Performance of deep-learning based approaches to improve polygenic scores

Abstract

Background/ObjectivesPolygenic scores (PGS), which estimate an individuals genetic propensity for a disease or trait, have the potential to become part of genomic healthcare. In maximising the predictive performance of PGS, neural-network (NN) based deep learning has emerged as a method of intense interest to model complex, nonlinear phenomena, which may be adapted to exploit gene-gene (GxG) and gene-environment (GxE) interactions. MethodsTo infer the amount of nonlinearity present in a phenotype, we present a framework for using NNs, which controls for the potential confounding effect of correlation between genetic variants, i.e. linkage disequilibrium (LD). We fit NN models to both simulated traits and 28 real disease and anthropometric traits in the UK Biobank. ResultsSimulations confirmed that our framework adequately controls LD and can infer nonlinear effects, when such effects genuinely exist. Using this approach on real data, we found evidence for small amounts of nonlinearity due to GxG and GxE which mildly improved prediction performance (r2) by [~]7% and [~]4%, respectively. Despite evidence for nonlinear effects, NN models were outperformed by linear regression models for both genetic-only and genetic+environmental input scenarios with [~]7% and [~]5% differences in r2, respectively. Importantly, we found substantial evidence for confounding by joint tagging effects, whereby inferred GxG was actually LD with due to unaccounted for additive genetic variants. ConclusionOur results indicate that the usefulness of NNs for generating polygenic scores for common traits and diseases may currently be limited and may be confounded by joint tagging effects due to LD.