Introduction

In utero stress exposure has been associated with adverse offspring mental and physical health outcomes, including internalizing symptoms [1], adiposity [2], asthma, and allergies [3], and has been hypothesized to put children in a disadvantaged position from early life onwards. Differential DNA methylation (DNAm) has been suggested as a putative mechanism underlying these associations, as DNAm has been linked to prenatal exposures such as maternal smoking [4], postnatal psycho-social stress [5], and to child outcomes such as body mass index [6], asthma [7] and cortisol reactivity [8]. Several multi-cohort studies have probed epigenome-wide associations of maternal prenatal stress with offspring DNAm, with varying results [9,10,11]. The largest study to date, including 5496 children from 12 cohorts, reported limited associations for DNAm sites located in genes that have been implicated in neurodegeneration, immune and cellular functions, and epigenetic regulation [10].

A growing body of research, however, shows that environmental factors in interaction with genetic variants are better predictors of DNAm than environmental factors or genetic variants alone, when looking into CpG sites within variably methylated regions (VMRs). For example, Teh, Pan [12] studied genome-wide interactions of 19 prenatal factors, including gestational age, maternal smoking and maternal depression, on highly variable neonatal DNAm sites. For 75% of the sites, DNAm was better predicted by the interaction between genotype and the environment than by either genotype or environment alone. Environment-only was never the best predictor of DNAm in that study. Similarly, a study by Czamara, Eraslan [13] in four cohorts, examining 10 prenatal factors, showed that gene-environment interactions best predicted DNAm in variably methylated regions in 38-60% of analyses, while genotype-only models were best in 11-30% and environment-only models were best in only <1-4%. However, while these studies analyzed which type of model worked best, they did not aim to identify specific genetic variants, environmental variables or DNAm sites. Knowing which genetic variants interact with prenatal stress in relation to DNA methylation would help to better understand the biological pathways underlying the gene-environment effects on health.

We therefore aimed to study genome-wide interactions between genetic variants and cumulative prenatal stress in relation to epigenome-wide DNAm at birth. We also aimed to test the hypothesis that DNAm is better predicted by the interaction of genetic variants and stress than by either factor alone. We used a comprehensive cumulative measure of psycho-social maternal stress during pregnancy, which has previously been related to suboptimal neurodevelopmental, mental, and cardiovascular outcomes [14,15,16]. We meta-analyzed data from two population-based cohorts, the Generation R Study in the Netherlands (Generation R) and the Avon Longitudinal Study of Children and Parents (ALSPAC) in the United Kingdom and followed up associations to study unique stress domain contributions, as well as running mediation analyses of maternal prenatal smoking, alcohol use, gestational age and birth weight. Lastly, we performed enrichment analyses to gain insight into potential biological pathways.

Methods

Setting

We used three non-overlapping datasets from two prospective population-based cohorts: two datasets from Generation R and a third dataset from ALSPAC.

In the Generation R Study, pregnant women residing in the study area of Rotterdam in the Netherlands with an expected delivery date between April 2002 and January 2006 were invited to participate in the study [17]. The Generation R Study is conducted in accordance with the World Medical Association Declaration of Helsinki and has been approved by the Medical Ethics Committee of Erasmus MC, Rotterdam. Informed consent was obtained for all participants.

In ALSPAC, pregnant women resident in Avon, UK with expected dates of delivery between 1st April 1991 and 31st December 1992 were invited to take part in the study [18, 19]. The ALSPAC website contains details of all the data that are available through a fully searchable data dictionary and variable search tool (http://www.bristol.ac.uk/alspac/researchers/our-data/). Ethical approval for the ALSPAC study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees. Informed consent for the use of data collected via questionnaires and clinics was obtained from participants following the recommendations of the ALSPAC Ethics and Law Committee at the time. Consent for biological samples has been collected in accordance with the Human Tissue Act (2004).

Study Population

The full selection procedure is described in the Supplemental Information. In Generation R, 9778 pregnant mothers gave birth to 9749 live-born children and in ALSPAC, the initial number of pregnancies enrolled was 14,541. Participants were selected based on availability of genetic data (nGeneration R = 7502; nALSPAC = 8797), as well as cumulative prenatal stress information (nGeneration R = 5684; nALSPAC = 7483), and DNAm data as measured with the Infinium HumanMethylation450 BeadChip (Illumina Inc., San Diego, CA) in Generation R (GENR 450 K) and ALSPAC (ALSPAC 450 K) or with the Infinium MethylationEPIC v1.0 Beadchip in Generation R (GENR EPIC) (nGENR 450K = 1231; nGENR EPIC = 986; nALSPAC = 793). Additionally, one of each pair of children with cryptic relatedness (IBD > 0.15) were removed, based on data availability or otherwise randomly. As a result, GENR 450 K included 1224 children, GENR EPIC included 949, and ALSPAC 450 K included 790 – for a total of 2963 children.

