Integrated unbiased multiomics defines disease-independent placental clusters in common obstetrical syndromes

Deidentified demographic and clinical characteristics were obtained from the Steve N. Caritis Magee Obstetric Maternal and Infant (MOMI) Database and Biobank at Magee-Womens Research Institute and the Health Record Research Request Service at the University of Pittsburgh. We included only women with singleton live birth. The six study groups (four disease groups and two control groups) are detailed in the “Results” section. The severe features of PE were defined according to the guidelines of the American College of Obstetricians and Gynecologists [32] most recent to the participants’ delivery. FGR was defined by birth weight below the 3rd percentile for gestational age, based on the World Health Organization’s weight percentile calculator [33]. sPTD included women whose labor started with contractions or premature rupture of membranes and delivered before 37 weeks. Additional file 1: Fig. S1 describes the workflow of the study.

All placentas were collected by the Obstetrical Specimen Procurement Unit at the MWH. Placental biopsies (5 mm³) were obtained from a region midway between the cord insertion and the placental margin and between the chorionic and basal plates, as we previously detailed [34]. Within 30 min of delivery, biopsies from the same site were (1) placed in RNA preservation solution (RNAlater) for 48 h and then snap-frozen for RNA extraction, (2) immediately snap-frozen in liquid nitrogen and stored at −80°C until processing for proteomic or metabolomic analyses, and (3) processed for paraffin embedding. Out of the 348 original cases in our cohort, we obtained 318 samples in RNAlater and 343 snap-frozen samples. All 348 were also paraffin-embedded.

Whole blood samples were obtained from some of the participants during their admission to the labor and delivery unit. Among these, we randomly chose four participants, representing each SNF cluster, who were diagnosed with severe PE, FGR, or sPTD.

RNA was isolated from the placental biopsies using TRIreagent (Thermo Fisher, Waltham, WA) and processed using the RNeasy mini kit (Qiagen, Germantown, MD), following the manufacturer’s instructions as routinely performed in our lab [35]. RNA quality was assessed using an Agilent bioanalyzer (Agilent Technologies, Santa Clara, CA) and an Agilent HS Total RNA 15nt kit (Agilent, #DNF-472T33) on an Advanced Analytical 5300 Fragment Analyzer. RNA concentration was quantified with a Qubit HS RNA assay kit (Invitrogen, Waltham, MA, #Q32855) on a Qubit 4 fluorometer (Invitrogen, #Q33238). Total RNA-seq libraries were generated with the Illumina Stranded Total RNA Prep kit with Ribo-Zero Plus (Illumina, San Diego, CA, #20040529) according to the manufacturer’s instructions. Briefly, 100 ng of input RNA was used for each sample with a 2 min RNA fragmentation time. Following adapter ligation, 13 cycles of indexing PCR were completed, using IDT for Illumina RNA UD Indexes (Illumina, #20040553 & 20040554). We generated small RNA-seq libraries using Qiagen’s QIAseq miRNA library kit (Qiagen, #331505) according to the manufacturer’s instructions. Briefly, 100 ng of input RNA was used for each sample. Following adapter ligation, 16 cycles of indexing PCR were completed using QIAseq miRNA 96 IL indexes (Qiagen, #331565). Library assessment and quantification were done using Qubit 1 × HS DNA (Invitrogen, Q33231) on a Qubit 4 fluorometer and an HS NGS Fragment kit (Agilent, #DNF-474-1000) on an Advanced Analytical 5300 Fragment Analyzer. Libraries were normalized and pooled by calculating the concentration on the basis of the fragment size and library concentration.

Total RNA-seq libraries were sequenced on an Illumina NovaSeq 6000, using an S4 200 flow cell (Illumina, #20028313), with read lengths of 2 × 101 bp and an average of ~40 million reads per sample. Prior to sequencing, library pools were quantified by qPCR on the LightCycler 480 (Roche Diagnostics, Indianapolis, IN) using the KAPA qPCR quantification kit (Kapa Biosystems, Wilmington, MA). Small RNA-seq libraries were sequenced on an Illumina NextSeq 2000, using a P3 50 flow cell (Illumina, #20046810) with read lengths of 1 × 75 bp and an average of ~12 million reads per sample. Library generation and NextSeq sequencing were performed by the University of Pittsburgh Health Sciences Sequencing Core, Children’s Hospital of Pittsburgh. NovaSeq sequencing was performed by the UPMC Genome Center, Pittsburgh. The RNA libraries were aligned to the human reference genome GRCh38 using the RNAseq alignment tool STAR [36] and annotated with the latest GENCODE 30 [37]. We used STAR quantMode GeneCounts, a method counting reads overlapping with a single gene, to calculate the reads per gene for each RNAseq library, and these counts were used for further analysis [38].

