Unlike current methods, which focus on individually significant eQTL and SNP associations5,6,8,9,11,13,26,29, our approach captures the full cis-SNP signal and does not require any individual marker to be significant. This is underscored by the fact that TWAS substantially outperformed its cis-eQTL analog both in imputing expression and in association to trait. Our results show that the imputation approach is especially effective when multiple variants influence expression (which in turn influences trait). The large number of new associations identified in real data supports this phenomenon and suggests that it may be a strong contributor to common phenotypes46. Therefore, our approach can be seen as complementary to GWAS by identifying expression-trait associations that are not well explained by individual tagging SNPs. Future work could leverage the difference in performances of TWAS and GWAS to explicitly detect allelic heterogeneity. We note that it is still possible for some loci to have an independent SNP-phenotype and SNP-expression association driven by the same underlying variant though we consider this to be an infrequent biological model.
