LigBind

LigBind predicts ligand-specific protein binding residues using a relation-aware graph neural network framework to enable residue-level binding prediction across 1159 ligands.

Key Features:

• Relation-Aware Framework: Models shared binding preferences among ligands using relation-aware classifiers to leverage similarities across ligands.
• Graph Neural Network-Based Feature Extraction: Uses a graph neural network to extract features from ligand-residue pairs and capture interaction patterns.
• Pre-Training and Fine-Tuning: Performs graph-level pre-training across ligands followed by ligand-specific fine-tuning using a domain adaptive neural network.
• Benchmark Datasets: Validated on datasets comprising 1159 ligands and an additional set of 16 unseen ligands.
• Application in Viral Proteins: Identifies binding residues in SARS-CoV-2 main protease, papain-like protease, and RNA-dependent RNA polymerase.

Scientific Applications:

• Drug target identification: Predicts residue-level ligand binding to facilitate identification of potential drug targets.
• Antiviral drug development: Maps binding residues in SARS-CoV-2 proteins to inform antiviral compound design.
• Generalization to sparse-data ligands: Enables prediction for ligands with limited known binding proteins by leveraging shared ligand relations learned in pre-training.

Methodology:

A GNN-based feature extractor is pre-trained on ligand-residue pairs with relation-aware classifiers developed for similar ligands, and the pre-trained model is fine-tuned per ligand using a domain adaptive neural network.