Having established the utility of TWAS, we applied the approach to identify novel expression-trait associations using summary data from three recent GWAS over more than 900,000 phenotype measurements: lipid measures (high-density lipoproteins [HDL] cholesterol, low-density lipoprotein [LDL] cholesterol, total cholesterol [TC], and triglycerides [TG])5; height7; and BMI6. Significantly cis-heritable genes across the three expression data sets were tested individually (6,924 tests) and together in an omnibus test that accounts for predictor correlation (1,075 tests; Methods), and we conservatively corrected for the 8,000 total tests performed for each trait. Overall, we identified 665 significant gene-trait associations (Supplementary Table 8). Of these, 69 gene-trait associations did not overlap a genome-wide significant SNP in the corresponding GWAS, residing in 60 physically non-overlapping cis-loci (Table 1, Supplementary Table 9). Averaging over the novel genes, the Z2 statistics from TWAS were 1.5x higher than the strongest eQTL SNP for the same gene(though this may be slightly inflated due to winner’s curse). Our previous simulations suggest that the substantial gain over testing the cis-eQTL is an indication of pervasive allelic heterogeneity40 at these loci, and analyses of the expression showed strong evidence for allelic heterogeneity at the TWAS genes (Supplementary Figure 17).
