An Antisymmetric Framework with Cross-Modal Information Decoupling for Predicting Missense Mutation-Induced Changes in Protein–Ligand Binding Affinity

Abstract Accurately predicting changes in protein–ligand binding affinity induced by missense mutations is important for understanding disease mechanisms, identifying drug resistance mutations, and supporting precision medicine. However, existing computational methods often directly integrate protein sequence and structure information, which may introduce background information, structural uncertainty, and redundant cross-modal information, limiting the identification of transferable mutation-related features. In addition, most existing approaches do not explicitly enforce the thermodynamic consistency of binding free energy changes upon mutation reversal. Here, we present MutBindPred, a cross-modal deep learning framework for predicting missense mutation-induced changes in protein–ligand binding affinity. MutBindPred jointly models wild-type and mutant protein–ligand complexes and introduces a cross-modal information decoupling module to separate modality-shared background information from modality-specific mutation-related signals. The framework further captures directional differences between wild-type and mutant states and incorporates an antisymmetric constraint to improve the thermodynamic consistency of predicted ΔΔG values. Under rigorous protein and ligand cold-start settings, MutBindPred achieves improved generalization performance compared with existing baseline methods. Further analyses show that sequence-derived heterogeneous representations provide the strongest transferable predictive signals. Structural representations show limited standalone predictive utility but contribute as an auxiliary modality for guiding cross-modal representation learning. Overall, MutBindPred provides a thermodynamically consistent and generalizable framework for predicting missense mutation-induced changes in protein–ligand binding affinity.