MPFIL: Multi-Level Protein Feature Integration for Predicting Mutation-Induced Protein-Protein Binding Affinity Changes

Predicting mutation-induced changes in protein protein binding affinity (∆∆G) is challenging because mutations can affect both local residue environments and the structural organization of protein complexes. Existing approaches often rely on handcrafted physicochemical, energetic, or evolutionary features, which introduce additional computational costs and require domain-specific knowledge. In this work, we propose MPFIL, a multi-level protein feature integration framework for predicting mutation-induced protein-protein binding affin ity changes. MPFIL integrates structural and sequential infor mation to characterize mutation effects from complementary perspectives. Its structural feature extraction module captures mutation-induced structural changes and differential information between wild-type and mutant complexes, while sequence feature extraction modules model local residue context and long-range dependencies. The resulting features are integrated for ∆∆G prediction. We evaluate MPFIL on three benchmark datasets. Experimental results demonstrate competitive and consistent performance across different protein complex settings, while ablation studies verify the contribution of the structural and sequential features. These results support the effectiveness of multi-level protein feature integration for modeling mutation induced changes in protein-protein binding affinity.