Fig. 1

Schematic overview of predictor benchmarking in population-based cohorts based on human gene-trait associations. A Participant-level genotypes and phenotypes were extracted from the UK Biobank and All of Us cohorts for the corresponding sets of gene-trait combinations, and (B) predicted functional scores were collected for a set of 24 computational variant effect predictors. C In order to assess predictor performance, the area under the balanced precision-recall curve (AUBPRC) and Pearson correlation coefficient (PCC) was measured for binary and quantitative traits, respectively. To estimate the uncertainty in these measurements, participants were resampled with replacement and performance measures recalculated for each resampled set. D For each gene-trait combination, predictors were ranked by mean performance (AUBPRC or PCC), and a false discovery rate (FDR) was calculated to assess whether performance differences were statistically significant. E To summarize comparisons across all gene-trait combinations, we (left) summed the number of combinations for which a predictor was either best performing or tied (FDR ≥ 10%) for best, and (right) compared the overall difference in performance measures between predictor pairs across all gene-trait combinations