Spatial atlas of the human brain vasculature reveals specialized cell ensembles

Summary

The brain vasculature comprises diverse specialized cells that are essential for brain function, yet their spatial organization remains poorly understood. Here, we construct a comprehensive cerebrovascular cell atlas encompassing 314,535 transcriptomes that captures the arteriovenous axis and defines consensus cell states. We then perform spatial transcriptomics to map 1,529,740 cells across the human temporal cortex and hippocampus, uncovering stereotyped micro-communities termed vascular cell ensembles. These ensembles comprise specialized subsets of endothelial cells, mural cells, fibroblasts, and perivascular macrophages that align with the arteriovenous architecture to coordinate segment-specific functions, such as neurovascular coupling, blood-brain barrier transport, and immune surveillance. By overlaying genetic risk and pharmacologic reactivity, we identify ensemble-specific susceptibilities and candidate therapeutic targets across neurological diseases, including small vessel disease and stroke. This study provides a Resource to dissect the spatial and functional logic underlying human cerebrovascular biology and establishes a blueprint for decoding neurological disease susceptibility and therapeutic response.

Introduction

The cerebrovasculature and its closely associated leptomeninges are a highly specialized network of blood vessels and brain barriers that deliver nutrients and oxygen, clear metabolic waste, and regulate neuroimmune surveillance to maintain the brain microenvironment and neuronal function^1–10. Organized as a continuous arteriovenous axis, the cerebrovasculature spans from large arteries to draining veins within the subarachnoid space, with arterioles, capillaries, and venules traversing the brain parenchyma¹¹. Within a vast capillary bed, the neurovascular unit, comprising endothelial cells, pericytes, and astrocyte end feet, couples with neurons and glia to establish the blood-brain barrier (BBB), a highly selective interface between the brain and circulation^1–8. Recent studies show that other vascular segments also perform specialized roles. Arteriolar endothelial cells, together with other cells of the neurovascular unit, couple blood flow with neuronal activity^12–14, while venules coordinate leukocyte trafficking and shape neuroimmune responses^15,16. These studies, and others^17–25, suggest that spatially organized vascular segments carry out distinct functions.

Recent single-cell and single-nucleus RNA-sequencing (sc/snRNA-seq) studies have revealed extensive diversity across cerebrovascular cells^26–30. Endothelial cells exhibit gradual shifts in gene expression, termed “zonations”, that correspond to their position along the arteriovenous axis^31–34. Pericytes, smooth muscle cells, and fibroblasts also display transcriptional heterogeneity, indicating cellular specialization^{26–29,35,36}. However, the dissociation required for single-cell profiling disrupts tissue architecture, preventing direct mapping of cell diversity in situ. In addition, methodological and annotation differences across atlases have hindered efforts to harmonize vascular cell definitions, and the cerebrovasculature has been excluded from large-scale initiatives, such as the Brain Initiative Cell Census Network (BICCN)^30,37. As a result, we lack a cohesive framework to understand how molecularly distinct vascular cell populations organize across arteriovenous segments and brain regions, and how this organization shapes vascular function, disease susceptibility, and therapeutic response in humans.

To address this gap, we integrated five vascular-rich sc/snRNA-seq datasets^26–30 from 39 donors to construct a unified, publicly available cerebrovascular cell atlas that comprehensively captures the arteriovenous axis and defines consensus vascular cell states. Using this atlas, we curated a 300-gene panel for cell-resolution spatial transcriptomics of the human temporal cortex (n = 6 donors) and hippocampus (n = 7 donors)³⁸, enabling high-resolution mapping of vascular cell subsets across cortical laminae, hippocampal subfields, and the arteriovenous axis. By repeating core spatial analyses across two distinct brain regions, we assess the generalizability of cerebrovascular organization. We then leveraged this framework to map cell type-specific genetic susceptibility to neurological diseases and predict arteriovenous segment-specific drug responses. Our findings reveal that specialized vascular cell types organize into stereotyped “vascular cell ensembles”, cohesive, spatially patterned micro-communities aligned with arteriovenous architecture that coordinate gene expression programs to support functional specialization, confer disease susceptibilities, and define therapeutic vulnerabilities in the human brain.

Results

An integrated cell census of the human brain vasculature

To systematically characterize vascular cell diversity in the adult human brain, we integrated five vascular-rich sc/snRNA-seq datasets from brain tissues without cerebrovascular disease (n = 39 donors) (Figure 1A–B, Table S1)^26–30. Following quality control, we recovered 314,535 high-quality transcriptomes, including 167,573 vascular cells (Figure 1A). To account for differences in batch or dissociation methodologies, we used scVI³⁹ and scANVI⁴⁰ with a tiered integration strategy, as previously described (Figure S1A–H)^41,42. Graph-based Leiden clustering and uniform manifold approximation and projection (UMAP) plots identified 11 transcriptionally distinct cell populations from the vasculature and brain parenchyma (Figure 1B–D, S1A–H). Using consensus marker genes, we classified these populations into major vascular and vascular-associated cell types, including endothelial cells (CLDN5 and PECAM1), pericytes (HIGD1B and KCNJ8), smooth muscle cells (CNN1 and ACTA2), fibroblasts (FBLN1 and COL1A1), and perivascular macrophages (LYVE1 and MRC1) (Figure 1C–D). No vascular cell type was restricted to a single individual or dataset (Figure S1A–H).

Figure 1. Integration of cerebrovascular cell atlases.

A. Schematic highlighting each individual dataset preparation used to generate the integrated cerebrovascular cell atlas. Isolation highlights differences in vascular isolation methods between datasets. FANS, fluorescent-activated nuclei sorting; FACS, fluorescent-activated cell sorting. B. UMAP of the integrated cerebrovascular cell atlas. EC, endothelial cell (n = 60,373); PC, pericyte (n = 48,209); SMC, smooth muscle cell (n = 42,536); FB, fibroblast (n = 12,612); Neu, neuron (n = 32,594); OL, oligodendrocyte (n = 38,773); OPC, oligodendrocyte progenitor cell (n = 7,032); AC, astrocyte (n = 21,540); Myeloid (n = 37,341); Lymphoid (n = 6,433); Epen, ependymal cell (n = 7,092). C. Dot plot highlighting marker specificity across canonical cell types. D. Hierarchical clustering of annotated cell subsets. Neut, neutrophil; TC, T cell; BC, B cell; MG, microglia; PVM, perivascular macrophage; Mono, monocyte; pvFB, perivascular fibroblast. E. UMAP with annotations after iterative clustering of each major vascular and vascular-associated cell population. Colors match those in (D). F. Dot plot of global markers for each annotated cell subset. G. Sankey plot showing allocation of each cell transcriptome from the five datasets across canonical and subset annotations. Myl, myeloid; Ly, lymphoid; Ep, ependymal. The width of the stream corresponds to the relative cell number. H. Pie charts showing the fraction of endothelial subsets and mural cells captured in the integrated dataset and in each source dataset. See also Figures S1, S3, and S4.

Beyond canonical cell types, distinct transcriptional cell states have been reported in cerebrovascular cells^26–29,31. However, inconsistent annotations and biases in arteriovenous capture have confounded the identification of these cell subsets across studies^26–29 (Figure S1H, S1I–M). To harmonize annotations and comprehensively capture cerebrovascular cell transcriptional specialization, we sorted vascular, leptomeningeal, and immune cells in silico and performed iterative clustering (Figure 1E–G). Using this approach, published cerebrovascular cell atlases showed different biases in arteriovenous coverage (Figure 1H), reflecting differences in cell isolation methods^26–30. Iterative clustering revealed 16 cell subsets from five major vascular and vascular-associated cell types (endothelial cells, pericytes, smooth muscle cells, fibroblasts, and perivascular macrophages). Cell annotations were reproducible across donors regardless of sex or age (Figure S1N–Y). Unlike prior atlases^26–29, our integrated analysis resolved endothelial cells spanning each arteriovenous segment, including arteries, arterioles, capillaries, venules, and veins (n = 60,373 transcriptomes) (Figure 1F–H, S1Z–AA). We identified 2 distinct pericyte subsets, previously labeled as ‘matrix’ and ‘transport’ pericytes on the basis of marker gene expression²⁷ (n = 48,209 transcriptomes) (Figure 1E–F). For smooth muscle cells, we distinguished 3 cell states hypothesized to represent arterial, arteriolar, and venous smooth muscle cells (n = 42,536 transcriptomes) (Figure 1E–F). One prior atlas²⁶ annotated fibromyocytes in the cerebrovasculature; however, these cells also co-expressed venous smooth muscle cell signatures³⁶ (Figure S1AB–AC). To maintain consistency with segment-specific descriptions, we re-annotated this cell state as venous smooth muscle cells. Finally, we identified 5 fibroblast subsets, including dural border, arachnoid barrier, inner arachnoid, pial, and parenchymal perivascular fibroblasts (n = 12,612 transcriptomes), and one subset of perivascular macrophages (n = 3,843 transcriptomes), which were transcriptionally distinct from parenchymal microglia and circulating myeloid cells (Figure 1E–F). Therefore, our integrated cerebrovascular cell atlas provides a comprehensive catalog of vascular cell heterogeneity across the arteriovenous axis and a unified framework for molecular classification of vascular cell populations in the adult human brain (Table S2–3).

