FamGeo-DDG: A Mutation-Centered Local Geometric-Semantic Framework for Protein–Protein ΔΔ G Prediction

Protein–protein interactions (PPIs) are central to cellular regulation, immune recognition, and therapeutic design, making accurate prediction of mutation-induced binding free energy changes (ΔΔG) important for variant interpretation and protein engineering. However, ΔΔG prediction remains challenging because mutational effects are often localized, while benchmark performance may be inflated by residual similarity between training and validation complexes. Here, we present FamGeo-DDG, a mutation-centered local geometric-semantic framework for protein–protein ΔΔG prediction. FamGeo-DDG constructs a local residue graph around the mutation site or the union of local neighborhoods surrounding multiple mutation sites and integrates residue substitution information, DSSP-derived structural descriptors, and ESM-2-derived contextual difference features (ΔESM). A local geometric encoder and mutation-centered readout capture structural context and mutation-induced perturbations. For evaluation, we adopt three complementary 5-fold partitioning protocols, the family split, which groups complexes by complex-signature similarity, the MpbPPI split, which directly reuses the partitions provided by MpbPPI, and the structure split, which uses ECOD-guided grouping based on available structural annotations. Under matched 5-fold partitions, we compare FamGeo-DDG with 13 recent and classical baselines, including TopNetTree, GeoPPI, 3D-ΔΔG, DSSA-PPI, MutPPI+, IGMI, Prot-ΔΔG, and Pythia-PPI, and report mean ± standard deviation across folds. On the family split, FamGeo-DDG achieves PCC values of 0.5471 ± 0.1212, 0.4956 ± 0.1299, 0.5486 ± 0.1830, and 0.5168 ± 0.0792 on S645, M1101, S1131, and S4169, respectively. Across multiple data sets and partition protocols, FamGeo-DDG achieves consistent predictive performance, demonstrating the effectiveness of mutation-centered geometric and semantic representations. Ablation analysis and external validation further support the contribution of individual model components and the transferability of FamGeo-DDG across different mutation prediction scenarios.