ReACT-Drug: Reaction-Template Guided Reinforcement Learning for de novo Drug Design

A complete reinforcement learning pipeline for de novo molecular drug design using reaction-template guided molecular transformations.

Overview

ReACT-Drug is an end-to-end RL framework that combines deep representation learning with multi-objective molecular optimization. Given a target protein, the system:

1. Identifies similar proteins using ESM-2 embeddings from the PDBbind database
2. Fragments known ligands using BRICS/RECAP to create starting molecules
3. Grows molecules using ChEMBL-derived reaction templates
4. Optimizes for binding affinity (AutoDock Vina), drug-likeness (QED), synthetic accessibility, and novelty

Features

• 🧬 ESM-2 protein embeddings for similarity search
• 💊 ChemBERTa molecular embeddings for state representation
• 🧪 ChEMBL-derived reaction templates for chemically valid transformations
• 🔬 AutoDock Vina integration for binding affinity scoring
• 🤖 PPO with dynamic action spaces for multi-objective optimization
• 🧩 Fragment-based starting state generation

Installation

# Clone repository
git clone https://github.com/YadunandanRaman/ReACT-Drug.git
cd ReACT-Drug

# Create conda environment
conda create -n react-drug python=3.10
conda activate react-drug

# Install dependencies
pip install -r requirements.txt

# Install AutoDock Vina (for docking)
conda install -c conda-forge autodock-vina

# Install Open Babel
conda install -c conda-forge openbabel

Directory Structure

ReACT-Drug/
├── data/                 # Create this directory (see Data Preparation)
│   ├── raw/              # PDBbind tar.gz and PDB files
│   ├── processed/        # Processed binding data (auto-generated)
│   └── templates/        # Place drug_templates.pkl here
├── src/
│   ├── __init__.py
│   ├── config.py         # Hyperparameters and paths
│   ├── encoders.py       # ESM-2 and ChemBERTa models
│   ├── chemistry.py      # Fragmentation, Templates, Rewards
│   ├── docking.py        # AutoDock Vina wrapper
│   ├── environment.py    # Gym-like RL Environment
│   ├── agent.py          # PPO Actor-Critic Agent
│   └── utils.py          # Data loading helpers
├── scripts/
│   └── generate_templates.py  # ChEMBL MMP extraction
├── models/               # Saved model checkpoints
├── results/              # Discovery results
├── main.py               # Entry point
├── requirements.txt
└── README.md

Data Preparation

Note: The data/ directory is not included in this repository due to file size and licensing restrictions. You must create it and download the required files manually.

Step 1: Create Data Directories

mkdir -p data/raw data/processed data/templates

Step 2: Download PDBbind Dataset

The PDBbind dataset is required for training. Due to licensing requirements, please download directly from the official source.

For Academic Use:

1. Visit: https://www.pdbbind-plus.org.cn/download
2. Register for a free academic account
3. Download: PDBbind_v2020_refined.tar.gz
4. Place in data/raw/

For Commercial Use:

1. Visit: https://www.pdbbind-plus.org.cn/download
2. Contact PDBbind for commercial licensing information

Step 3: Download Reaction Templates

Download the pre-computed ChEMBL-derived reaction templates:

📥 Download drug_templates.pkl

Place the downloaded file in data/templates/

Alternative: Generate templates yourself using ChEMBL data:

python scripts/generate_templates.py

Verify Setup

After completing the above steps, your data directory should look like:

data/
├── raw/
│   └── PDBbind_v2020_refined.tar.gz
├── processed/
│   └── (empty - will be auto-generated on first run)
└── templates/
    └── drug_templates.pkl

Usage

Basic Usage

python main.py --pdb_id 4nc3 --episodes 200

Advanced Options

python main.py \
    --pdb_id 6lu7 \
    --binding_site "10.5,20.3,30.1" \
    --episodes 500 \
    --max_structures 10000 \
    --seed 42

Using Local PDB File

python main.py \
    --pdb_file path/to/protein.pdb \
    --binding_site "x,y,z" \
    --episodes 300

Configuration

All hyperparameters can be modified in src/config.py:

CONFIG = {
    "discovery": {
        "episodes": 500,
        "max_steps_per_episode": 15,
        "success_criteria": {
            "binding_affinity": -8.5,  # kcal/mol
            "qed_score": 0.5,
            "sa_score": 4.0,
        },
    },
    "training": {
        "learning_rate": 1e-4,
        "batch_size": 32,
        "ppo_epochs": 10,
        "clip_ratio": 0.2,
    },
    # ... more options
}