PPAC: Predicting Protein–Protein Affinity Changes Induced by Amino Acid Mutations Using Protein Large Language Models

Understanding the impact of amino acid mutations on protein-protein binding free energy is fundamental to drug design and functional biology. We propose a novel approach utilizing Protein Large Language Models (PLMs) to characterize both wild-type and mutant proteins, enabling accurate predictions of mutational effects. We employed three state-of-the-art PLMsEsm2, EsmC, and ProtT5to generate sequence-based representations. These representations were subsequently integrated into seven distinct model architectures. Through a rigorous 5-fold cross-validation, we selected the optimal model and feature combination before training on a large-scale data set. Our results show that this PLM-based method significantly outperforms traditional approaches, achieving state-of-the-art (SOTA) predictive performance. The final PPAC model was evaluated on a test set of 9,558 data points and applied to two case studies. The results demonstrate that the model not only provides high-precision predictions but also exhibits a significant advantage in identifying key residues crucial for protein interactions, highlighting its effectiveness in protein interaction modeling.