Single cell atlas of the normal and malformed human brain vasculature

Abstract

Cerebrovascular diseases are a leading cause of death and neurologic disability. Further understanding of disease mechanisms and therapeutic strategies requires a deeper knowledge of cerebrovascular cells in humans. Here, we profiled transcriptomes of 181,388 single-cells from the adult cerebrovasculature. Our results define a cell atlas of the human cerebrovasculature, including endothelial cell molecular signatures with arteriovenous segmentation and expanded perivascular cell diversity. By leveraging this reference, we investigated cellular and molecular perturbations in brain arteriovenous malformations, a leading cause of stroke in young people, and identified pathologic endothelial transformations with abnormal vascular patterning and the ontology of vascular-derived inflammation. Herein, we illustrated an interplay between vascular and immune cells contributory to brain hemorrhage and cataloged opportunities for targeting angiogenic and inflammatory programs in vascular malformations.

The cerebrovasculature comprises an uninterrupted, arborized network of vascular conduits through which circulating blood flows (1–3). It is tasked with ensuring delivery of oxygen, energy metabolites, and other nutrients to the brain, while removing by-products of brain metabolism or preventing entry of circulating toxins (1, 3). Interruptions in cerebral blood flow or structural compromise and hemorrhage lead to stroke, which is a leading cause of death and disability worldwide (4, 5).

Like other vascular beds, the cerebrovasculature has functionally distinct, contiguous segments identified as arteries, arterioles, capillaries, venules, and veins hierarchically organized along an “arteriovenous axis” (1, 6, 7). Cell composition varies with these transitions, and each cerebral vessel is composed of endothelial cells, pericytes, smooth muscle cells, and/or perivascular fibroblast-like cells (referred to hereafter as perivascular fibroblasts) (1, 7, 8). Coordinated molecular interactions between vascular cells and surrounding neurons, glia, and perivascular immune cells, endow the cerebrovasculature with medically relevant, specialized properties. The blood-brain barrier in capillaries, for example, provides a basis for brain immune privilege and serves as an obstacle to pharmacologic treatment of brain diseases (3, 6, 9–11).

Single-cell mRNA sequencing (scRNAseq) in mice has suggested additional cell variation and provided a molecular basis for arteriovenous phenotypic changes known as “zonations” (7, 12–14). Due to biases in cell isolation, existing human brain cell atlas studies have overlooked the cerebrovasculature and its cellular heterogeneity has been largely unexplored in humans (15, 16). Neurologic diseases, such as stroke or Alzheimer’s disease, or brain aging display predilection for select arteriovenous segments (12, 17–19). Thus, large-scale single-cell profiling of the human cerebrovasculature should provide a translational reference to better understand molecular underpinnings of selective cell vulnerabilities and patterns of aberrant gene expression in human cerebrovascular disease.

Cellular and molecular profiles of the adult human cerebrovasculature

To profile cells of the cerebrovasculature, we obtained normal cerebral cortex tissue from patients undergoing tailored lobectomies for epilepsy (table S1). Large arteries and veins were micro-dissected, and smaller vessels (arterioles, capillaries, and venules) were isolated using established methods (20–22) (Fig. 1A and fig. S1A). Dissociated cells from five individuals were processed for scRNAseq using the 10X Genomics Chromium platform and generated high quality transcriptomes from 74,535 cells (fig. S1, B to E). Graph-based Leiden clustering was performed, and clusters annotated with differentially expressed genes to identify 15 cardinal cell populations, each with a unique set of enriched genes and present in multiple individuals (Fig. 1, B to D; fig. S1, F to G; and table S2).

Fig. 1. Cells of the human cerebrovasculature.

(A) Isolation and cell sampling from human cerebral cortex. scRNA, single cell mRNA. (B) UMAP visualization showing cell states from control adult cerebrovasculature (n=5 donors). (C) Dot plot showing expression of cell state markers. EC, endothelial cell; PC, pericyte; SMC, smooth muscle cell; FB, perivascular fibroblast; FbM, fibromyocyte; Mϕ, macrophage; TC, T cell; BC, B cell; Neu, neuron; AC, astrocyte; MG, microglia; OL, oligodendrocyte; OPC, oligodendrocyte precursor cell; RBC, erythrocyte. (D) Bar graph showing cell state proportion by donor. Number of cells sequenced by donor: control1, 6,033 cells; control2, 25,730 cells; control3, 22,816 cells; control4, 19,302 cells; control5, 654 cells. (E) UMAP visualization of endothelial cell states. Art, arterial; Cap, capillary. (F) Dot plot showing expression of endothelial cell state markers. (G) Correlation matrix of gene expression profiles between mouse and human cerebrovascular endothelial cell states. Mouse data obtained from a previously published database. Art, arterial; Cap, capillary; SS, shear stress. (H) UMAP visualization of perivascular cell states. FB, perivascular fibroblast; FbM, fibromyocyte; PC, pericyte; SMC, smooth muscle cell. (I) Dot plot showing expression of perivascular cell state markers.

Based on previously described gene expression patterns, we identified the major vascular cell classes: endothelial cells (CLDN5), pericytes (KCNJ8), smooth muscle cells (MYH11), and perivascular fibroblasts (DCN) (Fig. 1, B to C, and table S2) (7, 23, 24). Using our scRNAseq analysis to inform probe design, we spatially resolved vascular cell diversity in the adult human cerebral cortex with multiplexed spatial transcriptomics (Fig. 2, A to D, and fig. S2, A to F). Consistent with known variations of the human cerebrovasculature, the spatial distribution of cerebrovascular cells revealed reduced vascular cell densities in the white matter (fig. S2A) (25). Cerebrovascular cell classes were organized in known vascular cytoarchitectural structures – such as arteries, capillaries, and veins (Fig. 2D). Our data therefore defined cell classes across the major subdivisions of the cerebrovasculature by intersecting multiplexed spatial transcriptomics with cell-specific markers defined from scRNAseq.

Fig. 2. Spatial RNA analysis resolves the cells of the human cerebrovasculature.

(A) UMAP visualization of spatially defined cerebrovascular cell gene expression profiles identified by multiplexed, iterative single molecule fluorescent in situ hybridization (smFISH). RNA molecules were quantified and assigned to cells via automated spot detection and nuclei segmentation. aEC, arterial endothelial cell; cEC, capillary endothelial cell; vEC, venous endothelial cell; FbM, fibromyocte; PC, pericyte; SMC, smooth muscle cell. (B) Dotplot showing expression of cell state markers. (C) Representative high-magnification microscopy images of merged smFISH and expression distribution of cell state markers projected on UMAP embeddings from (A). DAPI (blue) stains cell nuclei. Scale bar: 10 μm. (D) Representative merged smFISH images showing cellular expression of CLDN5 (magenta, endothelial cells), VEGFC (cyan, arterial endothelial cells), MFSD2A (green, capillary endothelial cells), ACKR1 (yellow, venous endothelial cells), and PDGFRB (red, mural cells). DAPI (blue) stains cell nuclei. Left panel, artery; *, endothelial cell co-expressing CLDN5 (magenta) and VEGFC (cyan). Scale bar: 20 μm. Middle panel, capillary; *, endothelial cell co-expressing CLDN5 (magenta) and MFSD2A (green); arrow, PDGFRB-expressing pericyte (red). Scale bar: 15 μm. Right panel, vein; *, endothelial cell co-expressing CLDN5 (magenta) and ACKR1 (yellow). Scale bar: 15 μm.

