Immediate and durable effects of maternal tobacco consumption on placental DNA methylation: a replication and discovery study

Abstract

An increasing number of epigenome-wide association studies report tobacco smoking-associated DNA methylation levels. However, comprehensive replication studies remain scarce, particularly in placenta, despite their crucial interest in such a large-scale context. Using DNA methylation data from the EPIC array of 341 new placentas (85 smokers, 219 non-smokers, and 37 former smokers) from the EDEN cohort, we used a candidate approach to replicate maternal smoking-associated CpGs and regions previously identified using the 450K array, and an exploratory approach to discover new associations within EPIC-specific CpGs. Smoking-associated changes in DNA methylation in CpGs and regions were classified as either transient or persistent (indicating epigenetic memory), depending on the stability of their association with smoking status. Among candidate loci, 38% of probes and 9% of regions were replicated, providing robust evidence of effects of prenatal smoke exposure on methylation patterns of these loci. LEKR1 was the top hit in both the initial and replication studies. Most of the replicated loci were transient CpGs (i.e. current smokers), while persistent CpGs (i.e. former smokers) remained scarce and somewhat inconsistent with previous findings. The additional exploratory analysis identified 733 novel probes and 75 novel regions, including 18% and 30% of transient loci, respectively. Results suggested that most of the effects were related to in utero exposure only, supporting pregnant women’s efforts to quit smoking. This replication study also evidences the importance of reproducible work in omic investigations to provide a more in-depth and robust understanding of the effects of environmental exposures on health biomarkers..

Introduction

Pregnancy is a critical period for early development of the offspring. Many epidemiological studies report an increased risk for adverse pregnancy and birth outcomes [1] as well as child’s developmental difficulties later in life, associated with maternal tobacco smoking (MTS) during pregnancy [2]. Various measures to reduce the harmful effects of MTS on the health of the mother and foetus are debated, including stopping or reducing perinatal smoking and encouraging the use of nicotine substitutes [3]. Although the MTS prevalence has decreased over the past 5 years, still 12% of French pregnant women smoke tobacco during the third trimester of pregnancy [4]. Besides, the prevalence of tobacco smoking among French women from 18 to 75 years old remains high (28% in 2020, according to the last report from Santé Publique France [5]).

The underlying biological mechanisms of MTS on child development and health remain poorly understood [6, 7]. Epigenetic changes, in particular in DNA methylation (DNAm), are among the most explored pathways to explain such effects [8, 9]. Numerous epigenome-wide association studies (EWASes) have investigated MTS’s associated DNAm modifications in the perinatal period, mostly in cord blood [10, 11]. The placenta is an organ specific to the pregnancy, playing a leading role in the offspring’s development [12], representing a marker for environmental exposures during pregnancy [13, 14] and which is particularly relevant when investigating the effects of prenatal exposure to cigarette smoking [15]. Placenta, along with the brain, is also the tissue with the highest levels of imprinted transcripts [16]. Parental imprinting is an epigenetic process mainly controlled by DNAm and/or histone methylation, which can have strong and differential effects on the regulation of gene expression, depending on which parent the gene allele originates from. Apart from being biologically relevant for DNAm analysis, placenta is a tissue on which Illumina Infinium MethylationEPIC (EPIC) and Illumina Infinium HumanMethylation450 (450K) arrays have shown higher per-sample and median correlations at the CpG level than for cord blood or children whole blood samples [17].

Several recent (non-)systematic reviews have summarized the results on the effects of tobacco smoking exposure during pregnancy on DNAm changes in the perinatal period and have underlined the critical need for meta-analyses and replication studies to improve the robustness of the results [6, 8, 9]. Indeed, detection of statistical associations between high-dimensional epigenetic targets and environmental factors requires a large number of participants, while EWASes are conducted on limited sample sizes, leading to an increased risk of false positive findings. Currently, two meta-analyses from the Pregnancy and Childhood Epigenetics (PACE) consortium investigated the effects of prenatal exposure to MTS on placental and cord blood 450K-DNAm levels [18, 19]. Technical factors, such as batch effects and the different types of DNAm arrays, further challenge the reproducibility of the findings [20, 21], and thus scientific consensus and consistency [22]. To this end, replication studies represent a unique opportunity to test associations on different types of DNAm arrays using similar samples.

Overall, our cohort-based analyses enabled the first epigenome-wide replication study of DNAm MTS-associated levels in the placenta, the discovery of new MTS-associated CpGs on the EPIC array, and the investigation of the effects of quitting smoking on placental DNAm levels. Using new data from 341 participants in the French Étude des Déterminants pré et post natals précoces du développement et de la santé de l'Enfant (EDEN) mother–child cohort [23] (EPIC array), we aimed to assess whether previous findings involving 568 participants from EDEN (450K array) [14] could be reproduced, particularly regarding transient or persistent DNAm changes at specific loci after smoking cessation. We further conducted an exploratory analysis to discover new associations on 378 680 CpGs that were not included in the initial study. Among smoking-associated probes and regions, we looked for imprinted genes or germline differentially methylated regions (gDMRs) that could suggest further influence of MTS on gene expression regulation. Lastly, we used paternal tobacco smoking during pregnancy as a negative control to provide insights for an intrauterine effect of MTS in pregnancy on placental DNAm levels.

Materials and methods

Study population

This replication study is based on a subset of 382 participants enrolled in the EDEN mother–child cohort [23]. The EDEN cohort recruited a total of 2002 pregnant women between 2003 and 2006 within two different study centres (Nancy and Poitiers University Hospitals) before their 24th week of gestation. Exclusion criteria were pre-pregnancy diabetes, multiple pregnancies, French illiteracy, and willingness to live outside the region within the following 3 years at inclusion. The participants’ sociodemographic and health characteristics were collected by questionnaires and interviews at different time points during pregnancy and after childbirth [23]. Among the participants included in this replication study, seven women (four women who did not smoke tobacco during pregnancy and three women who smoked tobacco throughout pregnancy) were already involved in the initial analysis [14], which included 568 participants (Fig. S1). Methylation measurements for the seven participants common to both analyses were highly correlated (ranges from 0.977 to 0.985, P-value < .001).

The data supporting the findings of this study cannot be made available for ethical and legal restrictions purposes.

Maternal tobacco smoking status assessment

MTS during pregnancy was assessed by questionnaires completed by the mothers at different stages of the pregnancy. These questionnaires were self-reported or administrated by midwives during clinical examinations occurring before and after delivery. Women were asked, ‘Did you smoke during the 3 months preceding your pregnancy?’ and whether they smoked at each trimester of pregnancy as well as the average number of cigarettes they smoked on a daily basis (Fig. S2). Non-smokers were defined as women who neither smoked within the 3 months preceding pregnancy nor during any trimester of pregnancy. Former smokers were mothers who reported smoking more than one cigarette a day in the 3 months preceding their pregnancy but declared smoking no cigarette throughout the pregnancy duration. Finally, current smokers were defined as mothers who smoked at least one cigarette a day in the 3 months before their pregnancy as well as during the three trimesters of their pregnancy (Fig. S2).

Out of 382 participants, 59 (15.4%) did not fully report their tobacco smoking status or have quitted smoking during pregnancy. Among those 59 participants, 18 participants who declared smoking more than one cigarette per day in the 3 months before their pregnancy and during at least two pregnancy trimesters and reported smoking no cigarette during the remaining pregnancy trimester (either first, second, or third trimester) were categorized as current smokers. All of the remaining 41 participants were excluded from the analyses. Hence, the final data set contained a total of 341 participants with available data on tobacco smoking 3 months before and during pregnancy as well as DNAm levels (Fig. S1).

DNA extraction and DNA methylation measurement

Placental tissue from the foetal side was sampled by the midwives or the laboratory technicians of the study using a standardized procedure. Samples around 5 mm³ were collected a few centimetres from the cord insertion and immediately frozen at −80°C [14]. The depth of the sampling targeted at foetal villi.

