Introduction

Unbalanced nutrition in utero, such as hyperglycaemia, has been associated with the development of metabolic diseases in later life [1]. In particular, intrauterine exposure to maternal diabetes increases the risk that the child will subsequently develop obesity and diabetes [2]. Among Pima Indian siblings discordant for intrauterine exposure to diabetes, the risk was higher in those born after their mother developed diabetes than in those born before their mother developed diabetes, providing evidence that the increased risk is due to an effect of the diabetic intrauterine environment in addition to any effect conferred by inheritance of diabetes susceptibility alleles [3]. Studies in Pima Indians with normal glucose tolerance (NGT) have also shown that insulin secretion is lower in young adults whose mothers were diabetic during pregnancy compared with those whose mothers developed diabetes at an early age but after the birth of the child, suggesting that exposure to a diabetic intrauterine environment increases the risk of diabetes, in part because of decreased insulin secretion [4]. This reduction in insulin secretion may be mediated by epigenetic modifications during key developmental periods of the fetus [5].

One source of epigenetic modification is DNA methylation at CpG sites. In some cases, methylation at a specific site is stable; thus, assessment after birth can provide information on methylation patterns established in utero [6]. Two recent studies have reported epigenome-wide analyses that identified differentially methylated regions as a consequence of intrauterine exposure to maternal diabetes [7, 8]. One of these studies was our prior examination of promoter regions in DNA from 28 Pima Indians which did not identify any specific differentially methylated promoter that met statistical significance [7]. A third study which examined methylation of GNAS and IGF2 in 168 Chinese newborns found an association between maternal diabetes and methylation of GNAS, in which methylation levels were also associated with obesity [9]. However, these three studies had limited statistical power due to their small sample sizes. In the current study, we sought to understand the effect of maternal diabetes on DNA methylation in 388 Pima Indian offspring, which to our knowledge is the largest epigenome-wide report to date. These individuals were, on average, 13 years old when examined, so the current study sought to determine methylation changes due to exposure to diabetes in utero that were stable over many years. The Pima Indian population is at high risk of type 2 diabetes, but type 1 diabetes does not occur [10].

Methods

Study participants

Participants in the present study were a subset of individuals from a longitudinal study among residents of an American Indian community in Arizona who are predominately Pima Indian [2]. Biennial health examinations, beginning at age ≥5 years, included a 75 g OGTT; in a substudy of pregnant women, a 75 g OGTT was also offered in the third trimester. In the current study, data for the OGTT during pregnancy were obtained either at a biennial examination or at a third trimester examination. Diabetes was defined by World Health Organization criteria for non-pregnant individuals, regardless of whether a woman was pregnant, i.e. 2 h post-load plasma glucose ≥11.1 mmol/l, fasting plasma glucose ≥7.0 mmol/l or a diagnosis observed during clinical care [11]. DNA analysed in the current study was isolated from peripheral blood collected at the offspring’s first examination at which they were not diabetic. All available offspring of mothers with diabetes during pregnancy (OMD), defined as having a mother diagnosed with type 2 diabetes before the child’s birth, were selected for methylation studies (N = 187). Of these, 169 (90%) were offspring of a mother diagnosed more than 9 months before the child’s birth. Of the remaining 18 (10%) whose mother’s diabetes diagnosis was less than 9 months before the child’s birth, six were born to a mother with no previous examination, four were born to a mother with a non-diabetic examination <1 year before the child’s birth and eight were born to a mother whose last non-diabetic examination was ≥1 year before the child’s birth. Thus, the current study largely examines the effect on the intrauterine environment of type 2 diabetes that predates the pregnancy. For comparison, DNA samples (N = 201) were randomly selected from offspring of mothers with no diabetes during the pregnancy (OMND), where the mother did not have diabetes diagnosed before the child’s birth and the mother had a non-diabetic OGTT (2 h post-load plasma glucose <11.1 mmol/l) in the longitudinal study ≥12 months after the child’s birth. OMND were frequency-matched to OMD by sex, age (5 year strata) and self-reported Pima heritage (full vs partial), and all were required to have at least 50% American Indian heritage by self-report. All studies were approved by the institutional review board of the National Institute of Diabetes and Digestive and Kidney Diseases, and all participants provided written informed consent.

Characteristics of the 388 offspring analysed in the study are provided in Table 1. The mean age of the offspring when their DNA was obtained was 13.0 (range 5.2–41.6) years. The OMD group, on average, had a higher maternal age and BMI, and was more likely to have a sibling included in the analyses, compared with the OMND group (Table 1).

