Deep Learning for Polygenic Risk Prediction

Abstract

Polygenic risk scores (PRS) are relative measures of an individual’s genetic propensity to a particular trait or disease. Most PRS methods assume that mutation effects scale linearly with the number of alleles and are constant across individuals. While these assumptions simplify computation, they increase error, particularly for less-represented racial groups. We developed and provide Delphi (deep learning for phenotype inference), a deep-learning method that relaxes these assumptions to produce more predictive PRS. In contrast to other methods, Delphi can integrate up to hundreds of thousands of SNPs as input. We compare our results to a standard, linear PRS model, lasso regression, and a gradient-boosted trees-based method. We show that deep learning can be an effective approach to genetic risk prediction. We report a relative increase in the percentage variance explained compared to the state-of-the-art by 11.4% for body mass index, 18.9% for systolic blood pressure, 7.5% for LDL, 35% for C-reactive protein, 16.2% for height, 29.6 % for pulse rate; in addition, Delphi provides 2% absolute explained variance for blood glucose while other tested methods were non-predictive. 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 showed to generalize well when evaluated on individuals from non-European ancestries.