Spatial organization of vascular cell subsets in the human temporal cortex and hippocampus

To map the spatial distribution of vascular cell subsets in the human brain, we performed cell-resolution in situ spatial transcriptomics on coronal sections from the human middle temporal gyrus (n = 6 donors, 9 sections, Table S1) and hippocampus (n = 7 donors, 11 sections, Table S1) with the 10x Genomics Xenium platform³⁸ (Figure 2A). Using marker genes from the integrated cerebrovascular cell atlas, cortical laminar-specific neuronal genes from the Allen Brain Atlas⁴³, and hippocampal subfield-specific neuronal genes^44–46, we designed region-matched 300-gene panels to visualize vascular cell subsets and surrounding neuronal, glial, and leptomeningeal cells in the cortex and hippocampus (Table S4, Figure S2A, S2H). Following best practices for image-based spatial genomics⁴⁷, individual mRNAs were decoded and cell segmentation was performed to compile single-cell gene expression profiles that are highly reproducible between biological replicates (Figure S2B, S2I). After quality control and data integration, clustering yielded high-quality gene expression profiles from 1,155,946 cortical cells, including 101,929 vascular cells, and 373,794 hippocampal cells, including 52,739 vascular cells (Figure 2B–D, S2C–E, S2J–L).

Figure 2. Spatial analysis of the human cerebrovasculature in the temporal cortex and hippocampus.

A. Schematic showing preparation of Xenium sections for subcellular spatial transcriptomics of the cortical and hippocampal vasculature. Sect., sections; MTG, middle temporal gyrus; HC, hippocampus. B. Dot plot showing gene marker specificity across canonical cell types in the middle temporal gyrus (left) and hippocampus (right). EC, endothelial cell; PC, pericyte; SMC, smooth muscle cell; FB, fibroblast; Neu, neuron; OL, oligodendrocyte; OPC, oligodendrocyte progenitor cell; AC, astrocyte; ChP, choroid plexus. C. Hierarchical clustering of annotated cell subsets by middle temporal gyrus (left) and hippocampus (right). D. Spatial plot of canonical cell types in representative coronal sections from the middle temporal gyrus (left) and hippocampus (right). Scale bars: 2 mm. LM, leptomeninges; PM-L1, pia mater – layer 1; GM, gray matter; WM, white matter; PCL, pyramidal cell layer; NP, neuropil; GCL, granule cell layer. E. UMAP of low-lambda (λ = 0.2) BANKSY embeddings with reconciled annotations provided after iterative clustering of each major vascular and vascular-associated cell population in the middle temporal gyrus (top row) and hippocampus (middle row). F. Scatterplot of annotated vascular and myeloid cell subsets by value of Moran’s I calculated in the temporal cortex (CTX) and hippocampus (HC). Dashed line indicates the boundary where Moran’s I is equivalent between CTX and HC. Spearman’s correlation is calculated on the aggregated scatterplot. See also Figures S2, S3, and S4.

Based on marker gene expression, we identified 10 of the canonical cell populations revealed in our integrated cerebrovascular cell atlas in both regions, including all five major vascular and vascular-associated cell types (Figure 2B–D, S2C, S2J). Tissue sections captured spatially distinct cellular patterns in the temporal cortical gray and white matter, intervening sulci, adjacent leptomeninges, and vessel-dense subarachnoid space, as well as in the hippocampal laminae and subfields (Figure 2D, S2F–G, S2M–N). Iterative clustering analysis on in silico-sorted vascular cells identified the 15 vascular cell subsets found in the integrated atlas, including 5 endothelial subtypes (arterial, arteriole, capillary, venule, and vein; cortex: n = 57,826 transcriptomes; hippocampus: n = 28,043 transcriptomes), 2 pericyte subtypes (matrix and transport; cortex: n = 14,910 transcriptomes; hippocampus: n = 7,165 transcriptomes), 3 subtypes of smooth muscle cells (artery, arteriole, and venous; cortex: n = 7,518 transcriptomes; hippocampus: n = 4,229 transcriptomes), and 5 fibroblast subtypes (dural border, arachnoid barrier, inner arachnoid, pial, and parenchymal perivascular; cortex: n = 17,101 transcriptomes; hippocampus: n = 9,639 transcriptomes) (Figure 2E, Table S3). We also observed two spatially defined perivascular macrophage subsets (leptomeningeal and parenchymal; cortex: n = 4,574 transcriptomes; hippocampus: n = 3,663 transcriptomes) (Figure 2E). Cell-specific gene expression correlated well between the integrated cerebrovascular cell atlas and spatial transcriptomes (Figure S2O–U), and vascular cell states were spatially distributed similarly between temporal cortex and hippocampus (Figure 2F).

A well-described tenet of cerebrovascular organization is the graded alteration in endothelial cell gene expression that occurs with arteriovenous transitions, known as ‘zonations’^31,48–52. To examine arteriovenous zonations, we performed trajectory analysis of endothelial cells from our integrated cerebrovascular cell atlas to curate 33 high-fidelity arteriovenous zonation markers for spatial validation (Figure 1F, S1Z–AA, Table S4). Expression of transition markers such as MGP (arteries)⁵³, MFSD2A (capillaries)⁵⁴, and ACKR1 (veins)⁵⁵ reproduced well between the cell and spatial atlases, and trajectory analyses correctly assigned endothelial cells to arteriovenous identities in the spatial atlas (Figure S3A–H), which were highly conserved across the temporal cortex and hippocampus (Figure S3G–H). Newly identified endothelial arteriovenous markers, including DKK2 (arteries), PROM1 (capillaries), and TSHZ2 (veins), were validated with single-molecule fluorescent in situ hybridization (Figure S3I–J). Thus, this first-in-human spatially resolved cerebrovascular cell atlas of the temporal cortex and hippocampus delineates the distribution of molecularly defined vascular cell subsets and endothelial zonations in the human brain.

Because some endothelial zonation markers differ across species²⁷ and mice are widely used to study cerebrovascular structure and function^12,13,31, we curated and integrated 6 mouse neurovascular atlases (n = 206,654 transcriptomes) and harmonized mouse annotations to the human atlas to permit cross-species comparisons^{31,37,56–59} (Figure S4A–K). Consistent with prior work²⁷, we observed species-specific differences in endothelial cell gene expression and pathway enrichment (Figure S4L–N, Table S5). However, the majority (25 of 33 genes) of our endothelial zonation markers, including Dkk2, Prom1, and Tshz2, displayed conserved arteriovenous expression patterns (Figure S4M). Together, these analyses reveal that key components of endothelial arteriovenous zonation are conserved between the temporal cortex and hippocampus and across species.

