We performed a sensitivity analysis to assess whether the estimation of the metaGRS weights on the UKB derivation set led to over-fitting (upwards bias in apparent performance) of the score in the validation set. We developed a metaGRS based on four component GRSs (AS, IS, BMI, and SBP) in cross-validation on the derivation set. We compared this metaGRS with a score derived using smtPred28, which relies on the chip heritabilities and genetic correlations estimated from the GWAS summary statistics via LD score regression31,32, independently of the UKB (Supplementary Fig. 2). Overall, the two scores were highly correlated (Pearson r = 0.98), and had indistinguishable associations with IS in the UKB validation set, indicating that our metaGRS procedure did not lead to overfitting in the validation set.