Endothelial diversity and arteriovenous zonation in humans

Endothelial cells comprise the inner, blood-facing lining of the cerebrovasculature (1, 3). Identified by expression of CLDN5 and PECAM1, endothelial cells comprised six clusters (Fig. 1, E to F). Using a previously annotated cell atlas of mouse endothelial cells (13), we found that gene expression signatures corresponding to four arteriovenous segments: arteries, capillaries, venules, and veins, consistently mapped onto distinct clusters in our dataset (Fig. 1, F to G; fig. S3, A to H; and table S3). We also identified three clusters of endothelial cells within the arterial zonation (Fig. 1, E to F, and fig. S3D), including a cluster enriched for TXNIP, a regulator of glucose metabolism and oxidative stress (26), likely representing a metabolic state of arterial endothelial cells. By visualizing the spatial position of endothelial arteriovenous zonation markers and relationships with surrounding vascular cells, we confirmed VEGFC⁺, MFSD2A⁺, and ACKR1⁺ endothelium in arteries, capillaries, and veins, respectively (Fig. 2D) (7, 12, 13). Thus, our human vascular cell dataset captures the conserved distinctions of endothelial arteriovenous zonations in humans.

Diversity and distinction of brain perivascular cells

In addition to endothelial cells, we identified the major perivascular cell classes in the brain: pericytes, smooth muscle cells, and perivascular fibroblasts (Fig. 1, H to I; fig. S4, A to C; and table S2) (1, 7, 8). Our data serves as a reference for transcriptomic-based perivascular cell definitions based on correlated patterns of gene expression variation (cell identity scores), as opposed to a handful of marker genes with partially overlapping patterns of expression (fig. S4, D to F) (8, 27, 28).

Pericytes are found in capillaries, venules, and some arterioles, and induce and maintain the blood brain barrier (20, 29, 30). Although previously identified pericyte markers, such as ABCC9 and KCNJ8, together captured all putative pericyte clusters (Fig. 2, H to I, fig. S4G, and table S3) (7, 28), variations in their expression limited use of any single gene as a pan-pericyte marker. We therefore sought to nominate an alternative marker to capture a larger proportion of transcriptomically defined pericytes. Specifically, we identified that HIGD1B mRNA is highly enriched in pericytes and detected in 91.7% of pericytes (91.7% of cells, log₂FC = 3.10, P-adj<0.01) (Fig. 1I, fig. S4G, and table S2). We confirmed HIGD1B expression in PDGFRB⁺ or KCNJ8⁺ pericytes (Fig. 2, B and C).

Smooth muscle cells are contractile cells in arteries, veins, and most arterioles (7, 31, 32). We selected these cells based on expression of pan-smooth muscle cell markers CNN1, TAGLN, and MYH11 (Fig. 1, C and I, and fig. S3A). Iterative analysis of smooth muscle cell transcriptional variation suggested that additional axes of variation may exist (Fig. 1, H to I, and table S3). For example, one cluster was enriched for metallothioneins MT1X, MT2A, MT1M, MT1E, and MT1A, known to modulate smooth muscle proliferation and migration (Fig. 1I and fig. S5A) (33). Additional transcriptional variation included the perivascular cell chemokine ligand CCL2, which coordinates brain response to systemic infection (Fig. 1I and fig. S5A) (34), and RGS16, which regulates sphingosine-1-phosphate signaling implicated in smooth muscle cell proliferation (Fig. 1I and Fig. 2, B to C) (35). Thus, smooth muscle cells may represent a spectrum of transcriptional states and future studies will be necessary to identify their functional roles.

Fibroblasts loosely adhere to arteries, arterioles, venules, and veins within the perivascular space, express extracellular matrix proteins, and provide structural support (7, 23, 36). We identified two clusters of DCN⁺ and APOD⁺ perivascular fibroblasts (Fig. 1, H to I). None of the perivascular fibroblasts expressed markers of other brain fibroblasts, such as those in the meninges which were surgically excluded in this study (37), and DCN⁺ perivascular fibroblasts were visually confirmed to be associated with the cerebrovasculature (fig. S5B). Thus, our data confirms the presence of perivascular fibroblasts in the adult human brain.

Fibromyocytes in the human cerebrovasculature

Two cell clusters were not explained by known brain perivascular cell identities (Fig. 1, H to I, and fig. S4, D to F). We annotated these clusters as “fibromyocytes” based on lower expression of contractile proteins (TAGLN and ACTA2), and higher expression of fibroblast (DCN and LUM) and macrophage (LGALS3) genes as described in peripheral arteries, such as the aorta and cervical internal carotid artery (38–40). No expression of the smooth muscle transcription factor MYOCD was detected suggesting that fibromyocytes are distinct from smooth muscle cells (fig. S5C) (41). Differential gene expression identified IGFBP5, KCNT2, and CCL19 to be more specific to fibromyocytes (fig. S5C and table S2), and we validated KCNT2⁺ and CCL19⁺ fibromyocytes in the human cerebral cortex (Fig 2, A to C). Not identified in prior mouse cell atlases (7, 23, 24), our results demonstrate the presence of fibromyocytes in the human cerebrovasculature.

Fibromyocytes are thought to arise from smooth muscle cells in peripheral vascular beds (42). We therefore performed RNA velocity analysis, which infers transcriptomic trajectories based on the relative abundance of exonic and intronic reads (43). Based on inferred relationships of informatically predicted splicing dynamics, this analysis predicted that smooth muscle cells enriched for CARMN, a lncRNA associated with mesodermal differentiation (44), may give rise to fibromyocytes via upregulation of marker genes, such as LGALS3, KCNT2, and IGFBP5 (fig. S5, D to E). However, in the absence of direct evidence of lineage tracing, we cannot conclusively demonstrate that such a relationship indeed exists.

Retinoic acid signaling regulates smooth muscle-to-fibromyocyte transitions in the periphery (39). Fibromyocyte clusters were enriched for retinoic acid synthetic enzyme and receptor genes, which were also characteristic of perivascular fibroblasts (Fig. 1I and fig. S5, F to G), and we spatially confirmed ALDH1A1 and RARA expression in DCN⁺ perivascular fibroblasts and CCL19⁺ fibromyocytes, respectively (Fig. 2, B to C). Prior cell atlases have not documented non-meningeal sources of retinoic acid (fig. S3H). Therefore, our study identified that fibromyocytes and perivascular fibroblasts may be endogenous sources of retinoic acid in the adult human brain.

