This was a single-center, prospective cohort study designed and conducted in accordance with the principles expressed in the Helsinki Declaration. To avoid selection bias, subjects were enrolled sequentially from patients attending our clinics. All subjects provided written informed consent before enrollment. The protocol was approved by the Icahn School of Medicine at Mount Sinai Institutional Review Board (IRB) (HS#: 12-00714, GCO#: 12-1367). Consenting subjects received a full infertility work-up including measurement of serum anti-Müllerian hormone (AMH), FSH, and LH concentrations and transvaginal ultrasound (performed on day 3 of menstruation), and were deemed appropriate candidates for in vitro fertilization. Patients with irregular menstrual cycles and/or signs of hyperandrogenism had measurement of androgen levels and clinical evaluation of hirsutism through the Ferriman–Gallwey score. Twelve patients were recruited; six with PCOS and six without PCOS. The diagnosis of PCOS in study patients was based on the 2003 Rotterdam consensus criteria. Patients were diagnosed with PCOS by meeting two of the following three criteria: oligo/anovulation, hyperandrogenism (either clinical or biochemical), and sonographic evidence of polycystic ovaries. Clinical hyperandrogenism was defined clinically by the presence of hirsutism, alopecia, and/or acne, or subclinically by increased levels of serum testosterone and/or dehydroepiandrosterone sulfate. Polycystic ovaries were diagnosed by pelvic sonography according to the Rotterdam conference criteria. Patients who met Rotterdam criteria for PCOS were identified through a natural language processing query of the electronic medical record data base. They were approached in accordance with approved IRB protocols, and all signed informed consent for participation in the study. There were no exclusion criteria. The control group included ovulatory women matched for age and BMI without PCOS features and delivered a full-term singleton over the same time period (Table 1).

Umbilical cord blood was collected at delivery of the placenta in Purple Top BD Vacutainers (Franklin Lakes, NJ, USA), thoroughly mixed and aliquoted in 2 ml cryovials (Thermo Fisher Scientific, Waltham, MA, USA), snap frozen in liquid nitrogen and stored at −80°C.

The CXT 350 frozen sample aliquotter (CryoXtract, Woburn, MA, USA) was used to retrieve tissue subaliquots for downstream applications to prevent the thawing of the full UCB aliquots. DNA was extracted using the Maxwell 16 automated DNA/RNA extraction equipment (Promega, Madison, WI, USA) and quantified with Nanodrop spectrophotometer (Thermo Electron North America, Madison, WI, USA).

DNA samples were subjected to methylation analysis using the Illumina Infinium HumanMethylation450 BeadChip Kit (Illumina, San Diego, CA, USA) at the University of Cincinnati Genomics, Epigenomics and Sequencing Core (GESC). Native DNA was sent from our labs to GESC. GESC personnel were blinded about the samples’ diagnosis. Samples were treated following the HumanMethylation450 BeadChip consolidated protocol. Briefly, native DNA was bisulfite treated using the EZ-96 DNA Methylation-Gold Kit according to the manufacturer’s protocol (Zymo Research, Irvine, CA, USA). Bisulfite-treated DNA was then hybridized to HumanMethylation450 BeadChips that were later scanned. The intensities of the images were extracted using the GenomeStudio (v.2011.1), “Methylation Module” (1.9.0) software (Illumina, San Diego, CA, USA), which normalizes within-sample data using different internal controls that are present on the HumanMethylation 450 BeadChip and internal background probes.

Processed data were sent back to our lab for analysis. Probes with a low signal (detection p > 0.05) across all samples were excluded from analysis. For each CpG dinucleotide, we calculated a Beta-value representing the ratio of the methylated probe intensity and the overall intensity (sum of methylated and unmethylated probe intensities). Beta-values spans between 0 and 1, or 0 and 100%. Following a consolidated procedure, we transformed the Beta-values into M values using the following equation: M=log2[Beta/(1−Beta)].

This transformation provides a straightforward method for using the Beta-value statistic and obtaining the unique statistical properties of the M-value which are not compressed in the low and high range, provide better insights into the distribution of methylation, and can be better handled by statistical tools (16). To identify differentially methylated genes between UCB DNA samples taken from PCOS pregnancies vs controls, paired Linear Models for Microarray (LIMMA) test was performed between PCOS and controls. LIMMA is a statistical tool that was designed for gene expression analysis and it has found usage outside this area. It performs parallel linear regression analyses using an empirical Bayes approach to shrink the estimated sample variances toward a pooled estimate. This approach produces robust/stable estimates, especially for small sample size (17).

Because of the limited number of subjects, the analysis was focused on differentially methylated CpG dinucleotides with a minimum twofold methylation change obtained as: Fold change=2|(AVG(M)Cases−AVG(M)Controls)|.

