takes as input liberally-defined cis-eQTL associations (P < 10−3) and trans-eQTL associations (P < 10−5). For the trans-eQTL, we used a very strict definition to exclude putatively artifactual associations of SNP and gene expression, requiring, in addition to P < 10−5, that the trans-eQTL association also be present in the 206 controls-only Caucasian cohort, albeit with a P value as high as 10−3. Additionally excluding trans-eQTL where the eSNP was within 10 Mb of the associated gene (since such scenarios are perhaps instances of cis-eQTL for regions with larger LD blocks), yielded a final reduced subset of 13,114 trans-eQTL (~7% of all trans-eQTL SNP-gene pairs at P < 10−5) across 661 genes. This stricter filter increases the replication rate in HBCC to 36% at FDR ≤ 5% in both cohorts. For generating null GWAS P values, we used 100 permutations of random case-control assignments of the 2,504 individuals in the 1000 Genomes Phase 3 genotype data (http://www.1000genomes.org) 72, as suggested by the author of the Sherlock software (Xin He, personal communication). We also slightly modified the Sherlock source code, omitting the exclusion criterion for SNPs (and genes whose expression is associated with those SNPs) that could not be found in
