Evaluating the challenges and reproducibility of studies investigating DNA methylation signatures of psychological stress

Abstract

Psychological stress can increase the risk of a wide range of negative health outcomes. Studies have been completed to determine if DNA methylation changes occur in the human brain because of stress and are associated with long-term effects and disease, but results have been inconsistent. Human candidate gene studies (150) and epigenome-wide association studies (67) were systematically evaluated to assess how DNA methylation is impacted by stress during the prenatal period, early childhood and adulthood. The association between DNA methylation of NR3C1 exon 1F and child maltreatment and early life adversity was well demonstrated, but other genes did not exhibit a clear association. The reproducibility of individual CpG sites in epigenome-wide association studies was also poor. However, biological pathways, including stress response, brain development and immunity, have been consistently identified across different stressors throughout the life span. Future studies would benefit from the increased sample size, longitudinal design, standardized methodology, optimal quality control, and improved statistical procedures.

Plain language summary

Mental stress can increase the risk of a wide range of negative health results. Previous studies have been carried out to explore the epigenetic (related to changes in genes) effect of different types of mental stress. The authors reviewed the literature to look at consistencies in the results of these studies. With the exception of the first exon of NR3C1, the authors found few consistent findings in epigenetic changes related to stress with regard to DNA methylation at individual CpG islands but identified some important biological systems (stress response, brain development and immunity) related to psychological stress. Bioinformatics analysis showed that different types of psychological stress affect different biological systems. The epigenetic mechanism of these systems may control the effects of mental stress on health.

Biological impact of psychological stress

Psychological stress refers to the emotional, behavioral and biological responses of an individual to a perceived threat. The modern concept of psychological stress introduced by Walter Bradford Cannon refers to the fight-or-flight response based on bodily changes that occur when one faces a stressful condition. In 1950, Selye further defined stress as “a nonspecific response of the body to any demand made upon it” [1].

Psychological stress can be classified by duration as acute and chronic stress. Each of the two types may involve different biological processes and have different health impacts. Acute (time-limited) stress may be essential for successful adaptation to changing natural and social environments, particularly for young and healthy individuals. Short-term stress, such as tests, presentations or interviews, prepares one's brain and body to face potential danger and motivates an individual to perform well. However, with greater duration, chronic excessive stress, such as persistently challenging life events, trauma and chronic diseases, can increase a person's vulnerability to autoimmune and allergic diseases and have damaging consequences on one's physical and mental health [2,3]. However, owing to its broad and often subjective nature, psychological stress is difficult to measure accurately, which is further complicated by the temporal dynamics of chronic stress.

Psychological stress disrupts the homeostasis of biological processes. When a person is acutely stressed, both the hypothalamic–pituitary–adrenal (HPA) axis and the sympathetic–adrenal–medullary system activate and lead to increased levels of a cascade of hormones, including corticotropin-releasing hormone, glucocorticoids, catecholamines and other downstream hormones and growth factors. When stress persists and becomes chronic, the constant activation of the HPA axis and sympathetic–adrenal–medullary system results in elevated cortisol and catecholamines, interfering with control of many other downstream physiological systems [4]. Immunity is one of the frequently recognized systems that HPA axis and sympathetic–adrenal–medullary system hormones influence. The major difference between acute and chronic stress might be that the acute stress effect is more restricted and temporary, but the chronic stress effect is more widespread and long-lasting, involving more organs and biological processes. Healthy human individuals exposed to acute stress show an adaptive enhancement of natural immunity but a general suppression of specific immunity. By comparison, exposure to chronic stress is associated with a biphasic immune response in which partial suppression of cellular and humoral function coincides with low-grade, nonspecific inflammation [5].

Chronic stress is not only a robust risk factor for many medical conditions, including cardiovascular disease, obesity, cancer and immune disorders [4], but it is also tightly linked to the development of a broad range of psychiatric disorders, including post-traumatic stress disorder [6,7], mood and anxiety disorders [8–10], bipolar disorder and schizophrenia [11], substance abuse disorders [12] and addiction [13]. However, our understanding of the biological mechanisms of these psychiatric diseases associated with chronic stress remains incomplete.

DNA methylation as an epigenetic biomarker for psychological stress

Epigenetic alterations may be associated with the biological mechanisms of psychological stress, influence behavior and reflect disease susceptibility [14,15]. As one of the most studied types of epigenetic modifications, DNA methylation is essential for the regulation of gene expression and plays an important role in the etiology of many forms of stress-associated physical and mental illnesses.

DNA methylation is defined as the addition of a methyl group to the fifth carbon of a cytosine base (5mC), primarily occurring on CpG dinucleotides in the mammalian genome. Three forms of DNA methylation have been reported: methylation at cytosine before guanine sequences (5mC), methylation at cytosine before H (non-G nucleotide, CpH) and hydroxymethylation of cytosine (5hmC). Of these, 5mC is the most prevalent form; 5hmC occurs more frequently in the brain, heart and sperm [16] and less frequently in the spleen, liver and cancer cells [17]. Not only is 5hmC an intermediate process of active DNA demethylation, but it is also a key element in the regulation of gene expression and chromatin structure [18]. CpH methylation is commonly observed in human embryonic stem cells and brain tissues and comprises 0.02% of total methylcytosine in differentiated somatic cells [19]. Specifically, neurons seem to have the highest level of CpH methylation [20].

