Edge-aware GAT for Protein Binding Site Prediction

📘 Overview

Accurate identification of protein binding sites is crucial for understanding biomolecular interaction mechanisms and for the rational design of drug targets. We propose an Edge-aware Graph Attention Network (Edge-aware GAT) model for fine-grained prediction of binding sites between proteins and various biomolecules, including proteins, DNA/RNA, ions, ligands, and lipids.

🎯 Key Features

• Multi-class Prediction: Simultaneously predicts 5 types of binding sites (Protein, DNA/RNA, Ion, Ligand, Lipid)
• Edge-aware Attention: Incorporates interatomic distances and directional vectors into attention mechanism
• Atom-level Graph: Constructs detailed atomic representations with DSSP-derived features and RSA
• High Accuracy: Achieves ROC-AUC of 0.93 for protein-protein binding site prediction
• Interactive 3D Visualization: Real-time molecular structure visualization with probability-based coloring

🏆 Performance

Our model outperforms several state-of-the-art methods on benchmark datasets:

Method	ROC-AUC
Edge-aware GAT (Ours)	0.93
PeSTo	0.91
ScanNet	0.87
MaSIF-site	0.80
Sppider	0.73
PSIVER	0.64

🔬 Methodology

1. Feature Embedding Module

Integrates 88-dimensional features including element type, residue category, atom type, secondary structure (DSSP), and relative solvent accessibility (RSA) into a 32-dimensional latent space.

2. Edge-aware GAT Conv

Employs 32 GATConv layers with edge features (interatomic distances and directional vectors) to capture spatially directional dependencies.

3. Residue-level Pooling & Decoder

Attention-based pooling aggregates atom-level features to residue level, followed by MLP decoder for multi-class prediction.

📊 Dataset

• Source: Protein Data Bank (PDB)
• Clustering: 30% maximum sequence identity
• Split: 70% training, 15% validation, 15% test
• Benchmark: MaSIF-site (53 chains), PPI Benchmarks (230 structures), ScanNet dataset

🌐 Web Interface

This interactive web application provides:

• ✅ PDB file upload with example dataset (PDB: 1a01)
• ✅ Adjustable confidence threshold (0.0-1.0)
• ✅ Real-time prediction (2-5 seconds for typical proteins)
• ✅ Tab-based multi-class results visualization
• ✅ 3D structure viewer with color gradient (Green → Yellow → Red)
• ✅ Downloadable PDB files with B-factor encoded probabilities