Drug Toxicity Predictor 🧬

🚀 Features

• Dual AI Models:
  • ToxicityNet (Fingerprint-based): 94.5% accuracy
  • RichGCNModel (Graph Neural Network): 93.2% accuracy
• Batch Processing: Upload CSV files with thousands of molecules
• Smart Sampling: Optimized for large datasets (244K+ molecules)
• Uncertainty Quantification: Confidence intervals for predictions
• Compressed File Support: Handles .gz and regular CSV files
• Real-time Progress: ETA tracking for long-running jobs

🔬 Models

ToxicityNet (Fingerprint Model)

• Architecture: Dense neural network with dropout and batch normalization
• Input: 1024-dimensional molecular fingerprints
• Performance: 94.5% accuracy on Tox21 dataset

RichGCNModel (Graph Neural Network)

• Architecture: Graph Convolutional Network with attention mechanism
• Input: Molecular graphs with atom and bond features
• Performance: 93.2% accuracy with molecular graph representations

📊 Batch Processing Optimization

• Smart Sampling: Process representative subsets for faster results
• Progress Tracking: Real-time ETA and completion status
• Memory Management: Chunked processing to prevent timeouts
• Flexible Options: Balance between speed and completeness

🧪 Usage

1. Single Prediction: Enter a SMILES string for immediate analysis
2. Batch Processing: Upload CSV with 'smiles' column
3. Large Datasets: Use sampling options for datasets >10K molecules
4. Download Results: Get comprehensive CSV with uncertainty metrics

📈 Performance

• Processing Speed: ~100-200 molecules/minute
• Recommended Batch Size: 5,000 molecules (~5 minutes)
• Maximum Tested: 244,245 molecules with sampling
• Memory Efficient: Handles large files without crashes

🔧 Technical Details

• Framework: Streamlit + PyTorch
• Chemistry: RDKit for molecular processing
• Models: Pre-trained on Tox21 dataset
• Deployment: Optimized for Hugging Face Spaces

📋 Input Format

⚠️ Limitations

• Supports specific atom types: H, C, N, O, F, S, Cl, Br, I
• Complex salts are automatically simplified to largest component
• Large batch processing may require sampling for optimal performance

🏆 Citation