Biophysically Grounded Deep Learning Improves Protein–Protein ΔΔG Prediction

Predicting how mutations alter protein–protein binding affinity (ΔΔG) is central to protein engineering and therapeutic design, yet current deep learning models often fail to generalize beyond narrow benchmarks. We show that SKEMPI2, the field’s dominant training and evaluation dataset, suffers from widespread and previously underappreciated data leakage caused by deep sequence and structural redundancy, leading to inflated performance estimates. We introduce ProtBFF (Protein Biophysical Feature Framework), an encoder-agnostic module that injects five interpretable bio-physical priors into residue-level representations via cross-embedding attention. ProtBFF consistently improves performance under homology-based-clustering evaluation and enables general-purpose encoders, such as ESM, to match or surpass state-of-the-art specialized models or larger size models. These results demonstrate that incorporating biophysical priors could be key to robust protein property prediction.