correlations with FHB in the TO population (r = -0.73 for PH vs. FHB, and r = -0.65 for AE vs. FHB, p\<0.0001), and the corresponding correlations were also significant in the IO population but with lower r values (r = -0.63 for PH vs. FHB, and r = -0.48 for AE vs. FHB, p\<0.0001). In the TO population, 1,858 GBSs together with seven SSRs and the two dwarfing genes were used for LG construction. Thirty three LGs were generated, covering 4,053cM with an average density of 2.2 cM/marker. All but 1D chromosome were represented in this map and three LGs were not assigned to a chromosome due to a lack of anchored markers. Regarding the IO population, 1,986 GBSs, Rht-B1, Rht-D1 and four SSRs were used for linkage mapping and 35 LGs were obtained. Total length of the LGs was 4,430cM with a very similar density as that of the TO population. In this case, only chromosome 6D was not represented and six LGs were not assigned to chromosomes. Three QTL with major effects were identified in both populations, i.e. Rht-B1 on 4BS, Rht-D1 on 4DS and a QTL on 5AL. The latter was most likely at Vrn-A1, due to its strong effects on heading time, explaining 48.2% of the DH variation in the TO population and 14.9% in the IO population. The expression of the three QTL was more stable in the TO population, being associated with FHB parameters in most environments, accounting mostly 10–20% of phenotypic variation. Comparably, the magnitude of their phenotypic effects was similar in the IO population, but their expression was not detected in certain environments, e.g. the 5AL QTL was not identified in 2012 and the Rht-B1 QTL was significant mainly in 2012. The two dwarfing genes showed similar negative effects on FHB resistance in the two populations, although the phenotypic variations explained by Rht-B1 were often a bit higher than those by Rht-D1. A QTL on 2AL was also shared by the two populations based on common markers, but it was only a minor QTL accounting for phenotypic variations less than 10%. Additional minor QTL were found in the IO population, located on 1BL, 3BL (LG 3B_2), 3BS (LG 3B_3) and 5BL. It could be observed that for DON in the TO population, QTL on 2AL and 5AL in 2010 were significant, but the ones at Rht-B1 and Rht-D1 were identified in 2012, explaining the non-significant ‘genotype’ effect and low heritability estimate for this trait. Several QTL for AE were localized and three were shared by the two populations, viz. Rht-B1, Rht-D1 and a QTL on 2AL, all associated with FHB resistance. Additional QTL were found on 2BL in the TO population and on 2DS and 3BL in the IO population. The two dwarfing genes collectively explained around 20% of AE reduction in both populations and Rht-B1b was always more strongly associated with reduced AE than Rht-D1b. As for PH, Rht-B1 and Rht-D1 collectively accounted for about 60% variation in the two populations, while additional QTL were found on 5AL (Vrn-A1) and 7B in the TO population and on 5BS in the IO population. # Discussion FHB index after field spray inoculation was generally considered as for a combination of Type I and Type II resistance; but in our study it appeared that mainly the former took place considering the significantly high correlation of FHB with PH and AE that did not happen in point inoculated experiments for Type II resistance. Therefore, we considered that the results obtained in this study were based mainly on Type I resistance. Dwarfing genes Rht-B1 and Rht-D1 and the vernalisation gene Vrn-A1 were segregating in both populations used in this study, resulting in that most of the phenotypic variation for FHB parameters was explained by these three loci, whereas other QTL only explained a small fraction of the variation. The latter category comprised QTL with phenotypic effects below 10%, which were likely known QTL based on their locations. The association between the two dwarfing genes and FHB susceptibility has been reported in many studies, and three possible mechanisms including disease escape, pleiotropy and tight linkage have been proposed; but a conclusion has not been drawn as to which mechanism was actually taking place. Intuitionally, this could be ascribed to PH per se or escape since tall plants were farther from soil surface where the inoculum was present (in the case of natural infection or spawn inoculation where FHB infected grain kernels were scattered in the field as inoculum) and ventilation was reduced that lead to high humidity favourable to FHB development. This mechanism must have contributed to the association in this study since the correlation remained significant in the sub- populations with homozygous Rht-B1 and Rht-D1 alleles (data not shown); despite the utilization of spray inoculation in this study, huge quantity of Fusarium inoculum was present on soil surface due to the adoption of wheat/maize rotation and conservation agricultural practices, supporting the escape mechanism. With the accumulation of molecular evidences in the last decade, more researchers took pleiotropy as the main mechanism for this association. Being DELLA protein producers, Rht-B1b and Rht-D1b have shown association with reduced resistance to biotrophic diseases including Type I FHB resistance (although FHB is regarded as a necrotrophic disease, it behaves more like a biotrophic disease at the early stages when Type I resistance takes place) but increased resistance to necrotrophic diseases like Type II FHB resistance. Another evidence for the genetic effects of dwarfing genes instead of disease escape was that in sub-populations with homozygous Rht alleles, the correlation between PH and FHB disappeared or was significantly reduced, which was obviously not the case of the current study. However, this does not necessarily mean that the pleiotropic effects had no impacts on FHB in our study; it could function through controlling AE, which will be further explained below. Due to the limitation of map resolution, currently it is very difficult to separate pleiotropy and tight linkage; nevertheless clues supporting the latter have been reported. In the Soissons x Orvantis mapping population, Srinivasachary et al. found that the peaks of FHB QTL were constantly located in a short distance away from the Rht-D1 locus. Similarly in our previous research, a QTL for FHB in close linkage with Rht-D1 appeared when PH was used as covariate. So it appears that all the three mechanisms exist, but they are not necessarily simultaneously present in a single wheat line and their different combinations are expected. The importance of AE in FHB resistance has long been recognized, but genetic studies on AE were performed only in the last few years. In all these studies, the accumulated phenotypic variation explained by identified QTL for AE rarely exceeded 50% (usually around 30%), in accordance with the current study, demonstrating a typical quantitative inheritance of AE. In the aforementioned four studies, totally 18 AE QTL have been identified, but only the QTL on 7AL found by Skinnes et al. and Lu et al. may be the same, and the one found on 4AL in He et al. could be the same as reported by Buerstmayr and Buerstmayr. In the current study, six more AE QTL were found, and not unexpectedly only the one on 2DS might be the same as found in our previous study, whereas others were all from new chromosome regions. Similar to previous results, four out of the six AE QTL were associated with FHB resistance (Tables), supporting the phenotypic association of the two traits. The two dwarfing genes showed consistent effects on reducing AE in both populations across environments, collectively contributing about 20% of AE variation. The association may have its physiological basis. In Arabidopsis, the elongation of anther filament is stimulated by GA and repressed by DELLA proteins which are orthologous to wheat Rht-1 gene products. Thus it is reasonable to speculate that the GA insensitive mutants Rht-B1b and Rht-D1b in wheat have similar function in repressing anther elongation through over expression of DELLA proteins, resulting in the phenotype of anther retention or low AE. This finding partly explained the pleiotropic effects of Rht-B1b and Rht-D1b on Type I FHB susceptibility, i.e. the two dwarfing genes lead to low AE, which in turn caused increased Type I FHB susceptibility. The results have also implications for hybrid wheat breeding, in which the selection of male parent, the pollen provider, is very important. A good male parent is expected to have high AE and high pollen production, and these two traits were reported to be significantly positively associated with r = 0.82 by Joppa et al. and was later validated by Johnson and Patterson and Atashi-Rang and Lucken. Thus the utilisation of wild type Rht alleles Rht-B1a and Rht-D1a will improve both traits. Still more, the tall stature of such lines is favourable for efficient pollination, since it was suggested that the male parent be taller than the female parent in hybrid seed production. # Supporting Information Financial support by the CGIAR Research Program on Wheat and ‘Seeds of Discovery’-Sustainable Modernization of Traditional Agriculture project (MasAgro) is gratefully acknowledged. The first author was financially supported by a CGIAR scholarship, provided by the Research Council of Norway, through NFR project 208340/H30—Breeding for Fusarium resistance in wheat. The authors are grateful to Prof. Åsmund Bjørnstad, Norwegian University of Life Sciences, for his kind instruction on AE phenotyping, and to Francisco Lopez, Javier Segura and Nerida Lozano for technical assistance. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: PS ED XH. Formal analysis: XH. Methodology: XH PS ED SD SS. Resources: ED PS ML SD SS. Writing – review & editing: XH ML PS.
# Introduction Today, agriculture plays an important role in many developing countries, especially in Southeast Asia. It is estimated that these enterprises account for a considerable proportion of a country’s gross domestic product and a large proportion of its employed population. Vietnam and Thailand are the world’s largest agricultural exporters. Rice is one of the most famous agricultural products. Along with coffee, cocoa, corn, fruits, vegetables and other commodities, it contributes a large portion of the local GDP. Palm oil, for example, is an important agricultural product for Indonesia and Malaysia, both members of the Association of Southeast Asian Nations. Insects are the biggest hazard in agricultural production. Crops such as rice and wheat are vulnerable to pests, which bring huge economic losses to farmers. Therefore, in order to ensure that ASEAN countries can obtain more food from agriculture every year, the identification of crop diseases and pests in the process of agricultural production plays an important role in the early prevention and control of crops and damage to crops. Manually identifying pests in large-scale farming is a time-consuming and costly undertaking. Now, with the spread of high-quality image capture devices and advances in pattern recognition by machine learning, image-based automatic pest recognition systems promise to reduce human costs and perform the task more efficiently. When classifying insects, it is difficult to extract effective features. There are many kinds and forms of insects, which bring great difficulty to their classification and recognition. In recent years, artificial features (GIST, HOG, SIFT, SURF, etc.) have been thoroughly studied. However, the artificial characteristics lack the representation of large-scale morphological changes of multiple objects. At the same time, the deep learning based agroforestry data processing technology also provides a new way for researchers with strong generalization performance, and effectively avoids over- reliance on artificial features. In recent years, convolutional neural network, as an effective classification method, has been regarded as an effective classification method because it can automatically extract useful features from images without human guidance. Moreover, it performs well in recognizing complex features and different objects, and has high generalization performance. In addition to the new CNN architecture, Such as Alexnet, DenseNets, EfficientNets, GoogleNets, MobileNets, NasNets, ResidualNet-works(ResNets), SqueezeNet, eometric Group Networks(VGG), etc., are all hot spots in deep learning research. They have different advantages, for example Alexnet can take image classification tasks to new levels; EfficientNets enable the integration of multi-tier networks. GoogleNets achieves great success in the image classification task; MobileNets can recognize small objects; NasNets can realize image classification and image retrieval. ResidualNet-works(ResNets) enables end-to-end prediction models; SqueezeNet can implement multitasking learning; Visual Geometric Group Networks(VGG) enable object recognition in a number of different areas. Their appearance not only brings new changes to the field of deep learning, but also provides new ideas for solving more problems. The proposed fine-grained mode of simulation attention mechanism enables efficient reuse to provide a new entry point for solving a certain problem and greatly reduces the requirement for massive computing resources. This project is based on the existing PlantVillage database and adopts the network framework of Alexnet and GoogleNets to improve the accuracy and efficiency of vegetation pest identification methods. In addition, the technique can better handle higher- dimensional data and can more efficiently leverage correlations between data. In recent years, deep neural networks have played an increasingly important role in the classification of vegetation pests in some crops, For example, apple, cassava, corn, cucumber, grape, corn, mango, rice, millet, guava and so on, which benefit from their high recognition accuracy and strong robustness. Disease identification and classification can be carried out effectively. However, current methods are mostly focused on depth and complexity, rather than learning from traditional machine learning methods to improve detection accuracy. Therefore, it is of great significance to further study the application of deep neural network in crop vegetation insect classification.Cnn- based features are widely used to classify insects in the ImageNet Large Visual Recognition Competition (ILSVRC). WU and colleagues have demonstrated that CNN- based features are more efficient than hand-drawn ones in this work. Secondly, the developmental process of metamorphosis includes egg, larva, pupa, adult and so on. Also, there are strong similarities between the two species. For each category, efficient algorithms need to capture features that express a large number of morphologic variations. So far, there has been no research on insect identification. In order to solve this problem in the classification of bird species or insect models, a fine-grained image classification algorithm is adopted in this paper. Fine-grained image classification is to extract distinguishable features from the information area of objects and classify them in the form of vectors. Considering the influence of geographical environment and changeable weather on image information, the existing mainstream image segmentation methods are not applicable. Considering the main moving position of the target and the safety protection area of the track, combined with the actual background, this paper proposes an optimization algorithm based on the fusion of recursive attention convolutional neural network and adaptive particle swarm optimization. For the information loss caused by traditional convolution in the process of image classification and region sampling, attention learning is used to recursively operate under multi-scale, and the adaptive particle swarm optimization algorithm is used to synthesize the feature information of different scales, and continuously recursively generate regional attention from rough to fine, so as to realize the localization and detection optimization of the target region. Specifically, the attention module is used to plan the target area of the image, considering the information loss generated at each scale in the process, and introducing it into the cyclic convolution, so that it can more closely aggregate the convolution information before and after, improve the target segmentation performance of the network and reduce the amount of calculation. Attention every scales module division are on a level area location and the influence of information loss, to ensure that the final output is approaching the target location information of the image. In the whole cyclic convolution process, the above output results are processed through five pooling layers to obtain a number of attention maps that feedback the clustering correlation from different scales. Then the adaptive particle swarm optimization algorithm is introduced for weight ratio, and finally the Softmax function is used to identify and classify vegetation diseases and pests. In this study, the superiority of the proposed optimization method in the detection of vegetation pests and diseases is verified. Based on the traditional recursive attention convolutional neural network, an accurate identification method is proposed to gather the convolution information before and after, so as to perfectly segment the target from the background, and a high performance test is achieved in the dataset. In this paper, the main contribution of the summary is as follows: 1. First, we apply the CNN model with attention mechanism to create a feature extraction program focusing on insects. Since images of insects on crops often contain complex backgrounds of leaves, dust, and branches, the mechanism of attention is crucial. 2. Secondly, this project intends to use multi-scale convolutional neural network to capture insects of different sizes. 3. Third, we propose a fine-grained image classification algorithm based on multi-scale learning to solve the problem of high inter-class similarity. 4. Finally, we will adopt the "soft voting" method to integrate the above models to further improve the performance of the system. The rest of the article is arranged this way: In the related work section, we will explore the literature that is currently available to substantiate our motives. The experimental process section describes our proposed method in detail, and the results section gives some conclusions which show the correctness of the method we have introduced. In the conclusions section, we summarize the shortcomings of this paper and make suggestions for future researchers. # Related work In recent years, in agricultural production, a large number of visual research work is through a certain degree of training to solve the classification problem. In the field of insect classification, Cheng et al. has proposed a new method of crop feature extraction based on deep residual network. In the complex scene, the recognition accuracy of 10 kinds of insects is 98.67%, which is better than the conventional deep learning algorithm. Liu et al. proposed a new method, that is, using the obvious feature extraction algorithm to determine the object of the pest, and using the deep convolutional neural network to classify the pest. The test results show that this technique can be performed with an average accuracy of 0.951 for the calculation (" mAP "). Wang et al. adopted a deep convolutional neural network (Dependency Networks) based on crop pest images. They compared two alternative deep neural networks,Lenet-5,Alexnet,convolutional cores, and the number of alternative cores. The difference of test results has great influence on classification accuracy. studied the classification of crop pests and analyzed their performance. At the same time, several pre-learning deep learning systems (ResNet, GooleNet, VGG-Net) were also studied. Experiments show that the proposed method is much improved compared with other pre-training methods. In, the author uses DC-GAN to generate enhanced images for training. Another method described in is feature extraction of CNN based on prototype network, and Euclidean operation is carried out on it. This architecture uses triples to set up loss functions. Another recent study uses FSL to model small sample data. In, the author trained CNN to extract general plant leaf features, and combined Siamese network with triplet loss for classification calculation. In, two authors tested their data set using PlantVillage. Currently, there is only one ideal image of vegetation with distinct background, single leaf, unshaded and continuous light in the existing FSL data set. This method includes two core contents: 1) Embedding based on generalized samples; 2) Calculation based on the distance between the sample and the query sample. In many work-learning postures embedded, simple classifiers such as nearest neighbor method and linear classifier are used for further classification. In, the author adopts the nearest neighbor method. MetaoptNet adopted a Linear Sequencer SiamianI network uses a shared feature extractor and uses the shortest distance between the queried sample and the truth value for classification. Wu et al. proposed a new insect-based large-scale reference dataset (IP102). The database contains more than 75,000 categories, including 19,000 tagged target detection images. In IP102, it was tested on the basis of manual (GIST, SIFT, SURF) and CNN (ResNet, GooleNet, VGG-Net) based functions. Ren proposed a new method of interlayer features of residual data based on residual information, namely, to establish a residual network based on the interlayer features of residual data. Tests on the previous version of IP102 showed improved performance. Liu et al. also proposed a new block-based multi-branch fusion residual network (Dmf-ResNet) to learn multi-scale representation. The basic residue is bonded to the bottle and the residue is pressed into the residue so that the residue forms a residue of multiple branches. The output of these branches can be linked to new modules to achieve adaptive recalibration of the response so that it can be simulated. A multi-branch fusion residual net method based on deep learning was proposed and applied to classify pests. However, there are great similarities and differences between multiple species. Existing research algorithms are aimed at each category and need to capture features that can express a large number of morphologies. So far, there has been no research on insect identification. In order to solve this problem in the classification of bird species or insect models, a fine-grained image classification algorithm is adopted in this paper. Fine-grained image classification is to extract distinguishable features from the information area of objects and classify them in the form of vectors. # Experimental process ## A. Circular attention convolutional network The basic convolutional neural network performs convolution processing on the target image with the help of convolution kernels of different scales, and extracts various feature information including edges and textures, which provides help for subsequent information analysis and target recognition. Its basic structures include input layer, volume at the grass-roots level, pooling, the whole connection layer and output layer. In the convolution operation, the appropriate convolution kernel specification and the number of steps are selected to weighted sum the pixels, and the corresponding feature information is obtained. Then after pooling layer zone, processing, data dimension reduction, prevent over fitting phenomenon. After repeated volume base layer and pooling layer, image information was continuously simplified, and finally entered the fully connected layer to determine the results. On this basis, the introduction of attention mechanism formation of recursive convolution neural network. The mechanism of attention stems from the study of human vision. In cognitive science, due to the bottleneck of information processing, humans will selectively focus on some information while ignoring other information. The mechanism is often referred to as attention mechanism. In neural networks, the attention module is usually an additional neural network that can hard select some parts of the input, or assign different weights to different parts of the input. Through the repeated training of the neural network, the weight features are strengthened, so that the computer can identify the area that needs to be focused on in each image, so as to form attention. The proposed structure mainly includes three levels. The network structure of the three levels is the same but the parameter information is not related to each other. Each of these levels contain classification module (VGG) and regional sampling module (APN). Classification module of input image feature extraction and classification of regional sampling module based on attention to extract the characteristics of information area, and as the next level of input; Reciprocating operation under different level results output. When training data, it is a weakly supervised behavior that relies on information labels for classification and judgment, which consumes a lot of time and energy. The framework of attention region determination in the region sampling module is single, which is not suitable for processing all kinds of multi-shape feature information. It will cause the loss of image information during classification and region sampling, which will affect the classification and sampling of the next level. Based on this, this paper proposes an optimization algorithm based on the fusion of recurrent attention convolutional neural network and adaptive particle swarm optimization. In, a unified preprocessing operation is performed on the input image, and all the pictures of the training set and the test set are resized to the scale of 224×224. In order to make the classification information more accurate and reduce the information loss, the optimized model replaced the original classification module with the classification structure of multivariate comprehensive evaluation on the basis of maintaining the original three-layer network structure and regional sampling module. Take three-scale network structure as an example: the original input image (SCALE1) is used as the input to participate in the convolution operation of the next layer network structure (SCALE2) through one-scale image classification and feature information extraction. Reciprocating cycle three times, finally get output the results of three different scales. The specific results are shown in : In, the input image is continuously classified and sampled, and the fine-grained information of the image is constantly refined. Where P<sub>t</sub> is the output result after the pooling layer processing of the convolutional model. N is the sampling of attention area according to the extracted feature information, which is used as the input of classification and sampling at the next level. Y<sup>(m)</sup> is the classification label of the output result at the MTH layer; $P_{t}^{(m)}$ is the probability that the classification label of the m-th layer is accurate; L<sub>cls</sub> is the loss of the classification module; L<sub>rank</sub> is the loss of the region sampling module. ## B. Target segmentation The core operation of machine vision is to realize the directional segmentation of the target image, and the fully convolutional neural network is the first research paper that uses convolution calculation and multi-level image segmentation. Its completely made up of multiple volumes base, there is no connection layer to implement the mapping of the input image segmentation for different specifications. Moreover, researchers use a leap-forward structure when constructing the network, so that the network can obtain high-level and low-level image features to enrich the feature information of the target image, and many famous network structures are involved. However, full convolutional neural network also has certain defects. It cannot closely connect feature regions at different scales and limit its performance. On this basis, the relevant achievements of subsequent development are constantly exploring richer information extraction and aggregation. The size and specification of the convolution kernel determine the range of the received information domain. Richer information requires a larger acceptance range, but it also causes an increase in computational cost and has a negative impact on the timeliness of the network structure. The convolution kernel specification used in global cyclic convolution effectively saves the computational cost and the number of parameters. On the premise of not increasing the amount of calculation and the number of parameters, the extended convolution kernel can receive a wider range of information domain. Extended convolution introduces the expansion coefficient into the initial convolution kernel to determine the distance parameter between the weights in the convolution kernel. Similar methods are used in subsequent related works. The traditional encoder-decoder structure is a common method in image segmentation. The input image is converted into a probability map of pixel categories by using the encoder-decoder structure and the connection layer, which can realize the information fusion of the front and back background at different scales. Learning deconvolution network is the first image cutting using coding-decoder and achieves high performance test without using external data. SegNet is also a network structure using encoder-decoder, and it proposes an innovative method of upsampling to achieve a network structure with a smaller number of parameters under the same performance. U-net is widely used in the field of medicine, and many articles in this field are further based on it. In addition, the multi-level pyramid network structure is common used methods, characteristics of the pyramid network structure in the process of target detection has achieved good performance. In image segmentation, the pyramid scene parsing network extracts feature information through ResNet, and then inputs the feature information into the pyramid pooling module to process feature maps of different scales and perform fusion sampling operations. Traditional RNN is often used to solve natural language processing methods, but there are also papers that use RNN to realize the segmentation operation of image objects. With the rapid development of GAN, relevant researchers have realized various machine vision tasks, including image segmentation. Other Angle image segmentation methods include the DecoupleSegNets decouple the information feature and divide it into the main part and the edge part, and realize the optimization of the main part and the edge feature by the method of display modeling. SNE-RoadSeg use a surface normals estimator, from the depth of dense information to calculate the surface normals feature space partition. The classification module of the traditional model usually uses the output result P<sub>5</sub> of the fifth pooling layer to calculate the loss function and determine the classification label. However, the image feature information contained in P<sub>5</sub> is lost due to the change of the size of the sampling module area after multi-layer convolution. Therefore, the comprehensive multivariate algorithm is introduced on the basis of the VGG module, so that the network structure can select the appropriate convolution kernel for classification, and reduce the loss of feature information caused by the classification and sampling process. The specific results are shown in : ## C. Attention model Recently, researchers have begun to conduct in-depth research on the attention mechanism, and the proposal of attention module has been widely used in various fields of machine vision. Common attention modules include channel attention and spatial attention, in which channel attention enhances the network structure by strengthening the feature relation between different channels, and spatial attention enhances the network structure by extracting the feature information from a single pixel to a local area. GANet constructed a spatial gating attention module to realize the adaptive evaluation of multi-scale feature interaction mechanism. FocusNet the encoding—decoder module, the attention of the parallel branch association and generate a gradient flow segmentation mask to achieve optimization. The multi-modal fusion network uses multi-channel independent coding to extract feature information respectively, and designs an attention mechanism fusion module to fuse these information. GSANet using selective attention extraction under different spatial location and different levels of information of pixels. Researchers proposed a bilateral attention module, in which the foreground and background are respectively focused in different ways to obtain information. There is also a parallel reverse attention module to solve the polyp segmentation operation of colonoscopy images, and the reverse attention module is used to realize the division of target area and boundary area. Staff based on attention mechanism itself, by calculating the objective function space build focus area of each feature information. When determining the area to be sampled on the input image, the upper left point of the image is first assumed to be the origin of the coordinate, and the right and down are set as the positive direction of the X axis and Y axis, respectively. The upper left of the area to be sampled is t<sub>tl</sub> and the lower right is t<sub>br</sub>, and the specific Formulas and as follows: $$t_{x({tl})} = t_{x} - t_{l},\mspace{22mu} t_{y({tl})} = t_{y} - t_{l}$$ $$t_{x({br})} = t_{x} + t_{l},\mspace{22mu} t_{y({br})} = t_{y} + t_{l}$$ After determining the area to be sampled, the element multiplication method is used to cut and enlarge the operation, where ⊙ represents the element multiplication, X<sup>att</sup> is the area to be sampled, and M represents attention recurrence and the specific Formulas and as follows: $$X^{att} = X \odot M\left( {t_{x},t_{y},t_{l}} \right)$$ $$M = \left\lbrack {H\left( {x - t_{x({tl})}} \right) - H\left( {x - t_{x({br})}} \right)} \right\rbrack \bullet \mspace{9mu}\left\lbrack {H\left( {y - t_{y({tl})}} \right) - H\left( {y - t_{y({br})}} \right)} \right\rbrack$$ In order to reduce the information loss generated in the classification module and sampling module, the derivative mapping is used to reflect the direction of attention recurrence, and the darker the color is, the more consistent it is with the direction of attention recurrence. t<sub>x</sub> for example, which M′(t<sub>x</sub>) is the focus of recursive t<sub>x</sub> derivative, piecewise function is as follows: $$M^{\prime}\ \left( t_{x} \right) = \left\{ \begin{array}{l} \left. < 0\mspace{22mu} x\rightarrow t_{x({tl})} \right. \\ \left. > 0\mspace{22mu} x\rightarrow t_{x({br})} \right. \\ {= 0\mspace{22mu} otherwise} \\ \end{array} \right.$$ Considering that the mapping result of the attention derivative is a negative transformation from the image boundary to the interior, the $L_{rank}^{\prime}\left( t_{x} \right)$ result is positive, so t<sub>x</sub> becomes smaller in the recursion of the lower structure, which represents the direction of human attention. The specific results are shown in : Based on the attention recursion network model, the label classification within the level and the region sampling between the levels will cause information loss, in which the loss function is defined as follows: $$L = {\sum_{m = 1}^{3}{\left\{ {L_{cls}\left( {Y^{(m)},Y^{}} \right)} \right\} +}}{\sum_{m = 1}^{2}\left\{ {L_{rank}\left( {P_{t}^{(m)},P_{t}^{({m + 1})}} \right)} \right\}}$$ Particle swarm optimization(PSO) is a stochastic optimization algorithm, which simulates the behavior of birds flying and foraging, and makes the group achieve the optimal purpose through the collaboration between the flocks of birds. It is suitable for solving complex nonlinear optimization problems. The particle swarm optimization algorithm solves the optimization problem, and a single particle represents a feasible solution. The fitness of the particle is calculated according to the optimization objective function, and all individuals in the population are constantly moving to seek the optimal solution according to certain rules. The particle in the particle swarm moves according to the following rules and: $$v_{i}^{(t)} = \omega(t)v_{i}^{({t - 1})} + Mr_{1}\left\lbrack {x_{i}^{best} - x_{i}^{({t - 1})}} \right\rbrack + Lr_{2}\left\lbrack {x^{best} - x_{i}^{({t - 1})}} \right\rbrack$$ $$x_{i}^{(t)} = x_{i}^{({t - 1})} + v_{i}^{t}$$ Where, t is the time step of particle movement $\left( {t > 0} \right);\ x_{i}^{(t)}$ is the position vector of the particle at time step t, $x_{i}^{best}$ is the historical optimal position vector of the particle in the moving process, $x_{i}^{best}$ is the historical optimal position vector of the whole population, $v_{i}^{(t)}$ is the velocity vector of the particle at time step t. M and L correspond to the attention recurrence function and information loss function, respectively. r<sub>1</sub> and r<sub>2</sub> are random numbers with interval \[0.1\]; ω(t) for adaptive inertia weight coefficient, according to the Formula is as follows: $$\omega(t) = \left( {\omega_{max} - \omega_{min}} \right)P_{s}(t) + \omega_{min}$$ Where ω<sub>max</sub> and ω<sub>min</sub> are the maximum and minimum values of the inertia weight, respectively, and generally the initial values are 1.0 and 0.3. P<sub>s</sub>(t) is the proportion of particles that move to a better position, and f<sub>i</sub>(x) is the fitness value of the particle at time step t, which is calculated according to the following functional Formulas and: $$f_{i}(x) = \frac{1}{n}{\sum_{i = 1}^{n}{P_{s}(t)}}$$ $$f(I) = \omega_{1}^{c}f_{1} + \omega_{2}^{c}f_{2} + \cdots\omega_{5}^{c}f_{5}$$ In the convolution model, there are a total of five pooling layers to obtain five classification prediction results P<sub>t</sub>, in which the adaptive weights are calculated according to the recursive function of attention and the information loss function of each level, and then the target region is divided according to the five weight ratios corresponding to their respective self- fitness. Finally, the classification results of the input image are predicted by multi-level VGG-PSO labels. The classification label Y<sup>(m)</sup> of each layer is put into the fully connected layer and Softmax is used to obtain the final classification result. The specific process of weighted heat map is shown in : ## D. Data preprocessing and experimental equipment The experiment was carried out in Windows10 environment based on Tensorflow deep learning framework. Computer configuration: Inter Core i7, 1080Ti graphics card. In learning, we use a classification based on cross entropy. Using Adam algorithm, the initial learning rate can be obtained as 5x10−5, with 16 training batches and 40–45 training cycles. When the damping is 0.96, the learning rate decreases exponentially. If, after 20 cycles, it is confirmed that the performance of the set has not improved, the training phase is over. To solve the overfitting problem, we use dropout technology with a dropate value of 0.45. Data sets including experiments are divided into training sets, verification sets and test sets in a ratio of 6:3:1. The details are shown in below: The photo gallery contains 15,905 images in 10 categories. Among them, there were 1,888 palatine bugs, 973 palatine bugs, 1,759 wheat aphids, 715 noctuidae, 1,559 Oriental mole mantis, 1,269 rice planthopper, 363 rice locust bugs, 4,526 diasteridae bugs, 1,528 beet leucophora borer, and 1,325 palatine bugs. The image of the original data set is 224x224. Details are shown in below: In addition, we will add samples for each type and improve with a set of enhancement features including: randomly rotating different types of samples, mirroring (horizontal, vertical), varying brightness, etc. Details are shown in below: By optimizing multi-level attention region extraction, the extracted region can not only contain the overall target structure information, but also save the local location information. It can also be clearly felt that when the input image is extracted at the second and third levels, the difference in the information it contains is more obvious, and the extracted attention area is similar to that of humans.Sense of direction is consistent, which help the exact granularity classification. As shown in below: When considering the balance of information loss and clipping elements among multiple convolution layers, an adaptive optimization algorithm is needed to select the optimal solution for calculation. Genetic algorithm (GA), ant colony algorithm (ACO), particle swarm optimization (PSO), artificial bee colony algorithm (ABC), cuckoo search (CS) and so on are all based on population optimization algorithms. The probabilistic search algorithm of optimization step is implemented by iterative method. The experimental heat map obtained by the above five optimization algorithms combined with the proposed algorithm is shown in below: As can be seen from, the temperature curve of insulator identified by particle swarm optimization algorithm coincides with the actual target detection point, so particle swarm optimization algorithm is adopted in this paper. Particle swarm optimization (PSO) is a random optimization method that simulates the flight and foraging of birds. This method is suitable for solving complex nonlinear optimization problems. ## E. Evaluation index This metric consists of macromean precision (MPre), macromean recall (MRec), macromean F1-composition score (MF1), accuracy (Acc), and geometric mean (GM).To be equally important, we calculated the recall rates for each category and then averaged them to obtain the MRec shown below and: $${Rec}_{c} = \frac{{TP}_{c}}{{TP}_{c} + {FN}_{c}}$$ $${MRec}_{c} = \frac{{\sum_{c = 1}^{C}{Rec}}_{c}}{C}$$ C represents the number of a category. TP<sub>c</sub> and FN<sub>c</sub> denote pseudo-negative and pseudo-negative values.Also, we will carry out calculations. Pre<sub>c</sub> and MPre<sub>c</sub> look like this and: $${Pre}_{c} = \frac{{TP}_{c}}{{TP}_{c} + {FP}_{c}}$$ $${MPre}_{c} = \frac{{\sum_{c = 1}^{C}{Pre}}_{c}}{C}$$ Class C of FPc is a false positive. MF1 is the average of MPre and MRec, such as follow: $$MF1 = 2\frac{MPre \bullet MRec}{MPre + MRec}$$ Acc is computed by the true positive value among all classes as follow : $$Acc = \frac{TP}{N}$$ Where N is the number of samples. GM is calculated according to the sensitivity (S<sub>c</sub>). S<sub>c</sub> and GM are as follows and : $$S_{c} = \frac{{TP}_{c}}{{TP}_{c} + {FN}_{c}}$$ $$GM = {\prod_{c = 1}^{C}\sqrt[c]{S_{c}}}$$ If only one S<sub>c</sub> is 0, then GM is 0.To avoid this problem, we replaced the S<sub>c</sub> with 0.001. # Results It can be seen from the above figure that with the multi-level recurrent convolution operation, the deeper the network structure is, the stronger the ability to extract the feature model is. The changes of the recursive attention function and information loss function under different levels are shown in : In order to make the experimental effect more obvious, considering the results of target detection and recognition between different models, the classical model is introduced to compare the experimental results between different algorithms and verify the performance of the algorithm optimization algorithm. The specific results are shown in and : And the detection results between different scales of the optimal model, the three layers of scales are tested by permutation and combination, and the relevant experimental results are shown in : As can be seen from the above table, the target detection accuracy under different scales is different. The accuracy of single-scale first layer, second layer and third layer is 80.99%,83.02% and 80.62% respectively, which is 6.02%,3.86% and 6.63% respectively compared with the best 86.35% obtained by the complete three-layer scale (scale 1+2+3) connected network model. Although there is a 0.14% difference between MSSN(scale 2+3) and MSSN(scale 1+2+3) in the detection accuracy, MSSN(scale 1+2+3) has a macro mean accuracy (MPre) of 91.02% in addition to ACC index. The macro average recall rate (MRec) was 85.56%, the macro average F1-score (MF1) was 89.35% and the geometric average (GM) was 92.23%. In terms of the overall performance evaluation index, MSSN(scale 1+2+3) had a better effect. At the same time, the performance of different algorithms is compared in the Pest dataset, as shown in : It can be seen from that when other benchmark networks are used for comparison experiments, the indexes of 4.14FLOP/s, mAP72.3% and 26.5fps are the best, but it also results in the huge volume of the overall training network of 32.8M and the multifold increase of training time to 19.2hours.The results of agricultural pest detection on the data set achieved by the innovative algorithm proposed in this paper are shown in below: # Conclusions This paper proposes a recursive convolution neural network and adaptive particle swarm optimization algorithm of fusion, in view of the traditional convolution in the image classification and regional information loss during the process of sampling, through the study of attention under the multiscale recursive operations, with the help of adaptive particle swarm optimization algorithm integrated the characteristic information of different scales, from coarse to fine is recursive generation regional focus, Optimized by testing the positioning of the target area. The experimental results showed that the accuracy rate of the The test set reached 86.35%, which improved the accuracy of the detection of vegetation disease and insect pests entering the limited area. However, compared with other basic models, this method not only improves the recognition accuracy, but also reduces the memory capacity and computing speed of the model. Considering the balance between recognition accuracy and system speed is one of the future research directions of the author. In the future, this paper will further explore the diagnosis of plant diseases under the influence of various environmental factors such as image pollution, information defects and blurring, so as to ensure the safety of agricultural production. # Supplement In the supplementary material, in addition to the pest identification study involved in in the main text, the detection of agricultural pests can still be achieved in the case of different sizes, body types and appearances of pests, demonstrating the generalization ability of the test method, as shown in detail in : [^1]: The authors have declared that no competing interests exist.
# Introduction The discovery and design of effective drugs for infectious and chronic biological diseases, like cancer and cerebral malaria, require a deep understanding of behaviorial and structural characteristics of underlying biological entities (e.g., cells, molecules and enzymes). Traditional approaches, which rely on verbal and personal intuitions without concrete logical explanations of biological phenomena, often fail to provide a complete understanding of the behavior of such diseases, mainly due to the complex interactions of molecules connected through a chain of reactions. Systems biology overcomes these limitations by integrating mathematical modeling and high-speed computing machines in the understanding of biological processes and thus provides the ability to predict the effect of potential drugs for the treatment of chronic diseases. System biology is widely used to model the biological processes as pathways or networks. Some of the examples are signaling pathways and protein-protein interaction networks. These biological networks such as gene regulatory networks (GRNs) or biological regulatory networks (BRNs), are analysed using the principles of molecular biology. This analysis, in turn, plays an important role for the investigation of the treatment of various human infectious diseases as well as future drug design targets. For example, the BRNs analysis has been recently used for the prediction of treatment decisions for sepsis patients. Traditionally, biologists analyze biological organisms (or different diseases) using wet-lab experiments. These experiments cannot provide reliable analysis due to their inability to accurately characterize the complex biological processes in an experimental setting. Moreover, the experiments take a long execution time and often require an expensive experimental setup. One of the other techniques used for the deduction of molecular reactions is the paper-and- pencil proof method (e.g. Boolean modeling or kinetic logic). But the manual proofs in paper-and-pencil proof methods, become quite tedious for large systems, where several hundred proof steps are required in order to calculate the unknown parameters, thus prone to human error. Other alternatives for analyzing system biology problems include computer-based techniques (e.g. Petri nets and model checking). Petri net is a graph based technique for analyzing system properties. In model checking, a system is modeled in the form of state- space or automata and the intended properties of the system are verified in a model checker by a rigorous state exploration of the system model. Theorem proving is another formal methods technique that is widely used for the verification of the physical systems but has been rarely used for analyzing system biology related problems. In theorem proving, a computer-based mathematical model of the given system is constructed and then deductive reasoning is used for the verification of its intended properties. A prerequisite for conducting the formal analysis of a system is to formalize the mathematical or logical foundations that are required to model the system in an appropriate logic. Zsyntax is a recently proposed formal language that supports the modeling of any biological process and presents an analogy between a biological process and the logical deduction. It has some pre-defined operators and inference rules that are used for the logical deductions about a biological process. These operators and inference rules have been designed in such a way that they are easily understandable by the biologists, making Zsyntax a biologist-centered formalism, which is the main strength of this language. However, Zsyntax does not support specifying the temporal information associated with biological processes. Reaction kinetics, on the other hand, caters for this limitation by providing the basis to understand the time evolution of molecular populations involved in a biological network. This approach is based on the set of first- order ordinary differential equations (ODEs) also called reaction rate equations (RREs). Most of these equations are non-linear in nature and difficult to analyze but provide very useful insights for prognosis and drug predictions. Traditionally, the manual paper-and-pencil technique is used to reason logically about biological processes, which are expressed in Zsyntax. Similarly, the analysis of RREs is performed by either paper-and-pencil based proofs or numerical simulation. However, both methods suffer from the inherent incompleteness of numerical methods and error-proneness of manual proofs. We believe that these issues cannot be ignored considering the critical nature of this analysis due to the involvement of human lives. Moreover, biological experiments based on erroneous parameters, derived by the above-mentioned approaches may also result in the loss of time and money, due to the slow nature of wet-lab experiments and the cost associated with the chemicals and measurement equipment. In this paper, we propose to develop a formal reasoning support for system biology to analyze complex biological systems within the sound core of a theorem prover and thus provide accurate analysis results in this safety-critical domain. By formal reasoning support, we mean to develop a set of generic mathematical models and definitions, a process that is usually termed as formalization, of commonly used notions of system biology using an appropriate logic and ascertain their properties as formally verified theorems in a theorem prover, which is a verification tool based on deductive reasoning. These formalized definitions and formally verified theorems can then in turn be used to develop formal models of real-world system biology problems and thus verify their corresponding properties accurately within the sound core of a theorem prover. The use of logic in modeling and a theorem prover in the verification leads to the accuracy of the analysis results, which cannot be ascertained by other computational approaches. In our recent work, we developed a formal deduction framework for reasoning about molecular reactions by formalizing the Zsyntax language in the HOL4 theorem prover. In particular, we formalized the logical operators and inference rules of Zsyntax in higher-order logic. We then built upon these formal definitions to verify two key behavioral properties of Zsyntax based molecular pathways. However, it was not possible to reason about biological models based on reaction kinetics due to the unavailability of the formal notions of reaction rate equations (a set of coupled differential equations) in higher-order logic. In order to broaden the horizons of formal reasoning about system biology, this paper presents a formalization of reaction kinetics along with the development of formal models of generic biological pathways without the restriction on the number of molecules and corresponding interconnections. Furthermore, we formalize the transformation, which is used to convert biological reactions into a set of coupled differential equations. This step requires multivariate calculus (e.g., vector derivative, matrices, etc.) formalization in higher-order logic, which is not available in HOL4 and therefore we have chosen to leverage upon the rich multivariable libraries of the HOL Light theorem prover to formalize the above mentioned notions and verify the reactions kinetics of some generic molecular reactions. To make the formalization of Zsyntax consistent with the formalization of reaction kinetics in HOL Light, as part of our current work, we ported all of the HOL4 formalization of Zsyntax to HOL Light. In order to illustrate the usefulness and effectiveness of our formalization, we present the formal analysis of a molecular reaction representing the TP53 Phosphorylation, a molecular reaction of pathway leading to the death of cancer stem cells (CSC) and the analysis of tumor growth based on the CSC. # Related work In the last few decades, various modeling formalisms of computer science have been widely used in system biology. We briefly outline here the applications of computational modeling and analysis approaches in system biology, where the main idea is to transform a biological model into a computer program. Process algebra (PA) provides an expressive framework to formally specify the communication and interactions of concurrent processes without ambiguities. Biological systems can be considered as concurrent processes and thus process algebra can be used to model biological entities. Some recent work in this area includes the formalizations of molecular biology based on K-Calculus and π-Calculus. The main tools that support PA in biology are sCCP, BioShape and Bio-PEPA. Even though PA based biological modeling provides sound foundations, it may be quite difficult and cumbersome for working biologists to understand these notations. Rule-based modeling offers a flexible and simple framework to model various biochemical species in a textual or graphical format. This allows biologists to perform the quantitative analysis of complex biological systems and predict important underlying behaviors. Some of the main rule-based modeling tools are BioNetGen, Kappa and BIOCHAM. These tools are mainly based on rewriting and model transformation rules along with the integration with model checking tools and numerical solvers. However, these integrations are usually not checked for correctness (for example by an independent proof assistant), which may lead to inconsistencies. Boolean networks are used to characterize the dynamics of gene-regulatory networks by limiting the behavior or genes by either a truth state or false state. Some of the major tools that support the Boolean modeling of biological systems are BoolNet, BNS and GINsim. The discrete nature of Boolean networks does not allow us to capture continuous biological evolutions, which are usually represented by differential equations. Model checking has shown very promising results in many applications of molecular biology. Hybrid systems theory extends the state-based discrete representation of traditional model checking with a continuous dynamics (described in terms ODEs) in each state. Some of the recently developed tools that support the hybrid modeling of biological systems are S-TaLiRo, Breach toolbox and dReach. Recently, Petri nets have been widely used to model biological networks and some of the important associated tools include Snoopy and GreatSPN. However, the graph or state based nature of the models in these methods only allow the description of some specific areas of molecular biology. Moreover, the model checking technique has an inherent state-space explosion problem, which makes it only applicable to the biological entities that can acquire a small set of possible levels and thus limits its scope by restricting its usage on larger systems. In a system analysis based on theorem proving, we need to formalize the mathematical or logical foundations required to model and analyze that system in an appropriate logic. Several attempts have been made to formalize the foundations of molecular biology. The first attempt at some basic axiomatization dates back to 1937. Zanardo et al. and Rizzotti et al. have also done some efforts towards the formalization of biology. But all these formalizations are paper-and-pencil based and have not been utilized to formally reason about molecular biology problems within a theorem prover. In our recent work, we developed a formal deduction framework for reasoning about molecular reactions by formalizing the Zsyntax language in the HOL4 theorem prover. However, a major limitation of this work is that it cannot cater for the temporal information associated with biological processes and, hence, does not support modeling the time evolution of molecular populations involved in a biological network, which is of a dire need when studying the dynamics of a biological system. Reaction kinetics provide the basis to understand the time evolution of molecular populations involved in a biological network. To overcome the limitation of the work presented by Sohaib et al., we provide the formalization of reaction kinetics in higher-order logic and in turn extend the formal reasoning about system biology. # Higher-order-logic theorem proving and HOL Light theorem prover In this section, we provide a brief introduction to the higher-order-logic theorem proving and HOL Light theorem prover. ## Higher-order-logic theorem proving Theorem proving involves the construction of mathematical proofs by a computer program using axioms and hypothesis. Theorem proving systems (theorem provers) are widely used for the verification of hardware and software systems and the formalization (or mathematical modeling) of classical mathematics. For example, hardware designers can prove different properties of a digital circuit by using some predicates to model the circuits model. Similarly, a mathematician can prove the transitivity property for real numbers using the axioms of real number theory. These mathematical theorems are expressed in logic, which can be a propositional, first-order or higher-order logic based on the expressibility requirement. Based on the decidability or undecidability of the underlying logic, theorem proving can be done automatically or interactively. Propositional logic is decidable and thus the sentences expressed in this logic can be automatically verified using a computer program whereas higher-order logic is undecidable and thus theorems about sentences, expressed in higher-order logic, have to be verified by providing user guidance in an interactive manner. A theorem prover is a software for deductive reasoning in a sound environment. For example, a theorem prover does not allow us to conclude that “$\frac{x}{x} = 1$” unless it is first proved or assumed that x ≠ 0. This is achieved by defining a precise syntax of the mathematical sentences that can be input in the software. Moreover, every theorem prover comes with a set of axioms and inference rules which are the only ways to prove a sentence correct. This purely deductive aspect provides the guarantee that every sentence proved in the system is actually true. ### HOL Light theorem prover HOL Light is an interactive theorem prover used for the constructions of proofs in higher-order logic. The logic in HOL Light is represented in meta language (ML), which is a strongly-typed functional programming language. A theorem is a formalized statement that may be an axiom or could be deduced from already verified theorems by an inference rule. Soundness is assured as every new theorem must be verified by applying the basic axioms and primitive inference rules or any other previously verified theorems/inference rules. A HOL Light theory is a collection of valid HOL Light types, axioms, constants, definitions and theorems, and is usually stored as an ML file in computers. Users interacting with HOL Light can reload a theory and utilize the corresponding definitions and theorems right away. Various mathematical foundations have been formalized and stored in HOL Light in the form of theories by the HOL Light users. HOL Light theories are organized in a hierarchical fashion and child theories can inherit the types, constants, definitions and theorems of the parent theories. The HOL Light theorem prover provides an extensive support of theorems regarding Boolean variables, arithmetics, real numbers, transcendental functions, lists and multivariate analysis in the form of theories which are extensively used in our formalization. The proofs in HOL Light are based on the concept of tactics which break proof goals into simple subgoals. There are many automatic proof procedures and proof assistants available in HOL Light, which help the user in concluding a proof more efficiently. # Proposed framework The proposed theorem proving based formal reasoning framework for system biology, depicted in, allows the formal deduction of the complete pathway from any given time instance and model and analyze the ordinary differential equations (ODEs) corresponding to a kinetic model for any molecular reaction. For this purpose, the framework builds upon existing higher-order-logic formalizations of Lists, Pairs, Vectors, and Calculus. The two main rectangles in the higher-order logic block present the foundational formalizations developed to facilitate the formal reasoning about the Zsyntax based pathway deduction and the reaction kinetics. In order to perform the Zsyntax based molecular pathway deduction, we first formalize the functions representing the logical operators and inference rules of Zsyntax in higher- order logic and verify some supporting theorems from this formalization. This formalization can then be used along with a list of molecules and a list of Empirically Valid Formulae (EVFs) to formally deduce the pathway for the given list of molecules and provide the result as a formally verified theorem using HOL Light. Similarly, we have formalized the flux vectors and stoichiometric matrices in higher-order-logic. These foundations can be used along with a given list of species and the rate of the reactions to develop a corresponding ODEs based kinetic reactions model. The solution to this ODE can then be formally verified as a theorem by building upon existing formalizations of Calculus theories. The distinguishing characteristics of the proposed framework include the usage of deductive reasoning to derive the deduced pathways and solutions of the reaction kinetic models. Thus, all theorems are guaranteed to be correct and explicitly contain all required assumptions. # Results ## Formalization of Zsyntax Zsyntax is a formal language which exploits the analogy between biological processes and logical deduction. Some of its key features are that: 1) it enables us to represent molecular reactions in a mathematical rigorous way; 2) it is of heuristic nature, i.e., if the initialization data and the conclusion of a reaction is known, then it allows us to deduce the missing data based on the initialization data; and 3) it possesses computer implementable semantics. Zsyntax has three operators namely Z-Interaction, Z-Conjunction and Z-Conditional that are used to represents different phenomenon in a biological process. These are the atomic formulas residing in the core of Zsyntax. Z-Interaction (\)* represents the reaction or interaction of two molecules. In biological reactions, the Z-interaction operation is not associative. i.e., in a reaction having three molecules namely A, B and C, the operation (A ∗ B) ∗ C is not equal to A ∗ (B ∗ C). Z-Conjunction (&) is used to form the aggregate of the molecules participating in the biological process. These molecules can be same or different. Unlike the Z-Interaction operator, the Z-Conjunction is fully associative. Z-Conditional (→) is used to represent a path from A to B when condition C becomes true, i.e., A → B if there is a C allowing it. To apply the above-mentioned operators on a biological process, Zsyntax provides four inference rules that are used for the deduction of the outcomes of the biological reactions. These inference rules are given in. Zsyntax also utilizes the EVFs which are the empirical formulas validated in the lab and are basically the non-logical axioms of molecular biology. A biological reaction can be mapped and then these above-mentioned Zsyntax operators and inference rules are used to derive the final outcome of the reaction as shown in. We start our formalization of Zsyntax, by formalizing the molecule as a variable of arbitrary data type (α). Z-Interaction is represented by a list of molecules (α list), which is a molecular reaction among the elements of the list. This (α list) may contain only a single element or it can have multiple elements. We model the Z-Conjunction operator as a list of list of molecules ((α list) list), which represents a collection of non-reacting molecules. Using this data type, we can apply the Z-Conjunction operator between individual molecules (a list with a single element), or between multiple interacting molecules (a list with multiple elements). Thus, based on our datatype, Z-Conjunction is a list of Z-interactions for both of these cases, i.e., individual molecules or multiple interacting molecules. So, overall, Z-conjunction acts as a set of Z-interaction. When a new set of molecules is generated based on the EVFs available for a reaction, the status of the molecules is updated using the Z-Conditional operator. We model each EVF as a pair of data type (α list \# α list list) where the first element of the pair is a list of the molecules represented by data type (α list) and are actually the reacting molecules, whereas, the second element is a list of list of molecules ((α list) list), which represents a set of molecules that are obtained as a result of the reaction between the molecules of the first element of the pair and thus act as a set of Z-Interactions. A collection of EVFs is formalized using the data type ((α list \# α list list) list), which is a list of EVFs. Next, we formalize the inference rules using higher-order logic. The inference rule named elimination of the Z-Conditional (→E) is equivalent to the Modus Ponens (the elimination of implication rule) law of propositional logic. Similarly, we can infer introduction of Z-Conditional (→I) rule from the existing rules of the propositional logic present in a theorem prover. Thus, both of these rules can be handled by the simplification and rewriting rules of the theorem prover and we do not need to define new rules for handling these inference rules. To check the presence of a particular molecule in an aggregate of some inferred molecules, the elimination of the Z-Conjunction (& E) rule is used. We apply it at the end of the biological reaction to check whether the product of the reaction is the desired molecule or not. We formalized this rule by a function (: `zsyn_conjun_elimin`), which accepts a list `l` and an element `x` and checks if `x` is present in this list. If the condition is true, it returns the given element `x` as a single element of that list `l`. Otherwise, it returns the list `l` as is, as shown in. The Z-Interaction and the introduction of Z-Conjunction (& I) rule jointly enable us to perform a reaction between different molecules during the experiment. This rule is basically the append operation of lists, based on the above data types defined in our formalization. The function `zsyn_conjun_intro`, given in, represents this particular rule. It takes a list `l` and two of its elements `x` and `y`, and appends the list of these two elements on its head as shown in. According to laws of stoichiometry, we have to delete the initial reacting molecules from the main list, for which the Z-Conjunction operator is applied. Our formalization of this behavior is represented by the function `zsyn_delet`, given in and depicted in. The function `zsyn_delet` accepts a list `l` and two numbers `x` and `y` and deletes the x<sup>th</sup> and y<sup>th</sup> elements of the given list `l`. The function checks if the index x is greater than the index y, i.e., x \> y. If the condition is true, then it deletes the x<sup>th</sup> element first and then the y<sup>th</sup> element. Similarly, if the condition x \> y is false, then it deletes the y<sup>th</sup> element first and then the x<sup>th</sup> element. In this deletion process, to make sure that the deletion of first element will not affect the index of the other element that has to be deleted, we delete the element present at the higher index of list before the deletion of the lower indexed element. We aim to build a framework that takes the initial molecules of a biological experiment along with the possible EVFs and enables us to deduce its corresponding final outcomes. Towards this, we first write a function `zsyn_EVF`, given in and depicted in, that takes a list of initial molecules and compares its particular combination with the corresponding EVFs and if a match is found then it adds the newly resulted molecule to initial list after deleting the instance that have already been consumed. The function `zsyn_EVF` takes a list of molecules `l` and a list of EVFs `e` and compares the head element of the list `l` to all of the elements of the list `e`. Upon finding no match, this function returns a pair having first element as false (`F`), which acts as a flag and indicates that there is no match between any of the EVFs and the corresponding molecule, whereas the second element of the pair is the tail of the corresponding list `l` of the initial molecules. If a match is found, then the function will return a pair with its first element as a true (`T`), which indicates the confirmation of the match that have been found, and the second element of the pair is the modified list `l`, whose head is removed, and the second element of the corresponding EVF pair is added at the end of the list and the matched elements are deleted as these have already been consumed. Next, we have to call the function `zsyn_EVF` recursively, for the deduction of the final outcome of the experiment and for each of the recursive case, we place each of the possible combinations of the given molecules (elements at indices `x` and `y` of list `l`) at the head of `l` one by one. This whole process can be done using functions `zsyn_recurs1` and `zsyn_recurs2`, given in. In the function `zsyn_recurs1`, we first place the combination of molecules indexed by variables `x` and `y` at the top of the list `l` using the introduction of Z-Conjunction rule. Then, this modified list `l` is passed to the function `zsyn_EVF`, which is recursively called by the function `zsyn_recurs1`. Moreover, we instantiate the variable `p` of the function `zsyn_EVF` with the length of the EVF list `(LENGTH e - 1)` so that every new combination of the list `l` is compared with all the elements of the list of EVFs `e`. The function `zsyn_recurs1` terminates upon finding a match in the list of EVFs and returns true (`T`) as the first element of its output pair, which acts as a flag for the status of this match. The second function `zsyn_recurs2` checks, if a match in the list of EVFs `e` is found (if the flag returns true (`T`)) then it terminates and returns the output list of the function `zsyn_recurs1`. Otherwise, it recursively checks for the match with all of the remaining values of the variable `x`. In the case of a match, these two functions `zsyn_recurs1` and `zsyn_recurs2` have to be called all over again with the new updated list. This iterative process continues until no match is found in the execution of these functions. This overall behaviour can be expressed in HOL Light by the recursive function `zsyn_deduct_recurs`, given in. In order to guarantee the correct operation of deduction, we instantiate the variable of recursion (`q`) with a value that is greater than the total number of EVFs so that the application of none of the EVF is missed. Similarly, in order to ensure that all the combinations of the list `l` are checked against the entries of the EVF list `e`, the value `LENGTH l - 1` is assigned to both of the variables `x` and `y`. Thus, the final deduction function for Zsyntax can be modeled as the function `zsyn_deduct`, given in. The function `zsyn_deduct` accepts the initial list of molecules `l` and the list of valid EVFs `e` and returns a list of final outcomes of the experiment under the given conditions. Next, in order to check, if the desired molecule is present in this list (the output of the function `zsyn_deduct`), we apply the elimination of the Z-Conjunction rule presented as function `zsyn_conjun_elimin`, given in. More detail about the behavior of all of these functions can be found in our proof script. These formal definitions enable us to check recursively all of the possible combinations of the molecules, present in the initial list `l`, against each of the first element of the list of EVFs `e`. Upon finding a match, the reacting molecules are replaced by their outcome in the initial list of molecules `l` by applying the corresponding EVF. This process is repeated on the current updated list of molecules until there are no further molecules reacting with each other. The list `l` at this point contains the post-reaction molecules. Finally, the elimination of the Z-Conjunction rule `zsyn_conjun_elimin`, given in, is applied to obtain the desired outcome of the given biological experiment. In order to prove the correctness of the formal definitions presented above, we verify a couple of key properties of Zsyntax involving operators depicting the vital behaviour of the molecular reactions. The first verified property captures the scenario when there is no reacting molecule present in the initial list of the experiment. As a result of this scenario, the post-experiment molecules are the same as the pre-experiment molecules. The second property deals with the case when there is only one set of reacting molecules in the given initial list of molecules and in this scenario we verify that after the execution of the Zsyntax based experiment, the list of post-experiment molecules contains the products of the reacting molecules minus its reactant along with the remaining non-reacting molecules provided at the beginning of the experiment. We formally specified both of these properties, representing the no reaction and single reaction scenarios in higher-order logic using the formal definitions presented earlier in this section. The formal verification results about these properties are given in and more details can be found in the description of their formalization. The formalization presented in this section provides an automated reasoning support for the Zsyntax based molecular biological experiments within the sound core of HOL Light theorem prover. ## Formalization of reaction kinetics Reaction kinetics is the study of rates at which biological processes interact with each other and how the corresponding processes are affected by these reactions. The rate of a reaction provides the information about the evolution of the concentration of the species (e.g., molecules) over time. A process is basically a chain of reactions, called pathway, and the investigation about the rate of a process implies the rate of these pathways. Generally, biological reactions can be either irreversible (unidirectional) or reversible (bidirectional). We formally define this fact by an inductive enumerating data- type `reaction_type`, given in. In order to analyze a biological process, we need to know its kinetic reaction based model, which comprises of a set of m species, X = {X<sub>1</sub>, X<sub>2</sub>, X<sub>3</sub>,…, X<sub>m</sub>} and a set of n reactions, R = {R<sub>1</sub>, R<sub>2</sub>, R<sub>3</sub>,…, R<sub>n</sub>}. An irreversible reaction R<sub>j</sub>, {1 ≤ j ≤ n} can generally be written as: $R_{j}:s_{1,j}X_{1} + s_{2,j}X_{2} + \ldots + s_{m,j}X_{m}\overset{k_{j}}{\rightarrow}{\acute{s}}_{1,j}X_{1} + {\acute{s}}_{2,j}X_{2} + \ldots + {\acute{s}}_{m,j}X_{m}$. Similarly, a reversible reaction R<sub>j</sub>, {1 ≤ j ≤ n} can be described as: $R_{j}:s_{1,j}X_{1} + s_{2,j}X_{2} + \ldots + s_{m,j}X_{m}\underset{{k_{j}}^{r}} {\overset{{k_{j}}^{f}}{\rightleftarrows}}{\acute{s}}_{1,j}X_{1} + {\acute{s}}_{2,j}X_{2} + \ldots + {\acute{s}}_{m,j}X_{m}$. The coefficients $s_{ 1,j},s_{2,j},\ldots,s_{m,j},{\acute{s}}_{1,j},{\acute{s}}_{2,j},\ldots,{\acute{s }}_{m,j}$ are the non-negative integers and represent the stoichiometries of the species taking part in the reaction. The non-negative integer k<sub>j</sub> is the kinetic rate constant of the irreversible reaction. The non-negative integers ${k_{j}}^{f}$ and ${k_{j}}^{r}$ are the forward and reverse kinetic rate constants of the reversible reaction, respectively. In a biological reaction, we model a biological entity as a pair ($\mathbb{N}$, $\mathbb{R}$), where the first element represents the stoichiometry and the second element is the concentration of the molecule. We formally model a biological reaction as the type abbreviation `bio_reaction`, given in. The dynamic behavior of the biological systems is described by a set of ordinary differential equations (ODEs) and the evolution of the system is captured by analyzing the change in the concentration of the species (i.e., time derivatives): $\frac{d\left\lbrack X_{i} \right\rbrack}{dt} = \sum_{j = 1}^{n}n_{i,j}v_{j}$, where n<sub>i, j</sub> is the stoichiometric coefficient of the molecular species X<sub>i</sub> in reaction R<sub>j</sub> (i.e., $n_{i,j} = {\acute{s}}_{i,j} - s_{i,j}$). The parameter v<sub>j</sub> represents the flux of the reaction R<sub>j</sub>, which can be computed by the law of mass action, i.e., the rate (also called flux) of a reaction is proportional to the concentration of the reactant (c) raised to the power of its stoichiometry (s), i.e., c<sup>s</sup>. We define the function `gen_flux_irreversible`, given in, to obtain the flux of an irreversible reaction. A reversible reaction can be divided into two irreversible reactions with the forward kinetic rate constant and the reverse kinetic rate constant, respectively. The rate/flux of a reversible reaction is obtained by taking the differences of the fluxes of the two irreversible reactions. We formally define the flux of a reversible reaction by the function `gen_flux_reversible`, given in. Next, we combine the functions `gen_flux_irreversible` and `gen_flux_reversible` into one uniform function `flux_single` to obtain the flux of a single reaction. For all reactions from 1 to n of a biological system, the flux becomes a flux vector as v = (v<sub>1</sub>, v<sub>2</sub>,…, v<sub>n</sub>)<sup>T</sup> and the system of ODEs can be written in the vectorial form as: $\frac{d\left\lbrack \mathbf{X} \right\rbrack}{dt} = N\mathbf{v}$, where \[X\] = (X<sub>1</sub> X<sub>2</sub>,…, X<sub>n</sub>)<sup>T</sup> is a vector of the concentration of all of the species participating in the reaction and N is the stoichiometric matrix of order m × n. We can obtain the flux vector v for a chain of reactions of a biological system by the function `flux`, given in. Next, we formalize the notion of stoichiometric matrix N by the function `st_matrix` given in. Finally, in order to formalize the left-hand side of above vector equation, i.e., $\frac{d\left\lbrack \mathbf{X} \right\rbrack}{dt}$, we define a function `entities_deriv_vec` which takes a list containing the concentrations of all species and returns a vector with each element represented in the form of a real-valued derivative. We can utilize this infrastructure to model arbitrary biological networks consisting of any number of reactions. For example, a biological network consisting of a list of `E` biological species and `M` biological reactions can be formally represented by the following general kinetic model: $$\begin{array}{r} {\left( \left( \texttt{entities}\_\texttt{deriv}\_\texttt{vec }\mspace{720mu}\mathtt{E}\mspace{720mu}\mathtt{t} \right):\texttt{real}\hat{\mspace{720mu}}\mathtt{m} \right) = \texttt{transp}\left( \left( \texttt{st}\_\texttt{matrix}\mspace{720mu}\mathtt{M} \right):{\texttt{real}\hat{ \mspace{720mu}}\mathtt{m}}\hat{\mspace{720mu}}\mathtt{n} \right)\mspace{720mu}*\mspace{720mu}\texttt{flux}\mspace{720mu}\mathtt{M}} \\ \end{array}$$ We used the formalization of the reaction kinetics to verify some generic properties of biological reactions, such as irreversible consecutive reactions, reversible and irreversible mixed reactions. The main idea is to express the given biological network as a kinetic model and verify that the given solution (mathematical expression) of each biological entity satisfies the resulting set of coupled differential equations. This verification is quite important as such expressions are used to predict the outcomes of various drugs and to understand the time evolution of different molecules in the reactions of the biological systems. ### The irreversible consecutive reactions We consider a general irreversible consecutive reaction scheme as shown in. In the first reaction, `A` is the reactant and `B` is the product whereas `k1` represents the kinetic rate constant of the reaction. Similarly, in the second reaction, `B` is the reactant, `C` is the product and `k2` is its kinetic rate constant. We formally model this reaction scheme as a HOL Light function `rea_sch_01`, given in, and the formalization details are available as a technical report. We moreover verify the solution of its kinetic model in HOL Light. The formal verification results are given in. ### The consecutive reactions with the second step being reversible The second reaction scheme consists of the consecutive reactions with the second step being reversible as shown in. In the irreversible reaction, `A` and `B` are the reactant and product, respectively, whereas `k1` is the kinetic rate constant of the reaction. Since any reversible reaction can be written as two irreversible reactions, so the first irreversible reaction has `B`, `C` and the forward kinetic reaction constant `k2` as the reactant, product and the kinetic rate constant, respectively. Similarly, the parameters `C`, `B` and `k3` are the reactant, product and kinetic rate constant of the second irreversible reaction, respectively. We formally model this scheme as a HOL Light function `rea_sch_02` and then verified the solution for its ODE model given in. ### The consecutive reactions with the first step as a reversible reaction In this scheme, the first reaction is reversible and the second reaction is irreversible as shown in. The reversible reaction can be equivalently written as two irreversible reactions with `k1` and `k2` as their kinetic rate constants. In the first irreversible reaction, `A` and `B` are the reactant and product, respectively, whereas in the second reaction, `B` and `A` are the reactant and product, respectively. For the second step, `B`, `C` and `k3` are the reactant, product and kinetic rate constant, respectively. The verified solution of the ODE model corresponding to this reaction scheme (`rea_sch_03`, given in) is given in. ### The consecutive reactions with a reversible step In this reaction scheme, we consider the consecutive reactions with one reversible and one irreversible reaction step as shown in. The ODE model and solution corresponding to this reaction scheme (rea_sch_04, given) are given in. This completes our formal verification of some commonly used reaction schemes. The verification of these solutions requires user interaction but the strength of these theorems lies in the fact that they have been verified for arbitrary values of parameters, such as k<sub>1</sub> and k<sub>2</sub>, etc. This is a unique feature of higher-order-logic theorem proving that is not possible in the case of simulation where such continuous expressions are tested for few samples of such parameters. Another important aspect is the explicit presence of all assumptions required to verify the set of ODEs. For example, such assumptions for the above-mentioned reaction schemes are not mentioned in Korobov et al.’s paper. More details about the formalization of all above- mentioned types and functions and the formal verification of all above properties, and its source code can be found on our project’s webpage. ## Case studies In this section, we use our proposed framework to formally reason about three case studies: In the first, we formally analyse the reaction involving the phosphorylation of TP53 using our formalization of Zsyntax. In the second, we formally derive the time evolution expressions of different tumor cell types, which are used to predict the tumor population and volume at a given time instant, using our formalization of reaction kinetics. In the third, we take another model for the growth of tumor cells and perform both the Zsyntax and reaction kinetic based formal analysis using our proposed formalizations presented in the Result* section of the paper. ### TP53 phosphorylation TP53 gene encodes p53 protein, which plays a crucial role in regulating the cell cycle of multicellular organisms and works as a tumour suppressor for preventing cancer. The pathway leading to TP53 phosphorylation (p(TP53)) is shown in. The green-colored circle represents the desired product, whereas, the blued-colored circles describe the chemical interactions in the pathway. Similarly, each rectangle in contains the total number of molecules at a given time. It can be clearly seen from the figure that whenever a biological reaction results into a product, the reactants get consumed, which satisfies the stoichiometry of a reaction. Now, we present the formal verification of pathway deduction from TP53 to p(TP53) using our formalization of Zsyntax, presented in the last section. In classical Zsyntax format, the reaction of the pathway leading from TP53 to p(TP53) can be represented by a theorem as `TP53` `&` `ATP` `&` `Kinase` `⊢` `p(TP53)`. Based on our formalization, it can be defined as follows: Theorem 1. The reaction of the pathway leading from TP53 to p(TP53) ⊢ `DISTINCT` `[TP53;` `ATP;` `Kinase;` `ADP;` `pTP53]` `⇒` `zsyn_conjun_elimin (zsyn_deduct [[TP53];[ATP];[Kinase]]` `[([Kinase;ATP],[[ATP;Kinase]]);` `([ATP;Kinase;TP53],[[Kinase];[pTP53];[ADP]])]) [pTP53] = [[pTP53]]` In the above theorem, the first argument of the function `zsyn_deduct` represents the list of initial aggregate (IA) of molecules that are present at the start of the reaction, whereas the second argument is the list of valid EVFs for this reaction specified in the form of pairs and include the molecules (ATP, Kinase, etc.). These are obtained from wet lab experiments, as reported by Boniolo et al.. We use the HOL Light function `DISTINCT` to ensure that all molecule variables (from IA and EVFs) used in this theorem represent distinct molecules. Thus, the final list of molecules is deduced under these particular conditions using the function `zsyn_deduct`. Finally, if the molecule `pTP53` is present in the post-reaction list of molecules, it will be obtained after the application of the function `zsyn_conjun_elimin`, as previously described. Additionally, in order to automate the verification process, we developed a simplifier `Z_SYNTAX_SIMP`, which is based on some derived rules and already available HOL Light tactics that simplified the manual reasoning and thus allowed us to formally verify Theorem 1 automatically. It is important to note that formalization of Zsyntax was quite a tedious effort but it took only 6 lines of code for the verification of the theorem of pathway deduction from `TP53` to `pTP53` in HOL Light, which clearly illustrates the effectiveness of our foundational work. We have shown that our formalization is capable of modeling molecular reactions using Zsyntax inference rules, i.e., given an IA A and a set of possible EVFs, our proposed framework can derive a final aggregate (FA) B from A automatically. If it fails to deduce B, our formalism still provides all the intermediate steps to the biologist so that he can figure out the possible causes of failures, by carefully examining the intermediate steps of the reaction. ### Formal analysis of tumor growth based on Cancer Stem Cells (CSC) According to the Cancer Stem Cell (CSC) hypothesis, malignant tumors (cancers) are originally initiated by different tumor cells, which have similar physiological characterises as of normal stem cells in the human body. This hypothesis explains that the cancer cell exhibits the ability to self-renew and can also produce different types of differentiated cells. The mathematical and computational modeling of cancers can provide an in-depth understanding and the prediction of required parameters to shape the clinical research and experiments. This can result in efficient planning and therapeutic strategies for accurate patient prognosis. In this paper, we consider a kinetic model of cancer based on the cancer stem cell (CSC) hypothesis, which was recently proposed in Molina-Pena et al.’s paper. In this model, four types of events are considered: 1) CSC self-renewal; 2) maturation of CSCs into P cells; 3) differentiation to D cells; and 4) death of all cell subtypes. All of these types of reactions are driven by different rate constants as shown in. In the following, we provide the possible reactions in the considered model of cancer: Expansion of CSCs can be accomplished through symmetric division, where one CSC can produce two CSCs, i.e., $\text{CSC}\overset{\text{k}_{1}}{\rightarrow}2\text{CSC}$. A CSC can undergo asymmetric division (whereby one CSC gives rise to another CSC and a more differentiated progenitor (P) cell). This P cell possesses intermediate properties between CSCs and differentiated (D) cells, i.e., $\text{CSC}\overset{\text{k}_{2}}{\rightarrow}\text{CSC} + \text{P}$. The CSCs can also differentiate to P cells by symmetric division, i.e., $\text{CSC}\overset{\text{k}_{3}}{\rightarrow}2\text{P}$. The P cells can either self-renew, with a decreased capacity compared to CSCs, or they can differentiate to D cells, i.e., $\text{P}\overset{\text{k}_{4}}{\rightarrow}2\text{P}$, $\text{P}\overset{\text{k}_{5}}{\rightarrow}2\text{D}$. All cellular subtypes can undergo cell death (M), i.e., $\text{CSC}\overset{\text{k}_{6}}{\rightarrow}\text{M}$, $\text{P}\overset{\text{k}_{7}}{\rightarrow}\text{M}$, $\text{D}\overset{\text{k}_{8}}{\rightarrow}\text{M}$. In order to reduce the complexity of the resulting model, only three subtypes of cells are considered: CSCs, transit amplifying progenitor cells (P), and terminally differentiated cells (D) as shown in. This assumption is consistent with several experimental reports. Our main objective is to derive the mathematical expressions, which characterize the time evolution of CSC, P and D. Concretely, the values of these cells should satisfy the set of differential equations that arise in the kinetic model of the proposed tumor growth. Once the expressions of all cell types are known, the total number of tumor cells (N) in the human body can be computed by the formula N(t) = CSC(t) + P(t) + D(t). Furthermore, the tumor volume (V) can be calculated by the formula V(t) = 4.18 × 10<sup>6</sup> N(t), considering that the effective volume contribution of a spherically shaped cell in a spherical tumor (i.e., 4.18 × 10<sup>−6</sup> mm<sup>3</sup>/cell). We formally model the tumor growth model and verify the time evolution expressions for `CSC`, `P` and `D` that satisfy the general kinetic model. We formally represent this requirement in the following important theorem: Theorem 2. Time Evolution Verification of Tumor Growth Model ⊢ `∀` `k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `k`<sub>`4`</sub> `k`<sub>`5`</sub> `k`<sub>`6`</sub> `k`<sub>`7`</sub> `CSC` `P` `D` `M` `t` `k`<sub>`8`</sub>. `A1: ((−k`<sub>`1`</sub> `+k`<sub>`3`</sub> `+k`<sub>`4`</sub> `−k`<sub>`5`</sub> `+k`<sub>`6`</sub> `−k`<sub>`7`</sub> `)(−k`<sub>`1`</sub> `+k`<sub>`3`</sub> `+k`<sub>`6`</sub> `−k`<sub>`8`</sub> `)(−k`<sub>`4`</sub> `+k`<sub>`5`</sub> `+k`<sub>`7`</sub> `−k`<sub>`8`</sub> `)≠0) ∧` `A2: (k`<sub>`1`</sub> `−k`<sub>`3`</sub> `−k`<sub>`4`</sub> `+k`<sub>`5`</sub> `−k`<sub>`6`</sub> `+k`<sub>`7`</sub> `≠0) ∧` `A3: ∀t. CSC(t) = e`<sup>`(k`<sub>`1`</sub> `−k`<sub>`3`</sub> `−k`<sub>`6`</sub> `)`</sup> `∧` `A4:` $\!\forall\mathtt{t}.\mspace{720mu}\mathtt{P}\left( \mathtt{t} \right) = \frac{\left\lbrack \left( \mathtt{e}^{(\mathtt{k}_{1} - \mathtt{k}_{3} - \mathtt{k}_{6})\mathtt{t}} - \mathtt{e}^{(\mathtt{k}_{4} - \mathtt{k}_{5} - \mathtt{k}_{7})\mathtt{t}} \right)\left( \mathtt{k}_{2} + 2\mathtt{k}_{3} \right) \right\rbrack}{\left( \mathtt{k}_{1} - \mathtt{k}_{3} - \mathtt{k}_{4} + \mathtt{k}_{5} - \mathtt{k}_{6} + \mathtt{k}_{7} \right)}\mspace{720mu} \land$ `A5:` $\forall\mathtt{t}.\mspace{720mu}\mathtt{D}\left( \mathtt{t} \right) = \frac{\left( 2\mathtt{e}^{- \mathtt{k}_{8}\mathtt{t}}\left( \mathtt{k}_{2} + 2\mathtt{k}_{3} \right)\mathtt{k}_{5}\left\lbrack \left( - 1 + \mathtt{e}^{(\mathtt{k}_{4} - \mathtt{k}_{5} - \mathtt{k}_{7} + \mathtt{k}_{8})\mathtt{t}} \right)\mathtt{k}_{1} + \mathtt{k}_{3} + \mathtt{k}_{4} - \mathtt{k}_{5} + \mathtt{k}_{6} - \mathtt{k}_{7} \right\rbrack \right.}{\left( - \mathtt{k}_{1} + \mathtt{k}_{3} + \mathtt{k}_{4} - \mathtt{k}_{5} + \mathtt{k}_{6} - \mathtt{k}_{7} \right)\left( - \mathtt{k}_{1} + \mathtt{k}_{3} + \mathtt{k}_{6} - \mathtt{k}_{8} \right)\left( - \mathtt{k}_{4} + \mathtt{k}_{5} + \mathtt{k}_{7} - \mathtt{k}_{8} \right)}\mspace{720mu} +$ $\frac{\left( 2\mathtt{e}^{- \mathtt{k}_{8}\mathtt{t}}\left( \mathtt{k}_{2} + 2\mathtt{k}_{3} \right)\mathtt{k}_{5}\left\lbrack \mathtt{e}^{(\mathtt{k}_{1} - \mathtt{k}_{3} - \mathtt{k}_{6} + \mathtt{k}_{8})\mathtt{t}}\left( - \mathtt{k}_{4} + \mathtt{k}_{5} + \mathtt{k}_{7} - \mathtt{k}_{8} \right) + \mathtt{e}^{(\mathtt{k}_{4} - \mathtt{k}_{5} - \mathtt{k}_{7} + \mathtt{k}_{8})\mathtt{t}}\left( - \mathtt{k}_{3} - \mathtt{k}_{6} + \mathtt{k}_{8} \right) \right\rbrack \right.}{\left( - \mathtt{k}_{1} + \mathtt{k}_{3} + \mathtt{k}_{4} - \mathtt{k}_{5} + \mathtt{k}_{6} - \mathtt{k}_{7} \right)\left( - \mathtt{k}_{1} + \mathtt{k}_{3} + \mathtt{k}_{6} - \mathtt{k}_{8} \right)\left( - \mathtt{k}_{4} + \mathtt{k}_{5} + \mathtt{k}_{7} - \mathtt{k}_{8} \right)}\mspace{720mu} \land$ `A6: real_derivative M(t) = k`<sub>`6`</sub> `CSC(t) + k`<sub>`7`</sub> `P(t) + k`<sub>`8`</sub> `D(t)` `⇒ entities_deriv_vec [CSC; P; D; M] t =` `transp (st_matrix (tumor_growth_model CSC P D M k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `k`<sub>`4`</sub> `k`<sub>`5`</sub> `k`<sub>`6`</sub> `k`<sub>`7`</sub> `k`<sub>`8`</sub> `t))` ` ∗ ∗ flux (tumor_growth_model CSC P D M k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `k`<sub>`4`</sub> `k`<sub>`5`</sub> `k`<sub>`6`</sub> `k`<sub>`7`</sub> `k`<sub>`8`</sub> `t))` where the first two assumptions (A1-A2) ensure that the time evolution expressions of `P` and `D` do not contain any singularity (i.e., the value at the expression becomes undefined). The next three assumptions (A3-A5) provide the time evolution expressions for `CSC`, `P` and `D`, respectively. The last assumption (A6) is provided to discharge the subgoal characterizing the time- evolution of `M` (dead cells), which is of no interest and does not impact the overall analysis as confirmed by experimental evidences. Finally, the conclusion of Theorem 2 is the equivalent reaction kinetic (ODE) model of the CSC based tumor growth model. To facilitate the verification process of the above theorem, we developed a simplifier, called `KINETIC SIMP`, which sufficiently reduces the manual reasoning interaction with the theorem prover. After the application of this simplifier, it only takes some arithmetic reasoning to conclude the proof of Theorem 2. More details about the verification process can be found on our project’s webpage. The formal verification of the time-evolution of tumor cell types `CSC`, `P` and `D` in Theorem 2 can be easily used to formally derive the total population and volume of tumor cells. The derived time-evolution expression, verified in Theorem 2, can also be used to understand how the overall tumor growth model works. Moreover, potential drugs are usually designed using the variation of the kinetic rate constants, such as k<sub>1</sub>, k<sub>2</sub>⋯k<sub>8</sub> in Theorem 2, to achieve the desired behavior of the overall tumor growth model and thus Theorem 2 can be utilized to study this behavior formally. On similar lines, the variation of these parameters is used to plan efficient therapeutic strategies for cancer patients and thus the formally verified result of Theorem 2 can aid in accurately performing this task. ### Combined Zsyntax and Reaction kinetic based formal analysis of the tumor growth model In this section, we consider another model for the growth of tumor cells and formally analyze it using both of our Zsyntax and Reaction kinetics formalizations, presented in the Results section of the paper. Pathway Leading to Death of CSC The pathway leading to death of CSC is shown in. The green-colored circle represents the desired product, whereas, the blued-colored circles describe the chemical interactions in the pathway. We use our formalization of Zsyntax to deduce this pathway. In the classical Zsyntax format, the reaction of the pathway leading from CSC to its death can be represented by a theorem as `CSC` `&` `P` `⊢` `M`. Based on our formalization, it can be defined as follows: Theorem 3. The Reaction of the Pathway Leading from CSC to its Death (M) ⊢ `DISTINCT` `[CSC;` `P;` `M]` `⇒` `zsyn_conjun_elimin (zsyn_deduct [[CSC];[P]]` `[([CSC],[[CSC;P]]);` `([CSC;P],[[M]])]) [M] = [[M]]` In the above theorem, the first argument of the function `zsyn_deduct` represents the list of IA of molecules that are present at the start of the reaction, whereas the second argument is the list of valid EVFs for this reaction specified in the form of pairs and include the molecules (CSC, P, etc.). We use the HOL Light function `DISTINCT` to ensure that all molecule variables (from IA and EVFs) used in this theorem represent distinct molecules. Thus, the final list of molecules is deduced under these particular conditions using the function `zsyn_deduct`. Finally, if the molecule `M` is present in the post-reaction list of molecules, it will be obtained after the application of the function `zsyn_conjun_elimin`. We use the simplifier `Z_SYNTAX_SIMP` to formally verify Theorem 3 automatically. Reaction Kinetic based Formal Analysis of a Tumor Growth based on CSC We perform the reaction kinetic based formal analysis of a tumor growth model, which is shown in. In this model, two types of events are considered: 1) maturation of CSCs into P cells; 2) death of all cell subtypes. All of these types of reactions are driven by different rate constants as shown in. In the following, we provide the possible reactions in the considered tumor growth model: A CSC can undergo asymmetric division (whereby one CSC gives rise to another CSC and a more differentiated P cell), i.e., $\text{CSC}\overset{\text{k}_{1}}{\rightarrow}\text{CSC} + \text{P}$. All cellular subtypes can undergo cell death (M), i.e., $\text{CSC}\overset{\text{k}_{2}}{\rightarrow}\text{M}$, $\text{P}\overset{\text{k}_{3}}{\rightarrow}\text{M}$. In order to reduce the complexity of the resulting model, only two subtypes of cells are considered: CSCs and transit amplifying progenitor cells (P) as shown in. Our main objective is to derive the mathematical expressions, which characterize the time evolution of CSC and P. Concretely, the values of these cells should satisfy the set of differential equations that arise in the kinetic model of the proposed tumor growth. Once the expressions of all cell types are known, the total number of tumor cells (N) in the human body can be computed by the formula N(t) = CSC(t) + P(t). We formalize the reaction kinetic based tumor growth model and verify the time evolution expressions for `CSC` and `P` that satisfy the general kinetic model. We formally represent this requirement in the following HOL Light theorem: Theorem 4. Time Evolution Verification of a Tumor Growth Model ⊢ `∀` `k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `CSC` `P` `M` `t.` `A1: (k`<sub>`3`</sub> `−k`<sub>`2`</sub> `≠0) ∧` `A2: ∀t. CSC(t) = e`<sup>`−k`<sub>`2`</sub> ` ∗ t`</sup> `∧` `A3:` $\forall\mathtt{t}.\mspace{720mu}\mathtt{P}\left( \mathtt{t} \right) = \frac{\left\lbrack \left( \mathtt{k}_{3} - \mathtt{k}_{2} - \mathtt{k}_{1} \right)\mathtt{e}^{- \mathtt{k}_{3}\mathtt{t}} + \mathtt{k}_{1}\mathtt{e}^{- \mathtt{k}_{2}\mathtt{t}} \right\rbrack}{\left( \mathtt{k}_{3} - \mathtt{k}_{2} \right)}\mspace{720mu} \land$ `A4: real_derivative M(t) = k`<sub>`2`</sub> `CSC(t) + k`<sub>`3`</sub> `P(t)` `⇒ entities_deriv_vec [CSC; P; M] t =` `transp (st_matrix (tumor_growth_rk_model CSC P M k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `t))` ` ∗ ∗ flux (tumor_growth_rk_model CSC P M k`<sub>`1`</sub> `k`<sub>`2`</sub> `k`<sub>`3`</sub> `t))` where the first assumption (A1) ensures that the time evolution expression of `P` does not contain any singularity. The next two assumptions (A2-A3) provide the time evolution expressions for `CSC` and `P`, respectively. The last assumption (A4) is provided to discharge the subgoal characterizing the time- evolution of `M` (dead cells), which is of no interest and does not impact the overall analysis as confirmed by experimental evidences. Finally, the conclusion of Theorem 4 is the equivalent reaction kinetic (ODE) model of the CSC based tumor growth model. To facilitate the verification process of the above theorem, we use the `KINETIC_SIMP` simplifier, which sufficiently reduces the manual reasoning interaction with the theorem prover. After the application of this simplifier, it only takes some arithmetic reasoning to conclude the proof of Theorem 4. More details about the verification process can be found at. # Discussion Most of the existing research related to the formal analysis of the biological systems has been focussed on using model checking. However, this technique suffers from the inherent state-space explosion problem, which limits the scope of this success to systems where the biological entities can acquire only a small set of possible levels. Moreover, the underlying differential equations describing the reaction kinetics are solved using numerical approaches, which compromises the precision of the analysis. To the best of our knowledge, our work is the first one to leverage the distinguishing features of interactive theorem proving to reason about the solutions to system biology problems. We consider the concentration of the species of the biological systems in reaction kinetic based formal analysis as a continuous variable. Besides formalizing Zsyntax and the reaction kinetics of commonly used biological pathways, we also formally verified their classical properties. This verification guarantees the soundness and the correctness of our formal definitions. It also enables us to conduct formal analysis of real-world case studies. In order to illustrate the practical effectiveness of our formalization, we presented the automatic Zsyntax based formal analysis of pathway leading to TP53 Phosphorylation and a pathway leading to the death of CSCs in the tumor growth model, and reaction kinetics based analysis of the tumor growth model. Our source code is available online and can be used by other biologists and computer scientists for further applications and experimentation. The distinguishing feature of our framework is the ability to deductively reason about biological systems using both Zsyntax and reaction kinetics. The soundness of interactive theorem proving ensures the correct application of EVFs or the simplification process as there is no risk of human error. The involvement of computers in the formal reasoning process of the proposed approach makes it more scalable than the analysis presented in Boniolo et al.’s and Molina-Pena et al.’s paper, which is based on traditional paper-and-pencil based analysis technique. Another key benefit of the reported work is the fact that the assumptions of these formally verified theorems are guaranteed to be complete, due to the soundness of the underlying analysis methods, and thus enables us to get a deep understanding about the conditions and constraints under which a Zsyntax and reaction kinetics based analysis is performed. Also, we have verified generic theorems with universally quantified variables and thus the analysis covers all possibilities. Similarly, in the case of reaction kinetics based analysis, the theorems have been verified for arbitrary values of parameters, such as k<sub>1</sub> and k<sub>2</sub>, which is not possible in the case of simulation where these expressions are tested for few samples of such parameters. A major limitation of higher-order logic theorem proving is the manual guidance required in the formal reasoning process. But we have tried to facilitate this process by formally verifying frequently used results, such as, simplification of vector summation manipulation and verification of flux vectors and stoichiometric matrices for each of the reaction schemes, and providing automation where possible. For example, we have developed two simplifiers, namely `Z_SYNTAX_SIMP` and `KINETIC_SIMP`, that have been found to be very efficient in automatically simplifying most of the Zsyntax or reaction kinetic related proof goals, respectively. In the first case study, the simplifier `Z_SYNTAX_SIMP` allowed us to automatically verify the theorem representing the reaction of the pathway leading to TP53 Phosphorylation. Similarly, in the second case study, i.e., time evolution verification of the tumor growth model, the simplifier `KINETIC_SIMP` significantly reduced the manual interaction and the proof concluded using this simplifier and some straightforward arithmetic reasoning. These simplifiers are also used to automate the verification process of the third case study, i.e., the automatic verification of the theorem representing the reaction of the pathway leading to the death of CSC and a significant simplification of the verification of the theorem representing the time evolution for the growth of the tumor cell. In future, we plan to conduct the sensitivity and steady state analysis of biological networks that is mainly based on reaction kinetics. We also plan to integrate Laplace and Fourier transforms formalization in our framework that can assist in finding analytical solutions of the complicated ODEs. # Supporting information [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: OH US. Formal analysis: AR US. Funding acquisition: OH ST. Investigation: AR. Methodology: OH US AR. Project administration: ST. Software: AR. Supervision: OH ST. Validation: AR US. Visualization: AR. Writing – original draft: AR. Writing – review & editing: OH ST.
# Introduction Acinetobacter baumannii is a Gram-negative coccobacillus pathogen linked to severe nosocomial infections including pneumonia, bacteremia, urinary tract infections and necrotizing fasciitis. A. baumannii infections have been commonly associated with immunocompromised patients; however, cases of community-acquired A. baumannii infections in healthy individuals have also been reported. Reports have also associated A. baumannii with wound infections acquired by combatants deployed to Iraq earning it the popularized name ‘Iraqibacter’. Treatment of A. baumannii infections is exceedingly difficult due to increasing multi-drug resistance and the limited understanding of its virulence factors, conditions that have a paramount impact on human health worldwide. While the mechanisms of antibiotic resistance associated with this emerging pathogen have been extensively studied, there is a troublesome paucity of literature reporting the molecular mechanisms of virulence associated with A. baumannii pathogenicity. Among the more understood properties that make A. baumannii a successful pathogen is its versatility in acquiring iron. The majority of iron in a host is intracellular; thus the availability of intracellular iron-containing molecules such as hemin, hemoglobin and ferritin is dependent on the lysis of host cells and their subsequent release due to cell and tissue damage found in wounds. The liberation of intracellular nutrients may be accomplished by bacterial-mediated cell damage such as that described in V. cholerae infections, in which hemolysin-based cytotoxicity lyses intestinal epithelial cells and erythrocytes releasing intracellular iron compounds into the extracellular environment for bacterial utilization. One avenue by which bacterial pathogens can lyse host cells is by producing phospholipases, which act on phospholipids in host membranes resulting in membrane destabilizing products thereby leading to cytolysis and the release of host intracellular contents. The A. baumannii ATCC 19606<sup>T</sup> strain genome contains genes encoding proteins harboring phospholipase domains including four with a patatin- like protein (PLP) phospholipase domain, one outer membrane protein with a phospholipase A1 domain and two with a phospholipase C domain (<http://www.broadinstitute.org/>). A more recent report showed that the genome of this strain also includes three genes the products of which are proteins that harbor PLD domains. These phospholipases differ in the types of reactions they catalyze; PLP phospholipases are non-specific acyl lipid hydrolases that cleave the acyl ester bond of a phospholipid, phospholipase A1 specifically cleaves phospholipids through the hydrolysis of the fatty acyl ester bond at the sn-1 position of the glycerol moiety, and phospholipase C and phospholipase D cleave before and after the phosphate, respectively. Patatins are plant storage glycoproteins with lipid acyl hydrolase activity that account for 30–40% of the total soluble proteins in potatoes. The first, and one of the few, PLPs to be characterized in bacteria was the ExoU protein from Pseudomonas aeruginosa, which was shown to have phospholipase activity. While bacterial PLPs have not been linked to cytolysis, their presence in the genomes of animal and plant pathogens/symbionts is significantly higher than in the genomes of non-pathogens. The bacterial phospholipase A1 (PhlA) from Serratia marcescens has been implicated in hemolysis of human erythrocytes and cytotoxicity to cervical cancer HeLa and 5637 human bladder epithelial cells. The phospholipase C of Clostridium perfringens, which is also known as the α toxin, causes cytolysis, tissue destruction and necrosis. The phospholipase C produced by P. aeruginosa has been linked to hemolysis, tissue destruction and pathologies reminiscent of burn infections. Purified phospholipase D, such as that produced by Corynebacterium pseudotuberculosis, is dermonecrotic and fatal when injected into animals. While many of the phospholipases encoded within the A. baumannii ATCC 19606<sup>T</sup> genome have possible implications in cytolysis and the ultimate release of iron-rich intracellular contents, the roles of only a few of these phospholipases have been elucidated in this pathogen. Specifically, the role of a phospholipase C and a phospholipase D has been associated with cytolytic activity to the FaDu hypopharyngeal carcinoma epithelial cell line and serum survival and invasion into both human bronchial epithelial BEAS-2B cells and HeLa cells, respectively. A recent report showed that three phospholipase D proteins play a critical role in the pathobiology of the ATCC 19606<sup>T</sup> strain. Taken together, these observations indicate that bacterial pathogens can gain access to additional intracellular iron pools and other nutrients present in erythrocytes and tissues through the expression of hemolytic/cytolytic activities. Supernatants of A. baumannii cultures grown under iron-chelation are hemolytic to horse erythrocytes and possess phospholipase C activity. Our report describes the characterization of plc1 and plc2 and the involvement of the protein products of these two genes in the hemolytic, cytolytic and virulence phenotypes displayed by the A. baumannii ATCC 19606<sup>T</sup> strain and isogenic derivatives affected in the expression of these two genes. # Materials and Methods ## Bacterial strains, plasmids, media and culture conditions The bacterial strains and plasmids used in this work are listed in. All bacterial strains were routinely stored as Luria-Bertani (LB) broth/glycerol stocks at -80°C. Escherichia coli DH5α recombinant clones were cultured in LB broth or on LB agar (LBA) supplemented with appropriate antibiotics and incubated overnight (12–14 h) at 37°C. A. baumannii strains as well as the E. coli MG1655 strain were subcultured from LBA into Chelex 100-treated trypticase soy broth dialysate (TSBD) and grown for 24 h at 37°C with shaking at 200 rpm. These cultures were then used to inoculate fresh TSBD or TSBD containing 10% erythrocytes at a 1/100 ratio and grown for 24 h at 37°C with shaking at 200 rpm unless otherwise indicated. Bacterial cells were enumerated after 24 h using flow cytometry. Culture medium supplemented with erythrocytes was prepared by centrifuging whole blood at 1,000 x g, resuspending and washing the erythrocyte pellet three times in erythrocyte wash buffer (20 mM KH<sub>2</sub>PO<sub>4</sub>, 60 mM Na<sub>2</sub>HPO<sub>4</sub> and 120 mM NaCl, pH 8.0) and then resuspending the pellet to a final erythrocyte concentration of 10% in TSBD. Iron-repleted culture conditions were accomplished through the addition of 50 μM FeCl<sub>3</sub> dissolved in 0.01 N HCl, while iron-chelated conditions were generated by treating TSB with Chelex 100 (Bio-Rad Laboratories). Bacterial growth curves were determined in octuplet using 96-well microtiter plates containing TSBD under the aforementioned culturing conditions over a 24-h time period. OD<sub>600</sub> values of these cultures were recorded hourly. Defibrinated sheep and horse erythrocytes were obtained from Cleveland Scientific, Ltd., and sodium citrate-treated whole human blood was purchased from Bioreclamation, LLC. A549 human alveolar epithelial cells were passaged three times in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% heat-inactivated fetal bovine serum, 100 IU penicillin, and 100 μg/ml streptomycin at 37°C in the presence of 5% CO<sub>2</sub>. Approximately 1x10<sup>5</sup> A549 cells, as enumerated using a hemocytometer, were seeded into each well of a 96-well, white, opaque tissue culture plate for a fourth passage that was incubated at 37°C in the presence of 5% CO<sub>2</sub> for 24 h without antibiotics. A549 cell monolayers were infected with 10<sup>6</sup> bacteria suspended in DMEM and incubated at 37°C in the presence of 5% CO<sub>2</sub> for 24 h. The cell monolayers were washed three times with DMEM prior to performing cytolysis assays. ## General DNA procedures Total genomic DNA was isolated using an adapted mini-scale procedure from a previously published method, and plasmid DNA was isolated using commercial kits (Qiagen). Restriction digests were performed as suggested by the supplier (New England Biolabs) and size-fractionated by agarose gel electrophoresis. PCR primer pairs 3824/3826 and 3822/3827, which hybridize internally of plc1 and plc2, respectively, were used to confirm the presence of these two genes in 19 A. baumannii clinical isolates. ## Sequence acquisition and phylogenetic analyses Nucleotide sequences were analyzed with DNASTAR (DNASTAR, Inc.), BLAST and data available from the Broad Institute. In silico identification of potential ferric uptake repressor (Fur)-binding sites upstream of plc1 and plc2 was performed using a small training set of predicted Fur-binding sites, which were analyzed using MEME Suite 4.10.0. GLAM2 PSSM was used to generate a WebLogo representing the A. baumannii Fur-binding motif, and MAST was used to search nucleotide sequences upstream of plc1 and plc2. The PLC1 and PLC2 amino acid sequences from A. baumannii ATCC 19606<sup>T</sup> were used as queries for BLASTp to obtain similar sequences, excluding additional amino acid sequences from A. baumannii. A total of 101 amino acid sequences were retrieved by setting an arbitrary cutoff of approximately 40% amino acid identity for phylogenetic comparisons to PLC1, PLC2, and the hemolytic (PLCH) and non- hemolytic (PLCN) P. aeruginosa phospholipase C proteins (GI 489205171 and GI 489204069), which were added manually. These sequences, which were aligned with MUSCLE using default settings, were analyzed with the multiple sequence alignment (MSA) MEGA6 software package. The MSA analysis was done using an apparent-maximum-likelihood method encompassing the WAG (Whelan Goldman) model with a discrete Gamma distribution and rate calculations among invariable sites with FastTree. The phylogenetic tree, which was drawn to scale with the highest log likelihood (-714043.274), represents 109 residues analyzed with 1,099 positions in the final dataset. ## Site-directed insertional mutagenesis of plc genes The 2.2-kb plc1 gene (A1S_0043 following the A. baumannii ATCC 17978 genome annotation) was amplified from ATCC 19606<sup>T</sup> genomic DNA with Taq DNA polymerase and primers 3824 and 3825. The plc1 amplicon was cloned into pCR8/GW/TOPO generating pMU1042. Inverse PCR was performed using pMU1042 as a template, Phusion DNA polymerase (New England Biolabs) and primers 4017 and 4018, which hybridize within the plc1 gene. The aph-FRT cassette, which codes for kanamycin resistance (Km<sup>R</sup>), was amplified from pKD13 with Phusion DNA polymerase and primers 4003 and 4004 and ligated within plc1 using the inverse PCR amplicon described above to generate pMU1089. The ermAM cassette, which codes for erythromycin resistance (Em<sup>R</sup>), was amplified from pIL252 with Phusion DNA polymerase and primers 4046 and 4047 and ligated within the plc1 coding region using the inverse PCR amplicon described above to generate pMU1101. Phusion DNA polymerase and primers 3824 and 3825 were used to amplify plc1::aph-FRT and plc1:: ermAM, which were each subcloned into the SmaI site of pEX100T to generate pMU1091 and pMU1108, respectively. The 2.2-kb plc2 gene (A1S_2055 following the A. baumannii ATCC 17978 genome annotation) was PCR amplified from ATCC 19606<sup>T</sup> genomic DNA using Taq DNA polymerase and primers 3822 and 3823 ( and), and the resulting amplicon was ligated into pCR8/GW/TOPO to generate pMU1039. Phusion DNA polymerase and primers 3171 and 3172 were used to amplify the aph cassette from pUC4K, which was inserted into the unique NsiI site of the plc2 gene after end repair of NsiI-digested pMU1039 with the End-It Kit (Epicentre) resulting in pMU1040. Phusion DNA polymerase and primers 3822 and 3823 were used to amplify plc2::aph, which was subsequently cloned into the SmaI site of pEX100T to generate pMU1076. Electrocompetent ATCC 19606<sup>T</sup> cells were electroporated with pMU1091 and pMU1076 as described before to generate the 3452 plc1::aph-FRT and 3430 plc2::aph isogenic derivatives, respectively. For the generation of the 3494 plc1::ermAM/plc2::aph double insertion isogenic derivative, electrocompetent 3430 cells were electroporated with pMU1108. The ATCC 19606<sup>T</sup> 3430 and 3452 isogenic derivatives were selected on LBA plates containing 40 μg/ml kanamycin, while the 3494 derivative was selected on LBA supplemented with 40 μg/ml erythromycin. All isogenic derivatives were plated on LBA plates supplemented with 10% sucrose to ensure loss of pMU1076, pMU1091 and pMU1108. Proper allelic exchanges were confirmed with PCR using external primers 3905 and 3906 (plc1) and 3815 and 3918 (plc2). ## Genetic complementation of the plc1 and plc2 isogenic derivatives The A. baumannii ATCC 19606<sup>T</sup> plc1 and plc2 genes were PCR amplified from ATCC 19606<sup>T</sup> genomic DNA using Phusion DNA polymerase and primers 3894 and 3895 (plc1) and 3892 and 3893 (plc2), all of which included BamHI restriction sites. The respective plc1 and plc2 amplicons were ligated into pCR-Blunt resulting in pMU1073 and pMU1074. Both pMU1073 and pMU1074 were digested with BamHI and the plc1 and plc2 fragments were each subcloned into the cognate BamHI site of the E. coli-A. baumannii shuttle vector pWH1266 generating pMU1079 and pMU1080. Electrocompetent 3452 (plc1::aph-FRT) and 3430 (plc2::aph) cells were electroporated with the empty shuttle vector pWH1266 resulting in the 3452.E and 3430.E transformants, respectively, which were selected on LB agar supplemented with 1 mg/ml of ampicillin. Electrocompetent 3452 (plc1::aph-FRT) and 3430 (plc2::aph) cells were also electroporated with pMU1079 and pMU1080, respectively, and the 3452 and 3430 complemented derivatives (3452.C and 3430.C) were selected on LB agar supplemented with 1 mg/ml of ampicillin. The presence of pMU1079 and pMU1080 was confirmed by restriction analysis of plasmid DNA recovered from transformants grown in LB containing 150 μg/ml of ampicillin. ## Transcriptional analyses Bacterial strains were each grown as five independent 1-ml cultures for 24 h in TSBD or TSBD supplemented with 50 μM FeCl<sub>3</sub> at 37°C with shaking at 200 rpm. The 24-h time point was chosen for RNA isolation because there was not apparent hemolytic activity until this time. RNA isolation, cDNA synthesis and qRT-PCR analyses were performed as previously described. Briefly, bacterial cells were lysed in lysis buffer \[0.3 M sodium acetate (pH 4.0), 30 mM EDTA and 3% SDS\] previous to RNA purification following the manufacturer’s protocol included with the Maxwell 16 LEV simplyRNA Tissue Kit (Promega). Total RNA concentrations and the OD<sub>260/280</sub> ratios of each RNA sample were assessed using a NanoDrop 2000 UV-Vis spectrophotometer (Thermo Fisher Scientific). RNA integrity was assessed using a RNA 6000 NanoKit for the Bioanalyzer 2100 (Agilent Technologies) and the manufacturers’ protocols. Only RNA samples with OD<sub>260/280</sub> ratios \> 1.7 and RNA integrity numbers (RINs) \> 5 were further processed for qRT-PCR analysis. The iScript cDNA synthesis kit (Bio-Rad Laboratories) was utilized for cDNA synthesis from 100 ng of total RNA template following the manufacturer’s protocol, and iQ SYBR Green (Bio-Rad Laboratories) was used to examine gene transcription following the manufacturer’s recommendations. The 10-μl reaction mix included 0.4 μl cDNA, 5 μl iQ SYBR-Green supermix and 300 nM of forward and reverse primers. Primers 3966 and 3967 were used to amplify a 179-bp internal fragment of the 16S ribosomal RNA gene, which served as an internal control of gene expression, while primers 3970 and 3971 were used to amplify a 156-bp internal fragment of bauA, which served as a positive control for gene expression under iron chelation. Primers 3972 and 3973 or 3974 and 3975 were used to amplify a 197-bp internal fragment of plc1 and a 181-bp internal fragment of plc2, respectively. The cycling conditions for qRT-PCR assays, which were performed on a Bio-Rad CFX Connect real-time PCR detection system, were as follows: 95°C for 3 min followed by 40 cycles of 95°C for 10 s and 60°C for 45 s. Relative expression of plc1 and plc2 between iron-chelated and iron-repleted conditions was quantified by the standard curve method in which serial dilutions of cDNA samples served as standards. Samples containing no cDNA template were used as negative controls. qPCR efficiencies were as follows: 16S, 82.4%; bauA, 100.9%; plc1, 106.7%; and plc2, 97.7%. All samples were analyzed in triplicate and melting curve data were included in the analysis to confirm primers specificity. Data analysis also showed that there were no significant differences in 16S expression between iron-chelated and iron-repleted conditions. Therefore, the expression of plc1 and plc2 was normalized to that of the 16S ribosomal RNA gene. ## Cell-type phospholipid content The Bligh-Dyer method was utilized to extract total lipids from sheep, human, and horse erythrocytes, A549 pneumocytes, and G. mellonella. Erythrocytes were extracted by the addition of 2:0.8:2 parts of methanol, water and chloroform. A549 cells were processed with 2:1.8:2 parts of methanol, water and chloroform. G. mellonella larvae were homogenized and extracted with 2:0.8:2 parts methanol, water, and chloroform. The lipid-containing chloroform fraction was analyzed by normal-phase HPLC. Phosphatidylcholine and phosphatidylethanolamine levels in the samples were calculated using an Ultimate 3000 high performance liquid chromatography system (Thermo, Germering, Germany) coupled with a Corona charged aerosol detector instrument (Thermo, Chelmsford, MA, USA). Chromatographic separation was carried out using a Phenomenox Silica (150 x 4.6 mm; 3μ) column at a flow rate of 1 ml/min. The mobile phase consisted of (A) butyl acetate/methanol (4:1) and (B) butyl acetate/methanol/water (1:3:1) with the following gradient elution: 0%-100% B at 0–15 min, 100% B at 15–17 min, 100%-0% B at 17–21 min, and then equilibrated with 100% A for 4 min. The column temperature was set at 50°C and the injection volume was 10 μl. The nitrogen gas pressure and response range of the detector was set at 35 psi and 500 pA, respectively. The Chromeleon 6.8 software was used for data processing. Identification of isolated compounds was based on retention times of authentic standards. ## Phospholipase C and cytolysis assays The presence of phosphatidylcholine-specific phospholipase C activity in A. baumannii culture supernatants obtained after centrifugation at 15,000 x g for 30 min was tested with the Amplex Red PC-PLC assay kit (Molecular Probes) using lecithin as a substrate and following the conditions suggested by the manufacturer’s protocol. Erythrocytes incubated in the presence of bacteria were diluted 1:1000 into filter-sterilized FACSFlow sheath fluid (BD Biosciences) for differential interference microscopy (DIC) and enumeration using flow cytometry. Erythrocyte morphological changes were observed in these samples using DIC microscopy on a Zeiss 710 Laser Scanning Confocal System (Carl Zeiss Microscopy GmbH). The number of erythrocytes present in each analyzed sample was quantified using a FACScan flow cytometer (BD Biosciences). Flow Cytometry Absolute Count Standard beads (Bangs Laboratories, Inc.) were added 1:40 to diluted samples to standardize volumes amongst 5-second samplings. Erythrocyte populations were gated using the forward and side scatter channels. The relative number of A549 cells remaining in cell culture following incubation with A. baumannii strains was assessed using the CellTiter-Glo luminescent cell viability assay (Promega) following the manufacturer’s instructions. Briefly, the number of A549 cells remaining after incubation with bacteria was assessed by measuring the luminescence resulting from the reaction of the provided Ultra-Glo recombinant luciferase with ATP released from metabolically active A549 cells. The relative luminescence units (RLUs) produced, and thus the relative number of viable A549 cells remaining following infection, was quantified using a FilterMax F5 microplate reader (Beckman Coulter) and reported as a ratio of RLUs produced following lysis of infected A549 cells versus the RLUs produced following lysis of uninfected A549 cells. ## Galleria mellonella virulence assays Virulence assays were conducted using the G. mellonella model as previously described. Briefly, assays were performed by injecting in triplicate 10 randomly selected healthy final-instar G. mellonella larvae (n = 30) injected with 10<sup>5</sup> CFUs/larva (± 0.5 log) of the ATCC 19606<sup>T</sup> or its isogenic derivatives suspended in sterile phosphate-buffered saline (PBS). Non- injected larvae or larvae injected with five microliters of sterile PBS were included as controls. After injection, larvae were incubated in darkness at 37°C, and the numbers of dead larvae were assessed at 24-hour intervals over 5 days with removal of dead larvae at times of inspection. Trials were repeated if more than two deaths were observed in any of the control groups. ## Statistical analyses The Student’s t-test or one-way analysis of variance (ANOVA), both provided as part of the GraphPad InStat software package (GraphPad Software, Inc.), were used to analyze the statistical significance of data, as appropriate for the data set. Means of experimental data were compared to the means of the respective control groups using the Tukey-Kramer multiple comparisons post-hoc test. Survival curves were plotted using the Kaplan-Meier method and analyzed for statistical significance using the log-rank test of survival curves (SAS Institute Inc.). Significances for all data analyses were set a priori at P ≤ 0.05. # Results ## Membrane lipid preference for A. baumannii hemolytic activity DIC microscopy of horse erythrocytes incubated in the presence of ATCC 19606<sup>T</sup> cells shows that the presence of this strain significantly decreases the number of intact red blood cells remaining in TSBD culture after incubation at 37°C for 24 h (panels A and B). In addition, the horse erythrocytes showed morphological changes characteristic of cell membrane damage following incubation with ATCC 19606<sup>T</sup>. In contrast, the number and morphology of sheep erythrocytes did not change after co-incubation under the same conditions. These data prompted us to quantitatively determine the number of sheep, horse or human erythrocytes remaining as well as the number of bacterial cells present after 24-h co-incubations in TSBD medium. Flow cytometry analyses of samples obtained from A. baumannii ATCC 19606<sup>T</sup>, LUH 13000 or AYE cultures containing sheep erythrocytes showed that AYE was the only strain, of the three tested strains, to be significantly hemolytic (P \< 0.05) to sheep erythrocytes, as compared to E. coli MG1655, which was used as a hemolysis-negative control. A comparison of the mean amounts of sheep erythrocytes remaining after 24-h incubations with ATCC 19606<sup>T</sup>, LUH 13000 or AYE under iron-chelation demonstrated a 5%, 3% and 17% reduction in sheep erythrocytes, as compared to the E. coli MG1655 hemolysis-negative control, respectively. In contrast, flow cytometry analyses showed that ATCC 19606<sup>T</sup>, LUH 13000 and AYE were hemolytic to human erythrocytes as demonstrated by 41%, 23% and 41% reductions in the numbers of intact human erythrocytes, respectively, with only the hemolysis caused by ATCC 19606<sup>T</sup> and AYE being statistically different from the hemolysis- negative E. coli control (P \< 0.001). All three A. baumannii strains were significantly hemolytic to horse erythrocytes (P \< 0.001) with the percentage reduction of intact horse erythrocytes ranging from 95% after incubation with ATCC 19606<sup>T</sup> or LUH 13000 to 98% after incubation with AYE. Since A. baumannii was the most hemolytic to horse erythrocytes when grown in iron-chelated TSBD, horse erythrocytes were chosen to test the effects of iron-repletion on A. baumannii mediated hemolysis. The data demonstrated a significant decrease in the hemolytic activity of ATCC 19606<sup>T</sup> (P \< 0.001), LUH 13000 (P \< 0.01) and AYE (P \< 0.01) when TSBD was repleted with free inorganic iron. Together, flow cytometry analyses of the hemolytic activity of the three tested strains indicate that A. baumannii is poorly hemolytic to sheep erythrocytes, intermediately hemolytic to human erythrocytes and almost completely hemolytic to horse erythrocytes. This increasing hemolytic activity showed a direct correlation with 0.27, 0.99 and 3.43 phosphatidylcholine/phosphatidylethanolamine ratios for sheep, human and horse erythrocytes, respectively. Taken together, the data show that the extracellular iron concentration as well as the erythrocyte phosphatidylcholine content are critical factors in the detection of A. baumannii hemolytic activity. ## A. baumannii harbors two phospholipase C genes and produces phosphatidylcholine-specific phospholipase activity The direct correlation between the amount of phosphatidylcholine in the erythrocyte membrane shown above and the extent of hemolysis after incubation with A. baumannii suggests the potential role of a phosphatidylcholine- specific phospholipase C as the hemolytic effector. Analysis of the ATCC 19606<sup>T</sup> genome available through the Broad Institute website (<http://www.broadinstitute.org/>) showed that this strain has two genes predicted to code for phosphocholine-specific phospholipase C enzymes. One of them (annotated as HMPREF0010_03297 and referred to as plc1) has a 2169-nt open reading frame (ORF) coding for a potential 722-amino acid protein, which is located downstream of a gene transcribed in the same direction and predicted to code for an RNase PH. A gene coding for a putative nicotinate-nucleotide diphosphorylase is located downstream of plc1 and transcribed in the opposite direction. The plc1 gene corresponds to the plc ortholog reported as A1S_0043 in ATCC 17978, the expression of which is enhanced by 2.5-fold when bacteria are cultured in the presence of ethanol. The other ATCC 19606<sup>T</sup> phosphocholine-specific phospholipase C gene (annotated as HMPREF0010_00294 and referred to as plc2) encompasses a predicted 2229-nt ORF coding for a 742- amino acid protein. This gene corresponds to the ATCC 17978 A1S_2055 gene identified by Camarena et al.. A 48-nt intergenic region containing an inverted repeat resembling a Rho-independent transcription termination sequence follows plc2 and separates this coding region from a potential bicistronic operon, containing a thioesterase and a lactaldehyde reductase coding region, which is transcribed in the opposite direction of plc2. Based on ATCC 17978 genomic data, a 480-nt intergenic region separates plc2 from a predicted gene transcribed in the same direction and coding for the DNA polymerase III tau and gamma subunits. These observations indicate that the ATCC 19606<sup>T</sup> plc1 and plc2 are coded for by monocistronic operons as it was reported for ATCC 17978. Other A. baumannii genomes including AB0057, ACICU, ATCC 17978 and AYE show similar gene arrangements for the chromosomal regions harboring plc1 and plc2, an observation that suggests the conservation of this genomic region across different A. baumannii isolates. Currently, it is unknown if other nosocomial A. baumannii strains that have yet to be sequenced possess plc1 and plc2; therefore, the presence of these genes in additional A. baumannii strains that had not yet been sequenced was tested by PCR using total genomic DNA and primers which hybridize within the respective coding regions of both phospholipase C genes. Amplicons of the predicted sizes, 993 bp for plc1 and 1,167 bp for plc2, were obtained for all 19 tested isolates confirming the presence of these genes in a variety of A. baumannii strains. This screening study was further complemented by testing the expression of hemolytic, PC-PLC and cytolytic activities in the same 19 A. baumannii strains. Hemolysis assays using horse erythrocytes showed that all these strains had significant (P \< 0.001) hemolytic activity as a compared to the negative control. The Amplex Red PC-PLC tests showed that the PC-PLC activity of TSBD culture supernatants of all 19 strains was significantly higher (P \< 0.001) than the negative control. Finally, the CellTiter-Glo luminescent cell viability assays showed that all strains displayed cytolytic activity against A549 human alveolar epithelial cells. Interestingly, HPLC analysis of lipids extracted from these cells showed a phosphatidylcholine/phosphatidylethanolamine of 2.58, which is significantly higher than the 0.27 and 0.99 ratios detected in sheep and human erythrocytes, respectively, but lower than the 3.43 value detected in horse red blood cells. It is of note that the tested strains displayed significant variations in hemolytic activity, with the ATCC 19606<sup>T</sup> and AB3340 isolates producing the lowest and highest activities, respectively; PC-PLC activity, with the AB3560 and AB3806 isolates producing the lowest and highest activities, respectively; as well as cytolytic activity, with the AB4498 and AYE isolates being the most and least cytotoxic strains, respectively. All these data also indicate that there is no correlation between the hemolytic, PC-PLC and A549 cytolytic activities expressed by the tested strains. For example, strains ATCC 19606<sup>T</sup> and ATCC 17978, which are considered non-contemporaneous and antibiotic sensitive clinical isolates displayed less hemolytic activity than most of the tested strains, several of which are modern multidrug resistant isolates. However, the PC-PLC and cytolytic activities of ATCC 19606<sup>T</sup> and ATCC 17978 are comparable to that of most of the remaining tested strains. The comparative analysis of the AB3340-AB5197 strains, which were isolated from different infection sites of wounded soldiers and include the AB5075 strain considered a highly virulent strain, show that although there are significant variations in their hemolytic activity, PC-PLC and cytolytic activities among them are comparable to those detected in the non- military strains shown in. Finally, the comparative analysis of the ATCC 19606<sup>T</sup>, ATCC 17978 and the AB5075 isolates, with the two ATCC strains being considered less virulent than the latter military isolate when tested using the G. mellonella model, shows that although there is a correlation between their relative virulence and cognate hemolytic phenotype, such a correlation does not exist when their PC-PLC and cytolytic activities are compared. Taken together, these observations indicate that although PC-PLC seems to play a role in virulence, it is most likely that this is one of several bacterial factors responsible for the virulence of this pathogen. Preliminary assays showed that phospholipase activity was detected only when ATCC 19606<sup>T</sup> bacteria were cultured in TSBD, a liquid medium that was dialyzed against Chelex 100, an insoluble polymer that binds several metals, including iron. This observation suggested that the expression of the genes responsible for the production of phospholipase activity could be iron- regulated. Quantitative RT-PCR analyses using total RNA extracted from ATCC 19606<sup>T</sup> cells grown in iron-chelated or iron-repleted TSBD showed that transcription of plc1 and plc2 is indeed significantly higher (P \< 0.01 and P \< 0.05, respectively) in bacteria cultured under iron-depleted conditions when compared with TSBD supplemented with inorganic iron. The approximately 0.94-fold and 0.64-fold increase in the transcription of plc1 and plc2 when ATCC 19606<sup>T</sup> was grown in iron-chelated TSBD vs. iron-repleted TSBD may seem modest; however, since the product of plc1 and plc2 are enzymatic in nature these changes could have pronounced biological significance. These analyses also showed that the iron-regulated expression of plc1 and plc2 is similar to that of bauA, which is only increased about 0.82 fold in iron-chelated TSBD vs. iron-repleted TSBD and codes for the production of the BauA acinetobactin outer membrane receptor protein, which is the product of a proven iron-regulated gene. The observation that the addition of FeCl<sub>3</sub> to TSBD reduced the transcription of plc1 and plc2 to levels similar to those detected for bauA strongly indicate that iron indeed plays a critical regulatory role in the differential transcription of these genes. It appears likely that the iron-regulated expression of plc1 and plc2 is due to the presence of putative Fur-binding sites, which were located approximately 100 nt and 200 nt upstream of plc1 and plc2, respectively, and found to be significantly related (e-value \> 0.05) to the A. baumannii Fur motif recently reported. This possibility is strongly supported by the observation that hemolysin and siderophore production in V. cholerae is co-regulated by iron through a Fur-dependent transcriptional regulatory process. The possibility of compensatory expression between plc1 and plc2 was examined using ATCC 19606<sup>T</sup> isogenic derivatives harboring the appropriate plc mutation. Quantitative RT-PCR analyses of plc1 transcription in 3430 plc2::aph cells or plc2 transcription in 3452 plc1::aph-FRT cells, all grown in TSBD, showed that there is not a regulatory mechanism by which plc1 transcription compensates for the lack of plc2 transcription or vice versa. Taken together, these results show that the genomes of multiple A. baumannii strains contain two phosphatidylcholine-specific phospholipase C genes, the presence of which could correlate with their capacity to express hemolytic activity, preferentially toward human and horse erythrocytes. Furthermore, the production of this activity depends on the effect of free iron on the differential transcription of plc1 and plc2, which are expressed independently of each other at higher rates under iron limiting conditions. ## The A. baumannii PLC1 and PLC2 have diverged from a common ancestor protein Both PLC1 and PLC2 from ATCC 19606<sup>T</sup> cluster with phospholipase C proteins from other known pathogens. PLC1 is located within a clade that includes proteins produced by seven different Acinetobacter species with two of them, oleivorans and radioresistens not being commonly associated with human infections. Interestingly, two of these seven species have been reported as being hemolytic, A. beijerinckii sp. nov. and the A. calcoaceticus strain 1318/69 that was isolated from the urine of a 70-year old male patient. Our preliminary observations also indicate that A. nosocomialis M2 expresses hemolytic activity (data not shown). The hemolytic activity of A. nosocomialis M2 is not surprising since it is so closely related to A. baumannii that it was originally identified as A. baumannii M2. PLC2 groups with a more diverse clade that includes Achromobacter, Cupriavidus and Acinetobacter sequences, with A. gyllenbergii being reported to lyse both horse and sheep erythrocytes. Interestingly, PLC2 is also in the same clade as phospholipase C produced by bacteria belonging to different genera and species. Although most of these bacteria are non-pathogenic environmental microorganisms that share symbiotic relationships with invertebrates such as Verminephrobacter aporrectodeae, some of them have been isolated from human patients such as Massilia timonae, Bordetella hinzii and the Melioidosis agent Burkholderia pseudomallei. More distant branches from PLC1 and PLC2 contain clusters encompassing phospholipase C proteins from P. aeruginosa. The non-hemolytic PLCN of P. aeruginosa resides in a group containing Lysobacter antibioticus and Lysobacter capsici, which play roles in the rhizospheres of rice or peppers, respectively. The hemolytic PLCH of P. aeruginosa groups outside a cluster of environmental isolates with Burkholderia spp. reported to be involved with wound infections, bacteremia and hemolysis. ## Effect of plc interruption on phospholipase activity and cytolysis The role of the ATCC 19606<sup>T</sup> PLC1 and PLC2 proteins in the lysis of erythrocytes and human epithelial cells was tested using the 3452 (plc1::aph- FRT), 3430 (plc2::aph) and 3494 (plc1::ermAM/plc2::aph) isogenic insertion derivatives. Interruptions in one or both of these genes did not affect the growth of these derivatives; their growth kinetics were not statistically different from that of parental ATCC 19606<sup>T</sup> when cultured in TSBD under non-selective conditions. Flow cytometry analyses showed that the number of intact horse erythrocytes remaining after incubation with the 3430 or 3452 isogenic derivative is not significantly different from the number of erythrocytes remaining after incubation with the ATCC 19606<sup>T</sup> parental strain. However, the number of erythrocytes remaining following incubation with the 3494 isogenic derivative, which has interruptions in both plc1 and plc2, is more than 3-fold higher (P \< 0.001) than the number of erythrocytes remaining after incubation with either ATCC 19606<sup>T</sup> or the 3430 or 3452 isogenic derivatives. The decreased ability of the 3494 isogenic derivative to lyse erythrocytes correlates with that observed when these strains were incubated with A549 human alveolar epithelial cells. In these experiments, the number of A549 cells remaining after 24 h incubation in the presence of ATCC 19606<sup>T</sup>, 3430 or 3452 were not significantly different from one another; however, the number of remaining A549 cells following incubation with 3494 was significantly higher (P \< 0.001) than the number remaining after incubation with ATCC 19606<sup>T</sup>, 3430 or 3452. The role of plc1 and plc2 in the production of PC-PLC activity was further tested using the 3452.C (pMU1079) and 3430.C (pMU1080) derivatives, which harbor plasmid copies of the plc1 and plc2 parental alleles, respectively. As expected from data described above, the Amplex Red assays demonstrated that both 3452 (P \< 0.05) and 3430 (P \< 0.05) are significantly reduced in their ability to degrade phosphatidylcholine as compared to the parental ATCC 19606<sup>T</sup> strain. This figure also shows that the transformation of 3430 with pMU1080, which harbors the plc2 coding region under the control of the pWH1266 tetracycline resistance promoter, not only restored, but also significantly enhanced the production of PLC2 (P \< 0.01), with the latter effect being most likely due to a gene dosage effect. Unfortunately, several attempts to complement the 3452 mutant with the plc1 parental allele (3452.C) did not restore the phospholipase activity of this derivative possibly because of regulatory mechanisms due to the fact that pMU1079 was made by cloning only the plc1 coding region without any upstream sequences that could affect its expression in this recombinant derivative. The phospholipase activity of the 3452 and 3430 strains containing the empty A. baumannii-E. coli shuttle vector pWH1266 (3452.E and 3430.E, respectively) were not significantly different from that of the 3452 or 3430 strains demonstrating the pWH1266 shuttle vector used to clone the parental plc1 and plc2 alleles does not confer a phospholipase phenotype upon these strains. Taken together, these observations indicate that the activity of the phospholipase C proteins produced by ATCC 19606<sup>T</sup> are not host cell specific and have cytolytic activity against different cell types this pathogen could encounter during infection. ## Role of plc1 and plc2 in virulence The same strains used to test cytolytic activity were also used to examine the role of plc1 and plc2 in the virulence of A. baumannii ATCC 19606<sup>T</sup> with the G. mellonella experimental virulence model we have used previously to determine the virulence role of the acinetobactin- mediated iron acquisition system. shows that infection of caterpillars with ATCC 19606<sup>T</sup> resulted in a 47% mortality rate, which is significantly higher than the 16% rate scored (P \< 0.05) with animals that were not injected or injected with sterile PBS as negative controls. Interestingly, HPLC analysis of lipids extracted from larvae homogenates showed that the phosphatidylcholine/phosphatidylethanolamine ratio in this insect is 1.49, a value that is between the 0.99 and 2.58 values determined for human erythrocytes and A549 cells and much higher that the 0.27 ratio detected in sheep erythrocytes. There was also a significant difference in percent survival when larvae infected with ATCC 19606<sup>T</sup> were compared with larvae infected with either the plc1::aph-FRT 3452 or the plc1::ermAM/plc2::aph 3494 double mutant (13% mortality, P \< 0.01). Furthermore, the killing rates of these two mutants were not significantly different from each other as well as from the rates scored with animals that were not injected or injected with sterile PBS. In contrast, the death rates of caterpillars infected with the plc2::aph 3430 isogenic derivative, which actively expresses plc1, were very similar to those recorded after infection with the ATCC 19606<sup>T</sup> parental strain (43% vs. 47%). These observations together with those collected with the cytolytic assays described above indicate that while PLC1 and PLC2 seem to play similar roles in lysing different host cells when tested either under laboratory or ex vivo conditions, PLC1 appears to play a more critical role during the infection of a host that mounts an innate immune response that resembles that of vertebrate animals. # Discussion Acinetobacter baumannii has been generally considered a non-hemolytic pathogen because, according to our observations, the detection of such an activity depends on two critical factors. One of these factors is the type of erythrocytes used in the detection tests, which are normally conducted using Columbia agar plates containing 5% sheep red blood cells. Our work demonstrates that A. baumannii is poorly hemolytic to sheep and increasingly hemolytic to human and horse erythrocytes. These observations, which resemble those previously reported for the strains ACICU, AYE, ATCC 17978 and SDF using sheep and horse erythrocytes, are in agreement with the increasing phosphatidylcholine content we detected in these erythrocytes. Furthermore, the positive correlation between the A. baumannii hemolytic activity and the erythrocyte phosphatidylcholine content is reminiscent of a phospholipase C homolog in P. aeruginosa, where a hemolytic phospholipase C (PLCH) acts exclusively on phosphatidylcholine and sphingomyelin. The effect of the co-incubation of A. baumannii with sensitive erythrocytes, such as those from horse, is apparent not only because of their lysis but also because of the generation of schistocytes due to significant cell membrane damage. These fragmented red blood cells have been detected in blood smears obtained from infected neonates during an A. baumannii infection outbreak in a Saudi Arabian hospital. In this case study of seven neonates, five had a total erythrocyte count lower than controls throughout the course of A. baumannii infection, and two of the seven neonates succumbed to A. baumannii bacteremia. These observations clearly underscore the potential role of hemolytic activity in the pathobiology of A. baumannii. The second factor that determines the detection as well as the expression of A. baumannii hemolytic activity is the iron content of the culture media. Our data demonstrate that this activity is detectable when bacteria are cultured in an iron-chelated medium but not when the medium is iron rich. Columbia agar, which is used in standard clinical bacteriology methods to detect hemolytic activity, is considered a rich-nutrient medium that would have to be treated with a chelating agent such as Chelex 100 to properly detect hemolytic activity. Accordingly, the differential expression of PC-PLC activity in response to iron chelation correlates well with the differential transcription of the plc1 and plc2 gene orthologs when Chelex 100-treated media is supplemented or not with free iron, through a regulatory process that is most likely controlled by the interaction of the Fur transcriptional repressor that could bind to predicted iron boxes located upstream of these two genes. The iron-regulated expression of the plc1 and plc2 genes and the corresponding PC-PLC activity suggests that A. baumannii uses PLC1 and PLC2 for iron acquisition through the lysis of host cells and the subsequent release of iron-rich cytoplasmic contents. This possibility is further supported by the fact that iron co-regulates the expression of plc1 and plc2 genes as well as the expression of genes involved in the acinetobactin-mediated iron acquisition system, a response that is similar to that described in V. cholerae where the production of hemolysin and vibriobactin are regulated by a Fur-mediated process. Whether A. baumannii acquires iron from intracellular pools via a siderophore-mediated system or the expression of uncharacterized hemin utilization processes remains to be tested experimentally. Unfortunately, the role of Fur in the expression of the A. baumannii plc genes cannot be tested using a Fur deficient derivative since our efforts as well as that of others indicate that Fur is an essential gene. This possibility is strongly supported by a recent report showing that random mutagenesis of A. baumannii did not result in the isolation of fur mutants, an observation that led the authors to classify this gene as an essential genetic element in this pathogen. All previous considerations and the observation that plc1 and plc2 are present in all sequenced A. baumannii genomes as well as the genomes of the clinical strains tested in this study, but absent in the non-pathogenic A. baylyi ADP1strain, strongly indicate a role of these phospholipases in bacterial virulence. Until this report, only plc1 and its role in cytolysis had been tested experimentally where an A. baumannii ATCC 17978 plc1::aph isogenic insertion derivative was less effective in damaging FaDu hypopharyngeal carcinoma epithelial cells as compared to the parental strain. Our data not only support the virulence role of A. baumannii PC-PLC as established by this previous report, but also indicate that the modest cytotoxic effect reported is most likely due to the fact that only the inactivation of both plc1 and plc2 results in a significant reduction of host cell damage. Our work also shows that the product of these two genes target different types of host cells A. baumannii could encounter during the pathogenesis of systemic infections, as well as infection of the digestive and respiratory systems as revealed by the damage this pathogen causes to erythrocytes, and the FaDu and A549 epithelial cell lines, respectively. Our data collected using laboratory and ex-vivo experimental conditions suggest that the plc1 and plc2 genes code for potentially redundant cytotoxic functions. Only a double plc1/plc2 mutant showed a significant reduction in cytolytic activity when tested using horse red blood cells and A549 human alveolar epithelial cells. However, the G. mellonella experimental infection model showed that this is not the case in an in vivo infection model. This model showed that PLC1 but not PLC2 is critical for the virulence of the ATCC 19606<sup>T</sup> strain. Interestingly, plc1 but not plc2 proved to be transcribed at higher rates when ATCC 17978 bacteria were cultured in the presence of ethanol, a condition that also increases the virulence and the expression of virulence-associated traits including biofilm biogenesis and bacterial surface motility. Based on all these observations, it is possible to speculate that PLC1 and PLC2 play different roles during the pathogenesis of A. baumannii infections, with PLC1 being coded for by a gene the expression of which appears to be regulated at least by iron and stress signals. These signals are critical for the virulence of this pathogen when tested in an invertebrate host that mounts a complex defense response that mimics that of the human host. It is also possible that the biological role of these two enzymes depends on the nature of potential targets, which may reflect significant differences in phospholipid and fatty acid composition between insect and mammalian cells. Interestingly, the phylogenetic analysis of PLC2 shows that this protein clusters with a phospholipase C protein from the invertebrate endosymbiote Verminephrobacter aporrectodeae. This finding could explain the lackluster role of PLC2 in virulence using the invertebrate G. mellonella virulence model due to a host adaptation process. The phylogenetic analysis also showed that although many of the amino acid sequences used to construct the phylogenetic tree shown in are from environmental microorganisms isolated from soil, aquatic environments, or industrial sites, it is apparent that there is also a strong correlation of the bacteria producing these enzymes with plants in either antagonistic or synergistic ways (e.g., pathogenesis of blight disease in some plants or nitrogen-fixing bacteria present in the rhizosphere). Taking into account our experimental data together with the observations published by other investigators using different clinical isolates \[, –\], it is apparent that A. baumannii produces two PC-PLC and three PLD phospholipases. In the particular case of the ATCC 19606<sup>T</sup> strain, the three PLD phospholipases are not essential for the utilization of phosphatidylcholine as a carbon and energy source. This finding may indicate that the two PC-PLC produced by this strain could be responsible for the utilization of phosphatidylcholine as a nutrient source by the triple PLD deficient derivative. It is also apparent that the ATCC 19606<sup>T</sup> PC-PLC and PLD enzymes play a virulence role; although they may function differently during the infection process. The report by Stahl et al. shows that all single and double PLD mutants display a virulence phenotype indistinguishable from the parental ATCC 19606<sup>T</sup> strain and only the triple Δpld1-3 mutant showed a significant reduction in the killing rates throughout the course of the experiment when compared to the parental strain. However, the killing rate of this triple PLD mutant (74% for day 4 after infection) seems to be high if one considers data published by other investigators who have used the same experimental model and included in their Kaplan-Meier plots the control data collected with non-injected animals or animals injected with the same volume of sterile PBS, which unfortunately were not shown in this report. Furthermore, the lack of information regarding the CFUs injected per larva rather than OD<sub>600</sub> also impairs the proper comparison of our results with those recently reported using the same strain and comparable experimental conditions. Nevertheless, it is possible to speculate that there is a residual virulence activity in the triple Δpld1-3 ATCC 19606<sup>T</sup> mutant that could account for the activity of the PLC1 but not PLC2 as we describe in this report. This critical issue as well as the question of whether the A. baumannii PLD enzymes are differentially produced in response to extracellular signals and display selective cytolytic activity in response to differences in membrane lipid composition are critical topics that remain to be tested experimentally using the proper ATCC 19606<sup>T</sup> isogenic derivatives. Such knowledge will not only further our understanding of the role(s) of phospholipases in the pathobiology of A. baumannii, but also provide critical information needed to determine whether these enzymes could be used as alternative targets to treat the severe infections caused by this pathogen, particularly by emerging multi- drug resistant isolates. # Supporting Information This work was supported by funds from Department of Defense W81XWH-12-2-0035 grant, U.S. Public Health Service AI070174 award, and Miami University research funds. We thank Drs. D. Zurawski (Walter Reed Army Institute of Research, Silver Spring, Maryland, USA) and L. Dijkshoorn (Leiden University Medical Center, Department of Infectious Diseases, Leiden, The Netherlands) for providing the wound clinical isolates and the LUH and RUH strains, respectively, which are listed in. We are also grateful to Dr. E. Lafontaine (College of Veterinary Medicine, University of Georgia, USA) for providing the A549 cell line. The findings and opinions expressed herein belong to the authors and do not necessarily reflect the official views of the WRAIR, the U.S. Army, or the Department of Defense. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: SEF BAA LAA. Formal analysis: SEF BAA TT MKSJ RG LAA. Funding acquisition: LAA. Investigation: SEF BAA RES ACB TT MVC EJO MM MKSJ RG LAA. Methodology: SEF BAA RES ACB TT MVC MKSJ RG EJO MM. Resources: LAA. Supervision: SEF BAA LAA. Validation: SEF BAS RES ACB TT MVC RG EJO MM MKSJ LAA. Visualization: SEF BAA RES ACB TT MVC EJO MM MKSJ LAA. Writing – original draft: SEF BAA LAA. Writing – review & editing: SEF BAA RES ACB TT MVC RG EJO MM MKSJ LAA. [^3]: Current address: Department of Biological Sciences, Miami University Middletown, Middletown, Ohio, United States of America [^4]: Current address: Biology Department, Middle Tennessee State University, Murfreesboro, Tennessee, United States of America [^5]: Current address: Department of Pathology, Microbiology, and Immunology, Vanderbilt University, Nashville, Tennessee, United States of America [^6]: Current address: Department of Biological Sciences, University of Idaho, Moscow, Idaho, United States of America [^7]: Current address: Biology Department, Texas Christian University, Fort Worth, Texas, United States of America
# Introduction Many plant functions appear to cycle from low to high levels of activity on approximately a 24-hour schedule, a phenomenon recognized centuries ago. Virtually every physiological process in a plant undergoes at least one rhythmic cycle within a 24-hour period. Many of these diurnal cycles appear to be a response to external time cues, such as the light/dark or high/low temperature cycles that naturally arise from the alternation of day and night. These environmental time cues, termed Zeitgebers (German for “time givers”), tend to entrain the internal timing systems of plants to a period of about 24 h, corresponding to the period of the earth’s rotation. While some processes tend to reach peak activity in the middle of the subjective day or night, others reach peak activity at characteristic time points other than the middle. The alternating phases of the cycles are accompanied by extensive reengineering of the transcriptome, indicating a great many genes are primarily or secondarily influenced by the clock mechanisms. Cycles with a period of less than 24 h, or ultradian cycles, are becoming increasingly well-known in plants and are thought to have a number of functions including regulation of cell signaling, optimization of energy efficiency, optimization of responses to stimuli, spatial organization, temporal organization, the separation of incompatible processes occurring in the same subcellular compartment, and coordination of processes occurring in separate compartments. One school of thought holds that diurnal cycles arose through consolidation of ultradian cycles, presumably because of an evolutionary advantage associated with longer cycle times. Cycles with periods as short as a few minutes occur in plants as nutrients are taken up in the roots, and as part of carbon assimilation and partitioning during photosynthesis-related metabolism. Specifically, in relation to freezing tolerance in wheat (Triticum aestivum L.) plants, expression levels of many of the C-repeat binding factor (CBF) genes were observed to fluctuate in diurnal fashion in plants grown under a 16 hours light/8 hours dark photoperiod at 20°C. The CBF genes encode transcription factors that are intricately involved in the conditioning of the response to low temperature, as well as other stress factors. In Arabidopsis, CBF genes expression has been shown to cycle in diurnal fashion. Most of the 13 wheat CBF genes examined by Badawi et al. reached minimum expression at the end of an 8-hour dark period, then reached maximum expression 10–12 hours after the start of the light period. One of the CBF genes (TaCBFIIIa-D6) also showed a second, cyclic increase in expression at about the middle of the 8-hour dark period. At least 65 CBF genes are present in the wheat genome and involvement of these genes in both above- and below-freezing low temperature tolerance of wheat has been amply demonstrated. Campoli et al. observed that variation in expression levels of some of wheat, rye (Secale cereale L.), and barley (Hordeum vulgare L.) CBF genes were reduced during the dark phase of various photoperiod regimes, leading the authors to suggest their expression “may reflect a temperature-dependent, light-regulated diurnal response.” That observation, and the observations of Badawi et al. of diurnal cycles of expression of several wheat CBF genes suggests that low temperature tolerance may also show cyclic variation in expression. Others have reported that the photoperiod significantly affects the freezing tolerance of some wheat lines and references therein). The wheat genes CBF14 and CBF15 have become of particular interest to freezing tolerance investigations. These genes are part of a cluster of at least 11 CBF genes, collectively referred to as the Fr2 locus, found on the long arm of the group 5 homoeologous chromosomes of hexaploid wheat. While all of these CBF genes presumably contribute to freezing tolerance, CBF14 and CBF15 are expressed to the highest levels of all of the genes in the Fr2 locus on exposure to low temperature, perhaps indicating they condition a greater portion of the observed freezing tolerance. Copy number variation of CBF14 and CBF15 has been shown to be a significant factor influencing the effectiveness of these genes with greater numbers of copies associated with greater freezing tolerance. Each of these lines of evidence suggest that CBF14 and CBF15 play a major role in freezing tolerance. Indeed, Soltesz et al. demonstrated that genetically transforming these genes individually into freezing-sensitive spring barley plants resulted in highly significant improvement of freezing tolerance. Novak et al., found that light quality also strongly influences CBF14 expression and subsequently reported that “it can be assumed that temperature and light signals are relayed to the level of CBF14 expression via separate signalling routes”. Therefore, it appears that varying light conditions may influence the expression of the CBF genes, especially CBF 14 and CBF 15, even at constant temperature, implying that freezing tolerance also may vary over the course of 24 hours at constant temperature. Accordingly, the objective of this study was to determine whether fully cold-acclimated winter wheat plants grown under diurnal light/dark cycles were significantly more freezing tolerant at the mid-points, or at the end-points of the light and dark periods, and to assess the expression levels of CBF14 and CBF15 at these time points to look for a possible relationship of expression of these genes and freezing tolerance. # Materials and methods ## Freezing tolerance of wheat plants The winter wheat cultivars included in this study were ‘Eltan’, ‘Froid’, ‘Norstar’, ‘Bruehl’, ‘Centurk 78’, ‘Jagger’, ‘Lewjain’, ‘Madsen’, ‘Masami’, ‘Nugaines’, and ‘Tiber’, and germplasm line ‘Oregon Feed Wheat \#5’ (ORFW). These cultivars represent a range of freezing tolerance with Norstar the most freezing tolerant and ORFW the least. The freezing tolerance of most of these lines were described previously. The ORFW germplasm line appears to carry a major, dominant gene conditioning freezing sensitivity. Seeds were sown into Sunshine Mix LC1 planting medium (Sun Gro Horticulture, Bellevue, WA, USA) in 6-container packs (Model 1020, Blackmore Co., Belleville, MI, USA); the capacity of each container in the pack was about 100 ml. Each freezing tolerance trial was comprised of eight of the 6-container packs. Each of the 12 wheat lines were represented four times in each freezing trial. Each freezing trial was considered a replication and the four representations of each line were considered subsamples within trials. Seeds were germinated and seedlings grown at 22°C in a growth chamber (Model E15, Conviron, Pembina, ND) under cool, white fluorescent lights (about 300 µmol m<sup>-2</sup> s<sup>-1</sup> PAR at the soil surface) with a 12-hour photoperiod until the seedlings reached the three-leaf stage. Relative humidity was not controlled. The plants were then transferred to 4°C with a 12-hour photoperiod (about 250 µmol m<sup>-2</sup> s<sup>-1</sup> PAR at mid-plant height) for 5 weeks to induce cold acclimation prior to freezing survival tests. Lights were turned on at 6 a.m. and turned off at 6 p.m. Dawn is often defined as “ZT0” (Zeitgeber time zero;), thus, in this study, ZT0 refers to 6a.m. Plants were irrigated weekly with nutrient solution containing macro and micronutrients (Peters Professional, Scotts Co., Camarillo, CA, USA). Prior to freezing, the flats were drenched with a solution of 10mg/L Snomax snow inducer (Snomax LLC, International, Centennial, CO, USA) maintained at 4°C, and allowed to drain until drainage had essentially ceased, then freezing was carried out in darkness in a programmable freezer (model LU-113, Espec Corp., Hudsonville, MI, USA). The temperature of the plant growth medium in each container near the crowns of the plants was monitored using thermistor-based, food piercing temperature probes with expected accuracy of ±0.3°C, and an internet-enabled temperature monitor (Model E-16, Sensatronics, Bow, NH, USA). The temperature was recorded every 2 min using a data capture script running on a remote computer. Freezing tests were initiated at ZT0, ZT6, ZT12 and ZT18, i.e. at the start, midpoint, and end of the 12-hour light and dark periods. The trials initiated at ZT18 (middle of the dark period) were placed in the freezer at ZT12, at the start of the dark period, and the freezer was maintained at 4°C until the start of the freezing test at ZT18, thereby avoiding exposure to light before the start of the freezing trial. At the start of each freezing test, the temperature in the freezer was reduced from 4°C to -3°C and held at -3°C for 16 hours to allow all of the soil water to be converted to ice and the heat of ice formation to completely dissipate from the samples before the temperature was lowered to the target temperature at a rate of -2°C h<sup>-1</sup>. The temperature was held at the target for 2 hours, then raised to 4°C at a rate of 2°C h<sup>-1</sup>. Plants were exposed in separate trials to target temperatures of -13.5, -14.5, -15.5, or -16.5°C with freezing tests to each target temperature started at each of the four time points, ZT0, ZT6, ZT12 and ZT18. A total of 1,723 containers holding 37,236 plants were tested. After freezing, the plants were then transferred to a growth chamber at 4°C for 24 hours to allow the soil to completely thaw and the soil temperature to equilibrate, then were moved to 22°C with a 16-hour photoperiod for recovery. Survival was scored after 5 weeks of regrowth. Statistical analysis was carried out using the “Fit Model” platform of JMP version 12 (SAS Institute Inc., Cary, NC, USA). The proportions of plants surviving formed the response variable and wheat lines (genotypes), minimum temperature, freezing trial start time, replications, and subsamples within replications were the predictor variables. Means separation was determined using least significant difference with a significance level of 0.05. For statistical analysis of the main effects, the response was expressed as the arcsine of the square root of the surviving proportion as recommended but are expressed in the original scale in this report. For calculations of the LT<sub>50</sub>, the temperature expected to result in death of 50% of the plants, the response was expressed as the ratio of the number of plants surviving to the number of plants frozen in each trial, estimated with binomial distribution and probit link function specified. The minimum temperatures for the LT<sub>50</sub> calculations were specified as the minimum temperature recorded for each container, rounded to the nearest 0.5°C, then increased by 30°C to avoid negative numbers in the probit calculations. The LT<sub>50</sub> values are expressed in the original scale in this report. ## CBF gene expression analysis At each of the time points, ZT0, ZT6, ZT12 and ZT18, seedlings of the wheat lines Norstar, Tiber and ORFW that had been grown to 5 weeks of cold acclimation at constant 4° C under a 12-hour photoperiod as described above were quickly removed from the soil and plunged into liquid nitrogen. Three independent biological samples grown at different times were collected. Total RNA was extracted from the crown tissue (meristematic regions) using Trizol reagent and the supplier-recommended procedure (ThermoFisher Scientific, Waltham, MA, USA), and then treated with DNase I and further purified using the Direct-zol RNA MiniPrep kit (Zymo Research, Irvine, CA, USA). Purified RNA was quantified using a Nanodrop 1000C instrument (ThermoFisher Scientific, Waltham, MA, USA). Complementary DNA (cDNA) was synthesized using the ProtoScript First Strand cDNA Synthesis Kit (New England Biolabs, Ipswich, MA, USA) using 1µg RNA as starting template. The resulting cDNA was diluted 1:5 prior to use as template for quantitative real-time PCR. The qPCRs were conducted on an Applied Biosystems 7300 Real-Time PCR System (ThermoFisher Scientific, Waltham, MA, USA) using 25µl preparations that consisted of 1X Go-Tag Colorless Master Mix (Promega, Madison, WI, USA), 1µl cDNA, 200µM each primer, 0.85X SYBR Green I as a reporter fluor, and 300nM ROX dye (ThermoFisher Scientific, Waltham, MA, USA) as a passive, internal fluorescence standard. The PCR primers used for CBF14 were: forward, `5'-AACCGATGACGAGAAGGAAA-3'`, and reverse, `5'-AACTCCGAGTAGCACGATCC-3'`. The primers used for CBF15 were: forward, `5’-GTCGTCCATGGAAAATACCG-3’`, and reverse: `5’-ATGTGTCCAGGTCCATTTC-3’`. The PCR amplification profile was 3 min at 95°C, then 45 cycles of 95°C for 30 sec, 55°C for 25 sec, and 72°C for 30 sec. Two technical replications of each of the three biological replications were performed. The qPCR software available with the instrument was used to determine the Ct, the fractional cycle number at which the fluorescence intensity reached an arbitrary threshold, using the threshold determined by the software. Relative fold change was determined using the delta-delta Ct method, with Ct values normalized to the Ct values of house- keeping gene Ta30797, encoding phosphogluconate dehydrogenase, using the forward primer `5’-GCCGTGTCCATGCCAGTG-3’`, and reverse `5’-TTAGCCTGAACCACCTGTGC-3’`, as described by elsewhere. ## NMR spectra of plant extracts Wheat plants of the cultivars Norstar, ORFW and Tiber were grown and cold- acclimated for 5 wks at 4°C and 12-hour photoperiod as described above. After 5 wks, the plants (about 20 plants) were removed from the cell packs at each of the ZT0, 6, 12, or 18 time points, quickly rinsed in ice water, then dropped into liquid nitrogen. Each combination of wheat cultivar and time of day was represented by two biological replications grown and processed at separate times. Roots, residual caryopses, and shoots above the first node were removed while the plant tissue was frozen, using forceps, and the crown and “stem” tissue was ground to a fine powder with a mortar and pestle and additional liquid nitrogen. The powdered tissue was transferred to a 15 ml screw-top tube and transferred to a water bath at 85°C before the tissue had thawed. Samples were maintained at 85°C for 15 min to stop intrinsic enzyme activity, then held on ice until all samples were processed. An 18 gauge (1.3mm diameter) hypodermic needle was used to punch a hole in the bottom of the 15 ml tube, which was then placed within a 50 ml centrifuge tube. The samples were centrifuged at 8,000 xG for 10 minutes and the exudate recovered from the 50 ml tube. The residual plant tissue was transferred to a new 15 ml screw-top tube, 2 ml of pure water were added, samples were mixed by vortexing, then placed in a boiling water bath for 15 min. Samples were then cooled to room temperature, centrifuged at 10,0000 x G for 10 min, and the supernatant was transferred to a new tube. The recovered plant exudate and supernatant samples were frozen to -20°C and then lyophilized to complete dryness. The dried plant material was dissolved in deuterium oxide (D<sub>2</sub>O; Cambridge Isotope Laboratories, Tewksbury, MA, USA) to a concentration of 50 mg/ml. In preparation for nuclear magnetic resonance (NMR) analysis, samples were diluted to 15 mg/ml in 1mM TMSP (2,2,3,3-d4 sodium-3-trimethylsilylpropionate; Cambridge Isotope Laboratories) in D<sub>2</sub>O. <sup>1</sup>H NMR spectra were recorded on a Varian / Agilent 400-MR (399.763 MHz for <sup>\`</sup>H) instrument using a Varian OneNMR probe. The data were acquired using the standard PRESAT pulse sequence where the carrier was set to the residual water frequency. A relaxation delay of 13 seconds was followed by a weak presaturation field (30 Hz B1) for two seconds and the data collected using a 90-degree pulse and an acquisition time of 2.6 seconds. Sixteen NMR scans were completed for each sample. The receiver gain was kept constant for all samples. All spectra were processed using the DataChord Spectrum Miner software (One Moon Scientific, Westfield NJ). The raw FIDs were loaded into DataChord Spectrum Miner and processed as a batch with no apodization and the resulting spectra were aligned using the peak from TMSP as the standard located at 0 ppm. Spectral regions were defined manually and integrated peak areas determined. The compound peak areas of all spectra were standardized to TMSP as follows. The grand mean area of all of the TMSP peaks in the experiment were determined, then for each run, the area of the TMSP peak was divided by that grand mean, yielding a standardizing factor for the run. Each peak area in that spectrum was multiplied by the factor to yield standardized data. Further analysis was carried out using JMP software ([http://www.jmp.com](http://www.jmp.com/)). Identification of metabolites was undertaken by analysis of multidimensional NMR data and comparison of chemical shift data with database values. Two-dimensional <sup>1</sup>H-<sup>1</sup>H TOCSY (TOtal Correlation SpectroscopY), <sup>1</sup>H-<sup>13</sup>C HSQC (Heteronuclear Single Quantum Correlation) and <sup>1</sup>H-<sup>13</sup>C HMBC (Heteronuclear Multiple Bond Correlation) spectra were recorded using a sample that contained a compound associated with freezing tolerance based on PCA analysis. Two-dimensional data were recorded on a Varian / Agilent VNMRS 500 (<sup>1</sup>H 499.84 MHz, <sup>13</sup>C 125.69 MHz) using the standard pulse programs zTOCSY, gHSQCAD and gHMBCAD. For the TOCSY data 2 x 512 t1 increments (States-TPPI) using 16 scans each were acquired using non-uniform data sampling (NUS) at 50% density giving a total acquisition time of 4.25 hours. Residual water was reduced using a 20 Hz presaturation radio frequency field for 1 second and the TOCSY mixing was accomplished using a DIPSI-2 sequence for 80 ms. The sweep width was 5,660 Hz in both dimensions and the data were processed using DataChord Spectrum Analyst (One Moon Scientific, Westfield, NJ) using an iterative soft threshold Fourier Transform algorithm to produce a final 2Kx2K point 2-dimensional spectrum. HSQC data were acquired using gradient coherence selection and utilizing 2 x 256 t1 increments of 64 scans and NUS sampling (50%) for a total acquisition time of 7.75 hours. Residual water was reduced with presaturation and the sweep width in the <sup>1</sup>H dimension was 5660 Hz while for the <sup>13</sup>C dimension 20,736 Hz (165 ppm) was used. Data were processed similarly to the TOCSY data to yield a 2K x 2K final spectrum. HMBC data were recorded with 2 x 256 t1 increments using linear sampling and 64 scans per t1 increment to give a total acquisition time of 13.3 hours. A spectral with of 5,660 Hz was used for the <sup>1</sup>H dimension and 24,510 Hz (195 ppm) was used for the <sup>13</sup>C dimension. A long-range <sup>1</sup>H-<sup>13</sup>C J-coupling of 8 Hz was used to develop <sup>1</sup>H-<sup>13</sup>C correlations. Residual water was suppressed with presaturation and the data processed to produce a 2K x 2K final spectrum. The data were anodized in the t2 domain with a sine bell function and in the t1 domain with a gaussian function and the data were represented in the mixed phase mode (absolute value in F2 and phase sensitive mode in F1). Once identifications were obtained samples were compared with authentic samples of compounds. # Results ## Plant freezing tolerance Tabulated survival data are provided in. Analysis of variance revealed that wheat lines (genotypes), minimum temperature, and the time the freezing tests were initiated all were significant sources of variation at P\<0.0001. Replications and subsamples within replications were significant only at P\<0.05, and together accounted for less than 1.5% of the variation. Average survival of each of the wheat lines when freezing was initiated at each of the time points (ZT0, ZT6, ZT12, ZT18) is shown in. These times corresponded to the start and midpoint of the light period (ZT0 and ZT6, respectively), and the start and midpoint of the dark period (ZT12 and ZT18, respectively). With the exception of ORFW, which had very little survival in any of the tests, freezing tolerance was significantly greater when the plants were exposed to the subzero temperatures at the midpoint of the light period or the midpoint of the dark period, compared to the start of either of those periods. The improvement of survival in the tests conducted mid-period ranged from 7.8 to 32.4%, compared to the tests initiated at the start of the periods, with an average improvement of 21.3% (shown by the grey bars in). LT<sub>50</sub> values calculated from the combined data for the trials that were started at the midpoint of the light and dark periods, and for the trials initiated at the start of those periods are shown in. The differences in LT<sub>50</sub> between the periods ranged from 1.8 to 4.9°C, with an average of 2.5°C. All freezing tolerance tests were initiated with a 16-hour period at -3°C to allow the soil water to freeze and the heat of ice formation to dissipate, followed by controlled reduction in temperature to the target, potentially damaging, challenge temperature, as described above. This 16-hour “pre-freezing” period and the controlled reduction in temperature to the target resulted in the plants experiencing the potentially damaging temperature at about 24 hours after the start of the test. Consequently, the plants were exposed to the lowest temperature at about the same time of day as the test was started, but 24 hours later. ## CBF14 and CBF15 expression Expression of CBF14 in Norstar and Tiber showed upregulation from ZT6 to the subsequent ZT0, then downregulation from ZT0 to ZT6. In contrast, expression of CBF14 in ORFW was upregulated from ZT12 until the subsequent ZT0, then downregulated during the ZT0 to ZT12 time interval. Regulation of CBF15 in ORFW was very similar to regulation of CBF14. In contrast, in Norstar and Tiber, the regulation of CBF15 was essentially 180° out of sync with CBF14, i.e. CBF15 was downregulated from ZT6 to ZT12 and again from ZT18 to ZT0 while CBF14 was upregulated, and CBF15 was upregulated from ZT0 to ZT6 while CBF14 was downregulated during that time interval. ## NMR spectra of plant extracts An example of the <sup>1</sup>H NMR spectra obtained is shown in. A total of 70 compounds ranging from 1 ppm (the TMSP standard) to 9.5 ppm were identified in all of the samples, as shown in. For overall comparison of the plant lines and times of sample collection, the data from the two kinds of extractions (exudate vs. boiled) were considered as one data set and analyzed with principal component analysis (PCA). A plot of the first two principal components clearly distinguished the three plant lines, indicating Norstar, Tiber and ORFW differed from one another in relative content of some of the 70 compounds identified. The four sample collection time points also were associated with consistent positions on the PCA chart. The data points associated with the end of the light period/start of the dark period, ZT12, and middle of the dark period, ZT18, were clearly separated from the data points associated with the end of the dark period/start of the light period, ZT0, and the middle of the light period, ZT6 in each of the samples. These results suggested that the cellular contents of the plants were constantly in flux throughout the 24-hour cycle, at constant 4°C and 12-hour light/12-hour dark photoperiod. However, in contrast to the 12-hour cycling of freezing tolerance, cycling of the metabolites detected with NMR was in diurnal fashion with data points associated with the dark period (ZT12 to ZT18) consistently distinct from data points associated with the light period (ZT0 to ZT6) on the PCA plot. Thus, there did not appear to be an overall similarity of metabolite dynamics as detected by NMR, with the ultradian variation in freezing tolerance. This lack of association was most apparent when considering that the decreased levels of freezing tolerance were observed with freezing tests initiated at ZT0 or ZT12, which consistently were widely separated on the PCA plot of the NMR spectra. The PCA plot represents overall similarity of the composition of the plant extracts and may not represent the behavior of individual compounds. Further examination of the concentration dynamics of the 70 individual compounds found in the NMR spectra from each plant line at each time point revealed that “compound 15”, observed as a peak at 2.06 to 2.19 ppm, varied in a manner that was consistent with expressed freezing tolerance. Of the three lines examined with NMR, Norstar was the most freezing tolerant, Tiber was intermediate, and ORFW was the least freezing tolerant at each of the time points. At each of the time points, the concentration of compound 15 was greatest in Norstar, intermediate in Tiber, and lowest in ORFW. Also, in Norstar, the concentration of compound 15 varied in ultradian fashion, reaching local maxima at ZT6 and ZT18 and local minima at ZT0 and ZT12, coincident with levels of expressed freezing tolerance. In contrast, the concentration of compound 15 remained essentially unchanged in Tiber throughout the 24-hour period at 50–70% of the concentration seen in Norstar. In ORFW, with very little freezing tolerance, compound 15 reached maximum concentration at ZT6, at about 40% of the concentration seen in Norstar, but then declined to an undetectable level by ZT18, and increased only slightly between ZT18 and ZT0. Thus, there appeared to be an association of the concentration of compound 15 with the observed relative levels of freezing tolerance in these three wheat lines, especially apparent with consistently higher concentrations and cyclic variation in parallel with variation of freezing tolerance in Norstar. The identity of compound 15 was revealed by analysis of several 2-dimensional data sets starting with the TOCSY data where the signal at 2.144 ppm in the proton spectrum correlated to a signal at 2.452 ppm and further to another at 3.777 ppm. The HSQC data showed the signal at 2.144 ppm to be a methylene group with a carbon chemical shift of 29.0 ppm whereas the signal at 2.452 ppm was another methylene with a carbon chemical shift of 33.7 ppm and the signal at 3.777 ppm was a methine with a carbon chemical shift of 56.8 ppm. The HMBC data showed that the methylene at 2.144 ppm had long-range correlation to two carbonyl groups at 176.7 and 180.3 ppm. The methine at 3.777 ppm was also correlated by a long-range coupling to the carbonyl at 176.7 ppm. Long-range correlations were seen between the methylene groups and the methine group but no correlations could be seen from the methylene at 2.452 ppm to the carbonyl groups. The signal to noise ratio of the HMBC data was not high due to the overall low concentration of most of the metabolites in the mixture so even with the lack of correlation from one of the methylene groups to one or more carbonyls it was concluded that the patterns matched closely with those expected for glutamic acid or glutamine. HSQC spectra were collected on both glutamic acid and glutamine in D<sub>2</sub>O and compared to the spectra recorded on the plant sample. The overlay of the data sets for the plant sample and for glutamine are shown in. The chemical shifts in <sup>13</sup>C from glutamine were coincident with compound 15, while signals in the <sup>1</sup>H dimension from compound 15 were slightly shifted (2 Hz) from the authentic sample. This small of a difference easily could be due to minute pH or concentration differences, hence, we suggest that compound 15 is glutamine. # Discussion Freezing tolerance of young, fully cold-acclimated plants of the 12 winter wheat lines investigated here was expressed in a biphasic manner in plants grown at constant 4°C under a 12-hour photoperiod. Of the four time points studied, the greatest freezing tolerance occurred at ZT6, the midpoint of the light phase, and ZT18, the midpoint of the dark phase. This ultradian cycling of phenotypic expression of freezing tolerance suggested that genes involved in freezing tolerance also were expressed in ultradian cycles. The transcript expression levels of CBF14 and CBF15, genes shown to play a major role in freezing tolerance in wheat were shown to cycle throughout the day, but essentially 180° out of phase with one another, in freezing tolerant cultivars Norstar and Tiber. However, in wheat line ORFW, a line that develops very little freezing tolerance, the cyclic expression levels of CBF14 and CBF15 were essentially in phase. Thus, although both CBF14 and CBF15 appeared to be upregulated from ZT12 to the subsequent ZT0 in ORFW, there was no evidence of greater freezing tolerance as a result of that simultaneous elevated expression of both genes. In the case of Norstar and Tiber, freezing tolerance manifested most strongly at ZT6, when CBF15 was expressed to high levels, concomitant with low levels of expression of CBF14, and at ZT18, when both genes were expressed to high levels. These results suggest that precise coordination of expression of these genes, as opposed to simply higher levels of expression, was critical to manifestation of greater freezing tolerance in these two wheat lines. We previously found that ORFW is capable of responding to freezing stress in some ways similar to Norstar and Tiber. For example, lipid dynamics in response to six low-temperature treatments were very similar in ORFW and five other cultivars, including Norstar and Tiber, indicating that some components of the low-temperature response of the cold-sensitive ORFW function similarly to those components in more cold-tolerant genotypes. This observation suggests that different pathways may significantly contribute to freezing tolerance in the different genotypes. The evidence of the involvement of glutamine may point to one such pathway, suggesting that Norstar may make use of a glutamine-dependent pathway that is not functional in Tiber or ORFW. Others reported increased glutamine concentration in wheat plants on exposure to low temperatures. Kan et al, reported that supplying glutamine to rice plants grown in hydroponic culture resulted in upregulation of several genes involved in stress response including a number of transcription factors, leading the authors to suggest that “glutamine may also function as a signaling molecule to regulate gene expression in plants.” Kamada-Nobusada et al. also suggested glutamine may function as a signal molecule in rice plants, and Miranda et al. and Molina-Rueda and Kirby reported that glutamine may function as a signal molecule in stress response in Sorghum species and Pinus species, respectively. Thus, a growing body of evidence suggests that plants have the potential to activate a stress response pathway that may be regulated by variation in glutamine concentration. In the three wheat lines studied here, the concentration of glutamine in the cellular fluid varied in concentration in a manner consistent with freezing tolerance in that the concentration was greatest in Norstar, the most freezing tolerant line, intermediate in Tiber, which is intermediate in freezing tolerance, and lowest (sometimes undetectable) in ORFW, which is the least cold tolerant line. Also, the concentration of glutamine in Norstar varied with two highs and two lows in the 24-hour period, coincident with the freezing tolerance dynamics of Norstar. Others have found diurnal fluctuation in the concentration of some amino acids in wheat sap under warm (20–22°C) conditions and a 16-hour photoperiod, but glutamine concentration was not found to vary in that study. That result, considered with the results reported here, suggests that glutamine may function as a component of the pathway(s) responsible for cyclic variation of freezing tolerance under certain conditions of temperature and photoperiod in genotypes capable of using that pathway; constant low temperature (4°C) and 12-hour photoperiod in the cultivar Norstar in the present case. The observed sequential increase, then decrease in freezing tolerance is consistent with our previous findings showing that exposing the plants to -3°C resulted in sequential, extensive reengineering of the transcriptome, and of cellular composition, accompanied by significant increases in freezing tolerance. Others have shown that ice crystal formation begins at about -3°C in wheat tissues suggesting that the variation in expression of freezing tolerance at the time points studied probably originated in this initial response to -3°C and ice crystal formation. Plant biological rhythms entrained to day length usually persist for several cycles after the plants are moved to darkness and thus the biphasic response we observed resulted from interaction of gene products present at the time the plants were exposed to -3°C and darkness, interacting with the gene products and altered cellular component composition involved in the response that developed over the -3°C incubation period and subsequent exposure to lower temperature, likely while the two cycles observed within the 24-hour time frame continued to function. This biphasic stress response pattern is similar to that described for genes encoding enzymes involved in response to high light stress in Arabidopsis grown under a 16 hours light, 8 hours dark photoperiod. Of the genes investigated in that study, two peaks of expression were found, one in the light and a second in the dark. The authors suggested that the gene expression “was controlled by light and a second unidentified factor”. Our observations of expression dynamics of CBF14 and CBF15 under low temperature (4° C) and 12-hour photoperiod were consistent with the possibility that cohorts of CBF genes, and consequently genes of the CBF regulon, were expressed in a ultradian fashion, but with the expression dynamics of some out of phase with others in some genotypes, leading to the ultradian variation in freezing tolerance we observed. Under this hypothesis, freezing tolerance would appear to be under the control of “light and a second unidentified factor” as described in Arabidopsis, but the second factor simply may be the lack of light, to which some regulatory genes, including CBF genes, respond with upregulation. In barley, at least 20 CBF genes have been found, and have been classified into three phylogenetic groups, designated group 1, 3 or 4. Under a 12-hour photoperiod and constant 22°C, expression of CBF genes from group 1 was not detected. Expression of CBF genes from group 4 cycled in diurnal fashion, reaching peak expression 8–12 hours after subjective dawn. Expression of CBF genes from group 3 cycled in diurnal fashion, but one of them reached peak expression at the start of the dark period while expression of a second gene from group 3 reached peak expression near the end of the dark period. These observations are consistent with the ultradian cycling of freezing tolerance we observed, again suggesting that the onset of the light phase, and the onset of the dark phase were characterized by specific regulation of unique cohorts of genes conditioning freezing tolerance, with the tolerance ultimately reaching its greatest expression at the midpoint of the light, and of the dark phases. It is likely that this kind of regulation is dependent on the temperature at which the plants are grown in addition to the photoperiod, as has been shown with other wheat genes. For example, expression levels of wheat genes encoding lipocalins and similar proteins are correlated with the capacity to develop freezing tolerance; some of these genes in wheat plants grown under 16 hours light, 8 hours dark reached a peak of expression during the dark phase when grown at 4°C, but not when grown at 20°C. Whether the pathways active in the increased freezing tolerance observed after exposure to sub-freezing temperature beginning at the midpoint of the dark phase were the same pathways acting after exposure to sub-freezing temperature beginning at the midpoint of the light phase is unknown. However, Bieniawska et al., investigated the effects of low temperature and diurnal cycling on the Arabidopsis cold-responsive transcriptome and found “stronger, more abundant induction of TFs in the morning than in the evening,” (TFs = transcription factors); consistent with different transcription factors and presumably different target genes being involved in the response to cold at different time points of the light cycle. Thus, we suggest that the 12-hour cycle from low, to high, to low freezing tolerance in the light, and the similar cycle that completed in 12 hours of darkness, represent distinct, possibly overlapping signaling pathways that function as 12-hour ultradian cycles, and these pathways may well be components of the CBF regulon. In addition, specific genotypes may activate additional pathway(s), perhaps modulated by glutamine in some genotypes, that contribute to the realized freezing tolerance. This complexity of the freezing tolerance response, and apparently different pathways functioning in different genotypes, is not surprising considering the genetic constitution of the wheat plant. Wheat is an allohexaploid species formed by naturally-occurring interspecific hybridization of three diploid ancestors, first to form a tetraploid species from two of the diploids and then hybridization of the tetraploid with a third diploid to form the allohexaploid. The hybridization event that led to the allohexaploid occurred relatively recently, perhaps less than 10,000 years ago. Hence, each of the three genomes (designated A, B, or D) was contributed by a species that itself was subject to stress-related selection pressures for many generations, including low temperature stress, and presumably developed tolerance mechanisms independently of the other species. The three genomes comprising T. aestivum still contain “many (distinct) functional gene complexes” and interact in complex ways with myriad consequences. These complex interactions include “gene silencing, gene elimination, or gene activation and transposon activation via genetic and epigenetic alterations”. Transposon activation was observed during the cold acclimation process in Norstar, but not three other, closely-related, less cold- tolerant wheat lines. The uniqueness of this response to Norstar, among four closely-related wheat lines, suggests the specific kinds of complex interactions that occur among the three genomes is unique to each individual allohexaploid genome. This possibility is consistent with our observations of differing regulation of CBF14 and CBF15 and glutamine concentration among three wheat lines with differing freezing tolerance, perhaps indicating that the observed freezing tolerance was predominantly determined by a different genome (A, B, or D) in each line. Further study of the effects of environmental factors on the expression of the freezing tolerance that results from these pathways is needed to reveal the genes and mechanisms involved. From a practical standpoint, these results suggest that tests of freezing tolerance and of gene expression dynamics related to freezing tolerance in winter wheat lines should be standardized to the photoperiod in order to meaningfully compare the relative levels of freezing tolerance among wheat lines. # Supporting information Disclaimer: Mention of product names does not represent an endorsement of any product or company but is given only to clarify the methodology; other products may be equally effective. USDA is an equal opportunity provider and employer. This project was supported by USDA-ARS project 2090-21000-030-00D. [^1]: The authors have declared that no competing interests exist. [^2]: ‡ These authors also contributed equally to this work.
# Introduction The transmissible spongiform encephalopathies, or prion diseases (PrDs), are a family of fatal, transmissible neurodegenerative conditions. Characteristic features of prion diseases include deposition of abnormal prion protein within the brain, astrogliosis and spongiform vacuolation. The causative agents (prions) arise from mis-folding of a normal cellular protein, the prion protein (PrP), into disease-associated conformers that can then template further mis- folding and propagation of prions. Various indications of damage are found in terminal brain tissue, including signs of substantial oxidative stress. Reactive oxygen and nitrogen species (referred to collectively herein as ROS) are produced as byproducts of various cellular reactions and are known to have critical roles as cellular signaling molecules. Oxidative damage occurs when the mechanisms producing ROS exceed cellular detoxifying capacity. Markers of oxidative damage within the brains of people dying from PrD include lipid peroxidation and protein nitration. It is unclear whether oxidative damage plays a causative role in the disease pathogenesis or whether it is a consequence of failing detoxification pathways. There are various enzyme systems responsible for the detoxification of ROS, including the superoxide dismutase (SOD) family. The SOD family comprises three members; SOD1 or CuZnSOD, located ubiquitously throughout the cell cytoplasm and mitochondrial intermembrane space, SOD2 or MnSOD, confined to the mitochondrial matrix, and SOD3 or extracellular SOD found in the extracellular matrix. Failure of one of these enzymes to adequately detoxify ROS could permit oxidative damage to accumulate to toxic levels. Animal models have been used to consider the role of ROS in PrD and the importance of cellular antioxidant defenses. It has been shown that lipid peroxidation begins early during the course of murine prion infection, around the time of initial deposition of PrP and spongiform development, well before onset of symptoms. Reduction of SOD1 protein in mouse models of PrD accelerates disease, however, the role of SOD2 in prion infection was, until now, un- reported. SOD2 expression is essential for life and, in its absence, mice die perinatally of dilated cardiomyopathy. However, if SOD2 knock-out mice are kept alive for several weeks post birth using antioxidant therapy they develop a spongiform encephalopathy reminiscent of a PrD. At the terminal phase of murine PrD SOD activity and SOD2 expression were previously found to be decreased. In support of a loss of function of SOD2 during disease, a SOD2 mimetic has been shown to significantly extend the lifespan of prion infected mice and reduce spongiform vacuolation. Additionally, in cell culture systems modelling prion infection, increased ROS production and damage to cellular proteins and lipids have been found. Of interest, within this system, reduction of SOD2 protein was associated with cytosolic localization and degradation by caspases. Caspase activity is increased as mice approach terminal PrD and caspase 3 has directly been shown to cleave SOD2, therefore an increased turnover and resultant loss of function of SOD2 could be involved in disease progression or formation of spongiform damage. Potentially, if PrP influences SOD2 function and this function is critical for maintaining homeostasis during prion disease, its loss toward terminal disease could be responsible for the accumulating cellular damage including development of spongiform change. It is not possible to consider the influence of SOD2 by investigating disease in SOD2 knock-out mice due to their short lifespan. However, SOD2 heterozygous mice, which express approximately 50% of the SOD2 protein levels of wild type mice, develop normally and have a normal life expectancy with only a few reported effects of the reduced protein levels. The SOD2 heterozygous mice do show a clinically silent underlying mitochondrial oxidative stress, which renders them more susceptible to oxidative insults, such as what may occur during prion infection. We hypothesized that if SOD2 is critical for maintaining homeostasis and that this fails during disease, the reduction of SOD2 in SOD2 heterozygous knock-out mice should be sufficient to accelerate the disease course and exacerbate the spongiform change. Our results indicate that SOD2 reduction does not influence acute or long-term responses to three different scrapie prion strains. # Methods ## Animal ethics statement All mice were housed at the Rocky Mountain Laboratory (RML) in an AAALAC- accredited facility in compliance with guidelines provided by the Guide for the Care and Use of Laboratory Animals (Institute for Laboratory Animal Research Council). All experiments were approved by the RML Animal Care and Use Committee, protocol 2018-072-E. Mice were socially housed in groups of 3–4 mice per box in climate controlled disposable caging with wood shavings and nestlet bedding provided for enrichment. Room temperatures were kept at 70–72˚F and a 12-hr light-dark cycle was used. Mice had free access to food and water. ## Mice Heterozygous B6.129S7-Sod2tmLeb/J<sup>(+/-)</sup> (Jackson Laboratories) were crossed to produce B6.129S7-Sod2tmLeb/J<sup>(+/-)</sup> (SOD2<sup>+/-</sup>) and Sod2tmLeb/J<sup>(+/+)</sup> (WT) littermate wild type controls. Genotypes were confirmed by PCR analysis as described by Jackson Laboratory. Primers detecting the mutant, `TGT TCT CCT CTT CCT CAT CTC C` (oIMR0781) and `ACC CTT TCC AAA TCC TCA GC` (oIMR0782) along with wild type primers, `TGA ACC AGT TGT GTT GTC AGG` (oIMR0878) and `TCC ATC ACT GGT CAC TAG CC` (oIMR0888) were used to distinguish genotypes of littermate controls. ## Infections Prior to experimentation we used Graphpad Statmate software to determine reasonable group sizes to detect an estimated difference in incubation periods of 1 week. Groups of 10–12 male mice per genotype (total = 69 mice) were intracerebrally inoculated with 30 μl of a 1% (w/v) brain homogenate from 22L, RML or ME7 prion strains. Mice were anesthetized with isoflurane and inoculations were performed using a ½ inch long, 27-gauge needle on a 1 mL slip tip syringe. The inoculum was injected into the left cerebral hemisphere, approximately 2 mm lateral to the midline and 2–3 mm deep, near the region of the thalamus. The titers of these scrapie stocks had been determined previously in C57 mice and contained the following 50% infective doses (ID<sub>50</sub>) in each 30-μl volume: 22L, 6.0 × 10<sup>5</sup>; RML, 2.4 × 10<sup>4</sup>; ME7, 5.0 × 10<sup>4</sup>. Normal control mice were either ‘mock’ inoculated with uninfected brain homogenates (N = 4) or not inoculated (N = 4). Mice were monitored by evaluators blind to genotype twice weekly prior to onset of clinical signs, then every 1–3 days throughout the clinical phase (where they begin showing decrease in body condition, increased somnolence and kyphosis) and euthanized by isoflurane inhalation overdose followed by cervical dislocation when they developed terminal clinical signs (signs include ataxia, tremors, kyphosis, hyperactivity, somnolence, poor grooming, reduction in body condition). One SOD2<sup>+/-</sup> mouse from the 22L experiment was excluded from analysis. The mouse was euthanized at 67 dpi for unresolving dermatitis. This time point is far prior to any clinical signs of prion disease. Log-rank (Mantel-Cox) statistical analysis was performed for each prion strain using GraphPad Prism 8 software. ## H&E staining, immunohistochemistry and quantification Brain fixation, paraffin embedding, histology and immunohistochemistry for prion protein (antibody D13) and anti-GFAP were carried out as described previously. Positive pixel quantification was performed on brain sections from 5–6 mice per experimental group for D13 (produced in house as described in) and GFAP stained sections using the ImageScope positive pixel count algorithm (version 9.1) as described with the following modification. In the current manuscript, all positive pixels (including weak, moderate and strong) were included for the D13 analysis. ## Lesion profiling Lesion profiling was completed on 5–6 mice from each genotype from the 22L and ME7 infection experiments. Sagittal brain sections were cut approximately 0.5 mm from midline and stained by routine H&E protocols. Each section was scored for the severity of spongiform vacuolar degeneration in 4 gray matter areas (cerebral cortex, thalamus, superior colliculus and cerebellum). Spongiosis was scored as follows: 0, no vacuoles; 1, few vacuoles widely and unevenly distributed; 2, few vacuoles evenly distributed; 3, moderate numbers of vacuoles evenly distributed; and 4, many vacuoles with some confluences. An average score and standard deviation for each experimental group was calculated and shown in. ## Western blotting 10% (w/v) brain homogenates were prepared in 1x PBS, further diluted into 1% (w/v) for western blotting analysis. Proteins were denatured by boiling for 5 minutes in 1× sample buffer (containing 6% Beta-mercaptoethanol), resolved in Bolt 4–12% Bis-Tris Plus gels (Invitrogen), and transferred to PVDF membrane (Millipore). SeeBlue Pre-stained Protein Standard (Life Technologies) was resolved alongside the protein samples. The membrane was incubated in blocking solution (5% w/v non-fat milk in 1× TBS and 0.05% Tweens) for 1hr at room temperature and in anti-SOD2 antibody (1:5000 dilution in blocking solution; Abcam: ab13533), anti-SOD1 (1:4000 dilution in blocking solution; Abcam: ab13498) or anti-SOD3 (1:4000 dilution in blocking solution; Abcam: ab83108) overnight at 4°C. The primary antibody was coupled with the appropriate secondary antibody (HRP conjugate), and the protein bands were visualized using ECL Select (Amersham) and imaged by iBright imaging system (Invitrogen). Full blot images are shown in Sup File 1. ## Proteinase K digestion and PrP immunoblotting blotting 10% (w/v) brain homogenates in PBS were digested with Proteinase K, run on 12% tris-glycine gels and transferred to PBDF membranes as previously described. Immunoblots were probed as described with the two following modifications, anti- PrP antibody D13 was used at a 1:1,000 dilution and the secondary antibody (anti-human) was used at a 1:20,000 dilution. ## Electrophysiology As described previously, 300 μm thick hippocampal sections were collected from 12-week-old mice in an ice-cold cutting solution (3 mM KCl, 25 mM NaHCO<sub>3</sub>, 1.25 mM NaH<sub>2</sub>PO<sub>4</sub>, 206 mM Sucrose, 10.6 mM Glucose, 6 mM MgCl<sub>2</sub>.6H<sub>2</sub>O, 0.5 mM CaCl<sub>2</sub>.2H<sub>2</sub>O) using a vibratome (Leica VT1200S). The hippocampal slices were incubated for 1 hr at 32°C in carbogenated (5% CO<sub>2</sub>; 95% O<sub>2</sub>) artificial cerebrospinal fluid (aCSF: 126 mM NaCl, 2.5 mM KCl, 26 mM NaHCO<sub>3</sub>, 1.25 mM NaH<sub>2</sub>PO<sub>4</sub>, 10 mM Glucose, 1.3 mM MgCl<sub>2</sub>.6H<sub>2</sub>O and 2.4 mM CaCl<sub>2</sub>.2H<sub>2</sub>O). The slices were mounted on to 60MEA200/30iR-Ti-pr-T multi-electrode arrays (MEA; Multichannel Systems; Germany) for the electrophysiological recording while being continuously superfused with carbogenated aCSF. The Shäffer collateral pathway was stimulated by electric stimulation (2000–2500 mV) to evoke the field excitatory post-synaptic potential (fEPSP) at the CA1 region. The baseline stimulation intensity was determined by an input-output curve obtained from stimulating the pathway with increasing strength of electric stimulation (input), starting at 500 mV to a stimulation intensity (4000–5000 mV) that evoked the maximum fEPSP amplitude (output). This max response was indicated by the point where the input-output curve reached a plateau. The stimulation strength that evoked 30–50% of the maximum fEPSP amplitude was the baseline stimulation intensity. The baseline fEPSP was recorded for 30 minutes by applying the baseline stimulation intensity in 30 seconds interval. Following the first five-minute baseline, the slices were treated with 0.5% (w/v) brain extracts (normal brain homogenates or PrPSc-infected mouse brain homogenates in aCSF) for 10 minutes. The 0.5% (w/v) dosage was determined based on characterizations done for previous studies, we chose this dose after determining that higher concentrations caused non-specific tissue damage. After this treatment, the superfusion returned to aCSF for the rest of the recording. After the baseline recording, the LTP was induced by Tetani, consisting of three 100 Hz trains (500 ms long per train) delivered in 20 sec interval. The post- tetani fEPSPs were recorded for 30 minutes, with the last ten-minute recording was averaged and used for the statistical analysis of LTP. # Results As expected from previous reports, the SOD2<sup>+/-</sup> mice developed normally, lived a normal lifespan compared with their WT littermates and did not display any overt pathology related to the reduced SOD2 level. Within the brain, the SOD2<sup>+/-</sup> mice expressed approximately half the protein level of SOD2 that WT mice expressed. ## Reduced SOD2 does not influence electrophysiological responses following acute exposure to 22L, RML and ME7 prions Previous studies have shown that certain prion strains can cause acute synaptic toxicity, which is reflected in changes to neuronal long-term potentiation (LTP). To investigate whether the SOD2 reduction renders the SOD2<sup>+/-</sup> mice more vulnerable to acute synaptotoxic insult by prions, we measured LTP in hippocampal slices. shows the combined traces from the brain slices of 12-week- old SOD2<sup>+/-</sup> and WT mice exposed to one of the three strains, 22L, RML, or ME7 or control normal brain homogenate (NBH) in the artificial cerebral spinal fluid (CSF). For each mouse, slices were compared for their response to normal and infectious brain homogenate in parallel. Brain slices were assessed for LTP following 10 minutes exposure to the inoculum. No difference was seen in basal LTP between the genotypes in the artificial CSF, indicating that the reduced SOD2 protein was not causing a basal level of neuroelectrophysiological dysfunction. In the kinetic curves, LTP is significantly reduced in response to the infectious brain homogenate compared with NBH for SOD2<sup>+/-</sup> brain slices exposed to 22L. When the average of the last 10-minute LTP was considered for each slice, the SOD2<sup>+/-</sup> brain slices exposed to 22L remained the only significantly changed LTP response. Overall, exposure to scrapie prions had limited significant effects on LTP for any strain in either WT or SOD2<sup>+/-</sup> mice. ## SOD2<sup>+/-</sup> mice do not have reduced lifespan when challenged with 22L, RML or ME7 prions compared with WT mice We hypothesized that the lower protein levels of SOD2 in the SOD2<sup>+/-</sup> mice would leave them vulnerable to increased oxidative damage during disease pathogenesis and, therefore, they might demonstrate more rapid disease progression to death. When inoculated with 22L, RML or ME7 prions, with the evaluator blinded to their genotype, the reduced SOD2 expression made no difference to the disease course (P = 0.59, 0.58 and 0.88 respectively;). Furthermore, when the disease-associated PrP was examined by western blotting terminal brain tissue following protease digestion, accumulation was similar in both genotypes. Thus, reduced SOD2 was not influencing disease duration or the presence of disease-associated PrP. ## SOD2 protein levels are unchanged in SOD2<sup>+/-</sup> prion disease brains A possible explanation for the lack of difference across genotypes was that SOD2<sup>+/-</sup> mice maintained the ability to up-regulate SOD2 protein to WT equivalent levels when under stress. If this were the case, at terminal disease all mice could have the same SOD2 protein levels and, consequently, there would be no difference in disease parameters. Western blotting brain tissue from terminal mice showed that SOD2 protein levels were not increased in prion infected SOD2<sup>+/-</sup> mice relative to the NBH controls. Only the WT RML- infected mice showed a change in their SOD2 expression as a result of infection, with a significant \~2-fold increase over NBH mice, although the ME7-infected WT mice also showed highly variable SOD2 levels. It is unclear why SOD2 protein levels varied so widely in WT terminal disease mice across the strains, but this suggests the increase in the WT RML-infected mice has no relevance to pathogenesis and that fluctuations in SOD2 protein levels are unimportant for disease progression. We also considered protein levels of SOD1 and SOD3, finding only SOD1 significantly increased in the ME7 infected WT mice. No significant differences were apparent between genotypes suggesting that SOD1 and SOD3 are not changed as a compensatory response for the lower SOD2. ## SOD2+/- mice show no overt differences in brain pathology As SOD2 knock-out mice kept alive post birth using anti-oxidant therapy develop a spongiform encephalopathy, we examined the PrD pathology present in the SOD2<sup>+/-</sup> mice and WT mice infected with ME7 and 22L; representing the longest and shortest incubation periods of the strains tested respectively. No explicit difference was seen in PrP deposition, astrogliosis or spongiform change as a result of the reduced SOD2. ## Lesion profiling confirms reduced SOD2 does not influence spongiform pathology To look closer at whether the reduced SOD2 might influence the severity of spongiform change in the SOD2<sup>+/-</sup> mice, we examined four brain regions, grading their spongiform severity. The examiner was blinded to the mouse genotypes. The lesion profiles of the cerebral cortex, colliculi, thalamus and cerebellum were not different between the SOD2<sup>+/-</sup> mice and WT controls. This confirms the observation that reduction of SOD2 protein does not influence the development of spongiform pathology. Additional pixel counts on the 22L infected brains also confirmed that PrP deposition and astrogliosis was not different between SOD2<sup>+/-</sup> and WT mice. Overall, the reduction of SOD2 did not influence neuronal function upon challenge with infection, disease duration, or the appearance of the neuropathology and severity of the spongiform pathology across different brain regions in response to infection with different prion strains. # Discussion PrDs demonstrate increased oxidative damage within the brain that begins early in disease, and detection of SOD2 is changed in terminal brain tissue. The development of a PrD-like spongiform encephalopathy in SOD2 knock-out mice, treated with antioxidants to delay their death from cardiomyopathy, indicated SOD2 may be a contributing factor to PrD pathogenesis. However, in mice expressing approximately 50% of the SOD2 protein present in the brains of wild- type animals, we saw no effect of the reduced SOD2 on neuronal function or disease pathogenesis upon trial with three prion strains. As SOD2 is essential for life, the lack of difference in lifespan of the PrD SOD2<sup>+/-</sup> mice strongly indicates that SOD2 plays no role in the pathogenesis of PrD in mice. In previous studies, a decrease in SOD2 gene expression or protein has been observed in animals infected with scrapie strains. In Bourgognon et al., the infecting strain was RML; the same strain for which we observed an increase in protein levels of SOD2. The discrepancy may have arisen as Bourgognon et al. measured RNA levels rather than protein, as detected herein, or due to different genetic backgrounds of the mice. However, in human prion disease studies that examined protein levels in terminal brain tissue, increases in SOD2 protein were detected, indicating that the increased SOD2 in the RML infections does mirror what occurs in human disease. In Park et al., the strain studied was ME7 and the reduction in SOD2 measured by western blotting. We observed a highly variable response in SOD2 protein detection for the WT ME7 infected mice, with two of four brains showing a decrease. The variable SOD2 detection in our ME7 infections and the genotypes of this cohort were re-checked and a reason for the variability was not apparent. Since the WT protein intensities of SOD2 were variable across all infections and neither WT SOD2 protein levels nor the reduced levels in the SOD2<sup>+/-</sup> mice influenced the outcome of the disease, fluctuations of SOD2 protein levels appear to be a secondary effect of the disease process. We also observed discrepancies between our current LTP data and previously published electrophysiology studies looking at prion-induced synaptotoxicity where, in contrast with the previously published studies, we found few significant changes. These differences likely arise due to the different strains examined. The original studies used the mouse-adapted human M1000 strain, whereas in the current study we considered mouse-adapted scrapie strains. Differences in the aggressiveness of different prion strains and regions of the brain attacked are widely known. Even herein we found that the different strains produced different responses in the WT mice with a significant increase in SOD2 protein only in RML infected mice and an increase in SOD1 only in ME7 infected mice. The SOD2<sup>+/-</sup> mice showed the only significant change in LTP when challenged with 22L inoculum, possibly because this is the most aggressive strain we looked at with the shortest time to death. WT mice challenged with the same dosage of M1000 inoculum as we used here die around the same time as those challenged with 22L, although the timeframe is narrowed (145 dpi +/- 2 days for M1000 vs 145 dpi +/- 6 days for 22L herein). Therefore, an explanation for the discrepancy with our WT data is that 0.5% w/v does not contain enough scrapie prions to induce toxicity but 22L, causing the most aggressive disease, is closest to a toxicity threshold surpassed by the mouse-adapted human prions. The difference between the WT and SOD2<sup>+/-</sup> LTP in response to 22L was not sufficiently great that we believe it to be important to the disease process, especially when considered in the context of the lack of change in survival time or pathology. That the SOD2 protein levels were only increased in the RML infected and the SOD1 protein levels only increased in the ME7 infected WT mice is interesting. This raises the question of whether some strains cause greater oxidative stress than other strains or influence different redox producing/detoxifying pathways. Whilst there is ample evidence for oxidative stress in the brains of PrD mice infected with multiple strains, we are not aware of any studies that have found ROS levels to be higher in in one strain over another. Comparison of indirect markers of oxidative stress across several strains demonstrated no difference in the detection of these markers. More investigation may be required to understand if the strains do cause differing levels of stress and, if so, whether this is important to disease pathogenesis. While we showed no changes in the other SOD family members in the SOD2<sup>+/-</sup> mice to suggest the deficit was annulled by compensation mechanisms, this is not to say none were occurring. Potentially other redox balancing systems, outside the SOD family, may be providing some alleviation of increased ROS and PrP itself has been linked with protection against oxidative stress. Heightened ROS is associated with increased processing of PrP at the beta-cleavage site and during infection this becomes the dominant cleavage event. Beta-cleavage and especially the released N-terminal fragment, N2, that is not part of the protease resistant core, have been shown to have protective properties against oxidative stress. The N2 fragment has additionally be shown to localize to the mitochondria when administered to cells in their media, possibly effecting a protective response within the organelle. Therefore, many other cellular mechanisms may render the reduction of SOD2 redundant for survival or pathology during PrD. Our data indicates that SOD2 is not involved in PrD pathogenesis but this does not mean that other redox pathways and cellular antioxidant molecules or proteins are likewise unimportant. Indeed, SOD1 knock-out does reduce the lifespan of prion infected mice and, conversely, knock-out of the oxidant protein NADPH oxidase-2 (NOX2) was found to increase survival time and delay symptom onset. PrP has also been functionally linked to several redox signaling pathways, including NOX2 and neuronal nitric oxide synthase and any combination of these may become dys-regulated during disease leading to oxidative damage. Furthermore, a number of anti-oxidant based therapies have been found to extend survival time in mice, including the administration of a SOD2 mimetic in infectious disease and a combination therapy using a nanoformulation of pomegranate seed oil with neural progenitor transplantation in a genetic disease model. While SOD2 may contribute to the failings of these pathways in PrD, ultimately it appears that oxidative damage caused by other failing pathways is of greater significance to the development of disease pathology. A clear limitation of our study is that translation to humans is not necessarily guaranteed. For example, SOD1 knock-out mice have no overt phenotype up to six- months of age but develop age-dependent muscle atrophy later in life. However, human patients with mutations resulting in minimal SOD1 activity demonstrate signs of the deficit from around 6–9 months old with severe loss of motor functions in early childhood. Thus, the SOD1 knock-out mouse model does not fully recapitulate human SOD1 deficiency. However, a reported genetic SOD2 deficiency patient presented similarly to what was observed in SOD2 knockout mice, with death occurring a few days post birth. The causative mutation was recessive and the parents of the patient, who were heterozygous, showed no signs of the deficiency when just one functional allele was present. Consequently, the SOD2 knock-out model might provide a close indication of how reduced SOD2 protein influences human biology in the context of PrD, but this cannot be concluded without investigation within a human model of disease. We cannot exclude the possibility that a heterozygous reduction of SOD2 is insufficient to see any changes. While this is possible, and another heterozygous mouse experiment examining cysteine string protein alpha also showed no difference in incubation period, if the observed variations in SOD2 protein expression levels was a critical factor in PrD pathogenesis it is likely that some pathological effect would have been observed in the SOD2<sup>+/-</sup> mice. We conclude that reduction of SOD2 does not alter PrD pathogenesis and changes in SOD2 protein levels are likely a down-stream effect of disease rather than causative of damage. # Supporting information The authors would like to thank James Striebel and Dr. Christina Orru for critical evaluation of the manuscript, Jeff Severson for animal husbandry, and Lori Lubke and Nancy Kurtz for histology assistance. [^1]: The authors have declared that no competing interests exist.