Plasma was extracted from whole blood samples and 200ul were used for total RNA isolation using the miRNeasy mini kit (Qiagen #217004, Germantown, MD) following the manufacturer’s instructions. 60ug Glycogen (Thermo Scientific #R0551) and 300ng tRNA (Life Technologies #AM7119) were added per sample. cDNA was synthesized from 1μg of total RNA by using the High-Capacity cDNA Reverse Transcription kit (Applied Biosystems #4368813, Foster City, CA) according to the manufacturer’s protocol. RT-qPCR was performed using SYBR Select (Applied Biosystems #4472908) in QuntStudio5 (Applied Biosystems); cDNA templates were used to detect the relative expression of ALPP, PAPPA, LGR5, DUSP9, HTRA4, FLT1, LYVE1, and EDNRB. Analysis of qPCR data was performed using the delta-delta Ct method using GAPDH as a reference. The primer sequences are shown in Additional file 1: Table S1.

For the 343 snap-frozen samples, we extracted proteins using radioimmunoprecipitation assay (RIPA) buffer and measured the protein concentration using Versa max microplate reader (Molecular Devices, San Jose, CA). Placental proteins were analyzed using five Olink Target 96 panels (Olink Proteomics, Uppsala, Sweden): Cardiovascular II, Cardiovascular III, Development, Inflammation, and Oncology III. These were selected for their relevance to the placental biology addressed in our study. The proximity extension assay technology used for the Olink protocol was previously described [39]. Four hundred fifty-three unique proteins were measured in each placenta. Samples were processed in batches with pooled quality-control samples, which were included in each batch. All assay-validation data (detection limits, intra-, and inter-assay precision data) are available on the manufacturer’s website (www.​olink.​com).

Placental metabolites were analyzed by Metabolon (Morrisville, NC) using the Global Metabolomics platform. Two samples had insufficient material and therefore were not analyzed. Placental tissue samples (50 mg) were aliquoted and transported on dry ice to Metabolon. The detailed methods used by Metabolon were described by Ford et al. [40]. Metabolon’s informatic system was used for data extraction and peak identification, compound identification and quantification, curation, and data normalization. Samples were randomized across several batches and processed with pooled quality-control samples in each batch.

We used two information sets: pathology reports, obtained through standard clinical care and paraffin-embedded placental biopsies. The first dataset included pathology reports from the electronic medical records for 275 of our participants. A pathologist from the study team (WTP) reevaluated the reports for the presence of the major placental pathology patterns, as defined by the Amsterdam Placental Workshop Group Consensus Statement [19], including MVM, FVM, AI, and CI. For MVM diagnosis, we included cases with at least one MVM component, which included accelerated villous maturation (AVM), distal villous hypoplasia (DVH), villous infarct, decidual vasculopathy, and retroplacental hemorrhage and excluded cases with isolated placental hypoplasia. The components of CI were: villitis of unknown etiology, chronic deciduitis, chronic chorionitis, and eosinophilic/T cell chorionic vasculitis. Our second set of data was based on placental histopathological analyses of paraffin-embedded placental biopsies (n = 348) retrieved from the MOMI Biobank and examined by our study-team pathologist (WTP), who reviewed the slides while blinded to the clinical outcomes and determined the presence of AVM, DVH, and syncytial knots. These lesions are accessible for diagnosis when the biopsy is taken midway between the chorionic and basal plates: Other relevant lesions, including segmental avascular villi, delayed villous maturation, villitis of unknown etiology (VUE), diffuse villous edema, and chorangiosis, were rare and hence were excluded from the analysis.

