Beyond predictive R²: Quantile regression and non-equivalence tests reveal complex relationships of traits and...

Polygenic scores (PGSs) are genetic predictions of trait values or disease risk that are increasingly finding applications in clinical predictive models and basic genetics research. However, the predictive value of a PGS can vary within similar population groups, depending on characteristics such as the environmental exposures, sex, age, or socioeconomic status of the individuals. To maximize the value of a PGS, approaches to screen trait-PGS pairs for evidence of such heterogeneity without having to specify the relevant exposure or individual characteristics would be useful. Here, in analyses from the UK Biobank, we show that a PGS's predictive accuracy depends on the quantile of the phenotypic distribution to which the PGS is being applied. We quantify differences in predictive value across the phenotypic range using quantile regression linear models to estimate quantile-specific effect sizes for linear models of phenotype values as a function of PGS. Of 25 continuous traits, only three have no quantile-specific effect sizes that varied by at least 1.2-fold from the ordinary least squares estimate. Through simulation, we demonstrate that this heterogeneity of PGS predictive value can arise from gene-by-environment interactions. Our approach can be used to flag traits where the use of PGSs warrants extra caution, and perhaps stratification variables should be sought and used because PGSs perform substantially differently in portions of the sampled population than expected from quoted predictive R² or incremental R² values that represent average performance across a dataset.