Fig. 1

Illustration of the stepwise Mendelian randomization approach. a In a conventional EWAS, associations observed are potentially confounded (C) and the direction of the association between lipids (L) and DNA methylation (M) cannot be inferred. b Using Mendelian randomization, polygenic scores (P) are used to obtain an unconfounded proxy for lipid levels and, since M cannot influence P, an effect of L on M can be inferred. c An additional analysis is required to exclude a direct effect of P on M (i.e., cis-methylation quantitative trait loci (QTL) effect of polygenic score SNPs) not mediated through L. d Reverse causation, i.e., an effect of M on L, is excluded by evaluating the association of local genetic variation (S) affecting M (cis-methylation QTL) on lipid levels. e Pleiotropic effects are excluded using a multivariate model that incorporates all lipids and their polygenic scores