Major human placental cell types have highly specific methylation patterns
To characterize the dynamics of CpG methylation during human placental development, we performed microarray profiling (Illumina EPIC methylation array, n CpGs = 737,050 after removal unreliable probes) in samples of matched CV and 4 fluorescence-activated cell sorted (FACS) cell- types (Additional File 1: Figure S1A), from 9 first trimester (6.4-13 weeks gestational age) and 19 term (36.4-40.4 weeks) pregnancies (Table 1). Immunofluorescence staining of flow cytometry sorted cells (Additional File 1: Figure S1B-E) determined high purity for TB (KRT7+, 97%), HB (CD68+, 95%), and EC (CD31+, 88%) and lower purity for SC (VIM+, 73%). Several bioinformatic approaches, such as array-based sex inference (28), and genotype clustering, were used to identify contamination with maternal DNA (Additional File 1: Figures S2A-F, Additional File 2: Supplementary methods). We restricted analysis to samples with an estimated maternal cell contamination of less than 35%, with the majority of first trimester samples having less than 20%, and term samples less than 10% (Additional File 1: Figure S2G). This resulted in the exclusion of: 6 HB, 1 EC, and 4 TB from first trimester, and 1 HB from term samples. Final sample numbers in all downstream analyses are shown in Table 1.
To determine major factors that drive DNAm variation, we first applied principal components analysis (PCA) to all 126 CV and cell samples. Three distinct clusters were observed when samples were projected onto PCs 1 and 2 (total percent variation explained = 64%; Figure 1A). Samples in these clusters were i) TB and CV, ii) SC and EC, and iii) HB. Cell type was strongly associated with the first 3 PCs (p<0.001), while gestational age (i.e. “Trimester”) was the second strongest identifiable factor driving DNAm variation, being associated with PCs 4 and 5 (p<0.001, Additional File 1: Figure S3). Technical variables such as “Batch”, “Row”, and “Chip ID” explained less variation in comparison to biological variables. Sex was associated with PCs 6 and 8-11 (p<0.01). The close clustering of TB with CV (original unsorted tissue) is consistent with this being the predominant cell type in whole villi.
|
First trimester
|
Term
|
|
|
|
Chorionic villi
|
7
|
19
|
Trophoblast (EGFR+/KRT7+)
|
5
|
19
|
Hofbauer (CD14+/CD68+)
|
3
|
18
|
Endothelial (CD34+/CD31+)
|
8
|
19
|
Stromal (VIM+)
|
9
|
19
|
|
|
|
Mean Gestation age (mean and range in weeks)
|
10.8 (6.4-13)
|
39.0 (36.4-40.4)
|
Sex (n Males)
|
4
|
9
|
|
|
|
|
|
|
|
|
Table 1. Number of cell-specific and matched chorionic villi samples from first trimester and term placentas, measured on the Illumina EPIC methylation array. Surface markers for flow cytometry and immunofluorescence staining are shown in brackets.
We next wanted to define the extent and patterns of cell-specific DNAm. At a Bonferroni-adjusted p<0.01 and an absolute difference in mean methylation (Db) > 25%, we found 75,000-135,553 and 9,136-117,528 (term and first trimester, respectively) cell-specific differentially methylated CpGs (DMCs; Figure 1B; Additional File 3: first trimester DMCs, Additional File 4: term DMCs). The differences in the number of DMCs between first trimester (n=3-9) and term (n=18-19) comparisons are likely due to less power from the smaller sample size for first trimester samples compared to term. When comparing across term samples, we detected more DMCs for TB and HB (n=135,553 and 130,733) compared to SC and EC (80,153 and 75,525; respectively). This was also true for first trimester samples: there were more DMCs for TB and HB (117,528 and 78,309) than SC and EC (9,136 and 18,867). We further classified these DMCs by whether their methylation was in the “less than” (compared to all other cell types) or “more than” direction. Most TB DMCs were in the less methylated direction (61% - first trimester, 88% term), whereas HB DMCs were often more methylated than other cell types (74% - first, 72% term). A list of 38,656 - 86,355 differentially methylated regions (DMRs) were identified (FDR < 0.01) using the R package dmrcate (29) for each cell type and gestational age; these results are presented in Additional File 5 and 6.
