Title: Structure-Guided Adaptive Propagation for Protein-Protein Interaction Site Prediction

Abstract:

Identifying protein-protein interaction sites (PPIS) with high accuracy is a critical step in deciphering cellular functions, elucidating disease pathways, and discovering new therapeutic targets. While graph-based deep learning has significantly improved PPIS prediction by integrating residue-level structural data, existing models often employ rigid propagation schemes. These standard approaches treat residues uniformly, ignoring the inherent structural and functional diversity found at protein interfaces. This lack of adaptability hinders the model's capacity to tailor information diffusion to specific local geometric contexts, thereby complicating the distinction between genuine interaction sites and non-interacting neighbors with similar structural features.

To address this limitation, we introduce SGAP-PPIS, a novel model that utilizes structure-guided adaptive propagation for PPIS prediction. Unlike conventional fixed mechanisms, SGAP-PPIS employs an equivariant graph neural network to extract multi-scale geometric states. These states are used to calculate residue-specific propagation coefficients, enabling each residue to dynamically adjust the trade-off between preserving local features and diffusing information across its neighborhood based on its unique geometric microenvironment.

Our experiments demonstrate that SGAP-PPIS delivers performance comparable to leading state-of-the-art methods on the Test_60 benchmark. Furthermore, ablation studies confirm that these enhancements are driven by three key components: geometry-conditioned adaptive propagation, scale-aligned geometric guidance, and a multi-step propagation-state representation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC