Integrative analysis of the human brain mural cell transcriptome

Abstract

Brain mural cells, including pericytes and vascular smooth muscle cells, are important for vascular development, blood-brain barrier function, and neurovascular coupling, but the molecular characteristics of human brain mural cells are incompletely characterized. Single cell RNA-sequencing (scRNA-seq) is increasingly being applied to assess cellular diversity in the human brain, but the scarcity of mural cells in whole brain samples has limited their molecular profiling. Here, we leverage the combined power of multiple independent human brain scRNA-seq datasets to build a transcriptomic database of human brain mural cells. We use this combined dataset to determine human-mouse species differences in mural cell transcriptomes, culture-induced dedifferentiation of human brain pericytes, and human mural cell organotypicity, with several key findings validated by RNA fluorescence in situ hybridization. Together, this work improves knowledge regarding the molecular constituents of human brain mural cells, serves as a resource for hypothesis generation in understanding brain mural cell function, and will facilitate comparative evaluation of animal and in vitro models.

Introduction

Vascular smooth muscle cells (VSMCs) and pericytes, collectively termed mural cells, line the abluminal surface of blood vessels and regulate vascular development and function. VSMCs surround large vessels while pericytes line microvessels. In the brain, mural cells fulfill additional specialized roles as constituents of the neurovascular unit. VSMCs mediate neurovascular coupling,^1,2 while pericytes regulate blood-brain barrier (BBB) development and maintenance,^3–5 immune cell infiltration, ⁶ and potentially contribute to neurovascular coupling.^7–9 Further, pericytes provide trophic support to neurons via pleiotrophin (PTN) secretion, ¹⁰ and may regulate astrocyte end-foot polarization and the function of perivascular spaces.^4,11 Brain mural cell dysfunction or degeneration contributes to neurological diseases (reviewed in ¹² ) such as Alzheimer’s disease,^13–16 cerebral small vessel and associated white matter diseases, ¹⁷ and CADASIL (cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy), a disease caused by mutations in the mural cell-expressed NOTCH3 gene.^18–20 Despite the functional importance of mural cells, our knowledge of molecules that mediate key brain mural cell functions remains limited, with molecular and functional properties largely characterized in rodent models.^1–10 Therefore, improved knowledge of human brain mural cell gene expression profiles will expand understanding of the molecular mechanisms by which these cells contribute to physiological and pathological phenomena and facilitate molecular-level evaluation of animal and in vitro models.

Microarrays and bulk RNA-sequencing (RNA-seq) have been used to profile both mouse and human brain vascular cells. Most studies have focused on endothelial cells, although some studies characterized mural cells by marker-based isolation or subtractive comparison of microvessel and endothelial gene expression.^21–27 Single-cell RNA-seq (scRNA-seq) can provide comprehensive information about cellular diversity and cell type-specific gene expression, and lacks bias associated with isolation of cells based on expression of canonical markers.^28,29 The technique has been widely employed to profile cellular diversity in the mouse nervous system, including in recent studies that have provided detailed transcriptome profiles of mouse brain mural cells and related perivascular cell populations, such as fibroblast-like cells.^30–32 These studies have greatly advanced our knowledge of mouse brain mural cell biology and provided unambiguous molecular definitions of distinct mural cell populations (i.e., pericytes and multiple VSMC subtypes) that share expression of canonical marker genes and have historically been defined based predominantly on anatomical or morphological characteristics. An increasing number of studies have applied scRNA-seq to human brain samples,^33–37 but the scarcity of mural cells in whole brain samples is a major limitation in using these datasets to investigate human brain mural cell biology.

We reasoned that integrating the limited mural cell data from multiple human brain scRNA-seq datasets could reveal a more precise transcriptomic profile of human brain mural cells than is currently available. To this end, we made use of five human brain scRNA-seq datasets from the published literature^33–37 and employed computational tools that can identify common populations of cells across multiple independent scRNA-seq datasets.^38–40 Our results reveal high quality markers for human brain mural cells that are conserved across multiple brain regions and developmental stages. We also identify mural cell genes with human- or mouse-enriched expression, describe differences in gene expression profiles between in vivo and cultured human brain pericytes, and discover genes with enriched expression in brain mural cells compared to mural cells in other organs. Together, this work establishes high quality, consensus datasets for exploration of the human brain mural cell transcriptome.

Methods

RNA-seq datasets and analysis

We obtained scRNA-seq gene expression matrices and bulk RNA-seq FASTQ files from sources provided in Supplementary Table S8. We used R (version 3.6.2), the package Seurat (version 3.1.5),^38,39 and the package sctransform (version 0.2.1) ⁴⁰ for all analyses except where indicated. Statistical analyses of differential expression used the Wilcoxon rank sum test (a non-parametric test that does not assume normally-distributed data) with Bonferroni correction. Fold-change results are presented using the natural logarithm except where indicated. We also compared the expression of protein-coding genes in scRNA-seq data from in vivo pericytes and bulk RNA-seq data from cultured pericytes. A detailed description of all analyses and parameters is provided in Supplementary Methods.

Immunohistochemistry and RNA fluorescence in situ hybridization (FISH)

We validated a subset of brain mural cell-enriched genes using immunohistochemistry data from the Human Protein Atlas ⁴¹ (Supplementary Table S10). We performed immunohistochemistry on de-identified, normal human brain tissue and mouse brain tissue. For human tissue, the University of Wisconsin–Madison Institutional Review Board approved the experimental protocols in accordance with the guidelines of the federal Common Rule. All patients give informed consent for surgery at the University of Wisconsin Hospital which also includes a consent provision for use of leftover tissue removed during surgery for research purposes. For mouse tissue, the University of Wisconsin–Madison Institutional Animal Care and Use Committee approved the experimental protocols following National Institutes of Health (NIH) guidelines for care and use of laboratory animals. Results and methods are reported in compliance with the ARRIVE guidelines; inclusion/exclusion criteria, randomization, and blinding are not applicable to the study design. We also performed RNA FISH on human brain tissue obtained as described above. Statistical analysis of quantified RNA FISH data was performed using the Kruskal-Wallis test followed by Steel-Dwass test, non-parametric tests as data were not normally distributed as assessed by the Shapiro-Wilk test. The complete immunohistochemistry and RNA FISH protocols are provided in Supplementary Methods.

