Protein-protein and protein-nucleic acid binding site prediction via interpretable hierarchical geometric deep...

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning in protein science

Background:

• Protein-protein and protein-nucleic acid interactions are crucial for biological functions, disease mechanisms, and drug development.
• Accurate prediction of these binding sites is computationally challenging due to insufficient understanding of residue binding patterns.
• Existing methods struggle to integrate spatial residue distribution with physicochemical interactions.

Purpose of the Study:

• To develop a novel computational model, GraphRBF, for enhanced prediction of protein binding sites.
• To improve the characterization of residue binding patterns by incorporating spatial and physicochemical information.
• To provide a tool for advancing disease diagnosis and drug design.

Main Methods:

• Designed GraphRBF, a hierarchical geometric deep learning model.
• Utilized an enhanced graph neural network to model physicochemical interactions between residues.
• Employed a prioritized radial basis function neural network to capture spatial distributions of neighboring residues.

Main Results:

• GraphRBF demonstrated significant performance improvements over state-of-the-art methods.
• The model exhibited strong interpretability in its learned representations.
• Applied to SARS-CoV-2 omicron spike protein, GraphRBF successfully identified known epitopes and predicted novel binding regions for therapeutic development.

Conclusions:

• GraphRBF offers a powerful new approach for identifying protein binding sites.
• The model's ability to integrate spatial and physicochemical data enhances prediction accuracy.
• GraphRBF provides a valuable resource for biological research, diagnostics, and drug discovery, with an accessible online server.