DNA extraction

DNA was extracted with the DNeasy 96 Blood & Tissue Kit (ID:69581, QIAGEN) using the DNeasy® Blood & Tissue Handbook according to the manufacturer’s instructions (QIAGEN 2020). DNA concentration was derived using Nanodrop measurement and fluorescent quantification using Picogreen. Genome-wide DNA methylation was measured using the EPIC array. Covering over 850 000 CpG methylation sites, this microarray includes more than 90% (439 562 sites) of the CpG sites covered by the 450K array [24]. In the present replication study, although the data processing differs from that used in the initial 450K study, we have tried to use similar methods to the initial study as far as possible, except when better approaches or methods have become available or demonstrated greater efficacy. A comparison table of the methodologies used in the initial study [14] and in our replication study is provided (Table S1). Overall, samples were randomly allocated on the chips while balancing the male–female ratio on each chip. Raw signals from the EPIC array were extracted using the CHAMP pipeline. The DNAm level at each CpG site was calculated as the ratio of the intensity of fluorescent signals of the methylated alleles to the sum of methylated and unmethylated alleles [β-value, ranging from 0 (unmethylated) to 1 (fully methylated)]. DNA methylation measurement was performed by the Centre National de Recherche en Génomique Humaine (CNRGH, Evry, France).

Normalization and quality control steps

DNAm levels of the EPIC DNAm set were normalized using the InterpolatedXY method [25], which combines a noob normalization [26] with a functional normalization [27]. Normalizations on autosomes and X and Y chromosomes were conducted separately in order to avoid a sex bias [25]. Polymorphic probes [which are potentially influenced by single nucleotide polymorphisms (SNPs) underlying the entire sequence of the probe], cross-reactive probes (which are known to map several genomic loci), or probes located on sex chromosomes were excluded [28–31], leaving 758 269 CpGs. In order to limit the influence of probes with outlier fluorescent values, 0.5% of each side of the DNA methylation distribution for each probe was truncated to the nearest methylation value available. Finally, β-values were transformed into M-values, which is the log2 of the ratio of the intensities of methylated and unmethylated probes [32]:

where and are intensities measured at the ith (ranging from 1 to the number of probes with methylation measurement) methylated and unmethylated probes, respectively. is a constant offset set to 100 that helps regularizing the β-value when both methylated and unmethylated probe intensities are very low [32]. The use of M-values rather than β-values ensures the validity of further statistical analyses as well as a better detection power of differentially methylated CpG sites [32]. In order to allow for an easier interpretation of effect sizes, all regression coefficients were back-transformed using an intercept method [33] to express effect sizes on β-values, which were thus biologically interpretable in terms of change in DNAm [34]. This method consists in re-estimating the regression coefficients on a different scale, and therefore does not necessarily provide estimates that could be completely comparable to the coefficients obtained through a linear regression on β-values.

Among the 758 269 EPIC array CpGs that passed quality controls, 379 589 were common to the 450K array and passed quality controls and filtering steps in the initial study [14] (Fig. S1 and Table S2).

Adjustment factors

Adjustment factors were a priori selected and included potential confounding factors and predictors of placental DNAm levels. Predictors of placental DNAm levels comprised: season of conception (categorical variable; December–February, March–May, June–August, or September–November), child sex (categorical variable; male or female), study centre (categorical variable; Poitiers or Nancy), placental cellular heterogeneity (five continuous factors, see below), and technical factors (six continuous latent factors, see below). Confounding factors included paternal smoking status during pregnancy (categorical variable; non-smoker, smoker, or missing), maternal educational attainment (categorical variable; less than high school diploma or higher than high school), maternal age at conception (continuous variable; linear and quadratic terms), parity (categorical variable; none or more than one other child), gestational duration (continuous variable; linear and quadratic terms), and maternal pre-pregnancy body mass index (BMI; categorical variable; <25 or 25 kg/m² and more).

Cellular heterogeneity of placenta sample

The cellular composition of biological samples can confound associations investigated in epigenetic epidemiological studies [35]. We used a reference-based method to estimate the placental cellular heterogeneity from placental DNAm levels using the Placental DNAme Elastic Net Ethnicity Tool (PlaNET) package available in R [36]. This method is based on the identification of the six most important placenta-specific cellular subtypes: endothelial cells, Hofbauer cells, nucleated red blood cells, stromal cells, syncytiotrophoblast, and trophoblasts. The methylation profiles of these cellular subtypes were used as part of the reference-based method to form linear spaces on which placental DNAm levels were projected. In addition, to avoid multicollinearity and instability of our regression models, the six vectors of cell type proportions were transformed into five factors using an isometric log-ratio transformation (ILR) [37] using the compositions R package [38]. To ensure reliability of the ILR transformation, null estimated cell subtype proportions were imputed based on linear regression coefficients estimated using the impCoda function of the robCompositions package adapted to compositional data [39].

Technical factors

DNAm levels are sensitive to technical measurements features, when analysed on Illumina BeadChips [40]. Regression models were therefore adjusted for latent factors aiming at capturing technical effects and computed using latent factor mixed models (LFMMs) implemented in the lfmm R package using ridge regression [41]. LFMM allows to compute latent factors that correspond to the residual confounding factors of the association between the response and the explanatory variables. The DNA methylation levels measured on EPIC probes were regressed against all adjustment factors (see above) and maternal tobacco consumption in pregnancy. The number of latent factors was chosen based on a principal component analysis (PCA) that aimed at representing the structure of methylation data. Based on variance explained by PCA components, K = 6 latent factors were chosen to perform the LFMM estimation.

Imputation and censoring

Missing data on covariates represented slightly more than 4% of the overall data set. Five covariates had missing data: maternal pre-pregnancy BMI (n = 9), maternal educational attainment (n = 3), parity (n = 2), and maternal age at conception (n = 1). Simple imputation was performed using the mice R package [42]. Missing values were predicted using MTS status during pregnancy and all adjustment factors included in the main analysis. Fathers with missing data on tobacco exposure represented <10% of the overall data set and were not imputed, as in the initial study [14]. Gestational age at birth was censored at 42 weeks, since professional consensus among French obstetricians favours induction of labour after that point [43].

Statistical analyses

The associations between MTS during pregnancy and DNAm levels measured at each CpG were investigated using the following robust linear regression models:

where is the methylation level at CpG j for participant i and is the ith woman’s tobacco smoking status during pregnancy. The discovery analysis included j = 758 269 CpGs, while our replication analysis was based on the j = 379 589 CpGs common to both the 450K and the EPIC arrays. includes all covariates that have been described above. are independent and identically normally distributed residuals with a mean of 0 and a variance of . is the intercept for each CpG j.

Replication and exploratory analyses

The replication study was performed at probe and regional levels. At probe level, 1631 CpGs previously identified as significantly related to MTS during pregnancy and covered by the EPIC array [14] were candidates for replication (replicable probes). Probes whose uncorrected Wald test P-values of association with maternal smoking were <.05 in this replication study were considered replicated if the directions of association were similar to the initial study [14] in both current and former smokers as compared to non-smokers. At the regional level, differentially methylated region (DMR) identification was performed on 379 589 probes using the comb-p method [44], which was also used in the initial study [14] and has demonstrated a better sensitivity than over conventional methods for DMR identification [45]. A DMR was identified whenever at least two probes with Šidák multiple testing corrected P-values < .05 were included within a window of 2000 bp. We considered a DMR as replicated whenever it shared at least one common probe with a DMR identified in the initial study [14] and had similar direction of associations with MTS during pregnancy (in both current and former smokers as compared to non-smokers) in both studies. Exploratory analyses were conducted using all probes available in the EPIC array after quality controls and normalization steps (P = 758 269 CpGs) to identify potentially new pathways associated with MTS during pregnancy (Fig. 1). First, an EWAS allowed us to estimate the association between MTS during pregnancy and placental DNAm levels at each CpG. P-values were corrected for multiple testing using the Benjamini–Hochberg (BH) false discovery rate (FDR) procedure [46] at a level of 5%. The inflation of P-values was estimated using both the genomic inflation factor (GIF) and the Bayesian inflation factor (BIF), the latter being more appropriate for EWAS [47]. Based on this EWAS, an analysis identifying DMRs was conducted using the comb-p method, and the FDR significance threshold to initiate genomic regions was set to 10⁻³.

