Linking Prenatal Environmental Exposures to Lifetime Health with Epigenome-Wide Association Studies: State-of-the-Science Review and Future Recommendations

Abstract

Background:

The prenatal environment influences lifetime health; epigenetic mechanisms likely predominate. In 2016, the first international consortium paper on cigarette smoking during pregnancy and offspring DNA methylation identified extensive, reproducible exposure signals. This finding raised expectations for epigenome-wide association studies (EWAS) of other exposures.

Objective:

We review the current state-of-the-science for DNA methylation associations across prenatal exposures in humans and provide future recommendations.

Methods:

We reviewed 134 prenatal environmental EWAS of DNA methylation in newborns, focusing on 51 epidemiological studies with meta-analysis or replication testing. Exposures spanned cigarette smoking, alcohol consumption, air pollution, dietary factors, psychosocial stress, metals, other chemicals, and other exogenous factors. Of the reproducible DNA methylation signatures, we examined implementation as exposure biomarkers.

Results:

Only 19 (14%) of these prenatal EWAS were conducted in cohorts of 1,000 or more individuals, reflecting the still early stage of the field. To date, the largest perinatal EWAS sample size was 6,685 participants. For comparison, the most recent genome-wide association study for birth weight included more than 300,000 individuals. Replication, at some level, was successful with exposures to cigarette smoking, folate, dietary glycemic index, particulate matter with aerodynamic diameter $<10\mathrm{\mu }\mathrm{m}$ and $<2.5\mathrm{\mu }\mathrm{m}$, nitrogen dioxide, mercury, cadmium, arsenic, electronic waste, PFAS, and DDT. Reproducible effects of a more limited set of prenatal exposures (smoking, folate) enabled robust methylation biomarker creation.

Discussion:

Current evidence demonstrates the scientific premise for reproducible DNA methylation exposure signatures. Better powered EWAS could identify signatures across many exposures and enable comprehensive biomarker development. Whether methylation biomarkers of exposures themselves cause health effects remains unclear. We expect that larger EWAS with enhanced coverage of epigenome and exposome, along with improved single-cell technologies and evolving methods for integrative multi-omics analyses and causal inference, will expand mechanistic understanding of causal links between environmental exposures, the epigenome, and health outcomes throughout the life course. https://doi.org/10.1289/EHP12956

Introduction

Prenatal Exposures That Impact Health and Epigenetics

Environmental exposures, including chemicals and other exogenous factors, are prevalent and heterogenous. The United States produces and uses more than 85,000 different chemicals,¹ including more than 100 regularly detected in pregnant people’s biospecimens.² Levels of numerous chemicals are higher among pregnant Black and Hispanic people and those of lower educational attainment,³ representing an equity issue. Many chemicals cross the placenta, influencing fetal cellular function (Figure 1A). The prenatal period of rapid and exquisitely timed cellular differentiation and expansion confers heightened susceptibility to exposures. For example, in utero exposure to lead impairs neurodevelopment,⁷ arsenic is associated with impaired lung function and increased cardiovascular mortality,⁸ cigarette smoke causes reduced birth weight and reduced infant lung function,⁹ and folic acid deficiency causes neural tube defects.¹⁰ Exposures during pregnancy may impact the developing embryo/fetus with consequences for postnatal life,¹¹ known as the Developmental Origins of Health and Disease (DOHaD) hypothesis.¹² Investigating the totality of diverse prenatal exposures, termed the exposome,¹³ with later health and disease is an evolving area of etiological, prevention, and policy research. Specific and accurate retrospective markers of exposures experienced years prior to health outcomes are unfortunately rare, limiting available data for most prenatal exposures to prospective collection efforts.

Figure 1.

(A) Sources of prenatal exposures can include (clockwise from top left) dietary intake from food and supplements, pollution, psychosocial stress, and smoking and alcohol drinking behaviors. Exposures enter the pregnant person’s body typically through inhalation, ingestion, or dermal absorption. Once in the body, exposures are distributed and metabolized. They can interact with DNA in various tissues, including placental and fetal tissues. These are the major categories of exposure during pregnancy that were assessed for association with DNA methylation in this review article. (B) Common sources of information for exposure assessment include questionnaires, spatial sensors linked to participant residential history generally followed by exposure modeling, as well as blood- and urine-based exposure biomarkers. These methods are used by environmental health scientists to quantify individual levels of exposure to chemicals and to estimate the relevant time frame the exposure measure is applicable. (C) When we test for a relationship between an environmental exposure and DNA methylation, typically using multivariable regression methods, this is a test of association. Association tests do not indicate causation,⁴ because causal inference requires additional assumptions.⁵ In the case of environmental exposures and DNA methylation, a reverse causation scenario would mean the DNA methylation influences the exposure level. Although it is possible that for a given exposure level, differential methylation at genes essential for metabolism could influence measured concentrations of the contaminant (internal dose), methylation is unlikely to be causally related to being exposed. For this reason, reverse causation is less of a concern in studies of exogenous environmental exposures than in EWAS of disease or physiologic traits.⁶

Epigenetic marks may represent one such useful persistent biomarker of exposure, and potentially a mechanistic link between exposures and lagged health effects. Epigenetic mechanisms are essential for normal development: Errors in epigenetic processes can result in serious developmental disorders.¹⁴ DNA methylation, a type of epigenetic mark, is typically observed at cytosine residues upstream of guanine residues (CpG sites). Although other molecules, such as histones, can be methylated, henceforth we refer to DNA methylation as “methylation.” During normal reproduction and development, two major waves of methylation reprogramming occur shortly after fertilization and during gametogenesis.¹⁵ Cells maintain their differentiated lineage in part because of their methylation patterns.¹⁶ Prenatal environmental exposures can leave epigenetic marks.

Environmental epigenetics was pioneered through clever model system research.¹⁷ For example, in the Agouti mouse model,¹⁸ prenatal exposures including bisphenol A,¹⁹ dietary methyl donors,²⁰ and lead²¹ caused persistent methylation changes in offspring, demonstrating that diverse prenatal environmental exposures can impact offspring epigenetics. In humans, the Dutch Hunger Winter is a classic example linking prenatal exposures to postnatal health and methylation. Nazi blockades of food supplies produced famine in the Netherlands during the years 1944–1945. People pregnant then bore children with increased rates of cardiovascular, metabolic, and psychiatric disorders.²² Timing and degree of prenatal exposure to this famine were also associated with adult methylation differences.²³ Early-life epigenetic changes that persist into later life may provide biomarkers of prenatal exposures and/or a potential mechanism of later health effects.

Scope of This State-of-the-Science Review

This review of the current state of the science is motivated by exciting recent advances in both exposure and methylation assessment investigations in the field of prenatal environmental epigenetic epidemiology. Environmental exposures during pregnancy are estimated through multiple methods with varying accuracy, reliability, and cost (Figure 1B). The availability of high-throughput, low-sample-input arrays with reasonable epigenomic coverage²⁴ enables measurement of methylation in pregnancy cohorts.^25,26 Epigenome-wide association studies (EWAS) using these arrays test for associations between hundreds of thousands of individual CpG sites and environmental exposures (Figure 1C). Promising findings from groundbreaking studies in modest sample sizes have led to large, international, and collaborative consortia, which facilitate rigor and reproducibility of EWAS findings.^26,27 Recent mapping of the epigenome across tissues and cell types,²⁸ more frequent availability of genomic or other omic data in the same samples, and evolving bioinformatic tools are enabling stronger biological inferences on EWAS findings. Chemical exposure assessment and statistical methods advances are starting to allow evaluation of many environmental exposures simultaneously to comprehensively assess the exposome.^29,30 Together, these efforts are increasing the identification of reproducible epigenetic effects of the prenatal environment.³⁰

Here, we review the state of the science on associations between the prenatal environment and methylation. We take an epidemiological perspective to evaluate population-level studies. First, we summarize literature on perinatal environmental EWAS meta-analyses and studies with attempted replication. Consortium efforts have identified reproducible methylation signatures of some prenatal exposures, most notably smoking. Second, we evaluate the implementation of exposure biomarkers for those perinatal exposures with robust and reproducible methylation signatures, a major advance for identifying health impacts of exposures going forward. Third, we provide epidemiological considerations that can impact the power to detect associations in the EWAS framework. Fourth, we make recommendations for new cohorts considering prenatal exposure EWAS. Fifth, we discuss putative mechanisms linking exposures to methylation and challenges in determining their role in exposure-related health outcomes. Finally, we describe areas of needed future development and recommend approaches for implementing emerging tools to provide insights to advance the field. With these analyses and recommendations, this article illustrates the current strengths and weaknesses of perinatal environmental epigenetics research and highlights its future potential.

