DNA methylation as a mediator of associations between the environment and chronic diseases: A scoping review on application of mediation analysis

ABSTRACT

DNA methylation (DNAm) is one of the most studied epigenetic modifications. DNAm has emerged as a key biological mechanism and biomarkers to test associations between environmental exposure and outcomes in epidemiological studies. Although previous studies have focused on associations between DNAm and either exposure/outcomes, it is useful to test for mediation of the association between exposure and outcome by DNAm. The purpose of this scoping review is to introduce the methodological essence of statistical mediation analysis and to examine emerging epidemiological research applying mediation analyses. We conducted this scoping review for published peer-reviewed journals on this topic using online databases (PubMed, Scopus, Cochrane, and CINAHL) ending in December 2020. We extracted a total of 219 articles by initial screening. After reviewing titles, abstracts, and full texts, a total of 69 articles were eligible for this review. The breakdown of studies assigned to each category was 13 for smoking (18.8%), 8 for dietary intake and famine (11.6%), 6 for other lifestyle factors (8.7%), 8 for clinical endpoints (11.6%), 22 for environmental chemical exposures (31.9%), 2 for socioeconomic status (SES) (2.9%), and 10 for genetic factors and race (14.5%). In this review, we provide an exposure-wide summary for the mediation analysis using DNAm levels. However, we found heterogenous methods and interpretations in mediation analysis with typical issues such as different cell compositions and tissue-specificity. Further accumulation of evidence with diverse exposures, populations and with rigorous methodology will be expected to provide further insight in the role of DNAm in disease susceptibility.

Introduction

Epidemiology is the study of the distribution and determinants of health-related states and events in specified populations and the application of this study to the control of health problems. The ultimate purpose of an epidemiological study is to explore the causation between exposure (X) and health outcomes (Y) since such work is essential to establish suitable preventive methods to intervene in the exposure. There is accumulated epidemiological evidence on the causal relationship between environmental factors including behaviours and sub-clinical changes and disease onset. In particular, for non-communicable diseases, previous studies have focused on individuals’ lifestyle factors such as smoking, dietary habits and nutritional intake, alcohol consumption, physical activity and the chemical environment as preventable exposures [1–5]. In addition to these factors, epidemiological studies have been conducted to investigate genetic factors, socioeconomic status, social determinants, and climate change as potential causes of disease in recent years [6–9]. Of existing epidemiological research, much evidence has examined associations between exposures and outcomes, but less work has been done on causal pathways and mechanisms of observed associations. However, with recent advances in methodologies for causal inference [10] and high-throughput multiomics measurements [11], it is becoming the next frontier of the field to infer biological mechanisms and pathways underlying the associations between exposures and disease.

DNA methylation (DNAm) is one of the most studied epigenetic modifications and has crucial roles in regulating various cellular processes. DNAm refers to the addition of a methyl group to the 5th position of the cytosine base by enzymes such as DNA methyltransferases (DNMTs). In general, DNAm occurs predominantly at the promoter region of genes because ~70% of all genes have a genomic region of ∼1 kb that is rich in CpG dinucleotides, referred to as CpG islands [12]. Epigenetic modification occurring in promoter regions can regulate gene expression and may influence phenotypic outcomes. Since the late 1990s’, many researchers have investigated clinical associations between DNAm and various diseases. Some earlier studies showed that methylation status of a tumour-associated genes within promoter regions contributed to gene inactivation of tumour suppressors [13,14]. In the past decade, Epigenome-Wide Association Studies (EWAS) have investigated associations between DNAm levels across the genome and identified epigenetic marks of disease [15–17]. In addition, other epidemiological studies have also been performed to examine environmental factors associated with DNAm levels. Earlier studies have indicated that nutritional intake of folate plays a crucial role in one-carbon metabolism and changes global DNAm levels [18–21]. A broader range of environmental exposures such as smoking, dietary intake, alcohol intake, maternal behaviour, and environmental pollution have investigated whether these factors can alter global or gene-specific DNAm levels [22–30]. Referring to these previous studies, DNAm has the potential to be a suitable factor connecting biological mechanisms underlying a relationship between exposure and outcomes as a mediator of observed relationships.

Recent developments of causal inference methods have made significant contribution to infer causality in observational studies. Of the proposed causal inference methods, mediation analysis allows us to decompose the total effect from exposure to outcome into direct and indirect effects through a mediator [31].

Considering recent findings, in this scoping review, we will first briefly introduce the essence and implementation of statistical mediation analysis. We then review emerging epidemiological studies, by different types of exposures that applied mediation analysis to explore the causal pathways between exposure and outcome through DNAm (Figure 1). Finally, we will discuss the potential limitations of this method and further expansion and use in epidemiological studies.

Figure 1.

Conceptual figure of epidemiological studies for DNA methylation a) without mediation analysis and b) with mediation analysis.

Mediation analysis

The motivation of statistical mediation analysis is to evaluate whether an intermediate variable is on the pathway between the exposure of interest on an outcome. In mediation analysis, the total effect, which is the effect of exposure on the outcome, is decomposed into direct and indirect effects. A direct effect is the effect of exposure on an outcome in causal paths without mediators. An indirect effect is the effect of exposure on an outcome in causal paths through mediators. Considering a quantitative decomposition of effects, this methodology is matched by statistical analysis methods to estimate the mediated effects of DNAm. This section consists of the following four parts: 1) we will introduce several mediation approaches and their characteristics; 2) we will describe the effects estimated by mediation analysis in the context of DNAm; 3) we will illustrate aforementioned points with concrete examples; 4) we will summarize mediation analysis with potential pitfalls.

Approaches in mediation analysis

Mediation analysis can be classified into traditional methods, including the difference method, the product method, and structural equation models (SEMs); and causal mediation framework using potential outcomes [32,33].

Traditional methods are model-based approaches [34]. The difference method consists of two parametric models: an outcome model with mediator and an outcome model without mediator. Similarly, the product method consists of two other parametric models: an outcome model with mediator and a mediator model. Direct and indirect effects are computed from the estimates of each model in both methods. The product method is referred to as Baron and Kenny’s method and is equal to the difference method with continuous outcome and mediator using a linear regression model without exposure-mediator interactions [35]. SEM is an extension of the product method and generally used in the area of path analysis [36]. The path analysis can examine relationships among many variables with flexible modelling and handles nonlinearity, such as logistic regression models, and interactions. Path analysis defines models for each variable and defines the effects using the estimates obtained. These notations of each effect vary among models, and it is difficult to express it uniformly [37]. Therefore, the decomposition of the effects of path analysis remains ambiguous.

Causal mediation framework is a model-free approach [34], which extends the definition of effects using potential (counterfactual) outcomes [38] to a causal structure with the mediator. Like the potential outcome framework, this framework divides the process into three parts to obtain estimates for solving the research question (RQ). 1) Decision estimand: Define what parameters you want to estimate to solve the RQ. This is referred to as estimand. 2) Identification: Identify the assumptions under which estimand can be validly estimated and confirm that these assumptions are satisfied from data. 3) Estimation: Modelling the question and estimating the parameters. As a remarkable point of causal mediation framework, the process of deciding estimand provides a more explicit definition of effects and decomposition than traditional methods. We will dig into this process in the following subsection.

Effects of causal mediation framework

The Causal mediation framework defines several effects with the potential outcome framework. Although traditional decomposition is two effects: direct and indirect effects as shown in many reports [39], recent research has shown that it can be divided into three and four effects that may give us more insight [40,41]. In this subsection, we will describe the decomposition effects estimated by causal mediation framework using Fasaneli’s example [42].

Suppose that exposure (e.g., smoking) affects the outcome (e.g., lung cancer) and DNAm, and DNAm affects the outcome. In this case, DNAm is a mediator. Now, let our RQ be, ‘How much does smoking affect lung cancer?’ Then the effect is defined as the difference between the risk of lung cancer if the entire population smoked and the risk of lung cancer if the entire population did not smoke. This effect is generally referred to as the average treatment effect (ATE) or average causal effect (ACE), and in mediation analysis it is referred to as the total effect (TE). From here, we will decompose this total effect.