Genotyping

In the Generation R Study, children were genotyped with the Illumina HumanHap 610 or 660 quad chips. A full description has been published previously [20]. Data were imputed to the 1000 genomes reference panel (Phase 1 version 3). Phasing was done using MACH software, and imputation using Minimac software. The ALSPAC children have been genotyped with the Illumina HumanHap 550 quad chip [21]. The data were imputed to a phased version of the 1000 genomes references panel (Phase 1 version 3) from the Impute2 reference data repository.

In all (sub-)cohorts, we used best-guess genotypes. Quality control was done with PLINK 1.90 [22]. Autosomal variants were selected and variants with SNP call rates of <95%, with evidence for violation of Hardy-Weinberg equilibrium (p < 1×10-07), with a minor allele frequency <5%, or with low imputation quality (Rsq<0.3 in Generation R and info scores <0.8 in ALSPAC, according to local practices [20, 21]) were removed. Insertions, deletions, and multi-allelic positions were also removed. Samples were excluded in the case of sex mismatches, minimal or excessive heterozygosity, or a sample call rate of <97.5%.

This quality control procedure resulted in 5,584,862 SNPs in GENR 450 K; 5,627,497 SNPs in GENR EPIC; and 5,797,754 SNPs in ALSPAC 450 K. To reduce the multiple testing burden, SNPs were pruned based on linkage disequilibrium and haplotype blocks (window size=50 SNPs, step size=5 SNPs, VIF = 2) in the largest subcohort, GENR 450 K, which resulted in 447,713 SNPs. Of these, a final set of 374,152 SNPs was common to all three (sub-)cohorts.

Cumulative prenatal stress

The cumulative prenatal stress score was computed from ~50 stress-related items measured during pregnancy. The full item list and a detailed description of the score calculation can be found elsewhere (https://github.com/SereDef/cumulative-ELS-score [15, 16]). Briefly, in order to maximize data harmonization across cohorts, stress items were selected (based on closest item-similarity), dichotomized (0 = no risk; 1 = risk) and assigned to one of four stress domains: life events (e.g. death of a relative), contextual risk (e.g. financial problems), personal stress (e.g. depression), and interpersonal stress (e.g. family conflict). Stress domain scores (ranging from 0 to 1) were then computed by averaging items within each domain. A total prenatal stress score was obtained by summing all domain scores (range: 0 to 4). Individuals with >50% of all stress items missing were excluded. Missing data were imputed at the individual item level using predictive mean matching with 60 iterations, as implemented by the mice package [23] in R version 4.0 [24]. Within the selected samples of each (sub-)cohort, cumulative prenatal stress scores were standardized. To reduce the influence of extreme outliers, we winsorized values outside the range of (25th percentile - 3*interquartile range (IQR)) to (75th percentile + 3*IQR).

DNA methylation

For both cohorts, DNA extracted from cord blood was bisulfite converted. Samples were processed with the Illumina Infinium HumanMethylation450 BeadChip (Illumina Inc., San Diego, CA) in GENR 450 K and ALSPAC 450 K and with the Infinium MethylationEPIC v1.0 Beadchip in GENR EPIC.

In GENR 450 K and GENR EPIC, the CPACOR workflow [25] was applied for quality control. Arrays with observed technical problems such as failed bisulfite conversion, hybridization or extension as well as arrays with a sex mismatch were removed. Arrays with a call rate >95% per sample were carried forward into normalization.

In ALSPAC 450 K, quality control was done using the meffil package [26] in R version 3.4.3. Samples with mismatched genotypes, mismatched sex, incorrect relatedness, low concordance with samples collected at other time points, extreme dye bias and poor probe detection were removed and carried before normalization.

