Introduction

The public health burden of adverse pregnancy outcomes is substantial. An important example is preeclampsia, which affected 6.5% of all live births in the United States in 2017 and is characterized by high maternal blood pressure, proteinuria, and damage to other organ systems1. Adverse pregnancy outcomes may lead to myriad health complications including an elevated risk of chronic diseases throughout the life course2. The placenta, a temporary organ that develops early in pregnancy, promotes maternal uterine artery remodeling; mediates transport of oxygen, nutrients, and waste3; secretes hormones to regulate pregnancy; metabolizes various macromolecules and xenobiotics; and can serve as a selective barrier to some, but not all, pathogens and xenobiotics4. The executive summary of the Placental Origins of Adverse Pregnancy Outcomes: Potential Molecular Targets workshop recently concluded that most adverse pregnancy outcomes are rooted in placental dysfunction5. Despite this, the molecular underpinnings of placental dysfunction are poorly understood.

Placenta-specific cell types including cytotrophoblasts, syncytiotrophoblasts, extravillous trophoblasts, and placental resident macrophage Hofbauer cells are all essential for placental development, structure, and function6. Dysfunction of these specific cell types likely plays a role in placental pathogenesis. For example, extravillous trophoblasts are responsible for invading into the maternal decidua early in pregnancy to remodel uterine arteries and increase blood flow to the placenta3. Inadequate or inappropriate invasion of extravillous trophoblasts has previously been implicated in preeclampsia etiology7,8,9. Despite some knowledge of the roles of specific placental cell types in the development of preeclampsia, relatively little is known about how individual cell types contribute to placental dysfunction.

Existing research models used to investigate the function and dysfunction of individual cell types are limited. Protocols to isolate primary placental cells for experimental research are restricted to one or few cell types10,11,12,13,14,15. Cell type-specific assays are costly and require special techniques or training resulting in small sample sizes and have not yet been scalable to large epidemiological studies16,17,18. Furthermore, placental cell lines such as BeWo, derived from choriocarcinoma19, and HTR-8/SVneo, immortalized by SV4020, are typically derived by processes that alter the DNA of the cells, limiting their in vivo translatability. Consequently, the characteristics of even healthy placental cell type function and especially their connections to adverse outcomes such as preeclampsia are incompletely understood.

Measures of gene expression in bulk placental tissue are used to better understand the biological mechanisms underlying adverse pregnancy outcomes21,22,23 and are common in epidemiological studies24. Gene expression profiles differ systematically by cell type25,26. Thus, bulk placental tissue-level gene expression measurements represent a convolution of gene expression signals from individual cells and cell types27,28. Deconvolution refers to the bioinformatic process of estimating the distribution of cell types that constitute the tissue29,30. Deconvoluting tissue-level gene expression profiles is essential to account for effects introduced by unmodeled cell type proportions31 by disentangling shifts in cell type proportions from direct changes to cellular gene expression32. Reference-based deconvolution boasts biologically interpretable cell type proportion estimates with few modeling assumptions but relies on independently collected cell type-specific gene expression profiles as inputs32. Prior placental cell type-specific gene expression measures from term villous tissue16,17 had a limited number of biological replicates and included neither technical replicates nor benchmarking against physically isolated placental cell types. A robust, accessible, and publicly available gene expression deconvolution reference is currently unavailable for healthy placental villous tissue.

To advance the field of perinatal molecular epidemiology, our goal was to develop an accessible and robust gene expression deconvolution reference for healthy placental villous tissue at term. We generated single-cell RNA-sequencing data with technical replicates for integration with existing cell type-specific placental gene expression data16,17. In addition, we benchmarked these single-cell cell type-specific gene expression profiles against placental cell types isolated with more conventional fluorescence-activated cell sorting (FACS) followed by bulk RNA-sequencing. Finally, to assess links between preeclampsia and placental cell types and their proportions, we applied our placenta cell type gene expression reference to deconvolute bulk placental tissues in a secondary data analysis of a case–control study33 of preeclampsia, including a mediation analysis of the preeclampsia-associated genes FLT1, LEP, and ENG that quantifies the role cellular composition plays in explaining bulk gene expression measures.

Results

Single-cell gene expression map of healthy placental villous tissue

A conceptual layout of the laboratory methods and analyses contained within this manuscript is provided in Supplementary Fig. 1. From healthy term placental villous tissue, 9244 cells across a total of two biological replicates and two technical replicates were sequenced and analyzed (Michigan sample). These data were combined with single-cell RNA-sequencing data of 5911 cells from three healthy term villous tissue samples in a previously published study (Pique-Regi sample)17 and 25,339 cells from four healthy term villous tissue samples in another previously published study, two of which were subsampled with an additional peripheral placental villous tissue sample (Tsang sample)16 (Supplementary Table 1). Cells were excluded if they had low RNA content (<500 unique RNA molecules), few genes detected (<200), or were doublets or outliers in mitochondrial gene expression (Supplementary Figs. 2 and 3). Fetal or maternal origin of cells was determined by genetic variation in sequencing data. Fetal sex was determined by XIST expression (Supplementary Fig. 4). The final analytic sample included 40,494 cells and 36,601 genes across nine biological replicates, two of which had a technical replicate and another two included peripheral subsampling.

Uniform Manifold Approximation and Projection (UMAP)34 was used to visualize sequencing results in two dimensions with mutual nearest neighbor batch correction35 (Fig. 1a). Cells clustered into 19 fetal and 8 maternal cell types with 84.4% of all cells being of fetal origin (Table 1). Cell type clustering decisions balanced cluster stability, resolution, and biologic plausibility with prior knowledge. If desired, downstream analyses could collapse cell subtypes into a single, more general cell type cluster. We observed placenta-specific trophoblast cell types including cytotrophoblasts (KRT7), proliferative cytotrophoblasts (KRT7, STMN1 and other proliferation-related genes)36, extravillous trophoblasts (HLA-G)37, and syncytiotrophoblasts (PSG4 and other pregnancy-specific hormone genes) (Supplementary Fig. 5a)38. Proliferative cytotrophoblasts were distinguished from other cytotrophoblasts by overexpression of genes related to cytoplasmic translation (padj = 8.1 × 10−15) and mitotic sister chromatin segregation (padj = 1.5 × 10−12), indicative of their proliferative phenotype (Supplementary Fig. 6). Other fetal-specific cell types included mesenchymal stem cells (COL1A1lo, TAGLNlo, LUMhi), fibroblasts (COL1A1hi, TAGLNhi, LUMlo)39, endothelial cells (PECAM1)40, and Hofbauer cells (CD163)11 (Fig. 1b).

Fig. 1: Integrated single-cell gene expression map of healthy placental villous tissue.
figure 1

a Uniform Manifold Approximation and Projection (UMAP) plot of all cells (n = 40,494), with each cell colored by cell type cluster. b UMAP plot of fetal cells only (n = 34,165), with each cell colored by cell type cluster. c UMAP plot of maternal cells only (n = 6329), with each cell colored by cell type cluster.

Full size image
Table 1 Number of cells captured by single-cell RNA-sequencing in the final analytic dataset for each cell type by sample source.
Full size table

Fetal and maternal lymphocytes, B cells, and monocytes were also captured (Fig. 1b, c). We observed fetal and maternal B cells (CD79A)41 and maternal plasma cells (XBP1, IGHA and other immunoglobulins)42. We also observed fetal and maternal CD14+ monocytes (CD14+/FCGR3A−), maternal FCGR3A+ monocytes (CD14+/FCGR3A+)43, and a small population of fetal plasmacytoid dendritic-like cells (FLT3+/ITM2C+)44,45. We further observed fetal and maternal natural killer cells (NKG7), fetal GZMB+ or GZMK+ natural killer cell subtypes, and fetal natural killer T cells (NKG7+/CD3E+/CD8A-)46,47. Finally, we observed a variety of T cell subtypes: naïve CD4+ (CCR7, CD3E, CD4), naïve CD8+ (CCR7, CD3E, CD8A), memory CD4+ (S100A4, CD3E, CD4, IL2, CCR7lo), and activated CD8+ T cells (NKG7, CD3E, CD8A) (Supplementary Fig. 5b)48.