# Introduction Critical illness is marked by organ dysfunction, the need for vital support, and a high risk of death, occurring against a backdrop of systemic immune activation. This immune activation may begin as an adaptive response to the initial injury, however, as the disease progresses, the immune response may become maladaptive or paralyzed. Critical illness-associated immune dysregulation has been described as the interplay between pro- and anti- inflammatory responses, although recent evidence suggests a mixed inflammatory state is common. While this process has been qualitatively described, there are no quantitative diagnostic or prognostic tools that have been validated clinically to assess immune status in the critically ill. Consequently, infectious complications are not only common in intensive care units but also difficult to diagnose. This has contributed to inappropriate use of broad- spectrum antibiotics and the emergence of multi-drug resistant organisms. A few years ago, studies employing cultured human cells suggested that instead of a single molecule (e.g., IL-6), a constellation of molecules could be used to monitor the complexities of the inflammatory response, serving as markers of infection. Miniaturized, multiplexed assays provide a rapid method for the unbiased screening of thousands of molecular species in a single assay. These technological advances provided the potential for investigators to leverage high-throughput assays to better study the host response to and recovery from critical illness and injury. Improved molecular diagnostics and prognostics, a better understanding of the complexity of systemic inflammation, and new therapeutic targets are expected deliverables, as reviewed recently. Based upon our ability to diagnose abdominal sepsis in pilot mouse studies, we hypothesized that the host response to infection could not only differentiate between infected and non-infected states, but could also be used clinically to differentiate between the host response to infectious agents and to model the host response to and recovery from infectious perturbations. Pneumonia was chosen as an immune system perturbation, given its relative frequency and considerable cost in terms of patient morbidity and health care expense. A bench-to-bedside, translational approach was employed to study the host response to pneumonia in critically ill subjects, comparing the informational content of standard clinical parameters and plasma cytokines to changes in the RNA abundance in circulating leukocytes. # Methods ## Mice, experimental procedures, and samples Care and use of laboratory animals were conducted in accordance with a protocol approved by the Washington University Animal Studies Committee, in compliance with guidelines (N01-RR-2-2135) prepared by the Committee on Care and Use of Laboratory Animals, Division of Research Resources, National Institutes of Health. Seven to nine week-old, male C57BL/6 mice were purchased (Harlan, Inc. Madison WI) and allowed to acclimatize for at least one week in a temperature- and light-controlled, pathogen-free barrier facility. Treated animals and concurrently studied controls were observed at 24 hour intervals for survival over eight days. In additional cohorts, whole blood was collected at 24 hours after injury. The 5 experimental groups were selected to reflect clinically important distinctions relevant to care of ICU patients: high mortality Gram-negative pneumonia with Pseudomonas aeruginosa (40 µl of 0.3 McFarland culture, 90% 7-day mortality (∼2–4×10<sup>7</sup> organisms)), high mortality Gram-positive pneumonia with Streptococcus pneumoniae (60 µl of 0.5 McFarland culture, 85% 7-day mortality (∼1.8–3.6×10<sup>7</sup> organisms)), and moderate mortality Gram-negative pneumonia with Pseudomonas aeruginosa (20 µl of 0.1 McFarland culture, 50% 7-day mortality (∼2–4×10<sup>6</sup> organisms)). To induce severe systemic inflammation without infection, intratracheal (i.t.) injection of P. aeruginosa lipopolysaccharide (500 µg in 50 µl 0.9% normal saline; Sigma, St. Louis) was performed (LPS group, 90% 7-day mortality). Mice injected i.t. with normal saline vehicle acted as the concurrent control group (saline group, 0% 7-day mortality). Previously reported protocols were used to intratracheally instill fluid into the lung. The census of surviving mice was recorded at 24-hour intervals for seven days. In three additional cohorts of animals, blood was collected into an EDTA-coated syringe from the inferior vena cava being careful to avoid contamination of the needle with other tissues. Blood was diluted 1∶1 with normal saline, pooled for the 8 animals in each treatment group, and separated into cells and plasma. Plasma was stored at −80°C until use. Erythrocytes were lysed hypotonically and RNA from peripheral leukocytes was harvested using RLT (Qiagen) and stored at −80°C until use. The 24 hour time point after injury was chosen as a time before appreciable mortality develops in animals with significant lung injury. ## Target cRNA and gene expression signal Each RNA sample was run on one GeneChip (a total of 15 mouse blood GeneChips from 120 animals). Total RNA from mouse blood was extracted as previously described. cRNA target for GeneChip hybridization was prepared from total RNA (Affymetrix, Santa Clara, CA). Both total RNA and cRNA were electrophoretically assessed for quality (Agilent Bioanalyzer). The mouse blood cRNA samples were hybridized with the U74Av2 GeneChip (approximately 12,400 probe sets). Fluorescent hybridization signal was detected using a GeneChip Scanner 3000 (Affymetrix). These mouse microarray data (and those for patients, see below) have been deposited in NCBI's Gene Expression Omnibus (GEO, <http://www.ncbi.nlm.nih.gov/geo/>) and are accessible through GEO Series accession number GSE6377. ## Data analysis and statistical tests for differential expression Expression values were calculated from GeneChip.cel files using Robust Multichip Average (RMA) software. Differentially expressed genes were identified using a mixed-model analysis of variance (ANOVA) and linear contrasts (Partek® Infer™ software) as previously reported. Leave-one out cross-validation (k-nearest neighbors, k = 2) was used to determine the reproducibility within this experimental set. Principal components analysis (PCA) was used to visually explore global effects for genome-wide trends, unexpected effects, and outliers in the expression data (Partek® Pro™ software, [www.partek.com](http://www.partek.com)). ### Patient studies After obtaining informed consent, venous blood (7 ml) was collected from mechanically ventilated, non-septic patients according to a protocol approved by the Washington University Institutional Review Board (#2004-0294). Patients were candidates for enrollment if they were on mechanical ventilation in the surgical ICU medical ICU, neurological ICU, or cardiothoracic ICU (CTICU) for 48 hours, were expected by the attending ICU physician to need mechanical ventilation for at least another 48 hours at the time of enrollment, and could provide written informed consent (from the patient or legal surrogate). VAP was diagnosed by the ICU attending physician, consistent with recently reported recommendations, without input from the investigators. Clinical data were entered into a VAP database, including gender, ethnic background, age, admitting diagnosis, type of ICU, APACHE II score, airway sampling technique and culture results, initial antibiotic therapy, and maximal clinical pulmonary infection score (CPIS) calculated based upon available data (several patients lacked daily arterial blood gas measurements to calculate P<sub>a</sub>O<sub>2</sub>/F<sub>i</sub>O<sub>2</sub> ratios). Patient blood was processed as described previously to minimize red blood cell RNA artifact ; briefly, samples were centrifuged (400×g 10′ RT) to form a buffy coat and to separate plasma from cells. Plasma was withdrawn and stored at −80°C until use. Blood cells were diluted into EL buffer (90 ml) (Qiagen) and incubated on ice for 15′. Leukocytes were pelleted by centrifugation (400×g 10′ 4°C), washed with EL buffer (30 ml) and lysed into RLT buffer (Qiagen) containing 1% β-mercaptoethanol. Genomic DNA was sheared by repeated passage through an 18 gauge needle and the resultant material was stored at −80°C. ### Patient plasma cytokine analysis Cytokines (GRO-α, IFN-γ; IL1-β; IL1Ra; IL1sr2; IL4; IL6; IL8; IL10; IL12; IL18; MCP1; MIP1α; MIP1β; NGF; RANTES; TNFα; TNF-sr1; TNF-sr2) were measured using a microarray immunoassay as previously described. Procalcitonin was measured according to the manufacturer's instruction (BRAHMS PCT LIA kit, Product number 354.1). ### Patient blood leukocyte mRNA profiling RNA was extracted, amplified and assessed for quality as described for murine samples. cRNA was hybridized against the HG-FOCUS array (Affymetrix, \>8700 probe sets encoding ∼8400 genes) and imaged as described for murine samples. Orthologs of murine genes were identified by comparison of the GeneChip Identifiers using the NetAffx Toolkit (Affymetrix). Consistent with recently published consensus statements, , clinical data were judged to determine when (if ever) each patient developed ventilator-associated pneumonia, with each patient acting as her/his own control. The timeline for each patient was defined such that the day of VAP diagnosis by the ICU attending physician was defined as day 0. A seven day time window from the gene expression time series was chosen as days −3, −1, 0, 1, and 3, with day 0 being the day that a patient was diagnosed as having VAP by the attending physician. Because blood samples were collected every other day, patients' samples were collected either on days −2, 0, +2 or on days −3, −1, +1, +3 relative to the VAP day of diagnosis described as time 0. For the purpose of analyzing the data from those patients who had samples collected on “odd” days, the time 0 data for these patients were interpolated. Those mRNA species whose abundance changed concordantly among the patients during the 7-day window surrounding the date of diagnosis were identified using extraction of Differential Gene Expression (EDGE) software. Online databases were used to determine gene annotation and functional categorization (DAVID 2.0 accessed 16 November 2006). ### Clustering algorithm Patient genes identified by EDGE were clustered as described previously. Briefly, a gene co-expression network was constructed by connecting every gene to the top d genes (d = 5 in this study) to which its expression profile is most similar. The network then was partitioned into a set of communities, i.e., relatively densely connected sub-graphs, by a spectral graph algorithm. The genes within each community formed a cluster. The number of clusters was determined automatically by the algorithm in order to maximize a modularity score,. Gene expression data were normalized prior to clustering such that the expression levels of each gene for each patient had a mean of zero and a standard deviation of one. Similarity between two gene expression profiles was measured by Pearson correlation coefficient. ### Karhunen-Loeve decomposition of microarray gene expression data To determine the dynamics of the host response to pneumonia, we constructed first a raw gene expression matrix corresponding to the ith pneumonia patient after RMA normalization to be X<sub>i</sub> = \[x<sub>i</sub>(1) … x<sub>i</sub>(N<sub>T</sub>)\] where x<sub>i</sub>(k)∈ℜ<sup>N</sup>, (k = 1,…,N<sub>T</sub>, i = 1,…11, and N = 8793 genes) is the gene expression vector at k th time point. k = 1,…,N<sub>T</sub> where N<sub>T</sub> = 11 for most patients, correspond to sample collecting days 1, 3, 5, …, 21. Note that for few patients, only a portion of the time series (i.e., less than 11 points) was available. For the EDGE-selected genes, the data were projected onto a smaller dimensional space using the series expansion method similar to principal components analysis, Karhunen-Loeve Decomposition (KLD). In order to preserve the alignment of the time series with respect to day 0 of VAP, we first obtained the average expression vectors, k = 1,2,…9, by averaging the expression values at days corresponding to −3,−1,…,13 in all patients (corresponding to the nine time points that most patients had samples collected). The KLD method looks for a basis ψ<sub>1</sub>,ψ<sub>2</sub>,…ψ<sub>N</sub> so that one can expand aswhere and 〈·,·〉 stands for the standard inner product. The orthonormal basis ψ<sub>1</sub>,ψ<sub>2</sub>,…ψ<sub>N</sub> can be selected as the eigenvectors of the correlation matrix C<sub>1</sub>∈ℜ<sup>N×N</sup>, obtained asThe first principal mode ψ<sub>1</sub> corresponds to a constant bias term. Hence the most important variation is captured by ψ<sub>2</sub>,ψ<sub>3</sub>,…ψ<sub>N</sub> respectively in the decreasing order. Once the orthonormal basis is obtained, each patient data can be projected onto this basis as for patients j = 1, 2, …. The discrete derivative of the coefficients ### Validation of microarray results Select genes were subjected to real-time quantitative PCR (RTq-PCR) for independent confirmation of relative expression levels. cDNA was generated and 100 ng was subjected to routine SybrGreen RT-PCR as per manufacturer's instruction (Applied Biosystems, Foster City, CA). In addition, 85 genes were selected at random from the total number on the GeneChip. The gene expression signal from these genes was analyzed in a manner identical to that described above. This procedure was repeated 100 times to estimate the informational value of randomly selected genes. ## Validation of the riboleukogram After enrollment of the first 20 patients, a second cohort of 7 patients was analyzed to validate the informational content of the leukocyte RNA species (genes) that changed in abundance in response to critical illness complicated by VAP. The blood handling, processing, and GeneChip analysis protocols were identical to those described above. # Results ## Murine model Intratracheal (i.t.) installation of saline caused no deaths over 7-days whereas i.t. introduction of P. aeruginosa endotoxin resulted in death of 90% of the animals studied within 96 hours. The 7-day mortality caused by live P. aeruginosa was adjusted by varying the size of the inoculum and injuries causing 50% and 90% 7-day mortality were achieved. A dose of S. pneumoniae was given i.t. that resulted in 85% 7-day mortality. Once these injuries were established, three separate cohorts of mice were used for each experimental group in the subsequent studies. Peripheral blood was collected and pooled from groups of 4–7 mice 24 h after surgery, prior to appreciable mortality in any group. All murine RNA were of good quality based on the peak profiles of 18S and 28S ribosomal RNA. cRNA generated from these samples had a uniform size distribution. All hybridizations were of good quality; both the number of features present (35–40%) and the signals on each array fell within acceptable ranges. Analysis of normalized gene expression data identified 219 probe sets (40 unannotated ESTs, 10 redundant probe sets, 169 annotated genes whose expression levels differentiated between the five groups. Leave one out cross-validation using k-nearest neighbors (k = 2) resulted in a 93% classification accuracy. The single misclassified sample was from the low-dose P. aeruginosa infection and was classified as “saline”. These 219 genes differentiated between the host responses to Gram-negative bacteria, Gram-negative bacterial toxin (LPS) and Gram-positive bacteria. The probe sets clustered into six groups and these groups defined the gene expression cartography of the murine response to pneumonia. Genes that fell within clusters 2 and 3 exhibited increased RNA abundance in animals responding to high lethality insults. Genes that fell within clusters 4–6 were transcriptionally suppressed during high lethality insults. Cluster 1 bridged the two groups. Gene ontology assignments identified enriched molecular functions in the three distinct groups. The bridging cluster (cluster 1) was enriched for genes involved in the immune response (N = 10, P = 4×10<sup>−7</sup>) and genes with NTPase activity (N = 6, P = 5×10<sup>−4</sup>). Clusters 2 and 3 were enriched for genes involved in intracellular signaling pathways (N = 18, P = 6×10<sup>−7</sup>). Clusters 4–6 were enriched for genes that encode nuclear proteins (N = 24, P = 9×10<sup>−4</sup>). Principal components analysis of the microarray-derived transcript abundances of the genes selected based on cross-validation clearly differentiated the five experimental groups based on the 7-day mortality (principal component 1) and the type of agent used (Gram positive, sterile, Gram negative, principal component 2). ## Clinical study—training cohort After being mechanically ventilated for \>48 h, 27 patients were enrolled into the study. The first 20 patients enrolled were assigned arbitrarily to a training cohort; the other 7 were assigned to a validation cohort. Blood samples were taken at ∼48-hour intervals during the study period and then separated into plasma and leukocytes (see below). Of the 20 patients in the training cohort, eight patients either were extubated without developing VAP or withdrew from the study. Of the 12 patients in this cohort who developed microbiologically- confirmed VAP , 11 met our analysis criteria of having samples before and after the attending physician made a diagnosis of VAP (that is, one patient, \#9, was excluded from analysis for developing VAP on the study entry day). Clinical pulmonary infection scores (CPIS) increased in all 11 patients coincident with the diagnosis of VAP. Three of the patients were culture positive for a Gram positive agent (S. aureus) and the remaining eight patients were culture positive for one or more Gram negative agents. In every case, initial intravenous antibiotic therapy was appropriate for the cultured organism responsible for VAP, based upon cultured organism susceptibilities. Nine of the 11 patients developed VAP 3–6 days after enrollment in the study (“early VAP”) while two patients developed VAP after prolonged mechanical ventilation (“late VAP”). All of the patients survived and were discharged from the ICU. Patient- specific timelines were aligned for analysis by assigning “day 0” to the day that the attending physician diagnosed VAP. RNA isolated from patient samples was of high quality and hybridizations met standard performance criteria (vide supra). To assess whether the genes identified in the murine model conveyed information in the patient study, the microarray abundance of the human orthologs of the 219 genes that distinguished the murine pneumonias were numerically analyzed. Principal components analysis of the average RMA-normalized expression levels of these 109 ortholog genes resulted in gene expression trajectories that described the cohort of patients as they developed, were treated for, and recovered from VAP. Trajectory translocation along the X-coordinate (principal component 2) appeared to be informative with regard to the onset of VAP – beginning immediately before the diagnosis and ending approximately six days after appropriate antibiotic therapy was initiated. Principal components analysis of plasma cytokine abundance in these patients showed a qualitatively similar trajectory, but with large error bars. Nevertheless, translocation along the X-coordinate (principal component 2) again appeared to coincide with the onset of VAP. Independent analysis of patient microarray data resulted in the identification of 85 probe sets whose abundance changed significantly during the course of VAP. Of the 109 human orthologs that were used to calculate the trajectories shown in, probe sets (4.6%) were present in the list of human probe sets (lactotransferrin, cathelicidin antimicrobial peptide, phospholipid scramblase 1, inhibin beta A, and hydroxyprostaglandin dehydrogenase 15-(NAD)). Network analysis found that the expression behavior of these 85 genes segregated into four clusters. Transcript abundance in clusters 1 and 2 generally increased and transcript abundance in clusters 3 and 4 generally decreased around the time of VAP diagnosis. Molecular cartography of the human leukocyte transcriptional response to acute bacterial infection identified two densely connected networks of genes, the first containing clusters 1 and 2 and the second containing clusters 3 and 4. The 26 probe sets in clusters 1 and 2 are significantly enriched with GO biological process terms: “defense response to bacteria” (8 genes, p-value = 2×10<sup>−11</sup>), “response to biotic stimulus” (12 genes, p = 4×10<sup>−6</sup>), and “immune response” (8 genes, p = 0.002) and the cellular compartment term “extracellular region” (14 genes, p = 9×10<sup>−9</sup>). The 59 probe sets in clusters 3 and 4 are enriched with GO molecular function terms: “ATP binding” (13 genes, p = 9×10<sup>−4</sup>), “metal ion binding” (22 genes, p = 0.002), and “protein binding” (25 genes, p = 0.002) and the cellular compartment terms “cytoplasm” (23 genes, p = 0.007) and “plasma membrane” (13 genes, p = 0.02). Principal components analysis of the microarray expression profiles of these 85 genes defined a common response to pulmonary infection complicating critical illness. Importantly, trajectory translocation in the X-coordinate (principal component 2) occurred days prior to the clinical diagnosis of VAP. In addition, the informational content of 85 genes chosen at random was determined iteratively 100 times for the first 11 VAP patients. As shown in, only the 85 genes identified by EDGE as significant (FDR≤0.10) produced a trajectory; the other sets of genes were scattered randomly around the origin of the graph. Three patients exhibited contrary gene expression profiles within two of these four clusters. Patients 1, 7 and 11 showed decreased expression of genes in cluster 2 and patients 7 and 11 showed increased expression of genes in clusters 3 and 4. Based on patient demographics ( and data not shown), the only clear difference between patients 7 and 11 and the other patients in the study is that these two patients developed VAP later in their ICU course (study day 18 and 14 respectively) whereas the remaining patients developed VAP between study days 3 and 6. This was also evident in PCA analysis of the 11 individual trajectories (data not shown). The changes in transcript abundance of selected genes were validated by quantitative RT-PCR. Finally, we tested whether host ethnicity, host gender, host age, or the cell wall phenotype of the infectious agent had an effect on the number of informational genes. The most informational of these demographic variables was host ethnicity. We observed in that the aggregate riboleukogram variance in principal component 2 decreased as patients recovered from acute infection. This finding suggested that principal components analysis of microarray-measured gene expression described an attractor, as gene expression time series can be described in terms of dynamical system theory as trajectories in the phase space defined by the main principal components. By plotting the change in PC2 against PC2 over time, we found indeed that the mapped gene expression information appeared to converge toward a common point, suggesting that PC2 represents the expression of the infection-inducible genes. Consistent with differences in patient age, gender, ethnicity, pre-existing co-morbidities, and nature of injury insult, each patient's individual trajectory started at a different point and described a patient-specific arc (data not shown). ## Validation cohort A second cohort of 7 patients was analyzed to evaluate the informational value of the 85 genes that were identified in the first 11 patients with VAP. Two of these 7 additional patients were diagnosed with “late” VAP by the attending physician, while another 2 cases were described by the attending as “possible” VAP. The individual riboleukograms for these 7 patients demonstrate the existence of immune recovery (basins of attraction) as well as the heterogeneity of the host response. In general, the individual riboleukograms follow a path moving from left to right, that is, from critical illness to recovery (green and red shaded areas, respectively). The development of an infectious complication is typically associated with a change in riboleukogram trajectory. For example, the paths of patients 13, 14, 15, 16, and 17 change directions abruptly coincident with changes in clinical status and concern for VAP or sepsis. Patient 17 grew Staphylococcus aureus from both urine and tracheal secretions prior to withdrawal of therapy for cure (the only death in the study). In contrast, the riboleukograms of patients 13, 18, and 19 are atypical, in that their paths do not start and/or do not finish with the others. Both patients 18 and 19 had pulmonary contusions secondary to polysystem trauma, maximal CPIS scores of 7 and 9, Gram-positive cocci cultured from airway secretions, and were treated with antibiotics, but had a clinical course labeled by the attending physician as “possible” VAP. Their riboleukograms are in different portions of the graph in, but have a similar shape and slope. Both patients 13 and 18 had intracranial hemorrhage. Patient 13 was not diagnosed with VAP (no infiltrate on CXR) but was treated with antibiotics for a fever of 39.4°C and WBC of 31,300 (day 5), tracheal secretions that subsequently grew out Acinetobacter and Stenotrophomonas (CPIS 6), and concern for catheter-related sepsis. In, the aggregate riboleukogram for the training cohort of VAP patients (aligned for day of VAP diagnosis, see also) is compared to the aggregate riboleukogram for all patients aligned for day of study entry (that is, both training and validation cohorts, N = 11+7, irrespective of VAP day). Again noted are the PCA domains of critical illness and recovery. The aggregate VAP riboleukogram diverges from the aggregate critical illness riboleukogram and rejoins the latter at the point of recovery. # Discussion Using a bench-to-bedside approach, we have implemented a mouse model of pneumonia and found that RNA abundance profiles obtained from blood samples taken prior to appreciable mortality were able to distinguish between the two variables tested in the assay: lethality of the insult and type of infectious agent. These data extend our previous observations in a mouse model of abdominal sepsis, wherein microarray-measured expression profiles from circulating leukocytes distinguished between infectious and non-infectious etiologies of systemic inflammation in a de-identified cohort. Thus, the mouse circulating leukocyte transcriptional response to infection can not only distinguish between infectious and non-infectious inflammatory insults, but also the type of infectious agent and its associated mortality. Network analysis suggested that the pneumonia-induced transcriptional changes reprioritized mouse leukocytes for the initiation of an immune response, the transcriptional regulation of intracellular signaling cascades, and the induction of numerous transcription factors and other nuclear genes. Results from the mouse model suggested that the transcriptional activity of buffy coat-isolated cells may be used to monitor the onset of and recovery from acute infection. We tested this hypothesis by calculating principal components using microarray expression profiles of RNA isolated from mechanically ventilated patients at risk for pneumonia. Initially, we examined the behavior of the human orthologs to the genes identified in the mouse pneumonia study. The onset of acute infection corresponded with translation along PC 2 in. Translation along PC2 ceased 5–6 days after appropriate antibiotic therapy in the patients was started, consistent with recovery. These data show that there are specific transcriptional programs instituted by circulating immune cells during acute infection which have diagnostic potential in the setting of critical illness. Principal components analysis of plasma cytokine abundance generated a qualitatively similar trajectory; however, in line with previous reports, plasma cytokine abundance (including procalcitonin) was insufficient to diagnose acute infection in this small cohort of critically ill patients. As with other tissues, changes in RNA abundance observed in circulating leukocyte do not necessarily reflect changes in protein abundance, and vice versa. While principal components analysis of informational murine genes in authentic human disease showed there was a conserved and informational peripheral leukocyte transcriptional response to localized infection, the information contained in those genes would not appear to be more useful than current clinical criteria (that is, the translation in PC2 did not begin until the day the attending physician made the diagnosis). However, by explicitly accounting for variance over time, a set of 85 genes were identified subsequently in our first 11 VAP patients whose microarray expression levels changed consistently before the clinical diagnosis of VAP. These 85 genes clustered into four groups with their abundance either increasing or decreasing throughout the 7-day window bracketing the onset of infection. There was a significant association between genes known to play key roles in defense against bacterial pathogens and those genes that increased in apparent abundance (cluster 1 and 2 probes sets) coincident with the diagnosis of VAP. Consistent with activation of the host response to pneumonia, all of the genes with the “defense against bacteria” ontology were found in cluster 2 and encode primarily granulocytic, antimicrobial proteins, and adhesion molecules. In contrast, the genes that decreased in apparent abundance (clusters 3–4) showed different behavior depending on whether the patient developed VAP early or late in the study. Although no consistent biological theme emerged from this list of the 59 transcripts, this finding provides some insight into the transcriptional basis of differences in the critically ill host's response to early- versus late-onset VAP. Of interest, the RNA abundance of five genes were altered by pneumonia in both the mouse and human gene sets: lactotransferrin (LTF), cathelicidin antimicrobial peptide (CAMP), phospholipid scramblase 1 (PLSCR1), inhibin beta A (INHBA), and hydroxyprostaglandin dehydrogenase 15-(NAD) (HPGD). Pathway analysis indicates that these five genes connect in a network rich with interactions between important mediators of inflammation, as seen in and, (Ingenuity Pathway Analysis, Ingenuity Systems, Redwood, CA). Principal components analysis of the microarray expression data for the 85 transcripts generated a curve with qualitative similarities to curves derived from either cytokines or the 109 human orthologs; however, the patient-to-patient variance in the measured abundance of these genes was significantly smaller and translation along the second principal component preceded the diagnosis of VAP by 24 to 72 hours. In contrast, equivalent analysis of randomly selected sets of genes provided no information about the host response to critical illness complicated by pneumonia. Macroscopically, pneumonia is diagnosed when there are symptomatic changes in clinical status, as manifested typically by increased CPIS scores. However, microscopically pneumonia occurs at the transition from colonization to infection concomitant with failure of multiple host barriers protecting the bronchoalveolar epithelium. These events lead to tissue exposure to bacteria and bacterial products, and in many cases to toxicosis and bacteremia. Our data suggest that during VAP, circulating leukocytes are exposed to bacterial products 24–72 hours prior to diagnosis by an attending physician. Initiation of antibiotic therapy earlier during this time window is expected to significantly improve outcome. These findings were further evaluated in a small validation patient cohort, confirming the value of the 85-gene riboleukogram to quantitate the host response to and recovery from critical illness. Individual riboleukograms changed direction coincident with clinical findings of pneumonia or sepsis (frequently days before the clinical diagnosis and maximal CPIS scores). Nevertheless, marked heterogeneity was observed in some patient responses that could not be linked to monitored variables (e.g., patients 13 and 18). In these patients, perhaps the well-described influence of the underlying acute illnesses (pulmonary contusion and intracranial hemorrhage) on host immune responses could provide part of the explanation. As patients healed from critical illness, the riboleukograms converged, a finding consistent with the existence of an immunological “attractor” state. Comparison of the aggregate VAP versus critical illness riboleukograms indicated that a large portion of the signal for the 85 genes identified is a reflection of recovery from critical illness (PC2). Nevertheless, VAP signal became evident among the “noise” of critical illness once the individual riboleukograms were aligned for day of VAP diagnosis, evident as a deflection of the VAP riboleukogram upward in PC3. Thus, we submit that riboleukograms are a molecular analytical tool with substantial potential to improve diagnostics, prognostics, and our understanding of the host response to critical illness complicated by acute infection. An important consequence of our observation that patient-specific riboleukograms converged is that the variance in leukocyte gene expression for these 85 genes decreased significantly over the time course in patients with VAP. Studies of physical systems suggest that the stable states of networks can be represented as attractors, a set of points in the phase space to which the genetic network evolves over time. In particular, every trajectory initiated within the bounds of an attractor terminates inside the attractor. Recently this hypothesis has been confirmed experimentally in vitro. Our findings suggest that the immune system of a critically ill patient who recovers from disease returns back to a stable state and that the immune response trajectory can be a considered formally as a dynamic system. Borrowing from concepts in the physical sciences, we hypothesized that an individual's baseline immune system is within a basin of attraction prior to injury, nosocomial infection perturbs that system, and after the infection cleared the system would return to it's initial attractor. In phase space analysis, we found that the patient-specific trajectories appeared to converge and that the onset of pneumonia ‘pushes' the gene expression data away from that basin—a perturbation that is ameliorated by appropriate antibiotic therapy. This result provides what we believe is the first evidence in patients for a basin of attraction for the genetic network associated with immunological health. Because of the heterogeneity of this patient population (for example, age, gender, pre-existing health status, type and severity of critical illness), it is not surprising that the error bars on the PC analysis are initially quite large. The end-points of the trajectories, however, appear to converge in a smaller space, suggesting that the patients' immune systems are returning to a health attractor. As observed in the validation cohort, patient-specific trajectories are not smooth curves, but appear to have inflection points that may correlate with hospital intervention or the onset of infection. Our data further suggest that differences in host ethnic backgrounds and gender are more important than differences in infecting organism as determinants of the host leukocyte transcriptional response. These observations are consistent with a recent study comparing GeneChip signal from cells lines derived from Asians and Caucasians, indicating differences in the gene expression levels of 25% of the 4000 genes studied in these two ethnic groups. Moreover, as the rate of sepsis, the youngest age of sepsis onset, and the sepsis mortality rate are highest for African-American males, our data suggest that further study of riboleukograms is indicated to gain insight into these health disparities. These profiles also suggest that there is a transcriptional component that is conserved across that diversity of the human population. In summary, our analysis demonstrates the plasticity of the blood leukocyte response to bacteria in vivo, extending previous in vitro studies, into the clinical realm. For the first time, we provide evidence that riboleukogram gene expression analysis can be applied to a heterogeneous clinical population to monitor the host response to and recovery from critical illness complicated by acute infection. Moreover, we identify new, conserved gene targets that appear to be informational for recovery at the transcriptional level, many of which are involved with granulocyte maturation and chemokine (not cytokine) responsiveness. This conclusion is supported by the relative weak diagnostic information provided by the plasma cytokine data, consistent with conclusions reached at recent a consensus conference. In particular, plasma procalcitonin levels were not informational. There are, however, important limitations to our study, most notably, the preliminary nature of the work, the small number of patients, and the well-known difficulty of ruling out false-positive and false- negative diagnoses of VAP based on clinical parameters. This uncertainty is both the motivation for (and challenge of) developing novel molecular diagnostics for VAP. In addition, there is no consensus on how to leverage the dynamic nature of the clinical, RNA, and protein data collected to build hybrid models that improve diagnostics and prognostics. Thus, we conclude that our data demonstrate the technical feasibility and clinical potential of the riboleukogram approach, but proof of clinical utility will require further study. Thus, as graphs of myocardial electrical information (electrocardiograms) were tapped over a century ago to provide an objective means to aid heart diagnostics, we submit that riboleukograms will aid in the diagnosis and prognosis of acute infectious and inflammatory disease. The diagnostic potential of riboleukograms is supported by two very recent independent reports that corroborate our mouse data, indicating that circulating leukocyte RNA signatures in patients differentiate between the host responses to sterile versus infectious causes of systemic inflammation and between the host response to Gram-negative versus Gram-positive pathogens. Prospective clinical trials are indicated to validate our results and determine the value of this new technology; to optimize gene selection methods that account for differences in patient ethnicity, gender, and age; and to develop computational approaches that integrate clinical and molecular data to improve diagnostics. # Supporting Information The authors wish to thank Drs. Daniel G. Remick and Derek C. Angus for providing cytokine and procalcitonin measurements, respectively. We also acknowledge the technical and clinical expertise of Sandra MacMillan, Douglas Oppedal, Diane Salamon, Barbara Anderson, and Sharon Daube. [^1]: Conceived and designed the experiments: JC MC JM CC WD BG. Performed the experiments: JM KH. Analyzed the data: JC JM KH AP NL QL AB JR WZ BG DD WS. Contributed reagents/materials/analysis tools: JM AP NL QL AB JR WZ BG WS. Wrote the paper: JC JM AP AB WZ CC WS. Other: Designed the study: JC. [^2]: Current address: Pacific Northwest National Laboratory, Richland, Washington, United States of America [^3]: Current address: Department of Mathematics and Statistics, Texas Tech University, Lubbock, Texas, United States of America [^4]: The authors have declared that no competing interests exist.
# Introduction Democracies are supposed to produce policies that align with the public’s preferences. Although the ideal form of democracy provides ordinary citizens direct control over policy and governance, in most modern democracies policy making is left to politicians and their staffs. In recognition of the public’s putative centrality in democratic governance, studies of policymaking treat politicians and their staff as well-honed, career-minded machines who strategically select behavioral responses to electorally created incentives. Nonetheless, politicians (and their staff) do not always have an accurate perception of what their constituents want and fill in the gaps by projecting their preferences onto their constitutes. As a result, policymaking can end up reflecting politicians’ wants and desires, even if elected officials are trying to be faithful representatives to their constituents. It is, therefore, crucial to understand how politicians form policy preferences. Canonical models of policymaking presume that policymakers process information in a uniform fashion—given the same facts, they should reach the same conclusions. Differences in preferences, then, should simply reflect differences in political philosophies. However, recent research shows that psychological biases are more important for elite decision making than canonical models of policymaking presume. Although there is an understandable tendency to treat policymakers as special—after all, they are an elite group of policymakers—they are also only human. Rather than treating policymakers as well-honed machines who process information in a uniform way, we look to the study of how ordinary citizens form political preferences to gain insight into the psychology of policymaking. According to decades of research, ordinary people do not process information uniformly. Differences in genetics coupled with early life experiences lead to differences in the architecture of brain and, thus, the way in which people process external information. In the realm of politics, deep-seated differences in how individuals process information profoundly shapes political attitudes. A burgeoning area of research focuses particular attention to how people process perceived threats. Some people greet reports of terrorism, crime, and the like with a great deal of alarm, while others take them in stride. Recent work suggests that people who are highly sensitive to threats come to see the world as a dangerous and threatening place and tend to see conservative policies—such as adherence to conventional morality and support for a strong military—as a way to cope with and manage the dangerous and threatening world in which they live. These tendencies are evident in the association between conservatism and higher levels of electrodermal activity (i.e., sweating) in response to threatening images. Studies of neural activity suggest that the brains of people who hold conservative attitudes may be more attuned than liberals’ brains to potentially threatening stimuli. # Data and results This research was approved by Texas A&M IRB and was administered by Professor Johanna Dunaway who is on faculty at Texas A&M. We extend this line of research to the study of policymakers in the United States. We are aware of no study that investigates this central claim. Although one study shows that center-right political elites in Iceland offered more conservative policy opinions when induced to feel threat, it leaves unresolved whether politicians differ in how they process and respond to threatening stimuli. We evaluate the effects of threat sensitivity on policymakers’ policy preferences by recruiting 173 state legislators and their staff at the 2016 National Conference of State Legislators to participate in a brief study. The study consisted of a survey that measured state policymakers’ spending preferences and a standard protocol for measuring physiological responses to threatening stimuli. In particular, we measured threat sensitivity by sitting participants before a laptop, behind a privacy partition, and asked them to wear noise-canceling headphones to minimize distractions. Biosensors were attached to the first to third fingers on their non-dominant hand to capture skin conductance. Participants were then presented with four images in succession. The first two images were selected to be neutral and non-affective so that we could capture participants’ baseline physiological activity (a basket and a spoon, presented in random order). The second set of images were selected for their threatening content (a snake lunging at the camera and a close-up of a angry, barking guard dog, presented in random order). The skin conductance levels (SCL) captured during the images allow us to measure participants’ electrodermal activity when faced with threatening non-political images. Because electrodermal activity is difficult to regulate consciously, it offers a valid and reliable measure of people’s sympathetic nervous system response. See for details about the survey and threat sensitivity protocol. shows the distribution of our measure capturing threat sensitivity, which is the difference in SCL observed during the two threatening photos and the SCL observed during the two neutral photos. Higher values indicate greater sensitivity to threat. The distribution clearly leans to the right (i.e., towards threat sensitivity), but there is a good bit of variation as well. The survey asked policymakers to indicate what percent of their state budget they would devote to six policy domains that consistently divide liberals and conservatives: assistance to the poor, elementary and secondary education, counter-terrorism, health care, higher education, and police and public safety. Four of these tend to be liberal priorities: spending on education, higher education, health and the poor; two others tend to be conservative priorities: spending on defense and terror. Our measure of spending preferences captures the conservative-liberal difference by asking policymakers to indicate the percent of the state budget they would allocate to each domain; and then capturing policymakers’ spending priorities by taking the average spending preference on the two conservative priorities and subtracting from it the average spending preference on the four liberal priorities. We find suggestive evidence that threat sensitivity correlates with a preference for devoting a higher portion of the state budget to conservative spending priorities vis-a-vis liberal ones (β = 78.1, SE = 133.9, P = 0.28), even after controlling for participants’ partisan affiliation (β = 77, SE = 133.2, P = 0.29). (See Table B in for details). Although these findings corroborate findings from layperson samples that threat sensitivity correlates with conservative policy preferences, we cannot rule out sampling variability as a possible explanation. Although the six policies that are part of this general measure clearly relate to conservative-liberal policy priorities, we expect that some are at best only loosely related to threat sensitivity. Pushing beyond these preliminary but suggestive findings, then, we focused the analysis on the degree to which policymakers prioritize spending on counterterrorism to spending on assistance to the poor. We do so for three reasons. First, these policy domains are primarily national, not state-level, and, therefore, do not differ widely across states. Second, these issues lie at the center of partisan polarization in the US that previous scholars have traced to differences in threat sensitivity, and recent survey research shows that counterterrorism and welfare stand out in this regard. Third, terrorism is the policy domain that should be most clearly rooted in concerns about a looming, external threat. Welfare, in contrast, is the social domain that is most clearly involves an empathetic giving to others, even in the face of moral risk. Consequently, the terror-welfare tradeoff—capturing spending on threat versus spending on targeted income redistribution—should be more directly connected to threat sensitivity than the general spending priorities measure. The results reported in confirm these expectations. Here we observe a positive and statistically significant relationship between threat sensitivity and policymakers’ tendency to prioritize counter-terrorism over welfare spending. This relationship is unaffected when partisan affiliation is added to the model (Column 2 of). plots the estimated relationship between threat sensitivity and counter-terrorism spending priority, from Model 2 of. Moving across the observed range of threat sensitivity is associated with a shift in a preference for prioritizing counter-terrorism over welfare spending by 30 percentage points. On a measure that ranges from -56 to +33, with a standard deviation of 16.2, this is a rather striking relationship. We regard the tradeoff between counterterrorism and welfare as the most relevant for our purposes, but note that our results are not dependent on these two domains alone. One obvious change is to replace counterterrorism with crime prevention. There will more regional variation in crime-related spending; and crime prevention is less directly connected to threat sensitivity in the existing literature. Even so, results using this slightly revised measure produce roughly similar results. We include those results in Table C in. The empirical evidence that we present suggests that the genesis of policy elites’ spending preferences is more complicated than theoretical models assume. When these individuals are making budgetary tradeoffs, threat sensitivity correlates with prioritizing spending on counter-terrorism to welfare spending—the classic guns versus butter tradeoff. These findings challenge the notion that elites’ preferences can be simply reduced to strategically selecting behavioral responses to environmentally created incentives as well as the notion that strong incentives for policy elites to be cold and calculating obviates the influence of psychological needs. Instead we find that, like ordinary citizens, threat sensitivity plays an important role in state-level policymakers decisions to prioritize spending on government polices that are designed to minimize threats. To be clear, this is not to say that these elites are not being “rational.” They could certainly be behaving rationally in the sense that they are connecting their psychological needs in a logical way to their policy preferences. What these findings show is that elites—like all humans—vary in the way in which they evaluate the same information, leading them to adopt different policy preferences. Individual differences in information processing have not been incorporated into standard theoretical models of elite decision-making, and our findings suggest that perhaps they should. Doing so may provide insights into why some policymakers systematically perceive their constituents to be more conservative than they actually are. Like all studies, ours has several limitations. First, in order to preserve anonymity, we did not collect personal information about these participants, such as the districts that they represent. It could be that particular constituencies elect representatives with psychological needs that mirror the modal member of their district. If so, it would support an “electoral connection” between constituents and representatives, albeit one that does not require the representative to behave in a strategic way. Second, we do not know the level of professionalism of the legislature in which these policymakers inhabit or their level of progressive ambition. Some state legislatures are highly professionalized miniature versions of the US Congress, while others are filled with mostly part-time citizen legislators. Policymakers from less professionalized legislatures may behave in less strategic ways. On this point, individual legislators—irrespective of the type of legislature in which they work—vary in the degree to which they have ambitions to progress up the electoral ladder to a higher office. Perhaps more ambitious individuals will behave in a more calculating fashion. Furthermore, more research is needed regarding the psychophysiological measurement of threat sensitivity and, more generally, negativity bias in this domain of political science. We chose images that would directly tap into threat and, therefore, feelings of anxiety and fear. However, other researchers have used a variety of images to tap threat sensitivity that may also tap into disgust—making those measures better thought of as a general indicator of negativity bias that includes threat and disgust sensitivity. We find it noteworthy that our more tailored images point to the same conclusions as a more general measure of negativity bias. Nonetheless, we cannot speak to the specific effects of threat and disgust—whether both forms of negativity bias contribute equally to inducing conservative attitudes or whether they operate differently in different domains. Despite these limitations, we believe that our study makes an important contribution to the study of elite behavior. Scholars should take the psychological needs of policymakers more seriously. In short, we find that policymakers are people, too. Additional research should be directed at understanding the conditions that moderate and mediate the influence of psychological needs on elite decision making and behavior. # Supporting information We thank Dan Butler and the Laboratories of Democracies for giving us access to study participants, and Temple University, the University of Michigan and Texas A&M University for financial support that made this research possible. We also thank Sarah Anderson, Nicholas Ryan Davis, Mona Vakilifathi, and Laurel Harbridge Yong for their help in the design and implementation of the study. [^1]: The authors have declared that no competing interests exist.
# Introduction Proteins play a vital role in living organisms. They are the main players in metabolic pathways, and to understand how cells work requires insight into the structure and knowledge of the function of a protein. A plethora of research to determine the structure and function of proteins has been conducted, but the rate of knowledge generated has grown much more slowly than the universe of identified proteins has. For example, between two releases of the UniProt Knowledgebase (from release 2015_11 to 2015_12) only 287 sequences were added to the Swissprot section, which contains curated protein sequences with high levels of annotation. Over the same time period, the TrEMBL section (for automatically annotated sequences) has been augmented with more than 590,000 protein sequences. The need for more accurate and scalable automatic methods is compelling. Variation in proteins comes from recombination and mutations in evolutionary modules. These modules are generally known as domains. A single protein is made up of one or more domains. Thus, the detection of domains of protein sequences can be regarded as the initial step in domain family identification and protein clustering, which in turn can help in function and structure assignments. Accurate annotation of conserved regions, as building blocks of protein sequences, can also contribute to construction of evolutionary trees. Many current approaches perform the detection of sequence domains by detecting only putative regions and as a part of domain family identification, i.e., protein clustering based on domains rather than as a separate question. To the best of our knowledge there is no published research concentrating only on detecting conserved regions in proteins. Although correlation between the detection of domain regions and clustering based on domains appears to be natural, one can argue that a cluster is a global property of the universe of proteins, but a domain is a local property of each sequence which, as the definition suggests, needs to be common to many sequences. Global approaches to domain family identification also often involve the expensive operation of sequence alignment. Hence, fast and accurate detection of domain regions will contribute to the global operation of clustering proteins, both by improving the accuracy of detection and reducing runtime by removing expensive sequence alignment operations. While the unprecedented growth in numbers of sequenced proteins has been a challenge for integration of automated methods in proteomics, the abundance of sequences also presents us with an opportunity: as the number of protein sequences containing domains increases, the frequency of occurrence of a particular domain also increases, making it easier to detect. Our task, then, is to extract high frequency regions from within the protein sequences. As we explain fully below, for each sequence, our alignment-free method translates its level of similarity to other sequences into a vector and then looks through each vector to detect the pattern within its indices. While we introduced the idea of this translation into a vector in, our experiments showed that our simple mathematical approach resulted in a heavy inter-dependence between the parameters selected for the method and the protein sequences. Here, we improve upon this previous work by integrating a machine learning approach that reduces the dependence on parameters while improving the accuracy. We have also augmented our work with an extensive experimental results section in which we not only validate our proposed method but also provide comparison of our results with the results of other existing methods. ## Related Work While detection of structural domains as opposed to functional domains has been an area of active research featured as a part of CASP experiments, the detection of sequence level conserved regions has been mostly overshadowed by efforts to cluster proteins based on their domain families. In MKDOM2, the shortest sequence (without repetitions) is marked as a domain and then PSI-BLAST is used to detect regions of high similarity with that domain. Sequences containing regions similar to that domain are put in one cluster. The detected sections are removed from the sequences and the same operation is repeated to find other domains. The objective of MKDOM2 is expressed as identification of domain families rather than detection of domain regions themselves. MKDOM2 is an improvement over MKDOM which in turn succeeded DOMAINER. All three have been used to generate the ProDom protein family database. Parallelization of MKDOM2 has been implemented using a master-worker structure. In ADDA, an all vs. all BLAST is performed on the entire data set. Based on the results of BLAST, a tree of putative domain regions is generated for each sequence. An optimization target is then used to find the final partitioning for each sequence. Unification is performed to construct a new graph based on final partitions and to cluster similar proteins together. In ADDA, domain boundary assessments are performed using the trees of putative domains. However, the final goal being domain family enumeration, unification is a crucial step and can possibly improve the partitions. An implementation of construction of ADDA putative domain trees of protein sequences from their all vs. all BLAST results has also been presented with support for parallel/multi-threaded executions. In EVEREST an all vs. all BLAST on the data set is followed by runs of the Smith-Waterman algorithm on the selected sequences to generate a set of putative domain regions. After applying a series of different methods, including clustering the sequences based on detected putative domain regions, filtering low-scoring clusters using a boosting machine learning technique, and profile construction using ClustalW for each cluster, EVEREST finds a set of HMM profiles. The segments detected by these profiles replace the initial putative domain list and correspond to suggested domain families. This operation is repeated iteratively. As can be seen, all three of these approaches focus on domain family identification rather than domain region detection, but they also use some methods for detection of putative domain regions as a preliminary operation for their final objective. In this paper we present NADDA (No-Alignment Domain Detection Algorithm), an alignment-free, scalable method for detection of conserved regions in proteins. In the next section we present our approach to the problem. In the Results section we evaluate our method and compare its results with other methods. We conclude the paper by a brief discussion on what we achieved and prospective future work. # Methods Given a set S of n sequences, the problem of conserved region detection of protein sequences can be expressed as the demarcation of subsequences in each of the n sequences such that these subsequences are preserved over the evolutionary process. The preservation of these subsequences can be identified by a high local similarity score when the protein is aligned with other proteins of the same family. These conserved subsequences appear as motifs and sequence domains. Our method uses the frequency of short exact matching subsequences (k-mers) of a protein sequence in the data set as an indicator of the existence of a conserved region in the vicinity of that k-mer. Problem Definition: Notation: Given a sequence s, we denote the i<sup>th</sup> character of s by s\[i\] and the substring of length l that starts at index i of s by s(i, l). We denote the length of a sequence s by \\|s\\|. A k-mer is a string of length k. Additionally, let S = {s<sub>1</sub>, s<sub>2</sub>, …, s<sub>n</sub>} denote an input set of n protein sequences. We define: Definition 1 k-mer frequency: The frequency of a k-mer in S is defined as the number of sequences in S that contain that k-mer at least once. A conserved window is then a substring of length r, centered at index i, where the average frequencies of k-mers in that substring is above a certain threshold τ. However, this definition is heavily dependent on the values of the threshold (τ) and the window size (r). To compensate for these limitations, we consider a smaller resolution of regions—indices. For each index i we consider a window of size 2 × w + 1 centered at i, and based on the frequency of k-mers inside this window we decide whether or not index i is part of a conserved region. The decision based on k-mer frequencies is not dependent on a single threshold τ; rather we find a set of thresholds {τ<sub>1</sub>, τ<sub>2</sub>, …, τ<sub>2×w+1</sub>} where τ<sub>j</sub> is used as a threshold for the j<sup>th</sup> k-mer in the window, i.e., the k-mer originating at index i − w + j − 1 of the sequence. This set of thresholds is found using our machine learning algorithm, and the prediction is also made by a combination of comparisons based on the thresholds. Definition 2 Conserved Region: Given a sequence s ∈ S, a substring s(i, l) is said to be a conserved region if: 1. (conservation) every s\[j\] for j ∈ \[i, i + l\] is a conserved index, and 2. (maximality) neither the index s\[i − 1\] nor the index s\[i + l + 1\] (if either exists) is a conserved index. From the problem definition above, a simple approach is to compare each pair of sequences ($\begin{pmatrix} n \\ 2 \\ \end{pmatrix}$ pairs) and use the results to annotate the sequences with their conserved areas. To perform the alignment, one can use methods such as the Smith-Waterman local alignment or related heuristics such as PSI-BLAST. However, these approaches are not scalable when \\|S\\| is large. Instead we generate a new representative vector for each sequence—k-mer profile as defined below—and perform a binary classification using feature vectors obtained from this representation. Rather than dissecting the sequence itself we dissect the profile vector into putative conserved regions and map these conserved regions onto the same indices of the sequence. As we will see in the Results section, the detected putative conserved regions widely correspond to the sequence domain regions of the sequences as they are annotated in well-known protein domain databases. Definition 3 k-mer profile: The k-mer profile of a sequence s is a vector of the length of the sequence, where the value of each index is the frequency of the k-mer originating at that index in s. The problem of parsing a profile vector into putative conserved regions can be translated into a classification problem in which the task is to decide whether or not each index of the vector should be included in a conserved region. We generate an instance for each index of the profile vector and use a trained model to predict whether it belongs to a conserved region; this is a binary classification problem. Because the trained model depends only on the frequency values of each index in its profile vector rather than on the amino acid elements themselves, once we have a well-trained model it can be used for de novo detection of conserved regions. The details for each of these steps is explained below. ## Sequence k-mer Profile Generation A k-mer profile is an alternative representation of a protein sequence. To construct a k-mer profile we need to count the number of times every k-mer appears in different sequences in our data set and record the counts at the same index at which the k-mer originates in the profile vector. We approach the problem using a two-pass hashing procedure. First, the algorithm reads every protein sequence and for each position index of the sequence computes the k-mer starting at that index. The computed k-mer is then used as a hash key to store the position index number and sequence ID for that k-mer in the hash table. If a k-mer appears multiple times in a sequence, we store all appearances as one entry in the hash table. Thus, the frequency of each k-mer in the data set will equal the number of entries stored for that k-mer in the hash table. When reading of the sequences has been completed, we enter the second pass of the algorithm. The algorithm uses the sequence IDs stored as hash values in the first hash table as the keys to construct a second hash table. The value of each sequence ID (hash key) in the second hash table is the originating index of each k-mer in that sequence along with the frequency of that k-mer. By ordering these frequencies in the hash table based on the index numbers, we essentially generate the k-mer profile for each sequence. The serial computation time for this step is O(N), where N is the total length of the n sequences in the data set. Because the operation on one sequence is independent of another, the algorithm is easily parallelizable. Later in this section we present a MapReduce implementation. ## Classification Instance Generation At the conclusion of the first step we have a set of k-mer profiles that are vectors with index values equal to the number of sequences in the data set that contain the k-mer initiating on the same index in the sequences. The next step is to generate classificaton instances for each index in each sequence of the k-mer profiles. Conversion of k-mer profiles to classification instances allows us to first train a binary classificaton model using known domain data and then to use the model to predict new domains. Given the k-mer profile p for sequence s, we generate one classification instance per index i in s. For such a profile and index, our classification instance is a vector v consisting of 2w + 1 elements (features) where the 2w + 1 elements are derived directly from the profile vector by copying the values from the window centered on index i of p and extending from both sides for w indices. These features represent the frequency of k-mers initiating at indices \[i − w, i + w\] of the corresponding protein sequence. ## Classification Model Construction We train the classification model using classification instances that include a class label indicating whether or not the index associated with the instance is conserved. This is determined by querying domain databases. We also use these instances to test our model. Given the integer-valued features of our problem, the need to handle large data sets, and the non-linearity of the problem (due to the variable level of conservation among different domains), decision trees seem to be a natural choice to use for training the model. They support automatic feature selection and can learn different thresholds for different features. However, for small w the decision tree will not be able to see the overall picture of the k-mer profile and the model will lose robustness. The tree will be inclined to classify any short-lived increase in frequency as a domain index or a small local fall in frequency as a demonstration of absence of a conserved region. On the other hand, a larger w can result in overfitting the training data. To overcome the disadvantages of a simple decision tree classifier, we use a Random Subspace ensemble method for which the base estimators are decision trees. This ensemble method constructs multiple decision trees on random subsets of features and classifies an instance based on the majority vote of the decision trees. Because we are using k-mer profiles in our training set rather than actual protein amino acid sequences, our model depends only on frequencies (the values in the k-mer profiles) rather than on the amino acids. As a consequence, it can detect new conserved regions that were not present in our training set, given that there are enough sequences in the data set containing the conserved region to ensure the high frequency of k-mers in that region. In other words our algorithm is able to detect conserved regions that are present in multiple sequences in our data set, even if they were not annotated previously. On the other hand, because the model is trained only once and we require accurate performance from that model when presented with new data, the selection of a good training set is crucial to the algorithm. First, the presence of conserved regions in too few sequences in the training set will result in small values in the k-mer profile and can be misleading to the training algorithm. Second, as we increase the size of the training set, we will be adding robustness to the trained classifier because conserved regions with varying frequency profiles will be represented in the training set. Using a large data set acquired from any of the big databases of protein conserved regions as our training set will satisfy both our conditions for the training set and will enable our algorithm to detect new conserved regions. ## MapReduce Algorithm for Profile Generation We implement our algorithm in the MapReduce framework. A MapReduce program consists of two main functions of map and reduce. In the map phase a set of mapper tasks generate KeyValues based on the input. These KeyValue objects are hashed based on their key and redistributed based on the result of the hash. The intermediate hash-based grouping phase is called the shuffle phase. Each reducer task then processes the set of values that are hashed to the same key. A reducer, in turn, can also generate KeyValue objects. The MapReduce paradigm suits data-parallel applications naturally. For our purpose, we exploit the shuffle phase to implement many of our hash-based functionalities in parallel. In our MapReduce algorithm, initially each mapper reads the input set of sequences. It then generates KeyValue pairs using the k-mers of the sequences as keys and sequence ids and k-mer positions as values. The shuffle step subsequently sorts and redistributes these KeyValues such that each reducer receives one KeyMultiValue object, where all the values corresponding to the same key are sent to a single reducer. Based on the number of different sequence ids present in its MultiValue, each reducer determines the frequency of the k-mer mapped to it and emits a KeyValue, where the key is the sequence id and the value is the frequency of the k-mer and the position of the k-mer in that sequence. After this first complete stage of MapReduce, each reducer will contain frequencies of k-mers for each position in one protein sequence and can construct the k-mer profile for that sequence. The algorithm for this step is shown in Algorithm 1. Algorithm 1 Construction of k-mer profile (S:{s<sub>1</sub>, s<sub>2</sub>, …, s<sub>n</sub>}, k) for each sequence s<sub>i</sub> ∈ S with id $s_{id}^{i}$ do for each index j in s<sub>i</sub> do s′ = s<sub>i</sub>(j, k) emit KeyValue $KV = < s^{\prime},\left( j,s_{id}^{i} \right) >$ end for end for shuffle all KV’s for each KeyMultiValue $KMV = < s^{\prime},\left( j,s_{id}^{i} \right),\left( m,s_{id}^{k} \right),\ldots,\left( k,s_{id}^{l} \right) >$ do count = number of different s<sub>id</sub> in the KMV for each value $V = \left( j,s_{id}^{p} \right)$ in KMV do emit KeyValue $< s_{id}^{p},(j,count) >$ end for end for shuffle all KV’s for each $KMV = < s_{id}^{i},\left( j_{1},count_{j1} \right),\left( j_{2},count_{j2} \right),\ldots >$ do sort values based on j’s return the counts based on sorted j’s end for ## Implementation and Software Availability We have implemented our method, NADDA, in C++ and Python. The C++ code uses the MRMPI library for MapReduce and the BOOST library. The Python code uses the scikit-learn (v. 0.17) machine learning library. Software is available as open source at <https://bitbucket.org/armenabnousi/nadda>. # Results ## Experimental Setup Eleven different data sets were used for our experiments, one consisting of approximately 50,000 bacterial protein sequences and the remainder consisting of smaller sets of a few thousand protein sequences each. Some of the latter sets consist entirely of bacterial protein sequences; others are mixtures of sequences from bacteria, eukaryota, and archaea. The number of sequences and percentage of bacterial sequences in each of the smaller sets are listed in. More information regarding the domains in each of these is shown in Tables to. The smaller sets (#1-#10) were used to evaluate de novo detection of conserved regions with our model and to compare it with MKDOM2, ADDA, and InterPro. The large data set \#11 was used to train and test our classification model. Because both MKDOM2 and ADDA require sequence alignment, an expensive operation, it was infeasible for us to conduct a comparison using our largest data set. For example, MKDOM2 had not completed its run on data set \#11 after 20 days. It should also be noted that both methods focus on clustering proteins based on domain families rather than on detection of conserved regions. MKDOM2 clusters are generated based on regions in sequences with high sequence similarity. These regions represent conserved regions as defined earlier. Similarly, as an initial step ADDA selects some putative domain regions. An optimization objective is then used to generate final domain boundaries on each sequence. These boundaries are later used to construct a graph and perform clustering. While each identified region (indices enclosed between two consecutive boundaries) is expected to contain exactly one domain, because they are used as an intermediate result ADDA does not guarantee that a complete domain will be captured in each domain region predicted by the optimization procedure (i.e., it does not guarantee that there will not be any non-domain indices inside the predicted region). However, a comparison can still be made. To obtain our large data set (#11), we queried the SMART protein domain database (<http://smart.embl-heidelberg.de>) using eleven common bacterial protein domains listed in. This resulted in a list of 50,214 bacterial protein sequences that included one or more of the eleven domains. We then queried Pfam-A release 29.0 with each of these sequences and labeled the domain indices to use as ground-truth for our training instances. For the smaller mixed protein data sets, we randomly selected some protein families through the Pfam webpage (<http://pfam.xfam.org>; v.29) and then ran Pfam 29.0 on those sequences to mark their Pfam domain regions. Similarly, for the predominantly bacterial protein data sets, we either randomly identified bacterial protein families through the Pfam webpage or else used domains from the other data sets that had a high bacterial percentage. Again, we marked their domain regions using Pfam 29.0. Tables to show the protein families selected from the Pfam online database and the number of sequences present in each data set. Totals in these tables are computed after removing common redundant sequences. We then queried InterPro (v5.18-57) using each of our 10 smaller data sets. Comparisons were performed based on the coverage of the domain indices reported by Pfam-A in our findings and in the MKDOM2 and ADDA results as well as on the coverage of conserved region indices reported by InterPro in the NADDA results. We measured coverage using the three values of accuracy (AC), specificity (SP), and sensitivity (SN) of the detected conserved indices as defined respectively by: $$AC = \frac{\left\| TP \middle\| + \middle\| TN \right\|}{N}$$ $$SN = \frac{\|TP\|}{\left\| TP \middle\| + \middle\| FN \right\|}$$ $$SP = \frac{\|TN\|}{\left\| TN \middle\| + \middle\| FP \right\|}$$ where TP, true positives, indicate the indices in the data set that are marked as part of a domain region in Pfam-A, and the detection method also has classified them as conserved indices; TN, true negatives, are the indices that both Pfam-A and the conserved region detection method have not included in any domain/conserved region; FP, false positives, are the indices that are not marked by Pfam-A as contained in a domain, but the detection method has classified them as conserved indices; and FN, false negatives, are the indices that are included in a Pfam-A domain, but the detection method has fallen short of identifying them as conserved. Finally, N is the total length of the sequences in the set (N = \\|TP\\| + \\|TN\\| + \\|FP\\| + \\|FN\\|). For our experiments, unless otherwise stated, values for w (where 2w + 1 is the size of the feature vector), k (size of each k-mer), MSS (which indicates the level of pruning of the decision tree—minimum number of instances in each internal node), and max\_features (number of features used in each decision tree for the ensemble method) are set respectively to 10, 6, 100, and 7. We discuss parameter selection in greater detail in the Supporting Information. ## Evaluation of Results We evaluate our results in three phases. First, we evaluate our training method by partitioning data set \#11 into training and test sets. This is the common practice in the machine learning community for evaluation of a training method. The second phase includes comprehensive experiments evaluating overall performance of NADDA by testing it using data sets with domains that are new to NADDA (data sets \#1—#10). Here we also compare NADDA with ADDA and MKDOM2. Finally, in the third phase, we compare NADDA against InterPro, re-evaluating its ability to capture regions that are missed in Pfam. In addition, we provide some example sequences from data sets \#1-#10, comparing our detected conserved regions with the segments annotated by InterPro. Evaluating the training method: In the first phase of our experiments we perform two different types of training-set/test-set partitioning of data set \#11, training the model on a training set and measuring the results on a test set. First, we divide data set \#11 by randomly selecting 20% of the sequences to be included in the test set and the rest in the training set. The second partitioning is performed by selecting 20% of the domains from data set \#11 that are present in more than 50 sequences and including every sequence that contains one (or more) of these selected domains in the test set. The reason for setting a threshold for presence of a domain in a sequence in order to be considered a candidate for the test set is explained in the Supporting Information section. We refer to the first train/test partition as repetitive and the second one as non-repetitive. The reason for the two different partitions is that while our repetitive partitioning method is widely accepted in the machine learning community, here it can result in overfitting. Because our method depends on the frequency of repeating k-mers, we require that multiple sequences possibly very similar to each other be present in our data set. Dividing the data based on random selection of sequences might result in having similar sequences in the training and test sets. The results achieved might then represent the case when the domain in the query sequence has already been discovered in similar sequences and we are only trying to retrieve this knowledge. In contrast, for our non-repetitive partitioning, because we withhold a set of domains from the training set, the results are better representative of discovering new domains that were not previously known as well as domains that were previously known. This partition is better representative of the results we would obtain using the current knowledge-base of protein conserved regions for new sequences. The results for the two sets are shown in. For the repetitive case the accuracy is 83% with 96% recall (sensitivity). For the non-repetitive case, as expected, the numbers decrease. However, we can still predict with 80% accuracy whether or not a location in a protein sequence is included in a conserved region. Comparison with other methods against Pfam: In the second phase of the evaluation we compare our results with final sequence segments produced by MKDOM2 in clustering based on domain families and with final partitions found by the optimization method used in ADDA assuming that Pfam domains represent the ground-truth. As mentioned earlier, MKDOM2 performs clustering by selecting the shortest possible subsequence to be considered as a conserved region and aligning that subsequence with the remaining sequences. The final reported subsequences that have been aligned with each other represent conserved regions on the sequences and can be compared to our detected conserved regions. Similarly, ADDA performs clustering by first selecting a set of putative domain regions and optimizing the boundaries between these provisional domains to construct a protein similarity graph based on detected domains. Although clustering is performed after graph construction and additional filters in these later steps can affect the quality of ADDA clustering, we can use the optimized domain boundaries to compare their pre-computed domains with our method. We trained our model using data set \#11 and tested using data sets \#1-#10. Because other methods depend on pairwise sequence similarity and BLAST results, we were forced to pick smaller sets for testing purposes. As mentioned earlier, MKDOM2 ran for 20 days on data set \#11 without completion, and ADDA took long even for the small data sets as we will show later in Runtime Study section. We used MKDOM2 from the Xdom2.0 package and the ADDA implementation obtained from its webpage (using the default settings: K = 73.70676, C = 8.33957, E = 0.05273, M = 1.417, N = 0.008 and r = 1, except for 100 maximum iterations (i = 100); in many of the cases the program stopped after only a few iterations. We also used BLAST+ v2.2.31 to generate the input files for ADDA). represents our findings. The circle displays our three metrics of accuracy (AC), sensitivity (SN), and specificity (SP) for the three different methods: ADDA (blue), MKDOM2 (green), and NADDA (red). The outermost circle represents 100% of the related measure and the center of the circle represents 0%. For each section (AC, SN, and SP) the bacterial percentage of the protein sequences in the data set for each point decreases in the clockwise direction; the first four points in each section correspond to 100% bacterial data sets. shows the detailed results for the same measurements. NADDA shows higher accuracy for all but two of the test sets when compared to MKDOM2 and higher or comparable accuracy for many of the sets when compared to ADDA. MKDOM2 generally has better specificity scores (average 60.7%) than both ADDA and NADDA while its sensitivity scores are worse (average 51.5%). The advantage of specificity over sensitivity in MKDOM2 may be a result of two factors. First, MKDOM2 uses a high-cut off value in PSI- BLAST for the sake of runtime during its heuristic sequence alignment, thereby trading off sensitivity. Second, the greedy nature of the MKDOM2 algorithm inhibits precise detection of domain regions. ADDA shows higher than 99% sensitivity in all cases but very low specificity (4.6% on average). This is because the optimization method in ADDA tries to detect putative domain boundaries rather than domains themselves and the results are later used for graph construction together with some filters in order to achieve final clustering. As a result, ADDA does not care about mislabeling extra positions as long as the boundary between two domains is preserved. These extra indices could possibly be filtered in the graph construction and clustering process. In comparison to the other two methods, we can see that for 4 out of 10 test sets, NADDA exhibits higher than 80% sensitivity. The sensitivity is significantly low for data sets \#3 and \#4. It is noticeable that for these sets the mean and variance for the k-mer frequency are also smaller than for the other sets. In general there is a correlation between the accuracy of our method and the variance of the k-mer frequency. In fact, the frequency variance depends on our choice of k, and it will increase as k is decreased. However, if k is too small, the variance will be high due to random exact matches that do not signify real conservation. This is discussed further in the Supporting Information section. Comparison against InterPro: In the last phase of our evaluation, we compare conserved indices detected by NADDA with regions annotated by InterPro. InterPro uses predictive models (signatures) generated from multiple databases, including Pfam, SMART, ProDom, prosite, TIGRFAMs, etc. InterPro annotates domains as well as motifs (short conserved regions). As such, there should be more similarity between NADDA and InterPro. We compared our results with InterPro regions. The results are shown in. Because InterPro includes Pfam, replacing our Pfam annotations with InterPro annotations will only add to the indices marked as conserved in our data sets. This can be manifested by a decrease in our \\|FP\\| and \\|TN\\|. Indices that we had marked as conserved but Pfam had not might be detected as conserved in InterPro as well, resulting in a smaller \\|FP\\|. Similarly indices that neither NADDA nor Pfam detected as conserved might be marked as conserved in InterPro, resulting in a decrease in \\|TN\\|. A decrease in \\|FP\\| will result in improved SP. It should also be noted that for this comparison we are still using the model trained using Pfam annotations for data set \#11. Although this can add to the false examples in the training set, the results show an increase in SP as expected, indicating the robustness of our training method. shows some example results from our experiments. Training is performed using data set \#11, while the sequences shown are from the smaller data sets (#1-#10). The domains present in the smaller data sets are not present in the training set. These examples demonstrate that in many of the indices, the output from NADDA matches with the InterPro annotated regions. ## Parametric Study The parameters used in our method are k, the size of a k-mer; w, where 2w + 1 is the number of features used in a classification instance; MSS, indicates the level of pruning of the decision tree; and maximum\_features, number of features used in the ensemble method. The parameters w, MSS, and maximum\_features were set to avoid overfitting and underfitting as explained in the Supporting Information. The optimal value for k was found empirically from the set {3, 4, 5, 6, 7, 8}. We have shown the effect of varying k in the Supporting Information. Varying k affects the mean frequency of k-mers as well as their variance and, thus, is important in the performance of the method. ## Runtime Study We performed a runtime study by comparing the time for a serial run of NADDA with that of ADDA and MKDOM2 and also by a scalability evaluation of NADDA in a parallel environment. We ran the algorithms on our in-house Linux cluster which includes 8 nodes of 64 AMD processors (2.29GHz), each node having 128GB shared memory. We used ADDA and MKDOM2 implementations as described earlier. lists a comparison of runtimes for NADDA, ADDA, and MKDOM2 for data set \#7 on a single processor. The runtime for our method does not include the one-time training time. The ADDA runtime presented in includes the time required for running BLAST. We can see that even for a small set of about 2,500 sequences, NADDA finishes much more quickly than the other methods. We ran the parallel implementation of our code on data set \#11 (50,000 sequences) and showed that it scales nearly linearly with an increasing number of processors. NADDA took 79 seconds using 32 processors to complete this data set given an already trained model. Training was not parallelized and on data set \#1 took 61 minutes. Parallel constructions for decision trees are proposed. A detailed study of our parallel runtime is presented in the Supporting Information section. # Conclusions and Future Work We presented the problem of detection of conserved regions in protein sequences. In the past, multiple methods have been proposed for detection of putative domain regions from protein sequences. However, to the best of our knowledge there is no prior work that focuses on the detection of conserved regions, given a large collection of protein sequences. The method proposed in this paper fills this gap using an alignment-free approach. We showed that rather than using an amino acid sequence of a protein, we can utilize its vectorized representation for our computation. We presented k-mer profiles of proteins as a new representation for protein sequences which can be useful for increasing the scalability of computation. We presented a MapReduce algorithm for generation of k-mer profiles. We used a random subspace ensemble learning method to improve the accuracy of conserved region detection. Our experiments show competitive accuracy, sensitivity, and specificity measures for our method when compared to other methods. We also showed that our parallel implementation is scalable and works on large data sets. Our experiments show near-linear speedup. We showed that our method is able to detect conserved regions of bacterial protein sequences as well as conserved regions of eukaryota and archaea protein sequences. The category of the species does not affect the results of our method. Moreover, we showed that variance of the k-mers from their mean has a correlation with the ability of our method to detect conserved regions. When the variance is low, we may be able to decrease k to obtain higher variance and consequently higher accuracy. Extremely small values of k, however, can result in higher variance due to random exact matches and should be avoided. Use of structured prediction techniques might increase the accuracy of our algorithm. However, a structured prediction model will likely give higher weights to the context (previous predictions) rather than to other features. Another problem with use of different machine learning algorithms is the large scale of our data. The iterative MapReduce algorithm fits best in the Spark parallel computing model which also provides machine learning tools. In future work, we will implement our algorithm using the Spark paradigm. # Supporting Information The authors would like to thank Prof. Janardhan (Jana) Rao Doppa (Washington State University) for his advice on the use of machine learning algorithms. This work was funded by support from the U.S. National Science Foundation under the Advances in Biological Informatics program, Award DBI 1262664. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: AK. Formal analysis: AA. Funding acquisition: SLB AK. Investigation: AA. Methodology: AA AK. Project administration: SLB AK. Resources: AK. Software: AA. Supervision: AK SLB. Validation: AA. Visualization: AA AK SLB. Writing – original draft: AA. Writing – review & editing: AA AK SLB.
# Introduction The treatment paradigm for metastatic renal cell carcinoma (mRCC) has undergone a dramatic evolution over the past two decades. In 1992, interleukin-2 (IL-2) was approved for the treatment of mRCC. Although IL-2 has been shown to lead to durable responses in a small proportion of patients, the vast majority of patients either derive no clinical benefit or are physically too debilitated to receive this intensive therapy. As an alternative, monotherapy with interferon-α (IFN-α) was frequently employed. A meta-analysis of data from IFN-α trials showed modest results at best, with a median time to progression (TTP) of 4.7 months and a median overall survival (OS) of 13 months. At the time these data were published in 2002, it was suggested that IFN-α serve as a reference standard for future clinical trials in mRCC. The introduction of targeted therapies for mRCC shattered this reference standard. A total of seven targeted agents have been approved to date by the US FDA on the basis of phase III data – four vascular endothelial growth factor- tyrosine kinase inhibitors (VEGF-TKIs; sunitinib, sorafenib, pazopanib, and axitinib), one VEGF-directed monoclonal antibody (bevacizumab), and two inhibitors of the mammalian target of rapamycin (mTOR; temsirolimus and everolimus). With the advent of these therapies, IL-2 and IFN-α are presumably utilized to a lesser extent in the mRCC paradigm. It has been repeatedly observed that survival in more recent trials in mRCC has gone far beyond the landmark of 13 months proposed in association with IFN-α. For instance, in the randomized phase III study comparing sunitinib and IFN-α in treatment-naïve patients, a median OS of 26.4 months was observed with sunitinib therapy. Long-term survivors are also increasingly recognized with targeted therapy; in a phase II study of axitinib, up to 20% of patients were still alive 5 years beyond the time of treatment initiation. Although these data provide compelling rationale to suggest that survival has improved since the advent of targeted therapies, this hypothesis has not been definitively proven. In the current study, we queried the Survival, Epidemiology, and End Results (SEER) dataset and performed generational analysis of survival amongst patients with mRCC. With data extending from 1983 to 2009, we segregated our analysis using two clinically relevant time points: (1) the approval of IL-2 in 1992 and (2) the approval of the first targeted therapies (sunitinib and sorafenib) in 2004. # Methods ## Patient Selection The SEER dataset was analyzed for the current study, a registry encompassing approximately 28% of the US population. The SEER Program has extensive data pertaining to demographics, stage, tumor histology, and grade. The current analysis was restricted to patients 18 and older who had a diagnosis of RCC between 1992 and 2009 (n = 60,587). The analysis was further limited to patients with stage IV disease at the time of diagnosis, had a known surgical status, had a known cause of death if deceased, and had a clinically relevant histology (n = 5,150). Notably, the SEER Registry does not allow for capture of patients who progressed from localized or regional disease to metastatic disease, thus confining this analysis to those with de novo metastatic disease. ## Tumor Classification Reporting of tumor histology, grade and stage was done in accordance with the International Classification of Diseases for Oncology (ICD-O), version 3. Given the heterogeneity of histologic classification of kidney tumors in SEER, we limited our analyses to the three most clinically relevant histologic subtypes – clear cell (ICD-O 8310), papillary (ICD-O 8260), and chromophobe (ICD-O 8317). In our analyses, papillary and chromophobe tumors were combined in a category termed ‘non-clear cell’. Tumor grade was characterized as well differentiated, moderately differentiated, poorly differentiated or undifferentiated (a Fuhrman grade is not delineated in the SEER Registry). Stage IV disease was identified according to criteria specified within the American Joint Committee on Cancer (AJCC) TNM staging system, 7<sup>th</sup> ed. ## Statistical Analysis To test the a priori hypothesis that patients diagnosed in the targeted therapy era have a longer survival as compared to patients diagnosed in the cytokine era, we assessed our primary outcome, disease-specific survival (DSS), across two time periods: (1) 1992–2004, and (2) 2005–2009. These time periods reflect (1) the approval of IL-2 in 1992, and (2) the first approval of a targeted therapy for mRCC (sorafenib) in 2005. Clinicopathologic characteristics were compared between the groups using the student's t-test and chi-square test for continuous and categorical variables, respectively. Disease-specific survival was assessed, and defined as the time elapsed between date of diagnosis with RCC and date of death, if attributable to RCC. Patient data were censored at the time of last follow-up if the patient was still alive at last contact. Patients with an unknown cause of death were excluded. The Kaplan-Meier method and log-rank test were used to compare survival across the two time periods. Relevant clinicopathologic characteristics were evaluated for their association with disease specific survival using both univariate and multivariate Cox proportional hazard models. After a thorough review of the SEER methodology, we felt it appropriate to exclude T-stage and N-stage from our analyses. Specifically, in 2004 SEER adopted the Collaborative Stage Data Collections System (CS). Prior to CS, patients with M1 stage disease were rarely coded with specifics on T- and N-stage other than Tx and Nx, respectively. Hierarchical rules in SEER classification implied that T- and N-stage were “trumped” by M1 status. In contrast, from 2004 onwards, rules were set in place that called for recording of T-stage and N-stage despite the notation of M1 disease. All analyses were performed using SAS (SAS Institute, Cary, NC, USA). P-values reported herein are two-sided. P-values of 0.05 or less were deemed statistically significant. # Results ## Patient characteristics by time period Utilizing the aforementioned selection criteria, a total of 2,382 patients were identified from 1992–2004 and 2,768 patients from 2005–2009. Characteristics of the study population are noted in. The mean age was similar across the two study periods (approximately 62 for both) and, as anticipated, a male preponderance was observed in both groups. A greater representation of minority groups was seen in the latter study period as compared to the earlier study period, with an increased proportion of blacks (8.0% vs 7.5%) and Hispanic whites (12.4% vs 15.0%). A larger proportion of patients in the latter study period were noted to have poorly differentiated or undifferentiated tumors (69.0% vs 58.7%). Nephrectomy rates were similar across time periods, with 60.0% of the study population receiving this intervention. As expected, the duration of follow-up was substantially shorter for those patients assessed in the later time period as compared to the earlier time period (14 months and 24.5 months, respectively; P\<0.0001). ## Analysis of survival by time period To test the a priori hypothesis of this work, DSS was compared across the two time periods of interest. As noted in, median DSS was 13 months in patients diagnosed from 1992–2004 compared to 16 months in patients diagnosed between 2005–2009 (P\<0.0001). At both 1-year and 5-year landmarks, survival also appeared to be superior amongst patients diagnosed between 2005–2009 as compared to patients diagnosed between 1992–2004 (57% vs 52% at 1-year, and 22% vs 18% at 5-years, respectively). Given that the majority of systemic therapies have been assessed in clear cell mRCC, we then compared survival in clear cell and non-clear cell subsets. As noted in, reflecting the substantial proportion of patients with clear cell disease, the survival trends were akin to those observed in the overall study population. Amongst patients with non-clear cell disease, as was anticipated, no significant difference in DSS was observed (P = 0.32;). ## Univariate and multivariate analysis As shown in, univariate analyses demonstrated a number of factors beyond date of diagnosis that were associated with DSS. Relative to patients diagnosed aged 18–49, older patients (specifically, patients between 65–79 and ≥80) had shorter survival. Amongst clinicopathologic criteria, poorly differentiated or undifferentiated tumors were associated with shorter survival; no significant difference was noted in survival based on histology (i.e., clear cell vs non- clear cell). Survival was improved in patients who had received cytoreductive nephrectomy as compared to those who had not. On multivariate analysis, time period was independently associated with survival, favoring patients diagnosed from 2005–2009 (HR 0.88, 95% CI 0.83–0.94; P = 0.0001). Complete results of multivariate analysis are displayed in ; as noted therein, older age, female sex and black race were amongst the clinical characteristics independently associated with shorter survival. Higher tumor grade and absence of nephrectomy were also independently associated with shorter survival. # Discussion The results described herein suggest that survival has improved in the era of targeted therapies as compared to the era of cytokine therapy. This suspicion has been strongly held in the academic community for some time, given the substantial improvement in overall survival seen in recent studies assessing systemic therapy in treatment-naïve populations. For instance, in the recently reported COMPARZ study assessing sunitinib and pazopanib in the front-line setting, a median OS of 28.4 months and 29.3 months was observed in each treatment arm, respectively. These data stand in sharp contrast to estimates of survival generated a decade ago where, in the era of cytokine therapies, a median survival approaching 1 year was anticipated. Our review of the SEER registry encompasses a highly heterogeneous array of patients, and suggests a similar survival trend in survival – diagnosis with stage IV RCC between 2005–2009, during which time several targeted therapies were approved by the US FDA (including sorafenib, sunitinib, and everolimus), is independently associated with improved survival. Another recent epidemiologic study utilizing data from the California Cancer Registry has pointed towards a similar improvement in survival; however, this study included non-metastatic patients and further capped analysis at 2007. The latter would likely limit the effect seen from targeted therapies, introduced from 2005 onwards. The magnitude of difference in survival – from 13 months between 1992–2004 to 16 months between 2005–2009 – may not be interpreted as substantial progress. However, several caveats of our analysis must be accounted for. First, patients with metachronous metastatic disease are not captured in the SEER registry. Thus, our analysis is restricted to those patients with de novo metastatic disease. Across the commonly used classification schema for mRCC (i.e., MSKCC and Heng criteria, etc.), a time from diagnosis to initiation of systemic therapy less than 1 year is identified as adverse prognostic factor. Patients with de novo metastatic disease inherently possess this characteristic, and by the MSKCC or Heng risk stratification tools, this precludes them from having “good-risk” disease. Thus, our dataset is reflective of a population of intermediate- and poor-risk patients. The most recent risk stratification tool, developed by the International Kidney Cancer Working Group, includes a compilation of patient level data for 3,748 patients with mRCC enrolled in clinical trials between 1975 and 2002. This dataset also identifies an abbreviated time from diagnosis to treatment as an adverse prognostic factor. Notably, in this most recent series, patients who started treatment within a window of 506 days of diagnosis (encompassing those patients with de novo metastatic disease) had a median survival of 1.31 years. These data, pooled over clinical trials conducted over a wide span of time, are similar to what we have observed in the current report. A recent update of this dataset, like ours, suggests a temporal trend towards improved survival. However, it is important to keep in mind that their analyses included patients involved in clinical trials until 2002 – well before the widespread utilization of targeted therapies. Several other key distinctions in the SEER data may explain the small magnitude of difference in survival observed in our study. When reflecting upon the aforementioned phase III studies in mRCC evaluating targeted agents, it is critical to keep in mind that stringent eligibility criteria were set in place for each of these studies. Patients included within the SEER registry, in contrast to those on clinical trials, were not limited by performance status, organ function, and comorbidity. A report from the International mRCC Database consortium suggested that approximately 43% of patients in their experience would not be candidates for clinical trials based on standard eligibility criteria (i.e. presence of brain metastases, performance status, etc.). Furthermore, patients in the SEER registry did not necessarily receive systemic therapy. Patients with poor global health and those that did not receive systemic treatments may account for the rather modest median DSS of 14 months noted from 2005–2009. Outside of the finding that survival was improved in the era of targeted therapies, there were several other notable clinical and treatment-related characteristics associated with survival. Perhaps most notably, nephrectomy was independently associated with improved survival. Nephrectomies identified within this cohort inherently represent cytoreductive procedures, as all patients had de novo metastatic disease. Although its role is well established in the setting of cytokine therapy, cytoreductive nephrectomy is controversial in the era of targeted therapies. A selection bias may confound the observed association with survival in the SEER dataset – patients with a poor performance status, greater comorbidity, or aggressive and invasive primary tumors are less likely to undergo surgery. Two ongoing studies, the French-led CARMENA study and a separate trial by the European Organization for the Research and Treatment of Cancer (EORTC), assess the role of cytoreductive nephrectomy as an adjunct to therapy with sunitinib for patients with mRCC. Other notable findings from our analysis include a shorter survival amongst black patients relative to non-Hispanic whites. These data point towards potential disparities in access to care, or perhaps to biological differences across ethnicity. Substantial research in this domain is currently lacking, although other reports allude to similar findings. A caveat of assessing the impact of race and ethnicity within the SEER database is that the catchment area of the database has evolved over time. Specifically, beginning in 2000, 6 new registries were added – great California, greater Georgia, Kentucky, Louisiana and New Jersey. If these areas had greater numbers of minorities, these could skew the results observed herein. With respect to pathologic characteristics, the finding of shorter survival amongst patients with poorly differentiated or undifferentiated tumors (as compared to well differentiated tumors) was expected. Also as anticipated, there was no substantial improvement in survival amongst patients with non-clear cell (specifically, papillary or chromophobe) histology. The vast majority of phase III studies assessing targeted agents (with the notable exception of the pivotal study assessing temsirolimus) required the presence of clear cell disease. To date, no assessment has been made as to whether or not the therapeutic advances made amongst patients with clear cell mRCC are applicable to those with papillary or chromophobe disease. Prospective efforts to characterize the activity of sunitinib in papillary mRCC, for instance, have yielded disappointing response rates. Several limitations of the study should be noted. First, we utilized a refined cohort within the SEER dataset based on ICD-O codes for clinically relevant histologies – clear cell, papillary and chromophobe. Separate codes do exist that may encompass these histologies. For instance, a search based on the ICD-O code 8312 (“Renal Cell Adeno/Ca”) retrieved a total of 12,155 records – median survival in this cohort was 7 months, leading us to suspect that the search term may encompass a hetergenous array of histologies, such as upper tract urothelial tumors. As such, we felt that it was essential in our analysis to delineate those individuals where clear cell histology had been specified (ICD-O 8310: “Clear Cell Adeno/Ca”), as the preponderance of targeted therapies approved between 2005–2009 (excepting temsirolimus) were assessed in such patients. A second limitation of our study is that the specific nature of systemic therapies rendered is not recorded. Our underlying hypothesis, suggesting that survival is improved in the era of targeted agents, is predicated on the assumption that patients diagnosed from 1992–2004 received immune-based strategies, while patients diagnosed from 2005–2009 received primarily VEGF- and mTOR-directed therapies. Beyond systemic therapy for mRCC, it is possible that improvements in palliative care may contribute to the survival trends observed in mRCC. Enhanced palliative care may ease the burden of toxicities encountered with systemic therapy, and there is an indication that early intervention with palliative care may intrinsically contribute to improved survival in other malignancies. In all likelihood, although targeted therapies first garnered approval in 2005, these agents likely took time to integrate into the standard treatment paradigm for mRCC. Thus, in the earlier part of the second time period, it is still possible that many patients received cytokine therapy. Patients treated around the cutoff employed in this analysis (2005) may have also been exposed to placebo control arms on pivotal phase III studies evaluating targeted agents. It is possible that this may have diluted the difference in survival noted between the cytokine and targeted therapy eras. Finally, as noted previously, the follow-up in the later study period was significantly shorter than in the earlier study period (24.5 vs 14.0 months; P\<0.0001). Although this discrepancy in duration of follow-up is substantial, this is accounted for by the statistical analysis utilized (i.e., Kaplan-Meier analysis with the log-rank test). These limitations notwithstanding, our data underscore that progress is being made in the management of mRCC. Population-based studies are critical, and there are few in the available literature. The data assembled by the International mRCC Consortium has provided key insights related to clinical outcome in the era of targeted therapy, and most recently, the group has provided data pertaining to conditional survival. However, extrapolating these data to the population with mRCC at large is challenging because (1) patients in the experience have all received first-line VEGF-directed therapy, and (2) the data is derived largely from experienced academic centers with robust RCC-focused programs. Estimates provided by SEER should provide reassurance that the overarching direction taken in mRCC therapy (most notably, a shift towards targeted therapies) appears to have improved outcomes globally. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SKP RAN NV. Performed the experiments: SKP RAN NV. Analyzed the data: SKP RAN NV. Contributed reagents/materials/analysis tools: SKP RAN NV. Wrote the paper: SKP RAN NV.
# Introduction The soil microbial community holds a central position in ecosystem processes like carbon and nitrogen cycling. The performance and shape of soil microbial communities on one hand depend on soil properties, such as pH, temperature, texture and moisture, but on the other hand the soil microbial community is closely linked to plant communities through complex interactions. Plants affect the soil microbial community through biomass production, litter quality, seasonal variability of litter production, root-shoot carbon allocation and root exudates. In turn, soil microbial communities mineralize organic matter and enhance nutrient release by mineral weathering. Both processes increase the availability of nutrients enhancing plant growth , and consequently accelerate the matter flow between the aboveground and belowground parts of ecosystems. Plant diversity influences a wide range of ecosystem processes, but the underlying mechanisms are not well understood ; for example the link between plant diversity and belowground processes is just fragmentarily explained. Increasing plant diversity modifies resource availability for microbial communities in soil, which might lead to higher niche differentiation and facilitation of the soil microbial community. Beside species richness, the number of plant functional groups, i.e. species with similar morphological, phenological and physiological traits, impact soil microorganisms. Plant functional groups, such as legumes and grasses, differ in litter quality and the amount of carbon and nitrogen released to the soil, thereby affecting microbial decomposition processes. Bacteria and fungi form most of the soil microbial biomass and represent the main drivers of organic matter turnover. Since both groups prefer different qualities of resources they might be differently affected by plant diversity and plant community composition. Fungi are able to decompose litter with high C:N ratios. Therefore, it has been suggested that plant communities providing litter with high C:N ratio favor decomposition by fungi, whereas plant communities producing litter with low C:N ratio favor decomposition by bacteria. Moreover, there are differences among bacteria: Gram- negative bacteria are mainly root-associated and thus decompose organic molecules of low molecular weight, whereas Gram-positive bacteria are decomposing more complex materials, such as soil organic matter and litter. As a consequence, the presence of certain plant functional groups is likely to promote distinct microbial groups. Therefore, higher plant diversity, as number of species or number of functional groups, might affect the composition of the soil microbial community by differences in litter input quantity and quality. Most studies investigating effects of plant diversity on soil microbial community focus on plant-originated inputs, often ignoring that differences in diversity and composition of plant communities also affect microclimatic conditions such as soil moisture. Conversely, studies considering plant mediated effects on soil moisture, usually do not account for root inputs or changes in the soil microbial composition. Identifying the relative importance of drivers changing soil microbial communities is needed to better understand the functioning of soils. We assessed the effect of plant diversity and functional group composition on soil microbial communities using phospholipid fatty acids (PLFAs). The study was conducted in the framework of the Jena Experiment, a biodiversity experiment established by sowing different combinations of grassland species on a fallow agricultural soil. In addition to experimental plots with different levels of plant diversity and vegetation-free bare ground plots, we studied long-term meadows and on-going arable plots as adjacent to the field experiment as control sites to assessed, how the soil microbial community developed five years after establishing the experimental site. We hypothesized that (1) higher plant diversity increases soil microbial biomass, caused by higher amounts of litter input as well as by improved microclimatic conditions for soil microbes, and (2) plant functional group composition drives composition of the soil microbial community, exemplified e.g., by changes of fungal-to-bacterial biomass ratio (F:B ratio). Due to the production of low quality litter, we expected plant communities containing grasses but not legumes to favor fungi, whereas plant communities producing litter of high quality to favor bacteria. # Materials and Methods ## Site description and experimental design The field site of the Jena Experiment is located close to the city of Jena (Germany) in the floodplain of the river Saale (50°55′ N, 11°35′ E, 130 m a.s.l.). No specific permission was required to work on “The Jena Experiment” and no endangered or protected species were involved in this study. The soil (Eutric Fluvisol) has developed from up to 2 m thick fluvial sediments presenting a systematic variation of soil texture. The sand content decreases with distance to the river from 40% to 7%, while the silt and clay content increase (silt: 44% to 69%; clay: 16% to 24%). Experimental plots were arranged in four blocks parallel to the river to account for these differences in soil characteristics. Before the establishment of the Jena Experiment in 2002, the site was used as arable land since the early 1960s and ploughed and fertilized regularly. The Jena Experiment comprises 86 plots (82 vegetated and 4 bare ground plots, each 20 m by 20 m). The experimental design manipulates a gradient in sown plant species richness from 1 to 60 (1, 2, 4, 8, 16, and 60) near- orthogonal with a gradient in plant of functional group numbers from 1 to 4 (1, 2, 3 and 4). All 60 species are typical for Central European mesophilic grasslands. They were grouped into four functional groups according to their morphological, phenological and physiological traits. The species pool included 16 grasses, 12 small herbs, 20 tall herbs and 12 legumes. To maintain the diversity levels, all experimental communities have been weeded manually twice a year. Plots are mown twice a year, in June and September and are not fertilized. In addition bare ground plots with four replicates were established. Furthermore, soil microbial community was determined on two adjacent regularly mown non-fertilized meadows and two arable plots on the experimental site. The arable plots were continuously managed according to conventional agricultural procedures, growing cereals. ## Soil sampling and phospholipid fatty acid (PLFA) analysis In early May 2007, six soil samples per plot were taken with a core cutter (inner diameter: 4.8 cm, Eijkelkamp Agrisearch Equipment, Giesbeek, The Netherlands) to a depth of 5 cm, pooled and placed immediately in cooling boxes. Within 48 hours after sampling the soil was kept at 4°C, sieved \<2 mm, remains of roots were manually removed and finally the samples were stored at −20°C until further sample processing. Soil samples were shaken with a mixture of chloroform, methanol and 0.05 M phosphate buffer (pH 7.4) to extract soil lipids. The lipids were split into neutral lipids, glycolipids and phospholipids by eluting with chloroform, acetone and methanol from a silica-filled solid phase extraction column. Subsequently, phospholipids were hydrolyzed and methylated by a methanolic KOH solution and the PLFA-methyl esters were identified and quantified by gas chromatography with atomic emission detector (GC-AED) (Agilent, Böblingen, Germany) and gas chromatography-mass spectrometry (GC/MS) (Thermo Electron, Dreieich, Germany). Peak areas and the resulting amount of PLFA were calculated relative to the internal standard PLFA 19:0. The sum of all PLFAs was taken as total soil microbial biomass. Furthermore PLFAs were assigned to microbial groups. The PLFAs 14:0, 14:0br, 15:0, 16:0, 17:0, 18:0 were used as general microbial markers. All monounsaturated and cyclic fatty acids were grouped as Gram-negative bacteria (Gram-), while all branched PLFAs were grouped as Gram-positive bacteria (Gram+). PLFA 18:2ω6 was used as a fungal biomarker. The F:B ratio was calculated using the molar weight of the fungal PLFA marker divided by the sum of molar weights of bacterial PLFA biomarker. ## Covariables Fine root standing biomass (termed as ‘root biomass’ hereafter), leaf area index (LAI) and soil moisture were considered as potentially meaningful covariables. Unfortunately, in 2007, the year of the PLFA sampling, root biomass was not determined, thus we used an average of 2006 and 2008 root biomass measurements. In both years root biomass was sampled to a depth of 30 cm. In addition, the sampling in 2008 was stratified, so that the 0–5 cm depth increment could directly related to the sample of the soil microbes. Based on the ratio of the top increment (0–5 cm) to the total root biomass in 2008, we calculated the specific root biomass in the top soil (0–5 cm). Furthermore, nitrogen concentration of fine roots was determined using root material from ingrowth cores from, sampled between 2007 and 2008. N concentrations in the biomass were determined with an elemental analyzer (Vario EL Element Analyzer, Elementar, Hanau, Germany). In the course of the PLFA soil sampling soil moisture was determined, too, as the gravimetric soil water content. Leaf area index (LAI) was measured approx. 5 cm above ground level using a LAI-2000 plant canopy analyzer (LI-COR) in late May 2007 (shortly before the first mowing of the year; see experimental design). ## Statistical analysis Using analyses of variance (ANOVA) followed by Tukey's HSD test we assessed microbial biomass and F:B ratio in experimental plots of the Jena Experiment, and their relationship to those from control plots (arable fields and meadows). ANOVA with sequential sum of squares (type I SS) was applied to test for effects of plant diversity on microbial biomass (total, Gram+, Gram− and fungal). The Jena Experiment is based on a factorial design with different combinations of plant species richness and number of functional groups, where all plots are arranged in a block design accounting for differences in soil texture among the blocks. Therefore, ‘block effect’ was included as random factor and was fitted first. The contrast between bare ground plots vs. sown plots was fitted next, before testing for the effect of richness (log-linear term) and functional groups (linear term) as continuous variables. Finally, the presence of each plant functional group (small herbs, tall herbs, grasses and legumes) was included into the model in a series of alternative models. Furthermore, non- metric multidimensional scaling (NMDS) was used to compare plot-specific patterns of PLFA profiles. The data used in the NMDS was normalized to the peak area of the highest peak (18.1n11) set at 100%. Bray-Curtis was used as dissimilarity index. To investigate which mechanisms underlie the effects of plant diversity, we used structural equation modelling (SEM, see also) with observed variables. In SEM, all diversity levels, except bare ground, were considered. For every group of PLFAs assigned to specific microbial taxa, a full model was set up including all experimental variables that were significant in the preceding ANOVA. As possible means by which the effect of plant diversity might be manifest, we included root biomass as measure of belowground plant input, root nitrogen concentration as measure of litter quality and LAI as a measure of plant community influence on evaporation and thus the microclimatic conditions (e.g., soil moisture and temperature). The categorical variable ‘block’ was substituted by the continuous variable ‘clay’ content of soil. We considered aboveground plant inputs as negligible, because all above ground biomass was harvested twice a year. The minimal parsimonious models were identified using specification search, based on the Bayes information criterion (BIC). The adequacy of the model was tested with Chi-squared tests (χ<sup>2</sup> tests) and root mean square error of approximation (RMSEA). ANOVA and Tukey's HSD test were performed using R 2.15.2 and structural equation modeling was performed using AMOS 18.0. # Results ## Soil microbial biomass The mean of the total PLFA concentration, henceforth termed total microbial biomass, was14.4±3.5 nmol g<sup>−1</sup> soil dry weight (mean ± sd) on experimental plots (vegetated plots and bare ground). This was significantly higher than measured on plots of arable land (5.3±2.5 nmol g<sup>−1</sup>) and significantly lower than on meadows (30.2±10.3 nmol g<sup>−1</sup>;). ANOVA revealed block as a significant predictor of the total microbial biomass. Furthermore, total microbial biomass was significantly lower on bare ground plots (8.7±1.3 nmol g<sup>−1</sup>) than on vegetated plots (14.7±3.4 nmol g<sup>−1</sup>). Plant species richness had a significant positive effect on the total microbial biomass on vegetated plots. The presence of individual plant functional groups did not affect total microbial biomass. Structural equation modeling (SEM) showed that block (represented by the continuous variable clay content of soil) and plant richness, the design variables with significant influence indirectly affected total microbial biomass. The minimal parsimonious model (χ<sup>2</sup><sub>13</sub> = 21.74, P = 0.060; RMSEA = 0.093, P = 0.147) explained 45% of the variance of total microbial biomass. Total microbial biomass was mainly explained by its positive relationship to soil moisture. Soil moisture increased with increasing plant richness and higher clay content. The major effect on soil moisture was attributed to increasing leaf area index (LAI), which itself was strongly correlated to plant richness. The negatively influence of root nitrogen concentration on total microbial biomass was driven by higher root biomass, which itself was increased at higher plant richness. ## Soil microbial community structure The F:B ratio did not differ among experimental plots, arable plots and meadow plots (experimental plots: 0.052±0.015; arable land: 0.041±0.003; meadows: 0.049±0.016;). In contrast, the F:B ratio was significantly lower on bare ground (0.034±0.006) than on vegetated plots of the biodiversity experiment (0.053±0.015). F:B ratio was positively affected by an increasing number of plant functional groups and negatively by the presence of legumes. Plant richness did not significantly affect the F:B ratio. However, the F:B ratio increased from low plant richness plots to medium ones with eight plant species, and decreased again in plots with high plant richness. Regression analyses of the relationship between F:B ratio and both, fungal and bacterial biomass revealed that the F:B ratio was more related to fungi (R<sup>2</sup> = 0.67, P\<0.001) than to bacterial biomass (R<sup>2</sup> = 0.043, P\<0.066). Considering biomass of Gram+, Gram- and fungi separately, all groups differed significantly among blocks and between bare ground and vegetated plots. Plant diversity positively affected both bacterial groups. Gram+ as well as Gram- bacteria were reduced on bare ground plots (Gram+ = 2.3±0.4; Gram- = 4.6±0.6) compared to vegetation plots (Gram+ = 3.9±1.0; Gram- = 7.7±1.7) and biomass of both bacterial groups were increased with increasing plant richness. However, the ANOVA also revealed differences: while Gram+ were positively influenced by the presence of grasses, Gram- were not affected by any of the plant functional groups. Fungal biomass was positively affected by the number of functional groups present, and negatively by the presence of legumes. The results of the NMDS, based on the PLFA composition, confirmed the strong dissimilarity between bare ground plots and vegetated plots. The dissimilarity between the vegetated plots was relatively small, though we found a clear effect of plant diversity, i.e. the higher the plant diversity on the plot the more different were the microbial communities compared to low diverse plots. SEM for Gram+ (χ<sup>2</sup><sub>17</sub> = 24.877, P = 0.098; RMSEA = 0.078, P = 0.231) and Gram- bacteria (χ<sup>2</sup><sub>13</sub> = 19.80, P = 0.100; RMSEA = 0.082, P = 0.217) explained to 41% and 44% of variance, respectively, and revealed high analogy between the groups. Both bacterial groups were mainly driven by soil moisture, which was mostly affect by LAI. Furthermore, both groups had a negative relationship with nitrogen concentration of fine roots, mediated by increased root biomass. The minimal parsimonious model explains 44% the variation in fungal biomass (χ<sup>2</sup><sub>5</sub> = 5.43, P = 0.365; RMSEA = 0.034, P = 0.475). In contrast to the bacterial groups, fungal biomass was neither affected by the amount of root biomass nor by its quality. Although there was a positive indirect pathway from functional groups and legumes via LAI and soil moisture to fungal biomass, strong direct paths from functional groups and legumes remained in the minimal parsimonious model. These direct paths indicate that the diversity effect was driven by mechanisms other than soil moisture or quantity and quality of root biomass. The F:B ratio was explained to 35% by the minimal model (χ<sup>2</sup><sub>2</sub> = 2.27, P = 0.322; RMSEA = 0.041, P = 0.390). In contrast to all other models, only direct paths connected the experimental variables to F:B ratio: it strongly decreased in the presence of legumes, but increased with increasing number of plant functional groups to almost the same extent. These relationships could not be explained by our measured covariables. # Discussion In the framework of the Jena Experiment we investigated how soil microbial communities are affected by plant diversity and the underlying mechanisms of these effects. In grassland with manipulated plant species richness and number of plant functional groups we showed that the soil microbial communities are strongly linked to plant diversity. Corresponding to hypothesis 1, the positive plant diversity effect on total microbial biomass was mainly driven by improved microclimatic conditions, while we found only a minor influence of the amount of root litter inputs on the soil microbes. Furthermore, number of plant functional groups and the plant functional composition, in particular the presence of legumes, highly impact the microbial community composition, referring to hypothesis 2. Below, we will discuss in detail, how plant diversity drives the soil microbial community. Five years after conversion from arable land to grasslands, increased soil microbial biomass indicates that the microbial community performs better. In addition to the growth of the soil microbial community, it has been reported that the community also has a higher metabolic activity compared to the initial conditions. Lower microbial biomass on arable land probably is due to soil disturbance by tillage and the tillage-induced changes of soil properties. The lower organic carbon concentration in arable soils is attributed to faster decomposition of soil organic matter, which in turn reduces the microbial biomass in the long term. However, even in plots with highest plant diversity, i.e., 60 species and 4 functional groups, microbial biomass was lower than in adjacent meadows. This indicates that the time since conversion of our study area from arable use to grassland was not sufficient to reach the state of microbial biomass of permanent meadows. However, as total microbial biomass significantly increased with increasing plant richness, higher plant diversity promotes the development towards the stage of permanent meadows. Confirming hypothesis 1, plant richness as well as clay content of the soil indirectly increased total soil microbial biomass. Interestingly, structural equation modeling suggests that this was mediated via soil moisture. Soil moisture itself holds a central position in the interplay between plant diversity, abiotic soil conditions and microbial biomass. The strong influence of soil texture on soil moisture is well known: with smaller particle size soil water holding capacity increases. Results of the present study suggest, however, that the positive effect of plant diversity on soil microbial biomass may exceed that soil texture via changes in soil moisture. Higher plant diversity increased canopy density of the plant stands, measured as LAI, which presumably reduced evaporation from the soil surface. Plant richness also affected soil microbial biomass via root inputs, namely via root nitrogen concentration. The detrimental effect of nitrogen concentration on microbial biomass was closely related to increased root biomass; with increasing plant richness root biomass increased, while at the same time nitrogen concentration decreased, which confirms earlier findings. Results of NMDS showed that the composition of PLFAs differed mainly between bare ground and vegetated plots, while in vegetated plots (1-60 plant species, 1-4 plant functional groups) the composition of PLFAs was similar. However, the dissimilarity of the microbial community composition was more pronounced in plots with different diversity levels, i.e., low diverse plots differed most from high diverse plots. Higher diversity in plant communities leads to more diverse organic matter input in quantity, quality and timing, and this likely is responsible for the observed changes in microbial communities along diversity levels. The plant diversity effect on microbial community composition was also reflected in the F:B ratio, but in contrast to total microbial biomass, the F:B ratio was more affected by functional groups than by species richness, supporting hypothesis 2. Moreover, this relation to plant functional groups reflects the stronger dependency of F:B ratios from changes in fungal biomass than in bacterial biomass. We further found legumes to be a strong predictor of F:B ratio, which is in line with previous findings. However, neither the underlying mechanisms of the positive effect of functional groups nor of the negative legume effect on the F:B ratio was mediated by the considered covariables. Similar results have been reported by Lamb et al., who studied the effect of plant species richness and evenness on soil microbial communities in a pot experiment. The lack of relationships between F:B ratio and root litter quantity and quality as well as soil moisture indicates that both microbial groups are similarly affected by these variables. The strong direct link between plant diversity and F:B ratio, however, points to other plant resources as major drivers of soil microorganisms, such as root exudates. Indeed, root exudates were reported to strongly influence soil microbial communities. In more diverse plant mixtures resource supply for microorganisms may be assumed to be higher and more diverse, while resource supply in monocultures is expected to be more one-sided and temporally limited. Furthermore, it is known, that the number of plant functional groups and presence of legumes may be related to turnover rates and decomposition of fine roots , which might cause changes in microbial community structure. Although bacteria and fungi were similarly affected by plant diversity, we found bacteria more related to plant diversity-controlled abiotic soil properties, while and fungi were more affected by the input of organic materials. As shown by de Vries et al. fungal-based soil food webs are more resistant to disturbances, while bacterial communities are more resilient due to their fast life cycle. This might explain why the bacterial community was in our study more related to fast changing abiotic conditions, such as soil moisture. # Conclusion We identified changes in microclimatic conditions, in particular increased soil moisture, as a main mechanism how plant diversity affects soil microbial biomass in the topsoil. Furthermore, the results indicate that shifts in the microbial community composition, namely in the F:B ratio, heavily rely on differences in the quality and quantity of root exudates. Changes in soil microbial biomass with plant diversity suggest that microbial communities of the established grassland systems develop towards permanent meadows, but that reaching the state of these meadows takes decades. Notably, however, differences in microbial biomass indicate that high diverse plant communities promote faster transition towards permanent meadows indicating that plant diversity is a key factor for restoring functional grassland systems on former arable land. # Supporting Information We thank Janine Seyfferth, Sibylle Steinbeiss and Steffen Rühlow for their help during sampling and for helping with the PLFA extraction, Alexandra Weigelt for providing the LAI data and Francesca Hopkins for her helpful comments on the manuscript. We gratefully acknowledge all the people that were involved in planning, set up and maintaining of the experiment and the German Research Foundation (DFG) for financial support of FOR 456 and Gl262-14. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GG SS WW EDS CR. Performed the experiments: MH GG HB. Analyzed the data: ML MH CR NE. Contributed reagents/materials/analysis tools: GG. Wrote the paper: ML MH NE CR HB CE YO SS WW EDS GG.
# Introduction The posterior cruciate ligament (PCL) is one of the major passive stabilizers of the knee joint, serving as the primary restraint to excessive posterior tibial translation, but is also thought to act as a secondary restraint to tibial rotation. It is likely that injuries to the PCL interfere with its normal function and consequently lead to joint instability, as well as to possible subsequent osteoarthritis. In order to better understand these processes, as well as improve diagnosis, reconstruction and rehabilitation techniques, a number of studies have investigated healthy PCL loading and injury biomechanics. Here, the key goal is to quantify the physiological loading patterns of the healthy PCL during sport and activities of daily living. However, despite a number of in vitro, in vivo, and modelling approaches to provide science based evidence, the normal loading conditions of the PCL are still controversially discussed. Although the PCL is the strongest ligament of the knee joint, some scientists have dismissed it as functionally superfluous. These ideas originated mainly from studies reporting either no or very small forces in the PCL during activities of daily living and partially from those stating that an isolated PCL tear may not interfere with patient movement. On the other hand, a number of studies have reported PCL loads very close to or even higher than the failure limit of the ligament during activities such as passive knee flexion, body weight squats and forward lunge. These contrasting reports most likely result from the difficulties in accessing the kinematics and kinetics of the ligament in vivo. Nevertheless, this controversy indicates a clear requirement for improved understanding of the PCL and its functionality, which is closely related to the biomechanical function and loading behaviour of its anterolateral (AL) and posteromedial (PM) bundles. Unfortunately, a comprehensive investigation regarding the contributions of each bundle in providing stability to the knee is lacking. Traditionally, the two bundles were mainly believed to have reciprocal functions during flexion and extension of the knee. This argument has been supported by a number of studies reporting lengthening of the AL bundle and shortening of the PM bundle during knee flexion. However, more recently, a co-dominant functional relationship between the two bundles has been suggested based on simultaneous elongation observed in both PCL bundles during a forward lunge. The reasons for the lack of knowledge regarding loading of the PCL are manifold. Most importantly, PCL injuries have been historically underdiagnosed due to the high possibility of asymptomatic damage, and therefore the necessity to investigate PCL biomechanics has been generally underestimated. Over the last two decades, it has become apparent that PCL injuries are more prevalent than previously believed, with its involvement in nearly 3% of all knee injuries, and 38% of acute knee injuries. Considering these findings, together with an increasing number of vehicular traumas as the main cause of PCL rupture, scientists and clinicians have become more interested in research regarding the functional loading of the PCL within the knee joint. A number of technical challenges have also limited the accurate in vivo assessment of PCL function during dynamic movement. Compared to other ligaments of the knee, investigation into the function and loading patterns of the PCL is difficult due to its relatively inaccessible anatomical location within the knee joint. Furthermore, the PCL possesses a complex fibrous anatomy and undergoes differing contact conditions with neighbouring bone and soft tissue structures, all leading to kinematics that are difficult to assess. As a consequence, a number of both direct (sensor-based) and indirect (predominantly image-based) approaches have been used to examine the behaviour of the PCL using in vivo, in vitro and modelling investigations. The most common in vivo approach to determine strain has been to examine the relative movement of the bony attachment sites of the PCL. Here, the length of virtual bundles are determined as the absolute distance between the centroids of the origin and insertion sites, while virtual bundle strain (VBS) has been defined as the elongation from the reference length, which is generally selected as the length of the ligament at full extension of the knee, relative to their length measured at different flexion angles. Here, CT and MR imaging, and even dual-plane fluoroscopy, are generally used to track the attachment sites of the bundles. Given the hurdles associated with in vivo studies, current knowledge of PCL loading has mainly been based on in vitro cadaveric investigations. Here, strain and force sensors combined with either mechanical jigs or robotic manipulators have proven to be successful approaches for investigating the PCL. However, due to the complexities of applying physiological loading conditions to in vitro set-ups, the measurement of force or strain in the PCL has been generally limited to passive flexion, flexion under externally applied loads, or with simplified muscle forces only. Compared to the assessment of strain, force is considered to provide a more realistic measure of the ligament load; however, studies with implanted force sensors suffer from the imposed size and stiffness of the sensors, which compromise the natural kinematics and loading within the ligament. To avoid any interference caused by implantable sensors, an indirect method using robotic imposed kinematics has also been introduced for estimating total ligament force based on the principle of superposition. Although this approach allows a contact-less measurement of the effect of the ligament force, the in situ forces obtained do not necessarily represent the exact ligament forces. Strain sensors have less influence on the normal ligament behaviour due to their small size and low stiffness and therefore have been frequently used in cadaveric investigations of PCL loading \[, –\]. The sensor output or real bundle strain (RBS) represents the relative elongation of the real curvilinear bundle path of the ligament. However, due to the locality of the sensor measurement site, together with the invasive nature of sensor implantation, researchers have recently turned to indirect approaches for strain measurement. When used in cadaveric studies, these indirect methods have the advantage that they are not limited to image-based assessment. To obtain more accurate results, 3D coordinate measuring machines and surgical navigation systems have also been introduced into cadaveric investigations of ligament strain. In addition to experimental studies, in silico musculoskeletal and finite element modelling studies have also reported loading behaviour of the PCL. Musculoskeletal models have been generally used to estimate the PCL force patterns during dynamic physiological activities such as walking and squatting. In contrast, finite element models of the knee have been mostly subjected to either static or quasi-static loading conditions. Despite their ability to strongly complement in vivo and in vitro approaches, modelling investigations have suffered from oversimplifications of the structural geometries, material properties and loading conditions. For example, in the vast majority of models reported in the literature, ligaments have been represented using one-dimensional elements which therefore preclude the prediction of non- uniform strain distributions within the three-dimensional structure of the ligament. Improvements in modelling and experimental techniques have provided greater insights into the functional loading of the PCL in the healthy knee joint. However, differing results from in vivo, in vitro and in silico studies remain controversially discussed and a clear unbiased consensus on the biomechanics of the functional bundles for the healthy knee is still lacking. In particular, the following three questions need to be comprehensively answered: (1) Do the AL and PM bundles have different loading patterns during loaded and unloaded flexion of the knee? (2) Is the loading behaviour of the functional bundles of the PCL activity-dependent? (3) Are the observed loading patterns dependent upon the investigation techniques used to assess them? This study therefore aims to generate evidence-based understanding of the functional loading of the PCL in the healthy knee joint, including its AL and PM bundles, through systematic review and statistical analysis of the current literature. # Methods ## Literature Search and Study Selection For the systematic review, the databases PubMed, Embase, and Cochrane Central Register of Controlled Trials (CENTRAL) were all searched from their inception up to January 2016 to recruit studies that reported load data (strain or force) for the PCL. Different combinations of the terms “knee”, ‘‘ligament”, ‘‘load”, ‘‘force”, “tension“, “length”, ‘‘strain”, “elongation” and ‘‘lengthening” were used. To facilitate collection of the systematic review manuscripts as well as identification of duplicate reports, all search hits were imported into the Eppireviewer software (version 4.5.0.1). Titles and abstracts of all search hits were screened for eligibility based on the inclusion/exclusion criteria as follows: ## Inclusion criteria - Subject characteristics Human living subjects or cadaveric specimens with healthy ligaments inside healthy knees - Bundle definition Common anatomical definition was used, which consisted of real or virtual anterolateral (AL), posteromedial (PM) bundles, and the mid-PCL fibre as a representative of the whole PCL - Activities Physiological activities including passive flexion, body-weight squat, forward lunge, walking, stair ascent, stair descent Clinical tests including passive flexion with 100 N or 134 N posterior tibial load (PTL) - Results Numeric strain or force data Length or elongation of the bundles when the reference length was specified - Report Journal articles and conference proceedings with full texts written in the English language ## Exclusion criteria - Subject characteristics Animal subjects or specimens, deficient or reconstructed ligaments, pathologic knees - Load measurement techniques Techniques that provided no quantitative data of ligament strain or force (e.g. arthroscopic probes to assess the slack/tense status of the ligament) - Bundle definition Virtual bundles that did not match anatomical structure - Results Data collected using non-calibrated sensor outputs Strain data where the reference length was not reported Duplicate results (including multiple studies that report on the same cohorts) - Report Articles without full texts, non-English reports, and review papers After initial screening to reduce the pool according to clearly met criteria for exclusion, the final decision to include or exclude a specific study was made after carefully reading the full text article. The reference lists of all included full text articles were double-checked to identify additional reports that were possibly missed in the initial systematic search (e.g. in cases where reporting ligament strains and forces was not the primary focus of the manuscript). ## Analysis of the Literature An analysis was conducted to obtain the average load trends of the PCL and its two functional bundles during passive and active knee flexion, as well as to compare the reported strain and force results between direct and indirect methods. Outcome measures used to assess the loading trends of the PCL predominantly included measures of strain and force. It should be noted that due to varying definitions of ligament physiological length, which relied upon lax versus taught, knee flexion angle and insertion site definitions, the term “strain” alone was not sufficient to report and compare the results from the different methodologies. Therefore, we introduce the term real bundle strain (RBS) to address the strain within the PCL that included the real curvilinear path of the ligament. RBS can be measured either using implanted strain sensors or by calculating the relative elongation of a curvilinear fibre that follows the centroid-axis of the bundle based on medical images. Similarly, the term virtual bundle strain (VBS) was used to describe the relative elongation of a virtual straight-line fibre connecting the centroids of the two bony attachment sites of the entire ligament or an individual bundle. VBS can be determined by tracking and assessing the relative change in distance between the ligament origin and insertion sites (using e.g. medical images or data from coordinate measurement systems). As with strain data, the reported force data were also obtained using different techniques. In this review, the term “force” was used to describe ligament tensile forces, as measured by force sensors or calculated by means of computational modelling, while “in situ force” was used to represent forces obtained based on the principle of superposition. Real and virtual bundle strains, force and in situ force data of the PCL were thus extracted from the included articles. If numeric data were not explicitly reported within the article, graphs were carefully digitized for data extraction. It should be noted that while the strain and force data during walking and stair activities were included in the systematic review for qualitative comparison of the literature, the data were excluded from the statistical analysis because of the inconsistent formats used for reporting results. In order to be able to compare the data from the different studies, the reported measures of length were converted into measures of strain by dividing the length changes by the given reference length (bundle length at full extension). To extract the strain patterns of the PCL bundles during passive flexion, a weighted-regression approach was used based on the number of knees tested in the various studies. The average load trends of the PCL bundles were then extracted for the different activities using the weighted mean and standard error of the mean (SEM; presented in parenthesis after the mean value) of the extracted force and strain data at the most frequently reported angles of 0°, 30°, 60°, 90° and 120° of knee flexion. To reduce the risk of bias, all studies with at least one author in common were assigned to a specific research group. The articles that belonged to specific research groups were then assessed carefully to ensure that subjects or cohorts were not reported on more than one occasion. Those articles with no evidence of duplicate data were removed from the specific research group and were analysed using the routine procedure as outlined above. The data from articles with duplicate results were averaged and only the mean trend of each group was included in the statistical analysis. # Results ## Study Selection A total of 3577 articles were found through the initial electronic search. Search results were imported into the Eppireviewer software, where 854 items were identified as duplicates. An additional 2232 studies were excluded after initial screening. The full-texts of the remaining 491 studies were retrieved for further evaluation. Here, evaluation was conducted to ensure inclusion of articles that contained PCL load data even if the main focus of a study was on another ligament. A comprehensive full-text screening revealed that 66 articles contained strain or force data regarding the PCL (including 9 eligible manuscripts that were found through double-checking the reference lists of the included full-text articles), and thus fulfilled all inclusion criteria (data classified according to the activity types in Tables –). ## Strain Patterns of the PCL during Passive Flexion Twenty-four studies that contained PCL strain data during passive flexion met all inclusion criteria; from those, only three studies assessed the RBS within the PCL, whereas in the remaining experimental studies a variety of indirect methods was used for measuring the VBS. Most studies (16 articles) presented in vitro cadaveric approaches, while modelling (6 articles) and in vivo (3 articles) methods were less frequently utilised. ### In vivo results Only three studies reported on strain of the PCL during passive flexion in vivo. Jeong and co-workers calculated the distances between the attachment sites of the AL and PM bundles using CT images of ten living subjects. At 90° of flexion, the AL and PM bundles were elongated 33% and 9% respectively compared to their reference lengths at full extension of the knee. From 90° to 135° of flexion, there was no considerable further lengthening in the AL bundle, but the PM bundle continued its elongation. In another study, based on MR images of twenty subjects in the sagittal plane, Nakagawa et al. reported lengthening of the virtual bundles that represented the mid-PCL. With the reference length taken at full extension, the VBS trend was upward, with a calculated strain of 29% at 90°. Further flexion beyond 90° caused a reduction of the PCL strain to a value of 24% at 120°. These findings are in general agreement with those reported by King and co-workers who used an open-bore MRI scanner to measure lengthening of the virtual mid-PCL bundle in seven subjects. In addition to the virtual mid-PCL bundle, the elongation of the curvilinear fibres on the anterior and posterior surfaces of the PCL was calculated. Their results indicated continuous elongation of the anterior fibre up to 22% at 120°, while the posterior fibre lengthened up to 40° but became shorter thereafter. ### In vitro results Two studies reported RBS data for the PCL during passive flexion of the knee. Arms and co-workers attached Hall Effect Strain Transducers (HESTs) to the anterior and posterior fibres of the PCL in cadaveric specimens. They found a rapid elongation of the anterior bundle after the first 10° of flexion, corresponding to 19% maximum strain at 120°. The posterior bundle was positively strained only after 45° of flexion and gradually elongated thereafter. In another sensor-based study, Dürselen and co-workers attached Ω-shaped strain transducers to the PM bundle of the PCL in nine cadaveric specimens. Since their study considered the zero-strain position to be at 60° flexion, this point was referenced to full extension in order to allow an objective comparison of their results. Given this transformation, the PM bundle was shorter than its reference length up to 30° flexion, whereupon the strain gently increased up to 1% at 110°, with the minimum strain observed at 15° flexion. Other in vitro studies have used indirect techniques for measuring the PCL strain in cadaveric knees. Dorlot and co-workers as well as Inderster and co- workers measured the distance between origin and insertion of the anterior and posterior fibres of the PCL using threads passed through the centres of the attachment sites with one end secured to the tibia and the other connected to a displacement transducer. They found a gradual elongation of the anterior bundle throughout the full range of knee flexion but the reported strain patterns of the posterior bundle were not consistent in these studies. While in the first study the posterior bundle was found slack throughout the first 50° flexion, the second study reported a continuous elongation of the bundle with increasing knee flexion. Using Rontgen Stereo-photogrammetric Analysis (RSA) to measure the VBS of the PCL in cadaveric knees, Garbelotti and co-workers, van Dijk and co- workers, as well as Blankevoort and co-workers reported lengthening of the virtual AL bundle with increasing knee flexion. Based on their findings, the PM bundle was generally lax throughout the studied range of flexion and had a shortening phase up to 40°- 60° of flexion followed by an elongation trend thereafter. Garbelotti and co-workers also found the length of the mid-PCL unchanged during the first 30° of flexion with an elongation phase thereafter. These results are not in agreement with those previously reported by Wang and Walker who used lateral and anterior-posterior radiographs from 12 cadaveric knees to calculate the relative displacement of the mid-PCL attachment sites marked with metal pins. They found a 10% negative strain during the first 30° of knee flexion, which remained nearly constant throughout the higher flexion angles. Nakagawa et al. used sagittal MR-images of six cadaveric knees to assess elongation of the mid-PCL. They reported a steady increase in the length of the mid-PCL bundle from full extension to 120° of passive flexion with a maximum of 26% VBS. One common approach to assess the kinematics of the PCL has been based on superimposing or registering the geometrical data obtained by digitizing bone surfaces and ligament attachment sites onto captured kinematic data. To capture the joint kinematics, Trent and co-workers took photographs of the cadaveric specimens surrounded by several mirrors. Repeating this procedure for different flexion angles, they reported the pattern of ligament lengthening during 105° of knee flexion. In the first 15° of flexion, the mid-PCL bundle experienced shortening, followed by lengthening at higher flexion angles. However, recent studies in this field have employed more advanced technologies to capture the joint kinematics including 3D scanners, surgical navigation systems, optical markers, and electrogoniometers etc. Using such equipment, the elongation patterns reported by Wang et al., Zaffagnini et al. and Cross et al. suggest lengthening of both the AL and PM bundles throughout the whole range of flexion. Belvedere and co-workers as well as Hsieh and Draganich found similar elongation patterns for the AL bundle but they reported a phase of shortening for the PM bundle starting from full extension. The results of Amiri and co-workers, however, suggest a very different loading pattern for the PCL, with negative strains for both bundles over the entire arc of flexion. ### Modelling results Seven studies used modelling approaches to calculate strain patterns of the PCL during passive knee flexion. As a simple method, four-bar linkage mechanisms formed by the tibia, the femur and the cruciate ligaments were used to simulate passive motion of the tibio-femoral joint and investigate ligament kinematics. Using a 2D linkage model of the knee, Zavatsky and O’Connor calculated strain patterns of different bundles within the PCL during passive knee flexion. Their results demonstrated that the length of the anterior bundle gradually increased with knee flexion until 90° and remained constant thereafter, but the posterior bundle was positively strained only after 110° of flexion. Chittajallu and Kohrt constructed a 2D dynamic model of the knee consisting of two coupled four-bar linkages driven by the cruciate and the collateral ligaments. Here, the governing equations of motion were solved to minimise the generated anterior- posterior force in the joint. The results indicated that until 60° of flexion, the mid-PCL strain had a gentle increase with a maximum of 1% strain but with a declining trend thereafter. Crowninshield and co-workers introduced a 3D kinematic model of the knee that was driven by the motion of the joint centres of rotation and calculated negative strain patterns for both the anterior and posterior bundles throughout the whole range of flexion. Taking into account the contact between tibia and femur, Beynnon and co-workers, Wismans and co-workers as well as Amiri and co-workers developed quasi-static analytical models of the knee with bones represented as rigid bodies and ligaments as flexible one- dimensional elements. Although in these studies joint flexion was consistently achieved by applying a small force to the femur, the obtained strain patterns of the PCL bundles were not consistent. For instance, Wismans and co-workers reported mid-PCL load-bearing only after approximately 40° of flexion but Amiri and co-workers found it tense only up to 50° of flexion. In general, considerable variation was observed across the literature for the strain estimations within the AL, PM and mid-PM bundles. Nearly all studies suggested a positive strain in the AL bundle throughout the range of flexion, while both positive and negative strains were frequently reported for the PM bundle. As a result, the weighted-regression line for the AL bundle showed the highest strain magnitudes. The data indicate a gradual increase in AL bundle strain up to 90°, with a maximum of 21%, after which it declines to 17% at 120° of flexion. The regression line for the mid-PCL strain suggests a similar behaviour to that of the AL bundle but with lower magnitudes, and with a maximum of 14% at 90°. However, the weighted regression trend for the PM bundle strain indicates a slightly different pattern from those of the AL and mid-PCL bundles. In the first 80° of flexion, the bundle is shorter than its reference length, suggesting a lax state. The bundle has its shortest length at 30° joint flexion, with a corresponding strain of -3%. At 100°, the PM bundle presents its maximum positive strain of only 2%, and further flexion reduces the strain once again. ### Difference between assessment techniques Average strain patterns obtained for the virtual mid-PCL bundle varied among the different assessment techniques. No in vivo strain data were reported for bundles at 30° or 60° of flexion. At 90°, modelling studies calculated mostly negligible strains (0%) for the mid-PCL, while in vivo investigations found mean strains of 24%; approximately twice the values obtained in vitro. However, all the assessment techniques suggested a declining trend for the bundle strain above 90° of knee flexion. Compared to in vivo and in vitro investigations, strain data estimated by means of modelling techniques had considerably higher variation. Mean VBS and RBS patterns exhibited similar trends. Both direct and indirect measurement techniques reported elongation of the AL bundle throughout the first 90°, and shortening of the PM bundle during the first 30° of flexion. Contrary to direct methods, however, indirect methods of strain measurement showed a declining strain after 90° of flexion for both bundles. Moreover, the strain magnitudes of the AL and PM bundles were found to exhibit different patterns throughout flexion, with considerably higher strains reported for the AL bundle. ## Force Patterns of the PCL during Passive Flexion The majority of force data reported for the PCL during passive flexion has been acquired by the same research group using cadaveric knees and attaching a load cell to the isolated femoral attachment site of the ligament. In these studies, the knees were mounted in multi-degree of freedom jigs and manually moved through knee flexion. Based on their measurements, the PCL was found to be under low tension throughout flexion. The average tension in the PCL at full extension, 30° and 120° of flexion was approximately 15 (3), 4 (1) and 19 (5)N respectively. Smaller forces were reported by Wang and co-workers who used a similar assessment technique, and were in a general agreement with a study using buckle force transducers implanted on the PM bundle. Miyasaka and co-workers detached the PCL from its femoral attachment site and reattached it to its anatomical position using a metal plate instrumented with 12 strain gauges. They reported only small forces in the PCL throughout the range of passive flexion with a maximum of 4 (1)N at 90°. In another study, a robotic manipulator together with a universal force sensor (UFS) was used to measure the in situ force of the PCL in nine cadaveric knees during passive flexion. Except for 90° flexion, the force patterns of the PCL were close to the average force trend extracted from the data measured in vitro. Three studies presented modelling techniques to calculate the resultant PCL force during simulated passive flexion of the knee. With a 3D quasi-static model, Amiri and co-workers found the PCL to be slack at full extension, but the model predicted an increase in the ligament force to 40N at 30° that remained nearly unchanged until 90° of flexion. However, between 90° and 120° of flexion, a dramatic increase to 290N was calculated. Different results were achieved using a 3D finite element model of the knee, where the PCL was reported to become load-bearing only after 20° of flexion, and a maximum of 35N was predicted above 90° flexion. Contrary to these studies, large tensions in both PCL bundles at full extension were calculated using a 2D quasi-static model of the knee. Although the average force trend extracted from modelling studies had a similar pattern to that obtained from in vitro cadaveric investigations, analytical models generally overestimated the PCL tension at all flexion angles. PCL forces in the presence of a posterior tibial load (PTL) have been examined using load cells attached to the isolated femoral bony attachment of the PCL in cadaveric knees. These studies indicated that the PCL is subjected to approximately 22N tensile force at full extension, with a steady increase to 128N at 90° of flexion, followed by only a small increase (approximately 10N) thereafter. These values were corroborated by Vahey and Draganich using an indirect measurement technique, who estimated in situ forces 95N and 138N at 30° and 90° of flexion, respectively. In the presence of a 134N PTL, the PCL in situ force has also been indirectly assessed using robotic manipulators together with universal force sensors. These investigations indicated a relatively small force (29N) at full extension that gradually increased to 123N at 90° of flexion, and that reduced slightly thereafter. In general, the application of PTLs produced higher PCL forces than passive flexion, with the greatest difference at mid-flexion. ## Strain Patterns of the PCL during Active Flexion All studies examining forward lunge or squat used image-based techniques for strain measurement. To analyse joint kinematics during forward lunge, 3D geometrical models of the subjects’ knees were registered to orthogonal fluoroscopic images captured at different knee flexion angles. Lengths of the virtual bundles were estimated by tracking the reconstructed motion of the ligament attachment sites, reporting an increase in AL, PM and mid-PCL VBSs throughout flexion. At 90° of flexion, the average VBS in the AL and PM bundles were 32% and 25%. Surprisingly, the corresponding value for the mid-PCL was reported as being outside the values presented for the AL and PM bundles at approximately 23%. At 120° flexion, VBS for AL and PM bundles were reported as 35% and 27% respectively. A comparison between the experimental VBS patterns of the PCL bundles during forward lunge (loaded) with those of passive flexion (unloaded) revealed a considerable difference in the mean strains that became larger at higher flexion angles. Furthermore, no relaxation phase for the PM bundle could be observed during forward lunge. Moreover, contrary to passive knee flexion, the strain magnitudes of the PCL during forward lunge did not decline after 90° of flexion. For body-weight squat, the only study that contained strain data was an in vivo investigation by Nakagawa et al., who included MR-imaging of the knees in thirteen healthy subjects. They found a consistent elongation for the mid-PCL bundle up to 31% at 120°, followed by a declining trend. At 90° of flexion, the average VBS for the mid-PCL was calculated as 28%, which was somewhat higher than the corresponding value for forward lunge (23%) reported by Defrate and co- workers. ## Force Patterns of the PCL during Active Flexion Using a 2D musculoskeletal model, Shelburne and Pandy calculated the PCL force up to 90° flexion during squats. The model predicted zero PCL force until 10° of knee flexion but a steadily increasing force thereafter to a peak of 650N at 80°. The estimated PCL force decreased slightly between 80° and 90° flexion. Using similar approaches, Toutoungi and co-workers reported comparably small forces in the PCL during early knee flexion but predicted a much larger maximum PCL force of 2432N at 100° compared to Shelburne and Pandy. Utilizing a two- stage procedure, Shelburne and co-workers reported peak PCL forces of 274N during squatting, suggesting inconsistent results using musculoskeletal modelling approaches. ## Strain and Force Patterns of the PCL during Walking and Stair Activities All studies reporting PCL load data during walking used modelling techniques, but with highly variable results. While some investigators calculated almost no tension in the PCL (ranging from 0 to 27N), others found relatively high maximum PCL forces (0.2–1 times body-weight; BW). The time-point at which the peak PCL force occurred was also inconsistent between different studies and varied from early stance to late swing phase. Two studies reported on the strain patterns of the PCL during stair ascent or stair descent in vitro. Mahoney and co-workers measured in vitro strain in the AL bundle using implanted HEST sensors while the knee specimens were loaded to simulate stair ascent and descent. Until 40° of flexion, no significant strain was detected within the PCL for either activity studied. During stair ascent, the maximum RBS in the AL bundle was measured as 3%, which occurred between 100° and 110° flexion. Similarly, during stair descent, peak strains of 3% were measured between 100° and 110°, but these results were somewhat lower than the results of Emodi and co-workers who reported strains of 9% in the AL bundle at 105° flexion during stair descent using implanted differential variable reluctance transducer (DVRT) strain sensors. # Discussion The PCL is known to be one of the four major stabilizers of the knee joint. Many aspects of PCL biomechanics have been investigated based on strain and force measures of the ligament during different activities. However, due to the high variability in the reported load data, controversy still surrounds the definitive role of the ligament and its functional bundles during normal daily activities. In this comprehensive analysis of the literature, new understanding of the force and strain patterns within the PCL of healthy knees have been gained that would not be possible from a single study alone. Here, we have extracted the mean strain and force patterns of the PCL from an initial screening of over 3500 scientific articles, to provide the foundations for understanding PCL loading patterns during different activities. In addition, the influence of different assessment techniques on variations of the reported load data has been investigated as a possible source of bias. This systematic review has revealed that current knowledge of PCL loading mainly suffers from an insufficient number of studies, particularly during active and deep flexion activities. For example, no in vivo sensor-based study exists that have measured PCL force or strain in healthy human subjects. Only few studies have reported lengthening patterns of the virtual PCL bundles in vivo using indirect assessment techniques; and these studies have been conducted by single research groups or were limited to quasi-static forms of passive flexion, body-weight squat or forward lunge. In vitro cadaveric investigations on PCL loading are more frequent, but the majority have assessed PCL loads using indirect methods. While some in vitro studies have utilised strain or force sensors to investigate the effects of externally applied loads or muscle contractions on the ligament load, only a few of them have accurately captured the physiological loading conditions. Information on PCL loading from musculoskeletal and finite element modelling studies is limited and seems to be highly variable. As such, this comprehensive systematic review of the literature represents the single best understanding of loading conditions within the PCL available. Based on the statistical analysis of the reported data, the loading conditions in the PCL and its two main functional bundles during passive flexion have now been characterised in detail. However, studies on PCL biomechanics during physiological activities of daily living, as well as the extreme conditions experienced during impact or intense sport, are clearly missing and need additional investigation. The average loading patterns of the PCL confirm that both PCL strain and force magnitudes are dependent upon the knee flexion angle. During passive flexion, strains of the virtual AL and mid-PCL bundles follow similar patterns, consisting of an upward trend in the first 90° and a declining trend thereafter. With the zero strain condition defined at full extension of the knee, the PM bundle remains relatively relaxed throughout flexion. Interestingly, the magnitudes of the mid-PCL strains were between those of the AL and PM bundles. This fact adds confidence to the performed statistical analysis as the strain data were almost entirely extracted from independent studies. Although VBS does not seem to decline after 90° of flexion during forward lunges or squats, the slopes of the average strain patterns decreased at higher flexion angles. It was also observed that a posterior tibial load affects the PCL force more notably in the mid-range of knee flexion compared to early or deep flexion (c.f. in situ force passive vs 134N PTL;), due to the reduced restraining function of the PCL at low and high flexion angles. Although we did not show PCL loading beyond 120° in this review, Li and co-workers reported only a small contribution of the PCL to knee stability at 150° of flexion. This behaviour can be partially explained by changes in the spatial orientation of the PCL bundles, which become more vertically oriented at higher knee flexion angles, and therefore lose their efficiency in restraining posterior shear forces. However, some studies have reported an increasing trend for the posterior shear force acting on the knee joint throughout the whole range of knee flexion. This raises the question of which structures compensate for the weak restraining function of the PCL against posterior tibial loads at higher knee flexion angles. Here, posterior structures of the knee including skin, fat, hamstring muscles, joint capsule and meniscus are thought to partially provide posterior stability to a hyper-flexed knee joint, even though their individual contributions remain unknown. It is known that knee kinematics are activity-specific, suggesting that PCL strain is also affected by activity type. Our study supports this idea as the extracted strain patterns of the ligament bundles differ between passive flexion and forward lunge. This difference, which is more notable at higher knee flexion angles, might originate from variations in the contributions of muscle forces to knee joint kinematics during different activities. Both the quadriceps and the hamstrings muscle groups are involved during a forward lunge. While, contraction of the quadriceps can lead to a slight decrease of the PCL force, contraction of the hamstrings has been shown to considerably increase the PCL load. Therefore, compared to passive knee flexion, it is entirely plausible that the PCL is exposed to higher loads during sports activities. Variations in the reported strain and force data of the PCL during physiological loading might originate from different sources. The anatomy of the PCL is subject-specific, and geometrical parameters of the ligament including length, cross sectional area and attachment sites of the bundles have been described inconsistently. Therefore, it is likely that ligament strain varies among different subjects performing the same activity. It is also known that the lengthening pattern of the PCL is sensitive to the definition of its attachment sites, which is likely accompanied by errors especially when image-based techniques are used. In case of cadaveric investigations, specimens were prepared in different ways depending on the type and volume of the surrounding soft tissues resected. Furthermore, joint kinematics was simulated using a variety of techniques including manual, mechanical and robotic approaches, as well as internal or external poses of the bones, possibly adding to the reported variation in outcome measures. Additionally, strain of the ligament was measured using different direct and indirect methodologies. Given the many techniques and methods employed in assessing PCL loading, it is therefore hardly surprising that the reported outcomes vary considerably between the different studies. This systematic review provided the opportunity to compare RBS of the PCL (via direct approaches) with VBS (via indirect approaches) during passive knee flexion. The results from the statistical analysis demonstrate that VBS patterns during passive knee flexion are similar to RBS patterns but differ in magnitude. This should be kept in mind when interpreting results from studies reporting VBS; in particular, large VBS magnitudes may falsely imply failure of a ligament that in reality experiences safe strain magnitudes. For example, at 90° of passive flexion, the PCL was shown to be very lightly loaded; however, the average experimental VBS of the AL bundle was calculated at 24%, which is higher than the failure strain measured in situ (18%). Surprisingly, the calculated maximum VBS during forward lunge (35%) is approximately twice the magnitude of the bundle failure strain. The difference between VBS and RBS likely originates from the unknown zero-strain condition. Given the technical difficulties in identifying the accurate zero-strain condition, nearly all the included studies chose full extension of the knee as an arbitrary reference position, while in reality the PCL is considered lax at this position. Therefore, the selection of the reference position may add a systematic error to the strain data obtained by measuring the relative elongation of the virtual bundles. Moreover, in cases where the ligament is lax or wraps around a neighbouring structure, the straight-line representation between the attachment sites–or “virtual bundle”–does not represent the curvilinear length of the real bundle. Another key factor that might intensify the difference between RBS and VBS (especially for ligaments with large cross sectional areas) is the distribution of the strain across the ligament. RBS is generally measured using sensors attached to the surface of the ligament and therefore represents the local superficial strain. In contrast, VBS tends to represent the average strain of a virtual fibre inside the ligament. King and co-workers showed that large strain differences exist between the superficial fibres and the virtual lines that connect the attachment sites. While VBS can be used for intra-study comparison of the ligament strains under different loading conditions, any inter-study comparison of VBS is limited to investigations that have used a common reference length. Given the difficulties in interpreting ligament strain data from different studies, ligament force may be a more useful measure for investigating PCL biomechanics, especially considering the role of ligament viscoelastic properties in modulating the force—elongation relationships. PCL force can be measured directly by implanting force sensors; however, current sensors are somewhat large and stiff compared to the PCL structure and therefore interfere with normal ligament function. An alternative approach includes the use of robotic technology to measure the in situ force of the ligament based on the principle of superposition, thus avoiding limitations induced by direct contact of the force sensor and ligament. In this review, only a small difference was found between the in vitro and in situ forces determined in the PCL during passive knee flexion. In general, the PCL was found to be very slightly tense throughout flexion, with the smallest tension at 30° of knee flexion. Applying a PTL was shown to increase the PCL force at all flexion angles, but the increase was more considerable at mid-range of knee flexion. The fact that a PTL does not strongly affect the PCL force at full extension may explain why an isolated rupture of the PCL does not often lead to knee instability. Previous studies confirmed the role of meniscofemoral ligaments as well as posterolateral and posteromedial structures of the knee in providing knee stability near full knee extension. In the majority of reported studies, PTL had the greatest effect on the PCL force at 90° of knee flexion. This could explain the so called “dashboard injury” as the most common mechanism of PCL injury. During a traffic accident, the knee hits the dashboard while it is at approximately 90° of flexion and the posteriorly directed impact force tears the ligament. This finding is in agreement with studies reporting that an isolated sectioning of the PCL causes the highest change in posterior tibial translation in response to an applied PTL at 90° of flexion. Therefore, clinical testing for the integrity of the ligament should be performed at this position to obtain the highest sensitivity. We found that the choice of assessment technique affects the resulting loading patterns in the PCL. Ideally, PCL force or strain would be measured directly in vivo using implantable sensors. However, such in vivo procedures are generally invasive and the outcomes are limited to relative strain at the implantation site. Moreover, the imposed size and stiffness of the available sensors interfere with the natural ligament kinematics. Indirect assessment of the ligament strain via image-based tracking of the ligament attachment sites would therefore seem to deserve more attention for in vivo investigation of the ligament strain. Despite using substantially different methodologies, the mean strain patterns extracted from data obtained using indirect techniques follow those of direct investigations. This finding confirms the ability of indirect techniques to estimate the ligament strains. However, a correlation study to determine the relationship between RBS and VBS during different activities, and based on an adequate number of specimens tested, appears to be missing. Cadaveric investigations have provided the major contribution to our current knowledge on PCL loading; however, physiological loading conditions are difficult to reproduce in mechanical jigs. Finite element and musculoskeletal modelling techniques allow the simulation of different loading conditions, but limitations remain with respect to the accurate definition of the geometry, and the material properties, as well as the contact conditions within the joint and specifically with the ligament. As a consequence, compared to the load data measured in vivo or in vitro, the highest reported variability in the load data was obtained using modelling techniques. This fact emphasises the need to account for the 3D geometry, viscoelastic material properties and wrapping of the ligaments around neighbouring structures in future modelling studies. Some limitations of this systematic review need to be recognized. Only a limited number of activities were investigated, and also within a limited range of flexion (0° to 120°). Due to a lack of reported strain and force data, the loading patterns of the ligament during hyperextension and hyperflexion of the knee were not assessed in this review. In addition to a shortage of studies, the various assessment techniques used for measuring the ligament loading, in combination with the large number of dependent variables (flexion angle, bundles, activity etc.), preclude an effective statistical comparison to reveal the role of individual parameters on PCL loading. In some studies, the standard deviations in the reported outcome measures were not given, while in some others, the data were difficult to extract from e.g. graphical presentations. Consequently, standard deviations could only partially be taken into account for proper weighting of data points in the present statistical analysis. Many of the included studies were performed by the same research groups; a fact that might increase the risk of bias. Finally, while we have investigated PCL loading in the healthy knee joint, PCL function and its relationship with loading and kinematics in pathological cases is still less well understood. Given the insight obtained from this review, it is clear that certain considerations could help improve comparability of future investigations towards understanding PCL force and strain patterns. We propose that future studies should: - Adopt a standard definition of PCL anatomy, including the anterolateral and posteromedial bundles. - Use a standardised knee joint coordinate system for reporting the simulated bone kinematics. - Clearly report the approach taken to define zero-strain position. - Use standardised terminology (consistent with this review) for presenting results. - Couple in vivo image-based studies with in vitro sensor-based studies to relate VBS and RBS. # Conclusions This systematic review has provided an overview of the content and limitations in the current knowledge on PCL loading in the healthy knee. In general, the results of this review confirm that the individual bundles within the PCL are subjected to different loading patterns that are affected by both activity type and knee flexion angle. Moreover, it was found that the choice in assessment technique might have an important effect on the resulting load data. In the context of injury prevention, the extracted strain and force trends now provide the foundations for improved risk of injury assessment. The extracted loading patterns can also help determine the critical positions of the knee joint at which clinical tests are more sensitive to ligament deficiencies. In the context of PCL reconstruction, efforts should focus on restoring the loading patterns of the healthy ligament as summarised in this review. Finally, for rehabilitation of PCL injury, the range of knee joint motion during prescribed exercises can be adjusted based on safe strain levels to prevent any possible rupture of the reconstructed ligament. # Supporting Information We would like to acknowledge the Iranian Ministry of Science who have partially funded this study. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: SHHN WRT RL. Data curation: SHHN. Formal analysis: SHHN. Investigation: SHHN KO. Methodology: SHHN WRT RL. Project administration: RL WRT. Resources: SHHN RL WRT. Software: SHHN KO. Supervision: RL SFF JGS WRT. Validation: SHHN WRT. Visualization: SHHN WRT. Writing – original draft: SHHN KO RL WRT. Writing – review & editing: SHHN KO RL WRT SFF JGS.
# Introduction Sprouty (Spry) proteins were first discovered in Drosophila melanogaster as inhibitors of fibroblast growth factor receptor-induced tracheal branching. Subsequently, four mammalian isoforms of Sprouty (Spry1, Spry2, Spry3, and Spry4) were identified that are transcribed from four different genes. The different Spry isoforms have been shown to modulate the actions of receptor tyrosine kinases (RTKs); therefore, Spry proteins play a role in processes that require extensive RTK signaling such as organogenesis and tumorigenesis. Specifically in development, Spry proteins have been shown to regulate the process of angiogenesis, patterning of the midbrain and anterior hindbrain, and development of the kidneys, lungs, limb buds, craniofacial features, and trunk. After development Spry proteins continue to regulate angiogenesis, cell proliferation, migration and survival. Likewise, the role of Spry proteins, mainly Spry1 and Spry2, in cancer has also been investigated. Previous research has shown that the levels of Spry1 and Spry2 are decreased in cancers of the breast, lung, liver, and prostate correlating to poor patient prognosis. Due to the important role Spry proteins play in development and tumorigenesis, it is crucial to understand how Spry levels are regulated. We focused on Spry2, which is ubiquitously expressed and most studied among the Spry isoforms. Prior studies have concentrated on the regulation of the Spry2 protein through a variety of posttranslational modifications such as ubiquitylation or phosphorylation (reviewed). However, early on Spry expression patterns during development were assessed and showed that the transcription of SPRY is upregulated by growth factors elevating Spry protein levels in the centers of growth factor signaling (e.g. limb buds), thereby optimizing the ability of Spry proteins to act as negative feedback inhibitors of the enhanced RTK signaling in these areas. Additionally, while many other transcription factors have been predicted to bind to the SPRY2 promoter, few have actually been shown to bind. Ding et al. performed a functional analysis of the SPRY2 promoter and identified that Ap2, Ets-GATA, and SP-1 bind to the SPRY2 promoter enhancing its transcription. However, the functional significance of the binding of these transcription factors to the SPRY2 promoter remains unknown. Because Spry2 levels are reduced in some forms of cancer, the regulation of Spry2 in cancer has been investigated. Most studies, however, have focused on the post-transcriptional regulation of Spry2. The few studies that have investigated transcriptional regulation of SPRY2 promoter have shown that FOXO and beta-catenin bind to the SPRY2 promoter and induce its transcription. In terms of epigenetic modifications, the promoters of Spry4 and Spry2 have been shown to be hypermethylated in prostate cancer, but not breast cancer. Two reports suggest that SPRY2 promoter is hypermethylated in hepatocellular carcinomas, but another report suggests otherwise. In both development and tumorigenesis, cells experience a hypoxic environment to which they adapt to by upregulating the transcription factors, hypoxia inducible factors (HIFs). HIFs are composed of an oxygen-regulated alpha subunit (HIF1α/HIF2α) and a beta subunit (HIF1β a.k.a. aryl hydrocarbon receptor nuclear translocator (ARNT)). Opposite to the actions of Spry2, HIFs promote proliferation, migration, and survival of cells by increasing the transcription of a number of genes that regulate these processes (reviewed in). Because Spry2 protein levels are decreased in hepatocellular carcinomas and given the opposite actions of Spry2 and HIF1α/HIF2α on cell proliferation and migration, we performed an in silico analysis of the SPRY2 promoter for hypoxia response elements (HRE) with the consensus sequence `5’-A/GCGTG-3’` and found 10 putative HREs; five in the proximal promoter and five in the first intron. Therefore, the purpose of this study was to determine whether HIF1α/HIF2α regulated the transcription of the SPRY2 promoter. Herein, we demonstrate that, in the hepatoma cell line Hep3B, endogenous HIF1α and HIF2α decreased the mRNA levels of SPRY2 with a concomitant decrease in the protein levels of Spry2. While the stability of the SPRY2 mRNA wasn’t altered by HIF silencing, inhibiting DNA methylation with decitabine (DAC) abolished the increase in SPRY2 mRNA when the expression of HIF1α/2α were silenced. Chromatin Immunoprecipitation (ChIP) assays revealed HIF1α/2α bind to regions in both the proximal promoter and first intron, each of which contains four and five HIF1α/HIF2α binding sites, respectively. Methylation of the proximal promoter of SPRY2 was also observed and HIF1α/2α silencing decreased this methylation. Furthermore, ChIP assays revealed association of DNA methyltransferase 1 (DNMT1) with the proximal promoter and first intron of SPRY2 and silencing of HIF1α/2α diminished this interaction. Finally, silencing of DNMT1 mimicked the actions of HIF1α/2α silencing in elevating SPRY2 mRNA and protein levels. However, simultaneous silencing of DNMT1 and HIF1α/2α did not elevate SPRY2 mRNA or protein levels additively suggesting that DNMT1 and HIF1α/2α work through a common mechanism. These data suggest that HIF1α/2α suppress SPRY2 transcription, in part by increasing SPRY2 promoter methylation by DNMT1. # Results ## HIF1α and HIF2α decrease the mRNA and protein levels of Spry2 To investigate if HIF1α and HIF2α regulated SPRY2 mRNA and protein levels, we used siRNAs to silence the expression of endogenous HIF1α and HIF2α proteins in the hepatoma cell line Hep3B. While the silencing of HIF1α or HIF2α separately elevated SPRY2 mRNA levels by about 100% each, silencing of both HIF1α and HIF2α together more profoundly (200%) elevated SPRY2 mRNA levels (left panel); the efficient silencing of both HIF1α and HIF2α by the siRNAs is also shown in (right panels). Consistent with the change in SPRY2 mRNA levels, an increase in Spry2 protein levels was observed with HIF1α and HIF2α silencing. Here again, silencing both HIF1α and HIF2α had a larger effect on Spry2 protein levels, increasing them by about 150%, while silencing either HIF1α or HIF2α alone increased Spry2 protein to a lesser extent (\~100%) (right panel). Conversely, in HEK293T cells, ectopic expression of HIF1α or HIF2α alone or both isoforms together decreased SPRY2 mRNA levels. Once again, the expression of both HIF1α and HIF2α had a more profound effect on SPRY2 mRNA levels by reducing them by about 50% while the expression of either HIF1α or HIF2α only reduced SPRY2 mRNA by 20% and 30%, respectively (, left panel). Notably, we found that the co-expression of HIF1β, which dimerizes with HIF1α and HIF2α, is necessary to observe the effects of HIF1α or HIF2α overexpression, probably because endogenous HIF1β levels were not adequate to dimerize with the expressed HIFα subunits. However, HIF1β expression by itself does not alter SPRY2 mRNA levels (left panel). As a positive control to ensure that overexpressed HIF1α was modulating transcription appropriately, we monitored the mRNA levels of the HIF1α responsive gene phosphoglycerate kinase 1 (PGK-1). As expected, co- expressing HIF1α and HIF1β elevated PGK-1 mRNA while expression of HIF1β alone or together with HIF2α had no effect (right panel). Together, these data suggest that both HIF1α and HIF2α contribute to the decrease of the mRNA and protein levels of Spry2. ## HIF1α and HIF2α do not alter the stability of SPRY2 mRNA, but HIF1α and HIF2α bind to the proximal promoter and intron of SPRY2 Since changes in mRNA levels may reflect either an alteration in the half-life of the mRNA and/or the rate of its transcription, we first determined whether silencing of HIF1α/HIF2α modulated the stability of SPRY2 mRNA. For this purpose, Hep3B cells transfected with control or HIF1α and HIF2α siRNAs were treated with actinomycin D to inhibit RNA synthesis and the mRNA levels of SPRY2 were then monitored over 2 hours by qRT-PCR. As shown in, during the 2 hours of actinomycin D treatment, SPRY2 mRNA levels were reduced by approximately 80%. However, the rate of SPRY2 mRNA degradation was not significantly different whether or not HIF1α and HIF2α were silenced; the half- life of SPRY2 mRNA was 0.54 versus 0.70 hours in control versus HIF1α and HIF2α siRNA transfected cells, respectively. Interestingly, over the entire time course the cells with HIF1α and HIF2α silenced had higher SPRY2 mRNA levels, which is to be expected given the observation that SPRY2 mRNA levels are elevated when HIF1α and HIF2α are silenced. Since these data suggest HIF1α and HIF2α do not alter the stability of SPRY2 mRNA, we next investigated if HIF1α and HIF2α regulate the transcription of SPRY2 mRNA. It is well established that both HIF1α and HIF2α bind to Hypoxia Response Elements (HREs) with the consensus sequence `5’-A/GCGTG-3’` in the promoter of the genes they regulate. We performed an in silico analysis to determine if the SPRY2 proximal promoter (-3850 to +1) or first intron (+1 to +3395) contained putative HREs. Indeed, we found 5 putative HREs in the proximal promoter located at positions -3271, -1360, -1319, -570, and -390 and 5 putative HREs in the first intron located at nucleotides 1770, 1811, 1830, 1858, and 1953. Furthermore, some of the HREs are conserved amongst primates as well as mice, rats, and rabbits. To determine if HIF1α and HIF2α could bind any of these putative HREs, we performed a chromatin immunoprecipitation (ChIP) assay using a HIF1β antibody to immunoprecipitate both HIF1α and HIF2α bound DNA. We then used primers targeted against 4 different HRE containing areas of the proximal promoter and intron of SPRY2 to quantify the amount of DNA that was immunoprecipitated by HIF1β. The location of the primers is shown in. Intriguingly, there was no significant enrichment of HIF1α/HIF2α/HIF1β on the HRE located at nt -3271 in the proximal promoter of SPRY2 as shown by primer set 1. However, primer sets 2 and 3, targeting the four HREs closest to the transcription start site of SPRY2, and primer set 4, targeting the intron of SPRY2, showed significant enrichment of HIF1α/HIF2α/HIF1β. Furthermore, silencing of HIF1α and HIF2α significantly reduced the DNA enrichment in ChIP assays with primer sets for these sites. As a positive control, by identical ChIP assays, we monitored the enrichment of the HREs located in the promoter of the HIF1α target gene phosphofructokinase (PFK) and HIF2α target gene erythropoietin (EPO). As expected, the immunoprecipitation of HIF1β was greatly enriched with the DNA corresponding to HREs in the promoters of both PFK and EPO genes, and silencing of HIF1α or HIF2α greatly diminished the DNA enrichment from PFK and EPO promoters, respectively. These latter findings authenticate that the combination of ChIP assays with HIF1β antibody and HIF1α/HIF2α silencing is a valid approach to study HREs on promoters. Together the data in suggest that HIF1α and HIF2α do not alter the stability of SPRY2 mRNA, but both HIF1α and HIF2α bind to the proximal promoter and first intron of the SPRY2 gene. ## HIF1α and HIF2α regulate SPRY2 mRNA levels by modulating the methylation of the SPRY2 promoter The data in Figs and suggest that HIF1α/HIF2α bind to SPRY2 promoter and first intron to repress the expression of Spry2. While studies have shown that hypoxia represses a set of genes, not many studies have shown that HIF1α or HIF2α specifically repress gene transcription. Furthermore, the precise mechanisms of repression are unknown or vary depending on the gene. Methylation of promoters is well known to repress gene transcription and hypoxia and HIF1α have been shown to modulate methylation of genes. Therefore, we investigated whether the methylation status of the SPRY2 promoter: (a) altered SPRY2 mRNA levels, (b) modulated the ability of HIF1α/HIF2α silencing to alter SPRY2 mRNA levels, and (c) was regulated by endogenous HIF1α/HIF2α. By treating cells with decitabine (DAC), an inhibitor of DNA methyltransferases, we first determined whether the CpG islands in the proximal SPRY2 promoter were methylated. Using bisulphite-treated genomic DNA and primers that specifically detect methylated (M-1, M-2) and unmethylated (U-1, U2) DNA corresponding to the regions 1 and 2 on the SPRY2 promoter shown in inset (also see), we determined the methylation status of the SPRY2 promoter after treatment with DAC or its vehicle. As shown in, DAC treatment significantly decreased the methylation of the SPRY2 promoter at primer sites 1 and 2 by 80% and 47%, respectively. Concomitantly, as expected, the amount of unmethylated promoter monitored by unmethylated DNA-specific primers for sites 1 and 2 was elevated by 78% and 80%, respectively. These findings demonstrate that the SPRY2 promoter is methylated and DAC treatment effectively reduces its methylation. Next, we determined whether treatment of cells with DAC altered SPRY2 mRNA levels and the ability of HIF1α/HIF2α silencing to further modulate the amount of SPRY2 mRNA. As shown in, DAC treatment of cells increased SPRY2 mRNA levels by nearly 2-fold. However, while HIF1α/HIF2α silencing in vehicle treated cells (control) elevated SPRY2 mRNA levels by \~100%, in the presence of DAC, silencing of HIF1α/HIF2α did not significantly increase SPRY2 mRNA levels; the efficient silencing of HIF1α and HIF2α is shown in the right panels of. These data in suggest that in Hep3B cells the reduction in methylation of the SPRY2 promoter by DAC increases SPRY2 mRNA levels and that DNA methylation plays a role in HIF1α/HIF2α- mediated decrease of SPRY2 mRNA levels. To directly assess whether silencing of HIF1α/HIF2α altered the methylation status of the SPRY2 promoter, using primers corresponding to regions 1 and 2 on the SPRY2 promoter that specifically recognize methylated (M-1, and M-2) vs. unmethylated DNA (U-1 and U-2), we determined the methylations status of the SPRY2 proximal promoter with and without HIF1α/HIF2α silencing. As shown in, silencing of HIF1α and HIF2α decreased SPRY2 promoter methylation detected by M-1 and M-2 primers by 52% and 42%, respectively, and increased the amounts of unmethylated SPRY2 promoter monitored by U-1 and U-2 primers by 111% and 82%, respectively. Overall, these data suggest that the SPRY2 promoter is methylated in Hep3B cells, methylation of the promoter represses the expression of SPRY2 mRNA, and endogenous HIF1α and HIF2α increase the methylation of the SPRY2 promoter to decrease Spry2 mRNA levels. ## HIF1α and HIF2α regulate binding of DNMT1 to the SPRY2 proximal promoter and intron and DNMT1 contributes toward HIF1α and HIF2α- mediated regulation of SPRY2 mRNA DNA methylation is a predominant epigenetic modification in mammals and is catalyzed by DNA methyltransferases (DNMTs), which are the enzymes that methylate the 5-position of cytosine in DNA primarily within a CpG dinucleotide. There are four DNMT isoforms: DNMT1, DNMT3a, DNMT3b and DNMT3L. DNMT1 is the most abundant of these enzymes and is involved in maintaining methylation pattern by methylating newly replicated DNA. DNMT3a and DNMT3b are considered de novo methyltransferases, since they add methyl groups to completely unmethylated DNA during development of an embryo. DNMT3L does not possess any inherent enzymatic activity. Intriguingly, hypermethylation of specific promoter regions has been implicated in promoting tumorigenesis (reviewed in). One potential cause of this hypermethylation is the upregulation of DNMT’s, in particular DNMT1. In fact, previous studies showed that expression of DNMT1 in cultured cells increased CpG island methylation and resulted in cellular transformation. With this in mind, we focused our studies on DNMT1. Given our observations that HIF1α/HIF2α alter methylation of the SPRY2 promoter and the recent findings demonstrating that laccaic acid (LCA) is a direct inhibitor of DNMT1 activity, we investigated whether inhibition of DNMT1 with LCA altered the ability of HIF1α and HIF2α to regulate SPRY2 mRNA levels. Consistent with the data shown in Figs and, silencing HIF1α and HIF2α resulted in a 170% increase in SPRY2 mRNA. Importantly, treatment of cells with LCA, the DNMT1 inhibitor, also increased SPRY2 mRNA levels by 87% and attenuated the increase in the levels of Spry2 mRNA observed with HIF1α and HIF2α silencing (170% in control versus 70% in LCA treated) (left panel); the efficient silencing of HIF1α and HIF2α is shown in (right panels). These data suggest that DNMT1 regulates transcription of SPRY2 mRNA and is partially involved in the ability of HIF1α and HIF2α to repress SPRY2 mRNA levels. To investigate if HIF1α and HIF2α altered the binding of DNMT1 to the SPRY2 promoter, we performed ChIP assays with the DNMT1 antibody and monitored the amount of the proximal promoter and intron of SPRY2 that immunoprecipitated with DNMT1 with and without silencing the expression of HIF1α/HIF2α. The location of the primer sets that bind to these regions is shown in. Similar to the ChIP assays shown in, DNMT1 immunoprecipitates were enriched with SPRY2 gene regions corresponding to the promoter near the transcription start site (primer set 3) and the first intron of SPRY2 gene (primer set 4). Additionally, silencing of HIF1α/HIF2α decreased the enrichment of DNA corresponding to these regions of the SPRY2 gene in ChIP assays performed with the anti-DNMT1 antibody. These data suggest that DNMT1, either directly or indirectly, binds to the proximal promoter and first intron of the SPRY2 gene and that HIF1α and HIF2α regulate the binding of DNMT1 to these regions. One mechanism by which DNMT1 could be recruited to the SPRY2 promoter and first intron is via interactions with HIF1α and/or HIF2α. However, despite numerous attempts using different conditions, we did not observe co- immunoprecipitation of HIF1α/HIF2α and DNMT1 (not shown) irrespective of which protein we immunoprecipitated. Thus, the precise mechanism by which HIF1α/HIF2α regulates binding of DNMT1 to the SPRY2 promoter and first intron remains to be defined. To further elucidate the role of DNMT1 in the regulation of SPRY2 mRNA levels by HIF1α/HIF2α, we monitored the levels of SPRY2 mRNA in cells transfected with either control siRNA or siRNAs targeting HIF1α/HIF2α when DNMT1 expression was either silenced or not. In these studies, we included two hepatocellular carcinoma cell lines, HuH7 and Hep3B, to demonstrate the generality of the mechanisms investigated in this report. As shown in and as observed previously (Figs, &), silencing of HIF1α/HIF2α in Hep3B cells increased SPRY2 mRNA levels by 130%. The silencing of DNMT1 alone also increased SPRY2 mRNA levels by \~50%. The latter increase is consistent with the data in with LCA, suggesting that DNMT1, in part modulates SPRY2 promoter activity. Of note, silencing of HIF1α/HIF2α had no effect on DNMT1 mRNA levels and vice versa (right panels). Most interestingly, when DNMT1 and HIF1α/HIF2α expression was simultaneously silenced, SPRY2 mRNA levels were not increased in an additive manner. Similar results were observed in HuH7 cells with the exception that silencing of HIF1α/HIF2α elevated SPRY2 mRNA levels by \~50% instead of the 100% increase observed in Hep3B cells. Nevertheless, the other changes described for Hep3B cells with DNMT1 silencing alone and together with HIF1α/HIF2α silencing were also similar and statistically significant in HuH7 cells. Consistent with changes in the SPRY2 mRNA levels, silencing HIF1α/HIF2α or DNMT1 alone significantly elevated Spry2 protein levels in Hep3B cells and HuH7 cells. However, as described above for SPRY2 mRNA levels, simultaneous silencing of DNMT1 and HIF1α/HIF2α did not additively increase Spry2 protein levels suggesting that DNMT1 and HIF1α/HIF2α regulate Spry2 protein levels by a common mechanism. Together, the data in Figs and suggest that HIF1α and HIF2α regulate SPRY2 mRNA and protein levels, in part, by regulating the binding of DNMT1 to the promoter and intron of SPRY2. # Discussion Ever since the discovery of the first SPRY gene in Drosophila, it has been clear that the Spry family of proteins, in a variety of species, play an important role in normal development of organs. Because Spry proteins modulate the biological actions of growth factors that mediate their signaling via Receptor Tyrosine Kinases, a number of studies have examined the role that Spry proteins play in disease states associated with enhanced Receptor Tyrosine Kinase activities. Essentially, these studies have shown that in certain disease states such as carcinomas of the breast, liver, lung and prostate, the levels of Spry proteins, especially Spry2, are decreased and probably contribute toward the pathogenesis of the disease. Indeed, in hepatocellular carcinoma and breast cancer, a decrease in Spry2 levels has been correlated with poor prognosis and a decrease in patient survival. For these reasons, it has been suggested to use Spry2 protein levels as a prognostic marker and Spry proteins have been dubbed “tumor suppressors”. Thus, Spry proteins play an important role in normal development and in tumorigenesis. During development and tumorigenesis, rapid cell proliferation that precedes angiogenesis exposes cells to hypoxia. The cells adapt to hypoxia by stabilizing the Hypoxia Inducible Factors, HIF1α and HIF2α, which are transcription factors that increase the expression of certain genes that promote cell survival and proliferation (reviewed in). As such, in cancerous states, HIFs can be considered tumor promoters. Recently, we demonstrated that one of the mechanisms by which Spry2 exerts its “tumor suppressor” functions is by decreasing the stability of HIF1α and HIF2α with a corresponding decrease in their ability to alter transcription of the HIF1α and HIF2α target genes. Herein, we asked the opposite question i.e. do HIF1α and HIF2α regulate Spry2 levels? The evidence presented in this report demonstrates that endogenous HIF1α and HIF2α decrease the amounts of SPRY2 mRNA and protein as silencing of endogenous HIF1α and HIF2α elevate both SPRY2 transcript and protein levels. Conversely, ectopically expressed HIF1α and HIF2α decrease the amounts of SPRY2 mRNA. The increase in SPRY2 mRNA levels is not the result of changes in stability of the transcript, but for the following reasons, is a function of the change in transcription of the SPRY2 gene. First, HIF1α/HIF2α bind to the proximal promoter and first intron of the SPRY2 gene that contains a total of nine putative HRE consensus sequences (4 in proximal promoter and 5 in first intron). Second, HIF1α and HIF2α increase the methylation of the SPRY2 promoter and repress SPRY2 mRNA expression. Third, the inhibitor of DNMT1, LCA, and DNMT1-specific siRNA augment SPRY2 mRNA and protein levels mimicking the effect of HIF1α/HIF2α silencing had to a lesser extent. As discussed later, simultaneous silencing HIF1α/HIF2α with DNMT1 resulted in a significant increase only in SPRY2 transcript and not Spry2 protein levels. However, this increase in SPRY2 mRNA with DNMT1 and HIF1α/HIF2α silenced was not additive suggesting DNMT1 and HIF1α/HIF2α regulate SPRY2 through a similar mechanism. Moreover, our studies also show that DNMT1 binds the SPRY2 promoter and silencing of HIF1α/HIF2α decreases this association of DNMT1 with the SPRY2 promoter. These findings suggest that HIF1α/HIF2α, in some manner, recruit DNMT1 to the SPRY2 promoter to alter the methylation state of the promoter and, therefore, transcription of the SPRY2 gene. One obvious mechanism would be the association between HIF1α/HIF2α and DNMT1 to recruit DNMT1 to the SPRY2 promoter. However, despite several attempts using different immunoprecipitation conditions to elucidate interactions between HIF1α/HIF2α and DNMT1, we have been unable to show that these proteins reside in the same complex. It is possible that HIF1α/HIF2α alter the expression of some other protein that then permits the recruitment of DNMT1 to the SPRY2 promoter. Since HIF1α/HIF2α recruit DNMT1 to the SPRY2 promoter and first intron, it would be expected that silencing of DNMT1 would increase SPRY2 mRNA to the same extent as silencing of HIF1α/HIF2α alone or together with DNMT1. However, in both Hep3B and HuH7 cells, the silencing of DNMT1 alone elevated SPRY2 mRNA levels to a significantly lesser extent than when HIF1α/HIF2α and DNMT1 were silenced simultaneously. These data suggest that other DNMTs may also contribute toward HIF1α/HIF2α-mediated methylation of the SPRY2 promoter. Given the large number of other members of the DNMT family and possible involvement of histone modifications, the identification of the other mechanisms (besides DNMT1) that may contribute toward HIF1α/HIF2α- mediated regulation of SPRY2 promoter methylation should be the subject of future studies. Notably, the extent to which DNMT1 silencing alone and in combination with HIF1α/HIF2α elevated Spry2 protein is not significantly different. However, this may be the result of the semi-quantitative nature of Western blot quantification as compared to the more rigorous quantitative analyses of mRNA levels by real time PCR. In the light of this report and our recently published findings, one very important aspect of the regulatory interactions between HIF1α/HIF2α and Spry2 that emerges is that Spry2 regulates the stability of the HIF1α/HIF2α proteins and thereby attenuates their ability to alter transcription of the HIF1α/HIF2α-responsive genes, such as those regulating glucose uptake and glycolysis that play a critical role in survival of cells in hypoxia. Conversely, HIF1α/HIF2α by regulating the methylation status of the SPRY2 promoter repress expression of SPRY2 mRNA and protein. Hence, there is a reciprocal cross talk between the “tumor suppressor”, Spry2, and “tumor promoters”, HIF1α/HIF2α. However, the extent to which one dominates over the other may rely on the expression of other pertinent proteins that play a role e.g. pVHL in terms of HIF1α/HIF2α stability regulation by Spry2 and forms of DNMT that facilitate HIF1α/HIF2α-mediated alterations in SPRY2 promoter methylation. These elements may account for the differences in the extent to which SPRY2 mRNA is elevated upon silencing of HIF1α/HIF2α or DNMT1 in Hep3B and HuH7 cells Nevertheless, the cross-regulation between Spry2 and HIF1α/HIF2α would allow equilibrium to be reached so that one protein does not overly regulate the other to alter biological outcomes. This scenario would be beneficial in normal development and one could envisage this cross talk to regulate growth of tumors to some extent. In this context, Lee et al. reported that in HuH7 and Hep3B cells cultured in normoxia the SPRY2 gene is not methylated. This would be expected since in normoxia HIF1α/HIF2α levels are low and, therefore, would not facilitate the recruitment of DNMT1 to the SPRY2 promoter. Extending this to tumors, it would be expected that in the hypoxic zones of tumors, elevated HIF1α/HIF2α protein levels would methylate the SPRY2 promoter and first intron to a greater extent than in normoxic areas of the tumors. Although the stability of the Spry2 protein is enhanced in hypoxia, over time the decreased transcription of the SPRY2 gene in hypoxic regions of tumors would be expected to decrease the protein levels of Spry2, diminish “tumor suppressor” actions of Spry2, and reduce the ability of Spry2 to oppose the “tumor promoting” actions of HIF1α/HIF2α. Hence, targeting HIF1α and HIF2α in tumors would not only suppress the “tumor promoting” actions of these transcription factors but by elevating SPRY2 gene transcription, elevate Spry2 protein levels and, therefore, the tumor suppressing actions of Spry2. Interestingly, the HREs in SPRY2 promoter are conserved in other mammalian species and the promoters and introns of other human SPRY genes (SPRY1, SPRY3, and SPRY4) also contain putative HREs. Also, SPRY1 and SPRY4 promoters have been reported to be methylated. Likewise, a previous study showed that SPRY4 mRNA levels are increased in hypoxia, while conflicting studies, perhaps due to cell type, showed SPRY1 mRNA levels either increased or decreased by hypoxia. In Hep3B cells, we observed that Spry1 protein levels were increased while Spry4 protein levels were decreased when HIF1α and HIF2α were silenced (data not shown); Spry3 levels were undetectable in Hep3B cells (not shown). Thus, it is tempting to speculate that SPRY1 gene is also regulated by HIF1α/HIF2α via mechanisms described in this report for SPRY2. On the other hand, SPRY4 may be regulated by HIF1α and HIF2α in an opposing manner to SPRY2 by an as yet to be identified mechanism. SPRY3 promoter does not contain CpG islands and is not methylated and is probably not regulated by HIF1α/HIF2α via a methylation-dependent mechanism. Overall, our findings described here unveil a new mechanism by which SPRY2 gene expression is regulated by HIF1α/HIF2α. By binding to regions of the proximal promoter and first intron of SPRY2, HIF1α and HIF2α increase the methylation of the SPRY2 promoter. We identified DNMT1 as a contributor toward this process as silencing or inhibiting DNMT1 attenuated HIF1α/HIF2α silencing mediated elevations in SPRY2 mRNA and protein. These findings demonstrate that HIF1α/HIF2α, by repressing the expression of Spry2, can decrease the anti- tumorigenic actions of Spry2 protein. # Experiment procedures ## Reagents and antibodies Actinomycin D was purchased from Calbiochem, and decitabine or 5-aza-2’-deoxycytidine (also called dacogen, DAC) was from Cayman Chemical. Laccaic acid A (LCA) was obtained from TCI America. All siRNAs and PCR primers, including general PCR primers, real time PCR primers and primers for methylation specific and non-methylation specific PCR, were synthesized by Integrated DNA Technologies Inc. The sequences of the primers are listed in. Antibodies used for Western blotting and chromatin immunoprecipitation were from the following companies: Sprouty2 (against N-terminus, Sigma), HIF1α (BD Transduction Laboratories), HIF2α (R&D Systems), HIF1β (Santa Cruz Biotechnology), and DNMT1 (AbCam). ## Plasmids Human full-length HIF1α is PCR amplified with primers carrying HindIII and NotI sites from HIF1α cDNA clone (OriGene Technologies, Inc.) and inserted in pcDNA3 at HindIII and NotI sites. HIF2α is PCR amplified with primers harboring BamHI and NotI sites from pOTB7-HIF2α (Thermo Scientific.) and inserted in pcDNA3 at the corresponding sites. Plasmid pcDNA3-HIF1β was kindly provided by Dr. Guo- Qiang Chen (Shanghai Jiaotong University, China). Plasmid pGL2-Pfkfb3/-3566 was kindly provided by Dr. Ramon Bartrons, University of Barcelona. ## Cell culture, hypoxia and treatments Hep3B and HuH7 cells were obtained from Dr. Basabi Rana, University of Illinois, Chicago. HEK293T were incubated in DMEM supplemented with 10% FBS, penicillin (100 units/mL), and streptomycin (100μg/mL). Hep3B cells were incubated in MEM supplemented with non-essential amino acids, sodium pyruvate and HEPES in addition to 10% FBS, penicillin (100 units/mL), and streptomycin (100μg/mL). HuH7 cells were incubated in DMEM F12 1:1 supplemented with HEPES, 10% FBS, penicillin (100 units/mL), and streptomycin (100 μg/mL). For normoxic conditions, cells were maintained at ambient O<sub>2</sub> levels (21% O<sub>2</sub>) and 5% CO<sub>2</sub> at 37°C. For hypoxic conditions, cells were maintained at 3% O<sub>2</sub> and 5% CO<sub>2</sub> in a Coy Hypoxic Chamber (Grass Lake, Michigan) at 37°C. All media used for hypoxia experiments were pre- equilibrated under hypoxic conditions overnight before use. To examine whether HIFs alter SPRY2 mRNA stability, Hep3B cells treated with 3μg/mL actinomycin D after siRNA transfection and 24 h hypoxic exposure were lysed at the indicated times with Trizol for RNA extraction. To investigate the involvement of DNA methylation in the regulation of SPRY2 mRNA expression by HIFs, Hep3B cells were incubated with DNA methylation inhibitors, decitabine (DAC) at 5μM or laccaic acid A (LCA) at 50 μg/mL, for 24 h the day after cells were plated. Subsequently, cells were transfected with siRNAs in fresh medium containing DAC or LCA and then maintained under hypoxia for another 24 h before use. ## Overexpression of HIF1α and HIF2α HEK293T cells were seeded in 3.5-cm dishes at 2 x 10<sup>5</sup> /dish and transfected next day with 250 ng each of pcDNA3-HIF1β, pcDNA3-HIF1α and/or pcDNA3-HIF2α as indicated using Transit2020 transfection reagent (Mirus) following the manufacturer’s instructions. The total amount of plasmids transfected into each dish was kept the same by adding empty vector pcDNA3. Cells were incubated under normoxic condition for 40 h after transfection before RNA extraction. ### Silencing with siRNAs Hep3B cells were plated in 3.5-cm dishes at 3 x 10<sup>5</sup> /dish or HuH7 cells were plated in 3.5-cm dishes at 2 x 10<sup>5</sup>/dish. Next day, cells were transfected with mutant siRNA or siRNAs against HIF1α, HIF2α or both at 20 nM each or for the experiments in, cells were transfected with mutant, HIF1α and HIF2α, DNMT1 alone, or HIF1α and HIF2α and DNMT1 siRNAs (20nM each) with TKO transfection reagent (Mirus). After overnight transfection, cells were incubated in the hypoxic chamber for 24 h (for mRNA detection) or 32 h (for Western blotting). The sequences of siRNAs are: mutant siRNA, sense `5'-GUC AGC AGA ACA AAA GUA GTT-3'` and antisense `5'-CUA CUU UUG GUU CUG CUG ACT T-3'`; HIF1α, sense `5’-GAA GGA ACC UGA UGC UUU AAC UUT G-3’` and antisense `5’-CAA AGU UAA AGC AUC AGG UUC CUU CUU-3’`; HIF2α, sense `5'-GCU GGA GUA UGA AGA GCA AGC CUT C-3'` and antisense `5'-GAA GGC UUG CUC UUC AUA CUC CAG CUG-3'`. ## RNA isolation and real time PCR Total RNA was isolated with Trizol reagent following the manufacturer’s protocol (Invitrogen). The extracted total RNA (500 ng) was then converted to cDNA with SuperScript VILO cDNA synthesis kit (Invitrogen) according to manufacturer’s instructions. To detect mRNA amounts for HIF1α, HIF2α, SPRY2 and PGK1, real time PCR was performed with specific primers and probes and FastStart Universal Probe Master Mix (Roche Life Science) using the CFX96 real-time PCR detection system (Bio-Rad). PCR conditions were optimized for the primers/probe for each gene. The mRNA amounts of each gene were normalized with 18S rRNA. ## Chromatin Immunoprecipitation (ChIP) Hep3B cells were plated at 5 x 10<sup>5</sup> /dish in 6-cm dishes. Next day, cells were transfected with siRNAs as stated above. An extra dish of cells were transfected in parallel with siRNA and trypsinized for cell counting before use. After a 32 h hypoxic incubation (3% O<sub>2</sub>), cells were crosslinked with 1% formaldehyde for 10 min at room temperature. The crosslinking reactions were terminated by incubating in 0.125 M glycine for 5 min at room temp. The cells were washed twice with cold PBS, scraped into cold PBS containing protease inhibitors (1 μg/mL aprotinin and 1 μg/mL pepstatin, 2 μg/mL leupeptin and 1 mM phenylmethylsulfonyl fluoride) and pelleted by centrifugation at 2000 rpm for 5 min. The cell pellet was resuspended in SDS lysis buffer (1% SDS, 10 mM EDTA, 50 mM Tris-HCl, pH 8.0) containing the above protease inhibitors (200 ml/1 x 10<sup>6</sup>) and incubated for 10 min on ice. The cell lysate was sonicated to shear DNA followed by centrifugation to remove pellet. The supernatant was diluted 10 times with dilution buffer (0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, 16.7 mM Tris-HCl, pH 8.1, 167 mM NaCl plus the above protease inhibitors) and pre-cleared with salmon sperm DNA/protein G beads. The cleared supernatant (from 1 x 10<sup>6</sup> cells) was incubated with 2.5μg anti-HIF1β antibody or rabbit IgG overnight at 4°C and immunoprecipitated with 25 μL protein G agarose beads the next day. For ChIP with anti-DNMT1 antibody, 1.5 x 10<sup>6</sup> cells were used per ChIP and mouse IgG was used as control antibody. The immunoprecipitates were washed sequentially with the following buffers: low salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH 8.0, 150 mM NaCl), high salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH 8.0, 500 mM NaCl), LiCl wash buffer (0.25 M LiCl, 1% IGEPAL-CA630, 1% Sodium deoxycholate, 1 mM EDTA, 10 mM Tris-HCl pH 8.0), and TE (10 mM Tris-HCl, 1 mM EDTA). DNA was then eluted with elution buffer (1% SDS, 0.1 M NaHCO<sub>3</sub>) and the crosslinking was reversed by adding 20 μl 5 M NaCl and incubated at 65°C overnight. After purification with PCR purification kit (Qiagen), the amount of DNA that was immunoprecipitated with a specific protein was quantified by real time PCR using the indicated primers (locations are shown in Figs) and SYBR master mix (Roche Life Sciences). The results are presented as % enrichment (% of the input DNA was immunoprecipitated with the indicated antibody). ## Methylation-specific PCR Genomic DNA was extracted from Hep3B cells using the DNeasy tissue extraction kit (Qiagen) following the manufacturer’s instructions. Subsequently, 0.5–1.0 μg of DNA from each sample was used for bisulphite conversion using the EpiTect fast bisulfite conversion kit (Qiagen). The converted DNA was then purified with the same kit from Qiagen. DNA methylation status of the SPRY2 gene was examined by PCR employing two sets of primers that match the same sites with one specific for methylated (M) and the other for unmethylated (U) sequences. The sequences of primers are listed in. Two rounds of PCR amplification were performed to detect the methylation status using FastStart PCR master kit (Roche Life sciences). The first round PCR amplification conditions used were one cycle of 95°C for 4 min followed by 25 cycles of 95°C for 30 s, 55°C for 30 s, 72°C for 60 s. The resultant PCR product was diluted 20 times and used as template for the second round PCR amplification, which employed the nested forward primers (n) and the same reverse primers as in the first round PCR. The PCR conditions used were one cycle of 95°C for 4 min followed by 35 cycles of 95°C for 30 s, 60°C for 30 s, 72°C for 60 s. The PCR products were loaded into agarose gels and the DNA methylation status was quantified by densitometry of the bands and normalized with beta-actin. ## Statistical analysis One-way ANOVA was employed for multiple-group comparisons using GraphPad 6 software. For two-group comparison, Student’s t test was performed. # Supporting information We thank Dr. Basabi Rana, University of Illinois, Chicago, for providing us with the Hep3B and HuH7 cells used in this study. We would also like to thank Dr. Nancy Zelenik-Le, Molecular Biology Program, Loyola University Chicago for advice on performing ChIP assays. We are also grateful to Dr. Guo-Qiang Chen, Shanghai Jiaotong University, China and Dr. Ramon Bartrons, University of Barcelona, Spain for kindly providing the plasmids pcDNA3-HIF1β and pGL2-Pfkfb3/-3566, respectively. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: KH XG TP. Data curation: KH XG PN TP. Formal analysis: KH XG TP. Funding acquisition: TP. Investigation: KH XG PN. Methodology: KH XG. Project administration: KH XG TP. Resources: TP. Supervision: TP. Validation: KH XG PN TP. Visualization: KH XG. Writing – original draft: KH XG TP. Writing – review & editing: KH XG PN TP.
# Introduction Staphylococcus aureus is a versatile gram-positive pathogen capable of causing a wide range of diseases, ranging from superficial abscesses to pneumonia, endocarditis, and sepsis. In the pre-antibiotic era, serious systemic staphylococcal infection was associated with 80% mortality. When penicillin was introduced in 1944, over 94% of S. aureus isolates were susceptible. However by 1950, 50% of S. aureus isolates were penicillin-resistant, further demonstrating the remarkable ability of S. aureus to rapidly adapt to antibiotic pressure. In 1960, outbreaks of virulent S. aureus that were resistant to penicillin occurred in many hospitals. These could be treated successfully with methicillin and other newly available penicillinase-stable penicillins. However, by 1961, two years after the introduction of methicillin, methicillin-resistant Staphylococcus aureus (MRSA) had emerged. Since then, a number of distinct MRSA strains have emerged and spread throughout the world. MRSA, in addition to an intrinsic resistance to virtually all β-lactams, has an ability to accumulate and develop resistance to other, unrelated antibiotics. Some variants of MRSA have developed resistance to glycopeptide antibiotics, including vancomycin, which is the sole antibiotic that can still be used to successfully treat most S. aureus. S. aureus resistance to methcillin has been attributed to the acquisition of the mecA gene, which encodes a penicillin-binding protein (PBP) that is resistance to β-lactam inactivation, namely PBP2a; the four native PBPs of S. aureus are sensitive to β-lactams. Despite the efficiency with which expression of PBP2a confers resistance to β-lactams, the PBP2a expression level does not correlate with high levels of methicillin resistance. Other factors are suggested to play essential roles in the phenotypic expression of methicillin resistance. Screening for methicillin-resistance factors has resulted in the identification of auxiliary genes including fem, fmtA, llm, sigma factor, pbpD, and vraSR. These genes reside outside the mecA determinants and have been shown to have direct or indirect roles in the biosynthesis or autolysis of peptidoglycan. The fem genes are involved in the biosynthesis of the peptide core of the peptidoglycan precursor subunit – and pbpD, which encodes PBP4 and is involved in the synthesis of highly cross-linked peptidoglycan. Genome-based studies of the S. aureus response to cell wall-specific antibiotics identified fmtA as being part of the core cell wall stimulon. The expression level of fmtA increases in the presence of cell wall inhibitors and when genes involved in biosynthesis of peptidoglycan are inactivated. Insertions in fmtA reduce MICs of methicillin, cefoxitin and imipenem 8 to 16 fold for different MRSA strains. This effect is more pronounced in the presence of Triton X-100. Insertions in fmtA also impair polysaccharide intercellular adhesion production and result in significantly reduced biofilm formation. Furthermore, the fmtA mutants exhibit reduced peptidoglycan cross-linking and reduced attachment of wall teichoic acids to cell wall. The fmtA gene product (FmtA) is capable of forming stable acyl-enzyme species with β-lactams, but the interaction is weak. Upregulation of the fmtA expression by perturbation of peptidoglycan biosynthesis suggests presence of a regulatory mechanism capable of coordinating fmtA expression with cell wall biosynthesis. Here, we investigate the factors involved in regulation of fmtA. In this study, we identified SarA as a transcription factor responsible for regulation of fmtA. SarA is a global regulator of S. aureus involved in the regulation of many virulence factors and was previously reported to be involved in methicillin resistance. We report the DNA-binding sites of SarA in the fmtA promoter. The binding specificity of SarA to fmtA promoter region was probed by mutating a key functional residue of SarA (Arg90). In addition, fmtA-lux operon reporter constructs were used to investigate the regulation of fmtA expression in vivo in response to antibiotic stress. The activation of fmtA transcription by SarA was confirmed by in vitro run-off transcription assays. Together, our results show that SarA binds directly to the fmtA promoter and plays a direct role in the regulation of fmtA expression. # Materials and Methods ## Growth Media and Chemicals Chemicals were purchased from Sigma (Oakville, Canada) or Thermo-Fisher (Whitby, Canada), unless otherwise stated. Chromatography media and columns were purchased from GE Healthcare (Quebec, Canada). The growth media was purchased from Fisher. Escherichia coli Nova Blue and BL21(DE3) strains, as well as cloning and expression plasmids were purchased from EMD4 Biosciences (New Jersey, USA). Restriction enzymes were obtained from New England Biolabs Canada (Pickering, Canada) or Thermo-Fisher. All primers, including biotinylated primers, were purchased from Sigma (Oakville, Canada). The \[γ-<sup>32</sup>P\] ATP (3000 Ci/mmol) was purchased from Perkin Elmer LAS Canada Inc. (Toronto, Canada) or GE Healthcare (Quebec, Canada). The ProteoExtract All-in-One Trypsin Digestion Kit was purchased from EMD4 Bioscience. ## Preparation of Cell Extracts S. aureus strain RN4220 (Cedarlane, Burlington, Canada) was grown to an optical density (OD) at 600 nm of approximately 0.9 in tryptic soy broth with or without oxacillin (1.2 µg/mL) and incubated for 1.5 h at 37°C. The cells were harvested at 4°C (11,000×g) and washed with cold TEG buffer (25 mM Tris-HCl, pH 8.0, and 5 mM EGTA) and re-suspended in TEG buffer. After two freeze-thaw cycles, cells were lysed with lysostaphin (0.3 mg/mL) followed by sonication. Lysates were centrifuged at 4°C (21,000×g) and 20% glycerol was added to the supernatants. The supernatants were dialyzed overnight in 1 L dialysis buffer (10 mM Tris-HCl pH 7.5, 1 mM EDTA, 1 mM DTT, 50 mM NaCl, and 20% (v/v) glycerol). ## Electro-mobility Shift Assay (EMSA) The predicted promoter region of fmtA (PfmtA) lying between the open reading frames (orf) denoted SAV1056 and fmtA, encompassing 540 bp, was divided into three DNA fragments of 270 bp, each designated seq1, seq2, or seq3. The seq1, seq2, and seq3 fragments were amplified from the S. aureus Mu50 genome (Cedarlane) using primers as follows: seq1 (−342 to −70; the numbering of the upstream and downstream elements in the fmtA promoter is based on a putative transcription starting point) was amplified using Dir<sub>seq1</sub> and Rev<sub>seq1</sub>, seq2 (−70 to +199) with Dir<sub>seq2</sub> and Rev<sub>seq2</sub>, and seq3 (−207 to +63) with Dir<sub>seq3</sub> and Rev<sub>seq3</sub>. The PCR amplified fragments were gel-purified, 5′-end labeled with \[γ-<sup>32</sup>P\]ATP (3000 Ci/mmol) using T4 polynucleotide kinase, and purified by passage through ProbeQuant G-50 columns (GE Healthcare). Binding reactions were performed by mixing oxacillin-induced or uninduced cell extracts with 2 ng 5′-\[<sup>32</sup>P\]-labeled double-stranded DNA fragments in the presence of 1 µg poly (dI-dC) and 200 ng sheared herring sperm DNA for 30 min in 10 mM Tris-HCl (pH 7.5), 50 mM KCl, 1 mM dithiothreitol (DTT), 5 mM MgCl<sub>2</sub>, and 2.5% (v/v) glycerol. The reaction mixtures were loaded into 6% native polyacrylamide gels. Following electrophoresis, gels were dried and imaged using an electronic radiography instant imager (Packard). Quantitative analysis of the bands was performed using ImageJ (v 1.3, NIH). Dissociation constants were determined as the protein concentration that resulted in 50% bound DNA. The results were obtained from three independent binding experiments. To ascertain which region of fmtA promoter is important for binding to the SarA transcription factor, two additional DNA fragments were generated from the fmtA promoter region. These sequences, referred here in as seqA (−207 to −70) and seqB (−70 to +63), were amplified by PCR using specific primers pairs seqADir and seqARev for seqA, and seqBDir and seqBRev for seqB. Target DNA (2 ng) was mixed with purified SarA at concentrations varying from 1 nM to 1.8 µM in binding reactions for EMSA as described above for the cell extracts. ## Screening Protocol for Isolation of Regulatory Proteins The PfmtA derived sequence seq3 (−207 to +63) was amplified by using 5′ biotinylated primers and their respective reverse primers. An aliquot of 200 µg streptavidin-coated magnetic beads (10 µg/µL) (Dynal® Biotech) was washed three times with 20 µL washing buffer (10 mM Tris-HCl buffer pH 7.5, supplemented with 1 mM EDTA and 2 M NaCl) for 15 min at room temperature. A 40-µL aliquot of the seq3 target DNA at 25 ng/µL was incubated with the beads for 15 min. The supernatant was removed and the beads were washed twice with the washing buffer. The process of loading target DNA on the beads was repeated two more times, with a washing step after each loading step. This process ensured 90% loading of the beads with the target DNA. Prior to incubation with DNA-bound beads, oxacillin- induced cell extracts were incubated with 200 µg streptavidin-coated beads in binding buffer (10 mM Tris buffer, pH 7.5, supplemented with 150 mM KCl, 0.1 mM EDTA, and 0.1 mM DTT) on ice for 30 min. The pre-treated cell extracts (40 µL) were then incubated with the DNA-loaded beads for 30 min at room temperature in the presence of 100 ng/µL herring sperm DNA. Unbound protein was removed from the beads by three washes with 100 µL binding buffer. The beads were then resuspended in 25 µL water and subjected to trypsin digestion (EMD4 Biosciences). The samples were analyzed by liquid chromatography-tandem mass spectrometry (LC-MS-MS) at the Advanced Protein Technology Centre of Hospital for Sick Children (Toronto, Canada). Alternatively, the beads were resuspended in 25 µL water, mixed with 0.5 µL 10% sodium dodecyl sulfate (SDS) and boiled for 5 min. The supernatant was loaded into SDS-polyacrylamide gels. ## Cloning of sarA The 375 bp sarA coding region was amplified from S. aureus strain MRSA252 by PCR using primers SarADir and SarARev which incorporate NdeI and HindIII restriction enzyme sites at the 5′-end and 3′-end respectively. The amplicon was digested with NdeI and HindIII and cloned into vector pET22b digested with MscI and HindIII. The resulting plasmid was digested with NdeI (to remove the pelB leader sequence between the NdeI site of pET22b and the sarA gene), re-ligated, and transformed into E. coli NovaBlue cells. The pET22b::sarA construct was confirmed by sequence analysis, and transformed into E. coli BL21(DE3) cells. ## Site-directed Mutagenesis of Arg90 to Ala in SarA The Arg90 residue of SarA was mutated to Ala using the QuickChange™ Site- Directed Mutagenesis protocol (Agilent, Mississauga, Canada). The recombinant DNA construct pET22b::sarAR90A was generated using pET22b::sarA as template, Pfu Turbo™ DNA polymerase (Agilent) and two mutagenic primers: SarAMDir and SarAMRev. The PCR product was treated with restriction endonuclease DpnI and the resulting mixture was used to transform E. coli NovaBlue cells. Successful mutation of arginine to alanine was verified by DNA sequencing and the pET22b::sarAR90A plasmid was used to transform E. coli BL21(DE3) cells. ## Purification of SarA and SarAR90A Mutant Proteins All steps of purification were carried out at 4°C. Seed cultures in BL21(DE3) cells were grown overnight in Luria-Bertani media (Thermo-Fisher). A 1% starter culture was used to inoculate 1 L Terrific Broth (TB), supplemented with 100 µg/mL ampicillin. The culture was grown at 37°C with shaking until the OD<sub>600 nm</sub> reached 0.7. SarA expression was initiated by adding 1 mM isopropyl-1-thio-β-D-galactopyranoside (IPTG; Rose Scientific, Edmonton, Canada). The induced cells were harvested and resuspended in 50 mM sodium phosphate buffer (pH 6.8). The protein was liberated by sonication and cellular debris removed by centrifugation at 20,000×g for 60 min. The clarified supernatant was loaded onto a SP-Sepharose FF column and the protein was eluted with a linear gradient of 1.0 M NaCl in 50 mM sodium phosphate buffer (pH 6.8). The fractions containing protein were concentrated and the buffer was changed to 20 mM Tris-HCl, pH 7.4, 1 mM EDTA, and 1 mM DTT using a HiPrep desalting column (26/10, GE Healthcare). The protein mixture solution was then loaded to a Heparin Sepharose 6FF column (26/10, GE Healthcare). The protein was eluted with a linear gradient of 1.5 M NaCl. The fractions containing protein were concentrated and loaded onto a Superdex<sup>TM</sup>75 (10/300 GL) column (GE Healthcare) equilibrated with 50 mM sodium phosphate buffer (pH7.0). Following elution by the same buffer, the protein-containing fractions were analyzed by 15% SDS-polyacrylamide gel electrophoresis (PAGE). The molecular mass of the purified SarA and SarAR90A mutant were confirmed by electrospray ionization mass spectrometry in the Advanced Protein Technology Centre of Hospital for Sick Children (Toronto, Canada). The oligomerization state of SarA was determined by size exclusion chromatrography using a gel filtration LMW calibration kit from GE Healthcare and Superdex<sup>TM</sup>75 (10/300) column (GE Healthcare). The column was equilibrated and run with 50 mM sodium phosphate buffer (pH 7.0) at a flow rate of 0.4 mL/min. ## DNase I Footprinting Assay The seqA and seqB promoter sequences were amplified by PCR from the S. aureus Mu50 genome using specific primers. The primer of interest (Dir or Rev) was 5′–end labeled with \[γ-<sup>32</sup>P\] ATP (3000 Ci/mmol). Binding reactions were performed as described above for the EMSA, but with 8 ng DNA. Binding reactions were incubated for 30 min at 37°C and then subjected to DNaseI (0.024 units) for 2 min at 25°C after supplementing binding buffer with 10 mM CaCl<sub>2</sub> and 5 mM MgCl<sub>2</sub>. The reaction was terminated by adding stop solution (200 mM NaCl, 1% (w/v) SDS, 20 mM EDTA, and 250 µg/mL tRNA). The digested DNA samples were extracted with phenol:chloroform and ethanol precipitated. They were then resuspended in formamide containing loading dye, denatured at 95°C and loaded onto 8% sequencing gel. The control A, T, C, and G were prepared by using Therminator DNA polymerase (New England Biolabs Canada) and a set of acyclonuleotides (acyATP, acyCTP, acyGTP, and acyTTP; New England Biolabs Canada). Four separate reactions (one for each acyNTP terminator) generated DNA sequencing termination ladders. Reactions contained template DNA, 50 nM \[γ-<sup>32</sup>P\] ATP labeled sequencing primer, 50 nM dNTP, 20 mM Tris–HCl, pH 8.8 at 25°C, 10 mM KCl, 2 mM MgSO<sub>4</sub>, 10 mM (NH<sub>4</sub>)<sub>2</sub>SO<sub>4</sub>, 0.1% Triton X-100, and 0.1 U/µL of E. coli RNA polymerase haloenzyme (Epicentre Biotechnologies, Madison, Wisconsin). ## In vitro Run-off Transcription Assay Single round transcription by E. coli RNA polymerase holoenzyme (RNAP; Epicentre Biotechnologies, Madision, Wisconsin) was carried out using a 580-bp fragment of the fmtA promoter region (P<sub>fmtA</sub><sub>,</sub> nucleotide position −342 to +239 of fmtA promoter region) as template, in the absence or presence of SarA. The P<sub>fmtA</sub> fragment was amplified using pfu Turbo DNA polymerase and primers Dir-P<sub>fmtA</sub>, (5′-GAGAACCAATGCTAGAAGGATCAA-3′) and Rev-P<sub>fmtA</sub> (5′-TCGATGAAAAATTAACGCTATAGAAA-3′). The in vitro transcription reactions were performed as described previously with some modifications. The DNA template (5 nM) was incubated with SarA (0.1 and 0.6 µM) in transcription buffer (40 mM Tris, pH 7.5, 150 mM KCl, 10 mM MgCl<sub>2</sub>, 2 mM DTT, 100 µg/mL BSA, and 0.05% Triton X-100) for 20 min at 30°C. One unit of E. coli RNAP (0.5 µg) was added to each reaction mixture and the reactions were incubated for another 20 min at 30°C. A control reaction containing only RNAP was prepared similarly. Transcription was initiated by adding 3 µL NTP mix containing 1 mM ATP, 1 mM CTP, 1 mM GTP, 100 µM UTP, 2 µCi of \[α- <sup>32</sup>P\]-UTP, 600 µg/µL heparin, and 40 U/µL Murine RNase inhibitor. The reactions were placed at 30°C for 20 min. The reactions were terminated by addition of 5 µL transcription stop buffer (20 mM EDTA, 1% SDS, 1 mg/ml bromophenol blue, 1 mg/ml xylene cyanol, and 90% formamide). The samples were heated at 90°C for 5 min and immediately loaded onto an 8% polyacrylamide gel containing 7 M urea. The φX174 DNA/HinfI dephosphorylated DNA fragments (Promega), labeled with \[γ-<sup>32</sup>P\] ATP (3000 Ci/mmol), were used as a molecular weight marker. The gels were dried and exposed on phosphor screen and visualized using a Typhoon Trio<sup>+</sup> Variable Mode Imager (GE Healthcare). A transcription start point in fmtA promoter was identified the T residue located at the 156<sup>th</sup> nucleotides upstream of the fmtA start codon. A −10 box sequence of 5′-TATAAT-3′ sequence is present relative to that start site, which is identical to the −10 sequence recognized by the E. coli σ<sup>70</sup> RNAP. To investigate the role of the proposed transcription initiation elements, we designed two mutations in the fmtA promoter sequence and assessed their role in transcription by transcription run-offs experiments. In the mutant P<sub>fmtA</sub><sub>T1G</sub>, we replaced the proposed transcription start site T (+1) with a G residue. In the P<sub>fmtA</sub><sub>T14G</sub> mutant promoter, we substituted the T at −14 with a G residue because this T/A pair at the beginning of the - 10 consensus sequence is critical in interactions with RNAP. To accomplish the mutagenesis, we cloned P<sub>fmtA</sub> sequence into the pSTBlue vector (Invivogen). The mutagenic primers are presented in. ## Construction of luxABCDE Fusion Strains with Derivatives of fmtA Promoter Region The three PfmtA-derived sequences (seq1, seq2 and seq3) were amplified by PCR from fmtA operon using specific primer pairs designed to introduce EcoRI and BamHI restriction sites at the 5′ and 3′ ends, respectively. The predicted full length fmtA control region encompassing nucleotides −342 to +199 (referred to as seq1–2) was also amplified using primers DirP<sub>seq1</sub> and RevP<sub>seq2</sub>. All four DNA fragments, P<sub>seq1,</sub> P<sub>seq2,</sub> P<sub>seq3,</sub> and P<sub>seq1–2</sub>, were PCR amplified and digested using restriction enzymes EcoRI and BamHI. The DNA fragments were ligated to similarly digested vector pXEN1. pXEN1 harbors the luxABCDE operon from Photorhabdus luminescens that has been modified for lux expression in Gram positive bacteria. The ligation mix was transformed into E. coli DH5α and transformants were selected on medium containing ampicillin (100 µg/mL). Putative clones obtained were confirmed to harbor the inserts by sequence analysis using a primer derived from luxA and the forward primer used to amplify the insert fragment. The pXEN1 plasmids carrying different inserts are referred to as lux fusion plasmids. The pXEN1 plasmid and lux fusion plasmids were introduced into the restriction-deficient S. aureus strain RN4220 by electroporation (26). Transformants were plated on tryptic soy broth (TSB)-agar plates supplemented with 5 µg/mL chloramphenicol (selectable marker in pXEN1). Positive colonies were confirmed by sequencing of the plasmids using the primers described above. The S. aureus strain RN4220 carrying pXEN1 is referred to as RN(::lux), and RN4220 strains carrying the pXEN1 fusion plasmids, constructed herein, are referred to as lux fusion strains RN(P<sub>seq1</sub>::lux), RN (P<sub>seq2</sub>::lux), RN(P<sub>seq3</sub>::lux), and RN(P<sub>seq1–2</sub>::lux). In S. aureus PC1839 (parent strain S. aureus RN4220), the sarA gene has been knocked out and replaced with a kanamycin resistance cassette. PC1839 was transformed with the lux promoter constructs P<sub>seq1,</sub> P<sub>seq2,</sub> P<sub>seq3,</sub> P<sub>seq1–2</sub>, and pXEN1 isolated from strain RN4220. The resultant colonies were resistant to kanamycin and chloramphenicol. The S. aureus strain PC1839 carrying the empty vector pXEN1 is referred to as PC(::lux), and PC1839 strains carrying the pXEN1 fusion plasmids are referred to as lux fusion strains PC(P<sub>seq1</sub>::lux), PC(P<sub>seq2</sub>::lux), PC(P<sub>seq3</sub>::lux), and PC (P<sub>seq1–2</sub>::lux). The effect of pXEN1 and lux fusion plasmids on the growth of S. aureus RN4220 and PC1839 fusion strains was investigated by monitoring the growth profiles. Briefly, lux fusion strains RN(::lux), PC(::lux), RN4220, and PC1839 were inoculated into TSB and grown overnight at 37°C with 5 µg/mL chloramphenicol, except for RN4220 and PC1839. Next, 1% culture was inoculated in fresh TSB supplemented with 5 µg/mL chloramphenicol for the lux fusion strains except for RN4220 and PC1839. Then the OD(<sub>600 nm</sub>) was measured at 1 h intervals for the next 7 h. ## Measuring Bioluminescence from S. aureus Strains After overnight growth, RN4220, PC1839, RN(::lux), and PC(::lux), were diluted into fresh TSB and grown at 37°C with shaking at 200 rpm. The strains were grown to an OD<sub>600 nm</sub> of approximately 0.3 and induced with oxacillin at 10 µg/mL or 100 µg/mL for 1 h. The optical densities were measured at 600 nm for all the samples. The cultures with higher OD(<sub>600 nm</sub>) were diluted such that the cell density in each culture was the same. A 300-µL aliquot from each sample (in triplicates) was transferred to opaque 96-well optiplates and analyzed in a HT-Analyst (Molecular Devices, California). Bioluminescence was measured immediately after dispensing the samples into the plates over a period of 10 min. The data points collected over 10 min were averaged for each strain at each oxacillin concentration, and the standard deviations were determined from three independent measurements. The promoter activity was plotted as luminescence over time using the mean values from the readings taken over a period of 10 min calculated by subtracting values from blank medium controls. These experiments were repeated three times. ## Viability Testing of the Induced Versus Uninduced Cultures Cultures of RN4220, PC1839, RN(::lux), and PC(::lux) were inoculated at 1∶100 with an overnight culture into TSB with chloramphenicol (no antibiotic was added to RN4220 and PC1839). When the OD<sub>600 nm</sub> of the cultures reached approximately 0.3, oxacillin was added to a final concentration of 10 or 100 µg/mL, and the cultures were grown for 1 more hour. The OD of each culture was normalized to the sample having the lowest OD. A 20-µL aliquot from each culture was diluted 5000-fold with TSB to dilute the antibiotic, and a 50-µL aliquot of the dilution was plated on TSB-agar plates without antibiotic and incubated at 37°C. These experiments were repeated three times. ## RNA Extraction and c-DNA Synthesis Overnight-grown wild-type S. aureus RN4220 and PC1839 were diluted (200-fold) separately in 30 mL tryptic soy broth (TSB) in 125-mL Erlenmeyer flask. Cultures were incubated at 37°C with shaking at 200 rpm and growth was measured at regular intervals until the optical density (OD<sub>600</sub> nm) reached to early-exponential phase (∼0.4). Cultures were then distributed in 2 flasks (10 mL each), where one culture flask was used as control, and other was treated with 10 µg/mL of oxacillin for 15- and 60-min. Control cultures were incubated along with the treated culture for 15-min and 60-min. Aliquots of 2 mL cultures from each flask were taken, and mixed with 4 mL of bacterial RNA Protect solution (Qiagen, Valencia, CA). The mixtures were further vortex briefly for 5 s and incubated for 5 min at room temperature, followed with centrifugation at 5000×g for 20 min in a swinging-bucket rotor centrifuge to collect the cells. Cells were washed with fresh TSB and used for isolation of RNA. Cells were further suspended in 200 µL of TE (30 mM Tris-HCl; 1 mM EDTA, pH 8.0) buffer containing lysostaphin (50 µg/mL), mixed by vortexing for 10 s, and incubated at room temperature for 10 min on a shaker–incubator. Subsequently, 10 µl Proteinase K (100 µg/mL, solution) was added in the solution and total RNA was extracted and purified using an RNeasy minikit (Qiagen) following the manufacturer recommendation for bacterial RNA isolation. For RNA isolation of both antibiotic treated and untreated cultures, two independent bacterial cultures were prepared. cDNA was synthesized using a high capacity RNA-to-cDNA kit (Applied Biosystems, Foster City, CA). RNA concentrations were determined by absorbance readings at 260 and 280 nm, using a Nanodrop ND-1000 UV spectrophotometer (Nanodrop Technologies, Wilmington, DE). ## Quantitative Real-time PCR (qRT-PCR) qRT-PCR was used to characterize the transcript levels in the cells in response to oxacillin treatment. qRT-PCR was performed with Rotor-Gene Real-Time PCR Cyclers (Qiagen) using SYBR green technology. RT-PCRs were performed in a 20 µL reaction volume containing 1 µL of template DNA (25 ng), 1 µL of gene specific primers (10 µM, Table1), 10.0 µL of Power SYBR green PCR Master Mix (Applied Biosystems, Foster City, CA) and 8.0 µL of H<sub>2</sub>O. The following PCR conditions were used: 50°C for 2 min, 95°C for 10 min, 40 cycle of 95°C for 15 s, 60°C for 30 s, 72°C for 30 s and one final extension step of 72°C for 10 min. The transcript levels of the fmtA and sarA genes were normalized using the 16S rRNA transcript level of S. aureus RN4220 as an internal control. Each gene assay was performed in triplicate. The data analysis was carried out with Rotor-Gene Q software (Qiagen). ## Determination of MIC Minimum inhibitory concentration (MIC) assays were performed as follows. Oxacillin dilutions were made using sterilized water and then aliquotted (1.5 µL) into 96-well plates. An overnight culture of RN4220 and sarA mutant were diluted 1/1000 into fresh TSB medium to give ca. 5×10<sup>5</sup> CFU mL<sup>−1</sup>, and 150 µL aliquots were added to each well of the 96-well plate. Plates were incubated at 37°C for ∼18 h. MIC determinations were carried out on two different times in duplicate. # Results ## Screening for the Transcription Factor(s) of fmtA We divided the predicted 540 bp fmtA promoter region into three DNA fragments of approximately 270 bp each, referred to as seq1, seq2, and seq3, with seq3 overlapping seq1 and seq2 at the 5′ and 3′ ends respectively. When radioactively labeled seq1, seq2, and seq3 were incubated with oxacillin- induced and uninduced cell extracts from S. aureus RN4220, electrophoretic mobility shifts were observed. Presence of oxacillin has a clear effect on the electrophoretic mobility profile of DNA considering the multitude of non- specific interactions that these DNA fragments could be subject of. Further, introduction of unlabeled DNA targets eliminated the mobility shifts of the labeled target DNA fragments. These observations suggest that the promoter region of fmtA is the target of regulation by transcription factors. We developed a screening protocol to isolate the transcription factor(s) of fmtA, using biotinylated seq3 (−207 to +63) immobilized on streptavidin- coated magnetic beads. These beads when incubated with oxacillin-induced cell extracts pulled down several proteins that were identified by analyzing LC-MS-MS data using the MASCOT program and a probability based Mowse score of \>45. The analyses showed the presence of staphylococcal accessory regulator A (SarA), pyruvate carboxylase, biotin carboxyl carrier protein of acetyl-CoA carboxylase, DNA-direct RNA polymerase alpha chain, DNA polymerase I, translation elongation factor Tu, translation elongation factor G, elongation factor TS, chaperone protein dnaK, chaperonin GroEL, cell division proteins FtsZ, DNA binding protein II, putative DNA-binding protein, and 30S ribosomal protein S5. To confirm that SarA is pulled down by this protocol we incubated the oxacillin- induced cell extracts with seq3 and boiled the magnetic beads in SDS-loading buffer after they were washed twice with washing buffer. Then, we loaded the supernatant to a 15% SDS-PAGE. The SDS-PAGE was stained by commassie blue, and the gel in the region of 14 kD was cut and subjected to trypsin digestion and LC-MS-MS. This process resulted in only one hit, SarA. Seven SarA peptides were identified, which covered 69% of the protein sequence (79/113 amino acids). We then focused on the SarA protein, a 14-kD global regulatory protein involved in regulation of virulence factors, to investigate the transcriptional regulation mechanism of fmtA. ## Purification of SarA and Characterization of its Oligomerization State SarA belongs to the family of transcription factors called the “wing-helix” proteins. Full-length SarA gene was cloned, expressed, and purified to homogeneity. The identity of the protein was confirmed by mass spectrometry. The observed molecular mass of SarA was 14,635 D (theoretical mass, 14,586 D). The oligomeric state of SarA was investigated using gel filtration chromatography. Based on the elution times of four standards, the molecular mass of SarA was estimated to be 30 kD, which shows that SarA is dimer in our solution as shown previously, hence we are working with a functional protein. ## SarA Binding to PfmtA To elucidate whether purified SarA binds directly to the PfmtA promoter, we performed EMSA experiments with seq1, seq2, seq3, seqA and seqB. The dissociation constant of each DNA sequence was determined as the concentration of SarA that resulted in 50% DNA bound. We estimated the dissociation constants of SarA with respect to seq1, seq2, seq3, seqA, and seqB to be 63±9 nM, 54±12 nM, 66±7 nM, 147±18 nM, and 139±17 nM, respectively. As a control, we used a 45-bp sequence derived from the promoter region of agr that harbors the SarA binding consensus sequence (underlined), Pagr: 5′-GTAAATTTT<u>TTTATGTTAAAATATTAAATACAAAT</u>TACATTTAAC-3′ (47). The SarA binding affinity for the Pagr sequence was about 400 nM. Supershifts were observed in the EMSA experiments with seq1, seq2, seq3, seqA, seqB, and Pagr at SarA concentrations higher than 300 nM. Interestingly, in the case of seq3, which covers the 3′-end of seq1 and 5′-end of seq2, the supershift starts at 150 nM SarA. This observation indicates that either there is more than one SarA binding site in these sequences, which are saturated at higher protein concentrations, or there are different oligomers of SarA binding to the fragments; SarA is known to form dimers and dimers of dimers in target promoters and is proposed to oligomerize at its binding sites (45). ## Identification of SarA Binding Sites on PfmtA by DNase I Footprinting Two fragments of PfmtA spanning the regions −207 to −70 (seqA) and −70 to +63 (seqB) were used to investigate SarA binding to PfmtA. The SarA protection sites on the bottom strand of each sequence are shown in. SarA provided a protection of the DNA at concentrations higher than 200 nM. A similar DNase I protection profile was observed on the top strand of each DNA sequence (data not shown). Alignment of these sequences against the proposed 26 bp SarA consensus binding site, 5′-ATTTGTATTTAATATTTATATAATTG-3′, revealed that the SarA-protected sites in PfmtA encompass the SarA consensus binding site. Interestingly, the SarA-protected sites in PfmtA also included the 7-bp DNA sequence 5′-ATTTTAT-3′, recognized by the winged-helix proteins, and also suggested to be recognized by SarA. Additional three DNA fragments (obtained through primer annealing) were derived from the fmtA promoter region to investigate the putative SarA binding sites, referred to as seqA1, seqA2 and seqB1. The seqA1 fragment harbors the 26-bp SarA consensus binding sequence identified in seqA, whereby the nucleotides −186 to −178 were omitted. The seqA2 fragment harbors two 7-bp SarA binding sequences and seqB1 is downstream of the putative transcription starting point. The SarA binding affinities determined for seqA1 fragment is ≥1.8 µM (no binding saturation was observed up to 3.6 µM) and 550 nM for seqA2 fragment. No SarA binding was observed to seqB1 at concentrations as high as 1.8 µM SarA. The lower binding affinity for seqA1 in comparison to Pagr could be due to the removal of the four nucleotides however binding of SarA to seqA1 suggests that the 26 bp SarA consensus binding site identified in seqA1 recruits SarA and these nucleotides may be involved in the SarA binding. ## Analysis of the DNA-binding Properties of SarAR90A Mutant Arginine at position 90 is essential for the SarA regulatory activity in vivo. It is located within the winged region of SarA and is part of the conserved basic region of SarA, the motif DER (45). We mutated this residue to alanine, purified the mutant protein, and confirmed the substitution by mass spectrometry. We then analyzed the binding affinity of the SarAR90A mutant protein to seq1, seq2 and seq3 by EMSA. The mutant protein failed to bind to these three DNA sequences. ## SarA Promotes in vitro Transcription of fmtA We used E. coli RNA polymerase holoenzyme (RNAP) in these experiments. Several lines of evidence lead us to conclude that E. coli RNAP is a good substitute for S. aureus RNAP in the case of fmtA. The E. coli RNAP is a complex of the RNAP core enzyme and the β<sup>70</sup> factor protein. The DNA sequence of the putative −10 box in the fmtA promoter is identical to the consensus −10 box recognized by E. coli β<sup>70</sup>, suggesting that the fmtA promoter may recruit a β<sup>70</sup>-like factor for transcription. The β<sup>70</sup> homolog in S. aureus is SigA factor, in turn suggesting that regulation of fmtA could be SigA-dependent. Incidentally, SarA regulates its own promoter in a SigA-dependent mechanism. Our hypothesis that E. coli RNAP is a good substitute for that of S. aureus was confirmed by in vitro run-off transcription experiments with PftmA and PfmtA mutated at the putative +1 and −14 sites. We observed that while E. coli RNAP could poorly initiate transcription from P<sub>fmtA</sub>, there was a complete lack of transcription for the T(+1) to G and T(−14) to G PfmtA variants. Furthermore, the presence of SarA increased the production of the transcript from the PfmtA fragment in a concentration-dependent manner (data not shown). ## SarA Regulates Expression of fmtA–lux Fusion in vivo To investigate the role of the fmtA promoter in recruiting SarA in vivo, we fused PfmtA-derived sequences (P<sub>seq1</sub>, P<sub>seq2</sub>, P<sub>seq3</sub>, and P<sub>seq1–2</sub>) upstream of the lux operon in the reporter vector pXEN1. The lux fusion plasmids and the control vector pXEN1 were introduced into S. aureus RN4220. We hypothesized that SarA exerts direct regulation of fmtA via the SarA binding sites identified in our in vitro experiments. If this hypothesis were correct, we would expect that the disruption of SarA would reduce lux operon expression. To test this hypothesis, we introduced the lux fusion plasmids into a SarA- deficient strain, PC1839, and compared the luminescence of the RN4220 (sarA<sup>+</sup>) strains and PC1839 (sarA<sup>-</sup>) strains under oxacillin-induced and uninduced growth conditions. Strains harboring different fusion constructs showed comparable growth rates (data not shown). The luminescence data are shown in. In the absence of oxacillin, a low level of luminescence signal over the background level, set as the RN4220::lux- strain, was observed in any lux-fusion strain. Presence of 10 µg/ml oxacillin increased the lux operon expression especially in the case of the seq3::lux and seq1–2::lux fusions, where a 5.8 and 4.5-fold increase in luminescence was observed, respectively, in comparison to the lux expression in the absence of oxacillin. The same trend was observed in the presence of 100 µg/ml oxacillin. Overall, luminescence of RN(P<sub>seq1–2</sub>::lux) and RN(P<sub>seq3</sub>::lux) exhibited the highest increase in luminescence in the presence of oxacillin. In contrast, none of the sarA mutant strains \[PC(P<sub>seq1</sub>::lux), PC(P<sub>seq2</sub>::lux), PC(P<sub>seq3</sub>::lux) or PC (P<sub>seq1–2</sub>::lux)\] showed any significant luminescence, even at 100 µg/ml oxacillin. ## Deletion of sarA Enhances S. aureus Sensitivity to Oxacillin To understand whether the lack of functional SarA protein alters the sensitivity towards oxacillin, we tested the oxacillin MICs of wild type S. aureus RN4220 and sarA mutant. The sarA mutant showed modest, but significant decrease (2-fold) in MIC. The measured MICs of oxacillin for wild-type RN4220 and sarA mutant were 0.3 µg/mL and 0.15 µg/mL, respectively. ## Transcript Levels of sarA and fmtA in S. aureus RN4220 and PC1839 Our data revealed that sarA transcript level in S. aureus RN4220 did not change upon oxacillin treatment compared with corresponding untreated control culture. However, fmtA transcript level was increased 3-fold upon oxacillin treatment (3.5±0.7). This finding is in good agreement with the previously published data where no change in sarA transcript level, but increased transcript level of fmtA was observed. Increased expression of fmtA in response to several cell wall active antimicrobials was observed in several studies. On the other hand, in the sarA mutant strain, PC1839, the transcript level of fmtA did not change upon oxacillin treatment compared with corresponding untreated control culture. # Discussion Genome-based studies have shown that fmtA levels in S. aureus are upregulated in the presence of cell wall inhibitors and by the knock-out of the genes involved in peptidoglycan biosynthesis. A study by Kuroda et al. suggested that fmtA is under the regulation of the two component signal transduction system VraSR. However, the mechanism by which fmtA expression is regulated remains unknown. In our study DNA fragments derived from PfmtA region showed retardation in their electrophoretic mobilities when incubated with cell extracts isolated from S. aureus RN4220 with clear differences from the instance when bacteria were subjected to oxacillin. The electrophoretic mobility shifts were eliminated in the presence of a competing DNA fragment. Herein, we report a screening protocol designed to identify the transcription factor(s) involved in regulation of fmtA expression. A 269-bp fragment of the fmtA promoter region (seq3) was immobilized on streptavidin-coated beads and used to capture proteins from extracts of S. aureus cells treated with oxacillin to induce fmtA expression. Protein binding and washing steps were performed at physiological ionic strength and in the presence of nonspecific DNA to reduce nonspecific interactions. A relatively small number of non-DNA-binding proteins were identified in the screen, including pyruvate carboxylase and biotin carboxyl carrier protein of acetyl-CoA carboxylase which both have biotin as a cofactor and could be binding to the beads through streptavidin. The presence of RNA polymerase alpha chain, DNA polymerase I, translation elongation factor Tu, translation elongation factor G, elongation factor TS, and 30S ribosomal protein S5 could be due to the presence of the fmtA ribosomal binding site in seq3. The presence of DNA binding protein II, a protein that recognizes polyA DNA regions, could bind nonspecifically to the seq3 fragment which contains several polyA sites, and proteins such as chaperone protein DnaK, GroEL and FtsZ could bind nonspecifically to DNA or streptavidin. SarA was the only transcription factor identified by our screening protocol. DNase I footprinting demonstrates that SarA binds to specific sites in the fmtA promoter region. Specificity of SarA binding to PfmtA was probed by mutation of one of the conserved residues located in the winged region of SarA, Arg-90. The mutant SarAR90A protein, which is unable to regulate the SarA target promoters in vivo, failed to bind to the fmtA promoter region. The role of SarA in regulation of fmtA was corroborated by in vivo studies. Fusion of the lux operon to the various regions of the fmtA promoter indicated that the fmtA promoter region harbors transcriptional regulatory elements and that SarA is involved in the regulation of fmtA expression. These data are corroborated by the qRT-PCR data, whereby the fmtA expression level increases by 3.5-fold in the presence of oxacillin, while sarA level remains unchanged. In the sarA mutant, fmtA level are not affected by the presence of oxacillin. Interestingly, the in vivo data suggest that SarA is involved in basal expression of fmtA as well as upregulation of fmtA in the presence of cell wall stress. Further, in vitro run-off studies suggest that transcription of fmtA may require the S. aureus primary sigma factor, SigA. This is in agreement with a previous report that shows that SarA regulates promoters that are SigA-dependent. Additional evidence of the involvement of SigA in the transcriptional regulation of fmtA is provided by mutagenesis studies in which mutation of the −14 position and the transcription starting site of PfmtA resulted in drastic reduction of fmtA transcription. Sequence alignment of PfmtA-seq3 against the SarA consensus binding site revealed two regions that contain this sequence, centered around −168 and −53 nucleotides. In addition, we identified three sites that harbor the 7-bp consensus binding sequence for the winged-helix-turn-helix DNA-binding proteins, which SarA is a member. The 7-bp consensus binding sequence appears in pairs in PfmtA; two pairs are within the 26-bp SarA binding consensus sequences and the third pair is centered on the −112 nucleotide. The EMSA studies with DNA fragments derived from PfmtA, harboring the 26-bp (seqA1) or the 7-bp SarA binding sequences (seqA2), showed that SarA is recruited to these sites. The DNA fragment that did not contain any of these sequences (seqB1) failed to recruit SarA, suggesting that SarA recognize specific sequences in PfmtA, known to interact with SarA, and that it may utilize both these sites to bind to the fmtA promoter. SarA is a global transcriptional regulatory protein linked to the regulation of numerous virulence factors. It is a 124-residue protein that forms dimers in solution and binds as a dimer to the target promoters. SarA also belongs to the winged helix-turn-helix family of proteins. Several S. aureus proteins share high sequence similarities with SarA and are grouped into a SarA family of proteins. Based on the structural information on the SarA family of proteins, four regulatory mechanisms are proposed for SarA : 1) bending of the target DNA to facilitate contact with the regulatory proteins; 2) the formation of three SarA dimers which hold the DNA in a closed configuration not amenable to transcription; 3) the formation of a heterodimer between compatible family members that may interfere with the function of the homodimer and 4) competitive displacement of one homolog by another. In the case of the fmtA promoter, it is possible that SarA may follow the first regulatory mechanism whereby SarA could bind as a dimer at the three identified sites. The SarA family of proteins has been recently linked to S. aureus autolysis, biofilm formation, and resistance to cell wall inhibitors. A study by McAleese et al. showed that a clinical isolate of S. aureus with intermediate resistance levels to vancomycin (a VISA strain) exhibited higher sarA expression levels than the vancomycin susceptible parent strain when grown in an antibiotic-free medium. The expression levels of fmtA were also increased in the VISA strain. Interestingly, the sarA levels in susceptible or methicillin- resistant strains are not reported to increase in the presence of cell wall inhibitors; our data show the same phenomenon. By contrast, fmtA expression increases in the presence of cell wall inhibitors despite the strain background. The question that arises is, how does SarA regulate the expression of fmtA when its protein levels are not altered in response to cell wall stress? It has been hypothesized that the pleiotropic regulatory capabilities of SarA could be due to posttranslational modifications. Recently, it was reported that SarA is a phosphorylation target by Stk1 Ser/Thr kinase. Perhaps regulation of fmtA in response to stress is mediated by phosphorylated SarA or other posttranslational modifications. However, we cannot exclude that another transcription factor may be involved in the regulation of fmtA during antibiotic-induced cell wall stress. The fmtA gene is reported to be part of the VraSR regulon. However, VraR was not isolated from our protocol. There could be two reasons for this: i) VraR is not involved in regulation of fmtA or ii) our screening protocol failed to identify VraR due to low binding affinity of VraR for the fmtA promoter. We have previously shown that VraR binding affinity to its own promoter is 1 µM, which is one order of magnitude higher than that of SarA for PfmtA. It is possible that under our screening conditions the occupancy of PfmtA by VraR is low. A recent study by Sengupta et al. used chromatin immunoprecipitation techniques to investigate the promoters under the direct control of VraR. Interestingly, VraR was not identified as the transcription factor for fmtA, but the study showed that other members of the VraSR regulon, such as pbp2, murZ, and sgtB, are under the direct regulation by VraR. In conclusion, this study links the global regulator SarA with a penicillin- binding protein that is also involved in autolysis and biofilm formation. Our findings further extend the multiple functions of SarA and establish a link between processes involved with S. aureus pathogenicity, i.e., virulence factor expression and the cell wall stress response. We are in debt to Professor Simon J. Foster at University of Sheffield (United Kingdom) for providing the PC1839 strain. The authors declare that this manuscript or a related manuscript has not been previously published and is not currently being considered for publication elsewhere. [^1]: The authors have declared that no competing interests exist. [^2]: Analyzed the data: YZ VV AS. Wrote the paper: DGK VV AB. Conceived the project: DGK. Performed the screening protocol, the EMSA experiments, DNase I footprinting, cloning and purification of SarA, construction of the SarA mutant: YZ. Designed the in vivo studies with the Lux reporter vector: VV. Designed the in vitro run-off experiments: AB. Performed the qRT-PCR, MIC experiments: AS. Performed the EMSA experiments: MF.
