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

Abstract

Protein-protein interactions (PPIs) play a fundamental role in shaping cellular functional networks and guiding therapeutic target discovery. Although models such as AlphaFold have achieved impressive results in protein structure prediction and PPI inference, they tend to overlook the structural and contextual importance of residue-level microenvironments, which limits their predictive capacity. Here, we present MicroEnvPPI, a microenvironment-aware optimization framework designed to improve the accuracy and generalizability of PPI prediction. MicroEnvPPI integrates residue-level physicochemical features and contextual embeddings derived from the ESM-2 language model with structural information predicted by AlphaFold, enabling a comprehensive characterization of residue microenvironments. Additionally, auxiliary tasks that incorporate graph contrastive learning and masking mechanisms optimize the residue microenvironment representation, enhancing both its quality and the model's generalization ability. Finally, MicroEnvPPI strengthens its advantage in PPI prediction by jointly training global PPI and microenvironment optimization tasks. Notably, MicroEnvPPI achieves strong performance under challenging data partition schemes, such as DFS and BFS, indicating its ability to generalize to previously unseen interactions. These findings underscore the potential of MicroEnvPPI to advance our understanding of protein interaction networks.