Vascular organization varies according to cortical laminae and subregions

Stereotyped variations in vascular density and topology occur through the laminar organization of the cortex, and vascularization decreases markedly in white matter^11,60,61. We therefore sought to map molecularly defined vascular cells to distinct spatial niches in the human temporal cortex. To investigate vascular cytoarchitecture, we used leptomeningeal- and laminar-specific genes from the spatial atlas of the middle temporal gyrus to define a one-dimensional cortical depth axis from the subarachnoid space to the sub-gyral white matter⁴³ (Figure S5A–D). This standardized methodology accommodates and demarcates geometrically diverse cortical laminar boundaries across tissue sections (Figure 3A, S5C–D) and recapitulates known region-specific distributions in excitatory neurons and glia across donors (Figure 3B–C, S5C–E). Consistent with histological studies^11,62, larger arteries and veins (macrovasculature) were most prevalent in the subarachnoid space and cortical surface (Figure 3D). Conversely, microvascular cell density was greatest in the gray matter and peaked near the L3/4 boundary, while remaining broadly distributed across all cortical laminae, similar to other mammals^63,64 (Figure 3D–E, S5F). Unlike rodents, the human cortex is relatively scarce in venules^11,62,65. By quantifying ratios of molecularly defined arterial and venous endothelial cells across cortical regions, we found a venous predominance in the subcortical white matter, which we validated with immunostaining in an independent cohort (n = 3 donors) (Figure 3F, S5G), and a slight arterial predominance within the subarachnoid space and gray matter (Figure 3F, S5G). Within each lamina, nearest-neighbor analyses revealed remarkable consistency in mean distances between neuronal soma and endothelial cells (Figure 3G). While these results suggest stereotyped spatial microvascular patterning to support neuronal metabolic needs, this analysis is based on segmented cell bodies and therefore does not account for neuronal processes that directly contact the microvasculature and contribute to neurovascular coupling⁶⁶. Nonetheless, our spatial atlas reveals stereotyped arteriovenous patterning across the laminar architecture of the human temporal cortex.

Figure 3. Vascular organization across cortical subregions.

A. Spatial plot of a representative coronal section of the middle temporal gyrus colored by cortical laminar identity. LM, leptomeninges; PM-L1, pia mater – layer 1; L2/3, layer 2/3; L4, layer 4; L5, layer 5; L6, layer 6; WM, white matter. Scale bars: 2.5 mm. B. Dot plot of hypergeometric enrichment between cell populations (rows) and laminar identity (columns). Arachnoid mater comprises dural border, arachnoid barrier, and inner arachnoid fibroblasts. Macrovasculature comprises arterial and venous endothelial cells, arterial and venous smooth muscle cells, leptomeningeal perivascular macrophages, and pial fibroblasts. Glia limitans comprises pial astrocytes. Ex, excitatory neuron; OL, oligodendrocyte. C. Density plot of neurons and glia for each cortical lamina. D. Density plot of macrovasculature, microvasculature (arteriolar, capillary, and venular endothelial cells, matrix and transport pericytes, arteriolar smooth muscle cells, parenchymal perivascular fibroblasts and parenchymal perivascular macrophages), and pan-neurons (both excitatory and inhibitory neurons). E. Boxplot of vascular-to-total cell ratio between gray matter (GM) and white matter (WM) by donor. ***p < 0.001. F. Boxplot of venous-to-arterial endothelial cell ratio between LM, GM, and WM. Dashed line indicates equivalence point. *p < 0.05 and **p < 0.01. G. Boxplot of the average distance to the nearest capillary endothelial cell across each neuronal population. n.s., not significant. H. Neighborhood complexity across cortical depth defined as the number of vascular cell types (left) or cell types of any class (right) within a 100 μm search radius of the endothelium. See also Figures S5 and S6.

Next, we investigated whether the distribution of the identified vascular cell subsets differed according to vascular architecture and cortical regions. Using established methods^67–69 (Methods), we quantified local neighborhood complexity in spatially segregated cortical regions. Considering only vascular cell populations, neighborhood complexity was greatest in larger pial arteries and veins of the subarachnoid space (Figure 3H). When neuronal and glial cell populations were included, we observed greater neurovascular neighborhood complexity within the microvascular-rich cortical gray matter, specifically L2-5 (Figure 3H).

Mural cell coverage of the endothelium contributes to regional variations in the BBB^70–72. We quantified ratios of mural and perivascular cell subsets in each cortical region, normalized to endothelial cell densities in each region. Consistent with prior reports^73,74, pericyte abundance was greatest in the cortical gray matter (Figure S5H). Between cortical laminae, pericyte abundance varied according to their molecular identity (Figure S5I). For example, transport pericytes were more prevalent in more superficial cortical gray matter laminae (L1-5), which we validated with immunostaining (Figure S5I–J). Most smooth muscle cell subsets were enriched in larger arteries within the subarachnoid space (Figure S5K). Notably, venous smooth muscle cells were exclusive to veins within the subarachnoid space (Figure S5K–L), whereas intraparenchymal venules lacked venous smooth muscle cells and were instead associated primarily with matrix pericytes (Figure 5C–D). Other perivascular cell subsets, including most fibroblasts and perivascular macrophages, were similarly enriched in the subarachnoid space (Figure S5M–P). Together, these data molecularly resolve two predominant cerebrovascular niches with distinct cellular microenvironments: the leptomeningeal subarachnoid space and the intraparenchymal microvasculature. Rather than redefining this established anatomical distinction, our analysis maps the vascular and perivascular cell composition of these niches and shows that intraparenchymal vascular cell composition varies with the laminar and subregional patterning of neuronal and glial populations. This spatial alignment suggests coordinated co-development of specialized neurovascular ensembles, which encompass endothelial cells, mural cells, fibroblasts, and perivascular macrophages, across the human temporal cortex.

Figure 5. Mural cell arteriovenous organization and specialization in the temporal cortex.

A. Slingshot trajectory of cortical endothelial cells (ECs) across the arteriovenous axis. B. Representative spatial plot of the middle temporal gyrus showing the distribution of ECs across the arteriovenous segments of the section shown previously (in Figure 2D). Scale bars: 2.5 mm. C. Density plots showing the frequency of neighboring cell types within a 10 μm search radius across the endothelial arteriovenous score defined by the Slingshot trajectory in (A). SMC, smooth muscle cells; PC, pericytes. D. Dot plot showing hypergeometric test results for mural cells in the 10 μm neighborhood of each arteriovenous segment. Only highly significant associations are shown (odds ratio [OR] > 5, false discovery rate [FDR] < 1 × 10–50). E. Violin plot of contractility (left) and migration (right) gene signature scores between arterial and arteriolar SMCs (gray) and venous SMCs (light yellow). ***p < 0.001. F. Violin plot of extracellular matrix (ECM) organization (left) and synaptic signaling (right) gene signature scores between arterial SMCs (light blue) and arteriolar SMCs (green). G. Top Gene Ontology (GO) enrichments between matrix PCs (yellow) and transport PCs (brown). Dashed line indicates an FDR significance threshold of 0.05. H. Top GO enrichments of aggregated endothelial and mural cell communities in the arterial (red), capillary (purple), and venous segments (blue). Dashed line indicates an FDR significance threshold of 0.05. I. Cartoon showing stereotyped vascular cell ensemble organization and function along the arteriovenous axis. aSMC, arterial smooth muscle cell; aaSMC, arteriolar smooth muscle cell; vSMC, venous smooth muscle cell; PVM, perivascular macrophage; AV, arteriovenous. See also Figure S7.

Hippocampal vascular organization reveals conserved spatial principles with region-specific specialization

To test whether the spatial organizing principles defined in the middle temporal gyrus generalize beyond the cortex, we mapped molecularly defined vascular cell subsets within the human hippocampus. Using computational approaches similar to those used for the cortex (Methods), we created a one-dimensional hippocampal transverse axis spanning the principal neuronal layers of each subfield (Figure S5Q). This approach accommodated the complex geometry of hippocampal subfields (Figure 4A–C, S5R–V) and accurately recapitulated established neuronal distributions across donors (Figure S5S–T).

Figure 4. Vascular organization across hippocampal subfields.

