Protein–protein interactions (PPIs) play fundamental roles in biological processes and therapeutic development. Accurately predicting PPI binding affinity is critical for understanding interaction mechanisms and guiding protein engineering. Recent advances in structure prediction like AlphaFold have enabled accurate modeling of protein–protein complexes, creating new opportunities for structure-based affinity prediction. However, existing methods predominantly rely on sequence information and fail to fully exploit structural insights at interaction interfaces. To address this gap, we propose PPAP, a novel deep learning framework that integrates structural features with sequence representations through an interfacial contact-aware attention mechanism. Our model demonstrated superior prediction performance across all evaluated data sets, outperforming strong sequence-based large language models on the internal test (R = 0.540, MAE = 1.546). On the external test set, our model achieved a higher Pearson correlation coefficient (R = 0.63) than all benchmarked models. In protein binder design, we further demonstrate that incorporating our model's prediction can enhance enrichment by up to 10-fold in comparison to the metrics based on AlphaFold-Multimer prediction. Given its robust performance, PPAP holds promise as a valuable tool not only for protein design but also for a wide range of protein interaction-related applications.
Qian et al. (2025) studied this question.