Epigenetic and transcriptomic alterations in offspring born to women with type 1 diabetes (the EPICOM study)

This study is a part of the EPICOM (Environmental Versus Genetic and Epigenetic Influences on Growth, Metabolism and Cognitive Functions in Offspring of Women with Type 1 Diabetes) study (ClinicalTrials.gov registration no. NCT01559181). From 1992 to 1999, pregnancies in women with pregestational T1DM were prospectively reported to a registry managed by the Danish Diabetes Association. Information regarding diabetes status and pregnancy outcome was reported to the registry by local obstetricians at eight hospitals in Denmark, who were responsible for antenatal care and delivery for pregnant women with T1DM. Coverage of cases from the reporting centers spanned from 75 to 93%, as described by Jensen et al. [19]. The Danish Diabetes Association Register consists of 1326 records of children born to women with T1DM. Only children born after 24 completed gestational weeks were registered. As part of the EPICOM study, we invited one child per mother (the oldest) to participate in a clinical study concerning metabolic risk. Controls were recruited from the background population and were matched by sex, age, and postal number. Details from this study have previously been described by Vlachova et al. [20].

Among the oldest participants of the EPICOM cohort, we selected 20 offspring of women with pregestational T1DM (index children) and their 20 matched controls (controls) as our exploratory cohort of DNA methylation and RNA expression (Table 1). Besides fulfilling the matching criteria for inclusion in the EPICOM cohort, we made sure that our exploratory cohort did not differ in body mass index (BMI), glucose tolerance, HbA_1c, cholesterol, or triglyceride levels and that the index child and their control were age matched. The index/control pairs were chosen representing the pairs with the smallest age difference, living in the same postal code, gender matched, and not suffering from any chronical illnesses. DNA methylation and RNA expression profiling were performed on leucocytes isolated from peripheral blood samples.

All measurements except from height were performed three times, and the mean value was used for the analyses. We measured height to the nearest 0.1 cm and weight (in kilograms) to the nearest 0.1 kg. Waist circumference was measured using a tape measuring to the nearest 0.5 cm midway between arcus costae and crista iliaca after exhalation and hip circumference corresponds to the widest measure around the hips. Preeclampsia was defined as blood pressure > 140/90 mmHg and proteinuria 2+ on a urine protein test strip (equal to 1.0 g/l).

We performed a standard 2-h OGTT by using a glucose load of 1.75 g/kg body weight up to a total of 75 g. Plasma glucose was measured at 0 and 120 min and serum insulin at 0 min (fasting).

Insulin sensitivity was evaluated by HOMA-IR and calculated using the original equation [21].

Glucose was measured in venous plasma with a hexokinase-glucose-6-phosphate dehydrogenase assay (Abbott Diagnostics, Abbott Park, Illinois, USA). Serum insulin was measured by ELISA using dual-monoclonal antibodies (ALPCO Diagnostics, Salem, New Hampshire, USA). Lipids were measured by enzymatic calorimetric analysis, end-up reaction (Abbott). Analyses of maternal HbA_1c between 1993 and 1999 were measured with local assays. Correction was made to a common standard (normal range of standard assay, 0.044–0.064) by multiplying the HbA_1c value with a correction factor as previously described (mean of the reference values for a standard assay divided by the mean of the reference values for the given assay) [19].

EDTA-treated peripheral blood samples were collected from the participants when they participated in the EPICOM study and were stored as EDTA treated at −80 °C until use. Genomic DNA was extracted from peripheral blood using QIAmp Mini Kit (Qiagen, Germany). For each sample, 1 μg of genomic DNA was bisulfite-converted using Zymo EZ DNA Methylation Kit according to the manufacturer’s recommendations. DNA methylation level was measured using the 450K-Illumina Infinium assay (Illumina, Inc.) at Aros Applied Biotechnology A/S.

All analyses were performed in R statistics, version 3.6.1, and R package minfi (version 1.32.0) was used for normalization, analysis, and visualization [22]. Detection p-values were calculated to identify failed positions with a p-value cut-off > 0.01. Probes were removed if they failed in more than 20% of the samples (n = 350). No samples were identified as failed, as the proportion of failed probes did not exceed 1% for any single sample. Density bean plots were used to identify outliers, and minfi’s inbuilt function was used to evaluate data with respect to extreme methylation outliers (> 3 SD away from the median). We performed background normalization and control normalization implementing the pre-processing choices of Genome Studio. Next, we applied subset-quantile-within-array-normalization correcting for technical differences between Infinium type I and II assay design allowing both within-array and between-sample normalization [23]. Cross-reactive probes (n = 29.541), probes with SNPs documented in C or G of the target (n = 18.284), and probes on sex chromosomes (n = 12,312) were excluded, leaving 415,009 probes. Methylation values were calculated as M-values (logit [beta]) (Equation (I)) [24].