# Introduction The SARS-COV-2 pandemic, which has gripped the world over the last three years, has resulted in more than 530 million reported infections and 6.3 million deaths worldwide so far \[<https://covid19.who.int/>\]. The pandemic also resulted in unimaginable suffering to individuals, families, communities and countries across the globe. At its peak, the pandemic stressed healthcare systems in different parts of the globe to their limit, disrupted supply chains, destroyed businesses, resulted in massive unemployment and poverty and a never-seen-before upward re-distribution of wealth; which will collectively continue to impact life on earth for possibly generations to come. Although it is arguable whether the pandemic was foreseen and/or could have been avoided or even better managed; still, the way in which it was handled speaks of utter incompetence and indisputable lack of preparedness at all levels from governments and healthcare authorities all the way to the scientific community. Therefore, the world needs to learn its lesson, not only in terms of how to deal with future epidemics and pandemics but also how to study and understand them and how to use cutting edge technologies in such endeavors. One of the puzzling questions about the COVID-19 pandemic that still lingers is how and why some, seemingly healthy (low risk) individuals, succumbed to the disease while others, possibly of poorer health, recovered and survived. Indeed, most people would agree that the worst of this pandemic is, more or less, behind us, but efforts to uncover infection and disease correlates that may have contributed to its outcomes remain timely and needed. In this context, there is a real need to develop and test disparate data-integrating approaches and data- based models to understand the various aspects of COVID-19 and to easily and quickly enlist such models in combating future epidemics and pandemics. Polymerase chain reaction (PCR) testing of the nasopharyngeal swab for the presence of SARS-CoV-2 RNA continues to be the gold standard in identifying infected individuals. Based on global data input, almost 80% of SARS-COV-2 individuals end up with no symptoms to mild-to-moderate symptoms. Serological testing in the form of differential blood count along with inflammatory marker testing has proven partially successful in identifying patients at high risk of disease severity or death. The experience with COVID-19 has demonstrated that testing for serum IL-6, D-dimer, lactate dehydrogenase (LDH) and other analytes is helpful in identifying patients at risk of sever or fatal complications. That said, there is still a need for more specific predictive parameters of severe infection beyond serum ferritin, prothrombin time, and fibrin degradation products (FDP). Serum/plasma metabolomics profiling using liquid chromatography-mass spectrometry (LC-MS) has proven useful in identifying diagnostic, prognostic and therapeutic biomarkers in infectious diseases. Studies have shown that serum levels of citrate, malate and succinate increase in response to S. aureus and S. pneumonia infections. In a study on COVID-19 intensive care unit patients, the plasma metabolites kynurenine and arginine ratio was reported to be helpful in predicting COVID-19 disease irrespective of age, gender or hospital admission. However, the role of these findings in COVID-19 prognosis remains limited given that only ICU patients were assessed. In another study, the metabolites cytosine and tryptophan-nicotinamide were reported to be moderately sensitive in discriminating COVID-19 patients from healthy individuals. It is predictable that metabolomic changes resulting from SARS- CoV-2 infection could vary widely among patients owing to differences in patient health profiles, vis-à-vis comorbidities, medications, diet, lifestyle, etc. Accordingly, the search for profiles of metabolic biomarkers may provide higher sensitivity and specificity in assessing disease prognosis. In this study, we retrospectively recruited COVID-19 patients with no known comorbidities and divided them into three groups based on disease severity: asymptomatic, mild and severe. We performed LC-MS metabolomics profiling in serum samples of these patients and identified eight predictive biomarkers of COVID-19 disease severity. We then integrated patients’ laboratory findings and metabolomics profiles to generate a predictive model of disease severity. # Material & methods ## Sample collection and processing In this retrospective cohort study, blood samples were collected from donors who tested positive for COVID-19 and presented with no, mild or severe symptoms between March 20 until July 17, 2020. Patients were diagnosed with COVID-19 using a nasal swab PCR test and later divided into three groups (asymptomatic, mild, and severe) based on their clinical presentation. Each donor gave a 10 mL blood sample, one half of which was collected in a plain tube and the other half in an EDTA vacutainer. A total of 85 samples were collected (30 COVID-19-positive asymptomatic, 10 COVID-19-positive with mild symptoms, and 45 COVID-19-positive with severe symptoms) for the purpose of this study. COVID-19-positive asymptomatic individuals were identified as a result of the national screening campaigns. Symptomatic COVID-19 patients were classified into mild or severe based on guidelines published by Abu Dhabi Department of Health (circular number 33, 19<sup>th</sup> April 2020). Patients with mild disease presented with upper respiratory tract infection and symptoms like fever, dry cough, sore throat, runny nose, muscle and joint pains without shortness of breath. Patients with severe disease presented with severe pneumonia and symptoms like fever, cough, dyspnea and fast breathing (\>30 per minute), in addition to oxygen saturation \<90%. Patient records showed that many of the patients with symptoms were self-medicating with aspirin prior to their hospital visit and that some of the patients with moderate-severe symptoms were placed on dexamethasone and/or heparin subsequent to hospital admission. Immediately upon sample collection, the hospital laboratory staff separated and tested the serum for CRP, D-dimer, ferritin, IL-6 and LDH; a complete blood count was also performed on each sample. Whole blood samples were also aliquoted and frozen at −80 °C for subsequent processing and analysis. The study was jointly approved by the Ministry of Health, Abu Dhabi and Dubai Health Authority (DOH/CVDC/2020/1949) on the understanding that samples will be number-coded to hide patient’s identity, that no personal information will be shared with a third party and that no sample analysis can be performed by entities other than the Research Institute of Medical and Health Sciences (RIMHS), the University of Sharjah (UOS) without prior written approval. No informed consent was required as per the ethical approval decision (DOH/CVDC/2020/1949); in compliance with the said decision, all samples were fully anonymized before accessing or receiving them. ## Serum levels of hepcidin and soluble transferrin receptor (sTfR), sCD163 and haptoglobin (Hp) concentration and phenotype distribution Upon receipt of frozen samples at RIMHS, UOS laboratories, whole blood samples were thawed and centrifuged; serum was separated and levels of hepcidin (Cat No.733228; MyBiosource, San Diego, California, United States), soluble sTfR (Cat No. 750294; MyBiosource), sCD163 (MBS508555) and Hp (MBS763395) were measured using commercially-available colorimetric assay kits; absorbance was read at 450 nm on a microplate reader. Hp phenotypes were determined by vertical polyacrylamide gel electrophoresis, and the bands were visualized by staining with benzedine solution as previously described. ## Liquid chromatography tandem mass spectrometry (LC-MS/MS) Plasma was obtained after the collection of samples into heparinized tubes followed by centrifugation for 5 minutes (3000g). The samples were stored at –80 ºC for long-term storage until further metabolomics analysis. An aliquot of plasma sample was first placed into a microcentrifuge tube where cold methanol was added into the sample at 3:1 v/v (i.e., 30 μL sample, add 90 μL cold methanol) vortex and was then stored at –20ºC for two hrs. Next, the samples were centrifuged at 20,817 x g for 15 min at 4ºC. Then, the supernatant was transferred to a new microcentrifuge tube. Usually, the original sample volume is transferred three times (i.e., for 30 μL sample, add 90 μL cold methanol, then transfer 90 μL supernatant). The sample was dried down using Speed vac at 30–40°C. The dried sample was then either stored in a –80ºC freezer for further use or dissolved in solvent for LC-MS/MS analysis. Metabolites were analyzed by HPLC-Q-TOF MS/MS using the auto MS/MS positive scan mode as per described in our recent publications. Briefly, samples were chromatographically separated using a Hamilton<sup>®</sup> Intensity Solo 2 C18 column (100 mm x 2.1 mm x 1.8 μm) and eluted using 0.1% formic acid in water (A) and 0.1% formic acid in ACN (B) using the following gradient: at a flow rate of 0.250 mL/min 1% B from 0–2 min, then gradient elution to 99% B from 2–17 min, held at 99% B from 17–20 min, then re- equilibrated to 1% B from 20–30 min using a flow rate of 0.350 mL/min. The autosampler temperature was set at 8°C and the column oven temperature at 35°C. The ESI source with dry nitrogen gas was 10 L/min, and the drying temperature was equal to 220°C with nebulizer gas pressure set to 2.2 bar. The capillary voltage of the ESI was 4500 V and the Plate Offset 500 V. MS acquisition scan was set at 20–1300 m/z and the collision energy at 7 eV. Sodium formate was injected as an external calibrant between 0.1 and 0.3 minutes. A total volume of 10 μL sample was injected into the TIMS-TOF MS. Processing analysis was performed using MetaboScape<sup>®</sup> 4.0 (Bruker Daltonics). Analyte bucketing and identification were done using the software provided available T-ReX 2D/3D workflow with the following parameters: intensity threshold greater than 1000 counts and peak length equal to 7 spectra or greater. Feature quantitation performed using peak area, for features present in at least 3 (of 12) samples (per cell type) were considered for statistical analysis. Analyte MS<sup>2</sup> spectra were averaged on import and only features eluting between 0.3 and 25 min with mz between 50 and 1000. For metabolite identification, both MS<sup>2</sup> spectra and retention time (RT) were used with the MS/MS spectra as the minimum criterion for a positive hit. For the set of compounds meeting this criterion (either MS/MS alone or MS/MS with RT), annotation using Bruker’s implementation of the Human Metabolome Database (HMDB-4.0) was performed; all selected compounds were matched against this library. Where a particular database entry was matched by multiple features, putatively matching features were filtered by considering each of the features against the highest annotation quality score (AQ score) among other putative matches for the same metabolite; i.e. features exhibiting the best fit across the greatest number of factors such as retention time, MS/MS, m/z values, analyte list and spectral library matching were ranked first for the associated identification as per previous publication. Pathway enrichment analyses were performed using MetaboAnalyst V5 (<https://www.metaboanalyst.ca>). Pathway enrichment evaluates overall pathway impact by considering the relative importance of altered metabolite based on their position in the pathway map. All data, including raw files, have been deposited in the Metabolomics Workbench Repository (<https://www.metabolomicsworkbench.org/>). ## Data analysis The metabolomics data were first tabulated in Microsoft excel format and then exported to the Statistical Package for Social Sciences (SPSS) software, version 27. Demographics, clinical and metabolites data were all merged into one SPSS dataset. Descriptive statistics was used to conduct univariate analysis; frequencies and relative frequencies were used to condense categorical data while measures of central tendency were performed for scale data. Normality of scale data was first tested graphically, using Q-Q plots and histograms and then statistically analyzed using the Kolmogorov Smirnov test. Mean and standard deviations (SD) were reported for scale variables showing normally distributed data, whereas median and interquartile range (IQR: Q1-Q3) were used to summarize variables with skewed data. Chi-square test was performed to test for associations between categorical variables where the strength of an association was measured using the odds ratio (OR). To study the relationships between a normally distributed outcome and a categorical dichotomous predictor, the independent t-test was used. If the predictor had more than two groups, one-way ANOVA test was used to compare three or more means. For skewed outcome variables, similar analyses were conducted using the non-parametric tests Mann- Whitney U test and Kruskal-Wallis test, respectively. Spearman correlation coefficient was performed to investigate the correlation between two variables with skewed scale data. P-values less than or equal to 0.05 indicated statistical significance. Bonferroni correction was used to adjust p-values in pairwise comparisons. The receiver operating characteristic curve (ROC) and the area under the curve (AUC) were performed to identify, from among the clinical and metabolite tests, significant diagnostic biomarkers for predicting the severity of COVID-19 infection. An ROC AUC value above 0.70 indicated moderate to high level of accuracy of prediction. For each test’s AUC value, statistical significance was assessed against chance by calculating its 95% Confidence Interval (CI) and associated p-value. For each significant diagnostic test/biomarker showing high/moderate accuracy prediction level (AUC \> 0.70), data-driven approach was used to determine the optimal cut-off value, specifically, by maximizing the Youden’s index (Sensitivity + Specificity– 1). Next, the sensitivity (SN) and specificity (SP), along with their 95% confidence intervals, were calculated for each diagnostic test. Optimal cut-off values were used to dichotomize each biomarker into low and high levels. A low level was coded as zero and a high level was coded as 1. After excluding biomarkers that were linearly related, predictors were identified to develop a risk scoring system to define a diagnostic model for COVID-19 severity based on a combination of important biomarkers used as predictors. The risk score was calculated by counting, for each patient, the number of biomarkers that were of high levels. The ROC and the AUC, using Youden’s Index, were then used to identify the optimal risk score for predicting the severity of COVID-19. Demographics, clinical and serum metabolite laboratory test results were first compared between the three levels of COVID-19 infection severity (asymptomatic, mild and severe). Preliminary analysis has shown that the asymptomatic and mild groups were comparable on most clinical and metabolite results; no significant differences were observed between the two groups. Accordingly, the two groups (asymptomatic and mild) were clustered into a single group (asymptomatic/mild), which was then compared to the severe COVID-19 group to conduct the analysis reported in this manuscript. # Results ## Study population demographics In this study, we analyzed data pertaining to a total of 85 COVID-19 cases, of whom 35.3% (n = 30) were asymptomatic, 11.8% (n = 10) were mild and 52.9% (n = 45) were severe. Males constituted the majority of patients in this study (84.7%, n = 72) as compared to (15.3%, n = 13) females. Mean age of patients was 42 years (SD = 7.73) with a minimum age value of 27 years and a maximum of 62 years. Age was categorized into two groups where 43.5% (n = 37) were 40 years or younger and 56.5% (n = 48) were older than 40 years. Age group and gender distributions were comparable between the two groups (asymptomatic/mild) vs. severe. ## Laboratory findings (clinical tests) and severity of COVID-19 disease In the study sample as a whole, inflammatory markers including the C-reactive protein (CRP) and D-dimer had median values of 18.2 mg/L (Q1-Q3: 5.45–115.49) and 0.60 μg/ml (Q1-Q3: 0.31–2.24) respectively and mean values of lymphocyte and neutrophil counts of 1.45 (SD = 0.72) and 7.56 (SD = 4.13) cells/μL respectively. The majority of the inflammatory markers were found to be significantly higher in the severe COVID-19 group relative to the asymptomatic/mild group. For example, CRP had a median value of 5.90 mg/L in the asymptomatic/mild group and 131.22 mg/L in the severe group (U = 1627.50, p-value\<0.001). Similarly, D-dimer had median values of 0.28 μg/ml in the asymptomatic/mild group and 1.27 μg/ml in the severe group (U = 161050, p-value\<0.001). While lymphocytes were significantly higher in the asymptomatic/mild group (mean = 1.94 cells/μL) than in the severe group (mean = 1.00 cells/μL; t = 7.880, p-value\<0.001), neutrophils were significantly higher in the severe (mean = 9.35 cells/μL) as compared to the asymptomatic/mild group (mean = 5.58 cells/μL; t = -4.773, p-value\<0.001). No significant differences were found between the asymptomatic/mild group vs. the severe COVID-19 group, vis-à-vis median values of serum hepcidin or sCD163. However, Hp and soluble sTfR levels were significantly higher (p-value\<0.001) in the severe vs. the asymptomatic/mild group; Hp median values were 138.02 vs. 115.73 ng/ml and sTfR median values were 31.61 vs. 21.46 ng/ml. ## Metabolomics profiles of COVID-19 patients according to disease severity To investigate the possibility of identifying serum metabolites that help in studying the prognosis of disease severity, metabolomics profiling of plasma samples from patients with no, mild and severe symptoms was performed. A total of 99 metabolites were measured and compared between the asymptomatic, mild and severe cases. Pairwise comparisons showed comparable results for asymptomatic vs. mild patients, hence the grouping of data obtained from asymptomatic patients and patients with mild symptoms as one “asymptomatic/mild” group; much the same as was done in the previous section. Out of the 99 metabolites , 68 have shown significantly different median values between the asymptomatic/mild group and the severe groups. Of these 68, eight (8) metabolites (K_4\_Aminophenol, Acetaminophen glucuronide, Cytosine, Elaidic acide, Glycine, Isobutyric, Paracetamol sulfate and Succinylacetone) were significantly higher in the severe group. Additionally, sixty (60) metabolites showed significantly higher values in the asymptomatic/mild group vs. the severe group. Next, we conducted an enrichment analysis of the Biological Process gene ontology terms linked with those metabolites. The enrichment pathway analysis using the "small molecule pathway database (SMPDB)" (available in MetaboAnalyst 5.0 software) revealed that the pathways, that the differentially abundant metabolites were most enriched for included the citrate cycle, phenylalanine metabolism, phenylalanine, tyrosine, and tryptophan biosynthesis, pantothenate and coenzyme A biosynthesis, tryptophan, glycine, and serine. Additionally, the same data set produced disease-enriched groups for Hartnup disease, acute seizures, critical illness (major trauma, severe septic shock, or cardiogenic shock), and hyperbaric oxygen exposure when it was searched against the "blood disease signatures database" (available in MetaboAnalyst 5.0 software). As further detailed in the text, the bulk of the diseases or conditions that emerged from this research have symptoms that are consistent with those listed among the most severe COVID-19 cases. ## Determining Cut-off values for Biomarkers To identify clinical biomarkers of disease severity, the ROC and AUC were performed for each clinical laboratory finding (CRP, ferritin, sTfR, LDH, etc.) and metabolite test. For the clinical laboratory findings, the highest accuracy level in predicting disease severity was for LDH (AUC = 1.000) followed by Ferritin (AUC = 0.988, 95% CI: 0.966 to 1.000), D-dimer (AUC = 0.936, 95% CI: 0.876 to 0.997), CRP (AUC = 0.904, 95% CI: 0.842 to 0.966), IL-6 (AUC = 0.919, 95% CI: 0.831 to 1.000), Hp (AUC = 0.792, 95% CI: 0.696 to 0.889), sTfR (AUC = 0.756, 95% CI: 0.653 to 0.858) and Neutrophils (AUC = 0.749, 95% CI: 0.643 to 0.854). Hepcidin and Lymphocytes showed either insignificant or low AUC values (\<0.70), indicating low accuracy in predicting COVID-19 severity. Optimal cut- off values for the laboratory findings as determined by Youden’s index were 226 for LDH (SN = 100%, SP = 100%), 365 for Ferritin (SN = 95.6%, SP = 100%), 0.545 for D-dimer (SN = 93.0%, SP = 90.0%), 33.95 for IL-6 (SN = 87.1%, SP = 100%), 58.35 for CRP (SN = 73.3%, SP = 100%), 124.37 for Hp (SN = 79.5%, SP = 72.5%), 24.67 for sTfR (SN = 82.2%, SP = 60%), and 8.94 for Neutrophils (SN = 59.1%, SP = 85.0%). Of all the 99 metabolite tests, five were identified as significant diagnostic tests in predicting COVID-19 disease severity; namely, K_4\_Aminophenol (AUC = 0.883, 95% CI: 0.803 to 0.962), Acetaminophen (AUC = 0.949, 95% CI: 0.894 to 1.000), Acetaminophen glucuronide (AUC = 0.791, 95% CI: 0.539 to 1.000), Cytosine (AUC = 0.784, 95% CI: 0.680 to 0.887), and Paracetamol sulfate (AUC = 0.836, 95% CI: 0.660 to 1.000). Cut-off diagnostic value for K_4\_Aminophenol was 381.5 (SN = 82.2%, SP = 90.0%), for Acetaminophen was 1595.5 (SN = 90.9%, SP = 91.2%), for Acetaminophen glucuronide was 1416.0 (SN = 78.0%, SP = 85.7%), for Cytosine was 818.0 (SN = 68.2%, SP = 90.0%), and for Paracetamol sulfate was 652.5 (SN = 81.0%, SP = 88.9%). Each biomarker was then dichotomized into two groups, low and high, based on its determined cut-off value. ## Predicting COVID-19 disease severity Predicting the severity of COVID-19 was done at two levels, first by using a single biomarker and then a combination of biomarkers. All clinical and metabolites tests were significantly and strongly associated with the severity of COVID-19. The proportion of patients with severe COVID-19 in the high group of each clinical/metabolites test was significantly higher than that in the low group. The strength of association, as measured by the odds ratio, between disease severity and the clinical and metabolites groups, was lowest for CRP (OR = 4.3, 95% CI: 2.6 to 7.1) and highest for D-dimer (OR = 120.0, 95% CI: 25.1 to 572.9). A risk scoring system was developed to define a diagnostic model for COVID-19 disease severity based on a combination of important biomarkers used as predictors. All clinical laboratory findings and metabolites that were significantly associated with disease severity were considered as important biomarkers. After excluding biomarkers that were linearly related, and based on the statistical and clinical importance of all identified diagnostic biomarkers, we selected six predictors to conduct the risk-scoring predictive model. This model included the five lab findings (D-dimer, Ferritin, Neutrophils, Hp, and sTfR) and one metabolite (cytosine). A score was calculated for each patient. This score corresponded to the number of biomarkers that were of levels above their respective cut-off values (high group). The accuracy of predicting disease severity in the risk-scoring predictive model was reflected in a highly significant AUC value of 0.996 (95% CI: 0.989 to 1.000). The optimal cut-off risk score for the risk-scoring predictive model, as determined by Youden’s Index, was 3 with a perfect sensitivity of 100% and a specificity of 92.5%. Accordingly, all patients with high levels of at least three of the six predictors would be predicted as developing severe COVID-19 and none of the severe cases would be missed out. ## Correlating ferritin with other laboratory findings and plasma metabolites Among all patients, ferritin values were significantly correlated with the values of all other clinical tests except for hepcidin and sCD163. The significant correlations were all positive except for Lymphocytes that showed a moderate indirect correlation with ferritin (rho = - 0.520, p-value\<0.001, 95% CI: -0.664 to -0.338). The strongest direct correlations of ferritin were found with LDH (rho = 0.785, p-value\<0.001, 95% CI: 0.682 to 0.858), followed by D-dimer (rho = 0.630, p-value\<0.001, 95% CI: 0.475 to 0.748) and CRP (rho = 0.618, p-value\<0.001, 95% CI: 0.461 to 0.737); the weakest correlation was between ferritin and sTfR (rho = 0.282, p-value = 0.009, 95% CI: 0.067 to 0.472). Moreover, there were significant, mild to moderate indirect correlations between ferritin and several metabolites with serotonin showing the strongest indirect correlation (rho = -0.697, p-value\<0.001, 95% CI: -0.836 to -0.474). Ferritin was also found to have a significant positive correlation with acetaminophen (rho = 0.670, p-value\<0.001, 95% CI: 0.521 to 0.780), K_4\_Aminophenol (rho = 0.573, p-value\<0.001, 95% CI: 0.405 to 0.704) and cytosine (rho = 0.416, p-value\<0.001, 95% CI: 0.215 to 0.583). # Discussion In this study, the concentration of several serum analytes and metabolites was measured in COVID-19 patients with no, mild or severe symptoms. Consistent with numerous previous studies, the serum concentration of analytes that are routinely measured in infected individuals including CRP, ferritin, IL-6, D-dimer, IL-6 and LDH was significantly elevated in patients with severe COVID-19. Also consistent with previous work was the observation that the levels of hepcidin, sTfR, and Hp were either slightly-moderately elevated or not changed in COVID-19 patients with severe disease. Contrary to the suggestion of Zhou et al, our analysis showed that hepcidin is not a true predictor of disease severity in COVID-19 patients. Additionally, while data presented here show that the levels of sCD163 were reduced in severely ill patients, other studies have shown that sCD163 levels increase with disease severity. This discrepancy could be a reflection of variations in methodology, sample collection timing, and/or differences in macrophage activity. Irrespective of these discrepancies, variations in sCD163 concentration seem to have little, if any, impact on COVID-19 disease severity. Our data also showed that Hp phenotype distribution was similar in severe vs. asymptomatic/mild groups, which is in agreement with previous work which has suggested that Hp phenotype has no bearing on COVID-19 disease severity. With regard to the metabolomics profiling of COVID-19 patients and as noted earlier, 60 metabolites decreased in the severe cases relative to asymptomatic/mild patients’ group. The list of identified metabolites included several amino acids, vitamins and few fatty acids. These are regarded as the fundamental elements that support the rise in cellular demands during illness. However, through catabolism pathways, they are also involved in innate and adaptive immune responses to infection. Therefore, the decrease in some of the reported metabolites is consistent with previous studies, that reported lower levels of amino acids in hospitalized COVID-19 patients compared to asymptomatic ones. The outcomes do in fact support the previously noted negative correlation between amino acids and immune responsiveness and hyper-inflammation indicators, which is characteristic of severe COVID 19. For example, the levels of Kynurenic acid in severe cases was found to be lower than that in asymptomatic/mild cases. Previous studies have suggested that the Tryptophan catabolite/ Kynurenine pathway may play a significant role in COVID-19 and critical COVID-19. Moreover, it appears that the increased level of kynurenine and the ratio of kynurenine to tryptophan is strongly correlated with the severity of COVID-19 patients. Interestingly, the ratio of Kynurenic acid/Kynurenine did not significantly differ between COVID-19 patients compared with non-COVID-19 controls, indicating no significant changes in Kynurenic acid activity, according to a systematic review. This is indeed consistent with our finding that in patients with severe COVID-19, tryptophan and Kynurenic acid levels were significantly lower than in the counter group. Over the last three years, much time and effort has been spent on identifying serum biomarkers that could predict disease severity in COVID-19 patients with high accuracy. Elevated levels of serum biomarkers such as ferritin, IL-6, D-dimer and lactate dehydrogenase among others were reported to be valuable predictors of disease severity and death. However, not all COVID-19 patients showing elevated levels in one or more of these biomarkers ended up with severe disease and death. In other words, relying on one or more serum analytes tends to yield low prediction accuracy as evidenced by the fact that such approaches could only account for only a significant percentage of cases. In this context, numerous predictive models that relied on overlapping sets of variables drawn from COVID-19 patients’ demographic data, clinical signs and symptoms, chest X-ray imaging and co-morbidities were proposed (reviewed in. However, many of these severity predictive models suffer from a high degree of subjectivity and high likelihood of bias. For example, a disease severity predictive model based on the static demographics (age, gender, occupation, urban vs. rural living, socio-economic status, profile, etc.) of \>50000 Irish patients showed that modeling such parameter could predict hospitalization \[(AUC 0.816 (95% CI 0.809, 0.822)\], admission to ICU \[AUC 0.885 (95% CI 0.88 0.89)\] and death \[AUC of 0.955 (95% CI 0.951 0.959)\]. In the same study, body mass index (BMI≥40) was shown to be a risk factor for ICU admission \[OR 19.630\] and death \[OR 10.802\]. Moreover, while rural living was found to associate with increased risk for hospitalization (OR 1.200 (95% CI 1.143–1.261)\], urban living was found to associate with increased risk for ICU admission \[OR 1.533 (95% CI 1.606–1.682)\]. Another study which developed an artificial intelligence (AI)-based model based on 41 variables relating to patient’s demographics, physical measurements, initial vital signs, comorbidities and laboratory findings in a cohort of 5628 Korean COVID-19 patients yielded a predictive power of \>0.93 when 6 variables were used. Besides the fact that this model could be skewed by the demographics components making it more population-specific than desired, achieving 93% accuracy by relying on 6 variables is cumbersome and difficult to apply in many poor countries and rural settings. Another AI-based model was developed by relying on laboratory findings including LDH, IL-6, D-dimer, fibrinogen, glucose, monokine induced by gamma interferon (MIG) and macrophage derived cytokine (MDC) levels in 60 COVID-19 Russian patients. The model described by the study relied on eight parameters (creatinine, glucose, monocyte number, fibrinogen, MDC, MIG and CRP) to yield a predictive power of 83−87%. In other words, this laboratory findings-based model failed to account for 13–17% of patients at risk of severe disease. With the availability of six metabolite predictive biomarkers at our disposal, we sought to develop a high accuracy prediction model based on disparate data- integrating approaches; namely, patient laboratory findings and plasma metabolomics profiles. The statistical model developed and tested was based on ROC-AUC values. Although some metabolomes including acetaminophen, acetaminophen glucuronide and paracetamol sulfate significantly predicted COVID-19 severity, it was unlikely that these metabolomes were involved in the pathophysiology of COVID-19. Severe cases of the disease were more likely to receive more paracetamol than asymptomatic/mild cases. Therefore, these metabolomes were excluded from the predictive models. Accordingly, the predictive model was developed, and its prediction accuracy was internally validated using six biomarkers that were not linearly-related; namely, D-dimer, ferritin, neutrophil counts, Hp, sTfR and cytosine. Prediction accuracy of disease severity using this model was 0.996 (95% CI: 0.989 to 1.000) with optimal cut-off risk score of three biomarkers. In other words, out of 100 patients with severe COVID-19 showing significant elevation in at least three of the six metabolites would predict disease severity in all 100 patients. The model has the advantage of yielding high predictive power with small set of variables (three laboratory findings) that can be easily and quickly acquired at minimal cost. Moreover, the predictive model can be dynamically applied independent of non-empirical clinical data (co-morbidities, signs and symptoms and loss of taste or smell among others) and can be dynamically applied as the disease progresses making timely and proper clinical interventions possible. That said, the utility of the model remains limited by the fact that the study was conducted retrospectively on a small number of samples. Another limitation in our study is that, with the number of COVID-19 cases gradually dwindling to almost zero in the UAE as in most parts of the world, we were not able to compile a new independent dataset with the same set of predictors as means of validating our prediction model. Future studies are recommended to test the validity of the suggested model on multiple datasets to ensure its generalizability. # Conclusion By integrating laboratory findings and metabolomic profiling data, a model to predict disease severity in COVID-19 patients was generated. The accuracy of the model was high (\>98%), and it has the advantage of requiring three biomarkers to yield high sensitivity and specificity in predicting disease severity. The suggested model may prove useful in better managing COVID-19 patients at high risk of severe disease. Lastly, the fact that the model included cytosine as a biomarker and that cytosine is not usually included in routine laboratory testing for COVID-19 patients, merit further work on developing reliable and highly sensitive, yet quick and easy to perform, assays to measure serum cytosine concentration. # Supporting information 10.1371/journal.pone.0289738.r001 Decision Letter 0 Trongtrakul Konlawij Academic Editor 2023 Konlawij Trongtrakul This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 6 Nov 2022 PONE-D-22-27539Plasma metabolomics profiling identifies new predictive biomarkers for disease severity in COVID-19 patientsPLOS ONE Dear Dr. Hamad, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Dec 21 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Konlawij Trongtrakul, MD, PhD Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. 3\. If you are reporting a retrospective study of medical records or archived samples, please ensure that you have discussed whether all data were fully anonymized before you accessed them and/or whether the IRB or ethics committee waived the requirement for informed consent. If patients provided informed written consent to have data from their medical records used in research, please include this information. 4\. Thank you for stating the following in the Acknowledgments Section of your manuscript: "This work was supported by research grant CoV19-0305 (MH), seed grant 2001110138, University of Sharjah, UAE. This research is part of the -Human Disease Biomarkers Discovery Research Group-study. The authors wish to acknowledge the generous support of the Research Institute for Medical and Health Sciences, University of Sharjah UAE." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: "This work was supported by research grant CoV19-0305 (MH), seed grant 2001110138 (NCS), University of Sharjah, UAE. This research is part of the -Human Disease Biomarkers Discovery Research Group-study. The authors wish to acknowledge the generous support of the Research Institute for Medical and Health Sciences, University of Sharjah UAE" Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 5\. We note that Figure 1 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figure 1 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. 6\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: <http://journals.plos.org/plosone/s/supporting-information>. Additional Editor Comments: The manuscript provides an interesting issue in determining COVID-19 severity using the patient's laboratory investigations combined with significant metabolites from LC-MS/MS. The models provided by the authors seem optimistic, with AUC nearly equal to 1. Unfortunately, there is a concern about needing more statistical analysis. I recommend performing multivariable logistic regression analysis or binary regression analysis as it is more appropriate to finalize the model. Afterward, the author can test the AUC of the model. Nonetheless, a suitable criterion for selecting metabolites that reflect more severe disease in the model is necessary, rather than including acetaminophen or related medication. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Partly \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: N/A \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: No \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors developed the predictive models of COVID-19 severity based upon the previously common biomarkers together with plasma metabolomes. I have a few questions and suggestions as of the following: 1\. The authors should give more details regarding sample collection and treatment history of the enrolled cases, including 1) the blood samples were collected at which day of their symptoms (please give the range), 2) what was the treatment for symptomatic patients (both mild and severe cases)?, and 3) did the patients receive treatment (by physicians or by somebody else) prior to the sample collection? In fact, the treatment and the onset of disease at the time of sample collection may confound the predictive models. 2\. Were there any patients who developed the progression of symptoms (from mild to severe) and was there mortality among the severe cases? If there were, the authors should identify the differences of biomarkers and plasma metabolomes between the progressive cases vs. the non-progressive cases and between the survival vs. the non-survival cases. This may contribute to the identification of prognostic markers in this study as well. 3\. Besides age, sex, and co-morbidities, did the authors include some demographic information of the patients in the predictive models? For example, body mass index, smoking history, alcohol drinking history, physical activity, and medications since these can affect pulmonary and immune function that may determine the disease severity of COVID-19. In other words, adding these parameters to the models may increase the accuracy of prediction. 4\. Please give the rationale why the authors ran only the positive mode on the LC/MS. According to my previous run, the negative mode of this LC/MS condition could provide a lot of additional endogenous metabolomes, especially free fatty acids and phospholipids that play a crucial role in the pathophysiology of several diseases. 5\. Considering the positive mode LC/MS used in this study, this protocol is robust for the determination of acylcarnitines and phospholipids in human plasma. Again, these endogenous metabolomes play an important role in pathophysiology of several diseases. However, most of these metabolomes were not included in the 99 identified metabolomes of this study. Therefore, the authors should quantify these metabolomes and add them to the models. 6\. Among 99 identified metabolomes, some of them are considered exogenous compounds, such as caffeine, chlorpheniramine, acetaminophen, acetaminophen glucuronide, and paracetamol sulfate. It is less likely that these metabolomes are involved in the pathophysiology of COVID-19, and hence I do not think that they should be included in the predictive models. Specifically, the authors’ results showed that acetaminophen, acetaminophen glucuronide, and paracetamol sulfate were 3 of 5 significant metabolomes that predicted COVID-19 severity. According to these findings, it was likely that the severe cases received more paracetamol than asymptomatic/mild cases. Therefore, the exogenous metabolomes should be reported only in terms of the difference between groups of patients, but they should be excluded from the predictive models. 7\. The authors demonstrated that LDL exerted the highest accuracy level in predicting COVID-19 severity among the clinical biomarkers (Page 13, Lines 285-286). This finding was really interesting and should be included in the discussion section. 8\. Since the authors showed that LDL exhibited that highest accuracy level in predicting COVID-19 severity, it was likely that lipid status and insulin sensitivity play a critical role in severity of COVID-19. Did the authors include triglyceride, total cholesterol, HDL, VLDL, glucose, insulin, and HOMA- IR in the models? If the author did, what were the results? If the authors did not, I suggest that the authors should include them in the models. Reviewer \#2: I have 4 major comments, added here. The remaining will be uploaded as a document. From my perspective those 4 points are very important blockers to a publication at this stage. 1\. As briefly mentioned at the end of the discussion session this study requires a validation. No test / validation population was curved out from the initial cohort and all reported results were calculated on the full set of data. As such they are unreliably optimistic. This is hinted by the unusually high AUC values. I would strongly recommend that the proposed model(s) performance is validated in a new cohort prior to publication. 2\. As only few metabolites resulting from LC-MS were retained for the models I recommend that prior to publication the identification of those is validated against a standard. This will give the high confidence level (MSI 1) required to be of use to the scientific community. More specifically in the case of COVID-19 patients a lot of cytosine containing compounds are elevated. For many of those compounds the cytosine sub-structure tends to break easily at the MS source resulting in multitude of cytosine events in an acquired TIC. Each of those events will have a good match to cytosine based on MS/MS, but not on RT. Please make sure that the metabolic feature you selected for you model overlap in RT with a cytosine standard for your method i.e., column and LC-MS conditions. 3\. I strongly believe that predictive models should not rely on detecting medication (K_4\_aminophenol, acetaminophen) in patient’s blood. This is highly subjective to patient’s lifestyle choices and hospital practices. 4\. The study states in the abstract that “almost all such models, which relied on serum/plasma biomarkers, clinical data or a combination of both, were subsequently deemed as cumbersome, inadequate and/or subject to bias”. However, very few predictive models of COVID severity were discussed in this work. I don’t believe that this statement was well defended. A more thorough discussion on the alignment / misalignment of this study findings with previous models will be beneficial to the community. The LC-MS findings more specifically were lacking coverage. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \\\\\\\\\\ \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0289738.r002 Author response to Decision Letter 0 26 May 2023 Reviewer \#1 (R1): The authors developed the predictive models of COVID-19 severity based upon the previously common biomarkers together with plasma metabolomes. I have a few questions and suggestions as of the following: Q1. The authors should give more details regarding sample collection and treatment history of the enrolled cases, including 1) the blood samples were collected at which day of their symptoms (please give the range), 2) what was the treatment for symptomatic patients (both mild and severe cases)?, and 3) did the patients receive treatment (by physicians or by somebody else) prior to the sample collection? In fact, the treatment and the onset of disease at the time of sample collection may confound the predictive models. Response (R1 Q1): We thank the reviewer for his/her suggest. More details were provided regarding sample collection as per the reviewer’s suggestion. Regarding the issue of patient management, please note that this was in the early stages of the pandemic when treatment options were very limited and treatment guidelines were constantly changing. Please see Material and Methods, sample collection, page (P) 6-7 Line (L) 131-134. Q2. Were there any patients who developed the progression of symptoms (from mild to severe) and was there mortality among the severe cases? If there were, the authors should identify the differences of biomarkers and plasma metabolomes between the progressive cases vs. the non-progressive cases and between the survival vs. the non-survival cases. This may contribute to the identification of prognostic markers in this study as well. Response (R1 Q2): While the reviewer’s comment is very relevant, patient follow up was not possible Aas this was a retrospective study. Q.3. Besides age, sex, and co-morbidities, did the authors include some demographic information of the patients in the predictive models? For example, body mass index, smoking history, alcohol drinking history, physical activity, and medications since these can affect pulmonary and immune function that may determine the disease severity of COVID-19. In other words, adding these parameters to the models may increase the accuracy of prediction. Response (R1 Q3): We thank the reviewer for his/her comment and we totally agree with the reviewer’s comment; smoking history, alcohol consumption, etc. do indeed complicate the clinical picture and can be useful in predicating disease progression. In fact, previous models have evaluated the utility of some such demographic variables and showed that they are indeed very useful. (please see Boudou, M., et al. Sci Rep 11, 18474 (2021). <https://doi.org/10.1038/s41598-021-98008-6>). In this study however, no demographic parameters were included as the focus of the study was to test whether combining key routine lab findings such as plasma ferritin levels plus differential metabolomics profiling can be used to predict patient severity. To accommodate the reviewer’s concern, the reasoning behind excluding demographic parameters was briefly described. Please see P11 L39-37 Q4. Please give the rationale why the authors ran only the positive mode on the LC/MS. According to my previous run, the negative mode of this LC/MS condition could provide a lot of additional endogenous metabolomes, especially free fatty acids and phospholipids that play a crucial role in the pathophysiology of several diseases. Response (R1 Q4). We agree with the reviewer on the ionization mode of the processed samples; however, the goal of the study was metabolic profiling against our data base (HMDB), which is not comprehensive for metabolites prone to ionization in negative mode, such as fatty acids and phospholipids; thus, switching to negative mode will not improve the identification of certain metabolites. Furthermore, we have previously conducted studies with negative mode, and the results were not satisfactory. Q5. Considering the positive mode LC/MS used in this study, this protocol is robust for the determination of acylcarnitines and phospholipids in human plasma. Again, these endogenous metabolomes play an important role in pathophysiology of several diseases. However, most of these metabolomes were not included in the 99 identified metabolomes of this study. Therefore, the authors should quantify these metabolomes and add them to the models. Response (R1 Q5). Typically, the formation of acylcarnitine (carriers of the acetyl group) as an intermediate will occur in the intermembranous space of the mitochondria, transported into the matrix of the mitochondria, then hydrolyzed into carnitine and Acyl-CoA, so the time frame of its availability is short, and therefore, is identified as carnitine. Q6. Among 99 identified metabolomes, some of them are considered exogenous compounds, such as caffeine, chlorpheniramine, acetaminophen, acetaminophen glucuronide, and paracetamol sulfate. It is less likely that these metabolomes are involved in the pathophysiology of COVID-19, and hence I do not think that they should be included in the predictive models. Specifically, the authors’ results showed that acetaminophen, acetaminophen glucuronide, and paracetamol sulfate were 3 of 5 significant metabolomes that predicted COVID-19 severity. According to these findings, it was likely that the severe cases received more paracetamol than asymptomatic/mild cases. Therefore, the exogenous metabolomes should be reported only in terms of the difference between groups of patients, but they should be excluded from the predictive models. Response (R1 Q6): The authors thank the reviewer for this important comment and totally agree him/her on excluding the exogenous metabolomes from the predictive models. Accordingly, the authors decided to remove Model 1 which includes K-4-aminophenol and acetaminophen as predictors. The manuscript now reports a single model for predicting disease severity. The reported model includes a total of six predictors (5 lab findings and one metabolite, cytosine). Edits are highlighted in the manuscript in the subsection entitled “Predicting COVID-19 disease severity’ under the Results section. Furthermore, under the Discussion section, a few statements were added to explain why these metabolomes were excluded from the predictive model despite showing significant predictive findings in their ROC-AUC values. All edits and additions were highlighted in yellow in the Results, Discussion and Conclusion sections. Q7. The authors demonstrated that LDL exerted the highest accuracy level in predicting COVID-19 severity among the clinical biomarkers (Page 13, Lines 285-286). This finding was really interesting and should be included in the discussion section. Response (R1 Q7): We thank the reviewer for drawing our attention to this point. Unfortunately, the term LDL was erroneously used a couple of times in place of the correct term (LDH); we sincerely apologize for this error. Q8. Since the authors showed that LDL exhibited that highest accuracy level in predicting COVID-19 severity, it was likely that lipid status and insulin sensitivity play a critical role in severity of COVID-19. Did the authors include triglyceride, total cholesterol, HDL, VLDL, glucose, insulin, and HOMA- IR in the models? If the author did, what were the results? If the authors did not, I suggest that the authors should include them in the models. Response (R1 Q8): The authors thank the reviewer. Unfortunately, lipid status and insulin sensitivity were not included in the patients’ dataset. This point was addressed as a limitation in the discussion. Reviewer \#2: I have 4 major comments, added here. The remaining will be uploaded as a document. From my perspective those 4 points are very important blockers to a publication at this stage. Q1. As briefly mentioned at the end of the discussion session this study requires a validation. No test / validation population was curved out from the initial cohort and all reported results were calculated on the full set of data. As such they are unreliably optimistic. This is hinted by the unusually high AUC values. I would strongly recommend that the proposed model(s) performance is validated in a new cohort prior to publication. Response (R2 Q1): The authors agree to this point and have addressed it as a limitation under the Discussion section. P22 L455-459 2\. As only few metabolites resulting from LC-MS were retained for the models I recommend that prior to publication the identification of those is validated against a standard. This will give the high confidence level (MSI 1) required to be of use to the scientific community. More specifically in the case of COVID-19 patients a lot of cytosine containing compounds are elevated. For many of those compounds the cytosine sub-structure tends to break easily at the MS source resulting in multitude of cytosine events in an acquired TIC. Each of those events will have a good match to cytosine based on MS/MS, but not on RT. Please make sure that the metabolic feature you selected for you model overlap in RT with a cytosine standard for your method i.e., column and LC-MS conditions. Response (R2 Q2): For untargeted metabolomics analysis we follow a standard protocol according to the manufacturer’s recommendations “Acquisition of high quality LC-MS/MS data follows sample preparation in non-targeted metabolomics workflows, with the T-ReX® LC-QTOF solution, no LC-MS/MS parameter optimization is required” and “To analyze large sample cohorts which require high retention time stability the Elute UHPLC in combination with the dedicated T-ReX® Elute Metabolomics-kit: RP was used”. The Reversed-Phase LC column kit enables matching of retention times to values in the Bruker HMDB Metabolite Library. So, we use the same column, Elute UHPLC System, mobile phase gradients, and all parameters and settings. The m/z measurements were externally calibrated using 10 mM of sodium formate before sample analysis. In addition, sodium formate solution was injected at the beginning of each sample run and used for internal calibration during data processing. TRX-2101/RT-28-calibrants for Bruker T-ReX LC-QTOF (Nova Medical Testing Inc.) was injected before sample analysis to check and test the performance of the column reversed-phase liquid chromatography (RPLC) separation, multipoint retention time calibration, and the mass spectrometer. Also, TRX-3112-R/MS Certified Human serum for Bruker T-ReX LC-QTOF solution (Nova Medical Testing Inc.) was prepared from pooled human blood and injected before sample analysis to check the performance of the LC-MS instruments. As a standard protocol, the test mixture data files were uploaded to Metaboscape 4.0 and confirmed the presence of the entire set of metabolites in the samples by choosing the set with a higher annotation quality score (AQ score) representing the best retention time values, MS/MS score, m/z values, mSigma. Based on the nearest retention time values registered in the HMDB library all metabolites are filtered Q3. I strongly believe that predictive models should not rely on detecting medication (K_4\_aminophenol, acetaminophen) in patient’s blood. This is highly subjective to patient’s lifestyle choices and hospital practices. Response (R2 Q3): The authors thank the reviewer for this important comment and totally agree about on excluding the exogenous metabolomes from the predictive models. Accordingly, the authors decided to remove Model 1 which includes K-4-aminophenol and acetaminophen as predictors. The manuscript now reports a single model for predicting disease severity. The reported model includes a total of six predictors (5 lab findings and one metabolite, cytosine). Edits are highlighted in the manuscript in the subsection entitled “Predicting COVID-19 disease severity’ under the Results section. Furthermore, under the Discussion section, a few statements were added to explain why these metabolomes were excluded from the predictive model despite showing significant predictive findings in their ROC-AUC values. All edits and additions were highlighted in yellow in the Results, Discussion and Conclusion sections. Q4. The study states in the abstract that “almost all such models, which relied on serum/plasma biomarkers, clinical data or a combination of both, were subsequently deemed as cumbersome, inadequate and/or subject to bias”. However, very few predictive models of COVID severity were discussed in this work. I don’t believe that this statement was well defended. A more thorough discussion on the alignment / misalignment of this study findings with previous models will be beneficial to the community. Response (R2 Q4): We thank the reviewer for his/her comment. Firstly, the reviewer’s comment made it clear to us that we perhaps overstated the case regarding the lack of utility of previous models which relied on lab findings, clinical symptoms and demographics. Our commentary on previous models both in the abstract and in the discussion were watered down not to overstate the case. Secondly, as the reviewer suggest, the discussion was expanded and re-organized to better argue our case. Please see Abstract, lines 2-4 and discussion section, page …, line … Q5 The LC-MS findings more specifically were lacking coverage. Response (R2 Q5): We understand the reviewer's criticism regarding the identification of specific metabolites; nevertheless, the identification was compared to the 800 metabolites present in our HMDB, therefore we chose to use a more stringent set of criteria as described in M&M and keep those metabolites assigned by MS/MS only. This reduced the risk of false positive identifications. Accordingly, the number of metabolites here reported is comparable to other studies using the similar workflow 10.1371/journal.pone.0289738.r003 Decision Letter 1 Trongtrakul Konlawij Academic Editor 2023 Konlawij Trongtrakul This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 26 Jul 2023 Plasma metabolomics profiling identifies new predictive biomarkers for disease severity in COVID-19 patients PONE-D-22-27539R1 Dear Dr. Hamad, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Konlawij Trongtrakul, MD PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: (No Response) \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \\\\\\\\\\ 10.1371/journal.pone.0289738.r004 Acceptance letter Trongtrakul Konlawij Academic Editor 2023 Konlawij Trongtrakul This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 3 Aug 2023 PONE-D-22-27539R1 Plasma metabolomics profiling identifies new predictive biomarkers for disease severity in COVID-19 patients Dear Dr. Hamad: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Konlawij Trongtrakul Academic Editor PLOS ONE [^1]: The authors have declared that no competing interest exist.
# Introduction A considerable amount of evidence has accumulated demonstrating that personality traits show important associations with health and longevity. Conscientiousness, a dispositional tendency to be industrious, norm adhering, planful, and to thoughtfully inhibit impulses, has emerged as the Big Five trait most consistently associated with these outcomes. Mechanisms linking conscientiousness to health are diverse and include health behaviors such as smoking, physical activity, and diet. Conscientiousness also operates on health outcomes via social environmental factors, such as educational attainment, and socioeconomic status. As a result, conscientiousness has the potential to offer a large cumulative health benefit across multiple pathways, which may account for the now well-established association between conscientiousness and mortality. Conscientiousness is related to patterns of states and behaviors across a wide variety of health related domains, and as a result plays an important role in models of successful aging. Many of the same mechanisms linking conscientiousness to health and mortality, including physical activity, smoking, and educational attainment, have also been found to predict leukocyte telomere length (LTL). Telomeres are nucleoprotein structures which cap the ends of chromosomes, maintaining their structural integrity and preventing end to end fusion. In the absence of telomerase activity, telomeres shorten with each cell division. Critically short telomere length has been proposed as a mechanism leading to cell senescence or cell death. As a result, LTL has been proposed as a marker of biological cell aging. Consistent with this, shorter LTL is prospectively associated with mortality from cardiovascular disease in elderly individuals. Evidence for LTL as causal mechanism in the aging process however is limited, and at the present time it is perhaps best viewed as a surrogate endpoint for cumulative cellular damage. Because LTL is a cumulative outcome resulting from biological insults and stresses across many years, dispositional traits, which may exert persistent effects on factors related to health, are expected to also be associated with LTL. While this view is generally accepted, most of the work to date focusing on dispositional characteristics and variation in LTL has focused on health damaging effects associated with depression, anxiety, and neuroticism in the context of stressful and traumatic life events. In contrast, few studies have considered that dispositional traits positively associated with health might be associated LTL. Despite a large body of work suggesting overlapping correlates across both conscientiousness and LTL, only one recent study in Japanese medical students (mean age 23) has reported a cross sectional association between longer LTL and conscientiousness. To our knowledge ours is the first study investigating a prospective association between conscientiousness in childhood and LTL in later life. Predicated on the notion that conscientiousness serves as a stable dispositional trait which increases the likelihood that individuals will participate in a variety of health promoting behaviors and select into social environments that are conducive to health, we propose that conscientiousness will have a protective effect on LTL. Accordingly, we hypothesized that conscientiousness will predict LTL prospectively, such that individuals higher on conscientiousness early in life will show longer LTL in adulthood. We present pilot results using data from the ongoing Hawaii Personality and Health Study. A small number of stored dried blood samples out of a pool of approximately 800 available for testing were selected to validate our methods of DNA extraction and assaying LTL; we additionally planned to use these pilot data for an initial test of the association between childhood conscientiousness and LTL. Previously we had demonstrated that childhood conscientiousness predicts a global measure of physical dysregulation comprised of 11 objective biomarkers. As LTL and physiological dysregulation may be complimentary outcomes, we tested the degree to which childhood conscientiousness was associated with LTL independent of physiological dysregulation. Because LTL has been demonstrated to vary across cultural and ethnic groups, we additionally report on differences in LTL across ethnic and cultural groups and include this important variable in our analyses. # Methods ## Ethics Statement Study protocols for locating participants in adulthood, the adult clinic visit and for follow-up questionnaires in adulthood were approved by the IRBs at both the Oregon Research Institute and the Kaiser Permanente Center for Health Research, Hawaii. Upon being contacted in adulthood, the participants included in these analyses gave consent for their childhood personality data to be used as part of an ongoing longitudinal study. Written consent was provided prior to the clinic visit in adulthood, and participants who agreed to provide blood samples gave additional written consent for these to be used for future genetic analyses. Prior to completing questionnaires in adulthood, all participants completed and returned written consent forms by mail. ## Participants We used existing data and stored dried blood spots from the Hawaii Longitudinal Personality and Health Study. This project began with teacher assessments of childhood personality between 1959 and 1967 when the child sample was an average age of 10. Elementary school teachers’ personality assessments of all the children in their classrooms were obtained from elementary schools on the Hawaiian islands of Oahu and Kauai (n = 2,418). The schools were geographically dispersed across these islands, and represented a wide range of socio-economic status (SES). Beginning in 1998, efforts were initiated to locate members of the original child cohort in adulthood. Currently, 84% of the 2,320 available participants have been located. Of the 1,904 individuals who were living and available at the time of initial contact, 1,387 (73%) have agreed to participate in further research. Participants have completed several questionnaires administered between 1999 and the present. Data from two questionnaires were used in this report. Objective measures of physical health were collected as part of a clinic visit conducted at the Kaiser Permanente Center for Health Research, Hawaii in Honolulu, and also in clinics on the islands of Kauai and Hawaii. Health assessments were conducted by qualified staff using standard protocols, and included the collection of dried blood spots (see below for details). A total of 807 participants completed the clinic visit between 2002 and 2011, and 699 of these participants agreed to provide dried blood samples. The average year of assessment was 2006; approximately 40 years after the childhood assessments. The mean age of the clinic sample at the time of assessment was 51 years. The full adult sample reflects the ethnic diversity of Hawaii, and includes 44% Asian Americans, 22% Native Hawaiians or part Native Hawaiians, 15% of European ancestry, and 11% of Filipino ancestry, 4% indicating “other”, and 4% declining to respond. Financial constraints limited the number of samples that could be tested in the pilot. To allow for the possibility of finding an association we performed this initial test of the hypothesis that childhood levels of conscientiousness would predict adult LTL using an extreme groups design. LTL tends to differ across sexes, such that women tend to have longer telomeres at any given age. Participants who were assessed in one school, the University of Hawaii Lab School (n = 78), were excluded because they were assessed using a different set of items in childhood. Given the concern that sex difference could confound our findings, we elected to limit the current study to only one sex. The pilot sample was selected from women rather than men, and the choice of women instead of men was arbitrary. gives a flow diagram of the selection process for the pilot sample. The subsample for the current study was selected from those adult women who had provided a dried blood spot during the clinic visit and had also consented to allow this material to be used for future genetic analyses (n = 323). We ranked this subset of participants on childhood conscientiousness scores. Of these we selected the 30 scoring highest, and the 30 scoring lowest on childhood conscientiousness. In terms of standardized scores, this corresponded to cutoff points at 1.23 standard deviations above, and 1.40 standard deviations bellow the mean for childhood conscientiousness. ## Measures ### Childhood personality For the sample included in our analyses, a common set of 39 items across the samples was assessed across schools. Attributes were selected for their observability and focus groups of teachers developed definitions for each item. For each childhood item, teachers were presented with the definition and a list of their students’ names. At the end of the school year, teachers ranked every student in their class using a fixed nine-step quasi-normal distribution. Students were in grades 1, 2, 5 or 6. The majority of the sample, including all those in the pilot sample, were assessed on only one occasion in childhood. Example items assessing childhood conscientiousness include Careful of personal belongings, Planful, and Persevering. Although the items were selected prior to the widespread acceptance of the Big Five model of personality, Big Five personality factors have been robustly identified in these data, as is described in detail by Goldberg. Of these, the conscientiousness scale demonstrates acceptable reliability with α =.77. The stability of personality traits tends to increase with age. For a subset of the child sample assessed on more than one occasion, conscientiousness demonstrated adequate one year stability for childhood traits with a test re-test correlation of r =.53. Over long spans of time, personality traits are moderately stable. The best available estimate for the expected stability of personality traits over a 40 year span derived from a meta-analysis of over 150 longitudinal studies of personality is moderate with an estimated correlation of r =.25. Conscientiousness demonstrates a level of stability in our full sample that is directly in line with this estimate. Across a variety of instruments used to collect self-ratings of conscientiousness in adulthood and the teacher assessments in childhood, we find a 40 year stability coefficient of r =.25. In more recent research, teacher ratings of conscientiousness in childhood converge with parent ratings at levels similar to that observed for multiple observers rating adults they know well. Moreover, childhood conscientiousness in the Hawaii study shows predictive validity with respect to outcomes decades later in adulthood including educational attainment, the selection of occupational environments, health behaviors, and objectively measured physical health. This provides strong evidence for the reliability and validity of the teacher assessments. ### Adult personality As part of a mailed questionnaire, participants completed 120 items from the International Personality Item Pool (IPIP; items, scoring information, scale reliability, and convergent validity estimates available at [http://ipip.ori.org](http://ipip.ori.org/).) designed to assess the Big Five personality factors. Scores on conscientiousness were based on 24 items selected to replicate the facet structure of the NEO-PI-R conscientiousness scale. Participants rated these on a 1 (not at all like me) to 5 (a lot like me) scale. Adult conscientiousness scores were the sum of these 24 items. Reliability of the IPIP-NEO conscientiousness scale is excellent (α =.90), and the scale shows high convergence with the corresponding NEO-PI-R scale (r =.80). The NEO-PI-R is one of the most widely used measures of the Five Factor model and has established reliability and validity. Similarly the IPIP-NEO has demonstrated concurrent and criterion validity. Recent reports linking IPIP-NEO conscientiousness measured in adulthood to health behaviors, objective health outcomes, and mortality further support the validity of the measure. Ten participants in the pilot sample failed to complete and return the adult personality assessment. ### Demographics Demographic variables including date of birth, sex, self-identified culture, and educational attainment were assessed by questionnaire. Educational attainment was measured on a 9-point scale ranging from 1 (eighth grade or less) to 9 (postgraduate or professional degree). Five participants did not respond to the educational attainment item. Participants self-selected their own cultural identity from 12 options, and were instructed to select the one cultural group that they most identified with in cases where individuals identified with more one group. For the current analyses, cultural groups were collapsed into five groups: Caucasian (n = 10), Asian American (n = 21), Hawaiian/part Hawaiian (n = 9), Filipino (n = 11), and other (n = 4). Three participants in the pilot sample declined to indicate their cultural identity. We performed a series of chi-square goodness of fit tests in order to evaluate whether the observed proportions for cultural groups in pilot sample, and in the selected extreme groups (both high and low) differed from the sample from which these were drawn (n = 323). The proportion of participants in each cultural group did not differ in the pilot sample (χ<sup>2</sup> = 7.34, df = 5, p =.197) in the low conscientiousness group (χ<sup>2</sup> = 5.59, df = 5, p =.349), or in the high conscientiousness group (χ<sup>2</sup> = 9.25, df = 5, p =.100) when compared to the proportions observed in the sample from which these were drawn (n = 323). ### Smoking Lifetime history of smoking behavior was assessed via questionnaire. Pack years, defined as smoking one pack a day for one year, were derived from these. To create a smoking variable incorporating cumulative pack years with current smoking status, participants’ self-reported smoking histories were coded into a five point scale based on the distribution of smoking history in the full sample (0 = ≤ 100 cigarettes smoked, 1 = previously smoked \< 8 pack- years, 2 = previously smoked ≥ 8 pack years, 3 = currently smoke ≤ ½ pack a day, 4 = currently smoke \> ½ pack a day). Eight participants provided incomplete smoking data. ### Physical activity Three questionnaire items assessed strenuous, moderate, and mild physical activity via questionnaire. Participants were asked to rate their activities over a typical seven day period in the past year, including only activities lasting 15 minutes or more. For each item, a list of example activities was provided. Frequency for each item was rated on a five point scale (0 = zero times, 1 = one or two times, 2 = three or four times, 3 = five or six times, 4 = seven or more times). Scores on these three items were aggregated by a weighted sum. Four participants failed to complete these items. Twelve additional participants were administered a shortened questionnaire that did not include the physical activity items. ### BMI Height and weight were recorded as part of the clinic visit using a stadiometer and a balance beam scale. Scores for BMI were calculated using the standard formula (kg/m<sup>2</sup>). ### Physiological dysregulation Global health status was indexed using an aggregate score of 11 clinical indicators assessed during the clinic visit. These are systolic and diastolic blood pressure (measured twice and averaged), HDL cholesterol, total cholesterol/LDL ratio, fasting triglycerides, fasting blood glucose, body mass index (BMI), waist/hip ratio, urine protein (log transformed), and whether or not the participant was using prescription drugs for cholesterol or blood pressure. Participants arrived at the clinic in a fasting state. Blood from a venous arm draw, as well as urine samples, were collected and immediately sent for lab analysis. Waist circumference was measured at the point midway between the lowest rib and the iliac crest. Measures of hip circumference were made at the point where the gluteal muscle showed the greatest protrusion. Blood pressure was measured after five minutes of rest in a seated position using a sphygmomanometer. Two measures were taken with a one minute interval. Measures were scored such that higher scores corresponded to worse health. For some biomarkers, scores at either extreme can be unhealthy. We evaluated the number of participants in the full sample falling into the low/unhealthy range for each biomarker, and found that in all cases less than 1.2% met this criterion. Each measure was standardized and converted to a deviation score from the mean. The 11 biomarkers were not highly correlated with each other, and factor analysis did not reveal a single factor solution. Following recommendations for aggregating uncorrelated indicators we constructed an index by summing standardized scores, such that higher scores corresponded to greater physiological dysregulation, and worse health. Participant in the full sample who had more than 3 biomarkers missing were not scored on dysregulation. Of the participants included in the pilot sample, all but one had complete data on the dysregulation measure, and one participant was missing 3 biomarkers. More detail on the derivation of the physiological dysregulation scale can be found in Hampson, et al.. ### Leukocyte telomere length During the clinic visit, blood from a venous arm draw (125 μL) was blotted onto Whatman FTA Micro-cards (cat# WB120210) specifically designed for long-term preservation of dried blood and nucleic acids for quantitative DNA analysis. Following the manufacturer’s instructions, the dried blood spots were air dried, placed in plastic bags with an appropriate desiccant, and stored in the dark at room temperature. Specimens for the current study were identified, then shipped overnight in a room temperature-controlled box to Dr. Côté’s lab at the University of British Columbia (UBC). Researchers who assayed the samples remained blind to all participant data. Six 2mm punches were removed from each blood spot. DNA was extracted from these punches using QIA Amp kit on a QIA Cube (QIAGEN), and LTL was assayed using qPCR following methods described by Cawthon which have been modified and recently extended for use on dried blood spot samples. Samples were assayed in duplicate, on two 96-well plates, using the undiluted DNA extracted from the dried blood spots. The DNA concentrations ranged from 3 to 7 ng/μL. The standard curve was built from a serial dilution of human genomic DNA pooled from frozen whole blood from 24 healthy donors. Three internal controls (one DNA from a human cell line, one from pooled genomic DNA, and the other from a single volunteer) were included in each run and used for quality control related to the Lightcycler run; these were not used to normalize, and their T/S ratio was not incorporated into the standard curve. All samples were amplified successfully and met our quality control criteria. All runs had a PCR efficiency between 1.88 (94%) and 1.95 (97.5%). The intra-run (n-12) coefficient of variation (CV) for this assay is 5%, and based on the internal control included in each of the 22 runs performed, the inter-run CV was 8.5%. The assay produces an arbitrary count for telomeric DNA (T), and for a specific single copy nuclear gene DNA (S). The ratio of these (T/S) was then used as a measure of average relative LTL. We previously showed that DBS LTL is highly correlated with whole blood LTL. ## Statistical Analysis SPSS version 19 was used for statistical analysis. Group comparisons were evaluated using t-tests and ANOVAs. Standardized group differences were estimated using Cohen’s d. Tests of normality were conducted using the Shapiro-Wilk statistic. Non-parametric tests were conducted using a Mann-Whitney U test. Effect sizes were estimated by Pearson correlations, biserial correlation, and partial biserial correlations where appropriate, and all significance tests (α =.05) were two tailed. To maximize the sample size, all analyses employed pairwise deletion in the presence of missing data, hence sample size varied across analyses. # Results Initial examination of the distribution of the study variables revealed two extreme outliers, each with LTL exceeding 3.5 standard deviations above the mean. The two dried blood spots used for DNA extraction for these participants had been stored on filter paper designed for neonatal blood samples rather than the Whatman micro cards used for the rest of the sample. Because these LTL measurements seemed implausible, and the difference could be attributed to the difference in filter paper, they were excluded from our main analysis. provides descriptive information on the full sample from which the subsample was drawn, as well as the high and low childhood conscientiousness groups after removing the two extreme outliers, one from each group. The subsample (n = 58) did not differ from the full sample (n = 323) on any of the variables tested, and LTL was normally distributed (W =.96, p =.09). Consistent with our previous work with the Hawaii cohort, higher childhood conscientiousness was associated with higher educational attainment, lower BMIs, and healthier (lower) levels of overall physiological dysregulation. The high and low childhood conscientiousness groups did not differ with respect to age at the time of the childhood assessment, or at the time of the blood draw in adulthood. The two groups did not differ in levels of adult conscientiousness, smoking, or physical activity levels. Due to the use of an extreme groups design based on childhood conscientiousness, childhood conscientiousness was not normally distributed (W =.81, p \<.00). Because of this, associations with childhood conscientiousness were estimated using biserial correlations. The correlation between childhood conscientiousness and adult conscientiousness was similar in magnitude to that observed in our full sample (r =.19, p =.20, n = 48). The correlation was not significant at the level of p =.05, likely due the small sample size. A t-test to evaluate differences in LTL between the two groups was significant (t = 2.38, df = 56; p =.02), such that adult women who had been described by their teachers as higher on conscientiousness 40 years earlier had longer LTL (Cohen’s d =.62). This difference is illustrated in. We obtained a similar result using a non-parametric test including the two previously removed outliers. Specifically, a Mann-Whitney U test demonstrated that the high childhood conscientiousness group (n = 30, median LTL = 7.32) differed from the low childhood conscientiousness group (n = 30, median LTL = 6.01) on adult LTL (U = 303.00, Z = -2.17, p =.03). We next tested for associations between potential covariates and LTL. provides correlations for all study variables and LTL. The only significant correlation reported is the biserial correlation between LTL and childhood conscientiousness (r =.30; p =.02). Age, physiological dysregulation, educational attainment, and adult conscientiousness were not significantly correlated with LTL. The subsample did not have sufficient power to use multiple regression to test potential confounding variables or mediators for the effect of interest. None of the potential covariates or mediators in were significantly associated with LTL. However, because we expect that these variables may account for some portion of the association between childhood conscientiousness and LTL, we report the partial biserial correlations between childhood conscientiousness and LTL controlling in turn for age, physiological dysregulation, educational attainment, adult conscientiousness, smoking, BMI, and physical activity separately in. We contrasted these with the raw biserial correlation between childhood conscientiousness and LTL reported in. Controlling for age did not attenuate the raw biserial correlation, likely due to the narrow age range in the sample. Similarly, controlling for adult conscientiousness had no appreciable effect. Partial biserial correlations controlling for physiological dysregulation and educational attainment tended to attenuate the association. In the case of educational attainment, the partial biserial correlation failed to reach significance (r =.27, p =.05). Separately controlling for smoking, BMI, and physical activity attenuated the association between childhood conscientiousness and LTL in adulthood in a similar manner. To evaluate the degree to which LTL might vary across cultural groups, we used a one-way ANOVA. We found differences in LTL across groups, F(4, 50) = 3.27, p =.02. A post hoc analysis using Tukey’s HSD test revealed that individuals who described themselves as Native Hawaiian tended to have shorter LTL in comparison to all other groups (p \<.05), with the exception of those classified as other. We next tested ethnicity and childhood conscientiousness in a two-way ANOVA using type II sums of squares to partially adjust for unequal groups. The overall ANOVA was significant F(9,45) = 3.39, p =.00. Both childhood conscientiousness F(1, 45) = 5.06, p =.03 and cultural identity F(4, 45) = 3.04, p =.03 remained significantly associated with LTL. There was no significant interaction between ethnicity and childhood conscientiousness. We additionally tested a series of one-way ANOVAs in order to evaluate the degree to which study variables might have differed across cultural groups. Two variables showed cultural differences. For educational attainment, a one way ANOVA F(4, 47) = 3.05, p =.026 showed group differences across cultural groups. Post-hoc analyses using Tukey’s HSD showed one significant differences across groups such that Asian Americans (M = 7.70, SD = 1.35) reported higher educational attainment than participants identifying themselves as Native Hawaiian (M = 5.29, SD = 1.25). In the case of, BMI a one way ANOVA F(4, 47) = 3.05, p =.026 showed group differences across cultural groups. Post-hoc analyses using Tukey’s HSD showed that Native Hawaiians (M = 35.99, SD = 6.35) had significantly higher BMI scores in comparison to Asian Americans (M = 25.60, SD = 8.43). # Discussion To our knowledge, this is the first study to demonstrate an association between a Big Five trait in childhood and LTL in adulthood. Our findings suggest that childhood conscientiousness assessed by teachers predicts LTL 40 years later in adulthood. Controlling separately for age, physiological dysregulation, educational attainment, smoking, and BMI, attenuated the biserial correlation between of childhood conscientiousness and adult LTL. The association remained significant when controlling for age and dysregulation, but failed to reach significance when controlling for educational attainment, smoking, BMI, or physical activity. While the partial biserial correlation was not significant in this case, the degree to which the observed association was attenuated was small. While it is tempting to view the transition of significant to non-significant as meaningful, this does not constitute a formal test of mediation. Partialing out adult conscientiousness slightly increased the observed biserial correlation (r =.32 versus r =.30). This difference is likely due to sampling error and the smaller sample size available for this analysis (n = 45). Partialing out physical activity similarly increased the magnitude of the observed biserial correlation slightly, although in this case the effect was no longer significant (r =.32, p =.05), which is likely attributable to the reduction in sample size due to missing data on physical activity (n = 43). We found significant variation in LTL across ethnic groups, such that those identifying as native Hawaiian tended to have the shortest LTL. Both childhood conscientiousness and cultural group identity independently predicted LTL with no evidence of an interaction. These results suggest a novel pathway linking dispositional traits in childhood to cumulative health that has not previously been incorporated into research on LTL. Childhood personality traits have moderate to low levels of stability across the life-course. Our best estimates of the rank order stability of conscientiousness over 40 years corresponds to a correlation of r =.25. Both trait level conscientiousness and cognitive features of self-regulation develop and improve from childhood to young adulthood. Despite this, childhood conscientiousness predicts outcomes in adulthood including physical health, substance use, career attainment, and financial outcomes. In the broader context of lifespan development, LTL complements the growing list of consequential objective outcomes in adulthood associated with childhood conscientiousness. We did not find an association between adult levels of conscientiousness and LTL. This must be due in part reality that selecting on child conscientiousness for an extreme groups design magnifies effects for childhood conscientiousness but, given the moderate stability of the trait over 40 years, does not necessarily do so for adult conscientiousness. Sadahiro et al. reported a cross sectional association between conscientiousness and LTL in a sample of young adults at an age intermediate to the two points of personality assessment in our sample. It is possible that conscientiousness assessed at different points in the lifespan may show associations of differing magnitudes with respect to LTL. Personality and health researchers investigate models of the underlying psychosocial and biological processes that mediate associations between personality traits and health outcomes over the lifespan. LTL assessed in adulthood has several advantages as an objective measure of health status when studying these models: LTL can reflect cumulative insults, changes dynamically over time, and can be readily measured. These features, along with demonstrations that biological stressors, physical morbidity, and subjective states of distress predict LTL, suggest that LTL may be viewed as a psychobiomarker. As such, LTL offers an outcome variable that reflects the accumulation of health effects associated with numerous psychosocial and demographic variables. An important next step for lifespan developmental researchers involves evaluating the degree to which associations with childhood conscientiousness operate independently of adult trait levels and if so, whether these depend entirely on chains of events following from child trait levels and self-regulatory capacities. In models of effects related to health and aging, LTL may play a unique role as an integrative objective outcome. Particular strengths of the current study are the use of a prospective design with a 40 year interval, and the well-executed assessment of child personality. Although these are preliminary results, they are consequential in suggesting future directions and methods for incorporating LTL research into lifespan developmental models of dispositional traits related to health. Consistent with a lifespan mediation hypothesis, we found that controlling for potential mediators tended to attenuate the strength of association between childhood conscientiousness and LTL. The current study has specific limitations. An extreme groups design was selected to enhance our ability to detect an association between childhood conscientiousness and LTL in a preliminary sample. However, extreme groups designs do not necessarily produce accurate estimates of effect sizes and tend to inflate effects sizes associated with the selection variable. As a result the raw biserial correlation between childhood conscientiousness and LTL is likely inflated. Similarly, effect sizes for variables correlated with the selection variable can be distorted. In our analyses, partial biserial correlations controlling for variables associated with childhood conscientiousness are likely to be overcorrected. We might expect the true effect size of this effect to be reduced in magnitude in a larger sample. We have secured funding to assay telomere length in our larger sample, and in the future we will test this directly. Additionally, our small sample did not allow us to test multiple competing predictors, account for confounding variables, or test mediation hypotheses. We limited our sample to women, because there is evidence that LTL differs across genders. As a result, our results cannot be generalized to men. Our pilot sample is unlikely to give accurate estimates of effects related to cultural group and these results need to be interpreted with caution. Despite this limitation, our results do suggest the importance of investigating differences in LTL across cultural and ethnic groups, and for controlling for these effects when investigating predictors of LTL. Future analyses using our full sample will allow us to accurately evaluate effects across cultural groups. For all of these reasons, we view the present results as preliminary. Finally, we measured LTL using a qPCR-based assay which may have a higher variability than some other methods but is the appropriate option given the type of sample studied here and the large number of samples to be assayed in the future. Despite these limitations, the finding of a putative association between childhood conscientiousness and adult LTL is remarkable. There is a paucity of research linking personality at any stage in the life course to biomarkers of health; more so for childhood personality traits. Our results suggest a variety of future directions involving lifespan developmental trajectories of dispositional traits, physical health, and LTL. In addition to providing a more accurate estimate of the effect size of the associations presented here, future research will allow for testing lifespan mediation models where behavioral and environmental mechanisms can be evaluated as explanations of prospective associations between personality traits and LTL. Pathways such as smoking, physical activity, and diet may explain associations between dispositional traits in childhood and LTL years later. Joint or interactive effects of personality traits and sex, socio-economic status, and ethnicity may further explain variability in LTL. The authors thank all past and present members of the Lifestyle, Culture and Health team at the Kaiser Permanente Center for Health Research, Hawaii, the members of Dr. Côté’s lab at the University of British Columbia, and the team at Oregon Research Institute, for their many efforts that have led to this report. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GWE SEH. Performed the experiments: GWE SEH HCFC. Analyzed the data: GWE. Contributed reagents/materials/analysis tools: HCFC. Wrote the paper: GWE SEH HCFC.
# Introduction From 2001 to 2006, Salmonella enterica serovar Enteritidis was implicated in three outbreaks linked to raw almond consumption. Epidemiologic and traceback investigations of a 2000 to 2001 (denoted 2001) salmonellosis outbreak in Canada and in the United States identified a rare phage type (PT) of Salmonella Enteritidis, PT 30, from clinical samples, raw almonds sampled at retail, environmental (swab) samples at an almond processor and a huller-sheller facility, and environmental drag swabs obtained from multiple orchards in California. Salmonella Enteritidis PT 30 was recovered in 2001 (three of six samples collected)) and, in a subsequent study, at least two times in each year between 2002 and 2006 from environmental drag swabs collected in one of these orchards (eight to 96 samples collected each year). Raw almonds were epidemiologically linked to clinical cases of Salmonella Enteritidis PT 30 reported in Sweden in 2005 to 2006 (denoted 2006) but Salmonella was not isolated from the implicated California almonds. Between 2003 and 2004 (denoted 2004), another rare phage type, Salmonella Enteritidis PT 9c, was linked to consumption of raw California almonds in the United States. Excluding the farms linked to the 2001 outbreak, the prevalence and levels of Salmonella were determined in a multi-year survey of raw almond kernels. Samples of individual lots of raw almond kernels were collected as they were received by handlers (processors) located throughout the almond-growing regions of California from 2001 to 2007, and in 2010 and 2013. Inshell almonds were included in the survey in 2006 and 2007. The prevalence of Salmonella in 14,949 lots of raw almond kernels was 0.98% ± 0.29% over the 9 years (146 positive 100-g samples). Levels of Salmonella were estimated for 118 of the positive lots; mean and median levels of Salmonella were 1.14 ± 1.69 and 0.79 most probable number (MPN)/100 g, respectively, with single lots at 9.25 and 15.4 MPN/100 g. Of the small number of raw inshell almond lots (455) evaluated in 2006 and 2007, 1.5% (seven 100-g subsamples) were positive. Using classical serological nomenclature, the Salmonella isolates retrieved from these surveys were serotyped into 45 different serovars, including Salmonella Enteritidis PT 30 and PT 9c. Pulsed-field gel electrophoresis (PFGE), multilocus variable-number tandem repeat analysis (MLVA), and comparative genomic indexing (CGI) were applied to Salmonella Enteritidis strains associated with the 2001, 2004, and 2006 almond outbreaks, including clinical and environmental isolates. The Salmonella Enteritidis PT 30 and PT 9c strains could be separated from each other and from other Salmonella Enteritidis phage types based on DNA enzyme restriction profiles, MLVA types, and genes identified by CGI. However, neither PFGE nor MVLA could discriminate among the Salmonella Enteritidis PT 30 isolates associated with the 2001 and 2006 almond-associated outbreaks. Salmonella Enteritidis PT 30 almond-associated outbreak strains could not be distinguished from epidemiologically unrelated Salmonella Enteritidis PT 30 clinical strains included in the study. Among Salmonella Enteritidis PT 30 strains with identical genotypes, metabolic analyses with Biolog revealed differences between clinical and environmental isolates, indicating the discriminatory limit of the then current genotyping methods. Environmental Salmonella Enteritidis PT 30 isolates, collected between 2001 and 2006 from one of the 2001 outbreak- associated orchards, clustered in two groups based on the separation by PFGE of their XbaI-digested DNA. Whole genome sequencing (WGS) has replaced PFGE for investigating foodborne outbreaks because of its higher resolution, and this methodology has been incorporated into routine public health surveillance since 2014. Clonality of pathogen strains determined by WGS analyses can provide information about contamination during food production and distribution. Different workflows have been developed to assess levels of genetic relatedness of foodborne pathogens by the National Center for Biotechnology Information (NCBI) with the Pathogen Detection Pipeline (<https://www.ncbi.nlm.nih.gov/pathogens/>), the U.S. Food and Drug Administration with the Center for Food Safety and Applied Nutrition CFSAN SNP Pipeline, and the U.S. Centers for Disease Control and Prevention with Lyve-SET, based on the specific needs of the respective agencies. The CFSAN SNP Pipeline creates high quality SNP matrices from WGS data that allow connection between clinical isolates to food or environmental isolates based on their evolutionary relationship. Resident pathogen strains in facilities will have closely related WGS profiles, whereas transient pathogen strains will have unique or unrelated WGS profiles. The objective of the present study was a comparative genomic analysis of 171 Salmonella (45 serovars) isolated from raw inshell almonds and almond kernels in 9 years of surveys conducted between 2001 and 2013; 30 Salmonella Enteritidis PT 30 isolates recovered between 2001 and 2006 from a 2001 outbreak- associated almond orchard were included in the analysis. The CFSAN SNP Pipeline was selected to measure the genetic distances among the isolates. WGS was used to predict antimicrobial resistance (AMR), virulence genes, and presence of plasmid DNA. # Materials and methods ## Isolate selection Isolates were retrieved from enrichment of raw almond kernels and inshell almonds (survey isolates), and from swabs that were dragged across the floor of one of the 2001 outbreak-associated orchards (orchard isolates). A total of 15,505 ∼1-kg samples from different lots of raw almond kernels (14,949) and inshell almonds (455) were collected upon receipt at several almond processors located throughout California from the 2001–2005, 2006–2007, 2010 (; present study), and 2013 (; present study) harvests, as described previously. All samples were coded to keep the identities of the processors confidential; the geographic origins of the samples were unknown. The Safe Food Alliance, Kingsburg, CA (formerly American Council for Food Safety and Quality, Fresno, CA) analyzed a 100-g subsample from each lot of almonds by enriching for the presence of Salmonella. In addition, the MPN of Salmonella was determined for 118 samples by enriching one or more additional samples weighing from ∼0.25 to 100 g. Salmonella isolates were stored at −80°C. A total of 171 Salmonella survey isolates were selected for the present study: for each individual positive almond lot, a single Salmonella isolate retrieved from the initial 100-g subsample (153 positive samples, 148 isolates; three, one, and one isolate from 2001, 2002, and 2003, respectively, were lost), any isolates recovered from MPN or secondary enrichments that differed from the initial serotype (12), and unique isolates recovered from secondary enrichments of initially negative samples (11) (Tables and S1). Traditional serotyping was done for each banked Salmonella by the California Animal Health and Food Safety Laboratory System (Davis, CA), and phage typing for Salmonella Enteritidis isolates was done by the National Veterinary Services Laboratory (Ames, IA). Salmonella Enteritidis PT 9c LJH1024 (obtained from Robert Mandrell, U.S. Department of Agriculture, Agricultural Research Service; RM4635 or G04-101, raw almond isolate from the 2004 outbreak) and Salmonella Enteritidis PT 30 LJH0608 (raw almond isolate from the 2001 outbreak that was deposited to the American Type Culture Collection as ATCC BAA-1045) were also included in some of the analyses (S2 Table). Several 2001 outbreak-associated almond orchards sampled by investigators during the 2001 outbreak investigation were positive for Salmonella Enteritidis PT 30. One of these orchards was sampled every year from 2001 through 2006. Briefly, sterile gauze swabs attached to a string and soaked in full-strength evaporated skim milk were pulled along the orchard floor in a standardized manner. Four individual swabs were pooled and a procedure designed for recovering Salmonella from poultry houses was used to enrich the samples. Three of six (50%) pooled swab samples collected in 2001 and 53 of 228 (23%) samples collected between 2002 and 2006 were positive for Salmonella; every isolate was identified as Salmonella Enteritidis PT 30. A total of 30 Salmonella Enteritidis PT 30 orchard isolates from 2001 (3), 2002 (12), 2003 (10), 2005 (2), and 2006 (3) were analyzed in the present study (S2 Table). One orchard isolate from 2002 and all orchard isolates from 2004 (25) were not available and thus not included. Using the pathogen detection tool associated with the NCBI database (<https://www.ncbi.nlm.nih.gov/pathogens>), the sequence read archive (SRA) data were downloaded for 12 Salmonella Enteritidis PT 30 clinical isolates from the 2001 almond outbreak and for four clinical isolates from the 2006 almond outbreak (S3 Table). ## Whole genome sequencing Isolates were retrieved from frozen glycerol stock and plated on tryptic soy agar (TSA). Following overnight incubation at 37°C, one colony was inoculated into 2 ml of tryptic soy broth (TSB) and incubated for 24 h at 37°C, with shaking at 120 rpm, before being pelleted by centrifugation at 14,000 × g for 2 min. Genomic DNA, for the isolates in S2 Table, was extracted with the QIAamp DNA minikit (Qiagen, Valencia, CA) following the manufacturer’s directions. The 150-bp paired-end libraries were constructed for each purified DNA with the Illumina Nextera DNA flex library kit following the manufacturer’s directions (Illumina Inc., San Diego, CA). Pooled samples were sequenced on an Illumina 4000 HiSeq system by the DNA Technologies and Expression Analysis Core at the UC Davis Genome Center. DNA, for all the survey isolates, was sequenced as described by Emond-Rheault et al. in 2020. All sequence data obtained in this study were deposited to the NCBI pathogen database under the BioProject accession number PRJNA941918 (Tables and S2) and PRJNA951760 (S2 Table). ## Quality control and genome assembly Raw read quality was assessed with FastQC (v0.11.8) (<https://www.bioinformatics.babraham.ac.uk/projects/fastqc/>). Illumina adapter sequences and low-quality sequences were trimmed using Trimmomatic version 0.36. Reads were assembled de novo with SPAdes version v3.13.0 Genome Assembler. Draft Salmonella genome assemblies were serotyped with SeqSero and by aligning with BLASTn against the nonredundant nucleotide sequence database. ## Core genome SNP typing De novo genomes were used to build core genome single nucleotide trees using the Parsnp aligner v1.2 with default parameters and the requirement for all genomes to be included in the analysis. The 171 Salmonella genomes from the survey were mapped to the complete reference genome Salmonella Typhimurium str. LT2 (NCBI accession: AE006468), resulting in an alignment of 50% of the core genome. The whole-genome phylogeny was constructed with FastTree2. iTOL (<http://itol.embl/de/>) was used to visualize the tree and annotate it with the AMR genes. ## Genetic distance Genetic distance between multiple isolates of the same serovar was evaluated as the number of SNP differences detected with the CFSAN SNP Pipeline. The CFSAN SNP Pipeline v2.0.2 was installed on a local ubuntu platform with all the executable software dependencies. Prior to analyzing our data, we used the data set provided for testing the reproducibility of the software to confirm that our installation of the CFSAN SNP Pipeline was correct. Reference-based alignments were created for a set of samples and used to generate the SNP matrix. Because the software was developed for closely related genome sequences, where available, complete assembled reference genomes were downloaded from NCBI (20; S4 Table). The phylogenies were inferred with MEGA7 using the neighbor-joining method based on the obtained SNP matrix formatted as a FASTA file generated by the CFSAN SNP pipeline v2.0.2. ## Virulence and antibiotic resistance (AMR) gene prediction The presence of resistance genes, as well as point mutations, were determined using ResFinder 4.1 (Center for Genomic Epidemiology, <https://cge.food.dtu.dk/services/ResFinder>) with a setting threshold of 90% and minimum length of 60%. The assembled draft genomes for the survey isolates (171) were used as an input to identify Salmonella AMR genes associated with resistance to aminoglycoside, β-lactam, chloramphenicol, colistin, fluoroquinolone, fosfomycin, glycopeptide, macrolide, sulfonamide, tetracycline, and trimethoprim antibiotics. The VF analyzer pipeline was used to screen the assembled draft genomes against the Virulence Factor Database (VFDB) for potential virulence factors. ## Plasmid detection and reconstruction Plasmids from genome assemblies were typed and reconstructed using MOB-suite v3.1.0 with the default parameters. To determine the AMR and virulence genes that were plasmid-borne, the reconstructed plasmids were screened against the CARD ([https://card.mcmaster.ca](https://card.mcmaster.ca/)) and VFDB databases, respectively, using Abricate v0.5 (<https://github.com/tseemann/abricate>) with the same parameters as described above. # Results ## Phylogenetic analysis of Salmonella isolates retrieved from raw almonds Genomes were initially compared with Parsnp because the CFSAN SNP Pipeline is not recommended for relatively distant bacteria (greater than a few hundred SNP differences). A maximum-likelihood phylogeny tree was constructed with Parsnp based on alignment of the 171 Salmonella assembled genomes. All the isolates belonged to the species enterica, with most (165) belonging to the subspecies enterica ( and). Three isolates were classified as subspecies diarizonae and three as subspecies arizonae. The phylogenetic tree clustered the isolates by serovar. The serovars identified by classical serotyping matched the serovars predicted by WGS for 92% of the isolates. However, 14 isolates clustered with serotypes that differed from those to which they were initially assigned by traditional serotyping. To eliminate the possibility of manipulation errors, these isolates were resequenced and their serotype was confirmed with SeqSero and with BLASTn against the nonredundant nucleotide sequence database. Results were consistent with the initial WGS serovar prediction. Based on the core genome SNP typing, a total of 32 unique Salmonella serovars were identified. Of these, 22 Salmonella serovars were isolated two or more times between 2001 and 2013: Enteritidis (n = 16), Muenchen (n = 16), Newport (n = 15), Montevideo (n = 14), Typhimurium (n = 12), Thompson (n = 11), Give (n = 10), 1,4,,12:i- (n = 9), Oranienburg (n = 7), Heidelberg (n = 6), Infantis (n = 5), Senftenberg (n = 5), Anatum (n = 4), Cerro (n = 4), Brandenburg (n = 3), Duisburg (n = 3), Horsham (n = 3), Manhattan (n = 3), Tennessee (n = 3), Agona (n = 2), Braenderup (n = 2), and Mbandaka (n = 2). Three isolates initially identified as Salmonella Othmarschen (LJH0752, LJH1143, and LJH1276-1) clustered with Salmonella Oranienburg. Both serotypes have a similar antigenic formula, 6,7,14:m,t:-, which makes them difficult to distinguish by serological methods. Serovar 1,4,,12:i:- is a monophasic variant of Salmonella Typhimurium, and three of 14 isolates initially identified as Typhimurium (LJH0738, LJH1662-1, LJH1676) clustered with 1,4,,12:i:-. One Salmonella initially identified as 1,4,,12:i:- (LJH1144) clustered with Typhimurium. Among the five isolates that were untypeable by classical serotyping, two clustered with serovar Muenchen (LJH1151, LJH1620-1), one with Enteritidis (LJH1029), and one with Manhattan (LJH1026). A single untypeable isolate (LJH1624-1) clustered with Salmonella diarizonae (LJH1664) and was identified as Salmonella diarizonae with BLASTn and SeqSero. Fifteen unique serotypes, predicted with BLASTn and SeqSero, matched the serological serotyping. It was difficult to predict some serovars due to the limited number of representative genomes in the SeqSero database. Salmonella Zerifin LJH0724 was identified as Salmonella Istanbul with SeqSero and clustered with the Istanbul isolates in the Parsnp phylogeny tree. Conflicting serotyping results between traditional methods and WGS have been reported in previous studies. Serotyping based on WGS has been increasingly used by public health laboratories and federal agencies to replace the current standard of phenotypic serotyping. In the present study, serotype was assigned based on the WGS analysis when the serovar assignment was identical among the phylogenetic relationship, SeqSero, and BLASTn analyses. Except for a single strain of Salmonella Zerifin, all isolates (170 of 171) were assigned a serotype based on the WGS analysis. ## Genetic distance within each serovar The CFSAN SNP Pipeline was used to evaluate the genetic distance among multiple isolates within each serotype. Salmonella Enteritidis isolates clustered in three groups by phage type: PT 8, PT 9c, and PT 30. All five Salmonella Enteritidis PT 8 isolates were retrieved in 2005 from separate almond lots and had less than three SNP differences, classifying them as clonal isolates. Salmonella Enteritidis PT 9c isolated in 2005 (LJH1028) and 2010 (LJH1272), differed by 5 and 13 SNPs, respectively, from a 2004 Salmonella Enteritidis PT 9c outbreak isolate (LJH1024;). The genomes of Salmonella Enteritidis PT 30 recovered from survey almonds (eight isolates: LJH0762 \[2003\], LJH1023 \[2005\], LJH1104 \[2006\], LJH1059 \[2006\], LJH1096 \[2006\], LJH1109 \[2006\], LJH1633 \[2013\], LJH1673 \[2013\]), the 2001 outbreak-associated orchard (30 isolates; S2 Table), and a 2001 outbreak-associated almond isolate (LJH0608) were compared to Salmonella Enteritidis PT 30 genomes of clinical isolates from almond outbreaks in 2001 (12 isolates) and 2006 (four isolates) (S3 Table). Salmonella Enteritidis PT 30 isolates formed two clusters. One consisted of a single survey isolate (LJH0762), recovered in 2003, that differed from LJH0608 by 48 SNPs ( and S5 Table). All other survey and clinical isolates (n = 38) clustered in a single group with LJH0608 that differed from each other by ≤18 SNPs indicating that the isolates are from a common origin. Almond isolates from 2001 to 2013 had 2 to 13 SNP differences compared with the 2001 outbreak- associated almond isolate Salmonella Enteritidis PT 30 LJH0608 ( and S5 Table in). Although this isolate was recovered from recalled almonds in 2001, the almonds were harvested in the fall of 2000, a span of 14 years (2000–2013). The orchard isolates from 2001 to 2006 differed by 0 to 12 SNPs within their genomes and by 3 to 13 SNPs with the clinical genomes. The SNP differences ranged from zero to eight within the 12 clinical isolates from the 2001 outbreak and from one to 13 within the four clinical isolates from the 2006 outbreak. Among the clinical isolates from 2001 and 2006, the SNP differences ranged from four to 13, indicating that the isolates are from a common origin. Almond, orchard, and clinical isolates of Salmonella Enteritidis PT 30 isolated from 2001 through 2013 are closely related strains. The persistence of Salmonella Enteritidis PT 30 in an almond orchard over 6 years was reported previously. The SNP analysis confirmed the PFGE results obtained for these isolates. Almonds from the 2001 outbreak-associated orchards were purposefully excluded from the raw almond survey. Because survey samples were coded and the sources unknown, it is possible that samples harvested from outbreak-associated orchards were inadvertently included. It is also possible that survey almonds were cross contaminated with almonds harvested from outbreak-associated orchards via harvest equipment or at a common almond huller-sheller, or that Salmonella Enteritidis PT 30 was spread over a broader geographic region than recognized as associated with the 2001 outbreak. Salmonella Enteritidis was not the only serovar for which clonal isolates were recovered from almonds in different years. Salmonella Montevideo survey isolates clustered into three groups, with more than 100 SNPs between them. Within each of these clusters there were isolates separated by one or more years (including isolates from 2001 and 2013) that differed by ≤9 SNPs, indicating that they share a common ancestor. Salmonella Newport (S6 Table) and Salmonella Muenchen (S7 Table) each clustered into two groups separated by more than 100 SNPs. A small number of SNP differences (\<3) between Salmonella Newport genomes (S6 Table) were identified in isolates retrieved in 2003 (LJH0751), 2006 (LJH1084), 2007 (LJH1133), 2010 (LJH1248, LJH1273 and LJH1278). Nine isolates of Salmonella Muenchen were retrieved in 2006; eight of these isolates from six different almond lots had nearly identical genomes, with ≤3 SNP differences (S7 Table), and differed from single isolates in 2007 (LJH1135) and 2013 (LJH1661) by ≤8 SNPs. Closely related genomes for single isolates of Salmonella Anatum (S8 Table) and Salmonella Thompson (S9 Table) were identified 2 and 1 years apart, respectively. Several Salmonella Oranienburg (S10 Table) and Salmonella Typhimurium were identified 5 and 3 years apart, respectively. Closely related genomes (≤13 SNPs) were identified for isolates retrieved from separate almond lots during the same year for Salmonella serovar Cerro (S11 Table), Give (S12 Table), Infantis (S13 Table), Heidelberg (S14 Table) and Tennessee (S15 Table). For Salmonella serovar Brandenburg (S16 Table in S1 File), Braenderup, Manhattan (S17 Table in S1 File), Mbandaka, and Senftenberg (S18 Table), isolates were separated by more than 13 SNPs. Multiple SNP-based approaches have been developed to analyze the large number of short reads produced by various sequencing platforms. The CFSAN Pipeline was selected because it is less sensitive to coverage changes and has good discriminative power. Its high resolution, however, strongly depends on an appropriate reference genome. In addition to outbreak source attribution, this tool can reveal similarity among isolates and their persistence in food facilities or in the production environment. Differences of more than 100 SNPs were sometimes observed among isolates within each serovar. Survey isolates that were genetically similar (≤13 SNPs) were recovered in multiple years, up to 13 years for Salmonella Enteritidis PT 30 and Salmonella Montevideo and 10 years for Salmonella Newport. Because the survey samples were collected after hulling and shelling and prior to entering a processing facility or storage, the contamination source would have to be at production or harvest (orchard), during postharvest handling (transportation to huller, during storage, or during hulling and shelling), or transportation from huller to processor. Although the exact geographic locations of the almond survey samples were unknown, clonal strains of Salmonella Enteritidis PT 8 (five; 2005) and Salmonella Muenchen (eight; 2006) were isolated from different lots of almonds in the same year. The diversity of serovars and unique strains within serovars was expected in a survey that likely reflects broad environmental contamination (e.g., almond orchard during production or harvest). Common harvest and postharvest practices may also lead to distribution of Salmonella in almonds from geographically diverse orchards that share the same equipment or facilities. At maturity, almonds are shaken to the ground where they dry for several days. They are then harvested by windrowing and sweeping off the orchard floor. The in-hull, inshell almonds are then transported to facilities where the hull and shell are removed. Kernels mix with hulls and shells before sorting, separation, and bulk transportation to processing facilities. Once kernels are delivered to almond processing facilities, commingling of almond lots may occur prior to or during storage. These practices may explain the clusters of Salmonella Enteritidis PT 8 and Salmonella Muenchen isolated from different lots in 2005 and 2006. U.S. regulations were implemented in 2007 that require all California-grown almonds sold in North America (U.S., Canada, and Mexico) to be processed with a treatment capable of achieving a minimum 4-log reduction in Salmonella. While there have been outbreaks associated with almond-containing products such as blended nut butters, none have been associated with contaminated almonds since 2006, likely due to effective implementation of these regulations. The persistence of Salmonella has been described for other pre- and postharvest scenarios. A narrow range of Salmonella serovars has been associated with California pistachio outbreaks, outbreak investigations, and industry and retail surveys. Pistachio-associated isolates of Salmonella Senftenberg and Salmonella Montevideo recovered over multiple years (2009 to 2017) and from multiple facilities differed (within each serovar) by 0 to 31 SNPs, and the authors suggested that the organisms may have established residence in the primary production environment or orchards. ## Plasmid prediction and characterization Plasmids contribute significantly to the emergence and spread of genes encoding AMR, virulence, and other metabolic functions in multiple scales across Salmonella serotypes. Plasmid carriage in the Salmonella strains under study were assessed by screening and reconstructing the plasmid sequences from assembled genomes using the clustered plasmid reference database-based pipeline. A total of 106 of the 171 Salmonella isolates (62%) carried one to seven plasmids (total plasmids 161;) with sizes of 1,030 bp to 303,322 bp ( and S19 Table). Using the presence of relaxase and mate-pair formation marker genes and/or oriT sequence, most of these plasmids were predicted to be either mobilizable (n = 51/161; 32%) or conjugative (n = 83/161; 52%). In total, 61 plasmid clusters with 27 different plasmid families were identified, with IncFII and IncFIB being the most predominant types in the collection. Thirty eight Salmonella strains distributed across six serovars (Duisburg, Enteritidis, Lomalinda, Muenchen, Typhimurium, and 1,4,,12:i:-) carried plasmids that contained two to seven virulence genes, but no AMR genes (S20 Table). The IncFIB-IncFII plasmid family combination was common among these strains, predominantly in Salmonella Enteritidis (n = 15), Typhimurium (n = 8), and 1,4,,12:i:- (n = 9). Comparative analysis of the plasmid sequences with known plasmids using mash and BLASTn revealed that these plasmids are similar (nucleotide sequence homology = 100%) to plasmid pCFSAN076214_2 (accession number CP033342.1) described previously in Salmonella Enteritidis strain ATCC BAA-1045 isolated from raw almonds (also LJH608 in this study) and to plasmid p11-0972.1 (accession number CP039855.1) reported in S. enterica serovar 1,4,,12:i:- recovered from a human stool sample. IncF plasmids are among the most common plasmids found in Salmonella and are reported to carry multiple antibiotic resistance and/or virulence genes, suggesting their role in the dissemination of these genes across Salmonella serotypes and Enterobacteriaceae by extension. Thirteen Salmonella strains belonging to seven serotypes (Agona, Anatum, Heidelberg, Istanbul, Newport, Typhimurium, and Zerifin) carried plasmids with at least one AMR gene but no virulence genes. The most common among these was within the IncC plasmid family and was associated with six to 10 AMR genes (S20 Table). The IncC plasmid identified in the present study was predicted to be conjugative, predominant in Salmonella Anatum and Salmonella Newport and was almost indistinguishable (99.99% by BLASTn) from Salmonella Anatum plasmid pSAN1-1736 (accession number: CP014658.1). IncC plasmids are known to be widely distributed across Salmonella serotypes from diverse sources and often carry multiple AMR genes. A unique Salmonella Kentucky strain LJH1044 carried a large conjugative plasmid (146 Kb) with IncFIB-IncFIC-rep_cluster_2244 plasmid replicons and contained multiple virulence and AMR genes ( and S20 Table in). This plasmid carried three AMR genes encoding resistance to aminoglycoside and tetracycline, a complete iroBCDEN operon that encodes salmochelin siderophore, and an iucABCD-iutA operon that has been described to be associated with aerobactin synthesis essential for virulence and stress response in Salmonella (S20 Table). Comparative analysis showed that this plasmid had 99.99% nucleotide sequence similarity to plasmid pCVM29188_146 (accession number CP001122.1) reported previously in several Salmonella Kentucky strains from poultry in the United States. ## Antimicrobial resistance profile Among the isolates retrieved from the almond survey, a total of 24 AMR genes were identified with the ResFinder and CARD database, which classified into nine different antimicrobial protein groups: aminoglycosides, β-lactams, colistin, fosfomycin, glycopeptide, phenicol, sulfonamide, tetracycline, and trimethoprim. One gene, aac(6’)-Iaa, which confers resistance to aminoglycosides, was detected in the chromosome of all Salmonella subspecies diarizonae (3) and enterica (165) but not in arizonae (3). The frequency of Salmonella enterica subspecies enterica isolates that carried an aminoglycoside acetyltransferase aac(6’)-Iaa has been reported to be high (\>95%) in multiple WGS analysis. However, in a surveillance study of non-typhoidal Salmonella enterica, 11 isolates out of 3,491 (0.3%) showed phenotypic resistance to an aminoglycoside antimicrobial. The other 23 AMR genes were detected in 35 isolates, with fosA7 being the most common. The gene fosA7, which confers resistance to fosfomycin, a broad- spectrum cell wall synthesis inhibitor, was first identified in the chromosome of Salmonella Heidelberg isolated from broiler chickens. In the present study, all Salmonella Heidelberg isolates (n = 6), and some isolates of Salmonella serovars Agona (n = 2), Meleagris (n = 1), Montevideo (n = 8), Oranienburg (n = 2), and Tennessee (n = 3), carried the chromosomal fosA7 gene. Ten aminoglycoside resistance genes were detected in 16 isolates: aac(6’)-Iaa, aadA2, aadA3, aadA7, aadA12, aadA13, ant(2”)-Ia, aph(3")-IIa, aph(3’’)-Ib, aph(6)-lc, and aph(6)-Id. The bla<sub>CMY-2</sub> and bla<sub>CARB-2</sub> genes, which confer resistance to β-lactams, and the floR gene, which confers resistance to phenicol, were found in nine isolates. Three tetracycline efflux resistance genes were identified in 16 isolates: tetA, tetB, and tetG. Sulfisoxazole resistance, encoded by sul1 or sul2, was detected in 11 isolates. The dihydrofolate reductase resistance gene, dfrA12, which confers resistance to trimethoprim, was detected in one Salmonella Newport isolate. The plasmid- mediated colistin resistance and phosphoethanolamine transferase mcr-9.1 gene was detected in one Salmonella Agona isolate. Resistance to bleomycin, encoded by the bleomycin-binding protein gene (ble), was predicted for one Salmonella Heidelberg isolate. All the isolates, except Salmonella serovars Enteritidis, Irumu, Lomalinda, Typhimurium, and 1,4,,12:i-, contained a missense mutation in parC associated with resistance to quinolone. Multiple mutations in the quinolone resistance determining region are usually required to confer resistance to ciprofloxacin, but one mutation confers resistance to nalidixic acid. One Salmonella Senftenberg isolate had a point mutation in the 16S rRNA that is thought to confer resistance to spectinomycin. Most of the predicted AMR genes were identified in a small number (16 out of 171; 9%) of survey isolates and were plasmid encoded in 11 of 16 cases. Multidrug-resistant isolates (putative resistance to at least one antibiotic in three or more drug classes; <https://www.cdc.gov/narms/resources/glossary.html>) were identified among Salmonella serovars Agona (LJH1100 and LJH1106), Anatum (LJH0720, LJH0787, LJH1012, and LJH1063), Heidelberg (LJH1622), Newport (LJH0656, LJH1138, and LJH1107), and Typhimurium (LJH0788 and LJH1141). Antibiotic resistance profiles by the calibrated dichotomous sensitivity method were determined for Salmonella isolated from 2001 through 2005 but not for isolates from 2006, 2007, 2010, and 2013. Ten of the Salmonella survey isolates from 2001 to 2005 were resistant to three or more antibiotics. Resistance genotype and phenotype correlated highly for five of these isolates: Salmonella Anatum (n = 3), Salmonella Istanbul (n = 1), and Salmonella Typhimurium var. Copenhagen (n = 1). ## Analysis of virulence factors Salmonella serovars infect a wide range of hosts with different degrees of disease severity, with Enteritidis, Newport, Typhimurium, Javiana, and 1,4,,12:i- being significantly more likely to cause illness in humans in the United States. Differences in virulence factors contribute to the severity and outcome of salmonellosis and can be specific to serovars. The Virulence Factor Database (VFDB) was used to detect a total of 303 virulence genes among the 171 Salmonella assembled genomes (Figs –). Genes were classified under major virulence factors, including the secretion system, fimbrial and non fimbrial adherence, macrophage inducible genes, magnesium uptake, serum resistance, stress proteins, toxins, immune invasion, and two component regulatory systems. The Salmonella pathogenicity island 1 (SPI-1) and 2 (SPI-2), responsible for the type III secretion system, are ubiquitous in S. enterica subsp. enterica and were common to all 171 survey isolates. Among the fimbrial adherence factors, the genes encoding the curli fimbriae, csgA, csgB, and csgE, the bcfABCDEFG operon, and the genes that encode type 1 fimbriae, fimA, fimB, fimC, fimD, fimH, fimI, fimW, and fimZ, were also present in all the isolates. The two- component regulatory system phoP-PhoQ genes and the magnesium uptake genes, mgtB-mgtC, (part of SPI-3) were present in all the isolates. All the isolates had the microphage inducible gene, mig-14, but mig-5 was mainly detected in isolates of Salmonella serovars Enteritidis, Typhimurium, and 1,4,12:I:-. The typhoid toxin genes, cdtB and pltA, originally identified in serotype Typhi, were found in all Salmonella serovar Brandenburg, Duisburg, Horsham, Montevideo, Sandiego, and Schwartzengrund isolates and in nine of 10 Salmonella Oranienburg isolates. However, pltB, required for forming holotoxin, was not present in any of these isolates. In Salmonella Horsham isolates, two genes were identified as homologs of the enterotoxin hemolysin genes of Escherichia coli (hylE/clyA). A virulence plasmid that harbored a combination of pef, rck, and spv virulence genes was identified in 23% of the isolates (S20 Table in). The assembly of the major Pef fimbriae depends on the pefBACDorf5orf6 operon, which encodes PefA fimbriae subunit, PefC usher protein, and the pefD periplasmic chaperone. Eight of 10 Salmonella Typhimurium and all nine Salmonella 1,4,12:I:- isolates carried a plasmid encoding pefA, pefB, pefC, and pefD genes ( and S20 Table). The spv genes play a role in suppression of the innate immune response and are often associated with invasive disease and increased virulence. The spv genes were detected in all Salmonella serovar Lomalinda, Enteritidis, and 1,4,12:I:- isolates, in eight of 10 Salmonella Typhimurium isolates, and one of 10 Salmonella Give isolates ( and S20 Table). The rck gene, which provides protection against the complement-mediated immune response of the host, was found in one of 10 Salmonella Give isolates, in all Salmonella Enteritidis PT 30 and PT 8 isolates, in eight of 10 Salmonella Typhimurium isolates, and all Salmonella 1,4,12:I;- isolates. This study provides one of the first in-depth longitudinal characterizations of Salmonella strains isolated from a single product (almonds) or production environment (almond orchard), in a single geographical region (Central California). This isolate collection is important for understanding Salmonella populations in a significant food production region of the United States. Several clonal strains of Salmonella were isolated over multiple years, adding to a growing body of evidence that enteric pathogens may persist over long periods of time (years) in agricultural environments and in postharvest or food processing facilities (<https://www.cdc.gov/ncezid/dfwed/outbreak-response/rep- strains.html>)). # Supporting information We thank Sylvia Yada for editing the manuscript and Vanessa Lieberman for technical assistance. 10.1371/journal.pone.0291109.r001 Decision Letter 0 Karunasagar Iddya Academic Editor 2023 Iddya Karunasagar This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 27 Jun 2023 PONE-D-23-14605Genetic diversity of * Salmonella enterica * isolated from raw California almonds and from an almond orchard over 13 YearsPLOS ONE Dear Dr. Harris, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Aug 11 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Iddya Karunasagar Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Thank you for stating the following in the Acknowledgments Section of your manuscript: "This research was supported by the Almond Board of California Grant 18-HarrisL536 AQFSS-01. The DNA Technologies and Expression Analysis Core at UC Davis Genome Center is supported by NIH shared Instrumentation Grant 1S10OD010786-01. L. Goodridge and RC Levesque were funded by Genome Canada and Genome Québec. We thank Sylvia Yada for editing the manuscript and Vanessa Lieberman for technical assistance." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: "LJH 18-HarrisL-AQFSS-01 Almond Board of California (<https://www.almonds.com>) LG/RCL Genome Canada (<https://genomecanada.ca>) and Genome Quebec (<https://www.genomequebec.com/en/home/>) The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 3\. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. 4\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Additional Editor Comments: Please see the reviewer comments. Please revise addressing all comments point by point \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: N/A \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: PONE-D-23-14605 Genetic diversity of Salmonella enterica isolated from raw California almonds and from an almond orchard over 13 years This study compared the whole genome sequences of 171 Salmonella isolates that included 30 isolates of S. Enteritidis PT30 involved in almond-associated outbreaks. SNP analysis, serovar identification, antimicrobial resistance genes, plasmid types and virulence genes were determined. The main goal of this study was to see if S. Enteritidis PT30 outbreak isolates could be distinguished from similar environmental isolates. The study emphasizes the discriminatory power of WGS over traditional genotyping methods. The study identified clonal isolates of Salmonella Enteritidis and Salmonella Montevideo. The persistence of clonal strains of certain serovars suggest contamination of post-harvest processing areas with these strains. 1\. The manuscript is straightforward presentation of the WGS comparison. The study assumes significance in few contexts; that Salmonella Enteritidis PT 30 isolates belong to a long time line of over 13 years, comprise of outbreak strains and their environmental counterparts, and as the result suggest, these strains were phylogenetically closely related. The study aimed to solve three important issues with the traditional genotyping methods as stated in the introduction section and listed below. 2\. Lines 80-82:. “However, neither PFGE nor MVLA could discriminate among the Salmonella Enteritidis PT 30 isolates associated with the 2001 and 2006 almond- associated outbreaks” Now, the WGS analysis clustered (lines 275-276) clustered them together in one group with ≤18 SNPs. Is this not similar to PFGE results? Or does a difference of 18 SNPs make them different from each other? 3\. Lines 82-84: Salmonella Enteritidis PT 30 almond associated outbreak strains could not be distinguished from epidemiologically unrelated Salmonella Enteritidis PT 30 clinical strains included in the study Again, the fact that the WGS analysis put them together in one group suggests the same thing. 4\. Lines 87-89: Environmental Salmonella Enteritidis PT 30 isolates, collected between 2001 and 2006 from one of the 2001 outbreak-associated orchards, clustered in two groups based on the separation by PFGE of their XbaI-digested DNA As could be seen from the results in lines 274-276, one survey isolate formed a separate group. 5\. Lines 294-295: With less than ≤18 SNPs, are they still termed “closely related strains”. In a �4 Mb genome sequence, what would be the NGS contribution to SNPs? Although this depends on the depth of sequencing, still there would be contribution from the sequencing itself. 6\. Although AMR genes were identified from the WGS, these are not correlated with the antibiotic resistance phenotype of isolates. 7\. Only two pathogenicity islands (SPI-1 & SPI-2) were identified in WGS. What about the rest? \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \\\\\\\\\\ \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0291109.r002 Author response to Decision Letter 0 4 Aug 2023 PONE-D-23-14605 Genetic diversity of Salmonella enterica isolated over 13 years from raw California almonds and from an almond orchard Editor and Reviewer comments are in italics. Author responses are not in italics. Please submit your revised manuscript by Aug 11 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: Editor c1. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. Response: Done Editor c2. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. Response: Done. Editor c3. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. Response: Done. Journal Requirements: When submitting your revision, we need you to address these additional requirements. JR 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> JR 2. Thank you for stating the following in the Acknowledgments Section of your manuscript: "This research was supported by the Almond Board of California Grant 18-HarrisL536 AQFSS-01. The DNA Technologies and Expression Analysis Core at UC Davis Genome Center is supported by NIH shared Instrumentation Grant 1S10OD010786-01. L. Goodridge and RC Levesque were funded by Genome Canada and Genome Québec. We thank Sylvia Yada for editing the manuscript and Vanessa Lieberman for technical assistance." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: "LJH 18-HarrisL-AQFSS-01 Almond Board of California (<https://www.almonds.com>) LG/RCL Genome Canada (<https://genomecanada.ca>) and Genome Quebec (<https://www.genomequebec.com/en/home/>). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." Please include your amended statements within your cover letter; we will change the online submission form on your behalf. Response: We have deleted the funding information from the Acknowledgement section in the manuscript. Amended Funding Statement: “LJH 18-HarrisL-AQFSS-01 Almond Board of California (<https://www.almonds.com>), LG/RCL Genome Canada (<https://genomecanada.ca>) and Genome Quebec (<https://www.genomequebec.com/en/home/>), NIH shared Instrumentation Grant 1S10OD010786 (<https://orip.nih.gov/construction-and- instruments/s10-instrumentation-programs>). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." JR 3. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. Response: We included all relevant accession numbers in the manuscript and have now released those data. JR 4. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Response: Reference list has been reviewed and is complete and correct. Reviewer \#1: This study compared the whole genome sequences of 171 Salmonella isolates that included 30 isolates of S. Enteritidis PT30 involved in almond-associated outbreaks. SNP analysis, serovar identification, antimicrobial resistance genes, plasmid types and virulence genes were determined. The main goal of this study was to see if S. Enteritidis PT30 outbreak isolates could be distinguished from similar environmental isolates. The study emphasizes the discriminatory power of WGS over traditional genotyping methods. The study identified clonal isolates of Salmonella Enteritidis and Salmonella Montevideo. The persistence of clonal strains of certain serovars suggest contamination of post-harvest processing areas with these strains. The manuscript is straightforward presentation of the WGS comparison. The study assumes significance in few contexts; that Salmonella Enteritidis PT 30 isolates belong to a long timeline of over 13 years, comprise of outbreak strains and their environmental counterparts, and as the result suggest, these strains were phylogenetically closely related. The study aimed to solve three important issues with the traditional genotyping methods as stated in the introduction section and listed below. Response: Thank you for your comments and manuscript overview. As stated in the last paragraph in the introduction, the objective of the study was a comparative genomic analysis of almond-related Salmonella isolates. The study did not aim to solve issues with traditional genotyping methods. However, we have addressed each comment related to statements in the introduction below. R1c1a Lines 80-82: “However, neither PFGE nor MVLA could discriminate among the Salmonella Enteritidis PT 30 isolates associated with the 2001 and 2006 almond- associated outbreaks”. Now, the WGS analysis clustered (lines 275-276) clustered them together in one group with ≤18 SNPs. Is this not similar to PFGE results? Or does a difference of 18 SNPs make them different from each other? R1c1b Lines 82-84: Salmonella Enteritidis PT 30 almond associated outbreak strains could not be distinguished from epidemiologically unrelated Salmonella Enteritidis PT 30 clinical strains included in the study. Again, the fact that the WGS analysis put them together in one group suggests the same thing. R1c1c Lines 87-89: Environmental Salmonella Enteritidis PT 30 isolates, collected between 2001 and 2006 from one of the 2001 outbreak-associated orchards, clustered in two groups based on the separation by PFGE of their XbaI- digested DNA. As could be seen from the results in lines 274-276, one survey isolate formed a separate group. Response: The statement regarding the PFGE or MVLA was specific to clinical isolates of Salmonella Enteritidis PT 30 which could not be distinguished from each other using PFGE or MVLA (refers to Parker et al. \[8\]). With one exception (LJH0762), we were also unable to distinguish the Salmonella Enteritidis PT 30 isolates assessed in the present paper by WGS (beyond 18 SNP differences). Our analysis consisted of 55 isolates: eight almond survey isolates, 30 orchard isolates and 16 outbreak isolates. We did not include some of the non-almond associated clinical isolates that were evaluated in the Parker et al \[8\] study and thus are unable to know if WGS could distinguish among all the isolates used in that study. To clarify, we made minor modification to the concluding paragraph of the introduction to specify that 171 isolates (45 serovars) were evaluated (not just Salmonella Enteritidis PT 30). We provided additional information under “Genetic Distance Within Each Serovar” (copied below) to highlight the number of Salmonella Enteritidis PT 30 isolates that were assessed. We added a new supplemental table (S5 Table) that complements Figure 2 and provides more granularity to SNP differences. We also further clarified that almond isolates from 2001 to 2013 differed from the 2001 outbreak-associated almond isolate LJH0608 (from almonds harvested in 2000) by 2 to 13 SNPs and that this represents isolates collected over a span of 14 years. We also modified the following sentence to remove the statement “with higher resolution”: “The persistence of Salmonella Enteritidis PT 30 in an almond orchard over 6 years was reported previously \[2\]. The SNP analysis confirmed the PFGE results obtained for these isolates.” The reviewer’s statement (R1c1c) regarding environmental Salmonella isolates was related to orchard isolates (refers to Uesugi et al., 2006 \[2\]). The comment refers to only the environmental samples retrieved from the outbreak-associated almond orchard. The orchard isolates from 2001 to 2006 differed by 0 to 12 SNPs in our analyses, which is now more clearly stated in the modified results section (copied below). We believe that the changes outlined above adequately address the reviewer’s comments. “The genomes of Salmonella Enteritidis PT 30 recovered from survey almonds (eight isolates (Table 1): LJH0762 \[2003\], LJH1023 \[2005\], LJH1104 \[2006\], LJH1059 \[2006\], LJH1096 \[2006\], LJH1109 \[2006\], LJH1633 \[2013\], LJH1673 \[2013\]), the 2001 outbreak-associated orchard (30 isolates; S2 Table), and a 2001 outbreak-associated almond isolate (LJH0608) were compared to Salmonella Enteritidis PT 30 genomes of clinical isolates from almond outbreaks in 2001 (12 isolates) and 2006 (four isolates) (S3 Table). Salmonella Enteritidis PT30 isolates formed two clusters (Fig 2). One consisted of a single survey isolate (LJH0762), recovered in 2003, that differed from LJH0608 by 48 SNPs (Fig 2, S5 Table). All other survey and clinical isolates (n = 38) clustered in a single group with LJH0608 that differed from each other by ≤18 SNPs (Fig 2) indicating that the isolates are from a common origin. Almond isolates from 2001 to 2013 had two to 13 SNP differences compared with the 2001 outbreak-associated almond isolate Salmonella Enteritidis PT 30 LJH0608 (Fig 2, S5 Table). Although this isolate was recovered from recalled almonds in 2001, the almonds were harvested in the fall of 2000 \[1\], a span of 14 years (2000–2013). The orchard isolates from 2001 to 2006 differed by 0 to 12 SNPs within their genomes and by 3 to 13 SNPs with the clinical genomes. The SNP differences ranged from zero to eight within the 12 clinical isolates from the 2001 outbreak and from one to 13 within the four clinical isolates from the 2006 outbreak. Among the clinical isolates from 2001 and 2006, the SNP differences ranged from four to 13.” R1c2 Lines 294-295: With less than ≤18 SNPs, are they still termed “closely related strains”. In a �4 Mb genome sequence, what would be the NGS contribution to SNPs? Although this depends on the depth of sequencing, still there would be contribution from the sequencing itself. Response: There isn’t a standard cutoff for SNP differences with respect to strain separation (Brown et al. 2019) (9). Parameters defined in the CFSAN pipeline for filtering the SNPs should minimize the contribution of NGS to SNPs. The CFSAN pipeline constructs a high-quality SNP matrix for closely related sequences (\<100 SNP differences) with a higher recovery rate of SNPs for datasets with 100x coverage compared to 20x, as described by Davis et al., 2015 (12). Our genomes were sequenced with an average coverage of 50x. We believe that modifications to this section (as described above) have addressed this question. R1c3 Although AMR genes were identified from the WGS, these are not correlated with the antibiotic resistance phenotype of isolates. Response: Throughout the section on AMR we specifically note the WGS refers to genotype and not phenotype. However, in the last paragraph of the “Antimicrobial resistance profile” section we discuss the information that is available on genotypic vs phenotypic expression of AMR genes in some of the isolates associated with the study (copied below). We believe this addresses the reviewer concerns and have not made changes to the manuscript. “Antibiotic resistance profiles by the calibrated dichotomous sensitivity method were determined for Salmonella isolated from 2001 through 2005 but not for isolates from 2006, 2007, 2010, and 2013. Ten of the Salmonella survey isolates from 2001 to 2005 were resistant to three or more antibiotics \[5\]. Resistance genotype and phenotype correlated highly for five of these isolates: Salmonella Anatum (n = 3), Salmonella Istanbul (n = 1), and Salmonella Typhimurium var. Copenhagen (n = 1).” R1c4 Only two pathogenicity islands (SPI-1 & SPI-2) were identified in WGS. What about the rest? Response: SPI-1 and SPI-2 were present in all the isolates and we made note of this. Genes from other known pathogenicity islands were present but we did not assess the presence of complete islands in every isolate. The magnesium uptake genes (mgtB-mgtC) were present in all isolates (line 502) and are part of SPI-3. The hlyE hemolysin (lines 511) is part of SPI-18 and was identified in S. Horsham. 10.1371/journal.pone.0291109.r003 Decision Letter 1 Karunasagar Iddya Academic Editor 2023 Iddya Karunasagar This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 22 Aug 2023 Genetic diversity of * Salmonella enterica * isolated over 13 years from raw California almonds and from an almond orchard PONE-D-23-14605R1 Dear Dr. Harris, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Iddya Karunasagar Academic Editor PLOS ONE Additional Editor Comments (optional): All reviewer comments have been addressed. Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The manuscript has been sufficiently revised, and all my querries in the previous review have been addresssed. The authors have made relevant changes in the revised manuscript. \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \\\\\\\\\\ 10.1371/journal.pone.0291109.r004 Acceptance letter Karunasagar Iddya Academic Editor 2023 Iddya Karunasagar This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 29 Aug 2023 PONE-D-23-14605R1 Genetic diversity of Salmonella enterica isolated over 13 years from raw California almonds and from an almond orchard Dear Dr. Harris: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Iddya Karunasagar Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction Internalizing and externalizing behavior problems are considered a substantial risk factor for students’ social and academic well-being. Both dimensions are consistently associated with difficulties on different levels in schools. Children and adolescents with externalizing and/or internalizing behavior problems are exposed to higher risks of academic failure. In addition to academic difficulties, subtypes of externalizing as well as internalizing behavior problems are associated with a range of short- and long-term developmental risks, such as higher rates of social exclusion/rejection, school suspension, lack of bonding to school, or criminal arrest. Therefore, the early and adequate identification of students at-risk for the development of severe externalizing and internalizing behavior problems in educational practice is of high relevance. The gained insights in students’ behaviors might at the same time serve as a reference for designing appropriate behavioral interventions. Teachers and school psychologists consequently play an important role in supporting and identifying students who suffer from behavioral problems. Among others, the Strengths and Difficulties Questionnaire (SDQ) is a frequently used screening instrument for behavioral and emotional problems in children and adolescents and is commonly applied in educational research and practice. The SDQ comprises 25 items that can be grouped into subscales for emotional symptoms, peer problems, conduct problems, hyperactivity, and prosocial behavior, containing five items each. The narrowband subscales for emotional symptoms and peer problems can be combined into the broadband scale internalizing behavior and the narrowband subscales of hyperactivity and conduct problems can be subsumed into the broadband scale externalizing behavior. Correlations of behavioral and emotional problems, measured by means of sum scores on the SDQ, with academic outcome variables have been addressed in recent research. Previous studies have highlighted negative associations between the SDQ scales and academic achievement. Metsäpelto et al. found that higher externalizing behavior (broadband scale) was related to lower reading skills. Similarly, Palmu et al. highlighted that externalizing behavior (broadband scale) was associated with lower academic grades. Similarly, DeVries et al. concluded that peer problems (narrowband scale) are negatively associated with academic grades. Higher internalizing and externalizing behavior (broadband scales) were linked to lower scholastic performance. Hyperactivity and conduct problems (narrowband scales) were negatively correlated with reading and mathematical skills. Not only are the associations between the SDQ scales and academic performance of interest from a pedagogical standpoint, but also their relation to social outcomes, such as social integration within the classroom, the well-being of students in their families, school absenteeism, and feelings of self-worth and self-efficacy. Higher emotional symptoms, peer problems, conduct problems, and hyperactivity (narrowband scales) are associated with a lower social preference within the classroom and with more difficulties within the family and lower feelings of self-worth. Students rejected by peers are more likely to show internalizing behavior problems (broadband scale) and controversial students are more likely to exhibit externalizing behavior (broadband scale). Higher emotional symptoms (narrowband scale) are associated with lower feelings of self-efficacy. Lenzen et al. highlighted the association of greater conduct problems and emotional symptoms (narrowband scales) with increased school absenteeism. School absence was also linked to internalizing behavior (broadband scale) while school discipline referrals were related to externalizing behavior (broadband scale). Moreover, students with special educational needs, especially those with learning disabilities, have greater peer problems (narrowband scale). Although the originally proposed 5-factor structure of the SDQ (5 narrowband subscales) is often applied in research, the narrowband subscales are sometimes subsumed into broadband behavior scales (internalizing and externalizing behavior), which has resulted in an ongoing controversy, not only in educational research, regarding the factor structure of the SDQ. This controversy implies a discussion about the usefulness of broadband and narrowband scales of behavior. # Broadband and narrowband scales of behavior Some studies on the psychometric properties of the SDQ have confirmed the original 5-factor structure (emotional symptoms, peer problems, conduct problems, hyperactivity, and prosocial behavior) proposed by Goodman. At the same time, several studies have highlighted potential model fit shortcomings of the 5-factor structure. As a reaction to these concerns, several authors have proposed a 3-factor model as a possible alternative to the original 5-factor model. In the revised 3-factor model, the narrowband subscales of emotional symptoms and peer problems are combined into the broadband scale for internalizing behavior, and the narrowband subscales of hyperactivity and conduct problems are subsumed into the broadband scale for externalizing behavior (the narrowband subscale of prosocial behavior remains unchanged). Recent research has provided partial support for the appropriateness of the revised 3-factor model. In studies comparing the different factor structures (3 vs. 5) in terms of model fit, the superiority of either the 3-factor structure or the 5-factor structure or also the adequateness of both factor structures have been documented. Goodman et al. ( p. 1189) conclude that “there may be no single best set of subscales to use in the SDQ; rather, the optimal choice may depend in part upon one’s study population and study aims.” Subsuming narrowband subscales into broadband scales takes place from a clinical perspective and represents the well-known hierarchical model of child and adolescent psychopathology. For example, the SDQ narrowband scales of hyperactivity and conduct problems describe distinctive psychopathological phenomena but are often co-occurring, which is why both scales form a broadband externalizing scale. The broadband perspective on child and adolescent behavior and emotions leads to a two-dimensional taxonomy of psychopathology distinguishing between internalizing and externalizing behavior. The question of whether child and adolescent psychopathology is best described by narrowband or broadband measures is the subject of ongoing debate: > “\[…\] the once plausible goal of identifying homogeneous populations > for treatment and research resulted in narrow diagnostic categories > that did not capture clinical reality, symptom heterogeneity within > disorders, and significant sharing of symptoms across multiple > disorders. The historical aspiration of achieving diagnostic > homogeneity by progressive subtyping within disorder categories no > longer is sensible \[…\]” ( p. 12) However, it must also be considered that different narrowband dimensions of behavior (e.g., aggressive vs. non-aggressive rule-breaking behavior) are related to different etiological factors. Vice versa, different narrowband dimensions of behavior might explain educational outcome variables to varying degrees, for example, conduct problems are associated with arithmetic skills (r =.20), while the association is stronger for hyperactivity (r =.38). Subsuming these narrowband dimensions of behavior into a broader category of behavior might therefore lead to a loss of information; that is, differentiated effects (between narrowband scales) in predicting educational outcomes are not described by a broadband scale. The usage of narrowband scales could provide a nuanced description of the association between dimensions of behavior and child and adolescent outcomes. # Aims Empirical evidence concerning the superiority of the narrowband scales of behavior in the prediction of child and adolescent outcomes is lacking. Comparisons between the narrowband scales (5-factor structure) and broadband scales (3-factor structure) with regard to the prediction of outcomes is sparse. The question remains: Which factors (narrowband vs. broadband) are better predictors? In addition, the vast majority of previous research using the SDQ has focused on parent- or teacher-reported student behavior. There is a lack of studies that examine the associations between self-reports of students on the SDQ and relevant outcomes. However, children and adolescents can be considered experts of their own well-being and consequently might depict a valid and important source of information on their own behavior. In the present study, therefore, we examined how behavioral and emotional problems, measured by means of the different self-report SDQ scores (narrowband and broadband scales), are associated with child and adolescent outcomes, such as measures of academic success (grades), well-being (school, friends, and family), and self-belief (self-esteem and self-efficacy). We thereby assume that narrowband scales of behavior are more informative predictors of outcomes than broadband scales of behavior. This comparison (narrowband vs. broadband dimensions of behavior) seems of particular importance for emphasizing the need to differentiate behavioral problems when examining associations with child and adolescent outcomes. # Method ## Study design and participants The analyzed cross-sectional sample was obtained from the baseline of the German Health Interview and Examination Survey for Children and Adolescents (KiGGS study). The KiGGS study is a nationally representative health survey comprising children and adolescents. The survey’s main objective is to obtain information on key physical and mental health indicators, risk factors, health service utilization, health behavior, and living conditions of children and adolescents in Germany. Study participants were not recruited in schools (non-nested data structure) but randomly selected from the official registers of local residents. The data were collected from 2003 to 2006. The KiGGS sample consists of 17641 children and adolescents aged 0 to 17 years. The children were given a physical examination and the parents as well as the children and adolescents themselves (from age 11 on) were interviewed via written questionnaires. The study was approved by the Charité/Universitätsmedizin Berlin ethics committee and the Federal Office for the Protection of Data. The present analysis represents a secondary data analysis with a focus on behavior/emotions and child and adolescent outcomes such as measures of well- being (school, friends, and family), academic success (grades), and self-belief (self-esteem and self-efficacy), which are factors of interest to professionals in educational contexts. We will focus on the child/adolescent-reported data and refer to children and adolescents of compulsory education age. Compulsory education in Germany usually ends with the completion of grade 9 (usually 15-year-old adolescents). The survey’s questionnaires were addressed to children and adolescents from the age of 11 years onwards (usually children in grade 5 and above). The present sample therefore includes children and adolescents with a minimum age of 11 years in grades 5 to 9 (N = 4642; 52% boys; age in years: M = 13.46, SD = 1.47, Min = 11.00, Q<sub>1</sub> = 12.17, Md = 13.42, Q<sub>3</sub> = 14.67, Max = 17.92). The distribution of the children and adolescents across grades 6 to 9 is nearly equal (approximately 21.5% in each grade, but 14% in grade 5). The proportion of children and adolescents in grade 5 (usually 10- and 11-year-olds) is small, as individuals under the age of 11 are not included in the present sample. ## Measures ### Child and adolescent behavior and emotions The German self-report version of the Strengths and Difficulties Questionnaire (SDQ) was used to assess child and adolescent behavior and emotions. This questionnaire (25 items) quantifies emotional symptoms, peer problems, conduct problems, hyperactivity, and, as a dimension of strength, prosocial behavior (original narrowband scales). Emotional symptoms and peer problems can be combined into a broadband internalizing behavior subscale, while conduct problems and hyperactivity can be subsumed into a broadband externalizing behavior subscale. The prosocial behavior subscale was not considered in the present study, because the focus was on a comparison of the two broadband subscales with the underlying narrowband subscales. SDQ items are rated on a three-point scale (0 for "not true," 1 for "somewhat true," and 2 for "certainly true"). High subscale scores indicate elevated behavioral problems. The children and adolescents with the highest subscale scores, which are the upper 10% of the normative sample, can be categorized as “abnormal” and are considered to be at risk for psychiatric disorders. We therefore refer to these individuals as at-risk children and adolescents. A conservative classification rule, which minimizes false positive cases by selecting a cutoff value below 10%, was used to identify at-risk children and adolescents (emotional symptoms ≥ 6, peer problems ≥ 5, conduct problems ≥ 5, hyperactivity ≥ 7, internalizing behavior ≥ 9, and externalizing behavior ≥ 10; calculations based on KiGGS baseline data,). With regard to the self-report version used in the KiGGS study (data at hand), the internal consistencies for the subscales of peer problems and conduct problems are insufficient (α and ω ≤.50), but moderate for the subscales of emotional symptoms, hyperactivity, and internalizing and externalizing behavior (α and ω ≥.60). ### Child and adolescent outcomes Health-related quality of life. The self-report version of the KINDL-R is a brief questionnaire to measure the health-related quality of life of children and adolescents. The subscales school (e.g., “doing my schoolwork was easy”), friends (e.g., “I played with friends”), self-esteem (e.g., “I was proud of myself”), and family (e.g., “I got on well with my parents”) describe the students’ well-being related to daily school life, friendship, family life, and feelings of self-worth. Each subscale consists of 4 items. Items are rated on a five-point scale (1 for “never,” 2 for “seldom,” 3 for “sometimes,” 4 for “often,” and 5 for “all the time”). High scores indicate a positive quality of life in the specific domain. With regard to the KiGGS study (data at hand), the internal consistencies of the mentioned subscales are mediocre (α and ω range from.53 to.69). The subscales physical and emotional well-being are not used in the present study. School grades. The school grades (math and German) received on the last report card (half-year term) were reported by the parents. Germany uses a 6-point grading scale. School grades vary from 1 (excellent) to 6 (insufficient), which were reversed so higher values indicate a better academic performance. General self-efficacy. The general self-efficacy scale is a 10-item questionnaire that was designed to assess optimistic self-beliefs in coping with a variety of difficult demands in life (e.g., “I can always manage to solve difficult problems if I try hard enough”). In the KiGGS study, the scale was only used in adolescents aged 14 years and older (N = 1750). Items are rated on a four-point scale (1 for “not at all true,” 2 for “hardly true,” 3 for “moderately true,” and 4 for “exactly true”). High scores indicate stronger self-efficacy. With regard to the KiGGS study (data at hand), the internal consistency of the scale is good (α and ω \>.80). ## Statistical analysis strategy Ordinary least square regression models will be formulated with regard to the prediction of child and adolescent outcomes (outcomes regressed on SDQ subscales). The term “prediction” and cognate terms are used here in a statistical sense and shall not be confused with the concept of predictive validity, which describes the ability of a measure to forecast outcomes in the future. To judge the statistical predictive performance of the different subscales of the SDQ (broadband vs. narrowband), the regression model with the broadband subscale (model 1: outcome regressed on broadband subscale, e.g., internalizing behavior) will be compared to the regression model with both underlying narrowband subscales jointly as predictors in one regression model (e.g., model 2: outcome regressed on emotional symptoms and peer problems). This model comparison (narrowband vs. broadband) will be conducted with regard to each predicted outcome and separately for internalizing and externalizing behavior (internalizing behavior vs. emotional symptoms and peer problems; externalizing behavior vs. conduct problems and hyperactivity). To evaluate the predictive performance of the different models (predictive performance of the broadband and narrowband subscales), we report two goodness-of-fit indices for each regression model. The adjusted R<sup>2</sup> is the proportion of variance in the outcome that is predictable from the predictors (SDQ subscales). The Akaike Information Criterion (AIC) takes into account both model complexity (total number of estimated model parameters) and goodness of model fit (maximized likelihood) and balances these two. The individual AIC values are not interpretable. However, the smaller the AIC value, the better the model fit. Consequently, models with less complexity (fewer predictors) along with a high goodness of fit are deemed to be good models. If the difference in AIC values between the models is less than 3 (model with broadband scale vs. model with underlying narrowband scales), then the model with the higher AIC value is almost as good as the model with the smaller AIC value. For the application of AIC model selection in the fields of psychology and psychometrics, see Vrieze. The SDQ subscales are used as dummy variables. The reference is the at-risk category (“abnormal”). Therefore, the intercept (constant) of each regression model is interpretable as the expected average outcome for the at-risk children and adolescents. Since all the outcomes are standardized (M = 0, SD = 1), the intercept represents the average outcome for the at-risk group as a deviation from the overall sample mean in units of standard deviation. The regression parameters (B) for all the other SDQ subscale scores are interpretable as the average difference in the outcomes (in units of standard deviation) between the at-risk group and the children and adolescents with the particular SDQ subscale score. These types of analyses emphasize the clinical category “abnormal” (at-risk). In some additional regression analyses, we will use the SDQ subscales as continuous predictors. If the SDQ subscale is a continuous predictor, the intercept represents the average outcome for children without behavioral problems (SDQ subscale score equals zero) as a deviation from the overall sample mean in units of standard deviation, and the regression parameter (B) is the average change (slope) of the outcome (in units of standard deviation) when the SDQ score increases on average by one unit. All statistical analyses were conducted in R 3.6.0. # Results ## Preliminary results ### Model fit and measurement invariance of the 3- and 5-factor structures of the SDQ Confirmatory factor analyses (weighted least square mean and variance adjusted estimation) reveal an appropriate model fit (RMSEA \< 0.08, for details see) for both the 3- and 5-factor structures of the SDQ, although the 5-factor structure shows a better model fit (RMSEA =.05, CFI =.89, TLI =.88, χ<sup>2</sup> = 3250.87, df = 265, p =.00) than the 3-factor structure (RMSEA =.06, CFI =.85, TLI =.83, χ<sup>2</sup> = 4540.10, df = 272, p =.00). However, CFI (\<.90) and TLI (\<.95) do not indicate good fit for both the 3- and 5-factor structures. Both the 3- and 5-factor structures meet the standards for metric invariance across gender and age groups (multi‑group confirmatory factor analysis; comparison of metric and configural model: difference in the models’ CFI ≤.01), which can be interpreted to indicate that the measured dimensions of behavior (narrowband and broadband scales) manifest in the same way across boys and girls as well as different age groups (quartile age groups in years: \[11,12.2\], (12.2,13.4\], (13.4,14.7\], and (14.7,17.9\]). As the goal of the present paper is to compare the statistical predictive performance of the broadband (3-factor structure) and narrowband (5-factor structure) scales of behavior and as it is not the goal to highlight differences between boys and girls or different age groups, sex and age are not considered as predictors of the child and adolescent outcomes. ## Descriptive results Based on the SDQ narrowband subscales, the proportion of at-risk children and adolescents ranges between 5.3% and 9.2% (conduct problems: 5.3%; emotional symptoms: 6.2%; peer problems: 6.5%; hyperactivity: 9.2%), while the SDQ broadband subscales reveal a proportion of 7.9% (internalizing behavior) and 10% (externalizing behavior) of at-risk children and adolescents. The correlations, means, and standard deviations of the SDQ subscales and the child and adolescent outcomes are displayed in. All outcomes are positively associated. Increased positive correlations are observed between the grades in math and German (r =.47), as well as between the KINDL-R subscales of family and school (r =.35). The general self-efficacy scale is likewise considerably correlated with the KINDL-R subscales of self-esteem (r =.40) and friends (r =.32). The different SDQ subscales were negatively correlated with all outcomes, which means that increased behavioral problems measured by means of the different SDQ subscales are associated with lower values for the outcomes, indicating adverse outcomes. The KINDL-R friends subscale is highly correlated with the SDQ subscales of peer problems (r =.49) and internalizing behavior (r =.48). The grades (math and German) were only weakly correlated with the SDQ subscales of peer problems and internalizing behavior (r ranges from.04 to.10). Also, the correlation between the German grades and emotional symptoms is close to zero (r =.03). Another small correlation is between the KINDL-R subscale for friends and hyperactivity (r =.08). ## Main results: Predictive performance of the SDQ subscales ### Internalizing behavior vs. emotional symptoms and peer problems Each child and adolescent outcome is regressed on the different SDQ subscales, which are the broadband subscale for internalizing behavior (model 1) and both underlying narrowband subscales, i.e., emotional symptoms and peer problems jointly as predictors in one regression model (model 2). The regression coefficients (B) and model fit parameters (R<sup>2</sup> and AIC) are displayed in. With regard to the results of models 1 and 2, it can be stated that the associations between the outcomes and the SDQ subscales are negative, which means that increased behavioral problems as indicated by high SDQ subscale scores are associated with lower values of the outcomes, indicating adverse outcomes. Taken as a whole, the at-risk children and adolescents have the lowest average outcome values (the intercept ranges from -1.61 to -0.11). With reference to the model fit parameters, the narrowband subscales of emotional symptoms and peer problems (model 2: jointly as predictors in one regression model) outperform the broadband subscale of internalizing behavior (model 1) in the prediction of all outcomes (comparable R<sup>2</sup> and lower AIC values), except for the prediction of general self-efficacy (AIC value favors the predictive performance of the broadband subscale). However, the predictive performance of the narrowband and broadband scales is poor with regard to the prediction of the grades (math and German), i.e., the proportion of explained variance is close to zero (R<sup>2</sup> ≤.01). Therefore, it is hard to judge the predictive superiority of one of the SDQ subscales (with regard to grade prediction), although the AIC values favor the predictive performance of the narrowband subscales of emotional symptoms and peer problems (model 2). Besides this, model 2 offers a deeper insight into the magnitude of the effect sizes of the two narrowband subscales. For example, in the prediction of the KINDL-R school subscale, the regression coefficients for emotional symptoms are remarkably higher (B ranges from 0.37 to 1.30) than the coefficients for peer problems (B ranges from 0.04 to 0.32). As well, in the prediction of the KINDL-R family subscale and in the prediction of the math grade, the emotional symptoms subscale shows noticeably higher coefficients than the peer problems subscale (family: B ranges from 0.22 to 0.86 vs. 0.06 to 0.26, math: B ranges from 0.04 to 0.34 vs. 0.03 to 0.06). The situation is reversed in the case of predicting the KINDL-R friends subscale, i.e., higher coefficients are observable for peer problems (B ranges from 0.62 to 1.66), while lower coefficients are detected for emotional symptoms (B ranges from 0.23 to 0.69). Also in the prediction of the German grade, the peer problems subscale shows higher coefficients than the emotional symptoms subscale (B ranges from 0.02 to 0.30 vs. -0.10 to 0.00). This detailed information about the differences in effect sizes between the two narrowband subscales (model 2) is not depicted when the broadband subscale for internalizing behavior is used as a predictor (model 1). The results are almost the same when the SDQ subscales are considered as continuous predictors. With regard to the model fit parameters, the narrowband subscales of emotional symptoms and peer problems (model 2: jointly as predictors in one regression model) outperform the broadband subscale for internalizing behavior (model 1) in the prediction of all outcomes (comparable or lower R<sup>2</sup> and lower AIC values), except for the prediction of self-esteem and general self-efficacy (AIC values favor the predictive performance of the broadband subscale). Differences in effect sizes (B) between the two narrowband subscales (model 2) are apparent. ### Externalizing behavior vs. conduct problems and hyperactivity Each child and adolescent outcome is regressed on the different SDQ subscales, which are the broadband subscale for externalizing behavior (model 1) and both underlying narrowband subscales, i.e., conduct problems and hyperactivity jointly as predictors in one regression model (model 2). The regression coefficients (B) and model fit parameters (R<sup>2</sup> and AIC) are displayed in. With regard to the results of models 1 and 2, it can be stated that the associations between the outcomes and the SDQ subscales are negative, which means that increased behavioral problems as indicated by high SDQ subscale scores are associated with lower values for the outcomes, indicating adverse outcomes. Taken as a whole, the at-risk children and adolescents have the lowest average outcome values (the intercept ranges from -0.93 to -0.18). With regard to the model fit parameters, the narrowband subscales for conduct problems and hyperactivity (model 2: jointly as predictors in one regression model) outperform the broadband subscale of externalizing behavior (model 1) in the prediction of all outcomes (comparable or higher R<sup>2</sup> and lower AIC values). In addition, model 2 offers a deeper insight into the magnitude of the effect sizes of the two narrowband subscales. For example, in the prediction of the KINDL-R friends subscale, the regression coefficients for conduct problems are remarkably higher (B ranges from 0.12 to 0.57) than the coefficients for hyperactivity (B ranges from -0.05 to 0.20). Similarly, in the prediction of the KINDL-R family subscale, the conduct problems subscale shows considerably higher coefficients (B ranges from 0.20 to 1.14) than the hyperactivity subscale (B ranges from 0.10 to 0.52). The situation is reversed in the case of predicting the math grade, i.e., higher coefficients are observable for hyperactivity (B ranges from 0.17 to 0.72), while lower coefficients are detected for conduct problems (B ranges from 0.18 to 0.31). In addition, in the prediction of general self-efficacy, the hyperactivity subscale shows noticeably higher coefficients (B ranges from 0.02 to 0.97) than the emotional symptoms subscale (B ranges from -0.07 to 0.33). This detailed information about the differences in effect sizes between the narrowband subscales (model 2) is not depicted when the broadband subscale for externalizing behavior is used as a predictor (model 1). If the SDQ subscales are considered as continuous predictors, the broadband subscale for externalizing behavior (model 1) outperforms the narrowband subscales (model 2) in the prediction of the KINDL-R subscales of school and self-esteem as well as in the prediction of the German grade (comparable R<sup>2</sup> and lower AIC values). In these predictions (KINDL-R school and self-esteem subscales as well as the German grade), there are no differences in effect sizes (B) between the two narrowband subscales, but they are apparent in the predictions of the other outcomes (KINDL-R subscales for friends and family, as well as math grades and general self-efficacy). # Discussion For the first time, the SDQ broadband and narrowband scales were compared with regard to their criterion validity in predicting child and adolescent outcomes. The results of the study indicated the relevance of the different SDQ subscales for the description of students’ socio-emotional and academic situation. At the same time, the results could not support a superiority of the broadband subscales with regard to prediction of the outcomes. This interpretation can be described for the internalizing and externalizing behavior subscales (except for the prediction of general self-efficacy, where the internalizing behavior scale shows the best model fit). If the SDQ scales are considered as continuous predictors, the broadband scale for internalizing behavior outperforms the narrowband scales in the prediction of general self-efficacy and self-esteem. The same holds true for the prediction of the KINDL-R subscales of school and self-esteem, as well as for the prediction of the German grade, where the continuous externalizing subscale shows the best model fit. At any rate, in all cases where a continuous broadband scale outperforms the underlying narrowband scales, the difference in AIC values between the models is less than 3, i.e., the models with narrowband scales are almost as good as the models with the broadband scales. The use of sum scores for the narrowband subscales of emotional symptoms and peer problems is more informative with respect to the range of predicted outcomes. The models using the narrowband scales indicate that there are differences between emotional symptoms and peer problems with regard to their effect on different outcome variables. This information is not depicted through the use of the broadband subscale of internalizing behavior, which might therefore lead to a loss of information. At the same time, it must be stated that children and adolescents might exhibit symptomatic behaviors related to emotional symptoms but not have major peer problems or vice versa. Similar observations can be made for conduct problems and hyperactivity. It can be assumed that different categories of behavioral problems (e.g. conduct problems and attention deficit/hyperactivity) also go in hand with the development of different outcomes over time. Similarly, different conditions and predictors might lead to either the one or the other narrowband behavior. Moreover, different developmental trajectories become clear when focusing different narrowband behaviors problems (e.g. emotional problems and peer problems), Aggregating scores of narrowband into broadband scales might therefore run the risk of blending distinctive behaviors associated with different developmental outcomes. This evidence seems to strengthen the assumptions of Tandon et al. ( p. 593) who argue that: > “\[…\] a major shift, and advance in this area, has been the study > of more discrete differentiated disorders instead of lumping of all > internalizing symptoms into one broad category of the two-dimensional > internalizing versus externalizing taxonomy of childhood > psychopathology.” Therefore, the differentiation between different narrow facets of internalizing behavior seems to be of particular importance in the description of child and adolescent psychopathology and the prediction of relevant outcomes. With respect to the broadband subscale of externalizing behavior, these assumptions can also be partially supported. In line with similar previous findings, differences between the narrowband subscales of conduct problems and hyperactivity with regard to their effect on the outcomes can also be described for most predictions. ## Limitations In this context, however, it must be noted that the chosen outcome variables do not fully describe the levels of educational development of children and adolescents. Further research is desirable that applies in-depth assessment of educational outcomes, such as domain-specific academic achievement, social integration, cognitive or self-regulation processes and uses a multi-informant approach (especially data reported by educators and teachers are of relevance). Compared to school grades, a domain-specific assessment (e.g., reading comprehension) would provide a more detailed picture of the academic performance. Besides this, the reporting of school grades from the last report card by the parents might be prone to recall bias. In addition, some of the chosen outcome variables (KINDL-R scales school and friends) show only low internal consistencies and might therefore not be reliable. Unsatisfactory internal consistency can also be described with regard to some of the SDQ scales used (conduct problems and peer problems). However, it is important to note that poor reliability does not necessarily affect the goodness of predictive analysis. ## Conclusion Despite the aforementioned limitations of the study at hand, the results shed light on the predictive abilities of different subscales of the SDQ. In addition to previous studies, these insights can be used for a further discussion of the advantages and disadvantages of different factor structures of the SDQ. In the sense of predicting educational outcomes, no advantage of the broadband scales (resulting from the 3-factor structure) become clear. The application of narrowband scales (resulting from the 5-factor structure), providing a more differentiated picture of the socio-emotional and academic situation of students, seems to be more appropriate for the prediction of child and adolescent outcomes. This interpretation is of course limited, as it refers to a selection of criteria and needs to be replicated with further educational outcome variables. Furthermore, future research should examine the validity of these results when using parent- or teacher-reported student behavior. In addition, the finding that differentiation of behavioral problems might be a benefit for the description of educational outcomes should be replicated using other behavior assessment tools than the SDQ. Nonetheless, for the purpose of identifying students at risk of struggling in educational contexts, using the set of narrowband dimensions of behavior seems to be more appropriate in educational research and practice. ## Implications The study at hand indicated the need of focusing narrowband behaviors in educational practice in order to gain the most differentiated insights into possible predictors of the emotional-social as well as academic development of children and adolescents. It becomes clear, that different narrowband behaviors are more or less associated with different outcome variables. Subsuming behaviors into broadband categories in educational practice might lead to the fact that students, who might have been identified as at-risk because of salient narrowband behavior, might not be identified as at-risk in broadband categories (classification accuracy). In extension to these results of our analysis, the question arises, whether in educational practice, focusing narrowband behaviors might also be the most appropriate approach with regard to educational planning. Consequently, Casale et al. argue, that the early identification of at-risk students and the provision of individualized intervention might be a key advantage of applying universal screening procedures (e.g. the SDQ) in schools. Gaining insights in specific behaviors might offer the most detailed information on educational needs, which might be addressed in subsequent behavioral interventions. Subsuming behaviors might however lead to a loss of relevant information. We thank Jannis Bosch, Linda Kuhr, Anja Schwalbe and the reviewers for comments that greatly improved the manuscript. 10.1371/journal.pone.0240312.r001 Decision Letter 0 Slobodskaya Helena R. Academic Editor 2020 Helena R. Slobodskaya This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 16 Jul 2020 PONE-D-20-15994 Prediction of School-Relevant Outcomes with Broadband and Narrowband Dimensions of Internalizing and Externalizing Behavior Using the Child and Adolescent Version of the Strengths and Difficulties Questionnaire PLOS ONE Dear Dr. Kulawiak, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. After thoroughly considering the reviews and reading the paper myself, I offer a number of points to consider in a potential revision, as well. 1. Please enter a Financial Disclosure statement: The author(s) received no specific funding for this work. 2. The use of the term “school relevant outcomes” for measures of subjective well-being, self-esteem and self-efficacy seems problematic. These constructs in themselves are the focus of important areas of research, therefore, for a multidisciplinary journal such as PLOS ONE, it would be preferable to use a wider term, for example “child outcomes” or “developmental outcomes”. I would also suggest to change the ambiguous term “social and personal factors” to more specific terms. 3. What are reliabilities and cut-offs for the externalizing and internalizing scales? 4. Please describe the KINDL-R in more detail: How many items are there? How many scales? What are the reliabilities? 5. Grades. What time period is covered by the last report card? 6. Please present the reliability of the general self-efficacy scale. 7. Please provide chi square statistics and CFI for the CFA. Reviewer 2 also addresses this issue. 8. Please provide the significance levels for the correlations in Table 2: Please submit your revised manuscript by Aug 30 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Helena R. Slobodskaya, M.D., Ph.D., D.Sc. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2.Thank you for stating the following financial disclosure: \[NO\]. At this time, please address the following queries: a\) Please clarify the sources of funding (financial or material support) for your study. List the grants or organizations that supported your study, including funding received from your institution. b\) State what role the funders took in the study. If the funders had no role in your study, please state: “The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.” c\) If any authors received a salary from any of your funders, please state which authors and which funders. d\) If you did not receive any funding for this study, please state: “The authors received no specific funding for this work.” Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 3\. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For more information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially sensitive information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: No Reviewer \#2: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors present an interesting question: how do the different factor-structures of the SDQ predict school grades. It is a question of interest to many educational researchers. However, it is difficult to assess how this question is answered based upon the unclear analyses used. It is necessary to further describe the methods in detail. It may be advised to conduct additional analyses to account for the complex data structure and to provide analyses that correspond more clearly to the goals of the article. Below are my more detailed comments: Introduction: 1\) The authors argue that the literature linking SDQ scales to academic achievement is sparse - there is research linking subscales of it to achievement, as well as SDQ to other related scales. Additional work describing some of these works may be useful for the reader. (e.g., DeVries, Rathmann, Gebhardt, 2018). Failing the specific comparisons of SDQ subscales to achievement, perhaps a discussion about the underlying constructs relationships to achievement could be expanded upon further. 2\) In general the literature review is a bit brief, but to the point. It may be worth adding a section about predictive validity of the SDQ to other variables - who made what predictions and did they use narrow or broad scales? Methods/Results: 3\) Alpha and omegas in the.6 range while technically considered acceptable are still quite low - it may be worth making a cautionary note somewhere about this. 4\) Also, why not give the exact values, as you do in other places (lines 170, 171, 173). 5\) Alpha of for the KiGGS study is also very low, do you have the available omega? (line 191). 6\) It appears you dropped prosocial behavior from all results. It may be worth a sentence somewhere explaining this and that it was done because it does not relate to the goals of the study. 7\) It is altogether unclear exactly what analyses were done. Judging from the text and output tables, the authors appear to have run separate regressions for each dependent and independent variables. A multiple linear regression technique would be more appropriate here. It may be advised to run multiple linear models where the predictors are combined into the model, instead of a separate model for each predictor. Regardless, more information is required as the precise models that were developed and tested. 8\) Is there no attempt to asses possible clusters and group effects? For this, a multilevel multiple regression may be advised. Discussion 8\) Without a clear picture to the exact analyses run, it is impossible to evaluate the validity of any conclusions made. The general argument appears to be that the more differentiated model(s?) has(have?) a better fit and moreover there is a variation in the betas for the subscores. This line of reasoning is possibly convincing, but it requires some additional details. can you provide some theoretical connection to this reasoning? The discussion itself is altogether brief, and would benefit from this as well as a deeper connection to previous work (which may be needed also in the Intro). 9\) Some discussion of the predictive power of variables with low reliability may be relevant - either as a limitation or a caution. Essentially, the low reliability of the scales may have a major impact on predictive validity, which is a major focus of the article. other notes: line 84: provide to provided Reviewer \#2: The study entitled “Prediction of School-Relevant Outcomes with Broadband and Narrowband Dimensions of Internalizing and Externalizing Behavior Using the Child and Adolescent Version of the Strengths and Difficulties Questionnaire” is of great interest in the field of Child and Adolescent Psychological Health. The questionnaire SDQ is one of the most frequently used in this field and the approach is very stimulating. It contains new scientific knowledge and provides comprehensive information for further development in this productive line of research. This paper is well-argued and clearly worthy of publication. It has several strengths, amongst others: The background is adequate and up to date. The sample used is adequate and cross-sectional self-report data was obtained. Regression models have been compared with regard to the prediction of school-relevant outcomes using narrowband vs. broadband scales of behavior. The data are presented in a clear and easy-to-understand fashion. The results are clear and related to main goals. The conclusions are supported by data. As minor comments I would like to say: Introduction The literature review could include some more recent research on the use of the SDQ questionnaire, as well as provide a stronger rational for the study of internalizing and externalizing behavior problems in this specific population. Method The age group is not clearly defined. It would be necessary to include frequencies and percentages of the age and sex distribution. Results In the result section, the authors informed that models for both 3 or 5 factors show good fit indices. Please include more than one indicator of model fit besides the RMSEA, such as Chi-square (χ2); degrees of freedom (df); p-value, CFI or TLI. In addition to this information, factor invariance of gender and age would be interesting to include in the analyses. In table 1, in addition to the percentages of "abnormal" scores, it would be interesting to include those that are within the normal range and those that are at the limit. Please, provide a better rational for choosing this type of regression and the AIC measure rather than, for example, other types of regression such as hierarchical regressions. Why did you test so many different regression models? Please clarify the purpose. Discussion In the discussion section, please describe in more detail the contribution of your study and its implications for practice in the educational context. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: Yes: J M DeVries Reviewer \#2: Yes: Inmaculada Montoya-Castilla \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0240312.r002 Author response to Decision Letter 0 14 Sep 2020 Points raised by the academic editor 1\. Please enter a Financial Disclosure statement: The author(s) received no specific funding for this work. We added the statement to the current cover letter: “The author(s) received no specific funding for this work.” 2\. The use of the term “school relevant outcomes” for measures of subjective well-being, self-esteem and self-efficacy seems problematic. These constructs in themselves are the focus of important areas of research, therefore, for a multidisciplinary journal such as PLOS ONE, it would be preferable to use a wider term, for example “child outcomes” or “developmental outcomes”. I would also suggest to change the ambiguous term “social and personal factors” to more specific terms. We changed the term “school relevant outcomes” to “child and adolescent outcomes”. We removed the term “social and personal factors”. Instead, we use the term “self-belief”, for example (l. 157): “measures of academic success (grades), well-being (school, friends, and family), and self-belief (self-esteem and self-efficacy).” 3\. What are reliabilities and cut-offs for the externalizing and internalizing scales? Cut-offs (l. 206): “A conservative classification rule (37), which minimizes false positive cases by selecting a cutoff value below 10%, was used to identify at-risk children and adolescents (emotional symptoms ≥ 6, peer problems ≥ 5, conduct problems ≥ 5, hyperactivity ≥ 7, internalizing behavior ≥ 9, and externalizing behavior ≥ 10).” Reliability coefficients (Cronbach’s α and McDonald’s ω calculated with the data at hand) are reported in Table 2. 4\. Please describe the KINDL-R in more detail: How many items are there? How many scales? What are the reliabilities? Items (l. 230): “Each subscale consists of 4 items.”. Additional information about the subscales (l. 234): “The subscales physical and emotional well-being are not used in the present study.” Reliability coefficients (Cronbach’s α and McDonald’s ω calculated with the data at hand) are reported in Table 2. 5\. Grades. What time period is covered by the last report card? l\. 235: “The school grades (math and German) received on the last report card (half-year term) were reported by the parents.” 6\. Please present the reliability of the general self-efficacy scale. Reliability coefficients (Cronbach’s α and McDonald’s ω calculated with the data at hand) are reported in Table 2. 7\. Please provide chi square statistics and CFI for the CFA. Reviewer 2 also addresses this issue. Additional model fit parameters have been included (l. 307). 8\. Please provide the significance levels for the correlations in Table 2. Table note (Table 2): “Significant correlations (p \<.05) are in bold.” Reviewer \#1 1\. The authors argue that the literature linking SDQ scales to academic achievement is sparse - there is research linking subscales of it to achievement, as well as SDQ to other related scales. Additional work describing some of these works may be useful for the reader. (e.g., DeVries, Rathmann, Gebhardt, 2018). Failing the specific comparisons of SDQ subscales to achievement, perhaps a discussion about the underlying constructs relationships to achievement could be expanded upon further. The point we want to address is not that the literature that links SDQ scales to academic achievement is sparse, but that the literature that compares the narrowband and broadband scales in the prediction of outcomes is sparse. We added more information for clarification. We hope this clarifies your query. We have updated the literature to describe the issue in more detail: (Aanondsen et al., 2020; DeVries et al., 2018; Essau et al., 2012; Jones et al., 2020; McAloney-Kocaman & McPherson, 2017; Niclasen & Dammeyer, 2016; Zarrella et al., 2018) 2\. In general the literature review is a bit brief, but to the point. It may be worth adding a section about predictive validity of the SDQ to other variables - who made what predictions and did they use narrow or broad scales? We now make it clear (introduction) which studies used narrowband scales and which broadband scales. 3\. Alpha and omegas in the.6 range while technically considered acceptable are still quite low - it may be worth making a cautionary note somewhere about this. We updated the limitations (l. 502): “In addition, some of the chosen outcome variables (KINDL-R scales school and friends) show only low internal consistencies and might therefore not be reliable. Unsatisfactory internal consistency can also be described with regard to some of the SDQ scales used (conduct problems and peer problems). However, it is important to note that poor reliability does not necessarily affect the goodness of predictive analysis (67).”. 67 = Smits, N., van der Ark, L. A., & Conijn, J. M. (2018). Measurement versus prediction in the construction of patient-reported outcome questionnaires: Can we have our cake and eat it? Quality of Life Research, 27(7), 1673–1682. <https://doi.org/10.1007/s11136-017-1720-4> 4\. Also, why not give the exact values, as you do in other places (lines 170, 171, 173). Reliability coefficients (Cronbach’s α and McDonald’s ω calculated with the data at hand) are reported in Table 2. 5\. Alpha of for the KiGGS study is also very low, do you have the available omega? (line 191). Reliability coefficients (Cronbach’s α and McDonald’s ω calculated with the data at hand) are reported in Table 2. 6\. It appears you dropped prosocial behavior from all results. It may be worth a sentence somewhere explaining this and that it was done because it does not relate to the goals of the study. l\. 199: “The prosocial behavior subscale was not considered in the present study, because the focus was on a comparison of the two broadband subscales with the underlying narrowband subscales.” 7\. It is altogether unclear exactly what analyses were done. Judging from the text and output tables, the authors appear to have run separate regressions for each dependent and independent variables. A multiple linear regression technique would be more appropriate here. It may be advised to run multiple linear models where the predictors are combined into the model, instead of a separate model for each predictor. Regardless, more information is required as the precise models that were developed and tested. Each outcome is predicted by the SDQ scales: Outcome regressed on broadband scale (Model 1) and outcome regressed on both underlying narrowband scales (Model 2). Hence, model 2 is a multiple linear model (two predictors: two narrowband scales). Both models are compared (AIC comparison) to judge the statistical predictive performance of the different subscales of the SDQ (broadband vs. narrowband), whereby this type of model comparison is conducted separately for internalizing and externalizing behavior (internalizing behavior vs. emotional symptoms and peer problems; externalizing behavior vs. conduct problems and hyperactivity). We reorganized the section “Statistical Analysis Strategy” (starts at l. 248) and added more information for clarification. We hope this clarifies your query. 8\. Is there no attempt to asses possible clusters and group effects? For this, a multilevel multiple regression may be advised. l\. 171: “Study participants were not recruited in schools (non-nested data structure) but randomly selected from the official registers of local residents.” A multilevel regression is not necessary because the data structure is non-nested (random sample). 9\. Without a clear picture to the exact analyses run, it is impossible to evaluate the validity of any conclusions made. The general argument appears to be that the more differentiated model(s?) has(have?) a better fit and moreover there is a variation in the betas for the subscores. This line of reasoning is possibly convincing, but it requires some additional details. can you provide some theoretical connection to this reasoning? The discussion itself is altogether brief, and would benefit from this as well as a deeper connection to previous work (which may be needed also in the Intro). Your summary of our research describes exactly the main point. We are glad that our main point is understandable. We added more information to highlight the research question and aims. In addition, we tried to address the need of a stronger theoretical connection as well as a stronger connection to previous work. Therefore, we extended the discussion with regard to previous findings on the independence of narrowband behaviors (ll.47 and 470) as well as the implications for educational planning (l. 520) 10\. Some discussion of the predictive power of variables with low reliability may be relevant - either as a limitation or a caution. Essentially, the low reliability of the scales may have a major impact on predictive validity, which is a major focus of the article. We updated the limitations (l. 502): “In addition, some of the chosen outcome variables (KINDL-R scales school and friends) show only low internal consistencies and might therefore not be reliable. Unsatisfactory internal consistency can also be described with regard to some of the SDQ scales used (conduct problems and peer problems). However, it is important to note that poor reliability does not necessarily affect the goodness of predictive analysis (67).” 67 = Smits, N., van der Ark, L. A., & Conijn, J. M. (2018). Measurement versus prediction in the construction of patient-reported outcome questionnaires: Can we have our cake and eat it? Quality of Life Research, 27(7), 1673–1682. <https://doi.org/10.1007/s11136-017-1720-4> 11\. line 84: provide to provided Done Reviewer \#2 1\. The literature review could include some more recent research on the use of the SDQ questionnaire, as well as provide a stronger rational for the study of internalizing and externalizing behavior problems in this specific population. We have updated the literature: (Aanondsen et al., 2020; DeVries et al., 2018; Essau et al., 2012; Jones et al., 2020; McAloney-Kocaman & McPherson, 2017; Niclasen & Dammeyer, 2016; Zarrella et al., 2018). We see the need of providing a stronger rational for the study of internalizing and externalizing behavior in this specific population. Therefore, we revised parts of the theoretical background (l. 47) and added some references emphasizing the need and rational for the study of internalizing and externalizing behavior problems in this specific population. 2\. The age group is not clearly defined. It would be necessary to include frequencies and percentages of the age and sex distribution. Minimum, maximum and quartiles have been included to describe the distribution (age), l. 185: “The present sample therefore includes children and adolescents with a minimum age of 11 years in grades 5 to 9 (N = 4654; 52% boys; age in years: M = 13.46, SD = 1.47, Min = 11.00, Q1 = 12.17, Md = 13.42, Q3 = 14.67, Max = 17.92).” 3\. In the result section, the authors informed that models for both 3 or 5 factors show good fit indices. Please include more than one indicator of model fit besides the RMSEA, such as Chi-square (χ2); degrees of freedom (df); p-value, CFI or TLI. In addition to this information, factor invariance of gender and age would be interesting to include in the analyses. Additional model fit parameters have been included (l. 307). Factor invariance (also referred to as measurement invariance): “Both the 3- and 5-factor structures meet the standards for metric invariance (47,48) across gender and age groups (multi‑group confirmatory factor analysis; comparison of metric and configural model: difference in the models’ CFI ≤.01), which can be interpreted to indicate that the measured dimensions of behavior (narrowband and broadband scales) manifest in the same way across boys and girls as well as different age groups (quartile age groups in years: \[11,12.2\], (12.2,13.4\], (13.4,14.7\], and (14.7,17.9\]).” (l. 309) 4\. In table 1, in addition to the percentages of "abnormal" scores, it would be interesting to include those that are within the normal range and those that are at the limit. Done (see Table 1) 5\. Please, provide a better rational for choosing this type of regression and the AIC measure rather than, for example, other types of regression such as hierarchical regressions. Why did you test so many different regression models? Please clarify the purpose. We have added an important source: “For the application of AIC model selection in the field of psychology and psychometrics, see Vrieze (53).” (l. 271). Vrieze, S. I. (2012). Model selection and psychological theory: A discussion of the differences between the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). Each outcome is predicted by the SDQ scales: Outcome regressed on broadband scale (Model 1) and outcome regressed on both underlying narrowband scales (Model 2). Hence, model 2 is a multiple linear model (two predictors: two narrowband scales). Both models are compared (AIC comparison) to judge the statistical predictive performance of the different subscales of the SDQ (broadband vs. narrowband), whereby this type of model comparison is conducted separately for internalizing and externalizing behavior (internalizing behavior vs. emotional symptoms and peer problems; externalizing behavior vs. conduct problems and hyperactivity). We reorganized the section “Statistical Analysis Strategy” (starts at l. 248) and added more information for clarification. We hope this clarifies your query. 6\. In the discussion section, please describe in more detail the contribution of your study and its implications for practice in the educational context. We have added a subsection on the “Implications” of our study (l. 520 discussing the advantages of applying broad- or narrowband categorizations of behavior in schools with regard to diagnostic practice and educational planning. 10.1371/journal.pone.0240312.r003 Decision Letter 1 Slobodskaya Helena R. Academic Editor 2020 Helena R. Slobodskaya This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 24 Sep 2020 Prediction of Child and Adolescent Outcomes with Broadband and Narrowband Dimensions of Internalizing and Externalizing Behavior Using the Child and Adolescent Version of the Strengths and Difficulties Questionnaire PONE-D-20-15994R1 Dear Dr. Kulawiak, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Helena R. Slobodskaya, M.D., Ph.D., D.Sc. Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: All recommended changes have been addressed. The exact analyses and procedures are now quite clear, as well as the conclusions and limitations. \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No 10.1371/journal.pone.0240312.r004 Acceptance letter Slobodskaya Helena R. Academic Editor 2020 Helena R. Slobodskaya This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 29 Sep 2020 PONE-D-20-15994R1 Prediction of child and adolescent outcomes with broadband and narrowband dimensions of internalizing and externalizing behavior using the child and adolescent version of the strengths and difficulties questionnaire Dear Dr. Kulawiak: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Helena R. Slobodskaya Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction Phenolic glycolipids (PGL) are found in a limited group of pathogenic mycobacteria including the members of the M. tuberculosis complex, M. leprae, and a few other slow-growing mycobacteria (M. kansasii, M. gastri, M. ulcerans, M. marinum, M. microti and M. haemophilum),. These substances, located in the outermost layer of the mycobacterial cell envelope, have been shown to play important functions in the pathogenicity of these bacteria. PGL are composed of a mixture of long chain β-diols, esterified by multimethyl-branched fatty acids, named mycocerosic acids or phthioceranic acids depending on the configuration of the asymmetric centres bearing the methyl branches. The β-diols chain is terminated by an aromatic nucleus, which in turn is glycosylated. The sugar moiety of PGL consists of 1 to 4 sugar residues, depending on the mycobacterial species, and most are O-methylated deoxysugars. In M. tuberculosis the carbohydrate domain of the major variant of PGL, named PGL-tb, is 2,3,4-tri-O-Me-α-L-Fucp(1→3)-α-L-Rhap(1→3)-2-O-Me-α-L- Rhap(1→). M. tuberculosis also secretes a family of smaller molecules that contain the same glycosylated phenolic moiety as PGL-tb, the glycosylated p-hydroxybenzoic acid methyl esters (p-HBAD). Although PGL-tb are absent from many M. tuberculosis strains, due to a natural frameshift mutation within pks15/1, a gene encoding a type-I polyketide synthase specifically involved in the biosynthesis of PGL-tb , several lines of evidence indicate that they may contribute to the pathogenesis of M. tuberculosis. For instance, it was reported that a M. tuberculosis strain producing PGL-tb induced more severe tuberculous meningitis in infected rabbits than did strains devoid of PGL-tb. Moreover, the production of PGL-tb in one M. tuberculosis isolate was associated with a hypervirulence phenotype in the mouse model. This observation was correlated with the finding that PGL-tb inhibits the production of several pro-inflammatory cytokines by infected mouse macrophages. This phenotype was dependent on the structure of the saccharide moiety of PGL-tb since the structurally related monoglycosylated PGL produced by M. bovis BCG, the so-called mycoside B, did not produce a similar effect. The putative role of the saccharide moiety in the immunomodulation activity of PGL- tb was further highlighted by the finding that p-HBAD inhibits the pro- inflammatory response of infected macrophages. Thus it appears that the carbohydrate moiety of PGL-tb and of p-HBAD likely plays an important role in the pathogenesis of mycobacterial infections. The genes involved in the biosynthesis of the lipid core of PGL-tb are clustered on a 70 kb region of the mycobacterial chromosome, the so-called DIM+PGL locus. This chromosomal region also contains three glycosyltransferase-encoding genes, namely Rv2957, Rv2958c and Rv2962c, and another gene (Rv2959c) encoding a methyltransferase required for the formation of the carbohydrate part of PGL- tb and p-HBAD. Two additional genes (Rv1511 and Rv1512) located outside of the DIM+PGL locus and responsible for the formation of L-fucose in M. tuberculosis, have been shown to be involved in the biosynthesis of the terminal fucosyl residue of PGL-tb. Despite these advances, the genes involved in the O-methylations of the fucosyl residue of PGL-tb remain uncharacterized. In this study, we report the identification of the three methyltransferases responsible for modifying the hydroxyl groups at positions 2, 3, and 4 of the terminal fucosyl unit of PGL-tb. # Materials and Methods ## Bacterial Strains, Growth Media and Culture Conditions Plasmids were propagated at 37°C in E. coli DH5α or HB101 in LB broth or LB agar (Invitrogen) supplemented with either kanamycin (Km) (40 µg/ml) or hygromycin (Hyg) (200 µg/ml). The M. tuberculosis H37Rv Pasteur wild-type strain and its derivatives were grown at 37°C in Middlebrook 7H9 broth (Difco) containing ADC (0.2% dextrose, 0.5% bovine serum albumin fraction V, 0.0003% beef catalase) and 0.05% Tween 80 when necessary and on solid Middlebrook 7H11 broth containing ADC and 0.005% oleic acid (OADC). For biochemical analyses, mycobacterial strains were grown as surface pellicles on Sauton’s medium. When required, Km and Hyg were used at a concentration of 40 µg/ml and 50 µg/ml, respectively. Sucrose 2% (w/v) was used to supplement 7H11 for the construction of the PMM115 and PMM116 mutants. ## General DNA Techniques Molecular cloning experiments were performed using standard procedures. Mycobacterial genomic DNA was extracted from 5 ml saturated cultures as previously described. PCR experiments for plasmid construction or genomic analysis were performed in standard conditions on a GeneAmp PCR system 2700 thermocycler (Applied Biosystem). PCR was performed in a final volume of 50 µl containing 2.5 units of Pfu DNA polymerase (Promega). ## Construction of M. tuberculosis H37Rv Gene-disrupted Mutants The ΔRv2954c and ΔRv2956 M. tuberculosis H37Rv mutants (PMM115 and PMM116) were constructed by allelic exchange using the Ts/sacB procedure. Two DNA fragments containing either the Rv2954c gene or the Rv2956 gene were amplified by PCR from M. tuberculosis genomic DNA using oligonucleotides 2954A and 2954B (for Rv2954c) and 2956A and 2956B (for Rv2956) and subsequently cloned into pGEM-T (Promega). An internal 463 bp Rv2954c fragment and an internal 487 bp Rv2956 fragment were removed by double-strand site-directed mutagenesis with the inverse PCR technique using oligonucleotides 2954IP1 and 2954IP2 for Rv2954c and oligonucleotides 2956IP1 and 2956IP2 for Rv2956 and substituted by a km resistance cassette formed by the Ωkm cassette flanked by two res sites from transposon γδ. The DNA fragments containing the disrupted Rv2954c and Rv2956 genes were inserted into pPR27, a mycobacterial thermosensitive suicide plasmid harboring the counterselectable marker sacB and the resulting plasmids were transferred by electrotransformation into M. tuberculosis for allelic exchange. PCR screening for disruption of Rv2954c and Rv2956 were performed with a set of specific primers (2954C, 2954D, 2954E, res1, and res2 for Rv2954c and 2956C, 2956D, 2956E, res1, and res2 for Rv2956, and) after extraction of the genomic DNA from several Km and sucrose- resistant colonies. Two clones giving the expected pattern for disruption of Rv2954c and Rv2956 were selected for further analyses and named PMM115 (Rv2954c::res-km-res) and PMM116 (Rv2956::res-km-res), respectively. To construct a Rv2955c M. tuberculosis mutant of H37Rv, we used the strategy described by Bardarov et al.. A 2369 bp fragment containing the Rv2955c gene was amplified from M. tuberculosis genomic DNA using primers 2955H and 2955I and inserted within the vector pGEM-T. The km resistance cassette was then inserted between the BstEII and EcoRI sites of the Rv2955c gene and the fragment containing the Rv2955c disrupted allele was cloned within the cosmid vector pYUB854. The resulting cosmid, was cut with PacI and ligated with the mycobacteriophage phAE87. The ligation products were encapsidated in vitro using the Gigapack III XL kit (Stratagene) and the mix was used to infect E. coli HB101 following the manufacturer recommendations. Transfectants were selected on LB plates containing Km. A recombinant phagemid containing the disrupted gene construct was selected and transferred by electroporation in M. smegmatis and phage particules were prepared as described previously. These particles were then used to infect M. tuberculosis H37Rv and M. tuberculosis allelic exchange mutants were selected by PCR analysis using primers 2955J, 2955K, 2955L, res1 and res2. One clone, named PMM126 (Rv2955c::res-km-res), gave the pattern corresponding to allelic exchange and was retained for further analysis. The recovery of the res-Ωkm-res cassette from M. tuberculosis PMM115, PMM116 and PMM126 was performed as previously described using the thermosensitive plasmid pWM19 that contains the resolvase gene of transposon γδ. Several clones were selected and analyzed by PCR using various primers (2954C and 2954D for Rv2954c, 2955J and 2955K for Rv2955c, and 2956C and 2956D for Rv2956). Three clones giving the expected pattern for excision of the res-Ωkm-res cassette in Rv2954c, Rv2955c and Rv2956, were selected and named PMM144 (Rv2954c::res), PMM145 (Rv2955c::res) and PMM122 (Rv2956::res), respectively. ## Construction of Complementation Plasmids To construct pRS18 and pRS19, a region covering the Rv2954c gene and a region covering the Rv2955c gene were PCR-amplified from M. tuberculosis H37Rv genomic DNA using oligonucleotides 2954F and 2954G (for Rv2954c) and 2955F and 2955G (for Rv2955c). The PCR products were digested with NdeI and HindIII endonuclease restriction enzymes and cloned between the NdeI and HindIII sites of pMV361e, a pMV361 derivative containing the pblaF\* promoter instead of the original phsp60 promoter and carrying a km resistance marker. For the construction of complement plasmids pRS26 and pRS27, two DNA fragments overlapping either the Rv2957 (pRS26) gene or the Rv2956 and Rv2957 genes (pRS27) were amplified by PCR from M. tuberculosis H37Rv genomic DNA using oligonucleotides 2957A and 2957B or 2956F and 2957B and cloned between the NdeI and HindIII sites of pMV361e. ## Extraction and Purification of Glycolipids Mycobacterial cells obtained from culture on Sauton’s medium were left in 60 ml of CHCl<sub>3</sub>/CH<sub>3</sub>OH (1∶2, v/v) for 48 h to kill bacteria. Lipids were then extracted twice with CHCl<sub>3</sub>/CH<sub>3</sub>OH (2∶1, v/v) for 24 h each, washed twice with water, and dried. Extracted mycobacterial lipids were suspended in CHCl<sub>3</sub> at a final concentration of 20 mg/ml and analyzed by thin-layer chromatography (TLC). Equivalent amounts of lipids from each strain were spotted on silica gel G60 plates (20×20 cm, Merck) and separated with CHCl<sub>3</sub>/CH<sub>3</sub>OH (95∶5, v/v). Glycolipids were visualized by spraying the plates with a 0.2% anthrone solution (w/v) in concentrated H<sub>2</sub>SO<sub>4</sub>, followed by heating. Crude lipid extracts were subjected to chromatography on a Sep-Pak Florisil cartridge and eluted with a series of concentrations of CH<sub>3</sub>OH (0, 10, 20, 30%) in CHCl<sub>3</sub>. Each fraction was analyzed by TLC on Silica Gel G60 using CHCl<sub>3</sub>/CH<sub>3</sub>OH (90∶10, v/v) as the solvent system and glycolipids were visualized as described above. For MALDI-TOF mass spectrometry analyses, glycolipids were additionally purified by preparative chromatography on silica gel G60 plates using CHCl<sub>3</sub>/CH<sub>3</sub>OH (90∶10, v/v) as the developing solvent and recovered by scraping silica gel from the plates. ## Matrix-assisted Laser Desorption-Ionization Time-of-Flight (MALDI-TOF) Mass Spectrometry MALDI-TOF mass spectrometry analyses were performed in reflectron mode, with an Applied Biosystems 4700 analyzer mass spectrometer (Applied Biosystems, Framingham) equipped with an Nd:YAG laser (wawe-length 355 nm; pulse\<500 ps; repetition rate 200 Hz). A total of 2500 shots were accumulated in positive ion mode, and mass spectrometry data were acquired with the default calibration for the instrument. ## Nuclear Magnetic Resonance (NMR) Spectroscopy NMR spectroscopy experiments were carried out at 300° K on a Bruker AVANCE spectrometer operating at 600,13 MHz with a 5-mm triple resonance TCI <sup>1</sup>H <sup>13</sup>C <sup>15</sup>N pulsed field z-gradient cryoprobe. Samples were dissolved in 99,9% CDCl<sub>3</sub>. <sup>1</sup>H NMR studies on the native and the per-O-acetylated glycolipids were performed using one- dimensional and two-dimensional chemical shift correlation spectroscopy (COSY). Chemical shifts are expressed in ppm using chloroform signal as an internal reference (7.265 ppm). # Results ## Rv2954c, Rv2955c and Rv2956 are Involved in the Biosynthesis of PGL-tb in M. tuberculosis In an attempt to identify the protein(s) responsible for the O-methylation of the fucosyl residue in PGL-tb, we searched the M. tuberculosis genome for genes encoding methyltransferases and mapping within the DIM+PGL locus. We identified three genes (Rv2954c, Rv2955c and Rv2956) encoding proteins sharing similarities to S-adenosylmethionine (SAM)-dependent methyltransferases. These genes are located in the same chromosomal region as those previously shown to be involved in the synthesis of the sugar moiety of PGL-tb. Proteins encoded by Rv2954c and Rv2956 are similar in length with 241 and 243 residues, respectively, whereas Rv2955c encodes a protein of 321 amino acids. Although their amino acid sequences share little identity (about 15%), secondary structure prediction analyses revealed that these proteins are composed of alternative α-helices and β-sheets that are characteristic to the SAM methyltransferase core fold (data not shown). In addition a glycine-rich sequence, close to the E/DXGXGXG motif (motif I) found in a number of SAM- methyltransferases and an acidic region (motif II) located 17–19 amino acid residues downstream from the end of motif I are present in the sequence of each protein. Interestingly, motif I starts at position 45 for Rv2954c and Rv2956 but is located at position 132 for Rv2955c suggesting that the latter protein has an additional N-terminal domain. To investigate the role of these putative methyltransferases in the biosynthesis of PGL-tb, we constructed three M. tuberculosis knock-out mutants with insertion within Rv2954c, Rv2955c and Rv2956. These genes were disrupted by insertion of a km resistance gene cassette flanked by two res sites to yield three recombinant strains, named PMM115 (Rv2954c::res-km-res), PMM126 (Rv2955c::res-km-res) and PMM116 (Rv2956::res-km-res). Since M. tuberculosis H37Rv and derivatives are naturally deficient in the production of PGL-tb due to a mutation in the pks15/1 gene, the mutant strains and the H37Rv wild-type strain were transformed with plasmid pPET1 that carries a functional M. bovis BCG pks15/1 gene. Lipids were extracted from the resulting strains and analyzed by TLC. These analyses showed that the production of PGL-tb was affected in the three mutant strains. The PMM115:pPET1 mutant produced a glycolipid (product A) that exhibits a relative mobility lower than that of PGL- tb. The PMM126:pPET1 mutant produced a major (product B) and a minor (product C) glycoconjugates, both exhibiting a lower mobility than that of PGL-tb. Finally the lipid profile of the PMM116:pPET1 recombinant strain contained a single spot (product D) with a lower mobility than that of PGL-tb. These data suggested that Rv2954c, Rv2955c and Rv2956 are involved and play different roles in PGL- tb biosynthesis. ## Rv2954c Catalyses the O-methylation of Position 3 of the Fucosyl Residue of PGL-tb To determine the nature of product A produced by the PMM115:pPET1 mutant strain, theglycolipid was purified and analyzed by MALDI-TOF mass spectrometry and <sup>1</sup>H-NMR. The MALDI-TOF mass spectrum of product A showed a series of major pseudomolecular ion (M+Na<sup>+</sup>) peaks at m/z 1850, 1864, 1878, 1892, 1906 and 1920, 14 mass units lower than those observed with PGL-tb from H37Rv:pPET1 suggesting that this compound may differ from PGL-tb by the absence of one methyl group. As expected, data from one dimensional (1D) <sup>1</sup>H-NMR analysis of compound A were in agreement with the hypothesis. Compared to the NMR spectrum of the native PGL-tb, only three signals (3.4–3.7 ppm), instead of four, attributable to the proton resonances of methoxyl groups linked to the sugar moiety were seen in the spectrum of compound A, confirming the loss of one methoxyl group in compound A. The other NMR data are consistent with the proposed structure, notably (i) the presence of a p-substituted phenolic nucleus (two doublets centered at δ = 6.9–7.1); (ii) polymethyl- branched fatty acids (1.14 ppm) esterifying a β-glycol (4.83 ppm); (iii) a methoxyl group borne by the aliphatic chain (3.32 ppm) that characterize a phenolphthiocerol structure. Finally, the presence of three deshielded anomeric protons (5.1–5.6 ppm) confirmed the occurrence of a trisaccharide in the glycolipid. That compound A lacks a methoxyl group was also supported by the occurrence, upon per-O-acetylation of four distinct methyl groups of acetyl substituents (signal resonances at 1.9–2.2 ppm), indicating the presence of four non-substituted hydroxyl groups in the native compound A. Moreover, the per-O-acetylation resulted in a shift of the resonance of protons linked to the carbons bearing the acetyl group downfield by about 1.0–1.4 ppm; it was thus possible to distinguish free from substituted hydroxyl groups. 2D-COSY NMR analyses of native and per-O-acetylated compound A were performed to assign the resonances (Table). On the basis of chemical shift correlations, we established that it was the position 3 of the fucosyl residue that was per-O-acetylated, indicating that this position was free in the native compound A. This suggested that Rv2954c is involved in the methylation of this position. To firmly establish that production of product A relies to the disruption of the Rv2954c gene, we performed gene complementation studies by transferring a wild-type allele of Rv2954c into the mutant strain. Due to the limited number of antibiotic resistance genes available for use in mycobacteria, the res-km- res cassette was first recovered from the PMM115 strain by site-specific recombination between the two res sites. A new strain, named PMM144 (Rv2954c::res) harboring only a res site at the Rv2954c locus was generated. This strain was transformed either with plasmid pPET1 or with both pPET1 and pRS18 carrying a wild-type allele of Rv2954c. TLC and MALDI-TOF analyses revealed that the PMM144:pPET1 strain produced the same PGL-like substance (product A) than PMM115:pPET1 whereas the production of PGL-tb was restored in the PMM144:pPET1:pRS18 strain ( and). Altogether these data established that Rv2954c is involved in the O-methylation of position 3 of the fucosyl residue of PGL-tb in M. tuberculosis. ## Rv2955c Catalyses the O-methylation of Position 4 of the Fucosyl Residue of PGL-tb The structures of products B and C, which accumulated in the M. tuberculosis PMM126:pPET1 mutant were solved by both MALDI-TOF mass spectrometry and <sup>1</sup>H MNR analyses of the purified compounds. The mass peaks from the major (product B) and from the minor (product C) products were, respectively, 28 and 14 mass units lower than those of PGL-tb from the H37Rv:pPET1 strain suggesting that the PMM126:pPET1 mutant produced two PGL-like substances that may differ from PGL-tb by the absence of two and one methyl group(s), respectively. All the proton resonances typifying a phenolphthiocerol structure linked to a trisaccharide were observed in the 1D <sup>1</sup>H-NMR spectra of native compounds B and C (data not shown). Importantly, only two and three signals attributable to proton resonances of methoxyl groups linked to the sugar part (3.4–3.7 ppm) were observed in the spectra of native compound B and C (data not shown), respectively. This was in agreement with the presence of five and four signals of distinct methyl groups of acetyl substituents observed in the spectra of per-O-acetylated compounds B and C (data not shown), respectively, indicating the loss, compared to PGL-tb, of two and one methoxyl groups in the native compounds B and C, respectively. The positions of these free hydroxyl groups were assigned by 2D-COSY spectroscopy of native and per-O-acetylated compounds B and C (Tables). The resonances of the protons 3 and 4 of the fucosyl residue were shifted after per-O-acetylation, indicating that the hydroxyl groups in positions 3 and 4 were free in the compound B. Likewise, it was possible to establish that position 4 of the fucosyl residue was the free hydroxyl group in the compound C. These data demonstrate that Rv2955c is involved in the O-methylation at position 4 of the fucosyl residue of PGL-tb. However the accumulation in the PMM126:pPET1 strain of a major glycolipid that differs from PGL-tb by the absence of two methyl groups (product B) was surprising because Rv2955c encodes an enzyme that was expected to methylate the fucosyl residue at only one position. One explanation could be that disruption of Rv2955c in M. tuberculosis exerts a partial polar effect on the expression of Rv2954c involved in the O-methylation of position 3 of the fucosyl residue. Alternatively, it is possible that methylation of position 3 catalyzed by the product of Rv2954c occurs only after the O-methylation of position 4 by the methyltransferase encoded by Rv2955c. To discriminate between these hypotheses, we generated a new recombinant strain, named PMM145, with an unmarked mutation in Rv2955c and we performed genetic complementation studies with plasmids pPET1 and pRS19 that contains a wild-type allele of Rv2955c. The PMM145:pPET1 strain synthetized two glycolipids that correspond to products B and C previously characterized in the lipids from the PMM126:pPET1 mutant. Introduction of Rv2955c in the PMM145:pPET1 mutant restored the biosynthesis of a major glycoconjugate, which exhibits TLC mobility identical to that of PGL- tb. The MALDI-TOF mass spectrum of this lipid showed a series of pseudomolecular ion (M+Na<sup>+</sup>) peaks at m/z values identical to those observed in the mass spectrum of PGL-tb purified from the wild-type strain. Since the introduction of Rv2955c in the mutant strain was sufficient to restore the wild-type phenotype, we concluded that the disruption of Rv2955c in M. tuberculosis did not exert a polar effect on the expression of Rv2954c. These findings strongly support the conclusion that i) the product of Rv2955c is responsible for the O-methylation of the hydroxyl group located at position 4 of the fucosyl residue of PGL-tb and ii) the O-methylation of position 3 is dependent on the O-methylation of position 4 by the product of Rv2955c. ## Rv2956 Catalyses the O-methylation of Position 2 of the Fucosyl Residue of PGL-tb Product D, which accumulated in the M. tuberculosis PMM116:pPET1 mutant, was analyzed by MALDI-TOF mass spectrometry. Surprisingly, the mass peaks from this glycoconjugate were 188 mass units lower than those of PGL-tb. This difference corresponded to the mass of the tri-O-methylfucosyl residue of PGL-tb, suggesting that product D was a diglycosylated glycolipid. We reasoned that the insertion of the km cassette within Rv2956 may exert a polar effect on the expression of the downstream Rv2957 gene that encodes a glycosyltransferase involved in the transfer of the fucosyl residue in PGL-tb, making it impossible to determine the putative function of Rv2956 in the methylation of PGL-tb. To circumvent this problem, we generated a mutant with an unmarked mutation in Rv2956. This strain, named PMM122 (Rv2954c::res), was transformed either with pPET1 or with both pPET1 and pRS26, a plasmid harbouring a wild-type copy of Rv2957. The PMM122:pPET1 strain produced a glycolipid that showed a mobility on TLC identical to that of compound D from PMM116:pPET1. MALDI-TOF analyses confirmed that it was indeed the same lipid (data not shown). In contrast the PMM122:pPET1:pRS26 strain accumulated four glycoconjugates. The major one was identified as product D by MALDI-TOF mass spectrometry indicating a partial complementation of the polar effect on Rv2957 expression by the wild-type Rv2957 gene carried by plasmid pRS26. The PMM122:pPET1:pRS26 strain also produced two new PGL-like compounds, named E and F, and a small amount of PGL-tb. The structures of products E and F were solved by both MALDI-TOF mass spectrometry and NMR analyses following purification. MALDI-TOF mass spectrometry analyses indicated that products E and F differ from PGL-tb by the absence of one and three methyl group(s), respectively. The 1D <sup>1</sup>H-NMR spectra clearly demonstrate the presence of phenolphthiocerol in compounds E and F. The presence of three deshielded anomeric proton signals (5.1–5.6 ppm) confirmed the occurrence of a trisaccharide moiety in both glycolipids. Focusing on the signals attributable to proton resonances of methoxyl groups linked to the sugar units (3.4–3.7 ppm), only one signal was observed in the spectrum of product F attributable to the methoxyl group linked on position 2 of the rhamnosyl residue linked to the phenol ring. This clearly indicates the loss in product F of the three methoxyl groups usually present on the PGL-tb fucosyl residue. For product E, three methoxyl signal resonances were observed, compared to PGL-tb, confirming the absence of one methoxyl group in this compound. The position of the resulting free hydroxyl group was then identified by 2D-COSY spectroscopy of native and per-O-acetylated product E (Table). The resonance of the proton 2 of the fucosyl residue was shifted after per-O-acetylation, indicating that the hydroxyl group in position 2 was free in product E. The finding that products E and F were not methylated at position 2 of the fucosyl residue indicated that Rv2956 is involved in the 2-O-methylation of the terminal fucosyl residue of PGL-tb. In addition accumulation of product F which lacks O-methylation on the fucosyl residue, suggested that Rv2955c and Rv2954c display partial activities when position 2 of the fucosyl residue is unmethylated. The PMM122: pPET1 strain was also transformed with plasmid pRS27 carrying the wild-type allele of Rv2957 and that of Rv2956 to complement both the polar effect on the Rv2957 expression and the deletion of Rv2956. The resulting strain exhibited two major glycolipids that were identified as product D and PGL-tb by TLC and MALDI-TOF analyses. The presence of product D indicated that the Rv2957 gene carried by pRS27 weakly complemented the polar effect on the expression of the Rv2957 chromosomal allele as observed with pRS26. However, the absence of PGL-tb like substance lacking a methoxyl group at position 2 of the fucosyl residue and the accumulation of large amounts of PGL-tb confirmed that Rv2956 is indeed involved in the O-methylation at position 2 of PGL-tb. Thus from these data it can be conclude that i) Rv2956 O-methylates the position 2 of the terminal fucosyl residue of PGL-tb and ii) Rv2954c and Rv2955c exhibit higher catalytic activity in the presence of Rv2956. # Discussion In this study, we provide genetic and biochemical evidences that Rv2954c, Rv2955c and Rv2956 encode enzymes capable of O-methylating the terminal residue of PGL-tb at the 3-, 4-, and 2-position, respectively. Indeed, disruption of Rv2954c and Rv2955c in M. tuberculosis results, respectively, in the production of a glycoconjugate with a fucosyl residue that is not O-methylated at position 3 and 4. Interestingly, the lack of a functional Rv2955c gene also prevents efficient O-methylation of position 3 in M. tuberculosis. This effect was not due to a polar effect on the expression of Rv2954c in the mutant strain since introduction of a wild-type allele of Rv2955c in the PMM126:pPET1 mutant was sufficient to fully restore the wild-type phenotype. Disruption of Rv2956 in M. tuberculosis led to the accumulation of a glycolipid lacking the terminal fucosyl residue (product D) suggesting a polar effect on the expression of Rv2957. Complementation of the PMM122:pPET1 strain with Rv2957 or with both Rv2956 and Rv2957 partially suppressed the polar effect, likely due to a weak expression of Rv2957 from plasmids pRS26 and pRS27. Interestingly, the PMM122:pPET1 strain complemented with Rv2957 synthesized faint amounts of PGL-tb raising the possibility that an unidentified methyltransferase would partially fulfil the role of Rv2956. Rv2956 shares a high degree of sequence identity with the product of Rv1513 (68.7% identity) in M. tuberculosis. Rv1513 is likely not required per se for the O-methylation of position 2 of the fucosyl residue because transfer of functional Rv1511, Rv1512 and Rv2958c genes from M. tuberculosis in M. bovis BCG that naturally lacks an Rv1513 ortholog, led to the formation of a glycolipid that is structurally identical to PGL-tb. However in a genetic context where Rv2956 is absent, it is possible that Rv1513 could partially O-methylate the 2-position of the terminal fucosyl residue of PGL-tb in M. tuberculosis. Accumulation of product E which contains a terminal 3,4-di-O-methyl-fucosyl residue in the PMM122:pPET1:pRS26 strain and its absence in the PMM122:pPET1 strain complemented with both Rv2956 and Rv2957 is consistent with the role of Rv2956 in the O-methylation of position 2. Moreover the presence of product F indicated that Rv2954c and Rv2955c have reduced enzymatic activities when the position 2 of the fucosyl residue is not O-methylated. Therefore, Rv2956 is likely to be the first methyltransferase that acts on the fucosyl residue of PGL-tb. Consistently the various PGL-like variants that accumulated in the Rv2954c and Rv2955c mutant strains were O-methylated on position 2. Altogether these data suggest that O-methylation of the fucosyl residue of PGL-tb is a sequential process: the product of Rv2956 catalyses the O-methylation of position 2, to yield a 2-O-Me-α-L-Fucp(1→3)-α-L- Rhap(1→3)-2-O-Me-α-L-Rhap(1→)phenolphthiocerol dimycocerosates; this latter product is then sequentially methylated at position 4 by the Rv2955c-encoded enzyme and later on at position 3 by the product of Rv2954c to give PGL-tb. Sequential O-methylation of sugar residue has been described in other bacteria such as Streptomyces olivaceus and M. smegmatis. For instance M. smegmatis produces cell-wall-associated components, named glycopeptidolipids (GPLs) which contain a lipopeptide core that is modified with O-methylated rhamnosyl units and an O-acylated 6-deoxy talosyl residue. The rhamnosyl residue of GPLs can be O-methylated with up to three methyl groups at positions 2, 3 and 4 by the methyltransferases Rmt2, Rmt3 and Rmt4. It has been shown that O-methylation by Rmt3 at the C3 carbon of the rhamnose was necessary for subsequent methylation by the 4-O methyltransferase Rmt4 and the 2-O-methyltransferase Rmt2. The proposed roles of Rv2954c, Rv2955c and Rv2956 are consistent with the organization of the DIM+PGL chromosomal region in the various sequenced mycobacterial species producing PGL. Indeed conservation of these genes within the DIM+PGL locus correlates with the ability of these strains to synthesize a PGL containing an O-methylated fucosyl residue. For instance highly conserved Rv2954c, Rv2955c and Rv2956 orthologs (BCG2975c, BCG2976c and BCG2977) are found in the DIM+PGL locus of M. bovis BCG. This mycobacterial species naturally produces a monoglycosylated PGL due to the lack of functional Rv1511, Rv1512 and Rv2958 orthologs; Consistently, transfer of these genes from M. tuberculosis in M. bovis BCG allows synthesis of PGL-tb in the resulting strain, indicating that M. bovis BCG possesses the enzymatic machinery necessary for the O-methylation of the fucosyl residue. These reactions are most likely performed by the products of BCG2975c, BCG2976c and BCG2977. In contrast no Rv2954c, Rv2955c and Rv2956 orthologs are present in the DIM+PGL loci of M. marinum and M. leprae, two mycobacterial species that produced PGL with no terminal O-methylated fucosyl residue. Finally, the involvement of Rv2956 in the O-methylation of position 2 of PGL-tb is supported by the genetic organisation of the DIM+PGL locus in M. kansasii. This chromosomal region contains several genes encoding putative O-methyltransferases but none of these display significant amino acid sequence homology to Rv2954c and Rv2955c. In contrast one gene encodes a protein exhibiting a high degree of sequence identity (about 84%) with Rv2956. Interestingly, the carbohydrate domain of the major form of PGL produced by M. kansasii consists of four sugar residues, the third one being a 2-O-Me-4-O-Ac fucosyl residue. The data presented in this study provide insights on the biosynthesis of compounds that are important for the pathogenesis of M. tuberculosis. The generation of mutants producing PGL-tb derivatives, and thereby p-HBAD derivatives, with a modified carbohydrate moiety could be useful tools to decipher the mechanisms by which these compounds act in the course of infection. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: RS GH PC CC. Performed the experiments: RS GH PC WM AL FL CC. Analyzed the data: RS GH PC MD CG CC. Contributed reagents/materials/analysis tools: RS GH PC WM AL FL. Wrote the paper: CC CG MD. [^3]: Current address: Present address: Unité de Pathogénomique Mycobactérienne Intégrée, Institut Pasteur, Paris, France [^4]: Current address: Present address: Laboratoire des Interactions Plantes-Microorganismes (LIPM), Castanet-Tolosan, France
# Introduction Tegumentary Leishmaniasis (TL) is a neglected tropical disease caused by different species of the genus Leishmania (Kinetoplastea: Trypanosomatidae), transmitted to vertebrate hosts by sand flies (Diptera: Psychodidae). TL is considered an emergent and re-emergent disease, since a worrisome increase in its incidence has been reported. On a the global scale, the number of new autochthonous TL cases reported annually to the World Health Organization (WHO) increased from 71,486 to 251,553 during 1998 to 2018. Several factors are involved with the spread of TL, such as human migration from rural to urban areas, conflicts and wars, disturbances in microenvironments due to climate change and human intervention and deterioration of socioeconomic conditions in endemic countries. TL comprises a broad spectrum of clinical manifestations ranging from single or multiple ulcerative skin lesions (cutaneous leishmaniasis—CL), to diffuse (diffuse leishmaniasis-DL) and mucosal (mucosal leishmaniasis—ML) lesions, with the last two being typical in the Americas. TL is associated with physical deformities and psychological alterations, affecting the health and wellness of the patient. The range of clinical manifestations can hinder rapid and accurate diagnoses, a key step to initiate treatment promptly and control the disease. Although several advances, TL-diagnosis remains based on the triad of epidemiological background, clinical signs and laboratory diagnosis, including direct and histopathological examination of skin biopsy and molecular detection of Leishmania DNA. Despite high specificity, low sensitivities have been described for direct and histopathological examination, especially in New World countries, where chronic cases and ML are frequent. Molecular techniques are complex, expensive, still without a standardized protocol for routine use and are restricted to reference and research centers. Therefore, these limitations make the TL-diagnosis scenario restricted, particularly in resource limited settings. In this sense, immunological tests may present remarkable advantages for TL- diagnosis, due to the use of less invasive sampling compared to skin biopsy and their potential to be automated, quantitative and used as point-of-care tests. The anti-Leishmania delayed-type hypersensitivity reaction, known as the Montenegro skin test (MST), has been the most used immunological test for CL- diagnosis in Brazil, even though it presents significant limitations such as positive results associated with previous leishmaniasis or asymptomatic infections. Nonetheless, the production of the MST antigen was discontinued in Brazil, hampering even more CL-diagnosis in the country. Other immunological tests, mainly Enzyme-Linked Immunosorbent Assay (ELISA), have presenting promising results in the Americas and beyond. Several studies using soluble Leishmania antigen (SLA) in ELISA for TL- diagnosis, have presented variable sensitivity especially due to antigen preparation and antigenic differences among Leishmania isolates and species. Moreover, reduced specificity due to the cross-reactivity with other infectious diseases has been frequently reported. Since CL-patients commonly produce low levels of anti-Leishmania antibodies, there is growing interest in high sensitivity antigens for immunological tests. Different methodologies have been employed, such as bioinformatics tools, cDNA expression library, phage display, immunoproteomic approach and isolation and purification of glycoconjugates to identify potential antigens. Furthermore, immunological tools have already been used to detect Leishmania antigens using monoclonal and polyclonal antibodies by immunochromatographic test (ICT) or immunohistochemistry (IHC), such as the CL Detect Rapid Test (InBios International Inc., Seattle, WA, USA), which detects peroxidoxin from Leishmania and has been used especially in Old World countries, with limited sensitivity. In this sense, we consider immunnodiagosis as potential tools to increase the access and improve TL-diagnosis. Although systematic reviews have been conducted on some aspects of this form of diagnosis, it is essential to identify potential antigenic targets that have been evaluated as TL-immunodiagnostic, point out knowledge gaps that still remain and encourage other studies to allow its application in clinical practice. In this way, we performed a worldwide systematic review to identify potential antigenic targets, with reported sensitivity and specificity, used as TL-immunodiagnostic. # Material and methods ## Protocol and registration The review protocol was registered in the International Prospective Record of Systematic Reviews (PROSPERO: CRD42020213311) and was developed based on the Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy. This review followed the Preferred Reporting Items for Systematic Reviews and Meta- Analyses (PRISMA). ## Information sources and study selection Structured searches were conducted in the following databases: MEDLINE, Virtual Health Library, Embase and Cochrane. A comprehensive list of key terms including tegumentary leishmaniasis and its different clinical forms AND immunological diagnosis or targets (antigens and antibodies) AND techniques or outcomes (sensitivity and specificity), was constructed in MEDLINE. Similar searches were adapted to each database. Complementary searches were performed by analysis of reference lists of selected articles. Searches were performed on 23<sup>rd</sup> March 2020, without restriction of publication date. ## Inclusion and exclusion criteria Original research articles reporting on the performance (sensitivity and specificity) of immunological tests based on the detection of antibodies or antigens using purified or recombinant proteins, synthetic peptides or polyclonal or monoclonal antibodies for diagnosis of human-TL, CL or ML were included. Exclusion criteria were: evaluation of serological tests based on SLA; only non-human samples were tested (e.g. canine samples); both sensitivity and specificity of the immunological tests were not presented or were impossible to be calculated; less than five samples were tested; the absence of information about the reference test and a non-specific Leishmania antigen was used. ## Selection process For each database, all publications were retrieved and duplicate citations were excluded by EndNote software. Based on the inclusion and exclusion criteria, two independent reviewers analyzed each publication by title and abstract using Rayyan software. Articles with no reason for rejection were included for full text reading. All discrepancies were solved by consensus after discussion. Selected studies were read in full to confirm their eligibility, to extract data or to exclude if exclusion criteria were identified during this step. ## Data extraction Data were independently extracted by two researchers (MLF and FDR) directly from full-length articles and were checked by a third researcher (EO). In case of disagreements, the final decision was reached by consensus. In this study, data were extracted and a 2x2 contingency table set up for immunological tests, containing the true positives, false positives, true negatives and false negatives. Furthermore, the following items were extracted: origin of the participants; the immunological test used; antigen or antibody types; Leishmania species and reference standard test used for disease confirmation. The phase of development of each study was classified according to Leeflang & Allerberger (2019). ## Study quality assessment The quality of the studies was assessed using the second version of Quality Assessment of Studies of Diagnostic Accuracy Approach (QUADAS-2). This tool allows a more transparent rating of risk of bias for studies included in systematic reviews on diagnostic accuracy. ## Data synthesis The performance of antigenic targets was presented in four groups according immunological tests and clinical form: 1) ELISA for TL; 2) ELISA for CL; 3) Other immunological tests for CL and 4) ELISA for ML. The performance outcomes for each antigen or antibody were sensitivity (probability of a positive test among cases or disease confirmed individuals) and specificity (probability of a negative test among controls or individuals without disease). Forest plots showing sensitivity and specificity values of all antigens, including 95% confidence intervals (CI) and Summary Receiver Operating Characteristic (SROC) curves were created using RevMan 5.3. Several studies considered a set of results for the same antigen (e.g. different cut-off points were available or different non-case groups were used in the analysis, such as healthy patients and those with other diseases). If possible, these results were grouped and only one sensitivity rate and one specificity rate including all evaluated patients. When impossible, we chose to present data that reflect the best field conditions (e.g. non-case group of patients with other diseases) or the better performance (e.g. cut-off point with best performance). # Results ## Literature search A total of 1642 articles from four databases were initially identified. Of this total, 261 were excluded due to duplicity (the same study was found in different databases). The title and abstract of each of the 1381 articles were checked and 139 were selected for full text reading. Finally, 98 articles presented exclusion criteria and so 38 were included. ## Descriptive analysis of included studies The characteristics of all included studies are presented in. In several studies, test performance was analyzed according to the clinical form (CL and ML) or globally (TL). In 19 studies, the antigenic targets were evaluated for TL-diagnosis, in 21 for CL and in 9 for ML. Sample size ranged from 26 to 500 patients. A total of three different immunological tests using purified or recombinant proteins, synthetic peptides or polyclonal or monoclonal antibodies were reported: ELISA, ICT and IHC. Different reference standard tests were used to confirm leishmaniasis cases. Thirty-one studies (81.6%) considered at least one parasitological method as a reference standard test, such as microscopy examination or in vitro culture for isolation of the parasite. On the other hand, seven studies (18.4%) considered some immunological or molecular tests as a reference standard. A total of 89.5% (34 out of 38) of the studies was classified as phase I (proof-of-concept), and the remaining 10.5% (4 out of 38) was classified as phase III. ## ELISA for TL diagnosis Nineteen studies used ELISA to evaluate the performance of a total of 56 antigens for TL-diagnosis, without specification of the clinical form (CL or ML). These studies evaluated 38 recombinant proteins, 14 synthetic peptides and 4 purified proteins. Forty-seven antigens were evaluated in studies that considered at least one parasitological method, such as microscopy examination or in vitro culture isolation of the parasite, as a reference standard test. The number of TL-patients ranged from 20 to 219 and the number of non-TL patients ranged from 8 to 281. The highest performance (100% of sensitivity and specificity) was reported for four recombinant proteins (cytochrome c oxidase; hypothetical protein XP_003886492.1; putative IgE histamine releasing factor; tryparedoxin peroxidase) and four synthetic peptides (A10, B7, C12 and H7) selected by the phage display technique. Nine other antigens were evaluated in studies that considered at least one immunological method as a reference standard test. For these antigens the sensitivity ranged from 39.8% to 76.9% and the specificity from 53.4% to 97%. The forest plots for sensitivity and specificity of ELISA considering parasitological methods and other tests as reference standard tests for TL-diagnosis are presented in ; more details about each evaluated antigen are available in. ## ELISA for CL diagnosis Seventeen studies used ELISA to evaluate the performance of 44 antigens for CL- diagnosis, which comprised 20 recombinant proteins, 13 synthetic peptides and 11 purified proteins. The performance of 35 antigens was evaluated considering at least one parasitological method as a reference standard test. Among these, the sample size for studies of CL-patients ranged from 12 to 74 and for non-CL- patients from 10 to 177. Peroxidoxin was the only antigen presenting 100% sensitivity and specificity. Nine antigens were evaluated considering at least one immunological test as a reference standard. Overall, HSP83 presented the highest performance (100% sensitivity and specificity). ## Other immunological tests for CL diagnosis The performance of ICT and IHC using different monoclonal and/or polyclonal antibodies is presented in and detailed in. Four studies evaluated the CL Detect Rapid Test (InBios International Inc., Seattle, WA, USA) in different countries. The sensitivity ranged from 35.6 to 67.6 and the specificity was higher than 80%. For IHC, two monoclonal antibodies were employed to detect antigens in fixed skin fragments. The highest performance was reported for IS22B4/XLVI5B8 mAbs, with 96% and 100% sensitivity and specificity, respectively. ## ELISA for ML results Nine studies used ELISA to evaluate the performance of 23 antigens for ML- diagnosis, which comprised 19 recombinant proteins, three synthetic peptides and one purified protein. The sample size from ML-patients in these studies ranged from 14 to 53 and from non-ML-patients from 20 to 92. At least one parasitological method was used as a reference standard test for the evaluation of sixteen antigens. The highest performance was obtained for Hypothetical protein XP_001467126.1, with 100% sensitivity and 98% specificity. Seven antigens were evaluated in studies considering at least one immunological test as a reference standard. As noted for CL-diagnosis, 100% sensitivity and specificity were reported for HSP83. The performance of these antigens is presented in and more details are available in. The SROC curves with the antigen performances for the diagnostic of different clinical forms, using parasitological or other tests (such as ELISA and MST) as a reference standard, are presented in. The antigens tended to have greater accuracy in studies that have used the parasitological methods as reference standard tests, regardless of TL-clinical manifestation. ## Quadas-2 based quality assessment Quality assessment of the study according to risk of bias and concern with applicability (low, high and unclear) is shown in. Of the 38 studies assessed, 21 had high risk of bias in patient selection. The risk was unclear for the index test in 34 studies and for flow and timing in 30 studies. Nineteen studies had high concerns regarding applicability of patient selection criteria. # Discussion TL is considered a multifactorial disease, responsible for psychological and social impacts due to scars and mutilating lesions generating stigma and self- deprecation in affected patients. Improvements in healthcare access and laboratory diagnosis are needed to overcome the impacts of this disease and should be encouraged. According to WHO’s Special Programme for Research and Training in Tropical Diseases (TDR), the ideal test must be affordable, sensitive, specific, user-friendly, rapid, equipment-free and delivered to end- users (ASSURED). Immunological tests may fill these criteria since they are usually easy to perform, accessible and require minimally invasive sample collection. Therefore, the identification of sensitive and specific antigenic targets seems to be a promising step toward the improvement of TL-diagnosis. The studies analyzed here were conducted from 1996 to 2019, however, almost 50% of them were conducted in the last five years, mostly in Brazil or another country in the Americas. The increase in the number of studies is coincident with the interruption of the production of MST antigen in Brazil in 2015, which extinguished the simple and rapid immunodiagnostic for TL. This fact may have boosted research aimed at finding new diagnostic tools. Parasitological diagnosis was considered a reference standard test in 89.5% of the studies. Despite this technique being highly specific for TL-diagnosis, its sensitivity is limited and inversely correlated with disease duration. However, no test seems to present sufficiently high sensitivity and specificity to be used as a gold standard test. We observed a tendency for index tests to be more accurate if parasitological tests were used as a reference standard than other reference test such as MST and histopathology. Polymerase chain reaction (PCR) was used as a reference standard test in 18 studies, generally in association with parasitological diagnosis. Overall, PCR appears to be a more suitable reference test, however, a standard protocol is urgently needed and encouraged, since distinct extraction methods, protocols and molecular targets have been used overtime. The ability to accurately identify TL-patients is essential for a diagnostic test, in view of the range of clinical forms, disease severity and treatment toxicity. Several studies have included patients with Chagas disease as non-TL cases, however, despite phylogenetic proximity, the inclusion of patients with clinical signs that do not resemble TL is at least questionable. For tests with diagnostic purposes, a better sample panel needs to be encouraged, including diseases such as sporotrichosis, paracoccidioidomycosis, hanseniasis, vasculitis, syphilis and other dermal or mucosal diseases, that represent confounding factors in clinical practice. Despite the distinct profiles in immune response usually reported for each clinical form of TL, some antigens presented high values of sensitivity, even for CL-patients. In general, higher levels of antibodies have been reported for ML-patients compared to CL-patients, the latter being characterized by a moderate Th1 immune response. In this way, it seems that problems related to antibody detection in CL-patients may be reduced by using sensitive targets and well-standardized procedures. Some antigenic targets were evaluated for TL- diagnosis without distinction of clinical form and, consequently, immune response profile. We believe that the accuracy of these antigenic targets may be improperly estimated in these specific cases. This systematic literature review found 79 different antigens, comprising 40 recombinant proteins, 24 synthetic peptides and 15 purified proteins. The identification and more refined selection of protein targets using recombinant proteins or synthetic peptides allows the development of more standardized techniques due to the possibility of generating the purest inputs. Some protein- families have been widely evaluated as antigenic targets for TL-immunodiagnosis, such as heat shock proteins (HSPs), histones and peroxiredoxins, with promising results. HSPs represent a highly conserved family of intracellular proteins of varying molecular weights in prokaryotic and eukaryotic cells, including cytosolic, mitochondrial, nuclear and endoplasmic reticulum resident proteins. They act as a chaperon in peptide folding and in the translocation of proteins to organelles, the prevention of protein aggregation, and the stabilization and degradation of proteins. HSPs have usually been identified by amino acid sequence homology and molecular weight, with HSP70 and HSP83 being the most abundant. These proteins are constitutively expressed throughout the life cycle of Leishmania, increasing expression in the vertebrate host due to variation in temperature and pH. The recombinant proteins HSP70 and HSP83, and the synthetic peptides extracted from those proteins, have been widely evaluated for TL-diagnosis. The performance of these targets seems to be promising, with HSP83 presenting sensitivity of over 90% and high specificity with few cross reactions. Histones are conserved proteins bound to DNA establishing chromatin structure in eukaryotes. Several biological functions have been described for histones during Leishmania infection in susceptible hosts. Core nucleosomal Leishmania histones have been proposed as prominent intracellular pathoantigens, since immunological responses against histones seem to be involved in the pathological mechanisms of visceral leishmaniasis (VL). In this way, this protein family has been extensively employed in ELISA for both human and canine VL. The presence of antibodies against rH2B of L. peruviana, rH1 of L. braziliensis and rH2A, rH2B, rH3 and rH4 of L. infantum have been detected in sera from CL or ML patients. CARMELO et al. (2002) demonstrated that the antibody against histone H1 was specific for the parasite without cross reaction with human histones. However, moderate cross reactivity has been observed in a sample panel composed of Systemic Lupus Erythematosus (SLE) and Chagas disease. Peroxidoxin, also known as thiol-specific antioxidant protein, as well as tryparedoxin peroxidase, are peroxiredoxins, an antioxidant enzyme family. This protein family has been described in a wide variety of organisms and several biological functions have been reported for Leishmania parasites, such as virulence factor and protection against reactive oxygen and nitrogen species. In this manner, they are directly associated with cell proliferation, senescence, apoptosis, and circadian rhythms. These proteins have been described in the secretome of L. braziliensis and the antigenicity of tryparedoxin peroxidase has also been evaluated for both human and canine VL-diagnosis. Peroxidoxin is the protein target identified by the CL Detect Rapid Test (InBios International Inc.) for CL-diagnosis. Variable performance has been reported for this ICT, according to endemic region and, consequently, the Leishmania species involved, with better results for infections caused by L. tropica, with sensitivity ranging 65.4–73% and specificity 92–100%. This test, however, has not been evaluated in Brazil. Other recombinant proteins, such as cytochrome c oxidase, putative IgE histamine releasing factor, prohibitin, eukaryotic initiation factor 5a, cathepsin L-like peptide and small myristoylated protein-3, as well as hypothetical proteins, were evaluated in preliminary studies demonstrating potential as candidates for TL-immunodiagnosis, and so more studies are desirable. Some promising synthetic peptides have been identified and employed in ELISA. The use of small fragments containing potent antigenic determinants is able to minimize non-specific reactions. LINK et al. (2017) identified three peptides by phage display, probably from GP63 glycoprotein, and presented 79% sensitivity in ELISA. COSTA et al. (2016) found high performance for three clones (A10, C12 and H7) in discriminating TL-patients from patients with other diseases and healthy individuals (100% sensitivity and specificity). However, these short linear peptides may have some drawbacks, such as limited passive adsorption on polystyrene titration plates (ELISA-standard procedure), inability to identify serum antibodies that recognize conformational epitopes and problems considering reproducibility due to variation in inter-assay reactivity producing different batches. Despite the advantages, the absence of post-translational modifications of bacterially-expressed and chemically synthesized proteins comprises an important limitation for the employment of this biotechnology for immunodiagnosis. In this way, purified proteins can represent significant advantages, especially regarding immunoreactivity. This review found iron-superoxide dismutase to be a purified protein with interesting results, with more than 80% sensitivity for CL or ML diagnosis. However, being purified proteins, sensitivity and specificity may vary according to the type, source, and purity of the antigen used. Three polyclonal and monoclonal antibodies were evaluated for detecting Leishmania antigen by ICT and IHC. The phase III studies included were ICT tests, that is, prospective studies in which the index and reference test were performed simultaneously in patients with clinical suspicion. This is a commercial test that, despite its low sensitivity, has been useful in some localities due to the simple realization and high specificity, reducing the number of CL patients referred for diagnosis confirmation. High performance was observed in phase I studies for species-specific monoclonal antibody (IS2-2B4—A11/ XLVI-5B8-B3) employed in IHC, with 96% sensitivity and 100% specificity. More robust studies using monoclonal or polyclonal antibodies for TL-diagnosis need to be encouraged evaluating the performance in clinical practice. The strength of the present literature review is that it employed a comprehensive search strategy with four databases. One of the meaningful limitations may be the limited number of studies evaluating the same protein target, and so results need to be interpreted with caution. For this reason, a meta-analysis was not performed here. Additionally, it is important to consider that the risk of bias for many of the included studies was unclear and/or was high for some of the evaluated parameters: “Patient selection”, “Flow and Timing” and "Index test". Here, we identified a large number of antigenic targets that could help clinical diagnosis. However, the high number of proof- of-concept and phase I studies highlights the need to move forward with more refined and mainly prospective studies including patients with clinical suspicion of TL from different endemic regions and the most sensitive reference standard tests, to evaluate the diagnostic accuracy of antigenic targets reported in clinical practice. # Supporting information We are grateful for support from the Programa de Pós Graduação em Ciências da Saúde of the Instituto René Rachou and Coordination for the Improvement of Higher Education Personnel (CAPES). [^1]: The authors have declared that no competing interests exist.
