🧪 Molecular Toxicity Prediction with Graph Neural Networks (GNNs)

Predicting molecular toxicity is crucial for drug discovery, chemical safety, and environmental research. This project explores both classical ML models and Graph Neural Networks (GNNs) on molecular graph data, comparing their performance on toxicity prediction.

🚀 Features

Dataset: Tox21 – benchmark dataset for molecular toxicity prediction

Models Implemented:

Random Forest – robust baseline

XGBoost – gradient boosting model

Graph Isomorphism Network (GIN) – deep learning model for graph data

Evaluation Metrics: ROC-AUC for classification tasks

📊 Model Performance

Model	ROC-AUC	Notes
Random Forest	0.84	Best baseline
XGBoost	0.79	Strong classical model
GIN (GNN)	0.819	Per-task ROC-AUC: 0.70 – 0.87

Key Insights:

Classical ML (Random Forest) still outperforms deep learning for this dataset.

GNNs are competitive and scale well for graph-structured data.

Visualizations (training curves, boxplots) reveal per-task variability and model stability.

🛠 Tech Stack

Python 🐍

PyTorch Geometric (GNNs)

Scikit-learn (Random Forest, XGBoost)

Pandas & NumPy

Matplotlib

📈 Visuals

Per-task ROC-AUC comparison

GNN Training Curve

Boxplot of per-task AUCs

🔮 Future Work

Explore other GNNs like GCN, GraphSAGE, GAT for better performance.

Apply hyperparameter tuning and model ensembling.

Add explainability to understand key molecular features.

Expand datasets and try data augmentation for improved generalization.

Build a web app/API for real-time toxicity prediction.

📝 Author

Priyanka