Figure 1.

Workflow of the study.

After methylation profiling, normalization, and quality control steps, two different types of analyses were performed: a probe-level analysis based on DMPs and a regional analysis based on DMRs. Both analyses were conducted as part of the main replication study (on 379 589 probes) as well as for each sensitivity and additional analyses (without gestational age adjustment, complete-case study, and negative control for paternal smoking). Additionally, an exploratory analysis, including both an EWAS and a regional analysis for DMR identification and based on the 758 269 probes included in the Illumina Infinium MethylationEPIC (EPIC) array, was conducted.

Abbreviations: BH: Benjamini–Hochberg; DMP: differentially methylated probe; DMR: differentially methylated region; EWAS: epigenome-wide association study; FDR: false discovery rate.

Characterization of probes and DMRs associated with smoking

To identify methylation profiles in the initial study, differentially methylated probes (DMPs) and DMRs were classified as ‘transient’ whenever they were associated with tobacco smoking in current smokers (vs. non-smokers) but not in former smokers (vs. non-smokers; Fig. S2). Although the methodology for probes and DMRs classification was not different from that in the initial study [14], a modification in terminology has been made. Probes initially referred to as ‘reverse’ were named as ‘transient’ in the current analyses. Conversely, alterations of placental DNAm found in both current and former smokers as compared to non-smokers were labelled ‘epigenetic memory’ (also called persistent associations), suggesting that placental DNAm alterations could persist even after smoking cessation (Table 1; Fig. S2). The label assigned to each DMR corresponded to the most represented label within its CpGs, which have to account for at least half of the CpGs included in the DMR. Otherwise, the label of the DMR was set as ‘undefined’.

Table 1.

Classification criteria for classifying DMPs and DMRs according to the smoking association patterns.

Class	EWAS P-values			EWAS betas
	Current vs. non-smokers	Former vs. non-smokers	Current vs. former smokers	Current vs. non-smokers	Former vs. non-smokers
Transient	<.05	≥.05	<.05	None	None
Epigenetic memory	<.05	<.05	≥.05	>0	>0
Epigenetic memory	<.05	<.05	≥.05	<0	<0

Classification criteria used to classify replicated DMPs and DMRs as epigenetic memory or transient. The transient profile corresponds to a significant association (P-value < .05) between maternal tobacco smoking during pregnancy and the probe levels (current smokers vs. non-smokers), but this association with methylation change is not observed in the association between former smokers and non-smokers whose methylation levels are closer to those of non-smokers (current smokers vs. former smokers). The term ‘transient’ used in our analyses is exactly what the authors of the initial study named as ‘reverse’. The epigenetic memory profile corresponds to a significant association (P-value < .05) between current smokers and former smokers and methylation levels of CpGs and methylation variation of similar strength and direction (EWAS betas). Results were adjusted for child sex, parity, maternal educational attainment, gestational age, season of conception, study centre, maternal pre-pregnancy BMI, maternal age at conception, paternal smoking during pregnancy, latent factors linked to technical measurements and estimated placental cellular heterogeneity.

Parent-of-origin-dependent DMPs and DMRs

The regulation of genomic imprinting involves various mechanisms, including parent-of-origin-dependent monoallelic differential DNAm in specific genomic regions named as gDMRs. These genomic regions show parent-of-origin-dependent monoallelic differential DNAm [48, 49]. In the placenta, gDMRs are assumed to play an important role in placental function, development, and physiology [16]. Due to the high variability of placental DNAm in imprinted regions, maternal gDMRs are suspected to arise from incomplete DNA demethylation. Placental smoking-related DMPs or DMRs located nearby placental imprinted regions would suggest that MTS may indirectly affect the placental function and development by inducing DNAm alterations in parentally imprinted regions regulating the expression of genes crucial for foetal development. Thus, we identified whether the smoking-associated DMPs and DMRs from both replication and exploratory analyses overlapped known gDMRs. Our list of candidate gDMRs consisted of (i) 80 known gDMRs [48] and (ii) 1088 other gDMRs [50].

Imprinted genes, for which parent-of-origin specific expression has been described, are suspected to regulate foetal development, as well as maternal nutrient allocation or behaviour [16, 49], through different mechanisms (including differential DNAm). Although parental imprinting in placenta has not been widely investigated and is not well understood, monoallelic expression of imprinted genes is known to be regulated by imprinted control regions such as gDMRs, whether they are located within or sometimes outside the gene locus [16]. We investigated whether MTS-associated probes and regions were annotated to known imprinted genes. These genes (n = 276) were retrieved from the Otago database [51] (86 imprinted genes reported) and/or in the GeneImprint database [52] (233 imprinted genes reported).

We performed enrichment analyses on DMPs and DMRs on both the replication and the exploratory analyses. We used the EWAS Toolkit (based on a hypergeometric test) from the EWAS Atlas database [53] to identify whether these probes or regions were characterized by specific traits or genomic locations. We also performed enrichment analyses for Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. To correct for the over-representation bias introduced by probes, which are unequally distributed among the EPIC array, we used both gometh [54] and goregion (for the set of sites included in replicated DMRs) functions from the missMethyl package [55]. These methods take into account the number of probes located on a given gene and available on EPIC to produce enrichment analyses based on the representativeness of each gene on the EPIC array.

Sensitivity and additional analyses

Firstly, we focused on participants without missing data on MTS (n = 323) to assess the influence of classifying the 18 mothers who reported quitting smoking during one trimester of the pregnancy and who were categorized as being current smokers in our main analysis.

Secondly, we removed gestational age from the adjustment factors, considering it could eventually be an intermediate factor in the relationship between MTS during pregnancy and placental DNAm levels.

Finally, a negative control analysis for paternal smoking during pregnancy was performed to help investigate the potential maternal-specific effect of tobacco smoking during pregnancy on placental DNAm. The largest source of second-hand smoking among non-smoking pregnant women being partner’s smoking [56], paternal smoking in pregnancy does not match the ideal negative control configuration [57]. Given the weaker adverse effects of partner smoking as compared to active maternal tobacco consumption [58], we expect paternal smoking to have a minimal or no impact on placental DNAm [59]. Considering paternal tobacco consumption (smoker, non-smoker, or missing; Table S3) as a negative control for the association of MTS with placental DNAm, we investigated whether paternal tobacco smoking was associated with placental DNAm at CpGs associated with maternal smoking in our main replication analysis. To account for assortative mating and behavioural convergence [60], the paternal model was adjusted for parental-shared adjustment factors (including study centre, child sex, season of conception, gestational age, six latent factors linked to technical effects, and five factors for placental cellular composition) as well as for paternal-specific confounding factors (including BMI, parity, age, and educational attainment) and the MTS status during pregnancy. Paternal-specific covariates were imputed using the same method as for the maternal model except for paternal parity, which was imputed to maternal parity as both covariates were highly correlated (0.90 χ²-correlation test coefficient).

Results

Study population characteristics

The current study population of 341 women (Fig. S1) was highly comparable to the study population of 568 women used in the initial study [14]. In both studies, median age at conception of the mothers was 29 years (interquartile range (IQR) [26–32]), median gestational duration was 40 weeks (IQR [39–41]), including 4.7% born preterm (i.e. before 37 weeks of gestation; 5.3% in the initial study [14]). More than half of the participants [219 mothers (64%)] were non-smokers, 37 women (11%) were former smokers, and the last quarter of mothers [85 participants (25%)] were current smokers (67.1% non-smokers, 12.3% former smokers, and 20.6% current smokers in the initial study [14]). Paternal-specific characteristics are provided in Table S3, including the prevalence of paternal smoking during the mother’s pregnancy, which was 35% (40% in the initial study [14]). Mothers’ and fathers’ baseline characteristics according to parental tobacco smoking status during pregnancy were highly similar. Among non-smoking fathers, the prevalence of non-smoking partner was substantial (80% had non-smoking pregnant partner; Table 2, Table S3). Baseline characteristics of the 41 women who were excluded from the analysis due to missing data on MTS during pregnancy did not present any specific pattern (Table S4).

