Reaction-Enzyme-Interaction-Prediction-Model

Reaction-Enzyme Interaction Prediction Model

A deep learning framework for predicting reaction center atoms in chemical reactions using multimodal learning that combines molecular graph representations with enzyme sequence embeddings.

Overview

This project implements a multimodal neural network that predicts which atoms in a reactant molecule will participate in a reaction (reaction center prediction) by using:

• Molecular Graphs: Graph transformer encoders for atom-level representations
• Enzyme Sequences: ESM-2 protein language model embeddings
• Cross-Attention Mechanism: Fuses molecular and enzyme information

Architecture

The model consists of three main components:

1. ESM-2 Encoder: Processes enzyme amino acid sequences using META FAIR's ESM-2 protein language model
2. Graph Transformer Encoder: Encodes molecular structures using graph attention layers
3. Cross-Attention Fusion: Combines enzyme and molecular representations to predict reaction centers

┌─────────────────┐     ┌─────────────────────┐
│   ESM-2 Model   │     │  Graph Transformer  │
│  (Protein Seq)  │     │   (Molecular Graph) │
└────────┬────────┘     └─────────┬───────────┘
         │                        │
         ▼                        ▼
┌─────────────────┐     ┌──────────────────┐
│  ESM Projection │     │  Atom Features   │
└────────┬────────┘     └─────────┬────────┘
         │                        │
         └──────────┬─────────────┘
                    │
                    ▼
         ┌─────────────────────┐
         │   Cross-Attention   │
         │      Mechanism      │
         └──────────┬──────────┘
                    │
                    ▼
         ┌─────────────────────┐
         │  Reaction Center    │
         │   Prediction Head   │
         └─────────────────────┘

Features

• Multimodal Learning: Combines protein sequence and molecular graph data
• Reaction Center Prediction: Identifies atoms involved in enzymatic reactions
• Pre-trained Models: Uses ESM-2 for protein embeddings
• Graph Transformers: Custom graph attention layers for molecular encoding
• Interactive GUI: Streamlit-based web interface for predictions

Data Sources

The model is trained on data from:

• Rhea: Expert-curated database of chemical reactions
• RetroRules: Reaction rules for retrosynthesis
• UniProt: Protein sequences and EC numbers

Training History

License

This project is licensed under Apache License 2.0 - Feel free to use and modify

Acknowledgments

• ESM-2 - Evolutionary Scale Modeling
• Rhea - Reaction database
• RetroRules - Reaction rules database
• RDKit - Cheminformatics toolkit
• Hugging Face Transformers - Model implementations
• PyTorch - Deep learning framework
• Kimi AI with K2.5 - Assistance in generating code.