Our proposed method shares conceptual similarities with 2-sample Mendelian randomization approaches that aim to identify causal relations between traits using genetic variation predictions as a randomizer43–45. However, while Mendelian randomization is intended to quantify the total causal effect, our method has the less strict goal of identifying significant associations and can operate on summary GWAS data. Importantly, our approach maintains the attractive feature of not being confounded by effects on expression and trait that are independent of the SNPs. Other recent work proposed to leverage summary statistics to estimate the underlying genetic correlation between traits at the genome-wide level37, but cannot be applied locally as it requires multiple loci to estimate standard errors (Methods). Recent work in parallel to ours also proposes gene expression imputation from individual-level data to find expression-trait associations and observes benefits from a reduced multiple-testing burden and increased interpretability16. In contrast, our approach does not require individual GWAS data and is applicable directly to GWAS summary data of very large sample sizes thus increasing discovery power.