We excluded, in downstream analysis, RNA with < 500 total counts or greater than 80% zero counts. The sex-specific genes located on the Y chromosome, as well as XIST and TSIX, were also removed from the analysis. For the miRNA dataset, analytes with an average count of < 10 were excluded. The count data of both the RNA and miRNA datasets were modeled with DESeq2 (v1.36.0) [38] and conditioned on clinical diagnosis, gestational age, infant sex, race, maternal pre-pregnancy BMI, maternal smoking status, delivery type, labor initiation, and the presence of labor. Using DESeq2, the count data were transformed into approximately normally distributed data on the log scale using a variance-stabilizing transformation [41], and principal component analysis (PCA) was used to identify outlier samples. Outliers greater than four standard deviations (SD) from the mean in either of the first two principal components of the RNA dataset were excluded from downstream analysis. Both the RNA and miRNA datasets were corrected for batch effects related to the biobank of origin while retaining variation associated with clinical diagnosis, gestational age, infant sex, race, maternal pre-pregnancy BMI, maternal smoking status, delivery type, and labor initiation using ComBat_seq from the sva (Surrogate Variable Analysis) package (v3.44.0) [42]. The batch-corrected data was again modeled by DESeq2 as described above, and variance-stabilizing transforms were performed to get batch-corrected log-scaled datasets.

The proteomics and metabolomics datasets were filtered to remove analytes with greater than 50% of samples below the detection limit. Measurements below the limit of detection for retained metabolites were imputed to the lowest measured value, whereas protein measurements were used as is for measurements below the limit of detection for retained proteins. For the proteomics dataset, PCA identified several outlier samples defined as samples greater than four SD from the mean in either of the first two principal components. These samples were removed from any downstream proteomics analyses. As with the RNA and miRNA datasets, both the proteomics and metabolomics datasets were corrected for batch effects related to the biobank of origin using ComBat [43] from the sva package [44].

Differential expression analysis with respect to clinical diagnoses for the RNA and miRNA datasets was conducted using the DESeq2 package [38]. The same comparisons across clinical diagnoses were performed while conditioning on the same covariates. The Benjamini-Hochberg procedure was applied to each differential expression analysis to control the false discovery rate (FDR).

DE analysis with respect to clinical diagnoses for the proteomics and metabolomics datasets was conducted using the limma package (v3.52.3) [45]. DE analysis was performed to compare between the two control groups and between each disease group and the two controls. All DE analyses were conditioned on gestational age, infant sex, race, maternal pre-pregnancy BMI, maternal smoking status, delivery type, labor initiation, and whether labor occurred to identify analytes that varied across clinical diagnoses, independent of the covariate effect. The Benjamini-Hochberg procedure [46] was again applied to each differential expression analysis to control the FDR.

Clinical characteristics were compared across the clinical diagnoses and the cluster labels. Continuous clinical features were compared across groups using the nonparametric Kruskal-Wallis test to determine if their distribution differed significantly across groups. To assess whether the frequencies of categorical clinical features vary with clinical diagnoses or cluster labels, chi-square tests were performed and controlled for FDR, as described above. For each significant test, additional post hoc tests were performed to assess which groups differed significantly from the others. For each significant continuous feature, Dunn’s post hoc test was performed. For each significant categorical feature, a post hoc pairwise Fisher’s exact test was performed for each possible 2 × 2 contingency table. The family-wise error rate for each post hoc test was controlled using the Holm-Bonferroni method [47].

We performed Bayesian model selection to compare the association between the histopathological features and disease conditions vs SNF clusters. While calculating the Bayesian Information Criterion (BIC) score for each regression model, we noticed that the BIC scores tend to penalize complex models more severely when sample size was large. This might negatively affect the BIC score for the model using the disease condition (n = 6) more than the model using the SNF clusters (n = 4). Therefore, we created a new disease condition variable by merging the two control groups, and FGR + HDP with the severe PE group. All models were tested for BIC scores.