Table 1 Characteristics of participants
Full size table

Statistical analysis to identify differential methylation

DNA methylation at CpG sites was measured using the Infinium HumanMethylation450K BeadChip (Illumina, San Diego, CA, USA; methods detailed in electronic supplementary material [ESM] Methods). To stabilise the variance for purposes of analyses, Beta-values (the ratio of methylated probe intensity and overall intensity [the sum of methylated and unmethylated probe intensities]) were converted to M-values (the log2 ratio of methylated probe intensity vs unmethylated probe intensity) and standardised to a mean of 0 and an SD of 1 [12]. The standardised M-value was adjusted by linear regression for age, sex, a genetic estimate of American Indian ancestry in the offspring, mother’s age when the child was born, array batch number, and the first four principal components (PCs) to account for cellular heterogeneity or other unmeasured confounders [13]. In these analyses, American Indian ancestry was estimated according to the method of Hanis et al [14] from 45 ancestry-informative markers [15]. The PCs were calculated using the R software package pcaMethods (R Foundation for Statistical Computing, Vienna, Austria), and the methylation Beta-values from 47,340 probes, with one probe selected from each gene and CpG island. The residuals of the standardised M-values obtained from this regression were tested for association with intrauterine exposure to diabetes using Procedure GLM in SAS, version 9.3 (SAS Institute, Cary, NC, USA), with M-value as the dependent variable. To estimate the effect size in biologically meaningful terms, the same model was analysed with the original Beta-value as the dependent variable. In Table 2, effect sizes are shown for the model with Beta-value as the dependent variable, while p values are reported for the model with M-value as the dependent variable (which generally conforms better to the assumptions of linear regression).

Table 2 Forty-eight differentially methylated CpG sites in OMD vs OMND with epigenome-wide significance
Full size table

The empirical false discovery rate (FDR) was estimated using a nested permutation procedure to generate the null distribution. In brief, the methylation data (M-values, PCs) were permuted independently from the phenotypic data (exposure to maternal diabetes status, age, sex, maternal age and ancestry estimate). To account for familial dependence, all phenotypic data for a sibship were permuted together; thus, permutations were conducted across sibships of the same size. With 1000 permutations, the empirical FDR p values were calculated using the method proposed by Millstein and Volfson [16].

Differentially methylated regions were analysed to determine whether there was a general tendency for these CpGs to be hypo- or hypermethylated in OMD relative to OMND. The number of hypermethylated regions (effect >0) and hypomethylated regions (effect <0) were counted and subjected to a binomial test for the methylation proportion using the R function binom.test (R Foundation for Statistical Computing), in comparison with the expected percentage of hypermethylated regions of 50%.

Assessing differentially methylated sites for prediction of impaired secretory function, obesity and diabetes risk

Secretory function could be assessed in 230 of the 388 offspring who had NGT (2 h post-load plasma glucose <7.8 mmol/l) at baseline. The 2 h corrected insulin response (CIR), a surrogate measure of insulin secretion, was derived from OGTT measurements at the time of the DNA sample collection [17]. Fourteen of these NGT offspring had further been assessed for their acute insulin response (AIR) to a 25 g intravenous glucose challenge at a later time in their life [4]. The association between the insulin secretory function (defined as standardised logarithm of CIR or AIR) and standardised M-value residuals for individual CpG sites was calculated with adjustments as detailed in the Table 3 legend.

Table 3 The association of the 48 differentially methylated CpG sites with CIR, BMI and diabetes risk in later life
Full size table

BMI could be assessed in 376 of the 388 offspring who had both height and weight measured at baseline, while risk of type 2 diabetes could be assessed in 303 offspring who had data from at least one outpatient follow-up examination (58 out of 303 had documented diabetes at follow-up). The association between the logarithm of BMI and standardised M-value, as well as the association between diabetes risk and standardised M-value, was calculated with adjustments as detailed in the Table 3 legend. A Cox proportional hazards model (Procedure PHREG in SAS, version 9.3; SAS Institute) was used for the analysis of diabetes risk. To assess the extent to which observed methylation differences may account for the increased diabetes risk in OMD, a formal mediation analysis was conducted as detailed in ESM Text [18, 19].

Pathway enrichment analysis

Genes with differentially methylated CpG sites (p ≤ 0.005) were analysed for pathway enrichment (KEGG pathways [20]; release date 21 March 2011; access date 4 November 2015) compared with Homo sapiens genetic background using the WEB-based GEne SeT AnaLysis Toolkit (WebGestalt) [21]. The enrichment p values were adjusted for multiple tests by the FDR method. Pathways with FDR p ≤ 0.001 and more than three differentially methylated genes were considered enriched.