# Introduction Colorectal cancer (CRC) is the third most commonly diagnosed cancer in males and the second in females, with over 1.2 million new cancer cases estimated to have occurred worldwide in 2008. Clinically, distant tumor dissemination and metastasis are the most important factors in prognosis: whereas non-invasive stage I carcinomas present a 90% five year-survival, stage IV carcinomas with distant metastasis correlate with a dramatic drop to a 10% survival rate. Consequently, new therapeutic strategies improving the efficacy against metastatic disease and accurate biomarkers for the follow-up of CRC patients are major challenges, together with early detection and screening in high-risk populations. The spread of cancer relies on the detachment of aggressive malignant cells from the primary tumor into the bloodstream as a principal source of the further metastasis. It is widely accepted that Circulating Tumor Cells (CTC) own or acquire the ability to evade the host immune system and to reach a distant organ, usually the liver in CRC, where they establish a secondary tumor growth site in a highly inefficient but dramatic process. Concordantly, the presence of CTC in peripheral blood has been associated with poor prognosis in different types of cancer, including CRC. As presumptive founders in the generation of metastasis, CTC are becoming a field of interest, and the understanding of their biology may open new perspectives in oncology. Concerning their molecular characterization, in the last few years a number of groups have presented expression data focused on specific genes or signalling pathways related to cancer for the improvement of sensitivity and specificity of the detection. Smirnov et al. approached the profiling of CTC through a miscellaneous breast, prostate and colorectal metastatic perspective. In addition, data is emerging on the possibility of studying the biology and the utility to evaluate targeted therapies based on the genomic profiling of CTC. Within this scenario, defined by a limited efficacy of current chemotherapies in the treatment of metastatic CRC (mCRC) and CTC as key players in the management of metastatic disease, specific molecular profiling of the CTC population was approached. The combination of CTC EpCAM-based immunoisolation and accurate extraction of RNA from the very small number of CTC plus whole transcriptome amplification, made it possible to hybridise cDNA from the CTC population onto gene-expression microarrays. By applying this procedure to a group of mCRC patients compared to the background of unspecific isolated hematopoietic cells, the population of immunoisolated CTC was profiled specifically. In addition to the molecular characterization of CTC for the understanding of the biology of a main source of metastasis in CRC, these data provided potential therapeutic targets and diagnostic/prognostic biomarkers. # Results ## CTC Immunoisolation and Molecular Profiling The procedures for CTC immunoisolation, RNA extraction and amplification for hybridisa;tion onto cDNA microarrays are depicted in. Briefly, CTC were immunoisolated from 7.5 ml of peripheral blood from stage IV mCRC patients (n = 6;). Magnetic beads were used, which were coated with a monoclonal antibody towards the human Epithelial Cell Adhesion Molecule (EpCAM), a surface molecule highly expressed in epithelium-originated tumors such as CRC. RNA from isolated CTC was purified using a kit specifically designed for low abundance samples. In parallel, the same protocol was applied to blood samples from healthy donors (n = 3) to establish the baseline of background from unspecific non-CTC immunoisolation. Prior to gene expression analysis, the presence of isolated CTC was confirmed by direct immunofluorescence visualization using a cocktail of antibodies against cytokeratins 8, 18 and 19, and by a combination of two biomarkers validated for the accurate quantification of CTC in mCRC patients (GAPDH-CD45;) (Mann-Whitney U, p-value \<0.05). Moreover, the accuracy of the methodology assessed as the recovery rate during immunoisolation resulted in a median of 91.56% (Methods S1). In order to characterize the CTC population isolated from mCRC patients, the methodology described by Gonzalez-Roca et al. was adapted, for accurate gene expression profiling of very small cell populations. Basically, purified RNA was further treated with DNaseI, amplified using the WTA2 whole transcriptome amplification method, and complementary DNA was labeled and hybridized onto Agilent expression arrays (Gene Expression Omnibus, GEO. Accession number: GSE31023). After the initial pre-processing of raw data, an average of 21,070 spots were filtered according to the criteria described in Materials & Methods, which represented 47.35% of the spots in the microarray with a maximum of 32,443 and a minimum of 13,247. Upon filtration, the % coefficient of variation (CV) for replicated probes ranged from 5.98% to 12.13%. Normalization within each microarray was carried out using the Loess method, which assumes that most genes in microarrays are not differentially expressed in comparison to the control, so normalization among all microarray data was performed by the Aquantile method implemented in the Limma package of the R statistical software. This method ensured that the A values (average intensities) had the same empirical distribution across microarrays whilst leaving M values (log-ratios) unchanged. Correlation analysis of RTqPCR GAPDH-CD45 levels before and after amplification indicated a robust linearity during the process of transcriptome amplification (Spearman coefficient: 0.8667; p-value \<0.005). The next step involved gene-expression profiling of CTC isolated from mCRC patients against the background of contaminating cells during the immunoisolation. For this, normalized gene-expression intensities from mCRC patients and from the group of controls were processed using MeV (MultiExperiment Viewer) software (see Methods S1). The signals obtained from healthy controls were considered as the background from non-specifically isolated blood cells, which were mainly lymphocytes. The signal obtained from mCRC patients represented the sum of this non-specific background plus the specific gene-expression pattern of the CTC. By subtracting this background, the contaminating non-CTC population was removed, and the resulting genes showing statistically significant expression were considered to characterize the CTC population from mCRC patients. Concordantly, all significant genes presented positive expression in mCRC patients upon subtraction of the background from healthy donors, which was consistent with the presence of CTC only in the mCRC samples. This strategy led to the identification of a final set of 410 genes that were specific to the CTC population, with hierarchical clustering analysis clearly discriminating between mCRC patients and controls. Bioinformatic analysis with Ingenuity Pathway Analysis (IPA) and Genecodis (for Gene Ontology) software served to interpret the resulting list of genes characterizing the CTC population. The main cellular functions defined by these CTC-specific genes were related to cell movement, cell adhesion, cell death and proliferation, cell-cell signalling and interaction, and cytoskeleton reorganization. Interestingly, the IPA analysis of canonical pathways, which was representative of these genes, also highlighted a number of known signaling pathways involved in cell migration/invasion and cell adhesion, as Protein Kinase A, RhoA, Integrins, ILK, or Actin cytoskeleton signaling molecules. Moreover, the analysis of gene-gene interactions rendered biological networks that characterize the CTC population from mCRC patients. Among them, cancer and cellular movement and morphology were found to be the principal events associated with a CTC phenotype, while this interpretation might be biased by the important contribution of cancer to databases. All of these analyses point to a balance among the genes underlying cell survival, interaction with the environment and cellular movement as the fundamental biological processes that must converge in the population of CTC for the successful development of metastasis in CRC. ## Real-Time Quantitative PCR Validation and Identification of Diagnostic and Prognostic Biomarkers The next step involved the RTqPCR validation of eleven genes presenting high log2 ratios in mCRC patients and a functional relevance in the biological processes characterizing the CTC population in an independent series of mCRC patients and controls. The specific expression of these genes was evaluated by comparing the CTC populations from the group of mCRC patients (n = 20) with the background from healthy controls (n = 10), once they had been normalized to CD45 as a marker of unspecificity during CTC isolation. These genes included APP, CLU and TIMP1, which are associated with cell death and anti-apoptotic activity; VCL, ITGB5, BMP6 and TGFβ1, which are genes involved in cell migration and invasion and cellular morphology; and TLN1, ITGB5, LIMS1, RSU1 and CD9, which are associated with cell adhesion. As observed in (black bars), all of the selected candidate genes were validated in this new series of samples by RTqPCR with significant differences between the group of patients and the group of controls (p-value \<0.0001). Importantly, to ensure that the expression of the selected genes was characteristic of CTC and not an artefact due to the variation of gene expression in lymphocytes in cancer patients, their expression was compared in the remaining non-isolated fraction upon CTC enrichment in a set of mCRC patients (n = 5) and controls (n = 5). As shown in (white bars), no differences were found between both groups for the eleven candidate genes, reinforcing their specificity of expression in the CTC population. When five samples in the CTC isolated fraction were randomly selected, significantly higher levels were consistently found in mCRC samples, demonstrating that the lack of differences between patients and controls in the non-isolated fraction was not due to sample size artefacts (data not shown). To further evaluate the involvement of the candidate genes in the metastatic potential of CTC in CRC, their expression was compared in a series of primary carcinomas (n = 14) and lung (n = 7) and liver (n = 7) metastases. As shown in, all of the selected genes were up-regulated in metastases compared to primary lesions, with seven out of the eleven genes presenting statistical significance. Among them, CLU and TIMP1 presented a specific up-regulation in liver metastasis, suggesting a potential role for these genes in the tissue-specific ability of mCRC CTC to colonize the liver. In addition, three out of the eleven selected genes demonstrated a significantly increased expression at the invasive front of the primary tumors compared to the paired non-invasive superficial area, suggesting a link to the acquisition of an aggressive phenotype. All of these results validated the strategy to molecularly profile the CTC in mCRC patients, confirming the presence of the CTC population and the efficient immunoisolation of CTC from mCRC patients, as well as the capability to specifically characterize these cells at the molecular level. Finally, the molecular profiling of CTC should result in the identification of potentially reliable biomarkers for the detection of the CTC population from mCRC patients. To evaluate this point, the accuracy of the RTqPCR validated genes in the detection of metastatic disease was evaluated in the series of mCRC patients and controls. As shown in, all of the validated genes presented excellent discriminating values in terms of AUROC (values ranging from 0.94 to 1; p-value \<0.0001). Moreover, and with the exception of CD9 and TLN1, the selected biomarkers demonstrated an excellent behavior as predictors of disease progression. These data guarantee further studies in large cohorts of patients for the use of these biomarkers in the clinical setting. # Discussion Molecular profiling is widely employed as a powerful strategy to characterize specific types of tumors, specific subtypes of carcinomas or specific events associated with carcinogenesis. In addition to contribute to the understanding of molecular events associated with the genesis and progression of cancer, gene- expression profiling has led to the identification of therapeutic targets and biomarkers in an effort to improve the management of cancer patients at the clinical setting. In this work, we aimed to tackle with a very specific and attractive population of tumor cells that are at the origin of metastasis, the circulating tumor cells (CTC). A combination of CTC immunoisolation, accurate RNA extraction from very low number of cells, whole-genome amplification and massive gene-expression profiling for the characterization and interpretation of the biology of CTC in metastatic colorectal cancer is presented here. To technically validate this strategy, the profiling of CTC from mCRC patients was approached by subtracting the background of non-specific isolation from a group of healthy controls. The next challenge is to compare the gene-expression profile of the CTC population with the primary CRC lesions and with the overt metastases, in order to better understand the mechanisms of adaption of tumor cells during the process of metastasis and the crosstalk with the environment. The subtraction of the background from a group of controls was also considered to be more relevant than the background from the same mCRC patients as it would require two rounds of CTC isolation within the same samples, representing a significant source of technical artefacts. Differences in the background could not be excluded due to the systemic metastatic disease, although the analysis of the remaining fraction after CTC isolation did not render any differences within the selected genes. Regarding immunoisolation, and although CTC enrichment with EpCAM-coupled antibodies has demonstrated to be superior to other cytometric methods and a reliable method for CTC detection in mCRC patients, this CTC capture procedure has raised some debate due to the reliance of this technique on the expression of EpCAM. Initially described 30 years ago as a dominant antigen in human colon carcinoma tissue, it is assumed that a decreased expression of epithelial markers occurs during the epithelial to mesenchymal transition (EMT) associated with tumor invasion. Importantly, EpCAM is apparently needed to maintain distinct cancer cell attributes and, potentially, the cancer stem cell phenotype. CD133+ cells, currently one of the best markers for the characterization of colon cancer stem cells and an independent prognostic marker that correlates with low survival, are positive for EpCAM. Likewise, antibodies against EpCAM can efficiently target colorectal Tumor-Initiating Cells, conferring a considerable value to the EpCAM-isolated CTC population in terms of therapeutic intervention. The data presented here conclusively demonstrated the effective isolation of CTC from mCRC patients. A main achievement of this work was the translation of the method described by Gonzalez-Roca et al. for the accurate expression profiling of very small cell populations, into a clinically relevant cell population. Whole transcriptome amplification allowed the characterization of the CTC population isolated from metastatic CRC patients at the molecular level. To date, the majority of studies have described the expression levels of a limited number of genes in different CTC populations, principally for detection purposes. The approach used here allowed the molecular profiling of CTC from mCRC, and their definition as a population of cells with presumed migratory and adhesive capabilities. This is consistent with a subpopulation of tumor cells that must acquire an aggressive and invasive phenotype allowing dissociation from the primary lesion, leading to the invasion of the surrounding stroma and their intravasation and survival in the blood flow. In addition, these cells may also extravasate and successfully implant in the distant tissue target to generate a micrometastasis. The list of genes phenotypically characterizing the CTC population from mCRC patients includes candidates related to all the above described functions, warranting further studies to define their implication in the metastatic behavior of the CTC population in CRC. Genes such as VCL, ITGB5, BMP6 or TGFβ1 have been associated with the acquisition of an invasive phenotype, in part through EMT. Likewise, TLN1, APP, CD9, LIMS1 and RSU1 have been related to adhesion and migration, with CD9 modulating the localization of TLN1, a critical regulator of integrin activation, to focal adhesions, or LIMS1 being associated with cell adhesion and integrin signaling, and, in particular, with RSU1 and the Ras pathway during cell migration. A key step in the process of cancer dissemination is the ability to migrate and to progress towards and intravasate nearby blood capillaries. Once in the circulation, CTC must overcome the host immune system by enhancing their apoptosis resistance among other mechanisms. Genes like TIMP1 and CLU are key molecules related to this process. Interestingly, TIMP1 has been linked to the process of anoikis resistance in different cancer models, enabling the evasion of cell death in the absence of substrate anchorage. Upon reaching a target organ, CTC must be able to extravasate and form colonies, and CD9 has been described as critical to the process of implantation and micrometastases formation associated with stem cell attributes, which is related to these events. Interestingly, a significant overlap was found with a liver-specific metastasis signature in CRC, suggesting an active implication in the tropism and micrometastasis potential of CTC to this organ in CRC. In particular, and consistent with the correlation of TIMP1 expression in CTC and liver metastasis, it has been described as a regulator of the liver microenvironment, increasing the susceptibility of this organ to tumor cells. Overall, the involvement of the selected genes in the metastasis capacities of CTC has also been reinforced by their up-regulation in lung and liver metastases compared to primary colorectal carcinomas. In vitro primary CTC cultures and in vivo models for CTC in CRC will definitely provide researchers with robust evidence to validate the role of these genes in the ability of CTC to generate micrometastases, and to define strategies specifically targeting this population of metastatic CTC. Likewise, in terms of biomarkers for the management of mCRC patients, the molecular profiling of CTC from mCRC patients rendered valuable tools for the detection and quantification of metastatic disease, as well as for the prediction of disease progression. The high specificity and sensitivity demonstrated by the above described validated genes in the CTC population from mCRC patients warrant the study of these biomarkers in the evaluation of therapeutic responses or in the selection of patients for targeted therapies. In conclusion, this study described, to our knowledge, the first specific molecular profiling of CTC isolated from mCRC patients. This gene-expression analysis was applied to the characterization of this specific population with presumed adhesive, migratory and invasive capabilities, as well as a modulated response to cell death for the successful completion of the process of metastasis. Moreover, the identification of therapeutic strategies specifically targeting the CTC population should improve the efficacy in the eradication and prevention of CRC metastasis, while an armamentarium of highly specific and sensitive valuable markers should impact on the management and follow-up of metastatic CRC patients. # Materials and Methods ## Patients All participants signed an informed consent specifically approved for this study by the Ethical Committee of the Complexo Hospitalario Universitario of Santiago de Compostela (code of approval: 2009/289). Inclusion criteria for mCRC patients were the presence of measurable mCRC and an Eastern Cooperative Oncology Group (ECOG) performance status not greater than 2. Healthy controls with an absence of a previous cancer episode and an age matched with patients were selected. Detailed information about patients included in the analysis is available in. ## CTC Immunoisolation CTC were isolated by using the CELLection™ Epithelial Enrich kit (Invitrogen, Dynal) according to manufacturers’ instructions. Briefly, 7,5 mL of blood from mCRC patients and healthy controls were incubated for 30 minutes at 4°C with 100 µl of magnetic beads. After washing, CTC coupled to the magnetic beads were directly resuspended in 100 µl of RNAlater® solution (Ambion) and stored at −80°C until processed for RNA extraction. ## Gene Expression Analysis Total RNA from CTC was extracted with the QIAmp viral RNA mini kit (Qiagen), specifically designed for very low cellularity samples. Purified RNA was next subjected to a whole transcriptome amplification reaction (WTA2, Sigma Aldrich), Cy3 labeled and hybridized onto Agilent 4×44 k gene-expression arrays. Signal processing and filtering as well as gene expression data analysis are described in detail in Supplemental Material. Gene expression data is accessible at the NCBI Gene Expression Omnibus (GEO) database (Accession Number: GSE31023). ## Real-Time Quantitative PCR Validation RTqPCR validation was performed in an independent set of 10 healthy controls and 20 stage IV metastatic colorectal cancer patients. CTC were isolated as described, and RNA was purified with a RNA carrier to improve yield and stability. cDNA was synthesized by using SuperScriptIII chemistry (Invitrogen) following manufacturer’s instructions. To further optimize the sensibility of detection, we performed a preamplification step by using the TaqMan® PreAmp Master Mix kit (Applied Biosystems) with 14 reaction cycles. Preamplified products were subjected to TaqMan® real-time PCR amplification for eleven candidate genes (see for assay details). In addition, we evaluated by RTqPCR the expression levels of the candidate genes in the cellular fraction remaining upon CTC immunoisolation, in 5 mCRC patients and 5 healthy controls. Briefly, we performed a red blood lysis in the fraction remaining after EpCAM immune-bead incubation and CTC isolation, and RNA was purified from the remaining nucleated blood cells with the RNeasy mini kit (Qiagen). RNA quantity and purity were assessed by NanoDrop measurement, cDNA was synthesized using MuLV reverse transcriptase system (Applied Biosystems) and TaqMan® qPCR was performed for the eleven candidate genes. Expression values for each gene were normalized to CD45 as a marker of non- specific isolation. Fold change differences between patients and controls were statistically analyzed with the GraphPad Prism software, applying the Mann- Whitney non-parametric t-test and considering a p-value\<0.05 as significant. ## Gene Expression Analysis in Primary Tumors and Metastases Primary colorectal carcinomas (n = 14) and metastasis (liver metastasis, n = 7; lung metastases, n = 7) were processed by the Pathology Department of the Complexo Hospitalario Universitario of Santiago de Compostela. The superficial non-invasive zone and the deep invasive area of the primary tumors were macroscopically dissected, ensuring similar tumor percentages. RNA was purified (TRIZOL reagent, Invitrogen; RNeasy kit, Qiagen), cDNA was synthesized (MuLV reverse transcriptase, Applied Biosystems), and gene expression was evaluated (TaqMan RTqPCR, Applied Biosystems). Data was represented as fold change relative to the expression in the superficial non-invasive area. GAPDH was used as loading control. Non-parametric Mann-Whitney t-test was used for statistical significance considering p-value \<0.05. ## Area under the ROC Curve (AUROC) and Kaplan-Meier Analysis The accuracy of the eleven candidate genes as biomarkers for diagnosis was evaluated with ROC curves, while Kaplan-Meier curves were built for their evaluation as markers for prognosis. Cutoff values were set as the best-fit values for each marker. Patients Progression Free Survival times were established as the time elapsed between chemotherapy line start and disease progression evaluated by imaging, or patient dead by any cause. # Supporting Information We gratefully thank the patients for their willingness to participate in the study and the nurses from the Medical Oncology Department for their collaboration. [^1]: Conceived and designed the experiments: JB RL-L MA. Performed the experiments: JB AA LA-A. Analyzed the data: JB MV AGT MAC RL-L LMR MA. Contributed reagents/materials/analysis tools: EG SC IA BF MA-N. Wrote the paper: JB MA. [^2]: The authors have declared that no competing interests exist.
# Introduction Humans often use un-observable variables like beliefs, desires, and intentions to disambiguate agents’ behavior, attributing mental states to other people and to oneself. These mentalizing abilities emerge during early childhood and variations in mentalizing skills appear to be related to social environmental factors. Among these factors, collaborative experiences of a child with adult group members might play a crucial role. These interactions might allow children to gradually construct knowledge of the world, as well as knowledge of other people’s mental states, by capturing cognitive regularities that cooperative agents try to make transparent to the child. Eventually, children start using this knowledge to manipulate the mental states of other agents during referential communicative interactions. For instance, 4-year-old children use presumed knowledge of an interlocutor to select linguistic behaviors designed to change those mental states, producing more explicit descriptions of a toy when speaking to a blind as compared to a non-blind addressee, and simpler utterances towards a toddler than an adult. Five-year-old children can produce verbal requests that take into account the presumed knowledge of their interlocutor. However, it remains largely unknown how children learn to adjust their referential communicative behaviors to their mental model of an addressee. Here we elaborate on the suggestion that the extent and nature of the social interaction children experience will influence the development of children’s social understanding. Humans are exceptional among existing hominids for experiencing early developmental exposure to cooperative nonkin, i.e. conspecifics that lack a genetic reason for collaborating, and it has been suggested that this developmental feature might boost motivational predispositions to share mental states with others. We quantify one aspect of this faculty through audience design, i.e. adjustments of communicative acts to the presumed abilities and knowledge of an interlocutor. Given that audience design presupposes control of the ability to share mental states with others, we focus on five-year-old children, i.e. children with fully-fledged theory of mind capacities. We quantify developmental exposure to two main sources of social interactions experienced by children between zero and four years of age, namely familial and non-familial experiences. The former were quantified in terms of years of experience with siblings, and parents’ level of education. The latter were quantified in terms of days per week of attendance to daycare. Audience design effects were quantified in a controlled experimental setting involving the production of referential non-verbal behaviors with a communicative goal, exploiting a protocol previously validated in adults. In contrast to linguistic communication, the communicative behaviors evoked under these experimental conditions could not be directly based on previous concrete experiences. Accordingly, the novel communicative situation experienced by the children in this study allowed us to directly tap into their ability to influence the mental states of others through behaviors generated ex-novo. Five- year-old participants were told they were playing an online interactive game with a 2-year-old toddler and with a same-age peer, in alternation. In fact, a confederate performed the role of both addressees, while remaining blind to which one of the two roles he was performing in any given trial. Accordingly, both performance and response times of the two presumed addressees were matched. This feature of the protocol allowed us to test whether the mere belief that the child is communicating with addressees of different ability induces internally generated adjustments in the child behavior, over and above performance-related mutual adjustments. Furthermore, the precise quantification of children behavior afforded by this protocol distinguished between belief-driven adjustments restricted to the communicative components of the actions, and generic priming effects. These procedures allowed us to test whether the social environment experienced by a child early during his development influences his ability to adjust a self-generated communicative behavior to his mental model of the addressee. # Materials and Methods ## Participants The experiment was approved by the local medical ethical committee (ECG, Nijmegen, The Netherlands). Parents with 5-year-old children (N = 24, 12 females, mean age 5.09, range 5.02–5.16) were recruited from a database of the Baby Research Center Nijmegen. The children’s parents provided written informed consent for participation of their children in the study, and all participants received a book or monetary compensation for their visit. ## Experimental Design The game involves a Communicator (a 5-year-old participant, displayed as a bird on the game board) and an Addressee (the confederate, displayed as a squirrel) interacting on a digital game board with a 3×3 grid layout. On each trial, their joint goal was for the Addressee to collect an acorn from the game board. Given that knowledge of the acorn’s location in the game board was available to the Communicator only (on a printed copy of the game board, visible throughout the trial, see), a successful trial of this game required the Communicator to inform the Addressee where the acorn was located. Given the experimental setup, the Communicator could inform the Addressee only by moving the bird across the game board (event 2). The Addressee could then move the squirrel to the acorn’s location only by interpreting the meaning of the Communicator’s movements on the game board (event 3). For details on the experimental procedure see the Supplemental Material. By touching a square on the screen with his/her finger, the Communicator could move the bird token to that square, and this movement was also visible to the Addressee. The bird could only move to the center of each of the nine grid squares, and only through vertical or horizontal displacements. This feature of the task was introduced to create a spatial disparity between the movements of the bird and the potential locations of the target object (any of the thirteen white circles, see). Namely, the bird could not be overlaid on the precise location of the acorn when a square contained more than one white circle (see Manipulation of task difficulty of the Supplemental Material for details). The Communicator had no restrictions on planning time (event 1 in) or on movement time (event 2). The end of the movement epoch was marked by the return of the bird on the central square of the game board (nest). At this point, the token of the Addressee (the squirrel) appeared, in the center of the digital game board, visible to both players. The Addressee moved the squirrel to the location deemed appropriate given the movements of the Communicator (event 3). The Addressee had no temporal or spatial restrictions on the movements of the squirrel on the game board. Successful trials, in which the Addressee had moved to the location of the target, resulted in the presentation of a large acorn on the screen (event 4). A red “no” icon was presented over a small acorn for unsuccessful trials. There were a total of 50 trials, subdivided in blocks of five trials (∼35 min). Each child was informed that he would be playing an interactive game with two addressees in turns; either a toddler (‘2-year-old’) or a same–age peer (‘5-year-old’). They were told that the game partners were sitting in other rooms and that they could see the bird token and the digital game board on their monitors. There were two pairs of fictitious child-toddler addressees, two presentation orders of child-toddler addressees, and two sets of target configurations, counterbalanced over participants. ## Quantification of the Social Environment Given that the extent and nature of the social interactions experienced by children is widely thought to influence the development of their social understanding, we considered two main sources of social interactions experienced by children, namely familial and non-familial experiences, reconstructed from interviews with the parents of the children. Familial experiences were indexed with the parents’ level of education (11 levels, 7.4±1.6, group mean ± SD, range 4.5–10.5) and years of experience with siblings (i.e. the product of age and number of siblings: 4.3±3.4, range 0–15.2; number of siblings: 1.2±0.7, range 0–3). Non-familial experiences were indexed with the time spent at daycare (days per week) between the age of 0 and 4 (mean over these four years; 1.7±0.9 days per week, range 0.25–3). We did not consider between ages 4 and 5 given that in the Netherlands it is customary to start primary school at age 4. ## Data Analysis Audio- and video-recordings of the participant’s behavior were analyzed offline. Those trials in which the child behavior revealed procedural uncertainties (e.g. failing to return to the nest within 15 seconds, or interrupting the bird movements to look at the location of the acorn in the instruction game board) were excluded, leaving 80.1±13.4% (mean ± SD) of the original trials for further analysis (∼40 trials; four participants interrupted their performance after 30 trials). This study builds on the findings of a previous report involving the same task and obtained in a group of women, showing that the communicator’s belief about age of the addressee changed communicative behavior. More precisely, these adults spent longer time on communicatively relevant locations of the game board when interacting with a presumed child addressee (vs. an adult addressee), i.e. using time as a tool to place emphasis on target information. The first goal of this study was to replicate this finding in a group of five-year-old children. Accordingly, we considered the same dependent variable (namely, Time spent on game board locations), using the same statistical comparison, namely a two-way ANOVA with factors Addressee (Toddler, Child) and Location (Target, Non-target). The Time spent on game board location by the Communicator was calculated as the time interval between the first contact of the finger on the touch screen within the area of a square of the game board (either a Target or a Non-target location) and the subsequent contact of the finger within the area of a neighboring square of the game board. We considered the mean time spent on those location types per trial. It should be emphasized that, given the absence of temporal restrictions on the total time the children could spend on the game board, the time spent on target locations and the time spent on non-target locations could vary independently. Having replicated the findings of in this group of five year-olds, we used a multiple linear regression analysis to assess the differential contribution of familial and non-familial sources of social interactions experienced by these children in the first four years of their life. These three independent variables (i.e. parents’ level of education, years of experience with siblings, and time spent at daycare, see above) were jointly considered in the multiple regression analysis, with the degree of communicative adjustment observed in each child as dependent variable (i.e. the relative difference, \[toddler – child\]/\[child\], in time spent on Target locations between presumed toddler and child Addressee). This statistical approach allows one to make specific inferences on the inter-subject variance accounted for one variable, over and above the variance accounted by the other variables included in the multiple regression model. # Results ## Communicative Success The percentage of successfully communicated trials was 63.4±8.0% (mean ± SD). This is well above chance level (7.7%; 13 potential target locations). ## Communicative Adjustments We tested whether 5-year-old children are able to adapt their referential communicative behavior (event 2) to the presumed age, or cognitive level, of their interlocutor. A two-way analysis of variance revealed a significant interaction of the factors Addressee (Toddler, Child) and Location (Target, Non- target) on the mean time spent on game board locations during the movement epochs, F(1,23) = 5.4, p = .03. This interaction was driven by the fact that the 5-year-old children spent more time on the Target locations (containing the acorn) when they thought to be interacting with the toddler Addressee as compared to the child Addressee, t(23) = 2.6, p = .014, two-sided paired t-test. There was no difference between the two Addressee types for the mean time spent on the Non-target locations (other visited locations), t(23) = 0.04, p = .97; see. ## Effects of Social Environment We evaluated whether quantitative indexes of developmental exposure to social interactions of the child could explain inter-individual variability in the communicative adjustment observed over the whole group. A multiple linear regression analysis indicated that daycare attendance (i.e. mean days per week spent at daycare before starting school) predicted the communicative adjustments made by the 5-year-old participants, R<sup>2</sup> = .34, F(3,23) = 3.4, p = .039 (full model), Beta = .598, p = .005, R<sup>2</sup><sub>adj</sub> = .24 (daycare attendance); see. Parents’ level of education (Beta = −.14, p = .45) and years of experience with siblings (Beta = .04, p = .84) did not significantly account for inter-subject variance in communicative adjustments. # Discussion We have tested whether the expression of audience design abilities in 5-year-old children is modulated by their previous history of social interactions. Participants were asked to influence the behavior of an addressee, in an experimental setting where no pre-existing communicative conventions were immediately available. In fact, the communicative means made available to the children were purportedly limited, challenging them to devise new communicative behaviors that could be understood by the addressees. There are three main results. First, 5-year-old children were able to influence the mental states of others even at their first encounter with a novel communicative setting. This communicative behavior was internally generated by the children, and motorically different from the behavior of the two presumed addressees. Second, the mere belief of communicating with addressees of different ages selectively influenced the communicative behavior of the participants. The children spent longer at communicatively relevant locations when interacting with a presumed toddler addressee as compared to a presumed child addressee. This communicative adjustment was not a generic priming effect, being absent in communicatively irrelevant locations of the game-board. Third, the communicative adjustment observed in the children was predicted by the time spent at daycare during the previous years of their life. This latter finding refines the notion that human communicative skills might be shaped early during development, emphasizing the fundamental role of non-familial interactions in the gradual construction of children’s social understanding and abilities to influence the mental states of others. It has been suggested that children gradually construct mental variables through the regularities they experience within social interaction. In contrast to a large body of work focusing on verbal reports of children’s ability to attribute mental states to other people, as during Theory of Mind tasks, here we considered children’s ability to influence the mental states of others through non-verbal behaviors, i.e. the magnitude of their communicative adjustments. These spontaneous adjustments provided a sensitive index for quantifying inter-individual differences in communicative abilities close to the onset of those abilities. This sensitivity might arise from the implicit nature of the index of audience design used in this study, in line with findings previously obtained during language comprehension in children of similar age. Namely, in contrast to previous work exploring how a child’s inhibitory control handles the conflict between the knowledge of the child and that of the addressee, in this study we manipulated the presumed abilities of the addressee, minimizing demands on the control abilities of the child. The magnitude of communicative adjustments in 5-year-old children was predicted by the time spent in daycare during previous years of their life, over and above the effects accounted for by measures of the familial social environment (sibling experience, educational level of the parents). One possible mechanism accounting for this observation might relate to the importance that overheard communicative interactions have on the linguistic development of a child. Namely, kindergarten attendance might considerably boost the variety of children’s experience with this source of pragmatic inputs, enhancing their communicative skills. More generally, the structured social interactions afforded by a daycare environment (e.g. cooperative play, frequent integration of new group members) might provide the child with a larger set of communicative challenges than those experienced within a relatively stereotyped familial environment. These challenges might differ substantially from those experienced in a familial environment. In kindergarten, a child needs to communicate with a multitude of agents, and those agents lack a genetic reason for collaboration. Finally, kindergarten provides children with caregiving ‘alloparents’ that might boost their socio-emotional development. It remains to be seen how the present findings, showing stronger effects of non- familial over familial experiences on the development of referential communicative adjustments, can be reconciled with previous reports, showing that measures of familial interactions predicted ‘false belief understanding’, as assessed with verbal reports. One possibility is that the communicative adjustments observed in this study might be mainly driven by children’s assumptions on the presumed cognitive capacities of the addressees, rather than by children’s understanding that the beliefs, desires, or intentions of other agents differ from reality. Differences in outcome measures might also play a role, e.g. implicit measures of knowledge about a communicative interaction (as gathered through eye movements, reaction time, or movement times) vs. explicit verbal reports requiring a degree of executive control. This study opens the way for systematic and sensitive investigations into the contributions of early social experiences towards children’s communicative abilities, raising the possibility to chart the developmental trajectories generated by familial and non-familial social interactions (e.g. siblings, parents, non-sibling peers, alloparents) through longitudinal studies with objective measures of the time spent on those interactions. # Supporting Information We thank the Baby Research Center Nijmegen, Annelies van Wijngaarden, Evelien Akker, and the families who participated. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AS SH IT. Performed the experiments: AS. Analyzed the data: AS IT. Wrote the paper: AS SH HB IT.
# Introduction Gallid herpesvirus 2 (GaHV-2), more frequently referred to as Marek’s disease virus (MDV), is an alphaherpesvirus (type species of the genus Mardivirus) and the causative agent of a highly infectious lymphoproliferative disease termed Marek’s disease (MD) affecting many birds in the Phasianidae family. Despite global vaccination campaigns that are effective to prevent disease development, MDV field strains continue to spread in poultry and appear to evolve towards increased virulence. The dissemination of MDV in poultry is mediated by infectious viral particles associated with dander and feather debris. With the exception of the feather follicle epithelium, the site where free infectious viral particles are shed, the virus remains strictly cell-associated and progression of the infection is restricted to viral cell-to cell spread. The MDV particle is composed of a 180-kbp double-strand DNA genome packaged in an icosaedric capsid surrounded by a tegument layer, which insures the morphological and functional continuity between the capsid and the host cell derived viral envelope. By homology with other alphaherpesviruses, a number of viral proteins composing the tegument have been identified, including a major tegument protein, VP22 (pUL49), various trans-activators and two protein kinases (pUL13 and pUS3). The UL49-encoded VP22 protein is abundantly expressed in infected cells and is essential for MDV replication. VP22 is a specific tegument protein of alphaherpesviruses and conserved among this subfamily. To date, the absolute requirement of the UL49 gene for viral replication was initially demonstrated for MDV and afterwards for Varicella Zoster virus (VZV). The deletion of VP22 in other alphaherpesviruses including Herpes Simplex virus 1 (HSV-1), Pseudorabies virus (PRV), Bovine herpesvirus 1 (BoV-1) still allows viral replication, even though viral spread is reduced in some cell types. While its role in virus infection remains unclear, it was demonstrated for HSV-1 that VP22 interacts with and recruits various viral proteins, such as the trans- activators ICP0, ICP4 and viral glycoproteins composing the infectious virions. Furthermore, VP22 was shown to interact with cellular proteins involved in the organization of microtubules and nucleosome assembly. The VP22 protein encoded by MDV shares common functional features with VP22 encoded by other alphaherpesviruses. It was previously shown that MDV-VP22 shows both a cytoplasmic and nuclear location in infected cells and accumulates in the nucleus upon overexpression in cells. Moreover, MDV-VP22 exhibits a strong affinity to DNA, especially heterochromatin, and to microtubules. We previously demonstrated the role of VP22 in MDV cell-to-cell spread, which could explain the necessity of VP22 in MDV replication. It was recently shown that recombinant MDV viruses expressing VP22 with a C or N-terminal GFP-tag are highly attenuated in vivo suggesting that VP22 might play a role in MDV-induced lymphomagenesis. However, the precise role of VP22 in MDV replication and MD pathogenesis remains unclear. Notably, the functional significance of the VP22 nuclear distribution is still unknown, even if previous reports on VP22 encoded by alphaherpesviruses evoke a possible regulatory function of VP22 within nuclei. Virus infection frequently results in the disturbance of key cellular processes within the host cell. The subversion of cell cycle pathways is a well- established mechanism by which viruses create the most suitable environment for their replication. Especially, the induction of S-phase is either mandatory or at least advantageous for lytic replication of a number of viruses. The eminent role of cellular factors from the DNA synthesis machinery in viral replication was demonstrated for viruses from different families such as the Flaviviridae, Retroviridae, Parvoviridae, and Polyomaviridae. In contrast, herpesviruses encode their own DNA polymerase and accessory proteins, and thus theoretically do not require an S-phase environment to support their replication (reviewed in). Nevertheless, several studies have demonstrated the importance of the S-phase in the life cycle of VZV and Epstein-Barr virus (EBV). For EBV, S-phase cyclin-dependent kinase activity is essential for the expression of immediate early and early viral proteins and is thus required for viral replication. Vice versa, EBV lytic replication is able to provide a S-phase-like cellular environment by modulating DNA damage pathways. The impact of the S-phase environment on the viral life cycle is not restricted to lytic viral replication but is also involved in the episomal genome maintenance during viral latency or reactivation processes, as was recently shown for EBV and the Kaposi’s Sarcoma- associated herpesvirus (KSHV). Strikingly, infections with oncogenic viruses (e.g., SV40, HPV, HTLV-1, EBV) are often associated with S-phase deregulation and genomic instability, preferentially occurring during this critical phase of the cell cycle. In relation with cell cycle delay, DNA damage signaling is often triggered upon viral infections (reviewed in). Particularly, DNA damage response (DDR) pathways are preferential targets of herpesviruses, including HSV-1, EBV, KSHV, human cytomegalovirus and murine gamma-herpesvirus 68. The role of DDR in herpesviruses life cycle is complex. On the one hand, recent evidence suggests that DDR acts as an efficient antiviral response. On the other hand, DDR modulation can be beneficial for herpesviruses by facilitating viral replication, viral genome processing or latency establishment. Moreover, during the course of cellular infections with large DNA tumour viruses, such as human papillomaviruses (HPV) or gammaherpesviruses (e.g. EBV and KSHV), the generation of DNA damage and/or activation of DDR were found to be associated with genomic instability which in turns can participate to virus-induced tumorigenesis (reviewed in). In the present study, we set out to elucidate an important aspect of MDV-host cell interaction by analyzing the impact of MDV and virus-encoded proteins on the regulation of the cell cycle. We demonstrate that MDV lytic infection activates the proliferation of chicken primary skin cells concomitant with a delay in S-phase. By studying the effects of transient vector-driven overexpression in a proliferating chicken cell line, we identified the VP22 tegument protein as a potent cell cycle modulator encoded by MDV. A comparative experimental approach employing VP22 variants with a C- or N-terminal eGFP-tag allowed us to show that an unmodified C-terminus of VP22 is required to elicit the observed S-phase arrest. Moreover, the cell cycle regulating activity of VP22 relies on its ability to be associated with chromatin in the nucleus. In order to define the mechanisms underlying the drastic S-phase arrest observed in VP22 expressing cells, we investigated the impact of VP22 expression on DNA integrity. Strikingly, we found that the DNA of cells expressing this viral protein showed significant DNA damage, as was assessed by comet assay. Together, these data provide new insights into the interaction of MDV with the host cell during lytic replication and pinpoint to a novel powerful function of VP22 that may help to better understand the pre-eminent role of VP22 in MDV replication and more generally in the life cycle of the virus. # Materials and Methods ## Cell Culture and Viruses Chicken Embryo Skin Cells (CESC) were prepared from 12-days-old chicken embryos (LD1 Brown Leghorn chicken strain) and maintained in culture as previously described. This procedure was carried out in strict compliance with the French legislation for animal experiments and ethics stating that the use of embryos from oviparous species before the last third of their normal development (i.e. before day 14 for chicken embryos) is not submitted to regulation (Art. R.214-88). Thus, the preparation of CESC from 12-days-old chicken embryos does not require the permission of governmental or local authorities. Embryos were sacrificed by opening the eggshell, cervical dislocation and immersion in William’s Medium E (Lonza) supplemented with collagenase as described by Dorange et al., 2000. The chicken hepatocellular carcinoma cell line LMH was cultured on gelatin-coated flasks in William’s Medium E (Lonza) supplemented with 2 mM glutamine and 10% fetal bovine serum (FBS) at 37°C in a 5% CO2 atmosphere. As positive controls for DNA damage analyses, LMH cells were treated for 24 h with 1.5 µM etoposide, a topoisomerase IIα inhibitor potent inducer of DNA double strand breaks. Recombinant viruses were generated from the avirulent MDV-BAC20 strain cloned as bacterial artificial chromosome (BAC). The recEGFPVP22 recombinant virus harboring the UL49 gene fused at its 5′ end with the eGFP gene was previously described. Parental BAC20 and recEGFPVP22 viruses were produced after transfection of BAC- DNA into CESC as previously reported. Infections were performed by co-culture of 7×10<sup>6</sup> fresh CESC in a 100-mm diameter plate with infected cells at a ratio of 10<sup>4</sup> PFU/plate. ## Plasmids The pcDNA3-UL49 and pcDNA3-UL48 plasmids encoding the wild-type (wt) VP22 and VP16 tegument proteins of the RB-1B oncogenic MDV-RB-1B strain, have been previously described. Two plasmids harboring the VP22 protein cloned in frame with the enhanced green fluorescent protein (eGFP) were used: (i) the peGFP- UL49, encoding a VP22 tagged with eGFP at its N-terminal extremity and (ii) the pUL49-eGFP in which VP22 is tagged with eGFP at its C-terminus. The latter construct was generated by PCR amplification of the UL49-eGFP fragment from the purified rUL49-eGFP BAC-DNA kindly provided by B. Kaufer (Institut für Virologie, Freie Universität Berlin, Germany). The primer pairs used for amplification were UL49FCLBamHI/eGFPendNotI. The PCR product was inserted into the PCR2.1 TOPO TA cloning vector (Invitrogen) and the BamHI/NotI fragment was then sub-cloned into the peGFP-N1 vector (BD Biosciences, Clontech) where the internal eGFP cassette was previously removed by BglII/NotI enzymatic cleavage. The pGE109 plasmid harbouring the UL49 gene encoded by HSV-1 was kindly provided by G. Elliott. The HSV1-UL49 gene was cloned in frame with eGFP at the Bgl II site in the peGFP-C1 vector (BD Biosciences, Clontech). The VZV-ORF63 encoding the VZV orthologue of UL49 was amplified from pcDNA63wt (kindly provided by C. Sadzot-Delvaux) with the primer pair 5FUL49VZVXhoI/3RUL49VZVBamHI. The PCR product was T/A-cloned into the pGEMT-easy cloning vector (Promega) and subsequently subcloned in fusion with eGFP in the peGFP-C1 vector at the Xho I and Bam HI sites. The genes encoding UL37 (pUL37), UL54 (encoding the ICP27 trans-activator), and the two viral kinases UL13 (pUL13) and US3 (pUS3) were amplified from RB-1B genomic DNA with the primer pairs UL37F/UL37R; UL13F/UL13R; UL54F/UL54R, and US3F/US3R, respectively. Amplification products were inserted into the pGEMT- easy cloning vector (Promega). The UL37, UL13, and US3 genes were sub-cloned under control of the cytomegalovirus immediate early promoter into the pcDNA3.1 vector (Invitrogen) at the NotI site, and the UL54 gene was cloned into the pcDNA3.1 vector at the EcoRV site. All intermediate and final constructs were verified by sequencing (Eurofins, MWG Operon). ## Transient Expression The different eukaryotic expression vectors were transfected into CESC or LMH by using Lipofectamine 2000, according to the manufacturer’s instructions (Invitrogen). Briefly, cells at 80% of confluency plated on 60-mm dishes were rinsed twice with OptiMEM (Fischer Scientific) and were transfected with 5 µg of the plasmid of interest. After 6 h of incubation at 37°C, the transfection mix was removed and serum complemented fresh medium was added. Cells were harvested 24 h or 48 h after transfection for further analysis. Each transfection was performed in triplicate. ## Cell Cycle Analysis At the time points indicated, 1.10<sup>6</sup> vector-transfected or infected cells were trypsinized and washed twice in phosphate-buffered saline (PBS) prior to fixation with 70% ethanol at 4°C for 24 h. Cells were then washed twice in cold PBS and incubated in PBS containing 500 µg/ml Ribonuclease A (Sigma- Aldrich) at 37°C for 1 h. After filtration through a 30-µm pore size membrane, cells were stained with 10 µg/ml propidium iodide (Invitrogen) for 15 min in the dark. Flow cytometry analysis was performed using a MoFlo high-speed cell sorter (Beckman Coulter, Fort Collins, CO, USA) equipped with a solid-state laser operating at 488 nm and 100 mW. Cellular DNA content was analyzed with a 740 nm long-pass filter. Doublets were discarded on the basis of combination of pulse width and area/peak fluorescence. eGFP autofluorescence was detected with a 530/40 nm band-pass filter and the cell cycle distribution was specifically analyzed for eGFP-positive versus eGFP-negative cells. Cell cycle profiles were analyzed with the MultiCycle AV software (Phoenix Flow Systems, California, USA). ## Reverse Transcription-Polymerase Chain Reaction (RT-PCR) and Real Time Quantitative RT-PCR (qRT-PCR) Total RNA was extracted from 10<sup>6</sup> cells with Trizol according to the manufacturer’s instructions (Sigma-Aldrich). RNAs were treated with RNAse-free RQ1 DNAse (Promega, France) and RNA concentration was measured with a NanoDrop spectrophotometer. One µg of each total RNA preparation was reverse transcribed using 100 µg/mL oligo(dT) primers (Promega) and M-MLV reverse transcriptase according to the manufacturer’s recommendations (Promega). The expression of the different cellular genes involved in cell cycle regulation was analyzed by qPCR. Amplification of the cDNA by qPCR (CFX96 Touch Real-Time PCR Detection System; Bio-Rad) was performed in triplicate, using 200 ng of cDNA, 7.5 µl 2×iQ Supermix SYBR green (Bio-Rad), 1 µl ultrapure water (Sigma- Aldrich) and 0.75 µl of each specific primer (10 µM) selected according to the EST data deposited in Genbank (described). The PCR program consisted of a 5 min activation step at 95°C, followed by 39 cycles of 95°C for 10 s and 60°C for 10 s. Expression of the chicken glyceraldehyde phosphate dehydrogenase (GAPDH) was used for the normalization of all target gene mRNAs to enable cross-comparisons among the samples. The relative changes in gene expression were determined by the 2(−ΔΔCT) method. The expression of MDV genes (ICP4, UL13, US3, UL49, UL37, UL54 and UL48) was assessed by RT-PCR performed with 100 ng of the synthesized cDNA prepared from LMH or CESC cells transfected with the corresponding expression vector and 10 µM of specific primers. The GAPDH gene was used as internal control. Specific PCR products were resolved by agarose (2%) gel electrophoresis. ## Immunofluorescence Microscopy At 24 h or 48 h post-transfection, cells grown on glass coverslips were fixed with 4% paraformaldehyde (PFA) for 20 min at room temperature (RT), permeabilized with 0.5% Triton X-100 for 5 min at RT and blocked with PBS, 0.1% Triton X-100, 2% Bovine Serum Albumin (BSA). Immunostainings were performed with monoclonal antibodies directed against phospho-histone H2AX (Ser139) (Millipore; clone JBW301) and tubulin (Sigma-Aldrich; catalog number T9026) at a dilution of 1∶250 and 1∶500, respectively. Goat anti-mouse IgG Alexa-Fluor 594 secondary antibody (Invitrogen) was used at 1∶2000. Cell nuclei were counterstained with Hoechst 33342 dye (Invitrogen). Cells were observed under an Axiovert 200 M inverted epifluorescence microscope equipped with the Apotome imaging system (Zeiss). Images were captured with an Axiocam MRm camera and analyzed by using the Axiovision software (Zeiss). To determine the cellular distribution of eGFP- tagged proteins, a minimum of 100 transfected cells were observed and the results were presented as percentage reflecting the nuclear and/or cytoplasmic distribution of the protein. ## Cell Sorting RecEGFPVP22-infected cells or LMH cells transfected with peGFP vectors were trypsinized 24 h post-transfection and filtered on a 30-µm-pore-size membrane. eGFP positive and negative cells were sorted with a MoFlo (Beckman Coulter, Fort Collins, CO, USA) high-speed cell sorter equipped with a solid-state laser operating at 488 nm and 100 mW. Damaged cells and debris were eliminated on the basis of morphological criteria. eGFP fluorescence was analyzed with a 530/40 nm band-pass filter. The sorting speed was around 15,000 cells/s and cells were collected in appropriate media supplemented with 10% of FBS. ## Alkaline Comet Assay LMH cells transfected with the peGFP, peGFP-UL49 or pUL49-eGFP were harvested 24 h post-transfection and eGFP positive and negative cell were sorted by flow cytometry. After sorting, 2.10<sup>5</sup> cells were used to prepare 3 slides for comet assays, realized as previously described with minor modifications. Electrophoresis was performed at 0.7 volts/cm for 26 min with the Sub-cell GT agarose gel electrophoresis system (Bio-Rad). DNA was then stained with a 20 µg/ml ethidium bromide solution and slides were observed using the Axiovert 200 M inverted epifluorescence microscope (Zeiss). Images were captured with an Axiocam MRm camera (Zeiss) and comets were analyzed with the CometScore software version 1.5 (TriTek). The Tail Extend Moment (TEM) was calculated on the basis of the comet tail length and the relative proportion of DNA contained in the tail. Experiments were carried out 3 times and for each experiment, a minimum of 50 comets was analyzed on each of the 3 slides. Results are presented as the mean (±SD) of the TEM calculated for each condition or as a distribution of the comets with respect to their respective TEM value. ## High Salt Extraction of Histones Salt extraction of histones from chromatin was performed as previously described. Briefly, 1.10<sup>7</sup> cells were resuspended in 1 ml extraction buffer (340 mM Sucrose, 10 mM Hepes pH 7.9, 10 mM KCl, 1.5 mM MgCl<sub>2</sub>, 10% glycerol) containing 0.2% Igepal (Sigma-Aldrich) and 1X protease inhibitors (Complete Mini EDTA free, Roche). After incubation on ice for 10 min, the soluble fraction was separated from the nuclei by centrifugation at 6,500×g for 5 min. Nuclei were resuspended in 1 ml no-salt lysis buffer (3 mM EDTA, 0.2 mM). After incubation at 4°C for 30 min, the chromatin was pelleted by centrifugation at 6,500×g for 5 min, and incubated in 500 µl of high-salt solubilization buffer (50 mM Tris-HCl pH 8.0, 2.5 M NaCl and 0,05% NP40) for 30 min at 4°C. Nuclear debris were pelleted by centrifugation at 16,000×g for 10 min and the supernatant containing the histones fraction was collected. The proteins included in this fraction were separated in a 10% SDS-PAGE gel and revealed with colloidal coomassie blue staining (Sigma-Aldrich). Detection of VP22 was accomplished by immunoblotting using the monoclonal anti-VP22 antibody (L13a) diluted 1∶1000 and an anti-mouse IgG conjugated to horseradish peroxidase (HRP) (Sigma-Aldrich). Specific protein signals were detected with the Pierce ECL2 Western Blotting Substrate (Thermo Scientific) and the Fusion-FX7 imaging system (Vilber Lourmat). Quantification was carried out using the Bio-profil 1D++ software (ChemiSmart 5000). ## Statistical Analysis All graphs and statistics were performed using the GraphPad Prism software version 5.02 (San Diego, USA). Data are presented as means and standard deviations (±SD). Significant differences were determined using Student’s t-test. P values \<0.05 were considered statistically significant. # Results ## MDV Infection Delays Cell Cycle Progression in S-phase In order to analyze the influence of MDV infection on cell cycle progression, chicken embryonic skin cells (CESC) were infected with the parental BAC20 virus. At days 1, 4, and 6 post-infection (pi), mock- and virus-infected cells were fixed in ethanol, DNA was stained with propidium iodide and DNA content was analyzed by flow cytometry. While no significant difference in the cell cycle progression was observed in the early steps of infection (1 dpi), at day 4 pi the cell population in S-phase in BAC20-infected cells was about 3-fold higher than in mock-infected cells. At day 6 pi, the proportion of cells in S-phase as well as in the G2-phase remained 3-fold higher in BAC20-infected cells, suggesting that MDV infection activates cell cycle progression of CESC that normally exhibit a low proliferating rate (3 to 4% of cells in S-phase) and that MDV may delay the cell cycle in S-phase. In order to ascertain viral replication, mRNA expression of the early ICP4 viral gene was followed by qRT- PCR ( lower panel). To confirm the activation of the cell cycle progression assessed by DNA content analysis, and to define the molecular mechanisms of this process upon infection, we examined the expression of key factors involved in cell cycle regulation including cyclins and cyclin dependent kinases (cdk). CESC were mock-infected or infected with BAC20, and qRT-PCR analyzes were performed on total mRNAs extracted at 1, 4, and 6 days pi. At 4 dpi, BAC20-infected cells showed an increase of the mRNA expression of cyclin D (of about 6 fold), cdk6 (2.3-fold), pRb (3-fold), E2F1 (3.5-fold) and c-myc (3.1-fold) compared to mock- infected control cells. A slight up-regulation of cyclin A (1.7-fold), cyclin B (1.5-fold) and cdk1 (1.6-fold) mRNA expression was also detected at 1 dpi in infected cells compared to non-infected cells, while the level of cyclin E mRNA expression was comparable to that in mock-infected cells. These observations are in good agreement with the DNA content analyses showing an activation of the proliferative capacities of infected CESC, since cell cycle progression markers, especially cellular factors involved in the progression into G1 and S-phases (cyclin D, cyclin A, cdk6, pRb, c-Myc, and E2F1) were up-regulated during MDV infection. Interestingly, analysis of the mRNA expression pattern of p53, a protein crucially involved in cell cycle checkpoints and DNA damage pathways, revealed a strong up-regulation (of about 5.9 fold) of its expression at 4 dpi. Of note, we also observed a down-regulation of the mRNA expression of cdk6, pRb, E2F1 and c-myc at 6 dpi that reflect the non-progression of the cell cycle in G1/S phase. To specifically determine the regulation of the cell cycle in infected cells and to preclude problems associated to asynchronous infection and moderate infectivity titers, CESC infected with the recEGFPVP22 virus, an MDV recombinant virus expressing an eGFP-VP22 fusion protein, were sorted by flow cytometry. Infected cells monolayers were harvested at 6 days pi, and the DNA content in non-infected (eGFP-negative) and infected cells (eGFP-positive) was analyzed by flow cytometry. Using this approach, we could observe that 34.4% of the eGFP- positive cells were delayed in S-phase, while in the eGFP-negative population the percentage of cells in S-phase was equivalent to that in mock-infected cells (up to 5%). In addition, a slight increase of cells in G2-phase was detected in infected cells (8% compared to 2.8% in eGFP-negative cells or mock-infected cells). All together, these data clearly demonstrate that lytic MDV infection drives primary avian cells into an active proliferating state. Furthermore, the significant increase of MDV-infected cells accumulating in S-phase is also indicative of a virus-mediated delay in S-phase progression. ## VP22 is a Major Cell Cycle Regulator To identify viral factors involved in the regulation of the cell cycle during MDV infection, we tested the impact of the overexpression of six different viral proteins in CESC (low rate proliferating primary cells) and LMH cells (a cell line with high proliferative rate). Putative candidates were selected either on the basis of their biological activities that might influence host cell-encoded cell cycle regulators and/or on the basis of their essential role in the MDV life cycle. Because of the central role of cellular kinases in cell cycle progression, we were interested to test the two kinases encoded by MDV, pUL13 and pUS3. The ICP27 protein, encoded by the UL54 gene, was also included in the study as a multifunctional viral regulatory protein that has previously been shown to contribute to cell cycle modulation during HSV-1 infection. Three tegument proteins were also tested: the UL48-encoded viral trans-activator VP16, as well as pUL37 and VP22, both of which were shown to be essential for MDV growth (J-F Vautherot, unpublished data;). Eukaryotic expression vectors harboring the viral candidate genes UL37, UL48, UL49, and UL54 (encoding pUL37, VP16, VP22 and ICP27, respectively) were transiently transfected into LMH or CESC cells. At 48 h post-transfection, the cell cycle status was analyzed as outlined earlier and the expression of each of the transfected MDV genes was verified by RT-PCR from total RNA extractions. No significant differences in the proportion of cells in each cell cycle phase was observed (left panel) for transfected CESC, suggesting that none of the overexpressed proteins was able to impact the cell cycle in quiescent CESC. In LMH cells, we also did not observe any cell cycle regulation in response to the expression of UL13, US3, UL37, UL54 and UL48, despite an effective expression of their respective mRNA ( lower panel). However, VP22 (pUL49) overexpression had a substantial effect on the cell cycle in the LMH cell line (right panel), as shown by the strong accumulation of cells in S-phase compared to control cells transfected with the empty vector pcDNA (35% versus 18% of cells in S-phase). Next, we tried to confirm our finding that VP22-expression alone results in an increase of cells in S-phase by transfecting LMH cells with plasmids encoding the VP22 protein fused to a eGFP-tag at its N- or C-terminus. Using an N-terminal eGFP-tagged VP22 protein (peGFP-UL49), we could confirm our finding that VP22 modulates the cell cycle, since more than 90% of LMH cells expressing VP22 (eGFP-positive cells) were blocked in S-phase. However, cells transfected with the plasmid encoding VP22 tagged at its C-terminus did not show any difference in cell cycle regulation compared to empty vector (peGFP)-transfected cells, which indicates that the location of the eGFP-tag at the carboxy-terminal extremity of the VP22 protein abrogates its activity on the cell cycle. Of note, the dramatic intra S-phase arrest observed with the N-terminal eGFP-tagged VP22 protein could be reproduced after overexpression of VP22 in two other avian cell lines: the chicken fibroblast cell line DF1 and the quail myoblast cell line QM7 (data not shown). To verify whether the S-phase promoting activity of the MDV-encoded UL49 is conserved in other alphaherpesvirus orthologues, we tested the ability of VP22 encoded by HSV-1 and VZV to regulate the cell cycle. The HSV-1 and VZV-UL49 genes were cloned in-frame with eGFP and transiently overexpressed in the LMH cell line. At 48 hours post-transfection, the flow cytometry-based cell cycle analysis targeting transfected cells (eGFP-positive population) showed a significant S-phase arrest upon expression of all VP22 orthologues tested. VP22 orthologues derived from MDV and VZV proofed to be equally efficient, as approximately 80% of the cells expressing these VP22 were blocked in S-phase. Although HSV-1-VP22 substantially blocked the cell cycle progression in S-phase (61.8% of the transfected cells), it appeared slightly less efficient than other VP22 orthologues (especially MDV-VP22) in this process. We thus identified a novel function for MDV-VP22 as a potent cell cycle modulator, with a strong S-phase promoting activity. We also revealed that an unmodified C-terminal extremity of VP22 is required for this process. Moreover this biological feature seems to be conserved among the human alphaherpesvirus, even though the two VP22 orthologues tested does not exhibit equal activity. ## Subcellular Localization of the VP22 Protein Encoded by MDV We took advantage of the differential cell cycle modulating activities of the C- or N-terminally eGFP-tagged VP22 fusion proteins to decipher which VP22 properties are crucial to mediate S-phase arrest. One hypothesis for different activity patterns could rest on differential subcellular distributions of the two proteins. To test this hypothesis, the two constructs peGFP-UL49 and pUL49-eGFP were transfected in LMH cells and the respective locations of the proteins were analyzed based on the eGFP signal by fluorescence microscopy at 48 h post transfection. In order to visualize more accurately the distribution of the two proteins, nuclei were stained with Hoechst 33342 and the cytoskeleton was stained with an anti-α-tubulin. Upon overexpression in LMH, the control eGFP protein (peGFP) was distributed all over the cells; the two VP22 proteins tagged at the N- or C-terminus did not show any significant difference in their cellular localization, with respectively 74,4% or 72,6% of eGFP-positive cells presenting an exclusive nuclear distribution and 17,2% or 25% showing a combined nuclear/cytoplasmic staining. Thus, the location of the eGFP tag, at the amino- or carboxy-terminus of VP22, does not seem to affect VP22 cellular distribution in LMH cells. Another interesting feature of VP22 is its ability to bind to chromatin, especially to histones as it has previously been shown for the VP22 encoded by BoHV-1. By performing a high-salt histones extraction protocol from cells transfected with either pcDNA-UL49 or pcDNA3.1 (empty vector), we found VP22 to be included in the histones fraction, as it is demonstrated by a 27 kDa band in the colloidal coomassie blue SDS-PAGE gel and by the VP22-specific antibody (L13a)-probed Western blot (shown in -left panel). This result indicates that MDV-VP22 shares the ability of the VP22 encoded by BoHV-1 to interact with histones. In order to investigate the impact of the position of the eGFP tag on the ability of VP22 to associate with chromatin, we carried out a similar experiment using LMH cells transfected with peGFP, peGFP-UL49 or pUL49-eGFP. We observed that the VP22 tagged at its amino terminus could be co-extracted with histones and visualized as a specific 55 kDa band in coomassie blue-stained SDS- PAGE gel (-right panel). However, the protein tagged at its carboxy terminus appeared to be significantly less retained in the histones fraction. These observations were confirmed by immunoblotting experiments using the anti-VP22 L13a antibody that show the presence of VP22 in the histone extracts prepared from cells expressing peGFP-UL49 and at a far lesser extent (about 4.5 fold) from the pUL49-eGFP transfected cells ( lower panel). All together, these data indicate that VP22 is predominantly targeted to the nucleus of LMH transfected cells independently of the eGFP tag location. However, the fusion of eGFP at the C-terminus of the VP22 protein affects its capacity to associate with chromatin. ## Accumulation of DNA Damages in VP22 Overexpressing Cells Arrest or delay in S phase can arise either from the occurrence of DNA damages, especially double strand breaks (DSB), or replication fork stalling. Since VP22 is able to drastically arrest the cell cycle in S phase and moreover seems to be associated to chromatin, we tested whether the overexpression of VP22 in LMH can induce DNA damages. LMH cells were transfected with pcDNA-UL49 or pcDNA3.1 (as a negative control) and DNA damages were analyzed by alkaline comet assay at 24 h post-transfection. This method, based on a single-cell gel electrophoresis, allows the detection of DNA breaks that are visualized as fragmented DNA exhibiting the shape of a comet’s tail. We could observe an increased number of comets in the population of cells transfected with VP22 compared to the cells transfected with the empty vector pcDNA3.1. To estimate the extent of DNA damages, a more precise analysis with the Comet Score software was performed and the tail extent moment (TEM) was calculated. This parameter is calculated on the basis of the tail length, reflecting the severity of the damages and the amount of DNA in the tail relative to the head, which is an indicator of DNA break frequencies. The calculation of the TEM could show that cells expressing VP22 presented a significant higher TEM (12.62±0.62) than the cells transfected with pcDNA3.1 (4.11±0.25), indicating that the expression of VP22 seems to be associated with the occurrence of DNA damages in LMH cells ( lower panel). However, it should be stressed that this result reflects the DNA damage analysis on the whole population of pcDNA-UL49 transfected and non-transfected cells. Consequently to corroborate these findings and to determine whether the VP22 protein tagged at the C- or N-terminus was also able to induce DNA damage in LMH cells, we transfected the peGFP (empty vector), peGFP-UL49 or pUL49-eGFP plasmids in LMH cells and examined the onset of DNA damages by alkaline comet assay at 24 h post-transfection specifically in the eGFP positive cells sorted by flow cytometry. As positive control, LMH cells were treated with etoposide and as negative control, non-treated and non-transfected LMH cells were analyzed. We could readily observe comets from cells treated with etoposide and most of the cells overexpressing eGFP-UL49 and, to a lesser extent, from cells expressing the UL49-eGFP protein, whereas cells expressing peGFP produced almost no comets or comets with a shorter tail similar to the non-transfected cells. Calculation of TEM revealed that the mean tail moments of cells expressing eGFP- UL49 (27,73±2,11) or UL49-eGFP (11,99±1,52) is significantly higher than for cells transfected with peGFP (4,739±0,54), indicating that the expression of both tagged-VP22 proteins increases DNA damage in cells ( left panel). However, the damages were significantly more pronounced in cells expressing the protein tagged at its amino-terminal extremity than in cells expressing the C-terminally tagged version of VP22. Of note cells treated with etoposide showed a TEM of 77.47±0.55 thereby affirming the drastic induction of DNA damages by this DNA topoisomerase II inhibitor. In addition, we were interested to analyze the frequency distribution of tail moments (i.e. the percentage of cells presenting a defined TEM), which is representative for the number of cells encompassing damages. About 63% of cells transfected with peGFP had a tail moment inferior at 5, indicating that the majority of the cells contain non-damaged DNA or DNA with very limited damages. However, this cellular population decreased when VP22 was expressed both with the eGFP tag at the N-terminus or C-terminus (10% and 29%, respectively), and we could observe a marked increase of the proportion of cells presenting TEM values above 5 (89,6% and 71%, respectively). In particular, the expression of eGFP-UL49 tends to increase the frequency of cells with highly damaged DNA, more than 50% of the cells having a TEM\>20 and 13,7% presenting TEM\>50. In comparison, 19,7% of cells expressing UL49-eGFP showed a TEM\>20 and only 1,9% a TEM\>50. These observations indicate that the expression of VP22 in cells leads to an increased incidence DNA damaged cells and that damages are more severe when the VP22 is fused to eGFP at its N-terminal extremity. It should be however stressed that although the expression of VP22 leads to the occurrence of significant DNA damage, those damages are relatively less heavy than the ones induced by drugs such as etoposide that are responsible of potent damages (more than 55% of the comet having a TEM\>50). In order to specify the nature of the DNA damages generated in cells expressing VP22, we monitored by immunofluorescence staining the expression and localization of γ-H2AX in LMH cells transfected with peGFP-UL49, pUL49-eGFP or with the empty vector peGFP. Because histone H2AX is rapidly phosphorylated (γ-H2AX) after generation of DNA double strand breaks (DSB), γ-H2AX is a preferential marker used to reveal these damages. As positive control, the expression of γ-H2AX was also examined in cells exposed to etoposide. We observed an overall increase of the staining intensity of the γ-H2AX DSB-marker in cells treated with etoposide and specifically in cells expressing eGFP-UL49 compared to non-transfected cells, peGFP transfected cells or UL49-eGFP expressing cells. Moreover, with higher magnification we could visualize that γ-H2AX formed discrete foci in the nucleus of eGFP-UL49 transfected cells as was also observed in etoposide-treated cells. This typical punctuated staining of γ-H2AX reflects its recruitment to sites of DNA damage and thus indicates that cells expressing VP22 tagged at its amino-terminus undergo multiple DSB. # Discussion In the present report, we show for the first time that MDV lytic infection leads to a dysregulation of the cell cycle progression of the host cell. MDV infection not only promotes the proliferation of primary embryonic skin cells, but also leads to an accumulation of infected cells in S-phase. This modulation of the cell cycle is accompanied by a significant up-regulation of cellular genes involved in G0/G1 transition (cyclin D, cdk6) and in G1 to S-phase progression (pRb, E2F1, c-myc, cyclins A). A substantial mRNA up-regulation of the cell cycle regulator p53 was also observed early after infection. Cell cycle modulation is a mechanism that is frequently exploited by viruses in order to facilitate viral replication. In contrast to small DNA viruses, a cellular S-phase environment is not mandatory for herpesviruses encoding their own DNA polymerase and accessory factors required for optimal viral replication. Consequently, for most of alphaherpesviruses it has been demonstrated that they prevent the S-phase entry and rather activate the G1/S checkpoint. MDV is an alphaherpesvirus that shares a number of biological features with gammaherpesviruses, notably the viral lymphotropism and the ability to induce tumors. With respect to cell cycle modulation, our data suggest that MDV has adapted a similar strategy than EBV and KSHV, both of which were shown to promote cell cycle progression, especially into S-phase. It is conceivable that the MDV-mediated cell cycle modulation might also play a role in the multi- factorial events eventually leading to transformation and tumorigenesis. The S-phase is in fact the most vulnerable period of the cell cycle and a defect or inactivation of the key components of the intra S-phase checkpoint may predispose cells to oncogenic transformation (for review). The most unexpected discovery of our study is the identification of the MDV-VP22 protein as a potent trigger of cell cycle arrest in S-phase, as evidenced by the observation that its overexpression in proliferating LMH cells lead to the enrichment of up to 90% of transfected cells in S-phase. VP22 is a major component of the viral tegument of the Alphaherpesvirinae. While VP22 orthologs exhibit functional homology, their significance for alphaherpesviruses life cycle varies according to the virus species. This is well illustrated by previous studies showing that VP22 is dispensable for in vitro replication of PRV, HSV-1, and BoHV1, whereas it is essential for MDV and VZV replication. However, the biological properties of VP22 that determine its key role in the life cycle of MDV remain still unknown. One hypothesis is based on the crucial function of VP22 in cell-to-cell spread. We can also not exclude that a rapid distribution of VP22 after viral entry might prepare an optimal environment for viral replication by inducing an S-phase arrest. Several viral proteins encoded by herpesviruses have been shown to have an impact on the cell cycle. Among the ones encoded by the Alphaherpesvirinae, the ICP0 protein is probably the best studied. This multifunctional protein required for efficient HSV-1 lytic replication and reactivation from latency, has been identified as a major cell cycle modulator that is able to act either on the G1/S or at the G2/M checkpoints. However, the observation that ICP0 deficient mutant viruses are still capable to elicit proliferation arrest indicates that other viral factors also impact the cell cycle. Notably, the immediate early protein ICP27 was shown to be essential for the G1/S cell cycle arrest triggered by HSV-1, with ICP4, ICP0, and the virion host shutoff protein acting as contributors. Hence, MDV and HSV-1 appear to employ differential cell cycle modulation mechanisms as MDV does not encode a functional ICP0 protein and we could not detect any effect of ICP27-overexpression on cell proliferation. It is an interesting speculation that MDV may have evolved a distinct mechanism for cell cycle modulation that crucially involves VP22 in order to compensate for the absence of ICP0 activities. It should be noted that the overexpression of the VP22 proteins encoded by HSV-1 or VZV also resulted in a dramatic arrest of the cell cycle in S-phase in transfected LMH cells. This finding suggests that VP22 might also contribute to the modulation of the cell cycle in the context of infections with human herpesviruses. While screening for viral factors that are involved in the MDV-associated cell cycle regulation, we also tested whether the activity of the two MDV-encoded serine-threonine kinases pUS3 and pUL13 could have a cell cycle regulatory effect. Indeed, it is well known that cell cycle progression is submitted to a tight regulation mediated by kinases and phosphatases. Overexpression of UL13 and/or US3 in low proliferating cells (CESC) or high proliferating cells (LMH) had no effect on the cell cycle, thus excluding a direct involvement of these kinases in the cell cycle modulation. However, pUS3 and pUL13 are able to phosphorylate various cellular and viral proteins, including the VP22 proteins encoded by HSV-1 and -2, as well as BoHV-1,. So far, the phosphorylation status of MDV-VP22 during infection has not been investigated, and we cannot exclude post-translational modifications of MDV-VP22 by UL13 and/or US3, as previously shown for other alphaherpesviruses. Intra-S checkpoints activation mainly reflects DNA breaks or stalled replication fork formation. In order to identify the molecular mechanisms underlying the VP22-driven S-phase arrest, we focused on the impact of VP22 expression on the generation of DNA damage in the host cell genome. Following overexpression of VP22 in proliferating cells, we could indeed show by comet assay that the presence of VP22 coincided with the occurrence of massive DNA damage. Moreover, VP22-expressing cells showed an increased staining of the phosphorylated form of H2AX, suggesting that the DNA lesions observed are double strand breaks. Interestingly, the VP22-mediated generation of DNA damages seems to be tightly associated to the cell cycle modulation property of VP22. This was evidenced by our comparative experimental approach using two versions of the VP22 protein tagged either at its N- or C-terminus. The data from this experiment show that virtually all cells expressing eGFP-VP22 (N-terminal eGFP-tag) are arrested in S-phase and present severe DNA damage, whereas cells expressing VP22-eGFP (C-terminal eGFP-tag) are not affected in their cell cycle progression and show significantly less DNA lesions. Of note, all our attempts to generate a LMH stable cell line overexpressing the MDV-VP22 protein failed due to a high level of cellular mortality. These observations raised the question of the potential toxicity of VP22, which might find an explanation in the induction of double strand breaks in cells overexpressing VP22. The mechanisms by which VP22 induces S-phase arrest and DNA breaks still remain to be elucidated. However, among the characteristics of VP22, we can speculate that its capacity to interact with chromatin and histones might participate to those processes. Interactions of VP22 protein with nucleosomes were previously demonstrated for the BoHV-1-encoded VP22, which physically interacts with nucleosome-associated histones and thereby causes an impaired acetylation of histone H4. In addition, for MDV-VP22, the regions allowing interaction with heterochromatin were previously defined. In the present study, we confirmed that MDV-VP22 is found predominately in the nucleus of cells following overexpression in LMH cells. We also found that an N-terminally eGFP-tagged MDV-VP22 can be extracted from chromatin preparations together with histones. However, for a C-terminally eGFP-tagged MDV-VP22, the efficiency of recovery from histones extracts was far less, suggesting that an unmodified C-terminal extremity of VP22 is necessary for the association of VP22 with chromatin. Together, these observations suggest that the abilities of VP22 to arrest the cell cycle in S-phase and to effect DNA damage are linked to its direct or indirect interaction with histones and/or chromatin. According to this model, it is conceivable that the interaction of VP22 with chromatin or histones may disturb the unwinding of DNA in a similar fashion than cellular helicases or topoisomerases by preventing access to the DNA replication machinery. Alternatively, it can be speculated that an association of VP22 with DNA/histones may cause physical tension of the DNA double helix eventually leading to DNA breaks. Activation of DNA damage response (DDR) pathways is well documented for a number of viruses, especially tumorigenic viruses and plays a central role in viral replication. While our data only provide clear evidence for a role of VP22 as a powerful inducer of DNA damage in a non-infectious context, it can be assumed that DDR activation might play a role in MDV replication and/or MD pathogenesis. Due to the critical role of VP22 for MDV-replication, it is so far impossible to evidence the function of VP22 as a major cell cycle regulatory factor during MDV infection by using a VP22-deleted virus. However, Jarosinsky et al. have demonstrated that a recombinant MDV harboring a VP22 protein tagged with eGFP at the C-terminus (a construct that is identical to the VP22-eGFP used in the present study) showed a drastic decrease in its ability to induce MD in infected chickens, with only 10% of the chicken developing tumors. In addition, we have recently observed that a recombinant virus with the VP22 protein tagged at the N-terminus is also attenuated, but in a lower extent, with 33 to 66% of the infected chickens developing MD lymphoma. Although the impairment of pathogenicity of the recombinant MDV studied by Jarosinski et al. could in parts be explained by a lower viral replication efficiency in vivo, in view of our data, it can also be speculated that the defect in tumor development observed by the authors might be due to the loss of the ability of the C-terminally tagged VP22 protein to induce S-phase arrest and DNA damage. In conclusion, our findings provide new insights into herpesvirus-cell host interactions by demonstrating that the oncogenic alphaherpesvirus MDV affects the cell cycle progression in infected cells. Moreover, we could assign a novel role to the VP22 tegument protein as a potent cell cycle modulator, property that seems to be associated to its ability to induce DNA damages in cells. Current efforts are under way to elucidate the detailed mechanisms of VP22-induced DNA damage response, and its role during viral infection, especially with respect to a possible involvement of DDR in MDV replication and/or the establishment of MDV-latency and subsequent lymphoma formation. We thank S. Trapp (INRA, Nouzilly, France) and J. Vignard for their constructive comments and corrections on the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LTF. Performed the experiments: LTF DB DCV YLV SR EBR. Analyzed the data: LTF JFV CD GM. Contributed reagents/materials/analysis tools: JFV CD EBR GM. Wrote the paper: LTF.
# Introduction Cynodon dactylon (Linnaeus) Persoon (Family: Poaceae, bermudagrass in English) is a perennial, creeping grass. Although it is widely found in the tropical and warm temperate regions, bermudagrass is predominantly distributed between 45° North and 45° South latitudes. Currently, C. dactylon is globally used as a turf grass, fodder and medicinal plant, and it was also used for removal of heavy metals from contaminated soils –. This grass is adapted to extremely variable environments, such as fertile fields, arid land, saline land, wet irrigation canals and even contaminated wastelands with high levels of Pb, Cd, Zn and Cu. In addition, the extracts from bermudagrass are well known for various medicinal properties including antimicrobial, anti-inflammatory, immunomodulatory, and anti-diarrhea activities. It has been therefore used to treat traumatic wounding, kidney calculi, hypoglycemia, depression and cancer. Seed germination percentage is a major criterion used for evaluating suitability of an environment for grass cultivation. Previous studies showed that extreme temperatures could lead to seed dormancy and significantly decreased germination percentage. Constructing a precise mathematical model that correlates the germination percentage with temperature may avoid failure of plantation due to inappropriate sowing timing or mismatch between the grass species and climate zone. Hence it would be very helpful for decision-makers to select grass species and sowing time for lawn, erosion control and forage cultivation. Using thermal time approach, Bradford constructed temperature/water potential based seed germination and dormancy models. This study showed that germination experiments under temperature regime with discrete (stepwise) changes could also be used to accurately predict thermal responses of seeds in field environment with continuously changing temperature. Hardegree and Van Vactor, used the Piece-wise linear (PWL) regression equations to confirms that constant-temperature experimental results derived mathematical models could be used to predict the germination/growth responses to the combined effects of multiple environmental factors in the field. Although various experimental procedures have been established to generate data for the development and validation of reliable prediction models, the R<sup>2</sup> values of temperature-germination percentage functions in some of the previous studies were relatively low (0.61–0.80), and more accurate predictions are desired. Recently, the Geographic Information System (GIS) based spatial prediction approach was introduced for the creation of quantitative and accurate grass suitability maps. The grid data of mean minimum and maximum temperatures were used for calculating suitability values of regions via the temperature- germination percentage functions. Therefore, the suitability of grass for different regions could be appropriately visualized on the map. However, the visualization of time scale (season) of these maps remains to be streamlined for readability to facilitate their practical application. The objectives of this research were to: (i) explain the correlation error between temperature and seed germination percentage caused by selection of functions (ii) provide a new program for optimizing the temperature-germination function, (iii) use temperature-based seed germination percentage function in combination with national temperature grids in China for predicting the suitability of three bermudagrass cultivars, C. dactylon, ‘Savannah’ and ‘Princess VII’. # Results ## Germination response to Diurnal Fluctuations of Temperature The germination response of the three bermudagrass cultivars was similar. All the three cultivars were capable of germinating in warm-period temperature (T<sub>2</sub>) from 25 to 35°C and cool-period temperature (T<sub>1</sub>) ranging from 5–40°C. The optimal temperature for seed germination is defined as that which is not lower than the maximum germination minus one-half of its confidence interval (P = 0.05). For example, the maximum germination percentage (the optimal temperature for seed germination) of C. dactylon was 87.3% at 15/35°C; and the germination percentage at 5/30°C was only 80.7%, but it was still accepted as the optimal temperature for seed germination since the one- half of its confidence interval was 7.0% and 80.7%\>(87.3%–7.0%). Fluctuating between 5–25°C cool-period temperature and 30–40°C warm-period temperature gave rise to the optimal temperature for seed germination. On the other hand, germination percentage was usually lower than 50% at constant temperature ranging from 20 to 40°C (except the “Savannah” at constant 25/25°C), for example the germination percentages of “Princess VII”, were 0, 36.0%, 5.3%, 12.7% and 20.0% at constant temperature regimes of 20/20, 25/25, 30/30, 35/35 and 40/40°C, respectively. Interestingly, germination percentage of C. dactylon seed at a constant 30°C only reached 1.3%. The maximum germination and the mean of germination of the three bermudagrass cultivars are presented in. The germination percentage of ‘Princess III’ appeared to be the highest (96.7%), but it was not significantly different (P\>0.05) from that of the other two cultivars (87.3% and 94.0% for C. dactylon and ‘Savannah’, respectively). Depending on the cultivar, 25%–27.8% of the temperature regimes supported the optimal temperature for seed germination. Only eight temperature regimes, 30/5, 30/10, 35/5, 35/10, 35/15, 35/20, 40/15 and 20/40°C, supported the optimal temperature for seed germination for all three tested cultivars. ## Performance of Different Regression Models All the regression equations of temperature treatments that assess the germination suitability of the three bermudagrass cultivars are summarized in. Quadratic and quintic general equations were used for data simulation. The bisquare and BP-ANN approaches were utilized for optimizing these equation coefficients. There was no significant difference between the quintic functions and bisquare optimized quintic functions. Bisquare quintic functions of cultivars C. dactylon, ‘Savannah’ and ‘Princess VII’ showed lower R<sup>2</sup> than quintic functions (data not shown). Similar optimization results were observed for all the cultivars tested. Generally, quintic equations performed better than quadratic equations. The highest R<sup>2</sup> values were generated by the back propagation artificial neural networks aided dual quintic equation (BP-ANN-QE) model (0.9439 to 0.9813). In contrast, the dual quadratic equation model (DQEM) generated the lowest R<sup>2</sup> values (0.6940 to 0.8177). Contour plot maps were then used to visualize the prediction errors generated from different simulation functions. The blue spot represents the experimental germination percentage associated with particular fluctuating temperature combinations. Obviously, DQEM is the one most prone to produce prediction errors among the compared models. However, some weakness in functional convergence of quintic equations was observed, a small portion of observed temperature- germination spots were still out of their equation derived surfaces. ## Spatial Mapping of Optimal Planting Times The temperature–germination functions derived from our experimental data were used to predict germination percentage for various regions in China in different seasons using the 25 years mean minimum and mean maximum earth surface temperature grids as input. The percentage reflected the likelihood of germination suitability at each grid cell of the maps via FreeMicaps. The results reveal that most of the Chinese regions are not suitable for seed germination of all the tested bermudagrass cultivars from November to March. Although ‘Savannah’ has the narrowest geographic range of germination with percentages arranging from above 0 to 100%, it was predicted to have optimal temperatures for germination in widest geographic range in China. In contrast, C. dactylon had the narrowest range for the optimal temperatures for seed germination. For both C. dactylon and ‘Princess VII’, the widest range with the optimal temperatures for seed germination was in the month of June, whilst for ‘Savannah’ it was in May. # Discussion Simulation models have been widely used to correlate cultivation conditions and plant germination/growth. These regression models were generally used to predict the plant suitability to particular regions where climate and soil environment information is available. These mathematical models could help grass cultivation in three major aspects. First, these functions combined with a visual suitability map may help decision makers in selecting a grass species and planning its seeding timing. Second, they may serve as useful tools for identifying and evaluating desirable quantitative characteristics for specific grass breeding objectives, which is helpful in coupling genotype and phenotype of a target cultivar, a technology with significant application value in the rapidly expanding turf industry. Third, these grass-condition models may improve our understanding of how changes in agricultural systems would quantitatively influence grass germination/growth. ## Comparison of Regressions Using Quadratic and Quintic Equations The quadratic equation was broadly accepted as an effective statistical tool for simulating continuous variable-response relationship, and quadratic response surface has been a predominant method to analyze germination performance of grass seeds under various temperature regimes, especially to test the seed germination affected by diurnal temperature treatments. However, the weakness of quadratic equation is also obvious. Since the two-dimensional quadratic response surface cannot display some errors in the global fitting between quadratic function and experimental data, the weakness of low R<sup>2</sup> values (0.61–0.80) was ignored. In this study, quintic equation was employed for regression to simulate diurnal temperature-germination responses of three bermudagrass cultivars. Compared to quadratic surfaces, quintic approaches had significantly lower fitting errors and higher confidence. This may be due to the nonlinear temperature-germination correlation. Hence these quintic equation models could provide more reliable predictions for field performance of grasses, further improving the reliability of the suitability map of grasses tinted with different color based on predicted germination percentage data. ## BP-ANN and Bisquare for Intercept and Coefficients Fitting Artificial Neural Network (ANN), an algorithm for simulating the thinking processes of the human brain which usually have multiple networks that are logically arranged as fundamental units. The information of any unit can be learnt, recalled, concluded and speculated. Hence, the ANN has many advantages such as distributed storage of information, self-adaptability, self-organization and fault-tolerance properties. As a computer based program, it could be used to perform large-scale parallel calculations to simulate nonlinear correlation. Therefore, ANN is broadly used in various of fields including biomedical research, optimization of soil nutrient distribution coefficient, forecasting of microorganism community assemblages, prediction of animal metabolism and diets correlation, and simulation of fruit post-ripening process. Currently, there are tens of ANN models. Amongst them, the back propagation (BP) network is the most widely used one for simulation of nonlinear relationship. The BP-ANN model belongs to supervised study and its training process has two phases, forward propagation and backward propagation. In the forward propagation, the weighted value and threshold value of each layer is calculated by iteration and passed into the BP three-layer network. Subsequently, the backward propagation uses the weighted value and threshold values for revision. These two phases commonly occur repeatedly for about 10, 000 times, and the weighted value and threshold values alternate until they converge. The target of these training processes is to generate a function that can globally ‘distinguish’ and ‘remember’ all the input raw data. Therefore, the simulated BP-ANN functions could be used to predict proposed parameter with appropriate input variables. The robust bisquare estimator, also known as Tukey's bi-weight function, was a popular choice for nonlinear function fitting. This estimator was often used for smoothing the nonlinear response surface. It was reported that bisquare estimator could be adapted to process noisy data with outliers. In our study, both BP-ANN and bisquare methods were used to optimize the temperature-germination functions for three bermudagrass cultivars. We did not find significant difference between general quintic functions and bisquare fitted quintic functions. Furthermore, most bisquare quintic functions in this study had slightly lower R<sup>2</sup> values than the quintic functions. Therefore, the BP-ANN optimized quintic equations were found to be the best option for fitting temperature-germination functions of the tested bermudagrass cultivars. To some degree, the BP-ANN based temperature-germination functions could be evaluated by the visualized suitability map and the observed temperature- germination results. Generally, the cultivar “Savannah” could germinate under the widest temperature regimes, from about 20/5°C to 40/40°C, it also present the highest suitability with widest geographic regions in China (eg.). On the other hand, the predicted suitability map worked in concert with the observations that all the three cultivars preferred warm temperature above 15/15°C, and little germination could be found during the cool seasons in China (eg.). # Conclusions This study tested the influence of diurnal fluctuations of temperatures on seed germination of three bermudagrass cultivars (C. dactylon, ‘Savannah’ and ‘Princess VII’). Eight temperature regimes, 5/30, 10/30, 5/35, 10/35, 15/35, 20/35, 15/40 and 20/40°C, supported the optimal temperature for seed germination for the three cultivars. However, the germination percentage of all the three cultivars was lower than 50% under conditions of constant temperatures ranging from 5 to 40°C. To simulate the grass germination-temperature response function, both quadratic and quintic equations were employed. Quintic functions performed significantly better (R<sup>2</sup> were around 0.9439) than the quadratic ones (R<sup>2</sup> ranging from 0.6940∼0.8177) for the tested cultivars. The main objective of this study was to test the nonlinear fittin\g approaches, bisquare and BP-ANN, for optimizing the regression functions. Our results suggested that BP-ANN has significant advantages over bisquare for fitting the intercept and coefficients of the temperature-germination functions. Based on the experimentally derived BP-ANN functions and available climate data, we prepared a seed suitability map of three bermudagrass cultivars for cultivation in the People’s Republic of China. We observed that most of the regions in China are not suitable for bermudagrass seed germination from November to March. The cultivar ‘Savannah’ had the widest geographic range of the optimal temperature for seed germination, whilst C. dactylon had the narrowest range of the optimal temperature for seed germination percentage. The month with widest range of the optimal temperature for seed germination for C. dactylon and ‘Princess VII’ is June, whilst for ‘Savannah’, it is May. # Materials and Methods ## Seeds and Conditioning The bermudagrass cultivars (C. dactylon, ‘Savannah’ and ‘Princess VII’) were planted widely in China. All the grass seeds were purchased from Shanghai Chunyin Turf Inc., (Shanghai, China) and stored at room temperature until use. Seed viability was tested on sterilized wet filter paper at 25°C in darkness. Briefly, seeds were surface sterilized in 0.01% HgCl<sub>2</sub> for 1 min followed by four rinses with distilled water. The seeds were subsequently placed on wet filter papers in Petri dishes. The Petri dishes were placed in incubators set up for 36 different regimes of diurnal fluctuations of temperature treatments: 16 h at temperature T<sub>1</sub> and 8 h at temperature T<sub>2</sub>. The T<sub>1</sub> and T<sub>2</sub> ranged from 5 to 40°C with 5°C increments. Germinated seeds were counted daily until no further germination occurred (about 15–20 days) for viability analysis (Supplementary file –). For each experiment, three replications of 50 seeds were used in a randomized block design. ## Statistical Analysis In this study, quadratic and quintic response surfaces were constructed with estimated means and confidence intervals. The quadratic equations were first used for estimating germination percentages. The generalized equation was : ![](info:doi/10.1371/journal.pone.0082413.e001) where Y<sub>1</sub>: predicted germination percentage, A<sub>0</sub>: intercept, A<sub>1</sub> through A<sub>5</sub>: coefficients, T<sub>1</sub> and T<sub>2</sub>: two temperature in the diurnal regime. Meanwhile, quintic equations with the following generalized equation were also used to simulate the temperature-germination function:, where A<sub>0</sub><sup>′</sup>: intercept, f(A): coefficient function. Temperature inputs were normalized by dividing the maximum value. Subsequently, the intercept and coefficients were optimized using Bisquare, a default curve fitting method of MATLAB 7.9-R2009b software and back propagation artificial neural networks (BP-ANN) approach described in, respectively. The BP-ANN (or feed-forward network) has the capability to learn arbitrary nonlinearity and great potential for adaptive control applications. In the BP-ANN, the correlations among the input variables do not need to be specified. Instead, they learn from the examples fed to them. In addition, they can generalize correct responses that only broadly resemble the data in the learning phase. ## Spatial Mapping The suitability of a grass is represented by seed germination percentage. The grass suitability maps were created using the FreeMicaps software (<http://bbs.121323.com/guojf/FreeMicaps20111001.rar>). Like the Surfer software, FreeMicaps also uses grid data at selected points (station) that is compatible with GIS. The temperature data of adjacent regions around the station are generated using regression functions. The data used to construct the temperature grid are curated in the database in National Aeronautics and Space Administration (NASA, <http://power.larc.nasa.gov/cgi- bin/cgiwrap/solar/[email protected]#s11>) which are composed of minimum and maximum daily temperature of earth surface from 313 Chinese weather stations for a period of 25 years (from 1983 to 2007). These mean values of minimum and maximum daily temperatures were used as the T<sub>1</sub> and T<sub>2</sub> variables, respectively in the BP-ANN-Quintic functions for calculating germination percentages (grass suitability). # Supporting Information We are grateful to Xi Tang and Meixia Ruan for their kind assistance during seed germination tests. Thanks are also due to Dr Jiang-Feng Guo (Agricultural Technology Extension Station of Qingfeng County, Henan Province, China) for providing the FreeMicaps software and for his kind guidance on software application. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HL EP. Performed the experiments: EP. Analyzed the data: EP. Contributed reagents/materials/analysis tools: HL EP. Wrote the paper: EP HL LD SN NM.
# Introduction Although it is well known that improvement of health behaviour is necessary for the management of non-communicable diseases (NCDs) such as hypertension, diabetes, and dyslipidaemia, it is challenging to modify health behaviour. However, there is no lack of opportunities for health behaviour improvement, and the theory of health behaviour recommends taking advantage of such cues to action. In particular, the situation of being diagnosed with an NCD and treated is recognised as a good opportunity for becoming aware of the health crisis and thus preventing the exacerbation of the disease and the development of comorbid diseases (i.e., what is known as the ‘wake-up call’ or a teachable moment). Despite various efforts to improve NCD patients’ health behaviour, there is currently insufficient evidence to suggest that their health behaviours are significantly better than those of the general population. In a study that followed the trends of three health behaviours (current smoking, alcohol consumption, and walking) in major NCD patients (hypertension, diabetes, and dyslipidaemia) and compared them with those of the general population by analysing the relevant Korea Community Health Survey data covering the period of 2008–2017, major NCD patients showed only marginally more health behaviour improvement compared with the general population, partially showing a lower level of improvement than the general population. Regarding nutrition-related behaviours, some studies reported that patients improved their dietary behaviour through education or counselling after diagnosis of NCDs. In other studies, patients have complained of difficulties in correcting and continuing to fulfil their dietary behaviours that were formed over a long period of time. The present study focused on major NCD patients’ nutritional behaviours as a follow-up analysis of prior research, in which smoking, drinking, and physical activity (i.e., three typical health behaviours) were analysed; however, no analysis of nutrition-related health behaviours was undertaken. The World Health Organization (WHO) emphasises the importance of reducing unhealthy food and drink consumption, along with reducing tobacco use, harmful use of alcohol, and physical inactivity; these are the ‘best buys’ of the strategies to reduce the burden of NCDs. Therefore, this study compared low-salt preference, which is a typical nutrition-related health behaviour, between major NCD patients and the general population, using the analysis methodology adopted in prior research. # Materials and methods ## Data source In this study, we used data extracted from the Korea Community Health Survey (KCHS) during the period of 2008–2019. The KCHS is a nationwide survey that was first conducted in 2008 by the Korea Centers for Disease and Control and Prevention to evaluate the health status of community residents and to establish evidence-based health-related statistics. This survey was conducted each year from August to October at 255 community health centres across the country to evaluate anthropometrical parameters, smoking status, alcohol use, physical activity, health behaviours (e.g., diet and use of healthcare services), and quality of life in Korean adults (≥19 years of age). The sample size of KCHS is about 900 people in each community health center, for a total of about 220,000 people. ## Participants Thirteen major types of NCD patients (hypertension, diabetes, dyslipidaemia, stroke, myocardial infarction, angina, arthritis, osteoporosis, asthma, allergic rhinitis, atopic dermatitis, cataract, and depression) were defined as participants who responded clinically diagnosed from physicians in the KCHS. Depression was subdivided into patients who had been diagnosed by a physician and those who had experienced symptoms. The general population was defined as all the participants in the Korea Community Health Survey. ## Low-salt preference Low-salt preference was measured using four items: the percentage of respondents who answered that they 1) usually eat less salty food, 2) do not add salt or soy sauce at the table, 3) do not dip fried food in soy sauce, and 4) stick to a low-salt diet in the three previous items, that is, they have a low-salt preference (Type III). ## Data analysis Descriptive statistical analysis of data was conducted to determine the low-salt preferences of NCD patients and the general population according to the year (2008–2019) and region (municipalities and provinces). A Chi-square test was performed to compare the low-salt preferences of patients with major NCDs and the general population. The 2018 Korea Community Health Survey data were excluded from the analysis because there were no survey items on low-salt preferences. Since the 2018/2019 data for the prevalence of dyslipidaemia were missing, we used the 2017 data. Joinpoint regression was applied to detect significant changes in annual low-salt preference, with a maximum of one joinpoint allowed. The annual percentage change (APC) (i.e., the average annual percentage change \[AAPC\] in the case of one joinpoint) was computed using joinpoint analysis, and each p-value is presented. Analysis data were tabulated using Microsoft Excel 2014 (Microsoft Corporation, Seattle, WA). Stata/SE13.1 (StataCorp, Texas, TX) were used for descriptive statistics and Joinpoint Regression Program (version 4.6.0; US National Cancer Institute, Bethesda, MD, USA) were used for joinpoint regression. ## Ethical considerations Since this study used a publically available secondary data source (Korea Community Health Survey, Available from: <https://chs.kdca.go.kr/chs/main.do>), we did not seek approval from the institutional review board. We also did not have to ask for the consent of the participants. # Results As of 2017, low-salt preference among major NCD patients was generally higher than that of the general population, however, overall statistical significance was not reached. The percentage of respondents who answered that they usually eat less salty food was 23.6% in the general population and 25.2%, 27.7%, and 24.4% in patients with hypertension, diabetes, and dyslipidaemia, respectively. The percentage of respondents who answered that they did not add salt or soy sauce at the table was slightly higher in patients with hypertension and dyslipidaemia (72.2% and 74.0%, respectively) than in the general population (71.9%), though the percentage was slightly lower (71.5%) in patients with diabetes. The percentage of respondents who answered that they did not dip fried food in soy sauce was slightly higher in major NCD patients than in the general population (36.0%). The percentage of respondents who indicated that they preferred a low-salt diet for all three items was highest in patients with diabetes (15.6%), followed by patients with hypertension (14.1%), dyslipidaemia (13.4%), and the general population (11.3%). ## Overall low-salt preference The overall low-salt preference rate (usually eat less salty food) showed an upward trend until 2013 and declined thereafter. In the general population, it peaked in 2013 at 24.9% (APC: 2.654, p = 0.063) and decreased to 21.6% by 2019 (APC: -2.347, p = 0.195). In patients with hypertension, it peaked in 2013 at 27.6% (APC: 1.857, p = 0.136) and decreased to 21.6% by 2019 (APC: -2.649, p = 0.126). In patients with diabetes, it peaked in 2013 at 29.8% (APC: 1.859, p = 0.122) and decreased to 23.6% by 2019 (APC: -2.849, p = 0.093). Lastly, in patients with dyslipidaemia, it peaked in 2013 at 27.2% (APC: 5.690, p = 0.048) and decreased to 24.4% by 2017 (APC: -2.528, p = 0.023). ## Low-salt preference at the table The rate of low-salt preference at the table (not adding salt and soy sauce at the table) gradually increased and tended to remain elevated; in patients with some diseases, the rate slightly decreased by 2019. In the general population, it peaked in 2013 at 74.4% (APC: 11.788, p-value: \<0.001) and decreased to 68.4% by 2019 (APC: -0.964, p = 0.010). In patients with hypertension, it peaked in 2013 at 73.5% (APC: 11.401, p = 0.001) and decreased to 68.1% by 2019 (APC: -0.523, p = 0.225). In patients with diabetes, it peaked in 2013 at 72.7% (APC: 10.914, p \<0.001) and then gradually decreased to 68.3% by 2019 (APC: -0.567, p = 0.113). In patients with dyslipidaemia, there was a slight decrease from 75.9% in 2011 to 74.0% in 2017 (APC: -0.390, p-value: 0.109). ## Low-salt preference for fried food The rate of not dipping fried food in soy sauce increased until 2013 and then decreased. No significant changes were observed in the general population between 2008 (36.8%) and 2019 (36.4%), and there were fluctuations in some years (APC: 0.214, p = 0.693). In patients with hypertension, the rate peaked in 2013 at 50.3% (APC: 2.993, p = 0.023) and then decreased to 44.9% by 2019 (APC: -1.585, p = 0.092). In patients with diabetes, it peaked in 2013 at 50.3% (APC: 2.494, p = 0.035) and then decreased to 45.2% by 2019 (APC: -1.634, p = 0.070). Lastly, in patients with dyslipidaemia, it peaked in 2013 (47.9%) and then decreased to 44.2% by 2017 (APC: -0.152, p = 0.847). ## Low-salt preference rate (type III) Low-salt preference (type III) increased until 2013 and decreased thereafter. In the general population, it peaked in 2013 at 13.2% (APC: 7.528, p = 0.009) and then decreased to 10.4% by 2019 (APC: -3.422, p = 0.71). In patients with hypertension, it peaked in 2013 at 16.3% (APC: 4.621, p = 0.011) and decreased to 12.8% by 2019 (APC: -4.638, p = 0.037). In patients with diabetes, the low- salt preference rate (type III) peaked in 2013 at 18.1% (APC: 4.761, p = 0.024) and decreased to 13.7% by 2019 (APC: -3.634, p = 0.027). Although no consistent rate was observed in patients with dyslipidaemia, a slight decrease was observed from 14.2% in 2011 to 13.4% in 2017 (APC: -0.759, p = 0.603). ## Regional variations Greater regional variations in low-salt preference were observed in major NCD patients compared to the general population. In 2019, Seoul and Jeonnam had the highest and lowest percentages of respondents who answered that they usually ate bland or less salty food (24.3% and 18.3%, respectively; standard deviation \[SD\]: 1.7%). For patients with diabetes, Jeju and Gyeongnam were the regions with the highest and lowest percentages (30.1% and 20.8%, respectively; SD: 2.4%). Gwangju and Daejeon were the regions with the highest and lowest percentages of the general population who answered that they did not add salt or soy sauce at the table in the same year (77.0% and 54.1%, respectively; SD: 5.7%). Similar regional variations were observed in patients with hypertension: Gwangju, 81.1%; Daejeon, 58.2%; SD, 5.8%. Jeonnam and Daegu were the regions with the highest and lowest percentages of the general population who answered that they did not dip fried food in soy sauce, at 55.8% and 23.6%, respectively (SD: 8.0%). For patients with diabetes, Geonnam and Sejong were the regions with the highest and lowest percentages (65.0% and 20.3%, respectively; SD: 11.4%). With regard to the low-salt preference rate (type III), Seoul and Daegu were the regions with the highest and lowest percentages, at 12.3% and 7.6%, respectively (SD: 1.4%). In patients with diabetes, the regions with the highest and lowest percentages were Jeonnam (17.0%) and Gangwon (8.6%), respectively (SD: 2.5%). # Discussion In this study, we investigated the low-salt preferences of major NCD patients (e.g., hypertension, diabetes, and dyslipidaemia) from various angles and compared them with those of the general population based on the relevant datasets extracted from the Korea Community Health Survey. Data analyses revealed that major NCD patients had a higher low-salt preference compared with the general population; however, the difference was not statistically significant. Despite an overall upward trend in low-salt preferences, the low- salt preference rates of patients with different major NCDs and the general population was gradually decreasing after peaking mainly in 2013; thus, the low- salt preference rates leave much room for improvement. In addition, we found greater regional variations in low-salt preferences among major NCD patients compared to the general population. This finding highlights the need to prepare a region-specific strategic approach to improve low-salt preferences. Excess sodium intake is a risk factor for developing NCDs such as hypertension and cardio-cerebrovascular disease, as well as alcohol use, smoking, and lack of physical activity. Reducing sodium intake can contribute to reducing the prevalence and mortality caused by these diseases. Low-salt preference is known as an index reflecting individual’s attitudes and reactions to salty taste. Therefore, low-salt preference influences an individual’s sodium intake behaviour, such as the amount of salt that can be adjusted during the cooking process or at the table, the frequency of eating out and consumption of instant foods, and the amount of soup in the daily diet. The significance of this study lies in the fact that it demonstrated the need to develop more effective strategies for the formation of low-salt dietary habits in major NCD patients by identifying the trend of low-salt preferences of major NCD patients using big data from the Republic of Korea. In this study, the low-salt preference of patients with major NCD was found to be higher than that of the general population. From this finding, it can be inferred that NCD patients were interested in a low-salt diet after being diagnosed with NCDs, or that they received nutrition education in healthcare facilities. However, no significant differences in low-salt preference were observed between NCD patients and the general population, and the low salt preference rates gradually decreased after peaking in 2013. In Republic of Korea, the multicomponent national programs to reduce salt intake has been steadily promoted since the 2010s. In particular, in 2012, a national plan was established to reduce sodium intake. However, we estimated that this effect appears to have gradually weakened. This highlights the need to establish and implement more active interventions for a low-salt diet in NCD patients as well as the general population. More specifically, among the items used to determine a low-salt preference, NCD patients showed higher rates for the item regarding the overall low-salt diet preference (usually eating bland or less salty food) when compared to the general population; however, the absolute value was lower than 30%. Sensitivity and familiarity to salty taste is affected by the concentration of sodium in foods that have been frequently exposed recently, to form the habit of eating less salty food, efforts are needed to reduce the sodium content of the foods provided at home, school, workplace, and at other places regularly visited. In addition, efforts should be made to reduce the sodium content in convenience foods and delivery foods whose consumption continues to increase due to the coronavirus disease 2019 pandemic. Government-led policies to reduce sodium intake through food service companies may be of great help in getting individuals to comply with a low-salt diet. However, given the lack of publicity for this policy and limited voluntary participation in the turnover-driven foodservice industry, stronger measures (e.g., a sodium tax) are required. Another noteworthy finding of this study is that the rate of not adding salt and soy sauce at the table was higher than 70%. Since Korean food is mainly seasoned with salt or soy sauce during cooking, sodium intake can be significantly increased if it is added to the table. The high rate of a low-salt preference at the table is a desirable trend because, according to the health belief model, NCD patients recognise the severity of their diseases and make efforts to adhere to a low-salt diet, which is reflected in their avoidance of using salt and soy sauce. However, due to the low percentage of a low-salt preference (type III), it may be inferred that they do not add salt or soy sauce at the table because they usually eat sufficiently salted food. On the other hand, the rate of not dipping fried foods in soy sauce was lower than 50%. Fried foods are served with soy sauce because they are not strongly salted during cooking, and people tend to dip them in soy sauce. Therefore, to promote a low-salt diet, it is necessary to create a supportive environment such as educating cooks and foodservice professionals on recipes that use less salt or providing publicity about the importance of low-sodium diets for health. It is also worth noting that greater regional variations were observed in low- salt preference among NCD patients compared to the general population. This gap may be attributable to the differences in disease prevalence, income level, and medical resources from one region to another, but may also be associated with sodium reduction policy programs run by local governments. Furthermore, it is necessary to pay attention to various characteristics related to salt intake, such as average temperature and diet, in order to explain these variations. Just as there is variation in salt intake between countries, within a country, there may be differences in salt intake or salt preference from region to region. However, it is difficult to find studies examining which factors are most strongly associated with these variations. In the future, it would be meaningful to conduct studies to identify the causes of the regional differences in low- salt preference among NCD patients and to narrow the regional gap. The findings from these studies would ultimately contribute to resolving the regional disparities in healthy life expectancy. This study has three limitations. First, only the variables included in the raw data were considered for the analysis because of the characteristics of the data source. In particular, the incidence of NCDs was identified by self-report, and the timing of diagnosis was unknown. In future research, it will be necessary to follow changes in dietary behaviours before and after diagnosis of the disease using other data sources such as panel data or patient cohort data. Second, although various policies regarding sodium intake reduction have been put in place in the Republic of Korea, their effects could not be examined in this study. To more effectively promote a low-sodium diet, it is necessary to conduct long-term follow-up studies that consider changes induced by such policies. Third, this study included 13 patients with NCDs, including hypertension and cardiovascular disease, which are closely associated with sodium intake. However, given that patients with cancer (e.g., stomach cancer) are also affected by a low-salt diet. The KCHS does not include a question asking whether participants have cancer. In future research, it is necessary to examine the health behaviours of cancer patients by using other data sources that include cancer patients. # Conclusion The results of this study highlight the need to improve the low-salt preference rate in patients with major NCDs as well as in the general population. Specifically, it is necessary to consider expanding various low-salt policies and programs such as education on a low-salt diet, salinity monitoring to create a low-salt environment, certification for low-sodium restaurants, and sodium tax. In these contexts, the low-salt preference rates among major NCD patients, as determined in this study, will serve as baseline values for evaluating the effectiveness of various low-salt policies and programs. These findings can also be used as reference values for setting project targets. Therefore, it is necessary to continuously monitor the low-salt preferences of patients with major NCDs and the general population. # Supporting information 10.1371/journal.pone.0276655.r001 Decision Letter 0 Mogi Masaki Academic Editor 2022 Masaki Mogi This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Aug 2022 PONE-D-22-18808Comparison of low-salt preference trends and regional variations between patients with major non-communicable diseases and the general populationPLOS ONE Dear Dr. Ock, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ============================== Two reviewers well assessed this manuscript.However, there are still corrections existed in the present form of the manuscript. Read the suggestions carefully and respond them appropriately. ============================== Please submit your revised manuscript by Sep 28 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Masaki Mogi Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Thank you for stating in your Funding Statement: "This work was supported by the Research Program funded by the Korea Disease Control and Prevention Agency. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." Please provide an amended statement that declares \all\ the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at [http://journals.plos.org/plosone/s/submit- now. ](http://journals.plos.org/plosone/s/submit-now. ) Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement. Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf. 3\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: I Don't Know \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Dear Authors, The article provides a good insight into the trends in low-salt preference and regional differences in Korea among patients with the main non-communicable diseases. For a better understanding of the work, I have some suggestions: Line 118. Can the authors explain the background or give a reference why the 13 NCD listed are the main ones? Line 208: To the reader who does not know Korea well, the names of the regions do not give useful information. Would it make sense to describe each region (in few words) background that could influence salt consumption? For example, -a coastal region, traditionally lot of fish are consumed,… -an inland region, … -rural region... intensive urban region.. -a region where traditionally few meat is consumed... Lines 234 235: The low-salt preference rates of patients with different major NCDs and the general population was gradually decreasing after peaking mainly in 2013 The trend is evident from figures 2,3,4,5, can you comment the reasons for the trend? Line 305: In the title of the article you "regional variations" are "highlighted Can you briefly comment the reasons for the differences in patient behavior between regions? Best regards. Reviewer \#2: This study deals with an important issue within Korea as well as worldwide. The study uses the results of a large survey to analyze trends over time in the differences in low salt preference between major NCD patients and the general population, and offers suggestions for future environmental improvement and education. However, there are some concerns in this paper. 1\. Abstract: Result: The authors specifically describe the low salt preference rate. Since the survey years for this study are 2008-2019, Please indicate the year of the survey. 2\. Introduction: 87: The survey years are 2008-2017, but is it 2008-2019? 3\. Materials and methods: What is the size of The Korea Community Health Survey? Please provide sample size information for each survey year. 4.Materials and methods: Participants: The authors do not include cancer as Major NCD Patients, although they mention it in the study limitations. Why not include cancer in this analysis? In addition, since the results and discussion focus on hypertension, diabetes, and dyslipidemia, it would be better to exclude other diseases. I think it would be better to have the data of Fig2-Fug5 to grasp the overall trend of NCD, but I feel it is difficult to confirm the data. 5.Results:178-179:　“In the general population, it peaked in 2013 at 74.4% (APC: -11.788, p-value: \<0.001)”. Isn't the APC value 11.788? 6.Results:184-185：Other NCD results are not mentioned in the results that follow. If the relevant data set is not always available, then it may be possible to remove it. 7\. Results:194：Please list the ACP values for 2013. 8\. Results:201：Please correct the p-value. 9\. Discussion: The low salt preference rate and the adherence rate of not adding salt or soy sauce at the table have changed since 2013, but the reasons for this have not been mentioned. Is there any possible reason for this change, such as a change in the survey methodology, the selection method of the target population, or a change in government policy policy regarding salt reduction? 10\. Discussion：260-261：“the reaction to a salty taste is influenced by the sodium concentration in the food”. It would be helpful to the reader's understanding if you could describe the specific impacts while presenting the reports of the cited references. 11\. Discussion：260-261：Please describe in detail the contents of your education to cooks and foodservice professionals. 12\. Disccussion：287-288：Please cite any reports showing that low-salt preference is related to temperature. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \\\\\\\\\\ \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0276655.r002 Author response to Decision Letter 0 23 Sep 2022 Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: I Don't Know 3\. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Dear Authors, The article provides a good insight into the trends in low-salt preference and regional differences in Korea among patients with the main non-communicable diseases. For a better understanding of the work, I have some suggestions: Response: We really appreciate the careful and thorough review of this manuscript and for the thoughtful comments and constructive suggestions, which help to improve the quality of the manuscript. Our responses are as follows. Line 118. Can the authors explain the background or give a reference why the 13 NCD listed are the main ones? Response: Thirteen major types of NCD patients (hypertension, diabetes, dyslipidaemia, stroke, myocardial infarction, angina, arthritis, osteoporosis, asthma, allergic rhinitis, atopic dermatitis, cataract, and depression) were defined as participants who responded clinically diagnosed from physicians in the KCHS. We have revised the sentences (line 117-120). Line 208: To the reader who does not know Korea well, the names of the regions do not give useful information. Would it make sense to describe each region (in few words) background that could influence salt consumption? For example, -a coastal region, traditionally lot of fish are consumed,… -an inland region, … -rural region... intensive urban region.. -a region where traditionally few meat is consumed... Response: Thanks for the constructive suggestion. As suggested, we have added regional characteristics of Republic of Korea, including sodium intake, as a Supplemental file. Sentences that further research on regional variation are needed have been added to the Discussion section (Line 292-297). However, we ask for your understanding that we cannot cover various regional characteristics due to data limitations. Lines 234 235: The low-salt preference rates of patients with different major NCDs and the general population was gradually decreasing after peaking mainly in 2013 The trend is evident from figures 2,3,4,5, can you comment the reasons for the trend? Response: In Republic of Korea, the multicomponent national programs to reduce salt intake has been steadily promoted since the 2010s. In particular, in 2012, a national plan was established to reduce sodium intake. However, we estimated that this effect appears to have gradually weakened. These contents were further reinforced in the Discussion section (Line 256-259). Line 305: In the title of the article you "regional variations" are "highlighted Can you briefly comment the reasons for the differences in patient behavior between regions? Response: As mentioned in the previous response, the need for additional studies on regional variation was emphasized in the Discussion section (Line 292-297). Reviewer \#2: This study deals with an important issue within Korea as well as worldwide. The study uses the results of a large survey to analyze trends over time in the differences in low salt preference between major NCD patients and the general population, and offers suggestions for future environmental improvement and education. However, there are some concerns in this paper. Response: We really appreciate the careful and thorough review of this manuscript and for the thoughtful comments and constructive suggestions, which help to improve the quality of the manuscript. Our responses are as follows. 1\. Abstract: Result: The authors specifically describe the low salt preference rate. Since the survey years for this study are 2008-2019, Please indicate the year of the survey. Response: In the Materials and methods section, we have already mentioned that the Korea Community Health Survey is conducted annually (Line 110). The Korea Community Health Survey results for each year were used for analysis. 2\. Introduction: 87: The survey years are 2008-2017, but is it 2008-2019? Response: In the previous study (Reference number 10), the results of the Korea Community Health Survey from 2008 to 2017 were analyzed to analyze health behaviors, such as smoking and drinking, among major NCD patients. This content is a description of prior study. 3\. Materials and methods: What is the size of The Korea Community Health Survey? Please provide sample size information for each survey year. Response: As suggested, we added information about sample size from the Korea Community Health Survey to the Materials and methods section (Line 113-114). 4.Materials and methods: Participants: The authors do not include cancer as Major NCD Patients, although they mention it in the study limitations. Why not include cancer in this analysis? Response: Unfortunately, the Korea Community Health Survey does not include a question asking whether participants have cancer. We emphasize this point further in the part on limitations in the Discussion section (Line 312-314). In addition, since the results and discussion focus on hypertension, diabetes, and dyslipidemia, it would be better to exclude other diseases. I think it would be better to have the data of Fig2-Fug5 to grasp the overall trend of NCD, but I feel it is difficult to confirm the data. Response: As you know, among various NCDs, hypertension, diabetes, and dyslipidemia are important diseases that act as risk factors for other diseases. Also, these diseases have a higher prevalence than other diseases. So, Figure 1 focused on hypertension, diabetes, and dyslipidemia. In addition, a Supplemental file 1 is attached so that readers can check the specific values for each disease. 5.Results:178-179:　“In the general population, it peaked in 2013 at 74.4% (APC: -11.788, p-value: \<0.001)”. Isn't the APC value 11.788? Response: As you pointed out, this is a typo. We have corrected that value (Line 182). 6.Results:184-185：Other NCD results are not mentioned in the results that follow. If the relevant data set is not always available, then it may be possible to remove it. Response: As suggested, we have deleted that sentence. 7\. Results:194：Please list the ACP values for 2013. Response: It has a single APC with respect to its content. That is, the APC is -0.152. 8\. Results:201：Please correct the p-value. Response: We have modified the p-value (Line 202). 9\. Discussion: The low salt preference rate and the adherence rate of not adding salt or soy sauce at the table have changed since 2013, but the reasons for this have not been mentioned. Is there any possible reason for this change, such as a change in the survey methodology, the selection method of the target population, or a change in government policy policy regarding salt reduction? Response: In Republic of Korea, the multicomponent national programs to reduce salt intake has been steadily promoted since the 2010s. In particular, in 2012, a national plan was established to reduce sodium intake. However, we estimated that this effect appears to have gradually weakened. These contents were further reinforced in the Discussion section (Line 256-261). 10\. Discussion：260-261：“the reaction to a salty taste is influenced by the sodium concentration in the food”. It would be helpful to the reader's understanding if you could describe the specific impacts while presenting the reports of the cited references. Response: As suggested, we revised the sentence (line 265-266). 11\. Discussion：260-261：Please describe in detail the contents of your education to cooks and foodservice professionals. Response: The content of the sentence you pointed out has been modified to be more specific (Line 286-287). 12\. Disccussion：287-288：Please cite any reports showing that low-salt preference is related to temperature. Response: We have corrected the sentence and added a sentence that additional research on regional variation is needed (Line 292-297). 6\. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer \#1: No Reviewer \#2: No 10.1371/journal.pone.0276655.r003 Decision Letter 1 Mogi Masaki Academic Editor 2022 Masaki Mogi This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 12 Oct 2022 Comparison of low-salt preference trends and regional variations between patients with major non-communicable diseases and the general population PONE-D-22-18808R1 Dear Dr. Ock, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Masaki Mogi Academic Editor PLOS ONE Additional Editor Comments (optional): The authors well responded to the Reviewers' comments. Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Dear Authors, Thanks for the comments and inclusion of reccomendations. I wish you successful further research work. Best regards Reviewer \#2: (No Response) \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \\\\\\\\\\ 10.1371/journal.pone.0276655.r004 Acceptance letter Mogi Masaki Academic Editor 2022 Masaki Mogi This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 17 Oct 2022 PONE-D-22-18808R1 Comparison of low-salt preference trends and regional variations between patients with major non-communicable diseases and the general population Dear Dr. Ock: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Masaki Mogi Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction Diffusion kurtosis imaging (DKI) is a new in vivo method for diffusion imaging that originated from diffusion tensor imaging (DTI). DTI provides a way for probing the microstructure of biological tissues by measuring the diffusion coefficient of water molecules using Gaussian models. Gaussian models can only be used to depict free diffusion processes. However, water diffusion in biological tissues is normally restricted, which constitutes the basis of DTI tractography, and therefore is not exactly Gaussian. To address this self- contradiction, DKI is proposed to model the Gaussian coefficient of diffusion as well as the deviation from the Gaussian model, thereby providing new insights into the microstructures,,. DKI has already been successfully used in a wide range of clinical studies, including studies in Parkinson’s disease, Huntington’s disease, epilepsy, aging, attention deficit hyperactivity disorder and cerebral gliomas. DTI data are composed of baseline images without applying a diffusion gradient and a series of diffusion-weighted (DW) images with diffusion gradients applied along different directions. Typically, DW images are acquired at one b-value, which is an index calculated based on the strength and duration of the diffusion gradient field, and the interval between the two opposite gradients. The diffusion coefficient along a certain direction can be calculated using the following equation: $$\text{S}_{\text{n}} = S_{0}\exp( - bD_{n})$$ where S<sub>0</sub> is the baseline signal, S<sub>n</sub> is the DW image acquired with diffusion gradient along the n<sup>th</sup> direction, b is the b-value of the applied diffusion gradient and D<sub>n</sub> is the diffusion coefficient along the n<sup>th</sup> direction. To construct a diffusion tensor, at least seven DW measurements are needed, including a baseline image. In DTI, only one DW image is needed for each direction. However, DKI requires multiple DW images at different b-values along each direction. DKI uses the following equation to depict the change of the signal intensity with respect to b-values: $$S_{n}(b) = S_{0}\exp( - bD_{n} + b^{2}D_{n}^{2}K_{n}/6)$$ where S<sub>0</sub>, S<sub>n</sub>, b, and D<sub>n</sub> remain the same as in Eq., and K<sub>n</sub> is the kurtosis coefficient, which depicts the deviation from a Gaussian model. For each direction, both D<sub>n</sub> and K<sub>n</sub> can be calculated by curve fitting using the baseline image and the corresponding DW images. Then mean diffusion (MD) and mean kurtosis (MK) values can be calculated by averaging the D<sub>n</sub> and K<sub>n</sub> along each direction. The MD and MK can also be calculated from diffusion and kurtosis tensor respectively, which are reconstructed using D<sub>n</sub> and K<sub>n</sub> along all directions. The two methods generate similar MD and MK values, while the direct averaging method makes no restriction on the number of diffusion gradient directions and therefore is more efficient in computation. To profile the kurtosis deviation, a typical DKI setting would employ the maximum b-value from 2000 to 2500 s/mm<sup>2</sup>, which is much higher than the 1000 s/mm<sup>2</sup> that is typically used in DTI. DW images in a DKI dataset therefore often suffer from heavier noise, resulting in low signal-to- noise ratios (SNR) at these higher b-values. The noise, which is Rician,, in turn, may significantly affect the reliability of parameter estimation. A previous study showed that Rician noise in DW images may lead to significant overestimation of the kurtosis coefficients. Various denoising methods have been developed to improve the quality of DW images, such as the Gaussian filter,, anisotropic diffusion filter,, , linear minimum mean squared error filter, and non-local means (NLM) filter. The NLM filter outperforms most other filters in both denoising and edge preserving, and thus has been used extensively in magnetic resonance (MR) image denoising,. While DW images are often denoised on an image-wise basis, correlation between DW images should also be exploited, as it contains spatial cues of the imaging data. Wiest et al. proposed a vector non-local means (VNLM) filter based on the NLM filter for denoising DTI data. It bundles all DW images in a DTI dataset into a vector image and applies NLM to denoise it as one whole entity. To identify the best denoising algorithm among the existing algorithms or to further tailor them for treating a specific type of image, a method to assess their performance is required. Both visual,, and quantitative,, comparisons can be used for this purpose. In a visual comparison, a good denoising algorithm should have (1) minimized image noise; (2) preserved image details; and (3) introduced no artifacts. Quantitative comparisons often use noise-free images as ground truth, to which noise of a known distribution is added for simulating phantom data for testing purposes. In this way, a denoised image can easily be compared quantitatively with the ground truth. Because DTI and DKI contain high dimensional information in each voxel, ground truth based on the maps of diffusion-derived parameters is often favorable for evaluating the performance of denoising algorithms for diffusion imaging data. In previous studies of DTI denoising, DTI phantoms were constructed for denoising DTI data by evaluating the reliability of estimating DTI-derived parameters. However, to the best of our knowledge, no such phantom has been reported for DKI denoising, although various DKI schemes were previously evaluated systematically using a specifically designed simulation dataset of diffusion parameters. To evaluate how denoising algorithms can affect the precision of DKI parameter estimation, we developed a pipeline for constructing DKI phantoms and consequently created a DKI phantom from real brain data. We used it to quantitatively evaluate NLM and two different VNLM schemes. The first VNLM scheme combined DW images at the same b-value as a vector whereas the second combined DW images along the same diffusion gradient direction as a vector. We conducted evaluations using our phantom to check which VNLM performs the best. # Materials and Methods ## Materials This study was approved by the Institutional Ethics Committee of East China Normal University. Four local volunteers were recruited to the study and informed written consents were obtained from all these volunteers. A 12-channel head coil was used in data acquisition. DKI data from the volunteers were collected on a 3T Siemens Trio system (maximum gradient strength = 40 mT/m, maximum slew rate = 200 mT/m/ms). A bipolar single-shot EPI sequence was used for DW image acquisition to minimize the eddy current artifacts. Dataset 1, which contained DKI data from one volunteer, was acquired using conventional acquisition parameters for DKI data, with DW images at 6 b-values (0, 500, 1000, 1500, 2000, 2500 s/mm<sup>2</sup>) along 30 diffusion gradient directions, 35 slices, NEX = 2 (averaged in image domain), spatial resolution = 2 × 2 × 3 mm<sup>3</sup>, FOV = 256 × 256 mm<sup>2</sup>, acceleration factor of parallel imaging = 2 (GRAPPA). The other parameters were: TR / TE = 6000 ms / 112 ms, diffusion time Δ = 39.1 ms, and diffusion gradient duration δ = 37.5 ms. The diffusion time Δ here is defined not exactly the same as that in standard Stejskal–Tanner (monopolar) sequence because we used the bipolar single-shot EPI sequence here, while Δ has been calculated according to the conventional expression b = -(γGδ)<sup>2</sup> (Δ – δ/3). The acquisition time was 30 min 32 sec. Thirty extra baseline images (therefore a total of 32 baseline images) were collected for generating one baseline image of high SNR. This dataset was later used for DKI phantom construction. Dataset 2 contained DKI data from the other three volunteers that were acquired with slightly different parameters. DW images were collected along 12 / 12 / 20 diffusion gradient directions, TR = 6100 / 12000 / 5300 ms, TE = 114 ms, Δ = 40.1 ms, δ = 38.5 ms, 35 / 40 / 40 slices, and NEX = 1. The other parameters were the same as those used for Dataset 1. In addition, we applied physical constraints to the participants during the data acquisition, and screened the acquired data afterwards to prevent motion-induced artifacts. All the raw data are available at <http://pan.baidu.com/s/1ntMD68x#path=%252F>. ## DKI Phantom Construction Because no noise-free DKI image is readily available as ground truth, we developed a pipeline to construct a DKI phantom based on a real DKI dataset of the human brain. The process consists of 8 steps: (1) denoise each DW image (non-brain region was removed through the BET tool of FSL (<http://www.fmrib.ox.ac.uk/fsl/>)) using a 3D UNLM filter (see next subsection) after eddy current correction and motion correction using ACID Toolbox (<http://www.diffusiontools.com>); (2) average the DW images from repeated scans; (3) estimate D and K value maps for each gradient direction using Eq.; (4) apply 3D Gaussian filter to D maps using a Gaussian kernel of 2 mm full-width-at-half-maximum (FWHM), and consequently obtain D’; (5) recalculate K’ with obtained D’ using Eq.; note that in step 4) and 5), we do not directly smooth K but recalculate K using a smoothed D as the noise in the D map may significantly influence K estimation, especially for voxels with small D values; (6) reconstruct diffusion tensors and kurtosis tensors using the D’ and K’ maps of all gradient directions, and the tensor data were then used to recalculate D and K maps, which are denoted by D” and K” respectively; (7) use D”, K” and baseline image to calculate DW images with nonzero b-values (again, using Eq.) which forms our noise-free DKI phantom; (8) add noise to the noise-free phantom, Rician noise can be added using Eq.: $$I_{n} = \sqrt{{(I + n_{1,\sigma})}^{2} + n_{2,\sigma}^{2}}$$ where n<sub>1,σ</sub> and n<sub>2,σ</sub> are both Gaussian distributed noise with standard deviation of σ. A fixed σ is used for all DW images when noises are added. A discussion on noise adding can be found in Section of Discussion and Conclusion. ## Non-Local Means (NLM) Filter Family NLM is a spatial domain filter that replaces each pixel P(i) in the image with a weighted average of every pixel P(j) in its “search region” Ω: $$NLM(P(i)) = Z_{0}{\sum\limits_{\forall j \in \Omega}{\omega(i,j)P(j)}}$$ where Z<sub>0</sub> is the normalization coefficient, defined as: <img src="info:doi/10.1371/journal.pone.0116986.e005" id="pone.0116986.e005g" /> Z 0 = 1 / ∑ ∀ j ∈ Ω ω ( i , j ) The weight ω(i,j) assigned to P(j) is based on the weighted Euclidean distance d between the neighborhoods of pixels i and j, named R<sub>f</sub>(i) and R<sub>f</sub>(j) respectively: $$d(i,j) = G_{\rho}\left\\| {R_{f}(i) - R_{f}(j)} \right\\|^{2},(i \neq j)$$ $$\omega(i,j) = \exp( - d(i,j)/h^{2}),(i \neq j)$$ where h is a parameter that controls the degree of smoothing and is normally set proportionally to the standard deviation of noise. G<sub>ρ</sub> is a Gaussian kernel of standard deviation ρ. In theory, the search region Ω in Eq. can cover the whole image (thus non- local). However, a limited radius t is commonly adopted with regard to computational efficiency,. When calculating the weight of the center pixel itself, the distance is simply set to the minimum distance found in the search region. Similar to other weighting average filters, larger weights are assigned to pixels with higher similarity. NLM is unique in that it uses the distance between neighborhoods of pixels instead of the distance between pixels themselves. Thus, it can make use of redundant information in texture patterns in the image for robust denoising. Manjόn and his colleagues proposed an unbiased non-local means (UNLM) filter to correct the gray level bias introduced by Rician noise that is typical in MR images. UNLM subtracts the bias from the NLM filtered image, which can be expressed as: $$UNLM(P(i)) = \sqrt{{(NLM(P(i)))}^{2} - 2\sigma_{r}^{2}}$$ where σ<sub>r</sub> denotes the standard deviation of Rician noise. In this work, all algorithms involved in the comparison use bias subtraction as in UNLM. Wiest et al. proposed the VNLM filter to denoise the DTI dataset. As previously mentioned, DTI acquires DW images using at least six different diffusion directions. VNLM groups these images into a vector image and denoises the vector image as a whole. Therefore, the distance in NLM is redefined as the distance d<sub>v</sub> between neighborhoods of two vectors: $$d_{\nu}(i,j) = G_{\rho}{\sum\limits_{\nu = 1}^{V}\left\\| {R_{f,\nu}(i) - R_{f,\nu}(j)} \right\\|}^{2}/V,(i \neq j)$$ where v and V denote the index and total number of DW images, respectively. We realize that when VNLM is applied to DKI dataset, there are two different ways to combine DW images into vector images. One is to combine images of the same b-value but of different directions of diffusion gradient as a vector (VNLM-b), and the other is to combine images of the same direction of diffusion gradient, but of different b-values (VNLM-d). When applying NLM to MRI, Coupe et al. extended it to 3D, in which both the neighborhood window and the search region become cubes centered at the pixel in concern. While this makes better use of the redundant structure information in the 3D MRI data, parameters of the 3D NLM filter should be carefully set for balancing the denoising effect and computational efficiency. # Experiments and Results We adopted 3D NLM-based filters in the evaluation. Regarding the parameter setting, previous work showed that no significant improvement can be achieved with a search region greater than 11 × 11 = 121 pixels for a 2D NLM filter, and using a larger search region will significantly increase computational time. Thus, we adopted a 5 × 5 × 5 search region (125 voxels in 3D instead of 121 in 2D case), which is the same parameter used in previous work. In addition, a 3 × 3 × 3 neighborhood window was employed in our work. A neighborhood window of such a size is common in 3D NLM processing,, which is normally smaller than the search region. The parameter ρ was set to commonly used value 1.0. The value of the parameter h for NLM, VNLM-b, and VNLM-d was set to 1.0σ, 0.8σ, and 1.2σ respectively according to their optimal performance based on an exhaustive search. The search result of NLM agrees with those in previous reports. The standard deviation of noise σ in real data was calculated from a background region of the image. Assuming that the signal in the background region consists of only Rician noise, σ can be estimated from: $$\sigma\text{=}\sqrt{\mu/2}$$ where μ is the mean value of squared signal intensity in the background region. A discussion on noise estimation can be found in Section of Discussion and Conclusion. We used Mean Square Error (MSE), Bias and standard deviation (Std) for quantitative comparison of the denoising methods, which reflect precision and accuracy of the denoising methods by their definitions: $$MSE = \frac{1}{N}{\sum\limits_{i}^{N}{(I_{i} - Q_{i})}^{2}}$$ $$Bias = \frac{1}{N}\left\\| {\sum\limits_{i}^{N}{I_{i} - Q_{i}}} \right\\|_{1}$$ $$\text{Std} = \sqrt{\frac{1}{N}{\sum\limits_{i}^{N}{(I_{i} - \frac{1}{M}{\sum\limits_{\text{j}}^{M}I_{j}})}^{2}}}$$ where I and Q denote noise-free and denoised image respectively, and N is the number of voxels. The MSE, Bias and Std are calculated only in the brain region. ## DKI Phantom Construction We created a DKI phantom from Dataset 1 using the aforementioned process. Compared with the original images, the constructed phantom is visually cleaner. In the phantom, noise is successfully suppressed, and anatomical structures are well preserved. ## Filter Comparison with DKI Phantom We compared the three NLM-based denoising algorithms mentioned earlier, NLM, VNLM-b and VNLM-d. First, five different levels of Rician noise (with standard deviations at 5, 10, 15, 20, 25) were added to DW images of the noise-free phantom. The resulting SNRs are different for DW images with distinct b-values, which are approximately 60, 30, 20, 15, 12 for baseline image (b = 0), or 12, 6, 4, 3, 2 for DW images with b = 2500 s/mm<sup>2</sup>. The noise-corrupted images were then denoised using the three filters, and results were evaluated both visually and quantitatively using the phantom and real dataset. Quantitative comparisons can be performed in multiple ways. For example, the denoised DW images can be compared with phantom DW images by calculating the MSE between them. Parameter maps calculated from denoised DW images can also be compared with those of the phantom. The latter approach allows a more comprehensive assessment of the filters with respect to parameter evaluation and should be favored in cases like DKI, in which producing a reliable parameter map is often the target of denoising. In this study, MD and MK maps, which were calculated by averaging D and K maps respectively along all diffusion gradient directions, are used to evaluate the denoising results. For robust statistical results, we repeated the above process that adds noise, denoises, and conducts performance evaluation for 500 times. The quantitative comparison of MD and MK maps shows that VNLM-d filter achieves lower MSE, Bias and Std than the original NLM and VNLM-b filters at almost all noise levels. Moreover, the MSE, Bias and Std values of MK produced by the VNLM-d filter increase most slowly with the increase of noise level. When standard deviation of noise reaches 25, the MSE, Bias and Std values from the VNLM-d filter are only 40.6% (MSE), 10.9% (Bias) and 60.6% (Std) respectively of those from the NLM filter. Meanwhile, it is interesting that comparing with the other two filters, the VNLM-d filter shows equivalent or even poorer performances concerning MSE and Std of the denoised DW images, which is contrary to its good performance for MD and MK maps as discussed above. A similar conclusion can be drawn from visual comparison. For MK maps, results from these three filters have different visual appearances. MK maps of NLM and VNLM-b filters contain black holes, which represent incorrectly estimated voxels,. The MK map from VNLM-d filter is almost free of black holes and has a consistent visual appearance with the phantom MK map (see middle row of). Moreover, the DW image denoised by the VNLM-d filter show clearer those finer structural details than do those denoised by NLM and VNLM-b filters (see upper row of). The VNLM-d filter produces visually satisfying results even for low SNR DW images acquired with b = 2500 s/mm<sup>2</sup>. ## Denoising of Real dataset To further validate the observation obtained from the phantom dataset, we also applied the NLM, VNLM-b and VNLM-d algorithms to Dataset 2 and evaluated the results for their performance. The parameters to these algorithms were set the same as described previously. Visual comparison of the real dataset has revealed an impression similar to the information shown from the simulation data. While all filters can significantly reduce noise in the DW images and parameter maps, VNLM-d produces DW images with clearer brain structure and MK map with fewer black holes for all three volunteers. In addition, we may see that the MK map provides microstructural information in gray matter, which is not visible in MD or fractional anisotropy (FA) maps of diffusion tensor imaging model. # Discussion and Conclusion In this work, we propose a processing pipeline for constructing DKI phantoms using datasets of real human brains. The pipeline produces a high-SNR DKI phantom with clear image structure that can be used as ground truth for evaluation of DKI denoising methods. In addition, the proposed pipeline can also be used to generate a DKI dataset with a customized combination of b-values and diffusion gradient directions, different from those used in acquisition. This flexibility will be useful to evaluate the performance of various DKI acquisition schemes, such as 3-b-value fast scheme and also non-uniform schemes with different gradient directions at each b-value. For these purposes, after the diffusion tensors and kurtosis tensors have been calculated in step (6), we can first apply a customized gradient table to recalculate D and K maps at the specified directions. DW images with desired b-values can then be calculated with D, K maps and baseline images using Eq.. We should pointed out that Rician noise was used in step (8), because magnitude of MR signal is intrinsically corrupted by Rician noise, which is a model frequently adopted in MRI denoising and parameter estimation studies,. Nevertheless, the potential use of our DKI phantom is not limited to removal of Rician noise, as discussed above. Furthermore, we quantitatively evaluated three NLM-based denoising algorithms using the constructed phantom. The simulation based on our phantom indicates that VNLM-d outperforms NLM, and the VNLM-b algorithms, generating more reliable MK and MD maps, with the lowest MSE, Bias and Std values for most of the noise levels. Visual comparison of these filters using a real dataset produces results consistent with this conclusion. While compared with the other methods, VNLM-d algorithm produces DW images with equal or higher MSE and Std. This can be explained by the fact that VNLM-d tends to smooth more conservatively the structure in gray matter regions of the DW images. This lowers the level of image denoising but preserves more fine structures (Figs.) which are helpful in the following parameter evaluation process. Thus, it yields less black hole effects and preserves structure better in MK maps. We think the good performance of VNLM-d filter can be attributed to the reason that the DKI parameter is calculated from diffusion decay curve, which means similar shape of the decay curves generate similar DKI parameters. VNLM-d filter treats DW images of different b-values at one direction as a unit in similarity calculation, thus voxels with similar decay curves, which means similar DKI parameters, tend to contribute more in the weighted average process. This efficiently exploits the similarity of DKI parameters. The structural preserving ability of VNLM-d filter may be attributed to the higher level of structural similarities between DW images, because VNLM is effective only when images with similar structures are grouped together, which have similar weights for averaging. We found DW images along the same direction but at different b-values (column b and c) generally demonstrate higher similarities than do the DW images acquired at the same b-value but along different directions (column a and b). Although the voxel intensities of DW images at the same b-values fall in a similar value range and are therefore more consistent with one another, it is the similarity between structures that leads to the superior performance of VNLM. To further improve this work, several considerations can be taken into account. Firstly, DW images are often acquired with partial Fourier techniques, which can produce correlation between noises in neighboring pixels. This may present a new challenge to the denoising algorithms because many algorithms make an assumption of non-correlated noise. Despite of this, NLM filter has already been successfully applied to denoising DW images. This is understandable since when NLM calculates the weighted average to denoise a pixel, it considers the similarity between the neighborhood window of this pixel and those of the contributing pixels. The correlation between noises in neighboring pixels does not necessarily lead to similar neighborhood of these pixels, thus NLM filters are more robust to process correlated noise. Recently, NLM filter has also been improved to address correlated noise, and it is also suggested that denoising filter be used before partial Fourier reconstruction is carried out to avoid correlated noise. Incorporating these results into our work may produce better results. Secondly, in the procedures of adding and estimating noise, we assumed that the levels, spatial and statistical distributions of noise are all the same for all DW images at different b-values and along different diffusion gradient directions, which has been a commonly adopted hypothesis in previous studies . However, this assumption may become invalid in certain cases. For example, the eddy current and off-resonance effects in a DWI sequence may potentially affect the noise, and these effects may substantially vary with b-value and diffusion gradient direction. In addition, the spatial and statistical distribution of noise can also be affected by the use of multi-element surface coils and parallel imaging. For example, the noise distribution of parallel imaging is associated with a geometry factor (g-factor), which depends on coil geometry, phase-encoding direction and its acceleration factor. Thus, to more accurately evaluating the effects of simulating and estimating the noise in DW images, we must do a more careful simulation and inspect the influences that may be imposed by these factors. Limited by the length of this paper, and considering that this paper’s focus on reporting the general framework of providing a DKI phantom system, we decided not to pursue it in this study but to include it in our next step work. Thirdly, motion artifacts, eddy current and geometrical distortion are major challenges in preprocessing of diffusion imaging data. Navigator based methods can be used to minimize motion artifacts prospectively. Multi-shot EPI or fast spin echo sequences can also be used to reduce geometry distortion due to susceptibility changes at tissue interfaces. Thus these methods should be considered in our further work to improve the quality of DKI estimation. In our constructed phantom data, signal dropout can be found in medial frontal and bilateral gray matter. We think this can be attributed to individual variations that this particular individual happened to introduce some motion between the averages of data, and consequently motion artifacts. Retrospective Motion correction may not completely eliminated these artifacts. [^1]: Author Xu Yan is employed by a commercial company, Siemens Healthcare, was a MR collaboration scientist doing technique support in this study under Siemens collaboration regulation without any payment and personal concern regarding to this study. All other authors declare no conflict of interest. There are no patents, products in development or marketed products to declare. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: MZ XY GY. Performed the experiments: XY HZ. Analyzed the data: MZ XY. Contributed reagents/materials/analysis tools: MZ XY HX HZ. Wrote the paper: MZ XY GY DX.
# Introduction The ubiquitously expressed non-receptor protein-tyrosine phosphatase (PTP1B) plays an important role in regulating diverse cellular signaling pathways, including those initiated by receptor tyrosine kinases (RTKs), cytokine receptors, integrins and cadherins. Major insights into the physiological role of PTP1B were gained through the generation of knockout (KO) mice, which showed that PTP1B is a critical regulator of insulin sensitivity and energy balance in vivo. However, other functions for PTP1B have also been suggested, including roles in regulating cell-matrix and cell-cell, interactions. Given the salutary metabolic effects of PTP1B deletion, it has emerged as a potential target for anti-diabetic and anti-obesity drug development. Consequently, it is important to understand its mechanism of action in detail. PTP1B is anchored to the cytosolic face of the endoplasmic reticulum (ER) via a hydrophobic C-terminal targeting sequence, which constrains its access to key substrates. Consistent with this localization, PTP1B dephosphorylates the activated epidermal growth factor receptor (EGFR), platelet-derived growth factor receptor (PDGFR) and insulin receptor (IR) only after endocytosis, as they transit past the ER. PTP1B activity also is spatially regulated in the cell, thus creating distinct microenvironments that enable RTK signal propagation, followed by signal termination. Recently, PTP1B has been identified as a potential regulator of RTK endocytosis. Eden et al. reported that PTP1B-EGFR interaction occurs through direct membrane contact between multivesicular bodies (MVB) and the ER, with PTP1B activity promoting the sequestration of EGFR to MVB internal vesicles. Consistent with these findings, Stuible et al. identified the endosomal protein STAM2, which regulates sorting of activated RTKs for degradation, as a PTP1B substrate. Collectively, these studies reveal that PTP1B is a major regulator of RTK endocytosis and signaling. Although activated RTKs gain access to PTP1B only after endocytosis, PTP1B also can interact with some plasma membrane (PM)-bound substrates. For example, PTP1B targets forming cell-matrix adhesion contacts and contributes to the stabilization of focal adhesions. This process appears to involve dynamic extension of the ER via a microtubule-dependent process. PTP1B also can access substrates at points of cell-cell contact, although how these interactions are regulated remains largely unexplored. In the present study, we assess the dynamics of PTP1B mobility and investigate its spatial-temporal regulation of signaling at regions of cell-cell contact. Through the combined use of PTP1B mutants, advanced cell imaging and mathematical modeling, we show that ER- anchored PTP1B can reach PM-localized substrates, but only at regions of cell- cell contact. These studies strongly suggest that the ER is structured and polarized towards cell-cell junctions, and identify these PM-proximal sub- regions of the ER as important sites of cellular signaling regulation by PTP1B. # Results ## PTP1B Substrate-trapping Mutant Localizes to Regions of Cell-cell Contact We transiently expressed green fluorescent protein (GFP)-tagged wild type (WT) PTP1B and its “substrate-trapping” mutant D181A (D/A), which retains substrate- binding but is catalytically impaired, in PTP1B-null fibroblasts, and monitored their sub-cellular localization using confocal microscopy. Consistent with previous reports, each localized to the ER network. However, PTP1B D/A, but not PTP1B WT, also accumulated at regions of cell-cell contact, labeled using anti-β catenin antibodies. Studies using another PM marker, Cherry-tagged G protein- coupled receptor 43 (Grp43-Cherry), confirmed that PTP1B D/A accumulated at regions of cell-cell contact in live cells (lower panel). Quantification of fluorescence intensity showed that about 10–15% of total cellular PTP1B D/A was found at cell-cell contacts. Similar findings were obtained using Cos-7, MDCK, 293T cells and 3T3-L1 preadipocytes (data not shown). ## The ER Extends to Regions Proximal to PM at Cell-cell Contacts These observations raised the question of how an ER-bound PTP might access substrates on or near the PM at regions of cell-cell contact. In platelets, PTP1B can be cleaved by calpain to release an active cytosolic fragment. However, serial confocal images of tissue culture cells expressing PTP1B D/A-GFP, acquired at successive focal planes, revealed a “honeycomb” pattern at regions of cell-cell contact which is characteristic of the reticular structure of the ER and consistent with its extension to these regions. Experiments in which PTP1B D/A was co-expressed in MDCK cells with the general ER marker, stress-related ER protein (SREP), confirmed that the ER extends out to the PM at regions of cell-cell contact (arrows). Comparable results were obtained when PTP1B D/A was co-expressed with a marker (Sec61) for rough ER. Exposure of cells to pervanadate, which oxidizes the essential cysteinyl residue found at the catalytic center of PTPs, abolished the localization of PTP1B D/A at cell-cell contacts without altering its ER localization (or that of PTP1B WT in the ER). The latter findings indicate that PTP1B D/A enrichment at regions of cell-cell contact is mediated by interactions with one or more substrate(s). Electron microscopy using immunogold-coupled anti-human PTP1B (FG6) antibodies revealed labeling of PTP1B D/A in the ER, providing direct evidence that PTP1B resides in this compartment (arrows, left image). Prominent labeling also was detected at regions of cell-cell contact (, right image and inset), where four adjacent membranes were evident: two inner “thick” membranes (arrowheads) corresponding to the PMs of adjacent cells and two flanking “thinner” membranes (arrows), corresponding to the ER. The few gold particles that appeared to label the PM probably reflect the antigen-gold particle linker distance: because PTP1B should be separated from the gold particle by the distance of the antibody molecule in three dimensions, the particle can lie directly over PTP1B or up to 10 nm away. No significant labeling was detected in PTP1B-null fibroblasts, confirming the specificity of the immunogold labeling. To ask whether the observed proximity of ER to PM at regions of cell-cell contact is caused by PTP1B D/A expression, we performed routine EM and high resolution cryo-EM on vitreous sections from HeLa cells. These studies confirmed that the ER is in close proximity to the PM at regions of cell-cell contact even in the absence of PTP1B D/A. Collectively, these studies show that the ER is in close proximity to the PM at regions of cell-cell contact and that ER-anchored PTP1B engages substrate(s) at these locations. ## PTP1B Mobility on ER Membranes To determine how an ER-bound phosphatase accesses PM substrate(s) at regions of cell-cell contact, we examined the mobility of PTP1B fused to photoactivatable GFP (PhAc), which can be used to mark a population of molecules in a region of interest and track them over time in live cells. PTP1B WT-PhAc and PTP1B D/A-PhAc were transiently co-expressed in PTP1B-null cells with the corresponding RFP-tagged PTP1B, which served as a general ER marker and helped to identify transfected cells. Prior to photoactivation, no significant fluorescence attributable to PTP1B-PhAc (excitation at 488 nm) was detected. Photoactivation (excitation at 413 nm) of selected regions resulted in pools of PTP1B WT or PTP1B D/A that became highly fluorescent upon excitation at 488 nm. Continuous imaging of the photoactivated pools revealed rapid, non-directional movement of PTP1B WT throughout the ER, with fluorescence intensity increasing gradually at regions distal from the photoactivation site (note intensity gain in distal region marked “4”). PTP1B D/A-PhAc (but not PTP1B WT-PhAc) that was initially photoactivated in the ER (away from sites of cell-cell contact) became detectable at regions of cell-cell contact, indicating that the ER is contiguous with these areas and is the source of PM-proximal PTP1B. By contrast, PTP1B D/A that was photoactivated at regions of cell-cell contact (rectangle) diffused away more slowly compared with pools activated in the ER (see below for detailed discussion of this apparent difference in speed). Thus, the movement of PTP1B is consistent with non-directional diffusion throughout the ER. Moreover, these studies show that the ER network is continuous, but lies proximal to the PM only at regions of cell-cell contact. ## PTP1B Mobility is Affected by Transient Interactions The photoactivation experiments were complemented with fluorescence recovery after photobleaching (FRAP) studies. GFP-labeled PTP1B WT or PTP1B D/A, expressed in randomly growing PTP1B-null cells, was photobleached irreversibly at the ER (circular area, 4 µm diameter) or at regions of cell-cell contact (rectangular area, 5 µm length/1 µm width), and the recovery of fluorescence in the photobleached area, caused by the diffusion of unbleached PTP1B, was monitored. Diffusion and binding parameters were determined from the acquired data by fitting the recoveries to an exact mathematical solution based on a simplified geometry ( and mathematical modeling section). For assessing intracellular diffusion, recoveries for each molecule were fit to a cylindrical model that had two free parameters: the effective diffusion constant and the immobile fraction. The effective diffusion constant measures the combined effects of actual diffusion speed (cytoplasmic or along the ER, in the absence of interactions), as well as any transient binding interactions that can slow molecular mobility. The immobile fraction reflects the fraction that does not recover (at least on the timescale of several minutes). The fitted effective cytoplasmic diffusion constant (D<sub>eff</sub>) of PTP1B WT was 0.23<u>+</u>0.01 µm<sup>2</sup>/s with an immobile fraction of 0.12<u>+</u>0.02 (mean <u>+</u> S.E.M.; n = 11 cells). This diffusion constant was comparable to that of the closely related, ER-anchored TCPTP (0.20<u>+</u>0.01 µm<sup>2</sup>/s) which, however, had a statistically insignificant immobile fraction (0.04<u>+</u>0.03; n = 12) and other ER-anchored proteins, but was significantly lower than the predominantly cytosolic Src homology-containing tyrosine phosphatase 2 (SHP2) (1.0<u>+</u>0.1 µm<sup>2</sup>/s with a statistically insignificant immobile fraction of 0.04<u>+</u>0.03; n = 9). The effective diffusion of PTP1B D/A in the ER was three-fold lower than that of PTP1B WT: 0.070<u>+</u>0.003 µm<sup>2</sup>/s, with a highly significant immobile fraction of 0.28<u>+</u>0.02 (n = 13). This lower diffusion constant (due to a slower turnover for the transient interactions) and higher immobile fraction (indicating an increased amount of stable associations) presumably reflect interactions of PTP1B D/A with substrate(s) in the ER. The slower, but still efficient recovery of PTP1B D/A - on a timescale of 10 seconds - automatically sets a lower limit on the dissociation rate for transient interactions of roughly s<sup>−1</sup> (a slower dissociation rate is irreconcilable with the observed efficient recovery). The recovery of PTP1B D/A fluorescence at regions of cell-cell contact was even slower. The slower recovery might be explained in a number of ways, including the peripheral location of the FRAP region (replenishing proteins must on average traverse a larger distance and can approach from only one direction), the amount of PTP1B D/A that accumulates at regions of cell-cell contact (see mathematical model and), and/or a slow off-rate from binding sites at regions of cell-cell contact. To help differentiate between these possibilities, we fit the recoveries to an exact mathematical solution that employed a simplified rectangular geometry. This model implicitly accounts for reversible binding of PTP1B to its substrate(s) in the ER by assuming the same effective diffusion constant determined above from the FRAP measurements for PTP1B D/A in the ER, but explicitly accounts for enzyme-substrate binding (both transient and immobilizing) of the accumulated fraction at the region of cell-cell contact (see mathematical model). Estimation of the exact immobile fraction at regions of cell-cell contact was hampered because we were unable to track recovery to completion due to photobleaching and cell movement (on timescales greater than one minute). Models assuming no immobile fraction or up to 30%-70% (depending on the cell) fit the data equally well (the fits shown in and assume no immobile fraction). This uncertainty concerning the exact immobile fraction did not, however, affect our ability to estimate any of the other parameters. Assuming an approximate accumulated amount of roughly ten percent based on the images, the model could be reduced to the determination of two parameters: the dissociation rate (of the transient interactions) and the cell size. Additionally, for most cells, only a lower limit on the dissociation rate could be determined (for all cells, the dissociation rates were consistent with s<sup>−1</sup>), implying that the recoveries were limited only by the cell size (diffusion-limited recovery). The average cell size determined solely from the model fits was 28 µm, which is comparable to the observed cell sizes and demonstrates the overall self-consistency of our model with the observed properties of the cells. The roughly factor-of-two differences in recovery speeds from cell to cell can easily be attributed to the observed differences in cell size, though they could also arise from slight differences in the amount of accumulation at the cell- cell contact region. Interestingly, the required rapid turnover of transient interactions at these cell-cell contact sites (\<10 seconds), and the fact that limits on the immobile fraction are compatible with the actual immobile fractions observed in intracellular regions, suggest that the affinity of PTP1B for its substrate(s) at regions of cell-cell contact is likely similar to that at other locations in the cell. The slower recovery at the cell-cell contact region can therefore be completely accounted for by the geometry of the bleaching and, in particular, the amount of accumulated PTP1B-D/A at the cell- cell contact region. The observed accumulation of PTP1B D/A is consistent with a higher concentration of PTP1B substrates at these regions (or generally on the PM) as compared with the distributed substrates found throughout the cellular interior, but is specifically not due to tighter binding of PTP1B to substrates found at regions of cell-cell contact. ## PTP1B Recruitment to Regions of Cell-cell Contact We next analyzed the dynamics of PTP localization at regions of cell-cell contact in response to stimulation of the EGFR, an established PTP1B substrate. Cos-7 cells were co-transfected with PTP1B D/A-dHcRed and EGFR-GFP, and the localization of the two proteins was monitored by confocal time-lapse microscopy. Although the EGFR was expressed uniformly on the cell surface, we observed accumulation of mutant PTP1B at regions of cell-cell contacts (arrowheads) even in the absence of EGF treatment. Stimulation with EGF led to rapid recruitment of additional dHcRed-labeled PTP1B D/A to these regions (arrowheads, 2, 10 min). Importantly, the rapid EGF-induced PTP1B D/A recruitment to regions of cell-cell contact was not paralleled by increased localization of the EGFR in these regions, as ratiometric analyses showed increased PTP1B/EGFR fluorescence ratios at regions of cell-cell contact in response to EGF stimulation (white regions in PTP1B/EGFR ratio images). The increased PTP1B/EGFR ratio at the PM was observed only at regions of cell-cell contact, even though EGFR concentration, and presumably the amount of tyrosyl phosphorylated EGFR (also see below), was the same at both locations. Quantification of PTP1B D/A localization at cell-cell contacts (contacts), compared with PM regions not in contact with neighboring cells (no contacts) in multiple cells, confirmed that the ratio of PTP1B/EGFR fluorescence intensities was significantly higher at regions of cell contact compared to other areas of the cell membrane; this difference between the data sets “contacts” and “no contacts” was highly significant (p\<0.001; Kolmogorov-Smirnov test). To determine whether ER localization of PTP1B facilitates its ability to access substrates at regions of cell-cell contact, we analyzed the localization of PTP1B D/A lacking its ER-targeting domain (PTP1B D/A CD-mCitrine) in MDCK cells stimulated with EGF for 5 minutes. Unlike full length PTP1B D/A, PTP1B D/A CD- mCitrine was found in the cytosol and could interact with substrates all over the PM, with no apparent preference for substrates at regions of cell-cell contact. This result provides further evidence that the density of PTP1B substrates is similar at regions of cell-cell contact and the rest of the PM. We then co-expressed PTP1B D/A CD-mCitrine with ER-bound PTP1B D/A mCherry, and stimulated cells with EGF for 5 minutes. Consistent with our previous observations, the ER-bound form of PTP1B D/A preferentially accessed the PM at regions of cell-cell contact. On the other hand, the ER domain mutant PTP1B (free), which is localized throughout the cell and at the cell periphery, was actually displaced from regions of cell-cell contact. We interpret these findings as indicating that the high local density of ER-anchored PTP1B (which is confined to a two-dimensional surface) can outcompete the freely diffusible (soluble) catalytic domain-only mutant in regions where the ER and PM are in close contact. Collectively, these data demonstrate that ER-bound PTP1B can only reach its PM-localized substrates at regions of close ER-PM contact, and indicate that the ER anchor plays an important role in restricting PTP1B interactions with PM substrates mainly to regions of cell-cell contact. ## The ER is Polarized Towards Regions of Cell-cell Contact The above findings indicate that the ER is specifically organized and oriented towards regions of cell-cell contact, which raised the possibility that PTP1B might play a role in this process. Microtubules contribute to the formation and stabilization of the ER network. To destabilize the ER network, we disrupted microtubules using nocodazole, and ER retraction at the periphery and regions of cell-cell contact was monitored using total internal reflection microscopy (TIRF). Cos-7 cells were co-transfected with PTP1B WT-mCitrine and RFP-tagged TK-Ras (a general PM marker to identify areas of cell-cell contact and account for cell shape changes). Consistent with previous reports, we observed ER collapse, specifically the loss of peripheral tubular ER through its conversion into sheet-like structures that retracted from the cell periphery, after nocodazole treatment. The extent of ER reorganization varied between experiments depending on cell confluence, cell type and shape. Nevertheless, the ER always retracted from peripheral regions but seemed to persist at regions of cell-cell contact. In isolated cells expressing PTP1B D/A, the ER retracted only partially after nocodazole treatment, with long stretches of tubular ER still attached to the PM at sites of focal adhesion to the glass coverslip. These observations raised the possibility that, rather than (or in addition to) the ER directing PTP1B towards substrates at points of cell-cell contact, PTP1B interactions with substrates might help polarize the ER towards these regions. We reasoned that if this was the case, then outcompeting endogenous PTP1B for substrate interactions by overexpressing PTP1B D/A CD should allow the ER to retract from regions of cell-cell contact after nocodazole treatment. PTP1B WT localization is not altered by PTP1B D/A co-expression. PTP1B D/A CD-mCitrine, monomeric Teal fluorescent protein (mTFP)-tagged Calreticulin (ER marker), RFP- tagged TK-Ras and EGFR were co-transfected into MDCK and Cos-7 cells. After starvation, cells were stimulated with EGF (100 ng/ml) for 5 minutes (to increase phosphorylation and recruitment of PTP1B D/A CD to its substrates), and then cells were treated with nocodazole. The ER retracted from peripheral regions as observed before, but failed to retract from regions of cell-cell contact. To assess whether the expression of PTP1B D/A CD was high enough to compete for substrates with the endogenous pool of PTP1B, we measured the absolute concentration of ectopically expressed PTP1B D/A CD across the cell by correlating the image intensity of cytoplasmic PTP1B D/A CD to its concentration, as determined by fluorescence correlation spectroscopy. A linear fit to this correlation enabled the estimation (by extrapolation) of PTP1B D/A CD concentrations at any point in the cell. The maximal concentration of PTP1B D/A CD was observed at regions of cell-cell contact and was approximately 9 µM. In the immediate cytoplasmic vicinity of these cell-cell contacts, the concentration of PTP1B D/A CD remained in the µM regime, which is comparable to the K<sub>D</sub> for PTP1B D/A-substrate interactions. We therefore conclude that over-expressed PTP1B D/A CD should efficiently compete with endogenous PTP1B for substrates at the PM of cell-cell contacts. However, endogenous PTP1B may have an additional advantage, as it is anchored to the 2D ER, which may give it a geometrical advantage in its search for substrates (higher k<sub>on</sub>) over the cytosolic PTP1B D/A CD (invalidating arguments based on the K<sub>D</sub>, which assumes a search in 3D). To even overcome the high-local concentration of ER-bound PTP1B, we performed competition experiments with PTP1B D/A lacking the ER-targeting domain but fused to the C-terminal membrane anchoring residues of K(B)Ras (TK-Ras) (hereafter referred to as PTP1B D/A-RFP- TK). We reasoned that anchoring PTP1B D/A to the PM should give it an enhanced geometric, and consequently kinetic, advantage in substrate interaction over the endogenous ER-bound enzyme. We then co-expressed PTP1B D/A-RFP-TK with TFP- tagged calreticulin and unlabeled EGFR, and treated Cos-7 cells with nocodazole as described before. Again, the ER did not retract from regions of cell-cell contact compared with other regions of the PM. Taken together, these findings clearly indicate that PTP1B-substrate interactions do not by themselves stabilize the polarization of the ER to regions of cell-cell contact. Rather, intrinsic polarization of the ER towards points of cell-cell contact likely directs PTP1B towards these regions. ## Regulation of Signaling at Regions of Cell-cell Contact by PTP1B The above findings indicated that there is an orientation of the ER towards regions of cell-cell contact, which raised the possibility that PTP1B may play a significant role in the regulation of signaling at these cites. To test this hypothesis, we determined the effects of PTP1B inhibition on the phosphorylation of substrates found at regions of cell-cell contact. Cos-7 cells were transfected with mCherry-tagged EphA2, a RTK that is closely related to EphA3 (a recently described substrate of PTP1B). Tyrosine phosphorylation was monitored using the co-transfected YFP fused to SH2 domain of pp60<sup>src</sup> (dSH2-YFP). Cells were treated with an allosteric PTP1B inhibitor (539741), which exhibits higher selectivity compared with active site inhibitors, and then monitored by confocal time-lapse microscopy for 60 minutes. Before inhibitor treatment, EphA2 showed typical membrane distribution of a RTK, with occasional accumulation at regions of cell-cell contact (, panel ii). In addition, dSH2-YFP showed expression in the nucleus, punctuate clusters surrounding the cell typical of focal adhesions (, panel iii), and in some cases at regions of cell- cell contact (panel iii yellow arrowheads). Minimal spatial overlap was observed between dSH2 and EphA2 prior to inhibitor treatment, even at regions of cell- cell contact ( panel iv cyan arrowheads). On the other hand, treatment with PTP1B inhibitor for 1 hr led to extensive increase in spatial overlap between dSH2 and EphA2 especially at regions of cell-cell contact (panel vii). The increased tyrosine phosphorylation of EphA2 upon PTP1B inhibition suggests that EphA2 is a potential PTP1B substrate at these sites. These findings reveal that the ER-bound PTP1B, by virtue of the organization and polarization of ER towards points of cell-cell contact, can access its potential substrates and regulate their activity at these regions. # Discussion One of the earliest discoveries about PTP1B was its localization to the cytosolic surface of the ER. This result was initially quite surprising, because biochemical and genetic studies soon established that PTP1B dephosphorylates several RTKs, including the PDGF, EGF and insulin receptors. We subsequently provided an answer to this topological dilemma with the finding that RTKs encounter PTP1B only after they are endocytosed, a finding confirmed and extended by several other groups. Nevertheless, other work indicates that PTP1B can access PM substrates at points of cell-cell contact. These results raise an apparent paradox: why can PTP1B dephosphorylate some substrates only after endocytosis, whereas others can be targeted while at the PM? The results herein provide a solution to this paradox and present a novel mechanism of how ER-PM signaling at regions of cell-cell contact is regulated. This regulation is important in that it is observed in different cell types, including those that do and do not form tight junctions. Our findings clearly demonstrate that interactions between the PM and the ER are not ubiquitous, but rather are restricted to regions of cell-cell contact. In these regions, the ER network appears to be specifically organized and oriented such that it is juxtaposed to the PM. An analogous mechanism might also account for PTP1B interactions with substrates at cell-matrix contacts. If we envision that the ER can reach the PM everywhere in a stochastic, dynamic way, then the local concentration of substrates should determine the amount of PTP1B D/A at regions of cell-cell contact. Conceivably, substrate concentrations at regions of cell-cell contact are higher than elsewhere in the PM, in which case a critical amount of substrate(s) must be present to enable the ER-bound phosphatase to access them while they are still localized to the PM. Our photoactivation and FRAP studies show that PTP1B D/A rapidly moves in and out of regions of cell-cell contact (replenished by the ER pool), with a residence time of \<10 s. These findings are consistent with the presence of a large number of PTP1B binding sites, presumably representing substrates in these regions. Nevertheless, when the EGFR is at the same concentration in the PM and regions of cell-cell contact, interaction (without prior endocytosis) occurs only at the latter location. Furthermore, expression of the soluble cytosolic domain of PTP1B D/A does not result in a specific enrichment of this protein to points of cell-cell contact after EGF stimulation; rather, this protein labels the PM uniformly, thereby excluding the possibility that substrate density alone dictates where PTP1B interacts with substrates at the PM. Varying rates of endocytosis (at cell-cell contacts versus other areas of the PM) also might account for the differential ability of PTP1B to access substrates at the PM. Arguing against this model, though, we previously blocked RTK endocytosis using dominant negative dynamin, but were unable to detect direct interaction between PTP1B and PM-bound RTKs using fluorescence lifetime imaging microscopy (FLIM). Taken together, these findings strongly suggest that additional factors are instrumental in enabling ER-bound PTP1B to access substrates specifically at regions of cell-cell contact. These factors are likely responsible for polarizing the ER towards regions of cell-cell contact, thereby creating specialized zones of ER-PM interaction at cell-cell contacts in which PTP1B (and conceivably, other ER proteins) may access the PM. There are precedents for the existence of specific ER-PM signaling compartments. Studies using S. cerevesiae indicate that membrane curvature may help to define functionally distinct sub-domains in the ER. Whether an analogous mechanism creates ER sub-domains near regions of cell-cell contact in mammalian cells remains to be determined. In skeletal muscle, the sarcoplasmic reticulum junctophilin proteins bind to PM components to form junctional contacts. A new paradigm for ER-PM signaling has been proposed for the ER calcium sensor; stromal interaction molecule 1 (STIM 1). After ER calcium depletion, STIM 1 oligomerizes, and due to enhanced avidity, binds to and activates store-operated calcium channels at the PM. Ultrastructural studies reveal that upon calcium depletion, STIM 1 induces the formation of cortical ER (cER), which is composed of large sheets that are closely opposed to the PM but still connected to conventional ER cisternae. The distance between cER and PM (on average 8.3 nm in Epon sections) is comparable to the distance between ER-bound PTP1B and the PM in our EM studies. Moreover, mammalian STIM proteins have a lysine-rich sequence that is similar to the yeast peripheral protein Ist2 (a transmembrane protein that is localized to cER in yeast). Remarkably, Ist2 enrichment at the cER of mammalian cells directly modulates the formation and maintenance of this ER subregion, and cER induction in vivo is dependent on microtubules and the coat protein complex I (COPI). It will be important to determine whether cER or some other higher order ER structure promotes PTP1B interactions with substrates at cell-cell contacts. The identity of these substrates and their contribution to cell-cell signaling requires further investigation. Recent studies identified interactions between PTP1B and EphA3 and revealed that the interaction can occur at the plasma membrane at areas of EphA3/ephrin-mediated cell-cell contact. In the current study, we report increased EphA2 tyrosine phosphorylation at regions of cell-cell contact upon PTP1B inhibition, suggesting that EphA2 is putative substrate of PTP1B at these sites. Other proteins known to reside at cell-cell contacts, such as p120 catenin and Zonula Occludens (ZO-1), are hyper- phosphorylated in PTP1B-null fibroblasts, and represent additional candidate PTP1B substrates in these regions. Collectively, our studies suggest that the ER plays a dynamic role in regulating signaling at regions of cell-cell contact via PTP1B and highlight ER-PM interactions as an emerging new paradigm in cellular signaling. # Materials and Methods ## Cell Culture, Antibodies and Reagents PTP1B-null fibroblasts (Haj et al. 2003), MDCK and Cos-7 cells (both obtained from ATCC) were cultured on 35 mm glass bottom culture dishes (MatTek Corporation) or 4-well Lab-Tek chambers (Nunc) in Dulbecco’s modified Eagle medium (DMEM) with 10% fetal calf serum (FCS). Unless otherwise indicated, cells were transfected using Lipofectamine 2000 (Invitrogen), as described by the manufacturer, and incubated at 37°C and 5% CO<sub>2</sub> for at least 12–18 hours before imaging. For live cell experiments, the culture medium was replaced with CO<sub>2</sub>-independent imaging medium. Antibodies against mouse β-catenin (#610153) were purchased from BD Biosciences, fluorescein-conjugated secondary antibodies were from Jackson Immunoresearch, and nocodazole was purchased from Sigma. PTP1B allosteric inhibitor 3-(3,5-Dibromo-4-hydroxy- benzoyl)-2-ethyl- benzofuran-6-sulfonicacid-(4-(thiazol-2-ylsulfamyl)-phenyl)-amide (Cat \#539741) was purchased from Calbiochem. ## Plasmid Constructs PM-anchored PTP1B D/A was generated by inserting cDNA encoding the catalytic domain of human PTP1B (residues 1–407) flanked by Age I restriction sites into pRFP-tkB-C1 vector (gift of Dr. M. Lackmann). pRFP-tkB-C1 (hereafter referred to as RFP-TK) consists of the C-terminus membrane targeting domain and part of linker domain of K(B)Ras downstream of mRFP. ## Immuno-electron Microscopy Cells were fixed with 4% paraformaldehyde in 0.2 M sodium phosphate buffer, pH 7.4 for 5 min, followed by 2% paraformaldehyde in 0.1M sodium phosphate buffer for 30 min, then rinsed with PBS and quenched with 50 mM glycine in PBS for 10 min. After adding 2 ml of 1% gelatin, the cells were collected by scraping the monolayer, centrifuged, resuspended in 10% gelatin, and then centrifuged again for 1 min. After cooling on ice for 30 min, the pellet was cut into small cubes, which were infiltrated with 2.3 M sucrose overnight, mounted on pins and frozen in liquid nitrogen. Cryosections were prepared using an ultra-cryomicrotome (Leica) and collected with a 1∶1 solution of 2.3 M sucrose and 2% methyl cellulose. Immunogold labeling with anti-PTP1B (FG6) antibodies (diluted 1/10–1/50) was performed and revealed with protein-A gold conjugate (Utrecht University). Cryosections were viewed at 100kV on a Biotwin EM scope (FEI) equipped with a SIS Keen View 1.3×1 K CCD camera. For cryo-electron microscopy of vitreous sections (CEMOVIS), samples were treated as previously described. Briefly, cells were grown on ACLAR film (Science Services, Eppelheim, Germany), fixed using HPM100 high pressure freezer (Leica Microsystems, Germany) and cut at –150°C. Samples were analyzed using a JEM-1400 electron microscope (JEOL Germany) at 120 kV equipped with F-416 CCD camera (TVIPS, Gauting, Germany). For vitreous sections, a cryo-holder model 626 (Gatan, Pleasanton, CA) was used. ## Confocal Imaging, Photoactivation and Time Lapse Microscopy Confocal imaging and photoactivation were performed using an Olympus FlowView FV 1000 microscope with a 63×/1.4 N.A oil objective. Experiments were performed in an environmental box in which live cells were maintained at 37°C and 5% CO<sub>2</sub>. Photoactivation was performed in the region of interest using a 413 nm laser in the region of interest, and continued imaging of the photoactivated pool was performed using a 488 nm laser. All studies were conducted on randomly growing cells. For confocal time-lapse microscopy, Cos-7 cells were transiently co-transfected with EGFR-GFP and PTP1B-dHcRed using Fugene 6 (Roche Biochemicals). Cells were serum-starved, stimulated with EGF (100 ng/ml) and then subjected to confocal time-lapse imaging. Images were captured using a confocal laser microscope (Leica TCS-SP5) with a 63×/1.4 N.A oil objective at 12-bit resolution. Image processing and quantitative analysis were performed using Image J. For ratiometric analysis, images were converted to 32-bit (floating point) format and thresholded. Background was defined as not a number (nan) and ratio images were then generated by dividing PTP1B D/A dHcRed by EGFR-GFP images. ## Fluorescence Recovery after Photobleaching Photobleaching was performed on the temperature-controlled stage of a Leica SP2 AOBS Sirius microscope equipped with a 63×/1.4 N.A oil immersion lens. GFP was excited using a 488 nm Argon laser, and fluorescence was monitored at 0.203 sec intervals. Fifty pre-bleach images were recorded with 4% laser power of the 488 line every 0.203 sec. Consequently, a region of interest (outlined in the figure) was photobleached using 100% laser power of the 456, 476, 488 and 496 lines. Recovery of fluorescence was monitored over the course of 300 whole-cell scans (with a scan interval of 0.203 sec). For bleaching of PTP1B WT-GFP and PTP1B D/A-GFP in the ER, a circular region of 4 µm diameter was used, whereas for bleaching at cell-cell contacts, we used a rectangular area of 5 µm length and 1 µm width. Cells with varying levels of transfection ranging from high to low PTP1B WT and D/A expression were studied. Mean fluorescence intensities in the FRAP region and for the whole cell (including the FRAP region) were recorded, and the background was subtracted. The FRAP region and whole cell profiles were then individually normalized to their pre-bleach values (obtained by averaging the 11 images immediately preceding bleaching), and the final FRAP recovery profile was obtained by dividing the normalized FRAP region intensity by the normalized whole cell intensity. This final step removed the global fluorescence decrease due to photobleaching during initial bleaching (∼5% for FRAP experiments in the ER and 5–20% for FRAP experiments in the cell-cell contacts), as well as any gradual bleaching occurring during acquisition of the recovery (less than a few percent for all experiments). Mathematical modeling was used to fit the recoveries (see Supplementary Materials). ## Total Internal Reflection Fluorescence Microscopy Cos-7 cells were transiently co-transfected with RFP-TK and PTP1B WT-mCitrine or PTP1B D/A-mCitrine using Effectene (Qiagen). Cells were plated onto glass-bottom 35 mm dishes (MatTek) and then were incubated in CO<sub>2</sub>-independent medium supplemented with 10% FCS and 2.5 mM L-glutamine. Images were acquired with an inverted Olympus CellˆR microscope configured for triple line total internal reflection fluorescence, using a 60× PLAPO/TIRFM-SP oil immersion objective, with 1.45 N.A at 16-bit resolution in each channel. ## Fluorescence Calibrated Confocal Time-lapse Microscopy Fluorescence-calibrated time lapse imaging was performed on a Zeiss LSM 510 Meta confocal microscope equipped with a ConfoCor 3 unit and a C-Apochromat 40×/1.2 N.A water immersion objective. Experiments were performed in an environmental box in which the sample and objectives were maintained at 37°C. mTFP and mCitrine were excited with 458 nm and 514 nm argon lasers, respectively, whereas RFP was excited with a 561 DPSS laser. The fluorescent light was passed through a NFT 565 beam splitter and detected with PMT detectors through a BP 475–525 band pass filter for mTFP, a LP 530 long pass filter for mCitrine, and a LP 575 long pass filter for RFP. To minimize cross-talk between individual channels, mTFP and RFP were recorded simultaneously, whereas mCitrine was imaged separately. Pinhole diameter was set to 192 µm for mTFP, 296 µm for RFP, and 1000 µm for mCitrine. Images were recorded at a resolution of 512×512 pixels (0.15 µm/pixel) and a bit depth of 12 bit. For time-lapse imaging, single-plane multi-color images were recorded at 1 minute intervals. To correlate the image intensity of mCitrine to absolute protein concentrations, we acquired single-channel images of cells expressing mCitrine-PTP1B D/A CD at exactly the same settings as for the time-lapse experiments. Fluorescence correlation spectroscopy (FCS) measurements were performed at randomly chosen positions in the cytoplasm of the same cells. Background-corrected average image intensities were determined in a 1 µm radius around the FCS measurement point. For FCS, mCitrine was excited with the 514 nm line of an argon laser, while an APD detector recorded the fluorescence through a LP 530 long pass filter and a 70 µm pinhole. The raw intensity data were correlated using the microscope manufacturer’s software, and for each measurement, correlation curves were averaged for ten consecutive 10-second intensity recordings. Averaged correlation curves were fit by using a model accounting for anomalous diffusion and a triplet term in order to determine the average number of particles in the confocal observation volume, using a non-linear least squares minimization algorithm implemented in Matlab (MathWorks). Confocal volume was determined from calibration measurements of an Alexa 546 dye solution, assuming a diffusion coefficient of 280 µm<sup>2</sup>/s at 25°C that was extrapolated to 308 µm<sup>2</sup>/s at 37°C. ## Mathematical Model for Fitting FRAP Data For the circular FRAP in the ER (PTP1B WT and PTP1B D/A), we used the analytic formula for a cylindrical region in an infinite medium :where denotes the FRAP recovery (recovering from 0 to 1), I<sub>0</sub>(x) and I<sub>1</sub>(x) are the standard modified Bessel functions, and is the diffusion timescale with r the radius of the FRAP region and D the diffusion constant. Considering the case of imperfect bleaching and the possibility of an incomplete recovery due to an immobile fraction leads to the following formula:where is the bleaching efficiency (with corresponding to complete bleaching) and is the immobile fraction. The initial FRAP bleaching in the region of interest in our cells removed less than a few percent of the total cellular fluorescence. This justifies the use of the “infinite cell” approximation implicit in Eq. 2. We obtained the following values for diffusion constants and immobile fractions for the four intracellular/ER FRAP data sets (taking µm for the radius of the FRAP region): D (µm2 s<sup>−1</sup>) n PTP1B WT-GFP 11 PTP1B D/A-GFP 13 TCPTP-GFP 12 SHP2-GFP 9 Each FRAP profile was fit independently, and the above values quoted are the mean and standard error of mean (S.E.M) for the indicated number of cells, n (individual cell profiles shown). The mean and S.E.M of the diffusion constant were based on the statistics of the logarithm of The above parameters also fit well the mean cell profile obtained by simply averaging all of the single cell recovery profiles (see in the main text). PTP1B D/A, therefore, has a three-fold lower diffusion constant and a two-fold greater immobile fraction than PTP1B WT, consistent with the “trapping” nature of its catalytic region mutation. The diffusion of PTP1B WT is similar to that of another ER-bound phosphatase, TCPTP; however, the immobile fraction of TCPTP is insignificant (consistent with the few percent removal of fluorescence from the initial FRAP bleaching). All of the ER-bound proteins diffuse much more slowly than the cytosolic SHP2 (which also has an insignificant immobile fraction). For the FRAP experiments at the cell-cell interface, we employed a Laplace transform approach similar to that used previously. In the latter, a 2D cylindrically-symmetric geometry was explored. Here, we have a 1D system with a non-trivial boundary condition (due to the location of the binding sites there). Because this important case was not examined before and, in particular, because the inclusion of boundary conditions is subtle, we provide a full derivation for the sake of clarity and completeness. The following equations describe the concentration of a freely diffusing fluorescent protein, free binding sites, and binding sites bound with the fluorescent protein, Because the binding sites are localized to the cell-cell interface, and, similarly, where is area of the cell-cell interface (at) and is the cell volume. This gives: At steady state, the concentration of free binding sites, is unaffected by the FRAP process. We therefore need consider only the time dependence of and Following bleaching, the free and bound fluorescent protein concentrations split into bright and dark fractions: Upon substituting back into the differential equations above, these equations can be dissected into separate sets of equations (of the same form) for both the bright and dark fractions. Of course, we are only able to monitor the bright fraction: These equations are best solved by the Laplace transform, which gives:where the bar indicates the Laplace-transformed function dependent on the Laplace transform variable p. Rearranging, The initial concentration of the free protein is just We assume an initial bleaching of the bound protein, giving where is the bleaching efficiency, with denoting complete bleaching. Substituting into the differential equation for now gives: For with general solution: The boundary condition at (using Gauss’s law) isand at is simply 0 (hard wall, Neumann boundary condition). Using the boundary conditions to solve for the coefficients of the general solution yields: These give the following for the specific solution (at): This can then be substituted into the equation for yielding Normalizing to the pre-bleach amount of bound protein and rearranging gives the following dimensionless form for the Laplace transform of the recovery of fluorescence at the cell-cell interface:where is the diffusion timescale and is the contrast, which is equal to the ratio of the total protein bound at the interface to the total free protein in the cell. The full recovery is then: with From the fitting, we find that the “tanh” term is effectively 1 in our case, implying:with the recovery effectively depending on only two unknown parameters: and (the contrast and diffusion timescale are degenerate). But this is just the recovery one gets assuming an “infinite” cell. Here, the general solution is just: The boundary condition at (using Gauss’s law) is giving: And, again,now with the exact equality: The recovery (normalized only to the pre-bleach fluorescence in the FRAP region) asymptotically approaches the following value:which for complete initial bleaching of the bound protein is just or the ratio of free protein to total protein. However, for our data analysis, we additionally normalized to the total cellular fluorescence at each time point:which removes the reduction in intensity caused by the bleaching of the total cellular fluorescence. The bleaching efficiency is directly observed, and the diffusion constant can be obtained by FRAP in the interior of the cell (hence, our FRAP experiments with PTP1B D/A in the ER, which yielded µm2 s−1). Assuming that the confocal plane we observe is representative, we find that for most cells (this value, true for the entire stripe, can also be used for our experiments wherein only one third of the interface was bleached). This implies that the additional normalization in is only a few percent. It is therefore safe to ignore this (the “ is small” limit where) and simply fit to Assuming zero immobile fraction (see below), we have independently fit the recovery curve for each of the 12 cells, obtaining mean and S.E.M values (based on the logarithmic values) for the two remaining free parameters s<sup>−1</sup> (residence half-life of s) and s<sup>−1</sup>. For most of the recovery curves, needed only to be greater than some minimal value in order to fit the data, hence the lower limit. The mean values again fit well the mean profile resulting from averaging the recovery profiles of all of the cells, as shown in. Using s<sup>−1</sup>, µm<sup>2</sup> s<sup>−1</sup>, and gives an effective cell length of µm, which is comparable to the observed cell sizes. The FRAP recovery of PTP1B D/A at the cell-cell interface is therefore consistent with rapid turnover ( s), the ER-determined diffusion constant, the observed level of PTP1B D/A accumulation at the interface, and the cell size. From the raw images themselves, it is already clear that the recovery we observe is not an artifact, but represents a real turnover of bound protein at the cell-cell interface. As mentioned above, an immobile fraction at the cell-cell interface is not required in order to fit the recovery (in all of our displayed fits, we assume no immobile fraction). Recoveries assuming an immobile fraction also fit the data (not shown), and on a cell-by-cell basis these fits provide a useful upper limit to the immobile fraction at the cell-cell interface ranging from \<30% to \<70%. Theoretically, a longer acquisition time would help eliminate most of the parameter degeneracy and place a tighter limit on the immobile fraction, but observation of the actual long-term recovery is severely complicated by cell movement and morphological changes occurring on minute timescales. All of the analysis was performed in Mathematica (Mathematica, Inc.), where we have used the inverse Laplace transform algorithms described in: <http://library.wolfram.com/infocenter/MathSource/4738/and> <http://library.wolfram.com/infocenter/MathSource/5026/>. # Supporting Information We thank Drs. I. Yudushkin, M. Bentires, R. Pepperkok, and B. Geiger for PTP1B-PhAc, PTP1B-GFP, SREP/Sec61 and dSH2YFP cDNA constructs, respectively. [^1]: Conceived and designed the experiments: FH BN PB. Performed the experiments: FH OS AK SWK VR HMH MG. Analyzed the data: FH OS AK VR MG CA BN PB. Contributed reagents/materials/analysis tools: MB CA. Wrote the paper: FH OS AK BN PB. [^2]: The authors have declared that no competing interests exist.