To perform an unsupervised clustering analysis while integrating information from all four analyte datasets, we used similarity network fusion (SNF), implemented in the R package SNFtool (v2.3.1) [31], to construct a sample similarity matrix that combined information from the RNA, miRNA, proteomic, and metabolomic datasets. This fused similarity matrix was then used to perform spectral clustering. Before performing SNF clustering, each of the four analyte datasets was filtered to contain only the top quartile of highly variant analytes and only participants with measurements across all four datasets. For each analyte dataset, the original similarity matrix was constructed by applying a Gaussian kernel on the Euclidian distance between samples, followed by constructing a k-nearest neighbor graph and setting the weights of all non-neighbors to zero. The number of clusters used for spectral clustering was selected according to the eigengap heuristic [48], while the hyperparameters used for SNF (the bandwidth of the Gaussian kernel and the number of nearest neighbors) were selected through a stability-based approach [49]. Over a grid of possible hyperparameters, SNF and spectral clustering were performed on the full dataset. Then, for each of the possible combinations of hyperparameters, SNF and spectral clustering were performed on 50 random sub-samples of the dataset containing 80% of the patients in the dataset. The stability of the clustering on each sub-sample was assessed using the adjusted mutual information [50] of the sub-sample clusters and the clustering was performed on the full dataset. The combination of hyperparameters that resulted in the most stable clustering (σ = 0.3, k = 20) was selected.

To define molecular indicators of the four clusters, we identified analytes that were uniquely upregulated in each cluster compared to the other three. We measured the AUROC of each analyte’s ability to distinguish each cluster from the other three. The AUROC provides a nonparametric approach to ranking how well a marker can distinguish each cluster from the other three and allowed us to identify the top ten significant markers for each cluster for each molecular dataset. The R package pROC (v1.18.0) [51] was used to construct ROC curves and calculate their AUC. We used a similar approach to select two markers of each cluster for testing whether gene expression signatures of different SNF clusters are detectable in the plasma by qPCR. First, we filtered our list of candidate markers to include mRNAs with a mean expression in the top quartile that are preferentially expressed in placental tissue based on the Human Protein Atlas [52]. We then excluded genes that were expressed in blood and immune cells in the Human Protein Atlas and selected the top two upregulated genes with the highest AUROC distinguishing each cluster from the other three.

To assess the distribution of different placental cell types across SNF clusters and conditions, we deconvolved the bulk RNA-seq data on the basis of a single cell reference using the InstaPrism package (v0.1.4) [53] for derandomized implementation of the BayesPrism model [54]. The single cell reference was constructed from a scRNA-seq dataset consisting of placentas from two control and two term preeclampsia cases (GSE173193) [55]. The raw UMI count matrices from GSE173193 were processed with the Seurat package (v4.3.0) [56]. First, we filtered the dataset to remove low-quality cells. We removed any cells containing fewer than 500 reads or 250 unique genes, as well as cells that consist of more than 10% mitochondrial RNA. Cells with a complexity (defined as the ratio of the log of the number of unique genes and the log of the number of UMI) less than 0.8 were also removed. The degree of contamination from ambient RNA was estimated using DecontX [57], implemented in the Celda package (v1.16.1), and cells with a contamination score greater than 0.2 were excluded. Finally, we removed doublets using the scDblFinder package (v1.14.0) [58]. Cells were then clustered in Seurat using the Louvain clustering algorithm with a resolution of 0.35, and the set of differentially expressed genes for each cluster was determined by MAST [59]. Each cluster’s cell types were manually annotated on the basis of differentially expressed genes and their similarity to the markers identified in the original study [55]. After the cell type annotations were assigned to each cell, the following cell types were used as a reference by InstaPrism for the deconvolution of our bulk RNA-seq samples: cytotrophoblasts, syncytiotrophoblasts, fibroblasts, Hofbauer cells, endothelial cells, NK cells, and granulocytes. The significance of the associations of cell type proportions with SNF clusters and clinical diagnoses was assessed with the Kruskal-Wallis test. For significant associations, a post hoc Dunn’s test was applied. The family-wise error rate of these tests was controlled using the Holm-Bonferroni method.