To characterize the functional relevance of placental cell-specific DMCs, we tested these CpGs for enrichment in various genomic elements (chi-squared test, FDR < 0.05; term DMCs in Figure 1C, first trimester DMCs in Additional File 1: Figure S4). Cell-specific DMCs were depleted in gene-related elements such as promoters, exons, 5’ UTRs, and 3’ UTRs. Instead, we saw significant enrichment in non-coding regions, such as open seas, CpG island shores, intergenic regions, introns, and enhancers. The level and direction of enrichment was highly consistent across first trimester and term cell DMCs. Less methylated DMCs were enriched for placental PMD regions (15) for TB but depleted for all other cell types. Functional enrichment analysis tested if GO or KEGG pathways were associated with cell-specific DMCs. We adjusted for the variable number of CpGs per gene to reduce bias in gene set analysis (30). EC and HB DMCs were enriched (FDR < 0.05) for terms related to intercellular interactions such as “cellular response to external stimulus”, whereas stromal DMCs yielded more intracellular processes related to maintaining tissue structure, such as “actin cytoskeleton” and “collagen binding”. Trophoblast DMCs were enriched for two KEGG pathways, “ECM-receptor interaction” and “Regulation of actin cytoskeleton” (Additional File 7: Table S5 and S6).
Figure 1. Genome-wide characterization of placental cell DNA methylation. A) Principal components analysis (PCA) was applied to all samples and CpGs. Samples are projected onto axes PC1 and PC2 which account for 41% and 23% total variance, respectively. B) Results from the differential methylation analysis using the R package limma are shown here. DMCs, defined as those tests passing a Bonferroni-adjust p-value < 0.01, and a difference in group means > 0.25, were divided into less methylated and more methylated compared to all other cell types. C) Enrichment analysis of term cell-specific DMCs was carried out on genomic elements using a chi-squared test and a Bonferroni-adjusted p-value < 0.01. The expected (background) frequency, which is the percentage of total tested CpGs in each genomic element, is shown as a black line. D) Average term placental cell-specific DNA methylation across TFAP2C transcripts on chromosome 6, and E) INHBA transcripts on chromosome 7. Differentially methylated regions (defined as regions with a high density of differentially methylated CpGs), are highlighted with a grey background. Y axis ranges from 0-100% DNA methylation.
Cell-specific DNAm occurs at highly functionally-relevant genes
A number of regions with a high density of DMCs were located in or nearby functionally- and pathology-relevant genes. TFAP2C transcripts, which encodes a pan-trophoblast marker, were highly methylated in TB compared to other cell types in the promoter and upstream region; whole CV showed a similar profile to TB (Figure 1D). This region contains several predicted enhancers (29), which may require DNAm for recruiting transcription factors. Alternatively, other regions more distal to TFAP2C may be responsible for regulation of this gene’s transcription. Other trophoblast-specific markers, such as GCM1, MMP2, SLC1A5, and GATA3, also had regions of highly cell-specific DNAm localized near their transcription start sites (Additional File 1: Figure S5). We also observed high DMC density regions in genes for which placental DNAm and/or expression differences have been associated with preeclampsia (30), including INHBA (Figure 1E), JUNB, TEAD3, NDRG1, and CGA (Additional File 1: Figure S6). Out of 540 preeclampsia-associated CpGs previously identified by Wilson et al. 2018 that were also captured in our processed data, a statistically significant (Bonferroni adjusted p<0.01) fraction ranging from 19.4-27.2% were also identified as exhibiting cell-specific DNAm for term samples (Table 2) (30).