To identify upregulated genes in each cell type, we compared the expression of a gene in one cell type against that gene’s average expression in all other cell types (Supplementary Data 2). Consequently, the same genes could be upregulated across several cell types of a similar lineage. FLT1 expression was highly upregulated in extravillous trophoblasts (log2 fold-change (FC) = 3.89, padj < 0.001). Trophoblast cell types had the largest and most diverse transcriptomes, characterized by the largest number of unique RNA transcripts and detected genes per cell (Supplementary Fig. 7). Functional analysis of upregulated genes revealed cell type-specific biological processes (Supplementary Data 3). For example, fetal extravillous trophoblasts were enriched for genes relevant to placental structure and function such as cell migration (padj < 0.001) and response to oxygen levels (padj < 0.001) and syncytiotrophoblasts were enriched for genes involved in steroid hormone biosynthetic process (padj < 0.001). Technical replication in Michigan samples 1 and 2 appeared high in UMAP space (Supplementary Fig. 8a, b). Indeed, the average intra-cluster gene expression between technical replicates had an average Spearman correlation (mean ± standard deviation) of 0.94 ± 0.14 for sample 1 and 0.88 ± 0.20 for sample 2 (p values < 0.001).

Single-cell RNA-sequencing deconvolution reference exhibits excellent in silico performance

Based on the single-cell data, we created a placental signature gene matrix that incorporated expression information across an algorithmically selected 5229 signature genes to estimate the cellular composition of 27 fetal and maternal cell types from whole tissue gene expression data (Supplementary Fig. 9). To test the performance and robustness of this placental single-cell RNA-sequencing deconvolution reference, we randomly split our analytic single-cell RNA-sequencing dataset into 50% training and 50% testing subsets with balanced cell type proportions49. The same training dataset was used for each comparison; test mixtures were generated from the testing half of the dataset. Using a signature gene expression matrix generated from the training data, we estimated cell type composition in in silico pseudo-bulk testing data mixtures of known cell type composition with varying contributions of fetal vs. maternal origin cells and male vs. female fetal cells (Fig. 2). In all mixtures, the 27 predicted and actual cell type proportions were correlated (p value < 0.001 for each test). In the primary deconvolution analysis of all cell types at their natural rates (n = 20,242), estimated and actual cell type proportions had a Pearson correlation coefficient of 0.956 (95% CI [0.904, 0.980]). The worst performance was under the unrealistic scenario that the mixture was composed entirely of maternal cell types (n = 3162) with a Pearson correlation of 0.734 (95% CI [0.491, 0.871]) between actual estimated cell type proportions. Our deconvolution reference was also robust to fetal sex when only male fetal cells (Pearson correlation = 0.893, 95% CI [0.776, 0.950]) were included (n = 8394), or only female fetal cells (Pearson correlation = 0.983, 95% CI [0.964, 0.993]) (n = 8394). Together, these results show that our reference panel can successfully deconvolute placental tissues, though some maternal cell types common to both mother and fetus may be erroneously labeled fetal in the absence of fetal cells of those cell types.

Fig. 2: In silico placental deconvolution testing.
figure 2

Scatter plots summarizing the performance of our single-cell deconvolution reference using CIBERSORTx with in silico mixtures of single-cell libraries from a 50/50 training/test split of the integrated single-cell RNA-seq dataset (n = 40,494). The same training dataset was used for each comparison; test mixtures were generated from the testing half of the dataset. Predicted deconvoluted cell type proportions for each of the 27 cell types are encoded on the x-axis. Actual cell type proportions from the test dataset are encoded on the y-axis. Correlation coefficients and root mean square error measures are presented for each comparison. A linear line of best fit overlays the results. The gray shaded area represents the 95% confidence intervals around the simple linear regression estimates. a The test mixture is the test half of the single-cell dataset (n = 20,242). b The test mixture sampled only fetal cells (n = 17,080). c The test mixture sampled only maternal cells (n = 3162). d The test mixture sampled only female fetal cells (n = 8394). e The test mixture sampled only male fetal cells (n = 8394).

Full size image

Fluorescence-activated cell sorting of major placental cell types yielded mixed cell type isolation results

We isolated whole bulk placental villous tissue, enriched syncytiotrophoblasts, and sorted five cell types (Hofbauer cells, endothelial cells, fibroblasts, leukocytes, extravillous trophoblasts, and cytotrophoblasts) via FACS from four healthy term, uncomplicated Cesarean sections for bulk RNA-sequencing, labeled Sorted 1 (same sample source as single-cell RNA-sequencing sample 1), Sorted 2, Sorted 3, and Sorted 4 (Supplementary Fig. 10 and Supplementary Table 2). For analysis, as recommended50, we excluded 19,048 genes that were not present in at least 3 samples and an additional 865 genes that did not have a cumulative library size-normalized count of at least 10. Principal component (PC) analysis of whole-transcriptome sorted-cell bulk RNA-sequencing normalized counts is provided in Supplementary Fig. 11.

To identify upregulated genes in each cell type, we compared the expression of a gene in one cell type against that gene’s average expression in all other cell types (Supplementary Fig. 12). Consequently, the same genes could be upregulated across several cell types of a similar lineage. All 38,468 uniquely mapping genes were tested. A total of 746 genes were algorithmically dropped from the syncytiotrophoblast contrast due to excessively low counts, low variability, or extreme outlier status. Large-scale gene expression differences were observed for each cell type (Supplementary Data 4). Functional analysis of upregulated genes revealed cell type-specific biological processes (Supplementary Data 5). For example, syncytiotrophoblasts were enriched for genes relevant to placental structure and function such as angiogenesis, cell-substrate adhesion, and regulation of epithelial cell proliferation (padj < 0.001). To compare sorted and single-cell differential expression and enrichment results, we tabulated the number of unique genes and pathways overlapping between the two analyses after collapsing the single-cell cell type cluster labels to the seven cell type fractions that we had targeted to isolate for downstream analyses (Supplementary Table 3). On average, 15.0% of single-cell upregulated genes and 5.9% of enriched pathways were also identified among the sorted-cell data. On average, 17.5% of sorted cell type upregulated genes and 39.2% of pathways were also identified among the single-cell data. Sorted endothelial cell results were limited due to the limited number of biological replicates.

We applied the single-cell deconvolution reference to estimate cell proportions in the 4 whole tissue (with 1 additional technical replicate) and 19 sorted or enriched cell type fractions. We collapsed the single-cell cell type cluster labels to the seven cell type fractions we targeted for isolation for downstream analyses (Supplementary Data 6, Sheet 1). All deconvoluted samples exhibited high goodness-of-fit between original bulk mixtures and the estimated cell type proportion mixtures (p values < 0.001). Among the signature genes, original bulk and estimated cell type fractions had a Pearson correlation (mean ± standard deviation) of 0.73 ± 0.11 and root mean square error of 0.88 ± 0.04 (Supplementary Data 6, Sheet 2). Deconvolution results (mean ± standard deviation) suggest we successfully isolated fibroblast- (n = 3, 74.7% ± 0.6%) and leukocyte-enriched (n = 4, 82.3% ± 24.8%) cell type fractions. Other cell type targets were less successful (range 0–26% estimated purity). The Hofbauer cell fraction was predicted to be mostly leukocytes (n = 4, 91.5% ± 0.5%).

