Background
Diseases during the 9 months of human pregnancy markedly impact maternal and fetal health and predispose the newborn to diverse developmental and functional disruptions with lifelong consequences [
1,
2]. Adverse pregnancy outcomes are also associated with a higher maternal risk for cardiovascular, metabolic, and renal diseases later in life [
3‐
5]. Pregnancy health largely depends on the placenta, which constitutes the maternal-fetal interface after implantation and governs gestational homeostasis and response to adversity [
5]. The placenta performs a set of vital functions that are indispensable for maternal-fetal health, including gas exchange, transfer of nutrients, waste clearance, hormone production, and mechanical and immunological defense of the semi-allogeneic fetus [
5]. Placental dysfunction, in association with aberrant maternal-fetal homeostatic response, may lead to multifaceted diseases during human pregnancy [
5‐
7].
Preeclampsia (PE), fetal growth restriction (FGR), and spontaneous preterm delivery (sPTD) are the most common, syndromic complications of human pregnancy [
5,
8,
9]. PE is characterized by maternal hypertension, often accompanied by maternal target organ damage and a secondary adverse effect on fetal growth, attributed to placental dysfunction [
10,
11]. FGR, which emanates from maternal, placental, or fetal causes, affects fetal development and is a significant contributor to stillbirth and short- and long-term neonatal morbidity and mortality, and may also lead to prematurity [
7,
12,
13]. Any birth occurring spontaneously before the 37th week of pregnancy is classified as sPTD, which risks neonatal survival and may expose the offspring to health challenges during childhood and beyond [
14,
15].
Notwithstanding the distinct clinical phenotype that delineates each of these syndromes, placental dysfunction likely plays a central role in all, with abnormal remodeling of the uteroplacental spiral arteries early in pregnancy and subsequent attenuated perfusion and ischemic stress in PE [
6,
11]; hypoxia, reduced functional capacity and nutrient availability in FGR [
7,
12]; and inflammation with uteroplacental injury in sPTD [
8,
16]. Yet, these processes are not unique to any of the syndromes, and it is not clear how shared placental pathobiological pathways lead to distinct clinical phenotypes. Underlying placental histopathology is commonly divided into maternal vascular malperfusion (MVM), fetal vascular malperfusion (FVM), and acute and chronic inflammatory lesions (AI and CI, respectively) [
17]. Not surprisingly, isolated or combined histopathological findings are shared among clinical syndromes and are even found in placentas from uncomplicated pregnancies [
18‐
20].
Recent technological and informatics-based advances enable deeper insights into complex, multifactorial clinical syndromes. Several research groups recently harnessed omics-based approaches to deepen our understanding of abnormal molecular processes underlying obstetrical syndromes, thus defining disease subclasses that were not apparent through clinical or histopathological data [
21‐
28], resulting in improved diagnostic and predictive tools [
29,
30]. Here, we aimed to better define the molecular signatures of placental dysfunction in common obstetrical syndromes. For this goal, we gathered single source, rigorously obtained sets of multiomic analytes, derived from placental tissues with well-defined clinical conditions, creating a valuable multiomic data resource on which to perform analysis. We applied similarity network fusion (SNF) [
31] to integrate these multiomic data types into a comprehensive single network and identified clusters of similar phenotypic patterns independent of clinical presentation. Integrating omics data with clinical and pathological information allowed us to identify molecular drivers of placental dysfunction in major obstetrical disorders.
Discussion
Using an unbiased approach, we performed a combined analysis of clinical data, placental multiomics, and histopathology, in a cohort of pregnant women diagnosed with common obstetrical syndromes: HDP/preeclampsia, FGR, and sPTD. Integrating analyses of placental RNAs, miRNAs, proteins, and metabolites, we identified four molecular clusters, with most dominated by a different clinical syndrome.
Pairwise comparisons among the four predefined syndromes and the two controls revealed that the number of DE analytes and the magnitude of the difference were higher for the FGR + HDP group. We deepened our analysis using SNF, a machine-learning method that integrates diverse, heterogeneous high-throughput data sources into clusters [
31]. Assuming that the clustered analytes are related to disease pathogenesis, an unbiased SNF approach serves to cluster cases by shared etiological processes, irrespective of predefined clinical subtypes. Indeed, SNF created four placental data clusters, each largely consisted of one clinical syndrome. Cluster III, dominated by placentas from pregnancies complicated by PE, was associated with the most severe outcomes. The placentas in this cluster presented a molecular pattern of placental dysfunction across the four omics datatypes, supported by the expression of various known markers of placental injury [
72‐
74,
76‐
79]. The PE subclass in this cluster matched the phenotype previously referred to as “canonical PE” [
21,
24,
80]. In contrast, Cluster II exhibited the weakest placental dysfunction pattern.