Table 2.

Characteristics of the study population for the 341 EDEN mothers.

Characteristics	All (n = 341), n (%)	Non-smokers (n = 219 (64%)), n (%)	Former smokers (n = 37 (11%)), n (%)	Current smokers (n = 85 (25%)), n (%)	P-value
	Categorical variables
Centre					.083
Poitiers	153 (45)	91 (42)	15 (40)	47 (55)
Nancy	188 (55)	128 (58)	22 (60)	38 (45)
Sex of offspring					.12
Male	185 (54)	122 (56)	24 (65)	39 (46)
Female	156 (46)	97 (44)	13 (35)	46 (54)
Parity					.24
0 other child	148 (43)	96 (44)	20 (54)	32 (38)
≥1 other child	193 (57)	123 (56)	17 (46)	53 (62)
Maternal educational attainment					<.001
High school diploma or less	157 (46)	83 (38)	14 (38)	60 (71)
Higher than high school diploma	184 (54)	136 (62)	23 (62)	25 (29)
Season of conception					.17
December–February	82 (24)	58 (27)	10 (27)	14 (17)
March–May	76 (22)	48 (22)	6 (16)	22 (26)
June–August	91 (27)	62 (28)	11 (30)	18 (21)
September–November	92 (27)	51 (23)	10 (27)	31 (36)
Pre-pregnancy body mass index (kg/m2)					.69
<25	265 (78)	170 (78)	27 (73)	68 (80)
≥25	76 (22)	49 (22)	10 (27)	17 (20)
Paternal smoking during pregnancy					<.001
Non-smoker	190 (56)	152 (70)	18 (49)	20 (24)
Smoker	121 (35)	51 (23)	17 (46)	53 (62)
Missing	30 (9)	16 (7)	2 (5)	12 (14)
	Continuous variables
	Median	Median	Median	Median	P-value
	[IQR]	[IQR]	[IQR]	[IQR]
Average amount of daily cigarettes smoked in the 3 months before pregnancy	0.00[0.00;6.00]	0.00 [0.00;0.00]	5.00[2.00;10.0]	15.0[10.0;20.0]	<.001
Average amount of daily cigarettes smoked during pregnancy	6.67[4.33;10.0]	0.00 [0.00;0.00]	0.00[0.00;0.00]	6.67[4.33;10.0]	<.001
Maternal age at conception	28.9[25.7;32.1]	29.3[26.6;32.1]	26.6[23.9;31.6]	27.7[24.3;32.1]	.019
Gestational duration (weeks)	39.9[38.9;41.1]	39.7 [38.9;41.1]	40.4[38.9;41.3]	39.9[38.9;40.9]	.36
Estimated cellular heterogeneity
Endothelial cells	0.07[0.05;0.09]	0.07[0.05;0.09]	0.07[0.05;0.10]	0.07[0.06;0.08]	.88
Hofbauer cells	0.04[0.03;0.05]	0.04[0.03;0.05]	0.04[0.03;0.05]	0.04[0.03;0.05]	.69
Nucleated red blood cells	0.01[0.01;0.02]	0.01[0.01;0.02]	0.01[0.01;0.01]	0.01[0.00;0.02]	.98
Stromal cells	0.08[0.06;0.10]	0.08[0.06;0.10]	0.09[0.06;0.10]	0.08[0.05;0.10]	.55
Syncytiotrophoblast	0.63[0.57;0.68]	0.62[0.57;0.68]	0.63[0.56;0.69]	0.64[0.60;0.69]	.34
Trophoblast	0.16[0.12;0.20]	0.16[0.12;0.20]	0.16[0.11;0.18]	0.15[0.12;0.21]	.54

Bivariate descriptive statistics: The independence between all categorical covariates and MTS during pregnancy was tested using a Yates’ corrected version of Pearson’s chi-squared statistics (null hypothesis: two each couple of covariate-MTS during pregnancy are independent). As all distributions were non-normal, the equality of the distribution of the continuous covariates according to MTS status during pregnancy was tested using non-parametric Kruskal–Wallis test under the null hypothesis that the three sub-populations made by tobacco smoking status during pregnancy were derived from the same population.

Replication study based on DMPs

Among the 1631 candidate probes, 642 were found significantly associated (Wald P-value < .05) with MTS during pregnancy (Fig. 1, Table S5 and S6A). Among these 642 smoking-related probes, 621 had the same direction of the association as in the initial study [14], implying a replication rate of 38%. MTS associations ranged from −7% to 8% changes in placental DNAm levels in current vs. non-smokers, and included 75% of positive MTS associations (median DNAm change of 0.7%).

The top hit was cg27402634, located in the intergenic region of LINC00086 and LEKR1 genes, and identical to the one of the initial study [14]. Current smoking was associated with lower DNAm at cg27402634 compared to non-smokers with a 0.4% (P-value < .0001) decrease in our analysis and a 20% (P-value < .0001) decrease in the initial study [14]. More generally, among the top 100 probes from the initial study [14] included in our replication analysis, 45% were also found smoking-related in our replication analysis (P-value < .05) with an identical direction of the association as in the initial study [14].

Among the 621 replicated DMPs, we identified 419 probes (67%) with a transient profile (i.e. associations in current smokers vs. non-smokers only), 6 probes with an epigenetic memory profile (i.e. persistent associations in both current and former smokers vs. non-smokers), and 196 undefined probes (Tables 1 and 3). These 419 transient epigenetic marks were consistent with the initial analyses [14] (initially named as ‘reverse’) for 374 methylation sites (89%), while 17 sites had an epigenetic memory and 28 sites had an undefined methylation profile in the initial study [14]. The six epigenetic memory probes from our replication analysis were not consistent with previous results [14] in which all these sites were labelled transient upon smoking cessation (Table 1).

Table 3.

Consistency of classifications for replicated DMPs according to the smoking association patterns in both analyses.

		Rousseaux et al. [14]			Total
		Reverse	Epigenetic memory	Undefined
Replication study	Transient	374 DMPs	17 DMPs	28 DMPs	419 DMPs
	Epigenetic memory	6 DMPs	0 DMP	0 DMP	6 DMPs
	Undefined	154 DMPs	22 DMP	20 DMPs	196 DMPs
	Total	534 DMPs	39 DMPs	48 DMPs	621 DMPs

Most of the 621 replicated DMPs were labelled as transient epigenetic marks (374 DMPs) in both our replication study and Rousseaux et al. [14] (initially referred to as ‘reverse’), whereas the six epigenetic memory replicated DMPs did not have consistent profiles between both studies.

When looking more closely at those discordant CpGs, the 17 sites classified as transient, in our results and epigenetic memory in the initial study [14] could have been classified as epigenetic memory, should we have had more statistical power when comparing former smokers (n = 37) to non-smokers (n = 219). Interestingly, in previous results [14], the current vs. former smokers’ P-values were borderline significant for most of these 17 CpGs, suggesting there was both a transient effect upon smoking cessation (borderline significant) and also an epigenetic memory effect (former smokers significantly different from non-smokers). The methylation levels of both current and former smokers were different from non-smokers (epigenetic memory, statistically significant), and the methylation levels of former smokers were also different from the ones of current smokers (transient, borderline significant).

Replication study based on DMRs

Among the 203 smoking-related DMRs identified in the initial study [14], we identified 188 replicable regions that contained at least two probes common to the 450K and EPIC arrays after quality controls (379 589 CpGs). Our regional replication analysis identified 78 smoking-related DMRs (Šidák P-value < .05), including 364 CpG sites (Table S6B), among which 17 DMRs (including 73 distinct probes) were replicated with identical direction as in the initial study, implying a replication rate of 9% (Table S5). The other 61 DMRs from our replication analysis at the regional level did not have any common probe with the DMRs identified in the initial study [14] and were thus not replicated.