Results

Thirty-nine unique differentially methylated regions are associated with intrauterine diabetes exposure

Methylation data from 423,311 CpG sites passed our quality control metrics. When analysed in aggregate, baseline DNA methylation mean Beta-values across all 423,311 sites did not differ (p = 0.70) between the OMD group (N = 187; Beta-value 0.513 [SD 0.006]) and the OMND group (N = 201; Beta-value 0.513 [SD 0.012]). Analysis of each individual CpG site in OMD compared with OMND is shown by chromosome in a Manhattan plot (Fig. 1a), and the corresponding quantile–quantile plot is also shown (Fig. 1b). Forty-eight CpGs representing 39 unique regions were significantly associated with intrauterine diabetes exposure (empirical FDR ≤0.05; Table 2). At these sites, the adjusted median absolute value of the difference in DNA methylation between OMD and OMND was 2.2% (range 0.6–4.6%). Among these 48 CpGs (17 in gene bodies or 3′ untranslated regions [UTRs], 20 in regions close to transcription start sites [TSSs] or 5′ UTRs, 11 in intergenic regions [with two CpGs being found in the same intergenic region]), 29 were hypermethylated and 19 were hypomethylated; the proportion of hypermethylated CpGs was not significantly greater than 0.5 (binominal p = 0.097). Since there was a large difference in pre-pregnancy maternal BMI between the OMD and OMND groups (Table 1), we further adjusted for pre-pregnancy BMI in the 296 individuals (154 OMD and 142 OMND) for whom maternal BMI data were available. The effect sizes were generally unchanged with the additional adjustment for maternal pre-pregnancy BMI (ESM Table 1).

Fig. 1
figure 1

The genome-wide differential methylation signals. (a) Manhattan plot of differentially methylated CpGs in OMND vs OMD. Genome-wide significant differences are highlighted in green. (b) Quantile–quantile plot of the differential signals. The red line is the identity line

Full size image

Methylation at some CpG sites which associate with intrauterine exposure also associate with impaired insulin secretory function, higher BMI or future risk of type 2 diabetes

We further assessed whether any of the 48 CpG sites that associated with exposure to intrauterine diabetes also associated with reduction in insulin secretory function, higher BMI or future type 2 diabetes. Three of the 48 sites were modestly associated with CIR (Table 3). The CpG site (cg24049468) located in the gene body of AK3 was hypermethylated in the OMD group, and hypermethylation at this site was associated with a reduced CIR in the 230 NGT participants. Similarly, CpG sites located in the gene body of PCDHGA4 (cg25949304) and close to the TSS of STC1 (cg04645534) were hypomethylated in the OMD group, and hypomethylation at these sites was associated with a reduced CIR in the 230 NGT participants. Methylation at these three CpG sites was further analysed in 14 individuals who remained NGT and had participated in a subsequent inpatient examination (median follow-up was 1.8 + 5.8 years) in our clinical research centre to characterise insulin secretory function via a 25 g IVGTT. In this very small sample, those who were hypomethylated at STC1 similarly tended to have a reduced AIR (effect 3.7% per SD; p = 0.058).

Among the 48 CpGs differentially methylated between OMD and OMND, 17 were associated with baseline BMI of the offspring (p < 0.05). However, only CpG sites in SBK1 (cg04413090, p = 4.1 × 10−4) and PRDM16 (cg12140144, p = 1.2 × 10−3) and two intergenic CpGs (cg10772621, p = 2.3 × 10−4; cg07464358, p = 2.3 × 10−4) remained significantly associated after correction for multiple tests (assuming 39 independent tests). Three of the CpG sites, cg04413090 (SBK1), cg12140144 (PRDM16) and cg07464358, were hypermethylated in OMD, and hypermethylation was associated with higher BMI (5.3%, 6.0% and 9.2% higher per SD of standardised M-value, respectively). Similarly, the CpG site cg10772621 was hypomethylated in OMD, and hypomethylation was associated with higher BMI (by 4.9% per SD of standardised M-value).

Among the 303 offspring with data from a follow-up examination, 43 OMD and 15 OMND developed diabetes later in life. The adjusted HR for developing diabetes for OMD compared with OMND was 5.48 (95% CI 2.85, 10.52; p = 3.2 × 10−7). Among these 303 individuals, methylation at cg12140144 in the TSS of PRDM16 was associated with a higher risk of diabetes (Table 3; HR 2.12 [95% CI 1.74, 2.50] per SD of M-value; p = 9.7 × 10−5). Another differentially methylated CpG site, cg08911291, which is located in an intergenic region, also carried diabetes risk with marginal significance (Table 3; HR 1.47 [95% CI 1.09, 1.85] per SD of M-value; p = 0.05). Among the 303 individuals with data on development of diabetes, the adjusted standardised M-value at the cg12140144 site in PRDM16 was 0.51 SD higher in OMD compared with OMND. After adjustment for methylation at this site, the HR for developing diabetes in OMD compared with OMND was 3.89 (95% CI 1.98, 7.65; p = 8.0 × 10−5). In mediation analyses, this corresponded to a percentage mediation of 20% (two-sided p = 0.026). The corresponding value for cg08911291 was 7% (p = 0.27).