Results

Identification of mural cell populations in human brain scRNA-seq datasets

We began by identifying and evaluating five human brain scRNA-seq datasets from the published literature.^33–37 These five datasets were generated using samples from several different developmental stages, brain regions, and scRNA-seq library preparation methodologies (Figure 1(a)). Thus, our analysis of these datasets permits us to define a mural cell gene expression profile that is conserved across developmental stage and brain region. The five datasets also vary considerably in the total number of cells analyzed and in average sequencing depth (i.e., the average number of different genes detected) per cell (Figure 1(a)). Therefore, simultaneous analysis of multiple datasets may be beneficial, as datasets with high sequencing depth in few cells can complement datasets with large numbers of cells but low depth.

Figure 1.

Identification of mural cell populations in human brain scRNA-seq datasets. (a) Overview of the five scRNA-seq datasets. For each dataset, the sampled brain region(s), developmental stage, total number of cells analyzed, scRNA-seq platform, and sequencing depth (mean ± SD of genes detected per cell), are indicated. GW: gestational week. (b) UMAP plots of all single cells arranged in columns below the corresponding source dataset overview in (a). Plotted points (cells) are colored by cluster identity (top row) or by expression of PDGFRB or RGS5 (bottom). Dashed circles mark putative mural cell clusters. Cluster numbers shown correspond to those in Supplementary Figure S1 and Supplementary Table S1. (c) Venn diagram depicting the number of genes identified as markers of mural cells (P < 0.05, average log(fold change) > 0.25, expressed by at least 25% of cells in the putative mural cell cluster(s)) in the five scRNA-seq datasets. Complete lists of marker genes are provided in Supplementary Table S1. (d) Number of genes identified as markers of mural cells in a single dataset (1/5) or multiple datasets. 26 genes were identified as markers of mural cells in all datasets (5/5). (e) Average log(fold change) of gene expression in mural cell clusters compared to all other cells in the corresponding dataset for the 26 genes identified as markers of mural cells in all five datasets. Points represent average log(fold change) from each dataset and bars indicate the mean. (f) Genes identified as markers of mural cells in 4/5 datasets.

We first performed normalization, variable feature identification, and dimensionality reduction on each dataset independently, using the Seurat and sctransform packages^38,40 (see Methods). We visualized the heterogeneity in cell transcriptomes in two dimensions using UMAP embedding ⁴² and performed unbiased clustering of single cells (Figure 1(b)). We identified putative mural cell clusters based on the expression of pan-mural cell canonical markers PDGFRB and RGS5⁴³ or the VSMC-enriched gene ACTA2, ⁴³ and observed at least one cluster enriched for the combination of PDGFRB and RGS5 and/or ACTA2, in each dataset (Figure 1(b); Supplementary Figure S1). We next identified markers of these clusters, defined as genes enriched in the putative mural cell cluster(s) relative to all other cells in the same source dataset (P < 0.05, log(fold-change) > 0.25) and expressed by at least 25% of cells in the putative mural cell cluster(s) (Supplementary Table S1). Comparing the results of this analysis for the five datasets revealed that 26 genes were identified as mural cell markers in all five datasets (Figure 1(c) and (d)). Several additional genes were identified as mural cell markers in four out of five datasets, and many others appeared in three or fewer datasets (Figure 1(c) and (d)). Importantly, in addition to the canonical markers used to identify the putative mural cell clusters, the list of mural cell markers identified in all five datasets included several known pericyte and VSMC genes, including PTN, MYL9, and NOTCH3 (Figure 1(e)). In the list of genes identified as mural cell markers in four out of five datasets, additional known markers of brain mural cells appeared, including ABCC9, COL1A2, FOXC1, MCAM (CD146) and ZIC1,^{21,23,44–46} and shared mural/endothelial markers B2M and COL4A1^23,30 (Figure 1(f)). The lack of enrichment of these genes in all five data sets is likely attributable to the low sequencing depth of the dataset from Han et al., ³⁵ which lacks enrichment of these markers (Supplementary Table S1), rather than biological differences. Therefore, genes enriched in 4 or more datasets are likely high quality markers of human brain mural cells. Together, these results support the identification of mural cell clusters in the selected datasets, suggest striking similarities in marker gene expression in human brain mural cells across the evaluated developmental stages and brain regions, and provide genes that should robustly identify mural cells in human brain samples.

Integrated analysis of human brain mural cell transcriptome profiles

To take advantage of the combined power of the five scRNA-seq datasets and build a consensus molecular profile of human brain mural cells, we used the “anchoring” methodology in Seurat to integrate the five source datasets ³⁹ (see Methods). This combined dataset comprises more than 150,000 human brain cells. Visualization in low-dimensional space revealed both cell clusters that are unique to a single source dataset and cell clusters comprising cells from multiple or all source datasets (Figure 2(a)). We performed unbiased clustering on this combined dataset (Figure 2(b)) and used canonical markers of mural cells to identify cluster 30 as the putative mural cell cluster, which contained a total of 1,489 cells, representing 0.9% of total cells (Figure 2(b) and (c)). Importantly, cluster 30 contained cells derived from all five source datasets (Figure 2(a) and (b); Supplementary Figure S2(a)), suggesting gross similarity in the transcriptional profiles of mural cells from multiple brain regions and developmental stages. This finding is further supported by the lack of substantial spatial segregation of cerebellum- and temporal lobe-derived mural cells from the source dataset of Han et al. ³⁵ (Supplementary Figure S2(b)). The majority (77%) of cells from the mural cell clusters identified by analysis of independent datasets (Figure 1) were included in the integrated cluster 30 (Supplementary Figure S2(a)). Cells in cluster 30 expressed KCNJ8, a pericyte marker, ²¹ and ACTA2, a VSMC-enriched gene, suggesting that both of these mural cell subtypes are present in this cluster (Figure 2(c)). Of note, the mural cell cluster did not contain a large number of COL1A1⁺ cells, and we identified a separate, very small cluster (cluster 40) with enrichment of COL1A1 that may comprise brain perivascular fibroblast-like cells^{30, 32} (Figure 2(c)). We identified markers (as defined previously) of mural cells in cluster 30 and fibroblast-like cells in cluster 40 (Figure 2(d); Supplementary Table S2). Among the mural cell-enriched genes were all 26 markers previously identified in all five individual dataset analyses, a majority of markers identified in four out of five datasets, and importantly, additional genes not identified in any individual analysis, including HCFC1R1 and CYB5R3 (Supplementary Figure S2(c); Supplementary Table S2). We performed Gene Set Enrichment Analysis (GSEA) using the results of this differential expression analysis, and identified the KEGG gene sets focal adhesion and ECM-receptor interaction as enriched in both mural and fibroblast-like cell clusters, while oxidative phosphorylation and vascular smooth muscle contraction were enriched only in the mural cell cluster (Figure 2(e); Supplementary Table S2). We also identified an endothelial cell cluster comprising 768 cells (cluster 32; 0.5% of total cells) with enrichment for known endothelial and BBB genes including CLDN5, SLC2A1, VWF, FLT1, MFSD2A, ABCG2, and PODXL (Supplementary Figure S3(a) to (e); Supplementary Table S2). Finally, we identified transcription factors enriched in these three vascular cell clusters; we observed mural cell-specific enrichment of FOXS1, EBF1, ZEB1, TBX2, and HEYL, several shared mural- and fibroblast-enriched transcription factors (including ZIC1, HES4, and FOXF2), and pan-vascular enrichment of intermediate early genes (including JUNB, NR4A1, FOS, and JUN), potentially suggestive of dissociation-induced transcriptional changes (Supplementary Figure S3(f)).

