Epigenome-wide association study of maternal hemoglobin A1c in pregnancy and cord blood DNA methylation

Abstract

Background:

Gestational hyperglycemia is associated with adverse perinatal outcomes and long-term offspring metabolic programming, likely through dysregulation of DNA methylation (DNAm).

Materials & methods:

We tested associations between maternal HbA1c and cord blood DNAm among 412 mother–child pairs in the genetics of glucose regulation in gestation and growth (Gen3G) and implemented Mendelian randomization to infer causality. We sought replication in an independent sample from Healthy Start.

Results:

Higher second trimester HbA1c levels were associated with lower DNAm at cg21645848 (p = 3.9 × 10^-11) near URGCP. Mendelian randomization and replication analyses showed same direction of effect between HbA1c and DNAm at cg21645848, but did not reach statistical significance.

Conclusion:

We found that higher maternal glycemia reflected by HbA1c is associated with cord blood DNAm at URGCP, a gene related with inflammatory pathways.

Lay abstract

An adverse metabolic environment in utero, such as elevated maternal blood glucose levels, may have long-lasting impacts on child development. Studies show that maternal glucose can cross the placenta and may impact regulatory markers acting upon genes through epigenetics. We investigated associations between 3 month maternal glucose levels, reflected by HbA1c, during pregnancy and cord blood DNA methylation (one type of epigenetic marker) in two prospective cohorts of mother–child pairs. We found that higher maternal blood glucose levels were associated with lower cord blood DNA methylation near a gene called URGCP, suggesting potential influence of elevated maternal glucose on inflammatory pathways in the newborn.

Hyperglycemia during pregnancy is associated with impaired glucose tolerance and adverse cardiometabolic outcomes in the exposed offspring later in life [1–3]. Current evidence indicates that maternal glycemia, as a continuous measure, is linearly associated with perinatal complications for the mother and her child [1,4–6], and with adverse metabolic programming in the offspring later in life [6–8]. Glycated hemoglobin (HbA1c) can be used to diagnose diabetes outside of pregnancy and to monitor glycemia, as it provides a standardized measure of average blood glucose over approximately 3 months period [9,10]. In a normal pregnancy, levels of HbA1c are generally lower than non pregnant women due to a faster red cell turnover, and a physiologic decrease in fasting blood glucose [9,11,12]. Still, higher levels of maternal HbA1c during pregnancy, in the range below the threshold for pre-existing diabetes diagnosis, have been associated with various adverse perinatal outcomes, including higher risk of pre-eclampsia, preterm birth, shoulder dystocia, congenital anomalies and perinatal loss [13,14]. Additionally, higher prenatal maternal HbA1c has also been associated with higher adiposity in the offspring at 11 years [7]. Thus, maternal hyperglycemia reflected by HbA1c may affect fetal programming at key stages during development.

Animal models and human studies with careful control for genetic effects [15,16] have demonstrated that maternal hyperglycemia and gestational diabetes mellitus (GDM) may influence fetal programming via epigenetic mechanisms [16–21]. The most studied of these mechanisms is DNA methylation (DNAm) [22,23]. Current evidence of associations between GDM and DNAm levels in fetal tissues (cord blood and placenta), suggests dysregulation of DNAm in or near genes involved in placental function, fetal development [19,24] and in disease-related metabolic pathways [25]. We have recently demonstrated that elevated 2 h glucose after a 75 g oral glucose tolerance test (OGTT) is associated with alteration in DNAm levels in placental tissue at key inflammatory genes [21]. Apart from this study, few others have addressed the association between maternal hyperglycemia along the continuous spectrum of glucose, and genome-wide DNAm variation in fetal tissues, particularly in cord blood. We hypothesize that maternal HbA1c would capture longer integrated hyperglycemic exposure during in utero development compared with traditional glycemic markers, and reveal novel epigenetic markers associated with offspring DNAm.

Therefore, we aimed to identify novel DNAm markers of fetal adaptation to intrauterine glycemia by conducting an epigenome-wide association study (EWAS) of HbA1c measured during pregnancy, and cord blood DNAm. In the genetics of glucose regulation in gestation and growth (Gen3G) cohort, we measured HbA1c in the first trimester (capturing glycemic exposure from pre conception to early pregnancy) and at the end of the second trimester (capturing the mid-pregnancy period). Additionally, we sought to replicate markers from our EWAS in an independent study (Healthy Start), and we implemented the Mendelian randomization (MR) approach [26] to investigate causality in the association between maternal HbA1c and DNAm at top differentially methylated sites. Robust DNAm markers detected at birth may inform a later in life risk of adverse metabolic programming and provide markers for early intervention.

Materials & methods

Study population

In this study, we included mother–child pairs from Gen3G, a Canadian prospective prebirth cohort study. Gen3G was designed to investigate the determinants of glucose regulation in pregnancy and their influence in fetal development [27]. Expecting mothers enrolled in the study were ≥18 years of age, mainly of Caucasian origin, had singleton pregnancies, and no records of pre-existing diabetes or diabetes detected in the first trimester [27]. Gen3G research staff collected biological samples and sociodemographic, lifestyle and anthropometric variables from the mothers in the first and second trimesters of pregnancy, at delivery, and from the newborns at birth. These variables have been described in detail elsewhere [27]. For this study, we selected a sample of mother–child pairs based on the availability of purified DNA from full-term (i.e., gestational age at birth ≥37 weeks) cord blood samples to conduct the DNAm analyses (Supplementary Figure 1). All participants provided written informed consent prior to enrollment in agreement with the Declaration of Helsinki. The institutional review board from the Centre Hospitalier Universitaire de Sherbrooke (CHUS) approved all study protocols.

