MicroEnvPPI: Microenvironment-Aware Optimization Enables Generalizable Protein-Protein Interaction Prediction.

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions and drug discovery.
• Existing models like AlphaFold may miss residue-level microenvironment details, limiting PPI prediction accuracy.
• Understanding these microenvironments is key to advancing PPI inference.

Purpose of the Study:

• To introduce MicroEnvPPI, a novel framework for enhanced protein-protein interaction prediction.
• To improve the accuracy and generalizability of PPI prediction by focusing on residue microenvironments.
• To leverage structural and contextual information for comprehensive microenvironment characterization.

Main Methods:

• Integrating residue physicochemical features and ESM-2 embeddings with AlphaFold-predicted structures.
• Employing auxiliary tasks, including graph contrastive learning and masking, for microenvironment representation optimization.
• Jointly training global PPI prediction and microenvironment optimization tasks within the MicroEnvPPI framework.

Main Results:

• MicroEnvPPI demonstrates improved accuracy and generalizability in PPI prediction.
• The model shows robust performance on challenging data splits (DFS, BFS), indicating strong generalization to novel interactions.
• Comprehensive characterization of residue microenvironments enhances predictive capacity.

Conclusions:

• MicroEnvPPI offers a powerful approach to advancing protein-protein interaction network analysis.
• The framework's focus on microenvironments significantly improves PPI prediction.
• This work has implications for understanding cellular networks and identifying new therapeutic targets.