Most of our replicated DMRs (12 out of 17) presented a transient profile (Tables 1 and 4). Among our 12 transient replicated DMRs, 10 were also labelled as such in previous findings, whereas the other two DMRs showed an undefined methylation profile [14]. For the five remaining replicated DMRs, they were undefined in our study and classified as transient methylation patterns in the initial study [14]. None of the replicated DMRs pointed towards an epigenetic memory profile neither in the initial analyses [14] nor in our study (Table 4).

Table 4.

Consistency of classifications for replicated DMRs according to smoking association patterns in both analyses.

		Rousseaux et al. [14]			Total
		Reverse	Epigenetic memory	Undefined
Replication study	Transient	10 DMRs	0 DMR	2 DMRs	12 DMRs
	Epigenetic memory	0 DMR	0 DMR	0 DMR	0 DMR
	Undefined	5 DMRs	0 DMR	0 DMR	5 DMRs
	Total	15 DMRs	0 DMR	2 DMRs	17 DMRs

Among 17 replicated DMRs, 10 had a consistent transient epigenetic profile in both our replication study and Rousseaux et al. [14] (initially referred to as ‘reverse’), while we did not identify replicated DMRs bearing an epigenetic memory mark.

DMPs and DMRs overlap with nearby located imprinted gene loci

Among the 621 replicated CpGs, 12 were annotated to 7 distinct known imprinted genes: SGK2, TMEM52, CYP1B1, KCNQ1, FASTK, LRP1, and HOXC4. Interestingly, the four latter genes were already described in the initial study as they overlapped with several DMRs identified in their regional analysis [14]. Out of these 12 DMPs, only cg07312641 (FASTK; TMUB1) was in a replicated DMR at the regional scale of our analysis. This region was nine probes long and located nearby FASTK (imprinted) and TMUB1 genes on chromosome 7. We also identified another replicated DMR overlapping with BLCAP and NNAT (imprinted) genes, which was the longest DMR identified (16 probes long). Both replicated DMRs were annotated to FASTK;TMUB1 and BLCAP;NNAT presented lower methylation levels significantly associated with MTS during pregnancy in current smokers (vs. both non-smokers and former smokers) in the initial analyses [14] as well as in our replication analysis (Table 5). In the initial study, more than half of the CpGs included in these regions were labelled transient, and the DMRs were therefore classified with a transient epigenetic profile [14]. In contrast, both genomic regions in our replication study lacked one transient site to pass the lower limit of 50% for transient categorization (P-value < .1) and were thus undefined regions.

Table 5.

Seventeen DMRs replicated from the initial study.

DMR in Masdoumier et al. replication analysis									DMR in the initial study [14]			Replicated DMR in Masdoumier et al. sensitivity and additional analyses
DMR location information						DMR composition information
chr	Start	End	Gene name	Location	IG/gDMR	nprobe	Šidák P-value	Epigenetic profile	nprobe	Šidák P-value	Epigenetic profile	Without gestational age adjustment	Complete-case	Negative control for paternal smoking
chr1	12251718	12251935	TNFRSF1B	Body	FALSE	4	.00156	Transient	4	.00000	Undefined	TRUE	TRUE	FALSE
chr1	36807363	36807506	STK40	Body	FALSE	3	.00029	Transient	4	.00000	Reverse	TRUE	TRUE	FALSE
chr2	10766314	101766587	TBC1D8	Body	FALSE	2	.00233	Transient	3	.00000	Undefined	TRUE	TRUE	FALSE
chr4	41540115	41540230	LIMCH1	Body	FALSE	2	.00000	Transient	2	.00004	Reverse	TRUE	TRUE	FALSE
chr5	14452105	14452156	TRIO	Body	FALSE	2	.00000	Transient	2	.00000	Reverse	TRUE	TRUE	FALSE
chr7	150778724	150779155	FASTK; TMUB1	Body; TSS1500	IG	9	.00011	Undefined	16	.00000	Reverse	TRUE	TRUE	FALSE
chr8	669177	669657	ERICH1	Body	FALSE	3	.00049	Undefined	8	.00000	Reverse	TRUE	TRUE	FALSE
chr8	142275649	142275723	SLC45A4		FALSE	2	.00119	Transient	2	.00008	Reverse	TRUE	TRUE	FALSE
chr8	142310085	142310277	SCL45A4		FALSE	3	.00000	Undefined	3	.00355	Reverse	TRUE	TRUE	FALSE
chr11	71145665	71146832	DHCR7	3′UTR	FALSE	6	.00000	Transient	5	.00000	Reverse	TRUE	TRUE	FALSE
chr17	27948259	27948753	CORO6	1stExon	FALSE	3	.00001	Transient	3	.00000	Reverse	TRUE	FALSE	FALSE
chr17	38075408	38075856	GSDMB	TSS1500	FALSE	4	.00001	Transient	5	.00022	Reverse	TRUE	TRUE	FALSE
chr17	74274956	74275138	QRICH2	Body	FALSE	3	.01062	Undefined	3	.00743	Reverse	TRUE	TRUE	FALSE
chr19	49055390	49055444	SULT2B1	TSS200	FALSE	5	.00171	Transient	8	.00000	Reverse	TRUE	TRUE	FALSE
chr20	36148604	36148929	BLCAP; NNAT	5′UTR; TSS1500	IG/gDMR	16	.00511	Undefined	35	.00002	Reverse	TRUE	TRUE	FALSE
chr21	35016787	35016874	ITSN1	5′UTR	FALSE	2	.01354	Transient	2	.00046	Reverse	FALSE	TRUE	FALSE
chr21	46378243	46378626	C21orf70	Body	FALSE	4	.00000	Transient	6	.00000	Reverse	TRUE	TRUE	FALSE

Underlined lines indicate replicated DMRs that were also found significantly associated with MTS in pregnancy in the regional exploratory analysis. Criteria such as chromosome (chr), position on the chromosome (start, end), gene name, and location on the gene (location) as well as overlap with imprinted genes or germline DMR (IG/gDMR, IG if the DMR overlapped a known imprinted gene, gDMR if the DMR overlapped a known gDMR, IG/gDMR if the DMR overlapped a known gDMR located on an imprinted gene, FALSE otherwise) were reported. DMR composition, such as the number of probes (nprobe), Šidák P-value, and epigenetic profile of the DMR, was also included for our replication approach and for Rousseaux et al.’s [14] analysis. DMRs labelled as ‘transient’ in our replication study were equivalent to ‘reverse’ categorization in the initial study by Rousseaux et al. Finally, results from sensitivity and additional analyses for each replicated DMR are given.

Abbreviations: DMR: differentially methylated region; gDMR: germline differentially methylated region; IG: imprinted gene; TSS: transcription start site; UTR: untranslated region.

DMPs and DMRs overlap with candidate gDMRs

None of our replicated CpG sites overlapped a known maternal gDMR. Interestingly, among the 78 MTS-related regions identified, one replicated region was both located on the promoter of the imprinted gene BLCAP and overlapped a known gDMR, suggesting that this replicated region may be maternally imprinted.

Enrichment analyses for replicated probes and replicated DMRs

According to the traits enrichment analysis, the 621 replicated DMPs were enriched in probes known to be linked to smoking (74 CpGs), nitrogen dioxide (NO₂), air pollution (31 CpGs), but also to adverse birth outcome traits [pre-eclampsia (28 CpGs) and preterm birth (37 CpGs)], carcinoma [papillary thyroid carcinoma (36 CpGs), and oral squamous cell carcinoma (33 CpGs)] as well as in auto-immune and inflammatory diseases [multiple sclerosis (41 CpGs), systemic lupus erythematosus (43 CpGs), and psoriasis (34 CpGs); Table S7A]. Similarly, the 17 replicated DMRs were enriched in probes linked to adverse birth outcomes such as preterm birth (6 CpGs) and pre-eclampsia (3 CpGs) and various maternal traits [maternal alcohol consumption (2 CpGs) and maternal depression (2 CpGs); Table S7B].