Phenotype measures

We measured maternal anthropometrics at each visit using standardized procedures. Weight was measured with an electronic calibrated scale, and height with a wall stadiometer. We calculated BMI as weight/height² (in kg/m²). Mothers reported their smoking status in the first trimester (coded as non smoker vs current smoker). We screened all women for GDM in the second trimester using a 75 g OGTT (mean ± SD weeks of gestation 26.4 ± 1.0). GDM was diagnosed following the International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria [27]. For our primary analyses, we excluded mothers who were diagnosed with GDM from these cord blood DNAm analyses to avoid bias by GDM treatment between time of HbA1c measurement and DNAm assessment (birth). We quantified HbA1c (%) from whole blood samples collected during the first and second trimesters using HPLC (Bio-Rad VARIANT, CA, USA) [27], a method certified by the National Glycohemoglobin Standardization Program [28], at the CHUS biomedical laboratory. We used the multiplexed particle-based flow cytometric assays (Human MILLIPLEX map kits, EMD Millipore, MA, USA) to measure TNF-α concentrations (pg/ml) from cord blood samples collected at delivery (within 30 min postpartum) by trained research staff.

Cord blood DNA extraction

Trained personnel collected cord blood samples from each delivery using a standardized protocol, and then stored the samples at -80°C until analysis. We extracted DNA from cord blood buffy coats using a protocol previously described [21].

Epigenome-wide DNAm measurement

We performed bisulfite conversion of genomic DNA from cord blood using the EZ-Methylation Kit (Zymo Research, CA, USA), followed by DNAm quantification utilizing the Infinium Methylation EPIC BeadChip (Illumina, CA, USA). To minimize batch effects, we randomly allocated samples to different plates and chips. We applied quality control (QC) of DNAm data at the sample and probe levels. At the sample level, we removed samples with low quality (n = 3), technical replicates (n = 10) and samples with genotype mismatch (n = 6 samples; based on SNPs on EPIC array from paired cord blood and placenta samples). At the probe level, we excluded probes with detection p > 0.05 in at least 5% of the samples [29] (n = 1,754), probes with null variance [30] (n = 125), probes in sex chromosomes [31,32] (n = 19,128 sites), non-CpG probes [31] (n = 2,836 sites), SNP-associated probes at the single base extension (n = 5,547 sites) [32] or at the target CpG site with MAF ≥5% [31] (n = 5668 sites), and cross-reactive probes [33] (n = 40,454 sites). In total, we retained for analyses 791,324 CpG sites in 412 cord blood samples from non-GDM mothers that passed all QC steps. We corrected for non specific background signals using the minfi R package [34], and controlled for additional unwanted technical variation using functional normalization with two principal components derived from control probes [35]. To correct for probe-type bias, we implemented the regression on correlated probes method, which uses the correlation between consecutive probes to adjust the distribution of type II probes [36]. Finally, we used ComBat [37] in the sva R package [38] to adjust for sample plate. We performed all the statistical analyses using DNAm on the M-value scale (logit transformation of methylation β-values) considering the statistical validity of this method [39]. However, to ease interpretability, we also showed effect estimates at our top findings in the β-value scale. Associations at the individual CpG sites were interpreted as the mean percent difference in DNAm (in a range between 0% or completely unmethylated, and 100% or completely methylated), per 1%-unit increase in HbA1c.

Replication in an independent study: the Healthy Start cohort

The Healthy Start cohort is a prospective prebirth cohort including mother–infant pairs of multiple ethnicities (non-Hispanic white, Hispanic and non-Hispanic African–American). Women in the study were recruited from the obstetric clinic at the University of Colorado Anschutz Medical Campus (CO, USA) from 2009 to 2014 [40]. Eligible women had singleton-pregnancies, were aged ≥16 years, and had no history of preterm births, stillbirths or chronic conditions including diabetes, cancer, steroid-dependent asthma or psychiatric conditions [40]. The standard clinical care was to offer women screening for GDM using a 1 h, 50 g glucose challenge test at about 26 weeks of gestation. If screening values of glucose during the glucose challenge test were ≥7.8 mmol/l, mothers were asked to undergo a 3 h, 100 g glucose test, with confirmation of GDM when Carpenter–Coustan criteria were met [41]. For this analysis, mothers with physician diagnosis of GDM were excluded to avoid confounding by GDM or its treatment. HbA1c was quantified from whole blood drawn in women that were in their 20–34th week of pregnancy (second–third trimester) using the DCA Vantage analyzer (Siemens, PA, USA) system, a method certified by the National Glycohemoglobin Standardization Program [28]. HbA1c quantification was performed at the University of Colorado Hospital Clinical and Translational Science Institute Core Laboratory. The study protocol was approved by the Colorado Multiple Institutional Review Board. Maternal participants provided written informed consent prior to study enrollment.

In the Healthy Start cohort, DNAm levels in cord blood were measured using the Illumina Infinium HumanMethylation 450K array. We excluded probes with detection p > 0.05, bead-count <3 in at least 5% of the samples, and samples with mismatch between predicted sex (by DNAm) and reported sex. Normalization of DNAm was conducted using the preprocessQuantile function in the minfi R package [34], with correction for batch effects using ComBat from the sva R package. We investigated in silico replication of CpG sites discovered in Gen3G with association p < 1.0 × 10^-6. In total, we performed replication analyses in 506 mother–child pairs (58% non-Hispanic white) with complete measures of second trimester maternal HbA1c, using normalized DNAm (M-values) from cord blood samples.

