Abstract
Deep learning has advanced our ability to assess the effects that individual mutations have on protein function; however, predicting the complex interplay between two or more mutations remains challenging. Here, we seek to address this challenge by building a deep learning framework that incorporates information related to protein dynamics. Namely, we build a neural network architecture using a physics-based metric called the Asymmetric Dynamic Coupling Index (DCI(asym)), which quantifies the degree to which each member of a pair of residues influences the flexibility of the other. DCI(asym) enables us to train models through an allosteric Graph Neural Network (GNN) in which each residue is linked to its distant dynamic influencers. Despite not being trained on experimental epistasis data, our GNN consistently outperforms existing approaches on deep mutational scanning datasets across four distinct proteins, highlighting its enhanced capacity to model epistatic interactions. Our GNN model was then challenged to predict the functions of 37 novel, computationally designed TEM-1 β-lactamase variants of unknown function, and it demonstrated excellent predictive accuracy for these variants. Thus, our GNN provides a pathway for better assessing the impact of multiple mutations on protein function, including epistatic relationships and mutations that have profound effects on activity despite being spatially far from the active site.