Replicated probes were mainly located in gene bodies (52%, P-value < .001) and open seas (63%, P-value < .001; Table S7A), whereas methylation sites included in replicated DMRs were mainly located in CpG islands (51%, P-value < .001) and promoter gene regions (41% for TSS1500, P-value < .001; Table S7B).

None of the enrichment tests in GO terms or KEGG pathways performed on the 621 replicated DMPs or on the 73 CpG sites included in the replicated DMRs reached the FDR 5% significance level.

Sensitivity analyses results

Sensitivity analyses based on complete cases or not adjusted for gestational age both led to the identification of 638 DMPs [among which 586/642 (91%) and 619/642 (96%), respectively, were concordant with the main replication analysis] with the same direction of the association as compared to the main replication analysis (Fig. 2).

Figure 2.

Venn diagram for the replication analysis at the probe level.

The main replication analysis at the probe level identified 642 smoking-related DMPs out of the 1631 candidate probes taken from the initial study. Among these 642 DMPs, 619 were also significantly associated with MTS during pregnancy in the sensitivity analysis without gestational age adjustment, with identical direction of associations in current smokers (vs. non-smokers and vs. former smokers); 587 DMPs were significantly associated with MTS in complete-case; and almost all (99.8%) with identical direction of associations in current-smoking mothers (vs. non-smokers and vs. former smokers). The negative control for paternal smoking revealed 88 DMPs significantly associated with paternal tobacco smoking during pregnancy. Only 27 of them were also found smoking-associated in the main analysis and at least one other maternal approach (without gestational age adjustment and complete-case).

Abbreviations: DMP: differentially methylated probe.

In the complete-case analysis (respectively, when removing adjustment for gestational age), 84 (respectively, 78) smoking-associated DMRs, including 380 (respectively, 353) CpG sites, were identified with overlap proportions of 79% (respectively, 85%) of DMRs with those from the main replication analysis. Among the 17 replicated DMRs identified in the main replication analysis, 15 also showed up in both sensitivity analyses (Table 5). Hence, the results of our replication study were robust to the exclusion of participants with missing values and to gestational age adjustment.

Negative control for paternal smoking

Polychoric correlation between the smoking status of both parents during pregnancy was 0.47 (P-value = .0017), indicating a moderate positive correlation. Among the 642 probes associated with maternal smoking, the additional analysis performed as a negative control for paternal smoking led to the identification of 27 DMPs, mostly with the same direction of the association between mothers and fathers (20 out of 27). Half of these 27 CpGs showed weaker associations in current-smoking fathers than in current-smoking mothers (vs. non-smokers).

None of the 17 replicated genomic regions identified in the main analysis showed significant association with regard to paternal tobacco smoking during pregnancy.

Exploratory analyses on the 758 269 probes of the EPIC array

In the EWAS of the discovery analysis, P-value distributions were close to the theoretical distribution as illustrated by the BIF value of 1.03 (relatively smaller than the GIF value of 1.26). We identified 858 CpGs associated with MTS during pregnancy (with a 5% FDR-correction level). Out of these 858 DMPs, changes in placental DNAm levels ranged from −7% to 8% changes in placental DNAm levels in current smokers (vs. non-smokers), and 85% were positive MTS associations (median DNAm change of 0.6%). More than half of these CpGs (67%) were specific to the EPIC array (i.e. not included in the replication analysis), and 125 (15%) were replicated probes already identified in the replication analysis at the probe level (Fig. 3), leaving 733 new smoking-associated probes.

Figure 3.

Manhattan plot of the −log10 P-values of smoking-induced methylation changes in the discovery EWAS.

A total of 858 CpGs were significantly associated with MTS (as compared to non-smokers) during pregnancy in the discovery analysis (dots and triangles above the dashed line), among which 125 were part of the 621 replicated CpGs of the replication analysis. Triangles illustrate EPIC-specific CpG sites that were not included in the replication analysis.

Abbreviations: EWAS: epigenome-wide association study.

Among these 733 new smoking-related probes identified from the exploratory EWAS, we did not identify any probe with an epigenetic memory profile. However, 133 (18%; Fig. 1) of these new smoking-related probes showed transient methylation patterns in smoking pregnant women (P-value < .05).

We further identified 84 DMRs associated with MTS, including 299 distinct CpGs. Among these, 9 DMRs were among the 17 replicated ones (underlined rows in Table 5). In addition to demonstrating the robustness of the replication results, this exploratory analysis also showed that among the 75 new smoking-related regions (including 261 CpGs), some were overlapping with DMRs from previous findings [14] despite they were not replicated in our replication approach (same region but different probes). For example, four DMRs, respectively, annotated to TINAGL1 (six probes long), TMEM136I (two probes long), SYNGR1 (seven probes long), and PARD3 (two probes long) were overlapping and with the same direction (current smokers vs. non-smokers and vs. former smokers) with as many DMRs in the regional analysis of the initial study [14] but were absent from our regional replication analysis (Fig. S3).

Among these 75 regions identified, 29% (22 DMRs) were labelled as transient, whereas no DMR was labelled epigenetic memory (P-value < .05; Table 1).

New loci overlapping with known imprinted genes among the new smoking-related DMPs and DMRs

Among the 733 new DMPs, 10 matched known imprinted genes: RB1, HNF1A, NNAT, FOXF1, THUMPD2, L3MBTL1, PRDM16, DLGAP2, GLIS3, and CYP1B1, some of which were already identified in the replication approach. At the regional scale, 3 DMRs among the 75 new regions identified overlapped with known imprinted genes, namely RB1 (2 CpGs), CYP1B1 (3 CpGs), and CALCR (8 CpGs).

New loci overlapping with candidate gDMRs among the new smoking-related DMPs and DMRs

As for the exploratory EWAS and regional analyses, none of the 733 new DMPs overlapped with known gDMRs [48]. At the regional scale, one region out of the 75 new MTS-related DMRs, annotated to DNAH7 gene (within the 5′UTR), overlapped with a maternal gDMR.

Enrichment analyses for new smoking-associated probes and regions from discovery analysis

The 733 new DMPs from the exploratory EWAS were enriched in probes linked to respiratory traits [fractional exhaled nitric oxide (31 CpGs) and allergic and childhood asthma (6 CpGs)], metabolic traits [BMI (5 CpGs) and fat-free mass index (2 CpGs)] among others (Table S7C). As for the CpGs included in the 75 new smoking-related DMRs (including 261 methylation sites), they were enriched in probes linked to smoking (23 CpGs), maternal NO₂ exposure during pregnancy (2 CpGs), pre-eclampsia (7 CpGs), ancestry (27 CpGs, including Werner syndrome) as well as ageing (28 CpGs; Table S7D). The smoking-related CpGs following the exploratory EWAS were mainly located in open seas (77%, P-value < .001) and genes bodies (51%, P-value < .001; Table S7C), whereas probes from the regional discovery analysis were enriched in promoter regions (TSS200 (23%), P-value < .001) and intergenic regions (17%, P-value < .001; Table S7D). None of the enrichment tests in GO terms or KEGG pathways performed on these probes reached the FDR 5% significance level.

Sensitivity analyses results

The sensitivity analysis performed on complete cases identified a total of 754 (72% concordance rate with the 858 hits in the main exploratory EWAS) smoking-related DMPs and 86 DMRs (including 303 CpG sites), among which 64 (76%) were concordant with main exploratory analysis at the regional level. The second sensitivity analysis performed on the model without adjustment for gestational age was completely concordant with the main exploratory analyses (at both probe and regional levels). None of the smoking-associated DMPs nor DMRs from the main discovery analyses showed statistically a significant association in the negative control for paternal smoking, witnessing the leading role of maternal tobacco consumption in the associations previously identified.

