Next, we employed TWAS to identify novel expression-trait associations using summary association statistics from a 2010 lipid GWAS17 (~100,000 samples), i.e. associations more than 500Kb away from any genome-wide significant SNPs in that study. We used all three studies (METSIM, YFS, and NTR;) as separate SNP-expression training panels. We then looked for genome-wide significant SNPs at these loci in the larger 2013 lipid GWAS5 (expanded to ~189,000 samples). We identified 25 such expression-trait associations in the 2010 study (Supplementary Table 7), of which 19/25 contained genome-wide significant SNPs in the 2013 study (P=1×10−24 by hypergeometric test, Methods) and 24/25 contained a more significant SNP (P=1×10−04), a highly significant validation of the identified loci. The validation remained significant after conservatively accounting for sample overlap across the studies (binomial P=3×10−16; Methods, Supplementary Table 7). As a sanity check, we compared direct and summary-level TWAS in the METSIM cohort, and found the two sets of imputed expression-trait Z-scores to be nearly identical, with summary-level TWAS slightly under-estimating the effect (Pearson ρ=0.96, Supplementary Figure 16). Overall, we find the TWAS approach to be highly predictive of robust phenotypic associations.
