Fig. 2: Predictive performance comparison between scPRS and baseline methods.

a, Bar plots of AUROC values of different models. The training and testing procedure was conducted for ten repeats with different random seeds. Training, validation and test dataset splits were kept identical across different methods to ensure a fair comparison. scPRS+, scPRS model integrating nonpeak PRSs; scPRS+covar, scPRS model integrating nonpeak PRSs and covariates (that is, age, sex and first ten PCs); C+T (nonpeak), LR model of nonpeak C+T PRSs; Covar, LR model of covariates. Performance comparison was conducted using a one-sided paired t-test. The mean and 95% confidence interval (CI) are annotated using the bar plot and error bar, respectively. b, ROC curves of different models evaluated on independent target cohorts. The performance of a random predictor is shown by the dashed gray line.