Figure 2.

Dataset integration and analysis of mural and fibroblast transcriptional profiles. (a) UMAP plot of all single cells after integration of the five scRNA-seq datasets described in Figure 1. Plotted points (cells) are colored by source dataset. An example of a cluster comprising cells from a single source dataset is circled; most clusters comprise cells from multiple source datasets. (b) UMAP plot of all single cells colored by cluster identity after re-clustering. Dashed circles mark the putative mural cell cluster (30) and fibroblast cluster (40). (c) Expression of known markers of mural cells (PDGFRB, RGS5), pericytes (KCNJ8), VSMCs (ACTA2), and fibroblasts (COL1A1) in the 42 clusters identified in (b). The position of the putative mural cell cluster (30) is indicated by an arrow below the x axis. (d) Genes identified as markers (P < 0.05, average log(fold change) > 0.25, expressed by at least 25% of cells in the putative mural cell or fibroblast cluster(s)) of mural cells (left) and fibroblasts (right). The average log(fold change) of gene expression in the mural or fibroblast cell cluster compared to all other cells is plotted for the 30 genes with largest average log(fold change). Complete lists of marker genes are provided in Supplementary Table S2. (e) Gene sets enriched and depleted (P < 0.05) in mural (top) and fibroblast (bottom) clusters compared to all other brain single cells as identified by GSEA. NES: normalized enrichment score. The GSEA input was a list of genes ranked from the highest-confidence mural- or fibroblast-enriched to the highest-confidence mural- or fibroblast-depleted gene, using the ranking metric –log₁₀(P)×log(fold change). Complete results of GSEA are provided in Supplementary Table S2. (f) UMAP plot of all mural cells (cluster 30 in b) after sub-clustering and cluster collapsing to two clusters, pericytes (cluster 0′) and VSMCs (cluster 1′) (Supplementary Figure S2). UMAP plots split by source dataset are shown at right. Plotted points (cells) are colored by cluster identity. The number of cells in the integrated mural cell dataset derived from each source dataset is shown above each plot. (g) Differential expression analysis comparing pericytes (cluster 0′) and VSMCs (cluster 1′). The heatmap shows expression of the 10 genes with the largest average log(fold change) for each cell type. Complete results of differential expression analysis are provided in Supplementary Table S3. (h) UMAP plots of all mural cells. Plotted points (cells) are colored by expression of PTN or TAGLN. PTN-expressing cells concentrate in the lower region identified as cluster 0′ (pericyte), while TAGLN-expressing cells concentrate in the upper region identified as cluster 1′ (VSMC).

We next asked whether sub-clustering the mural cell cluster could resolve distinct populations of pericytes and VSMCs. Initial sub-clustering yielded 12 clusters, some of which were largely driven by source dataset (Supplementary Figure S2(d)). We therefore collapsed the five ACTA2^high clusters into one putative VSMC-enriched sub-cluster (cluster 1′) and the remaining seven ACTA2^low clusters into a single putative pericyte-enriched subcluster (cluster 0′) (Supplementary Figure S2(e) and (f)). We also eliminated a small number of potential endothelial, neuronal, glial, microglial, and blood cell contaminants or multiplets based on expression of canonical markers (Supplementary Figure S2(g)), for a filtered mural cell cluster containing 1,182 cells (Figure 2(f)). When visualized in the original UMAP embedding, cells in the pericyte sub-cluster and cells in the VSMC sub-cluster appeared spatially segregated (Figure 2(f)). We performed differential expression analysis to identify genes that distinguish the two mural sub-clusters, and found that putative pericytes were enriched for genes including PTN, ATP1A2, PDGFRB, PLXDC1, SLC6A12, SLC6A1, MALAT1, and FN1, while putative VSMCs were enriched for genes including ACTA2, TAGLN, ADIRF, S100A6, and DSTN, many of which have been previously identified to discriminate between mouse brain mural cell sub-types (Figure 2(g) and (h); Supplementary Figure S4; Supplementary Table S3). We also evaluated a subset of these markers on the protein level using immunohistochemistry data from the Human Protein Atlas, ⁴¹ which confirmed robust expression of SLC6A12 and weak expression of SLC6A1 in smaller vessels, and enrichment of S100A6 and ADIRF in larger vessels across multiple brain regions (Supplementary Figure S5). While all source datasets contained cells classified as both pericytes and VSMCs, the vast majority of mural cells were classified as pericytes in all datasets except that from Han et al. ³⁵ (Figure 2(f)), potentially a result of the brain regions sampled, technical differences in tissue isolation, or increased pial vessel contamination. Taken as a whole, however, these combined human pericyte and VSMC datasets are derived from multiple independent studies and exhibit good representation of known pericyte and VSMC markers. We therefore propose that these datasets will be useful in identifying novel mural cell genes and in validating appropriate expression of these genes in model systems of these cell types.