DNA methylation is an important regulator of gene expression. A host of studies have already been conducted on the relationships between DNA methylation and gene expression, particularly those emphasizing inverse correlations [21,22]. However, positive correlations have also been found between DNA methylation and gene expression, mostly with regard to CpG in the gene body [23,24]. The relationships between DNA methylation and gene expression in diverse tissues across lifetimes have been investigated, and researchers have identified tissue- and development-specific DNA methylation–expression correlations [25]. The positive and negative correlations between DNA methylation and gene expression are associated with distinct epigenetic signatures.

DNA methylation is under the regulation of both genetic [26,27] and environmental factors [28,29]. Genetic variants such as single nucleotide polymorphisms can change the methylation status at a CpG site [30–32]. Even though DNA methylation patterns are established early in embryogenesis, environmental influences on DNA methylation are apparent across the life span [28]. Animal studies have been instrumental in showing a causal relationship between early life environment, epigenomic changes and subsequent behavioral changes [33,34]. Weaver et al. showed that maternal care levels resulted in functional and long-lasting changes in DNA methylation at the promoter of the glucocorticoid receptor in the offspring and persisted into adulthood [34]. As one of the most common environmental factors in humans, psychological stress has the potential to alter the DNA methylation of genes key to biological systems in an enduring manner and to be associated with psychiatric conditions.

It has been widely reported, based on candidate gene studies and epigenome-wide association studies (EWAS), that DNA methylation changes in stress responses are involved with global methylation and gene-specific methylation changes, respectively. Candidate gene studies comprised the majority of the earliest work in this research area. Candidate gene studies involve selecting genes based on prior knowledge of the dopaminergic, serotonergic, GABAergic and glutamatergic neurotransmitter systems as well as the immune system and inflammation. Candidate gene studies have the benefit of requiring a relatively small number of samples to achieve sufficient statistical power, making them cost effective, and many studies were completed. The hypothesis-free, data-driven EWAS gradually became the mainstream, with hundreds of thousands of CpG sites across the genome being tested simultaneously. Only associations that meet a certain threshold of significance, based on, for example, genome-wide Bonferroni threshold or false discovery rate, are typically highlighted. Compared with candidate gene studies, EWAS have the advantage of capturing epigenetic changes across the whole epigenome and are currently the most unbiased method for uncovering psychological stress-related epigenetic changes. DNA methylation studies of psychological stress examine both the brain and peripheral tissue.

DNA methylation changes of psychological stress can be carried out with cross-sectional or longitudinal studies. Cross-sectional studies measure variables of stress and DNA methylation at a single time point. However, data from the single time point make establishing causal relationships between observed DNA methylation alterations and stress exposure difficult. By contrast, longitudinal studies, where individuals are followed across different time points, allow for correction of the ‘baseline’ DNA methylation signature. Furthermore, the longitudinal design enables examination of how environmental adversity and DNA methylation change together over time, the timing of environmental effects and specific windows of biological vulnerability [35].

There are a large number of DNA methylation studies investigating one or more candidate genes associated with psychological stress, especially gene NR3C1. The designs of these studies vary widely. Some studies examined a single CpG site at the NR3C1 exon 1F area, whereas others focused on the average DNA methylation level of several sites over the entire promoter region of NR3C1. Additional studies looked at the total methylation level across the CpG sites. This diversity challenged the authors' comparative evaluation of the results of existing studies.

Reproducibility & specificity of published studies

Exploring DNA methylation signatures as biomarkers for psychological stress is crucial for improving the accuracy of diagnosis of stress-related disorders and helping to monitor clinical interventions. Dozens of studies have been carried out on this topic. However, the reproducibility and specificity of the reported DNA methylation changes have not been evaluated. In this review, the reported association between psychological stress and human DNA methylation was systematically assessed for three periods across the life span: prenatal, postnatal development (including childhood and adolescence) and adulthood. Prenatal maternal stress is considered psychological stress on the fetus. The results were summarized by demographic characteristics; stress types; DNA sample sources; main findings; statistical power; and statistical methods, including correcting for multiple testing, batch and positional effect (for EWAS) and other covariates as well as related biological pathways (for EWAS).