Cell proportion deconvolution of bulk placental tissue preeclampsia dataset

We applied the single-cell deconvolution reference to estimate cell proportions from bulk placental tissue in 157 preeclampsia cases and 173 controls33 compiled from eight previously published studies33,51,52,53,54,55,56,57. Mean gestational age was 2.2 weeks younger in cases than controls (p value < 0.001, Table 2). All deconvoluted samples exhibited high goodness-of-fit between original bulk mixtures and the estimated cell type proportion mixtures (p values < 0.001). Among the signature genes, original bulk and estimated mixtures had a Pearson correlation (mean ± standard deviation) of 0.70 ± 0.04 and root mean square error of 0.73 ± 0.03 (Supplementary Data 7). Fetal naïve CD4+ T cells and fetal GZMB+ natural killer cells were estimated to be at 0% abundance in all samples and were dropped from downstream analyses. Cytotrophoblasts were the most abundant (mean ± standard deviation) estimated fetal cell type (27.9% ± 4.3%) followed by syncytiotrophoblasts (23.4% ± 5.0%) and mesenchymal stem cells (10.3% ± 3.3%). The most common maternal cell types were naïve CD8+ T cells (2.8% ± 1.5%), plasma cells (1.4% ± 0.7%), and B cells (1.3% ± 0.8%). A comparison of deconvoluted whole tissue cell type proportions among healthy individuals (Supplementary Fig. 13) between the microarray dataset GSE75010 (n = 173 controls), our whole tissue bulk RNA-sequencing samples (sorted samples 1–4), and the single-cell dataset compiled here (single-cell samples 1–9) suggests that syncytiotrophoblasts and endothelial cells are underrepresented in the single-cell data. This is likely due to dissociation bias, which has been commonly observed in single-cell assays of other tissues58. Overall, the Pearson correlation of the average deconvoluted cell type proportion across the 27 cell types between healthy bulk RNA-sequencing and microarray controls was 0.80 (95% CI: [0.60, 0.91]).

Table 2 Demographic characteristics of eight previously published bulk microarray placental gene expression case–control studies (accessed through GSE75010) for deconvolution application testing.
Full size table

Differentially abundant cell type proportions in preeclampsia cases versus controls

To test for differences in cell proportions between preeclampsia cases and controls (Supplementary Fig. 14), we fit beta regression models for each cell type proportion adjusted for study source, fetal sex, and gestational age to estimate the prevalence odds ratio for each cell type (Supplementary Data 8). Among fetal cell types, extravillous trophoblasts (p < 0.001), memory CD4+ T cells (p = 0.007), CD8+ activated T cells (p = 0.005), and natural killer T cells (p = 0.006) were more abundant (Fig. 3) in preeclampsia cases relative to controls. The unadjusted median extravillous trophoblast abundance was 6.4% among cases compared to 2.1% among controls. Mesenchymal stem cells (median percent composition in cases vs. controls, 8.8% vs. 11.0%), Hofbauer cells (2.7% vs. 4.4%), and fetal naive CD8+ T Cells (4.2% vs. 4.5%) were all less abundant among preeclampsia cases compared to controls (p < 0.001). Among maternal cell types, maternal plasma cells (1.6% vs. 1.2%) were more abundant among preeclampsia cases compared to controls (p < 0.001).

Fig. 3: Preeclampsia case–control status and cell type proportion differential abundance analysis.
figure 3

Forest plot of multivariate beta regression models’ prevalence odds ratio estimates adjusted for study source, gestational age, and fetal sex tested for a difference in each cell type’s proportions in cases versus controls (n = 157 cases, 173 controls). Horizontal lines indicate the range of the 95% confidence interval.

Full size image

Differential expression between preeclampsia cases and controls attenuated by cell type proportion adjustment

To test whether microarray gene expression differences between preeclampsia cases and controls are partly driven by differences in cell type abundances, we fit linear differential gene expression models adjusted for covariates study source, fetal sex, and gestational age with and without adjustment for deconvoluted cell type proportions. To reduce the number of model covariates and account for dependence between deconvoluted cell type proportions, we applied PC analysis to deconvoluted cell type proportions. The first five PCs accounted for 87.2% of the variance in deconvoluted cell type proportions and were added as additional covariates to form the cell type-adjusted model. Variation in PCs 1 and 2 was largely driven by syncytiotrophoblasts (33.8%), extravillous trophoblasts (33.5%), and cytotrophoblasts (15.3%) proportions and provided some separation between cases from controls (Supplementary Fig. 15a, c). Variation in PC3 was largely driven by cytotrophoblasts (50.1%) and to a lesser extent syncytiotrophoblasts (16.6%), mesenchymal stems cells (14.5%), and extravillous trophoblasts (13.7%) (Supplementary Fig. 15b, d).

In the cell type-naïve base models (n = 14,651 genes, 173 controls, and 157 cases) adjusted for study source, gestational age, and fetal sex, 550 genes were differentially upregulated and 604 were downregulated in preeclampsia cases versus controls (Fig. 4a and Supplementary Data 9). Gene set enrichment analysis of biological processes identified 41 overrepresented pathways in the base model (Fig. 5a and Supplementary Data 10). Biological process pathways such as aerobic respiration (false discovery adjusted q < 0.001), mitochondrial respiratory chain complex assembly (q < 0.001), glutathione metabolism (q = 0.003), and ribosome biogenesis (q = 0.001) were overrepresented among downregulated genes. No pathways were overrepresented among upregulated genes, though intermediate filament organization (q = 0.26), keratinocyte differentiation (q = 0.77), and endothelial cell development (q = 0.42) had comparable enrichment scores. Remarkably, when the base model was additionally adjusted for the first five PCs of imputed cell type proportions, there were zero differentially expressed genes between preeclampsia cases and controls (Fig. 4b and Supplementary Data 9). Of the cell type-adjusted results, 19 pathways were overrepresented (Fig. 5b and Supplementary Data 10). Downregulation of mitochondrial respiratory chain complex assembly (q < 0.001), aerobic respiration (q = 0.001), ribosome biogenesis (q = 0.001), and glutathione metabolism (q = 0.02) were overrepresented among downregulated genes. Detection of chemical stimulus involved in sensory perception of smell (q = 0.04) and non-coding RNA processing (q = 0.04) were also overrepresented pathways among downregulated genes. Neuroepithelial cell differentiation (q = 0.04) was overrepresented among upregulated genes. Vascular endothelial growth factor receptor signaling pathway (q = 0.15), of which FLT1 is a member, had an enrichment score of 1.77 (up from 1.34, q = 0.43 in the base model) but did not meet the q value cutoff. Overall, downregulation of mitochondrial biogenesis, aerobic respiration, and ribosome biogenesis and related pathways were robust to cell type proportion adjustment.

Fig. 4: Preeclampsia case–control differential expression analysis.
figure 4

Volcano plots comparing differentially expressed genes in samples from 153 preeclampsia cases versus 173 healthy controls across two models: a the base model adjusted for covariates fetal sex, study source, and gestational age and b the model adjusted for fetal sex, study source, and gestational age and additionally adjusted for the first five principal components of estimated cell type proportions. Dotted line represents a false discovery rate-adjusted q value of 0.05. FLT1, LEP, and ENG are labeled as genes of interest in preeclampsia.