In addition to analyte-level DE and cluster marker identification, gene set enrichment analysis (GSEA) [60] was used to identify pathways enriched across clinical diagnoses and SNF clusters. Using the RNA dataset, we assessed the differential enrichment of canonical pathways in the Reactome Knowledgebase [61]. Due to the enrichment of large numbers of overlapping gene sets in the Reactome Knowledgebase, the GSEA for these pathways was conducted using SetRank (v1.1.0) [62], an advanced GSEA algorithm that corrects for pathways that are only significant due to their overlap with other pathways in the database. For each comparison of clinical diagnoses, genes were ranked according to their t-statistic from the corresponding DE analysis, allowing the GSEA to condition on the set of clinical covariates used in the DE analysis (gestational age, infant sex, infant race, maternal pre-pregnancy BMI, maternal smoking status, delivery type, labor initiation, and whether labor occurred). In addition, we applied this analysis to markers of each of the SNF clusters to identify pathways enriched in each cluster compared to the other three. In this case, the genes for each cluster were ranked according to their AUROCs, calculated as described above. The significance of a pathway in SetRank was determined by a corrected p-value, correcting for genes that were enriched because they belonged to another enriched pathway. The significance of canonical pathways was FDR-corrected at p < 0.05.

Analysis of differential metabolic pathway enrichment for each clinical diagnosis was conducted using Metabolon’s SUB_PATHWAY designations. The GSEA for these metabolic pathways was performed by the GSEA function in the clusterProfiler (v4.4.4) R package [63]. For each comparison of clinical diagnoses, metabolites were ranked according to their t-statistic from the corresponding differential expression analysis. This allows the GSEA to condition the same set of covariates used in the differential expression analysis. Metabolic pathways were considered significantly up- or downregulated if they had an FDR p < 0.05.

To determine whether a subset of analytes was strongly predictive of the SNF clusters, we used multinomial logistic regression with elastic net regularization (glmnet; v4.1.4) [64] to predict clusters from a concatenated dataset consisting of the clinical and omics datasets. Ten repetitions of 10-fold cross-validation were performed to assess predictive accuracy and feature selection stability and to select the regularization parameters (α and λ) for the elastic net. For each of the 100 training splits of the dataset, the top quartile of highly variant analytes was selected for each omics dataset and used in SNF clustering, using the hyperparameters described above. The cluster labels for each test set were then assigned using label propagation on the SNF affinity matrix [65] from their corresponding training set. Finally, the cluster labels for each training/test set pair were permuted to maximize their agreement with the original cluster labels from the full dataset so that cluster labels were consistent. This ensured that no data leakage occurred from the testing sets due to the SNF clustering and that the repeated cross-validation estimates of performance and feature selection frequency were unbiased. The α and λ regularization parameters that resulted in the sparsest model within one SD of the model with the highest balanced accuracy were selected. Model predictive performance was also assessed by multiclass AUC, as implemented in the pROC package [51]. Finally, the strength of the predictive features was assessed in terms of selection frequency by elastic net across the 100 training splits of the data. The features that were selected by elastic net in ≥ 50 of 100 training splits were strongly predictive of the cluster labels and were used in the cluster causal analysis.

We applied CausalMGM [66], implemented in the package rCausalMGM (https://​github.​com/​tyler-lovelace1/​rCausalMGM), to the set of predictive features identified above and the clusters labels to identify a subset of predictive analytes that were potentially driving cluster designation. CausalMGM learns a probabilistic graphical model from observational data that hypothesizes the direction of causal interactions, based on observed conditional independence relationships. First, an initial undirected graph (skeleton) was constructed using mixed graphical models (MGM) [67]. Next, causal discovery in the presence of possible latent confounders was performed with FCI-Max [68]. When constructing the MGM, the regularization parameter λ that minimized the model’s Bayesian information criterion score was selected [69]. FCI-Max was then performed, using the MGM as an initial skeleton of adjacencies. The FCI-Max search algorithm was performed while controlling the FDR of the adjacencies [70] at FDR < 0.05. The stability of the resulting causal probabilistic graphical model was assessed by bootstrapping the MGM-FCI-Max procedure, described above, on 100 resampled datasets. The resulting causal graph was displayed using Cytoscape (v3.9.1) [71]. Edge thickness indicated the stability of each hypothesized adjacency in the causal model, while different edge types indicated different causal information, inferred by the FCI-Max algorithm.

Participants were selected from our database and biobank, as described in the “Methods” section. The four disease groups included (1) women diagnosed with PE with severe features (PE; n = 75); (2) women delivering a growth-restricted fetus (FGR; n = 40); (3) women with FGR newborns and diagnosed with a hypertensive disorder of pregnancy (FGR + HDP; n = 33); and (4) women with spontaneous PTD (sPTD; n = 72). Two groups served as controls: (1) women delivering at term without PE, FGR, or chronic pre-pregnancy medical conditions (n = 113); and (2) women who delivered preterm for reasons unrelated to PE, FGR, preterm labor/premature rupture of membranes, or other forms of placental dysfunction (n = 16). Key demographic and clinical characteristics of the study groups are summarized in Additional file 1: Table S2.

