Early-life exposures and DNA methylation

To date, the prenatal exposure that has been most robustly linked to DNA methylation changes in the offspring is maternal smoking.Reference Joubert, Felix and Yousefi ^{18 –} Reference Rzehak, Saffery and Reischl ²³ In a recent large study based on data from 13 cohorts (n=6685), Joubert et al.Reference Joubert, Felix and Yousefi ¹⁸ found that sustained smoking during pregnancy was associated with altered DNA methylation patterns across thousands of CpG sites, an effect that was still observable – albeit to a lesser extent – in older children, although the sample size was around half that of the newborn analysis. This study’s top finding was a CpG in AHRR, which has been associated with both personal and maternal smoking in multiple other studies.Reference Joubert, Haberg and Nilsen ^{19 ,} Reference Kupers, Xu and Jankipersadsing ^{24 ,} Reference Zeilinger, Kuhnel and Klopp ²⁵ Pathway analyses found neurological and developmental pathways, which is in line with the maternal smoking-related child outcomes. Alterations in DNA methylation at birth have also been documented in relation to other chemicals, such as prenatal exposure to arsenic,Reference Green, Karagas and Punshon ²⁶ mercuryReference Cardenas, Rifas-Shiman and Agha ²⁷ and air pollution.Reference Gruzieva, Xu and Breton ^{28 ,} Reference Breton, Gao and Yao ²⁹

Maternal pre-pregnancy body mass index (BMI) has also been associated with offspring DNA methylation, with findings showing associations with both lower and higher maternal BMI in a recent meta-analysis of 9340 newborns.Reference Sharp, Salas and Monnereau ³⁰ Effects were generally small, with a median increase of 0.03% in methylation level per unit increase in maternal BMI, and strong evidence for a direct causal intra-uterine effect was lacking for most loci.Reference Sharp, Salas and Monnereau ³⁰ Furthermore, dietary factors have been examined, with a particular focus on folate intake, due to its established role in both epigenetic regulation and fetal brain development.Reference Irwin, Pentieva and Cassidy ³¹ The largest of these studies, based on data from two birth cohorts (n=1988), reported associations between maternal plasma folate levels during pregnancy and differential DNA methylation in cord blood across multiple genes involved in folate biology and neurodevelopmental processes.Reference Joubert, den Dekker and Felix ³² Interestingly, this study found reduced methylation with higher levels of folate, a methyl donor, at the vast majority of CpG sites, a finding which underlines the complex associations of the one-carbon metabolism pathway with DNA methylation.Reference Joubert, den Dekker and Felix ³² In contrast, other known risk factors for poorer offspring health have not been found to be strongly associated with neonatal DNA methylation. For example, a recent meta-analysis on maternal alcohol use found no evidence for an association with cord blood DNA methylation.Reference Sharp, Arathimos and Reese ³³ Similarly, a meta-analysis of prenatal maternal stress did not identify associations with DNA methylation in cord blood, although an enrichment for genes related to methyltransferase activity was identified.Reference Rijlaarsdam, Pappa and Walton ³⁴ Besides a true null effect, the lack of a strong signal in these studies may reflect low levels of more severe forms of exposure in the general population, and relatively small sample sizes, resulting in decreased power to detect effects, or heterogeneity between the included studies in terms of population, exposure definition or confounding structure.

DNA methylation and childhood health outcomes

Compared to prenatal exposures, fewer population-based studies have examined the relationship between DNA methylation patterns and childhood health outcomes. We highlight here two important areas of progress in the field of DNA methylation in population-based studies: neurodevelopment and adiposity.