Methods

We retrieved EWAS of prenatal exogenous exposures with designated search terms identified via PubMed, Web of Science, and Embase searches through 10 May 2022 (search terms in Supplemental Material, “Supplementary Methods”). Because a global term such as environmental exposures may fail to capture specific environmental exposures, we included a broad collection of search terms. Abstracts for all retrieved studies were screened using specified inclusion and exclusion criteria based on the research question of interest. To be included, studies had to be conducted in human populations evaluating the relationship between prenatal exogenous exposures and epigenome-wide methylation. To avoid including duplicated studies, we included only original research articles, excluding reviews and commentaries. Because we focused on epigenome-wide studies, we excluded studies that reported only associations with summary DNA methylation measures (e.g., global methylation, epigenetic clocks) or candidate CpGs, genes, or regions. Given our focus on exogenous exposures, we excluded EWAS evaluating nonexternal exposures, for example, maternal health constructs such as maternal depression, anxiety, or body mass index. All studies meeting inclusion criteria after abstract screening were briefly reviewed to determine the following key characteristics: the genetic ancestry of the population; the exposure; the tissue and method of DNA methylation measurement; the sample size; the number of significant CpGs, genes, or regions reported; the method of correcting for multiple testing; and the use of meta-analysis or replication. Abstract screening and brief review were conducted by F.B., with oversight by K.M.B. and S.J.L. In high-dimensional analyses such as EWAS, the myriad statistical tests require large sample sizes to limit false positive findings, which can also be minimized by replication testing in independent studies or meta-analyses across studies, which also provides evidence of consistency. From the earliest days of genome-wide association studies (GWAS), attempted replication has been regarded as an essential design element.^31,32 Although replication does not guarantee causation, it provides stronger statistical evidence for association, makes artifactual association less likely, and improves effect estimates. Associations might not replicate from one population to another for reasons related to true sources of heterogeneity, including genetic ancestry and exposure distribution differences, and these cannot be evaluated without attempted replication.^31,32 Therefore, as a final inclusion criterion, we restricted in-depth review to only studies that were either meta-analyses or single studies that attempted replication in an independent study (regardless of replication success).

Results

Our literature search identified 360 peer-reviewed publications. After excluding studies with only global or candidate methylation positions or examining nonexogenous exposures (Figure S1), we performed further review for 134 exogenous exposure EWAS conducted from 2012 to 2022. The vast majority ($n=128$) of these EWAS used the Illumina450K or EPIC arrays, and most measured DNA methylation in cord blood, peripheral blood, or placental tissue. Most of these EWAS were conducted in European or primarily European ancestry populations (Figure 2A–B). Meta-analyses or replication attempts were slightly more common among EWAS conducted in European or primarily European ancestry populations (43%) than in other ancestries (30%), likely reflecting greater availability of European ancestry cohorts in which to perform replication analyses, a situation that was also observed in GWAS.³¹ Sample size of EWAS varied across exposure categories, with the largest EWAS for prenatal maternal cigarette smoking ($n=\mathrm{6,685}$), but with many EWAS reporting sample sizes in the hundreds (Figure 2C; Table S1). Of 134 reviewed studies, only 19 (14%) have been conducted in cohorts of 1,000 or more individuals. Some studies with smaller sample sizes that did not attempt replication reported high numbers of significant differentially methylated positions (Figure 2D). Of the 134 reviewed studies, 51 reported a replication attempt or were a meta-analysis (Table 1). These studies were reviewed in depth and summarized below.

Figure 2.

(A) The number of epigenome-wide association studies of external exposures identified in this review ($n=134$ studies; one study appears twice in bar graphs because performed replication for one exposure tested but not the other) by year of publication, reported genetic ancestry, and replication or meta-analysis status. (B) The number of epigenome-wide association studies of external exposures with meta-analysis or replication ($n=51$ studies reviewed in depth) by reported genetic ancestry. (C) Sample sizes of prenatal environmental epigenome-wide association studies ($n=134$; studies involving multiple exposures represented as multiple points) by exposure category (x-axis) among meta-analyses (red squares), studies incorporating replication (blue triangles), and other studies (gray dots). Midline: median sample size, lower whisker: smallest sample size that is greater than or equal to the 25% – $1.5\times$ IQR, upper whisker: largest sample size that is less than or equal to the $75\%+1.5\times$ IQR, box limits: 25% and 75%; these statistics are reported in Table S1. (D) The number of significant differentially methylated positions reported in each study graphed against the sample size of each study and faceted by the category of exogenous exposure considered. Only studies using EPIC or 450K arrays that reported significance at either a $\text{FDR}<0.05$ or Bonferroni correction represented ($n=106$ of original 134 studies, some studies represented $>1$ time, because they tested multiple exposures). The y-axis range varies across panel of exposure category. Data used to generate Figure 2 is provided in Table S2. Note: FDR, false discovery rate; IQR, interquartile range.

Table 1.

Prenatal environmental exposure epigenome-wide association study literature review results. Studies ($n=51$) were included if they performed replication or meta-analysis.