Deconstructing the dysplastic cerebrovasculature in arteriovenous malformations

To showcase the utility of our dataset, we generated a scRNAseq dataset from arteriovenous malformation (AVM) samples, a type of vascular malformation (45). We obtained intraoperative, angiographically confirmed human brain AVMs from five patients (table S1). Using analogous dissection and scRNAseq techniques (Fig. 3A), we generated high-quality whole cell transcriptomes from 106,853 cells and identified 11 cardinal cell populations (Fig. 3, B to C, and fig. S6, A to D). Each cell population was identified in multiple specimens, except for astrocytes and choroid plexus (Fig. 3C and fig. S6E), and we spatially confirmed CLDN5⁺ endothelial cells, TAGLN⁺ smooth muscle cells, CCL19⁺ fibromyocytes, and COL1A2⁺ perivascular fibroblasts in AVMs (Fig. 3D). To identify endothelial and perivascular cell molecular changes in AVMs, we co-embedded control and AVM scRNAseq datasets (Fig. 3E and fig. S7A), identified differentially expressed genes (Fig. 3F; fig. S7, D to H; and table S4), and performed iterative clustering in each vascular cell class (Fig. 3E; fig. S8, A to E; fig. S9, A to C; and table S5). These findings confirmed presence of fibromyocytes in diseased human cerebrovascular tissues and cataloged cell-specific aberrant gene expression in AVMs.

Fig. 3. Cellular aberrancy in the malformed cerebrovasculature.

(A) Isolation and cell sampling from human brain arteriovenous malformations (AVMs). scRNA, single cell mRNA. (B) UMAP visualization showing cell states from AVMs (n=5 donors). (C) Bar graph showing cell state proportion by donor. Number of cells sequenced by donor: AVM1, 26,122 cells; AVM2, 28,868 cells; AVM3, 14,541 cells; AVM4, 26,660 cells; AVM5, 10,662 cells. EC, endothelial cell; SMC, smooth muscle cell; FB, perivascular fibroblast; FbM, fibromyocyte; Myl, myeloid cells; TC, T cell; BC, B cell; Neu, neuron; AC, astrocyte; MG, microglia; OL, oligodendrocyte; CP, choroid plexus. (D) Representative microscopy image of single molecule fluorescent in situ hybridization showing expression of CLDN5 (yellow, endothelial cells (EC)), TAGLN (magenta, smooth muscle cells (SMC)), COL1A2 (red, perivascular fibroblasts (FB)), and CCL19 (cyan, fibromyocytes (FbM)). DAPI (blue) stains cell nuclei. Boxes highlight representative cells. Scale bar: 50 μm. i, TAGLN⁺ smooth muscle cell. ii, CCL19⁺ and TAGLN⁺ fibromyocyte. iii, CLDN5⁺ endothelial cell. iv, COL1A2⁺ perivascular fibroblast. Inset scale bars: 5 μm. (E) Left, schematic describing computational pipeline. Endothelial cells (orange) are identified in silico by marker expression, co-embedded for downstream analytics, and iteratively clustered. An identical workflow was applied to perivascular cells (fig. S7). Middle, UMAP visualization of co-embedded endothelial cell states in control (gray) and AVMs (red). Right, UMAP visualization of iteratively clustered endothelial cell states. Art, arterial; Cap, capillary; Vn, venous; Nd, nidus. (F) Upset plot of differentially expressed genes (DEGs) (horizontal bars) by cell class. Number of DEGs exclusive to one cell class (black circles) or shared between multiple cell classes (linked black circles). Vertical bars show the number of genes per intersection. (G) Heatmap visualization of arteriovenous transcriptional identity in control (top, CTRL) and AVM (bottom) endothelial cell states. Art, arterial; Cap, capillary; Vu, venule; Vn, venous; Nd, nidus. Exp., expression, Blue, low expression; Yellow, high expression. (H) Upset plot showing intersections of DEGs in AVM endothelial cell states compared to controls. (I) UMAP visualization of AVM endothelial cell RNA velocity reveals two divergent trajectories from Nd1 (yellow). Upregulation of PLVAP and PGF occurs with endothelial Nd1-to-Nd2 transitions. Exp., expression. (J) Gene set enrichment analysis of DEGs in AVM endothelial Nd2. padj, false discovery rate adjusted P-value; NES, normalized enrichment score. (K) Dotplot showing top marker gene expression for control capillary and AVM Nd2 endothelial cells. Avg. Exp., average expression; Exp., expression (L) Violin plot of PLVAP expression showing specificity to AVM Nd2. (M) Representative-confocal microscopy analysis of PLVAP- (yellow) and ANGPT2-positivity (magenta) in PECAM1⁺ endothelial cells (cyan) in AVM nidus. Vessel shown in cross section. Colocalization of fluorescence results in white coloration. Scale bar: 50 μm.

Endothelial aberrancy in brain arteriovenous malformations

AVMs arise from pathologic molecular changes in endothelial cells (46, 47). This catalyzes direct connections to form between arteries and veins without intervening capillaries and results in tortuous, dysmorphic tangles of blood vessels referred to as the “nidus” (45). Joint analysis of control and AVM datasets revealed that endothelial subsets were enriched for arterial and venous, but not venular or capillary transcriptional identity scores in AVMs (Fig. 3, E and G, and fig. S9, D to G). Endothelial cell clusters with suppressed venule and capillary cell identities (nidus 1 (Nd1) and nidus 2 (Nd2)) showed greatest differential gene expression (Fig. 3H and table S6). RNA velocity analysis identified a consensus molecular trajectory from Nd1 to Nd2 (Fig. 3I and fig. S9H), and predicted a progressive upregulation of PLVAP, a marker of fenestrated endothelium normally confined to developmental angiogenesis, the brain’s circumventricular organs and choroid plexus, and PGF, a potent stimulator of brain angiogenesis (Fig. 3I) (13, 30, 48, 49). Gene set enrichment analysis confirmed pathogenic cascades, such as angiogenesis, inflammation, and epithelial-to-mesenchymal transition, enriched in AVM Nd2 endothelium (Fig. 3J and table S7) (50–52). Control capillary endothelial cells robustly expressed blood-brain barrier nutrient transporters – including MFSD2A, SLC16A1, and SLC38A5 (Fig. 3K). By contrast, AVM Nd2 endothelial suppressed nutrient transporter expression and upregulated pro-inflammatory (CCL14), pro-angiogenic (PGF and STC1), and pro-permeability (PLVAP and ANGPT2) genes (Fig. 3, K to L), and we confirmed the localization of Nd2 endothelial cells in the AVM nidus (Fig. 3M).