For neurodevelopmental outcomes, most of the evidence to date on this topic has been drawn from cross-sectional data in smaller, selected samples, for example, high-risk/clinical samples, although efforts to characterize these associations in the wider population are fast increasing. One such study, based on data from 537 newborns, found that epigenetic patterning of glucocorticoid genes in the placenta was prospectively associated with infant neurodevelopmental profiles.Reference Paquette, Lester and Lesseur ³⁵ Specifically, infant membership to a reactive, poorly regulated profile associated with increased placental methylation of the NR3C1 glucocorticoid receptor gene, and decreased methylation of the HSD11B2 gene involved in limiting fetal exposure to maternal circulating cortisol. Other population studies have examined DNA methylation in relation to the risk of neurodevelopmental disorders – particularly attention-deficit hyperactivity disorder (ADHD), but also other psychiatric problems with early developmental origins, such as oppositional defiant and conduct problem behavior.Reference Barker, Walton and Cecil ^{36 –} Reference Walton, Pingault and Cecil ³⁹ For example, a longitudinal epigenome-wide study based on over 800 children found that DNA methylation patterns at birth (but not age 7) were prospectively associated with trajectories of ADHD symptoms from childhood to adolescence.Reference Walton, Pingault and Cecil ³⁹ Of the 13 genome-wide significant loci identified at birth, several were annotated to genes involved in fatty acid metabolism and neural tube development. Importantly, one of the loci was annotated to a gene (ST4GAL3) that has been recently identified as the top genetic hit for ADHD in the largest genome-wide association study to date.Reference Demontis, Walters and Martin ⁴⁰ Studies are also beginning to integrate brain imaging data to investigate the links between DNA methylation, brain development and neuropsychiatric outcomes. For example, a recent EWAS found that SP6 methylation at birth was prospectively associated with lower amygdala–hippocampal volume and higher psychotic symptoms in young adulthood – an association that was replicated in an independent sample of patients with schizophrenia.Reference Walton, Cecil and Suderman ⁴¹

Most of the evidence regarding DNA methylation and adiposity stems from adult studies.Reference Agha, Houseman and Kelsey ^{42 –} Reference Wahl, Drong and Lehne ⁴⁷ The largest of these identified and replicated cross-sectional associations between DNA methylation at 187 loci and BMI in over 10,000 samples. Less is known about DNA methylation and childhood adiposity. Several studies have described associations of DNA methylation with birth weight, with a limited number of CpG sites identified in more than one study.Reference Agha, Hajj and Rifas-Shiman ^{48 –} Reference Simpkin, Suderman and Gaunt ⁵² No strong evidence was found for persistence of associations into later life in the two studies that explored longitudinal relationships.Reference Agha, Hajj and Rifas-Shiman ^{48 ,} Reference Simpkin, Suderman and Gaunt ⁵² In targeted analyses, a replicable association of RXRA methylation in umbilical cord tissue with childhood fat mass was shown.Reference Godfrey, Sheppard and Gluckman ⁵³ Several differentially methylated regions in newborn blood, such as the promoter of the long non-coding RNA ANRIL, have been associated with childhood adiposity measures based on epigenome-wide scans.Reference Lillycrop, Murray and Cheong ^{54 ,} Reference van Dijk, Peters and Buckley ⁵⁵ In cross-sectional analyses in children and adolescents, DNA methylation in several regions, including the SOCS3 locus, has been associated with adiposity measures.Reference Ali, Cerjak and Kent ^{56 –} Reference Wang, Song and Yang ⁵⁸ It remains unclear whether DNA methylation patterns are a risk factor for or consequence of adiposity. Detailed associations in the large adult EWAS show that changes in DNA methylation are mostly a consequence of obesity.Reference Wahl, Drong and Lehne ⁴⁷ This is supported by smaller studies in children and individuals before and after bariatric surgery.Reference Barres, Kirchner and Rasmussen ^{59 –} Reference Guenard, Deshaies and Cianflone ⁶²

DNA methylation as a mediator

Despite promising findings, very few population-based studies to date have integrated prospective exposure, epigenetic and phenotypic data in order to explicitly test the role of DNA methylation as a potential biological mediator of environmental effects on health outcomes. Notable exceptions include an EWAS study based on 321 cord blood samples, which showed that, in males, DNA methylation of the PON1 gene at birth associated with prenatal mercury levels and predicted cognitive performance scores during childhood.Reference Cardenas, Rifas-Shiman and Agha ²⁷ Using repeated measures of DNA methylation, the authors found that this effect persisted in early childhood (3–5 years) and was attenuated in mid-childhood (6–10 years). Furthermore, DNA methylation of the PON1 locus at birth was shown to be associated with gene expression levels in an independent set of cord blood samples. Another exception is a longitudinal EWAS study examining epigenetic correlates of adolescent substance use (tobacco, alcohol and cannabis use; n=244, age 14–18). This study found that neonatal DNA methylation levels in a tightly interconnected network of genes enriched for neurodevelopmental processes (n=65 epigenome-wide corrected loci) partially mediated the effect of prenatal tobacco smoking on later substance use risk in adolescence.Reference Cecil, Walton and Smith ⁶³ Relatedly, maternal prenatal smoking has been found by other studies to associate with lower birth weight partially through alterations in cord blood DNA methylation using both targeted and epigenome-wide approaches.Reference Kupers, Xu and Jankipersadsing ^{24 ,} Reference Bouwland-Both, van Mil and Tolhoek ⁶⁴

