Skip to content
This repository was archived by the owner on Mar 22, 2026. It is now read-only.

README.md

Latest commit

 

History

History
95 lines (63 loc) · 2.63 KB

README.md

File metadata and controls

95 lines (63 loc) · 2.63 KB

💊 Darwin PBPK Platform

DOI

"Ciência rigorosa. Resultados honestos. Impacto real."

AI-Powered PBPK Prediction Platform

State-of-the-art deep learning platform for physiologically-based pharmacokinetic (PBPK) parameter prediction using multi-modal molecular representations.

🚀 Features

  • ✅ Multi-Modal Embeddings: ChemBERTa 768d + Molecular graphs + RDKit descriptors
  • ✅ Advanced Architectures: GNN (GAT + TransformerConv), Multi-task learning
  • ✅ Large Dataset: 44,779 compounds (ChEMBL + TDC + KEC)
  • ✅ 3 PBPK Parameters: Fraction unbound (Fu), Volume of distribution (Vd), Clearance (CL)
  • ✅ PhysioQM Integration: Physics-informed constraints
  • ✅ Production Ready: Trained models, API endpoints

📊 Performance

  • Baseline MLP: R² > 0.30 (ChemBERTa only)
  • GNN Model: R² > 0.45 (Graphs + Attention)
  • Ensemble: R² > 0.55 (Multi-modal fusion)

🧬 Architecture

Embeddings:

  • ChemBERTa: seyonec/ChemBERTa-zinc-base-v1 (768d)
  • Pre-trained on ~100M molecules (ZINC, PubChem)

Graphs:

  • PyTorch Geometric
  • 20 node features (atom type, charge, aromaticity, etc.)
  • 7 edge features (bond type, conjugation, ring, etc.)

Descriptors:

  • RDKit: 25 molecular descriptors
  • MW, LogP, TPSA, QED, HBA, HBD, etc.

Models:

  • GAT: 4 attention heads, 3 layers
  • TransformerConv: 4 heads, 3 layers
  • Multi-task: Weighted loss (Fu, Vd, CL)

📚 Citation

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

Agourakis, D.C. (2025). Darwin PBPK Platform: AI-Powered Pharmacokinetic 
Prediction. Version 1.0.0 [Software]. Zenodo. 
https://doi.org/10.5281/zenodo.17536674

📖 Dataset

Large datasets (embeddings, graphs, ~1.7 GB) available at:

🚀 Quick Start

# Install dependencies
pip install -r requirements.txt

# Train baseline MLP
python apps/training/01_baseline_mlp.py

# Train GNN model
python apps/training/02_gnn_model.py

# Make predictions (after training)
python apps/prediction/pbpk_predictor.py --smiles "CCO"

📊 Data Sources

  • ChEMBL: Bioactivity and PK data
  • TDC (Therapeutics Data Commons): ADMET benchmark datasets
  • KEC: Curated literature extractions

📄 License

MIT License - See LICENSE file

🙏 Acknowledgments

Developed for computational drug discovery with Q1 scientific rigor at PUCRS.

"Rigorous science. Honest results. Real impact."