|
n cell-specific DMCs
|
n DMCs that are preeclampsia-associated
|
Proportion out of 599 preeclampsia CpGs that are also cell-specific DMCs
|
Odds ratio
|
Trophoblast
|
135,553
|
147 (0.11%)
|
27.2%
|
1.66
|
Stromal
|
80,153
|
105 (0.13%)
|
19.4%
|
1.98
|
Endothelial
|
75,525
|
109 (0.14%)
|
20.2%
|
2.22
|
Hofbauer Cells
|
130,733
|
131 (0.10%)
|
24.3%
|
1.49
|
Table 2. Number of preeclampsia-associated CpGs from Wilson et al. 2018 that are cell-specific DMCs for term samples. Enrichment for preeclampsia-associated CpGs was statistically significant for each term cell-specific set of CpGs at a Bonferroni-adjusted p<0.01.
We hypothesized that genome-wide differences in DNAm could in part relate to differences in the expression and DNAm at genes that regulate the deposition, maintenance, and removal of DNAm, such as DNMT1, DNMT3A, DNMT3B, DNMT3L, and TET1. In these genes, we found that a high proportion of CpGs in the promoter region (61%, 36%, 31%, 83%, 18%, respectively) were differentially methylated by cell type However, considering the variable number of CpGs associated with each gene’s promoter, these percentages were not significantly greater than genes of similar CpG coverage (Figure 2AB). Differential methylation within DNAm-regulating genes was highly localized (Figures 2C). The promoter of DNAm-maintenance gene DNMT1, which is known to be specifically imprinted in the placenta (31), shows the expected intermediately methylated (i.e. ~50%) pattern for all cell types except HB, which is completely unmethylated (Figure 2C). This suggests that DNMT1 is imprinted in TB, SC, and EC, but not in HB.
Figure 2. Differential methylation at DNA methylation -regulating genes. A) On a per-gene basis, the number of promoter CpGs that are differentially methylated by at least one cell type, out of the total number of promoter CpGs per gene. The y = x line is shown (blue), where genes with 100% of promoter CpGs are differentially methylated. The green line is a smoothed average. B) Distribution of the percentage of promoter CpGs per gene that are differentially methylated. The dotted line represents an array-wide average. C) DNA methylation at CpGs associated with DNMT1 for term placental samples (top). CpGs in CpG islands, imprinted regions, PMDs, and enhancers are indicated (middle). Associated UCSC transcripts and their genomic elements (promoter, 5’ UTR, exons, introns, 3’UTR) are displayed (bottom).
DNA methylation characterization of Syncytiotrophoblast and Hofbauer cells
We used the pan-trophoblast marker EGFR to isolate TB using FACS. Because mature EVTs exist primarily in maternal tissue, and STBs are structurally incompatible with FACS isolation protocols, our TB sample likely consists primarily of CTB. In order to better understand the relationship between STB and the isolated TB cells, we compared a subset of TB with matched STB from the same placenta that was obtained from enzymatic separation using Collagenase IA (referred to as eSTB; n=5) from term CV samples. This digestion protocol which extracts the outer layer of the CV, produces a sample enriched for STB, but is likely to also contain a proportion of non-STB cell types. To compare eSTB samples globally to other cell types, we projected eSTB onto PCs 1 and 2 to see where they cluster in relation to other samples. On PCs 1 and 2, eSTB clustered closely with TB and CV samples, indicating high similarity between these three populations (Figure 3A). Throughout gestation, the STB proportion increases, and is greater in nuclei number compared to CTB at term. To determine if TB or eSTB samples were more similar to CV, unsupervised hierarchical clustering was applied on the top 1000 most variable probes, and resulted in CV clustering with eSTB (Figure 3B), which is consistent with the expectation that CV consists primarily of STB. Supporting this, we found more DMCs (Bonferroni p<0.01, absolute difference in mean DNAm > 25%) between TB and eSTB (n DMCs = 4,666), than between CV and eSTB (n DMCs = 72). Differential methylation at specific CpGs localized to genes known to be expressed in STB, such as CGA, CYP19A1, PAPPA2, PARP1, SLC13A4, and SLC22A11 (Figure 3C) (32–35). The direction of DNAm at these CpGs was mostly consistent with expected patterns of genes that are more active in eSTB compared to TB and other placental cell types (i.e. more methylation at introns, less methylation at promoters).