Statistical analyses

EWAS of maternal HbA1c versus cord blood DNAm

We described baseline characteristics of 412 mother–child pairs in Gen3G included in this analysis using the mean and standard deviation (SD), or frequency and proportions. We compared baseline characteristics of participants in our study sample, with that of 456 individuals in Gen3G for whom HbA1c levels were measured at the second visit, but who were not included in our DNAm analysis. We conducted epigenome-wide assessment of differential methylation at the individual CpG sites in relation to levels of maternal HbA1c (%) using fitted robust linear regressions (RLR). For these analyses, DNAm in the M-value scale was modeled as the response variable, and maternal HbA1c at first or second trimester as the predictor. RLR was used to account for potential heteroskedasticity in the distribution of DNAm [21,42]. To control for differences in DNAm arising from cellular heterogeneity, we estimated cell counts for six white-cells (CD4⁺T, CD8⁺T, monocytes, granulocytes, natural killer cells & B cells) and nucleated red blood cells in cord blood using the Balkulski reference panel [43]. Models were adjusted for maternal age, maternal BMI, gestational week at HbA1c measurement, gravidity, gestational age at delivery, child sex, smoking status during pregnancy and estimated cord blood cellular composition. CpG-by-CpG associations were controlled for multiple testing using Bonferroni (α = 6.32 × 10^-8), but we considered as ‘suggestive’ associations CpG sites identified with p < 1.0 × 10^-6, and we took forward for replication and MR analyses this list of CpGs. To illustrate results of the EWAS, we used genome-wide Manhattan plots, volcano plots and Q-Q plots of p-values. We calculated the genomic inflation factor (γ) using the median values of the observed versus the expected chi-squared test statistic for the distribution of p-values in the EWAS. Finally, we used regional association plots to illustrate our findings at top differentially methylated CpGs. All analyses were conducted in R (version 3.5.1) [44].

To explore the effect of timing of glycemic exposure on our results, we took CpGs identified in the analysis with second trimester HbA1c and inspected their association with maternal HbA1c values in the first trimester among 410 mother–child pairs in Gen3G with complete measures of HbA1c, offspring DNAm and covariates. Regressions were adjusted for the same covariates as in the initial analysis with second trimester HbA1c. To determine if associations observed with offspring DNAm were driven by inflammatory processes, we tested the associations of maternal HbA1c and of DNAm at CpGs identified in the main EWAS with predicted cord blood cell types, and with the inflammatory marker TNF-α measured in a subset of cord blood samples (n = 406). For this analysis, HbA1c and DNAm were independently modeled as the exposure, while predicted count for cord blood cells and TNF-α were considered as the outcomes in multivariable linear regressions adjusted for maternal (age, gestational week at HbA1c measurement, gravidity, BMI and smoking) and newborn covariates (sex and gestational age at delivery). TNF-α values were log-transformed before the analysis. Associations between cells and HbA1c were considered statistically significant at p < 0.01 (α = 0.05/7 cell types), between cells and the CpGs at p < 0.002 (α = 0.05/4 CpGs tested*7 cell types), and between TNF-α and the CpGs at p < 0.01 (α = 0.05/4 CpGs tested).

Sensitivity analyses

We performed a secondary analysis to investigate if a non linear relationship was present between HbA1c and cord blood DNAm. For this analysis, we categorized HbA1c into quartiles and measured difference in DNAm at the individual CpG site (in M-values) when comparing the first (Q1 or lower HbA1c) versus the top quartile (Q4 or higher HbA1c) of HbA1c distribution, adjusting for same covariates as in our main analysis with continuous HbA1c. To rule out potential overfitting of the model by adjusting for maternal BMI, we tested correlations between BMI and HbA1c using Pearson correlations, and we inspected regression estimates for our top hits identified in the main EWAS using a model without adjustment for BMI.

Association between discovered CpGs & OGTT glucose levels

To determine robustness and specificity of our EWAS signals to HbA1c levels, we investigated associations between top differentially methylated CpGs and additional glycemic markers measured during the OGTT (i.e., fasting glucose, 1- and 2-h glucose levels). For this analysis, we used multiple RLRs with DNAm at the identified CpGs in the EWAS (in M-values) as the outcomes, and the glycemic markers as independent exposures. Associations with glucose levels were adjusted for the same covariates as in the discovery analysis using second trimester maternal HbA1c. To allow comparability in the interpretation of effect estimates obtained in the analysis with OGTT glucose (mmol/l) and HbA1c (%), we transformed values for these measures of glycemia to z-scores (mean = 0 and SD = 1) before the analysis. We interpreted association estimates at the individual CpGs as the mean percent difference in DNAm β-values, per SD increase in the glycemic measure. We considered associations with other glucose values at p < 0.02 (α = 0.05/3 CpGs tested).

Replication of top differentially methylated CpGs in the Healthy Start cohort

In the replication analysis, we adjusted associations for self-reported maternal ethnicity, maternal age, child sex, gestational age at visit, gestational age at delivery, gravidity (primigravid vs multiparous), smoking status during pregnancy (smoker vs non smoker), predicted count for seven cell types in cord blood [43], and maternal BMI (kg/m²) at the time of HbA1c measurement. Multivariable RLRs modeled DNAm as the outcome on the M-value scale, and continuous HbA1c (%) as the exposure. Effect estimates were reported using DNAm in the β-value scale to ease interpretability. In a sensitivity analysis, we used only samples of non-Hispanic white origin (n = 291) to investigate potential ethnic-specific effects at CpGs previously discovered. We considered associations to be replicated at p < 0.05/#CpGs tested.