Differentially methylated genes clustered in known pathways

The pathway analysis included 641 genes with a differentially methylated CpG site (p ≤ 0.005). The differentially methylated genes were enriched for: (1) metabolic pathways (37 genes, adjusted p = 0.0002); (2) Wnt signalling pathways (11 genes, adjusted p = 0.0004); and (3) protein digestion and absorption (eight genes, adjusted p = 0.0005) (ESM Table 2).

Database searches (www.ncbi.nlm.nih.gov/gene, accessed 1 September 2016) revealed that 11 of the 29 genes have known roles in embryonic development (ESM Table 3), and several genes have a physiological function consistent with a role in beta cell function/type 2 diabetes.

Discussion

We investigated genome-wide DNA methylation in peripheral blood in relation to intrauterine diabetes exposure. Compared with published studies of intrauterine exposure to maternal diabetes, our study featured a larger sample size. The Infinium (Illumina) methylation assay used in the current study covers 96% of CpG islands and 99% of reference sequence genes with an average of 17 CpG sites per gene. Forty-eight CpG sites were identified, with a statistically significant (FDR ≤0.05) difference in methylation among people exposed to maternal diabetes in utero compared with those not exposed. For differentially methylated sites, the median adjusted absolute difference in methylation between OMD and OMND was 2.2% (range 0.6–4.6%). Exposure to diabetes in utero is strongly associated with higher maternal BMI, and genetically raised maternal BMI is associated with birthweight, which in turn is associated with long-term developmental outcomes and adult diseases [22, 23]. However, the additional adjustment for maternal pre-pregnancy BMI did not materially change the effect sizes of our 48 top signals.

The strongest result in our study, in terms of statistical significance, is the differential methylation of the transcriptional activator LHX3. Six of the 23 CpG sites assayed on the array from this region were hypermethylated in OMD (p ≤ 1.1 × 10−5). Mutations in LHX3 have been associated with combined pituitary hormone deficiency (CPHD), which is often characterised by reduced growth and short stature [24]. Lhx3 expression is regulated by DNA methylation in mice, where decreased methylation has been shown to activate expression in pituitary cells that do not normally express Lhx3 [25]. In humans, LHX3 is broadly expressed in many tissues, including the pancreas. Although this particular transcription factor has not been previously reported to affect insulin secretory pathways, it does interact with several other transcription factors including insulin gene enhancer protein ISL1 and LIM homeobox transcription factor 1-alpha (LMX1A), which are known regulators of insulin gene expression [26].

Several of the genes identified as differentially methylated between OMD and OMND have established roles in beta cell function, including a CpG site in STC1 that was hypomethylated in the OMD group, where hypomethylation of the same CpG was also associated with both reduced CIR and AIR, as well as increased BMI at baseline. STC1 encodes the polypeptide hormone stanniocalcin-1 (STC-1), which co-localises with insulin in mouse pancreatic islets, and it has been proposed that STC-1 regulates beta cell mitochondria membrane potential, which is a crucial component in regulation of insulin release [27]. Another gene, CACNA1C, which was hypomethylated in the OMD group, has also been shown to be important for beta cell function [28]. However, hypomethylation at CACNA1C was not associated with reduction in insulin secretory response in the present data.

Two CpG sites that were significantly differentially methylated between OMD and OMND on an epigenome-wide basis also had nominally significant associations with diabetes risk in later life; however, only the CpG site in PRMD16 (cg12140144) remained associated after correction for multiple testing (assuming 39 independent differentially methylated regions). This CpG was hypermethylated in OMD by 1.1% on average, and a 1 SD increase in methylation was associated with a doubled diabetes risk. Mediation analysis suggested that differential methylation at this site could potentially explain 20% of the excess risk of diabetes for OMD compared with OMND, and this mediation effect was nominally statistically significant. This site maps within a CpG island (chr1:2983925−2987962) that spans both 5′ UTRs of LINC00982 and the promoter of PRDM16. In human placental tissue, methylation at two CpG sites in PRDM16 was associated with maternal fasting glucose in the second and third gestational trimesters [29]. CpGs in PRDM16 were also identified in another study which investigated differential DNA methylation in pancreatic islets of human donors with type 2 diabetes [30]. PRDM16 has been predominately studied in relation to brown adipose tissue (BAT), where it regulates differentiation from myoblastic precursors to BAT [31]. A SNP (rs12409277) in its 5′ flanking region was associated with lean body mass in a Japanese population, and the same SNP was demonstrated to affect the transcriptional activity of PRDM16 [32]. An essential role for PRDM16 in pancreatic islet development from a single fetal progenitor cell has recently been described [33], suggesting that this gene also has an important role in pancreatic development. In mice, the function of Prdm16 is regulated by epigenetic histone modification [34]. Taken together, multiple evidence shows that PRDM16 could be the regulatory target of hyperglycaemia and responsible in part for increased diabetes risk in OMD. The differentially methylated CpG site in PRDM16 identified in peripheral blood might be transferable to other tissues, making it a promising biomarker for diabetes risk screening.

