Abstract Accurate prediction of damaging missense variants is critically important for interpreting genome sequence. While many methods have been developed, their performance has been limited. Recent progress in machine learning and availability of large-scale population genomic sequencing data provide new opportunities to significantly improve computational predictions. Here we describe gMVP, a new method based on graph attention neural networks. Its main component is a graph with nodes capturing predictive features of amino acids and edges weighted by coevolution strength, which enables effective pooling of information from local protein context and functionally correlated distal positions. Evaluated by deep mutational scan data, gMVP outperforms published methods in identifying damaging variants in TP53, PTEN, BRCA1 , and MSH2 . Additionally, it achieves the best separation of de novo missense variants in neurodevelopmental disorder cases from the ones in controls. Finally, the model supports transfer learning to optimize gain- and loss-of-function predictions in sodium and calcium channels. In summary, we demonstrate that gMVP can improve interpretation of missense variants in clinical testing and genetic studies.

Predicting functional effect of missense variants using graph attention neural networks