Tissue specificity

DNA methylation patterns are largely tissue-specific. In large cohort studies, blood is typically the main source of DNA collected. In some cases, there may be DNA from buccal swabs, umbilical cord or placental tissue, but other tissues are generally inaccessible for population researchers. This poses challenges, as the extent to which findings may be generalized to the target tissue of interest is often unclear.Reference Jiang, Jones and Chen ^{65 ,} Reference Davies, Volta and Pidsley ⁶⁶ One way to address this limitation is to refer to publicly available datasets that show cross-tissue concordance in DNA methylation patterns, based on selected samples (e.g. comparison of blood and brain in postmortem samples or blood and adipose tissue in patients undergoing surgery), for example, http://epigenetics.iop.kcl.ac.uk/bloodbrain, and https://redgar598.shinyapps.io/BECon/.Reference Edgar, Jones, Meaney, Turecki and Kobor ^{67 –} Reference Huang, Chu and Loucks ⁶⁹ However, it is not just the concordance in DNA methylation levels between tissues, but also, and maybe more so, the variability of those DNA methylation levels in association with the exposure or outcome of interest that matters. Concordance between tissues in associations with specific phenotypes has been shown, for example, for BMI between blood, adipose and liver tissue, but overall evidence is limited.Reference Wahl, Drong and Lehne ^{47 ,} Reference Jiang, Jones and Chen ^{65 ,} Reference Stueve, Li and Shi ⁷⁰ The NIH Roadmap Epigenomics Project can be used as a reference for epigenetic marks in a large number of tissues (www.roadmapepigenomics.org).Reference Kundaje and Meuleman ⁷¹

Cell type adjustment

DNA methylation patterns vary not only between, but also within tissues. In blood samples, DNA is collected from leukocytes, which represent a heterogeneous population characterized by cell type-specific DNA methylation. As such, the relative proportions of cell types from which the DNA is extracted may influence epigenetic results. Ideally, detailed measurements of leukocyte subtypes would be available for correction in all study participants, but this is often not feasible. As an alternative, the use of reference-based approaches for estimating cell type proportions has become popular, such as methods developed for whole bloodReference Houseman, Accomando and Koestler ^{72 ,} Reference Reinius, Acevedo and Joerink ⁷³ and, more recently, cord blood.Reference Bakulski, Feinberg and Andrews ^{74 –} Reference Gervin, Page and Aass ⁷⁶ Reference-free methods are also available.Reference Kaushal, Zhang and Karmaus ^{77 –} Reference Teschendorff and Zheng ⁷⁹ These methods, however, are still under development and population-based studies with specific cell type measurements can play a pivotal role in developing and validating new analytic approaches.

Methodological considerations

DNA methylation data are multifactorial, high dimensional and inter-correlated, raising questions about how best they should be analyzed.Reference Almouzni, Altucci and Amati ⁸⁰ So far, studies have varied widely in study characteristics, methodology and analytic routines, limiting comparability of findings. Furthermore, many studies have been underpowered to identify effects, especially when examining exposures or outcomes with low prevalence in the general population. To address these challenges, large collaborative efforts of prospective cohorts, such as the Pregnancy And Childhood Epigenetics (PACE) Consortium, have emerged, similar to the field of genome-wide association studies although numbers are still much smaller.Reference Felix, Joubert and Baccarelli ⁸¹ Such collaborations have shown enormous potential in increasing statistical power and optimizing the use of resources,Reference Benke, Nivard and Velders ^{82 –} Reference Pappa, Pourcain and Benke ⁸⁶ as well as allowing replication and comparison of effects between studies with potentially heterogeneous confounding structures, thus strengthening the evidence for reported associations. In addition, they are a valuable platform for standardization of study design and analysis protocols to maximize comparability and help to delineate best practices in the field. Multidisciplinary teams, bringing together basic, population and clinical scientists, will enable easier exchange of knowledge and ideas, translation between bench, population and bedside, and strengthening of scientific conclusions by bringing together multiple lines of evidence.

