Fig. 2: Heritability, SNP-based heritability and variance explained in out-of-sample prediction.

Heritability (left panel) and SNP-based heritability (middle panel) estimates and the variance explained in out-of-sample prediction (right panel). Heritability and SNP-based heritability estimates are presented with 95% confidence interval. GCTA-GREML was used to estimate heritability from a UKB subset that included all pairs of individuals related with coefficient of relationship > 0.2 (N = 58,738 relatives). GCTA-GREML was used to estimate SNP-based heritabilities labelled GREML summer or winter using samples of ~50 K participants randomly drawn from the UKB. The SBayesR SNP-based heritability is estimated from the GWAS summary statistics (N = 417,580). In out-of-sample prediction into the QIMR and the UKB replication (UKBR) samples, polygenic risk scores (PRS) calculated by the standard P-value threshold method (P + T) were outperformed by using SNP effect estimates calculated from GWAS summary statistics using the SBayesR or SBayesS methods. Bars of the same colour used the same methodology (noting that SBayesR generates an estimate of SNP-based heritability as well as SNP effect sizes in prediction analysis). The numbers on top of the bars are −log10 P-value of the regression of 25OHD on 25OHD PRS. COJO conditional and joint, rg genetic correlation, s.e. standard error.