To characterize how pathologic endothelial gene expression may influence cell-to-cell communication networks, we used an in silico algorithm to predict reciprocal ligand-receptor interactions (53). The assembled interactome identified Nd2 as the strongest contributor to abnormal cell communications in AVMs (fig. S10A). Dysregulated communication pathways included established pathogenic cascades, such as angiopoietin, vascular endothelial growth factor, and transforming growth factor β signaling (54–56), and identified previously unrecognized immune activating and angiogenic communication networks in AVMs, such as CD99, SPP1, and CALCR (fig. S10, B to F). Thus, aberrant nidus endothelial gene expression is predicted to result in pathologic cell-to-cell communication networks in AVMs.

Immune cell microenvironment and cerebrovascular-derived inflammation

Inflammation is hypothesized to play a role in the formation of AVMs (45, 52). Iterative analysis of the immune cell populations associated with the cerebrovasculature identified 17 immune cell clusters in co-embedded cell populations (Fig. 4, A to B, and Fig. S11, A to B). Nine clusters comprised myeloid cells – including vessel-associated microglia, conventional dendritic cells (cDCs), three perivascular macrophages (pvMφ), and three monocyte (Mo) subpopulations, and we computationally separated myeloid cells with evidence of ex vivo activation (Fig. 4, A to B, and fig. S11C) (57). Eight clusters comprised lymphoid cells – including CD4⁺ T cells, CD8⁺ T cells, regulatory T cells (T_reg), B cells, natural killer (NK) cells, and plasmacytoid dendritic cells (pDCs) (Fig. 4, A to B). Resident pvMφs were the most abundant immune cell population comprising 31.2% and 28.3% of immune cells in controls and AVMs, respectively (Fig. 4C). Greater than 90% of circulating immune cells, such as CD8⁺ T cells, were confined within the resting cerebrovasculature, but infiltrated into the perivascular space or adjacent brain in AVMs (P<0.01) (fig. S11, D to E).

Fig. 4. Cerebrovascular inflammation with malformation.

(A) UMAP visualizations of co-embedded immune cells states in control cerebrovasculature and brain arteriovenous malformations (AVMs) (n=5 donors per condition). Top, colored by condition; control, gray; AVM, red; Bottom, colored by cell state. pvMϕ, perivascular macrophage; pvMϕ*, activated perivascular macrophage; Mo, monocyte; MG, microglia; ExV, ex vivo activated myeloid cells; cDC, conventional dendritic cells; pDC, plasmacytoid dendritic cells; CD8 TC, CD8⁺ T cells; CD4 TC, CD4⁺ T cells; T_reg, regulatory T cells; NK, natural killer cells; BC, B cells; Div, dividing immune cells. (B) Dot plot showing expression of immune cell markers. (C) Pie charts showing immune cell state proportions in controls (left) and AVMs (right). (D) Bar graph showing proportion of myeloid and lymphoid immune cells captured in AVMs (red) and controls (gray). n=5 donors per condition, Mean ± SEM, two-tailed t-test.*P< 0.05; ns, not significant. (E) Bar graph of relative immune cell state proportions normalized to total cells sequenced in AVMs (red) and controls (gray). (F) Representative confocal microscopy analysis of endothelial cells (cyan, podocalyxin (PODOXL)), smooth muscle cells (blue, smooth muscle α−actin (SMA)), macrophages (magenta, ionized calcium binding adaptor molecule 1 (IBA1) which is encoded by the gene AIF1 in (B)) and microglia (green, purinergic receptor P2Y12 (P2Y12)). Scale bar: 1 mm. i, layered ameboid perivascular macrophages. ii, perivascular microglial response. Inset scale bars: 50 μm. (G) Quantification of abundance (top left), cell ratio per image (top right) and perivascular distance (bottom) of IBA1⁺P2R12⁻ macrophages and IBA1⁺P2R12⁻ microglia (n=3 donors per condition; three non-adjacent sections per donor; 8–10 images per section). Mean ± SEM, two-tailed t-test.**P<0.01; ns, not significant.

Myeloid immune cells were more abundant and expressed gene signatures suggestive of activation in AVMs (Fig. 4, D to G, and fig. S11F), and we cataloged dysregulated immune cell communication networks (fig. S11G). Vessel-associated CD11c⁺ antigen presenting cells are potent activators of brain CD4⁺ T-cell responses in situ (58, 59). scRNAseq confirmed vessel-associated CD11c⁺ cells were composed of myeloid cells, including cDCs, pvMφs, and some microglia, and we identified a heterogeneous spatial distribution of vessel-associated antigen presenting myeloid cells in the AVM nidus (Fig. 4F). For example, discrete areas appeared to have more numerous IBA1⁺P2RY12⁻ macrophages or IBA1⁺P2RY12⁺ microglia (Fig. 4, F to G). A pronounced perivascular myeloid cell response was observed and IBA1⁺P2RY12⁻ macrophages were found at greater distances from the adjacent vasculature consistent with infiltration in AVMs (P<0.01) (Fig. 4G). Thus, this data describes a diverse cellular and spatially heterogenous cerebrovascular inflammatory response in AVMs.

Vascular immune cell crosstalk with brain hemorrhage

Hemorrhagic stroke is a devastating consequence of AVMs (60), and we sought to identify deleterious cell states associated with AVM rupture. We used scMappR to deconvolute cellular heterogeneity and to compute cell-specific gene expression signatures from AVM bulk RNAseq (n=39 AVMs; ruptured, 26 AVMs; unruptured, 13 AVMs) (Fig. 5A and table S1) (61). We first identified 871 DEGs associated with AVM rupture enriched in vascular developmental pathways (e.g., blood vessel development and morphogenesis) and inflammatory processes (e.g., cell recruitment) (fig. S12, A to B, and table S8). Using our scRNAseq data set, in silico cell abundance deconvolution resolved probable alterations in cell proportions (Fig. 5B). A subpopulation AIF1⁺ (encodes IBA1), P2RY12⁻ monocytes, identified as GPNMB⁺ Mo3 monocytes, was over-represented in ruptured AVMs (P<0.01) and expressed gene signatures consistent with activation (Fig. 5, B to D, and fig. S12, C to D). Thus, distinct infiltrating immune cell states become enriched with AVM rupture.

Fig. 5. Cell states implicated in brain arteriovenous malformation rupture.