While our previous in vivo clinical studies in Pima Indians have shown that exposure to diabetes in utero increases the risk of type 2 diabetes primarily via impairment of beta cell function, it is possible that reduced insulin sensitivity either as a primary effect or secondary to increased adiposity may also contribute to the higher risk of type 2 diabetes. For example, PPP1R3B, which was differentially methylated in our study, codes for the GL regulatory subunit of protein phosphatase 1 (PP1), which targets PP1 to activate glycogen synthase, thereby increasing insulin sensitivity of liver and skeletal muscle [35, 36]. A decreased rate of glycogen synthesis in the skeletal muscle in response to insulin is a characteristic feature of individuals with type 2 diabetes [37], and polymorphisms in another regulatory subunit (PPP1R3A), which also targets PP1 activation of glycogen synthase in the skeletal muscle, were associated with insulin resistance and type 2 diabetes in several different populations [38–42]. Differential methylation of a second gene, PM20D1, could also affect insulin sensitivity via its effect on body weight. In humans, methylation of PM20D1 in cord blood is associated with birthweight centile [43]. Methylation of PM20D1 has also been associated with several adverse health events during early life, indicating that PM20D1 may be a target gene for in utero epigenetic programming in multiple organs [44, 45]. Differential methylation of PM20D1 in obese and non-obese individuals at older ages suggests that in utero programming of PM20D1 has sustainable effects much later in life [46].

Among the 29 genes identified in the current study which are differentially methylated in OMD but not in OMND, only TBL1X has been previously implicated in exposure to maternal diabetes. In our study, a CpG in the 5′ UTR of TBL1X was hypermethylated in the OMD group, and others have reported a 27.22-fold decrease in expression following exposure to maternal diabetes [47]. TBL1X encodes a mediator of β-catenin proteolysis through ubiquitination in the canonical Wnt pathway, which was highlighted by our pathway analysis (ESM Table 2).

The strength of this study is its sample size and detailed clinical measures of a well-phenotyped, longitudinal cohort. There are, however, several weaknesses. First, since blood samples were not obtained at birth in our longitudinal study, the DNA analysed for methylation was from individuals with a mean age of 13 years. Thus, our study can only detect in utero differences in methylation that are stable throughout many years, and our results could also be influenced by differences in methylation that arise postnatally. Prior studies of in utero methylation have shown evidence of both transient and stable methylation sites. For example, a longitudinal analysis showed that sites that were differentially methylated at birth and associated with birthweight and gestational age were no longer different when the child reached 7 years of age [48]. By contrast, DNA methylation differences associated with in utero exposure to maternal smoking have been reported to be long-lasting [49, 50]. A second weakness of our study is that we analyse DNA from peripheral blood, although we believe that the pancreas and corresponding beta cell function are the key tissues responsible for the development of early-onset type 2 diabetes as a consequence of in utero exposure to diabetes. Therefore, our study will only detect differences in methylation that affect pancreatic development/function that are similarly differentially methylated in the blood. Fine mapping of CpG sites from DNA isolated from specific target tissues is required to further understand the detailed molecular mechanisms.

In conclusion, our study identified differentially methylated CpGs in 39 genomic regions that achieved epigenome-wide significance in their association with exposure to a diabetic intrauterine environment. Methylation at three sites also nominally associated with insulin secretion, while a fourth site associated with future risk of type 2 diabetes. However, our studies of insulin secretion and future diabetes risk were conducted in relatively small numbers of individuals and thus may be subject to considerable stochastic variation; therefore, replication in other cohorts will be very informative. In addition, further investigations on quantitative gene expression as a consequence of DNA methylation at these CpG sites, as well as causal CpG mapping, will lead to a better understanding of the mechanism of diabetes and its complications.