Using MR to infer causality in the association between maternal HbA1c & cord blood DNAm

Considering our observational associations may be biased by unknown confounders, we implemented a MR approach to test if higher maternal glycemia reflected by HbA1c levels may be causally implicated in epigenetic changes in cord blood DNAm. Briefly, MR uses genetic variants as instrumental variables (IVs) to proxy levels of the exposure (HbA1c), and to provide a causal estimate of the exposure-outcome association unaffected by reverse causation and unmeasured confounding [26,45,46]. Strong IVs for MR analysis are selected based on a prespecified F-statistic threshold >10 [45]. In this study, genetic variants used to construct a polygenic score were extracted from a recent transethnic GWAS meta-analysis of HbA1c measured outside pregnancy (n = 60 SNPs at p < 5.0 × 10^-8) [47]. We built the score for each participant by adding the number of HbA1c raising alleles for every SNP in the score, weighted by the GWAS effect for each SNP. To fulfill with specific assumptions of MR studies [26,45,46], we conducted several unadjusted linear and logistic regressions between the score as the exposure, and HbA1c, top differentially methylated CpGs and potential confounders of the main association as the outcome. To ease interpretability, we used the score in z-values (mean = 0, SD = 1), and association estimates were interpreted as the effect of one SD increase in the score, on a unit change in the trait. In total, we extracted information for 57 out of 60 HbA1c SNPs (three other SNPs were monomorphic in our sample) from 384 mothers in Gen3G with the availability of genetic, phenotype, and offspring DNAm data. All SNPs included in the score met our QC thresholds including MAF >0.01, genotype calling rate >0.8, r² <0.5 and correct allele assignment.

We conducted the MR using a two stage least square instrumental variable (2SLS-IV) analysis implemented in the AER R package [48]. We applied correction for multiple testing in MR results using the Bonferroni method (p < 0.05/#CpGs tested). Results were interpreted as the mean percent difference in DNAm β-values, per 1%-unit increase in genetically predicted values of maternal HbA1c. Additionally, we reported MR diagnostic tests to indicate strength of the IV (Weak instrument test), model fit (Wald test) and consistency in magnitude and direction of associations between the observational and causal estimates (Endogeneity test). Power of the MR was calculated using the online MR power calculator mRnd (http://cnsgenomics.com/shiny/mRnd/) [49]. For CpGs analyzed, we visually inspected consistency in effect-size and direction of effect between the MR and EWAS estimates using linear plots.

Results

Participant characteristics

All mothers in Gen3G included in these analyses were of European descent, the mean age was 28 years, 34% were primigravid and 92% were non smokers (Table 1). We observed very similar baseline characteristics between study participants (n = 412) and individuals in Gen3G not included in the DNAm analysis (n = 456 mother–child pairs without cord blood DNAm or relevant covariates), except for differences in the proportion of non-white participants. Based on the fact that we excluded women with GDM from our current analyses, we observed higher glucose values in the OGTT (1 & 2 h post glucose load) and more frequent GDM among excluded participants (Supplementary Table 1). In our study sample, levels of HbA1c were slightly higher in the first compared with the second trimester (mean difference = 0.3%; p < 0.001), consistent with previous literature showing that in normal pregnancies, levels of HbA1c gradually decrease from the first to the second trimester, when they reach their lowest values in pregnancy [9,12]. Maternal HbA1c was normally distributed in the two visits, mean ± SD for HbA1c in visit one was 5.2 ± 0.2 and for HbA1c in visit two was 4.9 ± 0.3. For this study, we included only term deliveries (i.e., gestational age at delivery >37 weeks); 53% of the offspring were males.

Table 1. . Characteristics of participants in genetics of glucose regulation in gestation and growth included in the cord blood DNA methylation analysis of maternal HbA1c.

Variable	n = 412 Mean (SD) or n [%]
Maternal age (years)	28.1 (4.1)
Gravidity [% primigravid]	138.0 [33.5]
Ethnicity [% Caucasian]	412 [100]
Smoking during pregnancy [% non smoker]	378 [91.7]
First trimester
Gestational age (weeks)	9.8 (2.3)
BMI (kg/m2)	24.6 (5.6)
1 h glucose challenge 50 g (mmol/l)†	5.4 (1.3)
Hemoglobin A1c (%)†	5.2 (0.2)
Second trimester
Gestational age (weeks)	26.4 (1.0)
Body mass index (kg/m2)	27.8 (5.3)
Glucose challenge 75 g
Fasting glucose (mmol/l)	4.2 (0.31)
1 h glucose (mmol/l)	6.9 (1.4)
2 h glucose (mmol/l)	5.6 (1.1)
Hemoglobin A1c (%)	4.9 (0.3)
Offspring variables
Gestational age at delivery (weeks)	39.6 (1.1)
Child sex [% males]	218 [52.9]
Birth weight (g)	3437.2 (431.1)

^†Smaller sample size was observed for measures of 1 h, 50 g glucose test (n = 384) and maternal HbA1c in first visit (n = 410).

Association between maternal HbA1c & cord blood DNAm (discovery EWAS)

In our CpG-by-CpG analysis adjusted for covariates (λ = 1.18; Supplementary Figure 2), we found one CpG site strongly associated with second trimester HbA1c after Bonferroni correction (p < 6.3 × 10^-8) (Figure 1 & Supplementary Figure 2). At cg21645848, we observed that a 1%-unit increase in maternal HbA1c was associated with an estimated 1.7% lower cord blood DNAm (95% CI: -2.2, -1.2%, p = 3.9 × 10^-11) (Figure 2 & Table 2). This CpG mapped within the first exon of the upregulator of cell proliferation or URGCP gene, located within a CpG island but showing a weak correlation in DNAm β-values with surrounding CpGs (Supplementary Figure 3).

Figure 1. . Manhattan plot for the epigenome-wide association study of second trimester maternal HbA1c and cord blood DNA methylation.

Horizontal red line is the Bonferroni significant threshold at p < 6.3 × 10^-8 or -log10(p) > 7.3; blue line is a borderline significance threshold set at p ≤ 10^-7 or -log10(p) > 6.0.

Figure 2. . Scatter plots for the correlation between second trimester maternal HbA1c and cord blood DNAm at the top hit identified in cg21645848 (URGCP), and at three other CpGs identified with suggestive association in the epigenome-wide association study.

Table 2. . Top differentially methylated CpGs in cord blood identified in the epigenome-wide association study of second trimester maternal HbA1c and inspected for their association with first trimester maternal HbA1c.

