The GPS far outperformed a score based only on the 141 variants most strongly associated with BMI, consistent with the highly polygenic nature of BMI and obesity. For example, in a direct comparison in 119,951 individuals, we observe a correlation with BMI of 0.29 for the GPS as compared with 0.13 with the 141-variant score. This improved performance using a genome-wide set of common variants was anticipated by a prior theoretical projection study based on early GWAS results and an analysis that indicated minimal ‘missing heritability’ of BMI when accounting for the full range of observed genetic variation (Chatterjee et al., 2013; Yang et al., 2015). Here, we use a recently developed computational algorithm that explicitly models the correlation structure between variants in calculating variant weights(Vilhjalmsson et al., 2015). This algorithm has been shown to outperform prior methods for a range of complex traits including cardiovascular disease, type 2 diabetes, and educational attainment (Khera et al., 2018a; Lee et al., 2018).
