Preeclampsia (PE) is a complex, heterogeneous disorder of pregnancy, diagnosed by the onset of maternal hypertension after the 20th week of gestation, with signs of maternal multi-organ dysfunction [1]. As with many pathologies of pregnancy, PE has no cure, robust predictive biomarkers, or effective treatments, other than the delivery of the infant to discontinue the pregnancy and remove what is thought to be the causative organ, the placenta. Repeated attempts to characterize the placental molecular pathology and identify biomarkers of PE by applying a binary approach (PE versus control) have not been clinically fruitful, and we hypothesized that this is due to the existence of multiple molecular subtypes of PE [2].

In support of this hypothesis, we recently published a large unsupervised clustering analysis of microarray data from a PE-focused placental cohort (N = 330), including 157 highly annotated samples purchased from a single biobank [3]. This revealed five clusters of placental gene expression containing at least three clinically significant etiological subtypes of PE: “maternal”, with term and near-term deliveries of average-sized infants and placentas that appear molecularly similar to normal healthy control samples; “canonical” with high placental expression of known PE markers, preterm deliveries, low fetal weights, and evidence of maternal malperfusion; and “immunological” with severe fetal growth restriction, enrichment of immune response genes, and histological signs of maternal anti-fetal/placental rejection [3], belonging to transcriptional clusters 1, 2, and 3, respectively. An additional subtype of PE placentas with chromosomal abnormalities was also discovered within cluster 5 (and supported by array-based comparative genomic hybridization (aCGH) analysis), but showed no strong clinical association [3].

However, despite our considerable progress towards understanding the molecular diversity observed amongst PE patients, RNA is relatively unstable, easily affected by technical variability [4], and rarely successful as a therapeutic target [5], limiting its clinical utility. We, therefore, propose that the integration of an additional level of molecular information in these placentas, such as DNA methylation, will compensate for these restrictions [4], as well as improve our understanding of the molecular pathology.

DNA methylation is a mitotically heritable epigenetic mark employed by the cell to control gene expression without altering the genetic sequence [6], although the relationship between the two data types is exceptionally complex [7–10]. Given the flexibility for modification in the epigenome, these methylation events may also serve to provide insight into the environmental exposures sustained by the cell [11], and as potential biomarkers of early cellular transformations [12]. In fact, many examples exist, particularly in the cancer field, for the exploitation of DNA methylation in the diagnosis, prognosis, and prediction of drug response in disease [12, 13], and as possible therapeutic targets [14, 15].

Here, we subject a subset of our highly annotated cohort samples to DNA methylation arrays and investigate differences in the placental methylome between our previously identified transcriptional clusters, as well as relationships between the two data types. Furthermore, by assessing epigenetic changes associated with the observed pathological gene expression, we also attempt to discover novel therapeutic targets for the various PE subtypes.

Our previous work unbiasedly investigating the placental heterogeneity observed in preeclampsia [3] revealed five transcriptional clusters, including four subtypes of PE placentas. However, while gene expression microarrays are an invaluable tool for understanding disease, it is also possible that, in some cases, an alternate level of molecular information is highly involved in the development of the pathology. Combined epigenetic and expression analysis of the same preeclamptic placentas has only ever been performed for a small number of samples [32, 33] or genes [34]. We, therefore, predicted that the integration of matched genome-wide DNA methylation data would further improve our understanding of these placentas, and allow us to investigate both the mechanisms underlying the formation of the transcriptional clusters and the associations between the multi-molecular data.

Overall, we found that the relationships between the transcriptional cluster 1–3 samples were still visible within the DNA methylation information, indicating a significant global relationship between the two data types in these samples. Cluster 5 placentas, on the other hand, no longer formed a distinct group by methylation. This is unsurprising given that this data type is known to be fairly robust to copy number abnormalities [35], the driving force behind the molecular formation of this cluster.

Within transcriptional clusters 2 and 3, controlling for fetal sex and gestational age in the identification of differentially methylated sites, compared to the healthier cluster 1, dramatically reduced the number of significant sites (66,837 to 8763 in cluster 2; 13,348 to 340 in cluster 3). However, it predominately corrected the observed imbalance in the direction of change (80% hypomethylated to 38% hypomethylated in cluster 2; 68% hypomethylated to 48% hypomethylated in cluster 3). Since both clusters 2 and 3 are significantly younger than cluster 1 (p < 0.01 and p = 0.02, respectively), this fits with the knowledge that placentas become progressively more methylated with time [36], while in cluster 3, a moderate bias in fetal sex (p = 0.12) may have also been involved. Additionally, controlling for fetal sex and GA substantially increased the proportion of significant sites that showed a strong linear relationship with gene expression (5% to 9% in cluster 2; 2% to 8% in cluster 3), thereby confirming that a large number of sites in the genome undergo DNA methylation changes in response to differences in these two factors that are independent of epigenetic regulation and gene expression [9, 36, 37].

An additional result of interest was the CpG distribution of significant positions found in transcriptional clusters 2 and 3. CpG islands are most commonly associated with the regulation of gene expression, especially when located in the gene’s promoter region [8, 38]. We discovered that substantially fewer of the significant sites were mapped into CpG islands than anticipated, based on the reference distribution of all potential CpG sites, although those that were annotated to islands were, unsurprisingly, often found in close proximity to each other. Instead, the majority of significant positions were associated with CpG open sea enhancer regions. This is consistent with a previous report of enrichment of altered DNA methylation at enhancers and low CpG density regions in early-onset preeclamptic placentas [33]. These open sea enhancer regions, when significantly associated with gene expression, were generally located in the gene body and exhibited a positive relationship. Sites with a strong negative correlation, on the other hand, were frequently located in the promoter region (TSS200, TSS1500, 5′UTR), as expected, but were annotated to CpG shore regions, not islands. Relationships between CpG shores and gene expression are thought to be in response to the binding of transcription factors and changes in the chromatin structure around the promoter [39, 40].