Using pairwise comparisons, we analyzed our placental multiomics data, including RNAs (both long and short RNA transcripts), proteins, and metabolites, to identify differentially regulated analytes among the six groups. As noted in Additional file 1: Table S3, the smallest number of differentially expressed (DE) analytes across all omics types vs controls was between the two control groups, and the largest number was between the FGR + HDP group vs each of the two control groups. Volcano plots for all DE analytes are shown in Fig. 1 and in Additional file 1: Figs. S2 and S3. Notably, Additional file 1: Fig. S4 shows that the two control groups shared many of the analytes when compared to the FGR + HDP group, supporting the distinct characteristics of this group when compared to the two control groups. Focusing on the DE analytes between FGR + HDP and the control groups (Fig. 1 and Additional file 1: Fig. S3), we noticed that several of the DE transcripts are known to play a role in PE, FGR, and placental dysfunction. These included enhanced expression of FMS-like tyrosine kinase 1 (FLT1) and endoglin (ENG) and reduced expression of placental growth factor (PGF), all characteristic changes of PE [72‐74]. Importantly, our proteomics analysis corroborated these findings (except for ENG, which was not included in our panels). Among the DE miRNAs, miR-210 was one of the most upregulated in the FGR + HDP group (4.4- and 2.7-fold change compared to the control-PT and control groups, respectively). Hierarchical clustering using the top 25 analytes of each molecular type performed well in separating the FGR + HDP from the control placentas (Fig. 1E–H) and even better in separating FGR + HDP from the control-PT placentas (Additional file 1: Fig. S3E-H). Analyte-based clustering on pairwise comparisons among other clinical diagnoses was less clear, pointing to potential overlapping molecular mechanisms that cause these conditions (Additional file 1: Fig. S5).

Applying Reactome canonical pathway analysis to our RNA results, we identified 42 and 47 significantly enriched pathways when comparing the FGR + HDP group to the term and preterm control groups, respectively (FDR < 0.05). Of the top-20 pathways, 12 were shared between the two comparisons. These included pathways related to interleukin and interferon signaling, protein processing, and glucose metabolism (Additional file 1: Fig. S6A, B). Most DE metabolites between the FGR + HDP group and the term control were lipids (31 upregulated and 6 downregulated) and amino acids (6 upregulated and 6 downregulated). Among individual metabolites, we identified many sphingolipid species to be upregulated in the FGR + HDP group. After grouping the metabolites by biochemical classes and applying pathway enrichment analysis, the sphingolipid pathways were again enriched (Additional file 1: Fig. S6C, D).

Realizing the inherent heterogeneity and overlaps among placental pathological pathways when segregated by clinical diseases, we deployed integrated omics tools for defining shared omics clusters, irrespective of the clinical diagnosis. For this goal, we employed SNF [31], a machine-learning method that combines diverse high-throughput data sources into a single similarity network used for clustering. SNF divided our cohort into four clusters, which were mostly dominated by a specific clinical syndrome (Fig. 2A): Cluster I, the biggest cluster, was dominated by the control placentas, Cluster II by sPTD, Cluster III by FGR + HDP and severe PE placentas, and Cluster IV, mainly comprised of term controls and FGR placentas (Fig. 2B). The control, control-PT, and FGR + HDP groups showed the most homogenous distribution across the clusters, with 75% of term control placentas allocated to Cluster I, 85% of the control-PT to Cluster II, and 81.5% of FGR + HDP allocated to Cluster III (Fig. 2B). Other clinical conditions presented higher heterogeneity. A univariate analysis identified significant differences in several important clinical variables across the clusters (Additional file 1: Table S4), with the earliest gestational age at delivery and lowest birth weight in Cluster III, followed by Clusters II, and with no difference between Clusters I and IV. Moreover, the number of early-onset PE cases, likely representing a greater placental involvement, was significantly higher in Cluster III (FDR < 0.001, Fig. 2C, and Additional file 1: Table S4).