Full size image
Fig. 5: Preeclampsia case–control differential expression enrichment analysis.
figure 5

Top Gene Set Enrichment Analysis pathways from the Gene Ontology: Biological Processes database results for the differential expression analysis by preeclampsia case–control status. Results arranged by descending magnitude of the absolute value of the normalized enrichment score. Pathways colored red are significant at a false discovery rate-adjusted (FDR) q value of 0.05 whereas pathways in blue are statistically insignificant. a Top pathways from the cell type-unadjusted analysis. b Top pathways from the cell type-adjusted analysis.

Full size image

Differential expression of preeclampsia-associated genes mediated by placental cell type proportions

Overexpression of FLT1 in placental tissue59,60,61,62, detection of a soluble isoform of FLT1 in maternal circulation63,64, and fetal genetic variants near FLT165 have implicated FLT1 in preeclampsia etiology. Because we observed cell type-specific expression patterns of FLT1 in trophoblasts, particularly in extravillous trophoblasts, we hypothesized that the observed attenuation of FLT1 differential expression may be due in part to the differences in cell type proportions observed between preeclampsia cases and controls. To test this hypothesis, we applied a unified mediation and interaction analysis to quantify the proportion of FLT1 expression differences mediated by deconvoluted cell type proportions. We did not observe an interaction between preeclampsia status and cellular composition (overall proportion attributable to interaction = −5.8%, 95% CI [−17.1%, 5.0%]). We therefore dropped interaction parameters from the model for the final analysis. In the model without interaction, 37.8% (95% CI [27.5%, 48.8%]) of the 1.05 (95% CI [0.89, 1.21]) log2 signal intensity increase in the association between preeclampsia and FLT1 expression was attributable to differences in placental cell composition between preeclampsia cases and controls (Fig. 6). Overexpression of LEP and ENG have also been associated with preeclampsia59,60,61,62. Mediation results were similar for LEP (total effect = 2.62 (95% CI [2.26, 2.97] log2 signal intensity increase; proportion mediated = 34.5% (95% CI [26.0%, 44.9%]) and ENG (total effect = 0.93 (95% CI [0.79, 1.07] log2 signal intensity increase; proportion mediated = 34.5% (95% CI [25.0%, 45.3%]).

Fig. 6: Placental cell composition as a mediator of FLT1 expression.
figure 6

Mediation of FLT1 gene expression by placental cell type composition (n = 157 cases, 173 controls). Placental cell composition was operationalized as first five principal components of estimated cell type proportions. 95% confidence intervals are provided after effect estimates for each model parameter. The same framework was also applied with LEP or ENG expression as the outcome.

Full size image

Discussion

To create the largest publicly available placental RNA deconvolution reference of 19 fetal and 8 maternal cell type-specific gene expression profiles, we newly sequenced placental villous cells, integrated those results with data from previously published studies, and built a signature gene matrix for deconvolution of bulk villous tissue gene expression data. In silico testing of our deconvolution reference demonstrated successful and robust deconvolution. To compare single-cell placental cell type expression profiles to more conventional sorting methods, we created a FACS scheme to enrich and sequence RNA from five important placental cell types as well as syncytiotrophoblasts. Deconvolution of sorted cell type fractions with the single-cell deconvolution reference suggested most conventionally sorted cell types are far less pure than what can be accomplished with clustering and aggregation of single-cell results, and at much lower cell type resolution. We applied the single-cell deconvolution reference to estimate cell type proportions in a previously published epidemiologic microarray study of the pregnancy complication preeclampsia, revealing placental cell type proportion differences between preeclampsia cases and controls at term. We then showed that large gene expression differences between preeclampsia cases and controls were markedly attenuated after adjustment for cell type proportions. Preeclampsia-associated pathways, including downregulation of mitochondrial biogenesis, aerobic respiration, and ribosome biogenesis were robust to cell type adjustment, suggesting direct changes to these pathways. Finally, to quantify the attenuation of differential expression of the preeclampsia biomarkers FLT1, LEP, and ENG, we applied mediation analysis to show cellular composition mediated a substantial proportion of the association between preeclampsia and FLT1, LEP, and ENG overexpression. Cell type proportions may be an important and often overlooked factor in gene expression differences in placental tissue studies.

By integrating our new single-cell RNA-sequencing results with those from a previously published study, our integrated dataset, to our knowledge, is the largest and possibly only reference available for healthy, term placental villous tissue to date. We document term cell type-specific gene expression patterns for well-characterized placental cell types, including syncytiotrophoblasts10, cytotrophoblasts13, and extravillous trophoblasts14. In addition, we provide gene expression markers for relatively understudied placental cell types such as endothelial cells, mesenchymal stem cells, and Hofbauer cells as well as maternal peripheral mononuclear cells recovered from the maternal-fetal interface. Compared to the previous analysis of the published samples17 which relied on predominately sex-specific gene expression markers to differentiate proliferative from non-proliferative cytotrophoblasts, we show that functional enrichment analysis revealed broad upregulation of proliferation pathways in proliferative cytotrophoblasts. The low representation of some cell types such as trophoblasts in our single-cell RNA-sequencing results from the Michigan study suggests that these cell types may be especially sensitive to dissociation and disintegrate before transcript capture, commonly referred to as dissociation bias58. Michigan samples 1 and 2 also included a cryopreservation step like those employed in large-scale epidemiological studies that may have exacerbated dissociation bias66; this applies to both single-cell and sorted cell type experiments. Future studies may propose alternative approaches to perform unbiased single-cell RNA-sequencing in placental tissues; indeed, single-nucleus RNA-sequencing has been used to characterize an in vitro syncytiotrophoblast model and may be more appropriate to assay such cell types sensitive to dissociation procedures67. We verified that our deconvolution reference exhibited strong performance even with extremely imbalanced and unlikely real-world test mixture distributions by fetal sex and maternal cell type representation.

Our preeclampsia findings are consistent with a prior pathophysiological understanding of the disorder, linking cell type proportion estimates and gene expression data in bulk tissue. Among preeclampsia cases, we observed an elevated proportion of extravillous trophoblasts and underrepresentation of stromal cell types, which may reflect an arrest in conventional placental cell type differentiation and maturation following insufficient uterine spiral artery remodeling implicated in preeclampsia68,69,70. A recent study of bulk placental gene expression across trimesters suggests that Hofbauer cells may more abundant in the second trimester compared to the third, possibly to support vasculogenesis, though this study involved a small deconvolution reference that contained a limited variety of cell types71. A better understanding of the evolution of temporal placental composition changes may yield greater insight into placental pathologies. In the cell type-naïve differential expression model, consistent with previous findings, placentas from pregnancies with preeclampsia overexpressed FLT1, LEP, and ENG59,60,61,62. In our cell type-adjusted model, FLT1 and LEP remained only nominally significant whereas ENG did not meet the nominal significance threshold. Mediation analysis confirmed that a significant proportion of FLT1, LEP, and ENG overexpression was attributable to differences in the cellular composition of the placenta. These results suggest that placental cell type proportion differences may be an overlooked factor in explaining the well-documented association between preeclampsia and FLT1, LEP, and ENG expression59,60,61,62. Downregulation of mitochondrial biogenesis, aerobic respiration, and ribosome biogenesis was robust to cell type adjustment, indicating direct changes to these pathways beyond shifts in cell type abundance. Indeed, disruption of the mitochondrial fission-fusion cycle72, malperfusion73,74, and inhibited protein synthesis secondary to endoplasmic reticulum stress75,76 have all previously been associated with preeclampsia. Interestingly, cell type adjustment increased the enrichment score results of vascular endothelial growth factor receptor signaling pathway, a mechanistic hypothesis in preeclampsia etiology63,73,77,78, from 1.36 to 1.77 (q = 0.43 to q = 0.15). This approach may reveal the biological mechanisms of other diseases beyond cellular composition differences. Because oxygen tension is a critical factor in trophoblast differentiation, inappropriate oxygenation may partially explain the elevated proportion of extravillous trophoblasts, though regulators of this process such as HIF1A and TGFB379 were not differentially expressed at the tissue level in our analysis. A recent single-cell RNA-sequencing case–control study of preeclampsia, however, identified upregulation of TGFB1 in extravillous trophoblasts, potentially indicative of altered trophoblast differentiation or invasion80,81. A similar study revealed decreased activity of gene network modules regulated by transcription factors ATF3, CEBPB, and GTF2B and decreased expression of CEBPB and GTF2B in preeclamptic extravillous trophoblasts compared to controls; follow-up in vitro experiments suggested CEBPB and GTF2B knockdown reduced extravillous trophoblast viability and invasion82. Consistent with our other findings, this study also observed a similar trend in cell type proportion differences and upregulation of FLT1 in extravillous trophoblasts and ENG in syncytiotrophoblasts between preeclampsia cases and controls80. Future work should consider and account for cell type proportions and the cell type-specific expression patterns of genes that regulate placental development or are associated with preeclampsia to better understand preeclampsia etiology.