# Introduction Several landmark studies have indicated that our behaviour can be biased or indeed driven by subliminal information. While the extent to which subliminal information affects behaviour is still a contentious topic (see for a review), there is a large body of evidence supporting the influence(s) of subliminal information on behaviour. Examples of such influences are (here named as originally described on the respective articles): implicit attitudes or biases, subliminal priming, unconscious evidence accumulation in decision making, unconscious learning, perceptual adaptation from invisible stimuli, and voluntary actions and choices triggered by non-conscious brain signals. Subliminal information thus can have a wide range of effects on our behaviour contributing to negative behaviour, in the case of unconscious biases to language and motor skills learning in the case of implicit and statistical learning and direct actions in the case of blindsight and agnosia. Research suggests that subliminal information might be processed at variable depths in several brain areas. ‘Invisible objects’ can still be processed through the visual system, especially in dorsal areas—see however for a critical review. Suppressed emotional facial expressions are processed by subcortical structures, in particular the amygdala (for a review see, while masked words are semantically processed in the temporal gyrus in the fusiform word area. Subliminal primes have been shown to facilitate or bias behaviour in several cognitive tasks. Subliminal information can accelerate reaction times for semantically related targets, improve the accuracy of motion discrimination and speed up reaction times at detecting spoken words after the presentation of subliminal speech, among others. Crucially, research thus far has focused on showing that decisions can lean towards subliminal primes. This begs the question of whether choices following subliminal priming are rigidly biased towards the subliminal information or could they be trained away, by reversing the bias from the prime. In other words, do we inevitably choose, on average, the same stimuli (or stimuli that share properties with the prime) that were presented subliminally? Another possibility is that we could learn to choose away from subliminal biases, which would mean that such biases and choices are separable. If choices invariably lean towards such subliminal primes, this will imply that the subliminal prime triggers a rigid cascade of steps towards a stereotyped behaviour, which determines choosing the primed stimuli. On the other hand, if behaviour can diverge away from subliminal biases, this would suggest that the intermediate steps from subliminal perception to behaviour are subject to influence (such as learning) and not always rigidly biased in the same manner. The present exploratory study aims to test whether decisions towards a subliminal prime can diverge from it after going through an implicit learning protocol. Specifically, we tested two hypotheses: (1) subliminal primes can bias imagery-content decisions, and (2) participants can be trained to reverse the bias from subliminal primes. We developed a paradigm in which primes were suppressed from awareness using Continuous Flash Suppression (CFS,). After the presentation of the subliminal prime, participants had to choose one of two gratings and imagine it as vividly as possible. In a subsequent experiment, we gave feedback to a small cohort of participants about their decisions to test whether we could reverse the priming away from the subliminal primes. # Materials and methods ## Participants Experimental procedures were approved by the University of New South Wales Human Research Ethics Committee (HREC#: HC17031). All methods in this study were performed in accordance with the guidelines and regulations from the Australian National Statement on Ethical Conduct in Human Research (<https://www.nhmrc.gov.au/guidelines-publications/e72>). All participants gave written informed consent to participate in the experiments. Participants were recruited during 2017. We tested a total of 29 participants (18 females, ages 18–39) with normal or corrected-to-normal sight for the main experiment. An additional 26 participants (22 retained) were tested in the no-imagery control experiment. Participants’ data were deidentified by assigning them a number so authors could not identify participants after data collection. ## Exclusion criteria We employed 3 independent selection criteria with the aim of monitor that participants did not consciously perceive the primes. First, we used realistic catch detection trials in which very faint gratings were presented to both eyes to measure participants ability (and willingness) to report subtle suppression breaks that were designed to be as similar as possible to real ones. In catch trials, the luminance of each one of the monocular gratings was lowered to 25% of the intensity of the standard monocular grating. The grating presented to the dominant eye was alpha blended with the CFS Mondrian masks and the other was presented on a black background. This design made catch trials perceptually similar to normal suppression breaks and thus hard to detect in some trials due to monocular suppression. This was done to verify that participants would report even the faintest breakthroughs of the primes. We rejected participants who failed to report at least 80% of catch trials (N = 8). Secondly, we discarded participants who did not report any broken suppression trials (N = 5). While we cannot be sure that some of these participants truly did not experience any suppression breaks, we imposed this criterion in the hope of making sure that participants were not refraining from reporting suppression breaks. Finally, we rejected participants who showed differences in priming between red and green higher than 40% (N = 1). This was done to ensure participants were not biased towards any choice (i.e., choosing red on most trials). Thus, we retained 17 participants who satisfied all these criteria in the main experiment for further analyses. For the no-imagery control experiment, we tested 26 participants. We rejected 4 participants (0 based on criterion 1, 2 based on criterion 2 and 2 others on criterion 3). Despite the use of these criteria, we cannot definitely rule out that participants were able to perceive the primes. While this shortcoming is not unique of our study, there are other controls currently regarded as better suited to detect suppression breaks which were not used in this study (see for a discussion on the subject). ## Power and effect size analyses We performed post-hoc power and effect size analyses using G\Power 3.1 on the main results of the study. For the subliminal priming experiment, we tested the priming of 17 participants (M = 0.6277, SD = 0.1253) against chance (0.5). The achieved effect size was d = 1.019 and power (1-β) = 0.98, one sample, two- tailed t-test against chance, α error probability = 0.05. For the priming differences post and pre-training, we tested a subset of participants that had significant priming on the main experiment (n = 7). While we acknowledge that this is a rather small sample size, the effect size and power analyses indicate that our results are robust, nonetheless. For the difference between pre- training and training, the effect size was dz = 1.788 (M = 0.298, SD = 0.167) and the power (1-β) = 0.973, two-sample, two-tailed t-test, α error probability = 0.05. Finally, for the difference between pre and post training, the effect size was dz = 1.118 (M = 0.2729, SD = 0.2442) and power (1-β) = 0.694, two- sample, two-tailed t-test, α error probability = 0.05. ## Subliminal prime experiment The tasks were carried out on a Windows 7 PC running Psychophysics Toolbox 3 in MATLAB (The MathWorks, Inc., Natick, Massachusetts, USA) on a 85Hz Dell Trinitron P1130 CRT monitor 1280x1024 resolution. We presented participants with gratings (red/green, horizontal/vertical, \~7 deg of visual angle, 0.33 monochromatic contrast at its maximum) suppressed using continuous flash suppression (CFS, \~10 deg of visual angle, flickering at 10Hz, square wave). We chose to use coloured gratings instead of grayscale as colour has been shown to promote subliminal priming, thus our stimuli had two dimensions that could generate subliminal priming: orientation (horizontal/vertical) and colour (red/green). We started testing perceptual isoluminance of the green and red gratings with the heterochromatic flicker test in which participants were presented with flickering red and green squares (temporal frequency 10Hz) while adjusting the luminance of green against red (constant at 0.33 contrast) until the perceived flickering was minimal. Perceptual isoluminance for green gratings was systematically lower than for red gratings. Subsequently, we tested participants’ eye dominance with the Miles eye dominance test and presented gratings to the non-dominant eye and the dynamic Mondrian patterns (CFS) to the dominant eye to maximize suppression. We used a stereoscope adjusted for each observer until the overlap of each monocular image was achieved while their heads were stabilized using a chinrest. We added a checkerboard pattern square frame around the stimuli (\~12 deg of visual angle, \~2 deg width) to aid vergence and thus overlapping of the monocular images. We presented grayscale Mondrian patterns (instead of coloured ones as normally used) and added random noise on them (60% amplitude, zero-centred), which increased the suppression of gratings. The CFS/grating presentation lasted for 6s. To avoid suppression breaks due to abrupt onsets, we linearly increased the opacity of the grating from 0 to a max of 33% (depending on the heterochromatic flicker test) for 2.5s. Importantly, at any moment of the prime presentation as well as in the decision waiting period (see below), participants could abort the trial whenever a prime broke suppression. We used empty trials (40%), where no gratings were presented as baseline for vividness and agency ratings. We also presented catch trials (20%) in which gratings were shown to both eyes to control for decision biases on reporting suppression breaks. We used realistic catch detection trials in which very faint gratings were presented to both eyes. The intensity of each monocularly presented grating was 25% of the gratings presented in a standard trial. This design was used to replicate real suppression breaks, because break- through gratings can be difficult to detect. Thus, we expected that catch trials could be missed in some trials. A total of 150 trials per participant were tested divided in 5 blocks of 30 trials each, with trials from different conditions pseudorandomized within blocks. After the prime presentation, participants had to decide which grating to imagine. During this waiting period, we presented grayscale plaids matching the position of the primes to mask any after images of the prime that could influence the decision. After participants decided which grating to imagine, indicated by pressing space bar (independently of which one of the gratings was chosen), they imagined it as vividly as possible for 5 seconds. After the imagery period, participants communicated their choice by pressing different keys on a computer keyboard. After this, participants were required to rate the strength of their visualizations or vividness on a scale from 1 to 4. Finally, they rated on the same scale the sense of agency felt while deciding which grating to imagine. ## Prime and decision dissociation training experiment We implicitly trained participants to switch their decisions away from the subliminal prime. For this experiment, we selected a subset of the retained participants from the main experiment. We selected participants with priming significantly above chance across the mean priming values within each of the 8 experimental blocks (one sample, one-tailed t-test against 50% p\<0.05). See the section for the rationale behind this choice. We thus selected 7 participants who showed significant above chance priming and trained them in 3 sessions performed on consecutive days plus a post-training session to test the persistence of the training effect a week after the training. Importantly, the selection criterion was defined, and participants were selected before the training session, so no participants were discarded after the training, nor we stopped testing participants after obtaining significant results. During the training sessions, performed weeks after the main experiment, participants were presented with the same paradigm as in the main experiment except for the following changes. After participants reported the chosen grating, written and audio feedback was given. For trials where participants chose the same grating as the prime (congruent trials), we presented the word “Wrong!” and played a beep sound. Inversely, in trials where participants did not choose the priming grating (incongruent trials), we presented the word “Right!” and played a bell sound. In catch and empty trials, decisions were labelled as right no matter the grating chosen. We asked participants to try to get as many right trials as possible. The instructions given to the participants were: “you will be presented with the same paradigm as in the previous experiment you performed weeks ago. Each trial will start with flashing stimuli followed by a decision imagery task as previously. As in the previous experiment, if you happen to detect a grating or red/green colour during the flashing stimuli, you must report it. This time, however, at the end of the trial there will be feedback: ‘right’ and ‘wrong’. Try to get as many ‘right’ feedback even if you don’t know how to do it". In post-experiment interviews, participants reported not knowing how to perform the task, however, priming ratios revealed that they could implicitly learn to choose the right grating (incongruent, Figs). ## No-imagery decision control We devised a control experiment to test to which degree the task requirement of producing a mental image after the decision influenced the subliminal priming. We thus presented to an independent set of participants the same task as described in the subliminal prime experiment, except for that there was no imagery period nor vividness rating. ## Priming analysis We calculated priming by dividing the number of trials where the choice was congruent with the prime by the total number of trials that were not aborted. These ratios were subsequently multiplied by 100 to denote percentages. Statistical analyses were thus focused on testing that the amount of priming was either different from chance level, i.e., 50% or comparing priming levels between different conditions/experiments. # Results ## Paradigm The paradigm consisted of an imagery-content decision task following the presentation of a subliminal prime suppressed using Continuous Flash Suppression: CFS. Details of the paradigm and stimuli generation are available in the Materials and Methods section as well as in ‘s caption. Importantly, participants were instructed to abort trials at any time when the gratings (primes) became visible by pressing a key. We included control trials to test for response bias (catch trials) and as a baseline for vividness and agency ratings (empty trials). For a schematic of what each trial type looked like, see S1 Fig in. ## Subliminal gratings prime decisions about the contents of future imagery We first tested whether the presentation of subliminal gratings biased participants’ choices towards subliminal stimuli. Our results indeed showed significant priming by the subliminal gratings on the subsequent imagery decision (mean 62.8% ±3.04 SEM, grey bar, one-sample t-test against 50%, C.I. = \[56.33, 69.21%\], t(16) = 4.204, p = 6.72 · 10<sup>−4</sup>, solid dots represent participants included in the training phase). We devised a control experiment to test to what degree the decision priming depended on imagery. The no-imagery control experiment was identical to the main experiment with the exception that there was no imagery period, nor vividness rating question. Average priming in the control condition was again significantly above chance (mean 57.65% ±2.07 SEM, white bar, one-sample t-test against 50%, C.I. = \[53.35, 61.95%\], t(21) = 3.6986, p = 0.0013). While the control priming was slightly lower than in the main condition (5.15% difference) this difference was not significant (p = 0.079, two-sample, one-tailed t-test). This result suggests that the decision priming is essentially independent of the subsequent imagery. We then compared the vividness ratings between primed and empty trials. We used empty trials (where no prime was presented) to measure the baseline vividness levels. This analysis revealed that, in primed trials, when the prime and the choice were congruent, vividness (the subjective imagery strength) was higher (white vs black bars, paired two-tailed t-test, C.I. = \[0.056, 0.474\], t(16) = 2.68, p = 0.0163), than when no prime was presented (empty trial). The vividness on incongruent decisions, while lower on average when compared to congruent decisions, was also significantly higher than the empty condition (, white vs grey bars, paired t-test, C.I. = \[0.0035, 0.3545\], t(16) = 2.52, p = 0.0227). No significant differences were found between vividness ratings for congruent and incongruent decisions. This suggests that, in this particular paradigm, showing subliminal primes boosted vividness regardless of the congruency with the imagined item. Analyses of agency ratings did not show significant differences among the different conditions of the main experiment: empty (baseline), primed and non- primed trials (S2 Fig). The lack of a difference in ratings of agency suggests that participants were largely unaware of the influence of the subliminal primes on their decisions, or at least they were not aware of any change in their agency. We used 3 exclusion criteria to strengthen the reliability of reports on the subliminal priming task: catch detection, suppression breaks and priming differences. Participants not satisfying any of these 3 criteria were excluded from further consideration ( open dots and in grey). First, we used realistic catch detection trials in which very faint gratings were presented to both eyes (i.e., not easily detectable, see for details) to measure participant’s ability (and willingness) to report subtle suppression breaks. As catch trials were designed to replicate real suppression breaks, they were expected not to be reported in some trials. We, therefore, excluded participants that failed to report at least 80% of the catch trials (N = 8). Mean catch detection ratios for excluded and included participants were 0.673 (±0.079 SEM) and 0.959 (±0.016 SEM) respectively and were significantly different (p = 4.32·10<sup>−4</sup>, C.I. = \[-0.4321–0.1389\], t(28) = -3.9896, two sample t-test). We also excluded participants who did not report any suppression breaks (N = 5). Mean suppression break ratios for excluded and included participants 0.074 (±0.034 SEM) and 0.042 (±0.008 SEM), respectively, but were not significantly different (p = 0.305). Lastly, we excluded participants (N = 1) that had differences in priming between red and green gratings larger than 40% (for example 80% red) as this imbalance could indicate that participants’ choices would be biased towards one particular option (e.g., choosing red). Differences in priming between excluded and included participants were 0.047 (±0.038 SEM) and 0.035 (±0.007 SEM), respectively, but these were not significantly different (p = 0.74). Mean priming ratios for excluded and included participants were 0.591 (±0.035 SEM) and 0.628 (±0.023 SEM) respectively but were not significantly different (p = 0.432), indicating that priming was not explained by differences in the exclusion variables. We performed additional tests to ascertain if the inclusion criteria for suppression break ratios could explain the priming results. We thus analysed whether priming scores were correlated with suppression breaks. Our analyses did not detect significant correlations (p\>0.1), although there was a slight positive trend, between suppression break ratios and priming (S3 Fig), thus suggesting that partial failure at reporting suppression breaks is not enough to explain priming results. ## Implicit training makes decisions diverge from subliminal primes To test whether decisions are bound to subliminal primes, we trained a subset of participants who had significant above chance priming across blocks (N = 7, see for details in the selection criteria details) to choose the grating that was not primed. We used the same paradigm as in the main experiment, except that we included a feedback section after they reported their decision. For primed trials, where the decision matched the subliminal grating, we presented the prompt “Wrong!” accompanied with a buzz sound. Inversely, when participants chose the grating that was not presented as a prime, we presented the prompt “Right!” accompanied with a bell sound. On empty and catch trials, the “Right!” prompt was shown independent of the answer. Systematic post-experimental verbal interviews after each session revealed that 100% of participants (7 out of 7) responded “no” to the question “Were you aware of how to perform the task?”. Despite this, participants learnt quickly how to make their choices diverge from the subliminal primes. We trained participants in 3 sessions of 5 blocks each. Pre-training priming on the participants’ subset was 69.89% ±4.66 SEM (one sample t-test against 50%, C.I. = \[58.48, 81.3%\], t(6) = 4.265, p = 0.0053). Training priming levels were significantly different from pre-training (paired t-test, p = 0.0104, 0.0082, 0.0082 for sessions 1 to 3, FDR corrected for multi- comparisons, q = 0.05), but only significantly lower than chance level (i.e., 50%) in session 2 (one sample t-test, p = 0.0901, 0.0473, 0.0595, FDR corrected, q = 0.05). On average, implicit training resulted in participants switching their decisions towards the incongruent grating significantly below chance level (average training priming = 40.11% ±4.75 SEM, one-sample t-test against 50%, C.I. = \[31.92, 48.3%\], t(6) = -2.95, p = 0.0255). ## Implicit training effects persist over time Finally, we investigated whether the effect of training was robust over time. We thus tested the same trained participants one week later. We used the same paradigm as in the main experiment, that is, without* any feedback. During the training phase, all participants switched their decisions towards the grating that was not shown as prime, in other words, they chose to make incongruent decisions. After one week, all but one participant (6 out of 7, 85.7%) kept their decision responses incongruent with the subliminal prime. Post-training decision priming (42.6% ±10.84 SEM) was significantly lower than pre-training levels (paired t-test, C.I. = \[-49.87, -4.71%\], t(6) = -2.96, p = 0.025), while training and post-training decision priming levels were not significantly different (paired t-test, C.I. = \[-21.23, 26.21%\], t(6) = 0.257, p = 0.806). Interestingly, agency ratings fell modestly but significantly from pre-training to post-training levels (S4 Fig in, C.I. = \[0.0335, Inf\], t(6) = 2.053, p = 0.043). This suggests that participants may have become more aware of the influence of subliminal information on their choices, even if they reported not being aware of the subliminal primes or the purpose of the task. On the other hand, we found no significant changes in vividness across the pre-training, training, and post-training experiments (S5 Fig). Participants’ catch trial and break suppression rates were not significantly different along training periods (S6 Fig), ruling out major changes in strategies acquired during the training. These results suggest that the associations made between the subliminal primes and decisions during the training can persist over time, even in the absence of reinforcement. # Discussion We developed a paradigm to test whether subliminal primes can bias decisions about the contents of future imagery. Our results indicate that participants’ choices were biased towards the subliminal prime. It is important to highlight that there is still debate surrounding the nature of non-conscious perception (for a review see), and there have been a number of studies questioning the use of CFS to study non-conscious perception based on methodological grounds. We used controls to minimize concerns about participants performing the task incorrectly. First, we used realistic looking catch trials where very faint (25% of the subliminal prime contrast) gratings presented to both eyes. We excluded participants failing to report at least 80% of catch trials. Additionally, we also excluded participants failing to report any suppression breaks during normal trials. While some of these participants could have genuinely experienced no suppression breaks, we used this criterion to filter out potential participants that could be systematically not reporting catch trials. Further, we excluded participants that were disproportionally primed by one of the coloured gratings as this can be evidence of a bias on their choices (see for inclusion criteria details). Finally, we verified that, for participants passing the inclusion criteria, the rate of suppression breaks was not significantly correlated with priming. This analysis suggests that priming is unlikely to be a result of partial reports of suppression breaks (S3 Fig). To the best of our knowledge, these selection criteria have not used in combination on previous studies, and due to the stringent nature of these criteria, we ended discarding an important number of participants (12 out 29 or 41%). While this is not unique to our study (see for example) we understand the need to replicate our results to assess their full impact. As noted before, however, priming on the discarded participants was not significantly different from priming on the included participants, indicating that our results cannot be explained by our selection criteria. It is also important to emphasize that despite our efforts to ascertain that participants were unable to see the gratings, we cannot definitively rule out this possibility. As epistemological and methodological views on non-conscious perception continually evolve to address the shortcomings of previous studies, it is up to future research to replicate these results with the latest methods to date to dissociate non-conscious perception. In the follow up experiment, we tested a subset of participants (n = 7) and found that choices could diverge from subliminal prime after implicit training. We gave participants feedback on their choices, rewarding divergence from the primes while penalising congruent decisions. Despite participants systematically reporting being unaware of how to perform adequately at the task (which was evaluated by systematic post-experimental verbal interviews), they quickly learnt to diverge from the primes, on average switching their decisions from the first session. Participants implicitly learnt to choose the grating that was not primed, leading to a significant decrease in choice priming compared to pre- training levels. We acknowledge that, due to the selection criterion (within- participant priming significantly above chance), we ended up with a rather small sample size (N = 7). Calculating significant priming within participants has not been done before in the literature, due to the variability across blocks/trials (in this respect, our sample was comparable to previous studies) and also because there has been no need to draw conclusions at the participant level. Instead, previous studies have used average within participant priming and calculated significance across participants. We however reasoned that using within-participant significant priming as a selection criterion would allow us to highlight changes produced by the implicit training at the participant level while maximizing the sensitivity of our experimental procedure by avoiding floor effects (i.e., if a participant’s priming was only barely above chance level, it would have been hard to detect a change after training if the effect size was small). While in hindsight this selection criterion might have been too stringent (another option could have been selecting participants with average priming above 50%, no matter whether significant or not), we chose it because we could not anticipate the effect size, which ended up being rather large, with most participants reversing their choices (i.e., priming below 50%). Therefore, a larger sample study will be needed to reveal the full scope of these findings, despite that our analyses showed remarkable large effect sizes and satisfactory power values (see for details). Our results suggest that decisions are not determined to be congruent with subliminal representations but can diverge through learning. This capacity of learning can seem surprising given that non-conscious behaviour has been classically regarded as rigid and stereotyped. However, recent studies have shown rather complex learning in the absence of awareness such as fear conditioning, sequence regularity learning and instrumental conditioning, see however for evidence against unconscious instrumental conditioning. Our results can be seen as another case of instrumental learning in the absence of awareness, this time implicit learning to diverge from subliminal primes. Implicit training effects persisted after a week despite the lack of reinforcement. This is consistent with reports on the robustness and long- lasting effects of the non-conscious learning. This result suggests that the ability to dissociate subliminal information from choices can last long enough to change behaviour in a meaningful way. This can be of potential interest for the treatment of mental conditions in which non-conscious material affects behaviour, such as in depression. Participants’ agency felt modestly but significantly when comparing pre- training with post-training (but not when comparing pre-training vs training nor training vs post-training), which could suggest that participants may have become more aware about the influence of subliminal information on their choices. Participants, however, consistently reported not being aware of the subliminal primes nor the purpose of the task. In addition, participants’ catch- trials detection remained high and didn’t significantly change across the training stages, indicating that participants were reporting suppression breaks appropriately. Further, the instruction given to the participants did not hint at the relationship to-be-learned (i.e., choose the grating opposite to the prime) and was basically “try to get more right feedback” (see for details on the instructions), suggesting that participants learnt the new association between the subliminal prime and their decision outside of awareness, thus learning implicitly. As for the agency reports falling through the training process, this is consistent with studies reporting dissociations between agency ratings and non-conscious information. That is, participants have been shown to rate their agency lower while not being aware of external influences on their behaviour. In summary, this proof-of-concept study opens new possibilities in the study of the effects of subliminal information on decision-making. # Supporting information We thank Gadiel Dumlao for his help with participant testing. [^1]: The authors have declared that no competing interests exist.
# Introduction In order to maintain homeostasis and promote survival, the circadian clock coordinates and maintains the 24-hour rhythmicity of various physiological processes, the most evident are the sleep/wake and fasting/feeding cycles. Disruption of the clock has been associated with an increased susceptibility and/or development of sleep and metabolic disorders. Proper circadian function occurs in response to synchronous expression of the molecular machinery of the circadian clock, composed of a central pacemaker in the suprachiasmatic nucleus (SCN) residing in the hypothalamus. While the master clock is entrained by light, almost all cells in the body harbor peripheral clocks that are entrained by both signals from the master clock and environmental cues. This molecular machinery is comprised of core clock proteins that control each other’s expression and function in a transcriptional/translation feedback loop; two transcriptional activators, BMAL1 and CLOCK, heterodimerize and drive the expression of Cryptochrome (Cry1, 2) and Period (Per1-3). Once Cryptochrome and Period proteins reach a critical threshold, PER:CRY complexes translocate to the nucleus to repress Bmal1:Clock transactivation. In addition to these core clock components, the nuclear receptors (NR) RORs (α, β, and γ) and the REV-ERBs (α and β) compete to activate and repress, respectively, the transcription of Bmal1 and Clock genes, completing an integral accessory loop which helps maintain circadian rhythmicity\[–\]. The REV-ERBs also participate in the regulation of a diverse set of metabolic processes, thus linking control of our daily rhythms with maintenance of metabolic homeostasis. Both REV-ERBs are ubiquitously expressed with very high expression patterns in the liver, adipose tissue, skeletal muscle, and brain, with both NRs exhibiting circadian patterns of expression\[–\]. The REV- ERBs are unique within the NR superfamily in that they lack the carboxy-terminal tail of the ligand binding domain (LBD called activation function 2 (AF-2, helix 12), which is required for coactivator recognition. As a result, both REV-ERBs are transcriptional repressors, recruiting corepressors such as NCoR in a ligand-dependent fashion. Both REV-ERBs bind to identical DNA response elements (termed RORE/RevREs) either as monomers to an AGGTCA “half site” with a 5’ AT- rich extension, or as homodimers to a DR2 element \[direct repeat (DR) AGGTCA sequence with a 2 base pair spacer\]. The retinoic acid receptor-related orphan receptors \[RORs (α, β, and γ)\] recognize the same DNA response element and are often coexpressed in the same tissues as the REV-ERBs. Due to the limited availability of genetic models to explore REV-ERBβ’s function, significantly more is known about REV-ERBα’s role in mammalian physiology. As a consequence, REV-ERBβ is considered functionally redundant to REV-ERBα and its role has, by default, been considered almost identical to REV-ERBα’s. Recent work has demonstrated that REV-ERBα is highly expressed in oxidative skeletal muscle and plays an integral role in mitochondrial biogenesis and oxidative function. Skeletal muscle accounts for \~40% of body mass, more than 85% of total insulin-stimulated glucose uptake, and is one of the most metabolically demanding major mass peripheral tissues. Consequently, skeletal muscle has a significant role in insulin sensitivity and the development of obesity. Due to REV-ERBα’s role in skeletal muscle, REV-ERBα- deficient mice display changes in daily energy expenditure, pre-disposing them to diet induce obesity. Work performed by Ramakrishnan and colleagues also demonstrated that REV-ERBβ was highly expressed in skeletal muscle, regulating genes involved in fatty acid and lipid absorption. Importantly, data presented by Ramakrishnan and colleagues implicated REV-ERBβ in the control of lipid and energy homeostasis in skeletal muscle. Given the overt skeletal muscle phenotype observed in REV-ERBα-deficient mice, coupled with their metabolic abnormalities and the general view held by some in the field that REV-ERBβ is functionally redundant to REV-ERBα, we sought to determine the extent of functional redundancy of REV-ERBβ to REV-ERBα in skeletal muscle and metabolism. Loss of REV-ERBβ resulted in increased mitochondrial biogenesis and genes involved in muscle metabolism both in vitro and in vivo. Characterization of REV-ERBβ-deficient mice demonstrated a metabolic profile opposite to that of REV-ERBα-deficient animals, including failure to gain weight despite increased food consumption during the day. This increased food consumption correlated with increased utilization of carbohydrates for energy as well as increased energy expenditure during the day. These data indicate that, at least in terms of muscle, REV-ERBβ is not functionally redundant to REV-ERBα, instead playing a distinct role in the regulation of skeletal muscle metabolism. Furthermore, these data suggest development of dissociated REV-ERB modulators (REV-ERBα-specific versus REV- ERBβ-specific) may be beneficial for the treatment of metabolic syndrome. # Materials and methods ## Animals The following mouse strains used were either purchased from the Jackson Laboratory and/or were bred at the Scripps Research Institute–Florida (Scripps Florida). Male B6.Cg-Nr1d1<sup>tm1Ven</sup>/LazJ, stock \# 018447 (REV- ERBα<sup>-/-</sup>); Male and female REV-ERBβ KO mice were generated as previously described. Wild-type (WT) littermates were used as controls for both knock out strains. This study was carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Animal care and experimental protocols used in this study were approved by the Scripps Florida Institutional Animal Care and Use Committee (Assurance number: D16-00726). Mice were housed in groups of 3–5 in 12 h light/12 h dark cycles at 23°C and fed standard chow (Harlan 2920X), unless otherwise stated. Mice had access to chow and water ad libitum. Mice were sacrificed by CO<sub>2</sub> asphyxiation for tissue processing. ## Food intake Food intake was monitored in normal chow-fed mice using the BioDAQ system (Research Diets) following vendor recommended procedures. The feeding patterns of pre-acclimatized, individually housed mice were continuously recorded for 3 consecutive days. The daily number of bouts, percent of time spent in bouts and total chow consumed were obtained. A bout constitutes a visit or set of visits (no more than 5 seconds apart) to the food hoppers. Each bout has a duration and amount of food obtained during it. Data represent the average day or night intake. Average values are compared by 2-tailed Student’s t test to determine significance. ## CLAMS Mice were individually placed and acclimatized in a Comprehensive Laboratory Monitoring System (CLAMS; Columbus Instruments) at 23°C for 48 h. Afterwards, VO2, VCO2, food intake, and spontaneous locomotor activity were measured for the indicated time periods. Respiratory exchange ratio (RER) and energy expenditure (EE) were calculated using the following equations: RER = VCO2/VO2; EE (kcal/h) = (3.815 + 1.232 X RER) x VO<sub>2</sub>. EE was normalized by lean body mass. Both raw and average values for RER are presented. ## Body composition studies Total, fat, lean, and fluid mass of mice was measured by Nuclear Magnetic Resonance using the Minispec LF-NMR (Brucker Optics) analyzer. ## Viral production, cell culture, and infection To generate murine REV-ERBα or REV-ERBβ retroviral vectors, mouse REV-ERB sequences were inserted into the MIGR1 vector (Addgene) using the XhoI site and further screened for orientation. MIGR1 empty vector was used as a control. MIGR1 was a gift from Warren Pear (Addgene, plasmid \# 27490). To knock down REV-ERBβ expression, 3 different shRNAmirs (TransOmic Technologies, Inc.) were inserted, individually into the pLMPd-Ametrine vector. The shRNAmirs used were top-scoring designs from shERWOOD analysis (TransOmic Technologies, Inc.). Each shRNAmir insert was PCR amplified using primers specific for common regions in the flanking miR-30 sequences, which included XhoI and EcoRI sites. pLMPd- Ametrine containing a CD8 shRNA was used as a control. Plat-E cells (Cell Biolabs, Inc.) were cultured in DMEM containing 10% fetal bovine serum, 2mM L-glutamine, and 1% penicillin/streptomycin at 37°C under standard culture conditions. Plat-E cells were seeded at 350,000 cells per mL in a 10cm dish the day before transfection. 5μg total plasmid DNA was transfected via Fugene6 reagent (Promega) according to manufacturer’s protocol. Viral supernatant was harvested 48 and 72 hours post transfection and stored at -80. C2C12 myoblasts were obtained from ATCC (Manassas, VA) and grown in Dulbecco modified Eagle’s medium (DMEM; 4.5 g/l D-Glucose; Gibco) supplemented with 10% fetal bovine serum, 2mM L-glutamine, and 1% penicillin (50units/mL)/streptomycin (50ug/mL). Prior to myogenic differentiation, C2C12s were plated at 30,000 cells/mL in 24 well plates and left to adhere overnight. The following day, cells were washed once with PBS and incubated overnight in 1mL total of 50% retrovirus conditioned media and 50% normal culture media containing Polybrene (8ug/mL, Santa Cruz Biotech). 24 hours later, retrovirus was removed and replaced with normal growth media for another 24 hours before myogenic differentiation was induced. One transduction was sufficient to obtain \>90% GFP/Ametrine-positive cells. Myogenic differentiation into myotubes was induced after cells reached confluency by adding DMEM supplemented with 2% horse serum (HS), 2mM L-glutamine, and 1% penicillin/streptomycin at 37°C in a humidified incubator under 5% CO<sub>2</sub> for 6 days. Overexpression and knock down was verified by qPCR analysis. ## MitoTracker and flow cytometry Retrovirally transduced C2C12 cells were washed with PBS, trypsinized and incubated at 37°C for 20 min with 100nM MitoTracker Red<sup>FM</sup> dyes (Molecular Probes). MitoTracker Red probe is a red-fluorescent dye that stains mitochondria in living cells and its accumulation is dependent upon membrane potential (Molecular Probes). To measure viability in the samples, a Fixable Viability Dye (Invitrogen/Thermo Fisher) was added to the cultures during the Mitotracker staining step. This dye can be used to stain both live and fixed cells to irreversibly label dead cells prior to analysis. Samples were washed three times in PBS and flow cytometric analysis was performed on a BD LSRII (BD Biosciences) instrument and analyzed using FlowJo software (TreeStar). ## mtDNA quantification Mitochondrial content was measured using methods previously described. Briefly, genomic DNA was extracted using the Blood & Cell culture DNA Mini Kit (Qiagen) per manufacturer’s instructions from C2C12 samples. DNA was quantified via a standard curve method in order to dilute all samples to 3ng/ul in TE buffer. Once normalized for concentration, samples were subject to RT-qPCR using a cocktail of SYBR green PCR master mix, template DNA, and mtDNA target specific primer pairs, plus H2O. Each sample was run in triplicate and data were averaged. This step was repeated in separate wells using nuclear DNA specific primers and the ratio between mtDNA and nuclear DNA was quantified. Dissociation curves were calculated for all samples to ensure the presence of a single PCR product. ## RNA isolation and quantitative real-time polymerase chain reaction (qRT-PCR) Muscle was harvested from WT and REV-ERBβ KO mice after a 5 hour fast, at ZT6 (Zeitgeber 6–6 hours into the murine nocturnal period). Total RNA was isolated from tissues by guanidinium thiocyanate/phenol/chloroform extraction. RNA concentrations were adjusted in order to load 500ng of RNA per cDNA reaction. Total RNA was isolated from C2C12 cells using a Quick-RNA MicroPrep kit (Zymo Research) using the manufacturers protocol. RNA concentrations were adjusted in order to load 1μg of RNA per cDNA reaction. cDNA was synthesized using qScript<sup>TM</sup> cDNA SuperMix synthesis kit (Quanta Biosciences). Quantification of each transcript by quantitative real time-polymerase chain reaction (qRT-PCR) was performed using SYBR Green dye (Quanta Biosciences) to detect dsDNA synthesis, and analyzed using cycling threshold (Ct) values. Relative expression was determined using the ΔΔCT method and normalized to the housekeeping gene 18s. Quantitative RT-PCR was performed with a 7900HT Fast Real Time PCR System (Applied Biosystems) using SYBR Green (Roche) as previously described. Primers were designed using Primer3 ([primer3.sourceforge.net](http://primer3.sourceforge.net)). Specificity and validation of the primers were determined using an In silico PCR software program ([genome.ucsd.edu](http://genome.ucsd.edu)) and melting curve analysis to eliminate the possibility of primer-dimer artifacts and check reaction specificity. Primer sequences can be found in. ## Statistical analysis All data are expressed as mean±S.E.M. Statistical analysis was performed using GraphPad Prism6 software. For multiple comparisons, two-way ANOVA was performed. Post-test analysis (ANOVA) was performed correcting for multiple comparisons using the Sidak-Bonferroni method. For all other independent data sets, since the data analyzed were not random, statistical significance was determined using two-tailed unpaired Student’s t-tests with no correction for multiple comparisons. ANCOVA analysis was performed using SPSS Statistics (IBM). (N per group per study is indicated in each figure legend). Significance was assessed as follows: \p\<0.05, \\* p\<0.01, \\\* p\<0.005, \\\\ p\<0.001. # Results ## Overexpression of the REV-ERBs drives mitochondrial biogenesis We previously demonstrated that REV-ERBα was a critical regulator of muscle mitochondrial biogenesis. To determine the cell autonomous role of REV-ERBβ versus REV-ERBα, we utilized the C2C12 cell line, a well-studied cell culture model commonly used to investigate myogenesis\[–\]. We retrovirally overexpressed REV-ERBα, REV-ERBβ, or empty vector retrovirus in proliferating C2C12 myoblasts two days prior to the induction of differentiation. MitoTracker Red staining, indicative of functional mitochondria, revealed that overexpression of both REV-ERBs enhanced mitochondrial biogenesis, demonstrated by increased fluorescence in the REV-ERB overexpressing cells relative to empty vector control. Compared to undifferentiated cells (Day 0; D0), the cells had differentiated and acquired a muscle specific phenotype by Day 6 (D6). Very little difference was observed in expression of Myogenin (Myog), Troponin I slow (Tnni1) and Troponin 1 fast (Tnni2), at Day 6 (D6) between groups. No overt viability issues were observed by day 6. Overexpression of both REV- ERBs also resulted in increased mitochondrial content, indicated by the increased expression of mitochondrial NADH dehydrogenase I (mt-Nd1), Cytochrome c oxidase I (mt-Co1,) and Cytochrome c oxidase II (mt-Co2), compared to empty vector control cells. These data suggest that the REV-ERBs can drive mitochondrial biogenesis in vitro. ## Overexpression of the REV-ERBs regulates the molecular clock and mitochondrial metabolism genes The REV-ERBs have been demonstrated to be core members of the circadian clock and participate in the regulation of a diverse set of metabolic processes, thus linking control of our daily rhythms with maintenance of metabolic homeostasis. Given the changes in mitochondrial content observed in the REV- ERB overexpressing cells, we next wanted to determine how expression of the clock genes correlated with expression of genes involved in mitochondrial function. Overexpression of REV-ERBα (gene name Nr1d1) significantly repressed almost all molecular clock genes, including REV-ERBβ (Nr1d2), Bmal1 (Arntl), Clock, the Cryptochromes (Cry1, Cry2), and the Period genes (Per1, Per2, Per3), which is consistent with its role as a transcriptional repressor. Overexpression of REV-ERBα also repressed expression of Cd36, a gene which is involved in lipid uptake. In contrast, RORα and PGC1α (Ppargc1a), the master regulator of mitochondrial biogenesis were upregulated with REV-ERBα overexpression. Additionally, genes encoding enzymes of fatty acid β-oxidation, notably carnitine palmitoyltransferase 1b (Cpt1b), long chain acyl-CoA dehydrogenase (Acadl), short chain acyl-CoA dehydrogenase (Acads), Ucp2, and Ucp3, genes involved in skeletal muscle metabolism, were also upregulated with REV-ERBα overexpression, consistent with previously published work. REV-ERBα also repressed expression of Sterol regulatory element binding-protein 1 (Srebf1), Stearoyl-CoA desaturase-1 (Scd1), Fatty acid synthase (Fasn), genes involved in lipogenesis, and Il6, a pleiotropic muscle myokine that has been shown to be involved in muscle growth, myogenesis, and regulation of energy metabolism. Overexpression of REV-ERBβ had similar effects on circadian clock and mitochondrial metabolism genes, with the exception that it potently repressed REV-ERBα expression. The negative regulation of REV-ERBα on REV-ERBβ expression and vice-versa is consistent their ability to negatively regulate their own expression through conserved ROREs/RevREs in their promoter regions. Interestingly, and in contrast to REV-ERBα overexpression, REV- ERBβ overexpression led to increased expression of Srebf1 and Fasn, but had no effects on expression of Scd1 or Il6. While it appears that REV-ERBβ was not as effective as REV-ERBα in the overexpression studies, this could be due to the differences in overexpression levels between the two receptors, with REV- ERBα being more highly expressed than REV-ERBβ. Regardless, the increased expression of genes involved in mitochondrial metabolism was consistent with the MitoTracker staining, mitochondrial content, and skeletal muscle metabolism genes. These results demonstrate an inverse relationship between the expression of the core circadian clock genes and the REV-ERBs and indicate that overexpression of REV-ERBα and REV-ERBβ can modulate the expression of genes driving fatty acid β-oxidation and lipid metabolism in C2C12 cells. However, under some circumstances, REV-ERBβ can not completely fulfill REV-ERBα’s role. ## Knock down of REV-ERBβ drives mitochondrial biogenesis Work previously performed assessing REV-ERBβ’s function in C2C12s suggested that it plays a significant role in the regulation of skeletal muscle lipid homeostasis. However, this analysis was performed in C2C12 cells using a dominant negative form of REV-ERBβ that lacked the ligand binding domain. Using this approach, Ramakrishnan and colleagues reported that dominant negative REV-ERBβ largely repressed genes involved in skeletal muscle energy expenditure and lipid catabolism. These results were in line with data demonstrating that loss of REV-ERBα in skeletal muscle led to decreased mitochondrial content and function. Given the overlap in function observed in our overexpression studies, we hypothesized that loss of REV-ERBβ would yield similar results as loss of REV-ERBα. Therefore, to fully assess how loss of REV-ERBβ affects myogenesis, we retrovirally overexpressed a REV-ERBβ shRNA or a CD8 control shRNA in proliferating C2C12 myoblasts 2 days prior to the induction of differentiation. Strikingly, knock down of REV-ERBβ enhanced mitochondrial biogenesis, indicated by the increased staining with MitoTracker Red. Analysis of Myog, Tnni1, and Tnni2 indicated that compared to undifferentiated cells (Day 0, D0), the cells had differentiated and acquired a muscle specific phenotype by Day 6 (D6). No overt viability issues were observed by day 6. Furthermore, loss of REV-ERBβ appeared to significantly increase expression of Tnni2. Knock down of REV-ERBβ also increased mitochondrial content, indicated by the increased expression of mt-Nd1, mt-Co1, and mt-Co2, compared to CD8 shRNA control cells. Interestingly, while knock down of REV-ERBβ led to increased expression of REV-ERBα and RORα, it also de-repressed all of the core molecular clock genes. Like overexpression of REV-ERBα, overexpression of RORα has also been shown to regulate the expression of genes involved in skeletal muscle metabolism. Knock down of REV-ERBβ also resulted in increased expression of almost all genes analyzed that were involved in lipid and skeletal muscle metabolism, including Cd36, Cpt1b, Acadl, Ucp2, and Ucp3, genes involved in skeletal muscle, which was consistent with the MitoTracker staining. No change was observed in Acads. However, knock down of REV-ERBβ resulted in decreased expression of PGC1α (Ppargc1a), Srebf1, Fasn, and Il6, but no change was observed in Scd1. While these results were largely in contrast to previously published work using a dominant negative REV-ERBβ construct, these results were at least consistent with the notion that in C2C12 cells, REV-ERBβ plays a role in the regulation of muscle metabolism and energy expenditure. Interestingly, these data are in contrast to the overexpression data and show a correlation between the circadian clock and skeletal muscle metabolic gene expression. ## Loss of REV-ERBβ drives skeletal muscle mitochondrial gene expression in vivo Given the surprising results we observed in C2C12 cells, we next wanted to determine whether loss of REV-ERBβ affected skeletal muscle gene expression in vivo in a similar manner. Using REV-ERBβ-deficient mice, we isolated (whole quadriceps) skeletal muscle from wild-type (WT) and REV-ERBβ knock out (KO) mice at Zeitgeber 6 (ZT6), 6 hours post lights-on and the time point in which peak REV-ERB mRNA expression occurs. qRT-PCR analysis indicated that, similar to the C2C12 data, loss of REV-ERBβ de-repressed most of the core clock genes, including REV-ERBα, RORα, Bmal1, and Clock. We observed no effects on Cry2 or Per3. Similar results were observed in REV-ERBα KO muscle. Loss of REV-ERBβ also led to increased expression of genes involved in lipid and skeletal muscle metabolism (Cd36, Cpt1b, Acadl, Acads) and energy expenditure (Ucp2, Ucp3). However, unlike the C2C12 cells, REV-ERBβ KO mice also had increased expression of PGC1α (Ppargc1a). Similar to the C2C12s, REV-ERBβ KO mice also exhibited decreased expression of Srebf1, Fasn, and Il6, with little effect observed on Scd1. Moreover, loss of REV-ERBβ appeared to shift the muscle towards a more oxidative phenotype as we observed increased expression of tropomyosin 3 (Tpm3), a marker of oxidative type 1 fiber and myosin heavy chain IIa (Myh2), a myosin heavy chain gene more associated with oxidative fibers than others. The expression of Myh4, a marker of type 2b glycolytic fibers, was unaltered. The shift towards a more oxidative phenotype in REV-ERBβ KO mice was in stark contrast to what was observed in REV-ERBα KO mice, which was largely a shift towards a less oxidative phenotype demonstrated by decreased expression of PGC1α, Cpt1b, Acadl, Acads, Ucp2, Ucp3, Myh2, and Tpm3 and is consistent with previously published results. However, not all changes in gene expression between REV-ERBβ KO and REV-ERBα KO mice were conflicting. REV-ERBα KO mice also demonstrated decreased expression of the lipogenic genes Srebf1 and Scd1 in muscle while loss of REV-ERBα de- repressed Fasn and Il6 However, unlike the muscle tissue, gene expression analysis of livers from REV-ERBβ KO and REV-ERBα KO mice demonstrated similar trends in gene expression (G6Pase, Cd36, Clock) or no effects (Pepck) on gene expression in vivo. These data are consistent with previously published work indicating that REV-ERBα and REV-ERBβ have largely overlapping roles in liver metabolism, with REV-ERBα acting as the dominant factor. These data suggest that in liver, REV-ERBβ appears to have overlapping functions to REV- ERBα. However, in skeletal muscle, the effects mediated by the loss of REV-ERBβ are in contrast to those of REV-ERBα. ## REV-ERBβ-deficient mice exhibit an altered circadian metabolism and feeding schedule Due to the increased expression of genes involved in mitochondrial biogenesis in the REV-ERBβ KO mice compared to WT controls, we wanted to determine whether this translated to effects on metabolism and any other circadian driven behavior. Real-time oxygen consumption (VO<sub>2</sub>), carbon dioxide production (VCO<sub>2</sub>), ambulatory activity and food consumption were monitored in twelve week-old male mice fed a normal chow diet using the CLAMS system. Energy expenditure (EE) and RER (Respiratory exchange ratio) were calculated based on the gas exchange as described in the Methods section. Body composition was obtained just prior to the CLAMS analysis, with no differences observed in lean, fat, and fluid masses or total body weight between groups. The RER is a surrogate indicator of the caloric source being utilized for energy production by an animal. It varies between 0.7, where lipids are the predominant fuel source, to 1.0, where carbohydrates are the main contributors. The REV-ERBβ KO mice showed a delay in the typical circadian drop in RER that occurs from the dark to light phase transition when mice shift from high activity and feeding to resting and sleep. This prolonged high RER cycle indicates an increased relative contribution of carbohydrates as an energy source in the KO animals. This could be due to their increased chow consumption, observed during the light phase while in the CLAMS. This light-phase specific increase in eating was further confirmed in REV-ERBβ KO mice tested in the BioDAQ system, which uses highly sensitive feeding monitoring hoppers adapted to regular housing cages. Additionally, an increase in energy expenditure was detected in association with the observed increased daytime chow consumption in the REV-ERBβ KO mice. Nighttime energy expenditure in REV-ERBβ KO mice was comparable to WT mice, despite the KO animals being less active during this time. Whether the increase in energy expenditure in the KO mice during the daytime can be attributed to elevated feeding and digestive processes per se remains to be tested. Ambulatory activity during the light-phase was similar between the groups and thus, cannot explain that. However, upon fasting, the KO’s demonstrated decreased RER relative to WT controls, which may be a result from a faster transition to fat burning. Regardless, the increased caloric output can, at least in part, account for the absence of weight gain by the KO’s in spite of their higher caloric intake. These data indicate that REV-ERBβ regulates circadian behaviors and metabolism in vivo. # Discussion REV-ERBα has recently been identified as a key regulator of skeletal muscle mitochondrial function. Due to the overlapping gene expression profiles and similarity in DNA binding domains, REV-ERBβ is thought to be redundant to REV- ERBα. Our studies were aimed at gaining a better understanding of REV-ERBβ’s function in skeletal muscle and overall metabolism. In this report, we utilized the C2C12 in vitro cell culture model and REV-ERBβ-deficient mice to investigate the role of this NR in comparison to REV-ERBα’s function in skeletal muscle in vitro and in vivo. Here we report that overexpression and knock down of REV-ERBβ drives mitochondrial biogenesis in C2C12 cells, exhibited by increased MitoTracker Red staining and increased expression of genes involved in skeletal muscle energy metabolism. We also showed that skeletal muscle from REV- ERBβ KO mice exhibited a similar gene expression profile as the C2C12 cells in which REV-ERBβ was knocked down, which was in contrast to skeletal muscle gene expression in REV-ERBα KO mice. Furthermore, we demonstrated that REV-ERBβ KO mice present with an altered metabolic phenotype and feeding behavior schedule, with KO mice eating more and utilizing more carbohydrates as fuel during their nocturnal/sleep period. In terms of muscle and energy expenditure, our data suggest that REV-ERBβ may not be as functionally redundant to REV-ERBα as originally hypothesized. Our initial hypothesis was based on the view, held by some in the field, that REV-ERBα and REV-ERBβ are functionally redundant, with REV-ERBα acting as the dominant factor. Overexpression studies are excellent tools to determine whether a protein, in this case REV-ERBβ, is able to perform certain cellular functions. Our initial studies in C2C12 cells supported this hypothesis as overexpression of each REV-ERB resulted in relatively similar functional outputs, largely increased mitochondrial biogenesis and expression of genes involved in skeletal muscle metabolism. However, the shRNA knock down of REV- ERBβ drove further inquiry into REV-ERBβ’s function in skeletal muscle. The increased MitoTracker Red staining and the changes in gene expression observed surprised us, particularly since evaluation of C2C12 cells using a dominant negative REV-ERBβ yielded opposite results to ours. However, this construct still possessed REV-ERBβ’s N-terminal region, which has been shown to interact with the transcriptional co-factor Tip60 in order to regulate gene transcription in a ligand-independent manner. Thus, REV-ERBβ was not completely eliminated and could have recruited co-factors to modify gene transcription in a ligand-independent manner. Alternatively, since RORα and REV-ERBα bind the same DNA consensus sequence as REV-ERBβ, the overexpression of the dominant negative REV-ERBβ could have blocked the ability of RORα or REV-ERBα to bind to the RORE and again, affect transcription. However, use of REV-ERBβ shRNA bypassed these issues and may account for the differences. Overexpression of REV-ERBα has been demonstrated to drive mitochondrial biogenesis whereas loss of REV-ERBα has the opposite effect. However, the effects rendered by overexpression and loss of REV-ERBβ were not opposite to each other. It is possible that the increased mitochondrial biogenesis in the absence of REV-ERBβ is a consequence of increased REV-ERBα or increased RORα. Indeed, we consistently observed increased expression of both NRs in the absence of REV-ERBβ. Like overexpression of REV-ERBα, overexpression of RORα has also been shown to regulate the transcription of genes involved in lipid metabolism and energy metabolism in skeletal muscle. REV-ERBα, RORα, and REV-ERBβ have been demonstrated to bind to conserved ROREs/RevREs in REV-ERBα’s promoter region, with RORα driving gene expression whereas the REV-ERBs repress REV-ERBα gene expression. This autoregulation is likely the reason for the dramatic downregulation of gene expression observed on REV-ERBβ when REV-ERBα was overexpressed and vice versa. It is possible that knock down or loss of REV- ERBβ relieved repression on REV-ERBα, hence the increased expression in vitro and in vivo. Another possibility is that the overexpression of RORα, which can also bind ROREs, drove gene expression and thus, increased mitochondrial biogenesis. Alternatively, loss of REV-ERBβ could have relieved repression on Bmal1, which is known to activate RORα and REV-ERBα transcription through binding to their respective E-box regions. Increased Bmal1 may account for the increased expression of REV-ERBα and RORα observed both in vitro and in vivo. Future studies need to be performed to determine the mechanism and which NRs may be driving the increased mitochondrial function in both REV-ERBβ- deficient C2C12 cells and in skeletal muscle in vivo. The circadian clock influences a broad range of physiological processes, including fasting/feeding cycles. Initial analysis of REV-ERBα KO versus REV-ERBβ KO mice indicated that while loss of REV-ERBα affected circadian output, loss of REV-ERBβ had minimal effect. However, our data indicate that REV-ERBβ does affect circadian gene expression in a manner similar to REV-ERBα. Overexpression and knock-out of REV-ERBβ generated a similar gene expression profile as REV-ERBα overexpression and knock-out. However, our data indicate that at least in vitro, the circadian clock is not as intimately linked to skeletal muscle metabolism as initially thought. While overexpression and knock down had different effects on expression of core clock components, the skeletal muscle metabolic genes were consistently upregulated and appeared to depend more on the expression of RORα and the REV-ERBs. While this could also be said for the in vivo data, we do know that what occurs in vivo is significantly more complex and REV-ERBα KO mice present with altered circadian rhythms, which could affect the muscle phenotype. Our data also indicate that REV-ERBβ affects circadian outputs, indicated by increased food consumption during the day. This increased food intake could account for the shift and increased RER and EE observed during the day when measured by indirect calorimetry. However, research has revealed that disordered eating, or eating when we are supposed to be sleeping, contributes to weight gain, metabolic syndrome, and obesity. Despite this, REV-ERBβ KO mice do not gain weight over time, instead maintaining similar weight profiles as WT littermate controls. It is possible that the metabolic demands of the skeletal muscle may compensate for the increased food intake. This may also explain the decreased activity observed in the REV-ERBβ KO mice during their “active” periods. That is, moving less compensated for the increased metabolism with no change in food intake. Alternatively, the food intake during the day may have disrupted their sleep schedule and the decreased movement may be a function of sleep compensation. Clearly, further in-depth analysis of these behaviors is needed to better understand the phenotype observed. # Conclusions REV-ERBβ plays a role in skeletal muscle mitochondrial biogenesis both in vitro and in vivo. This effect on mitochondrial gene expression in REV-ERBβ KO mice differs from that of REV-ERBα KO mice as each receptor appears to have effects opposite to each other. Furthermore, these differences translate to profound differences on metabolism in vivo. Our data suggest that REV-ERBβ may not be as functionally redundant to REV-ERBα as originally hypothesized. Thus, development of dissociated REV-ERB modulators (REV-ERBα-specific versus REV- ERBβ-specific) may be beneficial for the treatment of metabolic syndrome. # Supporting information We would like to thank Dr. Nelson Bruno for his helpful comments and critiques of experimental design and data analysis. [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Institute for Genomic Medicine, Columbia University Medical Center, New York, NY, United States of America [^3]: Current address: Center for Clinical Pharmacology, Washington University School of Medicine and St. Louis College of Pharmacy, St. Louis, Missouri, United States of America
# Introduction D-dimer is a degradation product of crosslinked fibrin that appears in the blood after a blood clot is degraded by fibrinolysis. Elevated D-dimer levels in the blood predict increased secondary fibrinolytic activity and are a principal marker of hypercoagulation and thrombosis. Many studies have shown that high D-dimer levels are associated with the risk of deep venous thrombosis (DVT) and pulmonary embolism, which are serious post-surgery complications with high mortality especially in patients with malignant tumours. Due to the injury to vascular endothelial cells caused by toxins released from fast growing tumour cells and the fibrinolytic activator on the surface of tumour cells, cancer patients often exhibit abnormal coagulation and fibrinolytic activities and their D-dimer levels tend to be higher than those in non-neoplastic populations. The D-dimer levels of almost all cancer patients exceed the recommended limits according to the existing reference range (0–0.5 mg/L). Therefore, the high risk of DVT according to this range might be overestimated. It is suggested that the current D-dimer reference range is unsuitable for cancer patients, which limits the application of D-dimer testing in laboratory diagnosis and the prevention of tumour venous thrombosis and venous thromboembolism (VTE). The present prospective study measured plasma D-dimer levels in a large sample of cancer patients to investigate the range of D-dimer concentration in the absence of VTE. In addition, changes in D-dimer levels in cancer patients were determined during the perioperative period and after chemotherapy or radiotherapy. # Materials and Methods ## Patients and controls ### Patients with malignant tumours A total of 1368 patients treated at Hubei Tumor Hospital and Wuhan Tongji Hospital between January, 2010 and August, 2013 were selected. The exclusion criteria comprised the following factors that may affect D-dimer level: hypertension; diabetes; personal or family history of thromboembolic disease; coagulopathy; cardiovascular/cerebrovascular disease; autoimmune disease (such as rheumatoid arthritis, systemic lupus erythematosus, or idiopathic thrombocytopenic purpura); infection in the previous 30 days with body temperature \> 37.5°C; surgery or trauma in the previous 30 days; blood transfusion in the previous half year; stressful state on the day before blood collection; and patient taking medications that may affect coagulation and the fibrinolytic system. Each patient was diagnosed with malignant disease, most based on pathology findings; some patients had advanced stage inoperable cancer confirmed with CT or another imaging method. The patients comprised the following: 142 liver cancer, 150 pancreatic cancer, 140 breast cancer, 120 stomach cancer, 120 colorectal cancer, 240 lung cancer, 120 gynaecological tumour, 123 oesophageal cancer, and 120 head and neck tumour. The average age was 55 years; 621 patients were male and 654 were female. Thirteen of the cancer patients (all male) were diagnosed with DVT and excluded from the study. ### Patients with benign tumours A total of 93 in-patients treated at Hubei Tumor Hospital from January, 2010 to August, 2013 were selected. The selection criteria were the same as above. Each patient received a pathological diagnosis of benign tumour. Their average age was 48.6 years; 20 were male and 73 were female. ### Healthy control group A total of 150 in-patients treated at Wuhan Tongji Hospital during the same period were selected. Their average age was 44.2 years; 103 were male and 47 were female. ### Ethics Statement This study was approved by the ethics committee of Hubei Tumor Hospital and Wuhan Tongji Hospital. All participants joined the study voluntarily and provided written informed consents. The committee approved the experiments and the methods were conducted in accordance with the approved guidelines. ## Methods ### Sample collection, instruments, and reagents Samples of 1.8 mL of elbow venous blood were collected in anticoagulant tubes with 109 mol/L sodium citrate for an anticoagulant to venous blood ratio of 1:9. After immediate mixing, the samples were centrifuged at 3000 × g for 10 min and then tested within 2 hours. D-dimer, activated partial thromboplastin time (APTT), prothrombin time (PT), thrombin time (TT), and fibrinogen (FiB) level were determined using a STA-R Evolution coagulant analyser with its specific reagents (Stago, Asnières-sur-Seine, France). Platelet levels were tested using a five-part blood cell counter with its reagents (Sysmex, Kobe, Japan). ### D-dimer, APTT, PT, TT, and Fib assay A STA-R Evolution coagulant analyser (Stago) was used to measure D-dimer, APTT, PT, TT, and FiB by immune turbidimetry. All samples were handled as routine clinical samples. ### Analysing the correlation between D-dimer levels and tumour stages or tumour pathological types D-dimer was compared among patients with tumours of stage I-II, or stage III-IV; moreover, in the same type of cancer, D-dimer was compared among patients with different pathological types. ### D-dimer levels in tumour patients during the perioperative period Of the 1275 patients with malignant tumours, 101 who underwent surgery were selected. These comprised 39 liver cancer, 10 breast cancer, 10 stomach cancer, 10 colorectal cancer, 10 lung cancer, 17 gynaecological tumour, and five head and neck tumour patients. Venous blood was collected on day 1, day 3, and 1 week after surgery for measurement of D-dimer level. ### D-dimer levels in tumour patients before and after radio/chemotherapy Of the 1275 patients with malignant tumours, 130 cases who underwent radio/chemotherapy were selected. These comprised 39 cases of liver cancer, 19 cases of breast cancer, 14 cases of stomach cancer, 18 cases of colorectal cancer, 35 cases of lung cancer, 2 oesophageal cancer, and 3 cases of gynaecological tumour patients. Of these 130 patients, 29 cases were treated with radiotherapy and 101 were treated with chemotherapy (detailed chemotherapy regiments see supporting information 1). All plasma samples were collected after one cycle of therapy to determine D-dimer levels. ### Analysing the correlation between D-dimer levels and cancer prognosis Of the 1275 patients with malignant tumours, 190 patients were tracked after being discharged from hospital to reveal the relevance of initial D-dimer levels to cancer prognosis. These cases comprised 15 cases of gastric cancer, 26 cases of colorectal cancer, 23 cases of liver cancer, 10 cases of pancreatic cancer, 30 cases of lung cancer, 15 cases of gynaecologic cancer, 48 cases of breast cancer, 21 cases of head neck cancer and 2 cases of oesophageal cancer. ## Statistical analysis SPSS16.0 (IBM, Armonk, NY, USA) was used for statistical analysis. First, the Kolmogorov–Smirnov method was used to test the normality of the D-dimer data, which were found to have a non-normal distribution. Therefore, simultaneous quantile regression was performed to construct a median, 5th percentile, and 95th percentile model of normal tumour D-dimer concentration. The D-dimer levels of the benign tumour group, the malignant tumour group, and the healthy controls were compared using the Mann–Whitney U test; P \< 0.05 indicated statistically significant differences. Comparisons between malignant tumours in different locations or the same malignant tumour at different stages or of different pathological types were performed using the Mann–Whitney U test; P \< 0.05 indicated statistically significant differences. D-dimer levels before and after treatment were also compared using the Mann–Whitney U test, with P \< 0.05 indicating statistically significant differences. # Results ## D-dimer levels in benign tumour group, malignant tumour groups, and healthy controls displays a diagram of the entire study. The baseline characteristics of the patients are reported in. The conventional coagulation measures (APTT, PT, TT, Fib, and platelet count) in all cases were normal. The D-dimer level in each malignant tumour group was higher than that in the benign tumor group (P \< 0.05) and in the healthy control group (P \< 0.05). There was no significant difference between the benign tumour group and the healthy controls (P = 0.11). D-dimer levels in participants who were older than 55 years were much higher than those in patients less than 55 years old in the healthy controls and the benign tumour group, independent of gender. In the malignant tumour group, participants over 65 years had higher D-dimer levels than those below 65 years, independent of gender. Thus, gender did not affect the D-dimer level, whereas age did. The D-dimer levels of older participants were higher than those of younger participants. ## D-dimer range and level in tumour patients and healthy controls The ranges of D-dimer concentration in malignant and benign tumour patients and healthy controls were represented as the median and 5th and 95th percentile values. D-dimer levels in liver, pancreatic, stomach, colorectal, lung, and oesophageal cancer patients were significantly higher than those in breast cancer patients (P \< 0.05). ## D-dimer level and tumour stage When D-dimer was compared among patients with tumours of differing stage in the same primary site, those with stage III/IV disease had significantly higher levels than those with stage I/II disease (P \< 0.05), as shown in. ## D-dimer level and pathological type When D-dimer was compared among patients with cancer in the same primary site but of differing pathological type, there were no significant differences for most tumour types (P \> 0.05). D-dimer levels were higher in bile duct cell carcinoma patients than in hepatocellular carcinoma patients (P = 0.018) and in squamous cell carcinoma compared with adenocarcinoma in patients with lung cancer (P = 0.035). ## Changes in D-dimer levels of tumour patients without thrombosis in the perioperative period Plasma D-dimer levels in 101 tumour patients without thrombosis were significantly raised (up to 20 mg/L) on day 1 after surgery but had significantly decreased on day 3; they had returned to within the recommended range for cancer patients 1 week after surgery. In some cases, D-dimer levels continued to increase after surgery with no sharp decrease after 1 week; this was observed in one stomach cancer patient, ten liver cancer patients, and two pancreatic cancer patients. Five of them had definitely confirmed VTE by CT films. Ten of them recovered and three died. Continued monitoring of the 10 recovered patients showed that their D-dimer levels gradually decreased to within the recommended range for cancer. ## Effect of radio/chemotherapy on plasma D-dimer levels There was no significant difference between plasma D-dimer levels before and after treatment in 101 patients who underwent chemotherapy and 29 who received radiotherapy (P = 0.94 and P = 0.72, respectively). ## Initial D-dimer levels correlated with prognosis Survival rate of some cancer patients were successfully tracked by the end of 2015. As shown in, cancer patients with higher initial D-dimer value were shown poor prognosis compare with those who had lower initial D-dimer value. # Discussion The kit used to measure D-dimer levels in this study is commonly employed in clinical laboratories and in its instructions defines a concentration of 0.5 mg/L as normal. This value has been shown to be appropriate for healthy people; however, more than 60% of cancer patients with normal coagulation exhibit D-dimer levels greater than 0.5 mg/L. In the present study we aimed to investigate the range of D-dimer concentration in cancer patients and its influencing factors. Even if such a range does not contribute to the diagnosis of VTE, it can provide data to support the exclusion of this disease. All steps in the study strictly followed the rules set out in the CLSIC28-A3 document and subjects were selected based on rational inclusion and exclusion criteria. To explain the differences in D-dimer level among the various cancers, we systematically analysed all patient data, including their basic characteristics, routine clinical laboratory tests, clinical treatment, and pathological diagnosis. Broadly, in the cancer population D-dimer level was independent of gender but was affected by the age of the patient and the stage of the tumour. D-dimer was also influenced by the pathological type of the tumour in bile duct cell carcinoma patients, who had higher levels than hepatocellular carcinoma patients, and, among lung cancers, in squamous cell carcinoma patients, who had higher levels than adenocarcinoma patients (P \< 0.05); these significant differences may have been due to the later stage of the cancers in these cases. Haase et al. showed that D-dimer levels increased markedly with age in healthy individuals. Consistent with this study, we found that patients aged over 55 years had higher D-dimer levels than those aged less than 55 years (P \< 0.05). Haase et al. also reported a certain degree of differences were found between the reference ranges for males and females among older people; yet they also indicated the difference was minor and the clinical relevance was highly questionable. Therefore, no difference of D-dimer levels between the sexes in our study is a comprehensible result. Interestingly, we found that the ranges of D-dimer concentration in breast cancer and head and neck tumour patients differed from those in the other malignancies. D-dimer levels in patients with breast or head and neck cancer were much lower than those in patients with liver, pancreatic, stomach, colorectal, lung, gynaecological, or oesophageal cancer (P \< 0.05). Some studies have reported D-dimer levels in cancer patients to be strongly associated with the number of metastatic nodes and patient prognosis. Compared with liver, pancreatic, stomach, colorectal, lung, gynaecological, and oesophageal cancer, breast and head and neck cancers have lower metastatic rates and a more favourable prognosis based on clinical statistics. Although the prevalence of lymph node metastasis might be high in the early stages of breast cancer, it is generally accepted that 5-year survival in breast cancer is much higher than that for other carcinomas. We also found that, among patients with the same cancer, D-dimer levels were markedly higher in stage III/IV disease than in stage I/II. The significant differences observed between tumour stages show that D-dimer level is also strongly associated with the grade of malignancy. Surgery may induce embolisms. Our findings demonstrate the importance of monitoring D-dimer levels in all perioperative cancer patients. D-dimer changes in such patients show a specific trend, as described above. Patients whose D-dimer changes resemble this trend may suffer VTE, and persistently elevated D-dimer levels may indicate a poor prognosis. It is therefore crucial to monitor the D-dimer levels of cancer patients at 3 days and 7 days postoperatively. Clinicians should take the necessary measures to prevent VTE or DVT if levels have not decreased to normal 1 week after surgery. The D-dimer levels of 130 patients with malignant tumours were measured 1 week after they underwent radio/chemotherapy. D-dimer did not appear to be influenced by radio/chemotherapy based on our results. The use of bevacizumab combined with chemoradiotherapy is associated with a higher risk of VTE compared with antiangiogenic therapy alone. Among 29 patients in our study who underwent radiotherapy and 101 who received chemotherapy, none was given bevacizumab at any point in his or her treatment. Therefore, we suggest that the effects of radio/chemotherapy on D-dimer levels should be evaluated in light of any drugs that the patient has been administered. The D-dimer level of patients whether can be served as an indicator of prognosis is also drawing us, even though so many factors might affect 5-year survival rate of cancer patients. As we expected, our study showed the initial D-dimer level of dead patients were obviously higher than that of the patients who were still alive at the end of 2015. Apart from gastric cancer patients, the initial d-dimer levels of other dead patients were higher than 1.00 mg/L, including patients in liver cancer, pancreatic cancer, lung cancer and gynaecologic cancer. These results consist with other reports before \[, and \]. The previous study with regard to lung cancer found the d-dimer median concentration was 0.84mg/L. As presented in our results, such level was 0.7mg/L. To date, our study firstly analyzed the d-dimer ranges on the other types of cancer. However, we couldn’t make a conclusion on the definite d-dimer range which can be served as an indicator of poor prognosis, because the cases of tracked patients are not idea. Hence, further studies are required to confirm our findings. # Conclusions We have discussed the cancer-specific concentration range for D-dimer and evaluated factors that could influence D-dimer levels in cancer patients. D-dimer level is independent of gender but dependent on patient age, tumour primary site, and tumour stage. In addition, monitoring changes in D-dimer level is critical for all perioperative cancer patients. Clinicians should take the necessary measures to prevent VTE or DVT if the D-dimer level has not decreased to the recommended range 1 week after surgery. The applicability of this research should be further investigated in a large prospective study. When sufficient cases and influencing factors are available, we may be able to establish an acceptable “normal” D-dimer reference range for cancer patients. # Supporting Information We are grateful to the Department of Microbiology, School of Basic Medical Science, Wuhan University, for supporting this work. We would like to thank Stago Diagnosis Company for reagent support. We also thank Lang Chen and Xianglei Wu for their excellent advices. [^1]: We have the following interests: Anming Yu is employed by Stago Diagnosis Company. Stago Diagnosis Company supplied partial D-dimer testing reagents and some technic supports. There are no patents, products in development or marketed products to declare. This does not alter our adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: D. Li JY YW. Performed the experiments: JY D. Lei FY FP HZ KW HC. Analyzed the data: JY XT. Contributed reagents/materials/analysis tools: AY XW LC. Wrote the paper: JY D. Li.
# Introduction Every year more than 50,000 women in the UK develop breast cancer, with 11,433 dying from the disease. Globally, it is the most common cause of female cancer death worldwide, with more than 500,000 women estimated to die every year from over 1.65 million cases. Nearly one quarter of breast cancer patients in the UK die within 10 years, with the majority presenting with distant disease several years after their initial breast cancer diagnosis. This highlights an urgent need for prognostic, predictive and pharmacodynamic biomarkers to identify patients at increased risk of recurrent disease, for early identification of recurrence, to facilitate tailored treatment and to monitor treatment response. Circulating tumour cells (CTCs) isolated from venous blood samples could be a relatively non-invasive real-time liquid biomarker that allows detection, monitoring and phenotyping of breast cancer. CTCs are rare cells (few per 10 ml blood /100 million leukocytes and 50 billion erythrocytes), and therefore highly sensitive assays with techniques to enrich then characterise CTCs are required. Techniques to enrich for CTCs are based either on tumour cell antigen expression or cell size or density. However, all these techniques are limited by their ability to capture all CTCs given the presence of CTC heterogeneity. Once enriched, CTC identification and characterisation can be achieved on a cellular level through microscopy or flow-cytometry, or on a molecular level using RT- PCR, however once again tumour heterogeneity remains the major challenge. Currently the only FDA approved CTC technology is the antibody-based CellSearch system. This identifies epithelial cells through their expression of the epithelial markers Epithelial cell adhesion molecule (EpCAM), cytokeratins 8, 18 and 19, and the lack of CD45 leukocyte marker expression. A similar, widely used technology is AdnaTest BreastCancer (AdnaGen, Langenhagen, Germany), which uses both an anti-EpCAM antibody and an antibody against the epithelial cell surface associated glycoprotein, mucin-1 (MUC-1). However CTCs are heterogeneous, and such positive selection techniques based on expression of epithelial markers are limited by the potential loss of CTC expression of such epithelial phenotype as the CTCs develop a more stem cell-like phenotype and behaviour pattern due to the epithelial-to-mesenchymal transition (EMT) that occurs to facilitate tumour cell adhesion, motility and subsequent intravasation. Although CellSearch has widespread popularity in the metastatic breast cancer setting, its use in early breast cancer is more limited. A study of 2026 higher- risk early breast cancer patients (66% node positive) confirmed the presence of CTCs as an independent prognostic marker for disease-free and overall survival. However only 21.5% of patients had CTCs, despite a large blood volume (30ml) analysed. In metastatic breast cancer, a major proportion of CTCs show EMT and tumour stem cell characteristics, with such characteristics being associated with treatment- resistance and worse prognosis. Cytokeratin-negative CTCs may represent a more aggressive CTC subpopulation and a majority of blood-borne tumour cells. Therefore whilst more established CTC enumeration and identification technologies such as CellSearch and AdnaTest employ anti-epithelial marker antibodies, more recently there has been a push towards size-based isolation to improve the CTC capture rate and better capture CTC heterogeneity. There is a range of promising non-epithelial based methodologies under evaluation, none of which have yet risen to the prominence that CellSearch has in epithelial-based CTC detection. The ISET device (Rarecells) is a filtration based CTC enrichment technology that collects large blood cells (\>8μm) including CTCs and circulating tumour microemboli (CTM) on a filter membrane. Of the ‘large cells’ captured by this technique, CTCs are routinely identified by morphology. These CTCs and CTMs containing both epithelial and mesenchymal CTC subpopulations can then be subjected to immunomorphological, immunofluorescence, genetic, DNA or RNA analysis. ISET has allowed the identification of CTCs co-expressing cytokeratins and the mesenchymal marker vimentin. Higher rates of CTC positivity by ISET than by Cellsearch using a morphology-based approach have been reported in NSCLC, pancreatic cancer, metastatic prostate cancer and melanoma despite larger cell size criteria. This is consistent with the presence of non-epithelial CTC populations. CTC enumeration by ISET has been correlated with shorter survival in hepatocellular carcinoma. In this exploratory study, we aimed to determine the role of CTC enumeration by a non-epithelial marker dependent technique, using ISET in a high-risk early (non-distant metastatic) breast cancer population, with a view to developing novel biomarkers for response to treatment. ## Patients and methods ### Patient population Following written informed consent, treatment-naïve, early breast cancer patients undergoing surgical resection were recruited at University Hospital of South Manchester (UHSM). Sex matched healthy controls were also recruited. CTC- positive (by CellSearch) metastatic breast cancer patients with oestrogen receptor positive disease were recruited solely for the development of an ER- based multiplex immunocytochemistry (ICC) assay. This study was approved by Oldham Research Ethics Committee (Ref: 09/H1011/47) and NRES Committee North West—Greater Manchester Central Ethics Committee (Ref: 12/NW/0447) and sponsored by University Hospital of South Manchester. ### Blood sampling A 20ml sample of peripheral venous blood was collected using CellSave preservative and EDTA vacutainer tubes for CellSearch and ISET analysis respectively. Blood was collected preoperatively and at surgery shortly after tumour removal, to investigate any systemic tumour or epithelial cell release as a result of surgical tissue handling. Blood was stored at room temperature until analysis. ### ISET filtration Within four hours of venesection, whole blood was diluted 1:9 with Rarecells buffer, and left to stand for 10 minutes. This allowed lysis of erythrocytes and fixation of nucleated cells as previously described. The mixture was then filtered using the ISET filtration device at a pressure of 5–9 kPa as per manufacturer’s instructions (Rarecells) to allow capture of large cells (\>8μm) on the ISET filter membrane. The membrane was allowed to dry overnight at room temperature before storage at -20°C. ### Immunocytochemical (ICC) staining of ISET filters: Morphology based method Following thawing ISET ‘spots’ from filters corresponding to 1ml whole blood were rehydrated using TBS buffer and subjected to antigen retrieval in a 99°C water bath using pH6 retrieval solution (Cat# S1699, Dako). Following permeabilisation (TBS + Triton X-100) and peroxidase blocking steps, the ISET spots were incubated with a primary mouse CD45 (Monoclonal, C7230, Dako, Glostrup, Denmark, 1/30 dilution) and CD144 (Monoclonal, 14-1449-82, eBioscience, San Diego, USA, 1/50 dilution) antibody cocktail at 4°C overnight. Anti-mouse-HRP secondary antibody (K4001, Dako) was applied the next day, with DAB substrate subsequently applied and the spots counterstained with haematoxylin. Spots were mounted onto slides using Faramount mounting media and allowed to set overnight at room temperature. ### Multiplex immunocytochemical (ICC) oestrogen receptor alpha (ERα)-based staining The nuclear ER receptor represents a potential non-epithelial marker of CTCs in the ER positive subset of breast cancer patients. Enzo Life Science’s Multiview (Mouse-HRP/Rabbit-AP) multiplex IHC Kit (ADI-050-100-0001, New York, USA) was used in combination with a rabbit Oestrogen Receptor alpha (ERα) antibody (Monoclonal, ab108398, Abcam, Cambridge, UK, 1/100 dilution) in a method incorporating the antigen retrieval, permeabilisation, peroxidase blocking and CD45/CD144 staining steps described above. ERα/CD45/CD144-stained ISET spots were mounted onto slides using Faramount mounting media as above. Oestrogen receptor positive MCF-7 breast cancer cells (ATCC HTB-22, obtained directly) spiked into healthy donor blood were used in assay development. ### Scoring of CTCs The mounted ISET spots were scanned using a Bioview automated light microscopy scanning system based on an Olympus BX61 microscope, using Duet software (Bioview, Rehovot, Israel). A circle of radius of 4000μm from the centre of each ISET spot was scanned to capture all cells. Galleries of images were manually reviewed using the Bioview Solo software. In the morphology based approach, CTCs were identified based on a set morphology definition as per previous publications and under the guidance of a consultant breast-specialist histopathologist (SP). All images were reported by JC, with a subset of the images reviewed by SP for concordance. Cells greater or equal to 16μm diameter, with hyperchromatic nuclei and negative for CD45/CD144 brown chromogen staining were classified as CTCs. For the multiplex immunocytochemistry (ICC) ERα-based approach, ERα positive CTCs were defined as cells ≥16μm diameter with positive for ERα red chromogen staining and negative for CD45/CD144 brown chromogen staining. Four ISET spots corresponding to 4ml whole blood were ICC stained and scored for each blood sample. The number of CTCs detected was extrapolated to the equivalent of 7.5ml for comparison with concurrent CellSearch analysis. ### CellSearch CTC analysis CellSave blood samples (7.5ml) were processed by the CellSearch system as described elsewhere. Briefly, CTCs are immunomagnetically separated from other blood components by EpCAM (epithelial cell adhesion molecule) antibody- conjugated beads and then stained for cytokeratins (CKs 8, 18 and 19) and CD45 in a fluorescent-based approach. CTCs are defined as CK+ CD45- cells over 4μm in diameter; exact CTC diameters were not measured as the system software does not have this functionality. # Results ## Early breast cancer patients To maximise the chance of finding CTCs, a relatively higher risk group of early breast cancer patients were recruited. At final pathology, 75% were node positive. Of the 16 patients, six had CTCs by CellSearch and all had CTCs by the ISET morphology-based assay at least one time-point. Using ISET, CTCs were detected in all 27 early breast cancer patient blood samples tested, whereas CTCs were only detected in 9 (35%) of the samples analysed by CellSearch. ISET CTC number was greater than CellSearch number in 26/27 samples. The median (range) number of CTCs per 7.5ml whole blood detected by CellSearch and ISET respectively were CellSearch: 0 (0–13) and ISET: 21 (2–62). The CellSearch system does not allow accurate cell size measuring so sizes of CTCs detected by each method were not compared. There was no correlation between number of CTCs detected by ISET and detected by CellSearch, either preoperatively, postoperatively or when all sample time- points were analysed together. Number of CTCs (ISET or CellSearch) did not correlate with tumour size, node positivity or timing of blood sampling (preoperation versus postoperation); however the limited sample size is acknowledged. The high number of CTCs found by ISET, independent of clinicopathological risk factors or CellSearch CTC numbers (although possibly reflecting an underpowered sample) raised the concern of false positive identification of CTCs by ISET. ## High CTC false positive numbers are detected in healthy volunteers by ISET morphology method ISET filter membranes generated from three sex-matched healthy volunteer blood samples were analysed alongside the early breast cancer samples as presumed negative controls. Surprisingly, high numbers of false positive CTCs were detected, including one sample with 99 ‘CTCs’/7.5ml blood. These presumed false positives could not be morphologically distinguished from CTCs detected in the early breast cancer samples by the consultant breast histopathologist. ## ISET healthy volunteer CTC false positives remain high in successive blood draws It was hypothesised that epithelial cells may be shed into the bloodstream during the first draw of venous blood collection, which could partly explain the high numbers of CTC-false positives seen in healthy volunteer samples. In three further sex-matched healthy volunteers the number of CTCs detected in first, second and third blood draws were compared. A trend towards lower ‘CTC' numbers in successive draws was observed, but in all samples ‘CTC’ number was unacceptably high, highlighting the high false positive rate of the morphology based ISET technique. ## Low numbers of ER+ CTCs are detected in metastatic breast cancer patients Given the apparent high false positive CTCs identified by ISET-morphology, a similar but non-comparable ERα-based multiplex immunocytochemistry staining assay was employed with the aim of increasing specificity of CTC identification in ER positive breast cancer patients. The methodology was performed on blood taken from three ER-positive metastatic breast cancer patients. The samples from each patient were taken at the same blood draw, and processed for ISET (EDTA) and CellSearch CTC analysis. ISET samples from patients identified as CTC positive on CellSearch were examined. ERα positive CTCs were identified by ISET, however the number of CTCs identified by this methodology was markedly lower than by CellSearch, even in a high-CTC population indicating that this assay would be inappropriate for the low numbers of CTCs found in early breast cancer sample analysis. # Discussion Using the ISET platform and an epithelial-marker independent immunocytochemical assay we detected higher number of putative CTCs in early breast cancer samples than CTCs detected by the EpCAM/cytokeratin dependent CellSearch. However, in an unexpected result, by the same cell morphology based technique we also found high numbers of false positive ‘CTCs’ in healthy volunteer blood samples. By employing oestrogen receptor alpha as a positive marker we developed an immunocytochemical assay which we applied to a small number of ER+ metastatic breast cancer samples, however this technique identified minimal numbers of CTCs despite evidence of their presence on CellSearch. Much of the previously published literature on the use of the ISET device in detecting supposed CTCs using morphological criteria has found higher numbers of CTCs than are detected by epithelial antigens across a range of cancer types, despite larger size criteria. These studies have identified ‘CTCs’ by a range of cell and nuclear morphological criteria unlike the standard epithelial-marker based approaches. Vona et al. (2004) in the first clinical article using the ISET technology described that CTCs/CTMs were detected in 53% of liver cancer samples by cytomorphological analysis but none were detected in 107 non-cancer patients, including many with other types of liver disease. However, the \>25μm cell diameter criterion they used is in our view inappropriate for breast cancer. This size would exclude most of the breast cancer CTCs we have encountered by CellSearch and the most commonly used breast cancer cell lines. Furthermore, an independent albeit smaller hepatocellular carcinoma study carried out in our laboratory concurrent to this work found no significant relationships between ISET-derived CTC number and clinical characteristics, raising doubts about the utility of identifying CTCs by morphology alone. In a subgroup of twenty metastatic breast cancer patients Farace et al. used a nuclear size cut-off of ≥16μm to define CTCs in ISET, finding 17/20 patients to be CTC-positive. However, no healthy subjects were examined in their study as controls. Hofman et al. in large NSCLC-focused clinical studies that included a breast cancer patient subset have developed their own ISET cytomorphology-based scoring method. Identified ‘circulating non-haematological cells’ (CNHCs) were further characterised as CNHCs with malignant features (CNHC-MFs, ‘CTCs’) if they exhibited four of the following criteria: anisonucleosis, nucleus size \>24μm, irregular nuclei, presence of tridimensional sheets, and a high nuclear- cytoplasmic ratio. CNHC-MFs/CTCs were identified in 43% malignant disease but only 5% of non-cancer disease patients. Although inter-observer agreement for detection of CNHCs was an impressive 100% (κ = 1) between three assessing cytopathologists, it is noteworthy that inter-observer agreement of CNHCs with benign features was relatively low (κ = 0.35). El-Hilibi et al. replicated their methodology in an alternative filtration based technology (ScreenCell) and determined that morphological criteria alone were inadequate to distinguish malignant from non-malignant cells. As CellSearch cell size criteria is \>4μm, compared to our ISET definition of ≥16μm, it might be expected that CellSearch would identify higher CTC numbers. However, like several other authors, we found higher numbers with ISET. Although many presume this is because of identification of non-epithelial antigen expressing CTCs, CTC false positivity is an acknowledged problem in filtration based technologies. This is supported by our high CTC identification in normal controls. ISET-based studies have used large cell/nuclear size criteria to avoid misidentification of endogenous nucleated blood cells at the expense of identifying smaller CTCs. Small CTCs (≤90 μm<sup>2</sup>) have been described in prostate cancer with small-nuclear CTCs being associated with visceral metastases, and therefore may represent an important subgroup where the ISET technology produces false negatives. There are several possible causes of the false positives seen in this study. It is possible that larger leukocytes may have been misidentified, especially monocytes which can reach up to 20μm in diameter, although the CD45 immunocytochemistry method employed has previously been shown to be effective at staining these cells. Another candidate is megakaryocytes, responsible for platelet production with a 37μm mean cell diameter however these are usually discounted due to their round and pale nucleus and high reported CD45 expression. Epithelial cells, which can be collected during peripheral venous blood collection by intradermal needle could also have been present on the filters and misidentified as CTCs. In addition, El-Hilibi et al. suggested circulating endothelial cells as a probable cause of misidentified CTCs in their study. Endothelial cells are released during venepuncture, with numbers shown to decline on subsequent draws. This prompted us to examine ‘CTC’ numbers in successive blood draws as shown in. Although there was a trend of lower false positives in each subsequent draw, the high numbers seen in the third draw indicates that this is not the sole cause of this phenomenon. Also, by using VE-Cadherin, CD144 as a negative stain, we have gone further than the studies in in actively identifying endothelial cells using a specific protein marker. Platelet endothelial cell adhesion molecule-1 (PECAM-1/CD31) as a further endothelial marker and CD61 as a marker for megakaryocytes may provide additional support in correctly identifying CTCs. However, approximately 90% of all putative CTCs we observed on stained ISET filters created by standard device operation were sucked into the 8μm pores, which plainly affected their cell and nuclear morphology. This factor undoubtedly contributed to the high rate of false positives seen. Based on these results, morphology based ISET CTC enumeration is not appropriate for early breast cancer. Oestrogen receptor alpha (ERα) expression in the primary tumour is routinely assessed in breast cancer clinical management. Clinical studies have previously examined ERα expression in cytokeratin/EpCAM positive CTCs, but to our knowledge we are the first to identify CTCs primarily by using ERα as a positive marker. The lower cell count found by ISET compared to CellSearch may reflect loss of ERα expression by CTCs. This is supported by Babayan et al. who stated that CTCs frequently lack ER expression in ER+ metastatic breast cancer. Atkas et al. (2011) have also reported discordance between CTC and tumour ER expression. However the lower CTC numbers reported by ISET compared to CellSearch when ERα expression is included may possibly be a result of ISET failing to capture smaller ‘true’ CTCs. This further highlights the one of the challenges of the ISET technology. In addition, the number of ISET ERα positive identified CTCs in the setting of metastatic disease was surprisingly low compared to ISET ‘CTCs’ identified by just morphology in the early breast cancer subgroup, as it would be expected that CTC counts would be higher in more advanced disease. Our finding of low CTC numbers in metastatic disease compared to early disease further highlights the likely presence of false positives with the morphology- only technique used in the early breast cancer group. One of the particular challenges we encountered with this methodology is the time required for blood analysis. To ensure the integrity of nucleated cells, whole blood samples in EDTA vacutainers must be diluted in Rarecells buffer within four hours of blood taking, presenting challenges in sample transport and analyst availability. Blood processing and filtration using the ISET technology is non-automated and requires an average of 60 minutes of technician time per sample. However, even more challenging, the images produced from each spot corresponding to 1ml whole blood (up to 10 spots per sample) require 30 minutes to view and score, highlighting the limitations of this technology in its current form. In summary, we investigated the use of the ISET technology to enumerate CTCs by a non-cytokeratin / epithelial marker technique in early breast cancer patients. The high false positive rates based on morphology alone, and the low numbers captured (presumed high false negative) by ERα plus morphology raises doubts on its utility in this low burden disease. This study highlights the need for caution in the use of all non-cytokeratin based CTC enumeration methodologies, with a necessity for extensive examination of normal control samples. However, considering the current postulated plasticity between epithelial and mesenchymal states of migratory cancer cells the detection of non-epithelial CTCs remains an important scientific and clinical goal. # Supporting information [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: CCK. Data curation: JC CCK. Formal analysis: CCK JC. Funding acquisition: CCK. Investigation: JC CCK KM. Methodology: CCK KM JC. Project administration: CCK JC. Resources: CCK KM. Software: KM. Supervision: CCK KM. Validation: SP KM CCK JC. Visualization: JC CCK. Writing – original draft: JC. Writing – review & editing: CCK JC SP KM.
# Introduction Dengue viruses (genus flavivirus, family Flaviviridae) are responsible for dengue fever and severe dengue (previously known as dengue hemorrhagic fever, DHF and dengue shock syndrome, DSS), they are some of the most important arthropod-borne viral diseases worldwide. Increasing incidences of transmission in urban and semi-urban areas in the tropical and sub-tropical countries have promoted dengue to a major public health issue with a hefty economic burden. Dengue epidemics are largely attributed to three factors: (i) Increased urbanization, (ii) global dissemination of the major mosquito vectors, Aedes aegypti and Aedes albopictus, and (iii) the interaction and evolution of the four DENV serotypes. Inadequate sustained vector control may partially explain unsuccessful disease prevention in developing countries however variables related to the viruses, the human host and the environment should also be considered. Vaccine development is currently underway as a strategy for controlling this worldwide health threat. Due to the complexity of DENV dynamics, prevailing serotypes and differences in distribution of viruses within different regions of the globe, serotype and lineage replacement events may play an important role in the evolution of DENV viruses and, consequently must be addressed during vaccine development. Gene genealogy-based DENV evolutionary studies helped characterize the genetic diversity and distribution of different serotypes/genotypes in space and time. Also, they are important for exploring the role of selection on particular dengue proteins. Given that serotype-specific neutralizing antibodies confer limited, if any, protection from subsequent infection by the other three serotypes, the appropriate choice of nucleotide residues in the variable genomic regions is critical for the design of effective tetravalent vaccines. Brazil is the South American country with the highest economic impact of dengue and also accounts for the majority of reported cases in the continent as evidenced by a 27 year-long study. The highest dengue incidence was among young adults from 2000 to 2007. However, in 2006 the highest incidence rate of severe dengue increased dramatically among children and remained high during severe epidemics in 2007 and 2008 in the state of Rio de Janeiro. These circumstances highlight the urgent need for continued studies on dengue molecular epidemiology in order to help determine appropriate policies and effective public health campaigns. The municipality of São José de Rio Preto (SJRP), one of the 645 municipalities within the State of São Paulo according to the Brazilian Institute of Geography and Statistics (Instituto Brasileiro de Geografia e Estatística, IBGE), has a tropical savanna climate (Köppen climate classification), with an annual rainfall of 1410 mm and mean annual temperature of 23.6°C, making it a favorable breeding ground for Aedes mosquitoe vectors. Our previous study on dengue in SJRP using a 399 bp-long portion of the NS5 gene, suggested that DENV-3 was introduced into the city in 2005. Since SJRP is one of the municipalities in São Paulo with the highest number of dengue cases reported yearly according to the Epidemiological Surveillance Center of the state (Centro de Vigilância epidemiológica Alexandre Vranjac, CVE). We chose to further characterize the isolated DENV-3 strains from SJRP from 2006 to 2007 to gain insights into the viral epidemiology that may help control disease. # Materials and Methods ## Area of study The city of SJRP is located in the northwestern region of the state of São Paulo, Brazil (20°49′11″ S and 49°22′46″ W), with a total area of 431 km<sup>2</sup> and an estimated population of 415,769 inhabitants in 2012 (Data obtained from the IBGE). The infestation by Aedes mosquitoes was reported for the first time in 1985 and the first autochthonous cases were reported in 1990, with the introduction of DENV-1 in 1996 (Chiaravalloti-Neto, Superintendence of Endemic Disease Control, SUCEN: personal communication). Unpublished reports by the Instituto Adolfo Lutz (Brazilian Public Health Laboratory) point to the introduction of DENV-2 in 1998. Dengue viruses have been endemic in the municipality for almost ten years. ## Sequencing genomes A total of 33 plasma samples were collected from dengue patients during the 2006 and 2007 outbreaks as described elsewhere. Viral RNA was extracted directly from plasma using the QIAamp Viral RNA mini kit (Qiagen). cDNA was synthetized in a 20 µl reverse transcription reaction containing: 1 µl of Superscript III Reverse Transcriptase (Invitrogen), 1 µl of random hexamers (50 ng/µl stock), 1 µL of specific reverse primer 5′AGAACCTGTTGATTCAACAGCAC3′ (10 µM stock) and 5 µl of template RNA. Viral cDNA were diluted to 800 µl in DEPC-treated water and 20 µl were used as template for a 96 specific PCR reactions. Each 10 µl PCR reaction contained: 3 µl of template, 0.03 µl of pfuUltra II polymerase 1 (5 U/µl) (Stratagene), 100 mM dNTPs (Applied Biosystems) and 4 µl of a mixture of forward and reverse primers (0.5 µM stock). Primers were synthesized with M13 sequence tags (forward primers with 5′GTAAAACGACGGCCAGT3′ and reverse primers with 5′CAGGAAACAGCTATGACC3′) so that PCR amplicons could be sequenced using universal M13 forward and reverse primers. Each PCR reaction produced 96 overlapping amplicons, each 500–900 nucleotides in length, which were subsequently sequenced bidirectionally using the Big Dye chemistry on ABI3730xl DNA sequencers (Applied Biosystems). The 33 obtained genomic sequences were submitted to the Broad Institute genome project and deposited in GenBank ( shows all GenBank accession numbers and name of all samples). (<http://www.broadinstitute.org/annotation/viral/Dengue/>). ## Phylogenetic Inference The data set includes the obtained 33 genomic sequences from SJRP that were aligned together with 33 DENV-3 worldwide genomic sequences deposited in GenBank using the Muscle 3.8.31 program. Visual inspection of the alignment and manual editing were done with the Se-Al v2.0 program (<http://tree.bio.ed.ac.uk/software/seal/>). A Bayesian maximum clade credibility tree was inferred from a set of plausible trees sampled at the stationary phase of four independent Markov Chain Monte Carlo (MCMC) runs with 40 million generations each using a general time reversible (GTR) model of nucleotide substitution, a gamma distributed rate variation and a proportion of invariable sites (GTR + Γ + I). A relaxed (uncorrelated lognormal) molecular clock and a Bayesian Skyline coalescent tree were used as priors. Analyses were done with Beast v1.6.1. Convergence of parameters was assessed using Tracer v1.5 program (<http://tree.bio.ed.ac.uk/software/tracer/>) until all parameters estimates showed Effective Sample Size (ESS) values over 100. ## Selection detection analysis The SJRP dataset was used in a selection analysis to determine whether amino acid positions were subject to negative or positive selection pressures. Comparisons were made with the dengue reference strain Philippines H87/57, following Barrero and Mistchenko general approach. Both genealogy-based, codon- site models Single Likelihood Ancestor Counting (SLAC) and the Random-Effects (REL) available in the HyPhy package ([www.datamonkey.org](http://www.datamonkey.org)) were used to estimate the nonsynonymous (dN) and synonymous (dS) rates of substitution. Polymorphisms were analyzed with the program DNASP v5 and subjected to the Tajima D statistic test, as proposed by Tajima to evaluate deviations from the neutral expectation of molecular evolution. ## Epitope and immunogenicity prediction The amino acid sequences coding for the capsid, envelope, NS1, NS2a and NS5 proteins were aligned (<http://multalin.toulouse.inra.fr/multalin/multalin.html>) and a consensus sequence of each protein was submitted to B (<http://www.cbs.dtu.dk/services/BepiPred/>) and T (<http://www.cbs.dtu.dk/services/NetCTL/>) epitope prediction algorithms. The Allele Frequency Net Database (<http://www.allelefrequencies.net/>) was used to select the HLA classes most predominant in the Southeast of Brazil and used to set up the NetCTL server. The mean value of the epitope propensity scores for each sequence was then classified and plotted according to its potential immunogenicity. # Results and Discussion ## Complete genome analyzes define distinct DENV3 lineages circulating in SJRP The availability of high throughput, parallel DNA sequencing allows for comprehensive information on population genetics at the complete viral genome level. In this study, we determined the sequence of 3390 codon-sites of the complete polyprotein from 33 DENV-3 strains isolated during the 2006 and 2007 outbreaks. The inferred phylogenetic relationship among DENV-3 strains was summarized in a Bayesian maximum credibility tree (MCC) shown in. The majority of Brazilian samples grouped with genotype III, as did most taxa from Latin America isolated since 1994,. Likewise, the Brazilian strain EF629370 grouped within genotype I, as previously described by Schreiber et al.. DENV-3 isolates, other than genotype III, have been reported in Colombia and in the Brazilian states of Minas Gerais and Rondonia, likely imported from Asia. The successive circulation of different genotypes is of public health concern given that distinct combinations could have different epidemic potentials and risk burdens. The samples from SJRP all fell within genotype III. The MCC tree indicates that SJRP sequences group with Brazilian isolates from different regions in highly supported clades. Clade A included most of SJRP isolates and is more closely related to an isolate from Acre in 2004 (EF629367_BR_2004). Clade B includes four SJRP isolates grouping with (i) a sample from Central West Brazil (FJ913015_BR_2001) and (ii) the Brazilian sample FJ898446_BR_2001 of unknown exactly provenance. Clade C includes two SJRP isolates from 2007 that cluster with a sequence from Rio de Janeiro (EF629369_BR_2002) sampled in 2002. Moreover, in clades A and B there was no temporal separation, suggesting that the 2007 isolates apparently originated from previously circulating 2006 lineages. These results suggest distinct virus introductions into SJRP with one lineage becoming more prevalent and experiencing in situ evolution (clade A). Co-circulation of multiple DENV-1 lineages has been documented in Asian and other Brazilian cities. ## Succession of DENV-3 genotypes in Brazil Our data also provides information on the movement of DENV-3 (genotype III) in South America, because sample FJ850079_BR_2002 obtained in the northern part of Brazil grouped with Colombian and Venezuelan sequences suggesting a plausible transmission route. DENV-3 (genotype III) was isolated for the first time in Brazil from an autochthonous case of DF in December 2000 from Nova Iguaçu, a city located in Rio de Janeiro (RJ) State, which is the second largest metropolitan area in the country, in 2001- 02, a large outbreak of DENV-3 occurred in the neighboring city of Rio de Janeiro. Following its introduction into RJ, the virus spread to the neighboring State of São Paulo (SP), and eventually to the municipality of SJRP. A previous introduction of DENV-3 took place in Brazil in Rondonia State (Amazon region), but sample EF629370_BR_2002 position in the MCC tree indicates it belongs to genotype V with its sister taxa coming from Asia. Nevertheless, most genotype III isolates from the north of Brazil grouped with the Caribbean strains (samples from Puerto Rico and Trinidad and Tobago), which suggests the existence of preferential routes of spread into Brazil, possibly facilitated by anthropogenic factors. ## The dynamics of viral change Sequence comparisons using the reference strain Philippines H87/57 (accession no. M93130) as an outgroup, revealed a total of 893 nucleotide changes, of which 132 corresponded to non-synonymous substitutions. The nucleotide diversity π (that is the average number of nucleotide differences per site between two sequences) within SJRP strains was 0.0052 ± 0.001. Nevertheless, this π value was one order of magnitude lower than that for the complete data set used for phylogenetic inference (n = 0.038 ± 2×10<sup>−5</sup>), which suggests a genetic bottleneck following the introduction into a new human population, with the virus experiencing rapid in situ evolution afterwards. This is a likely scenario given that we sampled from the first documented DENV-3 outbreak in a city where the local human population is expected to be largely susceptible to this serotype. Likewise, the average value of Tajimas' D for all genes was −1.823 (p\<0.05) indicating an excess of low frequency polymorphisms, which would be the case in viral populations expanding following a selective sweep. The Tajima's D value near −2 also implies the outcome of negative selection, most likely due to functional constraints imposed on the viral proteins. In agreement with this hypothesis, the overall dN/dS value for the entire polyprotein coding region was 0.068 (estimated 95% CI from 0.057 to 0.080) that implied purifying (stabilizing) selection. Zanotto et al. have argued that at a coarse-grain, the most common pressure acting on DENV is purifying selection with little evidence of recent adaptive evolution when comparing distantly related sequences in space and time, which would be expected due to synonymous change saturation at the serotype level. Moreover, given that arboviruses have adapted to diverse cell types it is generally considered that most mutations produced are likely to be deleterious. Accordingly, the per-site SLAC dN/dS analyses revealed that only eight codons were non-neutral and negatively- selected (p\<0.05). Nevertheless, the presence of low frequency polymorphisms could account for adaptive changes, such as the 10 sites under significant positive selection (Bayes factor \>50) that we found with the per-site REL dN/dS analyses. Selected sites were located in both structural (S) and nonstructural (NS) proteins. Among the four non-neutral sites within the envelope protein, codon 355 coding for A Proline located in the central and dimerization domains, was under negative selection suggesting its structural importance. Likewise, positively-selected codons 404 and 581 in the envelope are near sites associated with disulfide bonds, suggesting adaptive changes near structurally important residues. Interestingly, Barrero and Mistchenko also identified codons 581 and 893 (within the nonstructural NS1 gene) as positively-selected. King et al. detected a statistically significant positively-selected site in the NS1 gene of DENV-3 genotype II. They also detected selective pressure in the C and NS4B genes of genotype I and in the E and NS3 genes of genotype II. We also found additional positively-selected codon sites; 106 (Endoplasmatic reticulum anchor for the protein Capsid), 2191 (Lumenal domain of NS4A) and 3129 (catalytic domain of the RNA-directed RNA polymerase) that fell within domains that are important for membrane interactions/rearrangements and virus replication. Polymorphism analysis of individual genomic regions indicated that the NS5 and the envelope proteins varied the most with 24 and 22 replacement changes respectively, followed by NS1 (17), NS2A (11), NS4B (8) and Capsid (8). Changes found in the glycoprotein M, the AdoMet-MTase region and the NS3 serine protease were mostly conservative, while replacements in the capsid were less conservative. Twiddy et al. found codon sites 169 and 380 to be under positive selection in DENV-3, with the latter being located within the central and dimerization domains of the glycoprotein, which is involved in cell tropism. Likewise, Rodpothong and Auewarakul reported one codon site in the prM gene and twelve residues in the E gene to be under positive selection. In summary, our findings along with others, suggest that genes other than the envelope may experience adaptive changes following introduction into a susceptible population, which implies a possible role for either host-immunity interaction or vector adaptations that require further study. ## Short-term evolution of DENV-3 genotype III Studies documenting selectively driven evolution in DENV are limited and mostly available for serotype 2,, further studies to better characterize possible specific adaptive strategies of each serotype are needed. Therefore we looked at allele replacement in DENV 3 (genotype III). The allele replacement process, from a population genetics standpoint, has an explicit genealogic representation, since frequency increase of an allele will be manifested by appearing deeper as a synapomorphy in a viral genealogy. Therefore, a synapomorphy indicates a fixed allele in the clade defined by it. To clarify this, we coded each of the ten positively-selected residues as a set of terminal unordered character states, represented as a single capital letter. The most parsimonious reconstruction sets of changes at each internal state in the phylogeny were calculated with MacClade v4.07 using the MCC tree rooted with the reference strain Philippines H87/57. The reconstructed changes are illustrated in. They indicate that positively-selected codons (404, 449, 581, 2191, 2879 and 3129) mapped to internal nodes as synapomorphies defining the genotype III. For example, codon 106 Met to Val, (within Protein C, the most non-conservative region in our study) characterized the SJRP clade A and codon 404 Pro to Leu characterized the SJRP clade B. Codon 893 appeared as an autapomorphy for the reference strain H87/57 (M93130). The synapomorphic changes differentiating clades A and B that we observed could reflect some aspects of the intra-genotypic interactions exposed to the human population. Mondini et al. described the pattern of spread of two DENV lineages circulating in SJRP, Brazil, during 2006. Based on 82 NS5 sequences, the phylogenetic analysis indicated that all samples were of DENV-3 and related to strains circulating in Martinique during 2000–2001. We have considerably extended that analysis by further showing that different lineages, closely relate to different Brazilian isolates, were co-circulating while experiencing distinct adaptive changes as indicated by their positively-selected synapomorphies. This finding is of relevance, since in several other viral systems, positively-selected changes are adaptive in nature and indicative of adaptive changes imposed by vectors, animal reservoirs or human hosts (i.e., immune escape and herd immunity). When we observed that DENV-3 sequences, previously clustered in clades A, B and C, we conducted in silico analysis to verify the influence of amino acids substitutions in the immunogenicity of capsid, envelope, NS1, NS2a and NS5 proteins. We found that amino acid substitutions diminished the B/T epitope propensity scores mainly in the DENV-3 sequences clustered in clades B and C. These analyses suggest that amino acid substitutions in sequences grouped in clade A increases its immunogenicity. As a consequence, DENV-3 strains grouped in clade A should expose more antigenic epitopes and more potentially recognized/accessed by the human immune system. We expected that the most immunogenic strains would be negatively selected in a given population. In this particular study we found exactly the opposite. Thus, one can hypothesize that the ADE phenomenon, which would be responsible for an increased pathogenicity and replication in host, can also affect the viral dynamics in a population level. We wonder whether this increased immunogenicity of clade A can induce an antibody dependent enhancement (ADE)-like response in the individual leading to high titers of the virus, which can facilitate the viral transmission to the mosquitoes and, consequently, increase this strain basic reproductive rate. Our findings are of relevance given that ADE may be manifested by challenges imposed by increasing viral immunogenic diversity. While most studies document the broad geographic expansion of dengue, urban outbreaks dynamics deserves investigation because they contribute to a better understanding of the genetic diversity of strains during local transmission; knowledge that could be exploited for antiviral and vaccine development. In our analyses of 33 complete polyprotein-coding regions we revealed distinct lineages and detected polymorphisms sites that correlated with changes in the immunogenicity of several epitopes. The study provides fine-grain information about the molecular epidemiology of dengue infection. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AM PMZ MLN IB. Performed the experiments: DS AM CVA PMZ CECS. Analyzed the data: PMZ IB MLN CVA CECS. Contributed reagents/materials/analysis tools: IB PMZ CECS. Wrote the paper: CVA AM DS MN PMZ CECS.
# Introduction With multidisciplinary professionals, diverse and complex medical equipment, vulnerable patients, time pressures, and extremely high tension, the operating room (OR) environment is susceptible to errors. Some major safety problems in the OR include addressing incorrect surgical site/patient/procedure, retained surgical items, medication errors, bedsores, hypothermia, burns, inadequate emergency responses, and improperly reprocessing surgical devices. Patient safety accidents related to surgery require particular precautions, as they can induce serious and irreversible injuries. Hence, the Joint Commission on Accreditation of Healthcare Organization (JCAHO) stressed the importance of teamwork, continuous quality control, smooth communication, and information sharing between medical professionals to ensure surgical patients’ safety. Furthermore, the Association of periOperative Registered Nurses (AORN) recommended quickly streamlining and standardizing work to detect and correct errors that occur during surgery. Despite such efforts, accidents continue to occur with surgical patients. According to a systematic literature review of adverse events in hospitals, surgery-related accidents accounted for 39.6%, the highest proportion of all such events. In Korea, surgery-related cases accounted for the highest proportion (35.1%) of all medical dispute claims filed between 2012 and 2016 and are gradually increasing. Therefore, surgical patients’ safety is of utmost importance. ## Literature review The theory of planned behavior (TPB) describes individual-level predictors of actions. This theory states that individuals’ conduct consists of their attitudes toward behavior, subjective norms, perceived behavioral control, and behavioral intention. The TPB is widely used not only in social sciences but also in various healthcare fields, as it effectively predicts individuals’ behavior despite involving only a few simple constructs. However, no studies have applied TPB to patient safety management activities. Accordingly, we applied Ajzen’s TPB to establish a model for patient safety management activities in the OR. Human errors can be viewed at the individual or system level. System-level human errors are lapses in safety behaviors attributed to conditions of the work environment, which cannot be altered by an individual. We need to understand how systems, which include organizational culture and policies, interact with individuals. Thus, patient safety management activities in the OR should be examined considering both individual and organizational factors, given that social behaviors result from their interactions. The most frequently examined organizational factors related to patient safety are job and systemic factors such as the safety management system. Job factors include excessive work demands and job complexity, which increase the physical and cognitive burdens on healthcare professionals, thereby decreasing their ability to engage in safety management activities. The safety management system includes safety training, participation in safety policy, management supervision, communication, and feedback. Therefore, when the level of safety management system is insufficient, accurate information on safety is not delivered, and education and management are neglected, thus resulting in lower awareness and performance of patient safety management activities. Therefore, we developed a hypothetical model that encompasses individual and system dimensions of safety management activities in the OR by adding organizational factors (i.e., job factors and safety management system) to Ajzen’s TPB. Structural modeling studies investigating factors related to operating room safety management activities will be useful in developing effective strategies to enhance patient safety management activities of OR nurses. ## Study aim The objectives of this study were to develop a structural model for patient safety management activities, identify the factors influencing organizational and individual dimensions that promote patient safety management activities, and suggest effective intervention plans. # Materials and methods ## Design A cross-sectional research design was used. A hypothetical model was developed based on Ajzen’s TPB. Job factors and the safety management system were used as organizational factors, and attitude, subjective norms, perceived behavioral control, behavioral intention, and patient safety management activities were used as individual factors. ## Participants and data collection Data was collected from August 1 to October 31, 2017, using self-report questionnaires. The recommended sample size in the structural equation model was 10–20 per observation variable. The expected number of observation variables was 32; thus, 320 to 640 participants were required. The questionnaires were distributed by convenience sampling to 360 perioperative nurses in 12 general hospitals in the Republic of Korea, and 347 questionnaires were returned (response rate = 96.4%). Questionnaires that were missing \>10% of responses were excluded (17 questionnaires). The remaining questionnaires were processed with mean substitution. Consequently, 330 questionnaires were included in the final analyses. ## Instruments ### Job factors The Job Content Questionnaire developed by Karasek et al. is a commonly used instrument to assess organizational job factors, and its validity and reliability were verified in a previous study by Song, which examined job factors for Korean nurses. Therefore, in this study, the instrument modified by Song for perioperative nurses was used. As for job demand, each of the 10 items in the questionnaire was measured on a five-point Likert scale. Cronbach’s alpha was.87 in both Song’s study and the current study. ### Safety management system Safety management system was evaluated using an instrument developed by Vredenburgh and translated and adapted for use in the OR by Song and Jang. This included management supervision, communication and feedback, and participation system. The 9 items in the scale were measured using a five-point Likert scale. Cronbach’s alpha was.73 in Song and Jang’s study and.82 in the current study. ### Attitude, subjective norms, and perceived behavioral control A 12-item scale was developed based on Ajzen’s study and a scale developed by Moon and Song for hospital nurses. Attitude was measured with three questions regarding positive or negative feelings about certain behaviors. The subjective norm is the perceived social pressure imposed on conduct, and was measured using five questions. Perceived behavioral control is an individual’s confidence or controllability of a behavior, and was measured with four questions. Each item was rated on a seven-point Likert scale. Cronbach’s alpha for attitude was.69 in Moon and Song’s study and.77 in the current study. Cronbach’s alpha for subjective norms was.76 in Moon and Song’s study and.91 in the current study. Similarly, for perceived behavioral control, Cronbach’s alpha was.81 in Moon and Song’s study and.88 in the current study. ### Behavioral intention A 4-item scale was developed based on Ajzen’s study and a scale developed by Moon and Song for hospital nurses. Each item was rated on a 7-point Likert scale for willingness, planning, and thinking. Cronbach’s alpha for behavioral intention was.80 in Moon and Song’s study and.90 in the current study. ### Patient safety management activities Safety management activities were measured using an instrument developed by Kim and Jeong based on six international patient safety goals. The scale consisted of items pertaining to infection management, specimen management, patient identification, medical equipment and product management, surgical counting, and injury prevention. The 36 items were measured on a five-point Likert scale. Cronbach’s alpha was.95 in Kim and Jeong’s study and.94 in the current study. ## Data analyses Collected data were analyzed using SPSS 25.0 (SPSS; IBM, Armonk, NY, USA) and AMOS 21.0 (SPSS Amos; IBM, Chicago, IL, USA). Participants’ general characteristics were analyzed using descriptive statistics. Data normality was tested using skewness and kurtosis. Correlations between measurement variables were analyzed using Pearson’s correlation coefficient. The hypothetical model’s goodness of fit was tested using the following: χ<sup>2</sup> statistics, standard χ<sup>2</sup> (CMIN/DF, Normed χ<sup>2</sup>), standardized root mean square residual, goodness-of-fit index, the normed fit index, the Tucker-Lewis index, the comparative fit index, and root mean square error of approximation. A covariance structure analysis was performed using the maximum likelihood method to determine the model’s goodness of fit and test the hypotheses. The statistical significance of the direct, indirect, and total effects of the model was analyzed via bootstrapping. All statistical analyses with p \<.05 were considered statistically significant. ## Ethical considerations The study was approved by the institutional review board at Konyang University Hospital (approval number: KYUH 2017-07-011) and conducted in accordance with the Declaration of Helsinki. Questionnaires were placed in the nurses’ break rooms for nurses to complete voluntarily. Completed questionnaires were then placed in a collection box in the break room. All participants were provided with an information sheet explaining the study purpose and method, management of collected data, protection of personal information, and participants’ right to withdraw from the study. Participants who provided written consent were enrolled in the study. # Results ## Participants’ general characteristics The valid response rate was 91.7% (N = 330). Participants’ demographic characteristics are presented in. There was a difference in the mean of patient safety management activities between the group with less than 5 years of career experience and the group with more than 5 years of career experience (F = 5.98, p =.001). No other statistical significance of the small group mean according to general characteristics was confirmed. ## Verification of normality and validity of the measurement variables The absolute value of skewness was between 0.40–1.35 and the absolute value of kurtosis was between 0.17–2.03. As the absolute values of skewness and kurtosis did not exceed 3 or 10, respectively, the data satisfied univariate normality. The correlation coefficients for the measurement variables did not exceed.80; therefore, multicollinearity was not a concern. Discriminant validity was established as the average variance extracted for each observed variable was greater than its coefficient of determination (r<sup>2</sup>). ## Confirmatory factor analysis of the conceptual model Although the goodness-of-fit and normed fit indices were slightly lower than the required values, we determined that the model showed a good fit considering the other indices. Eight out of the eleven paths were significant. The safety management system showed a significant path to attitude, with an explanatory power of 14.5%. Safety management system showed a significant path to subjective norms, with an explanatory power of 34.5%. Job factors and safety management system showed significant paths to perceived behavioral control, with an explanatory power of 10.1%. Attitude, subjective norms, and perceived behavioral control showed significant paths to behavioral intention, with an explanatory power of 50.0%. Behavioral intention showed a significant path to patient safety management activities, with an explanatory power of 38.1%. shows the direct and indirect relevance of the hypothesis model. The safety management system has a direct influence on attitudes and subjective norms. Job factors and the safety management system had a direct influence on perceived behavioral control. Attitude, subjective norms, and perceived behavioral control were directly related to behavioral intention. Perceived behavioral control showed indirect relevance to patient safety management activities, and behavioral intention showed direct relevance to patient safety management activities. # Discussion The modified TPB model explained patient safety management activities in the OR relatively well. The explanatory power of the model for behavior was high; adding organizational factors as antecedents to personal factors further increased the model’s explanatory power. The higher the job demands, the lower the perceived behavioral control of patient safety management activities. The physical and cognitive burdens of excessive job demands undermine one’s problem-solving abilities related to safety performance and are, therefore, associated with increased accident occurrence. Regarding organizational factors, higher scores for safety management system were associated with more positive attitudes, stronger subjective norms, and perceived behavioral control in patient safety management activities. Organizational factors, including management values, the safety system, safety practice, education and training, and communication, could impact individual factors such as safety motivation and knowledge. Further, stronger behavioral intention regarding patient safety management activities was associated with more positive attitudes toward patient safety management activities, stronger subjective norms, and greater perceived behavioral control. These findings were similar to those of previous studies based on the TPB, in which attitude, subjective norms, and perceived behavioral control predicted behavioral intention. These results confirmed that the modified TPB is a valid model for explaining patient safety management activities in the OR. Perceived behavioral control was the most influential factor on the behavioral intention for patient safety management activities in the operating room, followed by subjective norms and attitudes. These results were contrary to a study examining alcohol abstinence in patients with chronic liver disease and one that examined hidden agendas in the use of mental health services for depression. In these two studies, attitude appeared to have the greatest influence on behavior. In predicting behavioral intention, the influence of attitude, subjective norms, and perceived behavioral control could vary depending on the extent to which behavior and situations are controlled by an individual. In previous studies, factors influencing behavioral intentions showed varying results depending on the characteristics and type of behavior. Unlike individual behavior, perceived behavioral control and subjective norms could be key factors in behavioral intention related to social behavior; these factors are difficult to control through an individual’s will alone. Moreover, the results showed that perceived behavioral control did not directly relate to patient safety management. Thus, perceived behavioral control may not be directly related to behavior if one’s perception is not consistent with actual behavioral control; hence, the relationship between behavioral control and behavior is indicated by dotted lines in the TPB. A meta-analysis of the TPB also showed conflicting results for perceived behavioral control depending on the type of behavior involved. Nurses expect to be able to control patient safety management activities, but they may not be able to do so if there are uncontrollable environmental factors such as a heavy workload and a lack of necessary supplies. Conversely, since they may not deliberately perform safety management activities as a result of excessive trust in their skills or reckless behavior, an analysis of the specific path between attitudes, subjective norms, and perceived behavior control needs to be researched. Organizational actions required to improve patient safety management activities in the OR include reducing job demands and enhancing the organizational safety management system. Individual actions required include fostering a positive attitude and increasing one’s behavioral intention by strengthening subjective norms and perceived behavioral control. Hospitals should recognize that individuals comprise the organization and devise strategies accordingly to improve patient safety management activities. Specific and practical education tailored to the conditions of the OR should be provided, and standardization of the OR patient safety management protocol and information management are necessary to enhance the efficiency of communication systems. ## Limitations The data for this study were collected from perioperative nurses working in large hospitals; therefore, future studies should include nurses from small and medium-sized hospitals with varying OR sizes and types of work. In addition, since the research was conducted using self-reported subjective data, casual effects between variables could not be confirmed. Therefore, further research using objective data on factors such as reporting of patient safety accidents (near miss, adverse events, sentinel events) is necessary. Moreover, we established a model based on a modified TPB to explain patient safety management activities in the OR; additional studies that examine other factors associated with patient safety management activities in the OR are needed. # Conclusions Crucial influencing factors on patient safety management activities in the OR were the safety management system, subjective norms, perceived behavior control, and behavior intention. Therefore, it is necessary to prepare hospital-level support and nursing policies to reinforce these factors. Organizations as well as individuals and medical staff should work together to strengthen OR patient safety management activities. We thank the operating room nurses who participated in this study. 10.1371/journal.pone.0252648.r001 Decision Letter 0 Delcea Camelia Academic Editor 2021 Camelia Delcea This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 15 Feb 2021 PONE-D-20-31121 Perioperative patient safety management activities: A modified theory of planned behavior PLOS ONE Dear Dr. Jeong, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. The paper addresses an interesting topic. In the revised version of the paper please consider the reviewers' comments listed in the following. Additionally, please consider adding a similar analysis in which you are considering smaller groups created based on the general characteristics in Table 1 and discuss whether there are differences when certain groups are analyzed compared to the whole sample. Please submit your revised manuscript by Apr 01 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Camelia Delcea Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please provide the names of all the participating hospitals, and a further details regarding how these hospitals were selected. Furthermore, Please include additional information regarding the survey or questionnaire used in the study and ensure that you have provided sufficient details that others could replicate the analyses. For instance, if you developed a questionnaire as part of this study and it is not under a copyright more restrictive than CC-BY, please include a copy, in both the original language and English, as Supporting Information. 3\. In the submission please specify whether IRB approval was obtained from all participating hospitals. 4.We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. 5\. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. Please see the following video for instructions on linking an ORCID iD to your Editorial Manager account: <https://www.youtube.com/watch?v=_xcclfuvtxQ> \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: No \\\\\\\\\\ 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \\\\\\\\\\ 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The research question is interesting and cogent, and the research is based on a sound literature ground concerning safety behaviors. I think the research is worth publishing, but I address some aspects that could be clarified by the authors. 1\. Line 83: the authors make reference to the safety management system. It si not clearly explained how these aspects could hinder safety. The authors do not explicitly refer to organizational culture factors, like the kind of leadership, the blame culture, the safety climate (which are implicit in the safety management tool used in the research). 2\. Line 125: The authors should explain why they chose the Job Content Questionnaire developed by Karasek et al. \\\n and no other tools. What is the rationale of this choice? 3\. Line 138: The 12-item scale is based on an unpublished study by Moon. The authors should provide evidence of the validity of the scale 4\. Line 155. Also, the safety management activity instrument, developed by Jeong, is an unpublished research and the authors should provide evidence of its validity 5\. Line 279: The issue of behavioral control is controversial. The internal locus of control is generally considered to be a better predictor of safe performance, but in extreme situations it could also represent an excessive trust in one’s own skills and a deliberate exposure to reckless actions. The authors could provide a deeper explanation of these results. Line 302: the authors mention among the limitations of the research the fact that it was based on self-report data. Self-report tools may be biased by social desirability, especially when they are related to errors, violations, and safety issues. The authors mention objective measurements such as observational surveys, however, I think it could have been useful to add other kind of objective data to the analysis, for instance concerning the rate of adverse events, injuries, near misses, etc. Reviewer \#2: The current article attempts to tackle the important topic of OR safety from the view of OR RNs. The authors provided OR RNs with surveys and matched the results to the TPB model. However the authors' concluded cause and effect from the survey data, rather than acknowledging that survey correlations cannot imply causation and TPB model fit. The authors need re-structure their conclusions to reflect this, rather than assume the survey results fit the TPB model. Essentially, their conclusions outreach the survey data results. \\\\\\\\\\ 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: Yes: Fabrizio Bracco Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0252648.r002 Author response to Decision Letter 0 9 May 2021 Response to Reviewer’s Comments Reviewer(s)' Comments to Author: Reviewer \#1: The research question is interesting and cogent, and the research is based on a sound literature ground concerning safety behaviors. I think the research is worth publishing, but I address some aspects that could be clarified by the authors. 1\. Line 83: the authors make reference to the safety management system. It si not clearly explained how these aspects could hinder safety. The authors do not explicitly refer to organizational culture factors, like the kind of leadership, the blame culture, the safety climate (which are implicit in the safety management tool used in the research). Reply: Thank you for your comments. Corrected the sentence. It has been described how the safety management system relates to safety management activities. � Line 90: The safety management system includes training and participation in safety policy, management supervision, communication, and feedback. Therefore, when the level of safety management system is insufficient, accurate information on safety is not delivered, and education and management are neglected, thus resulting in lower awareness and performance of patient safety management activities \[15-17\]. 2\. Line 125: The authors should explain why they chose the Job Content Questionnaire developed by Karasek et al. and not other tools. What is the rationale of this choice? Reply: Thank you for your comments. Corrected the sentence. It was used in a study on the job factors of Korean nurses in previous studies, and the validity and reliability were verified and used in this study. In Song's study, the fitness index of the confirmatory factor analysis for job factors was χ2=88.949 (df=5, p\<.001), TLI=.84, CFI=.92, RMSEA=.21, and SRMR=.05. � Line 131: The Job Content Questionnaire developed by Karasek et al. \[20\] is a commonly used instrument to assess organizational job factors, and its validity and reliability were verified in a previous study by Song \[21\], which examined job factors for Korean nurses. Therefore, in this study, the instrument modified by Song \[21\] for perioperative nurses was used. 3\. Line 138: The 12-item scale is based on an unpublished study by Moon. The authors should provide evidence of the validity of the scale. Reply: Thank you for your comments. Corrected the sentence. In the paper published by Moon and Song, the validity of the tool was verified, so the references were revised. � Line 146: A 12-item scale was developed based on Ajzen's \[23\] study and a scale developed by Moon and Song \[24\] for hospital nurses. 4\. Line 155. Also the safety management activity instrument, developed by Jeong, is an unpublished research and the authors should provide evidence of its validity Reply: Thank you for your comments. Corrected the sentence. The reference was changed to a published article. � Line 164: Safety management activities were measured using an instrument developed by Kim and Jeong \[25\] based on six international patient safety goals \[26\]. 5\. Line 279: The issue of behavioral control is controversial. The internal locus of control is generally considered to be a better predictor of safe performance, but in extreme situations it could also represent an excessive trust in one’s own skills and a deliberate exposure to reckless actions. The authors could provide a deeper explanation of these results. Reply: Thank you for your comments. Additional discussion was written. � Line 294: Nurses expect to be able to control patient safety management activities, but they may not be able to do so if there are uncontrollable environmental factors such as a heavy workload and a lack of necessary supplies. Conversely, since they may not deliberately perform safety management activities as a result of excessive trust in their skills or reckless behavior, an analysis of the specific path between attitudes, subjective norms, and perceived behavior control needs to be researched. 6\. Line 302: the authors mention among the limitations of the research the fact that it was based on self-report data. Self-report tools may be biased by social desirability, especially when they are related to errors, violations, and safety issues. The authors mention objective measurements such as observational surveys, however, I think it could have be useful to add other kind of objective data to the analysis, for instance concerning the rate of adverse events, injuries, near misses, etc. Reply: Thank you for your comments. Corrected the sentence. � Line 313: In addition, since the research was conducted using self-reported subjective data, casual effects between variables could not be confirmed. Therefore, further research using objective data on factors such as reporting of patient safety accidents (near miss, adverse events, sentinel events) is necessary. 7\. The paper addresses an interesting topic. In the revised version of the paper please consider the reviewers' comments listed in the following. Additionally, please consider adding a similar analysis in which you are considering smaller groups created based on the general characteristics in Table 1 and discuss whether there are differences when certain groups are analyzed compared to the whole sample. Reply: Thank you for your comments. Further analysis and presented in the results and tables 1. � Line 198: There was a difference in the mean of patient safety management activities between the group with less than 5 years of career experience and the group with more than 5 years of career experience (F=5.98, p=.001). No other statistical significance of the small group mean according to general characteristics was confirmed (Table 1). Reviewer \#2: 1\. The current article attempts to tackle the important topic of OR safety from the view of OR RNs. The authors provided OR RNs with surveys and matched the results to the TPB model. However the authors' concluded cause and effect from the survey data, rather than acknowledging that survey correlations cannot imply causation and TPB model fit. The authors need re-structure their conclusions to reflect this, rather than assume the survey results fit the TPB model. Essentially, their conclusions outreach the survey data results. Reply: Thank you for your comments. The comment part was added as a limitation of the study, and the discussion and conclusion parts were entirely revised and described. � Line 276: Perceived behavioral control was the most influential factor on the behavioral intention for patient safety management activities in the operating room, followed by subjective norms and attitudes. These results were contrary to a study examining alcohol abstinence in patients with chronic liver disease \[30\] and one that examined hidden agendas in the use of mental health services for depression \[31\]. In these two studies, attitude appeared to have the greatest influence on behavior. In predicting behavioral intention, the influence of attitude, subjective norms, and perceived behavioral control could vary depending on the extent to which behavior and situations are controlled by an individual \[32\]. In previous studies, factors influencing behavioral intentions showed varying results depending on the characteristics and type of behavior \[33\]. Unlike individual behavior, perceived behavioral control and subjective norms could be key factors in behavioral intention related to social behavior; these factors are difficult to control through an individual’s will alone. � Line 289: Moreover, the results showed that perceived behavioral control did not directly relate to patient safety management. Thus, perceived behavioral control may not be directly related to behavior if one's perception is not consistent with actual behavioral control; hence, the relationship between behavioral control and behavior is indicated by dotted lines in the TPB \[32\]. A meta-analysis of the TPB also showed conflicting results for perceived behavioral control depending on the type of behavior involved \[28\]. Nurses expect to be able to control patient safety management activities, but they may not be able to do so if there are uncontrollable environmental factors such as a heavy workload and a lack of necessary supplies. Conversely, since they may not deliberately perform safety management activities as a result of excessive trust in their skills or reckless behavior, an analysis of the specific path between attitudes, subjective norms, and perceived behavior control needs to be researched. � Line 313: In addition, since the research was conducted using self-reported subjective data, casual effects between variables could not be confirmed. Therefore, further research using objective data on factors such as reporting of patient safety accidents (near miss, adverse events, sentinel events) is necessary. � Line 322: Crucial influencing factors on patient safety management activities in the OR were the safety management system, subjective norms, perceived behavior control, and behavior intention. Therefore, it is necessary to prepare hospital-level support and nursing policies to reinforce these factors. Organizations as well as individuals and medical staff should work together to strengthen OR patient safety management activities. 10.1371/journal.pone.0252648.r003 Decision Letter 1 Delcea Camelia Academic Editor 2021 Camelia Delcea This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 20 May 2021 Perioperative patient safety management activities: A modified theory of planned behavior PONE-D-20-31121R1 Dear Dr. Jeong, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Camelia Delcea Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \\\\\\\\\\ 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: (No Response) \\\\\\\\\\ 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: (No Response) \\\\\\\\\\ 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: (No Response) \\\\\\\\\\ 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: (No Response) \\\\\\\\\\ 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: (No Response) \\\\\\\\\\ 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: Yes: Fabrizio Bracco 10.1371/journal.pone.0252648.r004 Acceptance letter Delcea Camelia Academic Editor 2021 Camelia Delcea This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 18 Jun 2021 PONE-D-20-31121R1 Perioperative patient safety management activities: A modified theory of planned behavior Dear Dr. Jeong: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Camelia Delcea Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.
# Introduction The first completed eukaryotic genome sequence was that of the budding yeast Saccharomyces cerevisiae strain S288C, completed through the effort of a worldwide sequencing consortium. Since that time, many S. cerevisiae genomes have been sequenced, encompassing a wide variety of commercial and laboratory strains, as well as wild isolates. With next-generation sequencing methods becoming ubiquitous, whole genomes are now being analyzed en masse. This has led to interesting work on the relationship between genotype and phenotype. For example, in the studies of the adaptive evolution of freezing tolerance, Fay et al. determined that an isolate taken from the soil beneath an oak tree in a natural woodland area in southern Pennsylvania (YPS163) is freeze tolerant, a phenotype associated with its increased expression of aquaporin AQY2. Similarly, Doniger et al. studied an Italian vineyard isolate (M22), and confirmed the presence of a reciprocal translocation between chromosomes VIII and XVI (relative to the laboratory strain S288C); this translocation is common in wine strains, and results in increased sulfite resistance, an adaptive trait for the yeast since vineyards are routinely dusted with elemental sulfur as a fungicide. Argueso et al. determined that a widely used Brazilian bioethanol strain that is resistant to heat and oxidative stress contains well-characterized alleles at several genes known to be linked with thermotolerance and fermentation performance. Novo et al. studied a well-known commercial winemaking strain (EC1118) and found three unique regions on three different chromosomes containing 34 genes related to key fermentation characteristics, such as metabolism and transport of sugar or nitrogen. They also noted that \>100 genes in the reference strain S288C are absent from the EC1118 genome. Comparative genomics work has revealed patterns of genetic variation including single nucleotide polymorphisms, and large-scale insertions and deletions in several wine and ale strain genomes. Functional genomic analysis has also been undertaken in a saké yeast strain (K7), which has two large inversions and dozens of novel open reading frames (ORFs) compared to reference strain S288C. Genomic variation in S. cerevisiae genomes, such as single-nucleotide polymorphisms (SNPs), small insertions/deletions (indels), and structural variation, have been investigated. Despite much effort, the association of genomic variations with phenotype and functional annotations remains challenging, partly due to difficulties gaining accurate phenotypic information and obtaining genome sequences at high quality. Fortunately, because of its status as both a model organism and as an important industrial organism, many different S. cerevisiae strains have been intensively studied at the phenotypic, genetic and genomic levels and the resulting information has been extensively curated in the Saccharomyces Genome Database (SGD) (See). Genomics studies using the standard S288C yeast reference genome have produced many informative and interesting results. However, our understanding of yeast genetics and systems biology will widen and deepen if we can integrate new data into a pan-genome model to account for a greater proportion of the genetic and phenotypic variation exhibited by the global population of S. cerevisiae. A pan-genome is defined as the set of all genes in a species, and can be constructed from the union of gene sets over all S. cerevisiae strains. The development and rapid expansion in the use of Next-Generation Sequencing (NGS) technologies has created an increase in the volume of high-throughput data. The expanding use of targeted approaches such as DNA-seq, RNA-seq, and ChIP-seq has also increased the types of data available. These developments allow questions and assumptions in population genetics and evolutionary biology to be addressed directly, but fulfilling the potential of these approaches depends on accurate and reproducible data analysis. Many computational methods are designed to handle DNA-seq data for assembly, annotation, and variation detection. However setting up a pipeline for these computational analyses is a non-trivial task. Existing analysis software often produces incongruent results even when addressing the same problems with the same data. Pipelines for the pan-genome analysis of bacteria have been developed such as PGAP, but these are not suitable for eukaryotic genomes, even for unicellular eukaryotes such as yeasts, which exhibit more complex gene structures and non-genic regions than prokaryotes. The frenetic pace at which new genomes are being sequenced has laid the groundwork for great steps forward in our understanding of chromosomal evolution and the extreme variability of the eukaryotic genome. However, the sheer volume of data presents a clear challenge because it has been, and is being, produced by different research groups using different techniques for sequence assembly, feature annotation, and gene functional analysis. Before we can realize the full potential of these new data and derive maximal benefit from the ever-increasing number of sequenced genomes generated by disparate groups, we must address the pressing need for a common standardized approach to genomic data analysis. To that end, we report here the development of AGAPE: an Automated Genome Analysis PipelinE for S. cerevisiae. The pipeline includes assembly, annotation, and variation-calling steps for the genome sequence of a given strain and generates integrative analyses among strains. We have sequenced, or re-sequenced, and analyzed the genomes of 25 S. cerevisiae strains that are commonly used in yeast laboratory research to initiate analysis of the yeast pan-genome using AGAPE. Simple eukaryotes such as fungi evolve rapidly and show presence or absence of genes in different populations within a single species. Our initial work can accelerate the establishment of the yeast pan-genome using AGAPE as more genome sequences are released; assembly and annotation data from new strains can be used to continuously update the pan-genome, and the integrative analysis steps of the pipeline can be easily performed using the updated pan-genome. AGAPE can also be useful for biologists with limited bioinformatics expertise who can conduct computational analyses with their eukaryotic genomic data. Replacement software for a specific computational step can be easily plugged into the pipeline. All analyses, data, and the software pipeline reported here are freely available online, see. # Materials and Methods ## Strain sequences and genome assemblies Twenty-five strains were selected for analysis based in part on their frequent use in genetic research. The libraries were sequenced using Illumina HiSeq 2000, resulting in paired-end reads of 101 nucleotides each. We sequenced the libraries to high coverage ranging from 60- to 330-fold. Reads with low quality or ambiguous bases were discarded using the error correction program SGA (command line ‘sga correct-k 41—discard—learn’ version 0.9.35). In this error correction step, on average 2–3% of the raw reads were removed. Note that the preprocessing step before running the assembler program is important for assembly quality control. The filtered reads were assembled to contigs using the de novo assembler program ABySS (command line ‘abyss-pe aligner = map k = 41’, version 1.3.4). The resulting contigs were extended to scaffolds using an SGA scaffolding pipeline (command lines ‘sga-align; sga-bam2de.pl-n 5-m 100-mina 95; sga-astat.py-m 100; sga scaffold-m 100—pe; sga scaffold2fasta-m 100—write- unplaced—use-overlap’). If desired, alternative parameters can be specified for each program in this assembly process. ## Gene annotations Predictions of protein-coding genes (ORFs) were made using a combination of two methods: a homology-based approach and ab initio prediction. For the first approach we used the Chain and Net program to find all intervals in each strain that are homologous to the reference genome. Next, for each matching interval we ran a modified version of the annotation utility program included in CHAP2 (The Cluster History Analysis Package). CHAP2 uses LASTZ for aligning the matching regions to the reference ORF sequences. We used the thoroughly curated SGD reference annotations, and predicted gene structures of each homologous ORF using AUGUSTUS (<http://augustus.gobics.de>). We replaced a component of CHAP2 (<http://www.bx.psu.edu/miller_lab/dist/CHAP/README>), the Wise2 (<http://www.ebi.ac.uk>) program with AUGUSTUS, because Wise2 is no longer available. AGAPE also includes an ab initio annotation pipeline, called MAKER. Protein and expressed sequence tag (EST) data for S. cerevisiae, required for running MAKER, were downloaded from SGD (<http://www.yeastgenome.org>) and FungiDB (<http://fungidb.org>) respectively. Results from the CHAP2 and MAKER methods were combined as follows: ORFs predicted by either method were kept. Predicted ORFs that lacked start or stop codons were discarded. Overlapping ORFs with the same stop coordinates but with potential alternative start sites were treated as separate annotations. ORFs predicted to have multiple exons were verified to include either the highly conserved splicing branch point 5’-UACUAAC-3’ or any of the unusual branch points CACUAAC, GACUAAC, UGCUAAC, AACUAAC, UAUUAAC, and AAUUAAC. If no branch point consensus sequence could be identified within an intron, the ORF was discarded. The nucleotide sequences of the predicted ORFs were compared against the S288C reference protein database using BLASTX. Protein matches with e-values less than 1E-6, no more than 5% sequence length difference between the query and target ORFs, and sequence similarity greater than 90% were categorized as bona fide matches and were used to annotate the predicted ORFs. Predicted ORFs not matching these criteria were considered potential novel ORFs and were labeled ‘undefined’. Regions within the contigs, which remained un-annotated or that were labeled ‘undefined’ in the initial phase of AGAPE were analyzed with the MAKER pipeline using all available fungal proteins (downloaded from <http://fungi.ensembl.org>) and ESTs (downloaded from <http://fungidb.org>). The resulting expanded dataset allowed us to capture more potential ORFs which were labeled with corresponding gene names. We applied the same procedure described above for predicted ORFs with potential alternative starts and for examining splicing branch point consensus sites for predicted ORFs with multiple exons. The remaining predicted ORFs were subjected to BLASTX analysis as above, but this time against all fungal proteins and ESTs, and the cutoff stringency was reduced (similarity \> 80%). Predicted ORFs that remained ‘undefined’ were consolidated with overlapping ORFs, and only ORFs greater than 300 bp were retained. All annotations are available in GFF3 format (<http://www.sequenceontology.org/gff3.shtml>) and the BLASTX output is available as a text file for each strain (<http://www.yeastgenome.org/download-data/published-datasets>). ## Identifying novel sequences and ORFs Sequence reads for each strain in FASTQ format were aligned to the S. cerevisiae reference genome using Burrows-Wheeler Aligner (BWA) (‘bwa aln-q 15-l 35-k 2-n 0.04-o 2-e 6-t 1’ and ‘bwa sampe’). Unmapped reads were extracted using SAMtools programs ‘samtools sort’, and ‘samtools view’ (with parameter settings of ‘-u-f 4-F 264’, ‘-u-f 8-F 260’, and ‘-u-f 12-F 256’). Unmapped reads were assembled using ABySS with the same parameters as set in the whole genome assembly. The resulting contigs were aligned to the reference genome to confirm that they were not present in the reference. Contigs shorter than 300 bp, which is the length cutoff for predicted ORFs, were discarded because short contigs are more likely to be derived from reads of low quality, composed of ambiguous bases, or represent spurious ORFs. We consider the remaining contigs as new sequences that are not in the reference. These additional sequences were then aligned to their own strain’s whole genome assembly using LASTZ (version 1.03.02, with parameters ‘T = 2 Y = 3400’) to find the corresponding genomic region of each additional sequence in the whole genome assembly. We created a set of non-reference ORFs from each strain by collecting ORFs annotated for these additional genomic regions in the whole genome assembly. ## Integrative analyses of non-reference ORFs The set of protein sequences of all the non-reference ORFs was aligned to itself to identify potential homologs using BLASTP with cutoff values (e-value less than 1E-1, sequence identity greater than 75%, and sequence length similarity greater than 75%, note we tested different cutoff values to choose the most appropriate combination). We made a binary matrix based on the pattern of presence or absence of each homologue group in each strain and used the matrix to calculate distance among 18 of the strains. The matrix did not include all 25 strains because we found no non-reference ORFs in 7 strains that are very closely related to the reference strain. Then we constructed a dendrogram of the 18 strains using ‘dist.gene’ and ‘nj’ functions in the ape R library (<http://ape-package.ird.fr>). We also predicted molecular function associated with the non-reference ORFs using sequence similarity (BLASTP against NCBI Non-Redundant (nr) database, <http://www.ncbi.nlm.nih.gov>) and conserved protein domains using InterProScan. ## Variation identification and genome diversity To identify SNPs and indels relative to the reference genome, we used the HugeSeq pipeline that integrates multiple variant calling programs. We used the Phylogenetic Tree Galaxy tool (within Galaxy’s genome diversity section) to infer a phylogenetic relationship and population structure based on the SNP data obtained by HugeSeq. To run this Galaxy tool, the SNP data was reformatted to gd_snp format and used as input for generating the phylogenetic tree and population structure to estimate relationship of the strains. Note the Phylogenetic Tree Galaxy tool includes filtering steps for discarding SNPs of low quality or SNPs that are in low coverage regions and we used default settings for these options. ## Tree construction of non-reference MAL gene family The maltose catabolic, metabolic and transport genes (MAL) that are not part of the reference annotations were extracted from our non-reference features, from Bergstrom et al.’s dataset, and from the NCBI Non-Redundant (nr) protein database using BLASTP with queries of the MAL23, MAL43, MAL63, and MAL64 protein sequences. We constructed a maximum-likelihood tree of the non-reference MAL gene family using Phylogeny.fr with default parameters. # Results and Discussion ## Overview of AGAPE We created an integrated pipeline to discover the full set of genomic features of the S. cerevisiae species—the pan-genome—from whole-genome sequences of multiple strains. AGAPE consists of three main parts: assembly, annotation, and variation calls. Given the raw sequence reads of a given genome, a reference genome sequence, and reference genome annotations, the pipeline generates de novo assembly scaffolds and contigs, ORF annotations including non-reference ORFs, and sequence variation calls such as additional newly inserted sequences in the genome (not present in the reference genome) as well as SNPs relative to the reference. The whole pipeline is performed automatically as shown in (for a detailed breakdown see the section). AGAPE was designed to generate genome assembly, annotation, and variation data, with features extracted from newly analyzed genomes added cumulatively to previously generated data. Integrative analyses can be done easily with the updated data and features. Although some organisms may not have thoroughly annotated reference genomes available, AGAPE can still generate the assembly and annotation data as long as a protein database is provided for predicting gene structure. (Note: although the NCBI Non-Redundant (nr) protein database can be attached to the AGAPE workflow, the speed of this annotation step is related to the number of sequences; we therefore recommend selecting a smaller protein database that includes only those proteins that are expected to be similar to the organism of interest). For the variation-calling steps, users can treat a subset of their contig-level sequences as the reference genome. The components of the pipeline can be easily substituted with alternative software as long as the input and output formats are similar to those used in the original step. ## Running the pipeline with the reference assembly for validation We validated the annotation steps by running the pipeline with the reference assembly as input (rather than FASTQ reads). These data were chosen because the reference assembly annotations have been thoroughly curated and can therefore be used to evaluate the accuracy of our predictions. We excluded true reference ORFs shorter than 300 bp to simplify the analysis. Since annotation steps are designed to predict at most one ORF per locus, we also excluded some overlapping ORFs. When two ORFs have overlapping intervals and either one is classified as Dubious, the Dubious ORF was excluded. If both overlapping ORFs are Dubious, the shorter one was ignored. However overlapping ORFs are kept if both are classified as Verified ORFs. In total, we used 5684 reference ORFs as the “true” set. The annotation pipeline predicted 5638 ORFs, 5532 of which were identical to the reference annotations (98.1%). The FDR (False Discovery Rate) was therefore 1.88% ((5638–5532) / 5638). Our approach outperformed the use of either MAKER alone or the homology-based method alone, indicating that our pipeline can generate accurate annotation results if assemblies are of high quality. ## Genome sequences of 25 S. cerevisiae strains To expand the S. cerevisiae pan-genome model, including those ORFs not present in the reference strain S288C, we sequenced strains that are commonly used in experimental yeast studies, including laboratory, wine, environmental, and clinical strains. (The strains are identified in and and short descriptions may be found at <http://wiki.yeastgenome.org/index.php/Commonly_used_strains>). Note that some of the strains in our list overlap with strains analyzed in genotype- and phenotype-based studies. Some strains are diploid. Diploidy may not influence the identification of new features in the pan-genome, but other types of variation analysis may be affected by heterozygosity. We subjected the strain genomes to deep sequencing with coverage ranging from 60X to 320X. Although our assembly contigs are still fragmented with gaps in some genomic regions composed of repeat elements such as rDNA and subtelomeres (.), this high coverage improved the resulting assembly compared to previous yeast sequencing projects. The assemblies yielded N50 values ranging from 30 kb to 125 kb with the longest scaffold reaching 580 kb. ## S. cerevisiae non-reference ORFs and their functional predictions As expected, we did not observe any non-reference ORFs among the seven strains (BY4741, BY4742, FY1679, SEY6210, JK9, W303, and X2180) known to be closely related to the S288C reference genome. Among the remaining 18 non-S288C strains, however, we found a total of 314 non-reference ORFs. We grouped the non- reference ORFs by aligning their protein sequences to each other using BLASTP. As a result, we identified 80 homologue groups of non-reference ORFs, including 16 unique ORFs that appear only in single strains. Eight ORFs out of the 80 non- reference groups were already annotated as non-reference features in SGD: MEL1, RTM1, MPR1, BIO6, TAT3, XDH1, MAL64, and KHR1. Previous studies had shown the presence of the BIO6 gene in saké strains and the TAT3 gene in RM11; our AGAPE results recapitulate these results, showing BIO6 occurring in the saké strain K11, and TAT3 in RM11. To predict functional association of the 80 non-reference ORF groups, we searched the NCBI Non-Redundant (nr) protein database using BLASTP and used InterProScan with the predicted protein sequences. Comparing our set of non- reference ORFs to those found by similar studies, such as Bergstrom et al., was instructive in showing how much of the pan-genome, as indicated by non- reference ORFs, our investigation has uncovered using AGAPE. Since annotation accuracy can be influenced by the quality of the de novo assembly, the comparison can also indirectly serve as an evaluation for the assemblies. Note that in the Bergstrom et al. study, additional data such as low coverage paired-end Sanger sequences and genetic linkage were used to improve assembly while our pipeline used only de novo assembly. Of the 319 non-reference ORFs from Bergstrom et al., 77% are shared with the non-reference ORFs identified by our pipeline. Forty non-reference ORFs from our 18 “non-S288C” strain genomes are not present in the Bergstrom et al. analysis, while 72 ORFs from the Bergstrom et al. study (coming from 14 strains that were mostly natural isolates and not represented in this study) were not found by our analysis. This supports the reasonable expectation that further sequencing will extend the pan- genome, especially if natural isolates are sequenced. ## SNP variations in the S. cerevisiae strains SNPs identified relative to the reference genome for our 25 strains are shown in. Strains BY4741, BY4742, FY1679, and X2180 all have less than 5 SNPs per 100,000 bp, indicating that they are essentially identical to S288C (the SGD reference genome). This is particularly important as FY1679 contributed roughly 50% of the initial chromosomal sequence released in 1996. Strains BY4741 and BY4742 are S288C-derivative strains were constructed to make an ORF deletion collection. The variation between these strains and S288C was known to be miniscule (T. Yamaguchi and F. Roth, personal communication), and our results confirm this. These SNP-based results are also consistent with the fact that we did not find any non-reference ORFs in these four strains (see above section). In general, strains that have more non-reference ORFs also tend to contain more SNPs, especially in the laboratory strains. Interestingly, the two baking strains (YS9 and RedStar) have similar or lower numbers of SNPs relative to the S288C reference, compared to strains isolated from more natural environments (UWOPS, YPS163, YPS128, and DBVPG6044), indicating that the baking strains are less diverged from S288C than the natural environment strains. However, YS9 and RedStar contain the most non-reference features (26 ORFs among the 2 strains), i.e. they have more non-reference ORFs than any other environmental strains. A total of 15 non-reference ORFs are shared by both baking strains, and are not present in any other strains (Groups 51–64). ## Phylogenetic inferences and population structure of S. cerevisiae A binary matrix based on patterns of presence or absence of the non-reference ORF groups in the 18 “non-S288C” strains that contained non-reference ORFs was used to calculate distance and construct a tree of the 18 strains based on a neighbor-joining method. This tree displays the relationships among the 18 strains based only on non-reference features. We also generated a tree based on the genome-wide SNPs found in each strain (relative to the reference). This tree reflects genomic distance based on the divergence of each strain from the reference, within only reference-homologous regions. In both trees, strains isolated from similar environments are generally located closely together. For instance, the two baking strains (RedStar and YS9) are grouped together, as are the three vineyard/wine strains (RM11–1A, L1528, and BC187) and the two oak strains (YPS163 and YPS128). Lab strains that are close to S288C such as D273 and FL100 are grouped together in both trees close to the tightly-grouped S288C-related strains. Non-S288C-based laboratory strains, SK1 and Y55, used widely in studies of meiosis, appear as a branch off the lineage of environmental strain DBVP6044. Interestingly, K11 and YJM339 show different patterns in the two trees. A structure plot suggests the existence of mosaicism in several strains, with SK1 sharing cluster identity over most of its genome with baking and wine strains. This may be relevant to the unknown origin of laboratory strain SK1 and may indicate that SK1 has been mixed with other strains. Unlike the vineyard/wine strains RM11–1A, L1528, and BC187, which are grouped together in both trees, the saké strain K11 appears close to laboratory strains CEN.PK and 10560–6B in the non-reference ORF-based tree, but in the SNP-based tree it clusters more closely to other environmental strains like YPS163 and YPS128, similar to results reported by Liti et al.. Most non-reference ORFs of K11 are present in other environmental and wine strains. ## Case study of non-reference ORFs in strain K11 The distribution of 314 non-reference ORFs into 80 putatively homologue groups enables an exploratory analysis of ORFs that are absent from the reference strain. As a means to link genotypes with phenotypes, strains used in the production of alcoholic beverages are of particular merit given the intense interest in understanding the metabolism of fermentation in these strains. Saké is made from a rice ferment known as koji; before a saké strain of S. cerevisiae can produce alcohols, the rice undergoes saccharification by a mold (or filamentous fungus, viz. Aspergillus oryzae) that metabolizes complex carbohydrates (starch) into sugars (glucose). Saké yeasts form a clade within S. cerevisiae and possess distinct features such as the ability to synthesize biotin. In strain K11, a saké yeast, we have identified 10 non-reference ORFs belonging to 9 homologue groups. Consistent with biotin prototrophy in saké yeast strains, one of these (K11.ORF10) is identical at the DNA level (over its full length) with BIO6 (GenBank AB188681.1). The BIO6 gene is required for biotin biosynthesis and was identified in strain K7 from which K11 is derived. At an intermediate stage of saké fermentation maltose is produced, potentially selecting for the retention, evolution, or horizontal acquisition of maltose utilization genes. Mutagenized strains of saké yeast with low maltose utilization appear to generate higher levels of malate, an organic acid contributing to the flavor of the beverage. Genes for maltose permease (GenBank BAB59002.1) and maltase (GenBank BAB59003.1) have been identified in Aspergillus oryzae and appear to be in a gene cluster with a regulatory gene. Several maltose gene clusters are present in the S. cerevisiae pan-genome. A maltose gene cluster such as MAL6 typically consists of a maltose permease (MAL61), maltase (MAL62), and a MAL regulatory/activator gene (MAL63). Constitutively active forms of the regulatory proteins coded for by these genes have also been identified and appear to relate to loss of function mutations affecting C-terminal residues responsible for negative regulatory function. At the MAL6 locus an additional activator gene MAL64 has been described. A premature termination codon in MAL64 confers constitutive expression although the function of the wild-type allele is unclear. One of the homologue groups identified by LASTZ is comprised of reading frames similar to MAL activator loci. In saké strain K11, two ORFs fall into this group. K11.ORF1 shows partial similarity to maltose activator genes from multiple loci and its function therefore awaits further investigation, while K11.ORF9 shows substantial similarity (\~98% at the DNA sequence level) to MAL64. An alignment (not shown) indicates that, across its length, K11.ORF9 closely resembles wild-type MAL64 in other S. cerevisiae, and a phylogenetic tree indicates that the divergence between MAL63 and MAL64 regulatory genes preceded the divergence of multiple strains. Another interesting non-reference homologue group is represented in K11 by K11.ORF8 and in SK1 by SK1.ORF11, both of which have 100% sequence identity with an epoxide hydrolase-like protein previously identified in saké strain K7 (GenBank GAA21449.1;). This ORF was previously identified in K7 and S. paradoxus and has a presumed bacterial origin, thus representing a possible trans-kingdom horizontal transfer; it has also been identified in 2 commercial wine strains, a sourdough strain and a fuel ethanol strain. Given the toxicity associated with reactive epoxide compounds and the presence of a seemingly non- homologous epoxide hydrolase in Aspergillus oryzae (GenBank XP_001727603.2), it is tempting to suppose that this ORF is required in the saké environment. # Conclusion Rapid evolution and the mosaic structure of genomes in microorganisms makes adequately capturing the diversity of a taxonomic group a difficult task, and requires systematic analysis of multiple genomes. Information from multiple bacterial isolates is frequently combined into a pan-genome, which comprises all genes found within a particular taxon. We have adopted this approach with yeast and have created a flexible pipeline, AGAPE, that uses a variety of tools and sources of information to construct and update a pan-genome. Although AGAPE generates assemblies that can be used to examine between-strain differences, a critical additional output is non-reference ORFs, and AGAPE identifies these by combining prediction methods. We have explored the utility of this approach in yeast by using AGAPE to identify non-reference ORFs through analysis of high-throughput sequencing data from 25 S. cerevisiae genomes. This examination of a small set of non- reference ORFs within S. cerevisiae demonstrates that an updatable pan- genome model can be used as a starting point for analysis of function. We also found that contrasting patterns in SNP- and ORF-based phylogenies, combined with analysis of population structure, suggest that the dynamics of horizontal gene transfer, recombination, or gene gain and loss may be fruitfully investigated as more strains are sequenced and the pan-genome is expanded. Eventually it may be possible to characterize for a particular strain whether ORFs that are not part of the “core” genome (i.e., the set of genes shared by all S. cerevisiae strains) arose by retention and evolution (or duplication followed by divergent evolution) of ancestral genes, or by horizontal acquisition of “novel” genes, e.g., by mating with diverged S. cerevisiae strains or through interspecific hybridization. Despite the difficulties in assembling complete chromosomes, which complicates determination of the presence or absence of some genomic loci, AGAPE provides an expandable pan-genome. The process includes thorough annotation and variation steps, and thus opens a new window to genotype-phenotype association studies. Analysis problems caused by the difficulty of generating complete assemblies, particularly in examination of repetitive elements, can be ameliorated by incorporating improved methods such as using mate-pair libraries and genetic linkage. Beyond yeast, the AGAPE pipeline can be used for genome analyses of other eukaryotes. AGAPE can be modified to consider more complicated gene models and more sophisticated assembly methods can be used to investigate genomes rich in repetitive sequences. In addition, the steps defined in AGAPE can guide genomics studies for researchers who have little experience in computational biology. Our high-quality genome data and the analysis for 25 commonly studied strains are also important resources for furthering yeast genetics studies. The AGAPE package, genome annotation data, and the ongoing expansion of the yeast pan- genome model will facilitate genetic studies in this important model organism. # Supporting Information We thank SGD Project staff for the creation of the high quality and detailed database of S. cerevisiae genes and their products and Webb Miller for helpful comments. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GS JMC. Performed the experiments: GS BD. Analyzed the data: GS BJAD JD. Contributed reagents/materials/analysis tools: GS BD. Wrote the paper: GS BJAD JD SE BD JMC.
# Introduction In the control of infectious diseases by voluntary vaccination, individuals decisions on whether or not taking the vaccine will affect the vaccination coverage and, hence, the effectiveness of disease control, since the prevention of disease transmission requires the vaccination coverage of a host population to be above the level of herd immunity threshold. Existing studies on individuals vaccination decision making have typically focused on several determinants associated with the risks and benefits of vaccination, including the perceived risk of disease infection, the perceived safety and efficacy of vaccine, (e.g., vaccine side-effect rate and the related adverse complications), as well as the social financial costs associated with vaccination and disease infection (e.g., charge of vaccine administration, expenses for infection treatment, and absence from work). Besides these factors, individuals vaccination decisions are also subjected to the impact of social influence in that an individuals behaviors or opinions are affected by those of others. For example, the social influence on individuals vaccination decisions can come from the interactions among them, such as recommendations given by friends or family members, suggestions from health professionals, and advices given by trusted colleagues. The effects of social influence on human health related behaviors have long been observed. In the case of 2003 SARS outbreak in China, individuals avoidance behaviors arose as a response to the circulation of short messages about disease outbreaks. As for vaccination, health related newscasts would change individuals perceptions of vaccine safety and efficacy. The attitudes shared among parents would influence the vaccination decisions of their children. In this regard, modeling vaccination decision making should be treated as not merely a process of payoff optimization, but also a process of individuals response to the impact of social influence. In order to better understand individuals vaccination decision making, in this study, we take a dual-perspective view to address both the cost analysis of vaccination decisions and the impact of social influence. As illustrated in, we consider a group of individuals that make their vaccination decisions by both minimizing the associated costs and evaluating the decisions of others (i.e., social influence). Specifically, we consider that the social settings of individuals are structured with reference to their interaction relationships (i.e., connected individuals and their social closeness). Therefore, the impact of social influence among them will be heterogeneous with respect to the structure of their interactions. In addition, when individuals interact with those having similar choices, their decisions may be further affirmed; otherwise, their decisions may be weakened. In such a case, social impact theory (SIT) provides a computational approach to characterizing the impact of social influence with respect to individuals interaction relationships. Generally speaking, SIT describes how individuals change their attitudes/decisions in a structured social environment, and further suggests that the strength of the social impact be determined by the characteristics of the source (e.g., various attitudes/decisions), the closeness of their social relationships, and the number of sources holding similar attitudes/decisions. In our current work, we propose a novel modeling framework for describing individuals vaccination decision making by integrating an extended SIT-based characterization of social influence with a game-theoretical analysis of cost minimization. In this model, we use a conformity rate to describe the impact of social influence on vaccination decision making, in terms of individuals tendency of being affected by the social influence of others. Additionally, we represent individuals interaction relationships with reference to a social network structure, in which individuals are heterogeneously connected with different numbers of connected neighbors and the social closeness of their interactions. We parameterize the proposed model with an influenza-like disease as well as a real-world social network. By carrying out a series of simulations on voluntary vaccination, we examine the steady state of individuals decision making and evaluate the vaccination dynamics as well as the effect of disease control, in terms of vaccination coverage and the resulting disease infection rate, respectively. By doing so, we aim to investigate the interplay of cost minimization and social influence on individuals vaccination decision making, and examine the impacts of different levels of individuals conformity towards the impact of social influence. Furthermore, we provide a new modeling framework that incorporates the impact of social influence for investigating the effectiveness of voluntary vaccination for infectious diseases control. # Methods We consider a voluntary vaccination program for controlling an influenza-like infectious disease (e.g., seasonal flu), in which individuals need to decide whether or not to be vaccinated each season based on their perceived risk of disease infection. It is assumed that individuals will have some knowledge about the vaccine and the disease (e.g., acquired from their previous experience and/or from public media and health authorities), and about others vaccination decisions through their social interactions. For such a situation, we construct a computational model that describes how an individual arrives at his/her vaccination decision with respect to the cost analysis of vaccination decisions, and the social influence of others decisions. Based on the constructed model, we aim to investigate the impact of social influence on individuals vaccination decisions as well as on the disease control. ## Vaccination Decision Making We take a dual-perspective view on modeling individuals vaccination decision making that incorporates individuals evaluation of vaccination associated costs as well as the impact of social influence. In doing so, we introduce an individual-based model, as described in. In the figure, denotes an individual s vaccination decision. There are two possible decisions that an individual can make: corresponds to an acceptance of vaccination, and represents a rejection. We utilize a social network to characterize the structure of individuals interactions, in which the nodes correspond to individuals and te edges denote the interaction relationships among them. Each edge has a weight, which represents the closeness of interactions between individuals and. Individuals can evaluate the costs associated with their decisions and then arrive at their optimal choices by minimizing the costs. Meanwhile, individuals may also convert their decisions due to the impact of social influence (i.e., neighbors vaccination decisions). Thus, individuals vaccination decision making will be modeled here to include two aspects: (1) cost minimization and (2) the impact of social influence. The parameters used for modeling individuals vaccination decision making are summarized in. ### Cost minimization There are two types of costs associated with an individuals vaccination decision: (1) the cost of vaccination (e.g., the potential risk of vaccine side- effects or the expense of vaccine administration) and (2) the cost of disease infection if not vaccinated (e.g., disease complications, expenses for treatment, or absence from work). We let and denote the costs associated with vaccination and disease infection, respectively, and use represent the perceived risk of disease infection for individual. Then, we can introduce a cost function for individual with a decision, as follows:where denotes the cost associated with accepting vaccination, and denotes the cost associated with rejecting vaccination. Next, without loss of generality, we let describe the relative ratio of and. Thus, we can further transform the cost function in Eq. 1 into the following: Here, we assume that individuals can estimate the risk of disease infection based on their perceived disease severity, as reflected in the perceived disease transmission rate, as well as their neighbors vaccination decisions, as represented by and for the numbers of neighbors with the decisions of vaccination or not, respectively. In addition, vaccinated individuals are assumed to be successfully immunized from disease infection and unvaccinated individuals will be possibly infected and thus transmit disease. Therefore, the perceived infection risk, can be computed corresponding to the proportion of unvaccinated neighbors as follows: Based on the above formulation, an individual can arrive at an optimal choice by minimizing the cost function in Eq. 2. In our proposed model, individual will accept vaccination (i.e.) if, reject vaccination (i.e.) if, and keep his/her decision unchanged in the previous step if. We can write this cost-minimized choice of individual, in the following form: If all individuals follow the same strategy of minimizing their cost functions, after some iterations of decision making, they will reach a steady state, in which all individuals will have no incentive to change their decisions in the next step. ### Social influence In addition to the above-mentioned cost minimization, an individual may affect by those decisions of others, i.e., due to the impact of social influence, and then convert the cost-based choice to the social opinion of his/her neighbors. According to social impact theory (SIT), the strength of such a social influence will be subjected to the structure of individuals interactions, e.g., the types of opinions (i.e., acceptance or rejection), interaction relationships (i.e., social closeness), and the number of opinion sources (i.e., the numbers of vaccinated and unvaccinated neighbors, and, respectively). In our social network, for individual, the strengths of social influence for two opposite opinions (i.e., vaccination acceptance and rejection), described by and, can be accordingly computed based on as follows: We use to denote the formalized social opinion resulting from the social influence of individual s neighbors. As a modification of the standard SIT definition (where corresponds to the opinion with the larger strength of social influence), being either acceptance or rejection of vaccination will be determined by comparing the influences of two opposite opinions. We let denote the discrepancy between and. Then, we normalize as follows: Here, we use to denote the probability that social opinion is to accept vaccination, and to reject vaccination. Therefore, we can write in the following form:where is computed from the Fermi function as follows: The Fermi function is a sigmoid function that has been widely used for describing individuals behavioral changes as a response to the payoff discrepancy of two different choices. Here, describes individuals responsiveness to the impact discrepancy of two opposite opinions. As shown in, a larger value of means the choice with a higher social influence will be more inclined to dominate the social opinion even the discrepancy of the two opposite social influence, is relatively small. Next, we introduce a probability, called individuals conformity rate, that indicates the degree of individuals tendency towards adopting the social opinion of his/her connected neighbors, that corresponds to how likely individual will convert his/her cost-minimized choice to the social influence formalized opinion. Thus, corresponds to the case of a cost-based decision maker, whereas indicates that the individual is an absolute social follower (i.e., ignoring his/her own cost evaluation). In other words, the final decision of individual can be expressed as follows: ## Vaccination Threshold In order to evaluate the impact of individuals vaccination decision making on disease control, we further construct a disease model to describe the threshold of vaccine coverage for mitigating an epidemic (i.e., the reproduction number is less than one). The parameters used for estimating the vaccination threshold are listed in. For the sake of illustration, we use a standard SIR model to describe an influenza-like disease transmission in a group of individuals that are densely aggregated (e.g., students in a school), which can be treated as a homo-mixed population for disease transmission. Individuals are divided into three compartments with respect to their epidemiological states, i.e., susceptible (S), infectious (I), and recovered (R). In addition, the number of individuals in each compartment is denoted by, , and. When the natural birth and death of the population are not taken into account, the overall population size is calculated as. The disease spread dynamics is described by the following set of differential equations:Additionally,where is the risk of disease infection for susceptible individuals that is proportional to the percentage of infectious population size. denotes the disease transmission rate that is the probability of disease transmission between the mixing of infectious and susceptible individuals. describes the recovery rate that corresponds to the time period for an infected individual to be naturally recovered and thus immunized from secondary infection. Reproduction number (i.e., the number of secondary infections caused by a typical infectious individual in a completely susceptible population,) indicates a threshold for disease transmission; that is, if, disease transmission will naturally decay. In such a compartmental disease transmission model, is given as follows, : Therefore, the vaccination threshold for mitigating an epidemic is estimated as the reproduction number less than one (i.e.). The corresponding vaccination coverage, denoted by, can be estimated as follows: ## Simulation Setting For our simulations, we calibrate the parameters of individuals vaccination decision making based on the scenario of the 2009 H1N1 influenza epidemic, in which reproduction number was estimated as and recovery rate was set to 0.312 (i.e., a 3.2-day recovery period for disease infection). In order to focus our studies on the impact of social influence, we assume that the perceived disease transmission rate is equal to that of the actual disease transmission, i.e.. In addition, we construct a social network based on the data of individuals close proximity interactions (i.e., distance less than 3 m) at an American high school, where the social closeness between individuals and corresponds to the frequency of their interactions (i.e., the sum of all interactions between the two individuals during the day). The total number of nodes is and the average node degree (i.e., the number of connected neighbors) is 35. The average edge weight (i.e., social closeness) is 115 units. Based on our model parameterization, we carry out Monte Carlo simulations to experimentally study individuals vaccination decision making and the impacts of the resulting vaccination coverage on disease control. # Results Based on the proposed decision model, we have conducted a series of simulations on vaccination dynamics to estimate the vaccination coverage at the steady state of individuals decision making. As shown in, we first investigate the interplay of cost minimization and the impacts of social influence on individuals vaccination decision making with reference to three initial levels of individuals vaccination willingness:, and. Generally speaking, the level of vaccine uptake will be subjected to the cost ratio and individuals initial level of vaccination willingness, when individuals conformity rate takes different values. Specifically, the simulation results in show that when the impact of social influence is relative weak (i.e., conformity rate is relatively small), the cost of vaccination (i.e., cost ratio) fundamentally determines the resulting vaccination coverage in that increasing the cost of vaccination will lower individuals vaccination willingness (i.e., the steady state of individuals decision making). In our considered scenario, the vaccination coverage is around when cost ratio. Gradually, if is decreased and approaches, the vaccination coverage will become as high as. Based on our model design, when an individual perceives that all of his/her connected neighbors have decided for vaccination, the individual will keep his/her previous choice of non-vaccination even if the cost of vaccination is zero, due to the consideration that disease transmission will no longer exist. Furthermore, we can observe that the strength of social influence (i.e., conformity rate) can adjust the aforementioned impacts of cost ratio on individuals vaccination decisions. As in the extreme case that individuals are pure cost-based decision makers (i.e.), the resulting vaccination coverage will be completely determined by the relative cost of vaccination (i.e., cost ratio ). On the other hand, as in the extreme case that individuals are absolute followers of social opinion (i.e.), the impact of social influence will promote a universal vaccination coverage, the level of which depends on individuals initial level of willingness instead of the associated costs. In this case of simulation, when, the vaccination coverage at the steady state of decision making will converge to around for individuals vaccination willingness at the initial level of (i.e., as shown), at the level of (i.e., as shown in), and at the level of (i.e., as shown). In addition, the impacts of varying conformity rate (i.e., individuals tendency to adopting social opinions) are also observed as the adjustment of vaccination decisions with reference to different situations of vaccination associated costs (i.e., cost ratio). When individuals become more likely being affected by social influence (i.e., gradually increasing conformity rate), as shown in, the impact of social influence tends to increase the vaccination coverage when the cost of vaccination is low (i.e.). On the other hand, when the cost of vaccination is relatively high (i.e.), the impact of social influence will reduce the resulting vaccination coverage at the steady state of individuals decision making. Furthermore, when conformity rate approaches, the vaccination coverage will drop/increase sharply and finally converge to a fixed level that depends on individuals initial level of vaccination willingness. Based on the earlier-mentioned SIR model, we have investigated the impact of social influence on disease control by evaluating disease infection rates (i.e., the percentage of individuals being infected as a result of disease transmissions) with respect to different vaccination coverage resulting from individuals decision making. shows the disease infection rates with respect to the interplay of individuals cost minimization and the impact of social influence on vaccination decision making (i.e., the values of cost ratio and conformity rate ranging between and, respectively). With respect to our considered epidemic scenario (i.e., basic reproduction number), the simulation results in show that disease infection can be eliminated given a relatively lower cost of vaccination (i.e., cost ratio) and a moderate impact of social influence (i.e., conformity rate). Specifically, when individuals are less likely to being affected by social influence (i.e., conformity rate), the effectiveness of disease control is generally determined by the relative cost of vaccination (i.e., cost ratio) in that a lower vaccination cost can lead to a reduction in the disease infection rate due to a resulting higher vaccination coverage. Furthermore, as individuals tendency of being affected by social influence become strengthened (i.e., conformity rate), the effect of vaccination cost on disease control will be weakened accordingly, while individuals initial level of vaccination willingness matters. In the extreme case of (i.e., individuals are absolute followers of social influence), the disease infection rate is observed as high as for the initial level of vaccination willingness at, as shown in. If the initial level of vaccination willingness is set as (i.e., as shown), the disease attack rate will be relatively higher than the situation of the initial level at (i.e., as shown), where cost ratio and conformity rate. Besides, we have examined the steady-state vaccination coverage and the resulting disease attack rate with respect to different initial levels of individuals vaccination willingness prior to their decision making, the results of which are shown in. We can note that individuals initial level of willingness will affect the converged level of the steady-state vaccination coverage as well as the effectiveness of disease control when individuals are absolute followers of social opinions (i.e., conformity rate). In our simulations, when the initial level of individuals vaccination willingness is, the converged steady- state vaccination coverage is around. The vaccination coverage will reach and, if the initial levels of vaccination willingness are and, respectively. In addition, we can observe that there exists a critical phase transition in vaccination coverage when individuals initial level of vaccination willingness is between and. That is to say, in the situation of individuals being absolute social followers, there is a threshold value in terms of individuals initial level of vaccination willingness that can be used to evaluate the effectiveness of a voluntary vaccination program for eliminating the epidemic. ## Sensitivity Analysis In order to investigate the sensitivity of our results, in what follows, we further consider individuals vaccination decision making with respect to the different values of disease reproduction number: (1) ; (2) ; (3). shows the vaccination thresholds for eliminating the epidemic with respect to different basic reproduction numbers. shows the vaccination coverage at the steady state of individuals decision making with respect to different disease reproduction numbers. Here, we can observe the similar impacts of social influence in all three considered situations: the impact of social influence will increase the vaccination coverage when the relative cost of vaccination is small, decrease it when is relatively large (see.g, 8.h, and 8.i), and bring it to a certain level when individuals become followers of social influence (i.e., conformity rate approaches 1). The simulation results further show that when the impact of social influence is relatively weak (i.e., conformity rate), relatively severe disease transmissions in terms of a larger reproduction number (i.e.) will increase the vaccination coverage. While, if the impact of social influence is strengthened (i.e., conformity rate approaches 1), the vaccination coverage at the steady state of individuals decision making is mostly determined by individuals initial level of vaccination willingness, rather than the related costs and disease severity. # Discussion The phenomena of social influence that individuals behaviors or opinions are affected by their social environment have long been observed and studied, such as in the domains of political voting, and consumer purchasing decisions. In the context of vaccination, social influence can affect individuals vaccination decisions and thus the effectiveness of disease control in terms of the resulting vaccination coverage. In this study, we address the impact of social influence on individuals vaccination decision making, vaccination coverage, and disease control. Towards this end, we have provided a dual-perspective view on modeling individuals vaccination decision making by incorporating the impact of social influence with the game-theoretical analysis of vaccination cost minimization. In a group of individuals, the impact of social influence on an individuals decision making relies on the structure of how he/she interacts with others. In order to characterize the impact of social influence in such an interactive environment, we have used social impact theory (SIT) to characterize the strength of social influence on changing individuals vaccination decisions with respect to their interaction relationships. We have used individuals social network to represent the structure of their interaction relationships. Based on our proposed model, we have examined the impact of social influence on individuals decisions and on the effectiveness of disease control (i.e., vaccination coverage), with respect to three determinants: (1) the relative cost of vaccination decision, i.e., cost ratio ; (2) individuals conformity to social influence, i.e., conformity rate ; and (3) individuals initial level of vaccination willingness. By parameterizing the proposed model with a real-world contact network and with the epidemiological scenario of 2009 H1N1 influenza, we have carried out a series of simulations on individuals voluntary vaccination. The simulation results have confirmed that the relative cost of vaccination (i.e., cost ratio) is one of the determining factors in the voluntary vaccination coverage. In our simulations, such results can be observed if individuals are less likely to be affected by social influence (i.e., conformity rate is relatively small). While, if individuals become more susceptible to social influence (i.e., is large), the impact of social influence has been found to increase the vaccination coverage when the cost of vaccination is small and, conversely, reduce the vaccination coverage when the cost is large. In the extreme case where individuals are absolute social followers (i.e., conformity rate), the vaccination coverage at the steady state would converge to a certain level that merely depends on individuals initial level of vaccination willingness, instead of the vaccination associated costs. In modeling individuals vaccination decision making, several mathematical models have been earlier proposed that utilize payoff-based approaches to characterizing vaccination decision making with respect to individuals perceived costs and benefits of vaccination. Bauch et al. , characterized individuals vaccination decisions as a modified minority game by exploring the herd immunity effect; that is, in a group of mixed individuals, vaccinating a proportion of them would decrease the infection risk for the rest of individuals. In consideration of that, game theory has been used to describe individuals interactive decision making in favor of optimizing personal payoffs. Cojocaru extended the game-theoretical model of vaccination decision making by considering a finite number of heterogeneous population groups. Perisic et al., further incorporated individuals contact networks into the vaccination game analysis. Moreover, some studies have considered social and psychological aspects of decision making (e.g., social learning process and imitation behaviors). While, others have considered the issues of incomplete information by adding either the potential discrepancy between individuals perceptions and real situations (e.g., the perceived disease prevalence and the adverse effects of vaccine) or different sources of information (e.g., previous disease prevalence or vaccination programs –). Besides the payoff-based analysis, Salathe et al. in investigated the clustering of vaccinated and unvaccinated individuals with an opinion formation model. They proposed that the probability for an individual changing his/her vaccination opinion is proportional to the ratio of neighbors that have an opposite opinion. As an improvement over the above-mentioned existing models, we consider an individuals vaccination decision as a hybrid process balancing his/her self- initiated cost minimization (i.e., individuals minority-seeking-like behaviors by exploring the herd immunity effect) as well as the social influence of neighbors decisions (i.e., social conformity behaviors). Our model introduces a parameter (i.e., conformity rate) to modulate individuals tendency towards these two decision making mechanisms: an individual will adopt his/her cost-minimized decision, or convert to the social opinion of his/her connected neighbors. Different from the existing studies that address individuals vaccination decision making as a process of opinion formation (e.g., Salathe et al.), here we further take into account the heterogeneities of individuals interaction relationships by exploiting an extended SIT-based characterization of the strength of social influence. Additionally, by incorporating the impact of social influence, we are able to investigate the impact of individuals initial level of vaccination willingness on the vaccination coverage of individuals decision making. By computationally characterizing the impact of social influence, this study has practical implications for understanding individuals vaccination behaviors and for improving the effectiveness of adopted vaccination policies. In the recent years, the rapidly increasing use of new communication tools e.g., internet- based social media services, has further amplified such a social influence. For instance, the efficacy or the adverse effects of vaccines would be debated, and the opinions on either accepting or rejecting vaccination would fast spread among individuals. We have identified that individuals initial level of vaccination willingness as an important factor in determining the final vaccination coverage due to the impact of social influence (i.e., individuals social conformity). Our results have shown that when conformity rate approaches, the vaccination coverage at the steady state of individuals decision making will be polarized given different initial levels of individuals vaccination willingness. Moreover, the empirical studies that survey the determinants of individuals vaccination decisions in a social environment can readily provide us a practical means for measuring and evaluating individuals conformity to social influence. As has been shown in our study that individuals vaccination decisions can be affected by both the associated costs and their conformity to social influence, it becomes necessary and feasible for public health authorities to estimate the level of individuals acceptance of vaccine prior to the start of a voluntary vaccination program, as well as to timely assess and enhance the effectiveness of their adopted vaccination policies, e.g., providing certain financial subsidies to reduce the cost of vaccination. So far, our study has provided a general modeling framework for incorporating the impact of social influence into the individuals decision making and disease control. It should be pointed out that the obtained results of this study may be subjected to the considered social network (e.g., students interactions within an American high school). In our proposed model, the social influence accounts only for the localized interactions between an individual with his/her connected neighbors. Additionally, by utilizing the SIT-based characterization of social impact, an implicit assumption is that individuals are passive recipients of social influence and their active behaviors have not been taken into account. It would be interesting for us to further consider some of the related aspects in our future work: ## ### 1. The effects of public media Public media represents another type of information source that will affect individuals vaccination decision making. Due to broadcasting effect, the transmission of social influence through public media may be faster and wider. Related work by Breban discussed the effects of media on the fluctuation of vaccination coverage. For the future work, it is possible to extend the current model by incorporating the effects of public media, e.g., by adding a super node that interacts with a large portion of nodes. ### 2. Host population heterogeneity To focus on the SIT-based characterization of social influence, we have assumed that individuals are homogeneous in disease infection, e.g., susceptibility, infectivity, and infection risk. In this regard, our modeling framework will be further extended to incorporating individual variations in disease transmission as well as in their social characteristics (e.g., creditability). It would be desirable to further improve our simulations by differentiating physical contacts for infectious disease transmission from interaction relationships for social influence. Along this line, related work by Eames constructed a parent network for describing vaccination decision and a children network for representing disease transmission, and found that the impact of social influence would be influenced by the overlap of these two networks. ### 3. Dynamics of disease spread In this work, we have only considered individuals making vaccination decisions based on their perceived infection risk, which may come from either their previous experience of disease and vaccine or the awareness about the upcoming epidemic season. In the real world, real-time disease dynamics could also affect vaccination dynamics, i.e., disease outbreaks may increase individuals willingness for vaccination. In the future, we will extend our model by characterizing the interplays between individuals vaccination decisions and the dynamics of disease spread. The authors would like to express their gratitude to the three anonymous reviewers who provided constructive comments on further improving this article. They would also thank other members of the AOC Research Group at HKBU for their feedback on this study. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JL SX. Performed the experiments: SX. Analyzed the data: SX. Contributed reagents/materials/analysis tools: JL SX. Wrote the paper: SX JL.
# Introduction The crucial assertion of the neutral theory of community ecology is that diversity can be maintained in the absence of species differences as long as there is steady input of new species via speciation or immigration. More broadly, diversity can be maintained independent of niche divergence, or in the face of competitive dominance, given sufficient dispersal. Various models produce predictions of species abundance distributions as they depend on dispersal, and these have been tested against real forests, but the theory also includes a quantitative prediction of diversity as a function of speciation. None of the previous empirical tests of neutral theory, however, considered the speciation parameter. In the absence of novel species input, the neutral theory is irrelevant, and stabilizing mechanisms such as niche differentiation among species or competitive interactions must maintain diversity. Here we make use of repeated censuses of the Barro Colorado 50-ha forest plot in Panama to examine the rate at which novel species appear; we call this the rate of species input. There have been seven complete censuses over 30 years, and each of the last six provides a direct estimate of the rate of input. The simple and obvious test is whether there has been any at all: have any novel species recruited into the 50 hectares since the initial census of 296 species in 1982 ? The flora of Barro Colorado Island is well known, and we would certainly know new species. The more precise test is whether the observed species input is high enough to maintain the plot's observed diversity. The theory provides an exact prediction about what would be sufficient. Our aim does not end with a qualitative confirmation or rejection. We intend to measure a rate constant that is relevant in the ecological theory that accommodates stochastic births and deaths, dispersal, and species interactions. Whatever value we find, the rate at which new species immigrate leads to inferences about the forces that are key in maintaining species diversity. # Materials and Methods ## Theory In the basic neutral theory, a community of individuals is subject to random deaths and births. At each time step, one individual dies, with every individual equally likely, and then one of the survivors is chosen at random to become the parent of a replacement. At a constant rate, the newborn mutates and becomes a new species, hence is called the speciation rate and is equal to the probability that any newborn is a novel species. If is constant, then species diversity and the full species abundance distribution eventually reach a dynamic equilibrium around which they will subsequently fluctuate randomly. The equilibrium can be derived analytically, and a single parameter, the biodiversity parameter, fully describes it. turns out to be asymptotically equal to Fisher's diversity statistic, defined bywhere is the number of species in the community. Setting, this leads to a prediction about the speciation rate that would maintain species in the community, This formulation holds only for a metacommunity: a community into which there is no immigration and within which there is unlimited dispersal, meaning every individual is equally likely to be the parent of any birth. To relax both assumptions and accommodate limited dispersal, consider a small subset of the metacommunity termed the local community. Any subset will do, as long as the boundaries are unvarying so that immigration of newborn from outside has a consistent meaning: offspring whose parents reside outside the local community. We imagine the metacommunity as a continent of trees and the local community as a rectangular plot with precise but unchanging borders, but the theory accommodates more general arrangements. Define the migration rate m as the proportion of births in the local community whose parents are outside, or equivalently the probability that a newborn comes from outside. The remaining births are from local parents. Immigration alters the local species abundance distribution, and – no longer hold. There are various derivations of the abundance distribution in a local community, providing estimators of both and. Still however, refers to speciation in the entire community, and we need to know it for the local community, where both genetic variants causing true speciation and arrival of novel species via immigration must be considered. The former is simply, but we need a derivation for the latter: the probability that a recruit in the local community is the immigrant offspring of a novel species from outside. is the ratio of new species to all immigrants, equal to the probability that a randomly chosen immigrant is a species not present locally. We assume genetic speciation is very rare locally, so, and henceforth consider as the only species input parameter relevant to diversity in a small community. To find, first define as the probability that an individual randomly selected from outside the local community is a novel species, meaning it belongs to a species not currently in the local community. Then we can write, because to be a novel species, a recruit must be an immigrant , and the immigrant must be a new species. We first derive an explicit formula for using a previously published formula for the probability of any local abundance given the metacommunity abundance. The leads to an expression for as a function of and :where is the local community size and the biodiversity parameter defined above. Using Equation 7 in, we can remove from the formula, using instead, the number of species in the local community. Thenandwhere and is the lower branch of Lambert's W function. parallels as a way of estimating species input given species richness and community size, but with migration rate also needed because it is a local community. It turns out, however, that is only weakly dependent on m, which has an intuitive interpretation: larger m means more births are from parents outside the local community, however, it also means that fewer species are absent from the local community. In the Barro Colorado 50-ha tree census, there were individuals and species in 1990. Using, we find. This value of has been derived several times by various means, but always based on trees mm. Using the program Tetame (<http://www.edb.ups-tlse.fr/equipe1/chave/tetame.htm>), we applied the formulation from to arrive at for trees mm. This leads to, meaning that a near four-fold increase in m leads to a 50% increase in. Indeed, from to, varies only three-fold. Hence, if the Barro Colorado forest were a community of fixed size, all species were demographically identical, 38% of recruits came from parents outside, and one of every recruits was a novel species, the observed equilibrium diversity would be species. Our question is simply whether this predicted rate of species input is in fact observed. If it is, then local diversity can be attributed to species input, and no local diversifying mechanisms are needed. If we do not observe species input, then we must look to local processes for the maintenance of diversity. # Methods ## Ethics statement All research was conducted at the Barro Colorado Nature Monument, a forest preserve owned by the nation of Panama and managed by the Smithsonian solely for scientific studies and protected from all other uses. A long-term agreement with the Government of Panama assures the Smithsonian permission to continue the research indefinitely. No protected species were sampled during the forest census. ## Plot census Since 1982, a rectangle of forest 1000 m×500 m on Barro Colorado Island, Panama, whose southwest corner is at 9.15125°N. latitude, 79.85530°W. longitude, has been fully enumerated seven times. During the initial survey (1981–1983), every individual mm in stem diameter was given a numbered aluminum tag and measured; in 1985 and every 5 years since, tagged trees were remeasured or noted as dead, and new trees mm diameter were given tags. Trees were identified to species by three experts, and voucher specimens were collected and deposited in two herbaria in Panama (STRI, PMA). Since our results hinge on population changes in rare species, we considered carefully how misidentification would affect the calculations. In a random sample of re-identified trees, we found 0.85% misidentified (159 of 18694). In the rarest species – the 44 having individuals (in 1990) – we double-checked every individual (in 1995), and found just one out of 85 individuals was misidentified, a Hamelia axillaris mistakenly called the rare H. patens. Other H. patens were correctly identified, and in no case did a misidentification either remove the last record of a species or create a novel species. There were four rare taxa that we omitted from calculations due to taxonomic uncertainty: one unidentified tree in the genus Nectandra, possibly a novel species but never seen flowering before its death; a single tree in the genus Apeiba that could not be identified before it died (it appeared to be hybrid between the two well-known species); and both species in the genus Trema, originally identified as one species but later separated (we re- identified every living individual after 2000, but several trees that died earlier remain forever unidentified). The remaining species subject to extinction or invasion are very well known to us; most of them occur in our tree plots elsewhere in Panama, and we have observed every one of them outside the 50-ha plot. We are thus confident that our estimate of species turnover is based on true extinction and invasion, and that it is unbiased, since misidentifications could cause either errors of omission (misidentification of a rare species as a common species and thus failure to detect invasion or extinction) as well as commission (misidentification of a common species as a rare species and thus an apparent case of invasion or extinction when there was none). Criteria for including stems in the census were applied consistently and define the local community of our theory: free-standing, woody stems at least 10 mm in diameter. Species known to be lianas at maturity were never included. On the other hand, individuals of species known to be stranglers (hemiepiphytes) at least some of the time were tagged whenever they were free-standing. To be consistent with the liana method, we excluded all stranglers from analyses here (nine Ficus and one Oreopanax species). Reinserting them in the calculations had a trivial impact on the final estimates. Recruits were defined as newly appearing stems, those growing from mm stem diameter in one census to mm in the next. Deaths were trees with stems mm in one census but with no such stems alive in the next census, meaning we considered a tree ‘dead’ even if it maintained a living base. This definition is required to guarantee book-keeping of the population mm: adding recruits and subtracting deaths is how populations changed. ## Observed recruitment and species input Following the theory, we define a species input event as any case where a species absent from the plot in one census appeared in the next, and an extinction event as the opposite. An intuitive estimate of the rate of species input between any pair of censuses is the number of novel species divided by the number of recruits. Likewise, the extinction rate can be simply defined as the number of extinctions divided by the number of deaths. But recruitment, input, death, and extinction are continuous processes, and a more precise estimate can be generated by solving differential equations describing their rates. The estimated input rate based on the continuous solution differs only slightly from the intuitive estimate because the rates are low. Cases where species became locally extinct in one census interval, then reappeared in a later census, were counted once as extinction and later as species input. Such input does indeed maintain diversity. Hypothetically, there might be a time far in the future when all species in the region have passed through the 50 hectares at least once, when every input event would be a species that had already been in the plot. This would still comprise a community in which local diversity is controlled by the rate of species input from outside. # Results In every census interval, there was species input and extinction. A total of 308 non-strangler species mm stem diameter were observed in the 50-ha plot during at least one of the seven censuses, but only 275 species were present in all seven censuses. The other 33 species had some turnover: they were absent in at least one census. For example, in 2010 there were three new arrivals relative to 2005: Vasconcellea cauliflora, Vismia macrophylla, and Psychotria hoffmannseggiana. Vasconcellea is rare throughout the region but unmistakable even to novices, having large, extremely lobed leaves like those of the related papaya (Carica papaya). Prior to 2010, we had never seen it in the 50-ha plot. Vismia is also easy to recognize, since it is abundant along roadsides of wet and submontane forest nearby, but at Barro Colorado it is rare. It was found in the plot prior to 2010, but went extinct (twice) and has now (twice) re-invaded. The final invader, P. hoffmannseggiana, is a rare shrub that can only be identified by experts; it also went extinct then reinvaded. Those three immigrant species were found among 29243 recruits since 2005, thus the intuitive input rate per recruit. The dynamic rate estimate was slightly lower. Over six census intervals, varied from to new species per recruit, while according to theory, would maintain 291 species in the 50-ha plot. That assumes a migration parameter of, but with the lower migration rate, the input rate needed according to theory would be only slight lower. The observed rate was close, falling within the theoretical range twice and barely lower once. When differing most, the observed rate was higher than predicted, though by less than threefold. # Discussion Tree species have been continually input into the 50-ha plot at Barro Colorado over 30 years, and the observed rate was consistent through time and a quantitative match to the theoretical rate needed to maintain diversity. There is nothing circular in our estimates of observed and predicted input: the theoretical rate depends on local species richness , community size, and immigration (m), none of which depends on novel species. Indeed, given these values of, and m, an observation of no new species and thus zero input was completely plausible. We have known since the first census that the 50-hectare plot is a subset of a regional community, because Croat's flora of Barro Colorado Island includes 450 tree and treelet species, leaving nearly 150 species absent from the plot. A few of those are specialists in habitats not found within the 50 hectares, such as the pond apple (Annona glabra) of the lake shore, but most are upland species that could grow in the plot. Future censuses will capture more of those species, while others will continue to drop out. The 2010 census included 13 singletons (species with a single individual), and these are at obvious risk of local extinction: 10 of the 17 singletons in 1982 are now extinct. But it is not just singletons subject to turnover. Three species with individuals in 1982 are now extinct, and Cecropia longipes, which invaded the plot in 1995, now has 13 individuals. Had we observed a rate of species input substantially lower than the prediction for maintaining diversity we would have concluded that stabilizing mechanisms, i.e. rare species advantage, competition, or niche differentiation, are important in maintaining species richness. Had the rate been too high, we would have been forced to consider destabilizing mechanisms that drive rare species to extinction faster than expected by chance. We conclude instead that species input is maintaining tree richness in the Barro Colorado plot. Stabilizing forces may be present, and they may limit abundances, but they do not contribute to diversity. Indeed, the observed extinctions demonstrate that whatever stabilizing forces are present are insufficient to protect rare species. These conclusions conform with a variety of theoretical studies showing that dispersal can overwhelm niche differences as the driver of diversity and community structure, so that regional diversity can regulate local diversity,. The importance of species input explains the success of the neutral model in predicting abundances in spite of evident non-neutrality. When species input dominates, abundances resemble the neutral prediction, particularly in the long tail of rare species, even if there are species differences. In fact, the zero- sum multinomial abundance distribution predicted by the neutral theory generalizes to habitat-partitioned communities as long as there is species input. The neutral model predicts diversity and abundance at Barro Colorado because it properly describes what matters most – species input – while ignoring irrelevant details. The theory also predicts exactly how much species input is sufficient to maintain richness, and that when insufficient, diversifying mechanisms spawned by species differences must account for abundances and diversity. Diversifying mechanisms at wider scales are not addressed by these results. There is a regional species pool from which the 50-ha plot is drawing, and there may be niche differentiation maintaining diversity in the wider region. The role of species input (speciation of any kind) in maintaining diversity at larger scales remains untested, because precise measures of species interactions and species input are not now possible beyond local plots. The rate of immigration of novel species must decline as area increases, but the rate of input needed to maintain diversity also declines. Perhaps over the entire nation of Panama or the continent of South America, speciation is too rare to maintain diversity without niche segregation, or perhaps species input via true genetic speciation is what drives diversity and abundances at continental scales. # Supporting Information We acknowledge R. Foster as plot founder and the first botanist able to identify so many trees in a diverse forest, and we thank R. Pérez and S. Aguilar for species identification, S. Lao for data management, S. Dolins for database design, plus hundreds of field workers for the census work, now over 2 million tree measurements. T. Zillio first produced the calculations in. The Smithsonian Tropical Research Institute and Smithsonian Institution Global Earth Observatories provided logistical support, and we thank S. Davies, I. Rubinoff, and E. Bermingham for their support. [^1]: The authors have declared that no competing interests exist. [^2]: Analyzed the data: RC. Contributed reagents/materials/analysis tools: SPH RC RAC. Wrote the paper: RC RAC SPH. Conceived field work: SPH. Implemented field work: RC. Conceived database: RC. Implemented database: RC. Generated funding: SPH. Derived theory: RAC RC SPH.
# Introduction Myotonic dystrophy type 1 (DM1) can appear at any time in life and is regarded as the human disease which probably has the most variable clinical presentation, somehow affecting virtually all body systems. Although typically classified as a neuromuscular disease, besides its prominent muscular system defects (including cardiac, smooth, and skeletal muscle cell types), it also compromises cognitive, ocular, digestive, endocrine, respiratory, reproductive, cutaneous, haematopoietic, and immune systems to varying degrees. Characteristic muscular symptoms include cardiac problems such as malignant arrhythmias and conduction defects, and involvement of facial (ptosis), bulbar (dysarthria, dysphagia), limb (steppage, gait troubles), and smooth (constipation) muscle with associated muscular atrophy and myotonia. Patients also suffer from iridescent cataracts and insulin resistance with metabolic syndrome. Genetically, it is an autosomal dominant disease caused by unstable expansion of the CTG microsatellite in the 3’ untranslated region of the dystrophia myotonica-protein kinase (DMPK) gene and is a rare disease that afflicts one in 8000 people worldwide. Unaffected individuals carry between 5 and 37 CTG repeats whereas DM1 patients carry between 50 and thousands of CTG triplets. Importantly, CTG trinucleotide expansions are unstable both in the somatic and germinal lines, likely contributing to the heterogeneity in clinical symptoms and age of onset, which inversely correlates with the size of the triplet expansion. A further increase in the size of the CTG microsatellite occurs in most intergenerational transmissions of the expanded allele, which correlates with genetic anticipation. Despite the correlation between the size of the CTG expansions in blood cells with disease severity and age of onset, its predictive power is poor and it is not a good parameter for characterising the disease load. Forthcoming therapeutic trials urgently need good biomarkers to evaluate the therapeutic response to treatments. Alternative splicing changes in skeletal muscle have been described as potential biomarkers of disease severity and therapeutic response, but they involve invasive techniques and it would be difficult to routinely measure them in other sites (such as cardiac or cerebral tissues) which are strongly involved in DM1 pathophysiology. Expanded RNA transcripts containing CUG repeats are retained in the cell nucleus as insoluble RNA aggregates known as ribonuclear foci. These foci are able to sequester different RNA binding proteins that are prevented from performing their normal functions. The alternative splicing regulators Muscleblind-like1 (MBNL1) is among the recruited proteins, which result in its functional depletion. CUGBP, Elav-like family member 1 (CELF1), a splicing factor antagonist of MBNL1, is not sequestered in ribonuclear foci but becomes abnormally activated due to hyperphosphorylation. As a consequence, several alternative splicing events are misregulated in DM1 and in some cases these splicing defects contribute to DM1 symptoms such as myotonia, insulin resistance, or muscle weakness. The molecular mechanism leading to DM1 pathogenesis is complex and, in addition to splicing defects, also includes mispolyadenylation of pre-mRNA, a process that is also regulated by MBNL proteins, repeat-associated non-ATG translation (RAN translation), bidirectional transcription, defects in transcription and translation, epigenetic changes, and the silencing of cardiac and muscle transcripts by changes in miRNA expression levels. miRNAs are endogenous non-coding RNAs, approximately 21 nucleotides long, that function as post-transcriptional gene expression regulators by targeting the 3’ untranslated region of their complementary target mRNA. miRNAs regulate RNA stability and translation rates via degradation or inhibition of protein translation, respectively (reviewed in). Over 2000 miRNAs have been identified in the human genome and have been implicated in numerous biological processes including development, proliferation, differentiation, and stress responses (reviewed in). Because miRNAs can be readily detected in body fluids, and particularly in blood components, differences in serum miRNAs have been proposed as potential non-invasive biomarkers of disease progression for several conditions such as cancer, Alzheimer´s disease, hepatitis B infection, retinopathies, gestational diabetes mellitus, or Duchenne muscular dystrophy. Because several miRNAs have been detected to be altered in DM1 cardiac and muscle tissues, and there are numerous drugs that work in DM1 animal models pending accurate pharmacological development and clinical testing in humans, we explored the possibility that misexpression of specific serum miRNAs could be identified as non-invasive DM1 biomarkers. To this end, we profiled 175 miRNAs in the peripheral blood serum of DM1-affected individuals and healthy controls by real time qPCR. Even though none of them showed expression differences greater than 2.6 fold, the six miRNAs with the highest fold-change score (miR-103, miR-107, miR-21, miR-29a, miR-30c, and miR-652) were further investigated but no significant differences between the control and DM1 conditions were found for any of them. # Materials and Methods ## Sample collection and serum isolation This study was approved by the Ethics Committee at the University of Valencia. All blood samples were taken after specific written informed consent to participate in the present study. All individuals were Subjects with DM1 were ambulatory adults with proven CTG expansions. Peripheral blood samples were obtained by venous punctures with a fine bore needle (21 G ¾”) of 26 DM1 and 22 healthy individuals (Tables) and placed in serum collection tubes (BD VACUTAINER SST II ADVANCE). After 10 min centrifugation at 1200 g at room temperature, the serum was aliquoted and kept at -80°C until use. For CTG repeat size determination, genomic DNA was isolated from peripheral blood leucocytes and was processed for Southern blotting with a <sup>32</sup>P-labelled cDNA25 probe or, alternatively, the CTG-repeat region was amplified by PCR using DM101 and DM102 as primers. ## RNA extraction and cDNA synthesis We assayed for the presence of oxyhaemoglobin in the serum samples because haemolysis has been described to affect the levels of certain miRNAs. The absorbance at 414 nm was determined spectrophotometrically and samples with an absorbance higher than 0.2 were discarded, as this is the cutoff at which samples have previously been considered to be haemolysed. Independent total RNA extraction was performed for each serum sample using the miRNeasy Mini kit (Qiagen). Briefly, 500 μL of serum was thawed on ice, centrifuged for 5 min at 3000 g at 4°C and 200 μL of the supernatant serum was taken and mixed with 750 μL QIAzol containing 1.25 mg/mL bacteriophage MS2 RNA as a carrier. The extraction was performed according to the manufacturer’s instructions, except that the final wash (with RPE buffer) was performed three times instead of once. Total RNA was eluted with 50 μL water, and cDNA synthesis was performed with 4 μL of total RNA using the Universal cDNA synthesis II kit (Exiqon). ## MicroRNA expression profiling and validation The miRCURY LNA™ Universal RT microRNA PCR assay and the Serum/Plasma Focus microRNA PCR Panel (Exiqon) was used for miRNA expression profiling. These panels contain primers for the detection of the 175 most-expressed miRNAs in human serum. Each 384-well plate contained 2 complete panels of primers and 2 negative controls; real-time PCR was performed according to the manufacturer’s instructions, and cDNAs from a DM1 patient and a control sample were amplified in parallel in each plate. Expression values were calculated using the 2<sup>-∆∆Ct</sup> method using the mean Ct of miRNAs detected (Ct \< 34) for normalisation. During the validation step, the analysis of expression of these miRNAs used real-time PCRs with specific miRCURY LNA microRNA PCR primers (Exiqon). The GeNorm and Normfinder algorithms were used to find optimal reference genes to normalise the expression of the miRNAs being validated. Expression level determinations were performed using an Applied Biosystems 7900HT Fast Real-Time PCR System. ## Statistical analysis A logarithmic transformation (log2) was used to normalise the expression data in the profiling experiment. Expression differences were analysed using the Student t-test and different methods for multiple-testing correction were applied, including Bonferroni, Benjamini-Hochberg (False Discovery Rate), Westfall-Young, and Benjamini-Yekutieli corrections. Cluster software was used for hierarchical clustering analysis of genes and samples. Euclidean distances and the average linkage method were selected, using the normalised expression values of each miRNA to represent clusters. # Results ## MicroRNA expression profiling in human myotonic dystrophy type 1 serum A total of 175 miRNA expression levels were obtained from each peripheral blood serum total-RNA sample using commercial microRNA PCR panels (Exiqon;). miRNA profiling was initially carried out with peripheral blood from 10 male DM1-affected individuals (aged 51.3 ± 1.6; P01-P10), expressing between 333 and 1333 CTG repeats (in blood samples), and 10 sex and age-matched controls (aged 46.1 ± 1.7; C11-C20) that did not display any neuromuscular disorders. The absorbance at 414 nm was measured in all the samples to discard the possibility of haemolysis, which can occur during blood collection and has a potentially substantial impact on serum miRNA content. Because two samples, P3 and C18, generated a positive result for this parameter (absorbance \> 0.2) they were discarded during data analysis. Expression data from each sample was initially normalised to the mean values of all 175 miRNAs. Statistical analysis of the results (Student t-test) showed 35 miRNAs with a P-value lower than 0.05, of which 24 miRNAs were up- and 11 were downregulated when compared to controls (, ). However, only miR-21 was significantly downregulated in DM1 according to three different statistical corrections (Bonferroni, Benjamini-Hochberg (False Discovery Rate) and Westfall-Young). It is worth mentioning that all the differences in expression levels detected between controls and DM1 patients were relatively low (below 2.6-fold) compared to other described biomarkers. Given the controversy regarding the most appropriate way to normalise data when determining miRNA expression values, in addition to mean normalisation, we also normalised the data to specific miRNA expression levels. For that purpose we used two different algorithms, NormFinder and geNorm, to identify the most stable miRNAs from our study cohort. Therefore we normalised the miR-15a, miR-23a, miR-28-3p, and miR-484 expression levels to the mean of miR-15a, miR-23a, and miR-484 and the mean of miR-15a and miR-28-3p. In most cases, miR-21 was the only miRNA with significantly different expression between the controls and patients. We carried out additional statistical analyses using G\Power software to select additional candidate miRNAs to validate by qPCR. We chose miRNAs with the highest fold-change and with a Power value ∼1, which included miR-21. Considering these parameters, we selected the six most promising miRNAs for validation: miR-103, miR-107, miR-21, miR-29a, miR-30c, and miR-652. ## Expression quantification of six candidate miRNAs in serum We experimentally determined individual expression levels of miR-103, miR-107, miR-21, miR-29a, miR-30c, and miR-652* in the same nine DM1 and nine control serum samples used during the initial profiling. The data were normalised to miR-15a, the most and second-most stable miRNA from all of the samples, according to geNorm and NormFinder, respectively. However, we did not detect statistically-significant differences between DM1 and control samples, not even for miR-21, the only miRNA that was positive after the profiling. Considering that from among all of our results, data supporting miR-21 misexpression was the strongest, we decided to carry out further analyses on it. To prevent potential false-negative results because of the miRNA selected as a normaliser, we normalised miR-21 to the expression of the miRNA with the strongest alteration in the opposite direction because this ratio would be independent of any endogenous control. Taking into account data from the profiling, miR-21 expression was downregulated 2.4 times in DM1 samples while miR-130a was upregulated, with a 2.5-fold change. Next, we used serum samples from 21 DM1 male and female individuals and 17 counterpart controls to quantify expression levels of miR-21 and miR-130a. We confirmed the absence of haemolysis in all the samples by measuring absorbance at 414 nm. However, again, we were unable to detect any significant difference in the miR-21 to miR-130a ratio between controls and DM1 samples. # Discussion The only method available for monitoring the progression of DM1 is clinical assessment provided by semiquantitative scales, which correlates poorly with underlying biological defects. A more targeted strategy which characterises muscle involvement is the transcriptomic analysis of muscle biopsies via invasive techniques. These analyses have led to the recent proposal that suggests that alternative splicing events in skeletal muscle can serve as valid biomarkers for quantifying the severity of DM1 and its likely response to therapy. Nevertheless, the different patterns of muscle involvement in DM1, and the invasive nature of the approach, inherently limits this proposal as a good measurement of outcome. An alternative method for neuromuscular diseases is to use blood miRNAs as biomarkers. Cacchiarelli et al. described three miRNAs that correlated with disease severity in Duchenne muscular dystrophy where, as a consequence of muscle-fibre damage, muscle miRNAs are released into the bloodstream. However, cell membranes remain undamaged in DM1 muscle fibres and, consequently, levels of myomiRs in blood in these patients are not expected to be as dramatically increased as a result of the disease as in Duchenne muscular dystrophy. In the present work we profiled 175 miRNAs in serum samples but we did not observe differences greater than 2.6-fold between DM1 patients and healthy individuals. Owing to the novelty in the use of miRNAs as biomarkers, there is a lack of consensus regarding different technical aspects such as sample quality. However, it was demonstrated that haemolysis can occur during blood collection and this can have a substantial effect on miRNA content in plasma or serum. This fact shows the relevance of good sample quality control for the results obtained when searching for miRNA biomarkers. Initially, we performed the profiling with ten control and ten DM1 patient samples, however one sample from each group was removed because of unacceptable haemolysis levels. After the profiling we identified 35 altered candidate-miRNAs. However, after applying Bonferroni correction only one, miR-21, was statistically different. It was recently published that four muscle-specific miRNAs, miR-1, miR-133a, miR-133b, and miR-206, were altered in serum from DM1 patients. Of these, miR-1, miR-133a, and miR-133b were included in our profiling panel, however, we did not detect differences in their expression levels. There are two aspects that should be considered: Firstly, that in this aforementioned work the data were normalised to miR-16 expression, however other work has shown that levels of this miRNA vary as a function of haemolysis levels because it is one of the most abundant miRNAs in red blood cells, and thus miR-16 levels may be unacceptably influenced by haemolysis. Moreover, it was recently observed that miR-16 is sequestered by long CUG repeats and consequently the amount of free miR-16 in the bloodstream may not be equal in healthy and DM1-affected individuals. Secondly, different methods may identify different sets of altered miRNAs. Indeed, a recent review showed that plasma and serum miRNAs described as breast cancer biomarkers in different publications in the literature do not overlap with each other, and although the exact reasons remain unclear, methodological differences in experimental procedures may be one major cause. Another group, using a similar approach to ours, identified nine miRNAs that were differentially expressed between healthy controls and DM1 individuals. Of those, seven were included in our profiling panel, however, none of them were positive. Of note, Perfetti et al. used plasma samples in their research, whereas we used serum; this is worth mentioning because differences in miRNA and RNA levels in serum vs. plasma have been reported. The authors suggested that miRNAs are released from blood cells into serum during the coagulation process, although they did not identify the reason for this. Therefore, results regarding the biomarkers identified using serum vs. plasma are not comparable. It is also noteworthy that we used different measurement platforms to those employed by Perfetti et al.: the Taqman and Exiqon miRNA qPCR panels, respectively. In this regard, Wang et al. demonstrated that the consistency between results obtained using both platforms is low, finding that from 358 miRNAs, only ∼19% were detected by both platforms, and that Taqman measurements were 6.7 Ct-values higher than those from Exiqon. After miRNA profiling we identified only one miRNA that was differentially expressed: miR-21. We tried to validate five additional miRNAs by qPCR, although the differences between controls and DM1 were not statistically significant, and we did not observe differential expression in any case. qPCR data were normalised to miR-15a because two different algorithms identified it as the most stable miRNA from among all of the samples. In addition, miR-21 expression was assessed as a ratio to miR-130a, however, neither of these analyses revealed significant differences. We used serum samples from males in the profiling, and samples from both genders during the validation where a higher number of samples was needed. However, it is unlikely this had any significant effect on our results because the qPCR results for each gender were similar to the results obtained from the combined sample analysis. In summary, we conclude that, under our reported conditions, the miRNAs miR-103, miR-107, miR-21, miR-29a, miR-30c, and miR-652 are not useful serum biomarkers for DM1. Although the successful use of miRNAs from body fluids as disease-severity and progression biomarkers in other studies represents an encouraging advance, several technical aspects must first be standardised because methodological differences in the experimental procedures seem to be the main reason that data from different studies do not coincide. # Supporting Information We thank Maria Goicoechea for help with the CTG repeat size determination in DM1 patients. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JMFC BL MCAA MPA ALM ALC RA. Performed the experiments: JMFC BL. Analyzed the data: JMFC BL ALC RA. Contributed reagents/materials/analysis tools: ALM. Wrote the paper: AB JMFC. Managed clinical data and processed human samples: MZ.