Firstly, let our RQ be, ‘Is there a mediated effect? How much is this effect?’ As described above, the 2-way decomposition is the most common way, decomposing TE into direct and indirect effects. In Fasaneli’s paper, the natural decomposition for this question is as follows. The direct effect is defined as the difference between lung cancer risk if the entire population smoked and lung cancer risk if they did not smoke, with the mediator set to the non-smoking DNA methyl level. This direct effect is called the natural direct effect (NDE). The indirect effect is defined as the difference between lung cancer risk at methylation levels if the entire population smoked and lung cancer risk at methylation levels if the entire population did not smoke, with exposure set to smoke. The indirect effect is called the natural indirect effect (NIE). This decomposition holds the following equation: TE = NDE + NIE [33].

Secondly, it is reasonable to expect an interaction between exposure and the mediator in mediation analysis. Considering the exposure-mediator interaction, TE is decomposed into four parts, 4-way decomposition. These four components consist of a controlled direct effect (CDE), a reference interaction (INT_ref), a mediated interaction (INT_med), and a pure indirect effect (PIE) [41]. Again, try to think about 4-way decomposition with Fasaneli’s example. Each component is described in the context of Fasanelli et al. as in 2-ways [42]. CDE is defined as the effect of smoking on lung cancer due to neither mediation by DNAm nor interaction between smoking and DNAm. INT_ref is defined as the effect of smoking on lung cancer due to interaction between smoking and DNAm only. INT_med is defined as the effect of smoking on lung cancer due to both mediation by DNAm and interaction between smoking and DNAm. PIE is defined as the effect of smoking on lung cancer due to mediation by DNAm only and referred to as mediated main effect. Inoue et al. used directed acyclic graphs to show an intuitive interpretation of these four effects [43]. This decomposition holds the following equation: TE = CDE + INT_ref + INT_med + PIE.

Finally, these components allow us to combine and define interpretive effects, as shown in Figure 2. Pure direct effect (PDE) is represented as the sum of CDE and INT_ref, and is equal to the NDE shown in the 2-way. This combination leads to the 3-way equation as follows: TE = PDE + INT_med + PIE [40]. In addition, INT_med is possible to combine with either PDE or PIE. It combines with PIE to form a total indirect effect (TIE), which is equivalent to the 2-way equation shown: TE = PDE + TIE = NDE + NIE.

Figure 2.

Decomposition of total effect and each interpretation of 4-way, the total effect can be decomposed into 2-, 3-, and 4-way decompositions. The arrows show the relationship between each effect and illustrate the respective interpretation of the 4-way decomposition. TE: total effect; PDE: pure direct effect; NDE: natural direct effect; TIE: total indirect effect; NIE: natural indirect effect; INT_med: mediated interaction; PIE: pure indirect effect; CDE: controlled direct effect; INT_ref: reference interaction.

So far, we have seen the definition of effect. Estimation of the effect is necessary for practical epidemiological studies. Work by TJ VanderWeele provided an approach to calculate the point and interval estimates of each effect [41]. Note that the modelling differs depending on the scale of outcome, exposure, and mediator. The R packages of ‘mediation’ and ‘CMAverse’ allow us to estimate 2-way and 4-way decomposition effects, respectively [44,45]. In SAS, CAUSALMED procedure estimates 4-way decomposition effects [46]. We illustrate these decomposed effects with concrete examples in Supplementary materials.

Pitfalls of mediation analysis

We have focused mainly on the definition of effects, effect decomposition, and briefly on estimation. However, what we have explained so far may not estimate the defined effects, may estimate effects that are not answered in your RQ, or may give the reader the misinterpretation.

First, causal mediation analysis also has assumptions to estimate causally interpret effects, like common causal inference, namely identifiability. The important assumptions to achieve identifiability is conditional exchangeability [41]: 1) the effect of the exposure (X) on the outcome (Y) is not confounded given the covariates (C); 2) the effect of the mediator (M) on the outcome (Y) is not confounded given the exposure (X) and the covariates (C); 3) the effect of the exposure (X) on the mediator (M) is not confounded given the covariates (C); 4) none of the mediator-outcome (M-Y) confounders are affected by the exposure (X). The point is, even if the effect of the exposure on the outcome is unconfounded, mediation causal effects will be biased if there are unmeasured exposure-mediator (X-M) or mediator-outcome (M-Y) confounders; and/or if the exposure affects mediator-outcome (M-Y) confounders. Sensitivity analyses have been proposed to assess the degree of bias [39,47] from stringent assumptions of mediation analysis.

Second, INT_med can combine with either PIE or PDE when researchers define 2-way decomposition. When INT_med combines with PIE, it is assumed that TE = PDE + TIE. This decomposition would be preferred if the question is ‘Is there a mediated effect?’ On the other hand, when INTmed combines with PDE, it is assumed that TE = TDE + PIE, where TDE is total direct effect. This decomposition would be preferred if the question is ‘In addition to the mediated effect, is there a direct effect?’ The RQ in epidemiological studies seem to be more frequently the former question [34,41].

Third, it is helpful to explicitly present the outcome and mediation models required for mediation analysis in the paper to allow the reader to assess whether the decomposition is appropriate and strengths of relationships. If the exposure-mediator interaction term is included in the outcome model, not only 2-way decomposition but also 3-way and 4-way decomposition interpretations are possible.

Methods

Study inclusion/exclusion criteria

As inclusion criteria in this review, studies are population-based studies with the assumption of DNAm levels as mediators. The target populations were all age groups, from infants to adults. The eligible study design was cross-sectional, cohort, case-control, and interventional studies. In addition, we considered that descriptions in statistical analysis or results stated the use of mediation analysis. We applied the following exclusion criteria: 1) DNAm levels as either exposure or outcome variable, 2) experimental studies using cells and animals (no human study), and 3) study types of reviews, case studies, editorials, and conference abstracts. We adhered to the PRISMA extension for scoping reviews (PRISMA-ScR) [48] and provide a checklist as a Supplementary Figure 1.

Search strategy

Until the beginning of December 2020, we performed an initial literature search through four online databases (PubMed, SCOPUS, Cochrane, and CINAHL). We included only peer-reviewed studies published in English. We searched these databases with predefined keywords and terms as provided in Supplementary Table 1.

Study selection

After initial screening, duplicates were removed for reviewing titles and abstracts. Two independent reviewers (RF and YT) screened the title and abstract of the remaining articles and selected them for full-text reading based on predefined inclusion/exclusion criteria. Then, the same reviewers independently read through the remaining articles for full-text scan and data charting. Discrepancies between two independent reviewers (RF and YT) were solved by an additional review from the third individual reviewer (KS). Finally, we got a consensus regarding all articles from all reviewers (RF, YT, and KS).

Data charting

After completion of study selection, the independent reviewers (RF and YT) extracted and summarized information as shown in Table 1. Data are summarized using the information available in each study. The outlined items are basic information of study, demographic characteristics of eligible participants, examined tissue, and results of mediation analysis. If definition and information for demographic variables are not presented within a main document, we searched available information (e.g., supplementary information and published study profile, study website) as feasible. As a result, in some cases we obtained appropriate information, and in other cases we did not know, we note ‘Not Reported (NR)’ to avoid misunderstanding.

Table 1.

Summary of the all reviewed articles.