Studies included in the evaluation

Literature regarding DNA methylation changes associated with psychological stress was collected from PubMed, Medline and Web of Science. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses standard was followed. The Medical Subject Headings included ‘psychological stress’ AND ‘DNA methylation’. To include as many related articles as possible, the authors also filtered according to the following combination of Medical Subject Headings: ‘prenatal stress’ OR ‘prenatal depression/anxiety’ OR ‘prenatal psychopathology’ OR ‘battle/war during pregnancy’ OR ‘natural disaster’ OR ‘pregnancy intimate partner violence’ OR ‘early life stress’ OR ‘life adversity’ OR ‘child adversity’ OR ‘child maltreatment’ OR ‘child neglect’ OR ‘child trauma’ OR ‘child abuse’ OR ‘psychosocial stress’ OR ‘emotional stress’ OR ‘violence’ OR ‘bullying’ OR ‘low socioeconomic’ OR ‘socioeconomic’ OR ‘poverty’ OR ‘shift work’ OR ‘chronic stress’ AND ‘DNA methylation’. The authors also examined the reference sections from the selected articles to identify any additional relevant studies. The inclusion criteria included the following: published in English before June 2021, assessed DNA methylation changes in peripheral tissue (e.g., blood, saliva, buccal cells) or brain, involved at least one type of psychological stress, human research and original article or meta-analysis. Article titles and abstracts were carefully screened. Studies were excluded from the systematic review for the following reasons: not focused on the relationship between psychological stress and DNA methylation, animal models, articles on epigenetic aging and epigenetic clock or global methylation and other epigenetic mechanisms (e.g., histone modifications and ncRNAs). The flowchart of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature screening is shown in Figure 1.

Figure 1. . Screening of DNA methylation studies associated with psychological stress.

DNAm: DNA methylation.

In total, 217 original publications were identified (Table 1). A detailed summary was organized by candidate gene and EWAS and is listed in Supplementary Tables 1 & 2. Of the 67 EWAS, child maltreatment was the most investigated stress type (the total number was 17; of these, 15 studies focused on child maltreatment solely, whereas two investigated complex stress involving child maltreatment and other stressors). Chronic stress during adulthood was studied the least (eight total studies). The other stressors included low socioeconomic status (SES) during childhood or life span, prenatal depression or anxiety, psychological stress and occupational stress. Detailed information regarding the 217 articles is listed in Supplementary Tables 1 & 2.

Table 1. . Studies included in the review.

Time of stress	Stress type	Age at sampling	Candidate gene studies		EWAS
			Number of studies†,‡	Sample size range	Number of studies†,‡	Sample size range
Prenatal development	Prenatal depression	Infant	17	57–1233	10	23–7243
		Child	3	22–218	–	–
	Maternal psychological stress	Infant	14	18–977	6	10–1740
		Adolescent	4	29–84	2	34–36
	Prenatal SES	Infant	4	120–619	3	422–914
	Total		42 (38; 4)		21 (17; 4)	–
Postnatal development	Child maltreatment	Child	6	85–534	6	124–2474
		Adult	38	24–3965	11	34–385
	Psychological stress	Child/adolescent	12	46–1149	7	22–1700
		Adult	21	25–1668	4	109–1332
	Child SES	Child	5	33–468	3	192–1619
		Adult	8	46–1231	7	40–489
	Total		90 (80; 10)		38 (8; 30)	–
Adulthood	Work-related	Adult	11	49–3631	4	117–2574
	Chronic stress		7	47–1231	4	92–1264
	Total		18 (17; 1)	–	8 (/; 8)	–
Total			150		67	–

^†Cross-sectional studies.

^‡Longitudinal studies.

EWAS: Epigenome-wide association studies; SES: Socioeconomic status.

Reproducibility of candidate gene results

Several candidate genes, including those involved in the HPA axis, neurotransmitter and neuroendocrine signaling, neuron development and signal transmission, were implicated in the literature regarding psychological stress-related epigenomic modification. Most of these studies reported associations between psychological stress and the DNA methylation of candidate genes. The most commonly investigated candidate genes were NR3C1, SLC6A4, FKBP5, OXTR and BDNF. To reduce the noise of false-positive results from underpowered studies, the authors evaluated only the results from studies with power >60% (Supplementary Table 1). The reproducibility of results from these commonly studied genes was grouped and assessed by three life stages: prenatal development, postnatal development and adulthood (Table 2). For candidate gene study results, owing to the small number of studies among adults, the authors evaluated only the candidate gene results during prenatal and postnatal development.

Table 2. . Candidate gene DNA methylation associated with psychological stress across prenatal and postnatal development.