# Introduction Inflammation has been repeatedly associated with diabetes mellitus, both in cross-sectional and prospective studies. A meta-analysis of 20 prospective cohorts and cross-sectional studies found that raised leukocyte concentrations are associated with higher risk of type 2 diabetes. However, results were inconsistent across studies \[–, \], and the authors pointed out that results from the meta-analysis likely represent an overestimate due to publication bias and inability to control for potential confounders in all studies. Most studies are cross-sectional and few prospective studies have examined the role of leukocytes in the development of diabetes. The Atherosclerosis Risk in Communities (ARIC) study from the USA found elevated leukocyte count to be associated with development of diabetes among 12,330 non-diabetic individuals aged 45–64 years after a mean follow-up of 7 years. The National Health and Nutrition Examination Survey Epidemiologic Follow-up Study (NHEFS) found leukocyte count to be significantly and positively related to diabetes incidence with a dose-response relationship. A high leukocyte count was found to predict impaired glucose tolerance and type 2 diabetes in Pima Indians. In the Cardiovascular Health Study (CHS), only C-reactive protein (CRP) and not leukocyte count was associated with the development of diabetes. Since gene variants are inherited and usually not subject to confounding, it is possible to use genetic polymorphisms as instrumental variables (IV) to explore the relationship between a trait and a certain outcome, without confounding bias from measured and unmeasured risk factors. We used a missense polymorphism (R262W, SNP rs3184504) in the SH2B adaptor protein 3 (SH2B3) gene to examine the effect of leukocyte count on diabetes. The R262W polymorphism is known to be associated with increased concentrations of leukocytes and its sub-populations. The SH2B3 gene codes for a protein that negatively regulates the hematopoiesis in knock-out models. Loss of function of SH2B3 has been identified as a risk factor for leukemia. The aim of this prospective study was to explore the relationship between total and differential leukocyte count and incidence of diabetes in an urban population. # Subjects and Methods ## Study population The population-based cohort Malmö Diet and Cancer study (MDC), from the city of Malmö in southern Sweden, was used in this study. In brief, all women born between 1923 and 1950 and men born between 1923 and 1945 living in Malmö city were invited to the MDC study during the period March 1991 to September 1996. A total of 30 447 individuals participated in clinical examinations at the screening center and filled in a self-administered questionnaire out of an eligible population of ≈74 000 individuals <http://atvb.ahajournals.org/content/32/2/533.long-ref-17#ref-17>. DNA was available for 28 767 subjects. Subjects with history of diabetes (n = 1 311) at the baseline examination were excluded. In order to exclude individuals with severe inflammation, analyses were restricted to participants with information on total leukocyte counts less than 20.0 × 10<sup>9</sup> cells/L. In addition, 888 participants without complete information on covariates were excluded. Thus, the final study population in the project consisted of 26 667 subjects (10 364 men (38.9%) and 16 303 women (61.1%), aged 45–73 years. A random subsample from the MDC cohort, the MDC cardiovascular cohort (MDC-CV, n = 6,103), was invited to take part in a study of the epidemiology of carotid artery disease between October 1991 and February 1994. The additional examinations in this sub-cohort included measurements of fasting whole blood glucose, hemoglobin A1c (HbA<sub>1c</sub>) and CRP. The ethics committee at Lund University Lund, Sweden, approved the study (LU 51/90) and all participants provided informed written consent. ## Baseline examinations The examinations were performed by trained nurses at the screening center. Blood pressure (mmHg) was measured using a mercury-column sphygmomanometer after 10 minutes of rest in the supine position. Standing height (m) was measured with a fixed stadiometer calibrated in centimeters. Weight was measured to the nearest 0.1 kg using balance-beam scale with subjects wearing light clothing and no shoes. Body mass index (BMI) was calculated as weight (kg) divided by the square of the height (m<sup>2</sup>). Waist was measured as the circumference (cm) between the lowest rib margin and iliac crest. Information on family history of diabetes, current use of lipid-lowering, blood pressure-lowering or anti-diabetic medications, smoking habits, leisure-time physical activity, education level and marital status were obtained from a self- administered questionnaire. Family history of diabetes was defined as known diabetes in at least one first-degree relative. History of myocardial infarction or stroke at the baseline examination was retrieved from the Swedish Hospital Discharge Register and the Stroke register in Malmö. Subjects were categorized into current smokers (i.e. those who smoked regularly or occasionally) or non- smokers (i.e. former smokers and never smokers). Low level of leisure-time physical activity was defined as the lowest quartile of a score revealed through 18 questions covering a range of activities in the 4 seasons. The evaluation of the questionnaire has been previously reported. Educational level was divided into three groups: school year \<9, 9–12 and \> 12, respectively. Marital status was categorized into married or not. ## Laboratory measurements Total and differential leukocyte counts (neutrophils, lymphocytes, and a group of mixed cell types including monocytes, eosinophils and basophils) were counted in heparinized blood samples using a SYSMEX K1000 automatic counter (Sysmex Europe, Norderstedt, Germany). The analyses were performed consecutively at the time of the screening examination, at the central laboratory of Malmö University Hospital. HbA1c and whole blood glucose was measured according to standard procedures at the Department of Clinical Chemistry. HbA<sub>1c</sub> was measured by ion exchange chromatography, with reference values of 3.9–5.3% in non-diabetic individuals. Insulin was measured by a radioimmunoassay in mIU/ L and the HOMA index was calculated as fasting insulin\glucose/22.5. CRP was analyzed using a high-sensitive assay, Tina-quant<sup>®</sup> CRP latex assay (Roche Diagnostics, Basel, Switzerland). ### Incidence of diabetes All subjects were followed from the baseline examination until first diagnosis of diabetes, death, emigration from Sweden or December 31<sup>st</sup>, 2009, whichever came first. Cases of new-onset diabetes in the MDC cohort were identified from several sources. In short, incident diabetes was identified from the Malmö HbA<sub>1c</sub> register (MHR) (56% of all cases), the Swedish National Diabetes Register (NDR) (14%), the Swedish inpatient register (40%), the Swedish outpatient register (38%), the nationwide Swedish drug prescription register (65%) and the regional Diabetes 2000 register of the Skåne region (22%). In addition, 44% of cases were identified at re-examinations of the cohort. At least two independent sources confirmed the diagnosis for 71.6% of the cases, and 53% of cases were identified in three independent data sources. NDR and the Diabetes 2000 register required a physician´s diagnosis according to established diagnostic criteria (fasting plasma glucose concentration of \> = 7.0 mmol/L, which corresponds to a fasting whole blood glucose of \> = 6.1 mmol/L, measured on 2 different occasions). The MHR at the Department of Clinical Chemistry, Malmö University Hospital, analyzed and recorded all HbA<sub>1c</sub> samples taken in institutional and non-institutional care in the greater Malmö area from 1988 onwards. Individuals who had at least two HbA<sub>1c</sub> recordings \> = 6.0% in the MHR with the Swedish Mono-S standardization system (corresponding to 7.0% according to the US National Glycohemoglobin Standardization Program) after the baseline examination were defined as incident diabetes cases. ## Polymorphism in the SH2B3* gene The missense polymorphism R262W (rs3184504), which previously has been associated with concentrations of leukocytes, neutrophils and lymphocytes was genotyped in 24,489 subjects. Based on results from previous GWAS studies of leukocyte count in subjects with European ancestry, we also tested whether polymorphisms in the 17q21 locus (rs4065321, rs3859192, rs9916158, rs4794822 and rs17609240) were useful as instrumental variables in this study. These genotypes were studied in relation to leukocyte count in a subsample of 4600 individuals, but none of the SNPs from the 17q21 locus reached p\<0.5 x10<sup>-8</sup> for the association with leukocyte count. Since we also found a significant inverse relationship between insulin resistance and the 17q21 locus, both in the literature and in our data, we decided to use the R262W polymorphism only. DNA was extracted from peripheral blood cells and assigned to batches without regard to disease status or personal identity. Batches were genotyped using real-time polymerase chain reaction (PCR) with 2.5 ng DNA as PCR template for allelic discrimination on an ABI7900HT (Life Technologies, Carlsbad, CA, USA). Genotype calls were obtained using SDS 2.3 software (Life Technologies, Carlsbad, CA, USA) with manual inspection and curation of fluorescence intensity plots. The call rate was \>90%. The Hardy-Weinberg equilibrium (HWE) was calculated using an online calculator (chi-2 = 3.42, df = 2, p = 0.06). ## Statistical analysis Subjects were categorized into sex-specific quartiles of leukocyte concentrations, i.e., four groups with the same proportion of men and women in each group. One-way analysis of variance (ANOVA) and logistic regression was used to assess cross-sectional relationships of leukocyte count to diabetes risk factors. A general linear model was used to adjust glucose and HbA<sub>1c</sub> for potential confounding factors, in categories of the R262W polymorphism. Cox proportional hazards regression was used to examine hazard ratios (HR) with 95% confidence interval (CI) for incidence of diabetes by total and differential leukocyte counts (neutrophils, lymphocytes, and mixed cells) using the lowest quartile as the reference category. Time axis was follow-up time until death, emigration, incident diabetes or end of follow-up. The results were adjusted for age, sex, BMI and family history of diabetes in the basic model. Secondly, we also adjusted for waist circumference, systolic blood pressure, using of blood pressure- or lipid-lowering medication, history of cardiovascular disease, smoking habits, leisure physical activity, educational level and marital status. CRP was added to the covariates in an additional analysis in the MDC-CV sub- cohort. Possible interaction between leukocyte count and risk factors for diabetes was explored by introducing interaction terms in the fully adjusted multivariate model. The Kaplan-Meier curve was used to demonstrate incidence of diabetes in relation to total leukocyte count during the follow-up. The association between leukocyte count, fasting blood glucose and HbA1c, respectively, was analyzed using general linear model adjusted for age, sex and BMI. Model discrimination was estimated with Harrell’s C-statistics. Sensitivity analyses were performed after excluding individuals having a cold or lung infection within two weeks before the baseline examination. We also explored the effect of non-steroid anti-inflammatory drug (NSAID) medication on the relationships between leukocytes and diabetes. A Mendelian randomization analysis path diagram was presented in. R262W was used as an instrumental variable for leukocyte count to study incidence of diabetes. Mendelian randomization analysis was performed with two-stage least squares regression (2SLS) using the ivreg2 command in STATA. The genetic instrument was validated for association with total and differential leukocyte counts using linear or logistic regression models. A power analysis was performed for the association between R262W and incidence of diabetes. With α = 0.05, there was 64% power to detect a significant relationship, assuming a HR of 1.25 per standard deviation increment of leukocytes, i.e., the age and sex-adjusted HR in the present cohort. All analyses were performed using IBM SPSS statistics (version 20; IBM Svenska AB, Stockholm, Sweden) and STATA12 (Stata Corp, College Station, TX, USA). # Results ## Baseline Characteristics The mean total leukocyte count was 6.37±1.68 x10<sup>9</sup> /L, and the proportion of neutrophils, lymphocytes, and mixed cells were 61%, 31% and 8%, respectively. The relationships between sex-specific quartiles of total leukocyte count and risk factors for diabetes are presented for all subjects in, and separately for men and women in and. Increased leukocyte count was associated with age, BMI, waist circumference, systolic blood pressure, antihypertensive- or lipid-lowering medication, prevalent cardiovascular disease, smoking habits, low physical activity, education level and being married. In addition, HbA<sub>1c</sub>, glucose and insulin were positively and significantly associated with leukocyte count (p\<0.001) in the subgroup. ## Incidence of diabetes in relation to total and differential leukocyte counts During a mean follow-up of 14 years, 2 946 subjects (1 521 men and 1 425 women, 7.87 per 1000 person-years) developed diabetes. Kaplan-Meier curves of diabetes free survival in relation to sex-specific quartiles of total leukocyte count is shown in Figs and. Incidence of diabetes was significantly associated with total and differential leukocyte counts in the basic model 1. After adjustment for potential confounding factors, the association remained significant for total leukocyte (4<sup>th</sup> vs 1<sup>st</sup> quartiles HR: 95% CI; 1.37; 1.22–1.53), neutrophil (1.33; 95% CI: 1.19–1.49) and lymphocyte (1.29; 95% CI: 1.15–1.44) counts, but not for mixed cells (1.04; 95% CI: 0.94–1.15, not shown in table). Male sex, family history of diabetes, high BMI and waist circumference, high systolic blood pressure, use of antihypertensive and lipid- lowering medications, current smoking and low education level were all significantly associated with incidence of diabetes in the final multivariate model. No significant interaction was observed between total leukocyte count and other risk factors for diabetes. Use of non-steroid anti-inflammatory drugs (NSAID) (n = 822) at baseline was added to the multivariable adjusted model in a sensitivity analysis, but the results were essentially unchanged. NSAID was not a significant risk factor for diabetes in the model. We also performed a sensitivity analysis after excluding individuals who reported having a cold or lung infection within two weeks before the baseline examination (n = 20 906). After adjustment for potential confounding factors, the association remained significant for total leukocyte (4th vs 1st quartiles HR: 1.39; 95% CI; 1.22–1.58), neutrophil (1.37; 95% CI: 1.21–1.56) and lymphocyte (1.32; 95% CI: 1.16–1.50) counts, while mixed cells remained non- significant (1.03; 95% CI: 0.92–1.16). In the MDC-CV subcohort, 736 subjects developed diabetes during the follow-up. The HR (95% CI) for total leukocyte count (4<sup>th</sup> vs 1<sup>st</sup> quartile) was 1.95 (1.55–2.46, P for trend \<0.001) in the basic model. The HR was reduced to 1.53 (1.20–1.95; P for trend 0.004) taking possible confounders into account and remained significant also after adding CRP into the model (HR: 1.37; 1.05–1.77; P for trend 0.044). Among the leukocyte subpopulations, only neutrophils remained significantly associated with incidence of diabetes after adjustment for CRP (HR: 1.39; 1.08–1.78, P for trend 0.022). The C-statistics value for a model with age, sex and BMI was 0.717 (0.708–0.726) and increased to 0.722 (0.713–0.731) when leukocyte count was added to the model. Leukocyte count significantly improved the discriminatory value (C-statistic) for incidence of diabetes with 0.005 (0.002–0.008) (p\<0.001). ## Polymorphism in the SH2B3 gene The association between the R262W polymorphism and leukocyte count is shown in. The frequency of the minor allele (T) of R262W was 48%. The T allele was strongly associated with increased leukocytes (0.11x10<sup>9</sup> cells/l per T allele, p = 1.14 x10<sup>-12</sup>, F statistics = 50.6), lymphocytes (p = 4.3 x10<sup>-16</sup>), neutrophils (p = 8.0 x10<sup>-6</sup>) and mixed cells (p = 3.0 x10<sup>-6</sup>). There was no association between the R262W polymorphism and BMI (p = 0.154), waist (p = 0.495), systolic blood pressure (p = 0.231) and use of lipid-lowering medication (p = 0.723). We found no statistically significant association between the R262W polymorphism and incidence of diabetes, fasting blood glucose or HbA<sub>1c</sub>. These relationships remained non-significant after full adjustment for potential confounding factors in multivariable Cox regression (diabetes) or general linear models (glucose, HbA<sub>1c</sub>) (all p\>0.288). For glucose, the IV estimator was -0.210 (95% CI: -0.723–0.302) (p = 0.421); for HbA<sub>1c</sub>, the IV estimator was -0.201 (-0.558–0.157) (p = 0.272). For diabetes, the IV estimator was -0.005 (-0.047–0.057)(p = 0.837). # Discussion The present study showed a graded association between concentrations of leukocytes, neutrophils and lymphocytes and risk of developing diabetes among middle-aged subjects, taking many potential confounding factors into account. The results confirm that leukocyte count is a risk factor for incidence of diabetes. However, a missense polymorphism in the SH2B3 gene, strongly associated with leukocyte count, was not related to glucose, HbA<sub>1c</sub> or incidence of diabetes. This suggests that the relationship between leukocytes and diabetes might not be causal. Previous studies have reported that various inflammation markers, e.g., interleukin-6, tumor necrosis factor α (TNFα) and CRP are associated with diabetes It is believed that TNFα contributes to diabetes through its interaction with insulin signaling pathways and beta-cell function. Since human granulocytes secrete TNFα, this could be a possible link between leukocyte count and diabetes. A polymorphism in the IL-6 gene has been associated with total and differential white blood cell counts. Since IL-6 is produced by human mononuclear cells, this suggests that IL-6 might be a common link between leukocyte count and diabetes. Hence, the relationship between leukocytes and diabetes could be related to the actions of various pro-inflammatory cytokines. Observational studies could be limited by unmeasured confounding. However, as the alleles are randomly assigned at meiosis and fixed through the lifetime, genetic association studies are usually not subject to confounding. The R262W polymorphism, which is a non-synonymous SNP located in exon 3 of SH2B3, leads to an amino acid change in the pleckstrin homology domain. SH2B3 regulates cytokine receptor-mediated signaling implicated in leukocyte activation. The R262W polymorphism was strongly associated with leukocyte count (F statistics value = 50.6), but we did not find any association between R262W and diabetes, glucose or HbA1c. The results indicate that the relationship between leukocytes and diabetes might not be causal. However, it should be acknowledged that the leukocyte population is highly complex with many different subpopulations. It remains possible that specific populations of leukocytes could be causally associated with diabetes. In addition, even though the number of participants was high and the SNP can be considered a fairly strong instrument, it is still possible that the statistical power was too small in this study. The minor allele of the R262W polymorphism has been previously associated with increased risk of several autoimmune diseases including type 1 diabetes, multiple sclerosis, blood pressure and MI. Some of these disorders could increase the probability that diabetes is detected and that antihypertensive treatments are prescribed that could increase blood glucose levels. It is not possible to exclude the possibility of pleiotropic effects. However, a potential relationship between the minor allele and increased risk of type 1 diabetes, MI and hypertension should increase the risk of diabetes and cannot explain the negative results in this study. ## Strength and limitations The strength of the study was the large numbers of subjects and events during a long follow-up period. A limitation of the present study is lack of information on type of diabetes. Participants were 45–73 years old and non-diabetic at the baseline examination. It can be assumed that almost all incident cases developed type 2 diabetes, since type 1 diabetes usually has an early onset, and were excluded from analyses as prevalent cases. New cases of diabetes were identified from several independent sources. The registers of out- and in-patients cover all hospital visits in the country and the pharmaceutical register covers all filled prescriptions from all pharmacies in Sweden since 2005. The HbA<sub>1c</sub> register covers the population in the city of Malmö. The relationship between leukocytes and diabetes was largely the same for each of the data sources. Diabetes can go undetected for several years and individuals that do not seek medical care will be missed. However, the coverage of the registers is very good and we have no reason to question the case validity. Lack of information about change of exposure during the follow-up (e.g. weight change, quitting smoking, new medication, etc.) was another possible limitation. A further limitation is the potential pleiotropy for the SH2B3 missense variant, which is associated with many traits. Although this would be more relevant in the context of a positive finding, it could be hypothesized that a pleiotropic effect could attenuate any association with diabetes. So-called canalization, i.e., compensatory mechanisms that counterbalance the effects of the genetic instrument, is another possibility that hypothetically could explain the negative results for the R262W polymorphism. Although the leukocyte count significantly increased the model discrimination in terms of C-statistics, it is uncertain whether measurements of leukocytes could improve prediction of future diabetes in clinical practice. However, further studies are needed to confirm this. In conclusion, increased leukocyte counts are associated with incidence of diabetes. However, the negative findings for the R262W polymorphism suggest that the associations may not be causal, although limitations in statistical power and balancing pleiotropic effects cannot be excluded. Further studies are needed for replication of this finding in other cohorts. # Supporting Information The Swedish National Diabetes Register (NDR), the Malmö HbA<sub>1c</sub> register (MHR), the Diabetes 2000 registers and the National Board of Health and Welfare are acknowledged for valuable assistance in retrieval of diabetes end- points. The study was supported by grants from the Swedish Heart and Lung foundation and the Swedish Research council. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YB JGS PN OM BH GE. Performed the experiments: YB JGS GE. Analyzed the data: YB JGS GE. Contributed reagents/materials/analysis tools: PN OM BH. Wrote the paper: YB JGS PN OM BH GE.
# Introduction Generative models (GMs) are a class of Machine Learning (ML) model which excel in a wide variety of tasks. The optimization of a GM finds a function $\mathcal{G}$ that maps a set of M latent variables in latent space to a set of d variables in real space representing the data of interest (e.g., sets of images, music, videos, etc.), i.e. $\left. \mathcal{G}:\mathcal{R}^{M}\rightarrow\mathcal{R}^{d} \right.$ where d \>\> M \> 1. When building a GM, we first define the support of the latent variables, then obtain the function $\mathcal{G}$ by iteratively optimizing a loss function. Loss function choice depends on application, e.g., maximum log- likelihood is common in Bayesian statistics, Kullback–Leibler divergence is common for variational autoencoders (VAEs), and the Jensen-Shannon entropy and the Wasserstein distance are common with generative adversarial networks (GANs). The last two models are deep learning models. Deep learning is a field in artificial intelligence which has had great success in recent years and has pervaded many fields in science and health as well as our day to day lives. When we fit a latent variable model to a data set, we are finding a description of the data in terms of “independent components”. Latent variables, \\|z<sub>i</sub>〉, have a simple distribution, often a separable distribution (i.e., $P\left( \left\{ z_{i} \right\}_{i = 1}^{M} \right) = \prod_{i = 1}^{M}P\left( z_{i} \right)$). Often the latent representation of data lives in a simpler manifold than the original data while preserving relevant information. There are many examples of latent representation used to understand or describe more complicated features, ranging from statistical methods like Latent Class Analysis to examples in statistical physics and condensed matter such as order parameters for phase classification and even the long standing problem of the genotype-phenotype where the genome is taken as the latent representation of the phenotype. Other examples are, for instance, Ref. proposes a time-frequency representation of a signal that allows the reconstruction of the original signal, which relies in what they define as “consensus”. Their proposed method generates sharp representations for complex signals. Deep neural networks can function as surrogate propagators for time evolution of physical systems. While the latent variables are constructed to be independent identically distributed (i.i.d.) random variables, the training process entangle these latent variables. Latent variable disentanglement is an active area of research employing a wide variety of methods. For instance, in Ref., the authors train a GAN including the generator’s Hessian as a regularizer in the loss function, leading, in optimum conditions, to linearly independent latent variables, where each latent variable independently controls the strength of a single feature. Ref. constructs a set of quantized vectors in the latent space using a VAE, known as vector quantized variational autoencoder (VQ-VAE). Each quantized vector highlights a specific feature of the data set. This approach has been used in OpenAI’s jukebox. A major drawback of these approaches is the lack of freedom in relating specific features in real space with specific latent space directions. This can be overcome by conditionalizing the generative model. However, conditionalization can reduce the latent space smoothness and interpolation capacity, since the condition is usually enforced by means of discrete vectors as opposed to a continuous random latent vector. Diffusion- based models have shown they can equate to GANs in performance and have become highly popular in recent times. Here we propose a method to relate a specific chosen labeled feature with specific directions in latent space such that these directions are linearly independent. Having a set of linearly-independent latent vectors associated with specific labeled features allows us to define operators that act on latent space (e.g. a rotation matrix) and correspond to feature transformations in real space. For instance, suppose a given data set in real space corresponds to the states of a molecular dynamic simulation, i.e., the i-th data point in the data set can be the positions of the molecules at time t<sub>i</sub>, \\|x<sub>i</sub>〉 → \\|x(t<sub>i</sub>)〉, where \\|x(t<sub>i</sub>)〉 is a vector. Let us suppose that $\left\| x \right.\left( t_{i} \right){\rangle\left. = \mathcal{G} \right\|}z_{i}\rangle$ and $\left\| x \right.\left( t_{i} + \Delta t \right){\rangle\left. = \mathcal{G} \right\|}z_{j}\rangle$, as depicted in. How can we construct an operator in latent space, $\mathcal{O}_{\Delta t}$, such that $\|z_{j}{\rangle =}\mathcal{O}_{\Delta t}\left\| z_{k} \right\rangle$?. For this construction to be possible, we argue the operator $\mathcal{G}$ must be locally linear. Furthermore, in order to build the operator $\mathcal{O}$, we need a basis that spans latent space. While linearity might seem counterintuitive given how NNs work, growing evidence suggests such linearity in practice. For instance, on the one hand there is an ongoing debate on how deep should a NN be to perform a specific task, on the other hand, it has been proposed the equivalence between deep NNs and shallow wide NNs. For at least one image-related GAN, simple vector arithmetic in latent space leads to feature transformations in real space (e.g., removal of sunglasses, change in hair color, gender, etc.). However, a complete understanding on how specific features in real space map to latent space and how are these features arranged in latent space (latent space topology) or why some GANs’ latent space behave like linear operators is lacking. It is believed that the latent representation of data with a given labeled feature forms a cluster. However, the tools employed to show this clustering effect quite often consist in a dimensional reduction e.g., t-SNE which collapses the latent representation into two or three dimensions. Other methods include principal component analysis, latent component analysis and important component analysis. Our method does not collapse or reduce the latent space, allowing us to inspect latent space topology by spanning all latent space directions. We strongly believe the need of a set of basis vectors for understanding the topology of the latent space. Given the typical high- dimensionality of the latent space, we employ the Gram-Schmidt method to construct linearly independent vectors from a set of vectors that map to specific features. This approach enables us to visualize the feature entanglement in the latent space. We contend that our work contributes to a better understanding of latent space topology in two key ways: 1) through the method itself, which involves constructing a set of basis vectors in the latent space that map to specific features in the real space using Gram-Schmidt, and 2) by possessing the latent space basis vectors that map to specific features in the real space, which enables data manipulation in the latent space via linear algebra. As a proof of concept, we demonstrate the method by applying it to MNIST. In the next section we introduce our mathematical method and notation and apply the method to the MNIST data set. In the Results section we show how we can use this method to understand the topology of the latent space by performing classification via principal component analysis; we apply this method to denoise images; and finally we show how to perform matrix operations in latent space which map to image transformations in real space. We discuss future steps and limitations in the last section. # Methods and materials Assume a vector space which we call real space and denote the vectors in this space \\|x〉 with \\|x〉 ∈ ℜ<sup>d</sup>. Assume a set $\left\{ \middle\| x_{i} \right\rangle\}_{i = 1}^{N}$, which we call the dataset with N the dataset size. Similarly, we assume a vector space, which we call the latent space and denote these vectors \\|z〉 with \\|z〉 ∈ ℜ<sup>M</sup> (in general, M ≤ d). We also consider three deep neural networks, a Generator $\mathcal{G}$, an Encoder $\mathcal{E}$ and a Classifier $\mathcal{C}$. We can interpret $\mathcal{G}$ as a projector from latent space to real space, i.e., $\|x_{i}{\rangle\left. = \mathcal{G} \right\|}z_{i}\rangle$, and interpret $\mathcal{E}$ as the inverse of $\mathcal{G}$. However, this last statement has to be taken with a grain of salt, due to how variational autoencoders work. In fact, if $\|z_{a}{\rangle\left. = \mathcal{E} \right\|}x_{i}\rangle$ and $\|z_{a^{\prime}}{\rangle\left. = \mathcal{E} \right\|}x_{i}\rangle$, in general, \\|z<sub>a</sub>〉 ≠ \\|z<sub>a′</sub>〉, since these vectors are i.i.d. random vectors, sampled from a Gaussian distribution with mean and standard deviation dependent on \\|x<sub>i</sub>〉 (the correct mathematical notation to use would be $\|z_{a}{\rangle\left. \sim \mathcal{N}\left( \middle\| \mu \right\rangle, \middle\| \sigma \right\rangle)}$, where $\mathcal{N}$ is a multivariate Gaussian distribution with mean and standard deviation \\|μ〉 and \\|σ〉, respectively, which depend on $\left. \mathcal{E} \right\|x_{i}\rangle$). Finally, the Classifier projects real-space vectors into the label space, i.e., $\|y_{k}{\rangle\left. = \mathcal{C} \right\|}x_{i}\rangle$, where \\|y<sub>k</sub>〉 ∈ L, where L denotes the label space. We assume that each vector \\|y<sub>k</sub>〉 is a one-hot-vector. The length of \\|y<sub>k</sub>〉 equals the number of labels \\|L\\| = l and k = 1, …, l. Henceforth, we assume that l \< M. We define $\left\{ \middle\| \xi_{i} \right\rangle\}_{i = 1}^{M}$ to be a set of basis vectors in latent space such that the inner product between them yields 〈ξ<sub>i</sub>\\|ξ<sub>j</sub>〉 = Cδ<sub>ij</sub>, where C is the norm and δ<sub>ij</sub> is the Kronecker delta function. Henceforth we call the set of basis vectors $\left\{ \middle\| \xi_{i} \right\rangle\}_{i = 1}^{M}$ the quasi-eigenvectors since they form a basis and each one represents a feature state in latent space. Notice that we can define the operator $\mathcal{A} = \sum_{j = 1}^{M}\|\xi_{j}{\rangle\langle}\xi_{j}\|$ (here \\|κ〉〈γ\\| denotes the outer product between vectors \\|κ〉 and \\|γ〉), which implies $\left. \mathcal{A} \right\|\xi_{i}{\rangle\left. = C \right\|}\xi_{i}\rangle$. Any vector in latent space can be expressed as a linear superposition of these quasi-eigenvectors, viz, $$\begin{array}{r} {\left\| z \right\rangle = \sum\limits_{j = 1}^{M}c_{j}\left\| \xi_{j} \right\rangle\;.} \\ \end{array}$$ where \\|c<sub>i</sub>\\| = \\|〈ξ<sub>i</sub>\\|z〉\\| is the amplitude of \\|z〉 with respect to \\|ξ<sub>i</sub>〉 and gives a measure of \\|z〉’s projection with the quasi-eigenvector \\|ξ<sub>i</sub>〉. Constructing a set of basis vectors is straightforward. However, we wish each labeled feature to corresponds one-to-one with a quasi-eigenvector. Since we are assuming that l \< M, there will be a set of quasi-eigenvectors that do not correspond to any labeled feature. To obtain a set of orthogonal quasi-eigenvectors, we use the Gram-Schmidt method. Specifically: 1. We train the GAN, which is composed by two NNs, namely, the Generator and the Discriminator, using the training set $\left\{ \middle\| x_{i} \right\rangle\}_{i = 1}^{N}$ as in Ref.. 2. We train the Classifier independently, using the training set. 3. We train a VAE using the trained Generator as the decoder. We also use the Classifier to classify the output of the VAE. We include in the loss function a regularizer $\lambda \cdot \mathcal{L}_{class}$, where λ is a hyperparameter and $\mathcal{L}_{class}$ denotes the Classifier’s loss function. At this stage, we only train the Encoder, keeping the Generator and Classifier fixed. There are several options to choose from for the $\mathcal{L}_{class}$ loss function. In our case, we used the Cross Entropy with a softmax activation function, i.e., $$\begin{array}{r} {\mathcal{L}_{class}\left. \left( \middle\| y \right\rangle, \right\|y^{GT}{\rangle)} = - \sum\limits_{i = 1}^{L}y_{i}^{GT}\mspace{360mu}\text{log}\mspace{360mu}\frac{e^{y_{i}}}{\sum_{j = 1}^{L}e^{y_{j}}}\;,} \\ \end{array}$$ where y<sub>i</sub> and $y_{i}^{GT}$ are the ith components of the vectors \\|y〉 and \\|y<sup>GT</sup>〉, respectively, and \\|y<sup>GT</sup>〉 is the ground truth vector. 4. Define n to be an integer such that M = n × l. Then, for each label, we allocate n sets of latent vectors and we denote each of these latent vectors as $\|z_{\alpha,i}^{k}\rangle$, where α denotes the label, i = 1, …, n and k = 1, …, V. Here V is the number of elements (latent vectors) in each set corresponding to the pair (i, α) ∈ n × l. We build each of these sets $\left\{ \middle\| z_{\alpha,i}^{k} \right\rangle\}_{k = 1}^{V}$ in two ways: Using the training set, we encode each vector $\|x_{a}{\rangle\left. \rightarrow \right\|}z_{a}{\rangle\left. = \mathcal{E} \right\|}x_{a}\rangle$, then we decode the latent vector, i.e., $\|z_{a}{\rangle\left. \rightarrow \right\|}x_{a}{\rangle\left. = \mathcal{G} \right\|}z_{a}\rangle$, and then we classify the output, i.e., $\|x_{a}{\rangle\left. \rightarrow \right\|}y_{a}{\rangle\left. = \mathcal{C} \right\|}x_{a}\rangle$. For each label l, there is a set of latent vectors. The goal is to have a large number of the latent vectors representation of the data set arranged by label. Due to the large latent space dimensionality, we may require additional latent vectors besides those generated directly by encoding the training set. For this reason, we do the following. We generate random latent vectors and map each of these latent vectors to their labels using the Generator and the Classifier as in 4(a), i.e., once we generated the random latent vector \\|z<sub>a′</sub>〉 using a random multivariate Gaussian generator, we project it to real space $\|z_{a^{\prime}}{\rangle\left. \rightarrow \right\|}x_{a^{\prime}}{\rangle\left. = \mathcal{G} \right\|}z_{a^{\prime}}\rangle$, and then we classify the output, i.e., $\|x_{a^{\prime}}{\rangle\left. \rightarrow \right\|}y_{a^{\prime}}{\rangle\left. = \mathcal{C} \right\|}x_{a^{\prime}}\rangle$. Notice that with this approach we can generate as many latent vectors as desired. We denote as V the number of latent vectors per set (i, α). 5. We take the average over V for each set of latent vectors $\left\{ \middle\| z_{\alpha,i} \right\rangle\}_{k = 1}^{V}$ and denote that average \\|η〉<sub>α,i</sub>, i.e., $$\begin{array}{r} {\left\| \eta \right\rangle_{\alpha,i} = \frac{1}{V}\sum\limits_{j = 1}^{V}\left\| z_{\alpha,i}^{j} \right\rangle\;.} \\ \end{array}$$ It is worth noticing that since the latent vectors are sampled from a multivariate Gaussian distribution, the average \\|η<sub>α,i</sub>〉 is finite and unbiased. By defining operators in latent space in terms of outer products of the \\|η<sub>α,i</sub>〉 vectors, these latent space operators will have encoded in them the set of latent vectors $\|z_{\alpha,i}^{k}\rangle$. 6. To impose orthogonality, we use the Gram-Schmidt method. Thus, from the vectors \\|η<sub>α,i</sub>〉 we generate a set of quasi-eigenvectors \\|ξ〉<sub>α,i</sub>, i.e., $$\begin{array}{rcl} \left\| \xi \right\rangle_{1,1} & = & \left\| \eta \right\rangle_{1,1} \\ \end{array}$$ $$\begin{array}{rcl} \left\| \xi \right\rangle_{2,1} & = & {\left\| \eta \right\rangle_{2,1} - \frac{{}_{2,1}\left\langle \eta \middle\| \xi \right\rangle_{1,1}}{{}_{1,1}\left\langle \xi \middle\| \xi \right\rangle_{1,1}}\left\| \xi \right\rangle_{1,1}} \\ \end{array}$$ $$\begin{array}{l} \ldots \\ \end{array}$$ $$\begin{array}{rcl} \left\| \xi \right\rangle_{l,n} & = & {\left\| \eta \right\rangle_{l,n} - \sum\limits_{\alpha = 1}^{l - 1}\sum\limits_{i = 1}^{n - 1}\frac{{}_{l,n}\left\langle \eta \middle\| \xi \right\rangle_{\alpha,i}}{{}_{\alpha,i}\left\langle \xi \middle\| \xi \right\rangle_{\alpha,i}}\left\| \xi \right\rangle_{\alpha,i}\;.} \\ \end{array}$$ Such that: $$\begin{array}{r} {{}_{\alpha,i}\left\langle \xi \middle\| \xi \right\rangle_{\beta,j} = C\delta_{\alpha\beta}\delta_{ij}} \\ \end{array}$$ In, C is the value of the norm. The set of quasi-eigenvectors $\left\{ \middle\| \xi \right\rangle_{\alpha,i}\}_{\alpha = 1,i = 1}^{l,n}$ span the latent space and, as we will show, a subset of them map to specific features. The key point is that the set of quasi-eigenvectors form a basis set in latent space and each direction corresponds to a feature in real space. This structure allows us to give a better topological description of latent space, i.e., how does labeled features map to latent space similar to how molecular configurations map to the energy landscape. In addition, we can use the set of quasi-eigenvectors as tools for classification, denoising and topological transformations. We demonstrate these applications next using the MNIST dataset. ## Applying method to MNIST We trained a GAN, a Classifier and a VAE using the MNIST dataset which has 60k and 10k one-channel images in the training and test set, respectively, with dimensions 28 × 28 pixels. In we show a sample of the dataset. The MNIST dataset can be found in many machine learning packages (e.g., PyTorch, Flux for Julia, etc.) as well as in. We fixed the batch size to 25 and number of epochs to 500 during all training runs. We trained the GAN using the training set, used the Jensen-Shannon entropy as the loss function, the ADAM optimizer with hyperparameters η = 0.0002, β<sub>1</sub> = 0.9, β<sub>2</sub> = 0.999 for both the Generator and the Discriminator, fixed the latent space dimensionality to M = 100 and sampled the random latent vectors from a multivariate Gaussian distribution centered at the origin with standard deviation equal to 1 in all M dimensions. Independently, we trained a Classifier using the training set, used crossentropy as loss function and a softmax as the activation function in the last layer, the ADAM optimizer with hyperparameters η = 3 ⋅ 10<sup>−5</sup>, β<sub>1</sub> = 0.5, β<sub>2</sub> = 0.99. The accuracy of the classifier on the test set reached ≈98.9%. Using the training set, we then trained the Encoder in a VAE and used the trained Generator as the Decoder. We used as loss function the Kullback–Leibler divergence and the hinge loss function. We also added as a regularizer the Classifier’s loss function and the Lagrange multiplier, λ, as hyperparameter set to λ = 100. During the training of the Encoder, we kept both the Generator and the Classifier fixed. In we show the training results. To train the NNs we used Flux in Julia and the code can be found in Ref.. The latent space dimension is M = 100, while the number of labels is \\|L\\| = 10. Thus, following step 4, for each label we generated n = M/\\|L\\| sets of latent vectors, each set containing V = 5000 latent vectors. In we show a sample of latent vectors for labels 0, 1, 2, 6, 7 and 8, projected to real space using the Generator $\mathcal{G}$. Then we take the average over each set as in step 5. We checked that the average and standard deviation over each of the entries in the set of vectors {\\|η〉<sub>α,i</sub>}<sub>α,i</sub> converges. Interestingly, when taking the average over the set of latent vectors corresponding to a label and projecting back to real space, the label holds. For instance, in we show the projected image of the average over V for each set of latent vectors $\left\{ \middle\| z_{\alpha,i} \right\rangle\}_{k = 1}^{V}$ in the case where the latent vectors were obtained following step 4(a), whereas corresponds to the case following step 4(b). We have also plotted the probability density function (PDF) per label in latent space for both cases and added a Gaussian distribution with mean and standard deviation equal to 0 and 1, respectively, for reference. Notice that the PDF in is shifted away from the Normal distribution, whereas in all PDFs are bounded by the Normal distribution, because latent vectors generated directly from latent space are, by definition, sampled from a multivariate Gaussian distribution with mean and standard deviation equal to 0 and 1, respectively. On the contrary, encoding real space vectors yields Gaussian vectors overall (i.e., the PDF over all latent vectors over all labels yields a Gaussian distribution, by definition) but the mean and standard deviation can differ from 0 and 1, respectively. Step 4(a) gives robustness to this method and step 4(b) allows us to generate as many latent vectors as wanted with a specific label. Since the latent space dimension is M = 100, we need M averaged latent vectors \\|η〉<sub>α,i</sub> to generate M orthogonal latent vectors. Since the number of labels is α = {0, …, \\|L\\| − 1}, then n = 10. To this end, we generate one set (i.e., i = 1) following step 4(a) and nine sets (i.e., i = 2, 3, …, n) following step 4(b). shows the projection to real space of all the \\|η〉<sub>α,i</sub> vectors while shows the inner product <sub>α,i</sub>〈η\\|η〉<sub>α′,i′</sub> as a heatmap, which shows they are non-orthogonal. At this point, we have M vectors \\|η〉<sub>α,i</sub> in latent space each i) composed of the sum of V latent vectors, and ii) maps to a specific feature in real space (the image of a number). However, these vectors are not orthogonal. Using the Gram-Schmidt method described in step 6, we obtain a set of vectors, \\|ξ〉<sub>α,i</sub>, in latent space such that each \\|ξ〉<sub>α,i</sub> vector i) encodes V latent vectors, ii) maps to a specific labeled feature and iii) the \\|ξ〉<sub>α,i</sub> vectors are orthogonal, as shown). Since the Generator was trained using random vectors sampled from a multivariate Gaussian distribution centered at zero with standard deviation 1, the value of the norm of any random latent vector will be 〈z\\|z〉 ≈ M. Therefore, we fixed the norm of the quasi-eigenvectors to C = M. Notice that while the non-orthogonal vectors \\|η〉<sub>α,i</sub> for the MNIST GAN map to sharp images of easily-identifiable numbers in real space, not all quasi-eigenvectors map to images of numbers in real space. Only a few of the M linearly-independent directions in latent space (≈ 20) project to images of numbers in real space. We will show how to apply this property of the quasi- eigenvectors to the MNIST test set to classify images in latent space and to denoise real-space images. We also show how to build a rotation operator in latent space that generates feature transformations in real space. # Results ## Using LSD as a classifier in latent space We can express any latent vector \\|z〉, in terms of the quasi-eigenvectors, viz. $$\begin{array}{r} {\left\| z \right\rangle = \sum\limits_{k = 1}^{M}c_{k}\left\| \xi_{k} \right\rangle\;,} \\ \end{array}$$ where the coefficients c<sub>k</sub> are given by, $$\begin{array}{r} {c_{k} = \left\langle \xi_{k} \middle\| z \right\rangle/C\;.} \\ \end{array}$$ Similar to principal component analysis, we are interested in how much information about an image is encoded in the quasi-eigenvector with the largest amplitude \\|c<sub>i</sub>\\|. We encode images from the MNIST test set into latent space, then express the latent vectors in terms of the quasi-eigenvectors (we call this expression latent spectral decomposition or LSD) and find the maximum amplitude \\|c<sub>i</sub>\\| for each latent vector. Recall that the amplitude \\|c<sub>i</sub>\\| is a measure of the projection of the latent vector with respect to the quasi-eigenvector \\|ξ<sub>i</sub>〉. Thus, the largest amplitude corresponds to the quasi-eigenvector that contributes the most to the latent vector. Since the quasi-eigenvectors are associated with labeled features in real space, we use the largest amplitude as a way to classify the image. shows a sample batch of 25 images. The blue dots corresponds to the true labels (see y axis), while the green (red) dots correspond to the case where label associated with the quasi-eigenvector with the largest amplitude is the correct (incorrect) label. In this batch, only batch elements 9 and 22 have true labels that do not agree with the label of the quasi-eigenvalue of the image with the largest amplitude. Since each time the Encoder encodes an image it generates a new random latent vector, then we could obtain a different outcome for batch elements 9 and 22 as well as the rest of the batch elements for each trial. For this reason, we perform an ensemble average over 20 trials. For each trial we take the whole MNIST test set and compute the accuracy of the latent space decomposition (LSD) classifier (see red dots). We also computed the accuracy when the test set is encoded through the Encoder, then decoded through the Generator and finally classified (see blue dots). We have included the accuracy of the trained Classifier in as an upper bound. While the trained Classifier has an accuracy of 98.8%, the LSD classifier has an average accuracy of ∼92%. This difference in accuracy, however, should not be interpreted as showing that the latent-space classifier does a poor job, but that the dominant few quasi-eigenvectors carry most of the information in latent space regarding the individual test-set images. In fact, the encoded 99% of the test-set data requires only the 10 linearly-independent directions in set 1, i.e., the largest amplitude correspond to quasi-eigenvectors in the first set. Suppose that when we perform the LSD, we sort the amplitudes such that \\|c<sub>1</sub>\\| \> \\|c<sub>2</sub>\\| \> … \> \\|c<sub>M</sub>\\| and ask the position of the ground-truth label? As previously mentioned, in 92% of the cases the ground-truth label corresponds to the first position (i.e., \\|c<sub>1</sub>\\|). In 5% of the cases the ground truth label corresponds to the second largest amplitude (i.e., \\|c<sub>2</sub>\\|). In we have plotted the cumulative of the probability for the ground-truth label being any of the first n positions. The dashed red line corresponds to the trained Classifier accuracy. Notice that the probability of the label being in position 1, 2, 3 or 4 of the LSD equals the accuracy of the trained classifier, i.e., in 98.9% of the MNIST test-set images the ground truth label is associated to a quasi- eigenvector such that the associated coefficient is either c<sub>1</sub>, c<sub>2</sub>, c<sub>3</sub> or c<sub>4</sub>. In this sense, it is possible that even when the amplitude of the quasi-eigenvector associated to the ground-truth label is not the largest one, rather the 2nd or 3rd largest one, then \\|c<sub>1</sub>\\| ≳ \\|c<sub>2</sub>\\| or \\|c<sub>1</sub>\\| ≳ \\|c<sub>2</sub>\\| ≳ \\|c<sub>3</sub>\\|. To test this idea, in we have plotted the normalized amplitude (i.e., \\|c<sub>i</sub>\\|/max{\\|c<sub>j</sub>\\|}) vs the rank (i.e., sorted amplitudes from largest to smallest) for all images in the test set. corresponds to the images where the LSD amplitude of the quasi-eigenvector associated with the ground-truth label is the largest, whereas in the amplitude is the 2nd largest or 3rd largest, respectively. Given the large dataset, in we have plotted the PDFs of the 2nd, 3rd, and 4th largest amplitudes for each of plots. To be clear, from we generated PDFs for the second-, third- and fourth-largest amplitudes in each plot and show the PDFs in, respectively. Notice that when the largest amplitude corresponds to the ground-truth label, the second-, third- and fourth-largest amplitude PDFs are centered below 0.6. When the second-largest amplitude corresponds to the ground-truth label the PDF of the second-largest amplitude is shifted towards 1, while the PDFs of the third- and fourth-largest amplitude amplitudes are centered below 0.7. Finally, in the case where the third-largest amplitude corresponds to the ground-truth label, the PDFs of the second- and third-largest amplitude are shifted towards 1, while the PDF of the fourth-largest amplitude is centered below 0.7. The previous results give us a broad picture of latent space topology: the labeled features project to well-defined compact domains in latent space. Let us now consider how we can use this information to denoise images. ## Denoising with LSD The main issue when reducing noise in images is distinguishing noise from information. In this sense, a reliable denoiser has to learn what is noise and what isn’t. One reason deep generative models are promising for denoising data is that in optimum conditions the GM has learned the exact data distribution. Of course, if the data set has noise, the GM will also learn the embedded noise in the data set. However, by sampling the latent space we may find regions where the signal to noise ratio is sufficiently large. For large M, this sampling is computationally expensive. To avoid this cost, we propose to LSD as a denoiser. Recall that in the previous section we showed that with a 98% accuracy the information needed to assign a label to the image is stored in either the first-, second-, third- or fourth-largest amplitude of the LSD. Therefore, we propose that once the test set is encoded into latent space, we decompose the latent vector in terms of the quasi-eigenvectors and drop the contribution from quasi-eigenvectors with low amplitudes. In we show the results of this truncation for 125 random sample images. In we describe how to understand these images. shows 5 columns, where each column has 25 rows and each row has 7 images. In each row, the first image corresponds to the ground-truth image, the second image is the image decoded from all 100 LSD components of the ground truth image. The third, fourth, fifth, sixth and seventh images are the images decoded after truncating the expansion after 1,2,3,4 and 10 LSD components of the ground truth image. In this method, denoising maintains the identity of the labeled feature in the image, e.g., each row shows different representations of the same number. In most cases in , the denoised image looks clearer and sharper. However, sometimes the LSD components project back to the wrong number. However we can consider as many LSD components as the dimension of the latent space, so even if taking the first n LSD components yields the wrong number, taking the first n + 1 LSD components could yield the correct number. In the previous section we showed that using only the first 4 LSD components gave us a 98.9% chance of obtaining the right number. ## Operations in latent space Here we explore how to build operators in latent space that can yield feature transformations in real space. Having a set of orthogonal vectors that span latent space allows us to perform most operations in latent space as a series of rotations, since we can express the operator as a superposition of the outer product of the quasi-eigenvectors. If we construct a rotation matrix, $\mathcal{R}$, in latent space, we can then recursively apply $\mathcal{R}$ to a set of encoded images. After each iteration we project the output to real space to see the effect of the latent-space rotation. We can define a projection operator $\mathcal{B}_{\xi_{i},\xi_{j}}$, such that, $$\begin{array}{r} {B_{\xi_{i},\xi_{j}} = \frac{1}{\left\langle \xi_{i} \middle\| \xi_{i} \right\rangle}\|\xi_{j}{\rangle\langle}\xi_{i}\left\| \;. \right.} \\ \end{array}$$ This operator projects from \\|ξ<sub>i</sub>〉 to \\|ξ<sub>j</sub>〉, i.e., $\mathcal{B}_{\xi_{i},\xi_{j}}\|\xi_{k}{\rangle =}\delta_{\xi_{k},\xi_{k}}\left\| \xi_{j} \right\rangle$, where $\delta_{\xi_{k},\xi_{k}}$ denotes the Kroenecker delta function. Similarly, we define the operator $\mathcal{R}_{\xi_{i},\xi_{j}}(\Delta\theta,\theta)$ as $$\begin{matrix} {\mathcal{R}_{\xi_{i},\xi_{j}}(\Delta\theta,\theta) \propto \left( \left. \text{cos}(\theta + \Delta\theta) \right\|\xi_{i}{\rangle\left. + \text{sin}(\theta + \Delta\theta) \right\|}\xi_{j}\rangle \right)} \\ {\cdot \left( \langle\xi_{i}\left\| \text{cos}(\theta) + \langle \right.\xi_{j}\left\| \text{sin}(\theta) \right. \right)\;,} \\ \end{matrix}$$ which projects from cos(θ)\\|ξ<sub>i</sub>〉 + sin(θ)\\|ξ<sub>j</sub>〉 to cos(θ + Δθ)\\|ξ<sub>i</sub>〉 + sin(θ + Δθ)\\|ξ<sub>j</sub>〉. Starting from a set of images with label zero, we first encoded them to latent space, then we applied the rotation operator $\mathcal{R}$ recursively, as follows: First, we perform the rotation from the quasi-eigenvector associated with label zero to the quasi-eigenvector associated with label 1, viz., $\mathcal{R}_{\xi_{\alpha = 0,i = 1},\xi_{\alpha = 0,i = 1}}(\Delta\theta,\theta)$. Then, we performed a rotation from the quasi- eigenvector associated with label 1 to the quasi-eigenvector associated with label 2, viz., $\mathcal{R}_{\xi_{\alpha = 1,i = 1},\xi_{\alpha = 2,i = 1}}(\Delta\theta,\theta)$, and repeat mutatis mutandi until we reach the quasi-eigenvector associated with label α = 9. To keep the individual rotations in latent space small (and maintain the local linearity of the transforms), we fixed the rotation step size Δθ ≈ π/6 so transforming from a direction associated with one quasi-eigenvector to a direction associated with a different quasi-eigenvector requires three sequential rotations. In Alg. 1 we show the pseudocode. To ensure the rotated latent vectors have constant norm value as in, after each iteration we divide the latent vector \\|z〉 by $\sqrt{\frac{\left\langle z \middle\| z \right\rangle}{M}}$. After each iteration, we project the latent vector into real space. In we show this projection for a set of sample images. Notice how the numbers transform from 0 to 9. In principle, we could rotate through any other set of sequential features in this way. The key idea is that having a set of quasi-eigenvectors that span latent space each mapping to a specific label, we can define a metric in latent space defining the distance between the latent-space representation of each label. Algorithm 1: Latent space rotation pseudocode. initialization $\left\| z \right\rangle\left. = \mathcal{E} \middle\| x \right\rangle$ (initial condition) Δθ = π/3 (angular rotation step) α = 0 (initial label) i = 1 (set index) for α in {0, 1, 2, …, 9} do for r in {1, 2, 3} do $\left\| z \right\rangle = \mathcal{R}_{\xi_{\alpha,i},\xi_{\alpha,i}}\left( r \cdot \Delta\theta,(r - 1) \cdot \Delta\theta \right)\left\| z \right\rangle$ (rotation) $\left\| z \right\rangle = \left\| z \right\rangle/\sqrt{\frac{\left\langle z \middle\| z \right\rangle}{M}}$ (norm) $\left\| x \right\rangle\left. = \mathcal{G} \middle\| z \right\rangle$ (projection to real space) end end # Discussion We have shown that it is possible to build a set of orthogonal vectors (quasi- eigenvectors) in latent space that both span latent space and map to specific labeled features. These orthogonal vectors reveal the latent space topology. We found that for MNIST, almost all the images in the data set map to a small subset of the dimensions available in latent space. We have shown that we can use these quasi-eigenvectors to reduce noise in data. We have also shown that we can perform matrix operations in latent space that map to feature transformations in real space. On the one hand, the deeper the NN the better its capacity in learning complex data and as depth increases, the non-linearity increases as well. On the other hand, it has been proposed the equivalence between deep NNs and shallow wide NNs. From catastrophe theory, we know that in non-linear dynamical systems small perturbations can be amplified leading to bifurcation points leading to completely different solution families of these non-linear dynamical systems. The results in Ref. suggest a different picture with what the authors call vector arithmetics in which adding or subtracting vectors in latent space can yield a feature addition, removal or modification (e.g., hair color, sunglasses, facial hair in the case of a headshot image data set). This behavior hints at the possibility of building a vector basis in latent space. It is not obvious why or how the label embeddings cluster in latent space or why they do so in a linearly independent manner. To put it in different terms, it would appear that the training of the GAN is reminiscent of a symmetry breaking mechanism from a rotationally invariant latent space to one where the label embeddings are linearly independently clustered. We consider that understanding why this pattern of clustering occurs is of great relevance and we intend to explore it in future work. Our intuition behind using the Gram-Schmidt method comes from the latent-space vector arithmetic and the flexibility of the method whereby one first chooses a set of vectors from which the vector basis is built. Our work contributes to this discussion of the emergent effective linearity of NNs as transformations. While the NNs we used are intrinsically non-linear, they exhibit local linearity over a region of interest in latent space. This subspace maps to labeled features. In this sense, we say the non-linear NNs are effectively linear over the domain of interest. As a proof of concept, we have shown this for MNIST successfully, and our results serve as a proof of concept. Future work is aimed at testing this method in broader data sets, such as, CIFAR and ImageNet. Similarly, we plan to test this method for different latent space dimensionality and the effect it can have on feature entanglement. We have considered labeled data which is a strong assumption in real problems since it is usually difficult to have that type of information. However, having a set of quasi-eigenvectors potentially allows us to recreate unlabelled data through latent superposition. We have not tested this here and we leave it for further work as well as testing this framework in other well-known datasets. Fundamentally, we have shown that the data clustered in the GAN’s latent space is linearly independent by building a set of quasi-eigenvectors pointing to each of these clusters. Further work is needed to understand the relationship between labels and linearly-independence when the latent space dimensionality varies. The classifier and encoder were merely tools used to be able to span latent space and further work is aimed at simplifying this framework. From an application standpoint, mapping to dominant quasi-eigenvectors could be useful for medical imaging, diagnosis and prognosis if, e.g., the labels denoted the severity of a disease; for predicting new materials if the labels denoted specific material features or external physical parameters. [^1]: The authors have declared that no competing interests exist.
# Introduction Neurons in the cochlear nucleus (CN), differing in their anatomical and physiological properties, give rise to different ascending parallel auditory pathways, each concerned with the processing of specific aspects of acoustic information – Insights into the functional organization of the CN are therefore essential for further comprehension of auditory brainstem processing.For the anteroventral division of the CN (AVCN) three principal morphological cell types have been described by Osen : spherical bushy cells, globular bushy cells and stellate cells. Further subdivisions were postulated by Brawer et al. and Lorente de Nó. In parallel, electrophysiologists have commonly classified AVCN units in vivo based on their temporal responses properties to acoustic stimuli as seen in peri-stimulus time histograms (PSTH) – According to these studies, the major response categories in AVCN were termed primary-like, chopper and onset -units, where each category were subsequently further subdivided, Still, it was shown that units exhibiting a specific PSTH also share other physiological characteristics such as spontaneous rate, inter-spike interval histograms, spectral response bandwidth, etc.. Intracellular labelling of CN cells combined with in vivo recordings provided evidence for a relation between PSTH types and morphologically defined types of neurons. The relationships that emanated from these studies are that primary- like patterns are attributed to spherical bushy cells, primary-like with notch patterns to globular bushy cells, and chopper patterns to stellate cells. However, even studies in favour of the respective physio-morphological correlation repeatedly reported cases that do not support the idea of equating both classifications. For example, Rhode has recently shown a variation of different PSTHs for globular bushy cells, indicating that a classification which is exclusively based on PSTH types might not be sufficient for a comprehensive characterization of AVCN neurons. In an earlier attempt, Blackburn and Sachs have already evaluated a large number of physiological properties of AVCN units to define classifying boundaries between PSTH types. In our study, we followed along these lines, but made great efforts to not base the analysis on any presumptions regarding different physiological properties of the units. To achieve this goal, we used a multivariate statistic, which allows the classification based on a uniform evaluation of a large number of response properties. This enabled a much wider spectrum of physiological response properties to be taken into account and at the same time to consider potential relationships between all these parameters. The data were collected from Mongolian gerbils, a model system that has gained importance over the last years in auditory research due to their distinct low frequency hearing ability. The vast majority of the previous studies investigating cell types in the AVCN have been conducted in cats, while only little work was done in gerbils. So, the present report provides the first comprehensive examination of the physiological properties of neurons in the AVCN in the Mongolian gerbil. We find that while the neurons of the same PSTH type are generally closer to each other than across PSTH types, no clear boundaries emerge and neurons of different PSTH type can be quite similar with respect to their physiological properties. We therefore conclude that at the present resolution of the analysis and the presently observed properties, the AVCN neurons appear to form a physiological continuum. # Materials and Methods All experiments were performed at the Neurobiology Laboratories of the Faculty of Bioscience, Pharmacy and Psychology of the University of Leipzig (Germany). The experimental procedures were approved by the Saxonian District Government Leipzig (TVV50/06) and conducted according to European Communities Council Directive (86/609/EEC). ## Animals and animal care Adult pigmented (agouti) Mongolian gerbils (Meriones unguiculatus), aged 2–4 months and weighing 40–70 g, were used in the experiments. The animals were obtained from the animal care facilities of the Institute of Biology II of the University Leipzig. During the surgical preparation and the recording experiments, the animals were anaesthetized with an initial dose of a mixture of ketamine-hydrochloride (18 mg/100 g body weight, Ketavet®: Upjohn) and xyalzine- hydrochloride (0.5 mg/100 g body weight, Rompun®: Bayer). A constant state of anaesthesia was kept by supplementary injections of one-third of the initial dose (on average once per hour). Body temperature was maintained with a heating pad (Harvard Apparatus) at 37–38°C. ## Surgical preparation For supporting the animal in a stereotaxic recording device the skull was exposed along the dorsal midsagittal line, and a small metal bolt was glued to the bone overlaying the forebrain. Two 500 µm diameter holes were drilled in the skull 2000–2300 µm caudal to the lambda suture. Through the first drill hole, located over the midline, the recording electrode was inserted. The second drill hole, located 1500 µm lateral to the midline, was used to position the reference electrode in the superficial cerebellum. ## Acoustic stimulation Auditory stimuli were generated on a standard PC using custom written software (Rec_thor: M. Weick, University of Leipzig; and Spike: B. Warren, University of Washington, Seattle). The signals were then transferred to a D/A converter (RP2.1 Enhanced Real Time Processor, 97.7 kHz sampling rate, Tucker-Davis Technology; or DD1, 50 kHz sampling rate, Tucker-Davis Technology). Near-field acoustic stimuli were delivered through custom made earphones (acoustic transducer: DT 770 Pro, Beyerdynamic) fitted with probe tubes (plastic, 70 mm length, 4 mm diameter) which were placed close to the opening of the ear canal. The speakers were calibrated as described previously. The maximum intensity used in the recording experiments was 90 dB sound pressure level (SPL). ## Recording setup All experiments were performed in a sound-attenuated and electrically isolated chamber (Type 400, Industrial Acoustics Company). The animals were placed on a vibration-isolated table (T 251.SL, Physik Instrumente) and fixed in a custom made stereotaxic device by means of a metal bolt. Recording electrodes were glass micropipettes (GC150F-10, Harvard Apparatus) filled with 3 M KCl. Targeting the AVCN, the recording electrodes were advanced through the cerebellum and the brainstem with a piezo manipulator (PM101, Newport). The extracellular recordings were amplified and filtered (Modell 1600, Neuroprobe Amplifier, A-M SYSTEMS; PC1, Tucker-Davis Technology \[0.4 kHz high- pass filter and 7 kHz low-pass filter\]; HumBug, Quest Scientific \[50 Hz notch filter\]). For isolation of single units, action potentials thresholds were set at voltages of at least two times the noise envelope (corresponding to approx. 4 times the S.D. of the noise), and the crossing of this level was taken as the time of occurrence of action potentials. The crossing of the threshold level triggered TTL-signals (SD1, Tucker-Davis Technology) which were digitized (RP2.1 Enhanced Real Time Processor, Tucker-Davis Technology; or ET1, Tucker-Davis Technology) and stored on a standard PC for offline analysis. Simultaneously, the voltage signals were digitized (DD1, 50 kHz sampling rate, Tucker-Davis Technology) and stored on the PC. ## Multiunit mapping At the beginning of each experiment, the stereotaxic coordinates of the exact position and extent of the AVCN were determined by on-line analysis of acoustically evoked multiunit activity recorded with low-impedance electrodes (\<5 MΩ). In several electrode penetrations we tested for neuronal responses evoked by ipsilateral acoustic stimulation. The characteristic frequency of multiunits was systematically documented every 200 µm along the dorsoventral dimension and every 200 µm in the rostrocaudal and also in the mediolateral dimension (2° change in angle of electrode penetration). Referring to the known tonotopic organization of the cochlear nucleus – it was possible to identify the exact penetration coordinates for the CN and its subregions in each experimental animal. ## Single unit recordings The response properties of single units were examined with high-impedance micropipettes (10–30 MΩ). Single units were identified with regard to the bipolar shapes of their waveforms and signal constancy. ### Frequency tuning curve and related response characteristics A unit's response area was measured by presenting pure-tone pulses (100 ms duration, 5 ms rise-fall time, 100 ms inter-stimulus interval) within a predefined matrix of 16×15 frequency/intensity combinations. Stimuli were presented five times in a pseudo-random order. Spontaneous activity was acquired in silent runs interspersed with the stimulus runs. The number of spikes was measured during the 100 ms period of stimulus presentation. From these data characteristic frequency (CF, frequency with the lowest threshold), best frequency (BF, frequency with maximum discharge rate), and their difference (CF- BF) were obtained. Additionally, the following sound-evoked response characteristics were computed: threshold at CF (TRS), spontaneous (spontR) and maximum discharge rates (maxR), rate-level function at CF (RLF), dynamic range (DR), response bandwidth 10 dB above the unit's threshold (Q<sub>10</sub>-value, Kiang 1965), the slope of the frequency tuning curve at the respective high- frequency and low-frequency borders sites, and the occurrence of inhibitory sidebands. For the categorical properties RLF and inhibitory sidebands we translated each category to a number in order to include them in the cluster analysis. For RLF the assignments were strictly monotone to 1, monotone-plateau to 2, and non-monotone to 3. For inhibitory sidebands we assigned units that show no sidebands to 0, units that show a sideband at the low-frequency side of the FTC to 0.8, units with a sideband at the high-frequency side to 0.9 and units with both, low- and high-frequency sidebands, to 1. ### Peri-stimulus time histogram (PSTH) PSTHs were acquired in response to pure tone pulses (100 ms duration, 5 ms rise- fall time, 100 ms inter-stimulus interval, 50 repetitions) at the unit's CF and with 80 dB SPL. The classification of the PSTHs was based on the definitions by Pfeiffer , Rhode and Smith, Young et al. and Blackburn and Sachs. Primary-like (PL) PSTHs display phasic-tonic discharge patterns, similar to those obtained from auditory nerve fibres. Primary-like with notch (PL<sub>N</sub>) PSTHs are similar to the PL ones, but here the initial response peak is separated from the subsequent tonic activity by a short (\<2 ms) pause. Chopper PSTHs exhibit several regularly spaced peaks with interpeak distances unrelated to the stimulation frequency. The chopper units were subdivided into two subcategories: transient choppers (C<sub>T</sub>) which display the regular discharge pattern only at the beginning of the response, while in sustained choppers (C<sub>S</sub>) prominent periodic discharge peaks were found throughout the entire duration of the response. To differentiate between C<sub>T</sub> and C<sub>S</sub> units we used the coefficient of variation (CV) as introduced by Young et al. (1988). Units with CVs smaller than 0.35 were classified as C<sub>S</sub>, units with larger CVs as C<sub>T</sub>. For the cluster analysis the following parameters were quantified from the PSTH: peak-over-total value as the number of APs occurring during the stimulus vs. the number of APs in the onset peak (all 0.5 ms bins around the maximum peak with at least 2/3 height of the maximum peak), average inter-spike interval (ISI) with standard deviation (S.D.), the average first spike latency (FSL), and the jitter of the first spike. For the ISIs and their standard deviation only the first 20 ms of the stimulus are consider for better comparability with previous studies. We are aware that the ISI most likely correlate with the maximum discharge rate. However, the maximum discharge rate corresponds to the best frequency and the ISI to the characteristic frequency; thus both parameters could potentially carry different information and therefore were both included in the analyses. For the FSL the first spike of each trail is included in the analysis. Note, that the dependency of first spike latency on CF is implicitly taken into account by the multidimensional analysis, where covaritions between difference parameters would produce elongated clusters, which could be identified by the hierarchical cluster analysis we employed. Due to the often limited single-unit recording time we had to restrict the recordings to the minimum. We therefore choose to use only one stimulus level for evaluating PSTHs. We decided to use 80 dB SPL for several reasons: (i) A threshold independent stimulus level avoids presumptions about connections between single parameters. Later on a correspondence to the threshold is always possible and was considered in the cluster analyses, since there all parameters are considered in parallel. (ii) At 80 dB SPL most of the units responded close to their maximal discharge rate and thus in a comparably activated state. We are aware that there are few units for which 80 dB SPL lies still in their dynamic range (especially those with high thresholds) and some units which already decrease their response at this level (units with non-monotonic RLF). (iii) The shape of the PSTH is more likely to be fully developed at high levels. For example, Blackburn and Sachs showed for some cases that the notch becomes only visible for higher stimulation levels and would hence influence the differentiation between PL and PLN. ### Sinusoidal amplitude modulation (SAM) Temporal encoding of rapid amplitude fluctuations in the stimuli was quantified from the responses to sinusoidally amplitude-modulated tone bursts at CF, 20 dB above threshold. Modulation depth of the SAM signals was 100% with modulation rates from 20 Hz to 1000 Hz (stimulus duration: 100 ms, inter-stimulus interval: 100 ms, 50 repetitions). To exclude a contamination of the units' SAM coding by the onset response, only the steady-state response (10–100 ms) was included in the analysis. The data were quantified by calculating the vector strengths (VS) and entrainment (entr) of neuronal discharges. ### Waveform analysis The waveforms of neuronal discharges were collected by triggering the voltage trace at a visually determined, conservative level. The waveforms of the triggered potentials were collected in intervals from 2 ms preceeding to 2.5 ms following the trigger. The average waveforms of the units were then used to quantify the signal-to-noise ratio (SNR, the positive peak of the AP divided by the standard deviation of the noise) of the respective single unit recording and the time between the maximum and minimum amplitude of the bipolar signals as an indicator of the duration of the unit's AP (t(AP)). The averaged waveforms were also inspected for distinct P, A, and B components as described for spherical bushy cells. ## Verification of recording sites The recording sites were verified histologically in 6 animals using horseradish peroxidase (HRP, Sigma) and in 10 animals using biotinyled dextranamin (BDA, Moleculare Probes). HRP was injected iontophoretically (+1.5 µA, 4–5 min) at the recording site; BDA was applied by pressure injections. The animals were allowed to survive for 24 h or one week, respectively. Then, they were given a lethal dose of Na-pentobarbital (100 mg/100 g body weight i.p, Narcoren®, Merieux) and perfused via the left ventricle of the heart with 0.9% NaCl solution followed by fixative (2.5% paraformaldehyde in 0.1 M phosphate buffer, pH 7.4). The brains were removed from the skull and post-fixed in paraformaldehyde, and thereafter embedded in agarose. Serial transverse sections of the brains (50 µm) were cut using a vibratome. BDA brain sections were reacted using an avidin-HRP complex (Molecular Probes). Accordingly, the BDA sections as well as the HRP sections were treated with 3,3′-diaminobenzidine to visualize the HRP (Adams 1981). The HRP sections were counterstained with cresyl violet (Nissl stain). Electrode tracks and recording sites were reconstructed by examining the sections with the light microscope. ## Statistics For statistical analysis SigmaPlot 8.0, SigmaStat 3.0 (both SPSS inc.) and MATLAB 7.3 (Mathworks Inc.) were used. If data were normally distributed (Kolmogorov-Smirnov-test) the results are displayed as means±S.D., otherwise as medians and the respective 25% and 75% quartiles. The statistical significance (p\<0.05) of differences between groups was assessed using one way ANOVA followed by Holm-Sidak post-hoc test if the data was normally distributed. For data with other distributions a one way rank-based ANOVA followed by a post- hoc test according to Dunn's method was used. ### Multivariate statistics To test whether the sample can be divided into groups based on electrophysiological response properties, hierarchical cluster analysis was employed. This method allows a classification on the basis of a wide spectrum of parameters simultaneously. The number of expected groups does not need to be predefined, thus the result is not affected by prior assumptions. Before performing the cluster analysis all parameters were standardized to remove the effect of scaling differences between parameters. Dissimilarities between units were expressed as distances (linkage distance) in a space of as many dimensions as the parameters taken into account. In the present study Euclidean distance between objects was used. It is the most commonly used distance and simply figures the geometric distance in the multidimensional space. Classifications based of other distance metrics yielded qualitatively similar results. For joining smaller clusters into larger ones Ward's method was used, which attempts to minimize the variance within the groups. shows the aggregation procedure schematically. On the left side the units are presented as points in the multidimensional space (here only two dimensions). On the right side the developing dendrogram is shown for every aggregation step. It starts with the individual cells at the bottom. First, the two closest units in the multidimensional space were grouped together. Then, at each step, the number of groups is reduced by merging the two groups or units whose combination gives the least increase in the within-group sum of squared deviation. The linkage distance is a parameter for the heterogeneity of the joined groups. When linkage distance increases, branch points represent clusters of increasing size and dissimilarity. The final number of groups is determined at the stage where the maximal decrease in linkage distance is observed. To illustrate the position of the single units in relation to each other, a principal component analysis was performed. Principal component analysis projects the units from the multidimensional space to a space of lower dimensionality while attempting to preserve the distance relations between them. Therefore the multidimensional space was rescaled in a way that the first new coordinate (principal component 1, PC1) goes through the greatest variance by any projection of the data, the second coordinate (principal component 2, PC2) through the second greatest variance, etc. Besides suggesting a grouping of the sample, this method also indicates which set of parameters best explains the dissimilarity between groups. Note, that both hierarchical cluster analysis and principal component analysis can aid in uncovering structure in data sets and suggest a scheme for classification. Yet, they do not produce a measure of statistical significance for differences between suggested groups. To verify our results we employed (i) different subpopulations of our unit sample and (ii) different sets of parameters and compared the respective cluster analyses. Changes in the sample size and in the number of parameters should be tolerated by the cluster analysis as long as the sample provides enough information of the whole population and the chosen parameters provide enough potential to distinguish the population. Otherwise the analysis would result in small, only minimal separated groups. To avoid this, our analysis is based on a broad spectrum of parameters that potentially could separate the neuron population. Thereby we tried to not exclude parameters that, at first, seem not to have a major impact on the classification and also not to include parameters twice, in form of different expressions (e.g. the CV-value additionally to ISI mean and ISI standard deviation). We also aimed to include the parameters in their most raw form to avoid distortions. # Results The analyses were based on data acquired from 233 units recorded in the AVCN of Mongolian gerbils. The minimum requirement for a data set to be included in the analysis was a complete acquisition of a unit's response area based on recordings during tone burst presentations (5 repetitions of 16×15 frequency- intensity combinations) and 16 sec (5×16×200 ms) recordings of spontaneous discharge activity. For the majority of the units (n = 181, 78%) additional recordings with 50 stimulus repetitions at the respective characteristic frequency (CF) could be achieved, which enabled a refined analysis of the unit's temporal response characteristics. Furthermore, for 60 units (26%) the discharge activity to sinusoidal amplitude modulated signals (SAM) was registered for modulation frequencies from 20 to 1000 Hz. The waveform of the extracellular recorded single unit field potential was digitized in 154 units (66%). This data set formed the basis for the analysis of the units' frequency and sound level coding as well as their temporal coding (detailed response properties for all units are summarized). In analysing the data we used two strategies. First, we grouped the data based on temporal response patterns into the ‘classical’ AVCN PSTH groups and tested for co-variations with single other response features. Second, we performed a cluster analysis using all evaluated parameters. ## Sample characteristics This first part of the results gives an overview of the properties of the recorded AVCN units based on classical classification criteria which relate to current models of distinct pathways of auditory signal processing. These models proceed on the assumption of neurons in the AVCN with distinct morphological and physiological properties: spherical bushy cells, globular bushy cells and stellate cells. However, the variability of the physiological properties within the neuron population highlights the necessity of the cluster analysis presented subsequently. ### Temporal response patterns In previous studies the temporal response characteristic was an important feature for classifying CN units. Referring to this classification, which is based on the PSTH of responses to CF tone bursts at 80 dB SPL, our AVCN unit sample could be subdivided into 6 classes, 4 of which are shown in. One quarter of the units (n = 57, 25%) showed a primary-like (PL) response pattern. Primary-like with notch (PL<sub>N</sub>) responses were obtained in 20% (n = 48), transient chopper (C<sub>T</sub>) responses in 28% (n = 65), and sustained chopper (C<sub>S</sub>) in 16% (n = 37) of the units. Fourteen units (6%) displayed an onset-inhibitory response (O<sub>inh</sub>), characterized by a sharp onset peak followed by strongly reduced discharge activity which was even below the spontaneous rate (data not shown). Only a fraction (5%, n = 12) of the units phase-locked to pure tones at their CF up to 1 kHz. Such phase- locking made it difficult to assign these units to one of the before mentioned PSTH types; therefore they were assigned in a separate group (PHL). In the following analysis we focus only on the four main PSTH types, since the data sets of the O<sub>inh</sub> and PHL units are too small for statistical analyses. In all PSTH types the peak-over-total ratio varied widely from 1% to 20%, but C<sub>S</sub> units had the highest peak-over-total ratios among the units of the four main PSTH types (13±4%). In PL units the peak-over-total ratio ranged from \<1% to 19% (10±5%), in PL<sub>N</sub> from 1% to 22% (6±4%), and in C<sub>T</sub> from 1% to 19% (8±3%). The mean inter-spike intervals of both types of chopper units (C<sub>S</sub>: 2.0 ms \[1.6 ms; 2.8 ms\]; CT: 2.6 ms \[2.2; 3.1\]) were significantly smaller than those of PL (3.4 ms \[2.8; 4.2\]) and PL<sub>N</sub> units (3.9 ms \[3.1; 4.6\]). Compared to the other units, both types of chopper units also showed a tendency for a smaller variance in inter-spike intervals (C<sub>S</sub>: 0.6 ms \[0.4; 0.9\], C<sub>T</sub>: 1.3 ms \[1.0; 2.0\], PL: 2.2 ms \[1.7; 2.4\], PL<sub>N</sub>: 2.4 ms \[1.6; 2.9\]). Despite these differences the respective values overlap strongly in the different PSTH types. The first spike latencies were longest in PL (4.5 ms \[4.0; 6.6\]) units. Also the jitter of the first spikes was larger in PL units (1.15 ms \[0.6; 2.1\]) than in units of other PSTH types. The values of PL<sub>N</sub>, C<sub>T</sub>, and C<sub>S</sub> units do not differ significantly. For none of the PSTH a dependency of the first spike latency from the characteristic frequency could be found, also there was a slight trend to shorter latencies at higher frequencies. In summary these results show that the consideration of the PSTH type, the statistics of the inter-spike intervals and of latencies of the responses to acoustic stimuli does not lead to an unambiguous separation of the unit sample. ### Waveforms of action potentials The spike discharges acquired with electrolyte-filled glass micropipettes typically displayed bipolar waveforms. Complex waveforms composed of a presynaptic component (P) and two postsynaptic components (A/EPSP+B/postsynaptic AP, inset), were observed in 17% (4/23) of the PL units and in one PL<sub>N</sub> unit. These units exclusively responded to low frequencies (CF\<2 kHz). Only 2/7 units in this CF range lacked the presynaptic component. No such presynaptic components were seen in units with higher CFs. Still, in 32% (11/37) of the PL<sub>N</sub> units postsynaptic EPSP (A-component) and AP (B-component) could be separated in the averaged AP waveforms; the same holds for 8% (4/51) of the C<sub>T</sub> and 12% (4/32) of the C<sub>S</sub> units. The signal-to-noise ratios of the recordings of the whole unit sample varied between 5.2 and 16 with systematic differences in different PSTH groups. The highest signal-to-noise ratios were obtained in C<sub>S</sub> units (12.0 \[7.9;14.5\]), which were significantly higher than in PL (7.4 \[6.3; 9.1\]) and PL<sub>N</sub> units (6.3 \[5.5; 7.6\]). The C<sub>T</sub> units (8.4 \[6.9; 11.0\]) also had significantly higher signal-to-noise ratios than PL<sub>N</sub> units. However, the duration of the APs did not significantly differ between the different PSTH groups. This was inferred from measuring the time between the maximum and minimum of the bipolar APs signals which range from 0.11 to 0.48 ms. The data show only a slight tendency for PL<sub>N</sub> units to have longer APs. ### Tuning characteristics and spontaneous activity The characteristic frequencies (CF) of the units covered the range from 0.3 to 45 kHz. In all PSTH groups the CFs spread widely, and the only significant difference was between the lower average CF in PL units (2.3 kHz \[1.6; 7.8\]) and the rest of the unit types (PL<sub>N</sub>: 18.1 kHz \[10.1; 23.0\]; C<sub>T</sub>: 15.2 kHz \[2.6; 29.0\], C<sub>S</sub>: 14.6 kHz \[7.3; 19.6\]). As a whole, the threshold values did not differ between the PSTH groups (means: 20–35 dB SPL). Only in the subgroup of units with CFs\>10 kHz the thresholds of PL units were elevated (55 dB SPL \[43.75; 66.25\]) in comparison to the other PSTH groups. Units of all PSTH groups displayed V-shaped frequency threshold curves with low- frequency tails. The slopes of the low-frequency flanks of the tuning curves were typically less steep than the slopes of the high-frequency flanks. The respective maximal values were 1.64 octaves and 1.15 octaves frequency increase per 10 dB. There was a tendency for C<sub>S</sub> units to have more symmetric frequency tuning curves than units of other PSTH groups. In C<sub>S</sub> units the average difference between the slopes of the high- and the low-frequency flank was only 0.02 octaves/10 dB. In contrast, it measured 0.16 octaves/10 dB in PL units. The unit's frequency selectivity was quantified by Q<sub>10</sub>-values. In all PSTH groups Q<sub>10</sub>-values correlate strongly with the unit's CF. High-CF units show larger Q<sub>10</sub>-values, i.e. better frequency selectivity than low-CF units. When comparing units within specified CF-ranges, PL units showed the best frequency selectivity. Inhibitory sidebands could only be observed in units with sufficiently high spontaneous discharge rates (\>10 Hz). Among those, the incidence of inhibitory sidebands was highest in C<sub>S</sub> (77%; 14/18) and C<sub>T</sub> units (62%, 15/24), followed by PL (58%; 23/41) and PL<sub>N</sub> units (42%; 8/19). In most units (82%, 190/233), the CF was within the ±1 octave range of the best frequency (defined as the stimulus frequency eliciting the highest stimulus evoked discharge rate). In 42 of the remaining 44 units the best frequency was below CF by up to 5 octaves. No apparent differences were observed between the PSTH groups. However, the units' maximum discharge rates were significantly higher in chopper units (C<sub>S</sub>: 410 Hz \[284; 486\], C<sub>T</sub>: 298 Hz \[255; 365\]) than in PL (163 Hz \[126; 216\]) and PL<sub>N</sub> units (183 Hz \[140; 220\]). Still, as for the CFs, there was a wide overlap of the respective values between different PSTH types. The spontaneous discharge rates of PL units (32 Hz ;) were significantly higher than in any other PSTH group. In PL<sub>N</sub> (0 Hz \[0; 27\]), C<sub>T</sub> (0 Hz \[0; 17\]), and C<sub>S</sub> (7 Hz \[0; 26\]) units the distribution of spontaneous rates was mostly in the same range. Three different types of rate-level functions (RLF) were distinguished : monotonic RLF in which the discharge rates increase with increasing intensity level without saturation (strictly monotonic, ms, indicated as light grey); monotonic RLF which saturates at a certain level (monotonic-plateau, mp, medium grey), and non-monotonic RLF (nm, dark grey), in which the discharge rates reach a maximum in the mid-level range and then decrease towards higher intensities. In all PSTH groups, monotonic-plateau RLFs were the most frequent ones (50–69%), the monotonic and the non-monotonic RLFs had shares of 19–28% and 11–35%, respectively. The dynamic ranges differed between the PSTH groups: PL (30 dB ;) and PL<sub>N</sub> units (30 dB ;) had significant smaller dynamic ranges than chopper units (C<sub>T</sub>: 40 dB ;, C<sub>S</sub>: 45 dB ;). However, these differences did not circumscribe clear borders between the PSTH groups. ### Response to sinusoidal amplitude modulations (SAM) To quantify the ability of AVCN units to respond to rapid fluctuations in signal amplitude, the vector strength and the entrainment to SAM stimuli varying in modulation frequencies (f<sub>mod</sub>) from 20 to 1000 Hz were calculated. A vector strength value of 0.3 was chosen as a cut-off criterion for classifying a response to SAM as being phase-locked. On average, PL units phase-locked to SAM stimuli up to 400 Hz (200; 600), while PL<sub>N</sub> units phase-locked up to 1000 Hz (575; 1000). In PL units also the maximum vector strength values were lowest (0.52±0.16) and they were obtained at low f<sub>mod</sub> (50 Hz \[20; 100\]). The highest vector strength values were found in C<sub>S</sub> units (0.72±0.14) at higher f<sub>mod</sub> (200 Hz \[200; 300\]). Also the entrainment, i.e. the proportion of SAM cycles in which at least one spike was recorded, reached the highest values and encompassed the highest modulation rates in C<sub>S</sub> units (0.93±0.09 at 300 Hz \[300; 400\]). Again, the lowest respective values were found in PL units (0.59±0.15 at 100 Hz \[50; 200\]). Both SAM phase-locking and entrainment showed large overlaps between the PSTH groups. In summary, a separate consideration of different parameters for frequency, sound pressure and temporal coding by AVCN units did not lead to an unambiguous separation of physiologically defined unit types. In fact, there is a large variability of values within each PSTH group and a strong overlap between the groups. Therefore we pursued an alternative approach and analysed the whole data set in a multidimensional parameter space. ## Cluster analysis The following cluster analyses are based on the same sample of AVCN units and the same set of discharge properties described above. Cluster analysis can uncover structures in data sets and suggest a scheme for classifying the unit sample. It may result in the identification of distinct groups of units with a fixed set of distinguishing properties, or in identification of groups of units which differ in their properties, but which are not strictly separated from each other, or alternatively in the absence of distinct groups (continuum). ### Classification of AVCN units considering all evaluated properties To allow a complex physiological classification of AVCN units, all 22 evaluated properties were included in this analysis. For the cluster analysis an equalled sample, i.e. the same number of parameters for each unit is required. To sustain a sufficient sample size missing values, i.e. values that could not be acquired because of limitations of single unit recording time, were substituted by the mean of the respective property in the whole unit sample. The result of the hierarchical cluster analysis is shown as a dendrogram in. Individual units are lined up at the bottom of the graph. With increasing linkage distance between units, successive branch points represent clusters of increasing size and dissimilarity. This dendrogram suggests five main clusters (I–V) of which one (cluster V) is characterized by a stronger separation from the rest. For purpose of comparison, the values for all evaluated properties of the respective clusters are displayed in and summarized in. Cluster V is composed of 80 units which are particularly characterized by short inter-spike intervals (2.2 ms \[1.8; 2.8\]), a low variation of inter-spike intervals (0.9 ms \[0.6; 1.4\]), high maximum discharge rates (324 Hz \[248; 428\]), and low CF- thresholds (15 dB SPL \[0; 25\]). The units in this cluster never show waveforms with a presynaptic component. Units in cluster I (n = 54) are characterized by long first spike latencies (4.8 ms \[3.7; 7.0\]), relatively high CFs (21.8 kHz \[8.5; 31.9\]), high thresholds (45 dB SPL ;), and low spontaneous rates (0.8 Hz \[0; 13.4\]). Inhibitory sidebands were rarely seen (n = 2/54). The waveforms show two postsynaptic components in one third of the units (11/37), a P-component was never observed. The typical parameter combination of units in cluster II (n = 43) is low peak- over-total values (0.04 \[0.02; 0.06\]), long inter-spike intervals (4.5 ms \[3.8; 5.4\]), and a large variation of the latter (2.7 ms \[2.4; 3.5\]), low maximum discharge rates (138 Hz \[113; 171\]), and high frequency selectivity (Q<sub>10</sub> values: 2.05 \[1.38; 4.05\]). The signal-to-noise ratios of these units (6.3 \[5.5; 7.5\]) was the lowest in the entire sample. At the same time, most units in this cluster displayed complex waveforms; in one third of the units (11/34) prepotentials were recorded. In 9 other units (26%) the waveforms showed the two postsynaptic components. Units in cluster III (n = 20) are characterized by low spontaneous (0.1 Hz \[0; 8.8\]), but high maximum discharge rates (307 Hz \[238; 376\]). The frequency tuning curves are mostly asymmetric, i.e. they showed large differences in the slope of the low and high frequency flanks (Δ: 0.5 octaves/10 dB \[0.1; 0.9\]), large differences between BF and CF (Δ: 2.2 octaves \[1.1; 2.7\]), and between the discharge rates at the respective frequencies (Δ: 151 Hz \[92; 206\]). The RLF are mostly monotonic (47%; 9/19) or monotonic- plateau (37%, 7/19) rather than non-monotonic (16%, 3/19). The dynamic ranges of these units (45 dB±18\]) were the largest in the entire sample. Inhibitory sidebands were seen in only 2/20 cases (10%). Cluster IV units (n = 38) show high peak-over-total values (0.13 \[0.08; 0.18\]), short first spike latencies (3.8 ms \[3.2; 4.2\]), but at the same time high jitter values (1.4 ms \[1.2; 1.9\]) and high spontaneous rates (68 Hz \[36; 87\]). These units had the highest incidence of inhibitory sidebands (89%, 34/38), and they responded well to fast fluctuations in signal amplitude. In conclusion, the clusters establish well separated groups of units which differ significantly in a sizeable number of properties (see also). Still, the different clusters show a strong overlap in the parameter fields. Thus, it may not be appropriate to think of the identified clusters as rigidly segregated classes of AVCN units. To visualize the clusters in a lower number of dimensions a principal component analysis was performed. The principal component analysis linearly maps the high- dimensional representation to a lower dimensional space with its axes defined by weighted combinations of the original properties, the principal components. Clusters which are separated in the higher dimensions often remain separated in the lower dimensional projection. The weights of the single units on the two first principal components (PCs) are plotted in. The properties that contribute mostly to PC1 are the maximum discharge rate (13%), the mean inter-spike interval (11%), and variation in inter-spike intervals (13%). The further left a unit lies on the PC1 scale, the higher was its maximum discharge rate (r = −0.75), the lower was its mean inter- spike interval (r = −0.67) and the variation of its inter-spike intervals (r = −0.76). But other properties correlate with the PC1 as well, e.g. the entrainment (r = −0.64) and the peak-over-total values (r = −0.55). The PC2 is influenced mostly by CF threshold (12%). The lower the value was on this PC scale, the higher was the unit's threshold (r = −0.60). All other properties contribute only little to PC2 (each less than 10%) and show only weak correlations (r\<0.5). In the two dimensional plot, the individual clusters occupy different regions, but the separation of the cluster is not perfect. Some units of cluster I are located in an area which contains primarily units of cluster III or V, some units of cluster IV in an area predominantly containing cluster II units, etc. Especially in the middle of the plot the clusters strongly overlap. Correspondingly, the first two PCs only constitute 28% (PC1: 17%, PC2: 12%) of the overall variance. A potential problem of the analyses presented above is that missing values were replenished with the mean of the unit sample. This could be the reason for the accumulation of units in the centre. To address this issue, additional analyses were done with a reduced set of parameters, i.e. fewer properties. ### Reducing parameters The following ten properties were used in this analysis: peak-over-total, mean and standard deviation of the inter-spike intervals and the first spike latency, CF and CF threshold, spontaneous and maximum discharge rate, Q<sub>10</sub> value and dynamic range. These physiological properties are easy to quantify from a limited number of extracellular recordings thus reducing the number of missing values. Consequently, all units with missing values were removed from analysis, which reduced the sample size from 233 to 174 units. The results of this more restricted analysis are presented in. The dendrogram of the hierarchical-agglomerative procedure now results in four main clusters (a–d), but still the two analyses yielded partially comparable results (compare with). In the restricted analysis, again one of the clusters shows a stronger separation from the rest. Here, it is cluster ‘a’ which shares some properties with cluster V in. The units in this cluster show short inter-spike intervals and a low variation of it; high maximum discharge rates and low CF-thresholds. There are also partial overlaps of the other clusters with clusters of the extended analysis. Cluster ‘b’ has its counterpart in cluster I, and ‘c’ and ‘d’ in clusters II and IV, respectively. On average, the more restricted analysis shows smaller linkage distances within and larger between the clusters. Consequently, the resulting clusters are more homogeneous and differ more from each other. Still, the principal component analysis on the whole unit sample results in one big cluster of units instead of clearly separated groups (1+2). We verified that this is not only an artefact of the projection into two dimensions by considering the results in three dimensions. Similar results for hierarchical clustering and principal component analysis were obtained when other parameter combinations were used, confirming the reliability of the depicted results. ### Reconsidering the classification based on PSTH-types To evaluate potential links between the hierarchical cluster analysis and the classification based on the PSTH types, we assessed how the PSTH types fit into the clusters obtained by the previous analysis. For this, the distribution of the different PSTH types in the multidimensional parameter space was analysed. None of the clusters contains exclusively one PSTH type, but the distribution of the PSTH types differs between the clusters. Cluster I mainly includes PL (30%, n = 16), PL<sub>N</sub> (30%, n = 16), and C<sub>T</sub> (19%, n = 10) units, while the proportion of C<sub>S</sub> (11%, n = 6), O<sub>inh</sub> (6%, n = 3), and phase-locking units (3%, n = 2) was low. In cluster II the majority of units have a PL<sub>N</sub> (52%, n = 22) or a PL PSTH (24%, n = 10). In cluster III 58% (n = 11) of the units respond with a C<sub>T</sub> pattern and 16% (n = 3) with PL and PL<sub>N</sub> respectively. Cluster IV mainly consists of PL (45%, n = 17), O<sub>inh</sub> (24%, n = 9) and C<sub>T</sub> (18%, n = 7) units. The dominant response patterns in cluster V are C<sub>T</sub> (38%, n = 30) and C<sub>S</sub> (36%, n = 29). In the principal component analysis, the PSTH types show a stronger overlap than the clusters obtained from hierarchical clustering. Still, units of different PSTH types tend to occupy different regions of the parameter space. The chopper units generally were found at the upper left side of the plot, and PL and PL<sub>N</sub> units more at the right side. Moreover, the PL units tend to be located in the upper quadrant and the PL<sub>N</sub> in the lower one. To exclude the effect of possible misleading PSTH classifications caused by outliers, we checked the temporal response patterns of those units that were located separated relative to units of the same PSTH type. We could confirm most of our previous classifications, but few units we found with PSTHs that are not unambiguously to classify. For example the unit in : This unit could either be classified as PL unit, bases on the strong onset and the following sustained rate which is comparable to auditory nerve fiber responses, or as CT unit since there are two sharp peaks in the onset component and the CV value is above 0.35. This kind of PSTH shapes could somehow described as mixture of different ‘classic’ PSTH types. In conclusion, it can be stated that based on their physiological response properties, AVCN units can better be described as occupying specific areas in a coherent multidimensional parameter space, rather than forming clearly separable groups of units. # Discussion The present report provides the first comprehensive evaluation of the physiological properties of AVCN neurons in the Mongolian gerbil in vivo. More important, this study re-evaluates the question of classifying AVCN neurons. Based on our present results, we conclude that units in the AVCN form a continuum concerning their physiological response properties, with each unit having a fingerprint-like combination of several properties. Thus, the present results challenge the conventionally held idea of separable classes of AVCN units. However, we show that units, which stand out by their prototypic physiological response properties, are located towards the extremes of this continuum. The significance of a classification lies in the fact that the AVCN is the starting point of distinct ascending auditory pathways. Hypotheses about specific auditory processing along these pathways must consider the functional organisation of the AVCN. We will discuss possible causes for such a continuum and the consequences for the functional differentiation of the auditory pathways. ## Limitations of the study In the first part of the present study we demonstrate that our sample of AVCN units is in most respects comparable to those previously described in other mammalian species. There are only minor deviations, for example the relatively long latencies in PL units. Based on the (somewhat limited) data we have from these units for lower-level stimulation, we can rule out that level dependence or nonlinear RLFs are the determining factors for these differences. Also, it is unlikely that differences in spontaneous rates are casual for this result, since PL units tended to have higher spontaneous rates than the other cell classes, and such higher rates would cause a trend for latency shifts towards lower values. Finally, species differences can be ruled out, since Frisina et al. also observed shortest latencies for the PL units in the gerbil. However, in other respects the PL units are comparable with those previously described, but we could not finally rule out distortion effects on our analysis. Generally, it cannot be excluded that certain cell class distinctions would only become apparent, if a different or an even wider set of parameters/conditions were included in the cluster analysis. The limited recording time in in vivo single unit electrophysiology makes it necessary to seek a compromise between the use of a standardized stimulus repertoire and extensive variations of stimulus parameters. Still, in future studies it could be more revealing to analyze responses to more complex stimuli (i.e. broadband stimuli) or to study the system under continuous activation, e.g. via the use of long, natural stimuli. Such an approach would require the use of more general neuronal characterizations, e.g. spectro-temporal receptive fields or multilinear models as recently employed in the medial nucleus of the trapezoid body. Also, PSTHs were only collected at one common sound level (80 dB SPL) for the reasons explained in the section. If PSTH types were to systematically change for different relations of sound level and rate-level function of a given neuron, some cells which currently fall far away from the majority of units in their corresponding PSTH type might have been classified more consistently. However, for a subset of our neurons (n = 60) we also collected the PSTH at 20 dB above their threshold. The PSTH shape for low and high intensity stimulation was very similar when normalized for mean rate (r = 0.85). Similarly, the inter- spike interval histograms for the low and high intensity stimulation were highly correlated (r = 0.92). Therefore, including this low level stimulation into the cluster analyses was unlikely to provide a better distinction between the cells. Evidently not all possible parameters were included in the present analysis (e.g. phase-locking filter type could have been included if more data were available), but still we tried to include a wide range that could aid the classification. Additionally, we reran the cluster analysis and the PCA with different combinations of the acquired parameters to exclude distortion effects of any single parameter. None of the respective combinations revealed distinct physiological neuron types in our data sample. To support the physiological data, it would also be useful to have additional data about cell morphology and projection pattern. Unfortunately, single-unit labeling is not easy to implement in vivo for a substantial number of units. At the same time, including it would defy the purpose of achieving a physiological classification. The present focus on the Mongolian gerbil as a model system for human (low frequency) hearing inherits a potential caveat, as recent studies have reemphasized the potential deterioration of auditory processing due to microcysts forming also in the cochlear nucleus of the gerbil. The functional consequences of this gerbil-specific condition, which develops mostly after sexual maturity, are not clear. However, the present study used comparatively young gerbils (2–4 months) to avoid potential effects of altered CN morphology. Given that the Mongolian gerbil is a well-introduced animal model, future studies might consider the use of even younger specimen or switching to other species. ## Variability of response properties of AVCN neurons When dealing with physiological classifications of AVCN units in vivo, earlier studies mostly relied on the units' temporal response properties to pure tone bursts as visualized in the PSTHs. The present sample of AVCN units recorded in the Mongolian gerbil fits well into the ‘classical’ PSTH types described for a number of mammalian species (gerbil:, ; cat:, guinea pig:). However, the analysis presented here is based on a larger number of physiological features, all acquired from the same sample of AVCN units. Hence, the assessment of potentially distinct physiological neuron types rested on a broader base. With respect to their morphology, AVCN neurons are subdivided in at least three different classes; spherical and globular bushy cells and stellate cells (review:). Several studies attempted to relate these different groups of neurons to physiological response types yielded by acoustic stimulation (gerbil: ; rat:, ; guinea pigs: ; cat: –). Most studies agree that PL responses are best attributed to spherical bushy cells, PL<sub>N</sub> responses to globular bushy cells, and any kind of chopper responses to stellate cells. However, the same studies reported a number of cases where such a simple agreement between PSTH type and morphological cell type does not hold. Especially globular bushy cells are reported to show a variety of different response patterns. Besides the PL<sub>N</sub> pattern, PL and various onset patterns are found in this cell type. Similarly, also for spherical bushy cells and stellate cells a variety of PSTH types have been described. Thus, either the correspondence between morphology and PSTH types of AVCN neurons is very complex, or the PSTH alone is not sufficient to establish the quested structure-function relationship, or the two modes of classification are mostly independent from each other. A correlation of PSTH types with a number of other physiological properties, such as spontaneous rate and first spike latency, has been suggested in previous studies. But considering these correlations, it was not possible to separate the units into non-overlapping groups. This problem has already been addressed in a number of earlier studies, , which attempted to develop different schemes for classifying AVCN units based on PSTH types considering different features of the unit's response properties (e.g. spiking regularity and response latency). Still, the authors of these studies admitted that the “continuous filling of the parameter map” and the “overlap of the characteristics of neighbouring unit types” suggest that the classification schemes ”do not provide absolute borders between response types”. In contrast to previous studies, our study is based on a larger repertoire of physiological parameters. Additionally, we avoided to make any assumptions on the significance of specific parameters. By performing cluster analyses, an unbiased analysis was possible. Even with this precautionary measure it was not possible to identify clearly separated physiological unit types. The units rather have a fingerprint-like combination of several properties and show a wide distribution in the multidimensional parameter space. While this does not preclude a correlation of some physiological properties across units, it appeared hard to establish an unambiguous structure-function relationship on this basis. ## Possible causes for a physiological continuum It is widely accepted that the morphologically defined AVCN neuron types – establish the origin for several distinct ascending auditory pathways. This is tantamount to distinct projection patterns of the neuron types. Also, in in vitro preparations the units can be differentiated with respect to their basic physiological properties. Hence, the question arises, what could cause the higher degree of variability in physiological characteristics across different AVCN units in vivo? The more uniform features obtained in vitro might relate to the reduced complexity of the system. In vivo there are several levels at which the units' responses can be modulated to show a wider range of response characteristics. With respect to the afferent input, some variability already exists in the activity of auditory nerve fibres. Next, the degree of convergence of auditory nerve fibres onto single CN neurons varies, ranging from 3–5 in spherical bushy cells – to 9–69 in globular bushy cells. Even higher degrees in convergence were reported for stellate cells,. Furthermore, variations of the synaptic strength of auditory nerve fibers on AVCN neurons can determine the response properties of the postsynaptic cells. Additionally, there are secondary excitatory and inhibitory inputs, which also can affect the response. Furthermore, postsynaptic properties can determine a unit's response as well. The units show a certain degree of variability in their morphology, first of all the size of their somata and dendritic branching patterns. Moreover, the neurons differ in the expression of specific subsets of receptors, ion channels and membrane proteins – which influence their membrane characteristics. It is conceivable that a multifold of these pre- and postsynaptic influences contribute to the establishment of a continuum in response properties. ## Consequences for concepts of auditory pathways and functional implications As discussed above, it cannot be ruled out that future studies with a different set of stimuli will provide a basis for an unambiguous classification. However, if the continuum of physiological properties cannot be resolved, the notion of classifying AVCN neurons on a physiological level would have to be reconsidered. Classification schemes have a great value in organizing our understanding of the system, but they do not have to reflect actual boundaries. While such classifications are often based on the description of prototypical representatives of each class, many cells appear to fill the parameter space between the prototypes and only fit ‘more or less’ into the classification. In the ‘classical’ concept, the AVCN is composed of a set of distinct cell classes, each of which occupies a well-defined location in the space of all possible combinations of properties. While such a representation is favourable from the perspective of a human observer (since it allows an easier conceptualization), it need not to be favourable from a more general coding perspective. As has been shown theoretically (e.g. liquid state machines), that a broad representation of various parameter combinations considerably simplifies the task of extracting many complex aspects of the original signal. In this sense the presently observed continuum of cell properties could be advantageous for processing in subsequent higher-order nuclei of the ascending auditory system. The authors thank the anonymous reviewers for their dedication to improve the manuscript. We also thank Anita Karcz and Sandra Tolnai for comments on earlier versions of the manuscript and for their encouragement during the writing process. Special thanks are also dedicated to Doris Freyberg and Gudrun König for keeping the lab running. [^1]: Conceived and designed the experiments: MT CKS RR. Performed the experiments: MT CKS. Analyzed the data: MT BE SD MS. Contributed reagents/materials/analysis tools: BE RR. Wrote the paper: MT BE SD CKS RR. [^2]: Current address: Schulich School of Medicine and Dentistry, University of Western Ontario, London, Ontario, Canada [^3]: Current address: Neural Systems Laboratory, University of Maryland, College Park, Maryland, United States of America [^4]: Current address: Bernstein Focus Neurotechnology, Georg-August- University Göttingen, Göttingen, Germany [^5]: Current address: MRC Toxicology Unit, University of Leicester, Leicester, United Kingdom [^6]: The authors have declared that no competing interests exist.
# Introduction Since the early observations of Darwin and von Humboldt, ecologists have attempted to explain why species diversity increases towards the Equator. Relationships of biodiversity with net primary productivity (NPP) are frequently suggested as potential explanations, and NPP is usually the best correlate of biodiversity. Productivity-diversity relationships are assumed to be driven by the NPP of an ecosystem as a result of increased provision of vital resources. For example, climates which are highly deficient in water (such as desert) or energy (such as arctic) have both low NPP and low species richness. There is, however, little consensus on the mechanisms underpinning increases in diversity from intermediate to high levels of NPP. Furthermore, the shape of productivity-diversity relationships is typically dependent on the spatial grain of the analysis. For example, linear increases in species richness in large sampling units (such as degree grids) contrast with the multiple ways in which species richness within local habitats responds to NPP. Thus, although productivity-diversity relationships are widespread, the underlying mechanisms still need to be resolved. In other cases however, biodiversity is more strongly correlated to ambient energy than to NPP. In contrast to NPP, ambient energy does not include water availability and can be expressed as annual mean temperature (T). Possible mechanisms for increasing species richness with ambient energy include tropical niche conservatism, dispersal limitation after glaciation and metabolic theory. Based on the evolutionary origin of many taxa in tropical climates, their occurrence in cooler climates depends on the evolution of cold-tolerance. Within larger taxonomic groups, communities in warm climates often include many basal taxa, while communities in temperate to cold climates are increasingly restricted to few derived taxa. This results in a positive relationship of species richness and T. Analyses of the phylogenetic structure of communities provide evidence for this mechanism, notably in butterflies. Another effect of historical climate on contemporary richness patterns has been described for European dung beetles. Their limit of thermal tolerance during the last glacial maximum marks a strong change in current richness with low species numbers north of this limit. Thus, limited colonization of areas with historically unsuitable climate can contribute to current correlations between richness and ambient energy. Metabolic theory provides an additional mechanism for higher species richness in warmer climates based on the influence of temperature on metabolic rates and on rates of speciation. Metabolic theory predicts a direct, monotonic relationship of species richness with T, whereby the slope of log<sub>e</sub> –transformed species richness with inverse T is predicted to be −0.65. A third framework for broad-scale patterns of species richness is biological relativity to water-energy dynamics. It is based on the dependence of all life on the availability of water in a liquid form. This framework suggests that species richness is proportional to (a) the availability of liquid water (annual rainfall, R), and (b) the lowest monthly value of potential evapotranspiration (PET<sub>min</sub>). This “interim general model” explains almost 80% of the variation in species richness of trees and shrubs in eastern and southern Africa. At higher latitudes (Europe and North America) the effect of rainfall on species richness of trees dominates (r = 0.64), and effects of PET<sub>min</sub> or other energy variables are non-significant. The final determinant of species richness patterns considered here is spatial environmental heterogeneity. Its importance can be expected to increase with increasing size of sampling units (spatial grain). For example, only limited spatial climatic heterogeneity can be expected along a 21 m transect within one habitat type (the small sampling grain in our study, see Methods). In contrast, large differences in internal spatial climatic heterogeneity exist among countries (the large sampling grain in our study). For example, T differs by 13.3°C between the warmest and the coldest 10×10 arc minute square within Switzerland – more than ten times the difference within Denmark, which has a similar surface area. Effects of spatial environmental heterogeneity on species richness can be explained with niche theory, which assumes different environmental preferences and tolerances among species. Among environmental variables, climate is central to the distribution and persistence of species worldwide. In particular, the distribution of numerous taxa is influenced by T, for example plants, beetles, spiders and birds. Thus, heterogeneity of environmental conditions (such as T) within large sampling units can enhance overall species richness by providing suitable conditions for larger numbers of species with different ecological niches. Increased species richness due to an increased importance of spatial environmental heterogeneity will consequently shift the focus from alpha diversity (local species richness) to beta diversity (turnover of species in space). Here, we explore species richness-environment relationships of European arthropods. We combine the results of a continent-wide standardised sampling programme of local ground-dwelling arthropod communities (local grain) with existing coarse-grained country inventories (country grain) of comparable spatial extent. At the local grain ground-dwelling ants (Formicidae), beetles (Coleoptera), bugs (Hemiptera) and spiders (Araneae) were sampled in 32 habitats at 16 locations across Europe, ranging from boreal to Mediterranean in climate. At the country grain, inventories of 25 European countries were obtained for the same groups. We used these data to test each of the above hypotheses by first comparing the explanatory power of productivity, ambient energy, the interim general model, and the best possible statistical model (drawn from all variables) for biodiversity within local habitats (alpha diversity). Secondly, we compared the potential for environmental heterogeneity to explain species turnover across locations (beta diversity). Thirdly, we compared the explanatory power of productivity, ambient energy, the interim general model, spatial environmental heterogeneity and the best possible statistical model (drawn from all variables) for biodiversity within countries (gamma diversity). # Methods ## Ethics 1. Field sites were selected and established within the EU FP6 ALARM project to form a long-lasting research network 2. Each field site had a site manager, responsible for contacts to local authorities and/or land owners 3. In most cases the land belonged to regional research stations 4. More detailed descriptions can be found in reference 5. Protected areas or rare habitats were not included into this field site network 6. We did not include protected species. The employed pitfall traps capture invertebrates, with no protected species affected in the habitats we sampled. No permissions are needed to use pitfall traps outside of protected areas. ## Data Ground-dwelling arthropods were sampled in 32 habitats at 16 locations across Europe. Thirteen locations were part of the ALARM field site network, and three sites (Bern, Silkeborg, Wien) were added to fill geographic gaps. As far as possible, one forest as an example of a near-natural habitat and one cereal field as an example of an intensive agricultural habitat were sampled in each location. When unavailable, other near-natural habitats (scrubland or extensive grassland) and other intensive agricultural habitats (intensive grassland or olive grove) were sampled instead. Trapping took place in 2006 and started five days after the beginning of the vegetation period (the onset of growth in the majority of plant species) in each location. In each habitat, eight pitfall traps of 7 cm diameter were placed along a transect and separated by 3 m from each other. The traps were filled with 0.1 L of a 4% formaldehyde solution, to which sodium dodecyl sulphate was added as detergent. Three sampling periods of two weeks were separated by pauses of two weeks. Adult arthropods were identified to species level by specialists. The following groups were considered: ants (Hymenoptera: Formicidae), beetles (Coleoptera: Carabidae, Curculionoidea, and Staphylinidae), bugs (Hemiptera: Auchenorrhyncha and Heteroptera), and spiders (Araneae). In addition to the analysis of taxonomic groups, we divided the studied arthropods into trophic groups according to the dominant feeding type in the respective family. Herbivore families were all Curculionoidea, all Auchenorrhyncha and the heteropteran families Berytidae, Cydnidae, Lygaeidae, Miridae, Pentatomidae, Piesmatidae, Plataspidae, Pyrrhocoridae, Rhopalidae, Scutelleridae and Tingidae. Spiders, ants and the remaining beetle and bug families were carnivores. Detritivores could not be analysed because this feeding type did not dominate in any of the sampled families. Species numbers within 25 European countries were taken from the Fauna Europaea database. Countries smaller than 30,000 km<sup>2</sup> were excluded, as were all countries for which the known number of arthropod species lay below the 95% confidence interval of the species - log area relationship, indicating effects of insularity (Republic of Ireland) or incomplete knowledge of the arthropod fauna (Ukraine, Belarus). We further excluded islands such as the Balearic Islands, Corsica, Greek Islands, Sardinia and Sicily from the respective mainland areas. Environmental variables were extracted from a European gridded data set with a monthly time step and a spatial resolution of 10×10 arc minutes, which corresponds approximately to 16 km. The primary climatic variables temperature (T) and precipitation were constructed through interpolation from station observations. Annual rainfall (R;) was calculated as the sum of precipitation in all months with an average temperature \>0°C. NPP was estimated by running the LPJ-GUESS ecosystem model, with the same climate data, and parameterized for the potential natural vegetation of Europe. LPJ-GUESS and the closely related LPJ-DGVM have formerly been shown to reproduce observed variations in NPP across various types of vegetation and climates. PET<sub>min</sub> represents potential evapotranspiration of the coldest month of each year, and was calculated using the Thornthwaite equation, which only requires knowledge of air temperature. We used long-term annual means of NPP, T, R and PET<sub>min</sub> from 1971 to 2000. For analysing species richness within habitats and species turnover across habitats, we used NPP, T, R and PET<sub>min</sub> values of the grid in which the habitats were located. For analysing species richness within countries, explanatory variables were averaged across all grids of the country. In addition, spatial heterogeneity in NPP, T, R and PET<sub>min</sub> were calculated as ranges for each country by subtracting the minimal from the maximal value, respectively (i.e. difference between the grid cells with the highest and lowest value). ## Analysis Local species richness and species richness in European countries was analysed using linear models with standardised explanatory terms (mean = zero, standard deviation = 1) in the statistical environment R version 2.12.0. To account for possible differences in sampling efficiency between locations, we used the number of captured individuals N as a covariate in the analyses of local species richness. We accounted for possible effects of spatial autocorrelation of the habitats within the locations and among the locations and countries with generalised least squares with spatial simultaneous autoregressive error models. Models were based on neighbourhood matrices that linked the two habitats within a location and each location with at least one other location for the local grain analyses and allowed each country to be in the neighbourhood of at least one other country, i.e. at a maximum distance of about 850 km from centre to centre. For this we used the package spdep. In addition, we calculated Moran’s I correlograms for the residuals of models with and without correction for spatial autocorrelation to assess if the tested theories miss important spatially structured environmental variables. Missing crucial spatially structured environmental variables will lead to significant residual spatial autocorrelation of the uncorrected models. Separate models of local species richness were calculated according to productivity-diversity relationships, ambient energy and the interim general model, plus one “Best” model in which all explanatory variables relevant for the different theories (NPP, NPP<sup>2</sup>, T, R, PET<sub>min</sub> and PET<sub>min</sub><sup>2</sup>) entered the initial models. We included interactions of all linear terms with habitat to test if there are different responses in the different habitat types. We identified the minimal adequate models by a backwards variable selection procedure according to the second order Akaike information criterion (AICc). Linear terms were always kept in the model when the respective quadratic term increased the model fit. In cases of high collinearity (Pearsson r \>0.5) of linear terms we calculated separate models always containing only one of these terms, and the best model was chosen based on the AICc model selection criteria. This restricted the models to only one energy variable (either T or PET<sub>min</sub>). It also reduced the risk of overfitting, which is considerable given the low numbers of replicates (N = 16 locations and N = 25 countries). As relationships of species richness with NPP can be either linear or hump- shaped, we allowed the quadratic term of NPP to remain in the productivity- diversity relationship models if that resulted in lower AICc values. Ambient energy models were calculated using untransformed species richness and T. To test predictions made by metabolic theory, we calculated the slope of log<sub>e</sub>(species richness) with 1/\[0.0000862(273+T)\] for comparison with the predicted slope of −0.65. We used the first version of the interim general model (IGM1), where species richness is explained by a linear term of R plus a linear and quadratic term of PET<sub>min</sub>. At the country grain, R and PET<sub>min</sub> had a Pearson correlation coefficient of 0.6. Nevertheless, we also tested the full model including both variables for means of completeness. We also performed an influence analysis using Cook’s distance. If data points had a Cook’s distance \>0.5, indicating disproportional weight in the regression analysis, then the effect of excluding those data points from the model was examined. Relationships of species turnover with environmental variables were analysed using Mantel tests. We used presence-absence data of the trapped species per location (natural and disturbed habitat combined). Community dissimilarities were calculated as Morisita-Horn distances and related to Euclidean environmental distances between all possible pairs of sites. Each environmental variable was tested separately, and separate tests were calculated for spiders, beetles, bugs, ants, all groups combined, herbivores and carnivores using the function mantel in the package vegan (default settings). The significance was based on Monte Carlo tests with 999 permutations. Species richness of arthropods in European countries was analysed in a similar way as local species richness, with the following additions. We corrected for the area of the countries by including log<sub>10</sub>(area) as an additional explanatory variable in all models. In addition to productivity-diversity relationships, ambient energy and interim general models, we calculated a model containing spatial environmental heterogeneity. Variables considered were the ranges of NPP, T, R and PET<sub>min</sub>. However, all these range variables were highly intercorrelated and thus it was not possible to include them simultaneously in one model. Therefore, we calculated separate models always containing one of these terms and selected the best model according to AICc. Again, we calculated “Best” models in which all explanatory variables relevant for the different theories entered the set of initial models containing only one of the range variables characterising spatial environmental heterogeneity. # Results ## Species Richness in Local Habitats The samples contained 33223 individuals of our focal taxa that comprised 83 ant species (Formicidae), 444 beetle species (Coleoptera), 185 bug species (Hemiptera) and 354 spider species (Araneae). Relationships between species richness and environmental variables at the local grain were highly variable. All arthropods combined and carnivores considered separately showed a hump- shaped relationship with NPP, while lacking a significant effect of ambient energy or variables from the interim general model. The ambient energy model was best for spiders, while beetles conformed most to the energy term of the interim general model. Bugs and herbivores were not significantly affected by any environmental variable. Ants showed a negative relationship to R plus an interactive effect of T and habitat type. Ant species richness increased with T in near-natural habitats, but did not change significantly with T in intensive agriculture. A more detailed examination of the response of ants to climate and habitat type has been given elsewhere. Habitat type had a significant effect only on spiders, with higher species richness in near-natural than in intensive agricultural habitats. In contrast to ants, there was a significant interactive effect of T and habitat type on spider richness, whereby species richness increased with T in agricultural habitats but did not significantly change with T in near-natural habitats. With respect to metabolic theory, only ants in near-natural habitats and spiders in intensive agricultural habitats had negative slopes of log<sub>e</sub> (species richness) with T<sup>−1</sup> of −0.47±0.56 95% confidence interval (CI) and −0.43±0.27 95% CI, respectively. In all other cases, the slopes were positive and differed significantly from the predicted value of −0.65 (all arthropods: 0.20±0.18 95% CI; beetles: 0.12±0.30 95% CI; bugs: 0.05±0.21 95% CI; herbivores: 0.15±0.28 95% CI; carnivores: 0.21±0.17 95% CI; ants in intensive agricultural habitats: 0.22±0.43 95% CI; spiders in near-natural habitats: 0.12±0.55 95% CI). ## Species Turnover Species turnover across the 16 locations was most strongly correlated with differences in T. Correlations were highest for spiders, carnivores, beetles and all groups combined, followed by ants and herbivores. Only bugs showed no significant relationship of species turnover with environmental variables. In the remaining groups, correlations of species turnover with differences in T were at least 49% stronger than with any other environmental variable. Correlations of species turnover with differences in NPP and/or PET<sub>min</sub> were significant, but substantially less strong than those with differences in T. ## Species Richness in Countries All tested environmental variables showed some significant effects on species richness within countries. With respect to productivity-diversity relationships, species richness of spiders, beetles, bugs and all groups combined increased with NPP. The ambient energy-models revealed increased species richness of beetles, bugs, ants, herbivores, carnivores and all groups combined with T. However, for all groups combined this relationship was no longer significant when an overly influential data point (Portugal) was omitted from the analysis. As for metabolic theory, the slope of log<sub>e</sub> (species richness) versus T<sup>−1</sup> came close to the predicted value of −0.65 for ants (−0.64±0.38 95% CI), but was shallower in the remaining cases (all groups: −0.36±0.29 95% CI, spiders: −0.21±0.29 95% CI, beetles: −0.38±0.32 95% CI, bugs: −0.36±0.26 95% CI, herbivores: −0.34±0.29 95% CI, carnivores: −0.38±0.30 95% CI). With respect to biological relativity to water-energy dynamics, the full interim general model including both variables (R and linear and quadratic terms of PET<sub>min</sub>) had constantly higher AICc values than simplified models. The reduced interim general model for spiders revealed an unexpected negative response of species richness to rainfall. The remaining groups showed hump- shaped relationships with PET<sub>min</sub> in accordance with biological relativity to water-energy dynamics. Portugal was overly influential in the interim general model for ants, and no significant model remained after its removal. There were consistent positive relationships of arthropod richness with spatial environmental heterogeneity. T<sub>range</sub> gave a better model fit than NPP<sub>range</sub>, R<sub>range</sub> and PET<sub>minrange</sub> in all cases. Models with free variable selection always combined variables from several theories. They were statistically superior to any single theory according to their higher explanatory power and lower AICc values (ΔAICc \>6.1). Residuals showed significant spatial autocorrelation in the majority of single- theory models, but in none of the models with free variable selection (“Best”). This suggests that the models with free variable selection included the majority of relevant variables while single theories tended to miss crucial information. In accordance with a high importance of spatial environmental heterogeneity, species richness increased with T<sub>range</sub> in all models with free variable selection. # Discussion Although our results from the local habitat samples were variable with respect to environmental effects on species richness, there were numerous significant effects of environmental variables on species richness of the same groups at the country grain. This suggests that productivity-diversity relationships, ambient energy, the interim general model and spatial environmental heterogeneity all contribute to the explanation of arthropod species richness of European countries. However, model selection according to AICc identified T<sub>range</sub> as the strongest predictor of arthropod richness across all studied groups. Independent support for a strong role of spatial heterogeneity in T comes from the significant relationship of species turnover across locations with differences in T. If species turnover across locations is driven by T, then countries with a high T<sub>range</sub> will contain higher beta diversity and consequently more species in total than countries with more uniform temperatures. In the following, we will discuss the different theories for broad-scale gradients in species richness and what can be concluded from our data. ## Productivity-diversity Relationships The observed increase of species richness in countries with NPP is in accordance with the majority of studies on broad-scale relationships of species richness with climate. In contrast, relationships of species richness with NPP at the local grain were hump-shaped and restricted to carnivores and to the sum of all arthropod species. This accords with a generally reduced effect size, and with a transition from monotonous to hump-shaped productivity-diversity relationships towards small spatial grain. The differences between grains cannot be explained by differences in gradient length, because NPP varied only slightly more among locations than among countries (locations: 0.39 g C m<sup>−2</sup> a<sup>−1</sup> in Lesvos to 0.66 g C m<sup>−2</sup> a<sup>−1</sup> near Kraków; countries: 0.41 g C m<sup>−2</sup> a<sup>−1</sup> in Norway to 0.65 g C m<sup>−2</sup> a<sup>−1</sup> in France;,). Nevertheless, the NPP gradient was relatively short at both grains. When gradients include areas with very low NPP, stronger effects at the local grain would be expected. Thus, any conclusions with respect to small-scale productivity-diversity relationships from our data should be made with caution. ## Ambient Energy We found significant effects of T on species richness at both local and country grains. The increase of ant species richness with T in both near- natural local habitats and in countries accords with ambient energy theories. In contrast, spider richness increased with T in intensive agricultural habitats but not at the country grain. This suggests that the richness pattern of spiders in agricultural habitats is not indicative of other habitat types and thus of limited relevance for their overall species richness in countries. Beetles, bugs, carnivores and herbivores showed significant positive relationships with T at the country, but not at the local grain. T can influence species via its effect on NPP. Low temperatures limit terrestrial NPP in temperate to arctic climates, and large parts of our study area lie in the temperate to boreal region. However, the latitudinal gradient of NPP was unimodal in our study, with decreasing NPP from temperate to Mediterranean climate. At the country scale, all groups except spiders and beetles were significantly affected by T in addition to NPP, suggesting direct effects of ambient energy on species richness. Across countries, the slope of log<sub>e</sub>(species richness) versus T<sup>−1</sup> accorded with metabolic theory only for ants. Species richness of ants has been found earlier to conform with metabolic theory. Our results suggest that this may be an exception rather than the rule among terrestrial arthropods. Based on widespread nonlinearity, geographic and taxonomic dependence of temperature-richness relationships, metabolic theory has been more generally questioned. Deviations from metabolic theory can be due to violations of its assumptions. For example, the assumption of body size invariance with temperature is violated by the significant increase of spider body size across Europe with temperature. Tests of tropical niche conservatism and dispersal limitation after glaciation require phylogenetic analyses that exceed the scope of the current investigation. These historical climatic explanations predict more basal taxa in warm climates and high richness of few derived taxa in cooler climate. Thus, the richness of higher taxonomic categories such as families should increase more strongly towards warm climate than the number of species. Such a pattern is present in our spider data: while species richness shows no significant relationship with ambient energy at the country grain, the number of spider families increases with T across countries (t<sub>1,23</sub> = 4.9, p\<0.001). This indicates niche conservatism in warmer climates and encourages more detailed phylogenetic exploration of the distribution of European arthropods. ## Temperature, Species Turnover and Spatial Environmental Heterogeneity In contrast to its variable effect on species richness within local habitats (alpha diversity), differences in T had strong effects on species turnover across locations (beta diversity;). This role of T confirms that it represents an important niche dimension of European arthropods. The number of available niches in a given area thus correlates to the range of temperatures present in that area. Our results are consistent with the ideas that species richness is enhanced by (i) elevational range and (ii) habitat heterogeneity in an area. Elevation is a main driver of temperature variation in mountains, leading to correlations between elevational range and spatial heterogeneity in temperature (r = 0.92 for the countries studied here). The occurrence of the same ecosystems at similar temperatures across the world that have contrasting elevations demonstrates that temperature is more crucial for biodiversity than elevation per se. Being difficult to measure, habitat heterogeneity is often determined by the number of distinguishable vegetation types present in an area. In near-natural situations, vegetation types are in turn determined by the environmental preferences of their constituent plant species, including their temperature preference. Thus, spatial environmental heterogeneity, elevational range and habitat heterogeneity are interrelated, and we presume that climate often has the most direct influence on biodiversity. In our study, variability in T had a dominant effect on species turnover and gamma diversity. Apart from the existence of more niches along large temperature gradients, climatic heterogeneity can also buffer species extinctions by allowing species confronted with climatic fluctuations to relocate to suitable climatic refuges. The dominant role of T<sub>range</sub> for arthropod species richness in European countries is in accordance with both enhanced niche availability and reduced extinction during climatic fluctuations. ## Biological Relativity to Water-energy Dynamics Biological relativity to water-energy dynamics is expressed in the interim general model. It predicts increasing species richness with R (water term) and a unimodal relationship of species richness with PET<sub>min</sub> (energy term). We found only partial support for IGM, since reduced models always resulted in lower AICc values compared to the full model. Numbers of spider species per country decreased with R in the interim general model, but increased with R in the model with free variable selection. Dominant effects of the energy term in the interim general models at the country grain were replaced by group-specific positive, negative (ants) or absent (herbivores) effects of R in the models with free variable selection. Given these inconsistencies, the interim general model provided no robust explanation of arthropod richness in our study. Nevertheless, the numerous significant relationships with R and PET<sub>min</sub> suggest that the interim general model may apply to European arthropods, but that larger datasets are necessary to disentangle its components. The dominant effect of the water term in the models with free variable selection is in accordance with Hawkins et al., who found that the energy term of the interim general model becomes dispensable in temperate to arctic climate. ## Herbivores versus Carnivores The strength of latitudinal diversity gradients has been found to increase across trophic levels. In our study, the major difference between herbivores and carnivores was at the local grain, where carnivores showed a hump-shaped relationship to NPP and herbivores no significant relationship at all. At the country grain, the results for herbivores and carnivores were similar to each other and to those of bugs and all arthropods combined. This suggests that the observed differences between arthropod taxa are due to their different phylogeny or other life-history traits rather than caused by their trophic position. ## Countries versus Local Habitats The stronger and more consistent effects in countries versus locations are in accordance with the general decrease of species-richness environment relationships towards small spatial grain. Ecological processes are scale- dependent, and effects of the studied broad-scale environmental conditions may affect regional species pools rather than local assemblages. However, species pools can affect local species richness, especially in mobile organisms such as the studied arthropod groups. In addition, some of the mechanisms to explain broad-scale patterns in species richness include local processes such as resource partitioning (productivity-diversity relationships), metabolism (metabolic theory), or cold tolerance (tropical niche conservatism). Thus, contrasts such as the variable role of ambient energy at locations and its widespread positive effect across countries are remarkable and provide a starting point for further research. Differences in data quality could have contributed to the stronger effects in countries versus locations. First, our pitfall traps sampled only ground- dwelling arthropods over a limited time period. The 33223 sampled individuals of 1066 species already represent a major identification effort. Nevertheless, sampling intensity per habitat corresponded only to a minimal effort that is expected to encompass around 75% of all species attainable at the respective site with pitfall traps. Even more problematic can be differences in sampling efficiency between sites due, for example, to weather or habitat structure. We reduced these differences by applying strictly standardized sampling methods and by including the number of individuals captured per habitat as a factor in the models of species richness. By using (log) individual numbers as a factor, we assume that true abundances are similar across habitats and that observed differences in individual numbers are due to variation in sampling efficiency. However, true abundances may differ. Results at the local grain would change strongly if individual numbers were excluded from the models – the only consistency being increased ant richness with ambient energy and variable effects between groups (results not shown). This highlights that the difficulty to obtain large, representative arthropod samples from defined areas remains a main obstacle in community ecology. Accordingly, conclusions from the results at the local grain should be drawn with care. In contrast to these sampling issues at the local grain, species inventories of countries are the results of many decades of research and have reached asymptotes in almost all cases. A second difference in data quality between locations and countries relates to the environmental data. Interpolated climatic variables will result in relatively accurate values across large areas such as countries, but have only limited accuracy at the grain of local habitats. Sources of error include spatially and temporally unpredictable factors such as rainfall, as well as anthropogenic effects on productivity, especially in disturbed habitats. Thus, sampling error at the locations is probably much higher than at the country grain, underlining the need for additional high-quality inventories of invertebrates along environmental gradients of large spatial extent. # Main Conclusions Our study supports the scale-dependence of species richness-environment relationships. While relationships of local species richness with environmental variables were contingent on the arthropod group, species richness patterns at the country grain were more consistent and partly supported all tested theories. Niche theory provides a plausible link between the two grains: On the one hand, differences in temperature were the best correlate of species turnover across locations. On the other hand, spatial heterogeneity in annual mean temperature had the strongest effects on arthropod diversity within European countries. These two independent findings suggest that temperature is an important niche dimension and that countries with wider ranges in annual mean temperature provide a greater breadth of niche space and so can support larger numbers of arthropod species. Unless environmental heterogeneity is constant across sampling units (thinkable e.g. in marine environments), we strongly suggest that studies with large sampling units take into account environmental heterogeneity, just as studies with variable area of sampling units nowadays routinely consider area. We thank Monica Wyss-Lopez for logistic support during the field survey and Eva- Maria Gerstner from the LOEWE Biodiversity and Climate Research Centre in Frankfurt (BiK-F) for preparing climate data. The traps were operated by Riccardo Bommarco, Sara Bonzini, Eduardas Budrys, Torben Christensen, Dawid Moron, Thomas Frank, Michael Greenwell, Steen Hansen, Mari Moora, Joan Pino, Simon Potts, Laszlo Rakosy, Agnès Rortais, Jane Stout, Ivan Torres, Catrin Westphal, and Monica Wyss-Lopez. The ALARM Field Site Network was coordinated by Koos Biesmeijer and Bill Kunin. Arthropods were identified by Lina Almeida, Angelo Bolzern, Antonio D. Brescovit, Thorsten Burkhardt, Christoph Germann, Joachim Holstein, Marco Isaia, Pekka Lehtinen, Antonio Melic, Christoph Muster, Milan Pavouk, Stano Pekar, Cristina A. Rheims, Adalberto J. Santos, Jörg Wunderlich, and Monica Wyss-Lopez. We thank Michael Nobis for the kind provision of the modified R package hier.part. The manuscript benefitted greatly from suggestions by Jonathan Belmaker, Richard Field, Gary Mittelbach, Jason Tylianakis, and three anonymous reviewers. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WN. Performed the experiments: WN XE SK MHE OS SB BAW TH. Analyzed the data: MHE OS. Wrote the paper: MHE OS SB XE TH SK BAW WN.
# 1 Introduction Brain networks derived from neuroimaging data have been shown to quantify the level of brain atrophy, and hence the relative stage of neurological disease and identify disease related changes. Cortical networks from structural magnetic resonance imaging (MRI) consist of nodes which represent brain regions of interest (ROI) and edges that link two nodes if these regions have spatial correlation or similarity. Unlike networks from ROI volume or surface area, cortical thickness networks have been shown to be a more stable measure along the Alzheimer’s Disease (AD) continuum. This is because cortical thickness is a direct measure of cortical atrophy due to cytoarchitectural features of the cortex tissue. The resultant networks characterise alterations in the communication processes across multiple ROI associated with morphological changes due to disease onset and progression. Furthermore, network analysis of cortex connectivity maps allow for the detection of ROI that serve particular cognitive functions, thus providing a link between brain structure and function. Such links include spatial topographical patterns typically observed between those with and without neurological disease. An early approach to derive a cortical correlation network is the application of Pearson pairwise correlation (PPC) analyses for all possible pairs of ROIs. This approach quantifies the presence or absence of a linear relationship between two sets of observations, and a threshold (tuning parameters) is applied to the correlation values to produce the resultant network. In addition to being threshold dependent, another disadvantage of PPC networks is the reliance on correlations based on independent analysis among two ROIs. While these methods quantify the correlation between region pairs i and j, this correlation measure ignores any relationship region i may simultaneously have with regions other than j, potentially resulting in a loss of information. To overcome this limitation, partial correlation networks, such as the sparse inverse covariance estimation with the graphical least absolute and selection operator or gLASSO have become increasingly popular. The gLASSO approach is particularly useful in situations where the set of observations N is smaller than the set of possible network connections p (N \< p case). However, in order to accommodate for this case, the gLASSO enforces sparsity in the inverse covariance estimate, and the penalised likelihood expression that needs to be optimised is not a consistent estimator. While the gLASSO overcomes some the shortcomings of the PPC, it too relies on a tuning parameter, a sparsity index λ, which is often defined independent of the data and has a large effect on the resultant network. Methods to choose the optimal value of λ have been well- researched. One such method is the stability approach to regularisation selection (StARS) for high dimensional graphical models. However, this approach also relies on pre-defined tuning parameters independent of the data, such as the size and number of sub-matrices to sample which is required by the algorithm. For these reasons, a consistent statistical network approach that is robust to the choice of value for the tuning parameter is needed in order to deduce reliable data driven networks. Furthermore, in an era of neuroimaging “big data”, Smith and colleagues foresee the need to develop novel statistical methods, such as connectivity network estimators, which have desirable theoretical properties such as convergence to the true solution as the sample size increases (N \> p case). This unmet need follows from one of the most successful and largest studies in advancing AD research, the Alzheimer’s Disease Neuroimaging Initiative (ADNI) as well as several other large-scale studies which are in the process of recruiting thousands to hundreds of thousands of participants. ## 1.1 Theoretical background of the MNL algorithm Markov random fields (MRF) are a broad class of neighbourhood based formulations which are often included in neuroimage processing models to account for the spatial variation among voxels or ROI. Conditional autoregressive (CAR) spatial models are a type of MRF which assumes a known and fixed neighbourhood adjacency structure in the form of a binary symmetric matrix W. The covariance structure for the multivariate CAR model is a function of W, and, while the joint distribution of the well-known intrinsic CAR model is improper (the distribution does not integrate to one and the expected value is not defined in closed form), variations of the CAR model yield well defined multivariate distributions. For example, the Leroux et. al. (2000) multivariate adaptation of the CAR model was applied by Anderson et. al. (2016) in the context of aerial disease mapping. However these simple and fixed neighbourhood formulations of W may not adequately capture the complex spatial covariance patterns between regions of the brain. More recently, in the context of estimating neuroimaging covariance, Cespedes et. al. (2017) estimated the matrix W using a Bayesian hierarchical model. However, as this model consisted of 595 parameters, it was found to be too computationally intensive to estimate with Markov chain Monte Carlo methods. It is therefore desirable to develop a method to approximate W using a less computationally intensive approach, particularly for large data sets, as this can be very useful for exploratory purposes. The Leroux et. al. (2000) multivariate adaptation of the CAR model is a joint probability distribution of spatial observations, b<sub>i</sub>, conditional on the adjacency structure W and a spatial scale variance term, $\sigma_{s}^{2}$, which in this work will be referred to as the network likelihood. Maximum likelihood estimation (MLE) is a well-known statistical approach employed in many applications for parameter estimation. One of the advantages of this approach is that it only requires optimisation of the likelihood function conditional on the sample data, which is straightforward to implement in general. Furthermore, MLEs have been shown to be consistent under certain conditions, meaning that as the sample size increases, the MLE will converge with probability one to the true parameter value of the data generating process. In this work, we propose a MLE algorithm to estimate W in the network likelihood, as it represents the underlying covariance connectivity brain structure, while taking into account the variation among all participants. The approach presented will henceforth be termed as maximisation of the network likelihood (MNL). Unlike gLASSO and PPC networks, the MNL returns a single binary connectivity matrix based on a consistent network estimator and is robust to the choice of value for the tuning parameter. This avoids the threshold and sparsity issues discussed earlier and provides a simultaneous analysis on the connectivity of all regions, while providing network estimates whose accuracy increases proportional to the sample size. The layout of this manuscript is as follows. Sections 2.1 and 2.2 presents the case study used in this research. The MNL approach is described in detail in Section 2.3. The utility of this approach is then demonstrated through both a simulation study (Sections 2.4 and 3.1) and an application of cortical thickness covariance networks from structural MRI data (Section 3.3). Two network connectivity matrices are derived for groups of healthy controls (HC) and mild cognitive impaired (MCI), followed by a comprehensive discussion of the comparative merits of the MNL algorithm with the PPC and gLASSO alternatives presented in Sections 3.1 and 3.4. # 2 Materials and methods ## 2.1 Participants of the ADNI study The Alzheimer’s Disease Neuroimaging Initiative (ADNI) is a world wide data sharing collaboration project for AD research. ADNI is a multisite ongoing longitudinal study designed to assist researchers develop clinical, imaging, genetic and biochemical biomarkers for AD research. In this work, we compare cortical connectivity’s of normal healthy ageing (HC) individuals with those who have mild cognitive impairment (MCI). As cognitive impairment precedes dementia onset, individuals with MCI may include prodromal AD participants where cortical atrophy may already be present and/or in its early stages. For the current research, we used the participant’s first visit (baseline) data from 1,383 individuals; 761 male and 622 female. Written and informed consent was obtained from all participants and/or authorised representatives and study partners. All ADNI studies are conducted according to the Good Clinical Practice guidelines, the Declaration of Helsinki, US 21CFR Part 50—Protection of Human Subjects and Part 56—Institutional Review Boards, and pursuant to the state and federal Health Insurance Portability and Accountability Act (HIPAA) regulations. Refer to <http://adni.loni.usc.edu/> for full details of ADNI protocol and ethical requirements for each ADNI study. ## 2.2 Image analysis and data acquisition In this work, we consider structural MRI scans which were undertaken at baseline. The structural MRI T1.5 and T3 weighted images were first segmented into grey/white matter and cerebral spinal fluid using an in-house implementation of the expectation maximisation algorithm applied to a Gaussian mixture model. Cortical thickness was then computed along the grey matter based on the combined Lagrangian-Eulerian partial differential equations approach. The automated anatomical atlas (AAL) was used to parcellate the brain into 116 cortical and sub-cortical regions. In this work, we analysed 34 cortical regions from the left and right hemisphere (K = 68 regions total) for each individual. The remaining 48 sub-cortical regions were excluded as these analyses considers ROI cortex regions measured in mm, and sub-cortical ROIs such as the hippocampus are better represented by their volume rather than thickness. The anatomical regions are listed in. Once parcellated, the mean cortical thickness of the voxels in each ROI was computed and used in this analysis. ## 2.3 Maximisation of the network likelihood The MNL algorithm estimates the connectivity structure via maximising the network likelihood. In this work, the network likelihood is the Leroux et. al. (2000) multivariate CAR model, which is of the following form $$\begin{aligned} \mathbf{b}_{i} & {\sim MVN\left( 0,\sigma_{s}^{2}Q \right)} \\ Q^{- 1} & {= \gamma\left( \hat{W} - W \right) + (1 - \gamma)\mathbb{I},} \\ \end{aligned}$$ where the set of spatial observations for K ROIs on the i<sup>th</sup> participant is b<sub>i</sub> and $\mathbb{I}$ is the K by K identity matrix. The binary elements of the symmetric adjacency matrix W take values w<sub>jk</sub> = 1 to denote a network link, if regions j and k have spatial similarity, or w<sub>jk</sub> = 0 otherwise, which denotes the absence of a link. The diagonal elements are w<sub>jj</sub> = 0 as specified by Lee (2011) and Anderson et. al. (2014). Diagonal matrix $\hat{W}$ has zero off-diagonals, with the j<sup>th</sup> diagonal term equal to j<sup>th</sup> row sum of matrix W. The spatial scale variance is denoted by $\sigma_{s}^{2}$ which controls the amount of spatial variation among the K regions and is multiplied by the spatial covariance matrix Q, which is a function of γ and the adjacency matrix W. The value of γ represents the strength of spatial dependence on b<sub>i</sub> and, in this setting, it is the tuning parameter in the MNL algorithm. Values of γ close to zero imply the set of spatial observations are independent and Q becomes a diagonal matrix. Alternatively, as γ approaches one, it forces Q to be a covariance structure with non-zero off-diagonal terms. This suggests that b<sub>i</sub> has an inherent spatial covariance structure. In practice γ is seldom estimated and remains fixed as it is a difficult (in terms of identifiability) and computationally intensive parameter to estimate. In the context of brain connectivity estimation, γ is often set to 0.9 to enforce a relatively large spatial dependence among the observations. In this work, in addition to the simulation study described in Section 2.4, we also performed a simulation study to assess the ability of the MNL algorithm (with γ fixed at 0.9) to recover the connectivity network on data generated on a range of γ values. We found that the MNL algorithm adequately recovered the simulated connectivity structure from data with various levels of spatial dependence and is hence robust to the value of γ, and supports our choice for fixing γ to 0.9. Refer to Supporting Information for simulation results. ### 2.3.1 MNL algorithm implementation For N total participants, the likelihood function is $$\begin{array}{cl} {p\left( B \middle\| W,\sigma_{s}^{2} \right)} & {= \prod\limits_{i = 1}^{N}p\left( \mathbf{b}_{i} \middle\| W,\sigma_{s}^{2} \right)} \\ & {= \prod\limits_{i = 1}^{N}\left\| 2\pi\sigma_{s}^{2}Q \right\|^{- \frac{1}{2}}\exp\left( - \frac{1}{2\sigma_{s}^{2}}\mathbf{b}_{i}^{T}Q^{- 1}\mathbf{b}_{i} \right).} \\ \end{array}$$ Maximisation of is performed using 15 steps as shown in Algorithm 1. The MNL algorithm provides iterative updates on W\* and $\sigma_{s}^{2}$ M* times. The fast quasi-Newton algorithm implemented to update $\sigma_{s}^{2}$ was adapted from Byrd et. al. (1995). As this is a deterministic algorithm, Step 14 of Algorithm 1 repeats the search for W* for P sets of different starting values to mitigate being stuck in a local minima. Algorithm 1: MNL algorithm Input: Set of spatial observations B, random binary matrix W\* and small positive spatial variance $\sigma_{s}^{2}$ Output:* W and $\sigma_{s}^{2}$ estimates that maximise the network likelihood 1 Evaluate log-likelihood $\delta^{} = \log\left\lbrack p\left( B \middle\| W^{},\sigma_{s}^{2} \right) \right\rbrack$ 2* for M runs do 3 foreach w element of W\* do 4 Permute w<sup>th</sup> element in W\* to get W\\ 5 $\delta^{w} = \log\left\lbrack p\left( B \middle\| W^{*},\sigma_{s}^{2} \right) \right\rbrack$ 6 if δ<sup>w</sup> \> δ\* then 7 δ\* = δ<sup>w</sup> 8 W\* = W\\ 9 end 10 end 11 Update $\sigma_{s}^{2}$ conditional on W\ with a fast quasi-Newton algorithm 12 end 13 Retain final W\* and $\sigma_{s}^{2}$ 14* Repeat Steps 1 to 13 at different random starting values P times 15 Return W\* and $\sigma_{s}^{2}$ estimates corresponding to the highest δ\ ### 2.3.2 Data processing To estimate the model proposed in based on real data, a linear regression was applied to the set of ROI observations y<sub>.k</sub> = \[y<sub>1k</sub>, y<sub>2k</sub>, …, y<sub>ik</sub>, …, y<sub>Ik</sub>\] for each ROI over all I individuals. Covariates in these linear regression models included gender, apolipoprotein (APOE) ε4 carrier and non-carriers status and age in a similar manner as previous studies. The predicted values of the regression model $\left( {\hat{y}}_{ik} \right)$ were obtained, and the residuals for each set of ROIs were computed by $\varepsilon_{ik} = {\hat{y}}_{ik} - y_{ik}$. These residuals were standardised by $b_{ik} = \left( \varepsilon_{ik} - {\overline{\varepsilon}}_{k} \right)/s_{k}$, where ${\overline{\varepsilon}}_{k}$ and s<sub>k</sub> are the empirical mean and standard deviation of the residual for each region over all individuals. The MNL algorithm was applied to the final set of observations (b<sub>ik</sub>). This transformation allows for the residuals of the ROIs to be on the same scale, (as the variance of each ε<sub>.k</sub> is set to one) while maintaining the correlation structure of the data after accounting for covariates. Pairwise plots and histograms of the transformed residual showed linear relationships between certain ROIs and each ROI were approximately Normally distributed and centred at zero (not shown). ## 2.4 Simulation study of MNL algorithm The goal of the simulation study is to assess the ability of the MNL algorithm described in Section 2.3 to recover binary connectivity matrices based on simulated neural data at various sample sizes. A further assessment focused on comparing the results of the MNL algorithm with those obtained using the gLASSO and PPC methods applied to the same simulated data. Our simulation study comprised of two simulated networks, S<sub>1</sub> and S<sub>2</sub> as shown in. We combined a second order diagonal network with a random network model as described in Bien and Tibshirani (2011). As cortical networks, in general, have a diagonal structure, both binary solution networks had second order connections S<sub>i,i−1</sub> = S<sub>i−1,i</sub> and S<sub>i,i−2</sub> = S<sub>i−2,i</sub> = 1 and zero otherwise. A random network model was used to simulate semi-sparse (S<sub>1</sub>) and sparse (S<sub>2</sub>) off-diagonal elements, whereby the remaining off- diagonal elements had a probability of 0.1 and 0.05 respectively, of a connection being present. To convert these binary solution networks into covariance matrices, in a similar manner as Bien and Tibshirani (2011), the diagonals and off-diagonal elements of S<sub>1</sub> and S<sub>2</sub> were multiplied by different positive constants resulting in covariance matrices Ω<sub>1</sub> and Ω<sub>2</sub>. Data were generated from a multivariate normal distribution MVN(0, Ω) for each covariance matrix for sample sizes N = 100, 250, 500, 1000. Ten independent sets of simulated data were drawn for each sample size in order to allow for a rigorous comparison of performance of each method at every sample size. To assess the performance of the MNL algorithm we compared the rate of the true positive connections (sensitivity), which was summarised by the percentage of the connections which were correctly identified to be present. Likewise, the true negative rate (specificity) was summarised by the proportion of absent connections that were correctly identified by the algorithm. As the connectivity matrices are symmetric, we only consider the upper off-diagonal elements of each matrix. A network classifier which has a perfect recovery of the solution network will have both sensitivity and specificity percentages close to one. Alternatively, a poor performing algorithm will have the respective percentages close to zero. Comparison of the performance of the MNL, gLASSO and PPC methods for increasing sample sizes provides insight into the consistency of each approach. A consistent network estimator has a property that as the sample size increases, the estimated network converges to the true solution. Mathematically, we can demonstrate that the parametrisation of $1/\sigma_{s}^{2}Q^{- 1}$, and by extension $\sigma_{s}^{2}Q$, is a positive definite covariance matrix for all values of γ, refer to Supporting Information : Proos MNL is a consistent estimator for a proof of this result. It follows from fundamental theoretical results by Greene 2010 (Chapter 14) and Pourahmadi (2000) among others, that an MLE estimator of a positive definite covariance matrix is a consistent estimator, and will converge to the true solution with probability one as the sample size increases. In addition to assessing the MNL algorithm as a suitable candidate for network estimation, the simulation study also provides information on the performance of each algorithm according to different sample sizes based on two network configurations for a range of tuning parameters (for gLASSO and PPC methods). ## 2.5 Alternative brain connectivity methods As described in Section 1, the PPC and gLASSO are current and popular methods used to derive connectivity networks. In this section we provide a brief description of each approach in relation to the simulation study and its application to the case study. ### 2.5.1 PPC approach The PPC continues to be a popular approach to derive cortical connectivity networks and, for this reason, this approach will also be considered in our simulation study. The correlation between region i and j is denoted by ρ<sub>ij</sub> and −1 ≤ ρ<sub>ij</sub> ≤ 1. All possible sets of pairwise correlations (ρ<sub>ij</sub>) were used to populate the correlation matrix. A binary adjacency matrix A was derived from the correlation matrices, with elements a<sub>ij</sub> equal to zero if \\|ρ<sub>ij</sub>\\| \< τ and value one if \\|ρ<sub>ij</sub>\\| ≥ τ, where the threshold range or tuning parameter is 0 \< τ \< 1, similar to the threshold range described by He et. al. (2008). Fewer spurious correlations are included as τ approaches one, and this may result in disconnected networks determined by the strongest correlations. Alternatively, if τ is too close to zero then the highly connected network may include connections which arose due to spurious noise from the data. In practice, the threshold range in cortical correlation analyses is chosen such that the resultant networks have several organisational features such as small-world topology and the minimum clustering coefficient is above zero in order to make meaningful comparisons between networks. In contrast, the estimated W from the MNL algorithm does not rely on tuning parameters and organisational network features described above can be evaluated directly on this estimated network. In this work, values for τ in {0.1, 0.15, 0.2, 0.25, …, 0.75, 0.8, 0.85} were initially investigated for the PPC networks, and this was fine-tuned for the simulation study. ### 2.5.2 gLASSO algorithm The graphical LASSO is a fast approach to estimate a sparse inverse covariance matrix. For a set of observations b<sub>.1</sub>, b<sub>.2</sub>, …, b<sub>.I</sub> from a multivariate normal distribution b<sub>.i</sub> ∼ MVN(0, Γ) with precision matrix Θ = Γ<sup>−1</sup>, gLASSO aims to find $\hat{\Theta}$ such that $$\begin{array}{r} {\hat{\Theta} = \text{max}\left\{ \text{log}\left( \middle\| \Theta \middle\| \right) - \text{trace}\left( S\Theta \right) - {\lambda \parallel \Theta \parallel}_{1} \right\},} \\ \end{array}$$ where the sample covariance matrix is denoted by S and \\|\\|.\\|\\|<sub>1</sub> is the L<sub>1</sub> norm. The sparsity tuning parameter λ, also known as the penalizing parameter, determines the sparsity of $\hat{\Theta}$. For example, high values of λ implies that \\|\\|Θ\\|\\|<sub>1</sub> has a large contribution to the optimisation problem in. Conversely, when λ = 0, reverts to a simpler MLE problem. In terms of brain connectivity, the gLASSO is applied to estimate Θ, which is then used to derive the binary connectivity matrix. Values of this network matrix are equal to one if the corresponding values of $\hat{\Theta}$ are non- zero, and an absent connection is defined by the zero values of $\hat{\Theta}$. Brain connectivity networks estimated by include the work by Huang et. al. (2010) and Cho et. al. (2017) among others. Authors Huang et. al. (2010) focused on the investigation of network organisation and selected λ such that the networks had a fixed number of links. Alternatively, the StARS approach was used to derive an optimal value for the tuning parameter λ in Cho et. al. (2017) and the effect of λ on the results were not investigated. In addition to the StARS approach, extensive research and development in relation to the optimal λ value has led to several novel approaches, including cross validation among others, refer to Fan and Feng 2009 for a review. In this work, we are interested on the effect λ has on the performance of gLASSO and its ability to correctly identify the solution networks for S<sub>1</sub> and S<sub>2</sub>. Values for λ in {0.1, 0.15, …, 1.7} were explored for gLASSO networks in the simulation study, and a subset of this range was chosen for the real data analyses. Furthermore, we also explored the optimal value of λ which minimises the cross validation error, as this is one of many standard approaches to derive the value of λ (See Supporting Information for results). As the intention of this research is to present the MNL algorithm, further investigation of networks derived by gLASSO with other approaches to derive λ is beyond the scope of this work. ## 2.6 Statistical analysis Via exploratory data analyses, the demographic characteristics were compared between HC and MCI participants over age using an Analysis of Variance, Independent Sample t-test, Chi-squared tests (gender and APOE ε4 allele positive status), as well as Kruskal Wallis test (MMSE and CDR). All statistical analyses were performed using the R statistical environment (R version 3.4.2, R Core Team). On both simulation and real data application, a single application of the MNL algorithm took less than a minute to run on a single central processing unit (CPU) on a standard computer (four core 3.40GHz Intel i7-4770 processor). We expect this to vary for different N and K data sets. Step 5 of Algorithm 1 is executed in C<sup>++</sup> in order to improve the computational time taken to run the MNL algorithm. The remainder of the algorithm is implemented in R. We note that nested for-loops (Steps 2 and 3 of Algorithm 1) act as a bottle neck and future work to profile Algorithm 1 would speed up the implementation of the MNL algorithm. In this work, the MNL algorithm was found to be slightly slower (in terms of seconds) than the gLASSO and PPC. Refer to <https://github.com/MarcelaCespedes/MNL_algorithm> for the coded implementation of the MNL algorithm, full simulation study as well as a tutorial on the implementation of the MNL approach. # 3 Results ## 3.1 Simulation study: Comparison of MNL, gLASSO and PPC algorithms The aim of this simulation study was to evaluate the performance of the MNL algorithm to correctly recover the connectivity matrices for each configuration shown in from the simulated data described in Section 2.4. As the simulation study described in Section 2.3 shows that for a range of γ the performance of the MNL algorithm is relatively robust in terms of the the sensitivity and specificity of the recovered network, in this simulation study we assess the performance of the MNL algorithm against a range of threshold and sparsity values for the PPC and gLASSO. ### 3.1.1 Simulated semi-sparse network S<sub>1</sub> As described in Section 2.4, out of the two simulated matrices considered in this work, S<sub>1</sub> is the semi-sparse diagonal network. Each sample size comprised of ten replicates and Supporting Information shows the covariance plots for randomly selected covariance matrices for each sample size. shows the gLASSO and PPC results and their ability to correctly identify the elements of the S<sub>1</sub> matrix via the sensitivity and specificity for all simulated data over a range of tuning parameters. The sensitivity and specificity of the MNL algorithm are shown by the red and blue horizontal lines respectively, refer to for values. Supporting our theoretical results which show that the MNL is a consistent estimator, our simulation study shows that as the sample size increases, both sensitivity and specificity approach one. It is interesting to note that for all sample sizes the MNL algorithm in general has a specificity close to one, suggesting that the algorithm has a high chance of detecting no link when no such link exists. The simulation study in this work show that the MNL algorithm is better suited for applications where it is desirable to avoid over-interpreting incorrect links. While this trait has it obvious merits, the MNL algorithm may be unsuitable in applications such as gene regulatory networks, where it is desirable to over- select the network connections rather then underestimate them. Both gLASSO and PPC approaches show a trade-off between the ability to correctly detect the presence and absence of links over the range of values of τ and λ. In general, for all sample sizes and small tuning parameters, both algorithms show a sensitivity close to one but a specificity close to zero, suggesting that these algorithms largely overestimated the number of links of the networks. At the other extreme of the tuning parameters, this relationship switches, and resultant networks approach a zero connectivity matrix reflected by specificity close to one and a sensitivity close to zero. Our results illustrate the extent the tuning parameters can have on the gLASSO and PPC, making the correct choice in practice difficult; as the optimal performance of the gLASSO and PPC occurs when the specificity and sensitivity are at their highest and this occurs within a very narrow range of the tuning parameters. This simulation study also serves to show the benefits of an approach that is robust to the choice of value for the tuning parameter, as the MNL algorithm is not affected by such trade-off. In relation to the PPC, the value of τ can range between zero and one, however, a smaller range away from the extremes is often used in practice to avoid the issues described in Section 2.5.1. It is surprising to see in our simulation study that at small values of τ = 0.08, 0.09, 0.1 and 0.11, the PPC offers superior performance compared to the other two alternatives, particularly at a sample size of N = 1000. While it is unlikely that these values of τ are used in practice, we note that the approach to simulate the data favours the PPC approach. As the sample covariance matrices in Supporting Information show a clear difference between covariance values for present and absent links. This difference is mostly emphasised as the sample size increases to N = 1000. Across all sample sizes, the optimal performance of the gLASSO occurs at sparsity values λ = 0.55, 0.6, 0.65 and 0.8 as shown in. Outside of these values, gLASSO shows the trade-off between sensitivity and specificity giving poorer performance. In comparison to the MNL algorithm, particularly at λ values where gLASSO had optimal performance, the MNL algorithm maintained similar or superior specificity and sensitivity across all sample sizes. As the MNL algorithm relies on the evaluation of the full likelihood, its application is not always suitable for small sample sizes (N \< p case). As suggested in literature the gLASSO may be better suited for brain connectivity estimation in smaller clinical studies. Our simulation study support this result in the case of N = 100 and K = 68 ROIs, as the gLASSO results were comparable to those those from the MNL algorithm. We note that unlike the bounded tuning parameters of the PPC, the range λ can take are all positive values, making the choice of λ in practice more difficult to ascertain than the PPC approach. ### 3.1.2 Simulated highly sparse network S<sub>2</sub> The data generating network for S<sub>2</sub> has approximately half the connections than the S<sub>1</sub> network and for this reason it is of interest to see how the MNL and PPC algorithm perform, and in particular how they compare to the gLASSO which is specifically designed to estimate sparse networks. Refer to Supporting Information for covariance plots for data generated by the S<sub>2</sub> binary networks. Simulation results in follow similar trends as those described in Section 3.1.1 for the gLASSO and PPC. The trade-off between sensitivity and specificity across the tuning parameters remains and the PPC demonstrates superior performance over the MNL and gLASSO algorithm at similar τ values as described above, particularly at a sample size of N = 1000. In this scenario, there is a faster improvement of the MNL algorithm as the sample size increases and we believe this is due to the simulated covariance values are larger for S<sub>2</sub> than S<sub>1</sub>, and this is reflected in their respective covariance plots. Refer to for MNL algorithm results. As the connectivity matrix we are interested in recovering is a highly sparse network in comparison to S<sub>1</sub>, it is not surprising to see that the performance of the gLASSO also improves faster as the sample size increases. In this simulated scenario, optimal performance of the gLASSO occurs at λ = 1.55 on a sample size of N = 100 and this improves for an optimal sensitivity and specificity of 0.62 with λ = 1.1 at a sample size of N = 1000. The MNL algorithm shows higher sensitivity and specificity values compared to the gLASSO at sample sizes greater than 250. It is interesting to note, that in this simulation study the optimal values of τ remain, in general, unchanged for the PPC approach, whereas for the gLASSO, there is a large difference in the range of λ for S<sub>1</sub> and S<sub>2</sub> matrices where optimal performance occurs. ## 3.2 Case study: Characteristics of study participants The results of the exploratory data analyses of the demographic features of the participants in the study are shown in. A chi-squared test for independence found a significant association between gender and diagnosis levels (p \< 0.0001), as the MCI group had a considerably higher number of males compared to the HC group. Compared with HC, MCI participants were more likely to have the variant APOE ε4 allele (p \< 0.0001). Cortical thickness measures for all 68 ROI were significantly higher in HC participants (mean 2.71 mm) compared with MCI participants (mean 2.66 mm) (p \< 0.0001). Significant ordinal patterns of degeneration from HC to MCI were observed as follows: cognitive Mini Mental State Examination (MMSE) scores decreased from 29 to 28 (p \< 0.0001); Clinical Dementia Rating (CDR) score values increased from 0 to 1.5 (p \< 0.0001). As each individual has 68 ROI observations, here, the smallest sample size comprises of 35,156 observations which is greater than 2,278 potential links for a 68 × 68 connectivity matrix. ## 3.3 Case study: MNL analysis Prior to applying the MNL algorithm, we observed the histograms of the b<sub>.i</sub> values for each region. These plots confirmed that the transformed data follow an approximate Normal distribution centred at zero (plots not included). Inspection of pairwise plots of the transformed data showed the association between regions displayed various levels of linear relationships, suggesting there is a covariance structure in the data (plots not included). Representative samples from these plots, such as paired regions 23 and 49, 48 and 60, 42 and 52 suggest there is a linear relationship among these regions for all diagnosis groups, refer to for ROI names and Supporting Information for plots. Hence, the covariance structure of the MNL algorithm in show these regions to be connected. Likewise, the absence of a linear relationships was observed in pairwise plots between, for example, regions 21 and 30, 36 and 57, 51 and 68, across both diagnosis groups. The lack of association between these regions is indicated by the corresponding absence of links in the networks of. Furthermore, as a goodness-of-fit assessment of the MNL algorithm, we examined the residuals after we fitted the model to the transformed data from the two diagnosis groups. Histograms and scatter plots of the residuals show they were approximately Normally distributed, refer to Supporting Information. In order to assess if the MNL algorithm adequately modelled the spatial structure of the data, we computed the Moran’s I statistic on the set of residuals from the MNL model fitted to the data for each person within each diagnosis group. Synonymous to Pearson’s correlation, a Moran’s I value close to zero, contingent on spatial structure matrix W, indicates the data have low spatial correlation. The median Moran’s I value for HC and MCI groups were found to be equal to or less than 0.31. Correlation and partial correlation plots of the MNL algorithm residuals in general had values which were substantially small, refer to the Supporting Information for plots. In summary, the selected pairwise, partial correlation plots and Moran’s I values suggest the covariance structure of the data on all diagnosis groups was adequately modelled by expression, and the histograms support the Normality assumption in. The binary matrices from the MNL algorithm applied to the case study data are shown in. These matrices represent the estimated general connectivity structures for the HC and MCI diagnosis groups. The total number of potential connections on a 68 ROI network is 2,278. The total number of links in the diagnosis networks are 180 and 167 for HC and MCI networks, respectively. The networks in shows a large overlap in connectivity between the networks, with 136 connections in common. Most of the connections which are common to both HC and MCI groups include those within each lobe, while most of the differences tend to occur between lobe connectivity. While there was only a subtle reduction in connectivity along the diagnosis spectrum (from HC to MCI), all estimated networks were connected, suggesting that, at some level, all ROIs co-vary with each other in that there were no regions independent from the rest. Our initial investigation was performed on a clinical study with smaller sample sizes (HC: 171, MCI: 46 and AD: 29). However, based on the simulation study results in Section 3.1, it is clear that the performance of the MNL algorithm improves as the sample size increases. Hence in this work we applied our method on the ADNI case study on two large groups (HC and MCI), with expected pathological differences in connectivity. From the networks in, additional network analysis can be applied to the network matrices to determine small-world topology and organisational network features such as characteristic path length and clustering coefficient, however, this is beyond the scope of the present study. The results from simulation studies in Section 3.1 suggest that the obtained networks are relatively reliable, and as the sample size increases, the performance of the MNL algorithm improves in both the ability to correctly identify the presence and absence of connections. ## 3.4 Case study: gLASSO approach Our intention of applying a competing algorithm to the case study is to compare a single binary network between the MNL algorithm to a current known approach. There are several alternatives available for estimating such a network via the gLASSO, and for this reason we applied the gLASSO to the case study data. As we have yet to find methods to choose a single value for τ given the data, in this application we did not apply the PPC to the case study data. shows the connectivity matrices for selected sparsity (λ) values, whose total links were similar to the MNL results. Without knowing the correct value of λ we first applied the glASSO for the range of sparsity values λ = {0.1, 0.15, …, 1} to understand the effect λ had on estimated networks. The resultant networks ranged between 2,278 to 0 in total number of links, refer to Supporting Information for full results. In a similar manner as Huang and colleagues (2010), the sparsity value was chosen such that the resultant networks had a similar number of links to those from the MNL algorithm in Section 3.3. We note that the primary intention of applying the gLASSO to the case study data is to compare the change in connectivity, with less focus on finding the best model fit. shows the resultant networks for HC and MCI groups and the total number of links for each network were 171 and 122 for HC and MCI groups respectively. The networks in show clearer block diagonal matrices, in comparison to the MNL algorithm, suggesting that in this work, gLASSO networks detected higher inter- lobe connections rather than between lobe connections. No connections were detected within or between the limbic lobe. In a similar manner as the MNL algorithm, there was a large overlap between the connectivity within each lobe for the HC and MCI groups, with 66 links in common. The HC network is shown to have no connections between the parietal and occipital lobes, however, in the MCI network there is a large change in connections between the occipital and temporal lobes. We note that in order to appropriately assess the network organisation (by investigating the clustering coefficient, efficiency and small world topology of the networks) and thereby further discuss biological and neurological differences between the MNL and gLASSO networks, a suitable range of λ is required, and this is beyond the scope of this work. # 4 Discussion In this work, we propose a novel approach to estimate brain networks from neuroimaging data. Validated on a numerical simulation study, the sensitivity and specificity performance of the MNL algorithm was shown to improve as the sample size increases supporting our theoretical results that the MNL algorithm is a consistent network estimator. In the simulation study for sample sizes greater than 100, the MNL algorithm was shown to have a higher sensitivity and specificity compared to the results from gLASSO, over a range of sparsity values. At the range of 0.08 ≤ τ ≤ 0.11, the PPC was shown to outperform both MNL and gLASSO algorithms, particularly at a sample size of N = 1000. Application of the MNL algorithm to the ADNI case study identified a loss of connections between HC and MCI connectivity networks, suggesting evidence of atrophy along the neurodegeneration pathway, supporting biologically meaningful results. Our simulation studies found that the PPC and gLASSO analyses were sensitive to the tuning parameters in terms of the ability to recover the solution networks. A trade-off exists between the specificity and sensitivity rates in all sample sizes considered in this work, which showed that as the tuning parameters (threshold τ and sparsity λ) increase the specificity increases, but the sensitivity decreases and vice versa. Application of the MNL approach yields a single connectivity structure that is robust to the value of the tuning parameter (γ) which serves as a descriptive network statistic which is beneficial for exploratory purposes. As such, interpretation of the resultant network is limited to the specific sample used to derive the network. The brain wombling models applied to neuroimaging data by Cespedes et. al. (2017) utilise expression as part of a Bayesian wombling model to estimate the network connectivity and its associated uncertainty. To compare our exploratory approach with those from the Bayesian wombling models, we applied the MNL algorithm to HC, MCI and AD diagnosis groups and 35 ROIs selected in the work by Cespedes et. al. (2017), albeit to baseline data only. We found that the MNL diagnosis networks correctly identified over 83% of the links (correctly detected the presence and absence of connections) obtained in the wombling model (see Supporting Information), suggesting the MNL algorithm can provide results that are comparable to those of Bayesian probabilistic network models. ## 4.1 Extensions Despite the substantive appeal of the MNL algorithm described in this paper, there are several extensions that could be considered. Firstly, the current mean of the multivariate network distribution is zero and as such the MNL algorithm does not provide ROI mean estimates. Extending the MNL algorithm to include a non-zero region mean vector *μ* may be informative as not all ROIs have the same mean. In this work we compensated for this by applying linear regression models to each ROI and transforming the residuals such that they have a mean of zero (Section 2.3.2). We note that the added complexity of the proposed extensions to the MNL approach may result in a more difficult optimisation problem and may require more sophisticated numerical optimisation methods to estimate the additional parameters. Secondly, analyses on longitudinal neuroimaging studies are favoured in contrast with cross sectional analyses, as they could potentially include information on ROI changes over time. While the MNL algorithm presented in this work does not account for repeated measures, an extension of expression to account for repeated measures can be achieved by adding a random effects layer in the model. However, as the MNL algorithm is the first brain network algorithm of its kind whose connectivity estimates improves as the sample size increases, such an extension is left as future work. ## 4.2 Conclusion The potential application of MNL networks is not restricted to cortical thickness structural MRI data, and can easily be applied to any complete spatial set of observations from any neuroimaging modality. The objective for the methodology and application presented here is to introduce and demonstrate the utility of the MNL algorithm, as the application of MNL method can be applied to functional MRI, positron emission topography (PET) and electroencephalography (EEG) data. Other than the suggestions already discussed, an additional area for future work is the application of the MNL algorithm to assess for network robustness as described in Bernhardt et. al. (2011) and Hart et. al. (2016). Here, the authors investigate the loss of random or targeted nodes or edges removed from the network, representing deterioration due to pathology. Furthermore, additional validation of the MNL algorithm on other neurological applications such as epilepsy and schizophrenia, as well as healthy ageing studies over a wide age range, and analyses of network topological metrics are needed to better understand the performance and biological insight from the proposed MNL algorithm. # Supporting information We thank Dr Pierrick Bourgeat for processing the magnetic resonance images. We also greatly appreciate the efforts from three anonymous reviewers and Handling Editor Dr Hayasaka, whose comments and feedback has greatly improved this work. The authors wish to thank all clinicians, scientists, participants and the families involved in the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Data used in preparation of this article were obtained from the ADNI database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in the analyses or writing of this work. A complete listing of ADNI investigators can be found at: <http://adni.loni.usc.edu/wp- content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf>. Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defence award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health ([www.fnih.org](http://www.fnih.org)). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. MIC was funded by the Research Training Program (RTP) doctoral scholarship provided by the Australian Government and a top-up provided by the Commonwealth Scientific and Industrial Research Organisation (CSIRO). JDD is the Biostatistics Team Leader of the Australian e-Health Research Centre and is funded by the CSIRO. JF is the Acting Group leader in the Biomedical Informatics team at the Australian e-Health Research Centre and is funded by the CSIRO. JM is funded by the Queensland University of Technology (QUT) and is affiliated with the Australian Research Council and Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS). KM is jointly funded by QUT and the Australian Research Council (ARC) Laureate, project ID FL150100150 Bayesian Learning for Decision Making in the Big Data Era and is the Deputy Director of ACEMS. CCD is affiliated with QUT and ACEMS and was supported by an Australian Research Councils Discovery Early Career Researcher Award scheme DE160100741. The funders provided support in the form of salaries for the authors (MIC, JM, CCD, KM, JDD and JF), but did not have any additional role in the study design, decision to publish or preparation of the manuscript. The specific role of these authors are articulated in the author contributions section. [^1]: MIC was funded by the Research Training Program (RTP) doctoral scholarship provided by the Australian Government and a top-up provided by the Commonwealth Scientific and Industrial Research Organisation (CSIRO). JDD is the Biostatistics Team Leader of the Australian e-Health Research Centre and is funded by the CSIRO. JF is the Acting Group leader in the Biomedical Informatics team at the Australian e-Health Research Centre and is funded by the CSIRO. JM is funded by the Queensland University of Technology (QUT) and is affiliated with the Australian Research Council and Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS). KM is jointly funded by QUT and the Australian Research Council (ARC) Laureate, project ID FL150100150 Bayesian Learning for Decision Making in the Big Data Era and is the Deputy Director of ACEMS. CCD is affiliated with QUT and ACEMS and was supported by an Australian Research Councils Discovery Early Career Researcher Award scheme DE160100741. The funders provided support in the form of salaries for the authors (MIC, JM, CCD, KM, JDD and JF), but did not have any additional role in the study design, decision to publish or preparation of the manuscript. There are no patents, products in development or marketed products to declare. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: Current address: CSIRO Health and Biosecurity/Australian e-Health Research Centre Level 5, UQ Health Sciences Building, 901/16 Royal Brisbane and Women’s Hospital, Herston, Queensland 4029, Australia [^3]: ¶ Membership of the Alzheimer’s Disease Neuroimaging Initiative is provided in the Acknowledgments.
# Introduction Transforming growth factor beta (TGF-ß) signalling is involved in a wide range of processes during development e.g. cell adhesion, bone morphogenesis and cell motility. Upon binding of a ligand of the TGF-ß superfamily to a Type II receptor, the Type II receptor recruits a Type I receptor and activates the Type I receptor by phosphorylation. Then, the Type I receptor phosphorylates receptor regulated smads (R-Smads), which then can bind to the common mediator Smad SMAD4 and translocate as a R-Smad/Smad4 complex into the nucleus, where these complexes interact with co-activators to activate gene expression. Negative regulators of the TGF-ß signalling pathway are inhibitory Smads (I-Smads), Smurfs and the Ski/Sno protein family. Proteins of the latter group possess two structural domains: the Ski/Sno homology domain and the SMAD4-binding domain. With the help of these domains, Ski/Sno proteins can interact, among others, with R-Smads, N-CoR, Sin3a, SMAD4 and the histone deacetylase HDAC1 and this complex leads to transcriptional repression of target genes. By their expression domains, Ski/Sno proteins can be further subdivided into ubiquitously expressed genes (human Ski and Sno), and mainly neuronally expressed genes, the SKI Family Transcriptional Corepressors (Skor1 and Skor2). The Ski/Sno proteins fulfil a wide range of different physiological functions such as axonal morphogenesis, Purkinje cell development, myogenesis and mammary gland alveogenesis. However, the Ski/Sno proteins were not discovered by their physiological functions but via the transforming capability of the viral ski (v-ski) homologue found in the Sloan-Kettering virus. The first evidence that Ski/Sno proteins possess oncogenic capabilities came from overexpression experiments, where it was shown that not the truncation of v-ski is responsible for the transformation of chicken embryo fibroblasts, but that overexpression of v-ski, Ski or Sno is sufficient for this transformation. Despite this background, their role in carcinogenesis is still not fully understood, if not even contradictory at times. Ski and Sno have been found to be upregulated in different types of cancer e.g. oesophagus squamous cell carcinoma, melanoma, and colorectal cancer. Further evidence for a pro-oncogenic role was found in downregulation analyses of Sno or Ski. This downregulation resulted in decreased tumour growth in breast cancer cells and pancreatic cancer cells. But as stated before, there is some objection that Ski and Sno function purely as oncogenes. Mice, which were heterozygous mutant for Ski or Sno, showed an increased level of tumour induction after carcinogen treatment. In metastatic non-small cell lung cancer, Ski expression is significantly reduced, whereas increased expression of Ski in these cells reduced the invasiveness inhibiting epithelial-mesenchymal transition. Therefore, this could reflect that the outcome of Ski or Sno expression in cancer cells is dependent on the cell type or the actual status of the cancer cells and cancer cells often exploit Ski or Sno to inhibit the anti- proliferative effects of TGF-ß signalling. Whereas Ski or Sno have been found to be involved in a lot of different cancer types, there is sparse evidence for deregulation of Skor proteins in cancer cells. Endogenously, Skor proteins have been linked to neurodevelopmental processes. After Skor1 overexpression, genes involved in axonal guidance or post-synapse assembly were differentially expressed. Skor2 is important for cerebellar Purkinje cell differentiation as in Skor2 knockout mice dendrite formation of Purkinje cells was impaired. Pathophysiologically, Skor1 has mainly been linked to restless leg syndrome and localized scleroderma. In Drosophila melanogaster, only one homologue of Ski and Sno, which is designated Snoo, and one homologue of Skor1 and Skor2, which is designated Fuss, exist. We have recently shown that Fuss is interacting with SMAD4 and HDAC1. In overexpression assays, Fuss can inhibit Dpp signalling and endogenously, the Fuss/HDAC1 complex is required for bitter gustatory neuron differentiation and fuss mutant flies pause more often during walking. However, we were interested if the Skor/Fuss proteins retained their ability to inhibit differentiation and induce increased proliferation. For this purpose, we overexpressed Fuss in differentiating cells of the eye imaginal disc, an excellent model tissue to study regulatory gene function in the context of carcinogenesis. This overexpression impaired photoreceptor axon guidance and inhibited the differentiation of accessory cells such as cone cells and primary pigment cells, which are all transformed into a basal pigment cell type. In a second approach we generated fuss overexpressing clones early during development in the eye imaginal discs, when cells are still proliferating. This resulted in vast outgrowths of undifferentiated tissue of the eye imaginal disc because fuss overexpression most likely inhibited Dpp-signalling, a member of the TGF-ß superfamily. Our work shows that Fuss retained the ability of Ski/Sno proteins to inhibit the antiproliferative effects of TGF-ß signalling by analogous inhibition of Dpp-signalling, allowing proliferation to be sustained. # Results ## fuss overexpression leads to a smooth eye surface and impairs axonal pathfinding As Ski/Sno proteins have been found to be involved in cancer development and progression, we were interested if fuss overexpression can also inhibit differentiation and induce increased proliferation and hyperplasia, respectively. To answer this question, we chose the eye imaginal disc as a model tissue, because fuss is not endogenously expressed in the eye imaginal disc. Furthermore, Fuss and its homologues are negative regulators of BMP/Dpp signalling in overexpression assays and consequently the eye imaginal disc enables us to investigate fuss overexpression in a Dpp independent and dependent context. First, we overexpressed fuss via the GMR-GAL4 driver line. GMR-GAL4 is active posterior to the morphogenetic furrow, which is the source of the Dpp morphogen during eye development and thus Fuss cannot directly interfere with Dpp signalling. Interestingly, in adult Drosophila flies, where fuss was overexpressed with GMR-GAL4 during eye development, this overexpression leads to massive differentiation defects exhibiting a smooth, red coloured eye surface devoid of any typical eye structures such as ommatidia or bristles and we observed little to no phenotypic variability in these flies. To see if photoreceptor induction is also affected by fuss overexpression, eye imaginal discs of late third instar larvae were stained with antibodies against Elav, a marker for neurons and Chaoptin, a marker for photoreceptors. In eye imaginal discs overexpressing fuss, neither loss of Elav nor Chaoptin was detected (arrow). This shows that cells overexpressing fuss still acquired a neuronal- as well as a photoreceptor fate. In the central nervous system (CNS), fuss is expressed in postmitotic interneurons during development, and its expression is maintained in adulthood. Therefore, we focused in more detail on photoreceptor development. After acquiring photoreceptor fate, these cells start to protrude their axons into the larval optic lobe. Photoreceptors R1-R6 target the lamina neuropil (arrow), whereas R7 and R8 axons migrate deeper into the medulla neuropil (arrowhead). This leads to a very specific pattern, which can be observed in larval optic lobes of control larvae, but upon fuss overexpression in developing photoreceptors, this pattern is strongly disturbed (arrowhead). ## RNAseq reveals downregulation of the PAX2 homolog Shaven after fuss overexpression To achieve a better overview of genes and processes which become dysregulated by fuss overexpression in the larval eye imaginal disc, we extracted RNA from eye discs of late third instar larvae of controls (GMR-Gal4 \> w1118) and experimental flies overexpressing fuss (GMR-Gal4 \> UAS-fuss), reverse transcribed it and prepared NGS libraries, which then were sequenced. Comparing gene expression profiles of both genotypes, we found that Fuss was highly enriched in GMR-GAL4; UAS-fuss eye discs, which was a proof of principle that the experiment was successful. Among 360 genes which showed an altered expression in contrast to controls with an adjusted p-value \< 0.01, one gene, shaven (sv), was especially interesting. Sv, the Pax2 homologue in Drosophila melanogaster, is needed for the proper differentiation of cone, primary pigment and bristle cells and the loss of sv can result in the well-known glazed eye phenotype. This phenotype is strongly reminiscent of the phenotype observed of fuss overexpression with GMR-GAL4. Therefore, eye imaginal discs were stained with antibodies against Sv and its expression was indeed strongly reduced compared to controls. Another protein, which is crucial for cone cell differentiation, is Cut (Ct), a homeobox containing transcription factor. Ct expression is induced by the transcription factor Sv and therefore also ct expression is severely reduced upon fuss overexpression. As mentioned before, Fuss can also inhibit Dpp signalling, although we found no altered expression of TGF-ß pathway components in our RNAseq data. In addition, we knocked down Medea and the type I receptor Tkv with the GMR-Gal4 driver line and did not find any defects in adult eyes compared to controls. Besides that, we also knocked down sv and indeed, the phenotype is very similar to that of the fuss overexpression. The eyes have a strong reduction in the number of lenses and show a smoother surface than controls. The phenotype of the sv knockdown is not as strong as the fuss overexpression, the reason for this might be, that the downregulation of sv in the knockdown is not as strong as after fuss overexpression or that additional processes are disrupted as there are 360 genes significantly altered in expression after fuss overexpression. ## Fuss overexpression results in loss of cell types and increased apoptosis To understand the developmental defects behind the smooth eye phenotype, pupal development of the overexpression eyes was analysed. In control eye imaginal discs of pupae (GMR-Gal4; UAS-nGFP), which have pupariated for 40 h, a highly ordered pattern of different cell types can already be observed. In the middle of an ommatidium, four cone cells, which secrete the lens, are surround by two primary pigment cells. The single ommatidia are separated from each other by secondary and tertiary pigment cells as well as bristle cells. In pupal eye imaginal discs, where fuss was overexpressed via the GMR-GAL4 driver line most distinct cell types are lost on the surface of the pupal retina in contrast to controls, except for some single bristle cells which are visible (arrows) but most of the cells which are present are of indefinable cell fate. However, these cells might develop to pigment cells, because the adults develop flat, structureless, but red eyes. With a fluorescent apoptosis sensor called GC3Ai, the fuss overexpressing retinas show an increase in apoptotic events in a deeper layer of the developing retina which could also account for the reduced size of the adult eye field if compared to the sv knockdown. As observed in larvae, the photoreceptor axons show a perturbed arrangement compared to controls, but are still expressing the neuronal marker Elav, although the positioning of photoreceptor nuclei is deranged. Additionally, large gaps can be observed between individual photoreceptor cell groups, which could be the result of apoptosis.. The final adult eye differentiation pattern was analysed by paraffin sections of heads of GMR-Gal4; UAS-fuss flies as well as controls (GMR-Gal4; UAS-nGFP). In fuss overexpressing eyes the photoreceptors and their rhabdomeres, which can be observed by a bright fluorescent signal in the controls are completely lost. Although they were determined to become photoreceptors as observed in, they are probably removed by apoptosis during pupal development, which also affects the integrity of the adult lamina. Furthermore, vacuoles can be observed (arrows) which are probably also the outcome of the cell death observed during pupal stage. These sections show, that the indefinable cells observed in pupal retinas have a completely different shape than those of secondary or tertiary pigment cells in controls, but still contain pigment granules. Thus, these cells rather adopt a basal pigment cell fate but don´t acquire the correct shape. Obviously, overexpressing fuss in differentiating cells posterior to the morphogenetic furrow highly impacts their specification. Photoreceptors are determined but are abolished during development and residual cells are prevented to adopt their natural fate and nearly all are transformed to a basal pigment cell fate. Therefore, fuss overexpression is able to completely inhibit differentiation but, in this context, no striking increase in proliferation was observed. ## Early induction of fuss overexpressing clones inhibits eye differentiation and leads to eye disc outgrowths The ability of Ski/Sno proteins to function as proto-oncogenes is often linked to their capacity to inhibit the antiproliferative effects of TGF-ß signalling. To check for early proliferative defects, we decided to induce fuss overexpression clones in first instar larvae. 48 hours after egg laying, we heat-shocked the first instar larvae with the genotype (P{hsp70-flp}1/+; Fuss- GFP/+; GAL4-Act5C(FRT.CD2).P) for 12 minutes to induce fuss overexpressing clones. Late third instar larvae were dissected, and tissue was analysed for fuss overexpressing clones and their effects on development. Surprisingly, as shown in, we detected a completely different behaviour of fuss overexpression. fuss overexpressing clones in eye imaginal discs result in tissue outgrowths. To check the cellular identity of these outgrowing cells, we first tested for the neuronal differentiation marker Elav, which would be expected in photoreceptor cells. However, these outgrowths lack the late differentiation marker Elav completely and can form big bulbous structures which protrude from the eye imaginal disc. All heat shocked larvae die at the latest during pupal stages, therefore we decreased the time of the heat shock, which enabled us to obtain single adult survivors. Some of these survivors had undifferentiated tissue growing out from their eye, which apparently was generated by fuss overexpressing clones. To further understand the fuss overexpression defects, we were searching for early retinal differentiation markers. Dpp signalling is assumed to be required for the expression of early retinal differentiation markers sine oculis, dachshund and eyes absent (eya). Previous studies have shown, that overexpression of fuss can interfere with Dpp signalling and we could indeed confirm that fuss overexpressing clones lack eya expression and therefore, Fuss might inhibit Dpp signaling in the eye imaginal disc as well. All in all, initiation of photoreceptor differentiation is already impaired in fuss overexpressing clones. ## fuss overexpressing clones exhibit increased proliferation However, this does not explain, why fuss overexpressing clones lead to tissue outgrowths from the eye. Generally, if eye development is inhibited, this would lead to a transformation from eye to head capsule tissue only and not to additional head capsule tissue as observed. In contrast to eya mutant eye discs, where proliferation is strongly reduced, the fuss overexpressing clones seem to show increased proliferation. To further analyse cell proliferation, we used the Fly-FUCCI system, where degrons of E2F1 and CycB have been fused to GFP and mRFP, respectively, to visualize the cell cycle behaviour of cells. Cells from anaphase to the G1 to S transition are green, S phase cells are red, and cells in G2 and early mitosis are yellow. Expressing the Fly-FUCCI system via the Flipase technique in eye imaginal discs, we found that many cells posterior to the morphogenetic furrow are green or red, therefore in G1 or S-Phase and only some where yellow, thus in early mitosis. But when we coexpressed fuss with the Fly-FUCCI system many more cells where yellow posterior to the morphogenetic furrow and consequently where in G2-Phase or undergoing mitosis. Interestingly, the shape of fuss overexpressing clones were highly different to that of control clones. Whereas control clones integrate into the patterning of the developing eye imaginal discs, fuss overexpressing clones might not react to Dpp signalling as shown before, have a rather elliptical shape and are protruding from the eye imaginal disc. Because of this experiment we expected, that fuss overexpressing clones have a higher division rate than controls. We chose to use an antibody against phosphorylated Histone H3, a marker for mitosis, to specifically identify mitotic cells and count them. In this assay we only counted cells, behind the second mitotic wave (smw), because after the smw cells start to differentiate instead to proliferate. As controls we used tissue which was not overexpressing fuss, therefore we counted the number of pHH3 positive cells inside fuss overexpressing clones and outside of it, and measured the area in pixels of the fuss overexpressing clonal tissue and the wildtype tissue and divided the number of mitotic cells by the amount of pixels. We found significantly more pHH3 positive cells per pixel in the fuss overexpressing clones than in the surrounding wild type tissue, demonstrating that fuss overexpression leads to increased proliferation in developing eye imaginal discs in contrast to controls. ## wingless is expressed in fuss clonal outgrowths Besides eye differentiation, Dpp signalling is also required to inhibit wg expression in the eye imaginal discs. In eye imaginal discs of third instar larvae, wg expression can normally only be observed at the dorsal and ventral margins of the eye disc, where it is supposed to promote head capsule structures instead of eye tissue, but in earlier stages wg is expressed in the whole prospective eye and this is around the time we induce the fuss overexpressing clones. Loss of Dpp signalling leads to overgrowth and ectopic wg expression in the eye imaginal disc and loss of eya expression might not only be the result of fuss dependent inhibition of Dpp signalling, but also ongoing Wg signalling, as it has been shown by ectopic Wingless signalling clones. To test if fuss overexpression clones continue expressing wg from earlier stages on via a possible inhibition of Dpp signalling, we used a wg-LacZ reporter construct to visualise wg promotor activity in the eye imaginal disc and as shown in fuss overexpressing clones exhibit indeed LacZ expression supporting the hypothesis that the inhibition of Dpp signalling allows continuous wg expression in these cells. Wingless signalling is also involved in promoting the proliferation of cells anterior to the morphogenetic furrow and ectopic Wingless signalling leads to increased proliferation. One possible hypothesis for the overproliferation observed with fuss overexpressing clones might be that Fuss acts again as an inhibitor of Dpp signalling in the eye imaginal disc allowing continuous wg expression, which might lead to the excess proliferation in the eye imaginal disc. To test this hypothesis, we tried to generate clones, which besides fuss, also express a knockdown construct for wg, but this approach turned out to be highly lethal. In a second approach, we generated clones which expressed fuss and a dominant negative form of Pangolin (dnPan), because the effects of wg overexpression can be suppressed by dominant negative Pangolin, even when expressed from the same cell and it has been shown, that wg expression can be autoregulated endogenously or ectopically in some tissues. fuss/dnPan overexpressing clones were not only consistently smaller compared to fuss overexpressing clones, but the clonal tissue did also not outgrow anymore. fuss/dnPan overexpressing clones did not exhibit wg expression anymore, as observed with the absence of LacZ staining, supporting our hypothesis that Fuss, via inhibition of Dpp signaling, might be able to allow wg expression from earlier stages to continue and wg expression and Wg signalling in these clones might promote the outgrowths from the eye disc. Similarly, overexpressing dominant negative Pan together with nuclear GFP lead to small clones and wg expression was also not increased. # Discussion In this work, we addressed the question if Skor/Fuss proteins, members of the Ski/Sno family, retained the function of Ski and Sno to induce uncontrolled proliferation as observed in early stages of tumorigenesis. First, the overexpression of fuss posterior to the morphogenetic furrow with the GMR-Gal4 driver line resulted in a nearly complete loss of all cell types in the adult eye. During development, photoreceptor axons were not able to target the appropriate layers of the optic lobe anymore and cone cells, primary pigment cells and bristle cells were transformed into a basal pigment cell fate. This transformation was caused by the inhibition of sv expression, which is crucial for accessory cell differentiation. Additionally, increased apoptosis during pupal development lead to the removal of photoreceptors and lastly adult eyes only consisted of cells containing pigment granules. This lack of differentiation cannot be explained by the Dpp inhibiting role Fuss exerts, when overexpressed, because inhibiting the Dpp signaling pathway via knockdown of Tkv or Med had no effect. Photoreceptor axon guidance is impaired, if Dpp signaling is disrupted in photoreceptors by the expression of the inhibitory Smad Dad. Thus, the observed photoreceptor axon guidance phenotype, when fuss is overexpressed with GMR, could indeed be a result of Dpp signaling inhibition. However, the loss of nearly all eye cell types is due to other effects (e.g. downregulation of sv and apoptosis) than Dpp signaling repression alone, because loss of Dpp signaling behind the morphogenetic furrow only results in mild patterning defects of the pupal retina. Nonetheless, the inhibition of cell differentiation has already been shown in other cancer models e.g., when two copies of the constitutive active form of the receptor tyrosine kinase dRET<sup>MEN2B</sup> are expressed with the GMR-Gal4 line, pupal retinas are devoid of any distinguishable cell types. This phenotype is indistinguishable from the phenotype of the pupal retinas generated by the overexpression of fuss via GMR-Gal4. In a screen for novel oncogenes from breast cancer patients, human transgenes have been overexpressed with the GMR-Gal4 driver line. Overexpression of human RPS12, a subunit of the small ribosomal subunit, whose expression is increased in various cancer types, leads also to a glazed eye phenotype. Therefore, different oncogenes can result in different outcomes when expressed with the GMR-Gal4 driver line and are not always leading to massive tissue overgrowth like the Yorkie overexpression. Most importantly, with this approach to overexpress fuss in cells which already were destined for acquiring a cell fate and have left the cell cycle, we were not able to induce increased proliferation anymore, but could prevent cell differentiation. Consequently, we switched to a more pluripotent cell type in the eye imaginal disc and induced fuss overexpressing clones prior to the formation of the morphogenetic furrow. These results let us assume that in this context, fuss overexpressing clones do not react to the antiproliferative effects of the Dpp morphogen anymore. Instead, wg expression and thus, proliferation promotion might be maintained. This leads to outgrowths of clonal tissue from the eye imaginal disc of third instar larvae, which showed an increased number of mitotic events. If these flies survived to adulthood, undifferentiated, extra tissue was visible in the complex eye. An analogous mechanism can be observed in tumors which overexpress Ski or Sno. The TGF-ß signaling pathway also acts anti-proliferative, but this action is inhibited by the increased presence of Ski/Sno proteins. Therefore, the molecular mode of action is similar to the human Ski/Sno proteins. The function of Ski and Sno is highly context dependent, as they can fulfill an anti- oncogenic or pro-oncogenic role depending on the cancer type or status of the cancer. We also observed this with fuss overexpressing clones. Only when induced 48h after egg laying, we found additional tissue in late third instar larvae and only in eye imaginal discs, because here, Dpp counteracts the proliferative effects of Wg signaling. When fuss is overexpressed in the wing disc or after induction of the morphogenetic furrow differentiation is inhibited, this results in a wing with truncated veins or in a smooth eye surface (this work). This is also underlined by RNAseq data from eye and wing imaginal discs, where fuss was overexpressed with the GMR-Gal4 and Nub-Gal4 driver line, respectively. In the eye dataset, wg expression in eye imaginal discs is not significantly different from control eye discs, whereas wg expression in fuss overexpression wing discs is significantly reduced in contrast to control wing discs. Thus, we were able to show that the Skor protein Fuss in Drosophila melanogaster still retained the function of the Ski/Sno proteins by inhibiting differentiation but inducing hyperproliferation. But the hallmarks of real tumorigenesis are lacking, because at some point during pupal development, proliferation stops, and these cells become protruding head tissue as it could be observed in complex eyes of surviving flies. Furthermore, there was no evidence of an epithelial-mesenchymal transition because fuss overexpressing clones maintained their epithelial fate. It will be of high interest if future studies can find similar results in overexpression studies for the vertebrate Skor proteins or detect increased expression of these proteins in specific cancer types. # Material and methods ## Fly husbandry and stocks Flies were raised at 25°C under a 12 hr/12 hr light/dark cycle. Fly lines obtained from the Bloomington Drosophila Stock Center were: P{GAL4-ninaE.GMR}12 (BDSC \#1104), w1118; P{UAS-Stinger}2 (BDSC \#84277), w1118; snaSco/CyO, P{en1}wgen11 (BDSC \#1672), y1 w1118; P{UAS-pan.dTCFΔN}4 (BDSC \#4784), y w\[\\]; P{w\[+mC\] = GAL4-Act5C(FRT.CD2).P}S (BDSC \#4780), P{ry\[+t7.2\] = hsFLP}1, y w\[1118\]; Dr\[Mio\]/TM3, ry\[\\] Sb (BDSC \#7), P{AyGAL4}25 (BDSC \#4412), w1118 (BDSC \#3605), y1 v1; P{TRiP.HMS05834}attP2 (BDSC \#67973), y1 sc\* v1 sev21; P{TRiP.GL01313}attP40 (BDSC \#43961), y1 sc\* v1 sev21; P{TRiP.HMS04501}attP40 (BDSC \#57303), w\[1118\]; Kr\[If-1\]/CyO, P{ry\[+t7.2\] = en1}wg\[en11\]; P{w\[+mC\] = UAS-GFP.E2f1.1–230}26 P{w\[+mC\] = UAS- mRFP1.NLS.CycB.1-266}17/TM6B (BDSC \#55122), Tb, w\; KrIf-1/CyO; P{UAS- GC3Ai}3 (BDSC \#84343). Additionally, Nub-Gal4 (J.F. de Celis, Madrid) was employed. ## Immunohistochemistry For analysis of fuss* overexpressing clones 48 hrs after egg laying, the larvae were heat shocked at 37°C for 12 min. Then, for all experiments late third- instar larvae were used for dissection and immunohistochemistry. By pulling the mouth hooks, the anterior mouth part including the eye imaginal discs still attached to the brain were removed from the rest of the larva and then fixed by incubation in 4% PFA in PBS for 20 min. The specimen was washed three times with PBST (PBS with 0.1% Triton-X) for 20 min and incubated in PBST supplemented with 5% normal goat serum and primary antibodies over night at 4°C. The specimen was washed three times with PBST for 20 min and incubated in PBST supplemented with 5% normal goat serum and secondary antibodies over night at 4°C. The specimen was washed once with PBST for 20 min, then incubated in PBST supplemented with 1 mg/ml 4′,6-Diamidin-2-phenylindol (DAPI) for 20 min and washed three times with PBST for 20 min. The eye imaginal discs were dissected and mounted using VECTASHIELD Antifade Mounting Medium (Vector Laboratories). Developmental studies Hybridoma Bank (DSHB) antibodies were: LacZ (JIE7, 1:50), Eyes absent (eya10H6, 1:50), Elav (Rat-Elav-7E8A10, 1:50), Cut (2B10, 1:20), Chaoptin (24B10, 1:50). Additional antibodies were: Sv/Pax2 (1:100, gift from Markus Noll), GFP (rabbit 1:1000, ThermoFisher), pHH3 (rabbit 1:2000, Cell signaling technology). Secondary antibodies were used with a dilution 1:200 overnight at 4°C. Secondary antibodies were goat anti-mouse, anti-rabbit, anti-rat and anti- guinea pig Alexa Fluor 488, 555 and 594 (ThermoFisher). For anti-Cut and anti-Sv stainings we used the Anti-Mouse / anti-Rabbit HRP-DAB IHC kit (abcam) to increase sensitivity and reduce background. ## Generation of FLP-out clones In general, we crossed virgins carrying the P{hsp70-flp}1 allele homozygously to males carrying the GAL4-Act5C(FRT.CD2).P allele. Before the cross was placed for 2h on standard food supplemented with dry yeast at 25°C, flies were allowed to mate for at least three days. The adult flies were removed from the vial and progeny was allowed to develop for 46h at 25°C. Progeny was heatshocked for 12 minutes at 37°C and placed again at 25°C. Late third instar larvae were then dissected. ## Retina dissection and immunostaining White pupae were collected and aged at 25°C for 40 hrs. The brains with the attached eye discs were dissected in PBS and placed in PBS with 4% PFA on ice until all brains from one genotype were dissected. Afterwards the brains were fixed for another 20 min with 4% PFA in PBS at room temperature. The brains with the attached eye discs were stained with rat-anti-DE-cadherin (DCAD2, 1:50, DSHB) or rat-anti-ELAV (Rat-Elav-7E8A10, 1:50) and goat-anti-rat Alexa Fluor 555 (1:200, ThermoFisher) in PBST 0.1% with 5% normal goat serum (NGS). After staining, the eye discs were removed from the brains directly on the mounting slide in a drop of PBST 0.1% and mounted using VECTASHIELD Antifade Mounting Medium (Vector Laboratories). ## Quantification of pHH3 positive cells Only eye imaginal discs where the second mitotic wave was clearly detectable via pHH3 staining were used. Mitotic cells in fuss overexpressing clones and in wildtype tissue were counted. The area of fuss overexpressing clones and wildtype tissue was measured with the measurement tool of ImageJ. The number of mitotic cells inside a fuss overexpressing clone was divided by its area and the number of mitotic cells inside the wildtype tissue was divided by the wildtype tissue´s area. The acquired data was visualized with Python and the Matplotlib and Seaborn libraries. Statistics were calculated with the SciPy library. ## Paraffin sections Paraffin sections were performed from two-day old adult flies. Flies were fixed with carnoy (ethanol:chloroform:acetic acid at a proportion 6:3:1), dehydrated in ethanol, and embedded in paraffin. Paraffin sections (7 μm) from 10 flies of each genotype were analysed under a fluorescence microscope. ## RNA extraction, library generation and sequencing Per replicate and genotype 40 eye antennal discs or 30 wing discs from third instar larvae were dissected. RNA was extracted via peqGold MicroSpin Total RNA Kit. Library preparation and RNA-Seq were carried out according to the NEBNext Ultra RNA Library Prep protocol, the Illumina HiSeq 1000 System User Guide, and the KAPA Library Quantification Kit—Illumina/ABI Prism User Guide. Library preparation and RNA-Seq were performed at the Genomics Core Facility “KFB—Center of Excellence for Fluorescent Bioanalytics” (University of Regensburg, Regensburg, Germany). ## RNA-Seq analysis The reads were quantified with the R package Salmon using the release of the Drosophila melanogaster genome BDGP6.22. The data was imported using tximeta and analysed with DESeq2. Cut-off for significantly dysregulated genes was set with an adjusted p-value \< 0.01. Top 21 differentially expressed genes between control and overexpression replicates were visualized with the heatmap.2 package in R. # Supporting information [^1]: The authors have declared that no competing interests exist.
