T5Chem

A Unified Deep Learning Model for Multi-task Reaction Predictions with MCP (Model Context Protocol) support.

Features

• Retrosynthesis Prediction: Predict reactants from a product molecule

• Product Prediction: Predict products from reactants and reagents

• Reagents Prediction: Predict required reagents for a reaction

• Molecule Validation: Validate SMILES strings

• Molecular Properties: Calculate detailed molecular properties

• MCP Server: Integrate with AI assistants through Model Context Protocol

Related MCP server: MIST - Model Intelligence System for Tasks

Installation

# Clone the repository
git clone https://github.com/bugatti742/t5chem.git
cd t5chem

# Install with MCP support
pip install -e ".[mcp]"

# Or install all dependencies
pip install -e .

Download Pre-trained Model

Large model files are NOT included in the repository. Download them separately:

# Download USPTO multi-task model
wget https://yzhang.hpc.nyu.edu/T5Chem/models/USPTO_MT_model.tar.bz2
tar -xjvf USPTO_MT_model.tar.bz2 -C model/

Usage

As MCP Server

Start the MCP server:

# Using default model path (model/)
t5chem-mcp

# Specify custom model path
t5chem-mcp --model_dir /path/to/your/model

Available MCP Tools

1. predict_retrosynthesis: Predict retrosynthesis routes

2. predict_product: Predict product from reactants

3. predict_reagents: Predict reagents for a reaction

4. validate_molecule: Validate SMILES strings

5. get_molecule_properties: Get molecular properties

Command Line

# Batch prediction
t5chem predict --data_dir data/sample/reactants/ --model_dir model/

# Training
t5chem train --data_dir data/sample/reactants/ --output_dir model/ --task_type reactants

Requirements

• Python 3.10+

• PyTorch 2.2+

• Transformers 4.38+

• RDKit 2022.9+

• MCP SDK 1.0+

Citation

Jieyu Lu and Yingkai Zhang, J Chem Inf Model, 62, 1376 - 1387 (2022)

License

MIT License