Time of stress	Stress type	Candidate gene	Locus	Age at sampling	Number of studies†	Largest sample size	Consistency	PMID‡
Prenatal development	Prenatal depression/anxiety	NR3C1	Exon 1F	Infant	8	1233	Yes	18536531↑, 25875334↑, 23566423↑, 25942041↑, 24135662↑, 27040859-, 24135662↑, 33215541↑
				Child	3	176	Yes	29606180↑, 31438539↑, 32087522↑
		SLC6A4	Promoter	Infant	2	82	–	20808944↓, 30447571↓
				Child	1	167	–	29606180↑
		IGF2/H19	Intron 7	Infant	6	619	No	27023171↓, 22677950-, 22414206-, 29538358-, 26333472↓, 25102259↑
	Maternal psychological stress	NR3C1	Exon 1F	Infant	7	977	No	27462209↑, 33330307-, 27013342↑, 24690014-, 26327302 (meta)↑, 22832523↑, 22810058 ↑
		SLC6A4	Promoter	Infant	2	133	–	33330307 (mean)↑, 26401310-
				Adult	1		–	24937096↑
		OXTR	CpG island	Infant/child	3	743	–	27107296↓, 27520745-, 25199917↑
		FKBP5	Intron 7	Adult/infant	2	61	–	26410355↓, 27013342↑
Postnatal development	Child maltreatment	NR3C1	Exon 1F	Child	4	534	Yes	29162170↑, 29162187↑, 25997773↑, 32472381↑
				Adult	7	295	Yes	27620456↑, 29433075↑, 25048180↑, 22832351↑, 29793048↑, 19234457↑, 29275786↑
		SLC6A4	Promoter	Adult	5	158	Yes	22707942↑, 25781010↑, 23421829↑, 23333376↑20947778↑
		FKBP5	Intron 7	Child	2	231	–	29162173↓, 26535949↓
				Adult	6	3965	No	23201972↓, 32553385↓, 30700816-, 28961425-, 29793048-, 32488091-
		BDNF	Promoter	Adult	3	167	–	24801751↑, 29748862↓, 29781947↑
	Psychological stress	NR3C1	Exon 1F	Adolescent	4	1149	Yes	26080088↑, 25056599↑, 26822445↑, 29921868↑
				Adult	3	1668	No	25080589↑, 23449091↑, 30144780↓,
			Promoter	Adult	2	340	–	27378548↓, 22295073↑
		SLC6A4	Promoter	Child/adolescent	3	939	No	23217646↑, 27218411↓, 31082414↓,
				Adult	2	133	–	24937096-
		FKBP5	Intron 7	Child	2	208	–	29162190↑, 27218411↓,
				Adult	2	147	–	29182159↓, 30144780↓,
	SES	NR3C1	Exon 1F	Adolescent	3	468	No	24713862↑, 27651384↑, 21883162-
		SLC6A4	Promoter	Child/adolescent	2	388	–	21883162-, 27217150↑
		BDNF	Promoter	Child/adolescent	2	109	–	21883162-, 30771753↓

^†Includes candidate gene studies with statistical power >60% for evaluation of reproducibility.

^‡↑ = hypermethylation; ↓ = hypomethylation; - = no DNA methylation changes.

PMID: PubMed identifier; SES: Socioeconomic status.

NR3C1

Owing to the pivotal role played by the glucocorticoid receptor in HPA axis activity, the gene NR3C1, especially the region of NR3C1 exon1F, has been studied extensively (Table 2 & Supplementary Table 1). Hypermethylation of NR3C1 exon 1F had the strongest support for an association with psychological stress in 83% (15 out of 18) of prenatal maternal stress and 90% (19 out of 21) of postnatal stress (Table 2). Reproducibility with regard to the results of child maltreatment was the highest (11 out of 11 studies) regardless of the difference in age at sampling (child, adolescent or adult). This result was consistent with a previous systematic review showing that hypermethylation of NR3C1 exon variant 1F was significantly associated with adversity in early life in 89% of human studies and 70% of animal studies [36]. When these studies were grouped by sampling time, the consistency further improved. Hypermethylation of NR3C1 exon 1F was consistently associated with psychological stress when the time of sampling was close to the time of stress exposure. In addition to exon 1F, other loci, such as those located around the promoter area of NR3C1, were investigated by several studies. However, owing to the much smaller number of this type of study, consistency could not be evaluated. A previous meta-analysis found that maternal psychosocial stress during pregnancy can significantly alter DNA methylation of NR3C1 promoter in offspring [37]. Therefore, the association between hypermethylation of NR3C1 exon 1F/promoter and prenatal and early life stress, especially for the stressor of child maltreatment, is well supported.

SLC6A4

SLC6A4 is the second most studied candidate gene. There were a total of six, twelve and four studies that examined the DNA methylation of the SLC6A4 promoter area for its association with psychological stress during prenatal development, postnatal development, and adulthood, respectively. Results were inconsistent among three studies on the DNA methylation signature impact in offspring by prenatal maternal depression. Both hypomethylation [38,39] and no association [40] of SLC6A4 promoter in offspring were reported. Similarly, the DNA methylation changes of SLC6A4 with regard to adulthood were also contradictory [41–44]. However, when child trauma was examined, the authors found that hypermethylation of the SLC6A4 promoter was identified in five out of five studies regardless of stress type – namely, child sexual abuse [45,46], child trauma and physical abuse [47,48] and being bullied [49] (Table 2 & Supplementary Table 1).

Other candidate genes

Fewer studies have been published for other candidate genes, including FKBP5, OXTR, and BDNF, than for NR3C1 and SLC6A4. The results from studies on these genes were inconsistent, with the effect size and direction of associations varying widely across studies (Table 2). For example, hypomethylation of FKBP5 intron 7 was reported in four publications [30,50–52], whereas four other studies did not detect DNA methylation change for FKBP5 [53–56] despite the fact that these studies all involved childhood trauma/maltreatment (Table 2). The genes BDNF and OXTR did not have enough studies to evaluate reproducibility.