Exposures	Author & Publish year	Country	Study design	Total participants	Ethnicity	Age, Mean (SD)	Sex (Men/Women)	Exposure	Outcome	Tissue	Examined mediator	Summary for mediation analysis	Ref No
Smoking	Cardenas A, et al. (2019)	Canada	Longitudinal	441 pairs	Caucasian	Maternal 28.3 (4.3)	Maternal 0/441 Child 232/209	prenatal maternal smoking	birth weight	placenta	71 CpGs	Of 71 CpGs, 7 mediated the association between prenatal smoking and birth weight. The strong mediator, cg22638236 in PBX1, mediated the effect of maternal smoking on child birth weight.	49
Smoking	Wiklund P, et al. (2019)	Multiple*	Multiple*	2821	Multiple*	Multiple*	Multiple*	maternal smoking	disease outcomes	whole blood	MIR548F3 (cg15578140) GFI1 (cg09935388) Unknown (cg04598670) GNG12 (cg25189904)	Mediation analysis examining the indirect effect of maternal smoking during pregnancy on Bipolar II Scale and Hypomanic personality scale through differential methylation of cg25189904 in GNG12.	50
Smoking	Hannon E, et al. (2019)	Denmark	Longitudinal	1316 (1263)	Caucasian	Maternal NR Child 6.1 days (3.2 days)	661/655	prenatal maternal smoking	birth weight	dried blood spot sample	GFI1 (cg09935388) AHRR (cg05575921) EXOC2 (cg26889659)	Mediation analysis found that methylomic variation at three DMPs (GFI1, AHRR, and EXOC2) may link to exposure and outcome.	51
Smoking	Witt SH, et al. (2018)	Germany	Longitudinal	282 pairs	NR	Maternal smokers 28.0 (5.3) non-smokers 31.8 (4.8)	138/144	maternal smoking during pregnancy	birth weight	umbilical cord blood	30 smoking-associated CpG sites	The following were found to mediate the effect of maternal smoking on birth weight: cg25325512 (PIM1, p = 0.005); cg25949550 (CNTNAP2, p = 0.008); and cg08699196 (ITGB7, p = 0. 045).	52
Smoking	Morales E, et al. (2016)	Spain	Longitudinal	179	Caucasian	Maternal Unexposed 31.1 (3.9) Exposed 30.8 (4.4)	Maternal 0/179 Child 93/86	maternal smoking	birth weight	placenta	between LINC00086 and LEKR1 (cg27402634) TRIO (cg25585967, cg12294026) WBP1L (cg20340720)	Both cg27402634 and cgcg25585967 showed mediation of the association between maternal smoking and birth weight.	53
Smoking	Küpers LK, et al. (2015)	Netherlands	Longitudinal	255	European	Maternal Unexposed 31.1 (3.6) Exposed 29.7 (4.7)	Maternal 0/255 Children 136/119	maternal smoking during pregnancy	birth weight	umbilical cord blood	8 GFI1 CpG sites	With two other replications, three GFI1 CpGs (cg09935388, cg14179389, cg12876356) partly mediated the effect of maternal smoking on birthweight.	54
Smoking	Chen Z, et al. (2020)	Multiple*	Cross-sectional	507	Multiple*	Multiple*	Multiple*	tobacco smoking	gene expression	whole blood	AHRR (cg14817490) GPR15 (cg1985920) Other 372CpG sites	Decreasing AHRR and GPR15 DNA methylation mediated association between tobacco smoking and gene expression.	55
Smoking	Jordahl KM, et al. (2019)	US	Longitudinal	457	Asian/Pacific Islander, Black/African American, Hispanic, Non-Hispanic White, Other	NR	0/457	cigarette smoking	bladder cancer	whole blood	AHRR (cg05575921) F2RL3 (cg03636183) GPR15 (cg19859270)	The overall proportion of the excess relative risk mediated by cg05575921 was 92% (p-value = 0.004) and by cg19859270 was 79% (p-value = 0.02).	56
Smoking	Liu Y, et al. (2019)	China	Cross-sectional	500	East Asian	low 8-OHdG 39 (32–45) High 8-OHdG 41 (37–46)	445/55	smoking & occupational polycyclic aromatic hydrocarbons	oxidative DNA damage	whole blood	CYP1A1	Mediation analysis suggested CYP1A1 hypomethylation could explain 13.6% of effect of high 8-OHdG related to smoking and 1-OHP co-exposure (p = 0.047).	57
Smoking	De Vries M, et al. (2018)	Netherland	Cross-sectional	658	European	46 (22–79)	375/283	cigarette smoke	lung function	whole blood/lung tissue	AHRR (2 CpG sites) F2RL3 (1 CpG site) GFI1 (3 CpG sites) PRKAR1B (1 CpG site) NA (8 CpG sites)	The mediation analysis revealed that 10 CpG-sites were significantly associated with lung function levels.	58
Smoking	Wahl A, et al. (2018)	Estonia, Germany	Cross-sectional	KORA1594 EGCUT 148	Caucasian	NR	NR	smoking	IgG glycosylation	whole blood	GNG12 (cg25189904) ALPPL2 (cg05951221, cg21566642 and cg01940273) AHRR (cg05575921) Intergenic (cg06126421) F2RL3 (cg03636183)	Mediation analyses with respect to smoking revealed that the effect of smoking on IgG glycosylation may be at least partially mediated via DNA methylation levels at these 7 CpG-sites.	59
Smoking	Meng W, et al. (2017)	Sweden	Cross-sectional	393	Caucasian/Asian	NR	NR	smoking & rs6933349	antibody-positive rheumatoid arthritis	whole blood	cg21325723	cg21325723 methylation may be a potential mediator of the GE interaction between rs6933349 and smoking in the risk of developing of ACPA-rheumatoid arthritis.	60
Smoking	Fasanelli F, et al. (2015)	Norway	Longitudinal	264	European	NR	0/264	smoking	lung cancer	whole blood	AHRR (cg05575921) and F2RL3 probes (cg03636183)	Alterations of AHRR and F2RL3 may mediate the carcinogenic effect of tobacco exposure in lung cancer aetiology.	42
Dietary intake & famine	Jiang W, et al. (2020)	China	Longitudinal	2909	East Asian	Parent 52.7 (2.6)/53.1 (4.6) Offspring 26.2 (5.2)/26.4 (4.5)	Parent 1240/641 Offspring 596/432	prenatal famine (maternal & paternal)	ΔeGFR	whole blood (leukocyte)	AGTR1 (cg13528513, cg20906621) PRKCA (cg17160506)	The mediation analysis showed methylation alterations of AGTR1 and PRKCA (F2) mediates the famine (F1)-eGFR association (F2).	61
Dietary intake & famine	Shen L, et al. (2019)	China	Longitudinal	188	East Asian	54.2 (2.3)	91/97	early famine	total cholesterol	whole blood (buffy coat)	IGF2 (CpG1)	Mediation analysis demonstrated that the mediation path through the CpG1 explained 5% (p = 0.30) of the association between famine severity and TC.	62
Dietary intake & famine	Bustamante AC, et al. (2018)	US	Longitudinal	112	European American/African American/Other	50.7 (13.0)	50/62	child maltreatment	depression	whole blood	FKBP5	FKBP5 DNA methylation did not mediate the relationship between CM and depression.	63
Dietary intake & famine	Tobi EW, et al. (2018)	Netherlands	Longitudinal	811	Caucasian	Controls 58.0 (5.4) Famine exposure 58.9 (0.5)	Controls 199/264 Famine exposure 160/188	prenatal famine exposure	adulthood metabolic disease (BMI, TG, glucose metabolism)	whole blood	BMI: PIM3 TG: TXNIP, ABCG1, LOC100132354, PNPO, LRRC8D, SYNGR	DNAm at PIM3 (cg09349128) mediated the association between famine exposure and BMI. DNAm at six CpGs, including TXNIP (cg19693031) and ABCG1 (cg07397296) mediated the association between famine exposure and TG.	64
Dietary intake & famine	Wang Z, et al. (2020)	China	Longitudinal	235	East Asian	Unexposed 54.6 (2.6) Exposed 55.1 (0.8)	117/118	prenatal famine exposure	adulthood waist circumference	whole blood (leukocyte)	INSR (9 CpG sites) IGF2 (8 CpG sites)	Mediation analysis demonstrated that mediation path through the CpG7 explained 32% of the association between prenatal famine exposure and waist circumference in adulthood.	65
Dietary intake & famine	Fujii R, et al. (2019)	Japan	Cross-sectional	225	East Asian	40–91 years	108/117	dietary vitamin C intake	HDL-C	whole blood	ABCA1	Mediation analysis showed a significant indirect effect of vitamin C intake on HDL cholesterol levels through leukocyte ABCA1 DNA methylation levels.	67
Dietary intake & famine	Lai CQ, et al. (2020)	US & Spain	Cross-sectional	3954	European	GOLDN 48.3 (16.4) FHS 66.2 (8.9) REGICOR 63.2 (11.7)	GOLDN 468/510 FHS 1056/1275 REGICOR 316/329	dietary carbohydrate, fat intake	metabolic disease	whole blood (lymphocyte)	CPT1A (cg00574958)	Dietary CHO intake reduced the risk of metabolic syndrome through higher CPT1A DNA mehtylation, whreas dietary FAT intake increased the risk of Mets through lower CPT1A DNA mehtylation.	68
Dietary intake & famine	Sherwood WB, et al. (2019)	UK	Longitudinal	297	Caucasian	NR	107/190	breastfeeding duration	childhood adiposity (BMI trajectory)	whole blood	LEP (cg23381058)	Mediation analysis showed no significant effect of DNAm of LEP between brestfeeding and BMI trajectories in 10-year-old child.	69
Other lifestyle factors	Shu C, et al. (2020)	US	Longitudinal	1,435 (467)	Caucasian/African American/Others	48.8 (8.1)	1399/36	cocaine use	HIV severity	whole blood	MX1 (cg26312951) PARP9 (cg08122652, cg22930808) NLRC5 (cg07839457) CX3CR1 (cg22917487) EPSTI1 (cg03753191) IFIT3 (cg06188083) TAP1 (cg08818207) IFITM1 (cg03038262) TAP2 (cg22940798) UXS1 (cg00096307) RIN2 (cg26396492) C2orf67 (cg22385827)	Significant mediation effect of MX1, PARP9, NLRC5, and CX3CR1 DNA mehtylation between persistent use of cocaine and HIV sensitivity.	70
Other lifestyle factors	Liu X, et al. (2020)	China	Cross-sectional	1032	Asian	T2DM 59.4 (8.7)/ Control 59.6 (8.6)	448/584	sedentary time	Type 2 diabete mellitus	whole blood	SOCS3 (Chr17: 76,356,190)	The association between sedentary time and T2DM was mediated by some degree of methylation level at Chr17:76,356,190 site.	71
