A Hierarchical Surface Graph Framework for Protein–Protein Binding Affinity Prediction

Accurate prediction of protein--protein binding affinity is fundamental to understanding molecular interactions and has broad implications for protein engineering. We introduce ProtSurf, a novel protein surface representation learning framework for predicting binding affinity between protein complexes. While many surface-based methods rely only on locally defined neighborhoods, ProtSurf builds a hierarchical surface graph architecture, with a local encoder that aggregates chemical and geometric features from residue-level surface patches and a global graph module that models interactions across the residues. We use graph transformer with relative positional encoding to represent both local and global features and a permutation-invariant reconstruction module to infer node ordering and reconstruct local residue graphs. We further show that data augmentation with protein protein interaction data from AlphaFold improves affinity prediction. Evaluated across three protein–protein affinity tasks — SKEMPI, SAbDab, and HER2 — ProtSurf achieves state-of-the-art performance on these datasets.