This study has several strengths. We profile the parenchymal healthy term villous tissue in the placenta and integrate our dataset with samples from previously published studies to generate the largest, to the best of our knowledge, cell type-specific placental villous tissue gene expression reference to date. Single-cell RNA-sequencing allowed us to agnostically capture diverse placental cell types without a priori knowledge of cell types and their characteristics and tabulate gene expression patterns at high resolution and specificity. Our in silico deconvolution tests demonstrated robust performance to even extreme distributions of maternal or sex of fetal cells. We demonstrate technical replication of single-cell RNA-sequencing in placental villous tissue. We were able to apply our findings to a large target deconvolution dataset of preeclampsia that contained placental measures from hundreds of participants across eight different studies. Most importantly, we evaluate cell type proportion differences in an epidemiological study of placental parenchymal tissue and preeclampsia, and genome-wide gene expression differences accounting for cell type heterogeneity, a critical limitation in bulk tissue assays.

This study also has several limitations. Although our cellular sample size comprised of 40,494 cells is relatively large compared to previous single-cell RNA-sequencing studies of term placental villous tissue, this dataset still represents a limited biologic replicate sample size compared to epidemiologic scale studies. Our newly sequenced samples came from a convenience sample without available demographic information beyond uncomplicated and healthy Cesarean-section status. Similarly, the sample size of FACS-sorted tissues was limited, and some cell type fractions were excluded due to low RNA quality or exhibited poor estimated purity, likely complicated by degradation of cell surface markers from apoptosis characteristic of development and parturition83,84 and sample processing. This study did not include placental tissues for single-cell analysis from preeclamptic patients to confirm intra-cell type gene expression changes. Despite excellent in silico performance, we had no external gold standard to verify deconvolution performance. This deconvolution reference may not be sensitive enough to discriminate between cell subtypes such as proliferative vs. non-proliferative cytotrophoblasts that are clearly delineated in the single-cell analysis; in such cases, investigators may collapse cell type proportions counts into a single major cell type group, such as cytotrophoblasts. Future studies may verify whether cell type proportions estimated in diseased or vaginally delivered tissues are robust to a deconvolution reference generated from healthy villous tissue delivered via Cesarean-section. Residual confounding may remain in our statistical models due to the limited number of common covariates across all eight preeclampsia case–control studies. Due to the nature of villous tissue sampling, our study design is cross-sectional, limiting our ability to establish temporality between exposure and outcome to rule out reverse causation. As with any study conditioned on live birth, selection bias may affect our results. However, the effects of harmful exposures that lead to selection tend to be underestimated in these scenarios85,86. Therefore, our results likely represent a conservative underestimate of the effects of preeclampsia on inappropriate cell composition and preeclampsia status on gene expression.

In summary, we provide a cell type-specific deconvolution reference via single-cell RNA-sequencing in the parenchymal placental term villous tissue. We demonstrated this reference was robust to different distributions of maternal and fetal sex through in silico validation testing. In addition, we benchmarked these single-cell cell type-specific gene expression profiles against placental cell types isolated with more conventional FACS followed by bulk RNA-sequencing. We applied this deconvolution reference to an epidemiologic preeclampsia dataset to reveal biologically relevant shifts in placental cell type proportions between preeclampsia cases and controls. Once cell type proportion differences were accounted for, differential gene expression differences were markedly attenuated between preeclampsia cases and controls. Enrichment analysis revealed downregulation of mitochondrial biogenesis, aerobic respiration, and ribosome biogenesis were robust to cell type adjustment, suggesting direct changes to these pathways. A substantial proportion of the overexpression of the FLT1, LEP, and ENG in preeclampsia was mediated by placental cell composition. These results add to the growing body of literature that emphasizes the centrality of cell type heterogeneity in molecular measures of bulk tissues. We provide a publicly available placental cell type-specific gene expression reference for term placental villous tissue to overcome this critical limitation.

Methods

Placental tissue collection and dissociation

Placentas were collected shortly after delivery from healthy, full-term, singleton uncomplicated Cesarean sections at the University of Michigan Von Voigtlander Women’s Hospital. Pregnant women provided written informed consent for research use of discarded tissues. Study protocols for discarded tissue collection and research use were approved by the University of Michigan Institutional Review Board (HUM00017941, HUM00102038). Villous placental tissue biopsies were collected and minced for dissociation after cutting away the basal and chorionic plates and scraping villous tissue from blood vessels13. We subjected approximately 1 g minced dissected villous tissue to the Miltenyi Tumor Dissociation Kit on the GentleMACS Octo Dissociator with Heaters (Miltenyi Biotec) to yield single-cell suspensions of viable placental cells in 5 μM StemMACS™ Y27632 (Miltenyi Biotec) in RPMI 1640 (Gibco) according to manufacturer’s instructions for soft tumor type. Red blood cells were depleted using RBC lysis buffer (BioLegend) according to manufacturer’s protocol A. Single-cell suspensions were size-filtered at 100 μm to remove undigested tissue and subsequently at 40 μm12,13. To collect a syncytiotrophoblast-enriched fraction, the fraction between 40 and 100 μm was washed from the 40-μm strainers, adapting a previous protocol that collected syncytiotrophoblasts throughout this size range10. Single-cell suspensions <40 μm were cryogenically stored in 5 μM StemMACS™ Y27632 90% heat-inactivated fetal bovine serum (Gibco)/10% dimethyl sulfoxide (Invitrogen). For each placenta, additional whole villous tissue samples were stored in RNALater (Qiagen).

Previously published single-cell RNA-sequencing raw data of healthy, term placental villous tissue samples came from the Database of Genotypes and Phenotypes (Pique-Regi et al., accession number phs001886.v1.p187) SRR10166478 (Sample 3), SRR10166481 (Sample 4), and SRR10166484 (Sample 5)17. The collection and use of human materials for the study were approved by the Institutional Review Boards of the Wayne State University School of Medicine. All participating women provided written informed consent prior to sample collection17. Additional previously published samples came the European Genome-Phenome Archive (Tsang et al., accession number EGAS0000100244988) (Samples 6–9)16. The study was approved by the Joint Chinese University of Hong Kong-New Territories East Cluster Clinical Research Ethics Committee, and informed consent was obtained after the nature and possible consequences of the studies were explained. Pregnant women were recruited from the Department of Obstetrics and Gynecology, Prince of Wales Hospital, Hong Kong with informed consent; the subjects studied had consented to sequencing data archiving16.