Clinical endpoints	Ochoa-Rosales C, et al. (2020)	Germany, UK, Netherlands	Cross-sectional	8270	Caucasian, South Asian	SHIP-Trend 51.5 (13.8) ESTHER 62.1 (6.5) KORA F4 61.0 (8.9) LOLIPOP 52.0 (10.3) RS-III-1 59.9 (8.2) RS-BIOS 67.6 (6.0)	SHIP-Trend 117/130 ESTHER 500/498 KORA F4 844/882 LOLIPOP 2390/1459 RS-III-1 336/395 RS-BIOS 305/414	statin use	glycaemic traits Type 2 diabetes	whole blood	ABCG1 (cg06500161, cg27243685)	Mediation anslysis provided evidence of ABCG1 (cg06500161) methylation partially mediating the association between statin use and higher fasting insulin and HOMA-IR.	72
Other lifestyle factors	Wang X, et al. (2020)	China	Cross-sectional	981	Asian	Case 51.7 (9.3)/ Control 51.8 (10.3)	0/981	health lifestyle score (diet, alcohol use, physical activity, body mass index, and smoking)	breast cancer	whole blood (buffy coat)	RARβ	Healthy lifestyle is associated with the lower risk of breast cancer, and this association is partly mediated by RARβ methylation.	73
Other lifestyle factors	Wu D, et al. (2018)	US	Longitudinal	398	Caucasian	62.4 (12.7)	0/398	alcohol consumption	epithelial ovarian cancer (EOC)	whole blood (leukocyte)	LMO2 (cg09358725) TRPC6 (cg11016563)	Two CpGs are potential mediators of the relationship between alcohol consumption and EOC.	74
Other lifestyle factors	Burghardt KJ, et al. (2016)	US	Cross-sectional	120	Caucasians/African-American/Others	Discovery 45.1 (11.2) Validation 42.3 (12.1)	44/76	atypical Antipsychotics	insulin resistance (HOMA-IR)	whole blood	FAR2	Our results identified a mediating effect of this FAR2 methylation site on AAP- induced insulin resistance.	75
Clinical endpoints	Chen Z, et al. (2020)	US & Canada	Longitudinal	499	NR	NR	NR	HbA1c	complications in Type 1 diabetes	whole blood	186 HbA1c associated CpGs	DNA methylation mediates development of HbA1c-associated complication in Type 1 diabetes.	76
Clinical endpoints	Guay SP, et al. (2020)	Canada	Longitudinal	69 pairs	Caucasian French canadian	Mother 28.0 (3.8)	Mother 0/69 Newborn 33/36	maternal lipid change	newborn’s anthropometric profile	placenta	LPR1 LDLR SCARB1	Mediation analysis supports a causal relationship between maternal cholesterol changes, DNAm levels ath LRP1, and cord blood leptin concentration.	77
Clinical endpoints	Gagne-Ouellet V, et al. (2020)	Canada	Longitudinal	262 pairs	Caucasian	Maternal 28.6 (4.2)	Maternal 0/262 Child 144/118	maternal fasting glucose	childhood leptin concentration	placenta	LEP (cg15758240)	DNAm levels at cg15758240 mediates the association between maternal glycaemia and neonatal leptinemia.	78
Clinical endpoints	Huang JV, et al. (2018)	US	Longitudinal	265	White/Black/Hispanic/Asian/Other	7.7 (0.7)	134/131	mid-childhood adiposity	cardio-matabolic risk	whole blood	ATOH8, CDC37/MIR1181, COMT/TXNRD2, DDX10, TAS2R40, TBCD	6 CpGs are potential mediators of the association between mid-child BMI z-score and adolescence cardio-metabolic risk score.	79
Clinical endpoints	Peng H, et al. (2018)	US	Cross-sectional	119 twins	White	THS 55 MMS 36	168/70	childhood trauma	depression	whole blood (leukocyte)	stress-related genes (4 CpGs in BDNF and NR3C1)	Hypermethylation of 3 CpG sites (1 in the BDNF gene, 2 in the NR3C1 gene) showed significant individual mediation on the relationship between childhood trauma and depressive symptoms	80
Clinical endpoints	Côté S, et al. (2016)	Canada	Longitudinal	E-21: 133 (Case 33/ Control 100) Gen3G: 172 Controls	Caucasian	E-21: Case 29.2 (3.8) Control 28.6 (3.9) Gen3G: 28.0 (4.1)	Maternal E-21:0/133 Gen3G: 0/172 Child E-21: 68/65 Gen3G: 81/91	maternal glocose levels	cord blood leptin levels	placenta	PRDM16, BMP7, CTBP2, and PPARGC1α	DNA methylation variations at the PPARGC1α gene locus explain 0.8% of the cord blood leptin levels variance independently of maternal fasting glucose levels.	81
Clinical endpoints	Sadeh N, et al. (2016)	US	Cross-sectional	200 (144)	White	31.8 (8.4)	182/18	PTSD	prefrontal cortical thickness	whole blood	SKA2 (cg13989295; adjusted for the SNP)	Path analysis showed that SKA2 methylationadj signifi- cantly mediated the relationship between PTSD and neural integrity in prefrontal cortex.	82
Clinical endpoints	West NA, et al. (2013)	US	Longitudinal	21	Caucasian	Maternal DM 32.7 (4.4) non-DM 32.7 (4.1) Offspring DM 9.9 (1.4) non-DM 10.8 (0.6)	Maternal 0/21 Offspring 9/21	maternal diabetes	VCAM-1	whole blood	PYGO CLN8	Increased methylation of PYGO1 and CLN8 had the relative mediation effect on the impact of exosure to maternal diabetes in utero on VCAM-1.	83
Environmental chemical exposures	Yang W, et al. (2021)	China	Cross-sectional	177	Asian	26.9 (4.8)	114/63	Pb (Umbilical cord lead)	non-syndrome cleft lip and/or palate (NSCL/P)	umbilical cord blood	WNT3A	Hypermethylation of WNT3A mediates the association between exposure to Pb and risk for NSCL/P.	84
Environmental chemical exposures	Gao M, et al. (2020)	China	Cross-sectional	1124	Asian	47.6 (8.8)	413/711	fluoride in urine	bone mineral density (BMD)	whole blood	RUNX2 promoter	Urine fluoride concentration reduced bone mineral density of women in excessive fluoride group through higher RUNX2 DNA methylation.	85
Environmental chemical exposures	Zhao Y, et al. (2019)	China	Cross-sectional	249 pairs	Asian	Mother 27.9 (4.5)	Mother 0/249 Newborn 123/126	polybrominated diphenyl ethers	foetal growth retardation (FGR)	umbilical cord blood	LINE-1 (3 CpG sites) HSD11B2 (1 CpG site) IGF2 (4 CpG sites)	IGF2 methylation mediated 38.47% of the effects of BDE-17–190 exposure on risk of FGR.	86
Environmental chemical exposures	Lei X, et al. (2019)	China	Longitudinal	36	East Asian	24 (2)	14/22	PM2.5	TNF-alpha, sICAM-1, sCD40L, fibrinogen	whole blood	TNF-alpha, ICAM1, CD40L, F3	Reduced methylation of TNF-α significantly mediated from 19.89% (Cu) (P = 0.03) to 41.75% (Ca) (P = 0.05) of the association between personal exposure to PM2.5 components and the elevation in TNF-α protein.	87
Environmental chemical exposures	Fu Y, et al. (2019)	China	Cross-sectional	385	East Asian	40	326/27	polycyclic aromatic hydrocarbons	cell cycle, oxidative DNA damage	whole blood	OGG1	The association between urinary 1-OHP and G2/M phase arrest was not mediated by OGG1 methylation (P = 0.129). However, about 20% effect of cell cycle and oxidative DNA damage related to PAHs exposure was mediated by OGG1 methylation.	88
Environmental chemical exposures	Georgiadis P, et al. (2019)	Italy & Sweden	Longitudinal	611	European	52.2 (7.8)	215/396	polychlorinated biphenyl	chronic lymphocytic leukaemia	whole blood (leukocyte)	PCDH17 (cg03865667) BARHL2 (cg25026529) MIR196B (cg15912800) GATA4 (cg03007522) ZFPM1 (cg00352652)	Meidation analysis reinforced the suggestion of a causal link between exposure, chnages in DNA methylation of three genes (PCDH17, BARHL2, MIR196B), and disease.	89
Environmental chemical exposures	Bowman A, et al. (2019)	Mexico	Longitudinal	223	Hispanic	Maternal NR Child 10.3 (1.6)/10.2 (1.7)	109/114	prenatal phthalate exposure (MBP, MIBP, MEHP, and MBzP)	adiposity	whole blood	H19 (4 CpG sites) HSD11B2 (2 CpG sites)	No significant mediation of phthalte exposure on adiposity by DNA methylation of H19 or HSD11B2 was observed.	90
Environmental chemical exposures	Huang LL, et al. (2018)	China	Longitudinal	106 pairs	East Asian	Maternal 29.9 (3.7)	Maternal 0/106 Infant 53/53	maternal phthalate	birth outcomes (birth weight, birth length)	umbilical cord blood	Alu, LINE1	Mediation analysis did not support the mediation effects of Alu on the association between maternal phthalate exposure and birth outcomes.	91
Environmental chemical exposures	Gliga AR, et al. (2018)	Bangladesh	Longitudinal	551 (113)	South Asian	8.9 (0.1)	50/63	prenatal arsenic exposure	IGFBP3	whole blood (peripheral blood mononuclear cells)	12 CpGs in IGFBP3	Mediation analysis suggested that methylation of 12 CpG sites for all children was mediator of effect for the association between prenatal As and IGFBP3.	92
Environmental chemical exposures	Yang P, et al. (2018)	China	Longitudinal	106 pairs	East Asian	Maternal 29.9 (3.7)	Maternal 0/106 Infant 53/53	polycyclic aromatic hydrocarbon metabolites during late pregnancy (urinary 2-Hydroxynaphthalene)	birth outcomes	umbilical cord blood	Alu, LINE1	Mediation analysis failed to show a mediator effect of global DNA methylation in the association between prenatal urinary OH-PAHs and birth outcomes.	93
Environmental chemical exposures	Wang C, et al. (2018)	China	Longitudinal	36	East Asian	24	14/22	PM2.5	inflammatory proteins	whole blood	TNF-alpha, ICAM1, CD40L, IL6	Reduced methylation at locus 2 mediated 14.0% of the asso- ciation between PM2.5 exposure and the elevation in TNF-α.	94
