ADMET Predictor

Multi-task ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) prediction API for drug discovery using deep learning.

Features

• Multi-task Prediction: Predict 22 ADMET endpoints simultaneously
• Deep Learning Models: Combines graph neural networks (AttentiveFP) with ChemBERTa encoders
• REST API: FastAPI-based service for easy integration
• Batch Processing: Efficient handling of multiple compounds
• Explainability: Attribution methods for model interpretation
• Uncertainty Estimation: Evidential deep learning for reliable predictions

Installation

Using Poetry (Recommended)

# Install dependencies
poetry install

# Or with pip
pip install -r requirements.txt

Manual Installation

python3.11 -m venv .venv
source .venv/bin/activate
pip install torch torchvision torch-geometric rdkit transformers pytorch-lightning mlflow \
    scikit-learn pandas fastapi "uvicorn[standard]" pydantic "redis[asyncio]" celery \
    faiss-cpu captum httpx pytest pytest-asyncio PyTDC molvs pyyaml scipy

Quick Start

Running the API Server

# Using Poetry
poetry run uvicorn admet_predictor.api.main:app --host 0.0.0.0 --port 8000 --reload

# Or using the Makefile
make serve

The API will be available at http://localhost:8000

API Documentation

Once the server is running, visit:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

Making Predictions

curl -X POST "http://localhost:8000/api/v1/admet/predict" \
  -H "Content-Type: application/json" \
  -d '{
    "smiles": "CCO",
    "include_uncertainty": true
  }'

API Endpoints

Health Check

• GET /health - Check if the API is running and model is loaded

Prediction

• POST /api/v1/admet/predict - Predict ADMET properties for a single compound
• POST /api/v1/admet/batch - Predict ADMET properties for multiple compounds
• POST /api/v1/admet/explain - Get attribution explanations for predictions

Model Information

• GET /api/v1/admet/model/info - Get metadata about the loaded model

Training

Download Data

make download-data

Preprocess Data

make preprocess

Train Model

make train

Model Architecture

The ADMET Predictor uses a fusion architecture:

• Graph Encoder: AttentiveFP for molecular graph representation
• Sequence Encoder: ChemBERTa for SMILES-based representation
• Task Heads: Separate prediction heads for each ADMET task
• Uncertainty: Evidential deep learning for uncertainty quantification

ADMET Tasks

The model predicts the following ADMET endpoints:

• Absorption: Caco2, HIA, Pgp
• Distribution: BBB, PPBR, VDss
• Metabolism: CYP2C9, CYP2D6, CYP3A4 substrates/inhibition
• Excretion: Clearance (hepatocyte, microsome), Half-life
• Toxicity: AMES, hERG, DILI, LD50

Development

Running Tests

# Using Poetry
poetry run pytest tests/ -v

# Or using the Makefile
make test

Code Quality

# Format code
poetry run black src/ tests/

# Lint code
poetry run ruff check src/ tests/

# Type checking
poetry run mypy src/

Configuration

Model and data configurations are stored in the configs/ directory:

• configs/data/admet_tasks.yaml - Task definitions and data sources
• configs/model/attentivefp_base.yaml - Model architecture parameters

Environment Variables

• ADMET_CHECKPOINT - Path to the trained model checkpoint
• ADMET_DATA_CONFIG - Path to data configuration file
• ADMET_MODEL_CONFIG - Path to model configuration file

License

This project is licensed under the MIT License.

Citation

If you use this code in your research, please cite:

@software{admet_predictor,
  title = {ADMET Predictor},
  author = {ADMET Team},
  year = {2026},
  url = {https://github.com/AmirhosseinOlyaei/admet-predictor}
}

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgments

• PyTDC for ADMET datasets
• RDKit for cheminformatics
• PyTorch Geometric for graph neural networks