The distinct DNAm profiles observed in placental HB suggests a distinct developmental trajectory. Indeed, the functional role and phenotypic diversity of HBs is complex and thought to vary across gestation, however, they show similar morphological and cell marker characteristics as adult and fetal monocytes (7). Therefore, to compare placental HBs to other immune cells, we compared their DNAm profiles to a curated 450k DNAm database of flow-sorted cord blood cell types (n=263) (36). We included only term HBs in this comparison since the available cord blood data was collected from term samples. To determine which cord blood cell types HB are most similar to, we applied unsupervised hierarchical clustering on the top 1000 most variable CpGs across each dataset. We observed that HB form their own distinct cluster (Figure 3D), indicating they likely have unique functional properties compared to other immune cells at similar developmental stages. This finding supports previous reports of distinct DNAm between of HBs to fetal/maternal monocytes, and decidual macrophages (6). HBs cluster most closely with monocytes and granulocytes, consistent with them having a common developmental origin.
Figure 3. Characterization of enzymatically-separated syncytiotrophoblast and Hofbauer cell DNAm to closely related cell types. A) Syncytiotrophoblast samples (n=5) were projected onto principal components PC1 and PC2. Original samples used for constructing these PCs (Figure 1A) are shown (chorionic villi: dark red, trophoblast: yellow, all others: grey). Syncytiotrophoblast (orange) cluster with the chorionic villi and trophoblast samples. B) Clustering on the top 1000 variable CpGs between chorionic villi, syncytiotrophoblast, and trophoblast samples. Hierarchical clustering with Euclidean distance was used for both CpG-wise (rows) and sample-wise (columns) clustering. DNAm is shown as a range between 0-100%. C) Density plots are shown for select differentially methylated CpGs, which were identified using limma, with a Bonferroni adjust p < 0.01, and a mean difference in DNAm > 25%. CpGs are shown along the y-axis with their locational relationship (shown in brackets) to their associated gene (left). DNA methylation is shown on the x-axis. D) Clustering on the top 1000 variable CpGs between Hofbauer cells and cord blood cell types. Hierarchical clustering with Euclidean distance was used for both CpG-wise and sample-wise clustering. WBC: whole cord blood, nRBC: nucleated red blood cells, NK: natural killer cells, CD4T: CD4+ T cells, CD8T: CD8 T cells, Gran: granulocytes, Bcell: B cells, DNAm: DNA methylation.
Canonical placental epigenetic features are not always present in all constituent cells
To determine if previously identified placental specific features of DNAm are cell specific, we compared cell-type specific DNAm at partially methylated domains (PMDs), genomic imprinting, and repetitive elements (15,18,37). PMDs are large (>100kb) regions of lower average methylation (<70%) compared to surrounding regions. Placental PMDs are thought to contribute to the observation that placental DNAm on average is much lower than other human tissues (10). To characterize their cell-specificity, we calculated the percentage of CpGs that are found in previously defined placental PMDs (15) with DNAm falling into 20% intervals (0-20%, 20-40%, 40-60%, 60-80%, 80-100%). We observed that DNAm levels in PMDs is highly cell-specific (Figure 4A). TB, like CV, have more CpGs with low levels of DNAm in PMDs (0-40%) compared to other cell types. HB show a strong bias towards higher DNAm levels, with over 43% of CpGs in PMDs exhibiting >80% DNAm. We observed some changes within cell types between trimesters. All cell types have lower levels (0-40%) of methylation in term compared to first trimester. All cell types except TB have less intermediately (40-60% intervals) methylated CpGs at term compared to first trimester. HB, in contrast, have more intermediately (40-60% intervals) methylated CpGs in third trimester. In summary, the methylation levels at CpGs in PMD regions were at the expected levels (relatively low methylated compared to surrounding regions) for CV and TB; sometimes hypermethylated for EC and SC; and were almost always highly methylated for HB, at levels typically found in somatic cells.