A. Spatial plot of a representative coronal section of the hippocampus colored by hippocampal subfield identity. LV, lateral ventricle; WM, white matter; CA, cornu ammonis; NP, neuropil; GC, granule cells. Scale bars: 2 mm. B. Dot plot of hypergeometric enrichment between cell populations (rows) and subregion identity (columns). Arachnoid mater comprises dural border, arachnoid barrier, and inner arachnoid fibroblasts. Macrovasculature comprises arterial/venous endothelial cells, arterial/venous smooth muscle cells, leptomeningeal perivascular macrophages, and pial fibroblasts. Microvasculature comprises arteriolar, capillary, and venular endothelial cells, arteriolar smooth muscle cells, matrix and transport pericytes, parenchymal perivascular fibroblasts, and parenchymal perivascular macrophages. Glia comprises oligodendrocytes, oligodendrocyte progenitor cells, and astrocytes. IN, interneuron; Sub, subiculum; Epen, ependymal cell; GC, granule cell; ChP, choroid plexus. C. Density plot of neurons and glia for each hippocampal subfield, ranging from CA4 to subiculum. D. Density plot of microvasculature and pan-neurons (both excitatory and inhibitory neurons) across hippocampal subfields. E. Boxplot of vascular-to-total cell ratio between the leptomeninges (LM), stratum granulosum (SG), stratum pyramidale (SP), neuropil (NP), and white matter (WM) by donor. **p < 0.01 and ***p < 0.001. F. Boxplot of venous-to-arterial endothelial cell ratio between LM, SG, SP, NP, and WM. Dashed line indicates equivalence point. *p < 0.05. G. Boxplot of the average distance to the nearest capillary endothelial cell across each neuronal population. n.s., not significant. H. Neighborhood complexity across hippocampal subregions defined as the number of vascular cell types (left) or cell types of any class (right) within a 100 μm search radius of the endothelium. See also Figures S5 and S6.

As with the cortex and consistent with histologic studies^75,76, larger arteries and veins were most prevalent within the subarachnoid space (Figure 4B). Consistent with prior histologic analyses^61,77–79, the hippocampus exhibited lower vascular cell density and greater neuron-to-capillary distances than the cortex (Figure S5W–X). Unlike the cortex (Figure 3E), microvascular distributions in the hippocampus did not parallel neuronal soma density but were instead more prevalent in the synapse-dense neuropil-rich layers (Figure 4D–E). This configuration may reflect proximity to vascular entry routes into the hippocampus, including arterial penetration along the hippocampal sulcus^75,76,80, and may contribute to the well-established vulnerability of the hippocampus to ischemic injury^81–83.

Despite these region-specific angioarchitectural differences, many spatial features of vascular organization were conserved at the cellular level. Each vascular cell subset defined in the integrated cell atlas was identified in the hippocampus (Figure S2P–U). Similar to cortex, quantification of molecularly defined arterial and venous endothelial cells demonstrated venous enrichment within hippocampal white matter tracts, which was validated with immunostaining in an independent cohort (n = 3 donors) (Figures 4F, S5Y). Within each hippocampal subfield, nearest-neighbor analyses showed consistent mean distances between neuronal somata and endothelial cells (Figure 4G). Neighborhood complexity analyses across hippocampal domains showed that vascular neighborhood complexity was greatest in larger arteries and veins within the subarachnoid space (Figure 4H). When neurons and glia were incorporated to assess the neurovascular unit, the greatest neurovascular complexity localized to principal neuronal layers within hippocampal gray matter (Figure 4H). Thus, although hippocampal neuronal cytoarchitecture differs markedly from neocortex, the same two predominant cerebrovascular niches were observed: the leptomeningeal subarachnoid space and the intraparenchymal microvasculature (Figure S5Z–AG).

While vascular cell identity remains unchanged between cortex and hippocampus, prior studies have demonstrated region-specific transcriptional specialization of endothelial cells and pericytes^84,85. Using the integrated cell atlas, we performed differential gene expression analysis between vascular cells from the temporal cortex and hippocampus with established methods⁸⁵. Consistent with prior reports^84,85, vascular differential gene expression was greatest in microvascular pericyte and endothelial subsets (Figure S6A). In matrix and transport pericytes, gene ontology analysis showed hippocampal enrichment of ionic transport and homeostasis pathways (Figure S6B), suggesting a possible regional shift in pericyte composition. Using the spatial atlases, we quantified relative microvascular cell abundance between cortex and hippocampus (Figure S6C). This analysis showed reduced matrix pericytes and an increased transport-to-matrix pericyte ratio in the hippocampus, which was independently validated by immunostaining in a separate tissue cohort (Figure S6D–E). This shift is consistent with altered hippocampal pericyte contractility reported in mice⁷⁷. Collectively, these findings demonstrate that while the fundamental spatial logic of vascular cell organization into ensembles is conserved across brain regions, microvascular transcriptional programs and mural cell composition are regionally tuned. Extending this spatial framework to additional brain regions in future studies will reveal how conserved vascular ensembles adapt to support distinct neuronal microenvironments across the human brain.

Mural cells are distributed according to arteriovenous specialization

Mural cells occupy different segments of the arteriovenous axis across diverse organ-specific vascular beds and they exhibit distinct morphologies and gene expression programs^{31,36,71,86–90}. Prior efforts to zonate mural cells were not anchored to endothelial arteriovenous identities^27,31, and therefore, arteriovenous localization of mural cell subsets has remained imprecise and controversial, particularly in the human cerebrovasculature^{18,20,31,86,91–96}.

To address this, we used spatially resolved, arteriovenous zonated endothelial transcriptomes from our spatial atlas to construct a one-dimensional axis spanning the arteriovenous axis in both the temporal cortex and hippocampus (Figure 5A, Figure S7A). This method accounts for varied geometric orientations in endothelial vascular fragments within tissue sections and reconstructs endothelial zonations in silico (Figure 5A, Figure S7A). Using a nearest-neighbor analysis, we spatially mapped each molecularly defined mural cell subset to its nearest endothelial cell. The distribution of each mural cell was then plotted along the endothelial arteriovenous axis to identify segment-specific mural cell enrichments (Figure 5B–D, Figure S7B–C). This analysis confirmed that arterial, arteriolar, and venous smooth muscle cells reside near their respective endothelial cells in both brain regions (Figure 5B–D, Figure S7B–C). Matrix and transport pericytes spatially occupied discrete arteriovenous segments: matrix pericytes were localized to arterioles and venules, whereas transport pericytes were more prevalent in capillaries (Figure 5B–D, Figure S7B–C). Thus, molecularly defined smooth muscle cell and pericyte subsets are not randomly distributed across the vasculature as previously suggested²⁷ but instead align with specific arteriovenous segments in the human temporal cortex and hippocampus.

Given these findings, we postulated that mural cells are uniquely specialized to support the arteriovenous segments to which they localize. Gene ontology analyses of biological processes showed that arterial smooth muscle cells were enriched in the expression of contractile, cytoskeletal, and extracellular matrix genes, congruent with their role in modulating the diameter of larger arteries (Figure 5E, S7D, Table S6). Conversely, venous smooth muscle cells showed heightened migratory signatures (Figure 5E, S7D, Table S6), possibly reflecting a higher angiogenic potential of veins in the postnatal brain⁹⁷. Arterioles regulate regional cerebral blood flow to meet neuronal metabolic needs, a process known as neurovascular coupling^12,18,21.

Prior studies in mice have demonstrated direct neuro-arteriolar smooth muscle junctions in which perivascular neurons form synapse-like contacts with smooth muscle cells^21,98–101. In alignment with these anatomical observations, human arteriolar smooth muscle cells were enriched for genes implicated in synaptic signaling pathways (Figure 5F, S7E, Table S6). Cross-species pathway enrichment analyses between mouse and human smooth muscle cell subsets were largely conserved (Figure S7F, Table S5), suggesting relative preservation of smooth muscle cell subset transcriptional programs across species compared with other vascular cell types. However, dedicated functional studies in human arteriolar smooth muscle cells will be required to determine whether these pathways or specific genes contribute to neurovascular coupling.

Within the microvasculature, matrix pericytes of arterioles and venules were enriched in the expression of extracellular matrix or basement membrane genes, e.g., COL4A1 and COL4A2, suggesting stabilizing properties (Figure 5G). Transport pericytes in capillaries, the site of most molecular exchange across the BBB, were enriched in the expression of multiple transporters, such as the solute carrier genes SLC20A2, SLC6A1, and SLC12A7 (Figure 5G). To determine whether transcriptionally defined matrix and transport pericytes correspond to described morphological subclasses, e.g., mesh and thin-strand pericytes^86,93,94, we performed single-molecule fluorescent in situ hybridization for ADAMTS1 (matrix pericytes) and CA4 (transport pericytes) with co-immunostaining for platelet-derived growth factor receptor β (PDGFRβ), a pan-pericyte marker used to assess pericyte morphology^102–105. ADAMTS1⁺ matrix pericytes exhibited a mesh-like morphology, whereas CA4⁺ transport pericytes displayed thin-strand morphologies (Figure S7G), demonstrating concordance between transcriptional identity and morphology in humans. Thus, mural cells show specialization to support vascular structure and/or function congruent with their arteriovenous localization.

