To investigate the potential utility of a transcriptome-wide association (TWAS) based on imputed gene expression we first estimated the cis- (1Mb window around the gene) and trans- (rest of the genome) SNP-heritability ( hg,cis2,hg,trans2) for each gene in our data18,19. These metrics quantify the maximum possible accuracy (in terms of R2) of a linear predictor from the corresponding set of SNPs20,21 (Methods). We used 3,234 individuals for whom genome-wide SNP data and expression measurements were available from METSIM (adipose), YFS (blood), and NTR (blood) data sets22–24 (Methods, Supplementary Table 1). All expression measurements were adjusted for batch confounders, and array probes were merged into a single expression value for each gene where possible (Methods). Consistent with previous work24,25, we observed significantly non-zero estimates of heritability across all three studies, with mean hg,cis2 ranging from 0.01–0.07 and mean hg,trans2 ranging from 0.04–0.06 in genes where estimates converged (Supplementary Figure 1, Supplementary Table 1). Although we observed large differences in the average hg,cis2 estimates between the two blood cohorts, the estimates were strongly correlated across genes (Pearson ρ=0.47 for YFS-NTR, as compared