In examining specific regions containing PMDs, a strong bimodal pattern of methylation was observed, where regions of lower methylation (overlapping known placental PMD regions), which were surrounded by regions of higher methylation (Figure 4BC). TB DNAm levels followed closely the levels measured in CV, supporting that placental PMDs are likely reflecting mainly TB-specific DNAm patterns. In contrast, DNAm in HB often deviated from the other cell types, typically showing higher levels of methylation within PMDs. SC and EC often “followed” CV DNAm levels, but were not nearly as consistent as TB cells in this respect.
We also looked at imprinted differentially methylated regions (DMRs) that are covered by the EPIC array. While many imprinted DMRs are maintained in somatic tissues, others are highly specific to the placenta (18–21). To evaluate whether placental-specific imprinting is maintained in constituent placental cell populations, we first combined the results from four studies (18,19,21,38) to form a list (Additional File 7: Table S7) of placental-specific (n CpGs = 981; n genes = 111) and non-placental specific (i.e. imprinted in other tissues) DMRs (n CpGs = 1,085; n genes = 307). To determine if CpGs were intermediately-methylated, as would be expected for an imprinted DMR, we counted the proportion of CpGs with an average DNAm across both alleles that were in a range between 25-75% methylation. For CpGs in non-placental specific imprinted DMRs, the mean percentage of CpGs in the intermediate range across each cell type and in CV in term samples was 69% (Figure 4D). For placental-specific imprinted CpGs, the percentage of CpGs falling into this DNAm range was much more variable. As expected, in the term placental samples, TB and CV had a high percentage (76%, 81%, respectively) of CpGs in this DNAm range. SC and EC had a lower, but still a majority, percentage of CpGs in this range (64%, 64%, respectively). In contrast, HB cells had almost no CpGs (12%) in this range; almost all CpGs were unmethylated (<25%). These proportions were similar in first trimester samples, except with EC and SC showing less intermediate methylation and more CpGs with less methylation at placental-specific imprints (Additional File 1: Figure S7A). These results suggest that placental-specific imprinting is maintained primarily in TB, and to a lesser degree EC and SC, and is virtually absent in HB. When considering the parental origin of imprinted DNAm (18–21,38), paternally- methylated regions had more CpGs falling within 25-75% as compared to maternal ones (Additional File 1: Figure S7BC). We only estimated this in non-placental specific imprinted DMRs, since almost all validated placental-specific imprinted DMRs are maternally methylated.
DNAm at specific imprinted DMRs was examined. As described above, TB and CV had intermediate (>25%, <75%) DNAm at nearly all CpGs located in placental-specific imprinted regions (Figure 4D). Most of these CpGs, in contrast, are hypomethylated for HB cells, consistent with this cell type having a different developmental origin than other placental components (embryonic versus extraembryonic). However, at the imprinted DMR associated with the placental-specific expressed microRNA cluster C19MC, this pattern is reversed: HB have hypermethylation at this region (Figure 4E) as is reported for somatic adult/fetal tissues (18). For SC and EC, these cell types generally show lower levels of DNAm than TB/CV at the placental-imprinted DMRs, sometimes matching that in HBs and other times showing levels somewhere between HB and TB/CV. Such patterns are observed for genes such as DCAF10 (Figure 4F), fibroblast growth factors FGF8, FGF12 (Additional File 1: Figure S8AB), and at epigenetic regulator JMJD1C (Additional File 1: Figure S9A). However, for a few DMRs, levels of DNAm in SC/EC matched that of TB/CV, such as ones associated with the DNAm maintenance gene DNMT1 (Figure 2C) and FGF14 (Additional File 1: Figure S9B). Higher DNAm than TB/CV was only observed for 1 gene (RASGRF1, Additional File 1: Figure S10).
