SNPDogg: Feature-importances in the identification of harmful missense SNPs

Jeff Gaither (May 8, 2020)

Please install the Flash Plugin

Abstract

Recent years have seen an explosion in the use of machine-learning algorithms to classify human mutations. There are now at least 30 scores designed to identify mutations likely to be deleterious to humans, but almost all are "black boxes" that provide no explanation of how they arrived at their predictions. In this talk I'll introduce a new mutational pathogenicity score, SNPDogg, that is transparent, insofar as every prediction can be decomposed as a sum of contributions from the model's features. SNPDogg's feature-importance ​values are computed via a game-theoretic approach implemented in the "shap" python package.