Mouse-human species differences in brain pericyte gene expression

We next applied our combined human brain pericyte single cell transcriptomic dataset to evaluate potential mouse-human species differences in brain pericyte gene expression. We used a mouse brain vascular scRNA-seq dataset^30,31 for initial comparison (Figure 3(a)). We chose to focus on pericytes given that all five human source datasets contributed substantially to the pericyte sub-cluster, and the average sequencing depth of human and mouse pericytes was more closely matched than that of human and mouse VSMCs (Supplementary Figure S6(a)). We visualized the mouse dataset by UMAP embedding and overlaid the authors’ cell type annotations (Figure 3(b)) and expression of the mural cell marker genes Pdgfrb, Rgs5, and Acta2 (Figure 3(c)). Mural cells, including pericytes (PC) and arterial, arteriolar, and venous VSMCs (aSMC, aaSMC, and vSMC, respectively), formed a cluster distinct from endothelial cells, fibroblast-like cells, and small populations of other cell types (Figure 3(b)). The VSMC cluster was spatially distinct from the pericyte cluster and was Acta2⁺ (Figure 3(b) and (c)), consistent with the authors’ report. ³⁰

Figure 3.

Human-mouse species differences in brain pericyte gene expression. (a) Overview of the mouse brain vascular scRNA-seq dataset. The sampled brain region, developmental stage, total number of cells analyzed, scRNA-seq platform, and sequencing depth (mean ± SD of genes detected per cell), are indicated. OB: olfactory bulb. (b) UMAP plot of mouse brain vascular single cells. Plotted points (cells) are colored by cluster identity. Cluster identities are as assigned by Vanlandewijck, He et al. ³⁰ , ³¹ OL: oligodendrocytes; FB1, FB2: fibroblast-like cells; aEC: arterial ECs; vEC: venous ECs; capilEC: capillary ECs; EC1, EC2, EC3: other EC subtypes; MG: microglia; PC: pericytes; AC: astrocytes; aSMC: arterial VSMCs; aaSMC: arteriolar VSMCs; vSMC: venous VSMCs. (c) UMAP plots of mouse brain vascular single cells. Plotted points (cells) are colored by expression of Pdgfrb, Rgs5, or Acta2. (d) Volcano plot illustrating differential expression analysis of human brain pericytes (Figure 2(f), cluster 0’) and mouse brain pericytes. Mitochondrial genes, ribosomal genes, genes with log(fold change) < 0.25, and genes expressed by fewer than 50% of human or mouse pericytes were excluded. Complete results of differential expression analysis are provided in Supplementary Table S4. (e) Expression of selected genes in human and mouse brain pericytes and VSMCs. Genes with expression consistent across species (top row), mouse-enriched expression (P < 0.05; middle row), and human-enriched expression (P < 0.05; bottom row) are shown.

We used differential expression analysis to compare homologous gene expression in human brain pericytes (cluster 0′ in Figure 2(f)) versus mouse brain pericytes (cluster PC in Figure 3(b)). We identified 541 mouse-enriched genes, including Atp13a5, Slc6a20a, and Sod3, and 168 human-enriched genes including DCN, PPIA, FRZB, SLC6A1, FN1, and MGLL (P < 0.05, log(fold change) > 0.25, expressed by at least 50% of cells in either the human or mouse pericyte cluster), excluding mitochondrial and ribosomal genes (Figure 3(d); Supplementary Table S4). A majority of human and mouse pericytes expressed PDGFRB, RGS5, and PTN, a majority of human and mouse VSMCs expressed ACTA2, and a similar proportion of human and mouse VSMCs expressed CNN1 (i.e., putative arterial VSMCs) (Figure 3(e)). RGS4, SNX33, VTN, PLXDC2, and SLC22A8 were strongly enriched in mouse pericytes compared to human pericytes, and all except SLC22A8 were also expressed by some mouse VSMCs and nearly no human VSMCs (Figure 3(e)). Vtn is robustly and selectively expressed in mouse brain pericytes, and absence of VTN transcript in human brain pericytes is consistent with our previous observations. ²⁷ DCN (encoding decorin), FN1 (fibronectin), PPIA (proinflammatory cytokine cyclophilin A), FRZB (Wnt-binding frizzled-related protein), and SLC6A1 (GABA transporter) were strongly enriched in human pericytes compared to mouse pericytes, and similar to the mouse-enriched case, many were also enriched in human VSMCs compared to mouse VSMCs (Figure 3(e)). In adult mouse brain, Dcn and Fn1 are enriched in fibroblast-like cells and endothelial cells, respectively. ³⁰ While DCN is also expressed by human fibroblast-like cells (Figure 2(d)) and FN1 by human brain endothelial cells (Supplementary Figure S3(c)), human-enrichment of these genes in pericytes suggests that the pericyte contribution to the vascular basement membrane may differ between human and mouse.

We validated a subset of human- and mouse-enriched genes using an additional mouse scRNA-seq dataset, ³² identifying the authors’ cluster PER3 as containing a relatively pure population of brain pericytes (Pdgfrb⁺Rgs5⁺Cldn5^–Acta2^–Col1a1^–) (Supplementary Figure S6(c)). Consistent with the mouse dataset used for differential expression analysis, cells in the PER3 cluster had little to no expression of Slc6a1, Slc6a12, Frzb, Dcn, and Fn1 and did express Slc6a20a and Vtn (Supplementary Figure S6(c)). Some pericytes in the PER3 cluster (∼43%) expressed Ppia, however, suggesting that populations of both human and mouse pericytes express this gene, consistent with previous protein-level and functional observations. ¹⁵ Together, these data validate several human brain pericyte markers previously identified in mouse, but also highlight key transcriptional differences between brain pericytes in these two species. These species-specific pericyte genes motivate future examination of functional differences between human and mouse brain pericytes, for example, in regulation of Wnt signaling (FRZB), GABA uptake/transport (SLC6A1), and vascular extracellular matrix composition. Finally, this human brain mural cell dataset can aid in the design and interpretation of experiments in murine models.

