Using the case-control cohorts in the meta-analysis, we conducted a leave-one-cohort-out GWAS meta-analysis for 42 European ancestry cohorts that had provided individual-level data. Polygenic scores (PGSs) were generated in each cohort using SNP weights for the multi-ancestry and the European ancestry meta-analyses derived using SBayesR.20 Other PGS methods, including the standard p value clumping and thresholding, gave similar results (Table S6). Across all European ancestry cohorts, the variance explained on the liability scale rl2 was 5.8% (SE 0.2%) (see key resources table and Table S6), with an area under the receiver operating characteristic curve (AUC) statistic of 0.625 (see key resources table). Adding functional annotations into the algorithm to generate SNP weights for PGSs (SBayesR) increased prediction accuracy by 0.1% (i.e., rl2 of 5.9%). The rl2 was more than 1.4 times greater than that reported in the PGC MDD 2018 analysis7,21 (Figure 4). The OR for being a case per standard deviation (SD) increase in PGS was 1.57. The OR for being a case in the tenth compared with the first decile of PGSs was 4.92 (95% CI 4.57–5.29) (Figure