Multidimensional scaling plots were evaluated to identify clusters of samples behaving differently than expected. Finally, the probes were annotated to the human genome version 19 using the enhanced Illumina annotation method developed by Price et al. [25].

We adjusted for differences in cell proportions using minfi’s estimateCellCounts that implements Houseman et al.’s regression calibration algorithm [26]. This algorithm returns relative proportions of CD4+ and CD8+ T-cells, natural killer cells, monocytes, granulocytes, and B-cells in each sample.

To identify positions where methylation is associated with intrauterine exposure to maternal T1DM, we fitted a linear model, which utilizes a generalized least squares model (lmFit of R package limma) allowing for missing values [27]. Significance was evaluated using F-test. The sample variances were estimated using an empirical Bayes approach with shrinkage towards the means. A Benjamini-Hochberg FDR below 0.05 was considered significant. We applied the model both without and with adjustment for estimated relative cell proportions (CD4+ and CD8+ T-cells, natural killer cells, monocytes, granulocytes, and B-cells).

We used DMRcate to identify any autosomal DMRs [28]. DMRcate identifies and ranks the most differentially methylated regions across the genome based on kernel smoothing of the differential methylation signal. The model performs well on small sample sizes and builds on the well-established Limma package, allowing us to incorporate estimated cell proportions as covariates. A FDR below 0.05 with a |ΔM-value| < 1 was considered significant.

Blood samples were drawn using PAXgene Blood RNA Tubes and placed 2 h at room temperature, sequentially stored overnight at −20° before being stored at −80° until analysis.

Whole transcriptome, strand-specific RNA-Seq libraries were prepared from total RNA using the Ribo-Zero Globin technology (Illumina, Inc.) for depletion of rRNA and globin mRNA followed by library preparation using the ScriptSeq technology (Illumina, Inc.). Depletion and library preparation were automated on a Sciclone NGS (Caliper, Perkin Elmer) liquid handling robot. The total RNA (1.7 μg per sample) was subjected to Baseline-ZERO DNase prior to depletion. Total RNA was purified using Agencourt RNAClean XP Beads before and after DNase treatment followed by on-chip electrophoresis on a LabChip GX (Caliper, Perkin Elmer) and by UV measurements on a NanoQuant (Tecan). Cytoplasmic and mitochondrial rRNA as well as globin mRNA were removed from 400 ng DNAse-treated total RNA using the Ribo-Zero Globin Gold Kit (Human/Mouse/Rat, Illumina, Inc.) following the manufacturer’s instructions, and quality of the depleted RNA was estimated on a LabChip GX (Caliper, Perkin Elmer). Synthesis of strand-specific RNA-Seq libraries were conducted using the ScriptSeq v2 kit (Illumina, Inc.) following the recommended procedure, and the qualities of the RNA-Seq libraries were estimated by on-chip electrophoresis (HS Chip, LabChip GX, Caliper, Perkin Elmer) of a 1 μL sample. The DNA concentrations of the libraries were estimated using the KAPA Library Quantification Kit (Kapa Biosystems). The RNA-Seq libraries were multiplexed paired-end sequenced on a NextSeq 500 (75 + 6 + 75 bp) using a high-output flowcell.

Paired de-multiplexed fastq files were generated using bclfastq (v.2.20 Illumina) and initial quality control was performed using FastQC. Adapter trimming was conducted using the GATK ReadAdaptorTrimmer tool followed by mapping to the human genome (hg19) in addition to transcripts from databases on non-coding RNAs (mirbase, mitranscriptome, rfam, snornabase, tjumirna, and trnascanse) using Bowtie and then further analyzed using Tophat, and Cufflinks and HTSeq-count (union method) [29, 30]. HTSeq-count (union method) was applied to produce raw counts which were then submitted for differential expression analysis in R using edgeR [31]. All non-informative features were filtered out by removing features with less than one count per million (CPM) in 39 samples removing 98,334 features, leaving 13,842 for downstream analysis. A generalized linear model was fitted, and p-values and log fold changes (Log2FC) were retrieved from the individual comparisons of index vs. controls. A Benjamini-Hochberg FDR below 0.05 was considered significant.