Transcriptional alterations in brain pericytes cultured in vitro

Primary cell culture is widely employed to study human brain pericytes, and we therefore asked whether our integrated human brain pericyte scRNA-seq dataset could be used to reveal potentially important molecular differences between in vivo and cultured pericytes. Cultured primary human brain pericytes from commercial sources (see Methods) are widely used and have been analyzed by bulk RNA-seq in several independent studies.^47–49 To compare these data to the gene expression profile of brain pericytes in vivo, we constructed mock bulk RNA-seq datasets from the single cells identified as pericytes (cluster 0′ in Figure 2(f)) from each source dataset (see Methods). The Pearson correlation between average log-transformed transcript abundances from cultured and in vivo pericytes was moderately high (r = 0.81), although numerous dysregulated genes were apparent (Figure 4(a); Supplementary Table S5). For example, while PDGFRB was moderately downregulated in vitro, the well-established pericyte marker KCNJ8 was highly and consistently downregulated in cultured pericytes from all three studies (Figure 4(a) and (b)). P2RY14, COL9A1, SLC6A12, and HIGD1B were similarly downregulated in cultured pericytes (Figure 4(a) and (b)). Among genes consistently upregulated in cultured pericytes were COL1A1, which is a marker of fibroblasts in vivo (Figure 2(c)), several additional collagens (COL8A1, COL11A2), and the protease inhibitor SERPINE1 (Figure 4(a) and (b)). GSEA revealed that the KEGG gene set ECM-receptor interaction contained genes that were both highly upregulated and downregulated in cultured pericytes, effects driven by collagens, integrins, perlecan, and laminins, among others (Figure 4(c); Supplementary Table S6). The KEGG gene set neuroactive ligand-receptor interaction was strongly enriched in the in vivo pericyte dataset, an effect driven by metabotropic glutamate receptor genes (GRM3, GRM8), the purinurgic receptor P2RY14 identified above, the endothelin receptor EDNRB, and the sphingosine-1-phosphate receptor S1PR3, among others (Figure 4(c)). Collectively, these data suggest that while cultured human brain pericytes retain expression of some key markers and have moderately well-correlated global gene expression profile, they downregulate receptors for neuroactive ligands when compared to pericytes in vivo and have markedly dysregulated extracellular matrix (ECM)-associated gene expression.

Figure 4.

Transcriptional alterations in brain pericytes cultured in vitro. (a) Comparison of gene expression in cultured human pericytes as quantified by bulk RNA-seq and in vivo human pericytes as quantified by constructing a mock bulk RNA-seq dataset from in vivo pericyte scRNA-seq data (Figure 2(f), cluster 0’). The Pearson correlation coefficient (r) is shown. Red lines indicate log₂(fold change) = ±2. Complete expression data are provided in Supplementary Table S5. (b) Comparison of expression of selected genes in mock bulk RNA-seq data from in vivo pericytes (Figure 2(f), cluster 0’) and bulk RNA-seq data from cultured pericytes. Average expression for each of the five in vivo pericyte datasets is indicated with a gray point. Expression in five bulk RNA-seq datasets from cultured pericytes are indicated with green points. Bars indicate the mean. (c) GSEA enrichment plots for the ECM-receptor interaction and neuroactive ligand-receptor interaction gene sets. The GSEA input was a list of genes ranked from cultured pericyte-enriched to in vivo pericyte-enriched gene based on log(fold change). The 10 genes in each gene set with highest and/or lowest rank are indicated at right. Complete results of GSEA are provided in Supplementary Table S6. (d) Overview of the cultured human brain pericyte scRNA-seq dataset. The sampled brain region, developmental stage, total number of cells analyzed, scRNA-seq platform, and sequencing depth (mean ± SD of genes detected per cell) are indicated. (e) Expression of selected genes in in vivo pericytes (Figure 2(f), cluster 0’) and cultured pericytes as described in (d) and Supplementary Figure S7. Genes with culture-induced downregulation (left) and genes with culture-induced upregulation (right) are shown.

To address the possibility that upregulation of COL1A1 and other ECM-related genes actually reflects proliferation of contaminating fibroblast-like cells in pericyte cultures, rather than culture-induced dedifferentiation of pericytes, we analyzed scRNA-seq data from cultured human brain pericytes ⁵⁰ (Figure 4(d)). We selected control (non-transfected) pericytes that were acutely isolated and cultured for two days in vitro for further analysis (Supplementary Figure S7(a)). The control cells formed two spatially distinct clusters when visualized by UMAP embedding, but both clusters expressed PDGFRB and RGS5 (Supplementary Figure S7(b) and (c)). We compared expression of key genes in human brain pericytes in vivo (cluster 0′ in Figure 2(f)) and these cultured human brain pericytes. While PDGFRB and RGS5 were downregulated in cultured pericytes, nearly all cells still expressed these genes; conversely, the pericyte marker KCNJ8 was downregulated in cultured pericytes to the point where most cells did not have detectable expression of this gene (Figure 4(e)). Similarly, most cultured pericytes did not express detectable levels of SLC6A1 and SLC6A12 (Figure 4(e)). Consistent with previous observations that pericytes rapidly upregulate α-smooth muscle actin (α-SMA) in vitro, ⁵¹ the cultured pericytes had elevated levels of ACTA2 transcript (Figure 4(e)). Finally, all cultured pericytes expressed COL1A1 and many expressed COL8A1 and other collagens that were not expressed by pericytes in vivo (Figure 4(e)). The presence of RGS5⁺COL1A1⁺ cells suggests that the dysregulated ECM gene signature identified in bulk RNA-seq of cultured brain pericytes is not the result of a population of contaminating COL1A1⁺ fibroblast-like cells. Together, these data support the notion that culture induces dedifferentiation/activation of brain pericytes marked by downregulation of neuroactive receptors and upregulation of fibroblast and VSMC gene signatures, among other transcriptional changes. Thus, these datasets should be useful for detailed evaluation of in vitro models based on presence or absence of relevant genes.