DNAm at repetitive elements, such as Alu and LINE1 elements, can be placental-specific and have been hypothesized to often be important regulatory components of placental processes (38,40). To determine if DNAm at repetitive elements is consistent across placental cell populations, we analyzed the subset of 850k CpGs that map to Alu (n=15,289) and LINE1 (22,006) elements. Compared to CV, TB had lower LINE1 DNAm (mean difference in DNAm = -1.5%, p=0.04), and HB had much higher DNAm (+9.7%, p<0.001; Additional File 1: Figure S11A). Similar relationships are seen for Alu elements (Additional File 1: Figure S11A). TB had lower (-1.2%, p=0.02), HB had higher (+7.0%, p<0.001), and EC had higher (+2.1%, p<0.001) DNAm in Alu CpGs, when compared to CV. To explore large-scale DNAm differences, we averaged DNAm across all 850k probes and compared each cell type to CV. We found these relationships to be similar to those with the subset of repetitive elements probes (Additional File 1: Figure S11A). HB had higher DNAm compared to CV (+5%, p<0.001), and all other cell types had lower DNAm (Additional File 7: Table S8). The relationships we found for repetitive elements and global DNAm between cell types and CV were also largely consistent in our first trimester samples (Additional File 7: Table S8, Additional File 1: Figure S11B). To determine genome-wide repetitive element DNAm, we used the random forest -based ‘REMP’ algorithm (41) to predict 438,664 Alu CpGs and 39,136 LINE1 CpGs that are not covered by the EPIC array. Relationships between cell types and CV for predicted and non-predicted repetitive elements were mostly the same, except TB DNAm in predicted Alu and LINE1 CpGs was not significantly different compared to CV (Additional File 7: Table S8, Additional File 1: Figure S11C).
Figure 4. DNA methylation at partially methylated domains (PMDs), and imprinted differentially methylated regions (DMRs). A) The percentage (y-axis) of CpGs in placental PMDs, falling into specific methylation intervals (0-20%, 20-40%, 40-60%, etc.) is shown for each cell type and trimester. B) DNAm across specific regions on chromosome 21 (B) and 4 (C). PMDs are highlighted with a grey background. D) Density plots (y-axis) of imprinted DMRs in term samples, divided into those that are imprinted in multiple tissues, (i.e. non-placental-specific; 1,085 CpGs total; top) and placental-specific (981 CpGs total; bottom). The percentage of CpGs falling within 25%-75% is shown above each plot. E) Cell-specific DNAm at the C19MC placental-specific imprinted DMR. This placental-specific imprint overlies a CpG island upstream of the miRNA cluster. F) DNAm at placental-specific imprinted region for DCAF10.
Cell-specific DNAm dynamics across gestation
To determine how DNAm changes in placental cell populations over gestation, we compared first and third trimester cell samples at 737,050 CpGs. We found 108,814 (TB); 94,619 (SC); 63,433 (EC) and 1,550 (HB) significant cell-specific gestational-age dependent DMCs (Bonferroni p<0.01, Db> 0.05). Strikingly, almost all of the TB DMCs show an increase in DNAm from first trimester to term (98.2%; Figure 5A). Most gestational-age DMCs for HB and SC also show an increase in DNAm from first trimester to term (75.6% and 56.6%, respectively). In contrast, EC DMCs show less DNAm in the term compared to first trimester (77.1%).