To investigate possible biological functions of the RNA expression changes in our cohort, we performed gene set enrichment analysis (GSEA) using Clusterprofiler [32]. The input genes were ranked based on the magnitude of changes in the index vs. controls comparison (log2FC). Disease associations were identified by disease ontology (DO) enrichment analysis, while Gene Ontology Biological Processes (GOBP) and REACTOME were used for functional and pathway enrichment. The top-enriched terms for each enrichment analysis were sorted according to p-value and presented as barplots. Furthermore, gene-concept networks that depict linkages of the genes and biological terms (DO, GOBP, and REACTOME terms) were made by evoking the cnetplot function of Clusterprofiler.

For methylation data, GOBP and REACTOME enrichment analysis were performed on all differentially methylated positions (p < 0.005) using the methylGSA package for R [33]. The probes were annotated to the human genome version 19 as previously described.

DNA methylation has been shown to play an important role in modulating gene expression. Therefore, we correlated changes in methylation and gene expression of the gene closest to the methylation site. Shared differentially expressed genes (DEGs) and DMPs (closest gene to methylation site) between index and controls (p < 0.05) were plotted in a scatter plot based on log2FC and ΔM-value. In the second part of the analysis, we used Spearman’s rank correlation to analyze the correlation between the specific methylation level (M-value) and RNA expression level (normalized counts) for shared genes. We considered the correlations to be interesting (either negative or positive) if p < 0.1.

Statistics was done using R 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria) with Bioconductor 3.9 [34]. DNA methylation data was analyzsed using the minfi [22], DMRcate [28], and Limma [35] package 1.32.0, and RNA-Seq data using edgeR [31]. Functional gene enrichment analysis was carried out using the R packages Clusterprofiler and MethylGSA [32, 33]. Graphics were made using basic R functions ggbio, Gviz, DEseq, DEXSeq, and ggplot2. The package knitr was used for data documentation.

Performing whole-blood genome-wide DNA methylation analysis using the 450K-Illumina Infinium assay, we identified 13,867 DMPs with a p < 0.05 between index and controls. To reduce the number of false positives, a criterion of p < 0.005 and |ΔM-value| > 1 was applied, resulting in 14 DMPs (9 hypomethylated and 5 hypermethylated) (Fig. 1a and Additional file 1: Table. T1) However, after correction for multiple testing, no single DMP reached the threshold of a Benjamini-Hochberg false detection rate (FDR) (< 0.05). DMRs defined as methylation in groups of nearby positions have been proposed to be involved in transcriptional regulation. Therefore, we extended our analysis to the regional level. Applying a p < 0.005, 37 DMRs were found (Additional file 2: Table T2). However, no DMRs were identified applying a FDR < 0.05.

Subsequently, we analyzed gene expression by RNA-Seq to test if any DEGs, coding or non-coding RNA, were present when comparing our two cohort groups (index vs. controls). We identified 39 coding DEGs (38 upregulated and 1 downregulated) (Table 2, Fig. 1b and Additional file 3: Fig. F1) (p < 0.005 and a log2FC ≥ 0.3). In addition, we also found 20 non-coding DEGs (7 upregulated tags and 13 downregulated tags) (p < 0.005 and a log2FC ≥ 0.3) (Additional file 4: Table T3 online). No DEGs reached an FDR < 0.05.

To gain mechanistic insight into functional importance of the coding DEGs found above, we used enrichment analysis to associate these changes to disease ontology (DO), biological processes (GOBP), and biological pathways (REACTOME). We observed enrichment in disease ontologies relating to acquired metabolic disease, glucose metabolism disease, diabetes mellitus, carbohydrate metabolism disease, disease of metabolism, and clustering around genes including CXCL5, EGF, and SPARC (Fig. 1c–g). For biological processes and pathways, the analysis revealed enrichment relating to striated muscle tissue development, muscle tissue development, growth, and developmental growth and biological pathways including ECM proteoglycans, RAF/MAP kinase cascade, MAPK1/MAPK3 signaling, and MAPK family signaling (Fig. 1c–g and Additional file 5: Fig. F2 online).

Functional enrichment analysis was also carried out with identified DMPs (p < 0.005) as input. Again, enrichment in disease ontologies and biological pathways relating to carbohydrate metabolism, glucose transportation, and pathways including glucagon-like peptide-1 (GLP1) regulated insulin secretion were in the top-enriched terms (Additional file 6: Table T4 and Additional file 7: Table T5 online).