Discussion

Summary and interpretation of the results

Following up on an initial study using the 450K array [14], we performed a replication study of maternal smoking-related methylation loci (probes and regions) in 341 new placenta samples (85 smokers, 219 non-smokers, and 37 former smokers) using the EPIC array and a similar study design for statistical analyses. From probes available on both the 450K and the EPIC BeadChips, we replicated 38% and 9% of smoking-related DMPs and DMRs identified in [14], most of which (60% of DMPs and 59% of DMRs) had transient methylation patterns witnessing no persistent association after smoking cessation in our replication analysis as well as in the initial study [14]. Among the 621 replicated DMPs, 12 overlapped with known imprinted genes (SGK2, TMEM52, CYP1B1, KCNQ1, FASTK, LRP1, and HOXC4), while 2 out of the 17 replicated DMRs were annotated to imprinted genes (namely NNAT and FASTK). Finally, a discovery EWAS performed on 758 269 CpG sites on the EPIC array identified 733 new smoking-associated DMPs, most of which were EPIC-specific probes not included in previous analyses as well as 75 new smoking-related DMRs. Among these new DMPs and DMRs, some overlapped with imprinted genes, and none was classified as epigenetic memory. All these analyses were subject to a negative control for paternal smoking, which led to a very limited number of smoking-associated DMPs and DMRs, suggesting a specific intrauterine effect of tobacco consumption during pregnancy on placental DNAm levels.

The top-hit CpG in both our replication study and the initial study [14] was cg27402634 located in the intergenic region of the LINC000086 and LEKR1 genes with a transient methylation profile in both studies [14]. In addition, cg27402634 did not show any association with paternal cigarette smoking in our negative control for paternal smoking, which is consistent with another study [61] in which the authors argued in favour of an intrauterine effect of MTS in pregnancy on its DNAm changes. cg27402634 was in the top smoking-related methylation sites in several other epidemiological studies investigating the relation between MTS in pregnancy and placental DNAm [15, 18, 61–63] witnessing the robustness of this result. In previous studies, including the PACE meta-analysis, cg27402634 was inversely associated with prenatal tobacco smoke exposure (9.3% [61] lower DNAm, 29.8% [62] and −25% [18] DNAm changes), which was consistent with the initial findings [14] (regression coefficients of −20%, P-value < .0001, in current smokers vs. non-smokers) but with higher effect sizes than in our replication study (regression coefficient of −0.4%, P-value < .001). However, alterations in methylation levels of cg27402634 did not show any correlation with neighbouring CpGs, impeding the possibility of identifying any DMR, including this locus, which is consistent with the observation of a previous study [61]. Placental DNAm patterns at this CpG site were also found to mediate the relationship between MTS during pregnancy and lower birth weight [61], but this mediated effect was not found in the EDEN cohort [64].

Overall, the results of our replication analysis at the probe level were consistent with those of the initial study [14] and included a high number of replicated DMPs (268 hits, 43%) already significantly associated with maternal smoking in the PACE consortium’s meta-analysis on 450K array data [18]. In a recent study aiming at validating placental DNAm targets for prediction of MTS during pregnancy [15], five CpGs were selected in two different models (one EPIC-based model and another 450K-based model) as they represented strong predictors of MTS. These five probes have shown biological relevance to tobacco smoking (cg27402634 and cg08103568) as well as to preterm birth and pre-eclampsia (cg07168214), ancestry (cg04233054), or asthma (cg08621277) [15]. It is worth noting that all of these 5 CpGs were part of our 621 replicated loci, suggesting the strong relevance of placental methylation changes in the relation with MTS in pregnancy at these specific sites.

Among the 17 replicated DMRs, a DMR with two probes located in TRIO gene was associated with higher methylation levels in our smoking pregnant women (median DNAm change of 32%, P-value < .0001) as well as in initial analyses [14] (median DNAm change of 7%, P-value < .0001) compared to non-smokers and showed a ‘transient’ effect upon smoking cessation in pregnancy. These two probes were identified in several EWASes [61, 62] and one meta-analysis [18] in which they were associated with an increased DNAm (around 8% change) with no evidence of a mediating role on birth weight [62, 64]. TRIO is assumed to be decreasingly expressed due to its interaction with benzo(a) pyrenes present in tobacco smoke, which can support the positive association of DNAm with maternal smoking in pregnancy observed in our replication study [14].

One of the largest smoking-related replicated DMRs was located on the so-called micro-imprinted domain BLCAP; NNAT, which presents an imprinted gene (NNAT) contained within the genomic structure of BLCAP, which was demonstrated to be non-imprinted [65, 66]. This replicated DMR was four probes shorter than in the initial study [14]. This DMR was associated with lower methylation in smokers as compared to non-smokers (median DNAm change of −12%, P-value < .01 in our replication analysis) and showed a transient methylation profile with regard to smoking cessation in pregnancy in the initial study only [14]. Our limited sample size compared to previous findings [14] may have induced a lack of statistical power that could explain the classification of this DMR as undefined in our replication study. Indeed, several CpGs included in this DMR had borderline-significant methylation changes in current smokers as compared to non-smokers and former smokers. The location of this DMR in the promoter region of BLCAP; NNAT placental micro-imprinted domain may suggest a direct effect of MTS on the gene expression, which might impact foetal growth programming [66]. Increased gene expression on BL CAP; NNAT was also previously associated with birth weight and large for gestational age [67].

As part of a discovery analysis, we explored any additional smoking-related changes in DNA methylation using all probes included in the EPIC array (758 269 sites). The presence of EPIC-specific CpGs allowed identifying three new DMRs overlapping with three regions identified in the initial study [14], but which did not show up in our replication approach, even though they were replicable regions. For example, the three-probe-long DMR located close to TINAGL1 gene in the initial study [14] was completely covered by a six-probe-long DMR in our exploratory analysis. Both regions shared three common probes, including one replicated DMP (cg16103203). These DMRs were associated with decreased methylation levels in current smokers (vs. non-smokers) and showed transient methylation patterns in both studies [14]. Placental expression of the TINAGL1 gene has been related to angiogenesis [68] and is down-regulated in pre-eclamptic women [69]. Overall, these three smoking-related regions in the exploratory analysis, overlapping with DMRs from the initial study [14] increased our replication rate by 9% to 11%.

We investigated epigenetic profiles of replicated DMPs and DMRs

Although most of the 419 ‘transient’ replicated probes identified in this study had consistent labels (89%) with the initial study, 17 replicated DMPs had epigenetic memory profiles in the initial study as the authors did not identify significant differences in DNAm patterns in current smokers vs. former smokers [14]. Regression P-values for several of these 17 DMPs were, however, close to the significance threshold (P-values ranging from .0543 to .808, with median P-value of .148) [14] indicating a trend towards a difference in DNAm changes in current smokers vs. former smokers, which would have resulted in a transient profile. In other words, this suggests the presence of ‘partially transient’ associations upon smoking cessation as the methylation levels of former smokers were not as high (i.e. low) as current smokers, and tended to reach those of non-smokers. As for our regional analysis, most of the replicated DMRs had consistent transient epigenetic marks in both studies (10 out of 17 regions) [14].

The ‘partially transient’ associations suggested in our study may result from DNAm marks identified in placenta samples that were not directly exposed to tobacco smoking. The underlying biological mechanisms by which the placenta could bear epigenetic marks of past exposure prior to its development are unknown, although several hypotheses have been put forward [70, 71]. These include the transmission of epigenetic marks to the oocyte, which would persist through placental cell differentiation, as well as other processes unrelated to epigenetics, such as modification of histological features through either direct cell contact or endocrine signalling [14].