Placental single-cell RNA-sequencing

Villous tissue single-cell suspensions were thawed and sorted via FACS with LIVE/DEAD Near-IR stain (Invitrogen) for viability and forward-scatter and side-scatter profiles to eliminate cellular debris and cell doublets. Viability- and size-sorted single-cell suspensions were submitted to the University of Michigan Advanced Genomics Core for single-cell RNA-sequencing. Single cells were barcoded, and cDNA libraries constructed on the Chromium platform (10X Genomics, Single Cell 3’ v2 chemistry). Paired-end 110 base pair reads were sequenced on NovaSeq 6000 (Illumina).

Single-cell RNA-sequencing preprocessing

Raw reads were processed, deconvoluted, droplet filtered, and aligned at the gene level with the Cell Ranger pipeline using default settings (v4.0.0, 10X Genomics) based on the GRCh38 GENCODEv32/Ensembl 98 reference transcriptome with STAR v2.5.1b89. Previously published single-cell RNA-sequencing raw data of healthy, term placental villous tissue samples from the Database of Genotypes and Phenotypes (Pique-Regi et al., accession number phs001886.v1.p1) SRR10166478 (Sample 3), SRR10166481 (Sample 4), and SRR10166484 (Sample 5)17 and from the European Genome-Phenome Archive (Tsang et al., accession number EGAS00001002449) (Samples 6–9)16 were processed identically. The freemuxlet program in the latest version (accessed December 5, 2021) of the “popscle” package was used to assign fetal or maternal origin and identify 736 mosaic doublets for removal based on single nucleotide polymorphisms with minor allele frequency greater than 10% from the 1000 Genomes Phase 3 reference panel (released May 2, 2013)90. Per cell quality control criteria were calculated using the quickQCPerCell() function (scater R package, version 1.18.6) with default settings91 (Supplementary Figs. 2 and 3) and included total unique RNA transcripts (also called unique molecular identifiers), unique genes, and percentage of reads mapping to mitochondrial genes92. According to the current recommended best practice, each batch was quality-controlled separately93. We excluded 6497 low-quality outlier cells defined as cells with less than 500 unique RNA molecules, less than 200 unique genes, or that were outliers in mitochondrial gene mapping rate. Mitochondrial mapping outliers exceeded four median absolute deviations in samples 1 and 2 (mitochondrial reads >9.2%) or three median absolute deviations in samples 3, 4, and 5 (mitochondrial reads >8.9%) and samples 6, 7, 8C, 8P, 9C, and 9P (mitochondrial reads >9.1%). To generate normalized gene expression data for visualizations and analyses that required normalization, single-cell gene counts were library size normalized by dividing the number of counts by the total number of counts expressed in that cell, multiplied by a scale factor of 10,000, and log-transformed with the NormalizeData() function (Seurat R package, version 4.1.1).

Single-cell RNA-sequencing clustering and cluster annotation

Maternal and fetal cells were split into separate datasets for clustering. To integrate data from cells across study sources and visualize clustering results with uniform manifold projection34, we used the mutual nearest neighbor batch correction approach via FastMNN from “SeuratWrapper” with default settings (R package, version 0.3.0)35. Supervised iterative clustering and sub-clustering with “Seurat” (R package, version 4.0.1) function FindClusters at different resolution parameters were evaluated using cluster stability via clustering trees in “clustree”94,95. A priori canonical cell type marker gene expression patterns and cluster marker genes were used to assign cell types to cell clusters (see results). Cells that fell outside cell type clusters or outlying in doublet density calculated with computeDoubletDensity were removed as putative doublets and doublet clusters were identified with findDoubletClusters for removal in “scDblFinder” (R package, version 1.4.0)96. 723 maternal-maternal or fetal-fetal putative doublets were excluded after integration and clustering. Using the manually annotated Michigan (this study) and Pique-Regi (phs001886.v1.p1) cell cluster labels as the reference data, Tsang sample (EGAS00001002449) cells were algorithmically annotated with “SingleR” (R package, version 1.6.1)97 with default settings, followed by manual review. Cells with low prediction certainty (assignment score lower than three median absolute deviations of all cells assigned) were excluded as putative maternal-maternal or fetal-fetal doublets. Fetal sex in Michigan (this study) samples was determined with average normalized XIST expression; fetal sex in Pique-Regi and Tsang samples was determined by annotation and confirmed with average normalized XIST expression (Supplementary Fig. 4). The final analytic sample included 40,494 cells and 36,601 genes across nine biological replicates, two of which had a technical replicate (Samples 1 and 2) and another two included peripheral subsampling (Samples 8 and 9).

Single-cell RNA-sequencing differential expression and biological pathway enrichment statistical analysis

Technical correlation was assessed by Spearman correlation after averaging the normalized expression for each gene by cluster and by technical replicate. Cluster marker genes were identified in “Seurat” with the FindAllMarkers function with default settings on single-cell gene expression counts92,95. Specifically, including both maternal and fetal cell types, the expression level in each cell type cluster was compared against the average expression of that gene across all other cell types using the two-tailed Wilcoxon rank sum test with significance defined at a false discovery rate-adjusted p value less than 0.05 and a log2 FC cutoff of 0.25. Pairwise cluster markers were identified in “Seurat” with the FindMarkers function with an identical testing regime. Overexpressed genes were ranked by decreasing log2 FC for functional enrichment analysis with “gprofiler2” (R package, version 0.2.0, database version e102_eg49_p15_7a9b4d6) using annotated genes as the universe, excluding electronically generated annotations, and with the g:SCS multiple testing correction method applying a significance threshold 0.0598.

In silico testing of deconvolution performance

To test the performance and robustness of our placental single-cell RNA-sequencing deconvolution reference, we randomly split our analytic single-cell RNA-sequencing dataset into 50% training and 50% testing subsets with balanced cell type proportions49. We applied the test subset with the CIBERSORTx Docker container (accessed December 7, 2021) to create a signature gene expression matrix to test deconvolution performance with default settings99. To evaluate the reference’s robustness to fetal sex and ability to discriminate immune cell types of fetal versus maternal origin, we generated in silico pseudo-bulk test mixtures with known distributions of fetal and maternal cells, as well as male and female placental cells. Test mixtures included all of the 50% testing data, only fetal cells from the test data, only maternal cells from the test data, only female fetal cells from the test data, or only male cells from the test data. For the female and male fetal cell test mixtures, the baseline distribution of maternal cells was maintained by randomly down-sampling the maternal cells and randomly down-sampling the male fetal cells to the number of female fetal cells. We used the signature matrix generated from the training data to estimate constituent cell type proportions in these test mixtures using CIBERSORTx with cross-platform S-mode batch correction and 50 permutations to evaluate imputation goodness-of-fit. Pearson correlations and root mean square error between the test set predicted and actual cell type proportions in the test mixtures were used to assess deconvolution performance.

Fluorescence-activated cell sorting of major placental cell types from villous tissue

Villous tissue single-cell suspensions were quickly thawed and stained according to manufacturer’s instructions with five fluorescently labeled antibodies (CD9-FITC, CD45-APC, HLA-A,B,C-PE/Cy7, CD31-BV421, and HLA-G-PE) as well as LIVE/DEAD Near-IR stain (Invitrogen) to isolate six viable populations of placental cells by FACS at the University of Michigan Flow Cytometry Core Facility. Initial flow cytometry experiments included fluorescence minus one, single-color compensation, and isotype controls. Isotype controls were found to be the most conservative and were consequently included in all sorting experiments, as well as single-color compensation controls due to the large number of colors used in sorting. The six populations of cells were Hofbauer cells, endothelial cells, fibroblasts, leukocytes, extravillous trophoblasts, and cytotrophoblasts. We developed a five-marker cell surface FACS scheme to sort cytotrophoblasts (HLA-A,B,C-), endothelial cells (CD31+), extravillous trophoblasts (HLA-G+), fibroblasts (CD9+), Hofbauer cells (CD9-), and leukocytes (CD45+/CD9+) from villous tissue (Supplementary Fig. 10)11,12,14,37,100,101,102,103,104,105,106. Syncytiotrophoblast fragments were enriched from villous tissue digests. We isolated cell type fractions and whole villous tissue from four healthy term, uncomplicated Cesarean sections, labeled Sorted 1 (same sample source as single-cell RNA-sequencing sample 1), Sorted 2, Sorted 3, and Sorted 4. We subjected 24 cell type fractions with sufficient RNA content to RNA-sequencing, including two cytotrophoblast, one endothelial, three extravillous trophoblast, three fibroblast, four Hofbauer cell, four leukocyte, and two syncytiotrophoblast fractions, and five whole tissue samples (Supplementary Table 2).