Reproducibility of EWAS

In recent years, a growing number of EWAS have investigated the DNA methylation mechanisms caused by different types of psychological stress. EWAS results were presented at individual CpG sites, differentially methylated genes and biological pathways that were enriched for the genes with DNA methylation changes. Individual CpG sites had to reach a predefined threshold (e.g., p < 1e-6) to be considered significant. By contrast, the pathway analysis could include genes with a loose statistical criterion for DNA methylation changes. For example, pathway enrichment analysis could include all genes with loci of p < 0.05 instead of 1e-6 and therefore involve genes that did not meet genome-wide significance.

Power estimation is critical to ensure robust findings. Few of the EWAS evaluated here assessed statistical power. The authors calculated the statistical power with G power for each study based on the effect size of the strongest signals each study reported, which was susceptible to the winner's curse. 47 EWAS with power >75% were used for assessment of the reproducibility of individual CpG sites, whereas 20 could not be used because they were underpowered (power <75%) or did not provide information for power estimates (Supplementary Table 2).

Individual CpG sites

Several relatively large EWAS failed to detect a significant association between DNA methylation and psychological stress when several types of adversity were combined [57,58]. In an EWAS of 1656 adolescents from the E-Risk Longitudinal Twin Study, no significant differential methylation gene (DMG) associations were identified with poly-victimization (e.g., maltreatment, neglect, witnessed violence) across childhood and adolescence [57]. Similarly, when prenatal maternal stress, including life stress, contextual stress, personal stress and interpersonal stress, was investigated in the Generation R cohort (n = 912) and Avon Longitudinal Study of Parents and Children (ALSPAC) cohort (n = 828), no association survived Bonferroni correction [59].

The results of reported stress-associated individual CpG sites had poor consistency. A total of 11 EWAS related to child maltreatment accumulated only 13 DMGs, which were reported by two studies. A study of 96 maltreated and 96 control children identified 2868 differentially methylated positions (DMPs) associated with child maltreatment (p < 5.0 × 10^-7), which is the largest number of DMPs reported for stress to date [60]. Another study with 94 maltreated and 96 healthy, non-traumatized children reported 33 DMPs, including 11 DMPs that withstood Bonferroni correction [61]. However, there were no DMPs in common between these two studies. Moreover, out of the 2868 DMPs reported by Yang et al. [60], only nine were also associated with childhood sexual abuse in another study of 1656 adolescents [57]. The 0.3% overlap is likely by chance. In another EWAS, DNA methylation of cg27512205 in the KITLG gene in blood from 85 adults was linked to childhood trauma and mediated the association between cortisol stress reactivity and child trauma. This effect was further validated in both blood and buccal samples from another 300 adults [62]. However, Frach et al. failed to replicate the mediation effect of cg27512205/KITLG methylation in CD14+ monocytes [63].

Several CpG loci at genes were replicated by one independent study, which may claim limited success in replication [64]. For example, EWAS of the ALSPAC and NSHD cohorts, with 1332 subjects, reported DNA methylation of cg19335412 (near ACTA2) and cg27347930 (near LIG3 and CDS) associated with sexual abuse. DNA methylation of cg00973947 (near C3orf58) was associated with adverse childhood experiences. These three loci were also reported by Suderman et al. [65], Marzi et al. [57] and Prados et al. [66].

Pathway level

The authors compiled the biological functions or pathways reported by each EWAS (Table 3) and found that psychological stress in pregnant women during the prenatal period had a broad impact on infants' DNA methylation with regard to biological pathways of brain development, transcription regulation and stress response. However, for one specific type of prenatal stress, either by maternal depression/anxiety, or by low SES or general psychological stress, the pathways where DMPs were enriched were not replicated very well. When psychological stress during the postnatal development period was examined, several reproducible pathways appeared.

Table 3. . Numbers and enriched function of epigenome-wide association studies associated with psychological stress across the life span.

Time of stress	Stress type	Age at sampling	Number of studies†	Number of studies‡	Functional pathway§
Prenatal development	Maternal depression/anxiety	Infant	10	5	Brain development (1); transcription regulation (1)
	Low SES		3	3	Transcription regulation (2); stress response (2); immunity (1); brain development (1); metabolism (1); cancer (1)
	Other psychological stress		8	7	Brain development (1); metabolism (1); transcription regulation (1); stress response (1); psychiatric disorder (1)
Postnatal development	Child maltreatment	Child/adolescent	5	4	Stress response (3); physical/psychiatric disorder (3); brain development (1)
		Adult	11	7	Brain development (4); transcription regulation (4); cellular signaling (3); stress response (2); immunity (2); metabolism (2)
	Low SES	Child/adolescent	3	3	Immunity (2); cellular signaling (2); tissue growth (1); metabolism (1)
		Adult	7	5	Immunity (3); cellular signaling (2); metabolism (1); brain development (1); growth development (1)
	Other psychological stress	Child/adolescent	8	6	Immunity (3); brain development (2); cellular signaling (2); stress response (2); metabolism (2)
		Adult	4	3	Brain development (1); cellular signaling (1); immunity (1)
Adulthood	Work-related stress	Adult	4	2	Cancer-related pathways (2); immunity (1)
	Low SES		4	1	Immunity (1); stress response (1)

^†Original EWAS included in PRISMA selection.