Reference	Cohort/consortia	$n$ a	Ancestry	Exposure	Exposure tissue	Epigenetic measure	Epigenetic tissue	Significant DMPs	Criteria for significance	Validation type	Replication successful yes/no
Cigarette smoking
Joubert et al.33	PACE	6,685	Multi – primarily European	Prenatal maternal smoking	NA	450K	Cord blood	FDR – 6,073; Bonferroni - 568	Bonferroni	Meta-analysis & replication	Yes, Kolmogorov $p\text{\hspace{0.17em}value}<2.2\times {10}^{-16}$ for replication look-up in older children’s blood
Sikdar et al.34	PACE	5,648	Multi – primarily European	Prenatal maternal smoking	NA	450K	Cord blood	5,547	$\text{FDR}<0.05$	Meta-analysis	NA
Everson et al.35	PACE	1,700	Multi – primarily European	Prenatal maternal smoking	NA	450K	Placenta	443	Bonferroni	Meta-analysis	NA
Hannon et al.36	MINERVA	1,263	European	Prenatal maternal smoking	NA	450K	Blood spots	110	Bonferroni	Meta-analysis & replication	Yes, 102 of 110 DMPs, sign test $p<2\times {10}^{-323}$
Vives-Usano37	HELIX	1,105	Multi – primarily European	Prenatal maternal smoking	NA	450K	Child peripheral blood	41	$\text{FDR}<0.05$	Replication	Yes, 17 of 18 loci
Joubert et al.38	MoBa	1,062	Multi – primarily European	Prenatal maternal smoking	NA	450K	Cord blood	26	Bonferroni	Replication	Yes
Rotroff et al.39	MoBA	1,062	European	Prenatal maternal smoking	Maternal plasma (cotinine)	450K	Cord blood	27	Bonferroni	Replication	Mixed, 9 of 15 genes
Markunas et al.40	Norway facial clefts study	889	European	Prenatal maternal smoking	NA	450K	Heel prick blood spots	185	$\text{FDR}<0.05$	Replication	Mixed, 22 of 26 DMPs
Küpers et al.41	GECKO Drenthe	255	European	Prenatal maternal smoking	NA	450K	Cord blood	35	$\text{FDR}<0.05$	Meta-analysis & replication	Mixed, 3 of 8 DMPs
Morales et al.42	INMA	179	European	Prenatal maternal smoking	Survey and maternal urine cotinine	450K	Placenta	50	$\text{FDR}<0.05$	Replication	Yes, 4 of 4 DMPs
Shorey-Kendrick et al.43	NA	59	Multi – primarily European	Prenatal maternal smoking	NA	EPIC	Placenta	726	$\text{FDR}<0.05$	Replication	Mixed, 105 of 726 DMPs
Howe et al.44	MACHS	20	European	Prenatal maternal Smoking	NA	Whole-Genome bisulfite sequencing	Cord blood $\text{CD}4(+)$	10,381	$\text{FDR}<0.05$	Replication	Mixed, 4 of 9 DMRs Replicated in adult $\text{CD}4+\text{\hspace{0.17em}cells}$
Maternal alcohol and diet
Sharp et al.45	PACE	3,075	Multi – primarily European	Maternal alcohol	NA	450K	Cord blood	0	$\text{FDR}<0.05$	Meta-analysis	NA
Küpers et al.46	PACE	2,802	Multi – primarily European	Mediterranean diet	NA	450K	Cord blood	1	$\text{FDR}<0.05$	Meta-analysis	No, 0 of 1 DMPs
Joubert et al.47	PACE	1,988	European	Folate	NA	450K	Cord blood	48	Bonferroni	Meta-analysis	NA
Gonseth et al.48	CCLS	343	Multi – primarily European	Folate	NA	450K	Blood spots	4	Permutation-based $P$ values test	Replication	Mixed, 3 of 4 DMPs
Amarasekera et al.49	NA	23	European	Folate	Maternal serum in last trimester of pregnancy	450K	Cord blood ($\text{CD}4(+)$ T-cells & antigen-presenting cells)	No DMPs/DMPs reported, only DMR	NA	Replication	Mixed, DMR only
Suderman et al.50	MoBa; ALSPAC	1,416	European	25-hydroxyvitamin D	Maternal plasma	450K	Cord blood	0	$\text{FDR}<0.05$	Meta-analysis	NA
Tauebert et al.51	INMA and Generation R	1,286	European	Iron	Maternal serum; cord blood	450K	Cord blood	3	$\text{FDR}<0.05$	Meta-analysis	NA
Størdal et al.52	MoBA	1,062	European	Iron supplements	NA	450K	Cord blood	0	Bonferroni	Replication	NA
Caramaschi et al.53	ALSPAC	641	European	B12	Genotype (proxy)	450K	Cord blood	3	$\text{FDR}<0.05$	Replication	Mixed, 1 of 3 DMPs
Geraghty et al.54	ROLO	60	European	Glycemic index diet intervention	NA	EPIC	Cord blood serum	0	$\text{FDR}<0.05$	Replication	NA
Yan et al.55	NA	24	Asian	Glycemic index dietary intervention	NA	450K	Placenta	365	$\text{FDR}<0.05$	Replication	Mixed, 4 of 10 DMPs
Air pollution
Gruzieva et al.56	PACE	1,949	Multi – primarily European	${\mathrm{PM}}_{10}$	NA	450K	Cord blood	6	$\text{FDR}<0.05$	Meta-analysis & replication	Mixed, 0 of 6 loci in newborns, 4 of 6 DMPs in older children
Isaevska et al.57	Piccolipin	384	European	${\mathrm{PM}}_{10}$	NA	EPIC	Cord blood	284	$\text{FDR}<0.05$	Replication	Mixed, 10 of 151 DMPs tested
Gruzieva et al.56	PACE	1,551	Multi – primarily European	${\mathrm{PM}}_{2.5}$	NA	450K	Cord blood	14	$\text{FDR}<0.05$	Meta-analysis & replication	No, 0 of 14 DMPs
Plusquin et al.58	ALSPAC and EXPOsOMICS (ENVIRONAGE, INMA, Piccolipiù and Rhea)	850	Multi – primarily European	PM	NA	450K	Cord blood; peripheral blood	0 – cross-sectional; 1 – longitudinal	$\text{FDR}<0.05$	Meta-analysis	NA
Breton et al.59	NA	240	Multi – primarily European	Air pollution PM	NA	450K	Blood spots	31	$\text{FDR}<0.15$	Replication	Mixed, 1 of 2 DMPs
Gruzieva et al.60	PACE	1,508	Multi – primarily European	${\mathrm{NO}}_2$	NA	450K	Cord blood	3	$\text{FDR}<0.05$	Meta-analysis & replication	Mixed, 1 of 3 DMPs
Peng et al.61	Project Viva	482	Multi – primarily European	Proximity to roadways	NA	450K	Cord blood	4	$\text{FDR}<0.05$	Replication	No, 0 of 4 DMPs
Metals and other elements
Lozano et al.62	PACE	1,462	Multi – primarily European	Methylmercury	Maternal blood; cord blood; maternal hair	450K and EPIC	Cord blood; childhood peripheral blood	2	$\text{FDR}<0.10$	Meta-analysis	No, 0 of 2 DMPs
Bakulski et al.63	Baltimore THREE	141	Multi – primarily African	Mercury	Cord blood	CHARM 2.0	Cord blood	0	$\text{FDR}<0.10$	Replication	Yes, 1 of 1 DMR region replicated
Tian et al.64	NHBCS and RICHS	484	Multi – primarily European	Selenium	Placenta	450K	Placenta	5	$\text{FDR}<0.05$	Meta-analysis	NA
Everson et al.65	NHBCS and RICHS	484	Multi – primarily European	Cadmium	Placenta	450K	Placenta	3	$\text{FDR}<0.05$	Meta-analysis	NA
Park et al.66	MOCEH	384	Asian	Cadmium	Maternal venous blood; cord blood	EPIC	Cord blood	2	$\text{FDR}<0.05$	Meta-analysis	NA
Gliga et al.67	NA	71	Asian	Cadmium	Maternal blood and child urine (age 9 y)	450K	PBMCs – cord blood and peripheral blood	Cord blood – 458; child blood age 9–6 y	$\text{FDR}<0.05$	Replication	Yes, 1 of 1 DMR tested
Kennedy et al.68	NA	447	European	Copper	Placenta	450K	Placenta	0	Bonferroni	Meta-analysis	NA
Park et al.69	NA	364	Asian	Lead	Maternal blood and cord blood	EPIC	Cord blood	0 (111 in male-only analysis)	$\text{FDR}<0.05$	Meta-analysis	NA
Wu et al.70	Project Viva	268	Multi – primarily European	Lead	Maternal blood	450K	Cord blood	4	$\text{FDR}<0.05$	Replication	No
Bozack et al.71	NA	120	Asian & admixed/Latino	Arsenic	Natural experiment and water samples	EPIC & 450K	Blood; buccal cells	PBMCs – 1; PBMCs & buccal cells – 3	$\text{FDR}<0.05$	Meta-analysis	NA
Kaushal et al.72	NA	64	Asian	Arsenic	Maternal urine	450K	Cord blood	579	$\text{FDR}<0.05$	Replication	No, 0 of 553 DMPs tested
Bozack et al.73	NA	44	Asian	Arsenic	Drinking water; maternal toenails	450K	Cord blood	380	$p<1\times {10}^{-6}$	Replication	Mixed, 2 of 3 genes tested using pyrosequencing
Zeng et al.74	NA	24	Asian	E-waste exposure (heavy metals)	Prenatal maternal blood (validate heavy metal exposure)	450K	Cord blood	125	$p<0.05$	Replication	Yes, 2 of 2 DMPs
Other chemicals
Liu et al.75	HOME	266	Multi – primarily European	PFAS	Maternal serum	EPIC	Cord blood; peripheral blood at 12 y of age	2 DMPs for PFOS, 12 for PFOA, 8 for PFHxS, and 413 for PFNA	$\text{FDR}<0.05$	Replication	Mixed, 6 of 315 DMPs
Miura et al.76	Hokkaido study	190	Asian	PFAS	Maternal blood	450K	Cord blood	4	$\text{FDR}<0.05$	Replication	No, 0 of 4 DMPs
Miura et al.77	Hokkaido study	203	Asian	di-2-ethylhexyl phthalate	Maternal blood	450K	Cord blood	2	$\text{FDR}<0.05$	Replication	No, 0 of 2 DMPs
Vilahur et al.78	NA	181	European	Xenoestrogens	Placenta	450K	Placenta	0	$\text{FDR}<0.05$	Replication	No, 0 of 2 DMPs
Miura et al.79	Hokkaido study	Male – 123; female – 154	Asian	BPA	Cord blood	450K	Cord blood	Male – 27; female – 16	$\text{FDR}<0.05$	Replication	No, 0 of 14 DMPs in females, 0 of 26 in males
McCabe et al.80	MMIP	69	Multi – primarily European	BPA	Maternal urine	EPIC	Umbilical cord blood leukocytes	38	$\text{FDR}<0.05$	Replication	No, non-significant and slight correlation between effect estimates of DMPSs in independent study
Yu et al.81	NA	24	Asian	DDT pesticide	Cord blood	450K	Cord blood	1,131	$p<0.05$	Replication	Yes, 2 of 2 DMPs
Social stressors
Alfano et al.82	ALSPAC	973	European	Socio-economic position	NA	450K	Cord blood; whole blood	Birth:4/childhood:0/adolescence:20	$\text{FDR}<0.05$	Replication	No, 0 of 4 DMPs
Other exposures
Caramaschi et al.83	ALSPAC and MoBa	2,644	European	Medically assisted reproduction	NA	450K	Cord blood	5	$\text{FDR}<0.05$	Meta-analysis	No, 0 of 5 DMPs

Note: ALSPAC, Avon Longitudinal Study of Parents And Children; BPA, bisphenol A; CCLS, California Childhood Leukemia Study; DDT, dichlorodiphenyltrichloroethane; DMP, differentially methylated position; DMR, differentially methylated region; ENVIRONAGE, ENVIRonmental influence ON early AGEing; FDR, false discovery rate; HOME, Health Outcomes and Measures of the Environment; GECKO, Groningen Expert Center for Kids with Obesity; HELIX, Human Early Life Exposome; HOME, Health Outcomes and Measures of the Environment; INMA, INfancia y Medio Ambiente; MACHS, Moderate Alcohol and Cardiovascular Health; MINERVA, MINimizE Right Ventricular pacing to prevent Atrial fibrillation and heart failure; MMIP, Michigan Mother and Infant Pairs Cohort; MoBa, The Norwegian Mother & Child Cohort Study; MOCEH, Mothers and Children’s Environmental Health; NHBCS, New Hampshire Birth Cohort Study; NA, Not Applicable; PACE, Pregnancy And Childhood Epigenetics consortium; PBMCs, peripheral blood mononuclear cell; PFAS, perfluoroalkyl substances; PFHxS, perfluorohexane sulfonate; PFNA, perfluorononanoate; PFOA, perfluorooctanoate; PFOS, perfluorooctane sulfonate; PM, particulate matter; RICHS, Rhode Island Child Health Study; ROLO, Randomised cOntrol trial of LOw glycaemic index diet to prevent macrosomia; THREE, Tracking Health Related to Environmental Exposures.

