ADMET Property Prediction System

A deep learning system that predicts Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties from molecular structures. The system combines Graph Neural Networks with a RAG-based LLM explainer to provide both predictions and mechanistic interpretations.

Project Overview

Component	Description
Domain	Preclinical Drug Development
Core Task	Predict ADMET properties from SMILES molecular representations
Model Architecture	Graph Neural Network (GNN) with message passing
Explainability	RAG pipeline with pharmacokinetic knowledge base
Output	Property predictions with evidence-based explanations

ADMET Properties Predicted

Property	Category	Description
Caco-2 Permeability	Absorption	Intestinal absorption potential
HIA	Absorption	Human Intestinal Absorption
BBB Penetration	Distribution	Blood-Brain Barrier crossing
Plasma Protein Binding	Distribution	Drug availability in bloodstream
CYP450 Inhibition	Metabolism	Drug-drug interaction risk
Half-life	Excretion	Duration in body
hERG Inhibition	Toxicity	Cardiac toxicity risk
AMES Toxicity	Toxicity	Mutagenicity

Project Structure

admet-prediction-system/
├── config/
│   └── config.yaml              # Model and pipeline configuration
├── data/
│   ├── raw/                     # Original TDC datasets
│   ├── processed/               # Featurized molecular graphs
│   └── knowledge_base/          # RAG documents
├── src/
│   ├── data/
│   │   ├── data_loader.py       # TDC data fetching
│   │   └── molecular_featurizer.py  # SMILES to graph conversion
│   ├── models/
│   │   ├── gnn_model.py         # Graph Neural Network architecture
│   │   └── trainer.py           # Training loop with MLflow
│   ├── rag/
│   │   ├── document_processor.py    # PDF and text ingestion
│   │   ├── vector_store.py      # ChromaDB embeddings
│   │   └── llm_explainer.py     # Explanation generation
│   └── utils/
│       └── helpers.py           # Utility functions
├── scripts/
│   ├── train_model.py           # Model training entry point
│   ├── build_knowledge_base.py  # RAG index builder
│   └── predict.py               # Inference script
├── notebooks/
│   └── 01_exploratory_analysis.ipynb
├── app/
│   └── main.py                  # Streamlit application
├── requirements.txt
└── README.md

Installation

git clone https://github.com/yourusername/admet-prediction-system.git
cd admet-prediction-system

python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

pip install -r requirements.txt

Configuration

Edit config/config.yaml to set:

data:
  dataset_name: "Caco2_Wang"
  test_split: 0.2

model:
  hidden_dim: 128
  num_layers: 3
  dropout: 0.2
  learning_rate: 0.001
  epochs: 100

rag:
  embedding_model: "all-MiniLM-L6-v2"
  llm_model: "gpt-4o-mini"
  top_k: 5

Usage

Step 1: Download Data

python scripts/train_model.py --download-only

Step 2: Build Knowledge Base

python scripts/build_knowledge_base.py --docs-path data/knowledge_base/

Step 3: Train Model

python scripts/train_model.py --config config/config.yaml

Step 4: Run Predictions

python scripts/predict.py --smiles "CC(=O)OC1=CC=CC=C1C(=O)O" --explain

Step 5: Launch Application

streamlit run app/main.py

Data Source

This project uses the Therapeutics Data Commons (TDC) benchmark datasets:

Huang, K., et al. "Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development." NeurIPS 2021.

Available at: https://tdcommons.ai/

Model Architecture

The GNN follows a message passing paradigm:

1. Atom features are initialized from molecular properties (atomic number, charge, hybridization)
2. Bond features encode connectivity (bond type, aromaticity)
3. Message passing layers aggregate neighbor information
4. Global pooling produces a molecular fingerprint
5. MLP head outputs property predictions

RAG Knowledge Sources

The explanation system retrieves from:

1. FDA Drug Labels (DailyMed)
2. PubChem Compound Summaries
3. Pharmacokinetic Textbook Excerpts
4. ADMET Review Articles

MLflow Tracking

Experiments are logged to MLflow:

mlflow ui --port 5000

Tracked metrics include:

• Training and validation loss
• AUROC for classification tasks
• RMSE for regression tasks
• Model hyperparameters
• Feature importance

Environment Variables

Create a .env file:

OPENAI_API_KEY=your_key_here

License

MIT License