For all genes carrying multiple differentially methylated CpG dinucleotides, the M methylation value was averaged in order to score the gene overall methylation level. Some 23 genes carried both hyper- and hypomethylated CpG dinucleotides; by averaging the CpG dinucleotides M methylation values we obtained 12 genes with an overall hypermethylated profile and 11 with a hypomethylated profile (Tables S1, S2, and S7 in Supplementary Material).

Additionally, because of the asymmetric distribution of sexes of the babies in the control and PCOS groups (Table 1), in our primary analysis we removed all sex-specific differentially methylated CpG dinucleotides. A subsidiary analysis was also performed to determine if any residual gender-specific difference in methylation profiles of male and female PCOS offspring vs their control gender counterparts existed.

Network and pathway analysis was conducted using the Ingenuity Pathway Analysis (IPA) engine for the analysis of “omics data” (http://www.ingenuity.com). The standard setup for network analysis provided by the IPA core analysis was employed with one exception; because of the confidence provided by the large amount of input genes, the molecules per network parameter was set at 140.

All genes carrying single and multiple, concordant and discordant differentially methylated CpG dinucleotides were used for the IPA analysis. A total of 918 unique genes were thus used (see Tables S1, S2, and S5 in Supplementary Material for their differential methylation status). Genes were fed to IPA complete of their differential methylation M values as: Differential M Methylation Value=(AVG(M)Cases−AVG(M)Controls).

Only significant genes and pathways identified by employing the built-in right-tailed Fisher’s exact test were used for the network-specific canonical pathway analysis and constructing the PCOS “superpathway.”

A total of 614 hypermethylated and 1,066 hypomethylated CpG dinucleotides were identified (Figure 1; Tables S1 and S2 in Supplementary Material). Some 209 of 614 hypermethylated and 280 of 1,066 hypomethylated CpG dinucleotides mapped to genomic regions not associated with any known gene (Table S3 in Supplementary Material). These CpG dinucleotides showed a distribution strongly associated with the length of the chromosome they map within (Spearman’s rho for non-parametric bivariate correlation: hypomethylated CpG dinucleotides = 0.720; p-value < 0.001—hypermethylated CpG dinucleotides = 0.739; p-value < 0.001) (Table S4 in Supplementary Material).

The remaining 405 of 614 hypermethylated CpG dinucleotides mapped in or near 335 genomic tracks banked in the UCSC genome browser (https://genome.ucsc.edu/cgi-bin/hgGateway—GRCh37/hg19 Assembly), and 786 of 1,066 hypomethylated CpG dinucleotides mapped in or near 606 genomic tracks banked in the same database (Figure 1; Tables S5 and S6 in Supplementary Material). Out of a total of 941 banked genomic tracks, 918 referred to unique genes; of these, 895 carried single or multiple CpG dinucleotides either hyper- or hypomethylated, while 23 genes carried both hyper- and hypomethylated CpG dinucleotides (Tables S1, S2, and S7 in Supplementary Material). The number of differentially methylated CpG dinucleotides per gene was correlated with gene density and length suggesting a direct role on these CpG dinucleotides in controlling the activity of the genes they map within (Wilcoxon Rank-Sum Test p-value = 0.047).

Ingenuity pathway analysis enlisted 908 of the 918 unique genes carrying differentially methylated CpG dinucleotides and clustered 794 of the 908 genes into 21 gene networks (Table S8 in Supplementary Material). Because the vast majority of unique genes (720 unique genes, or 91% of the total 794 genes distributed across all networks) resided in networks 1–10, we focused our analysis on these 10 networks (Table S9 in Supplementary Material).

Table 2 summarizes the main findings of the IPA analysis (see also Tables S8 and S9 in Supplementary Material). The highest scored network by IPA was Network 1, the “Carbohydrate Metabolism, Lipid Metabolism, Small Molecule Biochemistry” network (Figure 2A; Table S8 in Supplementary Material). This network included seven main nodes, five of which highlight three features of this network, namely hormonal regulation (ESR1), metabolic rate control through the control of the mitochondrial activity (APP, PARK2), and glucose metabolism [Proinsulin domain (INS is the focus gene), PAX6].