^a Largest $n$ reported when multiple exposure tissues, sum of all study $n$ for meta-analyses.

Maternal Smoking

Given smoking’s known impacts on offspring health,⁹ most birth cohort studies collect some smoking data, mostly via questionnaires. Thus, smoking EWAS were among the most frequent and largest of the EWAS we identified. Smoking EWAS have successfully identified robust and well-replicated methylation differences in newborn blood (Table 1), with larger studies revealing more differentially methylated CpGs (DM-CpGs). Smoking EWAS have considered both timing of exposure and of DNA methylation measurement. Changes to methylation appear to reflect sustained smoking across the pregnancy, not smoking limited to early pregnancy.⁸⁴ Many methylation differences observed at birth persisted into later childhood.^33,85 Substantial overlap existed in DM-CpGs from EWAS meta-analyses of sustained prenatal exposure to maternal cigarette smoking in newborns³³ and personal current cigarette smoking in adults.⁸⁶ However, many genes were uniquely differentially methylated in newborns,³⁴ underscoring the importance of timing on exposure influences and epigenetic plasticity.

The placenta provides oxygen, nutrients, and hormones to the fetus. Placental abnormalities have been linked to later health outcomes in offspring.⁸⁷ Maternal blood perfuses the placenta, exposing it to circulating environmental chemicals. An EWAS meta-analysis ($n=\mathrm{1,700}$, seven studies) identified widespread impacts of prenatal smoking on placental methylation.³⁵ Overlap with the most prominent DM-CpGs in newborn blood was minimal; reasons for this tissue specificity are unclear. Similar to results in newborn blood, placental findings were much more extensive for sustained smoking than for any smoking in pregnancy, supporting the importance of exposure timing.

Maternal Diet

Maternal nutrition can impact fetal development and child health.⁸⁸ Assessing food and nutrient intake requires detailed dietary assessments that are less frequently collected than smoking information. Some nutrients are more reliably assessed by blood measurements than by questionnaires, and these more expensive measures requiring appropriate samples are available in even fewer birth cohorts. Thus, fewer EWAS of maternal prenatal diet with smaller sample sizes have been conducted to date than smoking EWAS (Table 1).

Folic acid supplementation in pregnancy is associated with fewer neural tube defects.¹⁰ Widespread differential methylation by maternal blood folate levels was reported in an EWAS meta-analysis.⁴⁷ Prenatal alcohol consumption,⁸⁹ vitamin D levels,⁹⁰ iron levels, and Mediterranean diet⁹¹ are also associated with fetal and child health outcomes. However relatively large EWAS meta-analyses (n range 1,062–3,075) of these exposures have found few (range: 0–3) DM-CpGs.^45,46 For most nutrients and dietary intakes measured by questionnaires, substantial measurement error in dietary assessment will require large sample sizes to adequately test for reproducible associations.

Environmental Contaminants

Prenatal air pollution exposure has been associated with shorter gestation and impaired fetal growth.⁹² EWAS meta-analyses of ${\mathrm{NO}}_2$ and particulate matter (n range 850–1,949) reported between 1 and 14 DM-CpGs (Figure 2). As with diet, air pollution is heterogeneous and generally measured differently and with error across locations, decreasing power.

Prenatal exposures to metals have been associated with adverse offspring health outcomes.⁹³ The largest metals EWAS meta-analysis ($n=\mathrm{1,462}$) of prenatal methylmercury exposure and cord blood methylation reported 2 CpGs [$\text{false discovery rate\hspace{0.17em}}\left(\text{FDR}\right)<0.10$]; neither replicated. EWAS meta-analyses of other metals and elements have been small ($n$ range 120–484) and reported few significantly differentially methylated sites (range 0–5). Various other prenatal chemical contaminants (see Table 1) have been examined in EWAS of newborn blood or placenta in single studies, without successful replication.

Social Stressors

Psychosocial stressors before and during pregnancy can impact offspring health.⁹⁴ One EWAS of maternal socioeconomic position in a United Kingdom cohort ($n=914–974$) identified differential methylation in blood at 4 CpGs at birth, none in childhood, and 20 in adolescence, but no sites replicated.⁸² Exposure to psychosocial stressors can be challenging to harmonize across studies. Very large sample sizes will likely be required to find true associations. An ongoing Pregnancy And Childhood Epigenetics consortium (PACE) EWAS project is examining carefully harmonized prenatal stress measures across multiple cohorts.

Other Exposures

A large meta-analysis ($n=\mathrm{2,644}$) comparing children conceived using medical assistance to those conceived without assistance identified 5 DM-CpGs; none successfully replicated.⁸³

Summary of Prenatal EWAS Meta-Analyses and Potential Application as Biomarkers

When extensive, replicable, differentially methylated sites are available, a supremely useful application is the development of methylation biomarkers for prenatal exposures.⁹⁵ (See Figure 4) For example, robust methylation signals of exposure to prenatal cigarette smoke enabled creation of methylation biomarkers of prenatal smoke exposure in newborns,⁹⁸ older children,⁹⁹ and young adults.⁹⁹ For methylation biomarker generation, it is important to have a gold standard of exposure. In studies of prenatal smoke exposure, measurements of cotinine levels, a biomarker of recent smoking, enable identification of pregnant individuals who falsely report as nonsmokers.⁹⁸ Objective methylation biomarkers are a major advance for identifying exposure health effects.⁹⁵ For example, on questionnaires, some smokers falsely deny smoking, with more pronounced underreporting in pregnant persons,¹⁰⁰ leading to information bias. Smoking data may also be missing, and missingness is likely nonrandom—another source of information bias. Methylation exposure biomarkers avoid these sources of bias. In addition, methylation biomarkers in offspring offer information on both dose and duration of smoking across pregnancy that are difficult for respondents to accurately report.¹⁰¹ Epidemiological questionnaires often only considered any smoking during pregnancy. However, half of women smoking at conception quit soon thereafter.³³ Exposure timing is critical for health impacts, because sustained smoking across pregnancy reduces birth weight, whereas smoking ending early does not.¹⁰² This finding matches EWAS findings, where methylation signals reflect sustained smoking across pregnancy, not transient early smoking.^33,84 In contrast, cotinine, the previous primary biomarker of smoking, reflects only very recent exposure. By capturing additional parameters of exposure that are difficult to collect on questionnaires and identifying smokers who would be classified as nonsmokers, methylation biomarkers should improve researchers’ ability to accurately estimate health impacts of exposure. Because some methylation biomarkers of prenatal smoke exposure persist,⁹⁹ they are useful in childhood studies. Smoking strongly correlates with many harmful exposures. Methylation biomarkers allow thorough adjustment for confounding introduced by smoking exposure when studying other exposures. Clearly methylation biomarkers greatly improve on previous assessments of prenatal smoke exposure.

Figure 4.

Prenatal exposure EWAS findings can lead to multiple downstream research opportunities. First, prenatal exposure EWAS findings can be used to develop and test a DNA methylation-based exposure biomarker. Active methods development work is optimizing the DNA methylation site selection procedures and improving biomarker signals.⁹⁵ Second, once a DNA methylation exposure biomarker has been developed, this tool can be used to test associations between exposures and health outcomes. This approach is particularly useful in situations where other exposure measures are not feasible. Another potential application of prenatal exposure EWAS findings is to establish causal relationships. Instrumental variable analysis can be used to test potential causality between the exposure and DNA methylation. Mendelian randomization uses genetic factors as an instrumental variable, but this approach may be challenging for environmental exposures where genetic factors that contribute to exposure absorption, distribution, metabolism, and excretion are poorly characterized. Randomized controlled trials are feasible for nonhazardous dietary factors. Natural experiments featuring an event or policy that alter exposure levels may be a more realistic instrumental variable for prenatal exposures.^96,97 However, these would need to be drastic events, such as a policy banning a commonly used chemical. Last, prenatal exposure EWAS findings can be used in mediation analyses of exposures and health outcomes. In this case, DNA methylation is tested as a mediator linking exposure to a health outcome. This scenario is challenging when DNA methylation is a strong biomarker of the exposure, leading to overestimation of mediation effects. Note: EWAS, epigenome-wide association study.

Success of EWAS in generating biomarkers of maternal smoking provides a proof of principle for other pregnancy exposures. Routinely stored newborn blood spots become a goldmine of information about the prenatal environment for studies of later childhood health outcomes.¹⁰³ Even for exposures that can be queried in later childhood, methylation biomarkers avoid biased recall influenced by disease diagnosis.