Environmental chemical exposures	Kobayashi S, et al. (2018)	Japan	Longitudinal	177 pairs	East Asian	Maternal 29.8 (4.8)	Maternal 0/177 Infant 79/98	perfluorooctanoic acid & perfluorooctane sulphonate	ponderal index	umbilical cord blood	IGF2, LINE1, H19	Mediation analysis suggested that reduced IGF2 methylation explained approximately 21% of the observed association between PFOA exposure and reduced ponderal index of the infant at birth.	95
Environmental chemical exposures	Lim YH, et al. (2017)	Korea	Longitudinal	102	East Asian	72.8 (4.0)	8/94	ambient temperature	blood pressure	whole blood	ZNF (cg21194911, cg21761427)	Mediation analysis showed that a marginally significant proportion of the association between temperature and DBP can be attributed to changes in CG21761427 located in the promoter region.	96
Environmental chemical exposures	Lee J, et al. (2017)	China	Longitudinal	305 (217)	East Asian	Maternal 2002 Cohort: 25.3 (3.2) 2005 Cohort: 27.8 (4.6)	Infant 2002 Cohort: 75/742,005 Cohort: 86/71	prenatal airbone polycyclic aromatic hydrocarbon metabolites	child development (IQ score)	umbilical cord blood	LINE1	LINE1 methylation status did not mediate the relationship between PAH-DNA adducts and socres on GDS or IQ test.	97
Environmental chemical exposures	Peng C, et al. (2016)	US	Longitudinal	551	White/other	73.3 (6.9)	551/0	PM2.5	fasting blood glucose, diabetes	whole blood	ICAM1, IFNγ, IL6, TLR2	We found substantial mediation effects of PM2.5 on FBG through a decrease in ICAM1 methylation for the 28-day exposure time window.	98
Environmental chemical exposures	Wang C, et al. (2016)	China	Longitudinal	36	East Asian	24	14/22	PM2.5	blood pressure, ACE protein	whole blood	ACE	ACE hypome- thylation mediated 11.78% (P = 0.03) of the elevated ACE protein by PM2.5. Increased ACE protein accounted for 3.90 ~ 13.44% (P = 0.35 ~ 0.68) of the elevated BP by PM2.5.	99
Environmental chemical exposures	Chen R, et al. (2016)	China	Interventional	35	East Asian	23	10/25	PM2.5	cardiovascular biomarkers	whole blood	MCP1, TNF, TLR2, ICAM1, IL6, CD40LG, F3, SERPINE1, ACE, and EDN1	There was a significant mediated effect (17.82%, P = 0.03) of PM2.5 on sCD40L protein through CD40LG hypomethylation.	100
Environmental chemical exposures	Nylander-French LA, et al. (2014)	US	Cross-sectional	20	Caucasians/African-American/Asian/mixed ethnicity	35.3 (8.9)	20/0	1,6-hexamethylene diisocyanate	urine biomarker 1,6-hexamethylene diamine	whole blood (peripheral blood mononuclear cells)	LPHN3 (cg21814870), SCARA5 (cg24554944)	Mediation analysis confirm that no single or a set of CpGs as complete mediator.	101
Environmental chemical exposures	Bind MA, et al. (2014)	US	Longitudinal	777	White	median 72	777/0	air pollution (black carbon, sulphate, ozone)	fibrinogen ICAM-1 protein	whole blood (leukocyte)	F3, IFN-γ, IL-6, TLR-2, ICAM-1	Some significant mediated effects of black carbon on fibrinogen through a decrease in F3 methylation and sulphate and ozone on ICAM-1 protein through a decrease in ICAM-1 methylation.	102
Environmental chemical exposures	Tarantini L, et al. (2013)	Italy	Longitudinal	63	NR	44.0	63/0	PM10, PM1, aluminium, manganese, nickel, zinc, arsenic, lead, iron, and chromium	endogenous thrombin potential	whole blood	NOS3 and EDN1	Mediation analysis formally confirmed NOS3 and EDN1 hypomethylation as intermediate mechanisms for PM-related coagulation effects.	103
Environmental chemical exposures	Bozack AK, et al. (2018)	Bangladesh	Longitudinal	613	South Asian	NR	Maternal 0/613 Infant 320/292	maternal arsenic in drink water and toenails	gestational age	umbilical cord blood	miR124-3, MCC	The association between prenatal arsenic exposure and gestational age was nominally mediated by DNAm levels at miR124-3.	104
Environmental chemical exposures	Bozack AK, et al. (2020)	Bangladesh	Longitudinal	413	South Asian	NR	Maternal 0/413 Infant 218/196	maternal arsenic in drink water and toenails	birth outcomes	umbilical cord blood	DNMT3A	Cord blood DNAm levels at DNMT3A mediated the association between prenatal arsenic exposure and birth outcomes (gestational age and birth weight)	105
Socioeconomic status	Chu SH, et al. (2018)	US	Longitudinal	143	White/Non white	46.9 (1.7)	69/74	early life social disadvantage	adulthood BMI	whole blood (buffy coat)/adiose tissue	Male 22 loci Female 27 loci	In male, significant mediation site was in the genes of ABCC1, USP24, and other 25 loci. In female, significant mediation sites was in the genes of PKHG1, BCAR3, and other 20 loci.	106
Socioeconomic status	Loucks EB, et al. (2016)	US	Longitudinal	141	Caucasian/African American/Native American/Others	47	67/74	childhood socioeconomic status	adiposity (BMI)	adipose tissue	91 CpG sites in men 71 CpG sites in women	Methylation of FASN, STAT3 and TMEM88 in females, and NRN1 in males in adipose tissue mediated the association between childhood SES and adulthood BMI.	107
Genetic factor & race	Jordhl KM, et al. (2020)	US	Cross-sectional	836	Asian/Pacific Islander, Black/African American, Hispanic, Non-Hispanic White, Other	47.5 (*n = 742)	0/836	SNP (rs798766, rs401681, rs2294008, and rs8102137)	bladder cancer	whole blood (buffy coat)	cg00006948, cg27028750, cg26209169, SLURP1(cg06565975), LYNX1(cg03405983), LY6K(cg24023258), LYNX1(cg17888033), LY6D(cg17252645), CCNE1(cg16836589), CCNE1(cg27475126)	DNA methylation in these mQTL-associated CpG sites did not significantly mediate the association between SNP and bladder cancer.	108
Genetic factor & race	Coassin S, et al. (2020)	Germany & Austria	Cross-sectional	2208	Caucasian	KORA4 61.0(8.9) KORA3 53.2 (9.6)	KORA4 843/881 KORA3 252/232	SNP (rs76735376)	lipoprotein A	whole blood	LPA (cg1702867)	A formal mediation analysis revealed no significant causal mediation effect of the methylation level of cg17028067 on log-Lp(a) (p = 0.13).	109
Genetic factor & race	Dai JY, et al. (2020)	US	Cross-sectional	826	European American	35–74 years	826/0	147 prostate cancer risk SNPs	gene expression	tumour tissue	MLPH, HLA-DQB1, HLA-DRBS, IRX4, TUBA1C	Mediation analyses illuminate the likely intermediary role of CpG methylation of the five genes in eQTL regulation of gene expression	110
Genetic factor & race	Li D, et al. (2019)	China	Longitudinal	299	East Asian	64.6 (14.6)	109/190	SNP (rs3733890)	folate therapy efficacy	whole blood	BHMT BHMT1	The methylation levels of BHMT mediated the effect of rs3733890 on folate therapy efficacy (p = 0.03), and BHMT_1 methylation levels mediate the association between rs3733890 and folate therapy efficacy (p = 0.044).	111
Genetic factor & race	He Z, wt al. (2018)	China	Cross-sectional	304	East Asian	48.4 (9.7)	NR	SNP (rs174570)	FADS1, FADS2 mRNA expression	whole blood	FADS1, FADS2	DNA methylation levels of four CpG sites mediated the effect of rs174570 on FADS1 gene expression, while only DNA methylation of one CpG site mediated the effect of rs174570 on FADS2 gene expression	112
Genetic factor & race	Guerra S, et al. (2018)	Spain/Sweden	Cross-sectional	203 (172)	Caucasian	4.4 (0.2)	102/101	CHI3L1 SNPs (rs4950928 & rs7542294)	YKL-40	whole blood	CHI3L1 (cg07423149)	The effects of genetic variants of CHI3L on YKL-40 partly mediates the CHI3L1 DNA methylation	113
Genetic factor & race	Ma Y, et al. (2016)	US	Cross-sectional	1748	Caucasian	CHS 72 (5) GOLDN 48 (16) MESA 62 (10)	NR	SNP rs2246293/rs405509 Circulating fatty acid	HDL-cholesterol Triglyceride	whole blood (monocytes and lymphocyte)	ABCA1 (cg14019050) APOE (cg04406254)	ABCA1 cg14019050 (percentage) less likely to act as a mediator for the association between EPA and HDL cholesterol or the association between the rs2246293 C allele and HDL cholesterol or that the cg14019050 methylation mediated rs2246293 genotype-related differences in the EPA–HDL cholesterol association.	114
Genetic factor & race	Straughen JK, et al. (2015)	US	Longitudinal	87	Black/Non-black	Maternal Black 29.2 (6.6) Non-black 25.6 (4.6)	Maternal 0/87 Infant 48/39	race	birth weight	umbilical cord blood	IGF-1	IGF-1 methylation partially mediates the relationship between black race and birth weight, but not statistically significant.	115
Genetic factor & race	Liu Y, et al. (2013)	Sweden	Cross-sectional	691	Caucasian	51.9 (11.8)	197/494	524 SNPs within MHC region and 1 SNP outside of MHC region	rheumatoid arthritis	whole blood	DNAm within/outside MHC region (60 CpGs and 1 CpG)	Ten putative DMPs that mediates genetic risk for RA, nine in the MHC cluster, and one outside on the same chromosome.	116
Genetic factor & race	Huen K, et al. (2015)	US	Longitudinal	449 (176)	Latino	9.3 (0.3)	Unkknown	PON1 (rs705379)	PON1 expression	umbilical cord blood	PON1 DNA mehtylation CpG blocks	Statistically significant indirect effects of methylation providing evidence that DNA methylation mediates the relationship between PON1-108 genotype and PON1 expresion.	117