We next compared pericyte gene expression between mice and humans. Despite capturing 30,996 pericyte transcriptomes in the integrated mouse cerebrovascular cell atlas, we could not resolve distinct matrix and transport pericyte subsets in mice by transcriptomics or immunostaining (Figure S7H–I). Consistent with prior reports^27,106, cross-species comparisons of pericyte gene expression demonstrated human-specific enrichment in multiple transporters (Figure S7H), e.g., SLC6A1 and SLC20A2, and BBB transport pathways in humans (Figure S7J, Table S5). These findings suggest specialization of human pericytes for capillary transport functions; whether such pericyte specialization is unique to humans warrants future investigation.

Mural cell identity at arteriole-capillary transition zones is often debated^14,91. Morphological studies describe transitional cells with hybrid smooth muscle cells and pericyte features^93,94, whereas single-cell transcriptomic studies report clearer separation between these populations^{27,31,107,108}. Our spatial atlas showed that arteriolar smooth muscle cells and matrix pericytes both localize to arterioles in the human temporal cortex and hippocampus (Figure 5C–D, Figure S7B–C). To determine whether mural cell identity reflects a transcriptional continuum or discrete populations, we applied statistical testing for unimodality versus multimodality in the integrated cell atlas^109,110 (Methods, Figure S7K–L). This analysis revealed clear transcriptional separation between smooth muscle cells and pericytes, consistent with prior mouse and human single-cell studies^27,31,108. In human cortical arterioles, single-molecule fluorescent in situ hybridization with co-immunostaining identified ADAMTS1⁺ matrix pericytes as a discrete population embedded within the collagen IV⁺ vascular basement membrane (Figure S7M), supporting their identity as a distinct pericyte population rather than a transitional smooth muscle cell state¹⁴.

We next investigated whether endothelial and mural cells act in concert to confer the functional specialization of arteries, capillaries, and veins. Endothelial and mural cells were re-annotated by arteriovenous segment into vascular cell ensembles, and gene ontologies were calculated for each arteriovenous segment. This revealed striking synergy with established segment-specific functions, such as vasomotor responses in arteries and arterioles^18,95, BBB transport in capillaries^111,112, and leukocyte chemotaxis in venules and veins^15,16 (Figure 5H). Comparison between macro- and microvascular differences in the arterial and venous ensembles, respectively, showed a heightened synaptic signaling signature within the microvascular ensembles, suggesting nuanced tuning of these pericapillary segments to the dynamic microenvironment within the brain parenchyma^113,114 (Figure S7N–O).

Because cerebrovascular function is shaped by multidirectional interactions among vascular cells, neurons, and glia^14,115,116, we next asked whether patterned mural cell molecular specialization contributes to differential arteriovenous neurovascular interactions through reciprocal ligand-receptor interactions¹¹⁷ (Figure S7P–Q). Excitatory and inhibitory neurons are known to differentially interact with the cerebrovasculature^{100,118–120}, and we found that these neuronal subtype-specific interaction patterns extended to selective engagement with arteriolar, capillary, and venular vascular ensembles (Figure S7P–Q). Notably, some segment-specific interactions corresponded to differential mural cell gene expression (Figure S7P). For example, tenascin-R (TNR), an extracellular glycoprotein near neuronal synapses^121,122, was expressed by excitatory neurons and predicted to interact with integrin receptors, such as ITGB1, which was enriched in arteriolar smooth muscle cells, consistent with roles for integrin signaling in smooth muscle vasomotor responses and vascular stability^123,124 (Figure S7R–S). Conversely, netrin signaling, which regulates axonal guidance, vascular patterning, and arterial innervation, showed predicted interactions between smooth muscle ligands (NTN1/NTN4) and excitatory neuron receptors (UNC5/DCC) enriched in arteriolar ensembles^28,125,126 (Figure S7R–S). Together, these findings suggest that mural cell gene expression is adapted to complement endothelial cells in the same arteriovenous segment and contribute to segment-specific neurovascular communication. Thus, we propose that distinct arteriovenous molecular identities arise from coordinated endothelial-mural cell transcriptional modules in the human temporal cortex and hippocampus (Figure 5I).

Fibroblast subsets facilitate neurovascular and meningeal-vascular crosstalk

Brain fibroblasts are understudied cells in the brain that occupy perivascular spaces, meninges, and choroid plexus¹²⁷. Recently, brain fibroblasts were proposed to be more diverse than previously appreciated, particularly after neurological injury^57,128. However, brain fibroblasts are the most discordantly annotated cell population in prior sc/snRNA-seq atlases^27,29 (Figure S1H), and the lack of consensus molecular markers has hampered efforts to characterize them.

By plotting spatial coordinates of the five molecularly defined fibroblast subsets (KCNMA1⁺ dural border fibroblasts, STXBP6⁺ arachnoid barrier fibroblasts, SLC5A5⁺ inner arachnoid fibroblasts, NGFR⁺ pial fibroblasts, and ABCA8⁺ parenchymal perivascular fibroblasts) identified in our integrated sc/snRNA-seq atlas and the spatial atlases, we mapped fibroblast subsets to discrete niches within the human temporal cortex and hippocampus and across the arteriovenous axis in each region (Figure 6A). Dural border, arachnoid barrier, and inner arachnoid fibroblasts were enriched within arachnoid layers defining the subarachnoid space (Figure 6A–B). Pial fibroblasts lined cortical gyri or intervening sulci and ensheathed larger blood vessels (Figure 6A–B). Parenchymal perivascular fibroblasts were associated with smaller, penetrating blood vessels of the temporal gray and white matter (Figure 6A–B). Using endothelial-defined arteriovenous zonations (Figure 5A), pial fibroblasts were localized to larger arteries and veins (Figure 6C–D). Parenchymal perivascular fibroblasts were located along penetrating arterioles and venules, but not capillaries (which lack a perivascular space), similar to observations in mice^129,130. Dural border, arachnoid barrier, and inner arachnoid fibroblasts were not closely associated with the vasculature, supporting their localization to arachnoid barriers or trabeculae (Figure 6C). Despite distinct hippocampal and cortical cytoarchitecture, fibroblast subsets displayed similar spatial distributions and arteriovenous enrichments in the hippocampus (Figure S8A–C).

Figure 6. Perivascular cell organization and specialization in the temporal cortex.

A. Spatial plot of cortical fibroblast (FB) subsets and perivascular macrophages (PVM) with high-magnification insets indicating vascular cell niches located in the arachnoid mater (top, i), pial vasculature within a sulcus (middle, ii), and penetrating intraparenchymal blood vessels (bottom, iii). Main image scale bar: 2.5 mm. Inset scale bars: 250 μm. LM, leptomeninges; GM, gray matter; WM, white matter; PM-L1, pia mater – layer 1; AM, arachnoid mater; SAS, subarachnoid space; pvFB, perivascular fibroblast. B. Dot plot showing hypergeometric test results for perivascular cells within the 20 μm neighborhood of each cortical subregion. Only highly significant associations are shown (OR > 5, FDR < 1x10⁻⁵⁰). C. Same as (B), except for each arteriovenous segment. EC, endothelial cell. D. Density plot showing the frequency of neighboring perivascular cell types within a 20 μm search radius across the endothelial arteriovenous score. E. Representative confocal microscopy analysis of each fibroblast population in human temporal cortex. AM, arachnoid mater; SAS, subarachnoid space; PM, pia mater; BV, blood vessel. Scale bars: 100 μm. F. Top GO enrichments for each fibroblast subset. Black dashed line indicates an FDR significance threshold of 0.05. G. Scatterplot of outgoing (x-axis) and incoming (y-axis) interaction strength across all cell populations in the cell atlas. Dashed line highlights top fibroblast clusters along the “outgoing interaction strength” axis (i.e., the cumulative signal strength sent by the cell type to all other cell types). EC, endothelial cell; PC, pericyte; SMC, smooth muscle cell; FB, fibroblast; MG, microglia; PVM, perivascular macrophage; EX, excitatory neuron; IN, inhibitory neuron; OL, oligodendrocyte; OPC, oligodendrocyte progenitor cell; AC, astrocyte; Epen, ependymal cell. See also Figure S8.