To gain a deeper insight into the distribution of participants with PE across different SNF clusters, we examined the clinical characteristics that defined these participants: 56% of the PE cases in Cluster III were early-onset PE, whereas only 6%, 13%, and 40% of PE were diagnosed before the 34th week of gestation in Clusters I, II, and IV, respectively (Fig. 2D, and Additional file 1: Table S4). Participants with PE in Cluster III also delivered the smallest babies, at a mean birth weight of 1600 g (Additional file 1: Table S4). This is predictably consistent with the observation that 64% of all preterm deliveries (in any clinical syndrome) were found in Cluster III (Additional file 1: Table S4). Notably, considering only sPTD cases without other clinical syndromes, cases allocated to Cluster II delivered at the earliest gestational age and were associated with the smallest newborns (32 + 4 days and 1970 g). We conducted a comprehensive chart review and validated the presence of severe features in 67 of the 75 women in our severe PE group. Given the retrospective nature of our study and the possibility of missing information, we adhered to the clinical team’s determination and kept all 75 participants in the analysis. Six of the “misclassified” participants were part of the SNF analysis; interestingly, all were assigned to Clusters I and II. Notably, three participants in our FGR + HDP group were misclassified as having severe features, all allocated to Cluster III. Together, our SNF-based integration of placental analyte data identified omics-defined pregnancy subgroups that might have shared clinical diagnoses yet exhibit different outcomes.

To identify the molecular drivers of the different SNF clusters, we measured the area under the receiver-operator characteristic curve (AUROC) of each analyte’s ability to distinguish between the clusters. Figure 3 depicts the expression levels of the ten molecules with the highest AUROC in each data type for each cluster. The most striking finding was the involvement of PE/FGR-related analytes in Cluster III, where ENG, leptin, FLT1, and FSTL3 were among the top 10 RNAs, with the latter three also in the proteins’ top list (Fig. 3A, C). Several sphingolipids, miR-210, and miR-193 were among the highest analytes in Cluster III that were previously identified to be involved in the pathogenesis of PE. We also explored the expression level of known placental dysfunction analytes across all clusters. Cluster III had the highest expression of these analytes, followed by Cluster IV, while Cluster II had the lowest expression levels (Fig. 4A). Interestingly, most of the DE markers of placental dysfunction (27/38) exhibited an opposite expression pattern between Clusters II and III. We found a similar expression pattern when focusing only on placentas from women with PE (Fig. 4B). Clusters II and III showed a negative expression correlation throughout all omics datatypes, most robustly in the protein and RNA datasets (R² =  − 0.57; p < 0.001, and − 0.47; p < 0.001, respectively, Additional file 1: Fig. S7A). We also identified a negative correlation between Clusters I and IV which comprised most of the FGR placentas (Additional File 1: Fig. S7B and Fig. 2B). These data highlight the different molecular signatures that define the clusters and suggest discrete pathophysiological processes that underlie each cluster.

We applied BayesPrism, a deconvolution method, to evaluate the proportions of different cell types in our placental biopsies and their contribution to the clinical disease groups or to the SNF clusters. As expected, the primary cell type in our biopsies was syncytiotrophoblast, followed by cytotrophoblasts and fibroblasts (Additional file 1: Fig. S8). We found small, yet significant differences in the prevalence of the different cell types across the clinical diagnoses. In contrast, the differences in cell type representation were marked among the multiomics-defined clusters all detailed in Additional file 1: Fig. S8B, suggesting a contribution of cell type representation to the differences detected using multiomics.

Placental histopathology is commonly used to validate obstetrical diagnoses and is considered the gold standard in identifying placental injuries. We therefore examined the correlation of placental histopathology with clinical syndromes or SNF clusters. We found that MVM was significantly more common in the FGR + HDP group than in the sPTD, FGR, and term control groups (Fig. 5, p < 0.0001). The severe PE group was only different from the term control group. Considering our multiomic-defined clusters, we found a markedly higher rate of MVM in Cluster III compared to each of the other clusters (81.3% vs 37.6%, 46.4%, and 38.5% in Clusters I, II, and IV, respectively, FDR < 0.001). The rates of FVM, AI, and CI pathological diagnoses were similar across the clinical syndromes. There was also a higher rate of CI in Cluster III when compared to Cluster II (25% vs. 3.6%, FDR < 0.05). The rates of FVM and AI were similar across the clusters.