Standard and novel statistical and computational approaches to biomarker generation from EWAS findings have recently been reviewed.⁹⁵ Reliable biomarker generation is possible only when substantial and reproducible differential methylation is identified. To date, only smoking and folate meet this criterion (Figure 2). In the “Discussion” section, we highlight important factors influencing EWAS power and suggest advances that will lead to better-powered EWAS of many exposures and thus novel biomarker generation.

Discussion

Our review suggests that although the number of prenatal EWAS conducted has increased over recent years, reproducible DNA methylation signatures for prenatal exposures besides cigarette smoking and folate are currently lacking. We hold that this reflects a lack of power, rather than a failure of the approach. To expand on current research and potentially identify additional effects of prenatal exposure on methylation, better-powered EWAS are necessary.

The genome-wide nature of EWAS impacts power. In any high-dimensional genomic analysis, the large number of statistical tests results in greater testing burden and requires larger sample sizes. The burden of multiple tests creates a trade-off between higher coverage of the methylome and reduced power. As technology advances, more sites are tested for differential methylation. We saw this happen with the shift from the earlier Illumina450K array ($\sim \mathrm{480,000}$ CpGs) to the higher coverage EPIC ($>\mathrm{850,000}$ CpG) and the EPIC version 2 ($>\mathrm{935,000}$ CpGs) arrays and it may happen again with the implementation of reduced representation or whole genome sequence bisulfite sequencing. Each time, the multiple testing burden correspondingly increases. When holding all other factors constant, larger sample sizes increase power. Yet the largest EWAS conducted to date are small in comparison with genome-wide analyses. For example, the most recent birth weight GWAS has a sample size of 321,223,¹⁰⁴ whereas the largest EWAS of birth weight has a sample size of 8,825.¹⁰⁵ In GWAS, larger sample sizes have generally led to identification of many additional genetic variants with smaller effect sizes. The larger the expected magnitude of the association, the higher the power to detect it. For maternal smoking and newborn methylation, only a few genes, such as AHRR, GFI1, and MYO1G, have CpGs with large effect sizes—10%–15% differences by exposure group.³³ However, the average smoking effect sizes across all significantly differentially methylated sites have been small, in the range of 0.5% differences³⁴ but nonetheless are highly reproducible across studies. In EWAS meta-analyses of other exposures to date, effect sizes have generally been small but are potentially important for child health.²⁵ For studies to identify these small differences, or reliably conclude that they do not exist, samples sizes need to be large.

The power of any exposure effect study is also a function of the exposure prevalence (for a binary exposure) or variability (for a continuous exposure) and of the precision and validity of the measures of both outcome and exposure.¹⁰⁶ Optimizing these factors is often more difficult to achieve than simply increasing the overall sample size. We reflect on some of these factors and give recommendations for addressing them below.

New pregnancy cohort studies seeking to perform prenatal environmental EWAS studies in the future can take advantage of lessons learned in the field and prepare themselves to take advantage of emerging approaches. Each study will have specific research questions and resource constraints. Below, we walk through general recommendations to improve study power and validity at each stage of the sample and data collection process. Rapid developments are being made in environmental epigenetics. The greatest improvements in the ability of EWAS to detect prenatal exposure effects will come from advances in inclusion, methods, and biological interpretation.

Diversity, Equity, and Inclusion

Study participants are the foundation of every epidemiological research study. Most prenatal (Figure 2) and adult EWAS identified in this review were from participants of European genetic ancestry.¹⁰⁷ GWAS faced a similar limitation and took steps to increase coverage of diverse populations.¹⁰⁸ Multiethnic studies enable confirmation of signals in diverse groups, while identifying those unique to one group who would otherwise escape discovery. Methylation is partially heritable and levels at specific loci may differ across populations.¹⁰⁹ As in GWAS, inclusion of diverse populations will extend the informativeness of EWAS.^107,110 Partnering with communities to ensure research participants are represented from diverse backgrounds is another essential area. We believe that expanding the participation in epigenetic research will allow for the identification of signals that are either generalizable across groups or specific within groups, which is important for biomarker validation and mechanistic investigation.¹⁰⁷ Similarly, it will be important to expand the representation of participants included in reference databases that are used for functional annotation and interpretation of EWAS findings, such as the Genotype Tissue Expression (GTEx) database.¹¹¹

For environmental epidemiology studies, it is important to recruit participants who may have a wide range of exposure levels. Environmental health disparities across populations¹¹² can influence exposure patterns relevant to prenatal EWAS.^113,114 Including diverse populations who may have greater exposure variability will increase power to detect exposure related differential methylation. Understanding complex factors underlying exposure variation across geographic and racial/ethnic groups is critical to reducing disparities¹¹⁵ and should inform future EWAS study designs and interpretation of findings.

More diverse, interdisciplinary research teams lead to more creative, innovative, different, and broad research questions¹¹⁶ and increase equity in the field.¹¹⁷ Reducing structural barriers in training and funding, as supported by the recent National Institutes of Health UNITE initiative,¹¹⁸ will enhance prenatal epigenetics research.¹¹⁹

Journals should require authors to deposit entire EWAS meta-analysis results in public databases such as the EWAS Catalog¹²⁰ and should be annotated for ancestry. Deposition will support replication testing in new studies, enable creation of methylation risk scores, and stimulate mechanistic studies.

Tissue Selection and Biospecimen Storage for Epigenetic Measures Including Cell Type–Specific Samples

Most prenatal EWAS use cord blood because it is easily collected at birth and may be a surrogate tissue. To promote harmonization of epigenetic measures with existing cohorts, we recommend that new perinatal epidemiology studies collect newborn cord blood, regardless of the additional samples they are collecting. Collection of other available tissues should be based on health outcomes of interest and feasibility. For example, placental tissue has recently been an effective model of neurological and cardiovascular disorders and is an emerging tissue type in environmental EWAS.³⁵ If the newborns in a birth or pregnancy cohort will be followed into childhood, additional accessible tissues should be considered. In children, blood collection is challenging, and thus buccal cells or saliva are often used.¹²¹ Sampling disease-relevant target tissues can improve EWAS inferences.¹²² Nasal epithelium, a proxy for the lower respiratory epithelium,¹²³ is easily collected in children and has been used for EWAS of asthma and rhinitis¹²⁴ but could also help identify local effects of inhaled exposures. Skin cells can be obtained noninvasively in children using tape and could be useful for EWAS of eczema or sun exposure.¹²⁵ Follicles from plucked hairs can provide a source of stem cells.¹²⁶ Fibroblasts can be reprogrammed into induced pluripotent stem cells for further toxicological investigation.¹²⁷ Blood can be challenging to collect in population-based studies of children. However, for some exposures or health outcomes, blood may indeed be the relevant target tissue or an excellent surrogate.¹²² The postnatal tissues of interest may influence the selection of tissues at birth to enable longitudinal epigenetic measures within a common tissue type.

DNA methylation analyses can be performed on whole blood or whole tissue specimens, which can be stored frozen using standard practices. If the study’s goal is assessment of epigenetic marks beyond DNA methylation, assessment of specific types of cells (for example, mesenchymal stem cells¹²⁸) or measurement of single cells, careful investigation of the type of collection tubes and storage and processing is essential. Standard freezing lyses cell membranes; thus cell culture approaches or traditional cell counting methods, such as complete blood counts, or cell sorting with fluorescence-assistance, are possible only with fresh or specially cryopreserved samples. These analyses may mandate the use of cryopreservation tubes for blood collection. Use of cryopreservation tubes require centrifugation to isolate peripheral blood mononuclear cells (PBMCs) within a few hours of collection and aliquoting into cryovials containing freezing media. Sampling processing steps may be less feasible during field sample collection, such as during home visits. Smaller studies may be able to invest in the processing to focus on specific cell types, whereas larger studies may opt to use bulk tissue specimens such as whole blood. If biospecimens will also be used to measure other epigenetic factors, additional consideration is required. If RNA will be analyzed, RNA stabilizing solutions will be required. Ensuring higher quality RNA samples at collection and storage can save money during sequencing. Better-quality RNA samples can use more affordable poly(A) tail selection and library preparation methods than the more expensive ribosome depletion methods required for poorer quality samples. In our experience, routinely discarded term human placental tissue often has lower RNA quality, as expected, based on the physiology of pregnancy and labor,¹²⁹ relative to other tissues such as blood, with implications for RNA sequencing approaches. Many standard RNA extraction protocols omit small RNAs, and additional care must be taken if these are of interest. Epigenetic measures requiring chromatin (histone modifications, ATAC-seq) must also be considered at the sample-collection phase. Sample collection and storage methods can influence stability of epigenetic measures.¹³⁰ Samples can be processed fresh for these measures, and archiving recommendations for either snap freezing or slow freezing vary by tissue and lab.

