Delphi: Deep Learning for Polygenic Risk Prediction

Abstract

Polygenic scores (PGS) are relative measures of an individuals genetic propensity to a particular trait or disease. Most PGS methods use a regression framework for polygenic modeling and assume that mutation effect estimates are constant across individuals. While these assumptions simplify computation, they increase error, and PGS are particularly less predictive for under-represented genetic ancestries. We developed and provide Delphi (deep learning for phenotype inference), an individual-level deep-learning method that relaxes these assumptions to produce more predictive PGS. Delphi can integrate up to hundreds of thousands of SNPs as input and model non-linear SNP-SNP and SNP-covariate interactions. We compare our results with linear PGS models and a gradient-boosted trees-based method. We show that deep learning can be an effective approach to genetic risk prediction. We report substantial performance gains for a broad range of continuous phenotypes compared to the state-of-the-art. Furthermore, we show that Delphi tends to increase the weight of high-effect mutations. This work demonstrates an effective deep learning method for modeling genetic risk that also generalizes well when evaluated on individuals from non-European ancestries.