Several interesting KEGG pathways and GO terms were significant (FDR < 0.05) in our functional enrichment analysis (Figure 5BC). Immune pathways (“Cytokine-Cytokine receptor interactions”) and metabolism-related terms (“metabolic pathways”, “ATP binding”, “kinase activity”) for trophoblast gestational-age dependent DMCs suggest a highly active state throughout gestation affecting multiple placental functions. As expected, stromal terms were highly associated with cellular/tissue structure -related terms, such as “extracellular matrix organization”, and “Regulation of actin cytoskeleton”. No significant pathway or GO terms were found significant for HB gestational-age DMCs. Most gestational-age dependent DMCs were enriched with open sea regions, regardless of direction of methylation (Figure 5D). HB DMCs that increase in methylation with gestational age were the only cell type DMCs that were heavily enriched for enhancers (Bonferroni p<0.001). Trophoblast DMCs that increase in methylation with gestational age were enriched for CpG island shores, open seas, and intergenic regions (Bonferroni p<0.001). All cell type-specific gestational-age dependent DMCs were depleted (Bonferroni p<0.001) for promoter regions, suggesting that genome-wide promoter DNA methylation is mostly stable from first trimester to term.
Figure 5. Gestational-age dependent DNA methylation within each placental cell population. A) The distribution of the changes in DNA methylation between first and third trimester, within each cell type. Only statistically significant (Bonferroni p<0.01) and biologically relevant (mean change in DNAm > 5%) differences are shown. Number of gestational age associated DMCs are labelled above each plot. B) Functional enrichment analysis for GO terms tested with the R package missMethyl. C) Functional enrichment analysis for KEGG pathways tested with the R package missMethyl. D) Enrichment for genomic features: CpG island-related elements, enhancers, PMDs, and gene features.
Assessing cell composition in chorionic villi
Using placental cell DNAm profiles as a reference, we assessed cellular composition in CV samples using cellular deconvolution. To select cell-type discriminating CpGs, the pickCompProbes function from the R package minfi (42) was used, which takes the top 100 most hypo- and hyper-methylated CpGs ranked by F-test statistic for each cell type. Gestational-age specific references were created for first trimester and term. For first trimester samples, reference probes were selected from all first trimester cell samples, but also term nucleated red blood cells (nRBCs) and eSTB samples were used since these cell types are also present in early gestation (43–45). For nRBC samples, 11 DNAm profiles from umbilical cord blood from public databases were included (37,46,47). Reference CpGs determined from first trimester (Figure 6A) and term (Figure 6B) placental samples were highly cell-specific (Additional File 7: Table S9 and S10 for first trimester and term respectively).
To determine the best-performing cellular deconvolution method, 1,500 in silico bulk mixtures were generated based on our cell data with known cellular composition proportions. These deconvolution methods were compared: constrained projection (CP) (26), robust partial correlations (RPC) (25), and support-vector regression / CIBERSORT (CBS) (27). All three methods were tested using the implementation from the R package EpiDISH (39), and the constrained projection approach was used from implementations in both EpiDISH and minfi (40) R packages. Performance was high and consistent across algorithms and cell types (R2=0.88-0.99, RMSE=0.02-0.08, MAE=0.01-0.4; Additional File 1: Figure S12A; Additional File 7: Table S11). However, RPC slightly outperformed other approaches (R2=0.96, MAE=0.024, RMSE=0.045). Biases towards under-/over- estimation for certain cell types were small but were consistent across algorithms (Additional File 1: Figure S12B): SC tended to be overestimated (mean difference between estimated and actual = +0.33% to 0.98%), HB were underestimated (-0.03% to -0.38%). TB were underestimated (-0.07% to 0.94%), and nRBCs do not show as much bias (+0.03 to -0.21%).