As validation, we performed immunostaining in non-pathological human cortical specimens to localize the proteins encoded by the top gene markers of each fibroblast subset. As predicted by our spatial atlas, KCNMA1 (Dural Border), STXBP6 (Arachnoid Barrier), and SLC5A5 (Inner Arachnoid) co-localized in the arachnoid mater, but not pia mater, with the COL1A1 marker for pan-fibroblast identity^35,131, confirming these three fibroblast subsets as arachnoid fibroblasts (Figure 6E). NGFR (Pial FB) ensheathed the major blood vessels in the subarachnoid space, and failed to label the arachnoid mater and the parenchymal vessels (Figure 6E), matching the known anatomy of the pia mater¹³². Finally, ABCA8 (parenchymal perivascular fibroblasts) stained a subset of intermediate caliber vessels in the parenchyma but not all COL1A1⁺ vascular fragments (Figure 6E), consistent with the selective restriction of ABCA8⁺ fibroblasts to the perivascular spaces in the pericapillary vessels^129,130. Recent work has proposed an additional PROX1-labeled mesothelial layer within the arachnoid mater that partitions the subarachnoid space, termed the subarachnoid lymphatic-like membrane^133,134. Others have suggested that arachnoid PROX1 expression labels inner arachnoid fibroblasts¹³⁵. In our integrated human cerebrovascular cell atlas, PROX1 expression was sparse, whereas the integrated mouse cerebrovascular cell atlas showed greater Prox1 enrichment in inner arachnoid fibroblasts (Figure S8D). Consistent with this, PROX1 immunostaining co-localized with SLC5A5⁺ inner arachnoid fibroblasts in human cortical specimens (Figure S8E), supporting that inner arachnoid fibroblasts can express PROX1.

To understand the function of the five fibroblast subsets, we performed pathway enrichment analysis. Dural border and arachnoid barrier fibroblasts were enriched in the expression of contractile proteins and adherens junction complexes, respectively (Figure 6F). Inner arachnoid fibroblasts were enriched in the expression of solute carrier transporters, e.g., SLC5A5, SLC13A3, and SLC22A6 (Table S2), suggesting a role in regulating cerebrospinal fluid composition within the subarachnoid space^136,137. Pial fibroblasts were enriched in the expression of extracellular matrix proteins, e.g., FBLN2 and LUM, indicating a role in providing stromal support at the cortical surface and subarachnoid space (Table S2). Parenchymal perivascular fibroblasts expressed multiple ATP-binding cassette transporters, e.g., ABCA8 and ABCA10, and were enriched in the expression of genes associated with cell signaling and communication cascades such as NTRK3, a neurotrophic receptor tyrosine kinase, suggesting a contribution to neuronal survival and synaptic plasticity (Figure 6F, Table S2). We next investigated how fibroblast subsets contribute to cellular communication within the neurovascular unit^3,14,138. Using reciprocal ligand-receptor interactions in CellChat¹¹⁷, we assessed cell-to-cell communication pathways across cell subsets. Fibroblasts, especially parenchymal perivascular, pial, and inner arachnoid populations, emerged as top outgoing cell signalers (i.e., ligand producers) (Figure 6G). Extracellular signaling pathways involving laminin and collagen were key drivers of communication with endothelial cells, pericytes, smooth muscle cells, perivascular macrophages, and other brain fibroblasts (Figure S8F–G). Together, these data suggest that brain fibroblasts functionally specialize and fulfill underappreciated and spatially diverse roles in neurovascular and meningeal-vascular crosstalk in humans.

Brain fibroblasts are increasingly implicated in the pathogenesis of neurological diseases in mouse models^131,139,140. However, whether brain fibroblast heterogeneity is conserved across species remains uncertain, limiting the interpretation and translation of these findings to humans. We therefore performed cross-species comparisons of gene expression and pathway enrichment using integrated mouse and human brain fibroblast scRNA-seq datasets (Figure S8H–I). Conserved fibroblast populations showed concordant orthologous marker gene expression and similar functional specializations (Figure S8J–K, Table S5). Human fibroblasts were enriched for genes associated with neurotransmission, such as regulation of trans-synaptic signaling and neuronal projection development (Figure S8L). Thus, fibroblasts are uniquely specialized to support higher-order neurologic function within ensembles in the human brain, and our molecular and spatial mapping of human fibroblasts can guide the interpretation and translation of studies in mice.

Perivascular macrophages are specialized ensemble-associated immune cells

Perivascular macrophages are brain-resident border-associated macrophages positioned abluminal to the vascular wall within the Virchow-Robin perivascular spaces^141–144. These phagocytes contribute to brain immune surveillance and have been implicated in neurovascular dysfunction across multiple diseases, including Alzheimer’s disease (AD), stroke, hypertension, and CNS infections^{141,143,145–147}. Although identified in prior atlases of the human brain vasculature^26–28, perivascular macrophages have not been systematically characterized within a spatially resolved framework in humans.

Beyond established markers (LYVE1 and MRC1), both our integrated cell and spatial atlases confirmed enriched expression of CD163 and F13A1 (Figure S8M), supporting prior studies nominating these genes as perivascular macrophage markers^148–152. Iterative clustering did not reveal additional robust transcriptional heterogeneity, consistent with recent mouse analyses¹⁵³, although deeper sampling may resolve more nuanced niche-specific specialization. In our spatial atlases of temporal cortex and hippocampus, perivascular macrophages localized to two principal niches: the leptomeninges and the intraparenchymal perivascular space (Figure 6B, S8M). Spatial gene expression was highly concordant amongst perivascular macrophages (Figure S8N), consistent with the shared ontogeny of leptomeningeal and parenchymal populations¹⁵⁴.

To define their distribution along the arteriovenous axis, we mapped perivascular macrophages onto spatially resolved endothelial zonations. Consistent with mouse studies^153–155, leptomeningeal macrophages were enriched along larger arteries and veins within the subarachnoid space (Figure 6C–D, S8B–C). Parenchymal perivascular macrophages preferentially localized to penetrating arterioles and venules but were excluded from capillaries (Figure 6C–D, S8B–C). As reported in mice¹⁵⁴, perivascular macrophages were more prevalent along arteries and arteriolar segments than venous segments (Figure S8O), suggesting structured perivascular macrophage specialization along the arteriovenous axis in both the temporal cortex and hippocampus.

While perivascular macrophages are often studied in pathological settings, they signal to multiple vascular cell types, including endothelial cells, smooth muscle cells, and perivascular fibroblasts, to modulate vascular reactivity, barrier function, and immune responses^{145,146,154,156}. We therefore examined their role in neurovascular unit communication. Reciprocal ligand-receptor analysis showed that perivascular macrophages were not among the top incoming and outgoing signaling hubs in the assembled neurovascular interactome¹¹⁷ (Figure 6G). However, directed analysis revealed nuanced communication patterns: select fibroblast and smooth muscle cell subsets were prominent sources of incoming signaling, whereas perivascular macrophages exhibited outgoing interactions with multiple parenchymal cell populations (Figure S8P). We also cataloged ligand-receptor pairs patterned across arteriovenous vascular ensembles that may mediate perivascular macrophage communication (Figure S8Q) and can guide future studies.

Because mouse models are widely used to study perivascular macrophage biology^145,154, we compared gene expression across species. Human perivascular macrophages showed relative enrichment of ferritin-encoding genes, including FTL and FTH1, consistent with roles in iron handling and perivascular clearance of blood-derived products^157,158 (Figure S8R, Table S5). Immune-related pathways were conserved across species, but human cells exhibited broader enrichment of immune, oxidative stress, and biotic stimulus programs (Figure S8S), which may reflect greater cumulative and environmental immune exposure in humans^159,160. Collectively, these spatially resolved analyses establish perivascular macrophages as segmentally aligned, border-associated immune cells with defined niche localization, patterned communication, and species-nuanced molecular programs.

Genetic risk factors for neurological diseases affect specific vascular cell subsets

Cerebrovascular diseases are a leading global cause of death and disability^161,162. Although cerebrovascular contributions to neurological diseases are increasingly recognized^{1,2,5,8,163,164}, limited characterization of vascular cell subsets and how they are organized constrains our understanding of how vascular alterations contribute to neurological diseases.