Results

Search results

Figure 3 shows the PRISMA study flow chart in this review. From initial searches through online databases, 215 studies were identified. In addition to these studies, four studies were included through other resources (e.g recommendations from co-authors). After removal of 101 duplicated studies, 118 studies were screened at title- and abstract- levels and 76 studies were eligible for full-text screening. Following the full-text screening, seven studies were excluded from this review because these papers did not meet the eligibility criteria for this review. Ultimately, we included 69 papers in this review and summarized information of these studies (Table 1).

Figure 3.

Flow diagram for selection of the reviewed articles.

Basic information of reviewed papers

Graphical summary of the reviewed articles is shown in Figure 4. 87.0% of the reviewed articles were published in the past five years (2016–2021). Some studies have not been published, but available as an online material when we performed database search. Regarding the countries where the research was conducted by region, 24 studies were conducted in North America (34.8%), 25 in Asia (36.2%), 11 in Europe (15.9%), and 9 studies (13.0%) were international collaborative research (conducted in more than two countries). In relation to the research location, information of participant’s ethnicity was extracted as accurate as described in the methods of reviewed papers. There were 26 Caucasian/European studies (37.7%), 25 Asian studies (36.2%), 2 Hispanic/Latino studies (2.9%), 13 multi-ethnic studies (18.8%), and 3 for Not Reported (4.3%). The sample size varied widely between studies, ranging from 20 to 8,270. Given the nature of intergenerational inheritance of DNAm, 19 reviewed papers were focused on pairs of mother/father and child (27.5%). Among other 50 studies, 44 studies were conducted only in adults and seniors (63.8%), 6 studies not reported age distribution of participants (8.7%). There were different sampling tissues for DNA extraction; 59 studies from whole blood including 12 studies using umbilical cord blood, 5 studies from placenta, 1 study from dried spot blood, 1 study from tumour tissue, 1 study from adipose tissue, and 2 studies from multiple tissues (e.g., whole blood and lung tissue).

Figure 4.

Graphical information for the reviewed articles. a: Published year; b: Countries; c: Ethnicities; d: Exposures; e: Sample sizes; f Tissues.

Types of exposures

We classified all 69 articles into the following seven categories based on exposure variables: smoking, dietary intake and famine, other lifestyles, clinical status, environmental exposure, socioeconomic status, and genetic factors and race. The number (percentage) of studies assigned to each category was 13 for smoking (18.8%), 8 for dietary intake and famine (11.6%), 6 for other lifestyle factors (8.7%), 8 for clinical endpoints (11.6%), 22 for environmental chemical exposures (31.9%), 2 for socioeconomic status (SES) (2.9%), and 10 for genetic factors and race (14.5%), respectively (Figure 4).