^‡Subset of EWAS with statistical power >75%.

^§Numbers in brackets refer to the repeated number of DMGs enriched in the pathways. Biological pathways of immunity were repeated across three lifespan periods.

DMGs: Differential methylation genes; EWAS: Epigenome-wide association studies; SES: Socioeconomic status; PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Replicated biological pathways associated with specific stress during childhood

Low SES early in life leads to DNA methylation changes, mostly related to immune function in children and adolescents [67,68] and adults who were exposed to both low SES and other psychological stress during their childhood [69–71]. DMGs were enriched in pathways associated with immunity and inflammation in five of eight childhood low SES studies. The diverse study designs, including the participants' origins, SES of the corresponding society, measurement of SES and other technical and analytical variations (experimental platforms, choice of covariates), may explain why three of the studies did not show enrichment for immunity.

Child maltreatment contributed to DNA methylation changes associated with stress response and brain development pathways. When these data were further partitioned according to participants' sampling age, some interesting findings were disclosed. The samples from children and adolescents exposed to child maltreatment showed that DMGs were enriched in biological pathways of stress response and physical/psychiatric disorder in three studies [60,61,72]. By contrast, samples from adults who suffered from childhood trauma showed that impaired brain structure and plasticity [73–76], transcription regulation [75,77,78] and cellular signaling [73,74,78] were the most affected systems. Based on this observation, the authors formed the hypothesis that early life stress changes the stress response system, which gradually fades through time, but many other long-term effects related to signaling, regulation and brain structure become apparent over time.

Previous studies suggested that specific types of psychological stress might be associated with distinct adult outcomes [79,80] mediated by unique DNA methylation changes [78,81]. This notion is supported by the results of several EWAS [64,67,77,81]. Child abuse has broader impacts than poverty; poverty may have a more restricted impact on immunity. When three types of stressors (childhood abuse, life adversity and SES during childhood) were investigated separately in the same study, the authors found that each type of psychological stress was associated with distinct CpG loci [81]. Similarly, in another study, family income, parental education level and family psychosocial adversity were associated with 488, 354 and 102 DMPs, respectively, and only nine DMGs were shared across the three types of stressors [67]. Interestingly, a study comparing the biological and behavioral effects triggered by threat (e.g., violence) and deprivation (e.g., neglect or food insecurity) found that threat-related early life adversity was associated with accelerated DNA methylation age and advanced pubertal stage, whereas deprivation-related early life adversity was uniquely associated with delayed pubertal stage [82]. All of these results highlight the complexity and heterogeneity of stress. Stress is not one entity. It is therefore necessary to study the epigenetic biological mechanism of different types of stress separately.

Childhood stress consistently impacts several biological pathways. Childhood is known to be the most vulnerable time for detrimental effects associated with environmental stressors [83–85]. Stress during postnatal development affects the development of neuronal circuits that support complex behavior.

Biological pathways shared by distinct stress types across the life span

Shared biological pathways have been consistently identified with different stressors. The shared pathways include immunity, stress response, metabolism, brain development, cellular signaling and transcription regulation. Specifically, the function of immunity was the most widely implicated biological system across all three time periods with regard to almost all stress types (Table 3), including natural disaster exposure [86], child maltreatment [75,87], low SES [67–69,71], early deprivation [88] and chronic work stress during adulthood [89]. The impact of psychological stress on stress response and immunity is universal. In a study conducted by Needham et al., DNA methylation signatures of 18 genes related to stress or inflammation in 1264 adults were examined [44]. After controlling for covariates and multiple tests, low SES during adulthood was found to be significantly associated with DNA methylation in one stress-related gene (AVP) and five inflammation-related genes (CD1D, F8, KLRG1, NLRP12, TLR3). Accumulating data have demonstrated that early life psychological stress correlates with activation or blunting of neurobiological response systems, including the immune system, and HPA axis functioning. The HPA axis may link psychological stress to impaired physical and mental health and many other biological processes.

Cellular signaling is another pathway frequently enriched with DMGs as a result of various psychological stresses during the postnatal period, such as child maltreatment [73,74,78], low SES [90,91], maternal deprivation [88] and adversity in early life [92]. Cellular signaling may be tightly coupled with brain development, which is also widely linked to various early life stresses during the prenatal and postnatal periods (Table 3).

Candidate genes revisited in EWAS

The reproducibility of findings for candidate genes was also examined in EWAS and found to be poor. One EWAS showed significantly reduced DNA methylation of BDNF, NR3C1 and FKBP5 in association with child abuse [61]. Another EWAS reported an association between a few CpG sites of FKBP5 and BDNF and the psychological stress of poly-victimization early in life. However, the E-Risk study and two other EWAS failed to validate the DNA methylation changes of candidate genes when testing NR3C1, FKBP5, BDNF and SLC6A4 in poly-victimized youth [57] or DNA methylation of these candidate genes with parental stressors [93] or SES [68]. Nevertheless, this does not contradict the findings in the candidate gene studies because some specific loci in the candidate genes were not covered by the Illumina array [37,61].

