Accurately obtaining Protein-protein Interaction Sites (PPIS) information is crucial for understanding cell functions and drug development. In recent years, approaches based on Graph Neural Networks with a focus on local-global collaborative modeling have achieved significant advances in PPIS prediction. However, existing approaches typically process local and global structural information separately, leading to a lack of contextual awareness in local information. Furthermore, current fusion strategies for integrating these two types of information still have room for improvement. In this study, we propose a novel structure-aware protein-protein interaction site prediction with cascaded local and global information (RCLG-PPIS). Specifically, RCLG-PPIS employs residual cascade local-global (RCLG) module to strengthen the fusion between local and global information. In RCLG module, an E (n) equivariant graph neural network (EGNN) extracts local 3D structural features of proteins, while transformer captures long-range residue dependencies from a global perspective. By cascading these two components, transformer helps the EGNN expand its receptive field. Additionally, we integrate the Laplacian eigenvectors structural encoder (LapSE) into the node features to enhance the predictive capability of RCLG-PPIS. Experimental results on the benchmark data set (Test₆0) demonstrate that RCLG-PPIS outperforms state-of-the-art models across multiple metrics, achieving notable improvements of 6. 38% in AUPRC, and 8. 80% in MCC.
Chen et al. (Thu,) studied this question.