EGCPPIS: learning hierarchical equivariant graph representations with contrastive integration for protein-protein...

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions and disease mechanisms.
• Computational prediction of PPI sites offers an alternative to experimental methods.
• Existing methods often overlook protein chain hierarchy and graph equivariance.

Purpose of the Study:

• To develop an efficient computational method for identifying protein-protein interaction sites.
• To address limitations of existing methods regarding hierarchical protein structures and spatial transformations.

Main Methods:

• Developed EGCPPIS, an end-to-end Graph Neural Network (GNN)-based method.
• Constructed hierarchical graphs (residue-level and atom-level).
• Employed E(n) Equivariant Graph Neural Networks (EGNN) and GraphSAGE modules.
• Integrated hierarchical features using contrastive learning and gated multi-head attention.

Main Results:

• EGCPPIS achieved superior performance in identifying protein-protein interaction sites.
• Demonstrated significant outperformance compared to state-of-the-art methods on multiple datasets.
• The method effectively learns consistent residue- and atom-level embeddings.

Conclusions:

• EGCPPIS significantly advances the accuracy of protein-protein interaction site prediction.
• The method provides insights into decision-making patterns for PPI discovery.
• Available code and datasets facilitate further research and application.