Most tissues are complex mixtures of cell types with distinctive methylation patterns. Including cell type proportions in EWAS regression models can improve precision.¹³¹ Many exposures influence cell composition,¹³² however, which might mediate associations of exposures with methylation, and conditioning on a mediator can introduce spurious effects. Cell proportions in bulk tissue can be estimated from methylation measures using reference panels for many tissues including cord blood,¹³³ saliva,¹³⁴ and placenta.¹³⁵ Cell proportions estimated from methylation provide another relevant outcome measure.¹³¹ Newer methods leverage single-cell RNA sequencing (scRNA-seq) data to deconvolute bulk tissue methylation data.¹³⁶ ScRNA-seq identifies untargeted cells in the tissue, including cells in various states of activation or proliferation that might be influenced by exposure.¹³⁷ Improved cell estimation will also enable investigators to discern whether exposure–methylation associations are driven by specific cell types.¹³⁸ For example, smoking effects might be partially mediated by activating subpopulations of immune cells not captured in published reference panels.¹³⁹

Improved and Standardized DNA Methylation Measurements

As noted above, current methylation arrays cover only a fraction of the DNA methylome (935,000 CpGs on the IlluminaEPIC version 2.0 array vs. 28.3 million in the genome). Beyond the coverage-burden trade-off discussed above, the validity of DNA methylation measurements can also impact the ability to reliably detect associations with prenatal exposures. As expected, associations with exposures or outcomes have been disproportionately observed at CpGs that are measured more reproducibly.¹⁴⁰ Further, both arrays include many sites largely invariant in blood in general populations.¹⁴¹ At these invariant CpGs, even the most precise measurements will not be sufficient to reliably capture the tiny differences between individuals. Reproducibility of measurement tends to be lower at CpGs with either very low ($<10\%$) or very high ($>90\%$) levels of methylation.¹⁴² Removal of invariant probes prior to analysis has been suggested to decrease statistical testing burden and therefore increase power.¹⁴³ However, even at some loci with high interindividual variability, reproducibility has been shown to be low.¹⁴⁰ In the meta-analysis setting, it would be important to show that these probes are uniformly invariant or poorly reproducible across most populations, but such data are lacking. Increased precision of DNA methylation measurements will increase power. Illumina is developing a methylation screening array intended to be sold at a lower price with far fewer CpGs than the current EPIC version 2, targeting CpGs differentially methylated in prior EWAS plus other informative content. We anticipate that as the price of methylation arrays come down, using methylation exposure biomarkers in population studies will become more feasible and widespread. Like newer genetic arrays tailored for population-based studies (e.g., Illumina Global Diversity Array), methylation arrays tailored for using blood from diverse populations would increase power of EWAS.

Whole genome bisulfite sequencing is increasingly being recommended to increase genomic coverage. Although the increased genomic coverage enhances completeness of environmental methylation signatures, it also increases the statistical testing burden, decreasing power and requiring larger sample sizes. To improve the feasibility of implementing sequencing in epidemiological cohorts, decreases in sample input requirements and cost are needed. Alternative and more affordable sequencing approaches, including reduced representation bisulfite sequencing or sequencing following TruSeq Methyl Capture EPIC library preparation, can limit the genomic areas sequenced, thereby maintaining power. The TruSeq approach is designed to be compatible with microarray approaches and users have observed high technical correlations.¹⁴⁴ High sequencing depth is recommended for methylation quantification.¹⁴⁵ Low sequencing depths can result in measurement error, also decreasing power.

Like most laboratory assays, methylation arrays are subject to batch effects and lab drift. In early EWAS, expectation of similarity to genotyping, which is relatively robust to batch effects, contributed to underappreciating the importance of randomizing cases and controls, or exposed and unexposed, across plates and batches. Batch effects in sequencing data can be an even larger issue than for array-based measures.¹⁴⁶ A recent comparison of split sample reproducibility in measurements of methylation in newborn blood samples between the EPIC array and bisulfite sequencing found the array to be more reproducible.¹⁴⁷ Improved preprocessing and analytic methods for sequencing data will be needed to incorporate these into consortia. With either array- or sequencing-based methods to assess methylation, appropriate control for technical sources of variation is needed and improves power.¹⁴⁸

Even though sequencing approaches are becoming cheaper and offer better coverage, to promote replication and meta-analyses, investigators may strongly consider including an Illumina DNA methylation array for greatest harmonization opportunity with other cohorts. The explosion of software for all aspects of processing and analysis of array data makes them more accessible to those with less-specialized bioinformatic assistance or training. For the widest genome-wide discovery, investigators may select whole genome bisulfite sequencing approaches. Because there are few epidemiological studies with sequencing measures available for replication testing, investigators may need to adopt a testing and replication sample set within their study population. To balance costs, investigators may opt for an enriched sequencing approach, such as reduced representation bisulfite sequencing, but this approach may have limited overlap of coverage with previous studies. When performing laboratory epigenetic measures, we strongly recommend that investigators group multiple longitudinal specimens for measurement at the same time to avoid batch effects. This methodology may require waiting to measure specimens from an earlier collection period until specimens from a later collection period are available. This approach will allow for the analysis of longitudinal epigenetic associations, which enhances the ability to make causal inference. Stacking multiple omics measures on the same participants improves the biological interpretability of findings. For example, we recommend measuring DNA methylation, genetics, and RNA expression in the same participants.

Improved and Standardized Considerations for Exposure Assessment

Issues in exposure assessment for environmental epidemiology, including special considerations in pregnancy, have been extensively discussed elsewhere.^149–152 Here we briefly touch on points of special relevance to prenatal epigenetic studies. The power to detect exposure effects on an outcome, including differential methylation, depends on both its prevalence (binary exposures) or variability (continuous exposures) and on how well the exposure is measured. If the exposure is very rare or hardly varies, even the largest study will be poorly powered. For most exposures, populations differ in the prevalence and/or variability of any exposure, emphasizing the importance of diversity in populations as discussed above. Regardless of the exposure’s prevalence/variability, if an exposure is poorly measured, exposure misclassification can impact validity and power. Generally, if misclassification is nondifferential with respect to the outcome (as is often the case for prenatal exposures and DNA methylation levels), the resulting bias is toward the null.¹⁵³ In the setting of meta-analysis (or pooled analysis), exposure misclassification will be reduced when the exposure metric can be well harmonized across studies. For prenatal exposure EWAS, considerations of persistence of the chemical are important. For nonpersistent chemicals with shorter half-lives (such as phthalates with a urinary half-life $<24$ h), repeated measures during pregnancy better characterize exposure.^151,154 For some prenatal exposures, timing during gestation of exposure may matter. For example, for maternal smoking during pregnancy, methylation signatures in newborns predominantly reflect exposure that lasts throughout the pregnancy rather than smoking that ends early in the pregnancy.^33,84 About half of women smoking at the beginning of pregnancy quit early during the pregnancy, but it is very uncommon to take up smoking de novo later in pregnancy.^33,84 Thus, we cannot be certain that sustained smoking across pregnancy is required or whether new smoking late in pregnancy would lead to similar methylation patterns in newborns. Exquisitely detailed measurement of exposure will not increase the power to identify differential methylation if the timing captured during the pregnancy is not the relevant one. We recommend careful consideration of timing, frequency, and method of exposure measurement because these can impact the power and validity of an EWAS.

The decision of what method to use to measure an exposure often relates to pragmatic issues of availability, expense, and participant burden as well as the research question at hand. Some exposures, such as prenatal smoking, are feasible to assess using questionnaires, albeit with caveats expressed earlier, which decreases costs of data collection, facilitating larger sample sizes. When collecting exposure data via questionnaire, prospective data collection is important because after development or diagnosis of the condition, recall bias can influence participants’ exposure reports. Many prenatal exposures of interest cannot be reliably measured via questionnaire.

Sometimes multiple sources of exposure data can be collected and combined to improve measurement reliability. For example, outdoor air pollution exposure can be estimated by combining residential history with measured exposure levels at routine monitoring stations and/or satellite data, using modeling with land use features, point sources of exposure, and spatiotemporal factors, and that can incorporate corrections for measurement error.^155,156 Measures relevant to climate change, including flooding and heat islands, can also be obtained.¹⁵⁷ Residential history-based measures are least invasive for participants (no participant contact or biospecimens required) and can be selected for various periods of time to test windows of susceptibility. Additional information on how and where participants spend their time can be incorporated with additional participant burden. These strategies may not reflect what enters the body or a given target tissue.

External doses to chemical exposures can be measured using an emerging exposure assessment approach to monitor human contact. Silicon bracelets are worn by participants for a period (generally a week), and chemicals deposited on the bracelets can be quantified.¹⁵⁸ Bracelet approaches are appealing to many new studies for their ease of use, though exposure levels do not reflect the internal dose of an exposure, because the body’s barriers prevent absorption of a portion of external doses.