(A) Cellular deconvolution and cell-specific differential expression analysis of bulk RNAseq from ruptured and unruptured brain arteriovenous malformation (AVM). (B) Bar graph of change in cell proportion t-statistic in ruptured AVMs. Purple, increased cell abundance. Blue, decreased cell abundance. *P<0.05; **P<0.01. (C) Representative confocal microscopy imaging showing GPNMB⁺ monocytes (green), endothelial cells (cyan, podocalyxin (PODOXL)), and smooth muscle cells (magenta, smooth muscle α actin (SMA)) in unruptured and ruptured AVMs. Scale bar: 100 μm. (D) Quantification of GPNMB⁺ monocytes in unruptured (blue) and ruptured (purple) AVMs (n=3 donors per condition; three non-adjacent sections per donor; 8–10 random images per section). Mean ± SEM, two-tailed t-test.*P< 0.05. (E) Top, Confocal microscopy analysis of cleaved caspase-3⁺ (green, CC3) human primary smooth muscle cells following co-culture with GPNMB⁺ and GPNMB⁻ monocytes isolated from ruptured AVMs. DAPI (magenta) stains cell nuclei. White, colocalization of CC3 and DAPI. Arrow, CC3⁺ cell. Scale bar: 20 μm. Bottom, quantification of CC3⁺ smooth muscle cells (n=3 independent cultures per condition). Mo, monocytes. Mean ± SEM, ANOVA with Tukey post-hoc test. *P<0.05; ns, not statistically significant. (F) Scatterplot of dysregulated cell communication pathways in AVM GPMNB⁺ monocytes (Mo3) relative to controls by scRNAseq. Red, upregulated in AVM; Blue, upregulated in control; Gray, shared between conditions. Triangle, outgoing network; square, incoming network; diamond, outgoing and incoming network. (G) Top, Confocal microscopy analysis of cleaved caspase-3⁺ (green, CC3) human primary smooth muscle cells treated with osteopontin (OPN, encoded by SPP1) and CD44-neutralizing antibody, RGD integrin inhibitor or inhibitor cocktail. DAPI (magenta) stains cell nuclei. White, colocalization of CC3 and DAPI. Arrow, CC3⁺ cell. Scale bar: 20 μm. Bottom, quantification of CC3⁺ smooth muscle cells (n = 5–6 independent cultures per condition). Mo, monocytes. Mean ± SEM, ANOVA with Tukey post-hoc test. **P< 0.01; ns, not statistically significant.

Inflammation leads to a loss of vessel integrity and smooth muscle cells contribute to brain hemorrhage when depleted (52, 62–64). In silico abundance of GPNMB⁺ monocytes and smooth muscle cells correlated negatively in ruptured AVMs (r=–0.43, P<0.05). We therefore investigated whether GPNMB⁺ monocytes contribute to smooth muscle cell death. Co-culture of isolated GPNMB⁺ monocytes from ruptured AVM patients with primary brain vascular smooth muscle cells (VSMCs) increased apoptotic cleaved caspase-3⁺ VSMCs (P<0.01) (Fig. 5E). Cell-to-cell communication analysis identified SPP1 (encodes osteopontin (OPN)) as the greatest dysregulated outgoing signaling pathway from GPNMB⁺ monocytes in AVMs (Fig. 5F). The ligand OPN is predicted and previously shown to interact with CD44 and integrin receptors on smooth muscle cells (65). Soluble OPN induced a 2.7-fold increase in VSMC apoptosis which was ameliorated by pre-treatment with neutralizing CD44 antibody, an integrin inhibitor, or cocktail thereof (P<0.01) (Fig. 5G). Thus, GPNMB⁺ monocytes contribute to smooth muscle cell depletion and are associated with AVM rupture and brain hemorrhage.

Discussion

We present a cell resolution atlas describing the transcriptomic heterogeneity underlying cell function and interaction in the human adult cerebrovasculature. We identified conservation of endothelial molecular zonations essential to arteriovenous phenotypic change and expanded cellular diversity of brain perivascular cells, including fibromyocytes not previously identified in the cerebrovasculature (7, 23, 24). Smooth muscle cells are predicted to transform into fibromyocytes but requires validation with fate tracing methods. Fibromyocytes and perivascular fibroblasts may produce retinoic acid in the adult human brain. Retinoic acid signaling contributes to cortical vascular development and modulates smooth muscle plasticity and fibromyocyte speciation in other vascular beds (39, 66, 67). However, the functional significance of these findings in the adult cerebrovasculature warrants further investigation.

This atlas has important implications for neuroscience and clinical medicine. To exemplify its utility, we defined cellular and gene expression changes in AVMs, a leading cause of stroke in young people, and identified pathologic endothelial molecular transformations, spatially localized to the AVM nidus. Some molecular changes are shared with immature embryonic endothelium or angiogenic tip cells, but other developmental or angiogenic transcriptional programs are notably absent or altered (14, 30, 68, 69). We also describe the cellular ontology and communication networks of cerebrovascular-derived inflammation. The interplay between vascular and immune cells, such as GPNMB⁺ monocytes and smooth muscle, induced pathological changes associated with brain hemorrhage, and our findings may guide future therapeutic development.

We recognize that this atlas represents a first step towards a comprehensive census of the human cerebrovasculature. Limitations in unintended biases of cell capture or isolation and random sampling, such as relative proportions of small and large vessels from each individual, may alter relative cell proportions and require further validation in spatially resolved datasets. Additional work is also needed to ascertain distinctions between cell types and cell states, such as transient or metabolic variations. Nonetheless, our results will inform future studies in other brain regions or cerebrovascular diseases to accelerate mechanistic understanding and therapeutic targeting of the human cerebrovasculature.

Materials & Methods

Ethics statement and tissue acquisition

Human brain tissue specimens and clinical data were obtained from the University of California San Francisco with protocols approved from the institutional review board and ethics committee (IRB 10–01318 and 10–02012). All tissues were acquired from patients undergoing neurosurgical operations and written informed consent was obtained prior to the procedure permitting collection of tissue specimens for the purposes of research. Normal cerebral cortex was obtained as part of a neurosurgical operation to reach deep seated lesions causing epilepsy and uninvolved in the pathology. All specimens were >2cm from any radiographic abnormality on magnetic resonance imaging, showed no abnormalities on routine electrocorticography, and were histologically normal on a rapid hematoxylin and eosin stain. Diagnosis of human brain arteriovenous malformations (AVMs) were angiographically confirmed preoperatively, and fresh tissues were acquired as part of planned surgical resection. For single-cell RNA sequencing (scRNAseq), only unruptured brain arteriovenous malformations were enrolled into the study. Specimen orientation was maintained to ensure coverage of arteriovenous axis with surgical clips of different sizes with aide of intraoperative fluorescent angiography. All tissues were acquired in close collaboration with neurosurgeons trained in tissue isolation techniques to minimize tissue disruption such as avoidance of electrocautery. For bulk sequencing experiments, we utilized snap-frozen ruptured and unruptured AVM specimens as part of our biorepository. Patient demographic information for all tissues utilized is summarized in table S1.