As we were unable to robustly identify changes at the single-gene level, we used WGCNA to detect modules of correlated genes, and to test if any of these modules could be related to our two cohort groups and the associated clinical traits. Based on co-expression similarities, the genes were split into 25 modules (Additional file 8: Fig. F3). The eigengenes of these modules, representing a summary of the expression of all genes within the module, were correlated with group status, index or control, and the clinical and paraclinical measurements. The modules that were significantly associated to our two cohort groups, sex or clinical traits, were selected (Fig. 2a). The strongest association observed was a negative correlation of the salmon module with the gender female (sex, cor = −0.97, p = 4e−23). This module consisted of 153 genes, and as expected, almost exclusively came from the sex chromosomes (21 from the X chromosome and 13 from the Y chromosome). Other female-specific traits like increased hip circumference and fat percentage were also negatively correlated with eigengene expression in this module, while male traits like increased height, waist-hip ratio, and lean body mass were positively correlated. Interestingly, we also observed that the greenyellow module, consisting of 201 genes, was positively correlated with the T1DM offspring group (cor = 0.48, p = 0.002) and unaffected by sex. We used a signed network structure to create our modules, and thus, a general upregulation of all genes within this module was expected for the index group. This assumption was supported by the presence of 22 out of the 38 upregulated DEGs in the greenyellow module. The association of the individual genes to the index group revealed that a subset of the module genes was more strongly correlated with the index group (e.g., CLU, SDPR, BEND2, EGF, STON2, TFPI, SPARC, TUBB1) (Fig. 2b). Furthermore, the greenyellow module was negatively correlated with offspring systolic blood pressure at follow-up (cor = −0.35, p = 0.03) (Fig. 2c), gestational age at birth (cor = −0.52, p = 7e−04) and the 5-min Apgar score (−0.53, p = 7e−04). Maternal BMI was not correlated to any modules; however, maternal age was positively correlated to the lightgreen module, which was also positively correlated to increased offspring weight and length. The presence of preeclampsia was not significantly associated to the greenyellow group module (T1DM offspring). However, we observed that preeclampsia was positively correlated to the magenta, darkred, and turquoise modules and negatively correlated to the tan, blue, brown, and darkgreen module. Interestingly, almost the exact same correlations were observed for fat percentage. Thus, presence of preeclampsia may predispose to an increased fat percentage later in life.

All genes from the greenyellow module were used as input for functional enrichment analysis to associate these changes to disease ontology (DO), biological processes (GOBP), and biological pathways (REACTOME) (Fig. 2d–f). Enrichment in ontologies and pathways relate to platelet degranulation/activation, blood coagulation, smooth muscle cell contraction, and fibrin clot formation. Within disease ontologies, we observed enrichment in terms relating to coagulation disease, atherosclerosis, thrombosis, myocardial infarction, and hemorrhagic disease.

DNA methylation has been shown to play a regulatory role in mediating gene expression, and if located within a gene promotor, higher levels of methylation usually repress transcriptional expression [8]. To further explore the consequences of altered DNA methylation patterns, we therefore compared the DNA methylation data with our gene expression data (Fig. 3 and Additional file 9: Table T6). First, we visualized correlations between changes at the group level. That is, the mean change in methylation level plotted against the mean change in gene expression of the gene annotated to the methylation site. DEGs overlapping with DMPs (annotated to closest gene to the methylation site), between T1DM exposed offspring and controls (p-value < 0.05), were plotted in a scatter plot based on log2FC and ΔM-value.

This analysis revealed that 93 DMPs annotated to 76 DEGs showed a pattern where DNA hypomethylation was associated with an increased RNA expression (Fig. 3, lower-right (blue)). Fourteen DMPs annotated to 12 DEGs were found to show a pattern were DNA hypermethylation was associated with decreased RNA expression (Fig. 3, upper-left (red)). Two hundred one DMPs annotated to 129 genes showed a pattern where DNA hypermethylation was associated with increased RNA expression. As the second step of our correlation analysis, we performed a Spearman rank correlation for each DMP-DEG pair, using each subject’s methylation level at the involved DMPs and expression level (normalized counts) for the corresponding RNA expression data as input. Here, our analysis showed that 2 DMPs and their corresponding genes had significant inverse correlation between DNA hypomethylation and increased RNA expression (p < 0.1) (CIITA and TPM1, Table 3). DNA hypermethylation of CpG sites localized within the gene body has been described as being associated with increased RNA expression, and in our analysis, we found 4 annotated DMPs localized within the gene body and their corresponding genes displaying positive correlation between DNA methylation and RNA expression (p < 0.1) (PXN, ST8SIA1, LIPA, and DAXX, Table 3).