ToxicityAI: Research-Grade Drug Toxicity Prediction (Hackathon-Ready)

This project is a hybrid, explainable, end-to-end toxicity prediction system using:

• Classical chemical descriptors + Morgan fingerprints
• Graph neural networks (GNN) on molecular graphs
• Ensemble learning for performance and robustness
• Explainability with SHAP + substructure highlighting
• Streamlit demo UI for real-time prediction and insight

Quick Start

1. Create and activate a Python environment (3.10+).
2. Install dependencies:

pip install -r requirements.txt

3. Prepare data:

   • Place Tox21 CSV in data/tox21.csv with at least: smiles and all 12 endpoint labels: NR-AR, NR-AR-LBD, NR-AhR, NR-Aromatase, NR-ER, NR-ER-LBD, NR-PPAR-gamma, SR-ARE, SR-ATAD5, SR-HSE, SR-MMP, SR-p53.

4. Train:

python train.py --config configs/config.yaml

5. Run the app:

streamlit run app/streamlit_app.py

Project Structure

• data/ raw datasets
• features/ feature generation
• models/ baseline models, GNN, ensemble
• explainability/ SHAP + GNN explanations
• app/ Streamlit demo
• utils/ chemistry utilities, logging
• configs/ experiment config
• outputs/ trained artifacts and metrics

Scientific Rationale (Short)

• Descriptors capture global physicochemical properties (MW, logP, TPSA, HBD/HBA) strongly linked to membrane permeability, bioavailability, and off-target toxicity.
• Morgan fingerprints encode substructures that can reveal toxicophores.
• Graph neural networks model local atomic environments and bond topology for mechanistic signal.
• Ensembles reduce variance and improve generalization across chemical space.

Extra Scripts

• Reliability diagrams:

python calibration.py --config configs/config.yaml

• Ablation study:

python ablation.py --config configs/config.yaml

Notes

• GNN training uses PyTorch + PyTorch Geometric. If not installed, the pipeline still runs baseline + ensemble.
• All explainability modules are designed to be lightweight and hackathon-demo friendly.