Isolation of cerebrovascular specimens

No differences in vascular isolation methods were applied between normal cerebral cortex or AVM tissues. Tissue specimens were placed in chilled pre-oxygenated Dulbecco’s Modified Eagle Medium (DMEM, Fisher Scientific, Waltham, MA, catalog number: MT10017CV) and transported to the laboratory on ice. All tissue handling was performed with autoclave-sterilized equipment within a class II biological safety cabinet. Large arteries and veins were selectively isolated under 5X magnification with a Leica MZ75 dissecting microscope (Leica Microsystems, Inc., Wetzlar, Germany) with two #5/45 Dumont micro-forceps (Fine Science Tools, Foster City, CA, catalog number: 11251–35). Under 5X magnification, the lumen of each vessel was longitudinally opened with a #15 scalpel blade (Fine Science Tools) or flushed with DMEM. The large vessels were placed in an Eppendorf LoBind 5-ml tube (Eppendorf North America, Enfield, CT, catalog number: 0030122348) and kept on ice in chilled oxygenated DMEM. Following removal of all visible vasculature, the microvasculature was isolated using dextran gradient centrifugation followed by cell-strainer filtration (20–22). More specifically, tissue was cut into ~1–2-mm pieces with a scalpel and gently homogenized in a Dounce homogenizer containing oxygenated pre-chilled DMEM with 1% bovine serum albumin (BSA) (Sigma Aldrich, St. Louis, MO, catalog number: A9418). The homogenate was mixed in 18% dextran solution (MW: ~70,000 Da; Sigma-Aldrich, catalog number: 31390) at a volume ratio of 1:1 and centrifuged at 6000g for 20 min at 4°C. This resulted in a microvascular pellet and floating vascular-depleted brain. The floating vessel-depleted brain was gently aspirated and discarded. The vascular pellet resuspended in DMEM with 1% BSA and passed through a 40-μm cell strainer (Fisher Scientific, catalog number: 08–771-1) to remove circulating cells or other debris. Microvascular fragments remain trapped on top of the cell strainer and were subsequently collected by inverting with cell strainer and washing with pre-chilled oxygenated DMEM. A small aliquot was visualized at 10X magnification to confirm both purity and yield with brightfield microscopy. The microvascular fragments were pelleted by centrifugation at 500g for 5 min, the supernatant aspirated and then pooled with the microdissected arteries and veins from the same individual in pre-chilled oxygenated DMEM. These pooled preparations are referred to as isolated vascular preparations for subsequent steps and maintained in chilled oxygenated media on ice and immediately processed to create single-cell suspensions.

Generation of vascular single-cell suspensions

Isolated vascular preparations were incubated for 45 min in 0.2% collagenase type 2 (Worthington Biochemical Corporation, Lakewood, NJ, catalog number: LS004176) diluted in pre-oxygenated DMEM at 37° C with gentle agitation in an Eppendorf LoBind 5-ml tube (Eppendorf North America). Cell suspensions were filtered through a sterile 40-μm cell strainer (Fisher Scientific) to isolate undigested debris. Cells contained within the flowthrough were collected by centrifugation at 500g at 5 min and the supernatant carefully aspirated. To lyse any residual erythrocytes, the cell pellet was resuspended in Gibco ACK lysing buffer (Fisher Scientific, catalog number: A1049201) for 3 min at room temperature. Cells were then collected by centrifugation at 500g at 5 min and washed three times with sterile RNase-free phosphate buffered saline (PBS) (Sigma-Aldrich, catalog number: D8537–500ml) containing 0.04% BSA. Cells were pelleted with centrifugation at 500g for 5 min and resuspended in PBS with 0.04% BSA. To confirm cell viability and yield, a 10-μl aliquot of the cell suspension was mixed 1:1 with 0.4% trypan blue (Thermo Fisher Scientific, catalog number: T10282) to stain non-viable cells. Cells were then counted on a hemocytometer.

Single-cell RNA sequencing (scRNAseq)

All scRNAseq experiments were performed on freshly isolated, whole cells as described above. Droplet-based scRNA seq was performed with 10X Genomics Chromium Single Cell 3 prime reagent kits v3 as described by the manufacturer (10X Genomics, Pleasanton, CA, product code: 1000092; n=5 normal cortex samples and n=5 AVMs). Based on hemocytometer counts, single cells were loaded onto chromium chips with a capture target of 15,000 cells per sample. When cell yield was sufficient, two reactions per individual were performed. Libraries were prepared following the provided manufacturer’s protocol. Quality of prepared sequencing libraries were confirmed by electrophoretic analysis on an Agilent 4200 TapeStation System (Agilent Technologies, Inc., Santa Clara, CA). Libraries were sequenced with an Illumina NovaSeq 6000 (Illumina, Inc., San Diego, CA) with a targeted sequencing depth of 50,000 reads per cell.

Bulk RNA sequencing

Undigested, isolated vascular tissues from ruptured and unruptured AVMs were obtained from the operating room, snap frozen in liquid nitrogen and stored at –80°C. Snap frozen tissues were embedded in Optical Cutting Temperature Compound (Sakura Finetek USA, Inc., Torrance, CA, catalog number: 4583) and cryosectioned at a thickness of 20 μm. Tissue scrolls were collected in RNAase free Eppendorf LoBind-1.5 ml microcentrifuge tubes (Eppendorf North America, catalog number: 022431021). DNA/RNA Shield reagent (Zymo Research, Irvine, CA, catalog number: R1100) was added. Tissues were mechanically homogenized with a Squisher-Single homogenizer (Zymo Research, catalog number: H1001) and subsequently digested with proteinase K (Zymo Research, catalog number: R1057). Total RNA was isolated from the homogenized tissues with the Quick-RNA Miniprep Plus Kit as instructed by the manufacturer (Zymo Research, catalog number: R1057). The purified RNA was quantified with a NanoDrop (Thermo Fisher Scientific) and integrity confirmed with an Agilent 4200 TapeStation System (Agilent Technologies, Inc.). Ribosomal RNA was depleted using the NEBNext rRNA Depletion Kit as instructed by the manufacturer (New England Biolabs, Inc., Ipswich, MA, catalog number: E6310X), and sequencing libraries prepared with the SEQuoia Complete Stranded RNA Library Prep Kit (Bio-Rad Laboratories, Inc., Hercules, CA, catalog number: 17005710). Sequencing library quality was confirmed on an Agilent 4200 TapeStation System (Agilent Technologies, Inc) and quantified with a Qubit dsDNA high sensitivity assay kit (Thermo Fisher Scientific). Libraries were sequenced in batch with an Illumina NovaSeq 6000 sequencer (Illumina, Inc.) with targeted read depth of at least 5 × 10⁷ reads per sample.