Human mural cell organotypicity

The specialized functions of brain mural cells in regulating the neurovascular unit are likely established by brain-selective gene expression. We therefore attempted to use the integrated human brain mural cell scRNA-seq dataset (Figure 2(f)) to identify shared and distinct gene expression profiles by comparison to mural cells in human liver, lung, heart, and skeletal muscle scRNA-seq datasets.^52–55 Characteristics of these datasets are shown in Supplementary Figure S8(a). As before, we performed dimensionality reduction and unbiased clustering of the cells in each dataset, and identified putative mural cell clusters by visualizing expression of the canonical markers PDGFRB, RGS5, and ACTA2 (Supplementary Figure S8(b)). Because some datasets had too few cells or inadequate sequencing depth to fully resolve distinct mural cell subtypes (i.e., pericytes and VSMCs), we instead elected to make comparisons between the combined mural cell clusters from different organs. We identified markers of the mural cell clusters from each organ, defined as genes enriched in the putative mural cell cluster(s) relative to all other cells in the same organ (P < 0.05, log(fold change) > 0.25, expressed by at least 25% of cells in the putative mural cell cluster(s)) (Supplementary Table S7). 17 genes were identified as markers of mural cells in all five organs (Figure 5(a) and (b)): in addition to PDGFRB and RGS5, additional highly-enriched genes included IGFBP7, TPM2, CALD1, BGN, and SPARC (Figure 5(c)). An additional 37 genes were identified as markers of mural cells in four out of five organs, including known markers such as ACTA2, COL4A1, MYL9, NOTCH3, and TAGLN (Figure 5(d)). Further, KCNJ8 was identified as a mural cell marker in four out of five organs, extending the potential utility of this gene that was previously identified as a pericyte marker in the brain ²¹ (Figure 5(d)).

Figure 5.

Shared and distinct gene expression profiles of mural cells of multiple organs. (a) Venn diagram depicting the number of genes identified as markers of mural cells (P < 0.05, average log(fold change) > 0.25, expressed by at least 25% of cells in the putative mural cell cluster(s)) in the scRNA-seq datasets from five organs. The brain mural cells analyzed were from the integrated dataset (Figure 2(b), cluster 30). Overviews and cluster identification for the heart, liver, lung, and skeletal muscle scRNA-seq datasets are shown in Supplementary Figure S8. Complete lists of marker genes are provided in Supplementary Table S7. (b) Number of genes identified as markers of mural cells in a single organ (1/5) or multiple organs. 17 genes were identified as markers of mural cells in all organs (5/5). (c) Average log(fold change) of gene expression in mural cell clusters compared to all other cells in the corresponding organ for the 17 genes identified as markers of mural cells in all five organs. Points represent average log(fold change) from each organ and bars indicate the mean. (d) Genes identified as markers of mural cells in 4/5 organs. (e) Dot plot of gene expression in mural cell clusters from multiple organs. Genes with shared expression across mural cells of the five organs and genes with brain-enriched expression are shown. Color indicates expression level and dot size indicates the fraction of cells that express a given gene. (f) Expression of selected brain mural cell-enriched genes in pericytes from the five human brain scRNA-seq datasets.

We next compared gene expression in the mural cell clusters between the five organs to identify potential organ-specific differences. We focused on identifying genes with brain-enriched expression, as the mural cell clusters from some peripheral organ scRNA-seq datasets appeared to contain multiplets or non-resolvable populations of non-mural cells, exemplified by expression of the hepatocyte marker ALB in the liver mural cell cluster (Supplementary Figure S8(c)). We found several genes with enriched expression in brain mural cells (P < 0.05, log(fold change) > 0.25, excluding mitochondrial and ribosomal genes) including ATP1A2 (encoding a Na⁺/K⁺ ATPase subunit), SLC6A1 and SLC6A12 (GABA/betaine transporters), GPER1 (G protein-coupled estrogen receptor), ZIC1 (neural crest lineage transcription factor), and NTM (neurotrimin, a GPI-anchored cell adhesion molecule) (Figure 5(e); Supplementary Table S7). Importantly, the brain-enriched mural cell genes ATP1A2, SLC6A1, SLC6A12, and GPER1 were consistently expressed in pericytes from all five human brain scRNA-seq source datasets (Figure 5(f)), suggesting that these genes are expressed across multiple brain regions and developmental stages. Further, we observed vascular localization of SLC6A1, SLC6A12, and GPER1 immunoreactivity in Human Protein Atlas data from multiple brain regions (Supplementary Figure S5). We also found that a majority of brain-enriched mural cell genes were downregulated in cultured pericytes (Supplementary Figure S8(d); Supplementary Table S7), supporting the possibility that environmental cues are required to maintain organ-specific gene expression. Together, this analysis permits identification of brain mural cell-enriched genes, which may yield hypotheses for mechanisms underlying the unique functions of brain mural cells. Our combined brain scRNA-seq dataset should facilitate further identification of mural cell gene expression organotypicity as additional high-resolution scRNA-seq datasets from human peripheral organs become available.

Validation of human brain mural cell genes

We selected FRZB, ATP1A2, SLC6A1, and SLC6A12, genes identified in the bioinformatic analyses described above, for further validation. We assessed FRZB expression because it was identified as a brain mural cell marker, enriched in human mural cells versus mouse, and also expressed by human mural cells of other organs (Supplementary Table S7). This Wnt-binding protein might help modulate Wnt-mediated CNS angiogenesis and barriergenesis. We assessed ATP1A2 expression because it was identified as a brain mural cell marker, enriched in pericytes versus VSMCs, and was the gene with the highest enrichment in human brain mural cells compared to mural cells of other organs. Furthermore, mutations in ATP1A2 are associated with familial hemiplegic migraine and potentially other neurological disorders (reviewed in ⁵⁶ ) Thus, while ATP1A2 is also expressed by neurons and glia, ⁵⁶ confirming pericyte expression of this gene may advance understanding of cell type-specific contributions to these disorders. Finally, we assessed SLC6A12 and SLC6A1 expression given that both genes were identified as brain mural cell markers, highly enriched in pericytes compared to VSMCs, enriched in human mural cells versus mouse, absent in cultured pericytes, and absent in mural cells of other organs. These solute carriers might fulfill functional roles in GABA or betaine uptake by human pericytes. Importantly, all four of these putative mural cell markers were not identified as shared markers of endothelial cells (Supplementary Table S2).

