Predicting how mutations shape protein function remains a major challenge, particularly when multiple substitutions interact in non-additive ways. Traditional strategies, such as exhaustive mutational scanning or brute-force molecular dynamics simulations, are too costly to apply broadly. This motivates the need for generalizable physical principles that capture how mutations alter protein motion and long-range communication. Each residue is represented as a node, with edges defined not simply by correlated motion but by directional control. Using the asymmetric dynamic coupling index (DCI asym ), we identify cases where the motion of one residue disproportionately drives the motion of another. Edges are therefore assigned only from controlling residues to those whose dynamics are governed by them, capturing asymmetry in long-range allosteric communication. The model integrates convolutional neural network capturing local residue interactions and graph convolutional network capturing long-range allosteric communication. While remaining grounded in the physical principles of protein dynamics, it can predict the effects of tens of thousands of variants within an hour, which is orders of magnitude faster than conventional simulations. We trained and validated the model on deep mutational scanning data sets for β-lactamase, KRAS, GB1, and GFP. Across all four proteins, the dynamics-informed model consistently outperformed sequence-based model and revealed mechanistic insights into how distal mutations regulate activity through allosteric pathways. These findings demonstrate that integrating protein dynamics with deep learning provides a generalizable and efficient strategy for modeling mutational landscapes and allosteric regulation in diverse proteins.
Huynh et al. (Sun,) studied this question.