In order to minimize cohort effects, the data from GENR 450 K and ALSPAC 450 K have been previously normalized as a single dataset [27] and data from GENR EPIC were normalized using the same procedure. Functional normalization was performed (using 10 control probe principle components with slide included as a random effect) with the meffil package in R [26]. Probes were excluded if they had a detection p > 0.01 or low bead count (<3) in >10% of the samples. In total, 472,450 autosomal methylation sites (CpGs) passed these quality control filters in GENR 450 K and ALSPAC 450 K. To reduce the computational burden of the genome-wide analyses on the methylome, only probes that were previously identified as having epigenome-wide significant (p < 1×10-7) inter-individual variation in DNAm at birth in these cohorts were carried forward into analyses, leaving 100,687 CpGs [27]. Of these, cross-reactive probes (n = 14,451 CpGs) were removed [28]. Last, only probes present and passing quality control on both array types were selected, resulting in a final set of 86,236 CpGs (sample-size weighted median MAD-[median absolute deviation]-score [IQR] = 0.04 [0.03-0.06]; range=0.001-0.641). DNAm levels were represented as beta values, indicating the ratio of methylated signal relative to the sum of methylated and unmethylated signal per CpG. To reduce the influence of extreme outlying values, beta values of each CpG outside the range of (25th percentile - 3*interquartile range (IQR)) to (75th percentile + 3*IQR) were winsorized.

Covariates

All models were adjusted for sex of the child as determined at birth, the first 5 genetic principle components, estimated blood cell composition (CD4 + T-lymphocytes, CD8 + T-lymphocytes, natural killer cells, B-lymphocytes, monocytes, granulocytes, and nucleated red blood cells) as based on a cord blood reference panel [29], and DNAm batch effects (25 sample plates in GENR 450 K, 12 sample plates in GENR EPIC, and 20 surrogate variables in ALSPAC 450 K [26, 30]).

Statistical analyses

Analyses were performed with an adapted version [16] of the GEM software package [31] in R [24] version 4.0.5. This package applies large matrix operations allowing for fast analysis of genome-wide SNPs and CpGs and enabling us to test the following models: (i) a GxEmodel – our primary model of interest – in which the interaction effect of each SNP with cumulative prenatal stress was iteratively regressed on DNAm at each CpG site. For comparative purposes, we also tested (ii) a Gmodel, in which each SNP was iteratively regressed on DNAm at each CpG site, and (iii) an Emodel, in which cumulative prenatal stress was iteratively regressed on DNAm at each CpG site. A dominant model was applied to the GxEmodel and Gmodel, meaning that heterozygous and homozygous minor genotypes were contrasted against homozygous major genotypes, in order to make the models more robust against outlying values. The three models were performed in each (sub-)cohort separately, and results were meta-analyzed using inverse-variance weighted fixed effects with METAL [32]. The significance threshold of the meta-analyses was Bonferroni-corrected for the number of tests. For the GxEmodel and Gmodel (nSNPs = 374,152, nCpGs = 86,236: 32,265,371,872 tests) the threshold was set to p < 1.55×10-12, and for the Emodel (86,236 tests), the threshold was set to p < 5.80×10-07. To assess heterogeneity between (sub-)cohort results, the I2 statistic was used, with 75% taken as indication of considerable heterogeneity [33].

Model comparisons

First, we compared the occurrence of cis- versus trans-effects among suggestive findings (p < 5×10-08; i.e. genome-wide threshold) in the Gmodel and the GxEmodel. Second, earlier studies tested SNP-by-environment interactions with CpGs only among SNPs or CpGs for which an association was found in a genotype-only model. We tested whether SNPs/CpGs that showed suggestive associations in the Gmodel or Emodel had a higher chance of being part of a suggestive association in the GxEmodel. Associations in the Gmodel and the GxEmodel with p < 5×10-08 (i.e. genome-wide threshold), and associations in the Emodel with p < 1×10-05 were considered suggestive. We performed enrichment analyses comparing suggestive versus non-suggestive SNPs and CpGs using Fisher’s exact tests (significance threshold: p < 0.05).

As a sensitivity analysis, we performed similar enrichment analyses to assess if suggestive SNPs and CpGs had been identified in an earlier large-scale methylation quantitative trait locus (meQTL) study [34].