Internal exposure doses are generally measured in biospecimens from participants, such as blood, urine, hair, nails, or teeth. The feasibility of collecting each type of biospecimen may vary by participant group, and the half-life of exposure measures may vary by sample type, such as urine vs. blood. Exposure to many chemicals varies with calendar time due to secular trends in industrial, policy, or personal practices. This variation should be considered in deciding when to initiate or complete data collection. Accounting for time of day of collection, time since last consuming food, and delay from collection to freezing as potential confounders is often needed in environmental epidemiology even in cross-sectional studies. Exposure assessment in urinary specimens requires a correction for hydration status or dilution, which is most often done through measurement of urinary specific gravity or creatinine levels. Urinary exposure measure interpretation may be further impacted by a number of physiological changes that occur during pregnancy, including changes to glomerular filtration rate, urinary output, creatinine excretion.^159–161 Biomarker concentrations may also correlate with a participant’s physiology; for example, 1,1-dichloro-2,2-bis(p-chlorophenyl) ethylene (DDE), a pesticide metabolite, is correlated with adiposity,¹⁶² and numerous xenobiotics that are excreted in the urine are correlated with kidney function.¹⁶³ Genetic factors may influence absorption, metabolism, or excretion of xenobiotics, which influence measured biomarker concentrations.¹⁶⁴ Based on the target window of susceptibility of the chemical under investigation, different tissues may be selected. Laser ablation inductively coupled–mass spectrometry (ICP-MS) of shed baby teeth in children can simultaneously quantify exposure to multiple metals during the in utero period with precise timing information.¹⁶⁵ This method has been adapted for additional tissues, including placenta and brain, allowing for cell type specific exposure resolution.¹⁶⁶ Consideration must also be given to avoiding contamination in collection. For example, if trace metals will be measured, samples must be collected in metal-free tubes. Some analytes are impacted by freeze–thaw cycles, which influence the size of aliquots for storage. Whether the analytes of interest are stable in freezer storage needs to be ascertained in advance. If no prior data are available, pilot studies will be required.

When combining exposure data across cohorts for an EWAS meta-analysis, divergent exposure estimation methods can increase misclassification and reduce power. The Environmental Influences on Child Health Outcomes (ECHO) consortium combines U.S. birth cohorts to examine health effects of prenatal environmental exposures.²⁷ ECHO is applying common exposure estimation techniques for air pollution across birth cohort studies¹⁶⁷ and is measuring prenatal environmental exposures in central laboratories using common technologies. There is a similar European cohort harmonization project, Advancing Tools for Human Early Lifecourse Exposome Research and Translation (ATHLETE).¹⁶⁸ Refined exposure harmonization should reduce measurement error and increase power.

Improved exposure assessment (Figure 1B) should decrease exposure misclassification, thus increasing power to identify prenatal exposure-related differential methylation. Even for persistent chemicals where single measurements might be acceptable, most chemical exposures are currently measured individually at high cost. Technologies to measure hundreds of compounds simultaneously, including nuclear magnetic resonance and mass spectrometry,¹⁶⁹ enable assessment of the exogenous exposome. We believe that improvements in techniques to capture time-integrated multiple early-life exposures with high precision and low cost will revolutionize the field.

Replication, Meta-Analyses, and the Need for Improved Statistical Methods

Of note, the prenatal EWASs with the largest sample sizes in our review were meta-analyses or pooled analyses conducted through collaboration via consortia. An international consortium examining the prenatal environment’s impact on early-life methylation, and subsequent impacts on later child health outcomes is the PACE consortium.²⁶ PACE was modeled on highly successful GWAS consortia that have produced a wealth of reproducible and reliable results linking genetic variants to numerous phenotypes and diseases.^170,171 PACE and other EWAS consortia apply a multistep meta-analysis approach (Figure 3). This multistep approach can improve replicability (by verifying an association exists across multiple populations), reproducibility (by increasing sample size and implementing quality controls on analyses), and rigor (for example, by developing harmonized definition of exposure included in clear analysis plans that are tested before distributing to participating cohorts, by performing an independent shadow repeat of the meta-analysis from scratch). We note that some prenatal EWAS of relatively small sample size, which were neither performed as meta-analyses nor attempted replication, nevertheless reported large numbers of significant findings (Figure 2D), which may reflect false positives or otherwise nonreplicable findings. We believe this underscores the importance of meta-analysis and replication attempts. To increase the rigor and reproducibility of the research, we recommend that new studies partner with existing consortia where possible to take advantage of existing expertise and processes as illustrated in Figure 3 to enable better-powered, higher-quality studies. Continued increases in power will enable detection of weaker differential methylation signatures, as we might expect for exposures measured less precisely than smoking or those that vary less. Lower costs of both exposure assessment and methylation measures will also facilitate larger EWAS.

Figure 3.

Overview of the suggested process for conducting a prenatal exposure EWAS meta-analysis in the consortium setting. Note: EWAS, epigenome-wide association study.

The standard method to combine study-specific results in genome-wide analyses is fixed-effects meta-analysis,¹⁷² which weights by the inverse of the variances to produce an overall effect estimate and standard error. Although it is often stated that random effects models should be used when there might be between-study heterogeneity, Rice et al.¹⁷² show that inverse variance–weighted average estimates a reasonable and interpretable parameter, even under the assumption that effect sizes differ. They further point out that a fixed-effects meta-analysis does not require the assumption of homogeneity. Rather than relying on tests of homogeneity, it is important to evaluate meta-analysis effect estimates along with visualization both of study-specific estimates and leave-one-out analyses whereby the meta-analysis is repeated after leaving out each study. These more convincingly demonstrate whether meta-analysis results are driven by a single study than statistical tests of heterogeneity.¹⁷² In the setting of environmental EWAS, the exposure may have been measured in greater detail and with higher precision in some studies than others. Studies with more accurate characterization of the harmonized exposure metric will tend to have larger effect sizes and smaller standard errors, which would make them more influential in inverse variance–weighted meta-analysis.

Pooled analyses¹⁶⁷ are an alternative to meta-analysis that enable more flexibility in the analysis, including ability to perform additional analyses that might be suggested by reviewers. However, pooled analyses require sharing of underlying data, which requires more administrative person-time to coordinate human subjects issues and data transfer agreements as well as extensive data harmonization.

New studies should identify others with the most similar designs and sample-collection protocols to facilitate future collaboration and replication. Studies can partner for coordinated epigenetic and exposure measures, which makes pooling of data across cohorts more feasible.¹⁷³ If appropriate partnerships cannot be found, new studies performing prenatal environmental EWAS should look up their results in comparison with previously published results, a routine step even in consortium meta-analyses. Prior studies may differ in their exposure assessment approaches or tissue of DNA methylation measure, but comparing results across studies will help readers understand which findings are generalizable and which may either be population specific or potentially false positives.

The exposome is a complex mixture. Statistical mixture methods are currently computationally intensive, even for a few analytes, and thus not applicable for high-dimensional epigenomic analyses.¹⁷⁴ We expect that the development of computationally efficient methods will improve the ability to capture associations between the exposome, methylation, and health outcomes.

Statistical advances have improved EWAS, including consideration of bias and inflation across genome-wide models¹⁷⁵ and dealing with extreme methylation values prior to analysis, such as by winsorizing.¹⁷⁶ Improved methods are needed to address potential heterogeneity from incorporating multiple genetic ancestries in EWAS meta-analyses. New methods are also needed to optimize biomarker development and validation in the EWAS meta-analysis setting.

Many methods papers propose novel methods to address an EWAS limitation. However, when these methods are tested only in one or few publicly available data sets, which may vary in quality control, it becomes difficult for EWAS researchers to evaluate how widely applicable they may be. EWAS consortia facilitate testing new methods in multiple data sets. We believe that finding a superior method to standard approaches across many different studies is most informative.

Epidemiologic Approaches to Investigate Persistence and Causality of DNA Methylation Signals

Although it is possible that differential methylation at genes involved in metabolism of a xenobiotic could influence measured concentration, in general methylation is unlikely to influence the likelihood of exposure. Therefore EWAS of environmental exposure are less likely to result from reverse causation than EWAS of health conditions, where the direction of association can be difficult to determine. For nonharmful dietary exposures, sufficiently powered randomized controlled trials could be used to confirm causality of the relationships between the exposure and DNA methylation. Air pollution epidemiology studies have also used short-term controlled exposure chambers,¹⁷⁷ which have been deployed in a randomized design for DNA methylation research.¹⁷⁸ Natural experiments, whereby a drastic change in exposure occurs within a population because of policy changes or other unusual events, can provide an opportunity for exploring causal exposure effects.^96,97

When methylation data are available from both newborns and older children, persistence of signals present at birth can be studied. Assessment of the same exposure in childhood is required to assess whether persistent signatures reflect postnatal, rather than prenatal, exposure. Currently much more EWAS data are available regarding prenatal exposures than for later in childhood. The Avon Longitudinal Study of Parents And Children (ALSPAC) study measured methylation at birth, childhood, and adolescence in the same analytic runs. Accounting for postnatal smoke exposure, ALSPAC found persistence of prenatal exposure differential methylation varied substantially across loci,⁸⁵ a result mirrored in later cross-sectional meta-analysis.³³ Future studies with careful repeated measures are necessary to identify persistent signatures of other prenatal exposures. We expect that in addition to being useful biomarkers in later life of prenatal exposure, loci retaining these signatures might be more likely to be involved in the etiology of childhood and adult health outcomes. Based on the likely relationship of exposure influencing methylation, the next step is to determine if exposure-related methylation contributes to causation of exposure-related health outcomes.