CpG	DNAm (%), mean (SD)	Mapped gene	Genomic context	Maternal HbA1c second trimester (n = 412)			Maternal HbA1c first trimester (n = 410)
				Effect estimates†	95% CI	p-value	Effect estimates‡	95% CI	p-value
cg21645848	6.48 (1.64)	URGCP	Body	-1.7	(-2.2, -1.2)	3.97 × 10-11	-0.7	(-1.2, -0.1)	1.22 × 10-2
cg21825879	2.40 (0.60)	FAT3	Intergenic	-0.5	(-0.7, -0.3)	1.83 × 10-7	-0.1	(-0.4, 0.1)	2.96 × 10-1
cg21990700	41.67 (5.98)	C1RL	TSS200, Body	-4.5	(-6.2, -2.8)	2.97 × 10-7	-2.9	(-4.9, -0.9)	3.72 × 10-3
cg14332918	94.0 (1.36)	PICK1	TSS1500	-0.9	(-1.3, -0.6)	1.56 × 10-7	-0.6	(-0.9, -0.2)	4.59 × 10-3

Highlighted in bold is the association at cg21645848 (URGCP) surpassing the Bonferroni significance threshold at p < 6.3 × 10^-8. Associations with first trimester HbA1c were regarded statistically significant at p < 0.01.

^†Effect estimates interpreted as the mean percent difference in DNAm (β-values) per 1%-unit increase in second (or first) trimester maternal HbA1c.

^‡Regressions adjusted for similar covariates as in the discovery analysis with second trimester HbA1c, except for gestational age and BMI that corresponded to measures in the first visit.

In addition, results of our main analysis suggested associations of second trimester HbA1c with three other CpGs (with p < 1.0 × 10^-6) at cg21825879 (near FAT3), cg21990700 (C1RL) and cg14332918 (PICK1). Association estimates for these CpGs are reported in Table 2. In general, we observed small effect sizes in the associations with maternal HbA1c (second trimester); among top four differentially methylated CpGs, median estimated percent difference in DNAm was 1.3 (estimated percent change in DNAm ranged from 0.5 to 4.5) in response to a 1%-unit increase in maternal HbA1c.

We found modest correlation between maternal BMI and HbA1c in the second visit (r = 0.2, p = 1.5 × 10^-5). When excluding BMI from our regression models, we observed that the association at cg21645848 (URGCP) showed similar effect estimates and statistical significance as in the main model (p = 5.69 × 10^-11 in BMI-unadjusted model, and p = 3.97 × 10^-11 in main model). No other CpGs surpassed the threshold for Bonferroni significance in this sensitivity analysis. Overall, for all four differentially methylated CpGs previously identified in our initial analysis, excluding BMI from the models changed the absolute magnitude of the effect estimates by less than 10% (range = 3.0–8.8% decrease).

In sensitivity analyses, we used quartiles of HbA1c (second visit) as the exposure to investigate the assumption of a non linear relationship between HbA1c and cord blood DNAm. We found that an increase in HbA1c levels from Q1 (mean ± SD = 4.7 ± 0.1) to Q4 (mean ± SD = 5.3 ± 0.1) was associated with lower DNAm at CpGs in cg21645848 (URGCP), cg22322019 (intergenic) and cg04289992 (PALM) after correction for multiple testing (Supplementary Table 2). Thus, we confirmed our top hit in cg21645848 (URGCP) and found two other CpG sites that were not identified in our primary analysis, yet we recognized that this sensitivity EWAS showed greater inflation (λ = 1.4). Therefore, we remained prudent in the interpretation of these findings.

In multivariable linear regressions with predicted cord blood cells as the outcomes, we found no association between cell proportions and maternal HbA1c (p > 0.01), but specific cell types were associated with DNAm levels at cg21645848 (URGCP) (associated with CD8T, NK, Mono & Gran), cg21990700 (C1RL) (associated with CD4T, CD8T, B cells & Gran) and cg14332918 (PICK1) (associated with CD4T, Gran & nucleated red blood cell) at p < 0.002. In addition, none of the top four differentially methylated CpGs identified in the EWAS were associated with the inflammatory marker TNF-α measured in cord blood (p > 0.01).

Additionally, we investigated specificity of our top differentially methylated CpGs identified in the EWAS to timing of glycemic exposure by measuring their association with first trimester HbA1c. We observed that per 1%-unit increase in first trimester HbA1c was associated with lower DNAm in cg21645848 (URGCP), cg21990700 (C1RL) and cg14332918 (PICK1) (p < 0.01), consistent in the direction of association with the initial analysis using second trimester HbA1c, but effect sizes were smaller than for associations in the discovery analysis (Table 2).

For the EWAS conducted with first trimester maternal HbA1c (λ = 1.11), we did not identify any CpG sites where cord blood DNAm was robustly associated with HbA1c after adjusting for covariates and applying correction for multiple testing using Bonferroni (Supplementary Figure 4), but we found three CpGs with association p < 1.0 × 10^-6 (Supplementary Table 3). These three CpGs did not overlap with top associations identified by p-value in our main EWAS of second trimester HbA1c. We also assessed the impact of adjusting (or not) for maternal BMI in this EWAS. We did not find correlation between BMI and HbA1c in the first visit (r = 0.1, p = 0.18), and in regressions excluding adjustment for BMI we obtained similar results at top three CpGs identified in the main model. In adjusted regressions between quartiles of HbA1c (first visit) and genome-wide cord blood DNAm (λ = 1.12), we did not find evidence of a non linear association between a change in HbA1c levels from Q1 (mean ± SD = 4.9 ± 0.1) to Q4 (mean ± SD = 5.5 ± 0.1), and difference in DNAm.