Follow-up analyses

Multiple follow-up analyses were performed. Since the Gmodel has been tested extensively previously in search of methylation quantitative trait loci (meQTLs) [21, 34], we only followed-up results from the GxEmodel and Emodel. First, we looked up significant CpGs in the GxEmodel and/or Emodel in the EWAS Catalog for previously reported associations [35]. Second, we looked up genes annotated to significant SNPs and CpGs in the GxEmodel and/or Emodel via phenome-wide association studies (PheWASs) using the online GWAS Atlas tool (https://atlas.ctglab.nl/PheWAS) [36], including 4756 GWASs, using a Bonferroni corrected p < 1.05×10-05. Annotation of SNPs and CpGs was performed using ANNOVAR linking variants reported in the 1000 Genomes Project and SNPdb [37] and the Illumina HumanMethylation450 v1.2 Manifest (Illumina Inc.), respectively. Third, significant associations in the GxEmodel (p < 1.55×10-12) and/or Emodel (p < 5.80×10-07) were followed up with linear regressions containing the effects of the four stress domains (life events, contextual risk, personal stress, and interpersonal stress) in one model to identify unique associations of each of these stress types, independent of the other types. Here, associations with p < 0.05 were interpreted as a unique contribution to the GxE association on DNAm for that stressor. Fourth, significant associations in the GxEmodel and/or Emodel were tested for potential mediation of prenatal stress effects on DNAm by maternal prenatal smoking, maternal prenatal alcohol intake, gestational age, and birth weight (each modeled separately), using the Lavaan package in R [38]. Mediation was deemed to be significant if the AB path (predictor -> mediator, mediator -> outcome) has a p-value below a Bonferroni-corrected threshold of 0.0125 (corrected for the number of mediators). Last, functional enrichment analysis of associated biological pathways was performed with Gene Ontology for models for which results could not be explained by a mediator. Genes annotated to suggestive unique SNPs and CpGs were interrogated using the GOfuncR package [39] in R. The GOfuncR package compares associated pathways of candidate versus background genes using hypergeometric testing and a family-wise error rate correction for multiple testing. Genes annotated to all non-suggestive SNPs and CpGs included in the main analyses were used as the background set.

Mediators

In Generation R, mothers reported on prenatal tobacco smoking and alcohol consumption via questionnaires in the first, second, and third trimester. In ALSPAC, mothers reported via questionnaires on tobacco smoking in the second and third trimester and on alcohol consumption in the first and second trimester. For both cohorts, gestational age at birth was determined using fetal ultrasound examinations or last menstrual period, and birth weight was obtained from midwife and hospital registries.

Results

GENR 450 K, GENR EPIC, and ALSPAC 450K included 49.9%, 47.7%, and 49.1% boys, and mothers were 32.2, 32.0, and 29.7 years old at birth, respectively. After winsorizing, mean cumulative prenatal stress scores were 0.36 (SD = 0.28, min=0.00, max=1.48), 0.44 (SD = 0.35, min=0.00, max=1.48), and 0.51 (SD = 0.28, min=0.00, max=1.69), respectively, with a theoretical maximum score of 4 (Supplemental Fig. 1). There were differences between the (sub-)cohorts, as cumulative prenatal stress was higher on average in ALSPAC than in the GENR subcohorts (specifically contextual risk and personal risk; life events and interpersonal risk were highest in GENR EPIC), gestational age and weight at birth were somewhat lower in ALSPAC, as was maternal age at birth (Table 1).

Table 1 Sample characteristics.
Full size table

GxEmodel: SNP by prenatal stress interactions and DNA methylation

Five SNP-by-prenatal-stress interactions on DNAm were identified after Bonferroni correction, including five unique SNPs and two unique CpGs. Firstly, an association in cis of rs12901653 in CHD2 in interaction with cumulative prenatal stress was found for DNAm at a nearby (64862 bp) cg24317086 (B = -0.026, SE = 0.003, p = 4.07×10-16), for which CHASERR, or CHD2 Adjacent Suppressive Regulatory RNA is the nearest gene. Secondly, trans-associations of 4 SNPs in or near ORC5 in interaction with cumulative prenatal stress were found for DNAm at cg06592260, which is located in EPPK1 (rs7642426: B = -0.013, SE = 0.002, p = 1.12×10-13; rs10279675: B = -0.012, SE = 0.002, p = 2.76×10-13; rs2188287: B = -0.013, SE = 0.002, p = 1.85×10-13; rs10251976: B = -0.012, SE = 0.002, p = 7.32×10-13). The adjusted R2 of the significant SNP-by-prenatal-stress interaction terms ranged between 0.02 and 0.03 and had a median (IQR) of 0.02 (0.02-0.02). As these SNPs were in high LD (>0.9 in all [sub-]cohorts), they were not independent (Supplemental Fig. 2). Results are depicted in Table 2 and Fig. 1. For all associations except the interaction of rs12901653 with cumulative prenatal stress on cg24317086, heterogeneity between (sub-)cohorts was low (I2 = 0.0). Rs12901653 showed considerable heterogeneity (I2 = 93.5), as associations for GENR 450 K and GENR EPIC were negative, whereas it was positive (although p > 0.05) for ALSPAC 450 K (forest plot in Supplemental Fig. 3).