Problems & challenges in DNA methylation studies of psychological stress

The authors reviewed human DNA methylation signatures associated with psychological stress in studies of candidate genes and EWAS. The reproducibility of findings at individual loci was low. Only two out of a dozen EWAS detected the same CpG DNA methylation that was associated with stress. With the exception of NR3C1, studies of candidate genes did not yield replicable results and were rarely validated by EWAS. Small sample size, small effect size, lack of standardized objective measures of stress and statistical analysis, and other technical issues may be the major causes of the poor reproducibility.

Although the candidate gene approach was inefficient or ineffective for studying complex behaviors or diseases, EWAS were still underrepresented in the existing human studies. Moreover, many studies did not have adequate statistical power. Effect size and statistical power were sometimes ignored [58,94]. It has been suggested that to detect a 10% mean methylation difference in a case–control design, 112 cases and 112 controls would be the minimum sample size needed to reach 80% power at a genome-wide significance threshold of p = 1 × 10^-6; similarly, at least 98 monozygotic twin pairs would be needed [95]. However, only a few EWAS had a sample size that met these criteria [57,68,87]. Small sample size limited the ability to detect small changes that may be present. Even the few EWAS with large sample sizes failed to detect significant signals [57,59]. The actual DNA methylation changes caused by psychological stress may be minute, and it will require much larger sample numbers to detect these small but possibly important differences. The winner's curse may explain why few EWAS findings can be replicated. Ideally, the findings from a discovery data set should be replicated in other independent cohorts to validate results. However, only approximately 8.5% (four out of 47) of existing EWAS demonstrated replication with other data sets [57,64,96,97].

Inconsistent application of multiple testing corrections was observed in the DNA methylation literature. Most of the EWAS (44 out of 47; 94%) addressed multiple testing corrections according to the number of tested loci. Three studies used p = 0.05 [66,91,98]. The candidate gene studies tended to ignore the requirement of a stringent criterion when it was necessary. One-third of candidate studies did not apply multiple testing correction when multiple loci were examined simultaneously.

Regarding the practice of statistical procedures, covariates are a primary area of concern. Covariates such as sex, age, genetic background and ethnicity have been reported to affect the relationships between psychological stress, epigenetic patterns, gene expression and behavioral outcome [99–101]. Of the 47 EWAS studies reviewed here, most (39 out of 47) controlled some covariates, and 84% (37 out of 44) controlled batch effects when Illumina was used, but less than one-third (16 out of 44) controlled the positional effect [102], leaving space for false positives and irreproducible results.

Sex is another important covariate in DNA methylation studies. Sex-specific DNA methylation programming has been widely reported in rodents [103,104] and humans [105–107]. However, sex differences were largely ignored or neglected in the existing human studies, with notable exceptions, including some candidate gene studies involving OXTR [43,108], NR3C1 [109–112], SLC6A4 [40,113], IGF2/H19 [114], MAOA [115] and MEG3 [114]. Similarly, only a few EWAS considered the sex-stratified methylation effect [116,117], whereas most EWAS regressed sex out as a covariate.

Genetic factors can influence DNA methylation [118–120], but these are often overlooked. The genetic background was rarely considered in DNA methylation studies, with the exception of a few candidate gene studies [30,47,51,56,121–124]. Several studies used self-reported race or ethnicity as a covariate [60–62,67,69,72,74,75,81,96,98,116,125,126] or used cross-ethnicity comparison [62,87,114,127]. In the initial evaluation of 67 EWAS, 29 studies were conducted using European participants, and 38 studies were conducted using non-European participants. Some studies used participants of mixed racial origin. Genetic differences can mediate the stress response directly and can also interfere with analytical results and therefore should at least be considered a covariate.

Diverse methods were used to measure psychological stress. Of the 47 EWAS across the life span, 21 measured psychological stress by questionnaires, 14 used semi- or structured interviews and 12 used information from multiple data sources to create composite adversities. The Childhood Trauma Questionnaire score was commonly used for measuring child maltreatment [66,72,74,75,78]. Alternatively, the Maltreatment Classification System was used, which not only includes the subtypes but also takes frequency, subtype severity and developmental periods of child maltreatment into consideration [87]. Additionally, multiple informants and data sources of child maltreatment [60,61,67] were used. The same situation was true in candidate gene studies. Similarly, the psychological stress of SES was evaluated in different ways across studies, with maternal education level [44], a composite measurement of the parents' education, household income and asset status [70] being adopted. Although the consistency or correlation among these different measures remains to be established, these heterogeneous measurements across studies likely contributed to the poor reproducibility of results.

DNA methylation profiling also used diverse platforms, including EPIC, HumanMethylation450, HumanMethylation27, MassARRAY, and EpiTYPER, as well as methylated DNA immunoprecipitation and pyrosequencing. Different profiling methods do not cover the same CpG sites, precluding a direct comparison for many CpG sites. CpH and 5hmC have rarely been explored for DNA methylation response to stress, and 5hmC is rare in peripheral tissues but common in the brain. Whole-genome bisulfite sequencing (WGBS) and oxidative WGBS are currently the most promising technologies for profiling these DNA methylation forms to the best extent, but because of their high cost, they have not been used in stress studies as of yet.

