PSG-BAR

PSG-BAR predicts protein–ligand binding affinity using protein structure graphs and ligand graph representations to support computational drug discovery and binding-site identification.

Key Features:

• Integration of 3D structural information: Combines three-dimensional protein structures with two-dimensional graph representations of ligands to represent molecular complexes.
• Protein Structure Graphs (PSG): Represents proteins as structure-derived graphs to capture spatial and residue-level relationships.
• Residual Graph Attention Networks: Applies residual graph attention networks to assign attention scores that weight protein regions relevant for ligand binding.
• Attention-based pooling and residue identification: Uses attention-based pooling to highlight surface residues critical for protein–ligand interactions.
• Deep learning binding-affinity regression: Implements deep learning models for binding affinity prediction trained on extensive bioactivity data and evaluated on benchmark datasets.
• Avoids voxelization: Uses graph-based representations rather than voxelization or oversimplified protein encodings.

Scientific Applications:

• Drug discovery and development: Predicts ligand–target interactions to prioritize candidate compounds in computational screening workflows.
• Viral protease targeting: Provides predictive models validated on a viral main protease (Mpro) relevant to SARS-CoV-2 research.

Methodology:

Constructs Protein Structure Graphs from 3D protein structures and two-dimensional ligand graphs, applies residual graph attention networks with attention-based pooling within a deep learning binding-affinity regression framework trained on bioactivity data.