Smoking

Of 13 reviewed papers regarding smoking exposure, six papers addressed the biological mechanism underlying the association between smoking during pregnancy and health outcomes [49–54]. All of these studies were conducted in European populations (n = 179–2,821). Although these studies reported different CpG sites as a mediator, two papers reported that cord blood DNAm levels in Growth Factor Independent 1 Transcriptional Repressor (GFI1) mediated the association between maternal smoking during pregnancy and children’s birth weight [51,54]. In the earlier study, Küpers LK, et al. found that methylation at three GFI1 CpGs (cg09935388, cg14179389, and cg12876356) partly mediated the effect of maternal smoking on birthweight (mediation: 12.4–18.9%) [54]. Additionally, Hannon E, et al. reported novel mediation sites in Aryl-Hydrocarbon Receptor Repressor (AHRR) and Exocyst Complex Component 2 (EXOC2) for birthweight [51]. Other seven papers examined the indirect effect of smoking exposure on lung function, bladder cancer, gene expression, oxidative damage, IgG glycosylation, and rheumatoid arthritis in diverse populations [43,55–60]. Notably, five studies convincingly reported that cg05575921 or cg14817490 in AHRR mediated associations between smoking exposure and different outcomes with explained proportion of 13.4–57.4% [43,55,56,58,59]. In addition, several studies reported that common CpG sites in F2R Like Thrombin Or Trypsin Receptor 3 (F2RL3), G Protein-Coupled Receptor 15 (GRP15), and GFI1 could be a potential mediator of association between smoking and different health outcomes.

Dietary intake and famine

Five studies [61–65] considered the effect of maternal famine and malnutrition, as an extreme dietary pattern, on children’s health outcomes based on a well-known concept as the Development Origins of Health and Disease (DOHaD) hypothesis [66]. These studies focused on different health outcomes such as children’s kidney function, lipid metabolism, depression, clinical conditions, and adiposity. Therefore, researchers performed a candidate gene approach to examine whether DNAm levels in the different genes mediated the association between famine exposure and outcomes, but all of these were reasonable genes considering biological backgrounds. Of which, two studies in Asian populations examined the indirect effect of Insulin Like Growth Factor 2 (IGF2), a gene associated with growth factor activity and insulin receptor binding, on the associations between famine and children’s health outcomes. reported much less proportion (<5%) explained through DNAm levels of IGF2 for the association between famine severity and total cholesterol levels [62]. One recent study in China also suggested no significant mediated effect of IGF2 DNAm levels on the association between maternal famine and child waist circumference [65]. Two cross-sectional designed mediation analyses focused on the association between adulthood nutritional intake and lipid metabolism and metabolic syndrome. Our group showed mediated effect of ATP Binding Cassette Subfamily A Member 1 (ABCA1) DNAm on the association between higher vitamin C intake and lower HDL-C levels in Japan [67]. A recent study in European countries reported that a CpG site in Carnitine Palmitoyltransferase 1A (CPT1A) may mediate the association between dietary intake of carbohydrate and fat and metabolic diseases [68]. For duration of breast feeding and early transient overweight, no significant mediated effects of cg23381058 (LEP) was found in the Isle of Wight Birth Cohort [69].

Other lifestyle factors

Studies using mediation analysis on other lifestyles, such as exercise, alcohol consumption, and periodic drug use, were very sparse [70–75]. Liu X, et al. reported that DNAm levels of suppressor of cytokine signalling 3 (SOCS3) partially mediated association between sedentary time and type 2 diabetes (T2DM) in a rural Chinese [71]. For alcohol consumption, a longitudinal study with multiple mediators (LMO2 and TRPC6) represents a potential DNAm-mediated relationship between alcohol consumption and EOC risk [74]. With a combination of healthy lifestyles, a cross-sectional study in China showed significant mediation effect (14.3%) of RARβ methylation on a healthy lifestyle score-breast cancer association [73]. Interestingly, two studies were conducted in the U.S. to examine the mediated effect of DNAm on relationships between drug treatment and drug addiction and health outcomes. Shu C, et al. examined how DNAm could explain the association between cocaine use and human immunodeficiency virus (HIV) severity [70]. This study showed significant mediation effect of DNAm levels of four genes (MX1, PARP9, NLRC5, and CX3CR1) between persistent cocaine use and HIV severity. Another study addressed the mechanisms by which atypical antipsychotics (AAP) cause insulin resistance. The results suggested that DNAm levels in Fatty Acyl CoA Reductase 2 (FAR2) mediated the APP-induced insulin resistance with higher indirect effect (40%) [75]. Moreover, in diverse racial populations, a cross-sectional study showed that ATP Binding Cassette Subfamily G Member 1 (ABCG1) DNAm levels mediated the association between statin use and glycaemic traits [72].

Clinical endpoints

Six studies covered metabolic indices as exposure factors, including fasting glucose levels, adiposity, and lipid profiles [76–83]. Of which, many studies have focused on maternal clinical conditions as exposure in the nature of trans-generational feature of DNAm. A Canadian research group reported that several GpG sites in different genes mediated the association between maternal fasting glucose levels and cord blood leptin levels, but the explained proportion of variance was less than 1% [78,81]. Another study in the U.S. showed that PYGO1 and CLN8 mediated a high proportion of maternal diabetes and VCAM-1 levels in offspring [83]. Also, a longitudinal in Canadians indicated that the effect of maternal lipid change on children’s anthropometric traits was mediated by DNAm levels of LPR1, LDLR, and SCARB1 genes [77]. One study examined the indirect effects of DNAm on the association between mid-childhood BMI (7.7 yrs) and adolescence cardiometabolic risk (12.9 yrs) and found that six CpG sites accounted for about 10% of BMI-cardiometabolic trait associations [79]. In addition to metabolic conditions, two studies in Caucasians were performed to examine the mediation of relationships between early life psychological adversity and adulthood health outcomes. A twin study in the US demonstrated that DNAm levels in the BDNF and NR3C1 mediated approximately 20% of the association between childhood trauma and depression [80]. Additionally, an epidemiological study among trauma-exposed US veterans revealed that SKA2 DNAm mediated the association between PTSD severity and cortical thickness [82].

Environmental chemical exposures

A total of 22 manuscripts were included in this category with different types of chemical exposures such as air pollutants, organic pollutants, heavy metals which accounted for one-third of the reviewed papers [84–105]. Airborne monitoring is commonly used to determine exposure levels (eight papers, 36.4%), although blood (serum, plasma, and umbilical cord blood), urine, and toenail/drinking water are used in 6 (27.3%), 6 (27.3%), and 2 (9.0%) papers, respectively. Interestingly, 16 of 22 studies were conducted in Asian countries, mainly in China. Of all studies regarding chemical exposures, eight papers focused on air conditions, including particle matter 2.5 (PM_2.5), black carbon, or ambient temperature [87,94,96,98–100,102,103]. Four papers measured personal exposure levels of chemical substances using special equipment [87,94,99,100], while the rest of them identified exposure levels using a regional or a workplace-level monitoring system [96,98,102,103]. A study among 551 nondiabetic participants showed significant mediated effects of ICAM1 DNAm on the association between 28-day PM_2.5 and fasting blood glucose levels [98]. Other studies conducted in China and the US reported that DNAm levels of several genes mediated the association between air pollutants and coagulation factor [103], inflammatory proteins [87,94,100,102], cardiovascular-related proteins [99]. Ten studies investigated how DNAm mediates the association between exposure to organic pollutants and health outcomes. Although a Chinese study reported that the effect of polycyclic aromatic hydrocarbons (PAHs) on oxidative DNA damage and was mediated by DNAm of OGG1 gene (~20%) [88], two studies failed to show mediated effects of global DNAm on the associations between prenatal PAH exposure and birth outcomes [96] and IQ score in children [97]. Studies in Mexico and China reported no significant mediated effects of prenatal phthalate exposure-health outcomes [90,91]. Other five studies examined that DNAm levels of different genes mediated the associations between fluoride [85], diphenyl esters [86], polychlorinated biphenyl [89], 1,6-hexamethylene diisocyanate [101], and perfluorooctanoic acid [95] and health outcomes. In a longitudinal study in Bangladesh, DNAm levels at the different 12 CpG sites mediated the associations between prenatal arsenic exposure and IGFBP3 [92]. In another study in a Bangladesh population, the association between prenatal arsenic exposure and gestational age was mediated by DNAm levels at miR124-3 [104]. Furthermore, the same research group found that cord blood DNAm levels at DNMT3A mediated the association between prenatal arsenic exposure and birth outcomes (gestational age and birth weight) [105]. Another study for lead exposure suggested that WNT3A mediated 9.3% of total effect of Pb on non-syndromic cleft lip and/or palate risk [84].