The canonical pathways analysis of network 1 (Figure 2B) revealed two main pathway webs. The first includes Clathrin-Mediated Endocytosis, the key cellular entry port for metabolites hormones and proteins, the FXR/RXR, and the LXR/RXR activation pathways which regulate bile acid trafficking and cholesterol, fatty acid, and glucose homeostasis, respectively, and involve the activity of INS. This web connects with the Estrogen Receptor Signaling pathway centered on the activity of the estrogen receptor ESR1 which interfaces with the Wnt pathway (Figure 2B, left). The second web involves lipid metabolism and regulation of mitochondrial function determined by genes APP and PARK2 within network 1. This web is centered on the canonical pathways for the Triacylglycerol Degradation, Retinol Biosynthesis, Docosahexaenoic Acid (DHA) Signaling and Mitochondrial Dysfunction (Figure 2B, right). Interestingly, the IPA upstream regulator analysis also highlighted the effect of the Triacylglycerol Degradation pathway on the activity of the FGF1 growth factor, a known player for hepatic and peripheral insulin resistance.

Network 2 (Figure S1A in Supplementary Material) was dominated by the G-Protein Coupled Receptor Signaling pathway that directly interfaces with the Insulin Receptor Signaling pathway and other signaling pathways involved in inflammatory and regulatory metabolic functions (Figure S2A in Supplementary Material). These functions intersect with the activity of MOV10, a putative helicase gene related to the Phosphoinositide 3-Kinase (PI3K) family.

Network 3 (Figure S1B in Supplementary Material) highlights strong influence on lipid metabolism and neurotransmitter signaling. Network 3 included different forms of the d-myo-inositol Biosynthesis, key pathways for phosphatidylinositol metabolism and neurotransmitter signaling (Figure S2B in Supplementary Material).

Networks 4 and 5 (Figures S1C and S1D in Supplementary Material) feature nodes involved in cardiovascular system development and function. Specifically, Network 4 includes pathways related to inflammatory responses (Figure S2C in Supplementary Material) with upstream analysis showing modification of activity of interferon alpha, a key pro-inflammatory cytokine. Network 5 is populated by pathways with important links to cardiovascular system development (Figure S2D in Supplementary Material). The canonical pathway webs of both network 4 and 5 were linked to Type I and II diabetes pathways, and network 5 linked with hepatic functions. Network 5 canonical pathways also included the Androgen Signaling pathway. The IPA upstream regulator analysis also pointed at differential effects on the regulation of the peptidase SHH, a key gene for the embryo development programing that is tightly linked to cholesterol metabolism.

Networks 6–10 (Figures S1E–S1I in Supplementary Material) highlight the role of the glycosaminoglycan family of signaling carbohydrates regulating the activity of the transcription factor TGFB1 as highlighted by the IPA upstream regulator analysis. Additionally, differentially regulated pathways showed a role for control of amino acid biosynthesis centered on glutamate signaling and operated by the Urea Cycle, Arginine Biosynthesis and Citrulline biosynthesis pathways (Figure S2E in Supplementary Material). Of relevance is the significant scoring for the Thyroid Hormone Metabolism and the Estrogen Receptor Signaling pathways (Figure S2E in Supplementary Material). Among other relevant pathways are those that link the activity of networks 6–10 with those of networks 1–5 (Table 2) including Cholesterol Biosynthesis and Processing (ABCA7, ASGR1), FXR/RXR (RBP4) and LXR/RXR (RBP4); amino acids, fatty acids and glycogen metabolism (FASN, PCCA); estrogen receptor activity (ESR1); mitochondrial dysfunction (APP, Jnk domain, PARK2); G-Protein Coupled Receptors; Retinol pathways; and the PI3K family (PIK3R5, PIK3R6). These pathways highlight involvement of networks 6–10 in embryonic growth regulation suggested by co-existence with developmental pathways such as NANOG, NOTCH, eNOS, nNOS and OCT4 (Figure S2E in Supplementary Material).

Due to the gender imbalance between the PCOS and control offspring (Table 1) initial analysis included all patients and controls thereby excluding all gender-specific differentially methylated CpG dinucleotides. However, to ascertain whether gender differences in epigenetic reprogramming existed during pregnancy a subsidiary gender-specific methylation analysis was also performed. In the female PCOS offspring, most hypomethylated pathways were associated with both lipid metabolism and carbohydrate/glucose metabolism. In addition, numerous inflammation related pathways (e.g., “lymphocyte activation,” “regulation of cytokine secretion,” and “positive regulation of cell adhesion”) were also significantly hypomethylated (i.e., possibly upregulated). In contrast, in the male PCOS offspring, only glucose metabolism and insulin signaling pathways were hypomethylated (i.e., possibly upregulated) while the lipid metabolism pathways were not differentially methylated.

As in the females, several, albeit significantly fewer, inflammation related pathways were also hypomethylated in the male PCOS offspring (“B cell differentiation,” “lymphocyte differentiation,” “cell activation during immune response”), these data suggest a more pronounced effect on lipid metabolism and inflammation in female than in male PCOS offspring.