Because MVM was more discriminatory among the clinical syndromes and the SNF clusters, we further assessed paraffin-embedded placental biopsies taken from the same site as the snap-frozen specimens used for analyte measurements (n = 348). These biopsies were analyzed by a perinatal pathologist who was blinded to the clinical outcomes. We found that accelerated villous maturation (AVM) and syncytial knots clearly distinguished Cluster III from all the other clusters. Distal villous hypoplasia (DVH) was also higher in Cluster III compared to Clusters II and IV. The pathological lesions were again less distinctive of a specific clinical syndrome (Additional file 1: Table S5, Fig. 5). Given these observations, we performed Bayesian model selection to compare the association of disease groups or the SNF-defined clusters with placental histopathology. Specifically, we deployed logistic regression for the four main placental pathological injury patterns (MVM, FVM, AI, CI) and specifically, to the three MVM lesions (DVH, AVM, syncytial knots), using the disease group or the SNF cluster as the independent variable. Following the work of [75], we compared the Bayesian Information Criterion (BIC) scores of the SNF cluster models with either the 6-category disease group or the 4-category disease group models. Most histopathology findings correlated better with the SNF cluster model than with either disease group model. CI, DVH, and syncytial knots showed the most substantial difference. FVM, AI, and AVM showed only weak evidence (BIC score difference < 2, Fig. 6). Taken together, histopathological diagnoses and lesions, primarily those indicative of placental dysfunction, separated Cluster III from all other clusters, and correlated better with multiomics-based SNF clusters than with the disease groups.

To gain further insight into biological pathways that might underlie the SNF clusters, we performed gene set enrichment analysis (GSEA) in Reactome Knowledgebase [61], using the transcriptomics datasets, comparing each cluster to the other three. Cluster I presented the most diverse pathway domains but no biological process was dominant. Mitochondria-related pathways were identified in Cluster II. Cluster III enriched pathways were dominated by immune-related processes, led by interferon α/β signaling. Cluster IV also showed an immune-related signature, partly shared with Cluster III, as well as platelet-related pathways (Fig. 7).

We performed multinomial logistic regression with elastic net regularization, using the omics and clinical characteristics to identify a subset of features that are strongly predictive of the SNF clusters. We have selected the analytes chosen in at least 50 out of the 100 models trained across ten repetitions of 10-fold cross-validation as being highly predictive of the cluster labels (Fig. 8A). Analytes that previously contributed to Cluster III, including FSTL3, miR-210, miR-193, and miR-365, were predictive of the cluster labels. Of the clinical variables, only the gestational age was predictive of the cluster labels. Additionally, the predictive accuracy of these multinomial logistic regression models, as measured by the multiclass AUROC (0.966) and the AUROC for each cluster versus the rest (I = 0.949; II = 0.962; III = 0.954; IV = 0.971) (Additional file 1: Fig. S9), demonstrated that these clusters were highly separable even using a small molecular analyte subset to define the clusters. We used this subset of analytes to predict cluster assignment for placentas excluded from the initial analysis due to their missing at least one omics datatype. The clinical characteristics of these predicted clusters did not significantly differ from the clusters generated by SNF, based on all four analyte types. Of note, the participants that were excluded from the initial analysis and now predicted to belong to Cluster III, delivered earlier, and had a higher proportion of early-onset PE. Next, we used the set of 47 predictive analytes to conduct a causal discovery algorithm to identify a subset that might drive the SNF clusters. As depicted in Fig. 8B, based on the variables marked by a red square, gestational age was the only clinical variable found to be directly associated with the cluster labels, along with FSTL3 protein and several RNA transcripts. Using a similar analysis, we identified the most predictive features separately for each cluster (Fig. 8C and Additional file 1: Fig. S10). Notably, no analytes were found to be predictive of Cluster IV, which is the most heterogeneous and supported by the smallest number of placentas.

Lastly, to evaluate the feasibility of using our findings for plasma-based prediction of SNF clusters, we identified the two differentially expressed genes for each SNF cluster (see the “Methods” section). We found that the expression patterns of the majority (five of eight) of the maternal plasma mRNAs exhibited a pattern similar to that found in the placenta (Additional file 1: Fig. S11).