Using our integrated human cerebrovascular cell atlas and common genetic variants that predispose to diverse neurological diseases, we sought to map genetic susceptibility in vascular cell subsets and according to arteriovenous zonation^{103,104,165–169} (Figure 7A). We curated risk genes identified through genome-wide association studies (GWASs) across four disease classes: (1) cerebrovascular (stroke and small vessel disease), (2) neuroimmune, (3) neurodegeneration, and (4) headache^170–184 (Table S7). To identify disease-relevant cell populations, we applied multi-marker analysis of genomic annotation (MAGMA)¹⁸⁵ to map GWAS risk loci to relevant cell type(s) (Figure 7B). As a control, ambidexterity showed no vascular associations, as expected for a trait that is unrelated to cerebrovascular dysfunction or neurological disease¹⁸⁴ (Figure 7B). We also observed well-known myeloid cell associations with multiple sclerosis^177,186,187 (Figure 7B), supporting the specificity of the enrichment analysis.

Figure 7. Vascular cell- and ensemble-specific disease and therapeutic associations.

A. Schematic of the computational workflow for mapping common genetic risk to cell types using MAGMA. B. Heatmap with MAGMA-based enrichments across vascular (left) and parenchymal cell types (right). ICH, intracerebral hemorrhage; cSVD, cerebral small vessel disease; PVS, perivascular space; HC, hippocampus; FA, fractional anisotropy; WMHV, white matter hyperintensity volume; MS, multiple sclerosis; FTD, frontotemporal dementia; ALS, amyotrophic lateral sclerosis; AD, Alzheimer’s disease; NI, neuroimmune diseases; ND, neurodegenerative diseases; HA, headache disorders; MISC, miscellaneous traits. EC, endothelial cell; PC, pericyte; SMC, smooth muscle cell; FB, fibroblast; PVM, perivascular macrophage; MG, microglia; Mono, monocyte; Neut, neutrophil; TC, T cell; BC, B cell; Epen, ependymal cell; Neu, neuron; OL, oligodendrocyte; OPC, oligodendrocyte progenitor cell; AC, astrocyte. *, Bonferroni-corrected p < 0.05, **p < 0.01, and ***p < 0.001. C. Scatterplot of cell types shown in (B) plotted by log-transformed Bonferroni-corrected P values and the number of statistically significant associations with the diseases in (B). D. Same as (B), except with vascular cell types aggregated by vascular cell ensembles in different arteriovenous segments. E. Top, feature plots of annotated endothelial cell subsets and composite reactivity scores for ACE inhibitors and interleukin receptor antagonists. Bottom, heatmap of scaled Drug2Cell reactivity scores for individual ACE inhibitors and interleukin receptor antagonists across endothelial cell subsets. F. Gene set enrichments between the monogenic stroke panel and top gene markers defining each cell subset in the atlas. Cell abbreviations as in (B). LAA, large artery atherosclerosis; LAN, large artery non-atherosclerotic; cSVD, cerebral small vessel disease; CE, cardioembolic; Coag., coagulation; MB, metabolic; ICH, intracerebral hemorrhage; VM, vascular malformation; WMH, white matter hyperintensities; BGCs, basal ganglia calcifications. G. Heatmap of scaled Drug2Cell reactivity scores for major classes of drugs currently considered for cSVD treatment and atlas cell subsets. Cell abbreviations as in (B). See also Figure S9.

We examined cell-subset-specific vulnerabilities to cerebrovascular diseases. Each cerebrovascular disease showed significant mural cell enrichments regardless of ischemic or hemorrhagic etiology (Figure 7B). For instance, cerebral small vessel disease (cSVD) contributes to age-dependent diseases, including stroke and certain etiologies of dementia, but is heterogeneous and principally diagnosed radiographically¹⁸⁸. Intriguingly, GWAS risk genes for early cSVD (characterized by white matter perivascular space burden) were enriched in matrix and transport pericytes, whereas GWAS risk genes associated with additional MRI findings of white matter structural integrity related to cSVD (determined by fractional anisotropy) were enriched in smooth muscle cells, suggesting possible differential vascular cell contributions across cSVD progression¹⁷⁵ (Figure 7B). While these findings are consistent with existing experimental and clinical studies^189–201, longitudinal and functional studies are required to determine whether these enrichments reflect a temporal disease trajectory.

Other neurological diseases, such as neurodegeneration and headaches, have more complex vascular contributions^{163,181,202–206}. In AD and related dementias, where vascular contributions are multifactorial^207,203, our enrichment analysis identified perivascular macrophages as the vascular-associated cell population showing the strongest genetic association with AD risk (Figure 7B). This finding was reproduced in an independent AD brain single-cell dataset (Figure S9A–B), consistent with prior studies implicating perivascular macrophages in amyloid-β-associated neurovascular dysfunction^27,145,147. Genetic risk for amyotrophic lateral sclerosis (ALS), a neurodegenerative disease of motor neurons²⁰⁸, is selectively associated with both transport and matrix pericytes (Figure 7B), consistent with prior reports implicating pericytes in disease pathobiology^71,104,209. Genetic vulnerabilities were also found in transport pericytes and arterial/arteriolar smooth muscle cells in frontotemporal dementia (FTD) (Figure 7B), a neurodegenerative disease with significant clinicopathological overlap with ALS²¹⁰. Taken together, these results highlight previously unknown mural cell associations in ALS-FTD that warrant future investigation. The vascular basis of headache disorders remains debated^205,206. We found that genetic risk for migraine, but not cluster headaches, was selectively associated with arterial endothelial cells, transport pericytes, and multiple smooth muscle cell populations, supporting a vascular contribution^211,212 (Figure 7B). Despite the range of diseases analyzed, mural cells harbored the highest genetic vulnerability (Figure 7C), highlighting the need to better delineate disease-selective contributions of distinct mural cell subsets in humans.

We next investigated whether genetic susceptibility to neurological diseases segregates with arteriovenous architecture. We grouped endothelial, mural, and perivascular cells from our integrated human cerebrovascular cell atlas into arteriovenous cell ensembles and applied MAGMA-based genetic enrichment analysis. This segment-centric framework localized genetic predisposition to diseases known to arise in arteries, such as small artery occlusion and deep perforating arteriopathy^170,172, and resolved previously obscured cell–trait associations (Figure 7D). For stroke, arteriosclerotic cSVD, and migraines, arterioles harbored the greatest genetic vulnerability, reflecting their high cellular complexity (Figure S3C, S3F) and role in regulating local cerebral blood flow²⁰. Conversely, ALS showed the greatest susceptibility in capillary beds, consistent with reports that microvascular BBB breakdown contributes to neurodegeneration^71,104,209 (Figure 7D). Finally, genetic risk for neuroimmune diseases, such as vasculitis and multiple sclerosis, was localized to venules, consistent with their roles in leukocyte chemotaxis and transmigration^213–215 (Figure 7D). Thus, genetic risk for neurological diseases is not generalized across the vasculature, but localizes to distinct arteriovenous segments and cell subsets, likely reflecting pathogenic effects on specific vascular cell ensembles. Different therapeutic strategies may therefore be required depending on the affected ensemble.

Given the zonal segregation of cerebrovascular and neuroimmune diseases, we asked whether therapeutic agents used to treat these conditions also show preferential targeting by arteriovenous identity. Although in vivo phage display screens support organ-specific drug delivery^216–218, zonation-specific targeting remains unexplored. Using the integrated human cerebrovascular cell atlas and Drug2Cell to predict cell-type druggability²¹⁹, we found that angiotensin-converting enzyme (ACE) inhibitors, among the most prescribed antihypertensive agents, preferentially target arterial endothelial cells, with little predicted activity in capillary or venous endothelium (Figure 7E). By contrast, interleukin 1 receptor and interleukin 6 receptor antagonists, which are approved for the treatment of autoimmune disorders, such as neuromyelitis optica spectrum disorder (NMOSD)²²⁰, preferentially act on venous endothelial cells, with minimal activity in arterial or capillary beds (Figure 7E). These findings suggest that our integrated cerebrovascular cell atlas may guide development of arteriovenous-selective pharmaceutical therapies.

