Fig. 6: Single-trait versus multi-trait polygenic scores.

Area under the curve (AUC) from logistic regression model with the disease as an outcome (atrial fibrillation and flutter [AF], coronary artery disease [CAD] and heart failure [HF]) and the disease-specific candidate polygenic score as a predictor. The models including the single-trait polygenic scores (PGSsingleT) are represented in blue and the multi-trait ones (PGSmultiT) in red. On the left, the PGS univariate models are depicted and, on the right, the PGS models adjusted for age at last follow-up, sex and the four first principal components of genetic ancestry are shown. PGS developed by LDpred2, using the United Kingdom Biobank linkage disequilibrium reference panel (N = 362,320 individuals). The validation (N = 2500) and testing datasets (N = 12,677) consisted of unrelated individuals from the Montreal Heart Institute Biobank cohort. AUC shown are those from the testing dataset. For the female cohort, 997 individuals were included in the validation dataset and 5152 in the testing dataset. For the male cohort, 1503 individuals in the validation dataset and 7525 in the testing dataset. Representation of the AUCs corresponding to the best LDpred2 method, defined as the one with highest Wald-Z in the validation dataset in each disease. 95% confidence intervals are represented. Significant differences in the discrimination ability assessed using ∆AUC, Net Reclassification Index and Integrated Discrimination Improvement marked with an asterisk (*). The lack of model calibration in HF only-PGS models indicates caution in the interpretation of these results (Supplementary Data 36). Interestingly, PGS derived from the multi-trait summary statistics (PGSmultiT) improved prediction and discrimination in CAD in females (Supplementary Data 37). Abbreviations: PGS, polygenic score.