# Introduction There are few phytobacteria with the capacity to directly affect the health of humans or livestock. In the rare instances where they can, the pathogenic effects are often related to the production of toxins. One such toxin-producer is the Gram-positive bacterium Rathayibacter toxicus, the causative agent of annual rye grass toxicity (ARGT) in Australia. ARGT is an often-fatal toxicosis of forage animals caused by ingestion of infected hay or grain. Over 10 million hectares of Western Australian farmland has been affected and ARGT caused an estimated \$40 million AUD in direct losses in 2010. R. toxicus produces a highly lethal tunicaminyluracil class corynetoxin (LD<sub>50</sub> 3–5 mg/kg in sheep) that causes severe and often fatal neurological and hepatic disease. Sub- lethal doses are also damaging to livestock and diminish wool quality and quantity, meat quality, and cause fetal abortions in sheep. Symptom onset can occur up to 12 weeks after ingestion and a single exposure can cause lethality; toxin effects are cumulative. R. toxicus corynetoxins were identified as a new member of the tunicaminyluracil class of antibiotics, which inhibit an early stage in prokaryotic peptidoglycan cell wall assembly. In eukaryotes, tunicamycin reduces protein N-glycosylation by inhibiting uridine diphospho-N-acetylglucoseamine:dolichol-N-acetylglucoseamine-1-phosphate transferase. The dangers to U.S. agriculture presented by R. toxicus and tunicamycin production in forage resulted in the bacterium being listed as a U.S. Department of Agriculture (USDA) Plant Protection and Quarantine Select Agent in 2008 and relisted in 2012 ([www.selectagents.gov/SelectAgentsandToxinsL ist.html](http://www.selectagents.gov/SelectAgentsandToxinsList.html)). R. toxicus is most commonly found in annual ryegrass (Lolium rigidum) in association with Anguina funesta or other anguinid seed-gall nematodes. The infection cycle begins with R. toxicus adhering to the cuticle of compatible juvenile nematodes in the soil and being carried to the growing point of the forage grass. Once in a developing seed, the nematode and bacteria compete to form either a nematode or a bacterial gall. R. toxicus growth in developing galls can produce a yellow exopolysaccharide “slime” or gummosis; therefore, the plant infection is commonly called yellow slime disease. The trigger for toxin production is unknown but toxin generally appears late in the growing season as seed are senescing. Senesced seed, nematode galls, and bacterial galls dry and fall to the ground to repeat the disease cycle the following year. Host range of R. toxicus appears to be determined by the host range of the vectoring nematode. Tunicamycin production is often associated with the presence of an R. toxicus-specific bacteriophage NCPPB 3778, although toxin production has also been measured in the absence of phage. The NCPPB 3778 genome has recently been sequenced and is similar to siphoviral genomes. Although its role in nature is unclear, NCPB3778 infection of R. toxicus can restore tunicamycin production in the lab, where the ability to produce tunicamycin is otherwise rapidly lost in culture (A.J. Sechler, personal observation). Although complete genome sequences are publically available for two R. toxicus strains, FH-145 (NZ_CP010848.1) and WAC3373 (NZ_CP013292.1), neither sequence has been carefully annotated. In addition, the full genetic diversity of R. toxicus is not well represented by these two strains alone. Therefore, two additional strains of R. toxicus, FH-79 (the type strain) and FH-232 were sequenced. Because an established system for genetic modification of R. toxicus is not available, the analysis presented here uses comparative and structural genomics to identify the genetic basis of several previously described phenotypes including the production of tunicamycin. # Materials and methods ## Bacterial strains, culture, and DNA extraction Cultures of R. toxicus FH-79 and FH-232 were obtained from Dr. Ian Riley (University of Adelaide, South Australia); additional information about their origins is presented in. R. toxicus was maintained on modified YGM (mYGM). One liter of this modified media contained yeast extract 2 g, glucose 1.25 g, K<sub>2</sub>HPO<sub>4</sub> 0.25 g, KH<sub>2</sub>PO<sub>4</sub> 0.25 g, MgSO<sub>4</sub>·7H<sub>2</sub>O 0.1 g, and agar 16 g. Cultures were incubated at 25°C unless otherwise noted; cryogenic stocks were stored in 15% glycerol at -80°C. DNA was extracted using a modified Marmur method from 3 day old liquid cultures. DNA quality was estimated by OD<sub>260/280</sub> ratio as measured on a Nanodrop 2000 (Thermo Fisher Scientific) and only DNA with a ratio \>1.6 was used for sequencing. Purity of the cultures used for DNA extraction was confirmed by plating 50 μl on mYGM and monitoring for growth of non-R. toxicus colonies. 16S rDNA was sequenced using an Applied Biosystems 3130XL (Thermo Fisher Scientific) to test purity of extracted DNA prior to genomic sequencing; only extracted DNA yielding a single 16S sequence was sequenced further. ## Genome sequencing and assembly For R. toxicus FH-79, a shotgun DNA library was constructed for the 454 Junior (Roche) according to the manufacturer’s directions and three sequencing runs were performed. In addition, a library FH-79 was also constructed for the PacBio RSII (Pacific Biosciences); three SMRT cells were sequenced for FH-79 at the Washington State University Genomics Lab. The 454 sequence data was assembled using Lasergene Ngen v12.0 (DNAStar) and PacBio reads using Pacific Bioscience’s Hierarchical Genome-Assembly Process (HGAP); consensus sequences from the two methods were compared using Ngen. For R. toxicus FH-232, only a PacBio library was constructed and 3 SMRT cells were sequenced also at the Washington State University Genomics Lab; assembly was performed using Pacific Bioscience’s Hierarchical Genome-Assembly Process (HGAP). The putative tunicamycin gene cluster, vancomycin resistance genes, and 16S rDNA from FH-79 and the CRISPR region from FH-232 were resequenced by primer walking on an Applied Biosystems 3130XL (Thermo Fisher Scientific) to validate genome assembly. The genome sequences presented here have been deposited in GenBank under the following accession numbers: R. toxicus FH-79 BioProject PRJNA312185 and BioSample SAMN04495682; R. toxicus FH-232 BioProject PRJNA312185 and BioSample SAMN06040670. ## Genome annotation and analysis Initial automated genome annotation was obtained using the Prokaryotic Genome Annotation Pipeline (PGAP) at National Center for Biotechnology Information (NCBI). Custom gene models were constructed as necessary by aligning the selected input sequences using muscle (<http://www.drive5.com/muscle/>), followed by invocation of hmmbuild from the HMMer version 3.1.b2 package (<http://hmmer.org/>). The hmmscan tool from the HMMer suite was used for database scans. Predicted chromosomal origin of replication was identified using Ori-finder (<http://tubic.tju.edu.cn/Ori-Finder/>). Standard protein family and domain models were obtained from TBLASTN (<https://blast.ncbi.nlm.nih.gov/Blast>), Pfam (<http://pfam.xfam.org/>), TIGRFam (<http://www.jcvi.org/cgi-bin/tigrfams/index.cgi>) and TnpPred. Alien_Hunter and antiSMASH were used to identify regions with anomalous nucleotide composition and putative biosynthetic clusters, respectively; identified regions were manually annotated with special attention paid to transposases and known virulence factors in other Actinobacteria. Whole genome alignments were performed with Mauve. CRISPR analysis was performed using CRISPRFinder. ## Phylogenetic trees For the Actinobacteria phylogenetic tree, sequences for gyrB, secA1 and 16S rDNA genes were obtained for 15 representative species of Actinobacteria. Sequences were concatenated and aligned using three iterations of tree searching and realignment with the Clustal Omega algorithm in Megalign Pro (Lasergene). MEGA6 was then used to conduct model determination and maximum likelihood tree searches (default settings) with 100 iterations of bootstrapping analyses. A minimum bootstrap value of 50 was used as a cut-off level of support to determine valid branches. Rubrobacter radiotolerans was set as the outgroup. For the protease tree, amino acid sequences of serine proteases putatively secreted from R. toxicus FH-79 and Clavibacter michiganensis subsp. michiganensis NCPPB382 were aligned with MSAProbs. Aligned sequences were used to generate maximum-likelihood trees based on the Jones-Taylor-Thornton (JTT) model of MEGA 7.0 with bootstrapping repetitions of 1,000. ## GC content plot and statistics Percentage GC content was plotted using GC content calculator ([www.biologicscor p.com/tools/GCContent](http://www.biologicscorp.com/tools/GCContent)) with a sliding window size of 2,000 bp. Statistical significance of GC content differences was calculated by repeated random sampling of 1000 13.4 kb regions of the R. toxicus FH-79 genome excluding rDNA and the TGC itself. # Results ## Whole-genome sequencing, assembly, and annotation Sequence data resolved each genome into a single circular chromosome of 2,343,780 and 2,394,755 bp for R. toxicus FH-79 and FH-232, respectively; no plasmids or other extra-chromosomal sequences were found for either strain. The PacBio SMRT sequencing technology was especially important for evenly closing these high-GC genomes. compares these two genomes to the previously available R. toxicus FH-145 (NZ_CP010848) and WAC3373. All four R. toxicus strains have an average GC content of approximately 61%. Annotation using NCBI’s Prokaryotic Genome Annotation Pipeline yielded 2,078 open reading frames (ORFs) for R. toxicus FH-79 and 2,137 ORFs for FH-232. This PGAP annotation also contained a large number of genes with the \pseudo keyword due to variations in the placement of the stop codon. Manual comparison with carefully annotated genomes suggest that the observed variations in gene length are typical in Actinobacteria; therefore, the \pseudo keyword was removed. The two sequenced genomes presented here were aligned with the two available R. toxicus genomes using Mauve after rotating and/or reverse complementing sequences to place dnaA as the first gene on the positive strand. As shown in, the four genomes are essentially syntenic. The pink, yellow, and blue regions represent three locally collinear blocks (LCBs). The distinction between the pink and blue regions is an artifact arising from circular genomes being treated as linear by the Mauve algorithm. Therefore, there are only two physically distinct LCBs separated by short transpositions; location of transposition region is marked by a green line in. R. toxicus FH-232 has 12 insertions not present in the other genomes; this accounts for its larger genome size. Predicted and annotated open reading frames spanned the typical range of necessary biological functions, metabolism, cell wall biosynthesis, defense, etc. Importantly, no ORFs annotated as phage genes were present, indicating no prophages are incorporated into the bacterial genome and that samples were free from contaminating phage. R. toxicus FH-79 and FH-232 both have two 16S rDNA sequences and have 46 or 45 tRNAs, respectively. Because of the extensive similarity among the four sequenced R. toxicus strains, further analysis is only presented for R. toxicus FH-79 except for rare cases where significant differences exist. ## R. toxicus groups with the Microbacteriaceae A phylogenetic analysis based on three conserved genes clearly demonstrates that R. toxicus is a member of the Microbacteriaceae, most closely related to Leifsonia xyli and Clavibacter michiganensis. These three genes (gyrB, secA1, and 16S rDNA) are frequently used for resolving subfamilial relationships in Actinobacteria due to appropriate levels of within subfamily variation. Although L. xyli and C. michiganensis have slightly larger genomes than R. toxicus (2.6 Mb and 3.3 Mb, respectively, vs. 2.3 Mb), all three species have GC-rich genomes and all are plant-associated. ## Tunicamycin gene cluster A putative 13.4 kb tunicamycin gene cluster (TGC) was identified based on homology to proteins encoded by the TGC from Streptomyces chartreusis NRRL 3882. As shown in, the R. toxicus TGC has a GC content markedly lower than the genome as a whole (52% vs. 61%). Repeated random sampling of the genome demonstrated that only 0.2% of comparably sized genome segments have a GC- content that is lower than the TGC (p-value \< 0.002). Although the S. chartreusis TGC appears to be a single polycystronic operon consisting of either 12 (tunA-tunL) or 14 (tunA-tunN) genes, the R. toxicus TGC contains two operons, one monocystronic (tunC) and one polycystronic (tunA-tunF;). R. toxicus also lacks the tunM methyltransferase and tunN NUDIX hydrolase; however, these genes are not essential for tunicamycin biosynthesis. The TGC in R. toxicus does contain two novel ORFs: tunO, a hypothetical gene unique to R. toxicus, and tunP, a polyketide synthase with a beta-ketoacyl synthase domain. All the predicted TGC genes are present in the same order and orientation in all four sequenced strains. R. toxicus FH-145 and WAC3373 are identical at the nucleotide level to the FH-79 TGC except for the addition or deletion of 2 or 3 Gs in a highly repetitive, G-rich intergenic region upstream of tunC. The FH-232 TGC is more than 99% identical to the other TGC regions. FH-79 has been previously shown to produce tunicamycin; FH-232 and FH-145 also produce toxin. While tunicamycin production by WAC3373 has not been reported, biosynthesis is likely given the highly conserved TGC. The hypothesized tunicamycin biosynthetic pathway is shown in. ## Additional secondary metabolites To identify regions with anomalous nucleotide composition that may interfere with statistically based gene calling algorithms, Alien_Hunter was used to query the R. toxicus FH-79 genome. Such regions are also of interest because they may arise from horizontal gene transfer events and are more likely to contain biosynthetic genes for secondary metabolites or virulence factors. Forty-two regions, including the TGC described above, were identified and are listed in. To further aid in the identification of secondary metabolite biosynthetic clusters, antiSMASH was also used to query the R. toxicus FH-79 genome. As shown in, 21 of the 42 regions identified with Alien_Hunter were also identified within 14 antiSMASH regions. Regions vary from 5.2–28.7 kb and are predicted to encode a wide variety of functions: bacteriocins (lantibiotic), type III polyketide synthase (PKS) proteins, non-ribosomal peptide synthetase (NRPS) proteins, multidrug efflux permeases, serine proteases, exopolysaccharide- related proteins, Type VII secretion system (T7SS) proteins, and numerous YD/RHS-like repeat-associated proteins. Historically, R. toxicus has been defined based on several different biochemical characteristics. In addition to the production of tunicamycin as described above, these include yellow colony color, exopolysaccharide “slime” production, MK-10 as the predominant isoprenoid quinone, and a non-mevalonate pathway for isoprenoid biosynthesis. Although the exact biochemical nature of the yellow pigment has not been determined, the only candidate carotenoid biosynthetic cluster in the genome is shown in. It consists of six predicted genes: crtEb (AYW78_09695, UbiA-like prenyltransferase); crtYf (AYW78_09700, lycopene cyclase); crtYe (AYW78_09705, lycopene cyclase); crtBI (AYW78_09710, bifunctional phytoene synthase/oxidoreductase); crtE (AYW78_09715, geranylgeranyl diphosphate synthase); and ispH (AYW78_09720, isopentyl-diphosphate delta-isomerase, type I). The only predicted exopolysaccharide biosynthetic cluster in the R. toxicus genome is present on antiSMASH cluster AS-8. This cluster was identified based on similarity to proteins in the wcm, wcn, wco, and wcq exopolysaccharide biosynthetic clusters in Clavibacter michiganensis subsp. nebraskensis NCPPB 2581 (NC_020891.1). The carotenoid pigment and the secreted exopolysaccharide may account for the yellow slime observed during plant infection. The menaquinone profile, along with 16S rDNA sequence and cell wall amino acid composition, was used to justify moving the type strain from Clavibacter to Rathayibacter. The predominant menaquinone identified by Sasaki et al., MK-10, is also the expected product of a gene cluster from antiSMASH cluster AS-5. This cluster contains genes with similarity to menB-menF and ubiE, the core menaquinone biosynthetic genes first identified in E. coli, as well as several additional genes. The ORF labeled idsA is predicted to encode a geranylgeranyl pyrophosphate synthase that may be involved in both menaquinone and carotenoid production. Most organisms use one of two different pathways to synthesize the important isoprenoid building blocks isopentenyl pyrophosphate and its isomer dimethylallyl pyrophosphate, either the classical mevalonic acid (MVA) pathway or the non-mevalonate/methylerythritol phosphate (MEP) pathway. Although Gram- negative bacteria only use the MEP pathway, several Gram-positive organisms, including many in the Microbacteriaceae family, use the MVA pathway. Studies using the isoprenoid biosynthetic inhibitor fosmidomycin are consistent with use of the MEP pathway by several Rathayibacter species. The R. toxicus genome contains ORFs similar to the core MEP pathway proteins from E. coli: DXS 1-deoxy-D-xylulose 5- phosphate synthase, AYW78_05260; DXR/IspC 1-deoxy-D- xylulose 5-phosphate reductoisomerase, AYW78_03715; IspE 4-diphosphocytidyl-2-C-methylerythritol kinase, AYW78_07950; and a bifunctional IspD/IspF 4-diphosphocytidyl-2-C-methylerythritol synthetase and 2-C-methylerythritol 2,4-cyclodiphosphate synthase, AYW78_08320. The MVA pathway appears to be absent from R. toxicus. antiSMASH cluster AS-18 is predicted to encode a lantibiotic or class I bacteriocin, a heavily modified, ribosomally synthesized anti-microbial peptide. The predicted prepropeptide is encoded by the gene with locus tag AYW78_09457 and is serine and alanine rich. Neighboring ORFs AYW78_09425 and AYW78_09430 encode proteins containing lantibiotic dehydratase domains while AYW78_09455 encodes a putative peptide cyclodehydratase. AYW78_09440 and AYW78_09445 encode FMN-dependent oxidases that may act on the cyclized thioesters. The only R. toxicus gene that exhibits any significant similarity to the LanP-type peptidases involved in cleaving lantibiotic leader peptides is not part of this cluster (AYW78_08500). Although not identified by either Alien_Hunter or antiSMASH, it is notable that the R. toxicus genome encodes three predicted vancomycin resistance proteins: VanH pyruvate dehydrogenase, AYW78_09940; VanA D-lactate dehydrogenase, AYW78_09945; and VanX D-ala-D-ala peptidase, AYW78_09950. R. toxicus FH-79 is resistant to vancomycin experimentally. ## CRISPR arrays R. toxicus possesses a complete Type I-E CRISPR-Cas system (E.coli-type) with eight cas genes and an adjacent approximately 8.9 kb CRISPR spacer array. The four different sequenced strains have slightly different numbers of non- repetitive spacer sequences and conserved direct repeats. R. toxicus FH-79 and FH-145 both have 145 non-repetitive spacer sequences and 146 conserved direct repeats while WAC3373 has 144 and 145 and FH-232 has 139 and 140, respectively. Non-repetitive spacer sequences revealed no identity to known plasmid or phage sequences. ## Predicted pathogenicity-related genes Relative to the related phytopathogen Clavibacter michiganensis subsp. michiganensis, R. toxicus possesses a limited arsenal of plant-associated cell-wall hydrolyzing enzymes, consisting of only a single polygalacturonase (AYW78_01285) and pectate lyase (AYW78_01485). This is consistent with the life strategy of R. toxicus, which apparently cannot infect plant leaves or stems but most acquire nutrients in seed galls initiated by nematode infestation. However, R. toxicus does possess numerous secreted serine proteases that share sequence homology to the pathogenicity-associated protein Pat-1 of C. michiganensis subsp. michiganensis. A total of 11 secreted serine proteases were identified with an additional conserved pseudogene; all contain predicted signal peptides suggesting extracellular localization as described in C. michiganensis subsp. michiganensis. The corresponding genes were designated chpA-K (chromosomal homology to pat-1) and sbtA (subtilisin-like serine protease). In contrast to C. michiganensis subsp. michiganensis, the secreted serine proteases are dispersed throughout the chromosome, but several of the proteases are located in close proximity including: (i) chpG, chpH, chpK (pseudogene) and (ii) chpB, chpC. Phylogenetically, the serine proteases of R. toxicus and C. michiganensis subsp. michiganensis appear distinct with the majority of R. toxicus proteases (ChpB-E, ChpI-J) forming a subgroup. No R. toxicus serine proteases clustered with the C. michiganensis subsp. michiganensis Ppa family or plasmid-associated (PhpA-B) serine proteases. The subtilisin-like serine proteases of R. toxicus and C. michiganensis subsp. michiganensis were the only secreted proteases to cluster across species. # Discussion The key feature of R. toxicus is its ability to exploit a protected environmental niche, the developing grass seed, and produce tunicamycin, a potent toxin for grazing livestock. Prior to the work presented here, very little was known about the biosynthesis of tunicamycin by R. toxicus. Until the publication of the phage NCPPB 3778 sequence, it was hypothesized that tunicamycin production could reside in the phage rather than on the bacterial chromosome. However, no ORFs with similarity to known tunicamycin biosynthetic genes were found in the phage genome. The discovery of a tunicamycin gene cluster (TGC) in R. toxicus with similarity to the previously characterized cluster from S. chartreusis is an important first step in understanding toxin production in this bacterium. Both the lower GC content of the TGC and its similarity to Streptomyces indicate that R. toxicus probably acquired the ability to synthesize tunicamycin via a horizontal gene transfer event; however, the TGC does not contain identifiable transposases, nor is it adjacent to a recognizable tRNA or flanked by inverted repeats as is typical for a mobile genetic element. R. toxicus is regulated as a select agent because it is associated with the production of toxin that results in the death of foraging livestock. There are additional concerns about potential secondary effects that could manifest in humans consuming either contaminated plant material or the meat of ARGT affected animals. R. toxicus causes little in the way of disease symptoms on grasses, with the accumulation of exopolysaccharide “slime” as the primary sign of pathogen infection, and there is no indication that R. toxicus infections result in significantly reduced plant host fitness. The lack of phytopathogenesis-related genes in the R. toxicus genome further suggests that this bacterial species may not be a typical plant pathogen. Rather, R. toxicus, like other Rathayibacter species, has evolved a unique approach to reaching and exploiting a desirable niche, by utilizing gall forming nematodes as a convenient vector. A possible biological function for toxin production is the elimination of nematodes from the seed gall, thus eliminating competition for resources. Tunicamycin production increases drastically when R. toxicus is inside the seedhead at a tipping-point between the nascent gall progressing to either nematode or bacterial dominated growth. However, while all members of the Rathayibacter genus utilize gall forming nematodes as vectors not all members of the genus produce tunicamycin, although it is not yet known whether the TGC is present in all members of the genus. It should be noted that toxin production comes at a significant fitness cost to R. toxicus, as toxin producing bacteria reproduce at significantly slower rates than non-toxin producers. Alternatively, toxin production for R. toxicus may provide an advantage against competing microbial populations, both fungal and bacterial, at one or more points in the life cycle from soil to seed head. Microbial competition could also explain the repertoire and diversity of biosynthetic pathways encoding non-ribosomal peptide synthetase (NRPS) proteins, polyketide synthase (PKS) proteins, thiazole/oxazole-modified microcins, lantibiotics, and numerous efflux proteins present in the R. toxicus genome. Regardless, for the select agent R. toxicus, there would seem to be some selection pressure(s) acting to keep the TGC and associated machinery present and active in the bacterial genome. It is not known how any tunicamycin producer protects itself from the toxin. It has been hypothesized that tunI and tunJ, which are both similar to ABC transporters, export tunicamycin outside the cell immediately after synthesis. It is possible to express the S. chartreusis TGC in other Streptomyces species and thereby confer both tunicamycin production and resistance, implying that at least in the case of S. chartreusis, any export or detoxification mechanisms reside within the TGC itself. The R. toxicus strains sequenced here complement the two previously available complete genome sequences. A previous analysis of R. toxicus strains found three major genotypic groups based on amplified fragment length polymorphisms (AFLP) and restriction digestion patterns using pulsed-field gel electrophoresis (PFGE). As indicated in, the previously sequenced R. toxicus FH-145 falls in subgroup A while FH-79 is in subgroup B and FH-232 is the sole member of subgroup C. Many of the same R. toxicus strains, as well as some more recently collected, were also analyzed by multi-locus sequence typing (MLST) and inter-simple sequence repeats (ISSR). This analysis found three main populations, RT-I, RT-II, and RT-III, with strain FH-232/FH100 again forming an outgroup. R. toxicus FH-79 and FH-145 were not included in the MLST analysis. However, by in silico PCR, they both belong to RT-III. The four subgroup A strains also analyzed by MLST all fall into RT-III while the three subgroup B strains examined are in RT-II. This makes R. toxicus FH-79, which is the type strain for the species, somewhat unusual as it falls into subgroup B and RT-III. It is most common for bacterial chromosomes to be circular in topology. However, a number of both Gram-positive and Gram-negative bacteria have linear chromosomes and/or plasmids; they are especially common in the Actinomycetales. The R. toxicus chromosome was hypothesized to be linear based on its failure to enter a pulsed-field gel either before or after nuclease S1 treatment. Whether or not large circular DNA migrates during PFGE depends on the exact electrophoretic conditions; insufficient experimental detail is provided to assess the conclusions of Agarkova et al.. The genome presented here is most consistent with a circular topology. Virtual PacI digests of a circular genome generate the number and size of bands observed experimentally more closely than a linear genome. Additional bands are predicted but would not be expected to be visible on a pulsed-field gel due to their small size. Linear chromosomes have large terminal inverted repeats on the ends; these sequences can be up to 1 Mb each. Unless care is taken during genome assembly, these terminal repeats can be mis-assembled on top of each other and give the appearance of a circular genome. Terminal repeats have been observed to be under-represented in PacBio raw reads, perhaps because of the bias toward long DNA fragments during library construction. Therefore one clue that a genome is linear can be the presence of contigs made up of short (Illumina or 454) reads that do not map to PacBio consensus sequence; no such contigs were found in the 454 sequence from R. toxicus FH-79. If terminal repeats are incorporated into the PacBio library and therefore appear once in a circular consensus sequence, those regions would be overrepresented in short read libraries. However, no such regions of higher coverage were observed. Prior estimates of genome size match the sequence obtained here quite well (2.2–2.3 Mb predicted vs. 2.3–2.4 Mb observed); if two large terminal repeats were missing from the genomes reported here, the sequences reported here would be expected to be significantly smaller than previous size predictions. In general, the larger Actinomycetales genomes tend to be linear and the smaller ones circular, although there are exceptions. R. toxicus, at 2.3–2.4 Mb, is definitely on the smaller end of genome size. All of these factors taken together tend to support the presence of a circular chromosome in R. toxicus. R. toxicus is most closely related to the systemic xylem-dwelling Gram- positive phytopathogenic bacteria Clavibacter michiganensis and Leifsonia xyli. While L. xyli subsp. xyli is a fastidious xylem-limited bacterium of sugarcane, C. michiganensis subsp. michiganensis is an opportunistic pathogen of tomato and colonizes both vascular and non-vascular tissue. Regardless of differences in host and systemic lifestyles, C. michiganensis subsp. michiganensis and L. xyli subsp. xyli possess numerous canonical plant-associated cell wall-degrading enzymes (PCWDEs). C. michiganensis subsp. michiganensis utilizes a variety of PCWDEs including hemicellulases, xylanases, cellulases, polygalacturonases, pectate lyases, and endoglucanases. However, R. toxicus lacks many PCWDEs, possessing only a single copy each of pectate lyase and polygalacturonase. The relatively small arsenal of plant- associated enzymes is surprising for a plant pathogen, but could demonstrate its closer association and reliance on a nematode vector for plant colonization. Despite the small arsenal of PCWDEs, R. toxicus possesses numerous serine proteases with homology to the pathogenicity-associated protein Pat-1 of C. michiganensis subsp. michiganensis. C. michiganensis subsp. michiganensis harbors serine proteases on a putative 129 kb pathogenicity island and extra-chromosomal plasmids, which are necessary for effective disease development in tomato. However, the serine proteases from R. toxicus are dispersed throughout the chromosome and appear distinct from the C. michiganensis subsp. michiganensis disease-associated serine proteases. The putatively secreted R. toxicus serine proteases could possess alternative functions associated with nematode colonization, as opposed to plant colonization or disease development, since cuticle penetrating serine proteases are highly represented in nematode pathogenic bacteria and fungi. It is interesting to note that Bird et al. (1984 & 1985) document the destruction of the nematode epidermis and cortical structures shortly after Rathayibacter attachment. The relative lack of PCWDEs and differing serine proteases suggest that R. toxicus is not a typical vectored phytopathogenic bacterium. In summary, analysis of the complete genome of R. toxicus has identified a likely genetic pathway (TGC) for the production of tunicamycin, based on homology to other tunicamycin biosynthetic clusters. This represents a critical first step towards understanding the control of the key pathway that makes this Select Agent pathogen such a significant threat to agriculture and food safety. Sequencing the genomes of other members of the Rathayibacter genus, both toxin producers and non-toxin producers, would provide corroborative evidence implicating the TGC in tunicamycin production as well as providing some evolutionary context for the introduction of the TGC as a likely mobile element. The current genomic context, however, suggests that the TGC is no longer mobile in any of the sequenced R. toxicus strains. Ultimately, the connection between the TGC and toxin production must be assessed by expression studies, gene knockouts, and functional restoration experiments. # Supporting information The authors would like to thank Dr. Ian Riley (University of Adelaide) for providing Rathayibacter toxicus strains and for considerable assistance over the years with R. toxicus. Funding for this work was from the United States Department of Agriculture (USDA) Agricultural Research Service appropriated project 8044-22000-040-00D and from two 2008 Farm Bill grants, Section 10201 administered through the United States Department of Agriculture, Animal and Plant Health Inspection Service (13-8130-0247-CA and 14-8130-0367-CA). Mention of trade names or commercial products in this publication is solely for the purpose of providing specific information and does not imply recommendation or endorsement by the U.S. Department of Agriculture. USDA is an equal opportunity provider and employer. [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Global Institute for Food Security and the School of Environment and Sustainability, University of Saskatchewan, Saskatoon, Saskatchewan, Canada [^3]: Current address: Dept. of Biochemistry, Molecular Biology, Entomology and Plant Pathology, Mississippi State University, Mississippi State, Mississippi, United States of America [^4]: Current address: AIDS and Cancer Virus Program, Frederick National Laboratory for Cancer Research and Leidos Biomedical Research, Inc., Frederick, Maryland, United States of America
# Introduction An estimated 15% of couples are affected by infertility and male factor infertility is the suspected cause in 30–50% of these couples. Although the exact cause of many of these cases are unknown, chromosomal deletions, translocations and SNPs are associated with infertility in as many as 30% of these men. Identifying and understanding how genetic anomalies affect spermatogenesis and fertilization will improve the likelihood of overcoming male infertility. Conversely, effective male contraception may be developed based on data showing associations between genetic variation, nutrient metabolism and male factor infertility. Although the relationship between overall nutritional status and reproduction is well documented, the relationship between micronutrient metabolism and reproduction is understudied. There is some evidence that aberrant micronutrient metabolism may play a causative role in male factor infertility. Dietary deficiencies of vitamins A, C and E as well as trace metals such as zinc and selenium are associated with male infertility in animals and humans, Choline is an essential nutrient for humans and is important for normal fetal development. A link between choline metabolism and male fertility has been demonstrated in only one paper. Geer reported both normal mating behavior and sperm motility required adequate choline availability in Drosophila melanogaster and that carnitine, a proposed choline substitute, was unable to support male fertility. We discovered that choline dehydrogenase (CHDH, EC 1.1.99.1) is necessary for normal male fertility in mice. Male choline dehydrogenase knockout mice (Chdh<sup>−/−</sup>) are infertile due to severely compromised sperm motility; decreased motility is the result of abnormal mitochondrial structure and function in sperm cells. Betaine (N,N,N-trimethylglycine), a metabolite of choline, donates methyl groups for the formation of methionine from homocysteine and is an organic osmolyte used by cells for regulatory volume control. Dietary sources of betaine include wheat, shellfish, spinach and sugar beets. In addition, betaine can be made de novo via the oxidation of choline in a series of reactions catalyzed by CHDH and betaine aldehyde dehydrogenase (BADH, EC 1.2.1.8). Conversion of choline to betaine takes place in the mitochondrial matrix following the transport of choline across mitochondrial membranes. The betaine formed is a zwitterion at neutral pH and diffuses out of mitochondria for use in one-carbon metabolism. SNPs have been identified in the human CHDH gene; rs12676 (G233T) is a non- synonymous SNP located in exon 3 of the gene. Occurrence of the variant T allele results in the replacement of arginine, a polar, hydrophilic amino acid, with leucine, a hydrophobic amino acid. 38–40% of individuals are heterozygous and 2–9% are homozygous for rs12676. The CHDH minor T allele is associated with increased susceptibility to developing clinical symptoms of dietary choline deficiency (steatosis and muscle cell damage) as well as increased risk of breast cancer. Although not in the CHDH gene, a SNP in the adjacent interleukin 17 beta receptor (IL17βR) gene, rs1025689, is associated with increased risk of developing choline deficiency specifically in men (odds ratio 13.5; see for detailed description of this study). We now present data indicating that rs12676 genotype is also associated with dysmorphic mitochondrial structure, changes in sperm motility patterns and decreased energy status in human sperm. Sperm and hepatocytes harboring the TT rs12676 genotype have less CHDH protein, suggesting that this SNP is functional. In addition, rs1025689 is linked to changes in sperm motility patterns, suggesting that these SNPs may be contributing factors in the occurrence of idiopathic male infertility. # Materials and Methods ## Chemicals and Reagents All chemicals and reagents used were obtained from Sigma Aldrich (St. Louis, MO), unless otherwise noted. ## Ethics statement The study design and all procedures used in this study were approved by the University of North Carolina at Chapel Hill Office of Human Research Ethics Institutional Review Board. Subjects were at least 18 years of age and were recruited from the Charlotte-Kannapolis, North Carolina region. Informed written consent was obtained from all subjects at the initial clinic visit. Subjects were compensated up to \$125 for their participation in the study. ## Study design and recruitment This study was conducted at the University of North Carolina at Chapel Hill Nutrition Research Institute (Kannapolis, North Carolina) and was implemented in two phases – screening subjects for rs12676 and rs1025689 genotype, and analysis of semen and sperm specimens collected from subjects with these SNPs. Subjects were screened to determine the frequency of these SNPS in our study population and to allow for enrichment of the homozygous variant genotypes. Participants were recruited via mass email, advertisements in Craig's list, newspaper articles describing the study and in-person recruiting at local community colleges. ## Blood collection and DNA isolation Subjects were asked to complete the health questionnaire form at the screening visit. In order to screen subjects for the rs12676 and rs1025689 SNPs, blood was collected into Vacutainer Cell Preparation Tubes (CPT tubes; BD Diagnostics, Franklin Lakes, NJ) containing sodium citrate; lymphocytes were separated from other blood components for subsequent genomic DNA extraction. All CPT tubes were stored on ice if not processed immediately; all CPT tubes were processed within 1 hour of sample collection. Briefly, CPT tubes were centrifuged in a Sorvall RC-3B centrifuge equipped with an H-2000B rotor (Thermo Fisher Scientific, Waltham, MA) at 1500× g for 30 minutes at room temperature. Plasma was aliquoted into 2 mL microfuge tubes and stored at −80°C for choline metabolite analysis. The lymphocyte layer was washed with phosphate buffered saline (PBS), transferred to a 15 mL conical tube and was pelleted by centrifugation at 1000× g for 5 minutes at room temperature. Pellets were again washed with PBS, transferred to 1.5 mL microfuge tubes and pelleted by centrifugation in an Eppendorf 5415D microcentrifuge at 800× g for 5 minutes at room temperature. Genomic DNA was purified from lymphocyte pellets using a QIAamp DNA Mini Kit (Qiagen, Valencia, CA) according to manufacturer's instructions with some modification. Specifically, lymphocyte pellets were equilibrated to room temperature and resuspended in 500 µL PBS. The amounts of Qiagen Protease, Buffer AL and ethanol (96%–100%) were adjusted proportionally as indicated by the manufacturer's instructions. Two separate elutions of 100 µL with Qiagen Buffer AE were performed. Samples were incubated at room temperature for 5 minutes prior to each elution. Both elutions were collected in the same 1.5 mL microfuge tube for a final volume of 200 µL. DNA quality and concentration was determined using a Nanodrop 8000 Spectrophotometer (Thermo Scientific, Wilmington, DE). ## rs12676 genotyping rs12676 genotype was determined by direct sequencing. A 260 base pair region of the CHDH gene containing rs12676 was PCR amplified using Deep Vent<sub>R</sub>™ (exo-) DNA polymerase (New England Biolabs, Ipswich, MA) according to manufacturer's recommendations. The primers used for amplification were: CHDH forward 5′-ATTCCCCTCCGTGGATCAG-3′ and CHDH reverse 5′-TGTCGTCGCACAGGTTGG-3′. Each 50 µL reaction contained 600 ng of genomic DNA, primers at a final concentration of 200 nM, and 4 units of Deep Vent<sub>R</sub>™ (exo-) DNA polymerase. The PCR conditions were: an initial denaturing step at 98°C for 10 minutes followed by 30 cycles of denaturing at 98°C for 1 minute and annealing/extension at 72°C for 5 minutes. PCR products were purified from other reaction components using a QIAquick PCR Purification Kit (Qiagen, Valencia, CA) and the resulting DNA concentration was determined using Nanodrop 8000 Spectrophotometer (Thermo Scientific, Wilmington, DE). Purified CHDH fragments were sequenced using BigDye® Terminator chemistries (Applied Biosystems, Carlsbad, CA) by Eton Bioscience, Inc (Research Triangle Park, NC) using the CHDH forward primer. rs12676 genotype was determined by examining sequencing chromatograms using Sequence Scanner software (version 1.0, Applied Biosystems, Carlsbad, CA). ## rs1025689 genotyping rs1025689 genotype was determined using a TaqMan® SNP genotyping assay (Applied Biosystems, Carlsbad, CA) according to manufacturer's instructions. PCR reactions were performed using a StepOne™ Real Time PCR System and 2× TaqMan® Genotyping Master Mix (Applied Biosystems, Carlsbad, CA). ## Semen collection and processing Subjects were asked to refrain from sexual activity for 48 hours prior to semen donation. Semen was produced by masturbation and collected into 50 mL sterile sample cups. Olive oil was provided as a lubricant to use as necessary. Samples were incubated at room temperature for 30 minutes to allow for liquefaction. Semen volume was measured with a pipette. Sperm were separated from other seminal fluid components by layering the sample over a 45% ISolate®/human tubular fluid (HTF; 190 mM NaCl, 9 mM KCl, 0.7 mM KH<sub>2</sub>PO<sub>4</sub>, 0.3 mM MgSO<sub>4</sub>-7H<sub>2</sub>0, 4 mM CaCl<sub>2</sub> – 2H<sub>2</sub>0, 0.025 mM NaHCO<sub>3</sub>, 2.78 mM D-glucose, 21.4 mM lactate, 0.33 mM pyruvate and 5 mg/mL bovine serum albumin (BSA; Fraction V), 5 M NaCl was added as necessary to adjust osmolality) media gradient followed by centrifugation at 300× g for 20 minutes at room temperature using a Beckman- Coulter Allegra X-15R Centrifuge and SX4750A rotor. ISolate® was purchased from Irvine Scientific (Santa Ana, CA). The supernatant was discarded and the resulting sperm pellet was washed twice in 3 mL HTF followed by centrifugation at 300× g for 10 minutes at room temperature. Sperm were resuspended in 4 mL HTF and used for subsequent analyses. ## Sperm counts The total number of sperm per ejaculate and sperm concentration were determined by counting cells with a hemocytometer. ## Sperm motility Sperm were diluted 1∶10–1∶15 in HTF for motility measurements. 200 µL of diluted sperm were placed into a 35 mm glass bottom dish and covered with a coverslip. For each sample, video of 10 random, unique microscope frames were recorded using a Zeiss Axio Observer (Carl Zeiss, Inc, Thornwood, NY) inverted microscope equipped with a temperature controlled incubation chamber equilibrated to 37°C. Sperm were viewed under phase contrast conditions with a 20× objective lens. Motility parameters including mean velocity (MVUS), curvilinear velocity (VCL), straight distance velocity (VSL) and mean tortuosity (MT; MT = VCL/VSL) were determined using Zeiss AxioVision (release 4.7) image tracking software (Carl Zeiss, Inc, Thornwood, NY) as previously described. ## Electron microscopy Approximately 500 µL of washed sperm were transferred to 1.5 mL microfuge tubes; sperm were pelleted by centrifugation at 16,000× g for 5 minutes at room temperature. The supernatant was discarded and sperm pellets were fixed for transmission electron microscopy in 2% paraformaldehyde, 2.5% gluteraldehyde, 0.2% picric acid in 0.1 M sodium cacodylate, pH 7.2. The pellet was encapsulated in 2% agarose buffered with 0.1 M sodium cacodylate buffer, pH 7.2. The encapsulated pellet was post-fixed in 1% osmium tetroxide in 0.1 M sodium cacodylate buffer for 1 hour. Samples were washed in deionized water, dehydrated through an ethanol series, transferred to propylene oxide, infiltrated and embedded in Polybed 812 resin (Polysciences, Inc., Warrington, PA). 70 nm ultrathin sections were post-stained in 4% aqueous uranyl acetate and 0.4% lead citrate. Samples were examined and photographed using a Zeiss EM-10A transmission electron microscope (LEO Electron Microscopy, Thornwood, NY) with an accelerating voltage of 60 kV. ## ATP assay ATP concentration in sperm was measured using an ATP Bioluminescence Assay Kit CLS II (Roche Diagnostics, Mannheim, Germany) according to manufacturer's instructions. Luminescence was measured using a 1420 VICTOR<sup>2</sup> microplate reader (Perkin Elmer, Waltham, MA). ATP concentration was normalized to number of sperm analyzed. ## CHDH expression Sperm pellets were lysed in a buffer of 2% w/v SDS, 0.375 M Tris, pH 6.8, 10% sucrose. Lysates were boiled at 100°C for 5 minutes and clarified by centrifugation at 16.1× g for 10 minutes at 4°C. Protein concentration was determined by the BCA protein assay (Pierce/Thermo Scientific, Rockford, IL). 10 µg of protein lysate were resolved by SDS-PAGE and transferred to PVDF membrane (Bio-Rad, Hercules, CA). Primary human hepatocyte lots, screened for rs12676 genotype as described above, were purchased from Zen-Bio, Inc. (Research Triangle Park, NC). Three lots for each rs12676 genotype were obtained and lysates were made in radioimmunoprecipitation assay (RIPA) buffer and protein concentration measured using the BCA assay. Equal volumes of lysate (80–200 µg protein) were resolved by SDS-PAGE and transferred to PVDF as described above. The membrane was incubated with anti-CHDH antibody (1∶1000 dilution in 5% BSA/PBS-T; ProteinTech Group, Inc., Chicago, IL) at 4°C overnight. A 1∶10,000 dilution of HRP-conjugated anti-rabbit secondary antibody (Millipore, Billerica, MA) was incubated with the membrane for 1 hour at room temperature and bands detected with enhanced chemiluminescence (SuperSignal® West Pico ECL, Thermo Scientific, Rockford, IL). The number of pixels in the CHDH band was quantitated using the lasso tool in Adobe Photoshop (Adobe Photoshop CS3 Extended v.10.0.1) and the number of pixels per microgram of total protein loaded was calculated. As an external control for protein integrity, the abundance of ∝-TUBULIN (sperm) or ß-ACTIN protein (hepatocytes) was determined using mouse monoclonal antibodies (anti-alpha-tubulin antibody: 1∶2000 dilution in 5% BSA/PBST (Sigma Aldrich, St.Louis, MO); anti-beta-actin antibody: 1∶10,000 dilution in 5% BSA/PBS-T; Abcam, Cambridge, MA) and an HRP-conjugated goat anti-mouse secondary antibody (1∶5000 dilution in 5% BSA/PBS-T; Abcam, Cambridge, MA) ## Plasma and sperm choline metabolite analyses The concentrations of choline and its metabolites \[choline, betaine, glycerophosphocholine (GPCho), phosphatidylcholine (PtdCho), and sphingomyelin (SM)\] in plasma and sperm were measured by liquid-chromatography ionization- isotope dilution mass spectrometry (LC-ESI-IDMS) as previously described. Phosphocholine was not detected in either plasma or sperm. GPCho was not detected in plasma. SM was not detected in sperm. Plasma samples were collected from the entire screened population on Day 1. Three million sperm per subject were pelleted and flash frozen in liquid nitrogen on Day 2. ## Statistical analyses Statistical differences among genotypes were determined using JMP 9.0 software (SAS Institute, Cary, NC) using ANOVA and Tukey-Kramer HSD. Statistical tests were performed on log 10 transformed data for semen volume, total sperm per ejaculate and sperm concentration as these data failed a Shapiro -Wilk test of normality. Only data from sperm recorded for at least 3 seconds were included in motility statistical analyses. In order to address the intra-individual variation in the sperm motility data the following method was used to determine statistical differences. For each continuous measure (MT, MVUS, VCL and VSL) cutpoints for quartiles were determined from all observations having the wild type/wild type (WW) genotype for that SNP. Then using those cutpoints, all observations were placed into a quartile and an association between SNP level and the most extreme quartiles (1 and 4) was assessed via a repeated measures logistic regression with quartile (1 or 4) as the response and SNP as the predictor. The repeated measures on subject were taken into account by using a compound symmetric correlation matrix for observations within the same subject. P-values less than 0.05 were considered statistically significant. # Results ## rs12676 and rs1025689 distribution frequencies rs12676 and rs1025689 distribution frequencies were calculated for the population of men screened for inclusion in this study. For rs12676, 52% of subjects were GG, 41% were GT and 7% were TT. For rs102689, 22% were GG, 48% were GC and 30% were CC. These results are in agreement with published data regarding these two SNPs. It is important to note that 100% of men who were homozygous for rs12676 had at least one minor C allele of rs1025689; 83% of men with the TT rs12676 genotype were also CC for rs1025689. In addition, 91% of men who were GT for rs12676 had at least 1 C allele for rs1025689. ## Study population Average age, average number of biological children per subject and occurrence of abnormal semen characteristic or infertility were calculated from self-reported information provided by the health questionnaire form completed on Day 1. The average age of the entire screened population was 33.5 years and subject age ranged from 18 to 76 years. The average number of biological children per subject was calculated by dividing the total number of children per genotype by the number of subjects who answered this question. Men who were TT for rs12676 reported the lowest number of biological children per subject (0.33) while men who were wild type for rs1025689 reported the highest (0.94). 33% (2 of 6) of men with rs12676 TT genotype reported semen abnormalities or infertility; these were reported by 9.7% (4 of 41) of men who were wild type for rs12676. For rs1025689, 11.7% (2 of 17) of GG, 5.2% (2 of 38) of GC and 8.3% (2 of 24) of CC subjects reported semen abnormalities or infertility. ## Semen parameters rs12676 genotype was not associated with changes in semen volume, number of sperm per ejaculate or sperm concentration. Men who were homozygous for rs1025689 had decreased sperm concentration compared to men who were heterozygous for this SNP. Mean values for these parameters were all within the normal range expected of the general human population. All results were above the World Health Organization 5% lower reference limit for semen parameters from the general population except for total sperm per ejaculate and sperm concentration for men who were CC for rs1025689. ## Sperm motility characteristics Sperm from men who were homozygous for rs12676 had increased curvilinear velocity and tortuosity when compared to sperm from men who were wild type for this SNP. Sperm produced by men with the GT genotype for rs12676 traveled greater distances at a faster rate and had more tortuous paths than sperm collected from men who were GG for this SNP. No differences between sperm from heterozygous and homozygous subjects were detected. Sperm from men who were CC for rs1025689 had increased average velocity as well as curvilinear and straight line velocity when compared to sperm collected from men who were wild type for this SNP. Men who were homozygous for rs1025689 produced sperm that traveled in more tortuous paths (i.e. less progressively motile) as compared to sperm from men who were heterozygous for this SNP. Subjects who were GC for rs1205689 produced sperm with higher measures of average velocity, curvilinear and straight line velocity compared to sperm from men with the GG genotype. Tortuosity was decreased in sperm from heterozygous men compared to sperm from men who were CC for rs1025689. ## Sperm mitochondrial ultrastructure and energy levels Abnormal mitochondrial structure was observed in sperm collected from men who have two copies of the rs12676 minor allele. Mitochondria in the midpiece of these sperm appeared swollen with disordered cristae structure. Sperm from men with one allele of rs12676 (GT) have 40% less ATP than sperm produced by men who are GG for this SNP. Men carrying two alleles of the SNP (TT) produce sperm with 73% less ATP than men who are GG. 83% of subjects who were homozygous for rs12676 were also homozygous for rs1025689, but when mitochondrial morphology and ATP concentrations were analyzed in men who were homozygous for rs1025689 only, and we found there was no relationship between rs1025689 genotype and these measures. ## CHDH protein expression CHDH protein expression, relative to ∝-TUBULIN expression, was decreased in sperm produced by men who had one or two alleles of rs12676. This effect was observed in non-reproductive tissues as well; individuals who harbor the TT rs12676 genotype had a lower CHDH protein in their hepatocytes compared to individuals who were GG for this SNP. The CHDH protein levels in GT hepatocytes were not significantly changed from either GG or TT hepatocytes. Both ∝-TUBULIN and β-ACTIN protein expression was not changed among the genotypic groups (not shown). ## Plasma choline metabolite concentrations Choline metabolite concentrations in plasma and sperm were measured using LC- ESI-IDMS. Men who were homozygous for rs12676 minor allele had a small, but significant increase in plasma free choline concentrations. There were no differences in plasma betaine, PtdCho or SM among the other rs12676 or rs1025689 genotypes. There was no effect of rs12676 genotype on sperm choline metabolite concentrations. Sperm from men who were CC for rs1025689 produced sperm with decreased betaine, GPCho and PtdCho concentrations compared to sperm produced by men who were GC for this SNP. # Discussion We previously reported the novel finding that Chdh deletion in mice resulted in male infertility due to compromised sperm motility. We now present evidence demonstrating that this discovery translates, in part, to humans. A common SNP in the human CHDH gene, rs12676, is associated with altered sperm motility patterns, dysmorphic mitochondrial ultrastructure and a significant reduction in ATP concentrations in sperm. Humans who were heterozygous or homozygous for this SNP had less CHDH protein in their sperm. Interestingly, this association also was observed in hepatocytes from individuals who were homozygous for this SNP. Further studies are required to determine whether this is a result of decreased CHDH mRNA translation or increased CHDH protein degradation. Either way, this data suggests that rs12676 is a functional SNP, or that it is a tag SNP that marks a functional haplotype of the CHDH gene. In addition, rs1025689, a SNP located in the adjacent IL17βR gene that is highly associated with increased susceptibility to dietary choline deficiency in men, is correlated with changes in sperm motility patterns and reduction of betaine, GPCho and PtdCho concentrations in sperm. The allele frequency distributions of the screened population for both rs12676 and rs1025689 were in agreement with previously reported frequencies. The average age among the genotypic groups was not significantly different; therefore, we conclude that any changes detected in semen characteristics or sperm cell function were not due to differences in age among the groups. The rs12676 TT genotype group reported the lowest number of biological children per subject (0.33) and the highest rate of semen abnormality/infertility diagnosis (33%) of all groups, which suggests that men with the TT genotype may be subfertile. Interestingly, subjects who were homozygous for rs12676 had higher free choline concentrations in their plasma. Although plasma choline is not a direct measure of choline concentrations within tissues, it is a reflection of tissue concentrations. An increase in free choline would be expected with a decrease in CHDH activity as less choline would be converted to betaine. We did not detect a decrease in betaine in these men, probably because we did not require that the subjects to fast prior to their blood draw and betaine can be obtained from the diet. Sperm from men who harbor at least one minor allele of rs12676 were less progressively motile, as indicated by an increase in curvilinear velocity and tortuosity. Mean velocity was significantly increased in sperm from heterozygous and homozygous subjects. Hyperactivated sperm are those displaying motility patterns characterized by vigorous, non-linear trajectories and are in contrast to progressively motile sperm which move forward in a somewhat linear path. Sperm released from the cauda epididymis typically are progressively motile at first and transition to hyperactivated motility when incubated in media formulated to support this process. Sperm collected from the female reproductive tract are generally hyperactivated. Although the HTF medium used to wash the sperm in our experiments is designed to mimic the environment of the female reproductive tract, we cannot definitively state whether the observed changes in motility patterns with respect to rs12676 genotype were the result of premature hyperactivation or a decrease in progressive motility. HTF media provides all metabolic substrates required for capacitation and hyperactivation and it is reasonable to assume that the sperm used for the motility analyses could have achieved hyperactivation. Sperm produced by men who were GC for rs1025689 were more progressively motile than sperm from men who were either GG or CC for this SNP. rs12676, but not rs1025689, was associated with dysmorphic mitochondrial structure. ATP concentration was inversely correlated, in a dose-dependent manner, with the number of rs12676 minor alleles. Although the exact mechanism causing these changes remains unknown, these results are very similar to those we observed in the Chdh<sup>−/−</sup> mice. ATP is required for sperm to be motile, but considerable controversy surrounds the source of the ATP used for this function. Because mitochondria are only found in the sperm midpiece, there is some question as to whether ATP generated by oxidative phosphorylation (OXPHOS) can diffuse the length of the tail to supply substrate for the dynein ATPases. A creatine phosphate shuttle system has been proposed that may traffic ATP from mitochondria through the tail, but experimental evidence supporting the existence of this mechanism is still lacking. Alternatively, glycolytic enzymes are localized to the principal piece of the sperm tail, thus providing a source of ATP at the site in which it will be used. In most species there is evidence that both pathways are active in sperm cells; however, the relative importance of each pathway differs among species. For example, OXPHOS-derived ATP supports bull and ram sperm motility, while mouse sperm has a definitive need for glycolytic generation of ATP for motility. Surprisingly, sperm produced by Chdh<sup>−/−</sup> mice are characterized by a reduction in both mitochondrial oxygen consumption rates – an indication of OXPHOS activity – and glycolytic rates indicating that loss of CHDH function perturbs the energy homeostasis in these cells, thus resulting in an overall decrease in ATP concentration. It is possible that the betaine molecule itself plays an important role in maintaining testicular and sperm cell function and, in particular, spermatic ATP concentrations. Dietary betaine supplementation of Chdh<sup>−/−</sup> male mice resulted in total restoration of ATP levels in sperm and an increase in progressive motility of these cells. Betaine is an organic osmolyte used by cells for protection during times of osmotic stress. Sperm mature as they move from the lumen of the testis and through the caput, corpus and cauda regions of the epididymis. During transit, sperm accumulate molecules found within the epididymal environment including organic osmolytes such as glycerophosphocholine and carnitine. We measured betaine concentrations in the mouse epididymis and found levels 10 times greater than those measured in liver. The epididymal environment is relatively hyperosmotic (∼340 mmol/kg). In comparison, the osmolality of unliquified whole semen and of the fluid in the female reproductive tract is approximately 276–302 mmol/kg, indicating that epididymal sperm experience an osmotic “challenge" within the male urethra. An inability to regulate volume in response to the varied osmotic environments would render sperm susceptible to swelling which can impair motility. Evidence exists that links betaine to sperm motility. Preserving normal motility characteristics in sperm that have been frozen and thawed is an active area of research both in the fields of human andrology and veterinary animal husbandry. Kroskinen, et al and Sanchez-Partida et al reported increased sperm motility in thawed sperm when betaine was added to the cryopreservation media. It is hypothesized that betaine may directly interact with membrane lipids and proteins, altering the hydration status of these molecules, and thus protecting them through the freeze/thaw cycle. Changes in sperm membrane phospholipid composition may also contribute to abnormal sperm cell function. For example, Chdh<sup>−/−</sup> sperm have half as much PtdCho and GPCho (a metabolite formed from PtdCho) when compared to Chdh<sup>+/+</sup> sperm. Increased osmotic stress due to a lack of betaine may account for this via activation of phospholipase A<sub>2</sub> (PLA<sub>2</sub>). PLA<sub>2</sub>, highly expressed in sperm and activated by osmotic stress, catalyzes the hydrolysis of the fatty acid in the sn-2 position of PtdCho resulting in the release of the fatty acid and generation of lyso- PtdCho. The fatty acids in this position tend to be docosahexanoic acid (DHA) and arachidonic acid (AA). All of these molecules have been shown to inhibit sperm motility. Sperm produced by men who were CC for rs1025689 contained less betaine, PtdCho and GPCho. rs1025689 is a synonomous SNP located in the coding region of human IL17βR. In a separate study, men who were homozygous for the minor C allele, were more likely to develop signs of liver or muscle dysfunction when ingesting a choline- deficient diet. Because the presence of this SNP does not result in an amino acid change, it most likely tags a functional haplotype within this gene. CHDH and IL17βR are situated in a head-to-head orientation on opposite strands on human chromosome 3 and mouse chromosome 14. According to available data, rs12676 and rs1025689 are not in linkage disequilibrium; however, we noted a high degree of concurrence of these SNPs within our study population. Although we are not aware of any reports of copy number variations in the choline dehydrogenase gene locus, it is also possible this occurs and should be investigated in future studies. Because they share a promoter region, it is likely that transcriptional regulation of CHDH and IL17βR are similar. For example, transcription of these genes is enhanced by estrogen; an estrogen response element is located within the shared promoter region. Aberrant expression of CHDH and IL17βR has been associated with breast cancer survival prognosis. Ours is the first report linking the function of this chromosomal region to male sperm cell function. In this study, rs12676 is the primary predictor of abnormal sperm mitochondrial morphology and ATP concentration and this is strengthened by the finding that CHDH protein expression is decreased in sperm from men who were GT or TT, and hepatocytes from individuals who were TT, for this variation. No changes in mitochondrial ultrastructure or ATP level were detected in sperm from individuals who were CC for rs1205689, but not TT for rs12676. Together, this evidence indicates that altered CHDH activity due to rs12676 genotype may be an underlying cause of iodiopathic male factor infertility in men. This is an especially interesting finding because deficits in CHDH function may be overcome by dietary supplementation with betaine. Indeed, as noted above, sperm motility and ATP concentration were improved in Chdh<sup>−/−</sup> male mice ingesting betaine-supplemented drinking water. # Supporting Information We would like to thank Ms. Tondra Belevins and Ms. Julia Loewenthal, as well as the staff of the Nutrition Research Institute and the Kannapolis, NC police department, for their assistance in completing this study. The authors would like to extend special appreciation to the study subjects for their willing participation. We thank Dr. Kerry-Ann da Costa for the information regarding SNP rs1025689. [^1]: Conceived and designed the experiments: ARJ SHZ. Performed the experiments: ARJ SL TW. Analyzed the data: ARJ JAG SHZ. Wrote the paper: ARJ SHZ. [^2]: The authors have declared that no competing interests exist.
# Introduction Idiopathic pulmonary fibrosis (IPF) is usually a lethal lung disease; patient survival after diagnosis varies around 3 years, thus being worse than that in many cancers, although similar to that in lung cancer. Indeed, IPF has been occasionally referred to as a malignant lung fibrosis. The principal underlying molecular mechanisms of IPF have remained unresolved and no effective curative pharmacological treatment is available. IPF and lung cancer share common risk factors, like smoking, and patients with IPF have been shown to be at a greater risk to develop a lung carcinoma as compared to general population. Furthermore, a tyrosine kinase inhibitor, namely nintedanib is used in the treatment of both IPF and lung cancer and moreover, myofibroblasts, are believed to act as principal pathogenetic cell types both in IPF and lung cancer. The gene expressions of IPF and lung cancer have been previously evaluated mostly with microarray studies performed on RNA isolated from whole lung tissues. A few studies have used cultured stromal cells derived from IPF and normal lung tissues. Stromal cells from lung cancer have been investigated in only one microarray analysis. However, differences in gene expression profiles between cultured stromal cells derived from IPF and lung adenocarcinoma (ADC) have not been investigated previously in the same study. Comparing IPF with lung cancer may provide clues to develop new therapeutic strategies for both diseases. Thus, we aimed to quantify the gene expressions of fibroblasts isolated from the patients with ADC from tumor and the corresponding normal lung as well as from IPF by microarray analysis. We wanted to evaluate even slight differences in gene expression of extracellular matrix (ECM) associated factors, also called as the matrisome, between (myo)fibroblasts derived from IPF, ADC and normal lung. The expression of collagen type IV alpha 1 chain (collagen α1(IV), gene name COL4A1), periostin (PN, gene name POSTN) as well as matrix metalloproteinase-1 (MMP-1, gene name MMP1) and matrix metalloproteinase-3 (MMP-3, gene name MMP3) were further studied with quantitative real-time reverse transcriptase polymerase chain reaction (qRT- PCR). Furthermore, the cell-specific expression of all of the above-mentioned ECM associated proteins was analyzed by immunohistochemistry (IHC). # Materials and methods ## Patients, primary lung fibroblasts and lung tissue samples Stromal cells were cultured from tumor and corresponding tumor-free peripheral lung of the ADC patients undergoing lung cancer-resection surgery (n = 4), and from surgical lung biopsy samples of the patients with IPF (n = 3) in Oulu University Hospital. In addition, one IPF sample was derived from peripheral lung outside the tumor from a patient operated for lung cancer (n = 1). Control samples were derived from histologically normal lung tissues from non-smoking patients being operated for lung cancer. Tissue samples were processed, and cells were cultured as described previously. Briefly, the cells were cultured in medium consisting of Minimum essential medium Eagle α modification (Sigma-Aldrich, Steinheim, Germany) supplemented with 13% heat-inactivated fetal bovine serum (FBS-Good, Pan Biotech, Aidenbach, Germany), 2 mM L-glutamine, 100 U/ml penicillin, 0.1 g/l streptomycin, 2.5 mg/l amphotericin B and 10 mM HEPES (all from Sigma-Aldrich). These cell lines are composed of both fibroblasts and myofibroblasts as previously described in our electron microscopic analyses. Cells were passaged at near confluency and used for experiments in passages 2–6. In the microarray and qRT-PCR analysis, cells were plated at density of 2000 cells /cm<sup>2</sup> and cultured for 72 hours. For IHC, lung tissue samples were obtained from the same patients whose cells were used in microarray-analysis. In addition, lung tissue samples were analyzed from surgical lung biopsies of 10 IPF patients as well as from 10 lung ADC patients undergoing cancer-resection surgery. Eight out of 9 normal peripheral lung samples were taken from the same patients who provided the cancer samples. Thus, 14 IPF, 14 ADC and 13 normal control samples were analyzed by IHC. The Ethical Committee of Northern Ostrobothnia Hospital District in Oulu gave a favorable statement of the study protocol (64/2001, amendments 68/2005, 2/2008, 12/2014, 1/2015, 2/2018). All the patients gave their written informed consent. Paraffin embedded tissue samples have been approved for research use by National Supervisory Authority for Welfare and Health (reg. nr. 7323/05.01.00.06/2009 and 863/04/047/08). ## Microarray analysis Total RNA was extracted using the RNeasy Mini Kit (Qiagen, Hilden, Germany) according to the manufacturer’s instructions and the concentration was measured using the NanoDrop spectrophotometry system (Thermo Fisher Scientific, Vilnius, Lithuania). The quality of RNA was analyzed with a Qiaxcel electrophoresis system (Qiagen). Microarray experiment was performed in Biocenter Oulu Sequencing Center core facility. Biotinylated cRNA were prepared from total RNA by using a 3’IVT Express Kit (Affymetrix Inc., Santa Clara, CA) according to the manufacturer’s instructions. After labeling, cRNA was hybridized to Affymetrix human hgu133Aplus2 chips. After hybridization, the microarray chip was washed and stained on an Affymetrix GeneChip Fluidics Station 450, according to the manufacturer’s instructions. The chips were then scanned using the Affymetrix GeneChip Scanner 3000 7G. An analysis of the microarray expression data was done using the R/Bioconductor through a graphical user interface, Chipster (3.12.5, CSC, Finland, <http://chipster.csc.fi/>). Normalization was performed using a custom chip description file for hgu133Aplus2. Intensity data were log<sub>2</sub>-transformed, and quantile normalized using robust multi-array average (RMA). Statistical analysis was performed with Chipster using empirical Bayes. Benjamini-Hochberg was used as multiple correction for false discovery rate. Lists of differentially expressed genes between different groups were generated using a log<sub>2</sub> fold change lower than -1 or higher than 1. Microarray data has been deposited in the Gene Expression Omnibus repository (<http://www.ncbi.nlm.nih.gov/geo>) with accession number GSE144338. Differentially expressed genes were annotated using Matrisome annotator. With this tool each entry is annotated as being or not being part of the matrisome and will be tagged with matrisome division (core matrisome vs matrisome- associated) and category (ECM glycoproteins, collagens, proteoglycans, ECM- affiliated proteins, ECM regulators, or secreted factors). ## Quantitative real-time reverse transcriptase polymerase chain reaction Quantitative RT-PCR was performed to confirm the expression of differentially expressed genes in the same primary lung fibroblast cells as used in the microarray analysis. Total RNA was extracted as described above in the Microarray analysis experiments. One-μg aliquots of RNA were reverse-transcribed using RevertAid First Strand cDNA Synthesis Kit (Thermo Fisher Scientific). PCR amplification was performed in triplicate as previously described, and the threshold cycle values were averaged. Reactions were performed by using iQTM SYBR Green Supermix (Bio-Rad Laboratories, Inc., USA). Non-template controls were included for each gene. Samples were processed for qRT-PCR using CFX Connect<sup>TM</sup> Real-Time PCR Detection System (Bio-Rad Laboratories) in the following conditions: 95°C for 3 min, 40 cycles at 95°C for 10 s, annealing phase (temperature specific for each primer pair, see) for 10 s and 72°C for 15 s, and a final extension phase of 72°C for 2 min. The melt curve was created in the following way: 81 cycles from 55°C with 0.5 degree increments every 5 s to 95°C. Relative gene expressions were quantified by using the 2<sup>-ΔΔCT</sup> Livak method. Gene expression levels were normalized to glyceraldehyde 3-phosphate dehydrogenase (GAPDH) or to hydroxymethylbilane synthase (HMBS) to confirm that the data are reproducible. The normalized values were compared to the average ΔC<sub>T</sub> value of the control cell lines to calculate the fold changes. Two independent experiments were performed, and fold changes of each cell line were averaged. ## Immunohistochemistry IHC stainings were performed in serial tissue sections. The antibodies used in this study are listed in. Formalin-fixed and paraffin-embedded lung specimens were cut into 4-μm sections, de-paraffinized in xylene and rehydrated in a descending ethanol series. After microwave- or enzyme-stimulated antigen retrieval endogenous peroxidase was blocked with aqueous 0.3% H<sub>2</sub>O<sub>2</sub> (Peroxidase-Blocking Solution, Dako, Glostrup, Denmark) for 10 min. Stainings were performed using Envision+ System Kit (Dako) with DAB 3,3’ diaminobenzidine chromogen. Counterstaining was performed with Mayer’s hematoxylin (Sigma-Aldrich). Phosphate-buffered saline, mouse and rabbit isotype controls (Invitrogen, Carlsbad, USA) were used as negative controls. In order to identify the phenotype of the cells expressing collagen α1(IV), PN, MMP-1 and MMP-3, all cases were also studied for the markers of macrophages and monocyte lineage cells (cluster of differentiation 68, CD68), type II pneumocytes (thyroid transcription factor 1, TTF-1), endothelial cells (CD31) and myofibroblasts (alpha-smooth muscle actin, α-SMA). ## Scoring of the immunoreactivity The extent of the immunoreactivity for collagen α1(IV), PN, MMP-1 and MMP-3 was scored as negative or positive (+) in stromal cells with widened alveolar tips, fibroblast foci and tumor stroma. Additionally, immunoreactivity was evaluated in alveolar epithelium, smooth muscle cells, alveolar macrophages, endothelial cells, epithelial cells of the bronchiole and cancer cells. ## Statistical analyses Statistical analyses for qRT-PCR were performed by Statistical Package for the Social Sciences (SPSS; version 25.0, Chigago, IL, USA) using Mann Whitney U test when comparing IPF and control samples and Wilcoxon Signed Rank Test when comparing ADC and control samples. Values of P\<0.05 were considered as significant. # Results ## The different mRNA expression of several ECM genes in stromal cells of IPF, ADC and normal lung Using an Affymetrix platform (U133Aplus2), we determined gene expression levels in fibroblasts prepared from lung tissue of 4 patients with IPF, 4 patients with ADC as well as in 4 control samples consisting of histologically normal looking lung collected outside of the tumor. We observed modest differences between groups, but due to the small sample size, statistical significances were not achieved. With our selection criteria (log<sub>2</sub> fold change lower than -1 or higher than 1) 36 genes between primary fibroblasts derived from ADC and normal lung, 157 genes between IPF and normal lung and 152 genes between IPF and ADC were differentially expressed (– Tables). Fourteen genes (LINC01116, GPNMB, CSTA, S100A4, CHI3L1, TIMP3, VAMP8, HOXC6, MEOX2, CD36, POSTN, CRLF1, SERPINF1, ST6GALNAC5) were up-regulated in both IPF and ADC and six genes (CHRDL1, TRIM55, ACAN, CHRM2, PCOLCE2, RARRES1) were down-regulated in both IPF and ADC in comparison to control. Forty matrisome genes were altered in IPF and 15 in ADC as compared to control. Matrisome genes annotated with Matrisome Annotator are shown in and Tables. Of the matrisome genes COL4A1 was down-regulated in IPF compared to ADC and control, POSTN was up-regulated in IPF and ADC compared to control, MMP1 and MMP3 were up-regulated in IPF compared to ADC and control, and because of their differential expression profiles they were chosen for the present research. Microarray results were confirmed by qRT-PCR on selected ECM genes COL4A1, POSTN, MMP1 and MMP3. These factors were differentially expressed between IPF, ADC and normal lung and they might have a role in fibrosis. The mRNA levels of COL4A1, POSTN, MMP1 and MMP3 normalized to those of GAPDH or HMBS showed a trend in the direction of a change observed in the microarray results although statistical significance was not achieved due to the small sample size. As compared to controls, COL4A1 mRNA expressions were 41% in IPF and 85% in ADC when those of POSTN in IPF and ADC were 308% and 191%, respectively. MMP1 and MMP3 expressions in IPF were 238% and 221%, respectively, and in ADC 84% and 90%, respectively, compared to control. ## Collagen α1(IV), PN, MMP-1 and MMP-3 were localized in stromal cells of IPF and ADC in immunohistochemistry IHC was used to confirm the expression of collagen α1(IV), PN, MMP-1 and MMP-3 at the protein level in stromal cells within lung tissues. In particular, we evaluated the expression of these ECM proteins in fibroblast foci of IPF, tumor stroma of ADC, and widened alveolar tips of normal lung, since these areas contain abundant fibroblasts and myofibroblast. Widened alveolar tips were defined as widened endings of free interalveolar septa as previously described. shows the IHC expression of markers of macrophages and monocyte lineage cells (CD68), type II pneumocytes (TTF-1), endothelial cells (CD31) and myofibroblasts (α-SMA). Strong expression of collagen α1(IV) was observed extracellularly within stromal cells of the widened alveolar tips in control lung. Collagen α1(IV) was also localized in the basement membranes of alveolar epithelium and endothelium, smooth muscle cells and in a few cases, a weak expression was observed in bronchiolar epithelial cells and alveolar macrophages. In IPF, collagen α1(IV) was expressed extracellularly in the stromal cells of fibroblast foci, while a very weak immunoreactivity was observed in the hyperplastic alveolar epithelium lining fibroblast foci. A strong collagen α1(IV) expression was observed within stromal cells in ADC while a weak expression was detected in the cancer cells in a few cases. PN was observed extracellularly in some widened alveolar tips of normal control lungs, within alveolar epithelial cells and the bronchiolar basement membrane zone, but not in bronchiolar epithelial cells, smooth muscle cells or alveolar macrophages. In IPF, the strongest expression of PN was observed in stromal cells within fibroblast foci. In ADC, PN was observed in tumor stroma, but not in the cancer cells. A very weak extracellular MMP-1 immunoreactivity was observed occasionally in widened alveolar tips and smooth muscle cells in controls when a strong MMP-1 expression was detected in the alveolar macrophages, some alveolar epithelial cells, endothelial cells, and bronchiolar epithelium. Some stromal cells within fibroblast foci were positive for MMP-1 in IPF with the strongest MMP-1 expression being observed in the hyperplastic alveolar epithelial cells lining fibroblast foci. Both cancer cells and stromal cells were positive for MMP-1 in ADC. Almost negative expression of MMP-3 was detected in some of the widened alveolar tips within normal control lung. MMP-3 was mainly expressed in alveolar macrophages and monocyte lineage cells, and bronchiolar epithelial cells. Weak MMP-3 expression was observed in some alveolar epithelial cells and smooth muscle cells. Some stromal cells of fibroblast foci were positive for MMP-3 in IPF and in addition, hyperplastic alveolar epithelial cells were positive. Both cancer cells and stromal cells were mainly positive for MMP-3 in ADC. # Discussion We investigated mRNA expressions in stromal cell lines derived from IPF, ADC and normal lung by microarray analysis and we observed that 20 genes were similarly up- or down-regulated in IPF and ADC as compared to control, whereas most of the altered genes in IPF and ADC were different, including several ECM genes. We were particularly interested in the ECM associated genes and selected COL4A1, POSTN, MMP1 and MMP3 since they have been previously shown to be associated with pulmonary fibrosis. Furthermore, most of the previous studies have focused on epithelial cells, and not on stromal cells as investigated here. We concentrated on those genes that revealed a differential expression between IPF and lung cancer, such as higher expression of MMP1 and MMP3 in IPF than ADC, and lower expression of COL4A1 in IPF than in ADC. Because the expression of these factors in lung cancer and IPF has not been previously compared in the same study, we believe that this kind of protocol may provide novel information. POSTN was chosen as an example of a gene being equally highly expressed both in IPF and lung ADC. Previous studies have revealed differences in gene expression between IPF and lung cancer. A study comparing gene expression in lung tissues by microarray analysis from five patients with both lung cancer and IPF revealed five genes e.g. SMAD4, P21, MT1A, MMP7 and TIMP1; these were down-regulated in cancer tissue in comparison to IPF, in at least two of the patients. Cancer associated lung fibroblasts have been previously compared to paired normal fibroblasts; in that study, 46 differentially expressed genes were identified. Some of these genes e.g. CHI3L1, ST6GALNAC5, COL11A1, MFAP5, TNFSF4 were identified also in our analysis as differentially expressed genes between ADC and control or between IPF and control suggesting that at least some disease related changes in the transcriptome are maintained during in vitro culture. A previous study identified 178 differentially expressed genes in stromal cells derived from IPF and normal lung. Fourteen of these genes were also differentially expressed in our comparative analysis between IPF and control, and two of these genes (S100A4 and POSTN) were also differentially expressed between ADC and controls. Another study identified 547 genes which were differentially expressed in IPF compared to controls. Thirty-nine of these genes were also differentially expressed in our microarray analysis, but not all in the same direction since only 14 genes were similarly up- or down-regulated in IPF when compared to our study. Rodriguez et al. did not identify any statistically significant gene expression differences between cultured fibroblasts derived from either IPF or normal lung, but they observed that the gene expression of cultured fibroblasts differed from that of non-cultured, freshly isolated fibroblasts. Our study identified a down-regulation of COL4A1 in IPF derived stromal cells in comparison to control cells and ADC. This finding is different from published microarray-based and RNA sequencing studies, in which COL4A1 was found as an up-regulated gene in fibroblasts and tissues derived from IPF patients as compared to control. Collagen IV expression has been shown to localize in the fibroblast foci and early fibrotic lesions of IPF, findings which are supported by our IHC analysis which detected intense collagen α1(IV) immunoreactivity in the surrounding spindle shaped cells within fibroblast foci. COL4A1 expression in lung cancer has been previously detected in stromal fibroblasts surrounding tumor cells by mRNA in situ hybridization and discontinuous collagen α1(IV) protein expression has been detected around well-differentiated clusters in ADCs, results which are similar to the present findings i.e. strong collagen α1(IV) immunoreactivity in the stroma of ADC. Based on the results of our study one can speculate that the (myo)fibroblasts might have different abilities to produce collagen α1(IV) in IPF and ADC. PN has been previously studied in both IPF and lung cancer, and POSTN up- regulation in both cultured and non-cultured stromal cells derived from IPF as compared to normal lung fibroblasts has been reported. The gene expression level of POSTN has been shown to be higher also in IPF lung tissues as compared to normal lungs. We confirmed the previous findings showing that PN mRNA and protein were up-regulated in both IPF and ADC derived stromal cells. Similarly to our results, PN immunoreactivity has been observed in areas of active fibrosis and fibroblast foci in IPF. POSTN gene expression has been claimed to be up-regulated in lung cancer tissues compared to normal lung tissue, and it has been detected by IHC, in lung cancer stroma but not in cancer cells, in support of our findings. However, it is still a matter of debate whether the source of PN is cancer cells or stromal cells surrounding cancer tissues. Some investigators have claimed that the mRNA expression of POSTN occurs only in stromal cells, or in cancer cells. In lung cancer, POSTN mRNA expression has been detected mainly in the stromal cells surrounding the cancer cells, whereas very little expression was found in the cancer cells themselves. We observed that MMP1 was up-regulated in stromal cells from IPF compared to ADC, which is a novel finding. MMP1 gene expression has also been previously shown to be up-regulated in non-cultured fibroblasts derived from IPF lung. Several microarray-based studies have revealed that MMP1 gene expression was up-regulated in IPF in comparison to normal lung tissues. MMP1 gene expression has been shown to be up-regulated also in ADC and squamous cells carcinoma tissues as compared to normal lung. Previous IHC studies have shown that MMP-1 was expressed mainly in epithelial cells and alveolar macrophages in IPF. In turn, we observed MMP-1 immunoreactivity also in stromal cells in IPF and ADC, which results were in line with previous reports. Several microarray-based studies have reported up-regulated MMP3 gene expression lung tissues derived from IPF patients. In normal lung, MMP-3 was mainly expressed by alveolar macrophages while in IPF, it was expressed in macrophages, epithelial cells, intravascular leukocytes and fibroblasts. Our microarray analysis did not identify any differences in MMP3 gene expression between stromal cells derived from ADC and normal lung, at odds with a previous study in which an increased MMP3 gene expression in ADC and squamous cells carcinoma was observed. We observed also MMP-3 immunoreactivity in both tumor cells and stromal cells whereas previous studies have shown stronger MMP-3 expression in tumor cells than in stromal cells in ADC. The phenotypic changes in the fibroblasts during their passage in culture and differences in the culture methods might have affected the results as was previously speculated. Certain disease specific alterations in gene expressions are not maintained during cell culture conditions and the disappearance of these differences could be enhanced if cell culture conditions such as cell confluency are not controlled. Therefore, we adopted stringently controlled conditions from the very start of the sample collection. The most serious limitation of the study was the small number of cell lines examined in the microarray study. It is notable, however, that the numbers of cell lines in the previous published studies have also been limited. A recent single-cell RNA-sequencing study showed that human lung fibroblasts are a very heterogeneous population. The heterogeneity of the original cell populations before culture may also explain the differences between microarray studies. Despite the similarities found in gene expressions of IPF and lung ADC, there were also several differences, suggesting that the molecular changes occurring in these two lung illnesses are at least partly different. The comparison of IPF and lung cancer may, however, reveal new information about the pathogenesis of these severe diseases, which may help to develop new therapies for patients. # Supporting information The authors thank Riitta Vuento for her technical assistance. [^1]: I have read the journal’s policy and the authors of this manuscript have the following competing interests: R.K. has received a congress travel stipend from Orion Pharma, a consulting fee from Boehringer-Ingelheim, and lecture fees from Roche and Boehringer-Ingelheim. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: Current address: Research Unit of Biomedicine, Department of Pharmacology and Toxicology, University of Oulu, Oulu, Finland
# Introduction Considered the master regulatory cytokine of vasculogenesis and angiogenesis, vascular endothelial growth factor (VEGF) is evolutionarily highly conserved and identified throughout all developmental stages. The mammalian VEGF family of ligands, VEGF-A, VEGF-B, VEGF-C, VEGF-D and placental growth factor (PlGF), bind to homodimers or heterodimers of the transmembrane tyrosine kinase receptors VEGFR1 (Flt-1), VEGFR2 (Flk/KDR), or VEGFR3 (Flt-4). In mice, VEGF expression peaks embryonically in the yolk sac and embryo and steadily declines in all organs in adults. The loss of a single allele leads to in utero lethality between embryonic days 11 and 12. In contrast, VEGF expression in sheep jejunum is elevated in term animals compared to fetal stages, suggesting a greater role during postnatal development. Complex regulation of vasculogenesis and angiogenesis occurs through alternative splicing of VEGF ligands and receptors, producing pro-angiogenic and anti-angiogenic isoforms that are implicated in a host of healthy and diseased states. In mice, alternative splicing of VEGFR1 truncates the intracellular domain and creates a soluble receptor sFlt-1, which has a high affinity for VEGF-A, thereby reducing its bioavailability. VEGF signaling biodiversity leads to complex regulation of not only vasculogenesis and angiogenesis, but cell proliferation, migration, survival and permeability. VEGF regulates branching morphogenesis in mammalian vasculature, neurons, lung and pancreas epithelium. In human and mouse, VEGF-C activates quiescent neural stem cells through VEGFR3 to enter the cell cycle and generate progenitor cells. Additionally, VEGF-A influences differentiation of mesenchymal stem cells into osteoblasts and adipocytes by regulating the levels of the osteoblast and adipocyte transcription factors Runx2 and PPARγ, respectively. These observations suggest that VEGF has a crucial role in regulation of stem and progenitor cell populations, independent of vasculogenesis. The presence of VEGF in the gastrointestinal system of organisms lacking vascular systems suggests that VEGF may play a crucial role in the maintenance of homeostasis in multiple organ systems, including the gastrointestinal tract. Despite a lack of endothelium and blood cells, jellyfish (podocoryne carnea) express a VEGF homolog in their gastrovascular system, suggesting a role for VEGF in the development of the gastrointestinal system. Intrinsic platelet- derived growth factor, also known as vascular endothelial growth factor-like factor (Pvf), is required for homeostasis and differentiation of intestinal stem cells in the posterior midgut of Drosophila. Hyperactivity of Pvf/Pvr drives intestinal dysplasia, supporting its role as a regulator of intestinal stem cells. The necessity of VEGF homologs in gastrointestinal development and their augmentation causing dysplasia suggest an interaction between VEGF and the intestinal stem cell niche. Although embryonic VEGF augmentation in gastrointestinal epithelium has been implicated in the development of neoplasia, the role of VEGF in postnatal small intestinal development and homeostasis is currently unknown. The goal of this study is to elucidate the role of VEGF on the postnatal intestinal stem cell niche in a murine model. Triple transgenic mice were generated with the ability to augment VEGF in epithelial cells of the gastrointestinal tract in an inducible manner via the villin promoter (VEGF-Tg) or decrease VEGF bioavailability through villin-driven overexpression of an inducible soluble VEGFR-1 (sFlt-Tg). # Methods ## Generation of Villin<sup>Cre</sup>/rtTA<sup>flox/flox</sup>/tet(o)VEGF and Villin<sup>Cre</sup>/rtTA<sup>flox/flox</sup>/tet(o)s-Flt1 (sFlt-1 mutants) mutant mice, Transgene PCR, and Tissue Collection Transgenic mice on a C57/B6 background were maintained according to the animal care facility protocols of the institution with approval by the Children's Hospital Los Angeles Institutional Animal Care and Use Committee. All mice were housed in a controlled environment in clean cages, fed mouse chow or doxycycline chow ad libitum with an unlimited source of fresh water. Tail clips were collected from mice that were P14 or older under isofluorane anesthesia and were euthanized under CO<sub>2</sub> exposure at P21. Triple transgenic Villin<sup>Cre</sup>/rtTA<sup>flox/flox</sup>/tet(o)VEGF mutant mice (VEGF mutants) or Villin<sup>Cre</sup>/rtTA<sup>flox/flox</sup>/tet(o)s-Flt1 mutant mice (sFlt-1 mutants) were established. Intestine-specific VEGF or sFlt-1 overexpression was inducible with the administration of oral doxycycline. Villin<sup>Cre</sup> mice were mated with tet(o) VEGF or tet(o) sFlt-1 mice. Those positive for both genes were crossed with homozygous rtTA<sup>flox/flox</sup> mice. After birth of a litter, the mother was fed 625 mg/kg doxycycline chow (Harlan; Cat# TD.110720) ad libitum. The doxycycline chow induced overexpression of VEGF or sFlt-1 in the pups via the mother’s breast milk. The mice were genotyped by polymerase chain reaction (PCR) at P14. The pups were placed under general anesthesia and a small tail clip was acquired. The specimen was placed into Direct PCR Tail reagent (Viagen Biotech; Cat# 102-T) with 1:100 Proteinase K Solution (Invitrogen; Cat# 25530049) and moved to the 55°C incubator overnight. The temperature was increased to 85°C in the morning for one hour, then returned to room temperature. The genotyping PCR mix consisted of 10 μL MyTaq Red Mix (Bioline; Cat# BIO-25043), 0.1 μL 100 μM forward (F) primer, 0.1 μL 100 μM reverse (R) primer (Eurofins MWG Operon), 8.8 μL RNase-free water, and 1 μL mouse DNA for a total of 20 μL PCR reaction. This was placed in a 0.2 μL PCR tube. PCR was performed in a thermocycler with the temperature recommended by the manufacturer for the MyTaq Red Mix and appropriate annealing temperatures. A 1% agarose (Denville Scientific; Cat# CA3510-8) with 1:1000 ethidium bromide (Promega Corporation; Cat# H5041) electrophoresis gel was prepared and 10 μL of mix was run at 100V for 30 minutes and visualized under a UV light. The mice that contained one or two alleles were named littermates (LM) and those containing all three alleles were termed mutants. The duodenum of the mice was harvested at P21. After euthanization of the mice in a CO<sub>2</sub> chamber, three matched 5 cm segments of proximal duodenum were harvested. Samples were placed in 10% formalin (Fisher Scientific; Cat#305–510), RNAlater (Sigma-Aldrich; R0901), or flash frozen on dry ice and placed in the -80°C freezer. ## Histology and Immunofluorescence Formalin-fixed tissue was embedded in paraffin and sectioned at 5 μm. The tissue was dehydrated in 30% EtOH for 30 minutes, 50% EtOH for 45 minutes, 70% EtOH for 45 minutes, 95% EtOH for 1 hour, 100% EtOH for 1 hour twice. The tissue was then cleared with toluene twice for 45 minutes, then in 1:1 ratio of paraffin and toluene overnight at 65°C. The following morning the tissue was placed in pure paraffin for 1 hour twice and fixed in the cassette on a cooling plate to solidify. The tissue was then sectioned at 5 μm. Frozen sections were procured for confocal microscopy to evaluate angiogenesis after intracardiac FITC-dextran labeling of the vasculature. Tissue was protected from light and placed in 4% paraformaldehyde overnight at 4°C. The following day, tissue was placed in 30% sucrose until it sank to the bottom the tube. The tissue was then transferred to a 1:1 mixture of 30% sucrose:OTC compound overnight at 4°C. Tissue was subsequently transferred to OTC compound and slowly frozen at -20°C until the OTC was frozen. Blocks were stored at -80°C until sectioning. Frozen tissue was sectioned at 60 μm immunofluorescence staining with DAPI was performed as describe below. Hematoxylin and eosin (H&E) staining was performed on the sectioned tissue. Slides were placed in Histochoice (Amresco; H103) twice for 2 minutes, then in 100% EtOH for 2 minutes. The slides were hydrated in 70% EtOH for 30 seconds, 50% EtOH for 30 seconds, 30% EtOH for 30 seconds, and H<sub>2</sub>O for 2 minutes. The slides were placed in hematoxylin for 15 seconds and H<sub>2</sub>O until clear, followed by eosin for 5 seconds for counterstain. The slides were dehydrated in 90% EtOH for 1 minute, 100% EtOH for 1 minute and Histochoice for 2 minutes. The slides were imaged at 20x magnification on a brightfield microscope. Single and multiple clusters of red blood cells (RBCs) were counted by blind observers as previously described. The villus length, crypt depth, circumference and number of crypts per intestinal length were measured by a trained blinded observer with ImageJ. Immunofluorescence was performed to locate terminal cell markers and proliferative markers. The slides were placed in Histochoice for 10 minutes twice. The slides were rehydrated by soaking them sequentially in 100% EtOH, 90% EtOH, 75% EtOH, 50% ETOH, and 30% EtOH for 5 minutes each. A low pH, citrate- based Antigen Unmasking Solution (Vector; H-3300) was used to retrieve antigens in the microwave. The microwave heated the solution for 4 minutes at 50% three times, with 30 seconds in between each heating session. The solution was cooled to room temperature for 30 minutes and washed in phosphate-buffered saline (Gibco; Cat#10010) with 0.1% Tween (Amresco; Cat# 9005-64-5) (PBS-T) for 5 minutes. Universal blocking solution (1% BSA, 0.1% cold fish skin gelatin, 0.5% Triton-X 100 and 1x PBS) with 2% goat serum (Sigma-Aldrich; Cat# G9023) was applied for 30 minutes at room temperature. Primary antibodies were diluted in universal blocking solution with 2% goat serum and placed on tissue overnight at 4°C. In the morning, tissues were washed in PBS-T for 5 minutes three times. Appropriate Cy3 or Cy5 secondary antibodies diluted in PBS with 0.05% Tween and applied to tissue for 1 hour at room temperature. The slides were again washed with PBS-T for 5 minutes three times. The sections were mounted on Vectashield with DAPI (Vector; Cat# H-1200) and visualized under the fluorescent microscope. Lysozyme was quantified by percentage of immunofluorescent-positive cells per hemi-crypt by a trained, blinded observer with ImageJ. Ki-67-positive cells were counted in the crypt and amplifying zone by the number of positive cells per position by a trained, blinded observer with ImageJ. ## Confocal Microscopy and Three-Dimensional Volume Reconstruction VEGF mutant and littermate duodenal cross sections were prepared as described above and imaged for FITC and DAPI immunofluorescence using a Zeiss LSM 710 confocal microscope. Sixty micron sections were imaged at an individual z stack thickness of 1 μm to allow for accurate three-dimensional volume reconstruction using IMARIS software. Confocal images were imported into IMARIS version 7.7.2 and volume reconstruction analysis was performed using region of interest selection of FITC-labeled vasculature of individual hemivilli. Threshold values were obtained from littermate controls using autothresholding algorithms and these values were applied to VEGF mutants to allow for proper comparison of three-dimensional volume analysis. Sixty villi in total from four individual mice per group were analyzed. ## VEGF Protein Quantification Quantification of VEGF protein was executed via enzyme-linked immunosorbent assay (ELISA) on frozen tissue sections for the VEGF mutants and littermate controls or serum from VEGF mutant OU culture. Flash frozen duodenal sections stored at -80°C were selected. The Bradford protein assay method was employed to determine protein concentration. The frozen tissue was homogenized in 200 μL of extraction buffer (1M Tris-HCl, 5M NaCl, 10% Triton, 10% NaDeoxycholate, 0.5M EDTA, 1% PMSF, 10% phosphatase inhibitor, 1% protease inhibitor and distilled water). The homogenized tissue was centrifuged at 10,000 rpm for 10 minutes at 4°C and the supernatant was transferred to a new 1.5 mL eppendorf tube. The Bradford reagent was diluted as recommended by the manufacturer. The bovine serum albumin protein standards (BioRad; Cat# 500–02) and the samples were prepared in duplicates. These were placed in disposable cuvettes and absorbance was measured at a wavelength of 595nm in a spectrophotometer (Pharmacia Ultrospec III). A standard curve was prepared and the concentrations of the total protein were determined. Mouse VEGF ELISA was performed in duplicates with the suggested protocol of the manufacturer (Signosis; EA-2401). The absorbance for the wells were read at 450 nm (PerkinElmer; Cat# Victor<sup>2</sup> 1420) and a standard curve was plotted to find the concentration of VEGF present in the tissue. The total VEGF protein concentration (pg/mL) was divided by the total protein concentration (μg/mL) to determine the amount of VEGF present per mg of total protein (pg/mg) in tissue samples. VEGF protein concentration in OU cultures was expressed as ng/ul. ## Reverse Transcription Quantitative Polymerase Chain Reaction The duodenal or OU samples placed in RNAlater and stored in -80°C were thawed for RNA extraction. Tissue homogenization was performed with a rotor-stator homogenizer (Qiagen; TissueRuptor). RNA was extracted with the RNeasy Mini Kit according to the manufacturer protocol (Qiagen; Cat# 74104) with 350 μL as the appropriate volume. The concentration and purity of RNA was determined with a microvolume spectrophotometer (NanoDrop 2000; Thermo Scientific). All samples had a 260/280 ratio of 2.0 or higher (Mean 2.1 ± 0.01), indicating high purity of the isolated RNA. Reverse transcription was performed with the cDNA kit (BIO- RAD; Cat# 170–8896) and the mixture contained 1 μl RNA in a 20 μl reaction volume. The thermo cycler (BIO-RAD; Cat#C1000) temperatures were set according to the manufacturer’s recommendations. VEGF transcript overexpression was confirmed in VEGF mutants with RT-qPCR. SYBR green was performed for the transgenic VEGF mutants. The cDNA for the VEGF mutants was employed in this reaction. The reaction contained 7.5 μl of SYBR Green I Master Mix (Roche; Cat# 0707516001), 0.2 μl Tet(o)-VEGF forward and reverse primers, 6.1 μl RNase-free water and 1 μl of cDNA. The Roche Lightcycler 480 performed RT-qPCR with 1 cycle at 95°C for 10 minutes, 45 cycles at 94°C for 1 minute followed by 55°C for 1 minute and 72°C for 1 minute each, and 1 cycle at 40°C for 30 seconds. The reactions were performed in triplicate and single outliers were removed for quantification. Roche Lightcycler 480 was used to calculate the relative gene expression of Tet(o)-VEGF related to GAPDH. sFlt-1 transcript was confirmed in sFlt-1 mutants in a similar fashion. Quantitative PCR was performed with the resulting cDNA in the Roche LightCycler 480 system with hydrolysis probes in a multiplex reaction with GAPDH as a reference gene. Each reaction contained 0.2 μl GAPDH primers (Roche; Cat# 05046211001), 0.2 μl GAPDH probe (Roche; Cat# 05046211001), 0.2 μl 100 μM left primer, 0.2μl 100 μM right primer, 0.2 μl hydrolysis probe, 7.5 μl master mix (Roche; Cat#04707494001), 5.5 μl of RNase free water and 1 μl cDNA. Gene expression of Bmi1, Lgr5, Atoh1 and Hes1 (n = 6), as well as of Sox9, DII1, Wdr43, EphB2 and Bmp4 (n = 6) was analyzed for VEGF and mutants. Similarly, gene expression of Bmi1, Lgr5, Atoh1 and Hes1 (n = 6) and Sox9, DII1, Wdr43, EphB2 and Bmp4 (n = 6) was quantitated in sFlt-1 mutants. To investigate the effect of VEGF overexpression and suppression on angiogenesis in the postnatal mouse duodenum, quantitative RT-qPCR was performed for VE-cadherin, a marker of endothelial-specific cell-cell adherence junctions. The reactions were performed in clear, 96-well plates (Roche; Cat# 05102413001) under the following conditions: 1 cycle at 95°C for 10 minutes, 45 cycles at 95°C for 10 seconds followed by 60°C for 30 seconds and 72°C for 1 second each, and 1 cycle at 40°C for 30 seconds. Each PCR reaction was run in triplicate and single outliers that occurred in the technical replicates were removed for quantification. Roche Lightcycler 480 was used to calculate the relative gene expression of the target gene over the reference gene with advanced relative quantification calculations. ## Intravascular-Immunofluorescence Labeling by Intracardiac FITC-Dextran Injections Mice were anesthetized using standard IACUC procedures. Once fully anesthetized, mice were placed supine with arms and legs extended. Silk tape was placed across the abdomen and along the upper extremities to secure the animal firmly on the working surface. The chest was shaved to allow for visualization of the rib cage. The mice were then prepped and draped in a sterile fashion. A 2mg/ml FITC- dextran (Sigma FD2000S) solution was made in PBS and protected from light. Using a 1ml insulin syringe with 26-G needle, 200 μl of air was drawn up into the barrel followed by 200 μl of FITC-dextran solution. The needle was inserted into the second intercostal space 3 mm to the left of the sternum, directing the tip into the center of the chest (45° to the right, 45° to the horizontal plane and pointing towards the right flank of the mouse) to a depth of 6 mm. Pulsatile flow of red blood into the hub of the needle indicated correct placement of the needle into the left ventricle and 200 μl of FITC-dextran solution was injected per mouse while avoiding the introduction of air bubbles into the circulatory system. The needle was quickly withdrawn and pressure placed on the chest with alcohol wipes for 1 minute to prevent extravasation into the chest. Mice were sacrificed 3 minutes after FITC-dextran injection and intestinal tissue immediately harvested and fixed in 4% paraformaldehyde. ## Mouse Enteroid Cultures We isolated mouse enteroids from VEGF mutant mice using a previously establish protocol. Briefly, crypts were released from murine small intestine by incubation for 30 min at 4°C in PBS containing 2 mM EDTA. Isolated crypts were counted and a total of 500 crypts were mixed with 50 μl of Matrigel (BD Bioscience) and plated in 24-well plates. After Matrigel polymerization occured, 500 μl of crypt culture medium (Advanced DMEM/F12 (Invitrogen) containing growth factors (10–50 ng/ml EGF (Peprotech), 500 ng/ml R-spondin 1 (R&D systems) and 100 ng/ml Noggin (Peprotech)) was added. For sorting experiments, isolated crypts were incubated in culture medium for 45 min at 37°C, followed by trituration with a glass pipette. Dissociated cells were passed through a cell strainer with a pore size of 20 μm. Growth medium was changed every 4 days. ## Mouse Organoid Units The VEGF mutant duodenal samples were put into a 10 cm Petri-Dish with cold Hanks’ balanced salt solution (Life Technologies, Inc.) to clean the luminal contents. The tissue is then cut into 2x2 mm sections in the Petri-Dish. The tissue is placed in a 50 mm centrifuge tube and centrifuged for 1 minute at 8000 rpm. The tissue is washed 3 times with cold Hanks’ salt solution, centrifuging between washes and removing the supernatant. The tissues were then digested with 800 μg/mL collagenase (Worthington) and 0.12 mg/mL dispase (Invitrogen) for 20 minutes. The digestion was stopped by adding high glucose Dulbecco’s Modified Eagle Medium (DMEM) with 4% Sorbitol (Sigma, Inc.) and 10% fetal bovine serum (FBS) (Invitrogen). The digestion was triturated with a 10 mL pipette for approximately 10 minutes. Large debris sediment was removed by centrifugation and the supernatant was transferred to a 50 mL Falcon tube. The mixture was centrifuged at 500 rpm for 10 minutes and the supernatant was discarded. The pellet was resuspended in 2 mL high glucose DMEM with 10% FBS. This solution was mixed with Matrigel (BD Bioscience) at a ratio of 1:1. 100 μL of the Matrigel- organoid unit (OU) mixture was placed in each well of a 6-well plate. The plate was incubated at 37°C for 30 minutes. 2 mL DMEM was then added to each well. The cultures were incubated overnight at 37°C with 5% CO<sub>2</sub> and medium was changed on days 1, 3, 5, 7 and 9. Experimental OU had 2 μL of doxycycline added to the medium, while controls did not. The OU were imaged on days 1, 3, 5, 7, and 9. On day 10, each well was scraped and individually collected and placed in a 1.5 mL Eppendorf tube. These tubes were centrigued at 1000 rpm for 1 minute. The media was then removed and 200 μL of RNAlater was added to the Eppendorf tube. These were stored at 4°C until RNA extraction was performed as described above. Diameter of OU were measured on ImageJ by a trained, blinded observer. # Results ## VEGF overexpression resulted in elevated levels of VEGF mRNA and protein. sFlt-1 overexpression resulted in an increased expression of sFlt-1 in enteroid culture and negative feedback reduction of VEGFR1 receptor in duodenum Following 21 days of doxycycline induction, triple transgenic VEGF mutants (VEGF-Tg) were confirmed by demonstrating overexpression of VEGF in the postnatal mutant duodenum by RT-qPCR and ELISA. VEGF mutant mice displayed a 7.36 ± 1.49 SEM-fold increase in the transgenic VEGF transcript (p = 0.02) compared to littermates. Similarly, VEGF protein concentration was significantly higher in VEGF mutant duodenum (512.51 ± 95.69 pg/mg SEM) compared to littermates (36.81 ± 4.54 pg/mg SEM). This represents a 13.92 ± 2.60 SEM-fold change (p = 0.008). Following 21 days of doxycycline induction, the triple transgenic sFlt-1 mutants (sFlt-Tg) were confirmed by demonstrating expression of sFlt-1 in the postnatal mutant duodenum and epithelial enteroid culture by RT-qPCR. At 21 days of doxycycline treatment, sFlt-1 mutants had significantly decreased sFlt-1 expression compared to littermates (1.21 ± 0.19 SEM versus 0.24 ± 0.04 SEM; p\<0.001). RT-PCR and ELISA of VEGF-A in the duodenum failed to demonstrate a significant decrease between sFlt-1 mutant duodenum (0.90 ± 0.09 SEM; p = 0.6 and -0.117 pg/mg ± 0.004 SEM; p = 0.88, respectively) versus littermates (0.99 ± 0.14 SEM and -0.123 pg/mg ± 0.003 SEM, respectively). However, sFlt-1 mutant enteroid cultures demonstrated a significant increase in sFlt-1 expression upon doxycycline treatment compared to controls (0.958 ± 0.02 SEM versus 0.746 ± 0.09 SEM; p = 0.019). Doxycycline addition in sFlt-1 enteroid culture did not alter the expression of another VEGF receptor, Flk (0.156 ± 0.02 SEM versus 0.152 ± 0.02 SEM; p = 0.85) and did not promote endothelial cell growth in culture as enteroid cultures are devoid of endothelium. ## VEGF overexpression resulted in increased villus angiogenesis and taller villi. Reduced VEGF bioavailability produced a shorter, paler intestine with a diminutive cecum and smaller villi with longer, but fewer crypts per measured length The gross appearances of the duodenum in VEGF mutants and littermates were distinct. Littermate duodenum was a pink to yellow color compared to a deep red intestine seen in VEGF mutant mice. Compared to littermates, sFlt-1 mutant mice were smaller and had decreased body mass (p\<0.001), as well as a swollen anus. The duodenum of sFlt-1 mutant mice was paler in comparison to littermates (respectively). The length of the intestine was also shorter with a diminutive cecum in sFlt-1 mutant mice. The cecum of sFlt-1 mutant mice demonstrated decreased number of goblet cells compared to littermates (, respectively). Histological sections by H&E demonstrated significantly higher villus height in VEGF mutant duodenum at 594.1 ± 19.03 μm SEM compared to littermates at 497.4 ± 32.00 μm SEM (p = 0.03). There was no significant difference in the crypt depth (87.70 ± 4.47 μm SEM in VEGF mutants versus 77.41 ± 2.92 μm SEM in littermates; p = 0.08) or in the number of crypts per length measured in VEGF mutants compared to littermates (2.56 ± 0.14 μm SEM versus 2.77 ± 0.09 μm SEM, respectively; p = 0.3). No significant difference in duodenal circumference existed between VEGF mutants and littermates (5.72 ± 0.27 mm SEM versus 5.41 ± 0.19 mm SEM, respectively; p = 0.4). Duodenum of sFlt-1 mutants displayed shorter villi compared to littermates (293.3 ± 14.4 μm SEM versus 426.0 ± 11.7 μm SEM, respectively; p\<0.0001) and taller crypt depth (83.6 ± 3.7 μm SEM versus 64.7 ± 1.7 μm, respectively; p\<0.0001). There were also significantly fewer crypts per length measured in sFlt-1 mutants compared to littermates (2.49 ± 0.02 μm SEM versus 2.86 ± 0.12 μm SEM, respectively; p = 0.03). ## VEGF augmentation in intestinal epithelium leads to increased mucosal angiogenesis and vascular permeability H&E histological cross sections demonstrated a significant increase in the amount of single and multiple clusters of RBCs within the villi of VEGF mutant mice compared to littermates (respectively). Single RBC counts (34.28 ± 3.24 SEM versus 5.98 ± 0.54 SEM; p\<0.001) and multiple RBC clusters (9.13 ± 0.97 SEM versus 1.29 ± 0.21 SEM; p\<0.001) per villus were significantly increased in VEGF mutants compared to littermates, suggestive of increased vascular leak and angiogenesis. To further evaluate vascular leak and angiogenesis, intracardiac FITC-dextran injections were performed to label the vasculature and 60 μm cross sections were imaged using confocal microscopy (, respectively). Three- dimensional volume reconstruction demonstrated significant increase in intravillus FITC-dextran volume in VEGF mutants as compared to littermates (2259000 μm<sup>3</sup> ± 54200 SEM versus 1689083 μm<sup>3</sup> ± 75245 SEM; p\<0.001). ## VE-cadherin expression significantly increased in VEGF mutants and decreased in sFlt-1 mutants Given the distinct increase in intravascular volume and vascular leak in the hemivillus of VEGF mutants, we further explored the effects on endothelial cells in VEGF and sFlt-1 mutants by performing RT-qPCR for VE-cadherin, an endothelial cell marker. VEGF mutant mice showed a significant increase in VE-cadherin expression compared to littermates (p = 0.04), with a 16.5-fold increase in expression. In contrast, the littermates of sFlt-1 mutants demonstrated significantly increased expression of VE-cadherin compared to the sFlt-1 mutant mice (p = 0.01), with a 22.4-fold increase in expression in littermates compared to sFlt-1 mutants. ## VEGF mutants have increased Ki-67-positive cells per hemivillus extending higher into the transit-amplifying zone, whereas sFlt-1 mutants display Ki-67-positive cells residing lower in the crypt Duodenal sections of VEGF mutants displayed more Ki-67-positive cells per hemivillus compared to littermates. In the VEGF mutants, the Ki-67-positive cells extended to a higher position in the transit-amplifying zone compared to littermates. sFlt-1 mutants exhibited fewer Ki-67-positive cells per hemivillus compared to littermates. Ki-67-positive cells were identified lower in the crypts of sFlt-1 mutants compared to littermates. Caspase 3 staining revealed no significant difference between VEGF mutants, sFlt-1 mutants and their comparable littermates. ## sFlt-1 mutants demonstrated an increase in lysozyme-positive Paneth cell population within the intestinal crypts Immunofluorescence of duodenal sections revealed significantly more lysozyme- positive cells per hemicrypt in sFlt-1 mutants compared to littermates. The percentage of lysozyme-positive cells per hemicrypt was 5.70 ± 0.70% SEM for sFlt-1 mutants compared to 1.036 ± 0.05% SEM (p = 0.0005) for littermates. No significant difference in lysozyme staining was appreciated in VEGF mutants compared to littermates (2.14 ± 0.11% SEM versus 1.99 ± 0.29% SEM; p = 0.48). ## VEGF mutants demonstrate a significant decrease in expression of the intestinal stem cell marker Lgr5. sFlt-1 mutants demonstrate reduced Bmi1 and Wdr43 expression and a concomitant increase in EphB2 and Sox9 expression RT-qPCR was performed for stem/progenitor and early differentiation markers in VEGF mutant and sFlt-1 mutant mice. These genes included Bmi1, Lgr5, Atoh1, Hes1, Sox9, DII1, Wdr43, EphB2, and Bmp4. Bmi1 is a marker for slow cycling intestinal stem cells, typically located at the +4 position. Lgr5, or leucine- rich repeat-containing G-protein coupled receptor 5, is a marker for rapid cycling intestinal stem cells, also known as crypt base columnar cells (CBCs). Atoh1 marks progenitor cells of the secretory lineage, while Hes1 is expressed in progenitor cells of the absorptive lineage. Dll1 is a marker for secretory precursors. Sox9 marks CBCs when expressed at low levels or Paneth cells when expressed at high levels. Wdr43 is a marker for transit-amplifying cells. EphB2 marks differentiating cells at low levels of expression and intestinal stem cells and progenitor cells at high levels of expression. Bmp4 inhibits de novo crypt formation and is expressed in the intravillus mesenchyme. VEGF mutants demonstrated a 0.55 ± 0.12 SEM-fold reduction in Lgr5 expression compared to littermates (p = 0.04). There was no significant change in other stem/progenitor cell marker gene expression in VEGF mutants compared to littermates. sFlt-1 mutants exhibited a 0.42 ± 0.19 SEM-fold reduction in Bmi1 expression (p = 0.03) and a 0.60 ± 0.09 SEM-fold reduction in Wdr43 expression (p = 0.004) compared to littermates. EphB2 and Sox9 expression were increased 2.47 ± 0.24 SEM fold (p = 0.02) and 1.50 ± 0.09 SEM fold (p = 0.004) respectively in sFlt-1 mutants compared to littermates. ## VEGF mutant organoid units (OU) exposed to doxycycline are significantly larger than untreated controls. VEGF augmentation in murine OU increases expression of stem/progenitor cell markers Bmi1 and Atoh1 and decreases EphB2 expression To further isolate the effect of VEGF on the epithelium, we generated OU cultures from intestinal tissue and induced VEGF overexpression by exposure to doxycycline. OU contain both epithelium and mesenchyme, but lack a vascular supply. To evaluate the effect of VEGF overexpression on OU in vitro, VEGF mutant duodenum OU were cultured and the diameters were measured every other day for 10 days before tissue harvest. OU treated with doxycycline or medium alone grew over the 10-day period. The doxycycline-treated OU were significantly larger on day 5 compared to control (164.36 ± 14.43 μm versus 126.11 ± 6.57 μm; p = 0.04). This difference did not persist after day 5. These findings correlated with a temporal increase in VEGF protein expression by ELISA over a period of 5 days in culture. Doxycycline treated OUs from C57/B6 mice demonstrated no significant size difference over 10 days. After 5 and 10 days of in vitro culture with or without doxycycline, expression of stem cell markers was evaluated in VEGF mutant OU. At 5 days, a significant increase in Bmi1 (1.14 ± 0.13 versus 0.96 ± 0.13; p = 0.03) and Atoh1 (2.54 ± 1.07 versus 1.38 ± 0.60; p = 0.04) expression and decrease in EphB2 (0.68 ± 0.22 versus 1.11 ± 0.07; p = 0.001) expression was observed in doxycycline-treated VEGF mutant OU compared to controls. No significant difference in the expression of Lgr5, Bmi1, Sox9, Atoh1, Dll 1, Hes1, Wdr43, EphB2, or BMP4 was identified between doxycycline- treated VEGF mutant OU compared to controls at 10 days. # Discussion VEGF overexpression and reduced bioavailability had distinct effects on postnatal small intestine in a murine model. VEGF is excreted in breast milk and decreased in the intestines of formula-fed murine and human neonates that succumb to necrotizing enterocolitis. Mesenchymal-driven sFlt-1 mice demonstrate significant changes in body and organ size at 21 days. Given these findings and the dramatic phenotype demonstrated in the Villin-Cre VEGF and sFlt-1 mutant mice at the end of the weaning period, we examined mice at 21 days to address a physiologically important time period by which VEGF regulation may have the greatest impact on postnatal gastrointestinal development. VEGF augmentation, as demonstrated by increased VEGF expression by RT-PCR and ELISA, resulted in a deep red color of the intestine, which has been previously reported in transgenic mice that overexpressed VEGF via the villin promoter. The deeper red color is due to increased vascular leak and angiogenesis. VEGF mutant villi demonstrated a significant increase in single RBCs seen outside of the villus vasculature on H&E suggestive of increased vascular permeability or leak , which has been reported in other models overexpressing VEGF. Three-dimensional volume reconstruction of FITC-dextran labeled vasculature demonstrated increased FITC- dextran volume within VEGF mutant villi supporting enhanced angiogenesis and vascular leak. Taken together with increased VE-cadherin expression by RT-PCR, there is a notable increase in vascular permeability and angiogenesis in VEGF mutants. Although previous studies of VEGF augmentation in the small intestine during embryonic development resulted in the development of epithelial cysts within the crypts and an increased frequency of intestinal adenomas, we did not appreciate similar results in our postnatal model. However, postnatal VEGF overexpression in the colon demonstrated epithelial cysts throughout. The discordance between the phenotype seen in small intestine versus colon in our model may result from changes in VEGFR expression levels during postnatal development. We confirmed that in mice, VEGFR2 expression by immunofluorescence is greater in colonic epithelium compared to the small intestine. As a result, we speculate that VEGF augmentation may play a more significant role in the development of intestinal neoplasia and alteration of stem/progenitor cell populations during the postnatal period in the colon versus the small intestine. Our inducible VEGF model may be an important tool for further studies to investigate possible mechanisms by which VEGF promotes tumorgenesis in the adult colon. Decreased VEGF-A bioavailability was achieved by overexpressing sFlt-1 in the brush border of intestinal epithelium using a doxycycline-inducible tet(O) expression of sFlt-1 and rtTA expression of Villin-Cre in enterocytes. The sFlt-1 mutants were significantly smaller in size, which has been reported in transgenic mice that reduce VEGF bioavailability in the gastrointestinal tract by overexpressing sFlt-1 in the mesenchyme. Villin sFlt-1 mutant mice developed a shorter intestine and underdeveloped cecum. Further histological evaluation of the cecum revealed a significant decrease in the number of goblet cells; however, we did not appreciate a notable difference in goblet cells within the small intestine. VEGF induces Dll4 expression in endothelial tip cells. Dll1 and Dll4 are expressed in distinct patterns along the crypt-villus axis of the gastrointestinal tract with significantly less Dll1 expression and greater Dll4 expression in colonic crypts compared to small intestine. Loss of Notch signaling within intestinal epithelium via Dll1 and Dll4 blockade result in complete conversion of proliferating progenitors into goblet cells with a concurrent loss of intestinal stem and progenitor cells within the crypts of small intestine and colon. It is therefore surprising that we would see a loss of goblet cells within the cecum of sFlt-1 mutant mice and no appreciable effect in goblet cell populations within the small intestine. These findings could implicate changes in notch signaling ligand and receptor expression within intestinal stem/progenitor cells that are reflective of the postnatal period compared to embryonic development. RT-PCR expression of sFlt-1 was significantly decreased in sFlt-1 mutant duodenum after 21 days of doxycycline treatment compared to littermates with no significant change in VEGF-A transcript or protein levels. These data suggest a compensatory negative feedback relationship wherein chronic sFlt-1 overexpression decreases VEGFR1 expression, as the RT-PCR primer used can amplify both sFlt-1 and full length Flt-1. VEGF-A levels within the native duodenum were significantly low, which is likely why we were unable to see a significant reduction in VEGF-A by ELISA. Therefore, to confirm that sFlt-1 overexpression within the intestinal epithelium does indeed occur in our transgenic model, enteroid cultures were created from sFlt-1 mutant mice. After administration of doxycycline, a significant increase in sFlt-1 expression was demonstrated without altering expression of KDR, another VEGF receptor. Taken together, the molecular and phenotypic evidence supports that villin-positive cell overexpression of sFlt-1 leads to reduced VEGF-A bioavailability in sFlt-1 mutants, similar to our previously described mesenchymal-expressed sFlt-1 murine model. The murine postnatal period is important for the maturation of the intestine as upward migration of the crypt-villus axis promotes mature crypts formation in intervillus pockets, giving rise to intestinal stem cells. Histological analysis of VEGF mutants revealed increased villus height, but no change in crypt depth. In contrast, sFlt-1 mutants exhibited decreased villus height with longer crypts. In swine, increased vascular flow rate and resistance is crucial for normal development of villi during the postnatal period. The discordance in angiogenesis seen between VEGF and sFlt-1 mutant mice and its inherent effects on vascular flow and resistance might account for the significant changes in villus and crypt architecture; however, there is likely a direct effect of VEGF and additional angiogenic/vasculogenic signaling molecules on epithelial and other cell types. Platelet-derived growth factor (PDGF) acts with some coordination with VEGF in angiogenesis and contributes to proper crypt morphology and its inhibition leads to fewer and misshapen crypts. VEGF-A binds to PDGF receptors and induces signaling in mesenchymal cells. Therefore, the increased villus height in VEGF mutants may be occurring through PDGF receptor signaling in the small intestine mesenchyme, aiding in the formation and elongation of villi. This is in contrast to sFlt-1 mutant mice, which demonstrated shorter villi with longer crypts compared to littermates. As the crypt-villus axis migrates upward during crypt formation, the villi are shorter, but the crypts elongate, suggesting that VEGF could play a role in crypt-villus axis development and migration by directly regulating epithelium gene expression or indirectly through the mesenchyme. VEGF augmentation and reduction had antagonistic effects on the number and location of proliferating cells compared to littermate controls. In VEGF mutants, Ki-67-positive cells were greater in number and extended higher into the transit-amplifying zone compared to littermates and was not influenced by any appreciable induction of apoptosis via caspase 3. Absorptive and secretory progenitor cells typically reside in this region, though no change in Hes1 or Atoh1 RNA expression was identified. Therefore, Ki-67-positive cells may be proliferating under the direction of secondary mediators of VEGF signaling, such as the Notch signaling ligands Dll1 and Dll4 that are required for maintenance and of intestinal stem/progenitor cells. Moreover, increased proliferation in VEGF mutants could be due to a reduction in epithelial cell migration along the crypt-villus axis. In contrast, the sFlt-1 mutants displayed decreased proliferation in the transit-amplifying zone and an increase in the number of Paneth cells (Figs). Similar results of reduced body weight, shorter villus height and decreased proliferation of the transit-amplifying cells with an increase in Paneth cells have been identified in Krüppel-like factor 9 (Klf9)-deficient mice. Several signaling molecules involved in vasculogenesis were downregulated in Klf9-deficient mice, indicating the complex interplay between the mesenchyme and its effects on migration and proliferation of the vasculature, epithelium and intestinal stem/progenitor cell populations. These data also suggest a potential supportive role of angiogenic development and proper maintenance of intestinal stem/progenitor cell progenitor homeostasis. RT-qPCR identified differences in intestinal stem cell gene expression in VEGF and sFlt-1 mutants. Several experimental and biological limitations are encountered when performing gene analysis by RT-qPCR, including a discordance between transcript expression and protein translation. Therefore, downstream effectors with significant changes in expression still need to be identified, but the effect at the RNA transcription level suggests an important role of VEGF signaling in the maintenance and regulation of the intestinal stem cell niche. Lgr5 cells are rapidly cycling and sensitive to irradiation, while Bmi1 cells are slow-cycling and more quiescent. VEGF mutants had decreased Lgr5 expression, but no change in expression of Bmi1. Lgr5 reduction in the VEGF mutants did not alter the differentiated cell types in the intestinal epithelium, possibly due to replacement by differentiating Bmi1 cells which can compensate for the loss of Lgr5-positive cells. In contrast, sFlt-1 mutants had no change in Lgr5 expression, but a 0.42-fold decrease in Bmi1 expression compared to littermates. The sFlt-1 mutants display a number of characteristics that can be partially explained by decreased Bmi1 expression. Bmi1 knockout mice have shorter small intestines, similar to that in our sFlt-1 mutant mice. In sFlt-1 mutants, decreased proliferation per hemicrypt is observed. Similarly, the expression of Wdr43, a marker of transit-amplifying cells, was 0.60-fold less in sFlt-1 mutants than littermates. This decrease is consistent with the lower percentage of Ki-67-positive cells identified in the transit-amplifying zone of the crypt. While Bmi1 expression was decreased in the sFlt-1 mutants, the expression of Lgr5 was unaffected. VEGF-induced expression of Dll4 in vascular endothelium leads to the activation of Notch signaling. With reduced bioavailability of VEGF-A, Notch signaling potentially decreases, resulting in decreased Bmi1 and Wdr43 expression with a subsequent reduction in proliferation within the crypts of the small intestine. VEGF/neuropilin-2 signaling as been shown to repress insulin-like growth factor-1 receptor expression through Bmi-1. Elevated insulin receptor B (IR-B) levels in intestinal epithelial stem cells decreased proliferation in the crypts and enhanced epithelial barrier function. VEGF reduction could result in decreased expression of Bmi-1 and a subsequent increase in IR-B receptor signaling, leading to decreased intestinal stem cell proliferation and subsequent increase in Paneth cell differentiation within the crypts. The sFlt-1 mutants were found to have a 1.5-fold increase in Sox9 expression. Sox9 expression inhibits proliferation in vivo, which may have contributed to the decrease in the number of proliferating cells per hemicrypt, resulting in decreased villus height. Highly expressed in Paneth cells, Sox9 controls an early step in Paneth cell differentiation through interactions with the Wnt signaling pathway and affects the overall phenotype of the small intestine crypts. Consistent with the increase in Sox9 expression, there was an increase in the percentage of lysozyme-positive cells per hemivillus in the sFlt-1 mutants. Insulin-like growth factor 1 has been shown to enhance crypt regeneration and increase the percentage of intestinal stem cells in S-phase without expanding the population. Thus, IGF signaling primes intestinal stem cells to differentiate during conditions of cell loss or injury. Moreover, SOX9 has also been shown to be a transcriptional regulator of insulin-like growth factor-binding protein 4 (IGFBP-4), which is expressed highly in Paneth cells. SOX9-induced activation of IGFBP-4 is directly involved in the antiproliferative effects seen on intestinal stem/progenitor cells and may explain the phenotype we observe in sFlt-1 mutants as loss of VEGF signaling may induce increased expression of IGF-IR through Bmi1, thereby promoting Paneth cell differentiation and intestinal crypt regeneration. The sFlt-1 mutants also demonstrated a significantly increased expression of EphB2. EphB2 expression is under control of the Wnt pathway and is usually highly expressed in intestinal stem cells, with decreased amounts in more differentiated cells. Increased EphB2 is associated with increased proliferation and promotion of cell cycle re-entry. It is also found to be increased in wound healing in inflamed intestines. The increased expression of Sox9 and EphB2 and the decrease in proliferating cells is consistent with the ability of these progenitor cells to self-renew and differentiate into diverse cell types. In both VEGF and sFlt-1 mutant mice, there was no significant difference in expression of Atoh1 and Hes1, which mark progenitor cells in the transit- amplifying zone and regulate differentiation into secretory cell fates. Atoh1 is important in differentiation of intestinal stem cells into secretory cell fates, whereas Hes1 has been shown to promote proliferation and inhibit secretory cell development. We did not identify notable differences in the ability of VEGF or sFlt-1 mutants to differentiate into secretory cells lineages in the small intestine. To isolate the effects of VEGF signaling on epithelial cells without confounding interactions from native intestinal vasculature, OU were obtained from VEGF mutants and grown in culture over a period of 10 days. OU are comprised of epithelium and mesenchyme and are devoid of a vascular supply. We have demonstrated expression of VEGFR2 (Flk) within the epithelium of OU in vitro, suggesting that VEGF may have direct signaling effects on intestinal epithelium. Additionally, Caco-2 human intestinal epithelial cells are known to express VEGF receptors. Although VEGF mutant OU exposed to doxycycline were larger in diameter at day 5 and demonstrated a significant increase in VEGF protein expression, differences in OU diameter were not maintained at day 10 in vitro. To identify if the discordance in size at day 5 resulted from changes in stem/progenitor cell populations, we examined several stem cell markers by RT- PCR. Increased expression of Bmi1 and Atoh1 occurred, whereas EphB2 was downregulated in doxycycline-treated VEGF OU compared to controls at 5 days. Significant changes in stem/progenitor cell marker expression were lost by 10 days in vitro. Bmi1-positive slow-cycling ISCs have the capacity to self- renew, proliferate, and give rise to all the differentiated epithelial cell lineages of the small intestine. EphB2 expression within the intestinal crypt is highly expressed in ISCs and decreases as cells proliferate and contribute to the transit-amplifying zone. In VEGF OU, upregulation of Bmi1 and concurrent downregulation of EphB2 suggest that VEGF stimulates ISCs to proliferate and migrate within the transit-amplifying zone, supportive of increased Ki67-positive cells within the transit-amplifying zone of VEGF mutant mice at 21 days. Atoh1 upregulation suggests that VEGF may prime progenitor cells within the transit-amplifying zone to preferentially differentiate into cells of the secretory lineage. Our in vitro data suggest that the increase in epithelial cell proliferation seen in the transit-amplifying zone of VEGF mutant duodenum could represent a direct effect by VEGF on epithelial stem/progenitor cells, independent of effects from the native vasculature. However, the notable discordance between stem/progenitor cell gene expression between VEGF mutant duodenum at 21 days and VEGF OU at 5 days may occur primarily due to acute versus chronic VEGF exposure or secondary effects from native intestinal vasculature in vivo. Further exploration using OU culture as a system to discern between primary effects of VEGF on the intestinal epithelium and contributory secondary effects mediated through the native vasculature at earlier timepoints is needed. Furthermore, VEGF and sFlt-1 mutant mice demonstrate distinct phenotypic differences in stem/progenitor cells and terminally differentiated cell types between the small intestine and colon at the end of the weaning period. Elucidation of the mechanisms that account for differences seen between small intestine and colon using our inducible in vivo and OU culture models will provide a useful strategy to identify regulators of intestinal stem cell/progenitor cell maintenance, homeostasis and differentiation in discrete areas of the gastrointestinal system. Alterations in VEGF bioavailability had unique effects on postnatal small intestinal development in a murine model. Overexpression of VEGF led to increased angiogenesis and vascular permeability as well as increased proliferation higher into the transit-amplifying zone. In vivo, VEGF augmentation was found to decrease expression of Lgr5 without affecting expression of Bmi1. In OU culture, VEGF augmentation led to increased expression of Bmi1 and Atoh1 with a reciprocal downregulation of EphB2, suggesting primary effects of VEGF on intestinal epithelial stem/progenitor cell maintenance and homeostasis. In vivo, sequestration of VEGF caused inhibited vasculogenesis and restricted proliferation of intestinal epithelial cells within the crypts. Decreased availability of VEGF led to increased expression of Sox9 and EphB2, as well as a concomitant decrease in expression of Bmi1 and Wdr43. Future studies using concurrent in vivo and OU culture models to explore the role of VEGF signaling during early postnatal development may provide a useful platform to distinguish between the direct and indirect mechanisms by which VEGF bioavailability affects intestinal stem/progenitor cell populations and alters postnatal intestinal development. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: TCG. Performed the experiments: CRS SGM KAH XH KLF. Analyzed the data: CRS SGM KAH. Contributed reagents/materials/analysis tools: TCG. Wrote the paper: CRS TCG.
# Introduction Flicker Modulation Thresholds (FMT) can provide a sensitive measure of changes in the temporal responses of the visual system. Flickering stimuli generate trains of neural modulation and this load accounts for an increased demand on the blood supply and overall metabolic activity in the retina, when compared to static stimuli. Flicker modulation thresholds can, therefore, act as a useful photoreceptor-specific functional biomarker in early detection of diseases such as Age related Macular Degeneration (AMD) and Diabetic Retinopathy, where hypoxia plays an important role in disease pathogenesis. For the same reason, these thresholds can also be useful in assessing the severity of rod and cone dysfunction. For instance, flicker modulation thresholds measured at a single temporal frequency such as at 14 Hz is considered to be a sensitive indicator of cone function in patients with AMD, with regards to ease of testing and repeatability. It has also been shown that the differences in cone FMT between age-matched healthy individuals and patients with AMD are largest in the central 5° of fixation, which is within the region of maximum vulnerability to damage in these patients. The absolute rod detection threshold, which is typically measured after a period of dark adaptation (20–30 minutes), provides a useful measure of rod function. For rod-mediated vision, this test should also overcome the requirement of long dark-adaptation times that make the test results more variable and unattractive for use in a clinical setting. Given the previous literature, it is desirable to develop an efficient test to measure rod- and cone-specific sensitivities within the central vision and to account for the effects of normal, healthy ageing. Such data are needed to detect loss of sensitivity that falls outside the normal limits as one might expect in patients at risk of developing and quantify the earliest losses in patients with AMD or diabetes. It is well established that in normal, healthy ageing, photopic flicker modulation thresholds increase with increasing age, although this increase only starts to become significant beyond 40 to 45 years of age. An earlier study reported little or no change in absolute scotopic thresholds with age. However, the small sample size (n = 23) and the maximum age of the participants in that study (the oldest subject was 61 years of age), might have precluded the authors from establishing the true variation of scotopic thresholds with age. Subsequent studies reported greater increases in scotopic thresholds with increasing age. In particular, Jackson et al. found that the rate of age-related increase of scotopic thresholds (-0.08 log units/decade) is faster than that measured for photopic thresholds (-0.04 log units/decade). However, it is important to note that the previous studies tested absolute scotopic thresholds using a static target after a prolonged period of dark adaptation (20–30 minutes). In the current study, we have employed a rod / cone flicker test, which makes use of stimulus parameters such as background luminance, spectral composition, temporal modulation frequency and stimulus size to test for rod—and cone—specific responses at five discrete locations in central vision (5°), with minimal adaptation time. The findings from this study characterize the existing variability within rod- and cone-mediated thresholds in central and near peripheral vision and establish the upper, threshold limits one can expect in healthy, normal ageing (20–75 years). The latter can be used to detect when rod- or cone-specific thresholds exceed normal age-matched limits as a result of subclinical disease. In addition, the rod / cone flicker modulation threshold test can also be used to monitor the progress of a disease or the effectiveness of treatment. # Materials and methods ## Participants The study participants had normal vision with high-contrast logMAR visual acuity (VA) equal to better than 0.0 (20/20 Snellen fraction) and no history of ocular and systemic disease. Participants were recruited from the staff and student cohort at the L V Prasad Eye Institute (LVPEI), Hyderabad, India and from those patients who visited the institute for a regular eye examination, with no specific complaints. All the participants were of Indian ethnicity. The participants underwent ocular examination including, slit lamp biomicroscopy and fundus examination to rule out any ocular pathology. Intraocular pressure was measured using Goldmann applanation tonometry and was also within normal limits. Participants with any ophthalmic pathology or with lens status (nuclear, cortical or posterior subcapsular) graded as LOCS III, or higher were excluded. In addition, participants who were unable to complete successfully the ‘learning’ mode of the test were also excluded. The ethics and protocol of the study was approved by the Institutional Review Board of LVPEI and all the procedures were conducted in accordance with the tenets of the Declaration of Helsinki. Written informed consent was obtained from all the participants before they took part in the study. A sample of approximately 150 participants could be considered as a representative of the population to derive a normative database for visual functions. ## Flicker-Plus Flicker modulation thresholds (FMTs) were measured using the Flicker-plus module of the commercially available Advanced Vision and Optometric Tests (AVOT) developed at City, University of London. The AVOT tests run on a desktop computer with two displays, one for the experimenter and the other for generating the visual stimuli for a number of advanced vision tests. The display used by the experimenter to run the tests and the calibrated stimulus display viewed by the participant are separated by a black curtain, such that the participant can only see the stimulus display. The display monitor specifications were as follows: 24" IPS (in-plane switching) LCD monitor (EIZO, Model ColorEdge CS2420; EIZO corporation, Japan) using a 10-bit graphics card with 1920 X 1200 resolution at a frame rate of 60 Hz. The calibration of the monitor is performed using a photometer (Mavo-Monitor USB, Gossen, Germany) and a custom-built software (LUMCAL; City Occupational, Ltd., London, UK). The stimulus was a uniform disc of appropriate size (see for details) and was presented either at the centre (0°) of the display (as indicated by the four guides shown), or in one of the four quadrants (5° eccentricity from fixation) namely supero-temporal (ST), supero-nasal (SN), and inferonasal (IN) and inferotemporal (IT) quadrant, depending on the eye that was tested. The duration, temporal frequency, background luminance and spectral power distribution of the background and target were selected to produce either cone- or rod-enhanced stimuli. The chromaticities and other parameters were different between rod and cone—enhanced stimuli. However, in each of these conditions, the target and background had the same spectral composition. The differences in the spatiotemporal properties between rod and cones were exploited in this test such that the rod-enhanced condition had stimuli with relatively longer display duration, lower temporal frequency and light levels, all relative to the cone- enhanced conditions. The stimulus parameters for the two conditions were selected to produce approximately the same flicker thresholds for both rods and cones. Preliminary results in CNG-B3 patient with retinal achromatopsia confirm that the cone-enhanced thresholds can still be detected by rods, but with thresholds approaching the maximum that can be generated on the visual display (currently in preparation). Five, randomly interleaved 2-down 1-up adaptive staircases, one for each stimulus location, were used to measure the corresponding FMTs. During each test, the five interleaved staircases determine either the subject’s rod- or cone-mediated thresholds. Each threshold was estimated by averaging the last six reversals. The implementation of the 5AFC staircase with variable step sizes and the 2-down / 1-up approach reduces the chance probability of a correct response to 1/25. The measured thresholds correspond to just over 70% probability of a correct response. The participant viewed the stimulus display monocularly from 1m. Only one eye (dominant eye or eye with a visual acuity of at least 0.0 logMAR or better) per participant was tested. ## Procedure The Flicker-Plus test includes a ‘learning’ option which is used to familiarize the participant with the bespoke, numeric keypad buttons needed to indicate the position of the stimulus on display during the test procedure. The test stimulus was always preceded by a briefly presented fixation stimulus at the centre of the screen. The latter consisted of a dark square outline and a cross, designed to attract the participant’s point of regard. The participant was instructed to maintain their fixation on the location of this marker throughout the test to ensure that the peripherally presented stimuli were seen at 5° eccentricity. The raised, keypad buttons mirrored the five locations of the stimulus on display to make the task easier. Each participant was expected to achieve 100% correct responses in the learning mode before the full test was administered. The order of the rod and cone flicker testing was randomized. The rod-enhanced condition required the participant to adapt to the background field for 1½ minutes, wearing spectrally-calibrated, 1.0 log unit neutral density filter (Oakley Half Jacket 2.0 –Black Iridium, USA). Each test run took \~10 minutes to complete and breaks were rarely required. Pupil size was not measured during the experiment as the flicker modulation at equal luminance remains largely independent of pupil size. Previous research has also shown that small pupil size changes have only minimal effects on FMTs measured at photopic and high mesopic light levels. Pre-retinal filters such as the short-wavelength absorption of light by the lens and the macular pigment do not affect the stimulus modulation depth when the spectral power distribution of the test remains unchanged and the same as that of the adjacent uniform background. ## Data analysis The data were not normally distributed as tested by the D’Agostino & Pearson test (p \<0.05); therefore non-parametric statistical analyses were carried out. The age of 45 years was chosen as the demarcation point between the two groups based on previous literature for further analyses. We have examined the results using the point of separation as a free parameter in order to determine whether there was a justification for using a fixed cut-off point at 45 years. We examined all test conditions and found that the mean transition age was 46.5 ± 1.0 years. In view of this observation, the data were fitted with two separate linear functions for age group ≤ 45 years and for group \> 45 years using a linear regression model in GraphPad Prism (Graph Pad Software, Inc., CA, USA). Spearman correlation coefficient was computed to analyze how FMTs vary as a function of age separately for the two age groups (≤ 45 years and \> 45 years). # Results As mentioned in the methods, we aimed to recruit 150 subjects. However, we ended up recruiting a total of 145 participants (82 males and 63 females). Out of 145, five participants were excluded, as they had difficulties with the ‘learning mode’ and were unable to either see the flickering stimulus or to complete the test. Therefore, a total of 140 participants completed the study. The mean age (SD) of these participants was 41.81 (15.11) years. Nearly 4% (5/140) of the study participants were pseudophakic in the tested eye. The number of participants across each decade who participated in the study is shown in. Mann -Whitney test showed no significant differences between male and female groups for central (p = 0.80) and parafoveal cone FMT (p = 0.17). Similarly, Mann -Whitney test showed no significant differences between male and female groups for central (p = 0.40) and parafoveal rod FMT (p = 0.78). revealed no noticeable differences across the four parafoveal locations for cone thresholds and well as for rod thresholds. Therefore, the cone and rod thresholds in the four parafoveal locations were averaged separately, and only the mean values were used for each participant. The slopes of the regression lines between FMT and age were compared to determine if they were different from zero (i.e. if there was an age-related trend in the data) and if they were different from each other (i.e. if the rate of change in FMT’s were different across different test conditions). A p-value of ≤ 0.05 was the requirement for statistical significance. Except for cone thresholds (central and parafoveal) in the younger age group and parafoveal rod thresholds in the younger age group, all the other slopes were significantly different from zero (p \< 0.05). The rate of increase in central cone thresholds (i.e., 2.9% per decade–) above 45 years of age is not significantly different to the rate of increase in central rod thresholds (i.e., 4.5% per decade—; p = 0.15; See) for the same age group. The increase in parafoveal cone thresholds (i.e., 1.6% per decade–) above 45 years is also not significantly different than the measured increase in parafoveal rod thresholds (i.e., 2.4% per decade–; p = 0.19; See) for the same age group. The slopes and the y-intercepts of the regression lines were also compared separately as a function of stimulus location (between centre and parafovea) in cone and rod conditions. In the younger age group (≤ 45 years) cone threshold slopes and y-intercepts between the centre (0.4% per decade) and parafovea thresholds (0.05% per decade) were not significantly different from each other (p = 0.24 and p = 0.36 respectively). In the elder age group (\> 45 years), there are no significant differences in the slopes for the cone thresholds between central (2.9% per decade) and parafoveal locations (1.6% per decade) (p = 0.10). There was, however, a significant difference between the y—intercepts of central and parafoveal cone FMT (p = 0.004;) in the age group \> 45 years. Similarly, in the younger age group, the slope for rod thresholds in centre (0.7% per decade) was not significantly different from the rod thresholds measured at the parafoveal locations (0.2% per decade) (p = 0.15). The y-intercepts were, however, significantly different for rod thresholds between central and parafoveal locations (p \< 0.001) in the younger age group. In the elder age group, the rate of change in central rod thresholds (4.5% per decade) was significantly greater (p = 0.03) than that measured in parafoveal vision (2.4% per decade). The y-intercepts could not be determined as the slopes differed significantly. There is no significant correlation between flicker (rod and cone) thresholds and age up to the age of 45 years. However, beyond 45 years of age, there is significant correlation for both rod and cone FMTs with age (p \< 0.05) as shown in. Approximately 89% (125/140) of the participants showed higher rod FMTs compared to cone FMTs both centrally and parafoveally. The linear regression was performed between central cone and rod thresholds (central) for both age groups (≤ 45 years and \> 45 years) separately. The slopes (≤ 45 years: y = 0.60x+4.59, R<sup>2</sup> = 0.14, \> 45 years: y = 0.92x + 5.31, R<sup>2</sup> = 0.50; p = 0.12) were not significantly different from each other. However, parafoveally slopes were significantly different from each other (≤ 45 years: y = 0.44x+3.64, R<sup>2</sup> = 0.16; \> 45 years: y = 0.85x+3.00; R<sup>2</sup> = 0.45; p = 0.007). # Discussion This study reports measured flicker modulation thresholds as a function of age at five locations in central vision with stimulus conditions that favour either rods or cones. The new test employed in the study benefits from minimum times of adaption to the background luminance and the use of a statistically-efficient, 5-alternative, forced choice procedure to measure flicker modulation thresholds that are not affected significantly by residual refractive errors, higher order aberrations or moderate levels of scattered light The results show almost constant rod- and cone-mediated flicker sensitivities up to 45 years of age and a more rapid decline above this age. The age and stimulus location dependence of rod and cone data is discussed separately in the following sections. ## Cone flicker thresholds as a function of age and test location Our findings show that 15Hz cone FMTs measured at the point of regard remain mostly unchanged up to 45 years, followed by a steeper increase above this age. This finding is in agreement with earlier findings, which suggest that the variation of foveal flicker thresholds (\< 20 Hz) across the life span has three phases namely, the initial decrease in thresholds up to 16 years of age, the stable phase between 16yrs until 45–50 years and an increase beyond 50 years of age. It is well established that ageing causes significant changes in the optics of the eye as well as structural/anatomical changes in the retina, RPE and the choroid and neural changes in the post-receptoral pathways. Here we explore the possible factors that contribute towards the increase in cone FMT with increasing age. Firstly, the steady-state pupil size decreases with age, and this causes a reduction in retinal illuminance, which in turn can potentially affect FMT. Previous studies have, however, shown that the increase in FMTs in the elderly is higher than predicted by age-related miosis. Besides, the Flicker-plus test was designed to ensure that there is no change in the mean luminance level during the stimulus presentation. Both E<sub>b</sub> (background retinal illuminance) and E<sub>s</sub> (stimulus modulation retinal illuminance) are both directly proportional to pupil size. Therefore, pupil size affects the corresponding retinal illuminances but does not affect the temporal contrast modulation of the stimulus or FMT (i.e., the measurement variable). Similarly, this finding of increased FMT with ageing is unlikely to be associated with changes in cone density, as the reported decrease in cone numbers in the central ± 5° of the retina is small. Factors such as the decrease in ganglion cell density, especially parasol ganglion cells in the magnocellular (MC) pathway with increasing age may be attributed to the increase in FMT. This is supported by a previous finding that in patients with severe parvocellular lesions, contrast thresholds for temporal frequencies (1–32 Hz) is largely mediated by the MC pathway. However, the effect of ageing-related structural changes in neurons in the human post-receptoral pathways is unclear and is limited to animal studies. Previously, in monkeys and cats, it has been shown that cortical neurons’ (V1 area) sensitivity and the signal-to-noise ratio is decreased significantly in older age group compared to the younger age group. These results suggest that similar ageing changes in human cortical neurons are conceivable and which could account for an increase in FMT. Besides, the estimated sampling efficiency related to contrast threshold task in older human subjects is significantly poorer than younger subjects, which may also indicate a neural basis for loss in contrast sensitivity with increasing age. Although the results show that the rate of increase in cone thresholds with increasing age was comparable between central and parafoveal locations (p = 0.10), it is important to explore the possible reasons for the rate of changes in cone FMT. This finding of steeper slopes for central cone thresholds is in agreement with the results of a recent study, carried out under similar test stimulus conditions, where foveal photopic thresholds had steeper slopes for those over \> 40 years of age, compared to parafoveal locations (4°). Other studies failed to find a significant increase in thresholds in cone-mediated central vision (0°) with increasing age. The differences could be due to different stimulus parameters such as adapting luminance, size, temporal frequency and presentation time. More specifically, in this study, the size of the central flicker stimulus (30’) was smaller compared to the peripheral stimulus (60’) to account for spatial summation. ## Rod flicker thresholds as a function of age and test location In a way similar to cone-mediated thresholds rod thresholds (central and parafoveal) are significantly higher for age group \> 45 years compared to the other group (≤ 45 years) (Figs) and also significant differences are noted between rod thresholds in central and parafoveal locations. The former findings are in agreement with the previous studies that have studied rod sensitivity as a function of age. Rod sensitivity corrected for the expected absorption of light by the lens in the older group (mean age—70 years), show at least 0.4 log unit loss in sensitivity compared to the younger group (mean age—25 years). Similarly, Jackson and Owsley showed that the loss in scotopic sensitivity is 0.08 log units per decade. The loss in rod photoreceptors with ageing is well documented, however, that may not fully account for the loss in rod sensitivity with ageing and differences in sensitivity between central and parafoveal locations. It is because the latter shows poor correlation with the magnitude of loss in rod density at different retinal eccentricities. The neural basis for the increase in rod thresholds with ageing could be post-receptoral in origin, such as the steady loss of ganglion cells with increasing age. The decrease in scotopic pattern ERG, dark-adapted scotopic b-wave and VEP amplitudes associated with ageing, suggest the role of post-receptoral or post retinal circuitry in decreased contrast sensitivity with ageing. Also, the role of cortical neurons in the increase of rod FMT cannot be discounted as discussed earlier in the manuscript. ## Comparing rod and cone flicker modulation thresholds Although not statistically significant, the rate of increase in central rod FMTs for the age group \> 45 years was greater than central cone FMT (p = 0.15). This trend is consistent with the findings of Jackson et al., who also showed the rate of increase in scotopic thresholds (-0.08 log units per decade) was greater than photopic thresholds (-0.04 log units per decade) with increasing age in normal healthy participants. However, it is be noted that Jackson et al. used a linear model to describe their photopic and scotopic sensitivity data for the age range between \~ 20–80 years, contrary to the bilinear fit we used. The following potential differences between these studies may account for the discrepancies. The first major difference involves the state of background adaptation of rod photoreceptors. The earlier study measured rod thresholds when the retina was fully dark-adapted (40 minutes), whilst in the present study, we measured rod thresholds after adaptation to a uniform background of luminance (0.5 cd/m<sup>2</sup>) for a brief duration (90sec). The other differences include stimulus specifications such as briefly presented stimulus and not flickering stimuli, area of the retina (18°) tested compared to 5° (this study) and stimulus duration (200 ms vs 600 ms). The bilinear fit suggests that effects of ageing in a flicker test start to manifest after the age of 45 years unlike the functions such as photopic and scotopic flicker sensitivity measured using pulsed pedestal stimuli and mechanisms responsible for the detection of flicker detection and pulse stimuli are different. However, we do see an agreement with a couple of studies with regards to the bilinear model, used to describe the photopic flicker thresholds as a function of age. Besides, approximately 89% (125/140) of participants in this study showed higher rod FMTs compared to cone FMTs. This finding is similar to Jackson et al., who reported that 80% of the normal healthy participants had higher scotopic thresholds than photopic thresholds with increasing age. The finding of approximately similar rate of change for rod versus cone thresholds beyond 45 years of age (indicated by slopes), accompanied by markedly different slopes for each of these thresholds as a function of age suggests mediation of two different mechanisms or photoreceptors classes that may be responsible for detection these stimuli. We are aware of a couple of study limitations. One is that we had a limited number of participants \> 70 years and other was that testing was done in an Indian population, whose average life expectancy is expected to be lower than that of North American and European population. Therefore the results above 70 years of age rely largely on extrapolation. The results obtained in this study are more likely to reflect the true changes in flicker sensitivity because of the normal physiological ageing processes in the retina, as mentioned earlier in the manuscript. There is no obvious reason to assume that this ageing process would be different in the Indian population when compared to other populations. # Conclusions Cone-and rod-mediated FMTs remain relatively unchanged from 20 to about 45 years of age. Beyond this age, both rod and cone thresholds increase rapidly. The dataset obtained in this study describes the effects of normal ageing and will be a useful reference in the detection of rod and cone losses in preclinical disease, for the monitoring of disease progression and also help evaluate drug efficacy in patients with wet AMD, undergoing active treatment such as the use of the anti-vascular endothelial growth (VEGF) injections. We would like to acknowledge Hyderabad Eye Research Foundation, for providing infrastructure support to set up the AVOT unit. We would also like to acknowledge Ms Gayatri Yadav for assistance in data collection for this study. [^1]: AH and SRB received support in the form of salaries from the Hyderabad Eye Research Foundation. AS, JB, and JRES received support in the form of salaries from City University of London. JRES received support in the form of a salary from QinetiQ. The authors would like to declare the following patents/patent applications associated with this research: EU- EP08762401.1; US- US8,322,857: 'Vision testing apparatus and method.' This does not alter our adherence to PLOS ONE policies on sharing data and materials.
# Introduction The increasing likelihood of extreme climate events with ongoing climate change is expected to have major impacts on biodiversity at local scales. Extreme climate events will primarily consist of periods of heat, cold, drought and flooding with greater severity and less predictability than historical norms. These events will act as disturbance and are likely to decrease biodiversity at local as well as regional scales. Floods are projected to increase with global warming in the 21st century leading to rapid changes in soil conditions thereby detrimentally affecting soil microorganisms by limiting soil gas diffusion and oxygen availability thus reducing soil nutrient availability, mineralization and decomposition of dead organic material. As a consequence, anaerobic conditions develop quickly in flooded soils resulting in marked changes in soil chemistry including the accumulation of toxic substances. All these changes are likely to significantly affect the composition of soil food webs. To explore the effects of flooding on soil food webs and, more specifically, soil microarthropods, we investigated a severe natural flooding event in a grassland plant diversity experiment. The flooding, resulting from heavy summer precipitation, was accompanied by the input of sediments rich in nutrients and was associated by an unexpectedly fast recovery of soil microorganisms within three months. Fungal biomass increased, reflecting elevated availability of dead plant biomass. Further, flooding was associated with enhanced plant community productivity but decreased stability, particularly of plant communities of high diversity. Until today understanding how short-term disturbance events affect soil biodiversity is limited but is important as changes in soil biodiversity and community structure impact the functioning of soils. A major component of soil animal communities are microarthropods reaching high density and diversity in any kind of soil, and playing a crucial role in driving belowground ecosystem processes such as decomposition and nutrient cycling. Microarthropods, such as Acari and Collembola, are major animal groups interacting with soil microorganisms. Microarthropod species are likely to be differentially affected by changes in environmental conditions such as inundation events depending on physiological adaptations and life history traits. Although Collembola (Insecta) and Oribatida (Acari) are often grouped into the same trophic level and are considered to occupy similar niches in decomposition processes, the two groups differ in a variety of ecological traits including mobility, reproduction, level of predation pressure, and tolerance to abiotic conditions. Parthenogenesis may facilitate quick population establishment after disturbances and is most widespread in Oribatida. However, the general life-history traits of Oribatida have been considered typical of K-selected species, whereas Collembola species exhibit wider variation in life-history traits. In particular, compared to Collembola, Oribatida species are less mobile, characterized by low reproductive rates and recolonize disturbed habitats slowly. Collembola, in contrast, are more sensitive than Oribatida to abiotic microhabitat conditions and recolonize disturbed habitats more quickly. Other Acari such as Astigmata have short developmental time and excellent dispersal ability. They feed on fungi or bacteria, but may also consume plant tissue. Gamasida are mostly free-living predators but also parasites or symbionts. Prostigmata are predators, herbivores and parasites. In this study, we focus on the response of Acari and Collembola to flooding in grasslands of varying plant diversity. We expected the density and richness of Collembola and Acari communities to be reduced strongly by flooding with Collembola recovering faster than Acari due to higher reproductive potential and dispersal ability. We further expected that surface-living Collembola species with high dispersal ability will recover faster than species living deeper in soil. Among Acari we expected Astigmata, Prostigmata and Gamasida to recover faster than Oribatida due to generally faster reproductive cycles. We further expected the immediate effects of flooding to be similar in both Collembola and Acari and to be independent of plant species diversity. However, we expected the recovery to be facilitated by high plant species diversity in particular in Collembola. Collembola density and diversity have been shown to benefit from plant diversity due to increased root and microbial biomass, and elevated quantity and quality of plant residues serving as food resources. # Material and methods ## Experimental setup The Jena Experiment is a semi-natural temperate grassland on the floodplain of the Saale River close to the city of Jena (50°55´ N, 11°35´ E; Thuringia, Germany). Mean annual air temperature is 9.9°C and mean annual precipitation is 610 mm (1980–2010). The study site, a Eutric Fluvisol, has been used as an arable field for over 40 years before the experiment was established with typical Central European hay meadow plants in 2002. The experiment comprises 80 5 x 6 m plots arranged in 4 blocks to control for changes in soil texture with distance from the river. A gradient of plant species richness (1, 2, 4, 8, 16 and 60) and plant functional group richness (1, 2, 3 and 4) was established. Plant species are grouped according to the morphological, phenological and physiological traits into grasses (16 species), small herbs (12 species), tall herbs (20 species) and legumes (12 species). The established grassland is mown twice a year and weeded three times per year. No permission was needed to take the samples from this site. The site is rented and managed by the project and for taking samples for analysing soil arthropods from arable systems no permission from legal bodies is needed. The field studies did not involve endangered or protected species. ## Flooding The June 2013 flood in the Upper Danube Basin was one of the largest floods in the past two centuries. Rainfall in May 2013 in southeast Germany was exceptionally high. In Jena it amounted to approximatively 150 mm. High rainfall resulted in the flooding of the Saale River with the flood also covering the Jena Experiment field site and lasting for 25 days (30 May to 24 June). Flooding caused anaerobic soil conditions with redox potentials ranging from -121 to 193 mV in some plots. Water coverage was measured daily for each plot from 31 May to 24 June and ascribed to 5 levels: 0, 25, 50, 75 and 100% (percentage of the plot covered by water). Flooding severity was measured using a flooding index calculated as the sum of daily percentages for the whole flooding period (24 days). After the flood in August 2013, dead material, target species, weeds and bare ground percentage of the plot was measured. In general, we found that 78% of the plots was covered by target species, 14% of weeds, 8% of dead material and 23% of bare ground. Monocultures had only 41% of target species and 50% of bare ground. In contrast, plots with 16 plant species had 88% of the plot covered by target species and 7% covered by bare ground. In October the vegetation was recovered totally. ## Soil biota In November 2010, July 2013 (three weeks after the flood) and in October 2013 (three months after the flood), soil cores of 5 cm diameter and 5 cm depth were taken from each plot using a stainless steel corer (80 samples per date). Soil microarthropod species were extracted using a high-gradient extractor, increasing the temperature gradually from 25 to 55°C during 14 days. The animals were collected in mono-ethyleneglycol and transferred into 70% ethanol for preservation. Acari were sorted into Oribatida, Gamasida, Prostigmata and Astigmata. Oribatida were identified to species level using Weigmann and Collembola were identified to species level using Hopkin and Fjellberg. For identification a light microscope (Axioplan; Zeiss, Germany) with up to ×1000 magnification was used. For full list of species including authorities see supplementary material ( and Tables). Species richness (number of species; and Tables) and density (number of individuals per square meter; and Tables) for Collembola and Oribatida were calculated. ## Data analysis To improve homogeneity of variances, data on abundance (individuals per soil core) and species richness were log<sub>10</sub> (x+1) transformed prior to statistical analysis. The 60 plant species mixtures were excluded from the statistical analysis due to insufficient number of replicates (four replicates at the field site, each being differentially affected by the flood;). Linear models (type I sum of squares) were used to analyze effects of block (categorical variable, 4 blocks), flooding index (continuous variable, from 1 to 23 days), dead organic material, target species, weeds and bare ground (continuous variable, percentage of the plot, only used in October 2013), plant functional group richness (continuous variable, from 1 to 4), plant species richness (continuous variable, from 1 to 16, log-transformed) and presence/absence of grasses, legumes, small herbs and tall herbs (categorical variables) on the density and richness of Collembola and Acari, the density of Astigmata, Gamasida, Oribatida and Prostigmata (suborders of Acari) and the density of most abundant families of Collembola (Entomobrydae, Isotomidae and Tullbergiidae) for the data of 2010 and 2013 (three months after the flood). Due to very low density three weeks after the flood these data were not analyzed statistically. The full model with the lowest Akaike Information Criterion (AIC) was selected as the best starting model. This model was simplified in a stepwise manner by dropping non-significant variables. Although the experimental design was set up as orthogonal as possible, there is collinearity between functional group richness of plants and the presence/absence of individual functional groups, which we quantified using the inflation factor (VIF) from the car package. The analysis suggested to exclude functional group richness if there are two or more functional groups in the model (VIF \~ 4). Therefore, functional group richness was added after model simplification and was only included in the final model if it improved the model significantly (principle of Occam’s Razor, p \< 0.05). Generally, block and flooding index were fitted first followed by plant species richness; thereafter presence/absence of grasses, legumes, tall herbs and short herbs were fitted. F-values given in text and tables generally refer to those where the respective factor was fitted first. Statistical analyses were performed using R 3.2.1. Data on Collembola species were analyzed using non-metric multidimensional scaling (NMDS with Bray-Curtis distance) reducing the number of dimensions to four. To identify the factors which drive Collembola community composition, the four dimensions were further analyzed by MANOVA. In addition, discriminant function analysis (DFA) was carried out on four NMDS axes with Statistica 13 (Statsoft, Inc., Tulsa, Oklahoma, USA). Plant species and plant functional group richness were used as variables of discrimination. Squared Mahalanobis distances between groups were calculated to identify differences between plant richness levels. The community structure of Collembola and Oribatida was analyzed using principal component analysis (PCA) as implemented in CANOCO 5 (Microcomputer Power, Ithaca, NY;) using the abundance of species which appeared more than in three samples. Moreover, we correlate the factors and the axes of each PCA using Pearson correlation. # Results ## Collembola In November 2010 Collembola density was 22,310 ± 15,975 individuals m-<sup>2</sup>, whereas three weeks after the flood in 2013 it was only 515±1347 individuals m-<sup>2</sup>(mean ± SE). In contrast, three months after the flood Collembola density (23,220 ± 17,826 individuals m-<sup>2</sup>; mean ± SE) was similar to the level in 2010. Collembola density in 2010 was not influenced by experimental treatments, but three months after the flood in 2013 it increased slightly with plant species richness (F<sub>1,76</sub> = 2.97;). Moreover, dead organic material, target species, weeds and bare ground percentage of the plot was not significant in the density of October 2013. In 2010 a total of 27 species of Collembola were recorded, while only 16 species were recorded three weeks after the flood. However, three months after the flood species number increased to 22. In 2010 Collembola species richness increased marginally significant with plant species richness (F<sub>1,76</sub> = 3.76;) and plant functional group richness (F<sub>1,76</sub> = 2.82;). There was no significant effect of plant species and plant functional group richness on Collembola species richness three months after the flood in 2013, but increased significantly with the presence of tall herbs. Dead organic material, target species, weeds and bare ground percentage of the plot was not significant in the species richness of October 2013. In 2010 the density of the most abundant family of Collembola, Isotomidae, increased significantly with plant species richness and also in presence of grasses, but decreased in the presence of tall herbs. In contrast, the densities of Entomobryidae and Tullbergiidae were not significantly affected by experimental treatments in 2010. Three months after the flood in 2013 the density of Entomobryidae increased with flooding index (F<sub>1,73</sub> = 3.88). Further, the density of Tullbergiidae increased significantly with plant species richness and decreased slightly with the presence of grasses (F<sub>1,73</sub> = 3.28;). In contrast to 2010, Isotomidae were not significantly affected by experimental treatments in 2013. In 2010, Collembola community composition changed significantly with plant species richness (F<sub>1,76</sub> = 5.33, P \< 0.01) and plant functional group richness (F<sub>1,76</sub> = 4.35, P \< 0.01). Collembola community (number of species and species composition) was similar at higher plant species richness but less variable in the one and two species treatments. Similarly, community composition of Collembola differed between plant functional group one and four as well as two and four. PCA separated Collembola communities mainly along the first axis representing 25.28% of the variability in species data, whereas the second axis represented 15.96% of the variability. Separation along the first axis mainly represents differences between Collembola communities in 2010 and three months after the flood in 2013 (r = - 0.75). The most abundant species before the flood compared to 2013 were Parisotoma notabilis, Mesaphorura macrochaeta, Ceratophysella denticulata and Onychiurus jubilarius. After the flood the most abundant species compared to 2010 were Lepidocyrtus lanuginosus and Cryptopygus thermophilus. The second axis represents differences between plant species richness (r = 0.22), plant functional group richness (r = 0.13) and presence/absence of small herbs (r = 0.16). In general, at higher plant species richness Lepidocyrtus cyaneus and Stenaphorura denisi were more abundant. Moreover, there were some species present at each of the sampling dates including Lepidocyrtus lanuginosus, Lepidocyrtus cyaneus and Willowsia buski (Entomobryidae) as well as Isotoma viridis, Parisotoma notabilis and Isotomiella minor (Isotomidae). Other species like Hypogastrura manubrialis, Ceratophysella engadinensis (Hypogastruridae), Isotomurus fucicolus, Proisotoma minuta (Isotomidae), Paratullbergia macdougalli (Tullbergiidae), Protaphorura armata (Onychiuridae) and Sminthurus viridis (Sminthuridae) were present only three weeks after the flood. ## Acari In November 2010 Acari density was 21,500 ± 23,290 individuals m-<sup>2</sup>, but only 1,864 ± 3,059 individuals m-<sup>2</sup> three weeks after the flood in 2013(mean ± SE). In contrast, similar to Collembola, three months after the flood in 2013 (27,350 ± 30,040 individuals m-<sup>2</sup>; mean ± SE) it was similar to the level in 2010. In 2010 and three months after the flood in 2013, Acari density increased significantly with the presence of grasses. Moreover, three months after the flood in 2013 it was significantly higher with the presence of small herbs. The Acari density of October 2013 was not affected significantly by dead organic material, target species, weeds and bare ground percentage of the plot. The density of each of the suborders of Acari (Oribatida, Gamasida, Astigmata, Prostigmata) decreased significantly three weeks after the flood in 2013. In contrast, three months after the flood in 2013 the density of each of the Acari suborders reached a similar level than in 2010, except of Prostigmata which exceeded the density in 2010 by more than a factor of two. A total of 12 species of Oribatida were recorded in 2010, but only 9 species were recorded three weeks after the flood in 2013. However, similar to Collembola, Oribatida species also recovered quickly with 12 species being present three months after the flood in 2013. In 2010 Oribatida richness increased significantly with plant species richness and plant functional group richness. Further, in 2010 (F <sub>1,76</sub> = 2.94) as well as three months after the flood in 2013 (F <sub>1,76</sub> = 3.04) Oribatida richness increased with the presence of grasses, however, only slightly. Three months after the flood in 2013 Oribatida richness was slightly reduced at higher flooding index (F <sub>1,76</sub> = 2.86;). Furthermore, dead organic material, target species, weeds and bare ground percentage of the plot was not significant in Oribatida species richness of October 2013. Oribatida density increased significantly with plant species richness both in 2010 and three months after the flood in 2013. Moreover, in 2010 Oribatida density increased significantly with plant functional group richness and the presence of grasses. In 2010 Gamasida density decreased slightly with the presence of tall herbs (F <sub>1,70</sub> = 3.58), while in 2013 it increased significantly with plant species richness and plant functional group richness as well as the presence of legumes and slightly small herbs (F <sub>1,70</sub> = 3.01;). Prostigmata density increased slightly with plant functional group richness (F <sub>1,70</sub> = 3.75) and significantly in presence of grasses but only in 2010. In 2013, it was not significantly affected by experimental treatments. Astigmata density was significantly higher at higher plant species richness and in presence of grasses in 2010 (F <sub>1,70</sub> = 3.59) but not in 2013. PCA separated Oribatida species along the first axis explaining 48.2% of the variability in species data and the second axis representing 15.99% of the variability in species data. Separation along the first axis mostly represents differences between Oribatida communities in 2010 and three months after the flood in 2013 (r = 0.11) and the presence of legumes (r = 0.13) and grasses (r = 0.15). The second axis mainly represents differences with plant functional group richness (r = 0.14) and presence of grasses (r = 0.19). Separation along the first axis were due to e.g., Oppiella nova being more abundant three months after the flood in 2013. Separation along the second axis was due to e.g., higher numbers of Oribatula excavata in plant communities with grasses. Moreover, we found some species present at each of the sampling including Oppiella nova, Tectocepheus sarekensis, Oribatula excavata, Rhysotritia ardua and Punctoribates punctum as well as species present only three weeks after the flood like Zygoribatula frisiae and Schleloribates initialis. # Discussion Collembola and Acari density as well as Collembola and Oribatida species richness were affected drastically by the flood but recovered quickly, returning within three months to levels recorded three years earlier (2010), despite of the seasonal changes of the community, since generally the Collembola and Acari abundance in autumn is higher than in summer (;). This reinforces assumptions that microarthropods respond rapidly to environmental changes and recover quickly. In 2010 Collembola and Oribatida species richness increased with plant species and plant functional group richness, but these interrelationships were absent three months after the flood. This suggests that flooding resulted in homogenization of environmental conditions, eradicating effects of plant community composition established before the flood. However, three months after the flood Collembola species richness was increased in the presence of tall herbs, and Oribatida density and richness were consistently affected by grasses, indicating that effects of plant community composition were in a stage of being reasserted. As hypothesized, Entomobryidae, as epedaphic species, may benefit from their dispersal ability thereby quickly recovering after the flood, presumably taking advantage of increased fungal biomass in the most severely flooded plots. In contrast to Entomobryidae, the euedaphic and hemiedaphic Tullbergiidae and Isotomidae require habitable pore space and this likely contributed to the delayed recolonization of the clogged flooded soils. They also are more reliant on the recovery of the plant community. Isotomidae and Tullbergiidae are sensitive to soil quality and root exudates, and are assumed to benefit from increased root biomass and associated exudates in more diverse plant communities. In 2010 the density of Isotomidae increased in presence of grasses and tall herbs. Grasses increase root and microbial biomass, both likely contributing to increased food resource supply to Collembola. Oribatid mites are decomposers and have been used as indicators of soil stability and fertility. Generally, they have low metabolic rates, slow development and low fecundity. Consequently, they are vulnerable to disturbances and recover slowly thereafter. However, higher densities of Prostigmata and Oribatida than of Astigmata and Gamasida three weeks after the flood suggest that the former two taxa are more resistant or recovered faster to flooding than the latter. Detrimental effects of flooding may have been alleviated by dead plant biomass functioning as shelter for them. Gamasida are important predators of nematodes, which were negatively affected by the flood. Moreover, ant activity was not influenced by the flood and Gamasida are negatively affected by high ant activity. Notably, three months after the flood, the density of Prostigmata was twice that in 2010. Prostigmata are dominant Acari predators with a crucial role in soil food webs important as biological control agents. Some species are known to benefit from habitat disturbance. Prostigmata, Gamasida and Astigmata are assumed to be r-strategists frequently occurring in disturbed habitats and quickly colonizing new habitats due to high dispersal ability, high fecundity and fast development. However, in contrast to Prostigmata, Gamasida and Astigmata did not take advantage of the flood. The majority of Gamasida are mobile predators feeding on Nematoda, Collembola, Enchytraeidae, larvae of Insecta and Acari. Conform to their assumed sensitivity to environmental changes, their density increased with plant species and plant functional group richness as well as in presence of legumes three months after the flood in 2013. This likely resulted from increased availability of prey such as Collembola (Isotomidae) and nematodes both increasing significantly with plant diversity three months after the flood. As ecosystems develop after disturbances, changes in soil are likely to be associated by corresponding changes in the soil community. As indicated by the dramatic decline in the density and species richness of Collembola and Acari, flooding represented a strong disturbance for soil microarthropods. The community composition of Collembola markedly changed after the flood and this lasted for at least three months. On one hand some species were present at each of the sampling dates, including Lepidocyrtus lanuginosus, Lepidocyrtus cyaneus and Willowsia buski (Entomobryidae) as well as Isotoma viridis, Parisotoma notabilis and Isotomiella minor (Isotomidae). Each of these species had been recorded three weeks after the flood suggesting that they survived the flooding in some of the plots and then recovered quickly confirming their generalistic lifestyle. On the other hand, Hypogastrura manubrialis, Ceratophysella engadinensis (Hypogastruridae), Isotomurus fucicolus, Proisotoma minuta (Isotomidae), Paratullbergia macdougalli (Tullbergiidae), Protaphorura armata (Onychiuridae) and Sminthurus viridis (Sminthuridae) were present three weeks after the flood (as single individuals or in few numbers only) but neither three months after the flood nor in 2010 before the flood suggesting that a number of Collembola species were introduced by the flood. In contrast, Zygoribatula frisiae and Schleloribates initialis of Oribatida were only present three weeks after the flood. Furthermore, Oppiella nova, Tectocepheus sarekensis, Oribatula excavata, Rhysotritia ardua and Punctoribates punctum were present in all the dates and most of them quickly increased in density suggesting they are resistant against disturbances and respond in an generalistic way. In fact, these species are known as cosmopolitan generalistic species present in any kind of habitat. # Conclusion The present study demonstrated that soil microarthropod communities (Collembola and Acari) are affected heavily from summer flooding, but also that they are able to recover quickly. Recovery was based in large on ubiquitous and resistant species surviving the flood and able to form vigorous populations within short period of time (three months). Mobile surface living species were the quickest to recover and in part took advantage of resources made available due to flooding (dead plant material). Widespread microarthropod species with wide habitat niches recovered faster than those with more limited distribution and more narrow niches. Overall, Collembola were more affected by flooding and recovered faster than Acari. In contrast, recovery of community composition of Collembola after flooding was slower than that of Oribatida. These conclusions, however, were based on two sampling events following flooding and long-term studies are needed to uncover the resilience of microarthropod communities to extreme climate events such as flooding. # Supporting information We thank the gardeners and numerous student helpers for weeding and mowing of the experimental plots and helping with data collection. We thank Tamara Hartke and Laura Sánchez Galindo for support and helpful comments. [^1]: The authors have declared that no competing interests exist.
# Introduction Liaoning Province, located in the northeast of China, is known in Chinese as “the Golden Triangle” from its shape and strategic location. It was established in 1907 as the name of Fengtian and changed to Liaoning in 1929, with an estimated population of approximately 43.91 million in 2014 ([www.stats.gov.cn](http://www.stats.gov.cn/)). The population is mostly Han Chinese (83.94%) with minorities of Manchus (12.88%), Mongols (1.60%), Hui (0.632%), Koreans (0.576%) and Xibe (0.317%). Liaoning Han individuals mainly migrated from Shandong Peninsula during the hundred-year period starting at the last half of the 19<sup>th</sup> century. “Chuang Guandong” is a description that Han Chinese population, especially from the Shandong Peninsula and Zhili, entered Manchuria. During the first two centuries of the Manchu Qing Dynasty, Liaoning Province is the traditional homeland of the ruling Manchus with only certain Manchu Bannermen, Mongol Bannermen, and Chinese Bannermen allowed in. The region, now known as Northeast China, has an overwhelmingly Han population. After the establishment of the People's Republic of China at the end of the Chinese Civil War, further immigrations were organized by the Central Government to "develop the Great Northern Wilderness", eventually peaking the population over 100 million people. Thus, it is necessary and sufficient to investigate the genetic background of Liaoning Han population and compare the genetic distance with other population. Additionally, it is interesting to observe how much admixture took place over the past 100 years among Han Chinese and other groups. Y-chromosomal short tandem repeats (Y-STR) is a useful tool for inferring genetic genealogy evolution and ancient human migration trajectories and timing. The non-recombinant region of the Y-chromosome may play a potential role in revealing the ethnic and regional representation of the Han Chinese population owing to its significant phylogeographic information content. It can supply an informative reference for investigating patterns of genetic variation in the Han Chinese population across East Asia considering that the genetic and cultural diversity among East Asian populations is still not fully understood. Autosomal STR loci are usually applied in forensic personal identification and paternity tests, which can provide a mighty powerful discrimination capability without influenced by linkage disequilibrium. It can also be used to uncover the population genetic backdrop and structure. The population data of autosomal STR loci can be utilized to constructed the phylogenies and clarify the genetic structure using genetic distance measurements, neighbor-joining dendrograms and principal component analysis base on different genotyping frequencies. Therefore, we investigated the frequencies of 25 Y-STR and 15 autosomal STR loci in Liaoning Han population to expand the available population information for forensic medicine and human genetic diversity. Population comparison was performed between Liaoning Han population and different ethnic groups to better understand the genetic background of the Liaoning Han population. # Methods ## Study population Three hundred and five blood samples were collected from unrelated healthy male individuals living in Liaoning Province, Northeast China, after obtaining written informed consent. The blood was then stained onto filter papers. Samples were obtained and analyzed after approval from the Ethics Committee of China Medical University. ## Data extraction, PCR amplification, and genotyping Genomic DNA was extracted using Chelex-100. PCR amplification was performed using AmpFISTR<sup>®</sup> Yfiler<sup>®</sup> Plus and Identifiler<sup>TM</sup> PCR amplification kits (Thermo Fisher Scientific, CA, USA) in a GeneAmp<sup>®</sup> PCR 9700 (Thermo Fisher Scientific, CA, USA) thermal cycler, according to respective manufacturer specifications. The AmpFISTR<sup>®</sup> Yfiler<sup>®</sup> Plus amplification kit (Thermo Fisher Scientific, Waltham, MA, USA) can co-amplify 25 Y-STR loci with six dyes, including seven rapidly mutating loci. The AmpFISTR<sup>®</sup> Identifiler<sup>TM</sup> PCR Amplification kit (Thermo Fisher Scientific) can co-amplify 15 autosomal STR loci and the Amelogenin locus with five dyes. Fragments of the 25 Y-chromosomal and 15 autosomal STR loci were produced simultaneously. Separation and detection of amplicons was performed on an Applied Biosystems™ 3500 Series Genetic Analyzer (Thermo Fisher Scientific, Waltham, MA, USA). Data were analyzed using GeneMapper ID v4.1 software (Thermo Fisher Scientific, Waltham, MA, USA). Control DNA 007 was included as a standard reference in each batch of genotyping. We strictly followed the recommendations of the DNA Commission of the International Society of Forensic Genetics (ISFG) for Y-STR analysis. ## Data analysis For Y-STR loci, allele frequencies and gene diversity were calculated using PowerMarker v3.25. Haplotype frequencies, random match probabilities (sum of squares) and haplotype diversity were calculated using Arlequin Software v3.5. The discrimination capacity (DC) was determined as the proportion of different haplotypes in each sample. A cluster structure of Y-STR haplotypes was generated using the YHRD database (<http://www.yhrd.org/>). To compare data from the studied Liaoning Han population with other published data, genetic distance (Rst statistics) was measured by analysis of molecular variance (AMOVA) and visualized using two multi-dimensional scaling (MDS) Rst plots via YHRD online tools (<http://www.yhrd.org/Analyse/AMOVA>). For autosomal loci, sample allele frequencies and exact Hardy-Weinberg equilibrium (HWE) tests were calculated using PowerMarker v3.25. Values for power of discrimination (DP), polymorphism information content (PIC), power of exclusion (PE), and heterozygosity (He) were calculated using Power Stats v1.2 software that had been modified by Raquel, et al. to support and manage the large amount of samples. Pairwise genetic distance (Fst) and p values for each locus were calculated between populations using Arlequin v3.5 software. Furthermore, Nei’s standard genetic distance between populations was generated by the Phylip 3.69 package and visualized with Treeview software. Because the published relevant data is limited, the included groups for population comparison between Y-STR and autosomal STR are different. # Results and Discussion ## Y-chromosomal STR Two hundred and ninety-three different haplotypes were observed from 305 unrelated individuals. Among them, 281 were unique and 12 were shared by two individuals. Null alleles were found in nine individuals at DYS448 and one individual at DYS385, respectively. Haplotype diversity rendered a high value (0.9997 ± 0.0003). Likewise, a high random match probability (0.0035) was determined with a DC of 0.9607. Genetic diversity values of the 25 loci ranged from 0.4525 (DYS391) to 0.9617 (DYS385). Among them, allele frequencies ranged from 0.7016 (DYS438) to 0.0033 (DYS389I, DYS389II, DYS458, YGATAH4, DYS448, DYS391, DYS456, DYS439, DYS481, DYS533, DYS576, DYS627, DYS460, DYS518, DYS449, DYF387S1 and DYS385). Cluster analysis was performed for the 12 haplotypes that were observed twice. Ancestry information showed that the haplotypes of the Liaoning Han population most likely belonged to the East Asian-Sino Tibetan- Chinese culture, which corresponds with its history, culture, and geographical distribution. The powerful informative content of the 25 Y-STR loci in the Liaoning Han population will be useful and interesting in forensic medicine and enrich the Han Chinese population database. ## Autosomal STR The distribution of allele frequencies, forensic efficiencies, and statistical parameters across the 15 autosomal STR loci are presented in and Tables. Among the 157 observed alleles, allele frequencies ranged from 0.5164 (TH01) to 0.0016 (D8S1179, D21S11, D7S820, D3S1358, D13S317, D16S539, D2S1338, D19S433, D18S51, D5S818 and FGA). The DP ranged from 0.9621 (D2S1338) to 0.8177 (TPOX), with PE distributing from 0.7521 (D18S51) to 0.2988 (TH01). The span of He was 0.8787 (D18S51) to 0.6066 (TH01). Except for CSF1PO (0.6993), D3S1358 (0.6775), TH01 (0.5973), and TPOX (0.5872), all autosomal STR loci were highly polymorphic (PIC \> 0.7) with the most D18S51 (0.8435). No departures from HWE were observed after Bonferroni’s correction for multiple testing (p \< 0.05/15). ## Population comparison For Y-STR loci, we compared our haplotype data with that of the five populations that were submitted to the YHRD database (Release 51), which included Austrian, German, Polish, African and Native American. Rst values for genetic distance demonstrated that haplotypes of the Liaoning Han population were significantly different from those of the other five populations (all p values \< 0.05/5 after Bonferroni correction). As shown in the MDS plot, there were significant differences between Liaoning Han population and the five population. Furthermore, in order to comprehensively investigate the genetic substructure of Liaoning Han population, the population comparison using the 16 shared Y-STR loci except for DYS627, DYS460, DYS518, DYS449, DYF387S1, DYS481, DYS533, DYS576 and DYS570 was performed between Liaoning Han and 22 East Asian groups. They included Anhui Han Chinese, Beijing Han Chinese, Guangdong Han Chinese, Guizhou Han Chinese (YP001096), Jiangsu Han Chinese, Jiangxi Han Chinese, Jilin Han Chinese, Shandong Han Chinese, Shanxi Han Chinese, Yunnan Bai (YP000902), Xishuangbanna Dai (YP000903), Liaoning Hui (YP000819), Liaoning Korean, Liaoning Manchu, Hunan Miao (YP001038), Liaoning Mongolian, Gansu Tibetan (YP001032), Hunan Tujia (YP001037), Liaoning Xibe, Guangxi Zhuang (YP000591), Japanese and Korean. Rst values for genetic distance demonstrated that haplotypes of the Liaoning Han population were significantly different from those of the other 22 populations (all p values \< 0.05/22 after Bonferroni correction;). As shown in the MDS plot, minor differences were observed when the Liaoning Han population was compared to the Jilin Han Chinese, Beijing Han Chinese, Liaoning Manchu, Liaoning Mongolian, Liaoning Xibe, Shandong Han Chinese, Jiangsu Han Chinese, Anhui Han Chinese, Guizhou Han Chinese and Liaoning Hui populations; by contrast, major differences were observed when the Liaoning Han population was compared to Shanxi Han Chinese, Yunnan Bai, Jiangxi Han Chinese, Guangdong Han Chinese, Liaoning Korean, Hunan Tujia, Guangxi Zhuang, Gansu Tibetan, Xishuangbanna Dai, South Korean, Japanese and Hunan Miao populations. Additionally, the populations’ distributions in the MDS plot corresponded with their respective ethno-geographic origins. It is clear that Liaoning Han Chinese has a close genetic distance with Southern Han population and Liaoning native minorities, which indicated that Liaoning Han integrated gradually with natives, such as Manchu, Mongolian and Xibe, following its geographical migration. For autosomal STR loci, presents pairwise Fst and p values for differentiation tests between the Liaoning Han ethnic group and nine additional published populations; statistically significant differences (p \< 0.05/15) were found between the Liaoning Han population and the China Miao population at five STR loci, the China Bouyei population at four STR loci, the China Uygur and Jinan Han populations at three STR loci, the Japanese population at two STR loci, and the Korean and Shanghai Han populations at one STR locus. No statistically significant differences were detected at any STR loci between the Liaoning Han and the China Dong or Shaanxi Han populations. shows genetic distances between populations. indicates clusters of unrooted phylogenetic trees to mirror the historical and geographical backgrounds of the populations compared. In culture custom, because most people in Northeast China trace their ancestries back to the migrants from the Chuang Guandong era, Northeastern Chinese were more culturally uniform compared to other geographical regions of China. Therefore, people from the Northeast would first identify themselves as "Northeasterners" before affiliating to individual provinces and cities ([http://chinaneast.xinhuanet.com](http://chinaneast.xinhuanet.com/)). For Han Chinese population, the previous studies showed that it was intricately sub-structured and clustered roughly to two (northern Han and southern Han) or three (northern Han, central Han and southern Han) subgroups. The distinction between southern and northern Han populations were reported by Chu et al using the neighbor-joining method based on the data of STR loci. The Han Chinese group has the same predecessors, the Yan Emperor and the Yellow Emperor in the Yellow River Basin. However, the Han population has been forming a series of relationships with different groups and coexisted with other ethnic groups since thousands of years ago. Obviously, Liaoning Han population belonged to northern Han subgroup according to the geographic distribution and historical cultural. The population comparison based on Y-STR loci showed that Liaoning Han was an independent endogenous ethnicity with a unique subpopulation structure. The previous study showed that Liaoning Han had a close genetic distance with Manchu, which was not as near as Han population of Jilin and Beijing, but nearer than other ethnic groups. This result might indicate that the Liaoning Han integrated gradually with natives, such as Manchu, Mongolian and Xibe, following its geographical migration, which was corresponded with the historical records. However, autosomal STR population comparison presented that there was no significant difference between the Liaoning Han and the China Dong or Shaanxi Han populations, which seemed to be contradictory to Y-STR results. This might be due to the discrepancy of different genetic markers. Consequently, Liaoning Han population owns its unique genetic characteristics that are different from Han population from other provinces, except for Jilin Han population. There were two potential limitations in the present study. First, the analysis of the Y chromosomal and autosomal STR loci could not provide the precise and reliable data for population comparison with the absence of the whole genome data. Second, the included groups for population comparison between Y-STR and autosomal STR are different, due to the limited available relevant data. Thus, more genetic investigations need to do in order to better understand the characteristics of Liaoning Han Chinese population. # Conclusion The population comparison demonstrates that the Liaoning Han population is an independent endogenous ethnicity and still owns its unique genetic characteristics. In summary, the reported genetic characteristics of the 25 Y-STR and 15 autosomal STR loci allelic frequencies and haplotype distributions of the Liaoning Han population are informative for forensic investigation and paternity testing. The results could help inferring the genetic genealogy evolution and ancient human migration patterns. # Supporting Information STR short tandem repeat AMOVA analysis of molecular variance MDS multi-dimensional scaling HWE Hardy-Weinberg equilibrium DP power of discrimination PIC polymorphism information content PE power of exclusion He heterozygosity Fst Pairwise genetic distance [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: BW. Performed the experiments: JY. Analyzed the data: JY. Contributed reagents/materials/analysis tools: JY. Wrote the paper: BW.
# Introduction Streptococcus gallolyticus subsp. gallolyticus, formally known as Streptococcus bovis biotype I, belongs to the Lancefield group D Streptococci, is a normal inhabitant of the animal and human gastrointestinal tract, and appears in 2.5 to 15% of healthy humans. On the contrary, its frequency in the digestive tract of animals and the absolute frequencies in various species are not well described. To date, there is only one study, which estimated the percentage of S. gallolyticus subsp. gallolyticus in feces of turkeys. The detection rate in fecal samples of turkeys is 91%. It was also identified in pigeon, bovine and chicken as commensal bacterium. However, S. gallolyticus subsp. gallolyticus can also act as a facultative pathogen, causing sepsis, meningitis and infective endocarditis (IE) in humans and animals. Human IE is especially associated with colorectal cancer. The incidence of group D Streptococcus-associated diseases is increasing in the south of Europe. The detection in humans and animals as a causative agent producing the same clinical symptoms leads to the assumption that S. gallolyticus subsp. gallolyticus may be a zoonotic pathogen. Investigations in France and Spain suggest a correlation between a rural residency and the presence of the facultative pathogen. The transmission of the potential zoonotic pathogen may be directly by smear or droplet infections or indirectly from surfaces contaminated with S. gallolyticus subsp. gallolyticus. The transfer of the bacterium through a closer contact with colonized or infected animals is also discussed as a possible mechanism. It was described as an important risk factor for the transmission of Streptococcus suis between infected animals and humans. Epidemiologic analyses in a laying hen flock in North Rhine Westphalia also contribute to the assumption that a closer occupational contact with colonized laying hens may be a potential risk factor for the colonization of the gastrointestinal tract with S. gallolyticus subsp. gallolyticus, since the bacterium was identified as the causative agent of IE of the farm owner. In addition to the detection of S. gallolyticus subsp. gallolyticus in eukaryotic organisms, it was also identified in milk and raw milk products (especially in dairy cows with mastitis) and red meat. The detection in food leads to the assumption that the transmission of S. gallolyticus subsp. gallolyticus between animals and humans can be connected to dietary habits. Exemplarily, Streptococcus equi subsp. zooepidemicus was transmitted through the consumption of unpasteurized raw milk in an outbreak setting. There have been no investigations to date which systematically analyze the correlation between dietary habits or the contact with animals and the detection of S. gallolyticus subsp. gallolyticus in the human gut. Therefore, we conducted an epidemiological study which is comprised of two parts: Firstly, the case- control study to determine the prevalence of S. gallolyticus subsp. gallolyticus and the associated risk factors for the colonization of the human gut, and secondly, multilocus sequence typing (MLST) to characterize the S. gallolyticus subsp. gallolyticus population structure. This analysis identified a correlation between lifestyle habits and the human gastrointestinal colonization with S. gallolyticus subsp. gallolyticus. # Material and methods ## Sample and data acquisition A retrospective case-control study was conducted at the Herz- und Diabeteszentrum Nordrhein-Westfalen (Bad Oeynhausen, Germany) from December 2012 to July 2015. The case-control study used word of mouth to recruit people. All data are collected in pseudonymous form. A total of 135 volunteers from the north and west of Germany participated in this study. Written consent was required for the case-control study. Fecal samples were tested and a questionnaire was completed by each volunteer to analyze the correlation between the fecal presences of S. gallolyticus subsp. gallolyticus and potential risk factors. Furthermore, seven SGG-culture-positive tested healthy volunteers were selected and were analyzed two to three times to estimate the gastrointestinal presence of S. gallolyticus subsp. gallolyticus in a follow-up period (follow-up study). Participants were excluded from the study if there was no fecal sample, no completed questionnaire or no written consent. In addition, only healthy volunteers (without gastrointestinal diseases or IE) over 18 years were included for the identification of risk factors. These data were received from the questionnaires. People were also excluded if an antibiotic therapy was indicated six months prior to participation. The study was approved by the ethics commission of the Ruhr University Bochum Faculty of Medicine. ## Stool investigations ### DNA extraction DNA extraction of homogenized fecal samples was performed by using NucliSENS easyMAG (Biomerieux, Nürtingen, Germany). DNA extraction was generally performed according to the manufacturer’s instruction. The fecal samples were pretreated by inoculating material (about 0.1 g) in 1 ml PBS in a tube with Zirconia beads, mixed for 5 min, incubated for 10 min at room temperature and then centrifuged at 12000 × g for 2 min. A quantity of 200 μl of the supernatant was used for the extraction of the whole DNA. After prelysis within the NucliSENS easyMAG, 100 μl magnetic silica particles were added and extraction was performed as described by the manufacturer. DNA was eluted in 55 μl elution buffer. ### Real-Time PCR The detection of an internal fragment of the recN gene was used to screen fecal specimens for the detection of S. gallolyticus subsp. gallolyticus. The PCR amplification for the presence or absence of this gene was carried out within a 50 μl reaction volume containing 5 μl template DNA, 5 μl Platinum-Taq-buffer (ThermoFisherScientific, Darmstadt, Germany), 200 nM of each primer (F-recN SGG/P: `5’-GATTTTCAAGTCCAATTCACCAAAG-3’`, R-recN SGG/P: `5’-GGTTYGTTGAAATGTAAAATTCAACAG-3’`; LifeTechnologies, Darmstadt, Germany), 100 nm of the Pf-recN/SGG-probe (`5’-FAM-TTCAATCGTGATGGCAA-MGB-3’`; LifeTechnologies, Darmstadt), 240 μM dNTPs (Fermentas, Leon-Rot, Germany) and 0.25 μl Platinum-Taq-polymerase (ThermoFisherScientific, Darmstadt). The detection of S. gallolyticus subsp. pasteurianus was performed using the Pv-recN/SGP-probe (`5’-VIC-TCAACCGTGATGGAAA-MGB-3’`) and the same primers denoted above. The internal control used in the reaction mix was CMV-DNA (CMV- TM2-F: `TTYTTAGCACGGGCCTTAGC`, CMV-TM2-R: `AAGGAGCTGCATGATGTGASC`; CMV-TM2-S: `CY5-TGCAGTGCACCCCCCAACTTGTT-BHQ2`;). Diluted DNA extracted from a bacterial overnight culture of S. gallolyticus subsp. gallolyticus (ATCC BAA-2069) or S. gallolyticus subsp. pasteurianus (DSM 15351) was used as positive control and water as negative control to verify the specificity of the PCR reaction. A two-step PCR on the Rotor Gene Q platform (Qiagen, Hilden, Germany) was performed. Amplification of PCR products was carried out as follows: initial denaturation at 95°C (5 min) followed by 50 cycles, and a denaturation 95°C (15 s), annealing and elongation step at 60°C (60 s). ### Selective cultivation Real-time PCR S. gallolyticus subsp. gallolyticus positive-tested fecal samples were further selectively cultivated on modified trypton soya agar (TSA) (0.5% tannic acid pure (AppliChem GmbH, Darmstadt, Germany, 0.25 g/l sodium acetate, Merck, Darmstadt, Germany), as described previously. Briefly, the homogenized fecal sample was streaked out onto selective medium before weighing and suspending in PBS buffer. Then, 1 g of homogenized fecal sample was suspended in 1 ml PBS medium, mixed and streaked out with PBS (duplicate) and 100 μl was plated as triplicate onto sodium acetate tannic acid TSA. It was then incubated at 37°C and 5% CO<sub>2</sub> for 48 h. In parallel, an overnight grown culture of S. gallolyticus subsp. gallolyticus was plated onto modified TSA. Single putative S. gallolyticus subsp. gallolyticus colonies were selected and analyzed regarding species and subspecies level by matrix- assisted laser desorption ionization—time of flight mass spectrometry (MALDI- TOF-MS) and sodA sequencing. ## Multilocus sequence typing Multilocus sequence typing was performed, as described preciously. In brief, the total DNA of S. gallolyticus subsp. gallolyticus isolates was isolated by using a QIAamp Blood Mini Kit (Qiagen, Hilden, Germany) and 5 μl was used for each fragment amplification. Partial sequences of the housekeeping genes aroE, glgB, nifS, p20, tkt, trpD and uvrA were amplified, sequenced and analyzed. All detailed protocols can also be found on [www.pubmlst.org](http://www.pubmlst.org). The determination of sequence types (STs) was undertaken using the pubMLST database and Bionumerics Software 6.6 (Applied Maths, Sint-Martens-Latem, Belgium). For the characterization of the strains a minimum spanning tree was generated and eBURST version 3 (based upon related sequence types; [www.mlst.net](http://www.mlst.net)) was used to calculate clonal complexes. ## Questionnaire The questionnaire included 25 questions and sought data on the following aspects: personal characteristics (age, gender, gastrointestinal diseases, residence \[urban, rural, landscape—near the forest/farm\] and antibiosis), contact with animals (living or working on a farm, private or occupational contact) and dietary habits (consumption and handling of minced meat, raw milk or raw milk products). The exposure factors as well as the absolute frequencies can be found in. ## Statistical analysis The statistical analysis software SPSS version 21 was used. Binary logistic regression was utilized to establish a model to determine the simultaneous influence of potential risk factors. Six independent variables (age, gender, consumption of raw animal products, close animal contact, usage of animal waste) were tested within the multiple logistic regression model to verify adjusted odds ratios (ORs). Statistical tests were considered to be significant if the p-value was less than 0.05. A confidence interval of 95% was used for both calculations. Forest plots were generated using Microsoft Excel. The age was listed as mean plus/minus standard deviation. # Results A total of 134 participants (65 men and 69 women with a mean age of 48.4 ± 14.9 years) were included in the case-control study. After the application of exclusion criteria, 99 healthy volunteers were included to identify potential risk factors for the colonization of the human gut with S. gallolyticus subsp. gallolyticus. The fecal samples of healthy volunteers were screened by PCR for the presence of the facultative pathogen. S. gallolyticus subsp. gallolyticus was detected in 62.5% (n = 59) of the 99 fecal specimens of healthy volunteers. The presence of S. gallolyticus subsp. pasteurianus was also estimated using the VIC labeled probe in the PCR reaction mix. This bacterium was detected seven times out of 99 volunteers. Real-time PCR testing also identified three subjects who were colonized with S. gallolyticus subsp. gallolyticus as well as S. gallolyticus subsp. pasteurianus simultaneously. These volunteers are recognized in the S. gallolyticus subsp. gallolyticus-positive group. S. gallolyticus subsp. gallolyticus PCR-positive specimens were cultured on selective medium to isolate this bacterium for epidemiologic characterization by MLST. Three isolates, namely HDZ 1323, HDZ 1330 and HDZ 1332, were detected by culture and mass spectrometric analyses, and sodA sequencing confirmed S. gallolyticus subsp. gallolyticus. In addition to S. gallolyticus subsp. gallolyticus, three out of seven S. gallolyticus subsp. pasteurianus isolates were identified by MALDI-TOF MS and sequencing of the partial fragment of the sodA gene. In this regard, the real-time PCR demonstrated one inconsistent result: S. gallolyticus subsp. gallolyticus was detected instead of S. gallolyticus subsp. pasteurianus. The S. gallolyticus subsp. gallolyticus isolates selected were further typed using MLST. It revealed the sequence types ST 3 (HDZ1330), ST 7 (HDZ1323) and the newly defined ST 105 (HDZ1332). Bionumerics Software 6.6 was used to utilize for the construction of a minimum spanning tree. The minimum spanning tree of the strains revealed no phylogenetic relationship of these detected STs. The allelic profiles (STs) show no identical allelic numbers within the identified sequence types of the case control study (one exception: the number of the nifS allele from ST 3 and 7). The ST 3 was already identified in human heart valve cultures and from the intestine of a bovine, which was also detected for ST 7 isolates (, [www.pubmlst.org](http://www.pubmlst.org)). In order to identify the S. gallolyticus subsp. gallolyticus status in the gastrointestinal tract over time, a follow-up investigation of seven culture positive tested healthy volunteers was performed until the end of the study (a total of 2 to 3 samples per person). Initially, six fecal specimens were screened as real-time PCR positive for S. gallolyticus subsp. gallolyticus and one sample was tested as positive for S. gallolyticus subsp. pasteurianus. Selective cultivation offered three S. gallolyticus subsp. gallolyticus and one S. gallolyticus subsp. pasteurianus isolate. Further analyses of specimens revealed S. gallolyticus subsp. gallolyticus in four cases at any tested time point by using real-time PCR (volunteer 4 to 7). As an example, the first (January 2015) and second sample (March 2015) were tested as positive and the third specimen in July 2015 was tested as negative (volunteer 3). The fecal sample of the 7th volunteer was initially tested as positive (April 2013) and the second sample 26 months later was detected as positive for the presence of S. gallolyticus subsp. gallolyticus. At least one exception was identified. The PCR results of the second sample from volunteer 1 showed the presence of both subspecies, and S. gallolyticus subsp. pasteurianus was isolated using modified trypton soya agar. Based on the real-time PCR detection of the bacterium in fecal samples, cases and controls were defined: 59 cases (male/female \[m/f\] ratio: 25/34) and 40 controls (m/f ratio: 21/19). The cases included an age distribution from 20 to 70 years with a mean value of 44.2 ± 14.6 years and controls from 22 to 82 years with an average of 49.2 ± 15.0 years. The cases and controls were analyzed in terms of their nutrition habits, contact with animals and residence (rural, urban; forest or farm next to their residence). The frequencies of potential risk factors observed in cases and controls are listed in. Logistic regression was performed to analyze the simultaneous effect of risk factors and adjusted ORs were calculated and presented as a forest plot. Multiple expositions often characterize the outcome of a disease or different event such as the S. gallolyticus subsp. gallolyticus colonization of the human gastrointestinal tract. Adjusted ORs were calculated to assess the simultaneous effect of the variables. A closer animal contact between volunteers and animals (OR: 3.27, CI: 1.23–8.68; p = 0.02) and the usage of animal waste to fertilize plants (OR: 3.43, CI: 1.08–10.94; p = 0.04) demonstrate significant risk factors for the transmission between animals and humans and to colonize the gastrointestinal tract of healthy people. In conclusion, simultaneous testing of exposure factors indicate a higher risk of being colonized with S. gallolyticus subsp. gallolyticus of participants who have a direct contact with animals and utilization of manure. Furthermore, two out of three S. gallolyticus subsp gallolyticus isolates reveal the STs 3 and 7. These STs were previously isolated from human blood cultures and bovine (, [www.pubmlst.org](http://www.pubmlst.org)). # Discussion The knowledge of transmission pathways and the zoonotic potential of the facultative pathogen S. gallolyticus subsp. gallolyticus are quite limited. Thus, a systematic approach was conducted for the first time to determine the latter’s occurrence in the gut of healthy people and to describe the risk factors for the transmission of the bacterium and its colonization of the human gastrointestinal tract. S. gallolyticus subsp. gallolyticus is an opportunist of the gastrointestinal tract in humans and animals with varying prevalence in the healthy human population of 2.5 up to 15%, but the S. bovis fecal carriage increased three to five times in patients with colorectal cancer and inflammatory bowel disease. In comparison to the previous studies real-time PCR screenings of fecal samples offer a much higher prevalence in healthy volunteers of 62.5%. Spanish real-time PCR investigations of patients who underwent colonoscopy revealed 11.1% S. gallolyticus subsp. gallolyticus-positive and 13% S. gallolyticus subsp. pasteurianus-positive rectal swabs. It indicates a similar portion of both subspecies in the gastrointestinal tract, which cannot be confirmed by real-time PCR screenings of feces from participants of the case-control study. Similar results were detected for the presence of S. gallolyticus subsp. pasteurianus. The detection of both subspecies using real-time PCR may indicate a co-occurrence in the digestive tract, which was also suggested by Lopes et al.. The divergences observed between previous studies may be a resumé of the sample sets analyzed (colonoscopy, feces, rectal swabs) or the kind of screening methods (cultivation, molecular techniques) to identify or isolate S. gallolyticus subsp. gallolyticus. A higher sensitivity of the molecular screening method was demonstrated by positive real-time PCR results in comparison to selective cultivation and was also confirmed in the follow-up study. The complexity and characteristics of the sample type and difficulties in homogenization as well as the gut microbiota may influence the PCR results (e. g. inhibitors) and the selective cultivation. It explains not only false negative real-time PCR results, but also discrepancies between S. gallolyticus subsp. gallolyticus-positive culture and the S. gallolyticus subsp. pasteurianus-positive real-time PCR screening results. As suggested in a 17-year follow-up study, the prospective investigations of seven participants achieved shifts in the composition of the gut microbiome, which was e. g. demonstrated for the participant three. Nutrition and medicines shape the gut microflora (e. g. antibiotics), whereas antibiotics change the gut composition up to one year. This may also be transferred to our follow-up study. Although the participants with an antibiotic therapy were excluded it is not known if the composition of the gut microbiome affects the presence of S. gallolyticus subsp. gallolyticus in the gut. However, the sample age may be of more importance than the sample characteristics, microbiota or processing. The fecal samples were sent to the laboratory by mail. Consequently, samples could be in transit for three days before processing in the laboratory. Although a survival of S. gallolyticus subsp. gallolyticus was demonstrated in vitro in S. gallolyticus subsp. gallolyticus-negative tested human stool specimens for 14 d at RT (20°C) (real-time PCR and cultivation; data not shown) it cannot be ruled out, that the growth of other gastrointestinal bacteria may inhibit growth of S. gallolyticus subsp. gallolyticus or a less concentration of S. gallolyticus subsp. gallolyticus in the feces may effect a false negative cultivation. It is assumed that the presence of the same diseases in animals and humans may be a hint that S. gallolyticus subsp. gallolyticus is a zoonotic pathogen. This was supported by MLST, which typed a blood culture isolate of an animal farmer and excrements of the chicken of his laying hen farm with the same ST. Interestingly, the human fecal isolates which were identified in this case-control study were differentiated into the STs 3 and 7 and the new identified ST 105. ST3 and 7 were associated with human blood culture isolates and were also identified in cattle (unknown infective status). Thus, these results support the potential transmission of S. gallolyticus subsp. gallolyticus between animals and humans and highlight the zoonotic potential of the facultative pathogen. As described previously, it might be the case, as there seems to be some STs (as is the case of ST 7 or ST 3) that seem more associated with humans or animals whereas other isolates are more predominant in animals and humans (e. g. isolates of the clonal complex 45 and 6). A prevalence of 62.5% in human feces described in this study and a high prevalence in organic turkey flocks give rise to the question if S. gallolyticus subsp. gallolyticus belongs to the common gut microbiome of animals and humans. However, for this conclusion further systematic analyses have to be performed. Among other diseases, S. gallolyticus subsp. gallolyticus causes IE and is associated with colorectal diseases in humans. Both diseases are generally more often observed in male patients over 50 years old and various studies demonstrated the same positive association between the isolation of S. gallolyticus subsp. gallolyticus in men and the elderly population and IE, which was not identified in this study. To identify a relationship between the one-year increasing age and the detection of the bacterium in the digestive tract more people have to be tested. The facultative pathogen is the most frequently detected agent in cases of infective endocarditis in rural regions (especially in the cattle and milk production area) in the south of Europe, which cannot be observed in this case control study and should be figured out in further investigations. Joined together with living in the countryside, a close contact with animals was supposed to be a transmission source of the bacterium, it was demonstrated for the first time that a closer contact with animals is a significant exposure factor to be colonized with S. gallolyticus subsp. gallolyticus. It can be assigned as risk factors for the transmission of S. gallolyticus subsp. gallolyticus between animals and humans and its establishment in the human gut. Another interesting fact is the kind of animal species which come into contact with people. Although dogs and horses are described in the literature as a source of isolation and are common pets, it is not known whether these animals belonging to the volunteers are colonized with the bacterium. In this context, it is remarkable that fertilization of plants with the excrement of animals increases the risk of carrying S. gallolyticus subsp. gallolyticus significantly. Consequently, it is imaginable that S. gallolyticus subsp. gallolyticus may be transferred directly from animals to humans by smear infections and colonizes the gastrointestinal gut and, thus, can be accounted as a significant risk factor. However, the prevalence of S. gallolyticus subsp. gallolyticus in animals is still unknown. The bacterium has often been identified in, for example, pigeons, chicken and turkeys. Therefore, future studies should also include investigations of animal excrement in addition to human specimens to verify the prevalence in animals, too, and to estimate the real risk to human health. Derived from this hypothesis, not only the animal contact, but also the consumption of or the contact with contaminated food, such as red meat or milk products, may promote the colonization of the human gut and are propagated as exposure factors for the transmission of the bacterium from animals to humans. In total, statistical analyses demonstrate that raw food products play a minor role the colonization process. It was assumed that eating raw minced meat may promote the colonization of the human gut. More interestingly, because of the high frequencies of isolation in turkeys and laying hens, the consumption and processing of poultry and eggs and its function as a transmission source from animals to humans should be focused on in following research perspectives and participants should be ask about their habits in terms of processing and eating chicken meat and eggs. At the beginning of this case-control study the high prevalence was not known. The transmission from poultry to humans is well-known for Campylobacter. The main significant risk factor for the transmission of this species to humans is particularly bought fresh chicken (OR 5.80; 95%-CI: 2.11–15.93), whereby it was decreased by eating fruit, raw vegetables, high- fiber cereals, vitamins and acidified milk products. In summary, the case-control study conducted demonstrates a very high prevalence of S. gallolyticus subsp. gallolyticus in the gastrointestinal tract of healthy volunteers. In accordance with other researchers, it is essential to determine at least the subspecies or the biotype of the S. bovis strains to establish the identification of the frequency of S. gallolyticus in correlation with colorectal cancer, IE and other diseases, as well as its global impact to assess the risk to the human population. The data collected were evaluated with the help of multivariate statistical analyses to identify risk factors for the colonization of the human gut with the facultative pathogen. The simultaneous observation of exposure factors identified the closer contacts with animals and the usage of animals waste as significant risk factors for the detection of S. gallolyticus subsp. gallolyticus in human feces. Further investigations have to be performed to clarify the impact of chicken meat products and protective factors, such as vegetables. In addition, future studies should also include participants with e. g. gastrointestinal disorders associated with S. gallolyticus subsp. gallolyticus to detect the zoonotic pathogenicity of the Gram-positive bacterium to the health status of animals and humans and to determine the rate of S. gallolyticus subsp. gallolyticus fecal colonization. Another approach would be a prospective cohort study of people carrying S. gallolyticus subsp. gallolyticus or modifications of the gut microbiome along with environmental factors to detect the time-dependent influences on human health. In addition to the detected potential risk factors, the vitality outside the gastrointestinal tract is also important for the direct or indirect transmission of the bacterium between animals and humans or between the environment and animals or humans, and should be pointed out in future studies. # Supporting information This work was supported by Ruhr-Universität Bochum Medizinische Fakultät (FoRUM). We thank Philip Saunders (LL.B., B.Ed., FRSA) for his linguistic advice. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: J. Dreier CK TV. Data curation: J. Dumke. Formal analysis: J. Dumke. Investigation: J. Dumke. Methodology: J. Dumke OA. Project administration: J. Dreier CK. Resources: CK J. Dreier. Software: J. Dumke. Supervision: CK. Validation: J. Dumke TV. Visualization: J. Dumke. Writing – original draft: J. Dumke. Writing – review & editing: J. Dumke OA.
