Context-Dependent Dynamic Coupling in Residue Contact Networks Underlies Thermodynamic Non-Additivity in Protein Double Mutants

Abstract

Protein engineering often combines multiple mutations, yet predicting how these combinations affect stability remains a fundamental challenge. The double-mutant effect often departs from the sum of its constituent single-mutant contributions, a phenomenon known as mutation epistasis or non-additivity. Anticipating this is critical, as the number of candidate pairs grows combinatorially and each non-additive pair can either strengthen or cancel an intended design. Despite growing interest, current approaches to predicting epistasis remain limited, and systematically identifying which mutation pairs will behave non-additively remains an open problem. Here, we present a Dynamic Residue Contact Network Clique (DynRCNC) model to capture the non-additive mutation pairs. DynRCNC provides a protein native network with residue-residue dynamic coupling from molecular dynamics (MD) simulations and forms a three-tier prediction hierarchy with six criteria. The model was applied to 265 double-mutation pairs from seven proteins and achieved a weighted Matthews correlation coefficient of 0.422 under leave-one-protein-out cross-validation (weighted LOPOCV MCC) on the full benchmark. It transfers across proteins better than random-forest, gradient-boosting, and logistic-regression models trained on the same structural and dynamic features. Needing no training data from the target protein and reporting the criterion behind each prediction, DynRCNC gives an interpretable, mechanistic route to prioritizing candidate mutation pairs.