GeoGAT-site: A Face-Centered Geometric Graph Attention Network for Protein-Protein Interface Prediction

Abstract

Protein-protein interactions (PPIs) underpin the intricate machinery of cellular life, orchestrating processes from signal transduction to metabolic regulation, yet their precise interface prediction remains a cornerstone challenge in structural biology, demanding scalable computational paradigms that balance accuracy and efficiency. Herein, we present GeoGAT-site, a pioneering geometric graph attention network that leverages face-centered surface fingerprints extracted from three-dimensional protein architectures to forecast interaction sites. Diverging from conventional vertex-centric approaches, our face-centered methodology achieves a 3.64-fold acceleration in patch generation, mitigating computational bottlenecks while preserving granular surface descriptors. At its core, GeoGAT-site incorporates a bespoke attention mechanism that adaptively modulates inter-facial distances and normal vector orientations, synergistically integrating spatial geometries with physicochemical attributes for nuanced interface delineation. Harnessing a meticulously curated dataset of 150 million face-centered fingerprints from over 20,000 structurally diverse proteins, the model attains robust generalization across heterogeneous interaction motifs. Empirical validation on an independent cohort of 167 protein complexes yields a ROC AUC of 0.89, surpassing established benchmarks including MaSIF-site (0.845), SPPIDER (0.65), and PSIVER (0.63). By furnishing high-fidelity interface annotations, GeoGAT-site augments downstream structural modeling paradigms through targeted constraints, thereby providing a versatile scaffold for unraveling PPI dynamics with profound implications for therapeutic discovery, protein redesign, and molecular epistemology.