Table 2 Genome- and epigenome-wide associations of SNP by cumulative prenatal stress interactions and DNA methylation.
Full size table
Fig. 1
figure 1

Scatterplots of genome- and epigenome-wide associations of SNP-by-prenatal-stress interactions and DNA methylation.

Full size image

Gmodel: SNPs and DNA methylation

In the Gmodel, after Bonferroni correction, we found 691,202 associations between SNPs and DNAm, including 181,133 unique SNPs and 54,809 unique CpGs. As such, nearly half of all investigated SNPs (48%) could be marked as meQTLs, and more than half (59%) of the examined CpGs are under genetic control. In these results we find evidence of both polygenicity, i.e. multiple SNPs affecting the same CpG, as well as pleiotropy, i.e. the same SNP affecting multiple CpGs. Furthermore, 91% of SNP-CpG associations were in cis, 9% were in trans (distance of >1,000,000 bp; mean [SD] distance 136,731 [170,197] bp). The adjusted R2 of the significant prenatal stress terms ranged between 0.02 and 0.95 and had a median (IQR) of 0.04 (0.02-0.06). For 28% of associations, heterogeneity between (sub-)cohort results was considerable (I2 > 75).

Emodel: Cumulative prenatal stress and DNA methylation

Three DNAm sites at birth were associated with exposure to cumulative prenatal stress after Bonferroni correction, including cg05575921 (B = -0.009, SE = 0.001, p = 3.81 × 10-18), located in AHRR, cg09935388 (B = -0.016, SE = 0.002, p = 2.79 × 10-11) in GFI1, and cg04180046 (B = -0.007, SE = 0.001, p = 6.73 × 10-08) in MYO1G (Table 3; Fig. 2). The adjusted R2 of the significant prenatal stress terms ranged between 0.02 and 0.03 and had a median (IQR) of 0.02 (0.01-0.02). For cg05575921, there was considerable heterogeneity between (sub-)cohort results (I2 = 78.1), for cg09935388 and cg04180046 no heterogeneity was detected (I2 = 0.0; forest plot in Supplemental Figure 4).

Table 3 Epigenome-wide associations of cumulative prenatal stress and DNA methylation.
Full size table
Fig. 2
figure 2

Scatterplots of epigenome-wide associations of cumulative prenatal stress and DNA methylation.

Full size image

Enrichment of cis- and trans- associations

Suggestive findings in the GxEmodel (3327 associations with 3248 unique SNPs and 2613 unique CpGs) and Gmodel (1,088,683 associations with 223,254 unique SNPs and 62,826 unique CpGs) were compared to test whether GxEmodel and Gmodel results differed in distance between associated SNP and CpGs. Among suggestive GxEmodel findings, only 1% was in cis (mean [SD] distance 17,767 [256,557] bp), whereas 89% of suggestive Gmodel was in cis (mean [SD] distance 155,089 [183,495] bp). This difference was significant (OR = 1498.2 [95% CI = 983.0-2915.3], p < 2.23×10-308).

Enrichments of main effect model associations in GxEmodel

Suggestive SNPs and CpGs in the GxEmodel and Gmodel were compared to test whether a suggestive association in the Gmodel increased the chance of a suggestive association in the GxEmodel. This did not seem to be the case, as suggestive SNPs in the Gmodel were as likely to have been identified in the GxEmodel as other SNPs were (1 vs 1%; OR = 0.9 [95% CI = 0.9-1.0], p = 0.13). Similarly, suggestive CpGs in the Gmodel were as likely to be identified in the GxEmodel as other CpGs were (3 vs 3%; OR = 1.0 [95% CI = 0.9-1.1], p = 0.74). As a sensitivity analysis, we also checked for enrichment of SNPs and CpGs associated with meQTLs identified by others [34], and similarly found that suggestive SNPs and CpGs in the GxEmodel were as likely or even less likely to have been linked to an meQTL, whereas suggestive hits in the Gmodel were more likely to have previously been linked to an meQTL (Supplemental Results S1).

Furthermore, suggestive CpGs in the Emodel were as likely to have been identified as a suggestive CpG in the GxEmodel (7%) as other CpGs were (3%; OR = 2.5 [95% CI = 0.1-16.4], p = 0.35), although it should be noted that the number of suggestive findings for the Emodel was low with only 14 CpGs.

