limitations (for example, Supplementary Note). For the MR analysis, we opted to perform single-SNP MR instead of multi-SNP MR (such as inverse-variance weighted70), which requires multiple independent associations per gene. As this was the case for only a limited proportion of the tested cis-eQTLs and there were no genes with more than five independent eQTLs, we reasoned this would not provide for reliable inverse-variance-weighted estimation. Inverse-variance-weighted estimation could potentially be applied on a genome-wide trans-eQTL analysis, resulting in many more independent instruments per gene. However, such an approach would be more susceptible to confounding because of horizontal pleiotropy71, where a gene is affected by multiple indirect effects. Finally, our co-localization approach was based on the single-causal-variant assumption, which is not applicable to cis-eQTL genes with multiple independent associations (for example, TREM2; Supplementary Note), and therefore we may have failed to detect co-localizing signals in such loci. Recently published co-localization methods72 do not have this assumption, which may improve future co-localization results. Finally, it is possible that bulk RNA-seq eQTL studies generally capture eQTL effects for genes that are not dosage sensitive and do not cause disruptive downstream consequences73. Furthermore, many eQTLs can only be detected in certain contexts74 for which