Socioeconomic status

Two US studies within the same project explored mediated effects of DNAm levels on the association between childhood SES and adiposity [106,107]. The earliest study suggested that DNAm levels in FASN, STAT3, TMEM88 in males and NRN1 in females mediated the association between childhood SES and adiposity [107]. In their latest work, they also provided supportive evidence for the mediating role of DNAm in the effect of childhood SES on adulthood BMI [106]. This study suggested that the association between childhood SES and adulthood BMI was mediated by DNAm levels of seven genes in females and two genes in males.

Genetic factors and race

Nine studies [108–117] addressed the mediation effects of DNAm on the genetic variant-outcome associations. From a different viewpoint, this is a concept of mendelian randomization, one of the causal inference methods using genetic variants as instrumental variables. One of the earliest studies reported that a total of 11 CpG sites within/outside of the MHC region mediated the effect of single nucleotide polymorphism (SNP) on rheumatoid arthritis [116]. Following to this study, several studies in diverse ethnic groups performed mediation analysis with candidate gene (SNP) approach, including Paraoxonase 1 (PON1) [117], ABCA1 [114], Apolipoprotein E (APOE) [112], Chitinase 3 Like 1 (CHI3L1) [113], Fatty Acid Desaturase 1 (FADS1) [112], Betaine–Homocysteine S-Methyltransferase (BHMT) [111], and Lipoprotein(A) (LPA) [109], respectively. In a recent study, a patient-based study among 826 males examined to elucidate the role of DNAm on gene expression quantitative locus (eQTL) regulation using 149 prostate cancer-associated SNPs [110]. They found that intermediary role of DNAm of the five genes (MLPH, HLA-DQB1, HLA-DRBS, IRX4, and TUBA1C) in eQTL regulation of gene expression, but the proportions explained by DNAm has some gaps between different dataset.

Discussion

In this review, we first provided a methodological overview of mediation analysis. In the second part, we summarized practical examples of mediation analysis using DNAm as an intermediate factor for the associations between exposure and health outcomes. This review will be a useful primer for researchers to explore the biological pathway from exposure to outcomes using DNAm levels rather than examination of relationship either between exposure and DNAm or DNAm and health outcomes.

While there are differences in sample size, tissues, cell types, and statistical methods between studies, this review highlights the promising mediated effects of DNAm levels of specific genes on the exposure-outcome associations in a specific exposure, smoking. Several previous studies have consistently suggested that smoking during pregnancy is associated with cord blood DNAm levels. Based on reported evidence, many researchers examined the mediated effects of DNAm levels of specific genes on the associations between prenatal maternal smoking and health outcome. They found consistent results that DNAm levels of the AHRR and GFI1 in cord blood may be a strong mediator of the association between prenatal smoking and children’s outcomes. Undoubtedly, these are benefits through the application of mediation analysis and take one step forward from conventional epidemiological studies which assessed only exposure-mediator or mediator-outcome associations. Given that DNAm levels is a mediator of exposure-outcome association, DNAm can be used not only as a biomarker for exposures or outcomes, but also to contribute to causal biological mechanisms underlying the association between exposures and outcomes. However, concluding mediation effects of DNAm needs to be carefully considered because consistent mediated effects were observed only for smoking as an exposure. In other words, reproducibility and repeatability of these mediated effects remains an emerging issue to establish conclusive evidence for biological pathways and mechanisms. Additionally, among DNAm studies with exposure misclassification, it is known that DNAm could serve as better biomarkers of the exposure than self-report, thereby increasing the type I error rate, underestimating the direct effect, and overestimating mediated effects [118]. It is likely that effects have been overestimated and publication bias only reporting positive results exacerbating the winner’s curse.

We need to point out that different methodologies in mediation analysis were used among the reviewed articles. In some reviewed papers, because there was no description of the mediation analysis, we did not identify how they estimated and decomposed the effects. At least clearly stating what model was used, what confounders were included in, what software was used to estimate the results, and what the results of the sensitivity analysis were will certainly be important information to interpret the estimated results. As a different topic in methodology, the approach for multiple mediators will be a topic of debate for further applications in the epigenetic data. EWAS are becoming more common in large-scale epidemiological studies and have found multiple differentially methylated regions for several diseases and lifestyle factors. Additionally, other omics technologies such as genomics, transcriptomics, metabolomics, and proteomics are associated with DNAm levels. Therefore, a novel methodology for this biological complexity underlying diseases will be required to treat more high-dimensional mediators, where multiple CpGs are involved, using data from EWAS and other databases. In this context, Zhang H proposed a new method to estimate mediated effects in high-dimensional DNAm dataset [119]. However, these discussions on mediation analysis for multiple mediators including high-dimensional dataset have just begun in the past five years [120,121], and future methodological developments are expected.

In this review, we covered various exposure factors rather than focusing on a specific exposure. This concept allowed us to clarify when, in which country, and with what exposures mediation analysis using DNAm levels as a mediator was used. This scoping review may be a useful guide to boost epidemiological studies with DNAm using mediation analysis. This review also has several limitations to be noted, especially for substantial heterogeneity in methodologies (genomic approach, tissue-specific cells, and study designs) between the reviewed articles. First, this review has a mixture of the studies applying candidate-gene and epigenome-wide approaches. There are different things to pay attention in each approach from a methodological viewpoint. For example, in candidate-gene approaches, researchers are encouraged to validate the results in an independent dataset from the original population to deny that novel findings are not by chance. For example, among the reviewed article, Lai CQ, et al. examined whether CPT1A gene methylation mediated the association between carbohydrate and fat intake and metabolic diseases in two independent populations, the Genetics of Lipid Lowering Drugs and Diet Network and the Framingham Heart Study [68]. In addition to reproducibility, Epigenome-wide approaches also need to address the issue of multiple comparisons in statistical inferences. All epigenome-wide study in this review addressed using two common adjustments: Bonferroni correction or Benjamini–Hochberg correction [122]. It is necessary to mind these common and specific characteristics of each approach when interpreting the results. Second, DNAm levels were measured using different cell types for each study. Although we did not in detail summarize cell types used in each study, epigenetic status is highly variable across different cell types [123]. Many correction algorithms have been developed to adjust for heterogeneity of cell distribution [124–126]. Of the reviewed articles, some studies using whole peripheral blood adjusted for estimated cell mixture distribution based on DNAm array [127]. Another method, ReFACTor [128], was applied by Cardenas et al. [49]. However, a recent review compared performance of many correction algorithms and reported their different characteristics according to several simulations [129]. Biologically, one would also expect that different tissues would yield different candidate mediators. Therefore, even though the studies focused on the same exposure factors, the mediated effects of DNAm on exposure-outcome associations will be carefully interpreted. Third, in this review, five studies were conducted on the mixed exposure of multiple substances (i.e., compounds) in environmental chemical exposures. However, further studies are needed to broaden more exposures such as chemical compounds and dietary intake of multiple nutrients. This approach will be helpful to evaluate of mediated effects by DNAm in more realistic settings. Fourth, we excluded papers from this review that did not meet our predetermined criteria. This exclusion criteria helps to select the list of quality-serving literature, but it is also possible that some necessary papers in different databases or in languages other than English were excluded from our literature screening.

In conclusion, we provide an exposure-wide summary for the mediation analysis using DNAm levels as a mediator. In the reviewed articles, there are heterogenous methodologies in mediation analysis and typical issues in epigenetic studies such as different cell compositions and tissue-specificity. Further accumulation of evidence with diverse exposures, populations and with rigorous methodology will be expected to conclude these mediated effects of DNAm for future practical use.

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.

Funding Statement

This study was supported in part by the Japan Society for the Promotion of Science under Grant17K09139, 20K10515, and 20K18943.

Disclosure statement

Dr. Andres Cardenas was supported by the National Institutes of Health grant R01ES031259.

Author contributions

RF and SS contributed to study conception and design, and drafting and revising the manuscript. RF and YT contributed to literature searching, data extraction, and revising the manuscript. AC and KS contributed to critical revision of the manuscript.