To assess the validity of placental cell deconvolution estimates, we applied deconvolution to previously published placental samples that are enriched for specific cell populations. Deconvolution was applied to cultured trophoblast samples (n = 90) from Yuen et al. 2013 (12), that were cultured to 24 hours (predominantly CTB phenotype) or to 48 hours, after which many CTB cells have fused into STB; each set of samples was also subjected to varying oxygen levels (1%, 8%, 20%). Cultured STB had higher estimated STB relative to sample-matched cultured CTB (Figure S13A). The small changes in STB:CTB between culturing times are consistent with the small DNAm differences that were reported in Yuen 2013 (12), and suggest that although fusion of cytotrophoblast was achieved, further culturing would be required to produce a mature STB phenotype akin to term placenta. We then applied deconvolution to first, second trimester, and term enzymatically separated mesenchyme (n = 3) and matched samples of outer TB layer of chorionic villi (n = 3) from Hanna et al. 2013 (18), the latter of which were isolated in the same manner as eSTB in the present study. Despite batch effects and array differences (450k vs 850k), the term TB sample was estimated to be mostly syncytiotrophoblast (97%; Additional File 1: Figure S13B). Deconvolution estimates for trophoblast isolated from first and second trimester placentas were also mostly TB with some presence of the mesenchymal components, in particular some SC. Matched mesenchyme samples, as expected, were enriched for SC, and EC. Overall, these findings are consistent with our understanding that enzymatic separation enriches for certain populations but cannot produce homogenous cell populations. Lastly, we applied cell deconvolution to chorionic villi samples (n = 5) that were enriched for large visible stem villi. These samples had cell compositions that were heavily enriched for SC (mean = 51%, sd = 4%), compared to matched “normally” -processed chorionic villi (mean = 11%, sd = 2%; Additional File 1: Figure S13C).
RPC cellular deconvolution was applied to our 7 first trimester and 19 term CV samples. There was significant gestational-age specific variation in the estimated percentage of eSTB, TB, and SC (Table 3; Figure 6C). eSTB were the most abundant cell type in all (19/19) term samples (mean= 58%), whereas SC was the most abundant in most (5/7) first trimester samples (mean = 43%). There were significant changes from first trimester to term samples: there was a significant mean increase of 23% in eSTB (Bonferroni-adjusted p<0.001), a decrease in SC (-31%; adjusted p<0.001) and a small increase in EC (+5%; adjust p<=0.005). A detectable contribution of nRBCs was not estimated in any sample using RPC deconvolution. No significant (adjusted p>0.01) differences in cell composition were observed between male (n=9) and female (n=10) samples (Figure 6D; Additional File 7: Table S12), or between European/Caucasian (n=11) and East Asian (n=6) samples (Figure 6E; Additional File 7: Table S12) for term CV. Within-trimester gestational age (estimated and reported) was not significantly associated with cell composition (Additional File 7: Table S12), although numbers were small.
|
First (n=7)
|
Term (n=19)
|
|
|
|
Syncytiotrophoblast
|
35 (9)
|
58 (8)
|
Trophoblast
|
16 (12)
|
20 (6)
|
Stromal
|
43 (13)
|
12 (3)
|
Hofbauer cells
|
3 (2)
|
2.34 (1)
|
Endothelial
|
3 (2)
|
7 (1)
|
Nucleated red blood cells
|
0 (0)
|
0 (0)
|
|
|
|
Table 3: Mean of cell composition estimates (%) for first trimester and term CV samples using RPC cellular deconvolution. Standard deviation is shown in parentheses.
Figure 6. Assessing cell composition in first trimester and term CV samples. A) Mean DNAm across each cell type (columns) for 600 first trimester deconvolution reference CpGs selected by minfi::pickCompProbes. CpGs (rows) are hierarchically clustered using euclidean distance. B) Term reference CpGs. C) Cell composition of 7 first trimester and 19 third trimester CV samples, estimated with cellular deconvolution using RPC. D) Cell composition is similar between male and female term samples with respect to estimated percentage of each cell type (y-axis). F: female, M: male. E) Cell composition is similar between Asian and European/Caucasian third trimester samples.