To better define the contribution of altered cell composition to the multiomics changes, we used the BayesPrism for cell type deconvolution and found that syncytiotrophoblasts were the most prevalent cell-type in our biopsies. Unlike the relatively small differences in cell type representation across the clinical diseases, we found large differences among the multiomics-defined clusters. These data point to the contribution of both cell type composition and cell-specific gene expression changes to the multiomics phenotypes.
Through a comprehensive chart review, we validated the presence of severe features in women with HDP. Among the few misclassified participants, those from the severe PE group were assigned to Clusters I and II, while those from the FGR + HDP group were assigned to Cluster III. These findings emphasize that non-severe features are less indicative of placental dysfunction, whereas the co-occurrence of FGR and hypertension strongly suggests placental dysfunction. The omics-based clustering effectively captured the presence or absence of placental dysfunction in these cases. Notably, FGR, especially at term and without accompanying hypertension, is a challenging syndrome with many etiologies, and inconsistent clinical definitions [
81]. The SNF analysis allotted FGR placentas primarily to Clusters I and IV, suggesting that omics-based analysis may better define common features of this syndrome.
We used standardized pathological reports that were based on the widely accepted “Amsterdam criteria” [
19] and blindly reviewed our histological slides. Histopathological findings suggestive of placental dysfunction, such as MVM, correlated better with the omics-derived clusters than with the predefined clinical syndromes. This association was supported by previous studies, which identified vascular malperfusion lesions more frequently in early-onset PE and FGR [
82‐
84].
Although disease prediction was not our goal, we applied elastic net regression and causal probabilistic graphical models and identified a set of analytes that could accurately separate the clusters and predicted cluster allocation of placentas with incomplete omics measurements. Additionally, the repeated application of our SNF clustering procedure during the repeated 10-fold cross-validation of our predictive models confirmed both the robustness of our SNF cluster approach and the stability of our selected predictive analytes. These results suggest that discrete markers comprising the multiomics-defined clusters might be useful in predicting placental dysfunction. Indeed, despite using only plasma mRNA and no other analytes, we found concordant expression trends in six of the eight genes assessed between the placenta and maternal plasma, highlighting the feasibility of research into blood-based markers of molecular patterns that define placental dysfunction.
Our study is the first to interrogate a large number of placental multiomic analytes. Previous studies that applied hypothesis-free methods used a single datatype [
21,
23,
25,
80,
85] or several datatypes that were either separately analyzed or integrated using a stepwise approach [
22,
24,
84]. Moreover, we used placental biopsies, obtained through a most consistent protocol, to extract the omic analytes, thus minimizing the biological variability among different placental regions. Lastly, having clinical, histological, and molecular information from a single cohort of women minimized cofounding variables and simplified data integration. We corroborated our findings by identifying similar RNA and protein analytes and by validating our data using a curated list of known placental dysfunction markers, which were distributed across our clusters in a predictable way.
A limitation of our study is its retrospective nature, which prevented us from establishing causal relationships. Although we applied causal probabilistic graphical models to identify features linked to the SNF clusters, we note that in real-world datasets, with a finite sample size, this method cannot truly establish causality. Instead, the networks learned by CausalMGM can be considered a stringent test of association that can generate hypotheses of probabilistic causal interactions. Naturally, we accessed and analyzed each placenta at the end of pregnancy, yet some gestational age-dependent placental analytes that exhibited a transient change might not have been captured. Indeed, a significant challenge for placental research is the gestational age difference among the study cases. We addressed this challenge by accounting for the gestational age in our statistical models and by including women who delivered prematurely, and without evidence for placental abnormality, as a second control group. As expected, this control-PT group, although small, had similar proportions of histological findings as the term controls, had the lowest number of DE analytes compared to the term control group, and shared many DE molecules with the term control group when compared to the pathological groups. The clustering identified clear placental injury patterns in some clusters and less in others. This observation could be explained by the diverse pathways leading to the syndromes and by the varying placental contribution versus the contribution of other factors, which we could not capture in our data.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.