While the observed proportions of differentially methylated sites that were associated with corresponding changes in gene expression (2–9%) are in line with prior studies [9, 32, 41, 42], this indicates that a large number of significant sites in clusters 2 and 3, compared to cluster 1, show no meaningful relationship to gene expression. Some of these DNA methylation alterations could be the consequence of changes in gene expression or function [29, 43, 44], or an adaptive response to maintain stable or rebalanced expression. They could further be remnants of an earlier developmental process, or the result of environmental exposures or treatments, where the transcriptional evidence is no longer measurable [36]. Furthermore, methylation is involved in a range of functions outside of direct transcriptional regulation, such as genome stability [45], splicing [8, 46], and cellular development [47], while gene expression can be regulated by a number of other factors, such as microRNAs [48, 49], transcription factors [43, 50], and/or histone modifications [51, 52]. Therefore, it is expected that these two data types would not fully agree at the individual gene level, although altered methylation sites not associated with changes in gene expression could still provide important information about the overall status and gestational history of these pathological placentas.

When the transcriptome and epigenome data was utilized simultaneously in an integrated analysis, this revealed modifications in TGF-beta signaling, cell adhesion and migration, oxidative phosphorylation, and carbohydrate and lipid metabolism pathways in cluster 2 placentas, confirming that a significant global relationship exists between the two data types. Placental dysfunction encompassing dysregulation of these pathways has been extensively described in the classical paradigm of PE pathophysiology and fits with our characterization of cluster 2 patients as demonstrating a “canonical” early-onset form of PE [3, 53–59]. Additionally, a number of the top significant methylation and gene expression correlations in this cluster (cg23730027 and FLNB, cg13553455 and COL17A1, cg11079619 and INHBA, cg19140548 and SH3PXD2A, and cg26509870 and PHYHIP) have been previously described in a smaller sample set of early-onset PE placentas [33], thus validating these relationships. We also identified several methylation probes in the gene body of FLT1, one of the most frequently investigated markers of PE, with a strong positive correlation to expression, as well as one associated site in the IGR with a strong negative correlation. These methylation differences could be involved in or attempting to compensate for the pathologically elevated expression of this gene [2, 3], and are significant findings missed by prior studies that have focused only on FLT1 promoter methylation in early-onset PE [34].

In cluster 3 samples, integrated alterations were identified involving antigen presentation, allograft rejection, cytokine-cytokine receptor interaction, Jak-STAT and TGF-beta signaling, glycosaminoglycan biosynthesis, and metabolism. These are also in line with our prior transcriptional results in this “immunological” PE group [3], in which we characterized this cluster of patients as demonstrating evidence of maternal anti-fetal/placental rejection. While not as widely discussed in the literature, the primary involvement of heightened immune activation has been described in several previous studies of PE pathophysiology, along with these other metabolic pathways [60–66]. Interestingly, one of the most significant methylation and expression relationships observed in this cluster involved DNMT3A (one of the DNA methyltransferase enzymes responsible for de novo methylation): a CpG island site (cg05544807) was hypermethylated in the DNMT3A gene body, compared to cluster 1, and demonstrated a negative relationship to expression. While this likely has global implications for the DNA methylation pattern observed in these cluster 3 placentas, decreased expression of DNMT3A has been specifically implicated in immunological-associated disorders [67, 68] and abnormal placentation in preeclampsia [69]. Therefore, this CpG site may serve as a potential target for the epigenetic modulation of pathological gene expression in this PE subtype.

Our study also has some inherent limitations. In our previous gene expression analysis, we utilized a large cohort of over 300 placentas to identify clusters and dysregulated pathways between them. Despite our current study being the largest to integrate methylation and transcriptional information in PE, this analysis involved only 48 placentas. Therefore, it is likely to still be underpowered, thus restricting our discovery of epigenetic changes in these samples to those with large effect sizes. As such, a future direction will be the validation of these findings, and perhaps the identification of new significant sites, in a larger cohort of samples. Additionally, as with all investigations of delivered placentas, it is impossible to determine whether the observed molecular modifications are part of the cause or the consequence of the disease process. Finally, our analysis is based on the assumption that the cell composition is the same across all samples. This is probably not the case, as differences in cell ratios can occur for a range of reasons [4, 42, 70–72], including placental maturation or sampling variability. Therefore, some of the epigenetic changes that we are interpreting as being reflective of gene regulation in all cells may instead be due to shifts in cell numbers [29]. However, unfortunately, until individual methylation patterns for all possible placental cell types have been established, this limitation cannot be resolved. This investigation is currently ongoing in our groups.

Overall, we have improved our understanding of the portion of the divergent gene expression involved in the development of transcriptional clusters 2 and 3 that is associated with changes in DNA methylation, as well as confirmed the lack of true biological cohesion in cluster 5 placentas. Differentially methylated sites in clusters 2 and 3, compared to the healthier cluster 1, may have potential as biomarkers of these patient groups for early screening in maternal serum, whereas specific genes and sets of genes exhibiting a strong epigenetic and transcriptional relationship (either linear or integrated) may serve as therapeutic targets to modify or prevent pathological changes in PE placental groups. However, a further increase in sample size and an assessment of additional modes of gene regulation will be required to fully comprehend the mechanisms underlying these subtypes.