Quantifying Protein-Protein Interaction with a Spatial Attention Kinetic Graph Neural Network

Abstract

Accurate prioritisation of near-native protein-protein interaction (PPI) models remains a major bot-tleneck in structural biology. Here we present SAKE-PP, a physics-inspired, spatial-attention equivariant graph neural network that directly regresses interface RMSD (iRMSD) without any native references. Trained on docking decoys generated through our novel hierarchical sampling strategy applied to PDB-Bind dataset, SAKE-PP combines force-field-like attention with Laplacian-eigenvector orientation to couple local interaction forces with global topology. On the 2024PDB benchmark comprising 176 het-erodimers, SAKE-PP demonstrates effective optimization and selection of AF3 decoys, achieving improvements of 13.75% based on iRMSD statistics and 12.5% based on DockQ scores. It consistently outperforms the AF3 ranking score in multiple metrics, including overlap, hit rate, and correlation. In zero-shot evaluation of 139 antibody-antigen complexes, SAKE-PP improves the score-iRMSD correlation by 0.4. By unifying geometric deep learning with physics-based realism, SAKE-PP provides a robust, plug-and-play scoring function that streamlines reliable PPI evaluation and accelerates downstream structure-guided drug-design workflows.