# Introduction The measurement of levels of HIV viremia in blood is the cornerstone of laboratory monitoring of anti-retroviral therapy (ART). In resource-limited settings, where most HIV infections occur, a mere 50% of patients on ART are currently routinely monitored using viral load testing. Improved access to this critical laboratory assay is urgently needed as global and national initiatives aim to provide medical care to increasing numbers of HIV-infected individuals, including those living in hard-to-reach geographical areas. HIV viral load determination is best achieved in fresh plasma specimens, with viral RNA quantification performed in laboratories using sophisticated equipment. Plasma must be separated from the cellular elements of blood within 6 hours post venesection, and then stored at -20°C or -80°C until testing in order to prevent degradation of RNA. In resource-limited settings, especially in primary health care centers where most HIV-infected patients seek medical care, the separation of plasma and its timely transport to a usually small number of distant laboratories is an overwhelming task. This is mainly due to weaknesses in health systems, including shortage of human resources, lack of logistical capacity and poor cold chain infrastructure. Many countries have implemented dried blood spots (DBS) to simplify the logistics of transporting clinical specimens to laboratories for viral load testing. Whole capillary or venous blood is spotted onto filter paper cards to prepare DBS. This method of whole blood collection has been very successfully implemented for diagnosing HIV-infection among infants younger than 18 months through the qualitative detection of HIV-1 proviral DNA. The comparative advantages of DBS include their ease of transportation, lack of cold chain needs and low biosafety requirements. However, HIV viral loads performed on DBS specimens are poorly correlated to those determined in plasma since whole blood contains also cell-associated RNA molecules and proviral DNA. Hence, viral load results from DBS specimens at clinicaly relevant threshold of 1000 copies/ml are frequently inaccurate and can mislead clinical-decision making. Viral load results from DBS specimens are more acurate at the threshold of 5000 copies/ml The cobas<sup>®</sup> plasma separation card (Roche Diagnostics GmbH, Mannheim, Germany) (PSC) is a novel specimen collection and transport device, that allows for simultaneous collection of whole blood and plasma separation without the use of additional equipment. These cards maintain all the comparative advantages of DBS, while providing plasma as a matrix for laboratory assays. A recent study under controlled conditions in South Africa showed that viral load results determined using PSC correlated well with those generated using plasma. Our study aimed at evaluating the performance of the PSC for HIV-1 viral load testing in patients attending primary health care facilities under routine health system conditions in Maputo City, Mozambique. # Materials and methods ## Study design and participants This cross-sectional study consecutively enrolled 613 HIV-infected adults receiving antiretroviral therapy (ART) between August 2018 and October 2018. The study was conducted at two primary health care centers in Maputo City, the Polana Caniço and the Primeiro de Maio Health Centers. These health centers provide a range of HIV-related clinical services, including ART initiation and follow-up. Both clinics routinely collect DBS specimens for viral load monitoring and send these to the Instituto Nacional de Saúde laboratories located approximately one hour driving distance from the health facilities for testing. For each enrolled study participant, a single viral load testing was conducted on plasma, DBS and cobas<sup>®</sup> PSC prepared from venous blood; and cobas<sup>®</sup> PSC prepared from capillary blood. Technicians performing laboratory testing for one method were blinded to results generated by other methods. Standardised data collection forms were used to collect demographic data from all participants, including sex, age, ART regimen and length of time on ART. ## Specimen collection The cobas<sup>®</sup> PSC (Roche Molecular Systems, Pleasanton, CA, USA) is a blood collection and plasma stabilizing device that facilitates the filtration of whole blood into dried plasma spots. The device comprises a porous membrane which allows only plasma to filter through, while retaining all other blood components. After drying the specimen on the card, it can be transported under a wide range of environmental conditions to a central testing laboratory. Detailed description of how PSC works is available in a recent publication of Carmona et al (2019). For this study, the cobas<sup>®</sup> PSC (Roche Molecular Systems, Pleasanton, CA, USA) was prepared using both capillary and venous blood from each study patient. The capillary PSC specimen was collected using a single use safety lancet blue blade with a penetration depth of 2.0 mm for finger puncture. After puncturing, three capillary tubes marked 140μl provided with the test kit was used to collect a total of 420 μl of whole blood per card from each patient. After collection, blood in each capillary tube was transferred onto each of the three delineated areas of the card. In addition, 6.0ml venous blood was collected in a BD Vacutainer<sup>™</sup>K<sub>2</sub>EDTA (Becton, Dickinson and Company, 1 Becton Drive, Franklin Lakes, NJ 07417–1880, US) tube from each patient. Pasteur pipette was used to transfere one to two drops of whole blood onto each of delineated areas to get five full spots of the DBS card (Ahlstrom Germany GmbH) and an additional cobas<sup>®</sup> PSC was prepared using venous blood. The remnant anticoagulated whole blood was transported to Instituto Nacional de Saúde laboratories within 6 hours post venesection, for plasma separation and storage at -80°C until viral load testing. All three cards (2 cobas<sup>®</sup> PSC and 1 DBS) were placed overnight in a single-usage drying rack at room temperature. After drying, the cards were packed in separate gas impermeable zip-lock bags with desiccant and shipped to Instituto Nacional de Saúde laboratories for testing. All cards were stored at -80°C upon arrival at the testing laboratory. All staff involved in specimen collection for the study routinely collects DBS specimens for viral load, however in the context of this study were trained on PSC and DBS specimens collection and packing. Testing of study specimens (plasma and cobas<sup>®</sup> PSC) followed the routine schedule of the laboratory, with DBS specimens having the priority as part of the medical management of patients. ## Specimen preparation Whole blood anticoagulated specimens were centrifuged at 800-1600g for 20 minutes at room temperature to obtain plasma, of which 1100μl was used for viral load testing. One spot of DBS was eluted in phosphate buffered saline, pH 7.4 (1X) and incubated at room temperature for 30 minutes. After the incubation, specimens were manually homogenized and immediately loaded into the CAP/CTM v2 for testing using the free virus elution Roche protocol. For the cobas<sup>®</sup> PSC, one spot was eluted in Sample Pre-Extraction (SPEX) solution (Roche Molecular Systems, Pleasanton, CA, USA), incubated in a thermomixer at 56°C and 1000 rpm for 10 minutes and then immediately loaded into CAP/CTM96 for testing. ## Viral load testing Viral load testing on DBS specimens was performed as part of the routine patient monitoring, whereas plasma and PSC specimens were tested for study purposes. Testing for all specimens was carried out at the Instituto Nacional de Saúde using the Roche CAP/CTM 96 HIV-1 Quantitative Test v2 (Roche Molecular Diagnostics, Branchburg NJ, USA), according to the manufacturer’s instructions. The test definition files used for HIV viral load measurement was HI2PSC96 for cobas<sup>®</sup> PSC specimens, IFS96CDC for DBS specimens and HI2CAP96 for plasma specimens. Interpretation of viral load results was performed according to manufacturers’ instructions, which establish 20, 400 and 738 copies/ml of viral RNA as the low limit of detection for plasma, DBS and PSC specimens, respectively. These differences on limit of detection are prominent, but not clinically relevant on perspective that all specimens with VL bellow 1000 copies/ml will have the same approach. However, a lower limit of detection of VL in PSC specimens can be used to monitor VL in patients with low-level viremia. The reference laboratory routinely participates in and passed external quality assessment programs for viral load (provided by the US Centres for Disease Control and Prevention, Atlanta, USA) prior to and during the study period. ## Statistical analysis When viral load results (copies/ml) were log10-transformed, those with non- quantifiable viral load results with values below the limit of detection (LOD) for each specimen type, were assigned a value of 1 copies/ml to enable quantitative log10 transformation and for graphic visualization. Values from plasma testing were regarded as the reference for calculating sensitivity, specificity and misclassification of results obtained from DBS and cobas<sup>®</sup> PSC specimens. A threshold of 1000 (3 log) copies/ml were considered for these calculations. All samples with viremia below 1000 copies/ml, regardless of the specimen-type used for testing, were considered as being virally suppressed as per WHO guidelines. Concordance correlation and Bland-Altman analyses were performed to determine precision and agreement between plasma, DBS, and venous and capillary cobas<sup>®</sup> PSC. Specimen results that generated a reportable result across all specimen types were included in concordance correlation and these results \>1 000copies/ml were included in the Bland-Altman analysis. Scatter plots were presented for visual representation of log transformed and untransformed data across specimen types. Analyses were performed using STATA 14.2 (StataCorp, Texas, USA) and MedCalc Statistical Software version 16.4.3 (MedCalc Software bvba, Ostend, Belgium). ## Ethical considerations Ethical approval for the study was obtained from Mozambique’s National Health Bioethics Committee, with the reference number 297/CNBS, in July 2018. Written informed consent was obtained from each participant prior to conducting any study procedure. All patients were given a copy of the signed consent form, which contained information about the study as well as contacts details of the principal investigator and the ethics committee. # Results ## Study population A total of 613 consecutive patients attending the two health care facilities for routine ART monitoring between August 2018 and October 2018 were enrolled in the study. The median age of study participants was 41 years, and 65.0% were female. Most patients (93.6%, 574/613) were receiving first line ART and the median time on ART was 51 months. The plasma viral load among participants showed that 50.9% (312/613) had non-quantifiable virus, 2.9% (18/613) had a viremia between 1000 and 10,000 copies/ml, and 15.5% (95/613) had a viremia \>10,000 copies/ml. From the 613 patients included in the study, 2444 specimens including DBS, plasma, venous PSC and capillary PSC were collected, and 2407 (98.5%) generated reportable results. Overall, 0.41% (10/2444) specimens were rejected at the reference laboratory due to poor quality and 1.1% (27/2434) failed to report a result due to equipment failure. From the rejected specimens, four were DBS, four venous PSC and two capillary PSC. For all specimens, the reason of rejection was spot not completely full, which means that the quantity of blood transferred onto the delineated area of the card was not sufficient. ## Specimens testing At the time of the study, the standard-of-care for HIV viral load testing used for patient management in Mozambique was DBS, which took preference upon specimen arrival in the testing laboratory. The median number of days between specimen collection and testing for DBS was 48 days (range: 1–91), for plasma was 57 days (range: 17–129), for venous cobas<sup>®</sup> PSC was 66 days (range 13–130) and for capillary cobas<sup>®</sup> PSC was 58 days (range 13–130). No bias in batch testing for any particular viral load category was evident from internal review of measuring the difference between DBS, cobas<sup>®</sup> PSC and plasma over the time of storage at -80°C and testing. The analysis of viral load difference plotted against number of days to testing, to ensure -80°C storage did not affect the viral load yielded a random scatter, showing no trend in specimen type and result difference with the reference plasma. In addition, the laboratory work flow was not dependent on specimen type for loading specimens into the equipment. The CAP/CTM instrument software used in the laboratory was adaptable to plasma, DBS or cobas<sup>®</sup> PSC testing protocols, and various specimen types were tested in the same run on the same day. ## Performance of the cobas<sup>®</sup> plasma separation card illustrates the absolute HIV VL results from study participants and shows that DBS specimens with VL \>1000 copies/ml underestimate the true HIV viral load as measured by the reference plasma specimen. In addition, the DBS reports several HIV VL results \>1000 copies/ml when the plasma VL reference values are undetectable. The cobas<sup>®</sup> PSC prepared either from capillary or venous blood specimens, generate the same trend of higher values than plasma VL in specimens \>1000 copies/ml. The clinical relevance was further investigated by measuring the sensitivity, specificity and misclassification rate as outlined in and the absolute bias by Bland-Altman and concordance correlation in. The sensitivity of the cobas<sup>®</sup> PSC (99.8%, capillary and 100% venous) in identifying virological failure and success at the clinically relevant threshold of 1000 copies/ml, as determined by plasma, was statistically significantly better than DBS (97%), as shown by non-overlapping confidence intervals. This is similarly evident with the specificity, with DBS (81.4%) performing poorly compared to the cobas<sup>®</sup> PSC (97.3%, capillary and 98.2% venous). This performance is further compounded by the higher misclassification rate for DBS (5.9%), compared to lower rates of misclassification for the cobas<sup>®</sup> PSC (0.7%, capillary and 0.3% venous). Among the quantifiable VL results, the concordance correlation coefficient showed poorer accuracy and precision for measuring VL using DBS (0.732) than cobas<sup>®</sup> PSC (0.876 capillary and 0.880 venous). The bias among quantifiable results across all specimen types again outlines the reduced accuracy of the DBS compared to the reference plasma. DBS overall generates 0.86 log copies/ml lower results than plasma VL, with several outliers (n = 9) in the clinically relevant range at 1000 copies/ml across the 107 paired specimens, and high variability (standard deviation of the bias \>1.0log copies/ml) in the bias. The cobas<sup>®</sup> PSC (either venous or capillary collection) displayed minimal bias (\<0.09 copies/ml) compared to plasma with only 3 outliers across 107 data pairs. further illustrates the potential bias between DBS and cobas<sup>®</sup> PSC capillary, that would be anticipated within a testing program where cobas<sup>®</sup> PSC were to replace DBS in ART monitoring of routine patients. Approximately 1.0 log copies/ml difference in VL results would be reported across the quantifiable range of VL values for patient care, with the cobas<sup>®</sup> PSC overall generating higher VL values than DBS and therefore aligning the true VL result from a cobas<sup>®</sup> PSC more closely with plasma VL. Based on the data presented in, and the good agreement between cobas<sup>®</sup> PSC and plasma, it is evident that the cobas<sup>®</sup> PSC is a matrix suitable for use for VL testing. Based on the sensitivity, misclassification rate and absolute bias analysis, we identified two discordant results for venous PSC, three for capillary PSC and 36 for DBS. # Discussion The laboratory monitoring of ART in many sub-Saharan settings is currently performed on DBS, as this type of specimen overcomes many of the logistical challenges faced by resource-limited health systems. Our study shows that DBS constitutes a sub-standard alternative to plasma viral load, and demonstrates that accurate determination of viral load is feasible when cobas<sup>®</sup> PSC is used to obtain patient specimens under field conditions. The relatively low bias (\<0.09 copies/ml) displayed for venous and capillary cobas<sup>®</sup> PSC at the plasma 1000 copies/ml range is an acceptable bias, and in line with the manufacturer’s definition of acceptability of \<0.3log copies/ml. The bias generated by cobas<sup>®</sup> PSC is similar to the viral load differences reported previously comparing split plasma specimens across different testing platforms or assay versions, and similar to other plasma separation devices applied to HIV VL testing. On the contrary, the bias seen on DBS viral load (\>1.0log copies/ml) in our study is notably higher than the values considered to be acceptable, and similar to other findings of the performance of DBS for HIV VL testing. The better performance of the cobas<sup>®</sup> PSC is further evidenced by the very low misclassification rates of viral load values below 1000 copies/ml, with rates of 0.7% and 0.3% for capillary and venous cobas<sup>®</sup> PSC specimens, respectively, as opposed to 5.9% for DBS specimens. This is most likely because the use of the cobas<sup>®</sup> PSC eliminated the interference of cell- associated viral nucleic acid and the over-quantification of HIV genetic material, a factor that impacts the measurement of viral load around the critical threshold of 1000 copies/ml. The low limit of detection of 738 copies/ml in the cobas<sup>®</sup> PSC specimens is considerably higher than the 400 copies/ml and 20 copies/ml obtained with the DBS and fresh plasma, respectively. These limits are defined for the CAP/CTM system used in our study, but the differences are likely to be seen in other testing systems. Nevertheless, the limit of detection for the cobas<sup>®</sup> PSC is still below the threshold of 1000 copies/ml used to currently define viral suppression as per WHO recommendations. Moreover, the limit of detection defined for DBS specimens does not take into account the loss of specificity caused by cell-associated nucleic acid in whole blood. In most countries of sub-Saharan Africa, specimens for HIV viral load testing are collected at health facilities and sent to a relatively small number of reference laboratories for testing. For example, according to the DISA Lab Viral Load National Database, in Mozambique ten Laboratories were tasked with testing the 801,155 specimens collected in 2018. Consequently, most laboratories have a relatively long turnaround time for viral load testing. In this context, it is critical that specimens be stable for periods of at least two months. In this study, the cobas<sup>®</sup> PSC was tested between 13 and 130 days post venesection without a loss in RNA stability, but with storage prior to testing at -80°C. In this study we did not investigate the stability of PSC in a variety of temperature ranges, but previous stability data has shown that the cobas<sup>®</sup> PSC can be stored for 56 days in a temperature range between 2 to 30°C without significant variance on the HIV viral load result. Nurses in two primary healthcare centres performed the collection of the cobas<sup>®</sup> PSC specimens for this study, attesting to the feasibility of implementing this new method in resource-limited health systems. The procedures for collecting cobas<sup>®</sup> PSC specimens are similar to those for DBS, which health professionals throughout sub-Saharan Africa have been using in HIV viral load testing and early infant diagnosis for many years. Additionally, in this study viral load testing using cobas<sup>®</sup> PSC specimens was executed in a laboratory that routinely tests for viral load and early infant diagnosis, with no additional equipment required. # Conclusion In conclusion, the implementation of the cobas<sup>®</sup> PSC in primary healthcare centers is feasible, when prepared using either capillary or venous blood. HIV viral loads determined using this new technology for specimen collection and transport are accurate, and therefore, cobas<sup>®</sup> PSC constitutes a good alternative to fresh plasma for the laboratory monitoring of patients undergoing ART in resource-limited settings. We would like to gratefully acknowledge Lara Noble and Taskani Mhlongo from the South African iLEAD team within the Department of Molecular Medicine and Haematology at the University of the Witwatersrand in Johannesburg, South Africa for cobas<sup>®</sup> PSC data analytics and training. We also acknowledge the laboratory staff and nurses at the Primeiro de Maio and Polana Caniço Health facilities, Maputo, Mozambique, for specimen collection. [^1]: All authors declare no conflicts of interest.
# Introduction Hypertension (HTN) is a significant public health concern globally, with its prevalence continually rising in developing countries. According to the World Health Organisation, an estimated 1.28 billion adults aged 30–79 years worldwide have hypertension, with about two thirds living in low- and middle-income countries such as Uganda. Moreover, nearly half (46%) are unaware that they have hypertension, 42% are diagnosed and treated, and only 21% have it under control. In Uganda, the prevalence of hypertension varies with regions of the country, with the highest prevalence registered in the central region, estimated at about 34.3% and the lowest in Northern Uganda, estimated at 22.0% according to a national epidemiological study published in 2018. In addition, among private patients in the central region, the prevalence of hypertension was estimated at 41.6% and elevated blood pressure (BP) at 37.6%, but only 18.3% achieved control. Patients with rheumatic and musculoskeletal diseases (RMDs) such as rheumatoid arthritis (RA), systemic lupus erythematous (SLE), osteoarthritis (OA), gout and Sjogren’s syndrome are at an increased risk of developing HTN, with chronic inflammation as the main driver. Chronic inflammation causes damage to the blood vessels and contributes to salt retention in the body through the inflammatory cytokines such as interleukin-6 and tumor necrosis factor-alpha stimulating the renin-angiotensin-aldosterone system. Other factors associated with the development of HTN among patients with RMDs include traditional cardiovascular risk factors such as genetic predisposition, advanced age, obesity, physical inactivity and use of nonsteroidal anti-inflammatory drugs and corticosteroids. Anyfanti and colleagues, in a study conducted among patients with RMDs attending Rheumatology Outpatient Clinics in Greece, showed a high prevalence of HTN, estimated at 54.5%, with 21.7% being unaware that they had HTN. In addition, patients with RMDs such as RA are at a 50% increased risk of cardiovascular related morbidity and mortality compared to the general population. However, HTN among most patients with RMDs is undiagnosed, and therefore poorly controlled, yet it is the main risk factor for developing cardiovascular events. Mandatory regular screening for HTN among all patients diagnosed with RMDs is a reliable means for early detection, thus early and effective management of the condition. The burden of hypertension among patients with RMDs in Uganda remains unknown. However, many cases of hypertension go undiagnosed, and thus treatment is not availed, increasing mortality and morbidity in this population. Therefore, in the present study, we aimed to determine the prevalence of high blood pressure (HBP), awareness, treatment, and BP control among patients with RMDs seen in a Rheumatology clinic in Uganda. # Materials and methods ## Study design, setting & population Using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines, we conducted a cross-sectional analytical study. Data were collected prospectively between January and April 2022. The study was conducted in the Rheumatology Clinic of Mulago National Referral Hospital (MNRH), Kampala, Uganda. MNRH, located in the capital city, Kampala, is the largest public health facility in Uganda serving as a national specialized hospital with over 1800-bed capacity. The clinic is run by a rheumatologist (MK) assisted by other medical officers and registered nurses and has over 200 patients with various RMDs. The clinic is the largest in the country and serves as the national referral center for patients with RMDs across the country. We included patients of Ugandan nationality aged 18 years and above, with a confirmed diagnosis of RMD, irrespective of the gender, attending the MNRH Rheumatology Clinic after having given the consent to participate. The RMDs of interest were: RA classified according to 2010 European League Against Rheumatism (EULAR)/ American College of Rheumatology (ACR) criteria, SLE classified according to the 2019 EULAR/ACR criteria, gout classified according to the 2015 EULAR/ACR criteria, Sjogren’s syndrome classified according to the 2016 EULAR/ACR criteria, OA classified according to the 1986, 1990 and 1991 EULAR/ACR criteria osteoarthritis of the knee, hand and hip respectively, spondyloarthropathies classified according to the 2011 ASAS criteria and Anti- neutrophil cytoplasmic antibody (ANCA)-associated vasculitis classified according to the 2022 EULAR/ACR criteria. We excluded patients who declined to participate or were unable to give informed consent. ## Data collection Data classified into socio-demographic, clinical (RMD diagnosis, duration and treatment, BP readings, HBP and HTN awareness, HTN treatment and BP control) and anthropometric (weight, height, and waist-hip ratio) were collected using an interviewer administered semi structured questionnaire and direct measurements taken by the interviewers. The questionnaires were printed out and administered by the interviewers to the respondents at the Rheumatology clinic, to ensure that the respondents understand the questions. Three BP measurements were taken on the same day on two occasions, four hours apart, on both arms using a manual sphygmomanometer because it is the recommended method for accurately assessing BP. Each participant was at the triage station for at least five minutes to allow them to relax and for their BP to stabilize to avoid falsely elevated readings. BP was taken when the participant was seated upright with his/her arm at the heart level in a quiet triage room to ensure accurate readings, and it was done by routine care nurses to reduce the likelihood of white coat hypertension. Office BP was determined as an average of the three readings. Awareness was assessed by asking the patient whether they were hypertensive or not using a single question. ## Statistical analysis The sample size of 194 participants with RMDs was calculated using the modified Kish and Leslie formula for finite population size, with an estimated prevalence of HBP of 54.5% from a similar study conducted in Greece, a population size of 200, and type 1 error of 5%, and a standard deviation at 95% confidence interval (1.96). Adjusting for the finite population, using Slovin’s formula (N/(1+Ne<sup>2</sup>), where N = 150, and e is 5%, a final sample size of 130 participants was obtained. Data was cleaned, coded and entered into Microsoft 2016, and exported for analysis using the STATA version 17.0 analysis software. Numerical data were assessed for normality using Shapiro-Wilk test. Continuous variables were expressed as mean and standard deviation for parametric variables or median and interquartile range for non-parametric variables. Categorical data were presented as frequencies and percentages. Chi-square or Fischer’s exact tests were used to compare categorical variables as appropriate and student t-tests or Mann-Whitney U for numerical variables. Modified Poisson regression was used to determine the factors associated with HBP and association was measured using the prevalence ratio because the prevalence of HBP is high, which may bias odds ratio values. We conducted bivariate analysis and variables with p\<0.20 were considered for multivariate analysis. In the multivariate analysis, we used manual backward elimination method until all the variables in the model had p-value ≤ 0.05. We assessed for interaction by forming product terms and performed a chunk test. We assessed for confounding by considering a percentage change of \> 10% in the crude and adjusted prevalence ratios. The goodness of fit of the model was assessed using Hosmer-Lemeshow goodness of fit test. Variables with p\< 0.05 were considered statistically significant. ## Operational definitions HBP was defined as an office systolic blood pressure (SBP) ≥ 140mmHg and/or diastolic blood pressure (DBP) ≥90 mmHg. HBP awareness, as knowledge of the participant about being hypertensive based on previous diagnosis by a qualified health care worker. HBP treatment was defined as a patient being on antihypertensive medicine. Hypertension, as a patient self-reporting to be hypertensive or ongoing antihypertensive drug treatment. Uncontrolled hypertension, as a patient with a known diagnosis of hypertension on antihypertensive medicine whose blood pressures are ≥<u>1</u>40/90 on the day of the current clinic visit (day of the interview). ## Ethical consideration Ethical approval was sought from Mulago Hospital Research and Ethics Committee (Approval number: MHREC 2162). Approved by MHREC, a modification was made to the initial protocol from targeting only patients with RA to include patients with RMDs. All participants provided written informed consent by signing a consent form appended to the questionnaire. The study was conducted in observance of the Declaration of Helsinki-2013. # Results Of the anticipated 130 participants, 110 participants turned up at rheumatology clinic during the study period. However, only 100 participants met the eligibility criteria and were enrolled, resulting in a positive response rate of 100/110. No significant differences were observed between the participants and non-participants in terms of age, gender, or disease severity. The majority of participants were female (84%, n = 84), with mean age of 52.1(SD: 13.8) years and median BMI of 28 kg/m<sup>2</sup> (IQR: 24.8 kg/m<sup>2</sup>-32.9 kg/m<sup>2</sup>). Fifty-nine participants were employed (59%), from urban residency (69%, n = 69), majority had no history of smoking (96%, n = 96) or use of alcohol (67%, n = 67) and forty-nine had family history of hypertension (49%). Eight (8%) participants had diabetes mellitus. Overall, the median SBP and DBP were 129.3(IQR: 118.3–144.7) mmHg and 81.3(IQR: 74–90.7) mmHg, respectively. The trends of systolic and diastolic blood pressure in participants are shown in. summarizes socio-demographic and clinical characteristics of all the study participants. The prevalence of HBP was 61% (n = 61, 95% CI: 51.5–70.5). Out of the 61 participants with HBP, the majority (77%, n = 47, 95% CI: 66.5–87.6) were aware they had HTN. The prevalence of HTN was 47% (n = 47, 95% CI: 37.2–56.8), and none had it under control. The most frequent RMDs were RA (80.3% in patients with HBP versus 76.9% in those without) followed by OA (9.8% in patients with HBP versus 12.8% in those without), gout (8.2% in patients with HBP versus 0% in those without) and other RMDs which are SLE, spondyloarthropathies, sjogren’s syndrome and ANCA-associated vasculitis (1.6% in patients with HBP versus 10.3% in those without). Individuals with HBP were older (p = 0.001) and had higher BMI (p\<0.001) and median waist-to-hip ratio (0.86 vs 0.8, p = 0.009) compared to those without. Factors independently associated with HBP were age 46-55years (adjusted prevalence ratio (aPR): 2.5, 95% CI: 1.06–5.95), 56–65 years (aPR: 2.6, 95% CI: 1.09–6.15), \>65 years (aPR: 2.5, 95% CI: 1.02–6.00), obesity (aPR: 3.7, 95% CI: 1.79–7.52), and overweight (aPR: 2.7, 95% CI: 1.29–5.77). # Discussion In this study, we aimed to determine the prevalence of HBP, awareness, treatment, and blood pressure control among patients with RMDs seen in a Rheumatology clinic in Uganda. However, comparability of our findings with data from the literature of RMDs is limited because, apart from one study conducted in Greece, this is the first study on HBP among patients with RMDs in Uganda and sub-Saharan Africa. We found the prevalence of HBP to be 61%, with the majority (77%) of those aware that they had HTN, however, only 44.7% of participants had it under control. These findings align with results of the study conducted by Anyfanti and colleagues in Greece among patients with RMDs attending Rheumatology Outpatient Clinics. The study reported a high prevalence of HTN among patients with RMDs, estimated at 54.5%, with 21.7% being unaware that they had HTN. Similarly, a study by Protogerou and colleagues investigating HTN among patients with RA in Greece registered a high awareness level, estimated at about 80%. As such, HTN is a common comorbidity in patients with RMDs, with high awareness of the condition attributed to factors such as increased awareness campaigns, improved healthcare access and effective communication between patients and health care providers. The limited BP control observed in our study is comparable to other studies conducted in Greece, which showed low BP control rates estimated at only 48.6% and 29% among patients with RMDs and RA respectively. Poor BP control in patients with RMDs is attributed to various factors, including poor patient adherence to treatment, healthcare provider awareness, and suboptimal therapeutic approaches such as inadequate medication titration. In our study, obesity was independently associated with HBP, consistent with previous studies conducted in both patients with RMDs and the general population. Notably, we found that obesity exhibited a 1.5 times stronger association with HBP compared to overweight participants. The strong association between obesity and HBP underscores the importance of lifestyle modifications, including weight management, in the prevention and management of HTN in RMD patients. Obesity is closely linked to metabolic abnormalities and increased risk of cardiovascular diseases, including HTN. Moreover, the burden of high BMI-related deaths and DALYs increases with age, especially in males, as evidenced by the 26.8% and 12.7% rise in the age-standardized rate of high-BMI- related DALYs for males and females, respectively, between 1990 and 2017. Our study revealed a significant association between increasing age from 36 years and HBP among participants. This finding could be explained by the higher proportion of older participants in our study. Aging is characterised by physiological changes in the cardiovascular system, including arteriosclerotic structural alterations and calcification, leading to large artery stiffness. Consequently, the blood vessels become less flexible and more rigid, resulting in elevated resistance to blood flow and higher BP levels. Moreover, aging is associated with chronic inflammation and oxidative stress, which, in conjunction with the chronic inflammation associated with RMDs further augments the risk of developing HBP. Inflammation is known to be a major driver of HTN in patients with RMDs with elevated C reactive protein (CRP) an inflammatory marker and endothelial damage. Comparable to our findings, a recent study in China showed that risk of HBP increased from age 35 years in the general population. In addition, Anyfanti and colleagues also reported a high prevalence of HTN among the elderly patients with RMDs. Thus, our study underscores the role of age as a contributing factor to the development of HBP. Rheumatologists, allied health professionals, and stakeholders involved in the care of patients with RMDs should consider incorporating lifestyle changes in patient care to reduce body fat, thereby preventing obesity. In addition, strategies such as health education and improved communication between healthcare providers and patients have the potential to empower patients with RMDs to be aware of their HTN status and make informed decisions regarding the condition, as well as to adopt healthy habits. It is crucial for elderly patients with RMDs in the rheumatology clinic to be closely monitored with more frequent blood pressure measurements to ensure early diagnosis and timely treatment of HTN. Our study was not without limitations. Despite our study site being at a national referral hospital serving patients from across the country, it was a single centred study. Therefore, our findings are not generalizable to the Ugandan population. In addition, we used a non-standardized questionnaire which was however, developed based on expertise of physicians experienced in the management of RMDs in Uganda. We acknowledge that not measuring CRP levels among patients with RMDs limited our ability to directly investigate the potential link between inflammation in RMDs and HTN. Measuring CRP levels in future studies will be valuable in understanding the role of inflammation in the development and progression of HTN among patients with RMDs. In addition, we did not assess the causality of the associations observed because of cross-sectional study design. This was also the first study to assess the prevalence and factors associated with HBP in patients with RMDs in Uganda. # Conclusion This study shows a high burden of HBP among patients with RMDs in Uganda, with a considerable number being aware of their condition but experiencing poor BP control. The high burden was associated with high BMI and increasing age in this population. Thus, lifestyle modifications including weight management are crucial in preventing and managing HBP and its complications such as future cardiovascular disease and deaths. Continuous medical education and tailored guidelines for cardiovascular risk in RMDs are recommended. # Supporting information The authors acknowledge the administration of Mulago National Referral Hospital as well as the study participants and clinical care nurses. [^1]: The authors have declared that no competing interests exist.
# Introduction Structural aminopolysaccharide chitin is recognized to occur as the basic component in both non-mineralized and mineralized skeletal formations of the cell walls of diverse fungi, diatoms, sponges, corals, annelids, molluscs, and arthropods (see for review). This ancient biopolymer is typically cross-linked due to the complex linkage with pigments, lipids, other polysaccharides, peptides and proteins. As universal template in biomineralization, chitin plays a significant role in formation of calcium- (in molluscs) and silica-based (in diatoms and glass sponges) biominerals. Interaction between diverse organic and inorganic molecules listed above and chitin is often the key way to rigidification of broad variety of skeletal constructs in invertebrates. Mechanical stiffness of skeletons remains to be crucial for surviving of sponges as sessile and filtering organisms. Chitin plays important role in rigidification in some sponges in both non-mineralized and mineralized states have been recently reported in representatives of marine (see for review) and fresh water sponges. In some demosponges, chitin has been confirmed as template for formation of biomineralized structures in the form of aragonite-silica- chitin composites. Silica-chitin-based skeletal structures have been identified in glass sponges. Intriguingly, isolation of chitin-based structures have been never reported in calcarean sponges (class Calcarea), although chitin synthase genes have been already detected in two species. Furthermore, there is no reports in the literature about the existence of chitin in skeletons of the sponges belonging to the class Homoscleromorpha. Pure chitin, which was traditionally extracted at large scale from fungi and crustaceans, received special attention in the modern industry. The potential of chitin application as adsorbent and biomaterial for biomedical purposes is well known. Recent novel information concerning applied potential of chitin can be found in numerous reviews including few reports. It is to note here that industrially produced chitin can be mostly isolated in the form of powders and flakes. Interestingly, sponges originally produce 3D chitinous scaffolds, which are fibrous and macroporous due to their functional role in the skeletons of these filter-feeding invertebrates. This feature has been observed also in the Cambrian fossil demosponges including Vauxia gracilenta. Moreover, chitinous skeleton of the demosponges origin resembles the style and form of the source sponges. The nanofibrillar organization together with unique mechanical and thermal properties of chitinous skeletal scaffolds is the key to their successful applications in tissue engineering and modern biomimetics. However, up-to-date this progress has been based exclusively on chitin isolated from diverse representatives of only one order of marine demosponges, the Verongiida order. The idea to propose the use of this feature for systematics of all sponges related to Verongiida order logically appeared. However, our recent findings of chitin in fresh water demosponges of non-verongiid origin stimulated the monitoring of chitin also in other demosponges genera, especially in marine species which arose prior to fresh water sponges. Consequently, two years ago we started intensive monitoring of diverse non- verongiid marine demosponges with the aim to find, purify and characterize and identify chitin from different orders order of marine Demospongiae. Especially, we have taken advantage of the worldwide distribution of the sponges of the order Poecilosclerida which includes four suborders and 25 families. This order is recognized as the largest and most diverse among Demospongiae orders with species occurring in all oceans from shallow water habitats to deep seas. Preliminary investigations with respect to chitin identification in 60 diverse demosponges recently collected in the Red Sea showed that such representatives of Poecilosclerida as Acarnus woffgangi and Echinoclathria gibbosa should contain chitin within their skeletons due to their characteristic insolubility in 2.5 M NaOH solution. Members of the genus Acarnus (Gray, 1867) (Porifera: Poecilosclerida) belong to the family Acarnidae (Dandy, 1922) with 26 representative species. Acarnus wolfgangii was described for the first time by Conrad Keller in 1889 as a sponge having rigid fiber-based skeleton which show network-like architecture and rich on spongin. The diameter of skeletal fibers was measured as about 0.05 mm and the air dried sponge was stone hard. Echinoclathria (Carter, 1885) is a genus of demosponges belonging to the family Microcionidae (Carter, 1885). This family includes two subfamilies, Clathriinae and Ophlitaspongiinae, with nine valid genera and 524 valid species living worldwide in shallow waters with a few records from deeper seas. Unfortunately, with exception of one report on the identification of secondary metabolites from E. gibbosa no additional reports are available on this species. Here, we represent the first study on isolation of chitin from the skeleton of A. wolffgangi and E. gibbosa demosponges according to the step-by-step approach and identification of this structural aminopolysaccharide using corresponding bioanalytical methods in comparative modus. # Materials and methods ## Biological materials, sample collection and preparation Acarnus wolffgangi Keller (Demospongiae, Poecilosclerida, Acarnidae). The sponges was collected by SCUBA diving in July 2017 from the eastern side of the Small Giftun Island (N 27°11′12.9′′ E 33°59′03.1′′) in the Egyptian Red Sea at a depth of 28 m. The sponge is yellowish in color and forms a massive crust with clathrate surface. The skeleton is formed by a reticulation of skeletal fibers, cored by thick smooth styles and echinated by smooth cladotylotes. The fiber diameter measuring between 50 and 100 μm. At the surface, there are tangentially scattered tylotes. The spicules include: ectosomal tylotes with microspined tyles measuring 215–255 x 2–3 μm, smooth and curved choanosomal styles measuring 300–330 x 15–20 μm and the cladotylotes existing in two distinct shapes and sizes, both with smooth rounded tyles at one end, the larger with three strong hooks at the opposite ends with overall dimensions of 220–290 x 10 μm, and the smaller with four hooks, smooth or occasionally with spined shaft with overall dimensions 90–125 x 3–6 μm. The toxas exist in three distinct categories, including toxas I with oxhorn shape measuring 90–115 x 3–5 μm, toxas II, which is thin with shallow curve measuring 55–65 x 1 μm, and oxea-like toxas III which is barely curved measuring 500–620 x 3 μm; palmate isochelae, 15–20 μm. We were able to study the type material of Acarnus wolffgangi, kept in the collections of the Museum für Naturkunde Berlin, ZMB 1498 and 2922. The Red Sea specimen conforms closely in shape, skeleton and spicules, with this type. The Red Sea voucher was kept in the Naturalis sponge collection under registration number ZMA Por. 16636 measuring 10 x 5 by 1 cm in size. Another voucher was kept at the Red Sea Invertebrates Collection at the Department of Pharmacognosy of Suez Canal University under the code number RS-23. Echinoclathria gibbosa (Keller, 1889). The sponge was collected in July 2017 from Hurghada (N 27°17′0.53′′ E 46°22′0.8′′) in the Egyptian Red Sea at a depth of 30 m. The live sponge is blood-red in colour and forms a mass of long branches, which anastomose infrequently. The length of the branches in the voucher sample measuring up to 20 cm with varied thickness due to the irregular outline of the branches from 1 to 2 cm. The surface is pitted and clathrate. The skeleton displayed square- to round-meshed reticulation of skeletal fibers, cored by 1–5 spicules in cross section. The meshes measure 150–300 mm, while connecting fibers measure 10–25 mm in diameter, respectively. The surface skeleton showed a tangential arrangement of loose styles. The spicules are of choanosomal styles and ectosomal subtylostyles, existing in two or three diffeent sizes, but these are not clearly differentiated and measuring about 125–360 x 1–4 mm. The microseleres are very thin, shallow-curved toxas and measuring up to 20 mm in length. The specimen was compared with a slide of the Berlin Museum type and found to conform closely to it. The voucher is registered in the collections of the Netherlands Centre of Biodiversity Naturalis under number ZMA Por. 19793. Another voucher was kept at the Red Sea Invertebrates Collection at the Department of Pharmacognosy of Suez Canal University under the code no. RS-46. ## Isolation of chitin skeleton from A. wolffgangi and E. gibbosa The isolation of chitin-based skeletons from A. wolffgangi and E. gibbosa were carried out as previously reported. The protocol consists of four steps: firstly, the sponge skeletons were placed, separately, in deionized water at room temperature for 1 h in order to remove possible water-soluble sediment particles and salts. Then, the samples were treated with 20% acetic acid at room temperature for 12 h in order to remove residual carbonate-based debris (microfragments of mollusc shells and crustacean carapaces) from the skeleton of A. wolffgangi and E. gibbosa. Afterwards, the samples were washed several times with deionized water until achieving a pH of 6.5 followed by treatment with 2.5 M NaOH at 37°C for 72 h to remove pigments and proteins. The siliceous spicules were observed in the samples after 72 h of alkali treatment, thus thorough desilicification was needed. Consequently, after alkali treatment, samples were accurately rinsed with deionized water and placed in a plastic vessel containing appropriate amount of 2% hydrofluoric acid (HF) solution. The vessel was covered in order to prevent the evaporation of HF. Desilicification was conducted at room temperature for 12 h. The effect of alkaline and strong acidic treatments on the structure of skeletons of both demosponges was investigated using optical and fluorescence microscopy (Keyence BZ-8000). Finally, the isolated materials were washed several times with deionized water up to pH of 6.5. The fibrous scaffolds were put into the 250 mL large GLS 80 Duran glass bottles containing deionized water and stored at 4°C for further analyses. ## Light and fluorescent microscopy imaging Collected sponge samples and isolated chitinous scaffolds from A. wolffgangi and E. gibbosa have been observed using BZ-9000 microscope (Keyence) in white light as well as in fluorescence modes. ## Calcofluor White (CFW) test In order to evaluate the localization of chitin in the demineralized skeleton of A. wolffgangi and E. gibbosa, Calcofluor White (Fluorescent Brightener M2R, Sigma) was used as a fluorescent dye for staining of β-(1→3) and β-(1→4) linked polysaccharides. After binding to polysaccharides containing β-glycosidic bond, such as chitin, this flourochrome emits a bright blue light under UV excitation even using very short light exposure time (up to 1/1000 s) Selected fragments of demineralized skeletons of A. wolffgangi and E. gibbosa were placed in 0.1 M KOH-glycerine-water solution and few drops of 0.1% solutions of CFW were added and the mixture was placed in darkness for 60 min. Afterwards, the stained skeletons were rinsed 5 times with deionized water and dried at room temperature followed by investigation of the scaffolds under fluorescence microscopy. ## Scanning electron microscopy The surface morphology and microstructure of isolated chitinous scaffolds as well as untreated samples of both sponges were investigated on the basis of SEM images using ESEM XL 30 Philips scanning electron microscope. Before analysis, samples were covered with a carbon layer for 1 min using Edwards S150B sputter coater. ## Raman spectroscopy Raman spectroscopy was performed using a Raman spectrometer (RamanRxn1™, Kaiser Optical Systems Inc., Ann Arbor, USA) coupled to a light microscope (DM2500 P, Leica Microsystems GmbH, Wetzlar, Germany). For more details, see. The samples displayed intense fluorescence, which made the acquisition of a high quality Raman spectrum impossible. Therefore, the samples were bleached in 10% solution of hydrogen peroxide for 3 h. After three washing steps in distilled water, the samples were dried at room temperature. The Raman spectra were then acquired using an accumulation time of 3 s and summing up 50 accumulations. A baseline correction was finally applied in Matlab to remove the residual fluorescence signal from the spectra and display the Raman scattering. ## Fourier-transformation infrared spectroscopy FTIR spectroscopy is a powerful tool for the structural analysis of polysaccharides. This method is sensitive to the position and anomeric configuration of glycosidic linkages in glucans. It is worth to note that chitin, depending on its origin and function of tissue, occurs mostly in three main isoforms as α-chitin (fungi, sponges, arthropods), β-chitin (diatoms, molluscs) and rarely as γ-chitin (cocoons of some insects). This vibrational spectroscopy is also sensitive to the geometry of molecules, system of intramolecular and intermolecular interactions. Transmission spectra of chitinous scaffolds were made using a Nicolet 210c FTIR Spectrometer using ATR accessory. The investigation was performed over a wavenumber range of 4000–400 cm<sup>-1</sup> (at a resolution of 0.5 cm<sup>-1</sup>). The standard α-chitin was purchased from INTIB GmbH, Freiberg, Germany. ## Chitinase digestion test In order to carry out chitinase digestion test, the Yatalase<sup>®</sup> enzyme from culture supernatants of Corynebacterium sp. OZ-21 (Cosmo Bio, Japan) was used. One unit of this enzyme released 1 μmol of N-acetyl-D-glucosamine from 0.5% chitin solution and 1 μmol of p-nitrophenol from p-nitrophenyl-N-acetyl- β-D-glucosaminide solution in 1 min at 37°C and pH 6.0. The completely demineralized fibers of A. wolffgangi and E. gibbosa were incubated in enzyme solution containing 10 mg Yatalase dissolved in 1 mL of citrate phosphate buffer at pH 5.0 for 2 h. The effectiveness of enzymatic digestion was monitored using optical microscopy (Keyence). ## Estimation of N-acetyl-D-glucosamine (NAG) contents (Electrospray ionization mass spectrometry ESI-MS) The Morgan–Elson assay was used in order to estimate the N-acetyl-D- glucosamine content released after chitinase treatment, as previously reported. Preparation of the samples for ESI-MS: the demineralized organic scaffolds of A. wolffgangi and E. gibbosa were hydrolysed in 6 M HCl for 24 h at 50 <sup>o</sup>C. After hydrolysis samples were filtrated with 0.4 μm filter and freeze-dried to remove the excess of HCl. The dried samples were dissolved in deionized water for analysis. All ESI-MS measurements were performed on Waters TQ Detector ACQUITYuplc mass spectrometer (Waters, USA) equipped with ACQUITYuplc pump (Waters, USA) and BEHC18 1.7 μm, 2.1 × 50 mm UPLC column. Nitrogen was used as nebulizing and desolvation gas. Graphs were generated using Origin 8.5 for PC. # Results clearly indicates that the applied chemical treatment procedures (detailed presented) lead to purification of the fibrous scaffolds with well-organized anastomosing morphology from the skeletons of A. wolffgangi and E. gibbosa, respectively. The images presented in show that the overall shape and morphology of the extracted 3D scaffolds closely resemble the styles and forms of the investigated sponges. This means that, the isolation procedure does not lead to a breakdown of the–sometimes very fragile–demosponge structures, even after HF-based removal of the skeleton supporting spicules. SEM microphotographs of the skeletal fibers of A. wolffgangi and E. gibbosa prior to any treatment confirmed the complex character of their skeletons where, various forms of inorganic (spicules) as well as organic (fibres) structures are well visible.) show that glassy spicules were still present within skeletal scaffolds isolated from both sponges after NaOH and acetic acid treatment. Only HF-based treatment leads to dissolution of the spicules and purification of silica-free microfibers. The Calcofluor white staining (CFW) was the first step for the preliminary identification of chitin within isolated and demineralized skeletal samples. Fluorescence microscopy analysis of the scaffolds isolated from A. wolffgangi and E. gibbosa after CFW staining displayed very strong fluorescence even under light exposure time as short as 1/4800 s. Corresponding results were previously reported for chitin isolated from marine and freshwater sponges as well as in chitin-containing fossilized remnants. To confirm the presence of chitin in the isolated scaffolds, more sensitive analytical techniques were applied. FTIR spectra acquired for the fibrous scaffolds obtained from A. wolffgangi and E. gibbosa, as well for α-chitin standard are presented in. The region of the amidic moiety, between 1700 and 1500 cm<sup>−1</sup>, yields different signatures for chitin polymorphs. In this region, the spectra of the samples studied by us showed strong adsorption band associated with the stretching vibrations of C = O group characteristic of the amide band I. The amide band I showed twin peak at 1659 cm<sup>-1</sup> and 1626 cm<sup>-1</sup> for A. wolffgangi; 1659 cm<sup>-1</sup> and 1626 cm<sup>-1</sup> for E. gibbosa, as a result of the intermolecular C = O⋯H-N and the intramolecular hydrogen bonds C = O⋯HO- CH<sub>2</sub> which is characteristic for α-chitin polymorph. Additional feature, the characteristic intense band at 950 cm<sup>−1</sup> assigned to γCHx was observed in α-chitin standard as well as in the purified sponges chitin samples. Moreover, the α-chitin indicative band assigned to a ß-glycosidic bond is observed at a ν<sub>max</sub> 897 cm<sup>−1</sup> in the FTIR spectra of the scaffolds isolated from A. wolffgangii and E. gibbosa. Detailed analysis of the bands indicates that acquired spectra of both isolated chitinous scaffolds are very similar to those of the α-chitin standard. The results of Raman spectroscopy examinations showed that spectra of A. wolffgangi and E. gibbosa are very comparable with the spectrum obtained for α-chitin reference. For analytical investigations of the isolated scaffolds, prior and after HF-treatment, Raman spectroscopy was used. Consequently, for example, the Raman spectra of chitinous scaffold isolated from A. wolffgangi and E. gibbosa prior to demineralization display intense bands of biosilica at 443, 480, 599, 640, 805 cm<sup>-1</sup>. The bands of the organic matrices are visible in the ranges ≈ 900–1800 cm<sup>-1</sup> and 2700–3000 cm<sup>-1</sup>. These bands are comparable with those reported for α-chitin standard. Similar observations have been reported for chitin of demosponge origin previously. Chitinases possess the ability to degrade chitin directly to low molecular weight chitin oligomers including N-acetylglucosamine (GlcNAc). Consequently, such enzymatic treatment resulted in the loss of chitin integrity and in release of residual chitin microfibers of steadily decreasing size. The activity of chitinase is clearly visible using an optical microscope. Chitinase digestion test which have been previously utilized in the studies for the chitin detection in other sponges, definitely confirmed the chitinous nature of demineralized scaffolds isolated from both A. wolffgangi and E. gibbosa. The Morgan–Elson assay has been previously described in details and was used as the most accurate methods to estimate the GlcNAc released after chitinase treatment. Determination of GlcNAc in chitin-based scaffolds of A. wolffgangi and E. gibbosa showed, 750 ± 1.5 μg and 730 ± 1.5 μg N-acetyl-glucosamine per mg of chitinous scaffolds of these sponges, respectively. These results are similar to those reported for chitin isolated from the demosponge Spongilla lacustris. ESI-MS of D-glucosamine (GlcN) standard showed four main peaks at m/z = 162.08, 180.09, 202.07 and 381.15. The ion peak with m/z = 180.09 corresponds to the molecular ions \[M+H\]<sup>+</sup> of a species with a molecular weight 179.09 corresponding to GlcN (calculated: 179.1). The ion peak at m/z = 162.08 corresponds to a fragment ion \[M−H<sub>2</sub>O + H\]<sup>+</sup> after losing one molecule of H<sub>2</sub>O from DGlcN (calculated: 162.1). Finally, the ion peak at m/z = 381.15 corresponds to \[2M+Na\]<sup>+</sup> species which is sodium-bound GlcN non covalent dimmer. Similar ion peak for the proton-bound GlcN covalent dimmer was observed at m/z 359.17 \[M+H\]<sup>+</sup> in the spectra. The major peaks in ESI-MS spectra of the hydrolyzed samples of A. wolffgangi and E. gibbosa observed at m/z = 162.08, 180.09, 202.07 and 381.15 are comparable with the peaks of GlcN (Fig A). The ion peak of the sodium-bound GlcN dominated the spectra of these sponges as expected for marine-derived samples due to strong salt presence. # Discussion Previously, members of the genus Acarnus have been mostly investigated as a source of pharmacologically active compounds. For example, a group of compounds named acarnidines were isolated by extraction of the homogenized tissues of the sponge A. erithacus with toluene-methanol (1:3) and partitioning with 1 M sodium nitrate solution. They possess unique substituted homospermidine skeleton with diverse fatty acid substituents. The acarnidines showed antibacterial and antifungal properties and displayed significant antiviral activity against Herpes simplex type 1. Two cyclic peroxide-containing polyketide C22 methyl esters, peroxyacarnoic acid methyl esters A and B, have been isolated from the Red Sea marine sponge Acarnus cf. bergquistae. The methanolic extract of this sponge exhibited cytotoxicity against P-388, A-549, and HT-29 tumor cells with an IC<sub>50</sub> of 0.1 μg/ml. Two new cyclic peroxides have been reported in the organic extract of the sponge A. bicladotylota from India. Furthermore, sponges of the genus Echinoclathria are recognized as producers of biologically active compounds. Azaspiracid-2 was isolated from a marine sponge Echinoclathria sp. collected off Amami-Oshima area in Japan. It exhibited potent cytotoxicity against P388 cells with an IC<sub>50</sub> value of 0.72 ng/mL and caused S phase arrest on the cell cycle. The demosponge E. subhispida gave a new steroid sulfate, echinoclasterol sulfate with experimentally confirmed antifungal activity against Mortierella ramannianus, and cytotoxicity against PC-9 human lung cancer cells. Echinoclathrines A-C, a new class of pyridine alkaloids possessing a 4-aryl-2-methylpyridine moiety as a common structural element were isolated from an Okinawan sponge Echinoclathria sp. The procedure of the isolation of echinoclathrines have been patented recently. Some of echinoclathrines exhibited weak immunosuppressive activity in a mixed lymphocyte reaction assay. Studies on marine pharmacology potential of Echinoclathria demosponges habituated in Red Sea started only recently. Investigation of the Red Sea sponge E. gibbosa resulted in the isolation of three new compounds including β-sitosterol-3-O-(3Z)-pentacosenoate, 5α-pregna-3β-acetoxy-12β,16β-diol-20-one, and echinoclathriamide together thymine and uracil. β-Sitosterol-3-O-(3Z)-pentacosenoate showed weak activity against A549 non-small cell lung cancer (NSCLC), U373 glioblastoma (GBM), and PC-3 prostate cancer cell lines. New ceramide (icosanamide) was isolated from the Red Sea sponge Echinoclathria sp.. The in vitro growth inhibitory activity of this ceramide against different human cancer cell lines was evaluated. To our best knowledge there are no reports even about attempts to search for chitin in these species of demosponges. Till now, isolation protocols of diverse secondary metabolites from representative members of the genera Acarnus and Echinoclathria have followed traditional organic solvent-based extraction approaches. There are no data on isolation methods for such metabolites which are based on treatment with alkaline solutions as well as about structural stability of these biomacromolecules at alkaline pH levels. It is well known that, since the experimental work done by von Kölliker in 1864 the main skeletal protein of demosponges-spongin is quickly soluble in alkali solutions. This feature is crucial for extraction and isolation of poriferan chitin in purified form due to its exceptional resistance to the treatment with alkali up to concentration of 5% and temperatures not higher than 40°C for example, in the case of NaOH. Such treatment showed also no electron microscopically visible changes on the surface of siliceous spicules of the demosponges under investigation. Moreover, our observations showed with convincing support the localization of spicules within chitinous (Figs) and non-spongin based matrix. Similar results have been reported before in the case of chitinous skeleton of the fresh water demosponge S. lacustris (order Spongillida). The possible role of poriferan chitin as structural support for spicule-producing cells as well as in complete process of spiculogenesis in demosponges is still unknown. Complete desilicification of the spicules can be achieved using HF-based treatment. This study together with previously reported data showed that chitin remains to be well preserved even after such treatment. However, logical question about the possible presence of diverse secondary metabolites with alkaline, or HF-based extracts remain to be open. In contrast, the experiments with bromotyrosine- and chitin-producing demosponges representing the order Verongiida showed that bromotyrosines and chitin-based scaffolds could be isolated from the sponge skeletons using a stepwise extraction procedure mainly based on the use of NaOH. Recently, a patented method for isolation of both bromotyrosines and chitinous skeletal frameworks from selected sponges, without disruption of the skeletons in the mortar has been reported. Here, we propose a schematic view of the principal steps which can be now applied for isolation of secondary metabolites and chitin from the sponges of the order Poecilosclerida. There are no doubts about the necessity for the development of novel, more effective technologies for extraction of biologically active compounds together with chitinous scaffolds from sponges of the genera Acarnus and Echinoclathria. Especially those species which could be adapted for cultivation under marine farming conditions will possess high potential in this case. We suggest that the discovery of chitin within other representatives of Poecilosclerida order would be the next step in the evaluation of the possibility to accept these worldwide distributed demosponges as novel renewable source for both chitin and biologically active metabolites which are perspective for biomedicine and marine pharmacology, respectively. # Conclusions Chitin-producing marine demosponges are highly perspective invertebrates due to their ability to synthetize broad variety of secondary metabolites with antiviral, antibiotic, antidiabetic, cytotoxic and antitumor activities as well as chitin. Here, we showed for the first time that chitin is present as a structural component in skeletons of the Red Sea sponges Accarnus wolffgangi and Echinoclathria gibbosa demosponges. The question of chitin synthesis among representatives of the genera Acarnus and Echinoclathria should gain importance as a result of our findings. Consequently, the evolution, localization and functions of chitin in these demosponges as well as in other representatives of Poeciloscrerida order should be examined in the future. Additionally, separate studies should be carried out on the identification of chitin synthase genes within genomes of diverse representatives of the genera Acarnus and Echinoclathria as well. Also, additional investigations are necessary to obtain a better understanding of the nature and origin of spicules- containing skeletons of these demosponges with respect to the spongin-chitin relationship. It is still unclear how much spongin is present in the chitin- based skeletons of the sponges studied. Novel approaches must be proposed which will bring together molecular biology and modern bioanalytical methods for a better understanding of the poriferan chitins synthesis in diverse taxa on molecular level. The best way to solve this challenging task is to bring together coherent and synergetic collaborators and experts in marine biology, marine chemistry, marine pharmacology, marine biotechnology and biomaterials together with spongologists using their multidisciplinary knowledge and experiences to answer raised questions and develop new approaches in this interesting area of research. # Supporting information We thank the EEAA and the Red Sea Protectorate of Egypt for permission to make collections of the Red Sea sponges Acarnus wolffgangi and Echinoclathria gibbosa. Our thanks to Prof. Rob van Soest for the taxonomic identification of the sponge samples. [^1]: The authors have declared that no competing interests exist.
# Introduction The canonical Wnt signaling pathway is crucial for embryonic developmental processes and adult tissue homeostasis. Consequently, aberrations in this pathway were linked to human diseases and in particular cancer development. The key mediator of the canonical Wnt signaling pathway is β-catenin, whose protein levels are under tight control by a multiprotein complex known as the destruction complex. β-catenin is phosphorylated by this complex, which ultimately leads to its ubiquitin-proteasome-dependent degradation. In the presence of Wnt ligands the destruction complex becomes inactivated and β-catenin accumulates in the cytoplasm, translocates into the nucleus and initiates transcription of mitogenic target genes leading to cell proliferation. The core components of the destruction complex consist of Adenomatous Polyposis Coli (APC), axis inhibition protein 1 and 2 (AXIN1 and AXIN2) and the kinases glycogen synthase kinase 3 (GSK3) and casein kinase 1α (CK1α). In the majority of colorectal cancers, APC is found to be mutated and the destruction complex thereby inactivated. Interestingly, overexpression of AXIN1 or AXIN2 can compensate for APC mutations and leads to the degradation of β-catenin in APC- mutant cell lines, such as SW480 colorectal cancer cells. AXIN has been shown to be the rate-limiting factor for destruction complex function in Xenopus egg extracts and its protein levels are tightly regulated by APC and by the poly- ADP-ribosyltransferases tankyrase 1 and 2 (TNKS1/2). The tankyrase enzymes transfer ADP-ribose moieties onto AXIN1/2, marking it for degradation by the ubiquitin-proteasome system. Inhibition of TNKS1/2 by small molecule inhibitors (TNKSi) has emerged as a promising new cancer therapeutic approach as it leads to stabilization of AXIN1/2 and a concomitant reduction in β-catenin protein levels and transcriptional activity in vitro and in vivo \[, –\]. Of note, AXIN2 is also a target gene for β-catenin, adding another layer of AXIN2 regulation to the Wnt signaling pathway. In the current study, we sought to elucidate the consequences of combining TNKSi with proteasome inhibition, as proteasome inhibitors are extensively used in both clinical and research settings, often in combination with other inhibitors. # Materials and Methods ## Antibodies, plasmids, and chemicals The following reagents were used: rabbit anti-AXIN1 (C95H11), rabbit anti-AXIN2 (76G6) (Cell Signaling Technology), mouse anti-β-catenin (BD Transduction Laboratories); mouse anti-ubiquitin (Upstate / Millipore), mouse anti-active-β- catenin (05–665, Millipore); mouse anti-β-Actin (Sigma Aldrich), mouse anti- Calreticulin (Enzo lifesciences), mouse anti-Vinculin (HVIN-1, Sigma Aldrich), rabbit anti-FoxM1 (C-20, Santa Cruz), mouse anti-LaminA (Abcam), rabbit anti-p62 (MBL / Nordic Biosite). All secondary antibodies used for confocal microscopy studies were obtained from Jacksons ImmunoResearch Laboratories and secondary antibodies used for Western blotting were obtained from LI-COR Biosciences GmbH. Hoechst (Invitrogen). G007-LK (Gift from Stefan Krauss and Jo Waaler, Oslo, Norway); MG132 (Calbiochem); Dimethyl sulphoxide (DMSO), 3-Methyladenine (3-MA), Lactacystin, PhosSTOP (Sigma Aldrich); Epoximicin (Enzo lifesciences); Leupeptin (Peptanova Gmbh, Peptide Insitute, Japan). Quantitech mRNA primer pairs against TBP (QT00000721), AXIN2 (QT00037639) and FoxM1 (QT00000140) were obtained from Qiagen. FoxM1 siRNA (Sense: `5' GGACCACUUUCCCUACUUUUU-3'`, Antisense: `5' AAAGUAGGGAAAGUGGUCCUU 3'`, and control siRNA (cat: D-001810-01), Dharmacon. siRNA transfections were performed using RNAiMax (Invitrogen) according to the manufacturer's protocol. ## Cell-based assays SW480, COLO320, CaCo-2 and LS174T cell lines were purchased from ATCC. Upon receipt, cells were frozen, and individual aliquots were taken into cell culture, typically for analysis within 15 passages. Cells were grown in RPMI (SW480 and COLO320), DMEM (CaCo-2) or DMEM/F12 (LS174T) medium supplemented with 10% (SW480 and COLO320) or 15% (LS174T and CaCo-2) FBS and 1% penicillin/streptomycin. The stable SW480 cell line expressing GFP-TNKS1 was described earlier. Testing for mycoplasma contamination was performed every sixth week. For inhibition of TNKS activity, cells were treated with 0.5 μM G007-LK for 6 h. DMSO was used as a control. For inhibition of proteasomal activity, cells were treated with 10 μM MG132, 25 nM Epoxomicin or 10 μM Lactacystin. Other inhibitors used were: 10 mM 3-Methyladenine (3-MA, autophagy inhibitor), or 300 μM Leupeptin (protease inhibitor) for indicated time points, either alone or in combination with G007-LK. ## Western blot analysis Cells were rinsed in PBS and lysed in Laemmli lysis buffer (65.8 mmol/L Tris- HCl, pH 6.8, 2.1% SDS, 26.3% (w/v) glycerol, 0.01% bromophenol blue, dithiothreitol (DTT)). Equal amounts of whole cell lysate were separated by SDS- PAGE (Bio-Rad Laboratories) and blotted onto polyvinylidene difluoride membranes (Millipore). Immunodetection was performed with IRDye-conjugated secondary antibodies (LI-COR Biosciences). The Odyssey Imager system (LI-COR Biosciences) was used to scan all blots. Protein bands were quantified using the Odyssey software. ## Nucleo-cytoplasmic fractionation Cells incubated with or without MG132 were washed with PSB and lysed in lysis buffer (0.1 M NaCl, 10 mM Na<sub>2</sub>HPO<sub>4</sub>, 1% Triton X-100, 1 mM EDTA (pH 7.4), 10 mM protease inhibitor cocktail, 2 mM NEM) and the lysates were left on ice for 20 min. The cell lysates were centrifuged at 14000 rpm for 10 min at 4°C to separate cytoplasmic (supernatant) and nuclear fraction (pellet). The fractions were then subjected for Western blotting. ## Confocal fluorescence microscopy Cells grown on coverslips were permeabilized in PEM buffer (pH 6.8, 80 mM PIPES, 5 mM EGTA, 1 mM MgCl<sub>2</sub>x6H<sub>2</sub>O containing 0.05% saponin) for 5 min on ice before fixation in 3% paraformaldehyde for 15 min on ice and washed twice in PBS containing 0.05% saponin or permeabilized for 5 min with 0.5% Triton-X-100 in PBS, as indicated in the figure legends. The cells were then stained using the indicated primary antibodies for 1 h, washed three times in PBS/saponin, stained with secondary antibodies for 1 h, and washed three times in PBS. The coverslips were mounted in Mowiol containing 2 μg/ml Hoechst 33342 (Sigma-Aldrich). The cells were examined with a Zeiss LSM710 or LSM780 confocal microscope (Carl Zeiss MicroImaging GmbH) equipped with an Ar-Laser Multiline (458/488/514 nm), a DPSS-561 10 (561 nm) and a Laser diode 405-30CW (405 nm). The objective used was a Zeiss plan-Apochromat363/1.4 Oil DIC III. Image processing and visualization were performed with ZEN Software (Carl Zeiss MicroImaging GmbH), Photoshop CS4 (Adobe) and ImageJ (National Institutes of Health). All images were taken at fixed intensity settings below saturation. ## ScanR high-throughput microscopy Cells were grown on coverslips and further processed for antibody staining as described for confocal microscopy samples. Images were automatically taken using the Olympus ScanR system with an UPLSAPO 40x/0.95 objective. All images were taken with the same settings and below pixel saturation. The Olympus ScanR analysis software was used to detect and count the number of GFP-TNKS1 puncta in samples treated with DMSO, G007-LK, MG132 or a combination of these, and to measure the nuclear intensity of β-catenin, FoxM1 and active β-catenin in DMSO versus MG132 treated samples. ## Quantitative real-time PCR of mRNA expression mRNA expression analysis was done as described in. Primers used in the study were AXIN2 (QT00037639), FoxM1 (QT00000140) and TBP (QT00000721). ## SDS immunoprecipitation To investigate modifications of FoxM1 we did hot lysis immunoprecipitation as described in. Shortly, cells were lysed in SDS (1%)–containing PBS, and immediately incubated at 100°C for 5 min, chilled on ice and homogenized using QIA-shredder column (QIAGEN). The lysates were added to protein G-coupled magnetic beads (Dynabeads, Life technologies / Thermo Fisher) loaded with rabbit anti-FoxM1 (C-20) antibody dissolved in 2x IP buffer (2% (vol/vol) Triton X-100, 0.5% (wt/vol) sodium deoxycholate, 1% (wt/vol) bovine serum albumin (BSA), 2 mM EDTA, 40 mM NaF, 2 mM NEM, 10 mM protease inhibitor cocktail (Sigma)). The beads and lysates were gently mixed for 1h at 4°C before the beads were washed in 1x IP buffer and eluted in 2x sample buffer, then subjected for Western blotting. ## Electron microscopy SW480 were seeded on coverslips and treated with DMSO or MG132 for 6 hours before fixation in 2% glutaraldehyde in 0.1 M PHEM (240 mM PIPES, 100 mM HEPES, 8 mM MgCl<sub>2</sub>, 40 mM EGTA), pH 6.9, at room temperature for 40 min. Cells were post-fixed in osmium tetroxide, stained with tannic acid, dehydrated stepwise to 100% ethanol and flat-embedded in Epon. Serial sections (\~100 nm) were cut on an Ultracut UCT ultramicrotome (Leica, Germany) and collected on formvar coated mesh-grids. Sections were observed at 80 kV in a JEOL-JEM 1230 electron microscope and images were recorded using iTEM software with a Morada camera (Olympus, Germany). # Results and Discussion ## G007-LK-induced degradasome formation is counteracted by proteasome inhibition in SW480 cells Inhibition of the TNKS enzymes by small-molecule inhibitors has previously been shown to reduce the aberrantly high levels of β-catenin in colorectal cancer cells such as SW480 cells by re-establishing a functional destruction complex. Incubation of SW480 cells with the highly selective TNKSi G007-LK for 6 h results in the formation of cytoplasmic puncta (degradasomes), which contain the destruction complex components AXIN1, AXIN2, APC, GSK3, βTrCP, TNKS1/2, β-catenin and phospho-β-catenin and therefore most likely represent enlarged versions of the destruction complex, where β-catenin is phosphorylated and thereby earmarked for degradation in the proteasome. The formation of cytoplasmic puncta is most likely due to head-to-tail polymerization of AXIN molecules via their DIX domain and may also involve TNKS polymers. Surprisingly, the formation of degradasomes was reduced upon combination of G007-LK with the proteasome inhibitor MG132 for 6 h, as shown by high-throughput microscopy using an Olympus ScanR automated microscope. The number of GFP-TNKS1 puncta was quantified using the ScanR analysis software and revealed a rapid induction of degadasomes after 2 h of incubation with G007-LK, while the combination of MG132 with TNKSi severely impaired degradasome formation. To test whether this unexpected result could be reproduced with chemically unrelated proteasome inhibitors, we combined G007-LK with either 25 nM Epoxomicin or 10 μM Lactacystin, respectively. Like with MG132, we observed a decrease in degradasome formation with these alternative proteasome inhibitors. On the other hand, lysosome inhibition (300 μM Leupeptin) or phosphatidylinositol-3 kinase class III inhibition (10 mM 3-methyladenine (3-MA)) did not interfere with degradasome formation, indicating that specifically inhibiting the proteasome interferes with TNKSi-induced degradasome formation. ## TNKSi-induced AXIN2 stabilization is impaired upon proteasome inhibition TNKSi-induced stabilization of AXIN1 and/or AXIN2 is thought to mediate the re- establishment of functional destruction complexes. Therefore, we investigated AXIN1 and AXIN2 protein levels during 6 h of G007-LK incubation with or without MG132. During this incubation time the protein levels of AXIN1 remained largely unaltered, while the AXIN2 levels strongly increased upon G007-LK treatment. This increase in AXIN2 was abrogated when G007-LK was combined with MG132, indicating that the TNKSi-induced stabilization of AXIN2 is impaired upon proteasome inhibition. The chemically unrelated proteasome inhibitor Epoxomicin prevented TNKSi-induced stabilization of AXIN2, but not AXIN1, similar to MG132, while 3-MA or Leupeptin did neither influence AXIN1 nor AXIN2 protein levels when co-incubated with G007-LK. Next, we tested whether this observation was restricted to SW480 cells or whether other colorectal cancer cell lines showed the same response on AXIN2 protein level. Indeed, CaCo-2, LS174T and Colo320 responded similarly to SW480 cells to the combination of G007-LK and MG132 with abolished AXIN2 stabilization, as shown by Western blotting and verified by quantifications. We conclude that the lack of TNKSi-induced degradasome formation upon proteasome inhibition most likely depends on impaired stabilization of AXIN2. ## Proteasome activity inhibits transcription of AXIN2 To distinguish whether the lack of AXIN2 protein stabilization upon combination of proteasome inhibition and G007-LK originates from altered mRNA levels or is due to a regulation on the protein level, we measured relative mRNA levels of AXIN2 upon TNKS inhibition with or without MG132 by quantitative real-time PCR. MG132 led to a severe reduction in AXIN2 mRNA levels in SW480 cells and this was not changed in the presence of both MG132 and G007-LK. To verify whether MG132 had the same effect in other colorectal cancer cell lines with different APC and β-catenin mutation status, we performed similar experiments in LS174T, CaCo-2 and Colo320 cell lines. Indeed, we observed a reduction in AXIN2 mRNA levels in all three cell lines upon MG132 incubation for 6 h. ## Proteasome inhibition does not reduce nuclear β-catenin levels As nuclear β-catenin promotes transcription of AXIN2, we investigated β-catenin localization in SW480 cells treated with either DMSO or MG132 by confocal microscopy and demonstrated a distinct staining of β-catenin in the nucleus under both conditions. To investigate whether the amount of nuclear β-catenin in MG132-treated cells was reduced compared to control cells, we performed high-throughput microscopy that surprisingly revealed a slight increase in the mean fluorescence intensity of β-catenin in the nuclei of SW480 cells treated with MG132. To further interrogate whether the transcriptionally active fraction of β-catenin was altered upon MG132 treatment, we took advantage of an antibody specifically detecting non-phospho-β-catenin (i. e. active β-catenin, ABC). Quantitative fluorescence microscopy showed the same tendency for active as for total β-catenin. Next, we undertook biochemical fractionation experiments to verify our imaging results. As expected, MG132 led to an accumulation of both β-catenin and active β-catenin in total protein lysates. Moreover, nuclear accumulation of both total and active β-catenin verified the results obtained by immunofluorescence. Thus, changes in nuclear β-catenin levels cannot explain the decreased transcription of AXIN2 mRNA upon proteasome inhibition. ## The proteasome-regulated transcription factor FoxM1 regulates AXIN2 transcription Another regulator of AXIN2 transcription is Forkhead box M1 (FoxM1). FoxM1 was previously shown to be a positive regulator of AXIN2 mRNA levels by two different means: First, it can directly bind to and increase transcriptional activity of the AXIN2 promotor region in developing lung epithelium. Second, FoxM1 was reported to promote the nuclear localization of β-catenin and support β-catenin in binding to its target promotors, thereby indirectly controlling Wnt target-gene expression in glioma cells. Interestingly, FoxM1 was also shown to be negatively regulated by proteasome inhibition. Proteasome inhibitors such as MG132, MG115 and bortezomib were shown to inhibit FoxM1 transcriptional activity and FoxM1 expression, and this seems to be mediated by stabilization of a negative regulator of FoxM1, namely HSP70. However, Chen and coworkers report an increase in FoxM1 protein levels upon proteasome inhibition. Due to these conflicting reports, which most likely result from use of different cell lines and/or incubation protocols, we investigated FoxM1 mRNA and protein levels in SW480 cells with our experimental setup. We observed a substantial reduction of FoxM1 mRNA levels upon 6 h MG132 treatment. However, we detected a \~2.5 fold increase of FoxM1 at the protein level. We hypothesize that despite a decrease in FoxM1 mRNA transcription, the protein turnover of FoxM1 is severely impaired by the proteasome inhibition, which leads to an accumulation of FoxM1 protein. To investigate whether FoxM1 functions as a positive regulator of Wnt-target gene expression in our model system, we treated SW480 cells with siRNA against FoxM1, which efficiently reduced FoxM1 protein levels. We observed a significant reduction in AXIN2 mRNA levels upon knockdown of FoxM1, consistent with the published role of FoxM1 as both direct and indirect transcriptional activator of Wnt target gene expression. In order to test the contribution of FoxM1 to the TNKSi-induced AXIN2 stabilization, we depleted SW480 cells for FoxM1 and incubated with G007-LK. We observed a substantial reduction in the TNKSi-induced AXIN2 protein levels in cells depleted for FoxM1 as shown by Western blotting , and the same effect was true when G007-LK was combined with MG132 in FoxM1 depleted cells. This confirms the role of FoxM1 in the regulation of AXIN2 protein levels. ## Posttranslational modifications of FoxM1 are altered upon proteasome inhibition It was reported recently that USP5-mediated deubiquitination is required in order to promote translocation of FoxM1 into the nucleus, where it mediates interaction of β-catenin with target gene promotors. We therefore hypothesized that proteasome inhibition may alter the nuclear translocation of FoxM1. In order to investigate the subcellular localization of FoxM1, we performed immunofluorescence stainings for FoxM1 followed by high-throughput microscopy to quantify the nuclear fluorescence intensity of FoxM1. FoxM1 could be detected in the nuclei of SW480 cells after MG132 treatment. Knockdown experiments verified the specificity of the FoxM1 antibody in immunofluorescence stainings. To substantiate our findings we performed biochemical fractionation experiments, which showed the same tendency of a predominant localization of FoxM1 in the nucleus upon proteasome inhibition. These data indicate that FoxM1 localization cannot explain the reduced transcription of AXIN2 upon proteasome inhibition, irrespective of a possible ubiquitination of FoxM1. Besides ubiquitination, FoxM1 is regulated by various phosphorylation events and all phosphorylation events were shown to stabilize and activate FoxM1. To investigate the phosphorylation status of FoxM1 in the absence (DMSO) or presence of proteasome inhibition (MG132), we performed a hot-lysis immunoprecipitation of FoxM1. Lysing cells in PBS containing 1% SDS with an immediate incubation at 100°C for 5 min leads to the instantaneous denaturation of all proteins (including phosphatases) and enables the investigation of posttranslational modifications of the immunoprecipitated proteins. We observed a higher molecular weight band in the DMSO-treated conditions, which was absent in the immunoprecipitate of the MG132 treated cells. To investigate further whether this band could represent phosphorylated FoxM1, we incubated SW480 protein lysates with or without phosphatase inhibitor mix (PhosSTOP) for 1 h at room temperature. In the absence, but not in the presence, of the phosphatase inhibitor mix a high molecular weight band disappeared and the lower band increased in intensity, indicating that the higher molecular weight band is indeed phosphorylated FoxM1. We therefore suggest that despite FoxM1 localizing to the nucleus after MG132 treatment, it is dephosphorylated and thus transcriptionally less active upon proteasome inhibition. As most kinases phosphorylating FoxM1 are cell cycle kinases (CDK4/6, PLK1, CyclinA/CDK, Chk2), we speculate that proteasome inhibition might lead to cell cycle arrest, synchronizing the cell population in a cell cycle phase where these kinases are not active. Knockdown of FoxM1 has less effect on AXIN2 transcription compared to MG132 treatment (Figs). This could either mean that FoxM1 depletion was incomplete or that another still unknown factor influences AXIN2 mRNA transcription. Indeed, 72 h after siRNA transfection, a minor fraction of FoxM1 protein seemed to be left. In addition we noticed a redistribution of p62 and ubiquitin after MG132 treatment in immunofluorescence stainings. This prompted us to investigate the general cell morphology after 6 h of treatment with MG132 at an ultrastructural level by electron microscopy. Surprisingly, we found that various cellular organelles were clustered around the nucleus resulting in an organelle-depleted cytoplasm in a subset of MG132-treated cells. Such a drastic morphological change in a subpopulation of MG132-treated SW480 cells could possibly lead to an additional inhibitory effect on transcription and/or translation of proteins, e.g. AXIN2. However, transcription of house keeping genes was unaffected as judged by quantitative real-time experiments. Furthermore, protein levels were unaltered for AXIN1, arguing against a general full inhibition of gene transcription and protein translation upon proteasome inhibition. Taken together, a reduction of FoxM1 activity in combination with changed cellular morphology may explain the lack of TNKSi-induced AXIN2 stabilization upon proteasome inhibition and thus the lack of degradasome formation when TNKSi are combined with proteasome inhibitors. While our manuscript was in preparation, another publication reported that proteasome inhibition prevents degradasome formation and leads to decreased association of the PARsylated AXIN and TNKS proteins, but also to a perinuclear enrichment of AXIN. This is in agreement with our data, and our findings provide a further mechanistic explanation by showing that proteasome inhibition decreases the fraction of phosphorylated and thus active FoxM1 and that the cell morphology starts to change after 6 h of proteasome inhibition, which has implications for cell biological research utilizing proteasome inhibitors. Given that proteasome inhibitors (including MG132) are used as potential therapeutic agents in colorectal cancer, these findings are relevant regarding combination therapies with TNKSi. The fact that proteasome inhibition counteracts TNKSi-induced degradasomes argues against combining proteasome inhibitors and TNKSi in cancer therapy. # Supporting Information The authors thank Stefan Krauss and Jo Waaler for providing G007-LK. The authors thank The Core Facilities for Advanced Light Microscopy and Electron Microscopy at Oslo University Hospital for providing access to relevant microscopes. We also acknowledge Anne Engen and her co-workers in the cell lab facility for expert handling of cell cultures. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NMP TET SWS EMW HS. Performed the experiments: NMP TET SWS EMW. Analyzed the data: NMP TET SWS EMW HS. Contributed reagents/materials/analysis tools: HS. Wrote the paper: EMW NMP TET HS.
# Introduction There is increasing evidence that human immunodeficiency virus (HIV) infection is a risk factor for occlusive arterial disease including myocardial infarction (MI) – and peripheral arterial disease (PAD). A number of mechanisms may explain this relationship. Partly through antiretroviral therapy (protease inhibitors), patients with HIV infections may have an increased prevalence of conventional cardiovascular risk factors, the consequence being an enhanced degree of atherosclerosis. However, vascular pathology in HIV may occur in the absence of risk factors through arterial wall infections, producing a vasculitis either through direct effects or through co-infections. Whether an HIV-associated vasculitis accounts in-part for occlusive arterial disease nevertheless remains uncertain. Although a number of vasculitis subtypes have been described in HIV infected patients, the prevailing hypothesis is that HIV-associated occlusive arterial disease (MI and PAD) is largely attributed to premature atherosclerosis. Thus, cardiovascular risk prevention in HIV-infected patients has focused on targeting conventional cardiovascular risk factors. Nevertheless, there is some evidence that occlusive arterial disease in HIV-infected patients may be attributed to a vasculitis. In this regard, in the absence of atheroma, adventitial inflammation has been described in large vessels in a series of 16 HIV positive patients, most of whom had aneurysmal pathology, but 3 of whom also had occlusive arterial disease ; and in large vessels of 4 amputated limbs of HIV positive patients with critical lower limb ischemia (CLI). However, without an HIV-sero-negative group with matched pathology available for comparison, adventitial vasculitis in patients with occlusive arterial disease - could be attributed to an epiphenomenon. Hence, whether a vasculitis characterizes HIV-positive as compared to negative patients with occlusive arterial disease is unknown. In the present study we therefore aimed to compare large artery histopathological characteristics in untreated HIV-sero-positive patients with CLI requiring amputation, to large artery characteristics in HIV-sero-negative patients with CLI requiring amputation. # Materials and Methods ## Study groups and clinical data The present study was conducted according to the principles outlined in the Helsinki declaration. The Committee for Research on Human Subjects of the University of the Witwatersrand approved the protocol (approval number: M120739). Participants gave informed, written consent. Ten HIV positive and 10 HIV negative black African male patients with CLI requiring above knee amputations were recruited from the Division of Vascular Surgery at the Chris Hani Baragwanath Hospital, Johannesburg, South Africa over the period September 2012 to July 2013. None of these patients had clinical evidence of aneurysms. All patients had ischemic pain at rest and evidence of tissue loss (necrosis as evidenced by foot and leg ulcers and/or gangrene). Routine HIV serology (ELISA) was performed to determine HIV status. A CD4 count was obtained in all HIV positive patients with CLI. Participants were considered to have diabetes mellitus if they had a fasting plasma glucose concentration ≥7 mmol/l, or in whom glucose-lowering agents were prescribed. Brachial blood pressure (BP) was measured according to guidelines and taken as the mean of five measurements. Participants with a BP ≥140/90 mm Hg or in those receiving antihypertensive medication were considered to have hypertension. Dyslipidemia was diagnosed as the presence of either a raised triglyceride concentration (≥1.7 mmol/l) or a reduced high density lipoprotein (HDL) concentration (\<1.0 mmol/l). ### Histopathological assessment Specimens from the distal superficial femoral artery (SFA) were obtained at the time of above knee amputations due to PAD where reconstruction could not be performed resulting in end-stage CLI. One centimeter of the SFA was obtained at the level of transection for the amputation (from the viable stump). Specimens were fixed in 10% buffered formalin, processed according to standard methods and 4 µm sections cut and stained with hematoxylin and eosin. Sections were also stained with Elastic Von Giesson stain for elastic fibres; a Masson trichrome stain for collagen and an Alcian Blue stain for mucopolysaccharides. Immunohistochemical stains for CD3 (T-cell marker), CD20 (B-cell marker) and CD68 (histiocytic marker) were performed. The slides were assessed by a pathologist (SN) who was unaware of (blinded to) the HIV-status and clinical characteristics of the patients from whom the specimens were derived. The slides were evaluated for the presence of atheroma, fragmentation and reduplication of the internal elastic lamina, and adventitial vascular proliferation and perivascular inflammation. All parameters were graded semi-quantitatively, as 0 (absent), 1+(focally present), 2+(moderate) and 3+(extensive). ### Data analysis Continuous data are shown as mean ± SD. Proportions between groups were compared using a Fischer’s Exact test and continuous data were compared using a Mann- Whitney ranked sum test. # Results ## Participant characteristics The demographic and clinical characteristics of the participants are shown in. As compared to HIV negative patients with CLI, HIV positive patients with CLI were younger, and had a lower prevalence of traditional cardiovascular risk factors, including hypertension and diabetes mellitus. HIV negative patients had higher glycated hemoglobin values. A similar proportion of HIV positive patients with CLI smoked regularly as compared to HIV negative patients with CLI and similar average lipid concentrations were noted between the groups. A wide range (74 to 643×10<sup>6</sup>/l) of values for CD4 count were obtained and 5 (50%) HIV positive participants had values that would not have qualified for antiretroviral therapy based on the current thresholds for therapy in South Africa (\<350×10<sup>6</sup>/l), and 2 patients had CD4 counts \>500×10<sup>6</sup>/l. ## Histopathological evidence of femoral wall inflammation 90% of HIV positive patients, but no HIV negative patient had evidence of adventitial leukocytoclastic vasculitis of the vasa vasorum. In addition, 70% of HIV positive, but no HIV negative patient had evidence of adventitial slit-like vessels. Whilst in HIV positive patients T-lymphocytes were noted in the adventitia in 80% and in the intima and media in 30% of patients, in HIV negative patients T-lymphocytes were noted only in the intima and these occurred in 50% of patients. No B lymphocytes were observed in either group (data not shown). No significant differences in the mean (±SD) CD4 count were noted between HIV positive patients ranked as having 1+(366±218×10<sup>6</sup>/l) versus 2+(292±244×10<sup>6</sup>/l, p = 0.56) adventitial leukocytoclastic vasculitis. ## Histopathological evidence of arteriosclerotic changes The presence of calcified multilayered fibro-atheroma was noted in 40% of HIV positive patients with CLI with none showing eccentric changes, whilst calcified multilayered fibro-atheroma was noted in 90% of HIV negative patients with CLI, 60% of whom had eccentric changes. Ossification of plaque was noted in 30% of HIV negative patients with CLI, but no ossification was noted in HIV positive patients with CLI. A similar proportion of HIV-positive and negative patients had evidence of macrophages and the extent of intracellular and extracellular lipid deposits was similar in both groups. Lipid deposits were noted in HIV positive patients with multilayered fibro-atheroma only. A similar proportion of HIV positive and HIV negative patients with CLI had fragmentation and reduplication of the internal elastic lamina and mucoid degeneration. The combined thickness of the femoral artery intima and media tended (p = 0.20) to be lower in HIV positive as compared to negative patients with CLI. However, intima-media thickness was similarly increased in both HIV positive (1.38±0.28, p\<0.05) and HIV negative (1.56±0.79, p\<0.05) patients with fibroatheroma as compared to HIV positive patients without fibroatheroma (0.74±0.41). ## Relationship between risk factors and advanced atheroma All HIV positive patients with advanced atheroma were current or previous smokers. However, HIV positive as compared to negative patients with advanced atheroma were younger and a trend (p = 0.05 and 0.11) for the presence of fewer HIV positive patients to have hypertension or diabetes mellitus was noted. # Discussion The main findings of the present study are that as compared to femoral arteries from HIV negative African men with CLI, femoral arteries from HIV positive African men with CLI not receiving anti-retroviral therapy are largely characterized by the presence of leukocytoclastic vasculitis of the vasa vasorum and adventitial inflammation. In addition, a significant proportion of HIV positive men with CLI not receiving anti-retroviral therapy showed advanced femoral artery atheroma, despite a markedly lower age and a lower prevalence of conventional cardiovascular risk factors. To the best of our knowledge the present study is the first to compare histopathological changes in appropriate large vessels in patients with and without HIV and with occlusive arterial disease. Indeed, although a number of vasculitis subtypes have been described in HIV infected patients, few studies have reported on large vessel changes in patients with occlusive large artery disease, such as MI, PAD or stroke. In this regard prior studies have reported on the presence of leukocytoclastic vasculitis of the vasa vasorum and adventitial inflammation in large vessels of patients with HIV and aneurysmal changes in which 3 patients had occlusive large artery disease, and in large vessels of 4 amputated limbs of HIV positive patients with critical limb ischemia (CLI). In these studies, comparisons of histopathological features were not made with patients admitted for similar clinical events but without HIV. Hence, the histopathological changes reported on in HIV positive patients with occlusive arterial disease in these studies – may reflect an epiphenomenon and not HIV-associated pathology. However, the present study provides clear evidence that leukocytoclastic vasculitis of the vasa vasorum and adventitial inflammation in large vessels is a unique feature associated with HIV infection as compared to non-HIV-infection in patients with CLI. Our finding of a characteristic adventitial inflammatory change in HIV positive as compared to negative patients with CLI in the present study should be interpreted with caution. These relationships may not represent cause and effect. A previous study reporting on leukocytoclastic vasculitis of the vasa vasorum and adventitial inflammation in patients with mainly large artery aneurysms and in a small number of patients with occlusive arterial disease (n = 3), also showed adventitial and medial fibrosis and loss of medial muscle. The authors speculated that occlusion of the vasa vasorum leads to death of areas of large vessel walls and hence to aneurysm formation. However, in the present study we failed to note similar changes in any of the 10 patients with HIV and CLI. Nevertheless, to avoid the effects of ischemia on histopathological changes, we sampled femoral artery tissue proximal to the occlusion, rather than at the level of the occlusion. It is therefore possible that at the level of the occlusion, adventitial and medial fibrosis and loss of medial muscle may have been noted, the consequence being wall scarring, thrombus formation and ultimately vascular occlusion. The present study provides intriguing evidence to suggest that at least in a portion of patients with HIV not receiving antiretroviral therapy who develop CLI (4 of 10 in the present study), advanced large artery atheroma (calcified fibroatheroma) may contribute toward vascular occlusion. In this regard, although HIV positive patients with CLI with advanced femoral artery were younger and tended to have less risk factors than HIV negative patients with CLI with advanced femoral artery atheroma, all HIV positive patients with advanced femoral artery atheroma were current or previous smokers. This is entirely consistent with a number of studies that have reported a high prevalence of smoking in HIV positive patients with cardiovascular disease. These data therefore suggest that conventional risk stratification may not apply equally or as effectively among those with as compared to those without HIV infections. Indeed, non-human primate models of immunodeficiency virus infections are associated with atherosclerotic lesions in the absence of conventional cardiovascular risk factors. Large studies are required to evaluate whether current risk assessment charts adequately risk predict in HIV positive patients not receiving antiretroviral therapy in South Africa. The findings that fewer HIV positive as compared to negative patients with CLI had advanced calcific fibroatheroma, and that femoral artery intima-media thickness tended to be lower in HIV positive as compared to negative patients, must also be interpreted with caution. As previously emphasized, to avoid the effects of ischemia on histopathological changes, we sampled femoral artery tissue proximal to the occlusion, rather than at the level of the occlusion. Hence, we may have missed areas of fibroatheroma in HIV positive patients and consequently, it is still possible that atheroma is the major cause of CLI in HIV positive patients not receiving antiretroviral therapy. The clinical implications of the present study warrant consideration. In this regard, in two large vascular units in South Africa in which we have previously reported on very high admission rates for CLI, approximately 12% of these patients were HIV positive with low cardiovascular risk scores irrespective of whether or not they were receiving antiretroviral therapy. In this regard, the present study suggests that the pathogenesis of CLI in these circumstances may involve a combination of a vasculitis and/or advanced atheroma formation, despite a low cardiovascular risk. These findings may have implications for both risk prevention, where a more aggressive approach to risk management is required in HIV positive patients who smoke, and for decisions regarding revascularization procedures, where the causal lesions may in-part be atheromatous in nature despite a lower overall cardiovascular risk. Moreover, the present study suggests that a significant proportion of HIV positive patients (50%) may develop CLI when they have a CD4 count above the threshold that qualifies for antiretroviral therapy in South Africa. If confirmed in larger studies, these data would suggest that specialized screening procedures, such as measures of carotid intima-media thickness may be required in HIV positive patients with CD4 counts above the threshold for therapy, and that antiretrovirals are instituted before the CD4 count decreases to below the threshold for therapy in those patients considered at risk for CLI. There are a number of limitations to the present study that warrant consideration. The small study sample raises the question of false positive and negative findings and hence a much larger study is required to confirm aspects of the present data. Nevertheless, it is unlikely that a larger study will improve on our ability to show differences in the presence of a leukocytoclastic vasculitis in 9 of 10 HIV positive patients and none of the 10 HIV negative patients. In addition, because CLI is associated with disturbances of coagulation and fibrinolytic systems before, during and up to 30 days after surgery we were unable to assess the relationship between coagulation profiles and HIV in CLI. Further, as a consequence of limb ischemia and the presence of ulcers and/or gangrene (tissue necrosis) in all patients, which affects circulating inflammatory and immune activation markers, we were also unable to assess the relationship between inflammation or immune activation and HIV in CLI. Hence, we cannot exclude the possibility that a pro-coagulation state, well recognized to characterize HIV and which is associated with large artery changes in non-human primate models of immunodeficiency virus infections, is the main role player for CLI in HIV or whether inflammation and immune activation are important causes of HIV-associated CLI. In conclusion, in the present study we show that as compared to HIV negative patients with CLI, leukocytoclastic vasculitis of the vasa vasorum and adventitial inflammation characterize large artery pathology in HIV positive patients with CLI not receiving anti-retroviral therapy. Moreover, we show that a significant proportion of HIV positive patients with CLI not receiving anti- retroviral therapy have advanced femoral artery atheromatous changes, despite a lower cardiovascular risk. These data may have implications for both primary prevention and for the management of occlusive arterial disease in HIV positive patients. # Supporting Information This study would not have been possible without the voluntary collaboration of the participants. [^1]: Dr Simon Nayler is employed by Gritzman and Thatcher Inc. which is a private pathology laboratory. The authors specifically contracted Dr Nayler for this study because of his reputation and expertise as a histopathologist, especially in the cardiovascular field. Dr Nayler has no invested interests nor stands to gain from the outcomes of this study. Hence, the authors do not believe that he has a conflict of interest. Furthermore, this does not alter the authors’ adherence to PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: MB AJW FM SN MGV GRN. Performed the experiments: MB SN. Analyzed the data: MB AJW. Contributed reagents/materials/analysis tools: MB AJW SN GRN. Contributed to the writing of the manuscript: MB AJW FM SN MGV GRN.
# Introduction The association between birth weight and later risk of type 2 diabetes is U-shaped. One arm comprise men and women born small, due to maternal undernutrition or other factors. The other arm comprises of individuals born large, due to maternal adiposity and diabetes. South Asians have a higher prevalence of type 2 diabetes compared with Europeans, and are diagnosed at a younger age and with a lower BMI. The underlying mechanisms are only partly understood, but probably include that South Asian babies are born smaller, but relatively adipose, i.e. have a reduced lean mass, but higher fat mass, compared with Europeans. This “thin-fat” phenotype of South Asians tracks through life, and could result in increased insulin resistance. South Asian pregnancies present a mixture of characteristics: on one hand shorter maternal height, lower BMI and a higher risk of micro-nutritional deficiencies; factors which may constrain fetal growth, on the other hand higher adiposity, insulin resistance and hyperglycemia which may enhance growth. The usual description of a baby born to a mother with gestational diabetes (GDM) includes macrosomia and large for gestational age phenotypes in late pregnancy and at birth. Some studies indicate a stronger association between maternal hyperglycemia and offspring birth weight in ethnic South Asians than in Europeans. However, little is known about the fetal growth trajectories in GDM and non-GDM pregnancies, in populations universally screened for GDM, and if this differs between these two ethnic groups. The population-based multi-ethnic STORK-Groruddalen cohort provided a unique opportunity to explore these relationships. We compared maternal characteristics in early pregnancy, and fetal size and growth rate during the second half of pregnancy, in ethnic South Asian and European pregnancies with and without GDM in Oslo, Norway. ## Ethics statement Pregnant women attending the Child Health Clinics for antenatal care in three municipalities in Groruddalen, Oslo, from May 2008 to May 2010 were given oral and written information about the Stork Groruddalen project and invited to participate. The women who chose to participate gave informed written consent at inclusion, on behalf of themselves and their offspring. The study protocol and the consent-forms were approved by The Regional Committee for Medical and Health Research Ethics for South Eastern Norway, and The Norwegian Data Inspectorate. ## Population and design The study design has been presented in detail elsewhere. Prior to the study, 75–85% of all pregnant women in this area attended the local Child health Clinics during their pregnancy. Questionnaire data was collected by specially trained study midwives through interview, supported by a professional interpreter, using translated questionnaires when needed. Information material and questionnaires were translated to eight languages including English, Tamil and Urdu, covering the largest South Asian ethnic groups. Women were eligible if they were: (1) living in one of the districts, (2) would give birth at the study hospitals, (3) were in gestational week \< 20, (4) not suffering, at the time of enrollment, from diseases necessitating intensive hospital follow-up during pregnancy (i.e. pre-gestational diabetes and other substantial medical, psychiatric or obstetrical conditions) (5) not already enrolled with a pregnancy lasting \> 22 weeks, (6) could communicate in Norwegian or any of the other eight languages, and (7) were able to give informed consent. The final sample consisted of 823 pregnant women. Overall participation rate was 74% (81.5% in Europeans and 73.0% in South Asians) and the study cohort was considered representative for women attending the Child Health Clinics with respect to ethnicity and age. ## Maternal factors The participating woman and fetus were defined as ethnic European or South Asian if the participating woman or her mother were born in Europe or South Asia respectively. Europeans also included three women of European origin born in North-America. As few South Asians migrated to Norway before the 1970\`s, all potential participants with South Asian ancestry were either born in South Asia, or their mothers were born there. Glycated haemoglobin (HbA1c) and fasting glucose were measured at enrollment. HbA1c was measured with HPLC (Tosoh G8,Tosoh Corporation, normal reference range 4–6%). A standard 75 g oral glucose tolerance test (OGTT) was performed in gestational week 28±2. Women were diagnosed with gestational diabetes according to the 1999 WHO-criteria (fasting plasma glucose ≥7.0 mmol/l or 2-hour glucose ≥7.8 mmol/l). Glucose was measured on site (within 5 minutes after vein puncture) in venous EDTA blood according to a standardized protocol, using a patient-near method (HemoCue 201+, Angelholm Sweden) calibrated for plasma. According to national guidelines, women with fasting glucose \< 7.0 mmol/l and 2-hour values 7.8–8.9 mmol/l were categorized as “mild GDM”, and were given oral and written lifestyle advice and referred to their general practitioner for follow-up. Women with fasting glucose ≥ 7.0 mmol/l or 2-hour values ≥ 9.0 mmol/l were categorized as “moderate/severe GDM”, and referred to secondary care. Data related to GDM-treatment have been collected from hospital records. Maternal height was measured twice to the nearest 0.1 cm with a fixed stadiometer at inclusion, and the mean was used. Pre-pregnant body mass index (BMI, kg/m<sup>2</sup>) was calculated using self-reported pre-pregnant weight, to the nearest kg, and measured height. Weight gain (WG) in kg, was calculated from pre-pregnancy to enrollment, from enrollment to week 28 and from week 28 to birth (see footnotes Tables and). Parity was categorized as “primiparous” or “parous” (at least one previous pregnancy lasting \> 22 weeks). Except in 7% of women gestational age was derived from the first day of the woman’s last menstrual period (see footnote). ## Fetal- and neonatal measurements At each ultrasound visit, scheduled at 24, 32 and 37 weeks of gestation, participants were randomly allocated to one of four study-ultrasonographers. One Voluson Pro (GE-Healthcare) machine, with a AB2-7 scan head, was used. Methods have been described in detail elsewhere. Using standard anatomical landmarks, abdominal circumference (AC) and head circumference (HC) were obtained by applying computer-generated elliptical measurements to the outer surfaces, while femur length (FL) was obtained in a longitudinal section. Each biometric variable was measured three times, according to a study-specific standardized protocol, and the mean value was used. Estimated fetal weight (EFW) was calculated using Combs formula: EFW = (0.23718 x AC<sup>2</sup> x femur length) + (0.03312 x HC<sup>3</sup>), as used in clinical practice in the study- hospitals at that time. Birth weight was routinely measured on calibrated electronic scales immediately after birth. Within 72 hours after birth, study-specific anthropometric measurements were performed by trained study personnel, according to study protocol, unless contraindicated for medical reasons. Crown-heel-length (CH- length) was measured by a measuring rod, with the head firmly held, while stretching the legs. For circumference measurements a non-elastic plastic tape was used. CH-length, HC and AC were measured to the nearest 0.1 cm. ## Statistical methods Z-scores for fetal measurements (differences in standard deviations from gestational age specific mean) were calculated using formulas derived from an ethnic Norwegian reference population. National birth weight references were used to calculate birth weight z-scores in the total study sample. In the large sub-sample of term neonates with study-specific anthropometric measurements at birth, we have previously calculated individual z-scores, stratified by gestational week and gender for birth weight, HC, AC and CH-length. Statistical analyses were performed using SPSS version 22.0 for Windows (SPSS Inc., Chicago, IL, USA). The significance level was set to p\<0.05. Differences between pregnancies complicated by mild or moderate/severe GDM, compared with non-GDM pregnancies were explored using Chi Square test for categorical and t-tests for continuous variables. Differences in fetal growth from week 24 until birth were examined using linear mixed effects modelling, assessed separately for the outcomes weight, HC, AC and length. Z-scores through four different time points (week 24, 32, 37 and at birth) were entered. For the analysis of length we used z-scores for femur length at three time points in pregnancy and CH-length from birth. For the analyses of weight we used z-scores of EFW from three time points during pregnancy and birth weight. All models were run with a random intercept and a random slope. We first explored the impact of ethnicity on fetal growth in non- GDM pregnancies. Ethnicity was first entered as a fixed variable together with gestational age (weeks) and the interaction term between these two variables (to assess the impact of ethnicity on growth velocity). A priori, we added the covariates offspring gender and maternal parity to the adjusted models. As lower maternal height is often perceived to explain a substantial part of ethnic differences in fetal growth, maternal height was also included in the last supplementary model. A second degree term for gestational age was also included as a fixed variable in all models, due to a small, but systematic variation in the distribution of the fetal z-scores across the four time points. Interaction terms between other covariates and gestational age were tested for and included if significant. We then performed similar analyses replacing ethnicity with GDM as explanatory variable, first treated as non-GDM or GDM, thereafter stratified into non-GDM, mild GDM and moderate/severe GDM. We then added the covariates maternal ethnicity, parity and offspring gender to the adjusted models, and maternal height in a final supplementary model. Lastly, we performed the analyses in Europeans and South Asians separately. Estimated mean SD-scores from the adjusted mixed models from each of the four time points were extracted. ## Sample size The present study was restricted to women of European (n = 379) and South Asian (n = 200) ethnic origin. After excluding twins and those with missing newborn data, OGTT or ultrasound, the final study sample consisted of 349 European and 184 South Asian (63% Pakistani and 37% Sri Lankan/Indian) women (Flow chart). Of these, 72% were delivered at term and had study-specific measurements at birth. # Results ## Maternal characteristics At enrollment (mean gestational week 15), South Asian women were shorter, younger and less educated than European women. Their mean BMI was slightly lower, but they had more subcutaneous fat represented by thicker skin-folds. Furthermore, South Asian women had significantly higher fasting glucose and HbA1c, with a distribution skewed to the right. Nevertheless, few women had fasting glucose or HbA1c values at levels which may be considered indicative of impaired glucose control, irrespective of later GDM diagnosis. In total 39 (11%) European and 28 (15%) South Asian women were diagnosed with GDM in week 28 according to the WHO 1999 guidelines. Relatively more South Asian women had moderate/severe GDM (p = 0.04 for the difference between ethnic groups). Of these, two South Asian and one European woman received insulin treatment during pregnancy. None were treated with oral glucose-lowering drugs. European women later diagnosed with GDM were more likely to be overweight or obese when entering pregnancy than their non-GDM counterparts, in particular women diagnosed with moderate/severe GDM. However, they showed less weight gain throughout pregnancy than non-GDM women. A different pattern was seen in South Asian women diagnosed with GDM, as no differences compared with their non-GDM counterparts were observed for BMI, overweight and gestational weight gain. The majority of South Asian women with moderate/severe GDM (79%) were normal weight and only one was obese. ## Fetal growth In week 24, fetuses of South Asian non-GDM mothers had longer femurs, but smaller ACs than their European counterparts. From this time they grew significantly slower, and at birth they were smaller on all measures. Fetuses of mothers who were later diagnosed with GDM were smaller on all fetal measures in week 24 (mean difference in EFW z-score: -0.30 SD (-0.53, -0.07), p = 0.01), compared with fetuses of non-GDM mothers. From this time until birth they showed faster growth. However, when we categorized GDM into “mild” and “moderate/severe” GDM, and South Asian and European women were analyzed separately, different patterns emerged. South Asian fetuses exposed to mild GDM had a similar size and growth rate during the second half of pregnancy as South Asian non-GDM fetuses. In contrast, fetuses exposed to moderate/severe GDM were markedly smaller on all measures in week 24. EFW was -0.95 SD (-1.53, -0.36), p\<0.001) at this time point. From this time until birth, they showed a faster growth rate on all measures, and birth weight was 0.45 SD ((0.09, 0.81), p = 0.01) higher than in non-GDM pregnancies. This was also reflected in a 92 g larger placenta weight (p\<0.05). European fetuses exposed to GDM also tended to be smaller in week 24 and show faster growth than their non-GDM counterparts. # Discussion In this study we observed that fetuses of mothers later diagnosed with GDM tended to be smaller in week 24, compared with fetuses of non-GDM mothers. From this time until birth they showed a faster growth. This pattern was most pronounced in South Asian fetuses of mothers diagnosed with moderate/severe GDM, who were the smallest for all body measures in mid pregnancy. This study is unique, as we were able to combine longitudinal maternal and offspring data from early pregnancy to birth, including GDM-data from universal screening, in a well-characterized population-based cohort, in two different ethnic groups. The smaller fetal size in mid-pregnancy in GDM, compared with non-GDM pregnancies, was unexpected. However, an early fetal growth delay has been reported both in Type 1 and Type 2 diabetes. There are only few reports of fetal growth in pregnancies which are later diagnosed with GDM. A recent study from the UK did not observe a significant association between GDM and fetal size in week 20, in a predominantly ethnic white British cohort. They did, however, find an accelerated fetal growth velocity between week 20 and week 28, hence preceding the clinical diagnosis of GDM, which is in line with our study. The slower growth in early pregnancy in Type 1 and Type 2 diabetes has been associated with indicators of poor glucose control in early pregnancy and with lower levels of some biomarkers indicating impaired early placentation. Few women in our study had high HbA1c or high fasting blood glucose in early pregnancy, irrespective of GDM-severity later in pregnancy. For a given BMI, South Asians tend to have lower lean mass, including lower muscle mass and fewer insulin-producing β-cells, but relatively more fat mass than ethnic Europeans. We have previously shown from the same cohort that South Asian mothers had more subcutaneous fat and higher serum-leptin levels in early pregnancy compared with Europeans, despite having a lower BMI, indicating a thin-fat phenotype.. Hence, the substantial “growth delay” in mid pregnancy, observed in fetuses of predominantly normal weight South Asian women diagnosed later with moderate/severe GDM, could still indicate subtle metabolic and placental abnormalities from an early stage in pregnancy. Our results could also indicate that there may be different types of GDM. One being GDM in an obese women, with pre-existing insulin resistance and a sufficient capacity to produce insulin in a non-pregnant situation, but decreased capacity to compensate for the added insulin resistance caused by pregnancy. The other being GDM in a “thin-fat” woman, predominantly due to a reduced capacity to produce insulin but also diminished insulin sensitivity due reduced muscle mass. It has also been speculated that maternal hyperglycemia during the second half of pregnancy may in part be triggered by the fetus through the placenta, to promote intrauterine “catch-up” after early growth failure. If such a hyperglycemic response is aiming to compensate for early placental insufficiency, we could further speculate that very intensive treatment of GDM, primarily by a very strict glucose control, if followed by a low maternal weight gain, may aggravate growth failure, in particular in normal weight GDM women. The key challenge is that optimal fetal growth remains to be defined. Customized fetal growth charts aim to predict the growth potential, i.e. “optimal growth” for each baby, by adjusting for characteristics which influence birth weight, such as maternal height, parity and ethnic origin, and by excluding pathological factors such as smoking and diabetes. However, ethnic differences in fetal growth are observed, and only partly explained by key maternal factors, such as parity, socioeconomic position and height. Factors which are independently associate with fetal size at different time points during pregnancy, and across ethnic groups, await elucidation. Fetuses of South Asian non-GDM mothers had longer femurs, but smaller ACs in mid pregnancy, compared with their European counterparts. From this time until birth they showed a slower growth on all measures. Hence, at birth South Asian neonates were thinner, represented by a markedly smaller AC, while length and HC were modestly smaller. Whether this growth pattern is optimal for South Asians is not known. When using the Norwegian national birth weight references, more than 20% of the South Asian neonates in our cohort were defined as being small for gestational age, while few were large for gestational age, reflecting that the birth weight distribution is skewed compared with the distribution in the native Norwegian population. We did not observe any differences in birth complications between pregnancies with and without GDM, or between ethnic groups in our study. However, we have limited statistical power to answer these comparisons. In the short run, for perinatal outcomes, growth restriction diagnosed by ethnic- specific fetal growth charts seems to better predict adverse perinatal outcomes. Nevertheless, for long term outcomes, such as adiposity and type 2 diabetes, both “thinness” at birth and excess fetal growth have been associated with a higher risk. From this we could speculate that South Asian fetuses, both those exposed to GDM and those not, may be at increased risk of type 2 diabetes in later life. The study has some limitations. Most women diagnosed with GDM were treated non- pharmacologically, indicating that there were few cases of severe GDM. However, treatment could still have influenced the fetal growth patterns. Furthermore, the limited number of GDM-cases, further categorized into two groups, primarily based on elevated 2-hour-values, restricts our power to adjust for many covariates, and to explore interactions. However, as women in the two GDM- categories were given different follow-up after being diagnosed, the alternative of not taking this into account could cause bias. Almost all South Asian women with moderate/severe GDM were normal weight before entering pregnancy but had similar weight gain as non-GDM women. Most Europeans, however, were obese and had substantially less weight gain than their non-GDM counterparts. Adjusting for these factors could therefore potentially also cause bias. Our findings illustrate the complexity involved in the relationship between maternal factors and fetal growth, should be interpreted with caution. ## Conclusions To conclude, we observed differences in fetal size and body proportions from before diagnosis and treatment of glucose intolerance. The most prominent growth deviations was observed in fetuses of South Asian mothers diagnosed with moderate/severe GDM. Their babies were small in size in mid-pregnancy but but subsequently grew faster until birth compared with babies of non-GDM mothers. The mechanisms underlying these differences are likely to be at least partly present before the pregnancy. Our results highlights the need for large-scale comparative studies of serial fetal growth from early pregnancy. This could have implications for the timing of diagnosis and the treatment of pregnancy-related hyperglycemia in different ethnic groups. # Supporting information The authors thank all the study participants, and Nanna Voldner and Jorid Bech Barbøl for performing ultrasound examinations. We also thank study staff and general staff at the Child Health Clinics in Stovner, Grorud and Bjerke districts in Oslo and at the delivery- and post-natal wards at Akershus University Hospital and Oslo University Hospital, for help with collecting the data. [^1]: The authors have declared that no competing interests exist. [^2]: Conceptualization: LS AKJ. Formal analysis: LS. Funding acquisition: AKJ BN. Investigation: LS AKJ KM OHRJ SV. Methodology: LS AKJ CSY SV. Project administration: LS AKJ KIB. Visualization: LS. Writing – original draft: LS. Writing – review & editing: LS AKJ CSY KM BN OHRJ KIB SV.
# 1 Introduction Depression is one of the most common mental disorders in the United States. The 2017 US National Surveys on Drug Use and Health (NSDUH) indicated that approximately 13.2% (3.2 million) of adolescents aged 12 to 17-year-old and 7.1% (17.3 million) adults 18-year-old and older reported experiencing at least one major depressive episode. Although the prevalence of depression in the US population, especially among young adults has increased, and a variety of pharmacological treatments are available, a large proportion of individuals with depression delay seeking treatment or avoid it altogether. According to the 2017 NSDUH data, an estimated 35% of adults and 60.1% of adolescents who had a major depressive episode did not receive treatment. Current medical cannabis policies across the US do not include depression as a medical qualifying condition for medical cannabis use. However, emerging research indicates that coping with depression is often reported as an important reason for cannabis use. However, the potential causal relationship and directionality between cannabis use and depression remain uncertain. There is a paucity of research on the topic and existing studies mainly focus on treatment centers data. Furthermore, the on-going and rapid changes in the US cannabis legislative landscape along with the increased potency of cannabis over the past twenty years call for timely epidemiological monitoring of lay practices and therapeutic uses of cannabis products in order to assess the impact of policy changes, and identifying emerging issues and trends. In this context, social media platforms play an important role in uncovering experiences of individuals and their health-related knowledge. Social media data offers the possibility to indirectly collect information about those who do not receive treatment while having depression and using cannabis. Although user generated content area constitutes a rich source of unsolicited and unfiltered self-disclosures of attitudes and practices related to cannabis use, they have not been explored to derive insights about causal relationship between cannabis and depression. Despite the recent improvements in Natural Language Processing (NLP) techniques, scientific literature utilizing NLP to investigate this type of relationship and/or focus on cannabis use remains sparse. Research has investigated the relation between cannabis use and psychosis based on Electronic Medical Records using NLP techniques. Basic NLP techniques were also used to assess the frequency of experienced effects and harms of different generations of synthetic cannabinoids in drug-focused web forums. However, and to the best of our knowledge, no research using relation extraction has investigated the link existing between cannabis and depression based on social media data. Therefore, this research aims to design a relation extraction method facilitating the identification of the causal relationship between cannabis and depression as expressed by cannabis users using Twitter data. While we acknowledge that correlations so derived are not to be confounded with causation, they do provide insights on potential hypotheses that can be explored through RCTs in the future. We formulate this problem as the extraction of relationship between cannabis use and depression in terms of four possible relationships namely: Reason, Effect, Addiction, and Ambiguous (c.f.). Extracting relationships between any concepts/slang-terms/synonyms/street-names related to ‘cannabis’, and similarly those related to ‘depression’, requires a domain ontology. Here, we use Drug Abuse Ontology (DAO) a domain-specific hierarchical framework containing 315 entities (814 instances) and 31 relations defining drug-use and mental-health disorder concepts. The ontology has been utilized in analyzing web-forum content related to buprenorphine, cannabis, synthetic cannabinoid, and opioid-related data. The DAO included representations of mental health disorders and related symptoms that were developed following DSM-5 classification. These terms were collected from the medical literature related to substance use, abuse and addiction. In addition to medical terminology, DAO included commonly used lay and slang terms that were identified using prior clinical literature and social media-based studies on depressive symptomatology. The DAO was expanded using DSM-5 categories covering the most common mental health disorders by utilizing the study of for improving data collection about mental health and cannabis use on Twitter. The lexicon for DSM-5 has been constructed by utilizing publicly available knowledge bases, namely, ICD-10, SNOMED-CT, and DataMed, along with enriched Drug Abuse Ontology. For entity and relationship extraction (RE) task, previous approaches generally adopt deep learning models, in particular, Convolutional Neural Network (CNN) and Bi-directional Long Short Term Memory (Bi-LSTM) networks. However, Bi- LSTM/CNN model does not generalize well and performs poorly in limited supervision scenarios. Recently, several pre-trained language representation models have significantly advanced the state-of-the-art in various NLP tasks. BERT is one of the powerful language representation models that has the ability to make predictions that go beyond the natural sentence boundaries. Unlike CNN/LSTM model, language models benefit from the abundant knowledge from pre- training using self-supervision and have strong feature extraction capability. So we exploit the representation from BERT and CNN to achieve best of both the representations using novel gating fusion mechanism. Further, we tailored our model to capture the entities position information (using DAO knowledge) which is crucial for the RE as established in the prior research. We propose an end-to-end knowledge-infused deep learning framework (named, *Gated-K-BERT) based on widely adopted BERT language representation model and domain-specific DAO ontology to extract entities and their relationship. The proposed model has three modules: (1) Entity Locator, which utilizes the DAO ontology to map the input word sequence to the entities mention in the ontology by computing the edit distance between the entity names (obtained from the DAO) and every n-gram token of the input sentence. (2) Entity Position- aware Module, exploits the DAO to explicitly integrate the knowledge of entities in the model. This is done by encoding position sequence relative to the entities. Further, we make the attention layers aware of the positions of all entities in the sentence. (3) Encoding Module, jointly leverages the distributed representation obtained from BERT and entity position-aware module using the shared gated fusion layer to learn the contextualized syntactic and semantic information which are complimentary to each other. Contributions: (1)* In collaboration with domain experts, we introduce an annotation scheme to label the relationships between cannabis and depression entities to generate a gold standard cannabis-depression relationship dataset extracted using Twitter. (2) We propose an end-to-end knowledge-infused neural model to extract cannabis/depression entities and predict the relationship between those entities. We exploited domain-specific DAO ontology which provides better coverage in entity extraction. We further augment the BERT model into knowledge- aware framework using gated fusion layer to learn the joint feature representation. (3) We explored entity position-aware attention in the task to jointly leverages the distributed representation of word position relative to cannabis/depression mention and the attention mechanism. (4) We evaluated our proposed model on real-world social media dataset. The experimental results shows that our model outperforms the state-of-the-art relation extraction techniques. We further analyzed that enhancing neural attention with entity position knowledge improves the performance of the model to predict the correct relationship between cannabis and depression over vanilla attention mechanism. # 2 Related work Based on the techniques, recent existing works can be broadly categorized into the following: 1. Deep Learning (DL) framework: Several DL approaches primarily based on CNN and LSTM techniques has been proposed for RE. A study by develops a hybrid deep neural network model using Bi-Directional Gated Recurrent Neural Network (Bi-GRU), CNN, GRU, and Highway connection for classifying relations in SemEval 2010 and KBP-SF48 dataset. exploited the dependency tree by utilizing Graph Convolutional Neural Network (GCN) to capture rich structural information that has been demonstrated for the RE task. advanced the previous methods based on GCN by guiding the network through the attention mechanism. Another prominent work by explores the adversarial learning to jointly extract entities and their relationship. To further enhance the performance of the DL models, various techniques has also exploited latent features in particular the entity position information in the DL framework. 2. Pre-trained language representation model: Models such as BERT, BioBERT, SciBERT, and XLNet has shown the state-of-the-art performance on RE task. adapted the BERT for the relation extraction and semantic role labeling task. modified the BERT framework by constructing task-specific MASK that control the attention in last layers of the BERT. also modified the original BERT architecture by introducing a structured prediction layer that is able to predict the multiple relations in one pass and make attention layers aware of the entities position. 3. Knowledge-base framework: Study by saw the importance of external knowledge in improving the relation extraction from sentences. The study utilizes the parent-child relationships in Wikipedia and word cluster over unlabeled data into a global inference procedure using Integer Linear Programming (ILP). Experiments conducted on ACE-2004 dataset show that the use of background knowledge improved F-measure by 3.9%. A study by uses the attention model to traverse a medical knowledge graph for entity pairs which assist in precise relation extraction. jointly learn the word and entity embedding (obtained through the TransE) using the anchor context model to extract the relationship and the entities. Some of the other prominent work utilizing knowledge graph for relation extraction are. # 3 Resource creation and annotation scheme The corpus consists of tweets collected under the eDrugTrends project that aimed to analyze trends in knowledge, attitudes, and behaviors related to the use of cannabis and synthetic cannabinoids on Web forums and Twitter. Tweets were collected from January 2017 to February 2019 using Twitter data processing, filtering, and aggregation framework available through the Twitris platform, which has been configured to collect tweets with relevant keywords selected by the epidemiologists in the team and adapted to perform appropriate analysis. Domain specialists (RD and FRL) in the team selected the most adapted keywords to identify both cannabis and depression based on the DAO and prior research. From the available corpus of over 100 million relevant tweets collected so far, we further filtered tweets using DAO based on Cannabis and Depression entities and their respective instances specifically defined by domain experts (substance use epidemiologist) for this context. From that filtered corpus, a sample of around 11,000 tweets was sent for expert annotation to a team of 3 substance use epidemiologist co-authors who have extensive experience in drug use, abuse, and addiction research. Further processing was done on this corpus based on the tweets lacking one of the key concepts related to cannabis/depression and 5,885 tweets were annotated finally. The annotation scheme is based on the following coding: 1. Reason: Cannabis is used to help/treat/cure depression. 2. Effect: Cannabis causes depression or makes symptoms worse. 3. Addiction: Lack of access to cannabis leads to depression, showing potential symptom of addiction. 4. Ambiguous: Implies other types or relationships, or too ambiguous/unclear to interpret. The category “Addiction” is an intermediate between the first two as it indicates that feelings of depression would be resolved if one had access to cannabis (which relates to category 1) and also suggests potential presence of cannabis withdrawal symptoms, thus indicating that cannabis use could lead to depressive mood (as a part of withdrawal symptoms). Due to the brevity and ambiguity of information provided in the tweet content, the team decided to classify such cases as a separate category. The sub-samples of tweets were coded independently by each coder and an inter- coder agreement was calculated. The team went through 3 iterations of coding, assessing and discussing, disagreement, and improving coding rules until an acceptable level of agreement was reached among coders (Cohen’s kappa of 0.80,(c.f.)). Tweets that were coded differently by two primary coders were reviewed by a third coder to resolve the disagreement. This yielded a dataset containing 5,885 tweets out of which (1) 3243 tweets are annotated as ‘Reason’ (2) 707 tweets are annotated as ‘Effect’. (3) 158 tweets are annotated as ‘Addiction’ (4) 1777 tweets are annotated as ‘Ambiguous’. The mean tweet text length is 148 tokens (median 74). The university institutional review board (IRB) approved the study under Human Subjects Research Exemption 4 because it is limited to publicly available tweets. To protect anonymity, cited tweet content was modified slightly. We note that this dataset has some (inevitable) limitations: (i) the method only captures a sub-population of cannabis-depression related tweets in eDrugTrends campaign (i.e. those with terms defined in ontology), (ii) Tweets collected may not be a representative sample of the population as a whole, and (iii) there is no way to verify whether the tweets with self-reported cannabis related depression or cannabis related relief from depression are truthful. The team included researchers with extensive expertise in substance use epidemiology and community-based research, and they contributed to development of the annotated sample. Ethics: Our project involves analysis of Twitter data that is publicly available and that has been anonymized. It does not involve any direct interaction with any individuals or their personally identifiable data. Thus, this study was reviewed by the Wright State University IRB and received an exemption determination. # 4 Our proposed approach In this study, a knowledge-infused RE framework, Gated Knowledge BERT (Gated-K- BERT) is used to identify relations between entities ‘cannabis’ and ‘depression’ in a tweet. Our framework (c.f.) consists of three components discussed as follows: ## 4.1 Entity locator module Let S be an input tweet containing the n words {w<sub>1</sub>, w<sub>2</sub>, …, w<sub>n</sub>}. Extracting relationships between any concepts/slang-terms/synonyms/street-names related to ‘cannabis’ and similarly those related to ‘depression’ require heavy dependency on the domain knowledge model. We used domain-specific DAO to map entities in a tweet to their parent concepts in the ontology by computing the edit distance between the entity names (obtained from the DAO) and every n-gram token of the input sentence. Since, DAO provides much better coverage on the entities, it is assume that entity name will be mention in the sentence. Later, we perform masking on the extracted entities. The reason for masking is to explicitly provide the model with the entity information and also prevent a model from overfitting its predictions to specific entities. For instance, entities related to cannabis in a tweet are masked by ‘\<cannabis\>’. Similarly, entities related to depression are masked with ‘\<depression\>’. By this, we obtain a cannabis entity c and a depression entity d in the tweet, corresponding to two non-overlapping consecutive spans of length k and l: $S_{c} = \left\{ w_{c_{1}},w_{c_{2}},\ldots,w_{c_{k}} \right\}$ and $S_{d} = \left\{ w_{d_{1}},w_{d_{2}},\ldots,w_{d_{l}} \right\}$. In effect, this processing abstracts different lexical sequence in tweets to their meaning. ## 4.2 Entity position-aware module This module is designed to infuse the knowledge of the entity mention in basic neural models to effectively capture the contextual information w.r.t the entities. The module consists of following three layers as: ### 4.2.1 Position embedding layer Inspired by the position encoding vectors used in, we define a position sequence relative to the cannabis entity $\left\{ p_{1}^{c},p_{2}^{c},\ldots,p_{l}^{c} \right\}$, where $$\begin{array}{r} {p_{i}^{c} = \left\{ \begin{array}{ll} {i - c_{1}} & {i < c_{1}} \\ 0 & {c_{1} \leq i \leq c_{k}} \\ {i - c_{k}} & {i > c_{k}} \\ \end{array}\operatorname{} \right.} \\ \end{array}$$ Here, $p_{i}^{c}$ is the relative distance of token w<sub>i</sub> to the cannabis entity and c<sub>1</sub> and c<sub>k</sub> are the beginning and end indices of the cannabis entity, respectively. In the same way, we computed the relative distance $p_{i}^{d}$ of token w<sub>i</sub> to the depression entity. This provides two position sequences $p^{c} = \left\{ p_{1}^{c},p_{2}^{c},\ldots,p_{n}^{c} \right\}$ and $p^{d} = \left\{ p_{1}^{d},p_{2}^{d},\ldots,p_{n}^{d} \right\}$. Later, for each position in the sequence, an embedding is learned with an embedding layer to producing two position embedding vectors, $P^{c} = \left\{ P_{1}^{c},P_{2}^{c},\ldots,P_{n}^{c} \right\}$ for cannabis position embeddings and $P^{d} = \left\{ P_{1}^{d},P_{2}^{d},\ldots,P_{n}^{d} \right\}$, both sharing a position embedding matrix P respectively. Further, we map each of the tokens from the input tweet S to the pre-trained word embedding matrix $E \in \mathcal{R}^{V \times d}$ having the vocabulary size V and dimension d. We used FastText, a pre-trained word embedding. Word Embedding (Word2Vec) utilizes dense vectors to represent each word in the vocabulary by projecting into a continuous vector space and also captures both syntactic and semantic information associated with the words. However, in case of a tweet, incorrect spellings, slangs, and other word forms, are out-of- vocabulary (OOV) terms with respect to the word2vec model. In contrast, character-level embeddings have the ability to address the OOV issues by learning the vectors of character n-grams or parts of the words. FastText, unlike word2vec, is trained on the character-level corpus that enables the model to capture words that have similar meanings but different morphological word formations in a robust manner. We represent the input tweet after applying the word embedding as e = {e<sub>1</sub>, e<sub>2</sub>, … e<sub>n</sub>}, where $e_{i} \in \mathcal{R}^{d \times d}$. Finally each word i in the tweet S is represented as the concatenation of the word embedding and relative distance of position embedding with respect to cannabis and depression: $$\begin{array}{r} {x_{i} = e_{i} \oplus P_{i}^{c} \oplus P_{i}^{d}} \\ \end{array}$$ We denote the final representation of tweet as x = {x<sub>1</sub>, x<sub>2</sub>, … x<sub>n</sub>}. The word feature and position feature representations compose a position-aware representation. ### 4.2.2 Convolution layer A combined representation of word and position embedding sequence x is passed to the convolution layer, where filter $\textbf{F} \in \mathcal{R}^{m \times d}$ is convoluted over the context window of m words for each tweet. In order to ensure that the output of the convolution layer is of the same length as input, we performed the necessary zero-padding on the input sequence x. We call the zero-padded input as $\overline{x}$. $$\begin{array}{r} {f_{i}^{m} = tanh\left( \textbf{F}.{\overline{x}}_{i:i + m - 1} + b \right)} \\ \end{array}$$ where tanh is the non-linear activation function and b is a bias term. The feature map f is generated by applying a given filter F to each possible window of words in a tweet, Mathematically, $$\begin{array}{r} {f^{m} = \left\lbrack f_{1}^{m},f_{2}^{m},\ldots,f_{n}^{m} \right\rbrack} \\ \end{array}$$ We apply different length of context window m ∈ M, where M is the set of context window length. Finally, we generate the hidden state h<sub>i</sub> at time i as the concatenation of all the convoluted features by applying a different window size at time i. ### 4.2.3 Entity position-aware attention layer The intuition behind adding entity position-aware attention layer is to select relevant contexts over irrelevant ones. This position-aware representation of entities in a tweet is further modulated by an ontology developed by domain experts. This enhancement enables us to selectively model attention and weigh entities in a tweet. The position-aware attention layer takes as an input h<sub>1</sub>, h<sub>2</sub>, h<sub>3</sub>, …‥h<sub>n</sub> from the encoding module. We formulate an aggregate vector q mathematically as follows: $$\begin{array}{r} {\textbf{q} = \frac{1}{n}\sum\limits_{i = 1}^{n}h_{i}} \\ \end{array}$$ The vector q, thus, stores the global, semantic, and syntactic information contained in a tweet. With the aggregate vector, we compute attention weight a<sub>i</sub> for each hidden state h<sub>i</sub> as $$\begin{array}{r} {u_{i} = v^{T}\mspace{360mu}\text{tanh}\left( W_{h}h_{i} + W_{q}q + W_{c}P_{i}^{c} + W_{d}P_{i}^{d} \right)} \\ \end{array}$$ $$\begin{array}{r} {\alpha_{i} = \frac{exp\left( u_{i} \right)}{\sum_{j = 1}^{n}exp\left( u_{j} \right)}} \\ \end{array}$$ where, $W_{h},W_{q} \in R^{d_{a} \times d_{h}};W_{c},W_{d} \in R^{d_{a} \times d_{p}};V \in R^{d_{a}}$ are parameters of the network, where d<sub>h</sub> is the dimension of the hidden states, d<sub>p</sub> is the dimension of position embedding, d<sub>a</sub> is the size of attention vector. After applying the attention, the final tweet representation R is computed as $$\begin{array}{r} {R = \sum\limits_{j = 1}^{n}\alpha_{j}h_{j}} \\ \end{array}$$ ## 4.3 Encoding module In the encoding module, we aim to obtain the semantic and task-specific contextualized representation of the tweet. We leverage the joint representation through BERT language representation model and Entity position-aware module. Owing to its effective word and sentence level representation, BERT provide a task-agnostic architecture that has achieved state-of-the-art status for various NLP tasks. We use the pre-trained BERT model having 12 Transformer layers (L), each having 12 heads for self-attention and hidden dimension 768. The input to the BERT model is the tweet S = {w<sub>1</sub>, w<sub>2</sub>, …, w<sub>n</sub>}. It returns the hidden state representation of each Transformer layer. Formally, $$\begin{array}{r} {H_{b}^{1},H_{b}^{2},\ldots,H_{b}^{L} = BERT\left( \left\lbrack w_{1},w_{2},\ldots,w_{n} \right\rbrack \right)} \\ \end{array}$$ where, $H_{i} \in \mathbb{R}^{n \times h_{b}}$ and h<sub>b</sub> is the dimension of the hidden state representation obtained from BERT. We masked the representation of `[CLS]` and `[SEP]` tokens with zero. We obtained the tweet representation via BERT model as follows: $$\begin{array}{r} {B = \frac{1}{n}\sum\limits_{j = 1}^{n}H_{b}^{L - 1}\lbrack j,:\rbrack} \\ \end{array}$$ In our experiments, the representation obtained from the second last (L − 1) Transformer layer achieved the best performance on the task. The representation obtained from the last Transformer layer is too close to the target functions (i.e., masked language model and next sentence prediction tasks) during pre-training of BERT, therefore may be biased to those targets. We also experiment with the `[CLS]` token representation obtained from BERT but that could not perform well in our experimental setting. ### 4.3.1 Gated feature fusion The feature generated from CNN and BERT capture different aspect from the data. These features need to be used carefully to make most out of them. The joint feature obtained from concatenation or other arithmetic operations (sum, difference, min, max etc) often results in the poor joint representation. To mitigate this issue, we propose a gated feature fusion technique, which learn the most optimal way to join both the feature representation using a neural gate. This gate learn what information from CNN or BERT feature representation to keep or exclude during the network training. The gating behaviour is obtained through a sigmoid activation which range between 0 and 1. We learn the joint representation F using the gated fusion as follows: $$\begin{array}{r} \begin{aligned} h_{R} & {= tanh\left( W_{R}.R \right)} \\ h_{B} & {= tanh\left( W_{B}.B \right)} \\ g & {= sigmoid\left( W_{g}.\lbrack R \oplus B\rbrack \right)} \\ F & {= gh_{R} + (1 - g)h_{B}} \\ \end{aligned} \\ \end{array}$$ where, W<sub>R</sub>, W<sub>B</sub> and W<sub>g</sub> are the parameters. Finally, the joint feature representation F fed into a single layer feed-forward network with softmax function to classify the tweet into one of the relation classes, Y = {‘reason’, ‘effect’, ‘addiction’, ‘ambiguous’}. More, formally, $$\begin{array}{r} {p\left( \hat{y} \middle\| S \right) = {softmax}\left( W.F + a \right)} \\ \end{array}$$ where $\hat{y} \in Y$, W is a weight matrix and a is the bias. # 5 Experimental setup and results Here, we present results on the cannabis-depression RE task. Thereafter, we will provide technical interpretation of the results followed by domain interpretation of the results. We have chosen models’ hyper-parameters using 5-fold cross validation on entire dataset. The convolutional layer used filters of lengths 2, 3 and 4 and stride of length 1. The optimal feature size is turned out to be 128. We use Adam as our optimization method with a learning rate of 0.001. Hidden size of feed forward in relation classification layer is set to 100. The position embedding is set to 100 for position-aware attention model. The optimal value of d<sub>a</sub> is found to be 50 in all the experiment. We use the Adadelta optimization algorithm to update the parameters in each epochs. As a regularizer, we use dropout with a probability of 0.3. ## 5.1 Results The dataset utilized in our experiment is described in Section-3. We used Recall, Precision and F<sub>1</sub>-Score to evaluate our proposed task against state-of-the-art relation extractor. As a baseline model, we used *BERT, BioBERT, BERTweet* and its several variation such as: BERT<sub>PE</sub>: We extend the BERT with the position information (relative distance of the current word w.r.t cannabis/depression entities) obtained through ontology, as a position embedding along with the BERT embedding. BERT<sub>PE+PA</sub>: We introduced additional component to the BERT<sub>PE</sub> model by deploying position-aware attention mechanism. summarizes the performance of our model over the baselines. Our proposed model significantly outperforms the state-of-the-art baselines on all the evaluation metrics. In comparison with the BERT & BioBERT, our model achieves the absolute improvement of 2.9% & 3.69% F<sub>1</sub>-Score respectively. Second, the results shows that infusing entity knowledge in the form of entity position- aware encoding with attention can assist in better relation classification. Among all the BERT-based approaches, we found that BERT<sub>PE</sub> did not perform well. Thus merely including position-aware encoding in the BERT framework does not help model to capture the entities information. This may be due to the inbuilt position embedding layer in the BERT model which treats the explicit position encoding as a noise. Further, our observation shows that BioBERT did not generalize well for our task in comparison to the BERT with minor reduction of 0.79% absolute F<sub>1</sub>-Score. Although BioBERT is trained on huge corpus of biomedical literature (PubMed & PMC), however data being noisy hampered to performance. Interestingly, adding the entity position information in the form of the attention (BERT<sub>PE+PA</sub>) boosted the model performance. We report the performance absolute improvements of 0.92%, 2.03%, and 0.65% Precision, Recall, and F<sub>1</sub>-Score points in comparison to the BERT model. This shows that position encoding and position attention when used collectively can assist in capturing complementary features. Our final analysis reveals that solely concatenating two representation (CNN+BERT) may not be enough to capture how much information is required from both of these representations. Our method, which introduces the gated fusion mechanism can address this problem as validated by the improved F<sub>1</sub>-Score (c.f.). We also reported the class-wise performance of our proposed model in. The performance of the classes ‘Effect’ and ‘Addiction’ comparatively lower than other classes. It is because the classes ‘Effect’ and ‘Addiction’ have less samples (707, 158) in the dataset which inhibits to learn and generalize the model, in contrast to other classes. Also, in the real-life the explicit expression of being addicted to cannabis after depression can rarely be identified with a single tweets. It requires more contextual knowledge of users history of at-least 2 weeks tweets to capture the implicit sense of this class. ## 5.2 Ablation study To analyze the impact of various component of our model, we perform the ablation study (c.f.) by removing one component from the proposed model and evaluate the performance. Results show that excluding BERT from the model significantly drop the recall of the model by 5.27%, and F<sub>1</sub>-Score by 3.16%. This shows that contextualized representation is highly necessary for the cannabis- depression classification task. We further observed that entity position-aware attention is highly crucial for improving the precision of the model. We report a reduction of 1.47% in terms of precision after excluding the position attention as the model component. Similarly, removing the position encoding from the input layer also lead to a reduced performance. While, excluding convolution layer from the model leads to significant drop in precision, recall, and F<sub>1</sub>-Score by 5.56%, 9.84%, and 7.89% respectively. Thus, we show that every component in the model is beneficial for the cannabis-depression relation extraction task. We also evaluated our proposed entity locator module (based on edit distance) over simple string matching technique. The results show that though the string- matching technique performs well (94.36% F-Score), there are some cases like ‘smokin chronic’, and ‘marijuana candies’ that are not handled correctly by basic string-matching technique, since DAO contains the standard entity names as ‘smoking chronic’, and ‘marijuana candy’. Unlike in our proposed approach, DAO based entity locator module is build upon the domain-specific ontology which contain medical and slang terms related to substance use, abuse, and addiction. Further, as edit distance method allow the soft-string matching (with the insert, delete, and update operation) within the threshold, it captures to match the ill-formed entities (‘smokin chronic’ to ‘smoking chronic’) more accurately over string-matching method. Our DAO based entity locator is more focused and accurate (97.12% F-Score) in spotting entities related to cannabis and depression in tweets. We also compared the position-aware attention over vanilla (word-level) attention discuss in. We called it vanilla attention as it weighs each word equally regardless of the corresponding entities. Given the hidden states h<sub>1</sub>, h<sub>2</sub>, h<sub>3</sub>, …‥h<sub>n</sub>, in word- level attention, first the hidden state h<sub>t</sub> of each time step t is transformed into u<sub>t</sub> using one-layer feed-forward network. Thereafter, the importance α<sub>t</sub> of each token representation is computed using the softmax layer. Formally, $$\begin{array}{r} \begin{aligned} u_{t} & {= tanh\left( W_{u}h_{t} + b_{u} \right)} \\ \alpha_{t} & {= \frac{exp\left( u_{t}^{\top}v \right)}{\sum_{i = 1}^{i = n}exp\left( u_{i}^{\top}v \right)}} \\ \end{aligned} \\ \end{array}$$ where, W<sub>u</sub> v, and b<sub>u</sub> are weight matrix, context vector and bias respectively. The final tweet representation R is computed as $$\begin{array}{r} {R = \sum\limits_{i = 1}^{n}\alpha_{j}h_{j}} \\ \end{array}$$ The results (c.f.) shows that the position aware attention achieve the better performance (an improvement of 2.45 in F1-Score) over the vanilla attention. ## 5.3 Domain-specific analysis To assess the performance on our model, we examined a set of correctly and incorrectly classified, tweets and came up with the following observations: - Correctly classified tweets generally contained clear relationship words: For example, the following two tweets were correctly classified as expressing cannabis use to treat depression: “weed really helps my depression so much! i get less irritable, laugh, and so much more and people think it as the devil! f\\\ you mean”; “marijuana is seriously my best friend rn. it helps me sooo much with my depression and anxiety.” Both tweet contained word “help” that often times is used to convey a meaning indicating usage of a drug for the treatment of a certain condition. - The following correctly classified example represented a case where relationship indicating “treat” was expressed with a word “for”: “I was forced to tell my family i have a medical for weed bc someone been ratting me out, try explaining medical marijuana for depression to a traditional thinking family, i wanna die”. - Similarly, the following tweet were correctly classified as expressing situations where cannabis use is causing depression and/or making it worse: “me @ me when i realize weed is making me depressed but i keep smoking”. Both tweets contained clear relationship word expressing causation “make/making”. - The incorrectly classified tweets generally were more ambiguous and/or contained implied meanings. For example, the following tweet was labeled as expressing “cannabis use to treat depression” while our model classified it as “ambiguous”: “depression is hitting insufferable levels rn and hot damn i could use some weed.” This is an example, where relationship is implied, and there are no clear relationship word expressed in the text. - The same misclassification occurred with the following tweet: “me: wow i think im depressed i should really go to therapy: doesnt do any of that and instead uses weed to increase the dopamine in my brain*.” In this case, the expression “used weed to increase the dopamine…” implies use of marijuana to improve mood (in this cases depressive mood). Because DAO did not contain similar colloquial expressions to indicate depressive mood, our model failed to correctly classify this tweet. Overall, our model performs better than state-of-the-art algorithm in distinguishing depression as a result of cannabis use and cannabis use as a self-medication for depression. In turn, this model will help future works to collect relevant information specific to the behaviors, attitudes and knowledge of users who use cannabis to palliate their depression as well as information on the Twitter users who suffer from depression because of their previous cannabis usage. # 6 Limitations Limitations are noted. First, our work does not distinguish Cannabidiol (CBD) use from general cannabis use. This is of importance as several studies suggest that CBD could reduce anxiety and potentially depression. Although users tend to be more specific regarding whether they are consuming CBD specifically rather than other form of cannabis, future research on the topic of cannabis and depression using social media data needs to integrate this distinction. Second, although the goal of this study was to design a robust algorithm able to differentiate the causal relationship between cannabis and depression as expressed by Twitter users, our work did not aim to establish the objective causal “directionality” in between cannabis and depression (i.e., is cannabis causing depression or is cannabis a potential treatment for depression?). Third, the model has been trained on Twitter data that are rather short (280 characters maximum) form of text, and may not be as performing on other text format (e.g., blog, web forums pages). # 7 Conclusion This research explored a new dimension of social media in identifying and distinguishing relationships between cannabis and depression. We introduced a state-of-the-art knowledge-aware attention framework that jointly leverages knowledge from the domain-specific DAO, DSM-5 in association with BERT for cannabis-depression RE task. Further, our result and domain analysis help us find associations of cannabis use with depression. In order to establish a more accurate and precise Reason-Effect relationship between cannabis and depression from social media sources, our future study would take targeted user profiles in real-time and study the exposure of the user to cannabis over time informing public health policy. # Supporting information [^1]: The authors have declared that no competing interests exist.
# Introduction Populations exposed to mass conflict and persecution commonly experience extensive losses, experiences that are likely to provoke feelings of injustice and anger associated with symptoms of grief. Yet there is a dearth of research investigating a possible nexus between grief and anger amongst populations living in post-conflict environments. We attempt to identify a subpopulation experiencing combined symptoms of grief and anger amongst survivors of prolonged persecution and conflict in Timor-Leste, and whether that putative pattern is associated with particularly high levels of traumatic loss, persisting preoccupations with injustice, and ongoing family conflict. Anger as an unwanted and commonly dysfunctional emotional reaction has been associated with feelings of injustice amongst populations that have been exploited and persecuted. Having one’s human rights violated or economic goals systematically undermined can understandably lead to normal anger reactions, however, anger can also be associated with a loss of control, aggression and harm to others, including community members, intimate partners and children. Anger has also long been regarded as a core component of the normal grieving process. Moreover, clinical observations have suggested that a failure to resolve anger associated with a bereavement may contribute to the persistence of the grief reaction, presumably because of strong feelings of grievance and injustice associated with the loss. In that regard, it is notable that studies examining the factorial structure of the persisting grief reaction have consistently identified anger and bitterness as core components. For example, a confirmatory factor analysis (CFA) conducted amongst bereaved adults in the USA identified anger/bitterness as one of six symptom domains of the construct of prolonged grief. In keeping with this and other research, the constellation of anger-bitterness has been included in the categories of complex bereavement disorder (CBD), defined as a diagnosis requiring further empirical evidence in DSM-5, as well as in the proposed ICD-11 definition of prolonged grief disorder (PGD). Nevertheless, controversy continues about the nosological status of these categories, particularly insofar as they distinguish pathological from normative forms of grief. Studies amongst post-conflict populations exposed to repeated traumatic losses may shed further light on the role of anger in the grief response. Our past research in Timor-Leste identified what appeared to be a high rate of explosive anger in response to trauma exposure. Explosive anger can express itself as physiological arousal and either verbal or physical aggression, the response characteristically being out of proportion to environmental triggers and experienced as uncontrollable, the subject reacting without immediate thought to the consequences. Although in the aftermath of attacks, the person may feel a degree of relief or vindication, feelings of exhaustion, remorse and/or embarrassment are also common. A population study in a rural and an urban village of Timor-Leste undertaken in 2004 recorded a prevalence of explosive anger of 38%, based on the international threshold of at least one attack of explosive anger a month (noting that the majority of these persons experienced much more frequent episodes). In a six- year follow-up study, the prevalence of explosive anger remained high (36%), suggesting that, at a population level, the reaction had a strong tendency to persist over a prolonged period of time. Applying the stringent DSM-IV definition of intermittent explosive disorder (IED) which mandates the occurrence of acts of aggression in conjunction with anger, the prevalence if explosive anger was 8%, a high rate compared to other countries where the diagnosis has been studied at a population level. A consistent finding of our studies in Timor-Leste is that women reported higher rates of explosive anger and IED than men, the converse of the usual gender pattern recorded in other countries. Although a mixed methods study indicated that a range of experiences (exposure to conflict-related trauma and violent death of others, ongoing adversity, exposure to intimate partner violence) were associated with IED amongst women, these factors applied to other morbid mental health outcomes including post-traumatic stress disorder (PTSD) and depression, suggesting that the risk factors identified to date are not specific to anger. Doubts remain, therefore, about the origins and nature of explosive anger and its high prevalence in Timor-Leste, and why it is particularly common amongst women. In our endeavour to understand this phenomenon, we draw on the Adaptation and Development After Persecution and Trauma (ADAPT) model which highlights the core roles of interpersonal bond disruptions and experiences of injustice, amongst other domains, as major psychosocial challenges confronted by populations exposed to conflict. Although the model suggests that grief and anger represent the quintessential responses to disruptions in bonds and acts of injustice, respectively, these two experiences are likely to overlap given the inter-related nature and meaning of the traumatic events of conflict. Specifically, traumatic losses are likely to occur in settings of gross injustices, thereby provoking simultaneous reactions of anger and grief. Other forms of adversity, for example conditions of material deprivation during and in the aftermath of conflict, may compound and prolong anger and grief. Symptoms of grief and anger in survivors of trauma may lead to ongoing conflict within families, representing one of the more severe longer-term psychosocial consequence of earlier exposure to mass violence. The history of persecution and conflict in Timor-Leste provided a setting to investigate possible associations between grief and anger amongst a population exposed to extensive traumatic losses. The invasion and occupation of the territory by Indonesia in 1975 provoked a low-grade resistance war waged by members of the indigenous independence movement. During the period of conflict, which culminated in a humanitarian emergency in 1999, an estimated quarter of the indigenous population (of 600,000 persons at the time) died as a consequence of atrocities, warfare, burning of villages, murder, famine and untreated illness. In addition, there was widespread loss of property and livelihoods, and forced displacement of whole communities, with kinship and family groups being dispersed, some as refugees to other countries. In the post-conflict phase, further episodes of violence occurred, particularly in 2006–7, when a period of sustained internal conflict led to extensive injuries, deaths and displacement of communities into makeshift refugee camps. Socio-economic development in the newly independent country has been slow, with many families confronting extreme levels of poverty and deprivation. Our aim was to test whether it is possible to identify a combined pattern of explosive anger and grief symptoms (grief-anger) amongst the Timorese population. We hypothesized that a subpopulation with grief-anger would report high levels of traumatic losses, preoccupations with injustice and ongoing adversity including family conflict in the post-conflict environment. We also examined whether women were more likely than men to experience the putative grief-anger constellation. # Materials and methods ## Participants Between June, 2010 and July, 2011, we conducted a survey of all adults, 18 years and older, living in every household in two administrative villages (sucos), one in Dili, the capital, the other, a rural site an hour’s drive away. Each suco is defined by contiguous hamlets (aldeias) falling under the administration of one chief (chefe). GPS and aerial mapping produced by the government for census purposes allowed us to identify all households in a setting where there is an absence of street names and many dwellings are located in remote wooded and mountainous areas. Both study sites were extensively affected by mass violence during the Indonesian occupation (1975–1999) and the subsequent internal conflict (2006–7). ## Field team and procedure The team included 18 Timorese field workers with prior survey experience and/or psychology/public health degrees. They received a two-week training course followed by two months of field testing and piloting of survey measures under supervision. Pairs of interviewers were required to achieve a consistent 100 percent level of inter-rater reliability on the core measures. One hour long interviews were conducted in participants’ homes or another location if preferred by respondents, the procedure ensuring maximal privacy and confidentiality. In villages where families live in close proximity to each other, and where overcrowding is a problem, we sought to ensure privacy by taking participants to garden areas or away from the household to somewhere shaded and quite. We also arranged for children to be entertained by one of our colleagues if they were likely to cause a distraction to participants. Households were visited up to five times in order to meet potential participants. ## Ethics statement The study was approved by the ethics committee of the University of New South Wales, the Ministry of Health of Timor-Leste, and the chiefs of each village. The majority of respondents gave written consent prior to commencement of interviews. Verbal consent was obtained in some cases where respondents were illiterate, trusted witnesses co-signing the forms. The procedure was endorsed by the community and received ethical approval from the University of New South Wales and the Ministry of Health of Timor-Leste. ## Measures Our selection of constructs and the appropriate measures to assess them was based on theoretical considerations and the empirical findings in our past studies examining explosive anger in Timor-Leste. The protocol, including the grief measure, was iteratively field tested amongst communities geographically adjacent and similar in sociodemographic composition to the sites of the definitive survey. In piloting, we applied an iterative process of feedback in which responses and solicited comments by respondents in the field were analysed and considered by a committee comprising Timorese of diverse backgrounds (age, gender, education, position in the community) and expatriate researchers. Measures were reviewed and revised to ensure that the constructs were understood by the community, items were readily comprehended, both semantically and linguistically, and response options (such as likert scales) were appropriately graduated according to the language and culture. ### Exposure to conflict-related traumatic events The 17 conflict-related traumatic events (TEs) listed in the Harvard Trauma Questionnaire (HTQ) were modified to ensure their congruence with the historical context of Timor-Leste. TEs were recorded for two periods: the Indonesian occupation and the subsequent period (including the internal conflict) leading up to the study. We derived four broad TE domains based on their common nature and characteristics: conflict-related trauma, witnessing murder and atrocities, traumatic losses, and extreme deprivations. Each TE item was scored 0–2, the maximum score being assigned if participants endorsed a TE for both time periods. We then generated a summary index for each of the four TE domains based on the addition of endorsed items. ### Ongoing adversity An inventory of daily adversities was developed based on extensive community consultations and refinement of items during piloting. All participants rated each adversity item on a five point scale (1 = not a problem, 2 = a bit of problem, 3 = moderately serious problem, 4 = a serious problem, 5 = a very serious problem). The adversity items were assigned to thematic domains: 1.poverty (insufficient food, lack of money for school fees and to meet traditional obligations to family, poor shelter, unemployment); 2. conflict with family (spouse, children, and extended family); and 3. conflict with community (with young people, and the wider community). The score for each domain was based on the summary score of constituent items (0 for lower levels of seriousness, 1 for moderate through to a very serious problem) ### Preoccupations with injustice Respondents were asked to identify and describe the worst human rights violation or other event associated with injustice they had experienced during three defined historical periods: the Indonesian occupation, the period of internal conflict, and contemporary time. Ratings were assigned as 1 for assigning an unjust event; 2 for experiencing preoccupations relating to the event; and 3 for distress related to these preoccupations. The composite index of injustice reflected the addition of scores for each of the three historical time periods (range 0–3). ### Symptoms of explosive anger Our community measure of explosive anger was developed, tested and modified serially during piloting to ensure its cultural appropriateness and comprehensibility in the local language, Tetum. The screening questions inquired whether participants had ever experienced sudden episodes or attacks of anger and if so, how frequently these attacks occurred. Participants who endorsed attacks at a frequency of at least once a month were then asked about associated characteristics of loss of control, destruction of property, verbal aggression, and physical aggression towards others. We then applied an algorithm to derive a diagnosis of intermittent explosive disorder (IED) according to DSM-IV. In a convergence study, we compared our community index of IED with a blinded diagnosis made on the Structured Clinical Interview for the Diagnostic and Statistical Manual for DSM-IV assigned by experienced psychologists. There was a high level of concordance between the two measures: Area Under the Curve 0.90 (95% CI: 0.83–0.98). in the latent class analysis (described hereunder), we included the five core items of explosive anger as defined by IED each scored categorically (1 = present; 0 = absent): explosive anger attacks, loss of control of anger; destruction of property during attacks; verbal aggression during attacks; physical aggression towards others during attacks. ### Grief symptoms We inquired of all participants whether they had experienced a loss, defined as an event (since 1975) in which someone (e.g. family member, relative, or friend) close to the individual had died or been killed. Those who identified multiple losses were asked to identify the death that had the most impact on their lives, then recording the cause and time of the death. Almost all of these identified losses were related to traumatic deaths or untreated illness occurring during periods of mass conflict. Based on the identified loss event, participants were then asked to rate each of four grief items on a five-point frequency scale (0 = almost never, 1 = rarely, 2 = sometimes, 3 = often, 4 = always) as experienced in the past four weeks. The initial item pool was derived from the literature and contemporary criteria for assessing prolonged grief, the process of piloting reducing the number of symptoms to those that were widely recognised and regarded as core experiences of the Timorese people. The derived three symptom items were: persistent yearning/longing for the deceased, feelings of intense bitterness, and feelings emptiness in relation to the death. The fourth item assessed the level of functional impairment associated with the endorsed symptoms. For the latent class analysis, we assigned a score of 0 for symptoms scored not at all, rarely or sometimes and 1 if rated as often (3 on the scale) or always (4 on the scale). ### Post-traumatic stress symptoms and psychological distress Posttraumatic stress disorder (PTSD) symptoms and general symptoms of psychological distress (comprising depression, anxiety, somatic complaints) were assessed using the Harvard Trauma Questionnaire (HTQ) and Kessler-10 respectively, widely used measures applied in our previous studies in Timor- Leste. In our aforementioned convergence study using the SCID, a satisfactory level of concordance was achieved for PTSD (AUC 0.82,95% CI: 0.71–0.94) and severe distress (compared to major depressive disorder) (AUC 0.79,CI, 0.67–0.91). A score of ≥ 2.2 for PTSD, and ≥ 30 for severe psychological distress (matching the international cut-off) produced the best balance between specificity and sensitivity for each index. Cronbach’s alpha for the PTSD scale was 0.95 and for the K10, 0.90. All measures were translated into Tetum, the most widely spoken language in Timor-Leste. Minor inconsistencies were addressed during piloting and the final versions were translated and back-translated into English. ## Statistical analysis We calculated intra-class correlations to assess for possible clustering within households of indices of grief, psychological distress, PTSD and explosive anger. All correlations were low (\<0.05) indicating negligible clustering by households We used latent class analysis (LCA) to identify clusters of participants according to their pattern of symptoms of explosive anger and grief (each item scored in a binary manner as present or absent). We tested sequential models (one class, two classes, three classes, seriatim) examining a suite of conventional model fit indicators to assess for the best class solution: The Bayesian Information Criterion (BIC), sample size-adjusted Bayesian Information Criterion (SS-BIC), and the Akaike’s Information Criterion (AIC). Lower values of these indicators indicate a better fit in comparing successive latent class models. In addition, we applied the Vuong-Lo-Mendell-Rubin (VLMR) and the Lo- Mendell-Rubin (LMR) adjusted likelihood ratio tests, both of which compare the fit of a latent class model of n classes to one with n+1 classes. In judging the best-fitting model, we took into consideration the principle of parsimony, the degree of class separation, homogeneity of posterior probabilities within classes, and the interpretability of the classes yielded. We draw on conventional criteria in which conditional probabilities of 0.60 or above indicate a high probability of endorsing a particular symptom; values falling between 0.59 and 0.15, a moderate probability; and a value of 0.15 or less, a low probability. After selecting the best-fitting model, we examined for associations between class membership (with the low symptom class as the reference category) and a range of relevant predictors using multinominal logistic regression analysis. The covariates included: sociodemographic characteristics of gender, residency in urban or rural areas, educational attainment, and employment; traumatic domains comprising conflict-related trauma, witnessing murders and atrocities, traumatic losses, and extreme deprivations; current adversities including indices of poverty, family conflict, and communal conflict; and preoccupations with injustice (during the Indonesian occupation, the internal conflict, and in contemporary times). Analyses were performed in STATA version 13 and Mplus Version 7. # Results ## Sociodemographic characteristics Of the 3368 respondents approached, 2964 (men, n = 1451, 49%; women, n = 1513, 51%) completed interviews, a response rate of 83.6% (inability to contact residents was by far the major reason for non-participation). indicates the socio-demographic characteristics of the sample. The mean age was 36.4 years (SD = 14.4), and a larger portion (n = 1844, 62%) resided in the rural area. Two-thirds (n = 2013, 67.9%) were married, the remainder being single/never married (n = 756, 25.5%), widowed (n = 171, 5.8%), divorced (n = 5, 0.2%) or separated (n = 19, 0.6%). In relation to education, 11.6% (n = 343) had completed primary, 12.3 (n = 364) junior, and 26.3 (n = 779) senior high school, whereas 10.7 (n = 317) had received post-school education (college/university). Nearly half (n = 1278, 43.1%) engaged in subsistence farming, domestic duties, or were retired; 34.8% (n = 1032) were occupied with paid employment (in a range of enterprises including government and private sectors); and the remainder were students or unemployed (n = 654; 22.1%). ## Prevalence of explosive anger, prolonged grief, PTSD, and severe distress Two hundred and fifty persons (8.4%) met criteria for explosive anger according to IED criteria. A quarter (n = 779, 26.3%) endorsed one or more symptoms of explosive anger, including sudden anger attacks (n = 1074, 36.2%), loss of control (n = 662, 22.3%), verbal aggression (n = 637, 21.5%), destruction of property (n = 527, 17.8%), and physical aggression (n = 423, 14.3%). Over half (n = 1544, 52.1%) endorsed one or more symptoms of prolonged grief, including persistent yearnings or longings for the deceased (n = 2178, 73.5%), feelings of bitterness about the death (n = 1293, 43.6%), and feelings of emptiness (n = 1152, 38.9%). A third (n = 957, 32.3%) reported functional impairment associated with these symptoms. A similar number (n = 453, 15.3%) met the threshold for PTSD (\>2.2) and severe psychological distress (n = 447, 15.1%) (≥30). ## Exposure to conflict-related traumatic events and ongoing adversity Over half of participants (56.1%) reported experiencing one or more conflict- related traumas including political imprisonment, combat, physical assault, torture, and trauma related to involvement in the resistance movement. Four out five persons reported witnessing murders and atrocities and two fifths traumatic losses, including forced separations and disappearances. Ninety percent experienced extreme deprivations related to access to urgent health care (for self or family), food, water and shelter. ## Ongoing adversity shows the frequency of adversity items. In order, poverty-related items endorsed were: shortage of electricity (n = 1983, 66.9%), no access to clean water (n = 1872, 63.2%), insufficient food (n = 1617, 54.6%) and money (n = 11586, 53.5%), problems accessing transport (n = 1489, 50.2%), environmental problems (n = 1527, 51.5%), lack of shelter (n = 1372, 46.3%), being unable to meet traditional family obligations (n = 1333, 45%); conflict with spouse (n = 446, 15.1) and extended family members (n- = 397, 13.4); youth conflict (n = 574, 19.4%); and safety issues in the community (n = 579, 19.5%). ## Preoccupations with past and present experiences of injustice Distressing preoccupations with events associated with injustice were reported by 13.1% (n = 388) for the Indonesian occupation (1975–1999), 24.6% (n = 729) for the period surrounding the internal conflict (2002) and 18.5% (n = 549) in contemporary times. ## Latent class analysis Serial model testing concluded after assessing a four class LCA model. Fit indicators improved up to the three-class model, the gains then being only marginal when progressing to a four class model. Importantly, the VLMR and the LMR adjusted likelihood ratio tests showed no statistical changes in progressing from a three to four class model. Given these findings and the ready interpretability of the classes, we adopted the three-class model. shows the item probabilities for each class based on symptoms of grief and explosive anger. In the grief class (class 1, comprising 25% of the sample), item probabilities for preoccupations and bitterness were in the high probability range, and feelings of emptiness and functional impairment were in the moderate range. In contrast, all items of explosive anger items in this class fell into the low moderate or low probability range. In the combined explosive grief-anger class (class 2), comprising 24% of the sample, grief symptoms fell into the high (preoccupations) or moderate (bitterness, emptiness, functional impairment) ranges. In contrast to class 1, explosive anger symptoms fell into the high (episodes of explosive anger, verbal aggression) or high- moderate (loss of control, destruction of property, physical aggression) probability ranges. In the low symptom class (class 3), comprising 51% of the sample, there were low probabilities for the majority of symptoms of grief and explosive anger, with only two exceptions: preoccupations/yearning were in the moderate range and the generic item for explosive episodes was in the low/moderate range. ## Comorbidity In comparison to the low symptom class, both the grief and grief-anger classes were associated with PTSD (grief class: OR = 1.68, CI = 1.26–2.25; grief-anger class: OR = 1.99, CI = 1.49–2.67) and severe psychological distress (grief class: OR = 1.61, CI = 1.20–2.16; grief-anger class: OR = 2.42, CI = 1.82, 3.21). ## Associations with past trauma, ongoing adversity and preoccupations with injustice presents the findings of the multinomial logistic regression analysis testing for associations between the designated covariates (trauma, ongoing adversity, preoccupations with injustice) and the LCA classes. In comparison to the low symptom reference class, women and urban dwellers were more likely to be assigned to both the grief and grief-anger classes. The two TE domains of witnessing murder and atrocities, and traumatic losses were both associated with the grief and grief-anger classes (relative to the reference class). In addition, however, the grief-anger class reported greater exposure to traumatic losses than the grief class. Also, the grief-anger class alone reported greater exposure to extreme deprivations related to conflict in comparison to the reference class. In relation to ongoing adversities, both the grief and grief-anger classes exceeded the reference class on the index of poverty; the grief-anger class in turn reported higher rates of poverty than the grief class. Only the grief-anger class reported greater levels of family conflict, in comparisons with both the reference low symptom class and the grief class. Compared to the low symptom class, both the grief and grief-anger classes reported greater preoccupations with injustice for the two historical periods of conflict (the Indonesian occupation and the later internal conflict). Only the grief-anger class, however, reported a higher level of preoccupations with injustice for contemporary times compared to the reference class. # Discussion Our analysis in post-conflict Timor-Leste, identified a typology comprising three subpopulations including those experiencing grief, grief-anger and low symptoms, the first two categories affecting a quarter of adults in the sample. Women and urban-dwellers were more likely to be assigned to both the grief and grief-anger classes. Compared to the low symptom reference class, both the grief and grief-anger classes reported greater exposure to conflict-related murders/atrocities and traumatic losses, more extreme levels of poverty, and distressing preoccupations with injustice related to two successive historical periods of conflict. There were important distinctions between the two morbid classes however, in that the grief-anger class reported greater exposure to traumatic losses (compared to the grief class), greater deprivations during the period of conflict (compared to the reference low symptom class), higher stress levels related to poverty (compared to the grief class), ongoing family conflict (compared to both the reference and grief classes), and preoccupations with injustice for contemporary times (compared to the reference and grief class). Prior to discussing our findings, we consider the strengths and limitations of the study. The sample is one of the largest in the contemporary post-conflict mental health field and we achieved a high response rate. Although sampling was restricted to two localities, the sites were identified initially as being broadly representative of the socio-demographic profile of Timor-Leste as a whole. Nevertheless, replication of the study in other areas of Timor-Leste and in post-conflict countries worldwide will be needed to test the generalizability of our findings. Caution needs to be exercised in inferring causal relationships from cross-sectional data of this kind. Longitudinal studies may assist in delineating the chronological sequencing of the relevant symptom constellations, in particular, whether anger precedes and thereby acts to prolong symptoms of grief. Recall of traumatic events can be subject to amnestic bias, although there was a notable consistency in the pattern of traumas documented and the known history of Timor-Leste. A systematic approach was followed in the transcultural adaptation, translation and testing of measures. Although the majority of losses identified as triggers of grief symptoms occurred several years earlier, our measure did not record the course of grief symptoms (whether fluctuating or chronic) so that judgement is reserved as to whether the reaction was prolonged or not. Caveats notwithstanding, our findings cast new light on the high prevalence of explosive anger previously identified in community samples in Timor Leste, a phenomenon that has yet to be fully unexplained. Even though previous studies had shown associations between explosive anger and common stressors such as conflict, poverty and injustice, these factors were common to other patterns of mental distress including symptoms of PTSD and severe psychological distress. Yet there were reasons to suspect that explosive anger had distinctive (albeit unidentified) antecedents given that the reaction appeared to be relatively independent as a construct from those of PTSD and severe psychological distress. In that regard, the identification of a subpopulation comprising a quarter of the sample that manifested the constellation of grief-anger offers a potential explanation for the high prevalence of anger identified in this society. Notably, although a grief class (with low anger symptoms) of equal size emerged, there was no independent explosive anger class, further accentuating the close association between anger with grief. The grief-anger class reported the greatest exposure to traumatic losses, an important finding given that murder, atrocities and death by untreated illness and famine were widespread during the prolonged period of conflict in Timor- Leste. It may be that in collectivist societies such as Timor-Leste, losses that provoke strong and enduring feelings of injustice are particularly potent in generating the identified combined pattern of grief-anger. To confirm this, replication of our findings will be needed in other conflict-affected settings where traditional family and community values prevail. Importantly, our regression analysis involving relevant covariates added credibility to the distinction we found between the grief-anger and grief classes. Specifically, the grief-anger class stood out in reporting high levels of traumatic loss, extreme deprivations during the period of conflict, severe ongoing poverty and family conflict, and preoccupations with injustice extending over three contiguous historical periods. In relation to the latter finding, we have reported a similar association between the anger component of persistent complex bereavement (PCB) disorder and a sense of injustice amongst refugees from West Papua, a neighbouring territory that has experienced a comparable level of prolonged mass conflict under Indonesian occupation. The finding that a half of the population experienced relatively low levels of grief and anger symptoms offers some insights into the factors that protect post-conflict populations from these adverse psychological outcomes. It is notable that the low symptom group reported a similar level of exposure to the general traumas of conflict, indicating that they had not been sheltered from these events. It was only in the TE domains of witnessing murder/atrocities and traumatic losses that the low symptom group reported lower exposure, suggesting that protection from these salient forms of trauma may act to avert risk of developing the specific grief-anger constellation. Being male, living in a rural environment, experiencing lower levels of poverty and not experiencing family conflict were other factors that appeared protective, noting however, that cause-effect relationships involved remain to be confirmed given the cross- sectional nature of the study. Our findings have potential implications for the individual, the family and the society as a whole, not only in Timor-Leste but in other post-conflict settings worldwide. In particular, confirmation of a grief-anger class and the social factors associated with the pattern, has the potential to add support to a cycles of violence model which postulates that exposure to the traumas of past conflict (in this instance, specifically traumatic losses and deprivations) may contribute to risk of subsequent family conflict in the aftermath of the violence. We note however, that explosive anger associated with grief may be both a cause and a consequence of family conflict, resulting in a complex reciprocal and interacting effect that generates a vicious cycle of instability in the household. Our past qualitative data indicated that Timorese women with IED frequently recognized that their explosive anger led to harsh parenting behaviours which in some instances had an adverse effect on the health and well- being of their children. It is possible therefore, that the grief-anger pattern we have identified contributes to the transgenerational transmission of trauma in a manner that impacts adversely on the psychosocial development of the next generation. In relation to ongoing adversities, there appeared to be a stepwise relationship between the severity of poverty and the grief-anger, grief and low symptom classes respectively. These observations underscore the interaction between trauma-related mental health problems and socio-economic factors in post- conflict societies. Poverty places stress on individuals, families and communities, compounding past interpersonal and material losses in generating a sense of injustice and anger. In that sense, apart from the immediate hardship incurred by poverty, conditions of extreme material deprivations jeopardise recovery from trauma-related mental health conditions which in turn can impair functioning and reduce the capacity of survivors to engage in gainful employment or other opportunities to improve their economic well-being.. # Conclusions Our study identified a grief-anger constellation comprising a quarter of the study sample in post-conflict Timor-Leste. There were commonalities with the grief group in reporting greater exposure to witnessing murder, traumatic losses and poverty, and experiencing persisting preoccupations with injustice related to two consecutive historical periods of conflict. The grief-anger group was unique however, in reporting extreme levels of traumatic losses, exposure to material deprivations during the period of conflict, preoccupations with injustice in contemporary times and ongoing family conflict. It is a cruel irony that the traumatic rupture of interpersonal bonds during periods of mass conflict can generate a psychological reaction pattern (grief-anger) in survivors which in turn may undermine the survivor’s capacity to achieve a stable family environment in the post-conflict period. # Supporting information We thank Mr Natalino Tam our Research Manager and our research staff in Timor- Leste their input and contributions to the project. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SR DS. Performed the experiments: ES ZDC. Analyzed the data: AKT. Contributed reagents/materials/analysis tools: AKT. Wrote the paper: SR AKT DS.
# Introduction On March 11, 2011, a 9.0-magnitude earthquake (The Great East Japan Earthquake) struck the east coast of Japan, near Iwate, Miyagi and Fukushima Prefectures. The earthquake, together with the resulting tsunami, caused extensive damage to the Fukushima Dai-ichi Nuclear Power Plant (FNPP), such that a radioactive plume emanating from Units 1, 2, 3 and 4 of FNPP was dispersed into the atmosphere. The total quantity of radioactive materials released into the environment is estimated to be 10% \[1.6×10<sup>17</sup> Bq for <sup>131</sup>I (half-life: 8.0 d) and 1.5×10<sup>16</sup> Bq for <sup>137</sup>Cs (30 y)\] of that released following the accident at the Chernobyl Nuclear Power Plant (CNPP, 1.8×10<sup>18</sup> Bq for <sup>131</sup>I and 8.5×10<sup>16</sup> Bq for <sup>137</sup>Cs). Due to the possible health impacts of this nuclear accident, the Fukushima prefectural government decided to conduct the Fukushima Health Management Survey to assist in the long-term health management of residents, and this is being carried out by Fukushima Medical University. This survey consists of a basic survey to estimate the individual radiation exposure of residents, and four detailed surveys, including thyroid ultrasound examination, comprehensive health check-up, mental health and life-style survey, and a survey on pregnant women and nursing mothers. The thyroid ultrasound examinations on March 11, 2011, targeting all prefectural inhabitants aged between 0 and 18 years (approximately 360,000 inhabitants), have received much attention, since it is well known that the incidence of childhood thyroid cancer increased after the Chernobyl accident. However, for a valid comparison of the results of the ultrasound screening of the thyroid, it is important to understand the normal ultrasound occurring in childhood, since ultrasound technology, including imaging quality, has dramatically advanced in recent years. The prevalence of incidental thyroid abnormalities detected on ultrasound examinations in adults has been well documented. However, few reports have studied the prevalence, spectrum of appearance, and management of ultrasound- detected findings in children, except for the screening programs conducted around Chernobyl. Avula et al. conducted retrospective analysis of clinical and ultrasound findings in 287 Canadian children from 2006 to 2007, and detected incidental thyroid findings in 52 of them (18%). Among them, 35 were small (\<4 mm) well-defined cysts, and nine were hypoechoic, solid nodules with smooth, straight margins, and with echogenicity similar to the thymus, suggesting intrathyroid ectopic thymus. Given the need for large-scale ultrasound screenings of the thyroid gland, we recently performed ultrasound thyroid gland screening in children from three Japanese prefectures (Aomori, Yamanashi and Nagasaki Prefectures), in order to analyze the frequencies of specific ultrasound findings of the thyroid gland, with the same procedures for the ultrasound screenings as used in the Fukushima Health Management Survey, and identified thyroid cysts in 56.9% of the participants. This previous report showed only the categorized results of thyroid findings: “A” (“A1” and “A2”), “B” or “C”, and was based on the crude data only. It is anticipated that the detailed information presented in this report will provide greater disclosure. The current report aims to present further analyses of the results of ultrasound thyroid gland screening in children from these three Japanese prefectures. # Materials and Methods ## Ethics Statement This study was approved by the ethics committees of Hirosaki University, Yamanashi University and Nagasaki University, respectively. It was conducted in accordance with the guidelines expressed in the Declaration of Helsinki. Written informed consent was obtained from the parents of all surveyed children. ## Study Area The survey was conducted in Aomori Prefecture by Hirosaki University, in Yamanashi Prefecture by Yamanashi University, and in Nagasaki Prefecture by Nagasaki University, respectively. These areas were suitable for the current investigation, since they have thyroid ultrasound specialists and medical facilities enabling further examination, and they are located far from each other. More precisely, these areas are geographically dispersed throughout the eastern, central, and western regions of Japan, and are thought to have been unaffected by radioactive material from the FNPP accident. We did not consider the level of iodine intake in each prefecture, because Japan is generally an iodine-rich area. ## Study Population Ultrasound examinations were performed between November 2012 and January 2013. To perform ultrasound examinations in many children at once, and to match the ages of the participants in the Fukushima Health Management Survey, we selected a kindergarten, an elementary school, a junior high school and a high school from each prefecture for this study. We selected these kindergartens and schools because they are located near the center of each prefecture, and furthermore agreed to participate in this study. All children aged 3 to 18 years, from each kindergarten and school, were invited to participate in the study. Because we wished to clarify the general status of the thyroid gland in children, we did not define exclusion criteria in these participants. Children whose parents refused to participate were excluded from the study. Because we emphasized that potential subjects were free to decline to participate on any grounds, we did not collect the reasons why they refused. The overall participation rate was 85.0%. In total, 4,365 children (Aomori: 1,630 children; Yamanashi: 1,366 children; Nagasaki: 1,369 children) underwent the ultrasound examinations. The characteristics of participants in the three examination areas are shown in. The number of children aged 3 to 4 years who were included in the study was 41 in Nagasaki, 5 in Yamanashi and 25 in Aomori, respectively. The largest group of children in all examination areas was 10 to 14 years. In total, 47.5% (2,075/4,365) of the children were male and 52.5% (2,290/4,365) were female. ## Ultrasonography In all cases, ultrasonographic examination was conducted using 7.75-MHz probes (12L-RS linear array transducer \[GE Healthcare, Japan\] and LOGIQ e Expert ultrasound \[GE Healthcare, Japan\]), which is the same equipment as that used in the Fukushima Health Management Survey. The examiner in each prefecture was defined as being an expert in thyroid ultrasonography, and was someone who had experience conducting thyroid screening in the Fukushima Health Management Survey. The examination protocol was the same as in the Fukushima Health Management Survey of the Fukushima Prefecture, i.e. two or more images from each lobe (cross section and longitudinal section) were saved from all children, and the thyroid volume, structure, echogenicity, nodules, cysts, and the presence of pathologies, such as congenital abnormalities (e.g. congenital defect, ectopic thymus or ultimobranchial body) were recorded. In cases with nodules, the assessor recorded the number of nodules and the location of the largest nodule, and measured the greatest dimension of the nodules. In cases with cysts, the assessor also recorded the multifocality of the cysts and the location of the largest cyst, and measured the greatest dimension of the cysts. In cases of questionable findings, expert panels of thyroidologists from the three examination areas, together with thyroid specialists from related Japanese academic societies, discussed them and reached final agreement. ## Statistical methods In this study, the age distribution of the study population differed from that of the standard population of Japanese children. To estimate the general frequency, the age-adjusted prevalence was calculated by using the 2010 Japan standard population distribution for children aged between 3 and 18 years. We used logistic regression analysis to assess the associations between the prevalence of thyroid cysts and nodules and age, gender and the examination areas. We conducted all analyses with SAS for UNIX (SAS institute, Cary, NC). We considered P\<0.05 to be statistically significant. # Results shows the number of cases with thyroid cysts and nodules identified by ultrasound examinations in the 3 examination areas. Overall, thyroid cysts were identified in 56.88% and thyroid nodules in 1.65% of the participants. Most cysts were small in size, with a maximum diameter of 5 mm or less. Thyroid cysts and nodules with a maximum diameter of more than 5 mm were identified in 4.58% and 1.01% of the total participants, respectively. The maximum diameter of thyroid cysts ranged from 0.8 mm to 12.1 mm and most of the children who had cysts had 2 or more cysts (2,240/2,482, 90.25%). The prevalence of cysts varied among the examination areas (58.04% in Aomori, 69.91% in Yamanashi and 42.51% in Nagasaki), the difference being significant when adjusted for age and gender (P\<0.001). However, no significant difference among the examination areas was observed in the prevalence of cysts with a maximum diameter of more than 5 mm adjusted for age and gender (5.09% in Aomori, 5.12% in Yamanashi and 3.43% in Nagasaki, P\>0.2). The maximum diameter of nodules ranged from 1.9 mm to 23.5 mm. Two or more nodules were identified in 26.39% (19/72) of nodule cases. Forty-three participants had both nodules and cysts. The prevalence of nodules varied among the examination areas (2.15% in Aomori, 1.98% in Yamanashi and 0.73% in Nagasaki), the difference being significant when adjusted for age and gender (P = 0.03 for Yamanashi vs. Nagasaki, P = 0.01 for Aomori vs. Nagasaki). However, no significant differences among the examination areas were observed in the prevalence of nodules with a maximum diameter of more than 5 mm adjusted for age and gender (1.29% in Aomori, 1.10% in Yamanashi and 0.58% in Nagasaki, P\>0.1). shows the effect of age on the prevalence of thyroid cysts and nodules. The prevalence of thyroid cysts and nodules, especially those with a maximum diameter of more than 5 mm, significantly increased with age. Females had a significantly higher prevalence of thyroid cysts and nodules with a maximum diameter of more than 5 mm than males. We also estimated the prevalence of thyroid cysts and nodules adjusted for the Japanese population aged between 3 and 18 years. As reflections of the small fraction of children aged 3 to 4 years and the female preponderance of the study population, the age-adjusted prevalence of thyroid cysts and nodules (52.35% for thyroid cysts and 1.54% for thyroid nodules) were found to be less than the unadjusted prevalence (56.88% for thyroid cysts and 1.65% for thyroid nodules). Ultrasound findings other than cysts and nodules are shown in. The most frequent finding was ectopic thymus. The percentage of ectopic thymus decreased with age. Diffuse goiter, ultimobranchial body, lymph node swelling and thyroid agenesis were identified in participants aged 5 years or older. # Discussion Few reports have studied the prevalence, spectrum of appearance and management of ultrasound-detected findings in the thyroid glands of children, except for screening programs conducted around Chernobyl. Within the framework of the Chernobyl Sasakawa Medical Cooperation Project, ultrasound examination of the thyroid gland was conducted around Chernobyl in Ukraine, the Russian Federation and Belarus between 1991 and 1996. In this project, 120,605 children were examined at five centers in three countries, and 63 cases with thyroid cancer (0.052%) were identified. Of note, 38 cases (0.192%) of thyroid cancer were identified in the Gomel region (Belarus), which was the area that was most contaminated by the accident. Also, this study revealed that 42,470 (35.9%) children showed increased thyroid volume (goiter), ranging from 18% to 54% in the five examination centers. This relatively high frequency of goiter reflected the iodine deficient status of this area during the study period. On the other hand, only 502 (0.42%) showed cystic lesions, ranging from 0.19 to 0.63% at the five examination centers, which were much lower frequencies than those observed in our current study. Further, in 2000, we conducted ultrasound examinations of the thyroid gland in 250 Japanese schoolchildren in Nagasaki (Japan) and observed the presence of goiters in only 4 cases (1.60%), while 2 cases (0.80%) had cystic degeneration and a single thyroid cyst, and no cases had thyroid nodules. Of course, a big difference of detection sensitivity and image quality of ultrasound machine at the different time and place of examination should be deeply considered and a standardized method and protocol of ultrasound examination such as common diagnostic criteria should be implemented. In the current study, we conducted ultrasound examination of the thyroid gland in 4,365 Japanese children (from Aomori, Yamanashi and Nagasaki) in 2013 and identified thyroid cysts in 56.88% and thyroid nodules in 1.65% of them, learning from the same method and protocol of the Fukushima Health Management Survey. Following the first report that aimed at a rapid announcement of the crude data according to sociomedical needs, our current further study added the new information. For instance, the prevalence of cysts with a maximum diameter of smaller than 5 mm varied among the examination areas, and the most frequent finding other than cysts was ectopic thymus which showed a decreased prevalence with age. Focusing only on the subset data of Nagasaki in the current study, we identified thyroid cysts in 42.51% and thyroid nodules in 0.73% of the children, which indicates a gap in the frequencies of thyroid cysts and nodules between the studies conducted in 2000 and 2013 (the current study). These gaps between studies could be mainly due to the dramatic advances in ultrasound technology, including imaging quality. New ultrasound technologies, such as digital beam formers, tissue harmonic imaging and speckle reduction have greatly improved the image quality of diagnostic ultrasound machines. These advances enable us to perform detailed evaluation of the thyroid gland by ultrasonography, and consequently could have resulted in the relatively high observed frequency of thyroid cysts in our study, since most of the cysts we detected were less than 5 mm in diameter. In other words, if ultrasound technology advances further, there is a possibility that much more small cysts could be pointed out in the future screening. The prevalence of both thyroid cysts and nodules varied among the three examination areas, although no significant differences were observed in the prevalence of thyroid cysts and nodules with a maximum diameter of more than 5 mm. This suggests that the differences in the frequency of thyroid cysts and nodules among the three examination areas were due to different frequencies in relatively small thyroid cysts and nodules, which were perhaps affected by inter-observer differences. The thyroid ultrasound examination performed in the Fukushima Health Management Survey targeted all prefectural inhabitants who were between 0 and 18 years of age on 11 March 2011 (approximately 360,000 inhabitants). The information released by Fukushima Prefecture indicated that 74,216 (42.56%) of the 174,376 children who underwent thyroid ultrasound examination showed thyroid cysts and 2,014 (1.15%) showed thyroid nodules. Our sample size was much more limited, but the frequency of thyroid cysts and nodules was relatively high in our study, when compared with the data presented in the Fukushima Health Management Survey (56.83% vs. 42.56% for cysts and 1.65% vs. 1.15% for nodules). Since we used the same ultrasonography as was used in the Fukushima Health Management Survey, other factors, such as age composition, inter-observer differences, iodine intake of each region, socio-ecological status and family history or past history of each study participants may associate with the differences between studies. Further analysis, such as comparison with other conducting around Chernobyl or other region with the latest ultrasound machines will be needed. In this study, we observed 85 (1.95%) cases with ectopic thymus. Furthermore, the percentage of ectopic thymus decreased with age. The thymus originates from the 3<sup>rd</sup> pair of branchial pouches, with a rudimentary portion arising from the 4<sup>th</sup> pair, and descends to the superior mediastinum. Anomalies of migration can cause an ectopic thymus along the path of descent. The descents of the thymus and thyroid are closely related because of the proximity of the thyroid diverticulum and the 3<sup>rd</sup> branchial pouch. Thymic tissue can therefore get sequestered within the thyroid, giving rise to an intrathyroid ectopic thymus. The criterion for ectopic thymus in this study was “Intrathyroid lesions showing hypoechogenicity with multiple linear echogenic branching structures or punctate echogenic foci”, which is identical with the criterion of the Fukushima Health Management Survey. On the other hand, discrimination of ultrasonographic findings between ectopic thymus and thyroid nodule, especially for papillary thyroid cancer, is sometimes difficult, and the lack of knowledge has even led to unnecessary thyroid resection. When evaluating ultrasound findings of the thyroid gland, especially malignancies in children, careful observation is necessary to rule out the possibility of ectopic thymus. A large-scaled sample number was the strength of the current study even though the number was relatively small compared with the Fukushima Health Management Survey. There are several limitations to this study. The current study could not include 0 to 2 year old children. Further, the number of study participants aged 3 to 4 years was much smaller than other age groups. We also could not evaluate autoimmunity of the thyroid gland or iodine intake, both of which strongly influence the ultrasound findings of the thyroid gland. Although Japan is generally an iodine-rich area due to the Japanese diet which contains a large amount of seafood, changes in diet in children may affect iodine uptake in the three examination areas. A further follow-up study is definitely needed in order to evaluate the prognosis of the thyroid cysts and nodules obeserved in childhood. In conclusion, we conducted a sophisticated ultrasound examination of the thyroid glands of children from three Japanese prefectures besides Fukushima, and identified thyroid cysts in 56.88% and thyroid nodules in 1.65% of them, respectively. All the inhabitants of Fukushima Prefecture aged 0 to 18 years on March 11, 2011 (approximately 360,000 inhabitants) will undergo thyroid ultrasound examination every 2 years until the age of 20 years, and every 5 years thereafter. The reference data observed in the current study and the further data analyses (including follow-up surveys) of children in general, can provide relevant information for the Fukushima Health Management Survey when followed up in the future through additional population studies. We would also like to thank Mr. Yasuo Kiryu and Ms. Yoshie Hirose, Ms. Akemi Kiko, Ms. Kyoko Takemura, Ms. Misako Konta, and Ms. Michiko Kenmoku for assistance with study arrangement. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NH MI HS NO SN MK SS N. Taniguchi SY N. Takamura. Performed the experiments: NH HS NO YA TN SM SN AO TA SS N. Taniguchi N. Takamura. Analyzed the data: MI KK TA. Wrote the paper: NH MI KK N. Taniguchi N. Takamura.
# Introduction Immune responses include activation and clonal expansion of relevant lymphocyte populations to wipe up invading foreign antigens. Once the antigens are removed, circulating activated lymphocytes, especially T cells, should be eliminated to avoid unnecessary inflammation and to maintain immune homeostasis, while only a few of the activated lymphocytes are fated to develop into memory lymphocytes. In this process, apoptosis plays critical roles. Apoptosis in T cells can be initiated either extrinsically or intrinsically. The extrinsic pathway is initiated by ligation of cell surface death receptors, including Fas. The intrinsic pathway involves mitochondria, where activation of pro-apoptotic members of Bcl-2 family induces the release of cytochrome c (Cyt c) resulting in the formation of apoptosome composed with Cyt c, caspase (Casp) 9, and Apaf1. In both pathways, activation of effector caspases, including Casp3, 6, and 7, downstream either death receptors or apoptosome, leads to degradation of cellular proteins and resultant cell death. While caspase- dependent apoptosis occurs in activated T cells or in T cells deprived of survival factors, caspase activation is reportedly dispensable for thymocyte development and for T cell homeostasis, implying caspase-independent cell death also occurs in T cells. Apaf1, the mammalian homolog of C. elegans CED-4, participates in the formation and activation of the apoptosome. Absence of Apaf1, Casp9 or Apaf1-activating form of Cyt c, leads to the failure of mitochondria-dependent apoptosis. The result is remarkable accumulation of neurons during embryonic brain development, demonstrating the critical role of Apaf1 and other components of the apoptosome in development. Roles of Apaf1 in thymocytes/T cell apoptosis have been shown in a couple of previous reports; while Apaf1-deficient thymocytes were resistant to apoptosis induced by mitochondria-insulting stimulations, such as γ-irradiation and corticosteroid treatment, they were sensitive to Fas-induced apoptosis just like Apaf1-sufficient thymocytes. In addition, successful negative selection of Apaf1-deficient thymocytes showed that Apaf1-dependent apoptosis was not required for the negative selection of thymocytes. Although most of the Apaf1-deficient mice showed perinatal lethality due to the abnormal development of the brain, there were a few survivors, which showed no abnormal accumulation of peripheral lymphocytes, indicating that Apaf1 was not required for the elimination of auto-reactive T cells in the periphery or in the homeostasis of peripheral T cells in number (H. Y., unpublished data). The roles and functions of Apaf1 in the regulation of T cell-mediated immune responses, however, have yet to be clarified. In this study, we took advantage of T cell-specific Apaf1-deficient mice, Lck-Cre-Apaf1<sup>f/f</sup>, in which Apaf1 gene was disrupted with Lck promoter-driven Cre expression, to investigate the biological function of Apaf1 in T cells. Apaf1-deficient T cells showed resistance to mitochondria- dependent apoptosis but showed susceptibility to Fas-mediated apoptosis. We then performed the delayed-type hypersensitivity (DTH) assay, using ovalbumin (OVA)-specific T cell receptors (TCR)-expressing mice (OTII mice), and found that antigen-specific T cell activation leads to enhanced proliferation and Th1-type immune responses in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice, as compared with control Apaf1-sufficient OTII mice. Apaf1-deficient T cells showed higher percentages of cells expressing early activation markers after stimulation than control T cells. Surprisingly, Apaf1-sufficient T cells treated with a pan-caspase inhibitor, z-VAD-fmk (carbobenzoxy-valyl-alanyl- aspartyl-\[O-methyl\]- fluoromethylketone), did not reproduce the activation- related phenotypes observed in Apaf1-deficient T cells, indicating caspase- independent roles of Apaf1 during T cell activation. Our data suggested that Apaf1 in T cells is a negative regulator of immune responses. # Materials and methods ## Generation of T cell-specific Apaf1-deficient mice The design of the conditional targeting vector for Apaf1 is shown in, in which exons 2 and 3 are flanked by two loxP sites. The linearized targeting vector was electroporated into E14K ES cells and homologous recombinants were selected. The heterozygous Apaf1 mutant (Apaf1<sup>f-Neo/+</sup>) ES clones were injected into C57BL/6 blastocysts to generate chimeric mice. Apaf1<sup>f-Neo/+</sup> mice were selected and crossed into C57BL/6 mice more than 10 times; the PGK-neo cassette was removed by mating the heterozygotes with CAG-Flpe transgenic (Tg) mice (RBRC01834, RIKEN BRC). Mice heterozygous for Apaf1 mutation (Apaf1<sup>f/+</sup>) were then crossed with Lck-Cre Tg mice and transgene-positive Apaf1<sup>f/+</sup> mice were intercrossed to generate Lck-Cre-Apaf1<sup>f/f</sup> mice and Lck-Cre-Apaf1<sup>f/+</sup> mice. Lck-Cre-Apaf1<sup>f/f</sup> -OTII mice were similarly generated by crossing Lck-Cre-Apaf1<sup>f/f</sup> mice with OTII-Tg mice expressing OVA-specific TCR. Lck-Cre Tg mice and OTII mice were kindly provided by Dr. A. Yoshimura, Keio University, Japan. Successful disruption of Apaf1 gene was confirmed with genomic Southern blot analysis and absence of Apaf1 protein in Lck-Cre-Apaf1<sup>f/f</sup> T cells was confirmed with Western blot analysis using anti-Apaf1 antibody (PharMingen) ## Animal care This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of Saga University (Approval number 27-047-0). All efforts were made to minimize suffering. ## Apoptosis assay For induction of apoptosis in thymocytes, thymocytes (5 × 10<sup>5</sup> cells/well) were cultured in 96-well plates with or without the following agents or stimulation; dexamethasone (10, 30, 100 nM. Sigma), staurosporine (100, 300, 1000 nM. Sigma), γ-irradiation (100, 300, 1000 cGy), anti-Fas antibody (clone Jo2; 1 μg/ml. PharMingen) plus cycloheximide (0.2, 1 μg/ml. Sigma). After 20 hours, apoptotic cells were detected by staining with Annexin V- FITC (PharMingen) and propidium iodide (PI, BioLegend). For activation-induced apoptosis of peripheral lymph node (LN) T cells, LN cells were stained with FITC-labeled anti-Thy1.2 antibody (eBioscience) followed by positive selection procedure using anti-FITC magnetic beads (MACS; Miltenyi Biotec) and LS columns (Miltenyi Biotec). Purified T cells (\>90% T cells, 3× 10<sup>6</sup> cells/well) were activated with plate-bound anti-CD3ε (0.3 μg/ml; clone 145-2C11, eBioscience) plus anti-CD28 antibodies (3 μg/ml; clone 37.51, eBioscience) for 48 hours in 6-well plates. After removal of apoptotic cells by staining the cells with Annexin V-APC followed by anti-APC magnetic beads (MACS) procedure, remaining cells (Annexin V-negative cells \>97.6%) were cultured (2 × 10<sup>5</sup> cells/200 μl/well) either in fresh medium alone, or conditioned medium prepared from the supernatants of the primary stimulation culture (anti- CD3ε plus anti-CD28 antibodies), or with anti-CD3ε antibody in fresh medium. Twenty-four hours later, apoptotic cells were detected as above. ## Cell activation and cytokine production Mice were subcutaneously (s.c.) immunized with 200 μg of OVA emulsified in complete Freund’s adjuvant (CFA, Difco) at the tail base. Inguinal LNs were collected and single cell suspensions were prepared. Total LN cells (2 × 10<sup>5</sup> cells/well) were cultured in the presence of OVA protein (0, 1, 10, 100 μg/ml) or anti-CD3ε antibody (1 μg/ml) in RPMI1640 medium supplemented with 10% fetal bovine serum. For proliferation assay, cells were pulsed with 1 μCi/well of <sup>3</sup>H-thymidine (PerkinElmer) during the last 7 hours of 48 hours of culture and incorporated radioactivity was measured. Culture supernatants were removed before <sup>3</sup>H-thymidine pulse and IL-2, IFN-γ or IL-17 levels in the supernatants were measured using ELISA Ready-SET-Go (eBioscience). In some experiments, a pan-caspase inhibitor z-VAD-fmk (10 and 100 μM, MBL) was added into the culture. ## DTH assay Seven days after immunization with OVA as above, mice were challenged s.c. at right footpad with 200 μg of OVA in 20 μl PBS. As a control, the same volume of PBS was injected into left footpad. Footpad thickness was measured with a dial vernier caliper (Teclock) on day 1 and 2. The magnitude of the DTH response was calculated as follows; footpad swelling (μm) = thickness of OVA-injected footpad − thickness of PBS-injected footpad. For histological analysis of the DTH lesions, paws were removed on day 2 and fixed with 10%-formaldehyde neutral buffer solution (Nacalai). After decalcification by a standard protocol, specimens were embedded in paraffin and were stained with hematoxylin-eosin (H&E). For analysis of the tissue-infiltrating cells, paws were thoroughly minced with scissors and then were incubated at 37°C for 1 hour in Hank's solution containing 1.0 mg/ml collagenase II (Worthington), 1.0 mg/ml dispase (Sigma-Aldrich) and 40 μg/ml Dnase I (Roche). After removing debris with 70 μm cell-strainers, cells were re-suspended into 33.7% Percoll (GE Healthcare) and pelleted by centrifugation at 1,000 g, for 20 minutes at 24°C. Cells were stained with anti-CD45.2-APC (eBioscience), anti-CD11b-FITC (eBioscience), anti- Ly6G-PE (BD PharMingen), anti-CD3ε-APC (eBioscience) antibodies, and analyzed by flow cytometry for expression of respective molecules. ## Carboxy-fluorescein-diacetate succinimidylester (CFSE) staining Naïve LN cells were suspended in PBS at 2 × 10<sup>7</sup> cells/ml and incubated with CFSE (DOJINDO) at a final concentration of 5μM CFSE plus 0.1%BSA for 10 min at 37°C, followed by washing 3 times with RPMI1640 medium. Stained cells (2 × 10<sup>5</sup> cells/well) were cultured with anti-CD3ε or OVA Peptide (323–339, Peptides International) for 4 days, and evaluated for CFSE intensity. ## Flow cytometry For detection of apoptotic cells, cells were stained as above. For detection of CD69, CD44, and CD62L, LN cells isolated from naïve mice or mice on day 8 after immunization were stimulated with OVA protein (0, 1, 10, 100 μg/ml) or anti-CD3ε antibody (1 μg/ml) and stained with anti-CD69 (PharMingen), anti-CD44 (TONBO Biosciences), and anti-CD62L (BD Biosciences) antibodies on day 0 or 2. For detection of OTII-TCR-Tg-positive cells, cells were stained with anti-TCR Vα2-FITC antibody (clone B20.1, BioLegend). Stained cells were analyzed with BD FACSVerse<sup>TM</sup> (BD Biosciences) for detection of respective surface molecules. ## Western blot analysis For detection of capase cleavage during T cell activation, inguinal LN cells were isolated from mice on day 8 after immunization and were stimulated with OVA protein (0, 10, 100 μg/ml) or anti-CD3ε antibody (1 μg/ml) as above with or without pan-caspase inhibitor z-VAD-fmk (10 and 100 μM). Cell lysates were prepared, electrophoresed, and blotted. Tubulin, caspases 3, 7, and 9 were detected with respective antibodies (anti-tubulin; Sigma Aldrich and anti- caspases; Cell Signaling Technology) and visualized using an enhanced chemiluminescence procedure (ImmunoStar LD, Wako). ## Statistical analysis Experiments were repeated at least three times. Values were expressed as means + SD. Differences between control (Apaf1-sufficient) and Apaf1-deficient samples were analyzed using unpaired t-tests. Differences between two samples among multiple samples with various experimental conditions were analyzed using unpaired t-tests with Bonferroni correction (See figure legends). Alpha value was set at 0.05 and p values below the alpha value (0.05) were considered to be statistically significant. # Results ## Generation of T cell specific Apaf1-deficient mice T cell-specific Apaf1-deficient mice (Lck-Cre-Apaf1<sup>f/f</sup> mice) were generated as described in the Materials and methods. Genotypes of the mice were confirmed by Southern blot analysis of the genomic DNA using two external probes (not shown) and a probe at exon 4 and Cre-driven successful ablation of Apaf1 in T cells and thymocytes were confirmed by Western blot analysis. Mice sufficient or deficient for Apaf1 in T cells were viable, healthy, and fertile and showed no sign of brain deformity (data not shown). Lck-Cre-Apaf1<sup>f/f</sup> mice showed no sign of anatomical and pathological abnormality including lymphocyte accumulation in the lymphoid organs (not shown). In addition, total numbers of thymocytes and cellularity of CD4<sup>+</sup> and CD8<sup>+</sup> thymocytes were comparable between Lck-Cre-Apaf1<sup>f/f</sup> and control Apaf1<sup>f/f</sup> mice (not shown). In our previous report, Apaf1-deficient thymocytes showed resistant to drug-induced and irradiation-induced apoptosis but showed susceptibility to Fas- induced apoptosis over WT thymocytes. As expected, thymocytes from Lck-Cre-Apaf1<sup>f/f</sup> mice showed similar resistance to dexamethasone-, staurosporine-, and irradiation-induced apoptosis over control thymocytes from Apaf1<sup>f/f</sup> mice but showed susceptibility to Fas- induced apoptosis just like Apaf1-sufficient thymocytes. Activated Lck-Cre-Apaf1<sup>f/f</sup> peripheral T cells were refractory to the mitochondria-dependent passive apoptosis induced by growth factor deprivation as compared with Apaf1-sufficient Apaf1<sup>f/f</sup> T cells; when LN T cells that had been previously activated with anti-CD3ε antibody plus anti-CD28 antibody were placed in new medium, in which T cell-derived IL-2 was insufficient, control Apaf1<sup>f/f</sup> T cells showed significant reduction in viability whereas Lck-Cre-Apaf1<sup>f/f</sup> T cells showed higher viability. When activated T cells were re-stimulated with anti-CD3ε antibody in a new medium for activation-induced cell death via the extrinsic pathway of apoptosis, Lck-Cre-Apaf1<sup>f/f</sup> T cells showed substantial susceptibility to cell death just like control Apaf1-sufficient T cells. These data collectively demonstrated that Apaf1 is required for intrinsic pathway of apoptosis induced by mitochondrial insult in both thymocytes and T cells and also that Apaf1 is not required for controlling the number of T cells in the thymus and in the lymph nodes. ## Exacerbation of DTH responses in Lck-Cre-Apaf1<sup>f/f</sup> mice Given the role of Apaf1 in the regulation of intrinsic pathway of apoptosis in T cells, we then asked the pathophysiological roles of Apaf1 in vivo. To do so, we took advantage of OTII mice, in which T cells express OVA-specific T cell receptors. When control Apaf1<sup>f/f</sup>-OTII mice or Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice were subject to OVA-induced DTH responses, significantly higher levels of footpad swelling were detected in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice than in control mice. Histological examination also revealed higher numbers of infiltrating cells in the footpad lesion in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice than in control mice. Flow cytometric analysis showed that the numbers and the percentages of infiltrating neutrophils were higher in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice than control mice while those for lymphocytes and monocytes were comparable. These observations clearly showed that Apaf1 in T cells was involved in the pathogenesis of DTH responses. To address the cause of the exacerbated DTH response in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice, we performed OVA-specific recall response by T cells ex vivo. As shown in, an enhanced proliferative response was observed in Lck-Cre-Apaf1<sup>f/f</sup>-OTII T cells from the draining LNs, accompanied with significantly higher levels of IL-2 production over control Apaf1-suffucient T cells. DTH response is largely dependent on IFN-γ- producing Th1 cells and, to a lesser degree, on IL-17-producing Th17 cells. As expected, production of both IFN-γ and IL-17A was significantly augmented by Lck-Cre-Apaf1<sup>f/f</sup>-OTII LN cells as compared with control Apaf1<sup>f/f</sup>-OTII LN cells. The levels of IFN-γ and IL-17 indicated that Th1 response was obviously dominant in this experimental system as compared with Th17 response. Higher numbers of infiltrating neutrophils at the footpad lesion in Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice than control mice, however, implicated involvement of IL-17 in this system. These observations collectively demonstrated that Apaf1 in CD4<sup>+</sup> T cells critically participated in the attenuation of DTH response. ## Enhanced activation of Apaf1-deficient T cells We next examined the expression of early activation marker CD69, and CD44/CD62L as activation/effector markers on T cells from OVA-immunized Apaf1<sup>f/f</sup>-OTII and Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice. The percentages of CD69<sup>+</sup> OTII cells and those of CD44<sup>high</sup>CD62L<sup>low</sup> OTII cells as cells with activated (or memory) phenotype in Lck-Cre-Apaf1f/f-OTII mice were comparable to those in Apaf1<sup>f/f</sup>-OTII mice. Similarly, the total numbers of CD69<sup>+</sup> cells and CD44<sup>high</sup>CD62L<sup>low</sup> cells were comparable between the two groups. We then evaluated these cell surface markers following the re-stimulation of T cells in vitro. As expected, the expression of activated marker CD69 rapidly up-regulated upon re-stimulation on both Lck-Cre-Apaf1<sup>f/f</sup>-OTII and control cells in an antigen dose- dependent manner (left). Expression of CD69 was higher on Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells than on control cells (left and Panel A). The percentages of CD44<sup>high</sup>CD62L<sup>low</sup> cells also increased after in vitro re-stimulation and were higher in Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells than in control Apaf1-sufficient cells (right and Panel B). We further addressed the impact of Apaf1-deficiency in primarily activated T cells. When CD4<sup>+</sup> T cells from unimmunized mice were stimulated in vitro with either OVA peptide or anti-CD3ε antibody, Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells showed higher or more efficient cell division than control cells, as demonstrated by higher percentages of cells with lower CFSE intensity. Percentages of divided cells with lower CFSE intensity were significantly higher in Apaf1-deficient Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells than in Apaf1-sufficient control cells. Production of IFN-γ in the supernatants was also significantly higher by Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells than by control Apaf1<sup>f/f</sup>-OTII cells. These data shown in Figs and clearly demonstrated that Apaf1 in T cells plays a regulatory role during an initial encounter and activation of naive T cells as well as during recall activation of primed T cells with antigen. ## Caspase-independent enhancement of activation in Apaf1-deficient T cells The negative regulation of T cell activation by Apaf1 may or may not depend on apoptosis. T cells activated in an inappropriate condition may be deleted by Apaf1-dependent apoptosis. Apaf1 may, however, regulate T cell activation and proliferation independently of caspase activation/apoptosis. Unexpectedly, pan- caspase inhibitor, z-VAD-fmk, at 10 or 100 μM did not rescue the lower proliferation (thymidine incorporation) or lower viability of control Apaf1-sufficient cells over Apaf1-deficient cells from immunized mice after antigenic stimulation (, upper panels). Similarly, neither lower production of IFN-γ nor IL-17 production by Apaf1-sufficient cells were restored to the levels by Apaf1-deficient cells with z-VAD-fmk (middle panels). Additionally, percentages of CD69<sup>+</sup> and CD44<sup>high</sup>CD62L<sup>low</sup> cells in control Apaf1-sufficient OTII T cell population were still lower over Apaf1-deficient OTII T cells in the presence of z-VAD-fmk (lower panels). Dexamethasone-induced apoptosis and caspase 3 activation in thymocytes was completely suppressed by z-VAD-fmk at the same concentration (100 μM). To address whether or not caspase activation occurs during T cell activation, either in Apaf1-dependent or -independent way, western blot analysis was performed to detect caspase activation during antigen-induced T cell activation. Stimulation of LN T cells with anti-CD3ε antibody or OVA induced caspase 3 cleavage in Apaf1<sup>f/f</sup>-OTII cells. Perhaps counterintuitively, but in line with the data show above, caspase 3 cleavage was similarly detected in Lck-Cre-Apaf1<sup>f/f</sup>-OTII cells. The caspase inhibitor, z-VAD-fmk, did not inhibit caspase 3 cleavage in either Apaf1-sufficient or Apaf1-deficient cells, although the appearance of 17 kDa fragment was only partially inhibited by z-VAD-fmk at 100 μM (, cleaved Casp3). Cleaved form of caspase 7 was also detected in both Apaf1-sufficient T cells and Apaf1-deficient T cells almost similarly (cleaved Casp7) and z-VAD-fmk at 100 μM showed no inhibitory effects. Caspase 9 activation (cleavage), which should be shown by reduction of pro-caspase 9, was barely detectable in both Apaf1-sufficient or -deficient OTII cells; note that cleaved form of caspase 9 was not detectable with antibody used. The western blot analysis clearly demonstrated that antigen-specific activation of T cells induced caspase 3 and 7 cleavage, which was not mediated by Apaf1 (and presumably not by caspase 9) and also was not inhibited by z-VAD-fmk. As the caspase cleavage occurred similarly in both Apaf1-sufficient and Apaf1-deficient T cells, caspase activation did not account for the difference in proliferation and cytokine production between Apaf1-sufficient and Apaf1-deficient T cells. These data indicated that the negative impact of Apaf1 on antigen-stimulated T cells was largely independent of caspase activities, although the possible involvement of caspase/apoptosis was not formally excluded. Taken together, our data clearly demonstrated that Apaf1 is a negative regulator of T cell responses by attenuating proliferation, activation, and cytokine production. This negative regulatory role of Apaf1 plays a role during antigen- specific T cell response but not in the maintenance of T cell homeostasis and the role is caspase-independent. # Discussion Previous reports including ours demonstrated the role of Apaf1 in the functional shaping of some organs and adjusting the number of cells during development\[, –\]. While apoptosis is required for removal of unwanted or possibly hazardous immune cells, the negative selection of self-reactive clones in the thymus was not impaired in Apaf1-deficient mice. In addition, while almost all Apaf1-deficient mice were perinatally lethal due to the brain deformity, a few survivors observed showed no signs of lymph adenopathy (H. Y., unpublished observation), implying no role of Apaf1 in the T cell homeostasis in the periphery. In the current study, we addressed the role of Apaf1 in regulation of antigen-specific T cell responses. Perinatal lethality of Apaf1-deficient mice prevented us from examining the role of Apaf1 in the T cell activation. To circumvent this problem, we generated conditional Apaf1-deficient mice, in which Lck-driven expression of the Cre recombinase disrupted Apaf1 gene in a T cell- specific manner. Data shown in confirmed the requirement of Apaf1 for the intrinsic pathway of apoptosis, induced by mitochondrial damages including growth factor deprivation, but not for the extrinsic pathway of apoptosis, induced by, for instance, anti-CD3ε re-stimulation or Fas ligation (data not shown), in the thymocytes and peripheral T cells. To examine the role of Apaf1 in T cells during antigen-induced immune responses, we exploited OVA-specific delayed-type hypersensitivity assay using OTII mice expressing OVA-specific T cell receptors on T cells. Lck-Cre-Apaf1<sup>f/f</sup>-OTII mice showed remarkable exacerbation of DTH responses including footpad swelling, cell infiltration, and inflammatory cytokine production, clearly demonstrating the role of Apaf1 as a negative regulator of T cell responses. To our knowledge, this is the first to show the role of Apaf1 as a negative regulator of immune responses. Apoptosis occurs in mature T lymphocytes in an antigen-driven way or growth factor-deprived way and is required for homeostasis, tolerance, and proper regulation of immune responses. There are a couple of different (but not mutually exclusive) possibilities where Apaf1 plays its negative role during T cell responses. One is that Apaf1-deficiency leads to prolonged survival of T cells after activation, due to the impairment of Apaf1-dependent apoptosis, resulting in the augmentation of DTH responses. Although this possibility appeared plausible, we were unable to detect any difference in the percentages of apoptotic cell death in the infiltrating cells into the footpad lesion between Lck-Cre-Apaf1<sup>f/f</sup>-OTII and Apaf1<sup>f/f</sup>-OTII mice (data not shown). In addition, time course of the footpad swelling between the two groups of mice were almost comparable (data not shown). It was thus unlikely that Apaf1 attenuated T cells responses by controlling apoptosis, although the possibility was not formally excluded. Our ex vivo and in vitro analyses of antigen-specific cell activation supported another possibility that Apaf1 regulates T cell activation, independently of cell survival. Apaf1-deficient OTII T cells upon stimulation proliferated more efficiently and showed higher percentages of cells with activation phenotypes (Figs and). Although these differences in activation status could be explained by the impairment of Apaf1-dependent apoptosis in inappropriately activated T cells, addition of a pan-caspase inhibitor, z-VAD-fmk, did not affect the difference in activation and cell division of T cells in vitro. It was thus indicated that the effect of Apaf1 in regulation of cell activation and cell division was caspase-independent. As reported previously, caspase 3 was cleaved during T cell activation, which was not inhibited by z-VAD-fmk. We also detected caspase 3 and 7 cleavage during T cell activation, on which z-VAD-fmk showed little, if any, inhibitory effects. The cleavage of caspases 3 and 7 occurred similarly in both Apaf1-sufficient and Apaf1-deficient T cells. As caspase 9 activation was barely detectable in these cells, the cleavage of caspases 3 and 7 appeared independent of apoptosome activation. Although the precise mechanism for this caspase activation is not known yet, it is clear that the caspase activation was not the reason for the difference in T cell activation between Apaf1-sufficient and -deficient T cells. The roles of Apaf1 reside mainly but are not limited in the initiation of apoptosis. Zermati et al. reported the non-apoptotic role of Apaf1 in that Apaf1-deficiency resulted in the impairment of DNA damage-induced cell cycle arrest. This non-apoptotic function of Apaf1 requires nuclear translocation of Apaf1 and depends of checkpoint kinase-1 (Chk1) activity. Similarly, chemical inhibitor of Apaf1 suppressed Apaf1-dependent DNA-damage checkpoint. These reports supported anti-proliferation and/or anti-survival function of Apaf1. Oppositely, Apaf1 plays a pro-proliferation/pro-survival role by regulating centrosome stability. Ferraro et al. reported that Apaf1 was required for efficient centrosome and mitotic spindle formation. The non-apoptotic roles of Apaf1 are thus complicated and perhaps context-dependent. Although the precise mechanisms for these non-apoptotic functions of Apaf1 are poorly understood, Apaf1 exerts these functions through interaction with various molecules\[, –\]. Presumably, Apaf1 plays its anti- or pro-activation/survival function though interaction with the interacting molecule(s) depending upon cell activation status. In our hands also, Apaf1-deficiency in some situations resulted in enhanced cell death (e.g., Fas stimulation). This may reflect the context-dependent (non-apoptotic) roles of Apaf1. We were unable to detect any difference in cell cycle progression or cell activation status in the lesion-infiltrating T cells between Apaf1-deficient and Apaf1-sufficient cells in DTH assay (data not shown). It is possible that the difference in individual cells was subtle and/or only a small fraction of antigen-specifically activated cells showed difference. Alternatively, it is possible that Apaf1 exerts its negative regulatory effect at the priming phase where naive T cells are primarily stimulated in the draining lymph nodes, before the effector phase. Higher percentages of cells in activation status in Apaf1-deficient T cells at the primary stimulation phase in vitro may support this idea. In any case, the impact of Apaf1-deficiency on T cell function appeared limited. We have been unable to detect significant differences in pathological phenotypes in other disease models, including experimental allergic encephalomyelitis (EAE) and LPS-induced topical inflammation (data not shown). In DTH model in the current study, the differences were small albeit their significance and reproducibility. In our hands, we reported that the DTH response was exacerbated in the absence of EBI-3 of IL-27. The phenotypic difference was far worse in the EBI-3-deficient mice than in T cell-specific Apaf1-deficient mice (compared data not shown). The role of Apaf1 and Apaf1-mediated apoptosis in T cell function appears much limited in terms of the intensity and/or duration, as compared with, for instance, cytokines or other molecules directly involved in T cell activation/function. Further clarification of the exact point(s) where Apaf1 exerts it regulatory effect during T cell activation is a future challenge. Collectively, our data clearly demonstrated Apaf1 as a negative regulator of T cell responses. Further analysis of the underlying mechanisms may reveal Apaf1 as a potential target for immunosuppressive drug discovery. # Supporting information The authors thank Dr. Iizasa and Ms. Yoshida for technical help and animal husbandry, Dr. Kawaguchi for advices on statistics, and Ms. Furukawa for secretarial supports. [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Department of Immunology, Kagoshima University Graduate School of Medical and Dental Sciences, Kagoshima, Japan

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.