Association of OGTT glucose levels with cord blood DNAm at CpGs identified in EWAS of second trimester HbA1c

For glucose values measured during the second trimester OGTT, we found that greater maternal fasting glucose was nominally associated (at p < 0.05) with lower DNAm at previously discovered CpGs in cg21825879 (FAT3) and cg21990700 (C1RL) (p < 0.05) (Table 3). Associations between fasting or 1 h post-load maternal glucose and DNAm at cg21645848 (URGCP), cg21825879 (FAT3), cg21990700 (C1RL) and cg14332918 (PICK1), were in the same direction as in the discovery analysis with second trimester HbA1c, but only suggestive of associations (no significant p-values). Only cg21645848 (URGCP) and cg21990700 (C1RL) showed association in the same expected direction of effect with 2 h glucose values exposure, yet also non significant. The effect estimates identified in association with standardized glucose values measured during the OGTT were smaller than for associations with HbA1c in the second trimester at the four CpGs previously identified. All the glucose values standardized before the analysis were normally distributed. The correlations between second trimester HbA1c levels and OGTT glucose measures were weak (r range 0.2–0.3), but statistically significant (p < 0.001).

Table 3. . Association of maternal glucose values during a second trimester oral glucose tolerance test with top four CpG sites identified in the epigenome-wide association study of second trimester maternal HbA1c (all z-scores).

CpG	Gene	Fasting glucose z-score (n = 412)		1 h glucose z-score (n = 412)		2 h glucose z-score (n = 412)		HbA1c second trimester z-score (n = 412)
		Effect estimates (95% CI)	p-value	Effect estimates (95% CI)	p-value	Effect estimates (95% CI)	p-value	Effect estimates (95% CI)	p-value
cg21645848	URGCP	-0.09 (-0.23, 0.05)	0.24	-0.06 (-0.20, 0.09)	0.47	-0.13 (-0.28, 0.03)	0.10	-0.43 (-0.55, -0.30)	3.97 × 10-11
cg21825879	FAT3	-0.05 (-0.11, -0.0001)	0.04	-0.001 (-0.05, 0.05)	0.83	0.003 (-0.04, 0.05)	0.99	-0.12 (-0.17, -0.07)	1.83 × 10-7
cg21990700	C1RL	-0.56 (-1.06, -0.07)	0.03	-0.40 (-0.91, 0.12)	0.13	-0.34 (-0.83, 0.15)	0.18	-1.13 (-1.56, -0.7)	2.97 × 10-7
cg14332918	PICK1	-0.08 (-0.18, 0.02)	0.10	-0.03 (-0.13, 0.08)	0.44	0.002 (-0.10, 0.10)	0.98	-0.23 (-0.32, -0.14)	1.56 × 10-7

Measures of glycemia were transformed to z-scores (mean = 0 and SD = 1) before the analysis to allow comparability in the effect estimates among traits. CpGs analyzed were identified in the epigenome-wide association study of second trimester maternal HbA1c (n = 412). Highlighted in bold are associations identified with p < 0.05 with one measure of glycemia.

Replication in the Healthy Start cohort

In the multi-ethnic external replication cohort (n = 506), we sought to replicate top four CpGs identified with p < 1.0 × 10^-6 in the discovery analysis of second trimester HbA1c. Study characteristics of the replication cohort are presented in Supplementary Table 4. Of the four CpGs analyzed in the Healthy Start cohort, we identified a consistent direction of association with Gen3G analyses only at the top hit previously discovered in cg21645848 (URGCP) (effect estimate in Healthy Start = 0.84% lower DNAm β-values per 1%-unit increase in maternal HbA1c), but this association was not statistically significant (p = 0.09) (Supplementary Table 5). To address potential confounding by ethnicity at the CpGs investigated, we repeated the analysis using only samples of non-Hispanic white origin in the independent cohort (n = 291), but results were similar to those identified in the complete sample (data not shown).

MR analyses to support causality of associations between maternal HbA1c & cord blood DNAm

In an additional analysis, we evaluated the possibility that CpGs identified in our fully adjusted model were influenced by unmeasured confounders. For this analysis, we used a polygenic score (PS) with 57 SNPs as an IV for maternal HbA1c. In unadjusted regressions, we found that each SD increase in the PS was associated with a 0.06%-unit increase (95% CI: 0.03, 0.08%-unit) in maternal HbA1c in the second trimester, and this association was robust by statistical testing (p = 2.47 × 10^-6, n = 384). Distribution of the PS in the Gen3G sample and its correlation with measured maternal HbA1c is illustrated in Supplementary Figure 5. We did not identify a substantial change in the association of the PS with measured maternal HbA1c after including maternal age and the first five genetic PCs as covariates in the regression model (Supplementary Table 6). In the raw model, the PS explained 5.4% (r²) of the variation in maternal HbA1c in our sample, and this result is consistent with previous findings in Europeans [47]. The F-statistic of the regression model including the PS was 22.9, indicating that the genetic score was a strong IV for maternal HbA1c. In additional analyses, we demonstrated that the PS was independent of measured confounders of the exposure-outcome association (p > 4.0 × 10^-3 or α = 0.05/14 traits tested), and the PS was not directly associated with top four differentially methylated CpGs identified in the EWAS of HbA1c (p > 0.01 or α = 0.05/4 CpGs) (data not shown).

