Fig. 3

Predictor rankings across all gene-trait combinations in the UK Biobank (top) and All of Us (bottom) cohorts. A and C The number of gene-trait combinations for which a given predictor was either best performing (in terms of mean AUBPRC or PCC) or tied (FDR ≥ 10%) with the best-performing predictor in the UK Biobank and All of Us cohorts, respectively. In the UK Biobank cohort, 140 gene-trait combinations were considered; from this, 116 gene-trait combinations were matched in the All of Us cohort. B and D The overall difference in performance measures between predictor pairs was assessed using a two-tailed Wilcoxon signed-rank test comparing the distributions of mean performance measures across all gene-trait combinations for each pair; the predictors in a given pair are considered statistically different at an FDR < 10% (indicated in blue-grey). Where predictors were tied in the overall ranking (i.e., were best or tied for best in the same number of gene-trait combinations) ties were broken first based on the number of pairwise comparisons for which a given predictor statistically outperformed another across all gene-trait combinations; and second, where necessary, based on the number of comparisons for which a given predictor yielded lower q-values than the predictor with which it was tied. The overall ranking of predictors in the UK Biobank and All of Us cohorts showed significant positive correlation (Kendall’s Tau = 0.75; p-value = 1 × 10^–8)