We used RNA fluorescence in situ hybridization (RNAscope) to evaluate transcript expression in flash-frozen human brain neurosurgical samples. To identify mural cells and closely-associated endothelial cells, we employed the canonical markers PDGFRB and CLDN5, respectively. In resulting images, we observed sparsely distributed PDGFRB⁺ and CLDN5⁺ nuclei, which in some cases were closely associated (Figure 6(a) and (e)). We observed colocalization of FRZB and ATP1A2 with PDGFRB⁺, but not CLDN5⁺, nuclei (Figure 6(a) to (c); Supplementary Figure S9), with significant enrichment of FRZB and ATP1A2 mean fluorescence intensity in PDGFRB⁺ nuclei compared to parenchymal (CLDN5^–PDGFRB^–) nuclei (Figure 6(d)). While robust ATP1A2 signals were associated with most PDGFRB⁺ nuclei, we observed a comparative lack of ATP1A2 signal in PDGFRB⁺ nuclei in a large diameter vessel (Supplementary Figure S9), consistent with enriched expression in pericytes compared to VSMCs. We also observed colocalization of SLC6A12 and SLC6A1 with PDGFRB⁺ mural cell nuclei (Figure 6(e) to (g); Supplementary Figure S10), with significant enrichment of SLC6A1 mean fluorescence intensity in PDGFRB⁺ nuclei compared to parenchymal (CLDN5^–PDGFRB^–) nuclei (Figure 6(h)). These in situ data therefore support the scRNA-seq-based identification of FRZB, ATP1A2, SLC6A1, and SLC6A12 as novel human brain mural cell-expressed genes. Finally, to validate a human-mouse differentially expressed gene on the protein level, we stained human and mouse brain tissue sections for FRZB. Consistent with scRNA-seq data, we detected FRZB colocalization with PDGFRβ⁺ mural cells in human brain, but did not detect vascular FRZB immunoreactivity in mouse brain, although mouse olfactory ensheathing cells, which express Frzb transcript,^57,58 did have FRZB immunoreactivity (Supplementary Figure S11).

Figure 6.

RNA fluorescence in situ hybridization validates novel human mural cell genes. (a) Expression of CLDN5 and PDGFRB mRNA in a human brain tissue section. CLDN5⁺ endothelial cells are indicated with arrows; PDGFRB⁺ mural cells are indicated with arrowheads. DAPI nuclear counterstain is also shown. Scale bar: 100 μm. (b) Same image field as in (a) with expression of FRZB and ATP1A2 overlaid. Dashed boxes indicate regions displayed in (c). Scale bar: 100 μm. (c) Enlarged regions 1 and 2 of image as indicated in (b). Merged image displays all five channels with colors as defined in (b). Data from an additional human brain sample are shown in Supplementary Figure S9. (d) Quantification of DAPI, CLDN5, PDGFRB, FRZB, and ATP1A2 mean fluorescence intensity in endothelial cell nuclei (CLDN5⁺, n = 5), mural cell nuclei (PDGFRB⁺, n = 12), and non-endothelial/non-mural cell nuclei (CLDN5^–PDGFRB^–, n = 42). Mean ± SD is shown. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 versus CLDN5^–PDGFRB^–, Kruskal-Wallis test followed by Steel-Dwass test. (e) Expression of CLDN5 and PDGFRB mRNA in a human brain tissue section. CLDN5⁺ endothelial cells are indicated with arrows; PDGFRB⁺ mural cells are indicated with arrowheads. DAPI nuclear counterstain is also shown. Scale bar: 100 μm. (f) Same image field as in (e) with expression of SLC6A12 and SLC6A1 overlaid. Dashed box indicates region displayed in (g). Scale bar: 100 μm. (g) Enlarged region of image as indicated in (f). Merged image displays all five channels with colors as defined in (f). Data from an additional human brain sample are shown in Supplementary Figure S10. (h) Quantification of DAPI, CLDN5, PDGFRB, SLC6A12, and SLC6A1 mean fluorescence intensity in endothelial cell nuclei (CLDN5⁺, n = 6), mural cell nuclei (PDGFRB⁺, n = 15), and non-endothelial/non-mural cell nuclei (CLDN5^–PDGFRB^–, n = 35). Mean ± SD is shown. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 versus CLDN5^–PDGFRB^–, Kruskal-Wallis test followed by Steel-Dwass test.

Discussion

Brain mural cells play important roles in neurovascular function, but remain incompletely characterized, especially in humans. scRNA-seq has the potential to reveal detailed, cell type-specific gene expression profiles that will aid in classifying distinct cell populations and identifying mechanisms of cellular function, but existing human brain scRNA-seq datasets contain exceedingly few mural cells. Further, mural cell scRNA-seq profiles are frequently underanalyzed in studies focusing on neuronal or glial diversity, yet recent studies have begun to employ scRNA-seq to better understand roles of brain mural cells in off-target effects of CAR-T therapy ⁵⁹ and viral tropism. ⁶⁰ Here, we integrated multiple scRNA-seq datasets to increase sample size and mitigate the potential impact of technical artifacts in single datasets, exemplified by the exclusion of PDGFRB^low cells previously assigned to a mural cluster in the individual analysis of a low sequencing depth dataset (Supplementary Figure 2(a)). We analyzed the gene expression profiles of human brain mural cells in five scRNA-seq datasets that differ in brain region (neocortex, hippocampus, midbrain, cerebellum) and developmental stage (GW6 through adult).^33–37 Our analysis suggests broad similarity in mural cell marker gene expression across the analyzed developmental stages and brain regions. This observation is consistent with previous scRNA-seq analysis demonstrating that mouse brain vascular and oligodendrocyte gene expression profiles were largely brain region-independent, in contrast to neurons and astrocytes, ³² and that clustering of mural cells is driven predominantly by zonation along the vascular tree. ³⁰ While there may be brain region-dependent differences in mural cell gene expression not identified in our analysis, additional datasets, or scRNA-seq studies that directly compare brain regions, will be required to identify such differences. Similarly, additional studies will be required to understand changes in mural cell gene expression during development, ageing, and disease, factors which have begun to be examined in brain endothelium^26,61 and are key to potential therapeutic targeting of mural cells.