CpG look-ups

From the GxEmodel, variation at cg24317086 has been previously associated with gestational age [40], age in childhood [27, 41], tissue type [42], Down syndrome [43], and C-reactive protein levels [44]. Variation at cg06592260 has been associated with age in childhood [27] and tissue type [42].

Full results for the lookup of the previously reported EWAS associations for the three CpGs found in the Emodel can be found in Supplemental Table 1. In brief, variation at cg05575921, cg09935388, and cg04180046 was related to maternal smoking during pregnancy with reported associations stemming from 6, 8 and 9 studies, respectively, and to smoking behavior (not in pregnancy) with reported associations in 27, 19, and 10 studies, respectively. Other associations were found, among others, for age in childhood [27], tissue type [42], alcohol consumption [45,46,47], maternal educational attainment during pregnancy [48], lung function [49,50,51,52] and post-traumatic stress syndrome [53, 54].

Annotated gene look-up

The full results of the PheWASs are depicted in Supplemental Figures 5 to 10. In brief, genetic variants at CHD2 were related to use of sun/UV protection, resting heart rate, free thyroxine levels, educational attainment, measures of body composition, uric acid levels, processed meat intake, pork intake, napping during the day, (standing) height, and schizophrenia. Genetic variants at ORC5 have been related to risky behaviors, left and right entorhinal cortex volume, and drinking behavior. Genetic variants in or close to EPPK1 (annotated to several CpGs of the significant GxEmodels) have been related to resting heart rate, skin tanning, body composition measures, and height.

Genetic variants at AHRR were related to skin colour, hair colour, male balding patterns, ulcerative colitis, hematocrit, hemoglobin, aspartate, fat measures, and height. Genetic variants annotated to GFI1 were related to coronary artery disease, white blood cell measures, fat measures, multiple sclerosis, being a morning person, height, lung function, and asthma, eczema, and allergy related measures. Genetic variants at MYO1G were related to thyroid function, white blood cell measures, and height.

Stress-domain-specific results

In the GxEmodel, none of the individual stress domains (life events, contextual risk, personal risk, interpersonal risk) provided a unique SNP-by-prenatal-stress contribution (p < 0.05) to the association with DNAm, over and above co-occurring domains (Supplemental Table 2). In the Emodel, contextual risk provided a unique contribution to DNAm at cg05575921 (AHRR, B = -0.008, SE = 0.001, p = 4.42x-14), cg09935388 (GFI1, B = -0.015, SE = 0.003, p = 6.18×10-09), and cg04180046 (MYO1G, B = 0.005, SE = 0.001, p = 7.36×10-05). In addition, interpersonal risk provided a unique contribution to cg05575921 (B = -0.003, SE = 0.001, p = 7.53×10-03) and cg09935388 (B = -0.007, SE = 0.003, p = 5.44×10-03). Life events and personal risk did not provide unique contributions in the significant Emodel associations (Supplemental Table 3).

Mediation

In the GxEmodel, none of the significant SNP-by-prenatal-stress associations with DNAm were mediated by maternal tobacco smoking or alcohol consumption during pregnancy, gestational age, or birth weight. In the Emodel, all three cumulative prenatal stress associations with DNAm were mediated by maternal prenatal smoking (cg05575921: Bindirect = -0.007, 95% CI = -0.008;-0.006, p = 4.17×10-60; cg09935388: Bindirect = -0.010, 95% CI = -0.011;-0.008, p = 3.21×10-30; cg04180046: Bindirect = -0.006, 95% CI = -0.005;-0.007, p = 8.94×10-40), and not by any of the other mediators (Fig. 3).

Fig. 3
figure 3

Mediation of cumulative prenatal stress associations with DNA methylation by maternal tobacco smoking during pregnancy (prenatal smoking).

Full size image

Pathway enrichments

A Gene Ontology analysis of 3248 suggestive (p < 5×10-08) SNPs in the GxEmodel yielded 145 overrepresented pathways and 12 underrepresented pathways (Supplemental Table 4). The overrepresented pathways were predominantly linked to neuronal development and synaptic transmission. The underrepresented pathways were linked, amongst others, to DNA repair processes. A Gene Ontology analysis of 2613 suggestive CpGs in the GxEmodel yielded 35 overrepresented pathways (Supplemental Table 5), among which neuronal development-related pathways were predominant.