Our MR analyses demonstrated that the predicted effects were all in line with the observed effect in size and direction (Supplementary Table 7). Even though the effect size predicted was very similar to estimates obtained in the adjusted EWAS model, none of the MR estimates reached statistical significance (p > 0.01). Figure 3 shows a linear plot comparing effect-sizes and direction of association between predicted estimates of the MR and the observational analysis. We suspected that our MR analyses were underpowered to detect a causal effect of maternal HbA1c on difference in cord blood DNAm at top four CpGs identified in the EWAS. This assumption was supported by the results of a post hoc power calculation [49] that revealed that our analysis had 12% power to detect a 1.9% causal effect of maternal HbA1c on lower DNAm at cg21645848 (URGCP), our strongest signal of the EWAS, with statistical significance (α = 0.01). Furthermore, the endogeneity test (Hausman test) was underpowered to detect confounding in estimates of the observational analysis (p > 0.05). Genetically predicted values of maternal HbA1c explained between 2% and 5% of the total variation in DNAm at top four CpGs identified in the EWAS.

Figure 3. . Linear plot comparing association estimates between the observational and the Mendelian randomization analysis (2SLS-IV) for top four CpG sites identified in association with second trimester maternal HbA1c in the epigenome-wide association study.

X-axis represents estimates of the fully adjusted EWAS using measured maternal HbA1c as the exposure. Y-axis represents estimates of the MR analysis (unadjusted) using genetically predicted values of HbA1c as the exposure. Vertical error-bars represent the 95% CI for the MR estimates, while horizontal error-bars are the 95% CI for the observational estimates. CIs of the MR estimates overlapped with the red-dashed line of null associations, indicating no evidence of causality at these CpGs.

DNAm: DNA methylation; EWAS: Epigenome-wide association study; MR: Mendelian randomization.

Discussion

In this study we report, to our knowledge, the first EWAS of cord blood DNAm in relation to maternal hyperglycemic exposure reflected by HbA1c levels in the second trimester. We found one CpG in URGCP at which DNAm levels were strongly inversely associated with maternal HbA1c. The direction of association at URGCP was consistent with results we found in our replication analyses in an independent cohort. Three other epigenomic loci mapping to the genes FAT3, C1RL and PICK1 were suggestively associated with second trimester HbA1c in our discovery EWAS at a less stringent statistically significant threshold, but did not replicate. Our MR analyses predicted effects that were consistent with the observed effects in size and direction. However, we did not find statistical support for a causal effect of second trimester maternal HbA1c on cord blood DNAm at loci identified in the EWAS, likely due to limited power.

Cord blood DNAm in association with second trimester HbA1c

We demonstrated that maternal HbA1c in mid-pregnancy (as continuous or categorical) was associated with lower DNAm in cord blood at cg21645848 in URGCP. DNAm at this CpG was also associated with specific cord blood cells, indicating that HbA1c association with DNAm at cg21645848 may still reflect differences in the proportion of cells, despite our models being adjusted for cell types. URGCP or URG4 is the upregulator of cell proliferation, a recently discovered gene induced by hepatitis-B virus-encoded X-antigen (HbxAg) that has been identified as a putative oncogene involved in multistep carcinogenesis, and in various cellular processes [50]. In hepatocellular carcinogenic cells, overexpression of URGCP/URG4 correlates with upregulation of target genes for the NF-κB signaling pathway, including TNF-α, IL-6, IL-8 and MYC proto-oncogene [51].TNF-α is a proinflammatory cytokine that reduces insulin sensitivity by facilitating the phosphorylated state of the insulin receptor [52]. In our cohort, maternal TNF-α levels in pregnancy have been previously found independently associated with a higher maternal insulin resistance [53]. Similarly, our study previously identified an association between elevated 2-h glucose exposure in mid-pregnancy and increased placental methylation of a CpG mapping to the TNF receptor superfamily member 1b gene (TNFRSF1B), a high affinity receptor for TNF-α [21]. In non pregnant populations, overexpression of TNF-α in peripheral blood mononuclear cells was characteristic of patients with T2D [54], and expression of TNF was strongly correlated with HbA1c levels in the same study [54]. We demonstrated that cg21645848 was associated with cell types, but not with TNF-α in cord blood. Thus, our results provide some evidence that lower DNAm of cg21645848 (URGCP) in response to maternal hyperglycemia may be related with proinflammatory markers, yet further studies are needed to elucidate exact pathways and cytokines involved.

The CpG cg21645848 (URGCP) has not been reported before in EWAS of glucose traits [55] or T2D [56,57] in peripheral blood DNAm of middle-age adults, or in EWAS of GDM in fetal tissues (cord blood or placenta) [19,24,58]. However, we demonstrated the same direction of effect of HbA1c on cg21645848 (URGCP) in our replication study, and we showed a robust association of cg21645848 with first trimester HbA1c, and a consistent direction of association with other glycemic measures. Using a publicly available dataset, we found a methylation quantitative trait loci (meQTL) associated with DNAm at cg21645848 in cord blood according to data from the mQTL database (www.mqtldb.org/) [59]. This finding indicates that methylation at cg21645848 may be partially under fetal genetic control.

Lower DNAm at cg21990700 (C1RL), cg21825879 (FAT3) and cg14332918 (PICK1), showed some association with increased mid-pregnancy HbA1c in our discovery EWAS, but none of these signals were found with consistent effect size and direction of association in our replication study. Of these epigenetic loci, the largest effect size was observed at cg21990700 in the C1RL gene. In addition to second trimester HbA1c, cg21990700 was also associated with first trimester HbA1c, cord blood cell types, and there was a trend towards association with fasting glucose (second trimester). Previous EWAS have shown associations of cg21990700 with differential DNAm of fetal versus adult liver tissue [60], and with CRP [61], age [62] and smoking [63] in peripheral blood DNAm. The C1RL protein mediates the proteolytic cleavage of haptoglobins (cellular secretory proteins) in the endoplasmic reticulum [64]. C1RL is also a member of the complement system, which includes many proteins relevant to the innate immune response, and to the regulation of inflammation in adipose tissue [65]. Upregulation of C1RL expression was previously found in association with acquired obesity in subcutaneous adipose tissue of BMI-discordant twin pairs (lean vs heavy) [65]. In addition, C1RL expression was positively correlated with adiposity and hyperinsulinemia, and negatively correlated with the expression of insulin-signaling related genes in the same study [65].