Detailed antibody information: FITC, marker CD9: Mouse IgG1-kappa, clone HI9a (2.5 μg/mL), BioLegend #312103, lot B188319, BioLegend #312104, lot B232916; isotype control: clone MOPC-21 BioLegend #400107, Lot B199152 (2.5 μg/mL). APC, marker CD45: Mouse IgG1-kappa, clone 2D1, BioLegend #368511, Lot B215062 (0.125 μg/mL); isotype control: clone MOPC-21, BioLegend #400121, lot B216780 (0.125 μg/mL). PE/CY-7, marker HLA-ABC: Mouse IgG2a-kappa, clone W6/32, BioLegend #311429, lot B188649, BioLegend #3111430, lot B238602 (0.44 μg/mL); isotype control: clone MOPC-173, BioLegend #400231, lot B209000 (0.44 μg/mL);. BV421, marker CD31: Mouse IgG1-kappa, clone WM59, BioLegend #303123, lot B204347, BioLegend #303124, lot B232010 (0.625 μg/mL); isotype control: clone MOPC-21, BioLegend #400157, lot B225357 (0.625 μg/mL). PE, marker HLA-G: Mouse IgG2a-kappa, clone 87G, BioLegend #335905, lot B222326, BioLegend #335906, lot B199294 (5 μg/mL); isotype control clone MOPC-173, BioLegend #400211, lot B227641 (5 μg/mL). Mouse IgG1-kappa, clone MEM-G/9, Abcam #24384 Lot GR3176304-1 (2.5 μg/mL); isotype control: monoclonal, Abcam #ab81200, lot GR267131-1 (2.5 μg/mL). Validation information available on the manufacturer’s website under the catalog ID for each antibody.

A cutoff of 0.1% events was used to set a series of gates. Cells were first gated on size and granularity (FSC-HxSSC-H) to eliminate debris, followed by doublet discrimination (FSC-HxFSC-W and SSC-HxSSC-W). Ax750 was used to sort on viability. Extravillous trophoblasts were isolated based on Human Leukocyte Antigen-G (HLA-G) expression (Supplementary Fig. 10a). Cytotrophoblasts are HLA-ABC negative (Supplementary Fig. 10b). HLA-ABC-positive cells were then subjected to a CD45/CD9 gate to isolate Hofbauer cells and a heterogeneous population of leukocytes (Supplementary Fig. 10c). Finally, CD45-/CD9- population is sorted into the endothelial or fibroblast bins based on CD31 expression (Supplementary Fig. 10d).

Bulk placental tissue and sorted placental cell type RNA extraction and sequencing

Approximately 2 mg of bulk RNALater-stabilized (Qiagen) bulk villous tissue was added to 350 μL 1% β-mercaptoethanol (Sigma-Aldrich) RLT Buffer Plus (Qiagen) to Lysing Matrix D vials (MP Biomedicals). Samples were disrupted and homogenized on the MP-24 FastPrep homogenizer (MP Biomedicals) at 6 m/s, setting MP24x2 for 35 s. For the homogenized bulk villous tissue, syncytiotrophoblast-enriched fraction, and sorted cell types, RNA extraction was completed according to the manufacturer’s instructions using the AllPrep DNA/RNA Mini Kit (Qiagen) and stored at −80 °C. RNA samples were submitted to the University of Michigan Advanced Genomics Core for RNA-sequencing. Ribosomal RNAs were depleted with RiboGone (Takara) and libraries were prepared with the SMARTer Stranded RNA-Seq v2 kit (Takara). Paired- or single-end 50 base pair reads were sequenced on the HiSeq platform (Illumina). Raw RNA reads were assessed for sequencing quality using “FastQC” v0.11.5107 and “MultiQC” v1.7108. Reads were aligned to the GRCh38.p12/ GENCODEv28 reference transcriptome using “STAR” v2.6.0c with default settings89. featureCounts from “subread” v1.6.1 was used to quantify and summarize gene expression with default settings109.

Sorted placental cell type differential expression analysis and comparison to single-cell results

For visualizations or analyses that required normalized gene counts, sorted cell type gene counts were library size normalized with the median ratio method using the counts() function (DESeq2 R package, version 1.32.0). As recommended50, we excluded genes that were not present in at least three samples and did not have an expression of 10 library size-normalized counts. To visualize broad cell type-specific gene expression patterns, we used “DESeq2’s” (R package, version 1.32.0) plotPCA() function with the regularized logarithm transformation, blinded to experimental design. Upregulated genes in each cell type were identified using the negative binomial linear model two-tailed Wald test in “DESeq2” (R package, version 1.32.0) adjusted for biological replicate using default settings with contrasts comparing the expression of a gene in one cell type against the average expression across all other cell types at a false discovery rate-adjusted p value less than 0.05 and a log2 FC cutoff of 1.250. Overexpressed genes were ranked by decreasing log-FC for functional enrichment analysis with “gprofiler2” (R package, version 0.2.0, database version e102_eg49_p15_7a9b4d6) using annotated genes as the universe, excluding electronically generated annotations, and with the default g:SCS multiple testing correction method applying significance threshold adjusted p value of 0.0598. To compare sorted and single-cell results, we tabulated unique overlapping differentially expressed genes and overrepresented pathways by cell type (Supplementary Table 3) Peripheral fetal and maternal immune cell types from the single-cell RNA-sequencing data were collapsed to one leukocyte category, cytotrophoblast subtypes to one cytotrophoblast category, and mesenchymal stem cells and fibroblasts to one fibroblast category for this comparison.

We used the CIBERSORTx Docker container (accessed December 7, 2021) to create a signature gene expression matrix for deconvolution from the counts of the single-cell RNA-sequencing data with the following default parameters: differential expression q value < 0.01, no minimum gene expression cutoff, and a 300 gene feature selection floor and a 500 gene feature selection ceiling99. We used the signature matrix to estimate constituent cell type proportions in the 4 whole tissue (with 1 additional technical replicate) and 19 sorted or enriched cell type fractions using CIBERSORTx with cross-platform S-mode batch correction and 50 permutations to evaluate imputation goodness-of-fit. We collapsed the high-resolution single-cell cell type cluster labels to the seven cell type fractions we targeted for comparison with sorted cell type results.

Application: bulk placenta gene expression dataset and CIBERSORTx deconvolution

Bulk placental tissue microarray gene expression (previously batch-corrected and normalized) from eight preeclampsia case–control studies was downloaded from the NCBI Gene Expression Omnibus (accession number GSE75010) for deconvolution33. We used the CIBERSORTx Docker container (accessed December 7, 2021) to create a signature gene expression matrix for deconvolution from the counts of the single-cell RNA-sequencing data with the following default parameters: differential expression q value < 0.01, no minimum gene expression cutoff, and a 300 gene feature selection floor and a 500 gene feature selection ceiling99. We used the signature matrix to estimate constituent cell type proportions in GSE75010 using CIBERSORTx with cross-platform S-mode batch correction and 50 permutations to evaluate imputation goodness-of-fit.