Role of genetics

It has been reported that a substantial proportion of the variation in DNA methylation between individuals is due to genetic variants, acting in cis or trans (mQTLs).Reference Bell, Pai and Pickrell ^{87 –} Reference Grundberg, Meduri and Sandling ⁸⁹ As such, measured DNA methylation levels may reflect genotype and mediate associations of genetic variants and health outcomes, or, if genotype drives both DNA methylation and the outcome under study, changes in DNA methylation may be an epiphenomenon of the genetic risk rather than part of the causal pathway. Furthermore, genetic factors may moderate associations between environmental exposures, DNA methylation and outcomes (Fig. 1a). Thus, genetic variability should always be considered when examining DNA methylation – ideally directly, or if unavailable or underpowered, by referring to online mQTL databases (e.g. www.mqtldb.org.)Reference Gaunt, Shihab and Hemani ⁸⁸

Causal inference

DNA methylation is a dynamic process that may change under the influence of genetic and environmental factors, but also with disease states. Thus, DNA methylation may be a cause, a consequence, or an epiphenomenon of the phenotype under study (Fig. 1b). This issue of reverse causality is particularly problematic when using cross-sectional designs. While the use of prospective, longitudinal data partially addresses directionality – especially when DNA methylation is measured before the onset of disease – observational studies in general are subject to confounding. The application of advanced causal inference methods, such as two-step Mendelian randomization offers a valuable tool for probing causal pathways.Reference Relton and Davey Smith ⁹⁰ For example, a recent study applied this method to support a small, causal effect of prenatal maternal B₁₂ levels on childhood IQ levels, via DNA methylation at birth.Reference Caramaschi, Sharp and Nohr ⁹¹ The tissue specificity of DNA methylation is an additional challenge to causal inference analyses.

Determining whether DNA methylation plays a causal role in disease is important for understanding the etiology of the disease as well as potentially developing new therapies in future. Yet, even if DNA methylation changes are not causal to a disease, but rather a consequence or an epiphenomenon, they may still serve as biomarkers of the disease or its progression (Fig. 1b).Reference Ladd-Acosta and Fallin ⁹² In addition, if an association of DNA methylation with a phenotype is non-causal, the DNA methylation level may be used as a proxy for a specific exposure. This may be especially helpful for exposures that are difficult to measure (reliably). Previous studies have shown the potential of using DNA methylation as a biomarker in a number of these instances, including for alcohol consumption, maternal smoking during pregnancy and gestational age.Reference Bohlin, Haberg and Magnus ^{93 –} Reference Reese, Zhao and Wu ⁹⁶

Functional characterization

The buck does not stop at DNA methylation. To understand pathways leading to disease, we need to study the functional effects in terms of gene expression as well as the roles of other epigenetic marks. Effect sizes in EWAS are often small, for example, in an EWAS of maternal BMI, the largest effect estimates for the significantly associated CpG sites were an 0.1% decrease or increase in methylation level per unit increase in maternal BMI (kg/m²).Reference Sharp, Salas and Monnereau ³⁰ For a strong exposure such as sustained maternal smoking, the maximum effect size was a 10% decrease in methylation levels for sustained smokers compared to non-smokers.Reference Joubert, Felix and Yousefi ¹⁸ It is currently unclear to what extent statistical significance parallels biological meaningfulness. In future, studies that incorporate both DNA methylation and expression data, preferably within the same population and at the same time point, can shed more light on this question of functionality. Such data are not widely available in longitudinal birth cohorts yet, with some noticeable exceptions (INMA Study, Gambia Study).Reference Guxens, Ballester and Espada ^{97 ,} Reference Moore, Fulford and Darboe ⁹⁸ In addition, the ENCODE database can be used to explore whether identified DNA methylation sites are located in potential regulatory regions.Reference Barker, Walton and Cecil ³⁶ Intensifying collaborations and expanding them to researchers in neighboring fields, such as clinical and basic researchers, will make it easier to build chains of evidence across species, across research fields and across the life course, leading to strong and innovative research.

Starting at the very beginning

Most Early-life cohorts begin in pregnancy and thus generally do not have prospective and hands-on data about the preconception and early pregnancy periods. Some preconception cohorts do exist (Southampton Women’s Survey: www.mrc.soton.ac.uk/sws,Reference Inskip, Godfrey and Robinson ⁹⁹ Generation R Next Study: www.generationr.nl/next) and they, together with longitudinal birth cohorts with follow-up long enough to include the next generation, will provide valuable information on exposures in these very early life stages and their epigenetic associations. Furthermore, the vast majority of epigenetic research in birth cohorts is currently focused on maternal exposures during pregnancy. In contrast, the contribution of paternal epigenetics and exposures is currently understudied but has the potential to lend important new insights in the development of the offspring epigenome and the pathways leading from early-life exposures to child and adolescent health.