MissenseHMM: state-based annotations for missense variants through joint modeling of pathogenicity scores

Abstract

Many computational predictors of missense variant pathogenicity are available. To capture information across various predictors, we propose MissenseHMM, which learns states corresponding to combinatorial patterns of variant prioritizations. We applied MissenseHMM to 43 predictors, annotating over 70 million missense variants with 20 states that showed distinct predictor scores patterns, amino acid substitutions and other genomic annotation enrichments. MissenseHMM state annotations enhanced individual predictors' associations with clinical pathogenic variants and deep mutational scanning data, and also provided insight into the performances of various protein language models. Overall, MissenseHMM complements pathogenicity predictors and is an annotation resource for missense variant interpretation.