Although this was used in the initial study, we considered that the use of ‘reverse’ to describe smoking-related DMPs or DMRs in current smokers that were no longer significant in mothers who quitted smoking could be scientifically misleading. Placental DNAm is a highly dynamic epigenetic process that varies greatly throughout pregnancy, so placental DNAm levels measured at birth are not capturing the dynamics in DNAm levels during the whole pregnancy period. In this analysis, the ‘transient’ terminology was preferred over ‘reverse’ to define MTS associations in current smokers (vs. non-smokers) that were not found in former smokers (vs. non-smokers). In any case, the authors used the term transient to describe the dynamics of placental DNAM throughout pregnancy. Some non-replicated results classified as transient may also be the result of some unmeasured confounding or lack of statistical power between the two groups of women. Overall, we identified few replicated epigenetic memory loci (six DMPs), and all of them were initially classified as transient [14]. Given the relatively small sample of former smokers (n = 37) used here as compared to the initial study, which included 70 former smokers among 568 participants [14], we might have identified a lower number of DNAm patterns using this sub-population, thus contributing to an under-estimation of the proportion of epigenetic memory probes. Beyond the limited statistical power of our replication study, differential site-specific penetrance of tobacco smoke exposure might explain non-replication of some loci as well as the inconsistent epigenetic marks (transient or reverse, epigenetic memory, or undefined) observed between our replicated loci and the ones from the initial study [72]. Inconsistency in epigenetic profiles of some smoking-associated loci identified as smoking-associated from the initial study [14] might suggest that these sites show low and/or high instability with regard to tobacco smoke exposure penetration. Different profiles of replicated loci may suggest a strong but differential site-specific penetration of tobacco smoke. Indeed, as well as having fewer number of former smokers (n = 37), the former smokers included in our replication study smoked fewer daily cigarettes during the 3 months before the pregnancy (5.5 daily cigarettes smoked on average) than the 70 former smokers of the initial study (7 daily cigarettes smoked on average) [14]. This may suggest that replicated loci with different epigenetic profiles might show a high sensitivity to the dose of pre-conceptional tobacco smoke exposure.

Beyond investigating associations between placental DNAm and maternal tobacco consumption during pregnancy, we investigated a maternal-specific effect of cigarette smoking in pregnancy using a negative control for paternal smoking [59]. In all analyses, the negative control using paternal smoking led to a nearly complete drop in the number of smoking-related loci, suggesting an intrauterine effect of tobacco smoke exposure on placental DNAm. Among the 27 concordant smoking-associated DMPs for the two parents, 20 associations were in the same direction and in most cases with a weaker association with paternal than MTS. Even though polychoric correlation between maternal and paternal smoking during pregnancy was substantial in our data (0.47, P-value = .0017), these results pointed towards the existence of a maternal-specific effect of cigarette smoking in pregnancy on placental DNAm.

Strengths and limitations

Although several epidemiological studies have reported various effects of prenatal tobacco smoke exposure on DNAm in several tissues [11, 61–  63, 73], results are not always replicable, and the use of replication samples remains rare and insufficient [8]. In addition, evidence from large meta-analyses in diverse populations is needed to broaden the scientific understanding of tobacco-related DNAm damage further. Our results validate and bolster existing evidence on smoking-sensitive methylation sites in the placenta [14], while closely relying on the methodology outlined in the initial study [14].

The use of placenta samples rather than cord blood samples, which has been one of the most extensively studied tissues, is a strength. Indeed, the placenta is a transient organ highly involved in foetal development. The placenta ensures the proper exchange of nutrients and oxygen between mothers and their foetuses throughout pregnancy [12]. Investigating the associations between MTS and DNAm levels in placental samples is therefore fundamental, especially as the DNAm process has been shown to be highly tissue-specific [74, 75].

Despite efforts limiting the number of excluded participants, our replication study relied on a smaller sample of pregnant women than the previous work [14].

Parental smoking habits were self-reported, which could eventually lead to under-estimations of cigarette consumption and thus introduce a measurement error, which can lead to biased estimates in this replication study. However, in a subsample of 70 EDEN mothers included in our study population, urine cotinine levels were measured and validated most of self-reported smoking phenotypes attesting for a satisfying accuracy of pregnant women’s statements. Even though information on fathers suffered from missing data (10% of the overall participants), proportions of smoking and non-smoking fathers remained similar to the ones observed among mothers, and smoking data on fathers allowed us to conduct a negative control for paternal smoking [58].

Our replication study estimated cellular compositions using a reference-based method [36] to ensure accurate cellular heterogeneity prediction and our robust regression models were adjusted for latent factors representing technical effects. Although these methods are more relevant, they differ from the ones that have been used in the initial study, which might have influenced the results.

Besides, different arrays were used to measure DNAm levels, which might have affected the measured methylation levels [24] and limited the number of replicated DMPs or DMRs. However, the use of the EPIC instead of the 450K array enabled us to expand the analysis to new CpG sites not available in the initial study [14]. As a result, our exploratory analysis performed on a larger amount of methylation sites enabled us to identify and validate new genomic regions overlapping with previous findings [14].

Conclusions

This study replicated 38% of differentially methylated probes and 9% of differentially methylated regions associated with MTS in a previous study [14]. Our replicated findings point to genomic regions containing imprinted genes and genes associated with pregnancy outcomes as preferential tobacco targets, suggesting mechanisms by which tobacco could directly impact the foetus and future child. We further show that most replicated loci were classified as transient DNAm changes upon smoking cessation before pregnancy, which might support pregnant women’s efforts to quit smoking. Our results reinforce previous knowledge on placental epigenetic sensitivity to prenatal smoke exposure and demonstrate the importance of reproducible work in omic studies to provide a robust understanding of the variation of such measures.

Author contributions

Chloé Masdoumier (Conceptualization [equal], Formal Analysis [equal], Investigation [equal], Methodology [equal], Software [equal], Validation [equal], Visualization [equal], Writing – original draft [equal], Writing – review & editing [equal]), Lucile Broséus (Data curation [equal], Formal Analysis [equal], Methodology [equal], Validation [equal], Writing – review & editing [equal]), Florent Chuffart (Validation [equal], Writing – review & editing [equal]), Olivier François (Methodology [equal], Writing – review & editing [equal]), Ariane Guilbert (Validation [equal], Writing – review & editing [equal]), Barbara Heude (Data curation [equal], Funding acquisition [equal], Writing – review & editing [equal]), Saadi Khochbin (Validation [equal], Writing – review & editing [equal]), Sophie Rousseaux (Writing – review & editing [equal]), Emie Seyve (Data curation [equal], Writing – review & editing [equal]), Muriel Tafflet (Data curation [equal], Writing – review & editing [equal]), Jorg Tost (Data curation [equal], Writing – review & editing [equal]), Aurélie Nakamura (Conceptualization [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Software [equal], Supervision [equal], Validation [equal], Writing – original draft [equal], Writing – review & editing [equal]), and Johanna Lepeule (Conceptualization [equal], Formal Analysis [equal], Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Supervision [equal], Validation [equal], Writing – original draft [equal], Writing – review & editing [equal]).

Conflict of interest

The authors declared no competing interests.

Funding

This study was supported by grants from the National Institute of Cancer (INCa) and the French Institute for Public Health Research (IReSP) (INCa_13641). DNA methylation measurements were obtained thanks to grants from the Fondation de France (numbers 2012-00031593 and 2012-00031617) and the Fondation pour la Recherche Médicale (EPImEx project). The EDEN study is supported by the FRM, French Ministry of Research: Federative Research Institutes and Cohort Program, INSERM Human Nutrition National Research Program, and Diabetes National Research Program in collaboration with the French Association of Diabetic Patients (AFD), French Ministry of Health, French Agency for Environment Security (AFSSET), French National Institute for Population Health Surveillance (InVS), Paris-Sud University, French National Institute for Health Education (INPES), Nestlé, Mutuelle Générale de l’Éducation Nationale (MGEN), French-speaking Association for the Study of Diabetes and Metabolism (ALFEDIAM), National Agency for Research (ANR), and National Institute for Research in Public Health (IRESP: TGIR 2008 cohort in health program).

Data availability

The data supporting the findings cannot be made publicly available for ethical and legal restrictions purposes. The data included information that could be used to re-identify EDEN participants and could compromise the research participant’s privacy. However, data underlying the findings can be available from the corresponding authors on reasonable request and with the consent of the EDEN principal investigator and Steering Committee. The code used to generate the results is available upon request to the corresponding authors pending application and approval.