Discussion

In this study, we investigated SNP-by-prenatal-stress interactions on DNAm at birth, for the first time at the genome- and epigenome-wide level. From the GxEmodel, we report five SNP-by-prenatal-stress interactions on DNAm after multiple testing correction, including five unique, of which two independent, SNPs in CHD2 and ORC5, and two unique CpGs near CHASERR and in EPPK1. By comparison, the Gmodel yielded 691,202 associations of SNPs and DNAm, including 181,133 unique SNPs (48% of investigated SNPs) and 54,809 unique CpGs (59% of investigated CpGs), and the Emodel identified three associations between cumulative prenatal stress and DNAm at CpGs in AHRR, GFI1, and MYO1G, which are known DNAm loci for smoking exposure. Together, these results show that genetic main effects are by far stronger than prenatal stress effects alone or gene-by-prenatal stress interactions.

Significant results for the GxEmodel were scarce, which might in part be explained by the stringent Bonferroni multiple testing correction in which analyses were considered as independent. Due to the scale of the analyses, this resulted in a very low p-value threshold. However, a Bonferroni threshold might be overly stringent as there was a correlational structure among SNPs and among CpGs, and we also tested the same SNP 86,236 times, and the same CpG 374,152 times. Using a suggestive (genome-wide) significance threshold, Gene Ontology analyses pointed to enrichment of neuronal development related pathways. This may indicate that prenatal stress interacts with the child’s genome in many small ways to affect neuronal development, but that these associations only become visible with a less stringent threshold or a more powerful analysis. Moreover, previous GWASs have associated variation in ORC5, in or near which four of the associated SNPs were located with entorhinal cortex volume [55], an area important for memory processing, bordering the hippocampus and particularly rich in corticoid receptors [56]. These findings fit the developmental origins of health and disease perspective [57], which poses that the prenatal environment programs organ structure and function (in this case in interaction with the offspring genotype), as well as findings from our own lab that cumulative prenatal stress is related to childhood subcortical brain volumes [16], and that prenatal stress, beyond postnatal stress, predicts internalizing symptoms in childhood [15].

In contrast to the GxEmodel, significant associations in the Gmodel were abundant, indicating that many common SNPs are involved in epigenetic programming and in turn, that DNAm is under strong genetic control. This confirms earlier meQTL studies, which also identified numerous genetic effects on the epigenome [21, 34]. The abundance of genetic effects and far fewer GxE associations are contrary to the notion put forward by other studies that gene-environment interaction studies perform better than studies of genetic main effects alone in terms of predicting DNAm [12, 13, 58]. This may be explained by the different approaches to statistical inference used by these studies. In previous studies of genetic interaction effects with multiple prenatal environments [12, 13] and with adverse childhood experiences [58], G, E, and GxE models directly were compared for each CpG, using Akaike Information Criterion or R2, without adjustment for multiple testing. In contrast, we did not directly compare the different model fit per CpG. Instead, we compared the amount of Bonferroni-corrected significant results. It might therefore be possible that while GxE terms explain more DNAm variance than G- or E- main effect models do, these GxE effects are not large enough to survive the Bonferroni-correction applied here. However, Czamara et al. Czamara, Eraslan [13] also noted that ‘GxE models appear to be winning by a significantly lager AIC margin over the next best model, when compared to the other types of winning models’. If this had been the case for the current study, one might expect that the GxEmodel would have produced more hits than the Gmodel. Czamara, Eraslan [13] however, looked at a range of prenatal variables, so it may be that our findings are specific to prenatal stress.

What emerges from our results, however, is that the GxEmodel yields different results than when looking at genetic or environmental main effects alone. Follow-up analyses showed that SNPs and CpGs brought forward by the GxEmodel were not more likely to have been identified as, or related to an meQTL. For future GxE studies on DNAm, this means that only testing GxE interactions among significant findings in the Gmodel [12], would reduce the multiple testing burden, but might result in selective findings and may miss true GxE effects.

The Emodel yielded limited evidence of associations between cumulative prenatal stress and DNAm, which is in line with previous studies [9,10,11]. Moreover, whereas associations in the GxEmodel did not seem to be related to prenatal smoking and drinking behavior, gestational age or birthweight, the Emodel associations all could be largely explained by smoking behavior of the mother during pregnancy. Indeed, the look-up of related CpGs showed that these are top-hits in smoking EWASs [59,60,61]. Furthermore, whereas the GxEmodel results could not be explained by one of the types of stressors in particular, thereby ascribing to the notion that associations were due to the cumulative nature of prenatal stress rather than to the unique contribution of a specific stressor, contextual risk and interpersonal risk provided unique contributions to the results from the Emodel. These results fit with a recent EWAS meta-analysis of maternal educational attainment, often taken as an indicator of socio-economic position, which was also enriched for CpGs related to prenatal smoking [48] as well as an EWAS on victimization stress in children, of which results could also largely be explained by smoking [62]. It may be that prenatal stress in population-based samples does not provide enough variation to find true associations with DNAm at birth beyond those related to prenatal smoking, or simply that larger sample sizes are necessary to find small effect sizes. Alternatively, it may be that prenatal stress has limited direct effects on DNAm, but rather that its associations with DNAm are hidden in their dependency on genetic variation, meaning that interaction models are necessary to identify these associations. Our GxEmodel brought forward a similarly low amount of associations with DNAm, but potentially was burdened by a strict multiple testing correction. Again, larger sample sizes would be necessary to overcome the burden of multiple testing. Taken together, we conclude that our Emodel prenatal maternal stress associations with the offspring epigenome are not independent of maternal smoking behavior.