Cord blood DNAm in association with 1st trimester HbA1c

Early pregnancy is a critical period for the effect of adverse exposures on embryonic and placental development [66]. In early pregnancy, higher maternal glucose levels impair the development of the placenta and restrict fetal growth [67]. Some studies have reported an association of higher maternal HbA1c levels in early pregnancy with a lower rate of fetal growth in mid-pregnancy in the offspring of non-GDM mothers [68], and with an increased risk of adverse birth outcomes independent of GDM diagnosis later in pregnancy [69]. Because first trimester HbA1c covers hyperglycemic exposure during the peri-conception and early pregnancy period, we initially hypothesized that first trimester HbA1c will have more profound programming effects on DNAm at birth. However, our results do not support that first trimester maternal HbA1c levels were strongly associated with changes in cord blood DNAm. Several factors can explain this result. First, HbA1c levels in later pregnancy may be a better reflection of the exposure to hyperglycemic episodes, while first trimester HbA1c levels may be influenced to a greater extent by erythrocyte cells biology (faster turnover). Second, large-scale DNAm remodeling in fetal tissues occurs primarily during mid-pregnancy (from week 9th to 22nd) [70], with a gain of DNAm in genes related to developmental functions, but lowering DNAm in genes involved in tissue-specific functions [70]. Finally, cord blood DNAm may not reflect epigenetic changes that may have occurred in more target tissues early in the embryonic development.

Findings from MR analyses

We attempted to support causality in the association between maternal HbA1c in pregnancy and cord blood DNAm by implementing MR. We observed a strong consistency in results of the MR and the observational analysis with respect to effect sizes and direction of effect for the associations analyzed. However, findings from the MR were underpowered, even though an F-statistic >10 from the IV-exposure association confirmed the strength of our instrument to proxy HbA1c levels in our sample. We estimated that more than 2361 mother–child pairs would have been required in our study to detect a similar causal effect at our top signal of the EWAS in cg21645848 (URGCP), with statistical significance (α = 0.01) and 80% power.

Putting our findings in the context of other epigenetic studies of glycemic traits in pregnancy

We did not find overlap between our cord blood CpGs associated with maternal HbA1c, and CpGs previously reported for GDM in cord blood [19,71,72] and placenta tissues [19,72]. Also, these CpGs were not overlapping with those previously identified in the placenta in association with maternal 2 h glucose [21] and insulin sensitivity (Matsuda index) [73] in our cohort.

Strengths & limitations

Our study has both strengths and limitations. Our strengths include our relatively large sample size, our prospective study design and the use of the EPIC array that allowed us to inspect >790,000 CpGs across the genome. We controlled for many potential confounders, including cord blood cells. In addition, we tested for epigenome-wide associations with maternal HbA1c measured at two time points in pregnancy. Finally, we attempted to validate results in an independent study.

We also had limitations. First, we acknowledge that cord blood is not considered as a tissue with ‘direct’ metabolic relevance (i.e., adipose tissue, muscle, liver, pancreas) to reveal DNAm involved in metabolic programming in response to maternal hyperglycemia. Second, DNAm at some of the CpGs identified showed associations with cells proportions even though we adjusted for cellular heterogeneity in our main analysis, suggesting incomplete correction for cells types or reflection of underlying inflammation. Third, despite including multiple potential confounders in our main analysis, unmeasured confounders may still have biased the associations identified. Finally lack of power due to the small sample-size used in the MR analysis, limited our detection of a causal effect between maternal HbA1c in mid-pregnancy and DNAm at CpGs identified in the EWAS.

Conclusion

We found that exposure to higher maternal glycemia in mid-pregnancy reflected by HbA1c was associated with lower cord blood DNAm at cg21645848 in URGCP. The URGCP gene is implicated in cell survival and growth, and in inflammatory response. These functions relate to genes previously identified with differential DNAm in response to GDM and HbA1c in fetal and adult tissues, respectively. Future research should aim to confirm the association between maternal glycemia in pregnancy and cord blood DNAm at the CpG in URGCP in European and non-European populations, while functional and longitudinal analyses could help to elucidate whether this locus has a role in gene regulation in specific tissues, and in long-term offspring metabolic programming.

Future perspective

The identification and validation of DNAm signatures in cord blood associated with exposure to maternal hyperglycemia in pregnancy may help to identify candidate biomarkers for early intervention and prevention of adverse cardiometabolic outcomes in the offspring. These biomarkers will provide insights into the mechanisms dysregulated by exposure to hyperglycemia early during development. Future validation of these markers can be achieved through target-tissue studies, longitudinal studies and the use of well-powered causal inference methods.

Summary points.

• We identified a novel signal at cg21645848 in URGCP at which cord blood DNA methylation levels were inversely associated with maternal HbA1c levels in mid-pregnancy.

• We found consistency in the direction of association at URGCP in our replication study.

• We identified three other epigenetic loci mapping to C1RL, FAT3 and PICK1 with ‘suggestive’ association with mid-pregnancy maternal HbA1c levels, although none were replicated.

• Our Mendelian randomization analysis showed estimates consistent in direction of effect and in effect size with observational estimates for all four loci tested. However, the Mendelian randomization was underpowered and did not support a causal effect of maternal HbA1c on cord blood DNA methylation levels.

• We showed consistent direction of association with first trimester HbA1c, and with some of the glucose measures during the oral glucose tolerance test, for our top four CpGs identified by p-value in the epigenome-wide association study of second trimester maternal HbA1c.

• We hypothesized that for some of the epigenetic loci identified in association with mid-pregnancy maternal HbA1c levels (i.e., URGCP & C1RL), adverse metabolic programming in the offspring may involve inflammatory pathways.