Biophysical Priors Enhance Protein–Protein Binding ΔΔG Prediction

Predicting mutational effects on protein–protein binding affinity ($\Delta\Delta G$) remains a challenging task: current models often generalize poorly due to limited and biased training data. We show that SKEMPI2, the dominant training and evaluation dataset in this field, is affected by a subtle and pervasive data leakage due to sequential and structural redundancy, leading to inflated estimates of performance across models. We introduce ProtBFF (\underline{Prot}ein \underline{B}iophysical \underline{F}eature \underline{F}ramework), a lightweight, encoder-agnostic module that injects five key biophysical features (interface, burial, dihedral, SASA, lDDT) into residue latent representations via cross-embedding attention. ProtBFF consistently improves predictive power and, with ProSST, achieves state-of-the-art performance on clustered SKEMPI2, rivaling far more specialized models. These results point to a simple, general recipe for protein property prediction: integrate biophysical priors with machine learning.