Rebus esper spatial omics platform

Spatially resolved, multiplexed in situ RNA detection and analysis was performed using the automated Rebus Esper spatial omics platform (Rebus Biosystems, Inc., Santa Clara, CA). The system integrates synthetic aperture optics (SAO) microscopy (70), fluidics and image processing software and was used in conjunction with single-molecule fluorescence in situ hybridization (smFISH) chemistry. By intersecting a list of known or candidate cell type markers generated by scRNAseq, suitability for probe design, including gene length and relative abundance, and design constraints for compatibility with the Rebus Esper spatial omics platform using proprietary software, we generated the following 16-gene probe panel: MECOM, RGS16, RARA, MFSD2A, TAGLN, IL1R1, VEGFC, KCNJ8, DCN, CCL19, CLDN5, PDGFRB, HIGD1B, KCNT2 and ALDH1A1. Individual transcripts from target genes were automatically detected, counted, and assigned to individual cells, generating a cell by feature matrix that contains gene expression and spatial location data for each individual cell, as well as registered imaging data, as follows: Rebus Biosystems proprietary software was used to design primary target probes (22–96 oligos) and corresponding unique readout probes for each gene. The oligos were purchased from Integrated DNA Technologies and resuspended at 100 μM in TE buffer. Coverslips (24 × 60 mm, No. 1.5, Azer Scientific) were functionalized as previously published (71). Fresh frozen adult human brain tissue sections (7 μm) were cut on a cryostat, mounted on the treated coverslips and fixed for 10 min with 4% paraformaldehyde (Alfa Aesar) in PBS at room temperature, rinsed three times with PBS at room temperature, rinsed with 70% ethanol, and stored in 70% ethanol at 4°C before use. The sample section on the coverslip was assembled into a flow cell, which was then loaded onto the instrument. The tissue pretreatment, hybridization cycles and imaging were done automatically under the instrumental control software. Briefly, the tissue pretreatment helps to significantly reduce the autofluorescence such as a lipofuscin signal. Primary probes for all target genes were initially hybridized for 6 hours and unbound probes were washed away. Readout probes labeled with Atto594 and Atto647N dyes for the first two genes were then hybridized, washed, counterstained with DAPI and then imaged with an Andor sCMOS camera (Zyla 4.2 Plus, Oxford Instruments) through 20xc, 0.45NA dry lens (CFI S Plan Fluor ELWD, Nikon) with 365nm LED for DAPI, 595-nm and 647-nm lasers configured for SAO imaging. Multiple fields of view (FOVs) were imaged for each channel within the region of interest (ROI). Single z-planes with 2.8 μm depth of field were acquired for each field of view. After imaging, the first two readout probes were stripped and the readout probes for the next two genes were then hybridized, imaged, and stripped. This process was repeated until readout was completed for all genes. Using the Rebus Esper image processing software, the raw images were reconstructed to generate high-resolution images (equivalent or better than images obtained with a 100X oil immersion lens). RNA spots were automatically detected to generate high fidelity RNA spot tables containing xy positions and signal intensities. Nuclei segmentation software based on StarDist identified individual cells by finding nuclear boundaries from DAPI images (72). The detected RNA spots were then assigned to each cell using maximum distance thresholds. The resulting cell by feature matrix contains gene counts per cell along with annotations for cell location and nuclear size.

Tissue and Cerebrovascular Fragment Immunofluorescent Staining

Formalin fixed, paraffin embedded tissue sections were cut at a thickness of 6 μm, deparaffinized with xylene, and rehydrated to distilled water with serial ethanol washes. For immunostaining of isolated vessel fragments, cerebrovascular vessel isolation was performed as described above and immersion fixed in 4% paraformaldehyde overnight at 4°C. For antigen retrieval, all specimens were incubated with pH 9 Tris-EDTA buffer at 97°C for 15 min. Tissue sections were then blocked with PBS containing 0.2% gelatin, 1% donkey serum, and 1% triton for 1 hour at room temperature and incubated in the following primary antibodies overnight at 4°C. Primary antibodies included: podocalyxin (1:100, R&D Systems, Minneapolis, MN, catalog number: AF1658), smooth muscle α-actin clone 1A4 (1:100, Dako North America, Inc. Carpinteria, CA, catalog number: M085129), P2RY12 (1:500, Sigma Aldrich, catalog number: HPA014518), IBA1 (1:500, Synaptic Systems, Goettingen, Germany, catalog number: 234004), PECAM1 (1:50, Agilent Technologies, Inc., catalog number: M0882329), PLVAP (1:100, Sigma Aldrich, catalog number: HPA002279), angiopoietin-2 (1:100, R&D systems, catalog number: AF623), PDGFRB (28E1) (1:100, Cell Signaling Technology, catalog number: 3169S), GPNMB (1:100, R&D systems, catalog number: AF2550), and CD8 clone C8/144B (1:100, Dako North America, catalog number: M710301). Sections were washed with PBS containing 1% Triton and incubated with Alexa Fluor secondary antibodies for 2 hours at room temperature. Alexa Fluor 488-conjugated donkey anti-mouse secondary antibody (1:500, Thermo Scientific, catalog number: A32766), Alexa Fluor 546-conjugated donkey anti-rabbit secondary antibody (1:500, Thermo Fisher, catalog number: A10040), and Alexa Fluor 647-conjugated donkey anti-goat secondary antibodies (1:500, Thermo Fisher, catalog number: A32787) were used to detect mouse, rabbit, and goat primary antibodies, respectively. Sections were washed and autofluorescence quenched by incubating with 1% Sudan Black (Sigma Aldrich, catalog number: 199664) for 10 min at room temperature. Slides were mounted with 4-prime,6-diamidino-2-phenylindole (DAPI)-containing Fluoromount-G (SouthernBiotech, Birmingham, AL, catalog number: 0100–20). All imaging was performed with a Lecia TCS SP8 X confocal microscope with a 20X objective (Leica Microsystems, Inc.).

Fluorescence-activated cell sorting (FACS)

To isolate circulating GPNMB⁺ and GPNMB⁻ monocytes, 10 ml of whole blood was collected in standard 5-ml EDTA-containing vacutainer blood collection tubes obtained from adult patients with acutely ruptured AVMs. Cells were subsequently spun down at 500g for 5 min and erythrocytes lysed with incubation in Gibco ACK lysing buffer (Fisher Scientific, catalog number: A1049201). The resulting cell suspension was filtered through a sterile 40-μm filter to remove debris and washed with PBS containing 0.04% BSA. Cell suspensions were then resuspended in FACS staining buffer (Thermo Fisher Scientific, catalog number: 00–4222-26). Cells were blocked to prevent non-specific staining, and cells were stained with Alexa Fluor 647-conjugated mouse anti-human CD45 monoclonal antibody clone HI30 (1:200, Thermo Fisher Scientific, catalog number: 51–0459-42), FITC-conjugated mouse anti-human CD11b monoclonal antibody clone ICRF44 (1:200, Thermo Fisher Scientific, catalog number: 11–0118-42), and PE-conjugated mouse anti-human GPNMB monoclonal antibody clone HOST5DS (1:200, Thermo Fisher Scientific, catalog number: 12–9838-42) for 30 min at room temperature. Cells were subsequently isolated by FACS with a BD FACSAria II Flow Cytometer (BD Biosciences, Franklin Lakes, NJ). Viable CD45⁺CD11b⁺GPNMB⁺ cells and CD45⁺CD11b⁺GPNMB⁻ cells were separately collected for subsequent co-culture experiments.