Application: preeclampsia case–control differential cell type abundance, differential gene expression statistical analysis, and mediation analysis

To test for differences in estimated cell type proportions between preeclampsia cases and controls, estimated cell type proportions for GSE75010 were regressed on preeclampsia case–control status using beta regression models adjusted for gestational age, sex, and study source110 (Supplementary Data 8). Statistical significance was assessed using the two-tailed Wald test applying a nominal significance threshold of 0.05. Cell types imputed at zero percent abundance across all samples were excluded. For modelling purposes, zero percent abundance estimates were transformed to \(\frac{1}{2}/n\) where n is the number of observations (n = 330).

Differential expression analysis was conducted in limma111 with default linear models adjusted for gestational age, fetal sex, and study source with empirical Bayes standard error moderated t-test statistics. A cell type-adjusted model was built on the base model adjusted for gestational age, fetal sex, and study source and additionally adjusted for the first five PCs of deconvoluted cell type proportions (Supplementary Data 9). Statistical significance was assessed at false discovery rate-adjusted q value < 0.05 and a log2 FC cutoff of 0.1. Differentially expressed genes were descending-ranked by value of the moderated test statistic for gene set enrichment analysis in desktop version GSEA 4.1.0 with the GSEAPreranked tool with default settings against the c5.go.bp.v7.5.1.symbols.gmt gene set database112,113 (Supplementary Data 10). PC analysis was performed with prcomp() from “stats-package” (R, version 4.0.5) without scaling and with default settings.

A unified mediation and interaction analysis114 was conducted in “CMAverse” (R package, version 0.1.0)115 via the g-formula approach116 to estimate causal randomized-intervention analogs of natural direct and indirect effects117 through direct counterfactual imputation. The model was operationalized with preeclampsia status as the binary exposure, log2 transformed gene expression intensity as the continuous outcome, and the first five PCs of deconvoluted cell type proportions as continuous mediators. Baseline covariates included fetal sex and study source. Continuous gestational age was included as a confounder of the mediator-outcome relationship affected by the exposure. Confidence intervals were bootstrapped with 1000 boots with otherwise default settings. Statistical tests were two-tailed and interpreted at a p value significance threshold of 0.05.

Statistics and reproducibility

Technical replication measured by average intra-cluster gene expression between technical replicates was tested via the two-tailed Spearman correlation test within Samples 1 and 2 assessed across all 32,738 common genes. The number of cells contributing expression data for each cell type is available in Table 1. Single-cell cluster marker genes were identified in “Seurat” with the FindAllMarkers function with default settings on single-cell gene expression counts92,95. Specifically, including cells from both maternal and fetal cell types, the expression level in each cell type cluster was compared against the average expression of that gene across all other cell types using the two-tailed Wilcoxon rank sum test with significance defined at a false discovery rate-adjusted p value less than 0.05 and a log2 FC cutoff of 0.25 (n = 40,494 cells). The final analytic sample included 40,494 cells and 36,601 genes across nine biological replicates, two of which had a technical replicate (Samples 1B and 2B) and another two included peripheral subsampling (Samples 8P and 9P). Pairwise cluster markers were identified in “Seurat” with the FindMarkers function with an identical testing regime (n = 6132 cells for proliferative vs. non-proliferative cytotrophoblasts). Overexpressed genes were ranked by decreasing log-FC for functional enrichment analysis with “gprofiler2” (R package, version 0.2.0, database version e102_eg49_p15_7a9b4d6) using annotated genes as the universe, excluding electronically generated annotations, and with the default g:SCS multiple testing correction method applying significance threshold adjusted p value of 0.0598. Overexpressed genes per cell type cluster are available in Supplementary Data 2 and ontology results in Supplementary Data 3. Overexpressed genes and related enrichment results comparing proliferative to non-proliferative cytotrophoblasts are available in Supplementary Data 1.

Upregulated genes in each cell type were identified using the negative binomial linear model two-tailed Wald test in “DESeq2” (R package, version 1.32.0) adjusted for biological replicate using default settings with contrasts comparing the expression of a gene in one cell type against the average expression across all other cell types at a false discovery rate-adjusted p value less than 0.05 and a log2 FC cutoff of 1.250 (n = 19 cell type fraction samples with breakdown by cell type available in Supplementary Table 2). Overexpressed genes were ranked by decreasing log-FC for functional enrichment analysis with “gprofiler2” (R package, version 0.2.0, database version e102_eg49_p15_7a9b4d6) using annotated genes as the universe, excluding electronically generated annotations, and with the default g:SCS multiple testing correction method applying significance threshold adjusted p value of 0.0598. Differentially expressed genes per cell type available in Supplementary Data 4 and number of differentially expressed genes are summarized in Supplementary Fig. 12. Ontology results are available in Supplementary Data 5.

Bulk placental tissue microarray gene expression (previously batch-corrected and normalized) from eight preeclampsia case–control studies was downloaded from the NCBI Gene Expression Omnibus (GSE75010) for deconvolution (n = 330)33. We used the CIBERSORTx Docker container (accessed December 7, 2021) to create a signature gene expression matrix for deconvolution from the counts of the single-cell RNA-sequencing data with the following default parameters: differential expression q value < 0.01, no minimum gene expression cutoff, and a 300 gene feature selection floor and a 500 gene feature selection ceiling99. We used the signature matrix to estimate constituent cell type proportions in GSE75010 using CIBERSORTx with cross-platform S-mode batch correction and 50 permutations to evaluate imputation goodness-of-fit.

To test for differences in estimated cell type proportions between preeclampsia cases and controls (n = 330), estimated cell type proportions for GSE75010 were regressed on preeclampsia case–control status using beta regression models (n = 25 cell type proportion outcomes) adjusted for gestational age, sex, and study source110. Cell types imputed at zero percent abundance across all samples were excluded (n = 2 excluded: fetal naïve CD4+ T cells and fetal GZMB+ natural killer cells). Statistical significance was assessed using the two-tailed Wald test applying a nominal significance threshold of 0.05.

Differential expression analysis was conducted in limma111 with default settings using linear models (n = 14,651 genes) adjusted for gestational age, fetal sex, and study source (n = 330). A cell type-adjusted model was built on the base model additionally adjusted for the first five PCs of deconvoluted cell type proportions. PC analysis was performed with prcomp from “stats-package” (R, version 4.0.5) without scaling and default settings. Statistical significance was assessed at false discovery rate-adjusted q value < 0.05 and a log2 FC cutoff of 0.1. Differentially expressed genes were descending-ranked by the value of the moderated test statistic for gene set enrichment analysis in desktop version GSEA 4.1.0 with the GSEAPreranked tool with default settings against the c5.go.bp.v7.5.1.symbols.gmt gene set database112,113.

A unified mediation and interaction analysis114 was conducted in “CMAverse” (R package, version 0.1.0)115 via the g-formula approach116 to estimate causal randomized-intervention analogs of natural direct and indirect effects117 through direct counterfactual imputation. The model (n = 330) was operationalized with preeclampsia status as the binary exposure, normalized log2 gene expression signal intensity as the outcome, and the first five PCs of deconvoluted cell type proportions as continuous mediators. Baseline covariates included fetal sex and categorical study source. Continuous gestational age was included as a confounder of the mediator-outcome relationship affected by the exposure. Confidence intervals were bootstrapped with 1000 boots with otherwise default settings. Statistical tests were two-tailed and interpreted at a p value significance threshold of 0.05.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.