luxuan-lulu/Bioactivity_Final_project

Drug-Likeness Prediction Dataset (Based on DBPP-Predictor Data)

This dataset was created as part of a final project on drug-likeness prediction, based on the data from the DBPP-Predictor paper:

Gu, Y., Wang, Y., Zhu, K. et al. DBPP-Predictor: a novel strategy for prediction of chemical drug-likeness based on property profiles. J Cheminform 16, 4 (2024). https://doi.org/10.1186/s13321-024-00800-9

It includes curated molecular data, preprocessed RDKit descriptors, and training/test splits suitable for training classification models to distinguish drug-like from non-drug-like molecules.

Project Background

Drug-likeness refers to the potential of a small molecule to become a drug. Traditional rule-based approaches (e.g., Lipinski's Rule of Five) often fail to generalize across complex compounds. This project uses RDKit descriptors and AutoML (H2O) to construct a highly interpretable, generalizable classification model.

Dataset Description

The dataset is derived from the DBPP GitHub repository (https://github.com/yxgu2353/DBPP-Predictor.git). It includes:

• 5,147 drug-like molecules (FDA and globally approved drugs)
• 10,000 non-drug-like molecules sampled from ZINC

All molecules were standardized using MolVS(clean_data.py), the datasets were split by sklearn(split_dataset.py), and RDKit descriptors (216 features) were computed (Rdkit_descriptor.py).

Files:

• 'train_rdkit_descriptors.parquet'
• 'test_rdkit_descriptors.parquet'

Each file contains:

• 'label': 0 for non-drug, 1 for drug
• 'Standardized_SMILES'
• 216 numerical RDKit descriptors

Model Development

All models were trained using H2O AutoML with 10-fold cross-validation(model_constrcution.py). The top 3 models were also evaluated using an independent test set, and SHAP analysis was used to interpret the top structural features contributing to drug-likeness(model_analysis.py).