In the combined mural cell dataset, we observed separation between ACTA2-enriched and PTN-enriched cells, which we classified into VSMC and pericyte clusters, respectively, based on previous classifications in mouse scRNA-seq studies.^30,32 In addition to capillary pericytes, our pericyte cluster also contains a small number of PTN⁺ACTA2⁺ cells, which are likely the mural cells of postcapillary venules, which have been termed both pericytes ⁶² and venular VSMCs.^1,30 In mouse scRNA-seq data, these cells also express Ptn and are transcriptionally more similar to the cells identified as pericytes than those identified as arterial and arteriolar VSMCs. ³⁰ Though we observed clear pericyte-VSMC separation, we were not able to further resolve distinct VSMC subtypes (e.g., venous, arterial, arteriolar) that have been previously observed in mouse, ³⁰ likely because the combined dataset contains only a small number of VSMCs, with most derived from a low sequencing depth dataset. ³⁵ Though the expression of α-SMA (encoded by ACTA2) in brain capillary pericytes is a subject of debate in the field, ⁶³ as is the definition of a brain pericyte,^1,62,64 our data support the presence of a relatively large population of ACTA2^– cells. Deeper sequencing better able to detect transcripts with very low abundance and single cell proteomics approaches ⁶⁵ may aid in reconciling these observations. However, the ability of scRNA-seq to clarify questions surrounding mural cell identities is inherently limited by its lack of spatial information. Improved characterization of different mural cell populations demands additional spatial data to facilitate linking of mural cell functional properties (e.g., contractility) and protein expression, which vary along the vascular tree, to scRNA-seq-derived transcriptome profiles. Emerging spatial transcriptomics approaches^66–68 may facilitate such mapping. Once additional spatial, protein-level, and functional data exist for murine brain mural cells to facilitate correlations with existing mouse scRNA-seq data, our combined human scRNA-seq dataset will be useful in comparing and interpreting such results. Finally, we also identified a distinct population of COL1A1⁺ cells, which likely represent perivascular fibroblast-like (vascular leptomeningeal) cells or meningeal fibroblasts, which are extremely similar on the transcriptome level.^30,69 Although perivascular fibroblast-like cells are not mural cells, these cells share a proposed neural crest-meningeal developmental origin^69–72 and further molecular characterization of these cells is warranted given their role in scar formation after injury and disease^73,74 and poorly understood physiological function.

We applied the integrated dataset to assess differences between human brain mural cells in vivo and (i) mouse brain mural cells, (ii) human brain mural cells cultured in vitro, and (iii) human mural cells from other organs. We first identified species differences in gene expression, including in solute carrier-encoding genes, of which SLC6A1 and SLC6A12 were human-enriched and Slc6a20a was mouse-enriched, extracellular matrix-associated genes, of which DCN and FN1 were human-enriched and Vtn was mouse-enriched, and signaling-associated genes, of which FRZB was human-enriched and Rgs4 and Plxdc2 were mouse-enriched. These results corroborate and extend previous observations of species differences in brain microvessel gene expression. ²⁷ We also compared in vivo human brain pericytes to cultured primary human brain pericytes, which are widely employed as an in vitro model.^47–49 Notably, while cultured pericytes retained PDGFRB and RGS5 expression, they lacked expression of the pericyte marker KCNJ8, receptors for neuroactive ligands such as P2RY14, and the human-enriched genes SLC6A1 and SLC6A12. In addition to elevated expression of ACTA2, which has been previously observed, ⁵¹ cultured pericytes also had dysregulated expression of extracellular matrix-associated genes, with aberrantly high expression of fibroblast-associated genes such as COL1A1 and COL8A1. These results were consistent in scRNA-seq data from primary human brain pericytes that were acutely isolated and briefly cultured, ⁵⁰ suggesting that pericytes undergo fairly rapid dedifferentiation to a mixed mural/fibroblast-like phenotype. These findings of brain mural cell dedifferentiation in culture complement existing knowledge of brain endothelial cell dedifferentiation.^75,76 Last, we characterized human mural cell organotypicity using our integrated brain scRNA-seq dataset and scRNA-seq datasets from human heart, liver, lung, and skeletal muscle. Despite differences in developmental origin, this analysis identified mural cell marker genes conserved across all five organs. We also identified mural cell genes with brain-enriched expression, including ZIC1, a neural crest lineage transcription factor previously reported to be brain-enriched with expression in mural and fibroblast-like cells,^30,77 and SLC6A1, SLC6A12, ATP1A2, GPER1, and NTM.

Finally, we used RNA in situ hybridization (RNAscope) to validate brain mural cell expression of SLC6A1, SLC6A12, FRZB, and ATP1A2 in flash-frozen human brain neurosurgical samples. We observed colocalization of all four of these transcripts with PDGFRB⁺ mural cell nuclei. Furthermore, we observed vascular localization of FRZB protein in human brain, but not in mouse brain. Together, these data improve confidence in the many additional putative mural cell genes identified by the same integrative analysis of scRNA-seq data, and highlight protein-level validation as a necessary prerequisite to further studies. Additional work will be required to investigate potential functional roles for mural cell genes, but our results highlight the power of scRNA-seq to generate hypotheses of genes that may control cellular function. For example, mural cell-derived FRZB may modulate parenchyma-derived Wnt signaling to endothelial cells, and SLC6A1 or SLC6A12 may mediate GABA uptake by mural cells. Furthermore, given the known influence of pericytes on blood-brain barrier development and brain endothelial function in vivo^3–5,78,79 and in vitro,^48,80 scRNA-seq data may aid in identifying other putative pericyte-derived factors that might mediate interactions with endothelial cells or other neurovascular unit cells such as astrocytes and microglia.^11,81,82 Mural cells are also required for maintenance of normal cerebral blood flow (CBF) and mediate neurovascular coupling.^{1,2,7–10,83–86} Notably, the potassium channel K_ir6.1, encoded by the pericyte marker gene KCNJ8, plays a role in resting CBF. ⁸⁷ Thus, human scRNA-seq data may help generate hypotheses of additional mural cell receptors and channels with roles in these processes, and facilitate comparisons to recently published analysis of these genes in mouse. ⁸⁸

In summary, we present a comprehensive analysis of the human brain mural cell transcriptome based on scRNA-seq data from five independent scRNA-seq datasets. These data should be useful to evaluate animal and in vitro models, and advance understanding of human brain mural cell function.