Elucidating the role of maternal environmental exposures on offspring health and disease using two-sample Mendelian randomization

Abstract

Background: There is considerable interest in estimating the causal effect of a range of maternal environmental exposures on offspring health-related outcomes. Previous attempts to do this using Mendelian randomization methodologies have been hampered by the paucity of epidemiological cohorts with large numbers of genotyped mother-offspring pairs.

Methods: We describe a new statistical model that we have created which can be used to estimate the effect of maternal genotypes on offspring outcomes conditional on offspring genotype, using both individual-level and summary-results data, even when the extent of sample overlap is unknown.

Results: We describe how the estimates obtained from our method can subsequently be used in large-scale two-sample Mendelian randomization studies to investigate the causal effect of maternal environmental exposures on offspring outcomes. This includes studies that aim to assess the causal effect of in utero exposures related to fetal growth restriction on future risk of disease in offspring. We illustrate our framework using examples related to offspring birthweight and cardiometabolic disease, although the general principles we espouse are relevant for many other offspring phenotypes.

Conclusions: We advocate for the establishment of large-scale international genetics consortia that are focused on the identification of maternal genetic effects and committed to the public sharing of genome-wide summary-results data from such efforts. This information will facilitate the application of powerful two-sample Mendelian randomization studies of maternal exposures and offspring outcomes.

Keywords: DOHaD; Developmental Origins of Health and Disease; Fetal Insulin Hypothesis; Maternal effects; Mendelian randomization; birthweight; fetal effects; offspring genetic effects; type 2 diabetes.

Figure 1.

Structural equation model (SEM) used to estimate maternal and offspring genetic effects on birthweight. The three observed variables (in squares) denote the birthweight of a UK Biobank individual (BW), the birthweight of their offspring (BW_O) and their own genotype (SNP). The latent unobserved variables (in circles) represent the genotypes of the individual’s mother (their offspring’s grandmother; G_G) and the genotype of the individual’s first offspring (G_O). The total variance of the latent genotypes for the individual’s mother (G_G) and offspring (G_O) and for the observed SNP variable are set to Φ and are estimated from the data. The causal path between the individual’s own genotype and both their mother and offspring’s latent genotype is set to 0.5. The β_m and βo path coefficients refer to maternal and offspring genetic effects on birthweight, respectively. The residual error terms for the birthweight of the individual and their offspring are represented by ɛ and ɛ_O, respectively, and the variance of both these terms is estimated in the structural equation model. The covariance between the error terms is denoted by ρ. The model can be modified easily to include observed genotypes and/or the absence of one of the birthweight phenotypes.

Figure 2.

Power to detect maternal effects on birthweight as a function of variance explained. We assume a residual correlation of 0.2 between own birthweight and offspring birthweight and that the same locus exerts a maternal effect only. We compare power for N = 50 000 genotyped mother–offspring pairs (i.e. which is an estimate of the current number of available genotyped mother–offspring pairs worldwide that could be conceivably used for these analyses) with the current number of individuals contributing to the UK Biobank and Early Growth Genetics Consortium Analysis of birthweight (i.e. number of genotyped individuals who have data on their own birthweight and their offspring’s birthweight N = 85 518; number of genotyped individuals who have data on their own birthweight only N = 178 980; number of genotyped individuals who have data on their offspring’s birthweight only N = 93 842) at genome-wide significance (α = 5 × 10^–8). Asymptotic power calculations were performed using the ‘Maternal and Offspring Genetic Effects Power Calculator’ (Moen et al., 2019¹¹).

Figure 3.

Estimated maternal and offspring genetic effects on birthweight for 58 autosomal SNPs robustly associated with birthweight. Squares highlight the subset of SNPs that exert their effects predominantly through the mother’s genome (P_maternal < 0.001 and P_offspring > 0.5). Triangles highlight the subset of SNPs with both maternal and offspring genetic effects operating in opposite directions (P_maternal < 0.05 and P_offspring < 0.05); the SNPs with their gene names are those previously associated with type 2 diabetes. The figure is based on data presented in Warrington et al. (2018).

Figure 4.

Effect sizes and standard errors for 51 autosomal birthweight-associated SNPs, which have a minor allele frequency greater than 1%, estimated from a structural equation model using covariance matrices derived from GWAS summary results of own birthweight and offspring birthweight from the UK Biobank Study. Both GWASs used z-scores of birthweight in a subset of unrelated Europeans, after adjusting for ancestry informative principal components and sex for the individual’s own birthweight (sex was not available for the birthweight of the first child in the UK Biobank Study). The x-axis presents results when the sample overlap is known and the y-axis presents results when the sample overlap is estimated using bivariate LD score regression. The phenotypic correlation between own birthweight and offspring birthweight was assumed to be 0.23 (misspecifying this correlation by small amounts i.e. ρ = 0.1–0.3 did not appear to influence estimates nor their standard errors for these data—results not shown).

Figure 5.

Directed acyclic graphs illustrating the core assumptions underlying Mendelian randomization. Assumption (i) requires robust association between the genetic variants and the maternal exposure. Assumption (ii) requires that the genetic variants are uncorrelated with confounders. Assumption (iii) assumes that the genetic variants are only potentially associated with the offspring outcome through the maternal exposure of interest. Offspring genetic variants violate assumption (iii), as they allow a path to offspring outcome that is not through the maternal exposure (iv). However, conditioning on offspring variants (indicated by a box around offspring SNPs) blocks path (iv) and assumption (iii) holds.

Figure 6.

This figure illustrates the four possible ways in which maternal SNPs that are associated with offspring birthweight (conditional on offspring genotype at the same locus) can also be (unconditionally) associated with offspring cardiometabolic disease risk. The ‘X’ represents the effect of conditioning the association analysis on either the offspring or maternal genotype and therefore blocking the path between the conditioned genotype and the other variables of interest. The dashed path with the question mark indicates the potential pleiotropic effects of the offspring’s SNPs on their own cardiometabolic disease risk.