Although prenatal exposure biomarkers are a major advance, even without elucidating mechanisms,⁹⁵ determining whether methylation signals are on the causal pathway to exposure-related health outcomes is of primary interest. Mediation analysis, a statistical technique developed to assess causal relationships between exposures and outcomes in social psychology research,¹⁷⁹ is commonly used. Methylation is tested as a mediator of the exposure–disease relationship (Figure 4). For maternal smoking, large-effect size CpGs⁴¹ in three genes were tested as mediators of the well-established association between smoking and reduced birth weight. These CpGs were estimated to mediate nearly half of the effect of self-reported any smoking during pregnancy on birth weight. However, any smoking during pregnancy has a much weaker association with birth weight than sustained smoking,^98,102 and these smoking-related CpGs capture sustained smoking much better than self-report. One such CpGs is so powerful a biomarker of lifetime personal smoking history that it was patented for use in the insurance industry.¹⁸⁰ Because smoking CpGs are such strong biomarkers, false positive evidence of mediation between self-reported exposure and the outcome can result.¹⁸¹ Further, it seems biologically implausible that one, or few, CpGs in blood, among the thousands differentially methylated by maternal smoking,³³ could mediate so much of smoking’s impact on birth weight.^41,182

Another statistical technique used to assess whether exposure-related methylation causes exposure-related health outcomes is Mendelian randomization (MR). To evaluate whether exposure-related methylation is causal for an exposure-related health outcome, MR uses genetic variants related to the exposure-associated methylation and to the outcome only through the pathway of methylation.¹⁸³ Even when many CpGs are differentially methylated by exposure, there may be few genetic instruments meeting these criteria, and most genetic instruments for exposure CpGs are weak predictors. Fully half of CpGs on the Illumina450K platform are related to nearby genetic variants, but the variation in methylation explained is exceedingly low.¹⁸⁴ Further, local genetic correlation can lead to noncausal associations between genetic variants and methylation in cis,¹⁸⁵ obscuring inferences. Pleiotropy assumptions of MR can be difficult to verify.⁶ Another limitation of MR for interpreting blood EWAS findings is that genetic instruments predict methylation of the CpG site in blood, which is likely a proxy for epigenetic processes in the relevant target tissue.¹⁸⁶ Along with mediation, MR requires strong casual inference assumptions that are generally difficult to meet or verify.¹⁸⁷ Understanding whether exposure-related differential methylation causes exposure-related disease will require identification of the underlying epigenetic mechanisms.¹⁸⁵

Laboratory Approaches for Mechanistic Studies of DNA Methylation Signals

We currently lack insight into basic biological mechanisms linking exposures to sequence-specific methylation. Nutrients, such as folic acid, vitamin B12, and choline, provide the methyl substrate used in methylation.¹⁸⁸ Chemicals that cause oxidative DNA damage may inhibit DNA methyltransferase enzyme binding with DNA, resulting in hypomethylation.^189,190 Although one might predict that these chemicals would produce global, random shifts in methylation, they are often associated with altered sequence-specific methylation. Mechanisms for the sequence specificity remain largely unknown, limiting our ability to understand whether they cause health outcomes.¹⁹¹ Some hypotheses regarding mechanisms of exposure impacts on DNA methylation, as observed in EWAS investigations, have been proffered. The sequence-specificity suggests involvement of transcription factors.¹⁹¹ Exposures may activate or repress transcription factors that either hinder or facilitate gene-specific differential methylation.¹⁹² Alternatively, or additionally, exposure-related methylation changes may be proxies for histone modifications that alter gene function and contribute to exposure-related disease pathogenesis.¹⁸⁶ Indeed, during normal development, histone modifications often precede methylation changes.¹⁹³ Thus, differential methylation might be downstream impacts of exposure-induced histone changes rather than direct exposure effects.

Controlled, laboratory-based validation of population-based findings will provide important mechanistic insights. One approach involves examining downstream effects after experimentally modulating methylation levels. Traditional in vitro methods to modulate methylation use compounds such as 5-azacytidine,¹⁹⁴ or alter levels of methyl donors¹⁹⁵; both influence methylation genome-wide and often have off-target effects. Like CRISPR methods to modify genotypes at specific positions,¹⁹⁶ new epigenetic editing methods allow for sequence-specific alteration of methylation levels in animal and cell culture models.¹⁹⁷ Although current epigenetic editing approaches are less accurate than genetic methods, we observe that epigenetic editing is already revolutionizing investigation of the effects of methylation differences by helping identify the downstream biological processes resulting from exposure-induced differential methylation.

Precisely how methylation alters gene expression is poorly understood.¹⁸⁵ Methylation upstream of a gene in cis is generally assumed to be associated with reduced gene expression. In several EWAS of exposures or health outcomes, this canonical inverse association (higher DNA methylation with lower gene expression) holds for about two-thirds of DM-CpGs.^33,47,198 The correlation with gene expression depends on genomic context,¹⁹⁹ and many exceptions to the canonical association exist. By combining modern causal inference methods with genome-wide analysis,¹⁸⁵ specific and directional effects of gene expression on trans methylation ($\text{CpG}>10\text{\hspace{0.17em}MB}$ from gene) were identified at 818 genes. Most of these genes were neither transcription factors nor previously known to regulate methylation, pointing toward future research avenues. Understanding the fundamental epigenetic mechanisms influencing gene expression will enable meaningful interpretation of EWAS findings.

Multi-omic analyses for functional annotation, biological interpretation, and discovery.

Methylation is only one aspect of the epigenome. Additional levels include regulatory RNA and histone modifications. Storing samples for measurement requires steps currently too labor-intensive to be practical for large studies. Modified sample storage in future studies could enable better epigenome assessment. Each type of histone modification requires separate measurement.²⁰⁰ Development of high-throughput, low sample input methods to interrogate multiple epigenomic features on routinely stored samples will be a major advance.

Incorporating multiple types of omics on the same samples (e.g., gene expression, genetics, metabolomics, proteomics) can improve biological interpretation and mechanistic insights from EWAS findings.^{175,201–204} A common first level of integrated omics analyses with EWAS (Figure 3) is to look at correlation between gene expression and DNA methylation at differentially methylated CpGs. There are relatively few publicly available data sets with both Illumina methylation array data and gene expression in cord blood.^205,206 For this reason, prenatal epigenetic studies often also look up key findings in data sets with paired blood methylation and gene expression from up to a few hundred children²⁰⁷ or much larger studies of adults.^208,209 Given the tissue specificity of both methylation and gene expression, of considerable interest is the recent addition of Illumina EPIC methylation data to the GTEx gene expression resource across nine different tissues from 987 adults.²¹⁰ Colocalization methods^211–214 and integrative epigenomic²¹⁵ tools are increasingly aiding in EWAS biological interpretation. For example, eFORGE integrates multiple layers of epigenetic data to identify blood EWAS findings with potential functional impact in disease-relevant target tissues.²¹⁵ Such tools will become more valuable as the underlying epigenomic sample sets amalgamated by the International Human Epigenome Consortium²¹⁶ become larger and incorporate more diverse populations.¹¹⁰

Development of methods to use multi-omics data to discover novel loci or intermediate omic phenotypes, as opposed to interpreting EWAS²⁰⁴ loci, is an active area of research¹⁷³ and well-validated methods have not been established or widely employed in epidemiology. Multi-omics methods were reviewed, and their limitations were discussed in 2020.²¹⁷ A more recent approach is multi-set correlation and factor analysis, an unsupervised integration method to enable novel inference from multi-omic data.²¹⁸ Widespread future implementation of computationally efficient methods for multi-omics discovery, including epigenomic data, will help understand mechanisms of health impacts of prenatal exposures.

Summary

The early-life environment impacts health throughout the life course, and epigenetic mechanisms play a key role in programming responses to these exposures. Methylation is the epigenetic mechanism best epidemiologically studied to date, thanks to stable platforms with reasonable epigenome-wide coverage. Prenatal environmental exposure to smoking is associated with highly reproducible and specific methylation signatures that have great utility as biomarkers of early-life exposure. Improvements to study design could result in robust biomarkers for other exposures. Best practices for EWAS include careful study design, attention to tissue- and cell-type heterogeneity, rigorous exposure measurement, large and diverse study populations, and replication. Environmental epigenetics is an emerging area with promising research and clinical implications. Mechanisms underlying the specificity of exposure methylation signatures and consequences of the methylation differences are currently unclear but are a crucial future research area essential for developing clinical applications.