Cell culture

All cell culture experiments utilize primary human brain vascular smooth muscle cells (VSMCs, ScienCell Research Laboratories, Carlsbad, CA, catalog number: 1100). Cells were cultured in smooth muscle cell media in 5% CO₂ at 37°C. Early passage (P3, P4) cultures were used in the present study. Primary VSMCs were plated in equal number for all conditions. VSMCs were grown on 96-well tissue culture plates pre-coated with collagen. For coculture experiments, CD45⁺CD11b⁺GPNMB⁺ or CD45⁺CD11b⁺GPNMB⁻ cells were immediately cocultured with VSMCs following FACS isolation at an approximate ratio of 1:1 (73). Monocytes and VSMCs were co-cultured for 24 hours and then subsequently fixed with 4% paraformaldehyde for subsequent immunostaining. For osteopontin (OPN, encoded by SPP1) treatment studies, cells were pretreated with vehicle control or anti-human CD44 neutralizing monoclonal antibody (10 μg/ml, Thermo Fisher Scientific, catalog number: MA4400), RGD peptide to inhibit integrin receptors (10 μM, Sigma Aldrich, catalog number: A8052), or both in combination for 30 minutes as previously described (65). Cultures were then subsequently treated with OPN (200 ng/ml, Sigma Aldrich, catalog number: SRP3131) for 6 hours. The cell culture media was then changed, and cells fixed. with 4% paraformaldehyde for subsequent immunostaining.

Cell Culture Immunostaining

Paraformaldehyde, fixed cells were blocked with PBS containing 0.2% gelatin, 1% donkey serum and 0.1% triton for 1 hour at room temperature and incubated in primary antibodies overnight at 4° C. Primary antibodies included: smooth muscle α-actin clone 1A4 (1:100, Dako North America, Inc. Carpinteria, CA, catalog number: M085129) and cleaved caspase-3 (1:300, Cell Signaling, catalog number: 9661S). Alexa Fluor 488-conjugated donkey anti-mouse secondary antibody (1:500, Thermo Scientific, catalog number: A32766) and Alexa Fluor 546-conjugated donkey anti-rabbit secondary antibody (1:500, Thermo Fisher, catalog number: A10040) were used to detect mouse and rabbit primary antibodies, respectively. Cells were washed and nuclei stained with DAPI. All imaging was performed with a Lecia TCS SP8 X confocal microscope with a 20X objective (Leica Microsystems, Inc.).

Tissue and cell culture image analysis

For all quantitative imaging experiments, a tile-scan image was generated encompassing the tissue section or cell culture well with 10-to-12-μm z-stack and maximum projection z-stack images were reconstructed. For all tissue studies, 8–10 randomly selected fields in three non-adjacent tissue sections per tissue specimen were analyzed as previously described (20). Tissues from three donors per condition were used for all analyses. For quantification of IBA1⁺, P2RY12⁺, and CD8⁺ immune cells, cell bodies were counted with the NIH ImageJ multipoint tool and expressed as number of positive cells per cubic millimeter of tissue. For immune cell distance analysis, the distance between cell bodies and abluminal vascular wall was measured using the NIH ImageJ length measurement tool. For co-culture experiments, three independent cultures were analyzed per condition. For OPN experiments, 5–6 independent cultures were analyzed per condition. Cleaved caspase-3⁺ cells were counted with the NIH ImageJ multipoint tool, normalized to total cell number, and expressed per 1000 smooth muscle cells.

Data Analysis

Single-cell RNA-sequencing analysis

Salmon Alevin 1.3.0 was used to create a cell by gene matrix for spliced and unspliced counts using the annotation from GENCODE 34 for GRCh38 (74, 75). Solo was used for doublet detection and removal and enriched softmax values in clusters were used as additional criteria for another round of doublet filtering (76). A minimum of 1000 genes and 40% mitochondrial cutoff were used to remove low quality cells from all datasets. The SCTransform workflow was used for count normalization (77). Principal component analysis was computed on the residuals for input into Harmony for batch correction (78). Control immune cells were not batch corrected due to inability to resolve rare cell types after correction. The parameters of Harmony were set to use the top 30 principal components. Uniform manifold approximation and projection (UMAP) embeddings and neighbors for Leiden clustering used the batch corrected embeddings (79, 80). RNA velocity analysis was done using the scVelo package (81). HGNC labels replaced all corresponding Ensembl gene ID and non HGNC annotated gene IDs were left in the matrix as Ensembl gene IDs. Latent time from scVelo was computed to order the cells. RNA velocity analysis was performed on batch corrected embeddings. dittoSeq was used for color-blind friendly plotting (82). CellChat was used to infer cell communication analysis (53). scMappR was used to deconvolute gene expression between bulk gene expression datasets using the AVM perivascular, endothelial, and immune cell types as the reference (61). Correlations between mouse and human endothelial cell types were calculated on the space of shared orthologous marker genes of clusters. fgsea was used to look for enriched hallmark pathways from differential expressed endothelial genes (83). T-statistic for cell type proportions were based on deconvolution output from scMappR. All marker genes were using a wilcoxon rank-sum test with a minimum of 0.5 average log fold-change cutoff and filtering for genes with <0.05 FDR with Bonferroni correction. UCell was used for the cell type identity score and utilized the top 30 genes of the cluster (84). More specifically, gene lists were generated using a wilcoxon rank sum test and filtered for statistical significance (FDR < 0.01) for the top 30 genes. For endothelial cells, this list was also intersected with published mouse datasets based on their specificity after aggregating into 4 different subclasses: artery, capillary, venule, and vein (13). Genes used for UCell scoring for each cell type are located in table S10.

Spatial transcriptomic analysis

The RNA spot table was log normalized and scaled before PCA. The top 10 components were used for UMAP and neighbor analysis for Leiden clustering using the scanpy package.

Bulk tissue RNA-sequencing analysis

Salmon 1.3.0 was used to pseudo align all ruptured and unruptured bulk AVM samples (85). The output of Salmon was used to generate counts using tximport (86). edgeR was used to compute differential gene expression between rupture and unruptured samples using the exact test (87). Genes with a false discovery rate (FDR) adjusted q-value < 0.05 were considered to be differentially expressed.

Statistical analysis

For all immunostaining and cell culture experiments, statistical analysis was performed with Student’s t-test for two-way comparisons and one-way ANOVA with post-hoc Tukey tests for comparisons of three or more groups using GraphPad Prism. Data is presented as mean ± standard error of the mean unless otherwise indicated with individual data points shown.

Schematics

Schematic cartoons in Fig. 5A were created with BioRender.com.