Results of this study should be interpreted in light of several limitations. First, effect sizes were small and Gene Ontology enrichment analysis of suggestive hits seemed to indicate that subthreshold findings are informative – hence larger sample sizes will likely be necessary to identify relevant gene-by-prenatal stress interactions with greater statistical power. Another method to improve statistical power would be to study a cohort with higher, or more variable stress levels, as the occurrence of prenatal stress was relatively low in the population-based samples used in the current study. However, to the best of our knowledge, this is the largest effort thus far to identify gene-by-prenatal stress effects on the epigenome. Moreover, we included a comprehensive measure of cumulative prenatal stress, capturing multiple domains of stress that often co-occur together – which has been uniquely harmonized between Generation R and ALSPAC, making it difficult to include a larger sample size at this time point. Second, there was heterogeneity between the (sub-)cohorts in the GxE association of a SNP in CHD2 with a nearby CpG, which reduces the robustness of the finding. This finding was mainly driven by the results of the Generation R subcohorts, as the result in ALSPAC was nominally non-significant and even in the opposite direction. However, the finding was strong and consistent between the two Generation R subcohorts, that this still resulted in an overall significant finding. Moreover, as the I2 measure that was used for heterogeneity is relatively sensitive, this does not necessarily mean the finding is a false positive. Future studies are needed to examine this association in more detail, preferably in multiple cohorts. Third, whereas SNPs and CpGs included in these models span the full genome, we reduced the number of probes based on intercorrelation and/or variability to minimize the burden of multiple testing. This does mean, however, that it is possible that we missed associations. Fourth, since we took a genome-wide approach, including not only cis- but also trans-associations between SNPs and CpGs, we were computationally constrained and did not, as others [13, 58], include a G+Emodel with SNP and prenatal stress main effects. Given the weak evidence for associations in the Emodel, it is unlikely that a G+Emodel would have produced results too dissimilar from the Gmodel with only SNP main effects. Fifth, DNA methylation is tissue-specific and interactive effects of prenatal stress and offspring genotype on DNA methylation may differ between blood, which we used as an easily accessible tissue in population-based studies, and other, potentially more relevant tissues, such as brain – however, even in blood we found an epigenetic pattern of neurodevelopmental pathways. Sixth, the Generation R and ALSPAC are populations are generally selected towards being slightly healthier and more affluent than the general population, which may affect the generalizability of findings. It may also have reduced variation in prenatal stress and thereby the power to detect true associations. Also, as the epigenetic samples only include children of European ancestry, generalizability to populations of other ancestries may be limited. In the future, studies in populations of other ancestries are necessary to understand how genotype-by-prenatal-stress associates with DNAm at birth across different populations. Last, this is an observational study. Intervention studies on reducing prenatal stress [63,64,65] might help understand the degree to which gene-by-prenatal stress associations are causal in nature. Furthermore, more research would be necessary to understand the consequences of genotype-by-prenatal-stress associations found. The enrichment analyses indicated that neuronal development might be involved, yet more research is necessary to understand for which aspects of neuronal development this would be the case and to what degree.

In conclusion, in this comprehensive study of genotype-by-prenatal stress interactions on DNAm, we report suggestive findings that cumulative prenatal stress interacts with the child’s genome on DNA methylation in or close to genes related to neuronal development. Importantly, we found few significant associations in our environmental main effect model and our gene-by-prenatal stress interaction model, contrasting the many associations identified in our genetic main effect model. These results do not support the idea that gene-environment interactions on the epigenome are more abundant than gene effects alone, at least in the case of prenatal stress and when comparing the number of significant hits between different models. In the future, larger studies and studies including participants of different genetic ancestries are needed to identify associations with smaller effect sizes and generate results that are more generalizable.