Tissue and cell-type specificity of the response to stress are clear. A few rodent experiments have indicated that some candidate genes' DNA methylation changes from blood can be used to model changes in the brain [128,129]. The correlation of DNA methylation signature between peripheral tissues and the brain is, in general, low and varies greatly across loci [130]. Most existing epigenetic studies were based on DNA methylation from one or two peripheral tissues, with just a few comparing across different tissues or cells [64,76,131,132]. The most significant problem is that only 42% (20 out of 47) of EWAS adjusted for cell composition (Supplementary Table 2). Failure to address cell composition in EWAS could lead to false findings and poor reproducibility.

Recommendations for future studies

To reduce heterogeneity and identify reproducible DNA methylation biomarkers of psychological stress, the authors put forward several recommendations for future studies. Sample size and statistical power are the first issues to address. Large consortia or collaborations could facilitate sizable studies. Several large longitudinal cohorts have been created. The E-Risk Longitudinal Twin Study and ALSPAC in the UK have open policies for collaboration and have made some outstanding contributions in the field of DNA methylation due to psychological stress [57,59,64,68]. Collaborative projects can standardize laboratory protocols and analytical procedures to improve data comparability and consistency and reduce heterogeneity. Investigators with limited resources or small cohorts should consider joining or forming consortia, or at least following the same protocol used by large consortia, to avoid producing spurious findings. Based on the authors' evaluation, large cohorts with thousands of participants may be needed to obtain robust results.

The existing human DNA methylation literature is predominantly cross-sectional for both candidate gene studies and EWAS (Table 1). Longitudinal studies are preferred to minimize heterogeneity when paired tests can be performed on pre- and post-stress data. The unique impact of different types of psychological stress occurring during a lifetime will be better studied in cohorts with sampling at multiple time points throughout the life span.

Covariates, including technical variables, such as batch and positional effects in array methods, and biological variables, such as sex, age, race (or ethnicity), genetic background, and cell-type composition, need to be properly handled in statistical analyses. Future work would also benefit from increased attention to racial diversity. Epigenetics study of the stress response should provide more sex-specific and cell type-specific biology. At the same time, addressing other unknown covariates should also be considered. Methods such as surrogate variable analysis [133], probabilistic estimation of expression residuals [134], and hidden covariates with prior knowledge [135] are available to address these issues.

Experimental platforms like WGBS will cover all CpG/CpH sites in the genome, making data from different studies comparable. A dramatic reduction in the experimental costs of WGBS through methodological improvements is necessary. The significance criteria have also been a concern, especially with regard to studies like EWAS when a huge number of CpG loci are examined simultaneously. Following a strict standard with a false discovery rate or Bonferroni correction will reduce the probability of false findings.

Ultimately, the study data associated with psychological stress per se are critical, not only because different types of stress may have different biological effects but also, more importantly, because different data collection methods could result in data of variable quality. To date, the majority of studies in humans have relied on a questionnaire, which is subjective and less reliable than objective measures such as cortisol level. The objective measurement of stress will be the most challenging aspect of stress research to improve, but it is a critical component for future success in understanding the biology of psychological stress.

Conclusion

In summary, with the exception of the reproducibility of studies identifying hypermethylation of NR3C1 exon 1F in relation to child maltreatment and early life adversity, the existing studies produced mostly irreproducible results with regard to DNA methylation changes at genes or CpG sites linked to the psychological stress. Pathway analyses offered intriguingly consistent findings. Future large-sample, longitudinal studies with objective measures of stress, advanced experimental platforms, optimized analytical procedures that control covariates and strict significance criteria should enhance our understanding of the DNA methylation alterations associated with psychological stress and their contribution to mental health.

Future perspective

This review analyzed the strengths and weaknesses of previous studies, and it paves the way for future studies to map the epigenetic effects of psychological stress so that the relevant biological pathways may be revealed. More studies with large sample sizes, longitudinal study design, optimal quality control and statistical procedures will be needed to investigate the DNA methylation changes of each well-defined type of psychological stress. Reproducible findings should be the goal. Epigenome-wide methods are preferred until reproducible findings of specific loci are identified for follow-up functional studies. Future research should ultimately address the diagnostic and prognostic potential of DNA methylation changes as biomarkers for stress-related psychiatric disorders.

Executive summary.

• DNA hypermethylation of NR3C1 exon 1F associated with early life stress was reproducible.

• Associations between DNA methylation of other candidate genes and psychological stress were not replicated.

• The results of epigenome-wide studies of stress had poor consistency at individual CpG sites.

• Child maltreatment is associated with DNA methylation changes related to stress response and brain development pathways.

• Low socioeconomic status early in life is associated with DNA methylation changes related to immune function.

• The pathways of stress response, brain development and immunity have been consistently identified as being associated with psychological stress.

• The poor reproducibility of DNA methylation changes associated with psychological stress may arise from small sample size, diverse research design, lack of standardized statistical analysis and suboptimal quality control.