Detecting a locus was defined as follows. The single most significant trait associated SNP was taken as the GWAS association, considered detected if GWAS significance was <5×10−8. The single most significant eQTL in the training set was taken as the eQTL-guided association (eGWAS), and considered detected if GWAS significance was <0.05/15,000. The TWAS association was measured by training the imputation algorithm on the 1,000 held-out samples with expression and imputing into the GWAS summary statistics, and considered detected if significance was <0.05/15,000. The entire procedure was repeated 500 times (5 per gene) and power was estimated by counting the fraction of instances where each method detected the locus. As in the cross-validation analysis, training on the genetic component of expression instead of the overall expression consistently increased TWAS power by ~10% (Supplementary Figure 7). Two null expression models were tested by generating gene expression for the 1,000 held-out samples that was standard normal as well as heritable expression (cis-h2g=0.17) with GWAS Z-scores drawn from the standard normal (Supplementary Table 4). See Supplementary Note for detailed simulation setup.
