Predicting protein thermal stability changes upon single and multi-point mutations via restricted attention subgraph neural network

Abstract

Accurate prediction of protein stability changes due to mutations is instrumental for understanding mechanisms of disease and drug failure, as well as for engineering tailored protein-based materials. In recent years, computational tools using machine learning have been developed to predict stability changes and thereby supplement available experimental methods that may be time consuming and costly for bulk mutational analysis. Existing tools are limited, however, to predicting single point mutations and showing antisymmetric bias in direct/reverse mutations. Here, we develop structure and sequence-based models to predict protein stability changes via Gibbs free energy change upon single or multi-point mutations via two parallel augmented gated attention graph neural networks integrated with global attention blocks receiving subgraphs. These subgraphs are related to mutation sites and constructed through predicted protein contact maps to capture spatial structural information. We train our model on direct and reverse mutations obtained from the S5294 dataset and test on one independent test set and eight most commonly used test sets, including S350, P53, Ssym, S669, S1925, S250, and myoglobin. Our approach shows considerable improvement in estimating the impacts of stabilizing mutations, and consistently outperforms other methods by at least a 5.2% improvement in root mean square error. This approach can be employed for finding functionally important protein variants, helping to design new proteins with vast untapped potential for broad pharmaceutical applications.