Selective mural cell vulnerabilities in monogenic stroke syndromes

Stroke remains a formidable global health challenge, with mortality projected to increase 50% by 2050^161,162. Whole-exome and genome sequencing have identified monogenic alterations associated with stroke^221,222. To further define cerebrovascular cell-specific genetic susceptibility in stroke, we performed gene set enrichment analysis between the top cell subset markers curated from our integrated human cerebrovascular cell atlas and a recently curated monogenic stroke risk panel²²³ (Figure 7F). Gene sets for systemic conditions that increase the risk of stroke, e.g., cardioembolic and metabolic syndromes, were not enriched in vascular cells, supporting the specificity of the analysis (Figure 7F). Our analysis also recapitulated the endothelial cell specificity of causative genes for brain vascular malformations^224–227 (Figure 7F). Notably, gene panels associated with susceptibility to multiple stroke subsets, including non-atherosclerotic stroke, cSVD or its radiographic correlates (i.e., white matter hyperintensities and basal ganglia calcifications), lacunar stroke, and intracerebral hemorrhage, were significantly enriched in matrix pericytes, but not other mural cell populations (Figure 7F). Multiple stroke risk genes were also strongly expressed in matrix pericytes in the integrated cell and spatial atlases of the human temporal cortex (Figure S9C–D). These included COL4A1 and COL4A2, mutations in which cause Gould syndrome, a monogenic form of cSVD^228,229. These findings suggest that pathological extracellular matrix organization by matrix pericytes may contribute to multiple stroke subsets, but further investigation is needed to explore the underlying pathogenic mechanisms.

cSVD is a major cause of stroke and vascular dementia in older adults and is defined by subcortical microinfarcts and microbleeds that affect up to 35.7% of cSVD patients aged 80 and over^230,231. Current clinical strategies remain non-specific and show inconsistent effects on stroke risk and cognitive outcomes, likely reflecting an incomplete understanding of the cellular substrates underlying cSVD. To identify therapeutic strategies targeting matrix pericytes, we used Drug2Cell²¹⁹ to survey major drug classes currently considered for cSVD treatment^232,233. Matrix pericytes, and to a lesser extent some smooth muscle cell subsets, showed pronounced predicted reactivity to phosphodiesterase 3 (PDE3) inhibitors, whereas transport pericytes did not (Figure 7G, S9E). Cilostazol, a PDE3 inhibitor, has shown promising efficacy in preclinical and clinical trials of stroke prevention and BBB neuroprotection in cSVD^234–240, suggesting that some of its benefits may be mediated through matrix pericyte targeting. By contrast, drug classes with mixed clinical outcomes, such as lipid-lowering agents^232,233, or without disease-modifying effects, such as acetylcholinesterase inhibitors ^232,233, showed little matrix pericyte reactivity (Figure 7G), reinforcing the central role of matrix pericytes in cSVD treatment paradigms. Together, these findings support matrix pericytes as a therapeutic target in cSVD and demonstrate the utility of the integrated human cerebrovascular cell atlas as a resource to guide drug development or inform clinical practice when treating cerebrovascular disease.

Discussion

Here, we present integrated cellular and spatial atlases of the human cerebrovasculature in the temporal cortex and hippocampus. By combining vascular-enriched single-cell datasets with spatial transcriptomics, we define consensus vascular cell states across the arteriovenous axis and resolve their spatial distribution in situ^26–30. These analyses reveal that cerebrovascular cells assemble into spatially patterned microenvironments, termed vascular cell ensembles, composed of endothelial cells, mural cells, fibroblasts, and perivascular macrophages that are aligned with the arteriovenous architecture to coordinate segment-specific functions. Using this Resource, we identify cell type-specific genetic vulnerability to neurological diseases, nominate candidate therapeutic targets, and provide a framework to guide arteriovenous-selective therapeutic development. Data are available for future research at: https://brain-vasc-spatial.cells.ucsc.edu.

The relationship between pericyte identity and arteriovenous zonation remains unresolved^{27,31,86,93,94}. We show that transcriptionally defined matrix and transport pericytes correspond to mesh and thin-strand pericyte morphologies in humans. We did not detect comparable transcriptional specialization of pericytes in mice, suggesting that pericyte morphology may reflect microenvironmental or hemodynamic cues associated with distinct arteriovenous locations rather than intrinsic transcriptional programs. Mural cell identity at arteriole-capillary transition zones also remains debated^14,91. Our spatial analyses demonstrate that arteriolar smooth muscle cells and matrix pericytes co-exist within arterioles as distinct populations. Although additional functional or morphological transitions cannot be excluded, we propose that graded shifts in the relative abundance of mural cell subsets may contribute to physiological transitions across the arteriole-capillary interface in humans. However, future functional studies will be required to test this model.

Brain fibroblasts remain among the least characterized cerebrovascular cells, with growing recognition of their roles in response to brain injury^127,241. We identify five fibroblast subsets spanning the arachnoid mater to parenchymal perivascular spaces. Inner arachnoid fibroblasts localize to the reticular arachnoid layer and exhibit PROX1 immunoreactivity, although whether these correspond to the proposed subarachnoid lymphatic-like membrane (SLYM)^133–135 remains unresolved and warrants future study. Cell communication analyses identify perivascular fibroblasts as signaling hubs in large and intermediate vessels, and cross-species comparisons reveal conserved identities alongside human-specific transcriptional programs enriched for neuronal support pathways. Together, these findings suggest fibroblasts contribute to specialized vascular microenvironments beyond structural support.

We transcriptionally distinguish perivascular macrophages from microglia and map their distribution along the arteriovenous axis. Although present on veins and venules, we observed greatest enrichment around arteries and arterioles, consistent with a prior report¹⁵⁴. Genetic enrichment analyses identify these cells as the only vascular-associated population linked to AD risk, supporting experimental studies implicating them in amyloid-β-associated neurovascular dysfunction^145–147. Our findings establish perivascular macrophages as integral components of the vascular ensemble framework and motivate future studies to define their roles in neurovascular physiology and disease.

More broadly, our findings support a model in which the human cerebrovasculature is organized into segment-specific vascular cell ensembles whose coordinated transcriptional programs support specialized vascular functions. Despite inter-individual variations in arterial and venous anatomy^11,242, our analyses show that the microvasculature follows conserved spatial principles likely shaped by vascular integrity, neuronal homeostasis, and neuroimmune surveillance. Disruption of ensemble structure or communication may therefore drive segment-specific vulnerability to cerebrovascular and neurologic diseases. By linking vascular cell states, spatial architecture, and genetic risk, these atlases provide a blueprint for studying human cerebrovascular biology and for developing precision therapies directed at specific vascular cell ensembles.

Limitations of the Study

Vascular specializations extend beyond the vessel segments captured in the current atlas¹⁴, and future profiling of larger intra- and extracranial vessels will be required to evaluate disease associations arising from these vascular beds. Because the integrated human cell atlas combines independently generated datasets, residual technical and sampling differences cannot be fully excluded despite batch correction and covariate analyses. Current vascular isolation methods also exhibit inherent arteriovenous sampling biases, underscoring the need for standardized cerebrovascular tissue acquisition. Moreover, some cortical and hippocampal tissues used for spatial transcriptomics were obtained from neurosurgical epilepsy resections. Although rigorous selection criteria and comparisons with postmortem tissue suggest that the main spatial findings are reproducible, subtle molecular effects related to epilepsy or surgical resection cannot be excluded. Additionally, while vascular cell ensembles were conserved between the temporal cortex and hippocampus, further studies are required to determine how vascular cell states and spatial organization vary across other brain regions. It also remains unknown whether vascular cell ensembles differ according to age, sex, or ancestry, which will require larger, more diverse cohorts. Finally, because these analyses rely on RNA expression, future studies combining functional experiments and epigenomic profiling will be needed to test predicted mechanisms and cell communication networks and to refine the roles of specific cell states and arteriovenous segments in neurological disease pathogenesis and progression.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Ethan Winkler (ethan.winkler@ucsf.edu).

Materials availability

This study did not generate any new, unique reagents.

STAR★Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Human brain tissues

Fresh and postmortem human brain tissue specimens and clinical data were obtained from the University of California, San Francisco, using institutional review board and ethics committee-approved protocols. Fresh specimens were acquired from donors undergoing neurosurgical operations, and written informed consent was obtained prior to the procedure for collecting tissue specimens for research purposes. These temporal cortical and hippocampal specimens were obtained from neurosurgical resections for remote epileptic lesions and carefully curated to minimize potential pathology involvement. Specifically, specimens were >2 cm from any radiographic abnormality on magnetic resonance imaging, showed no electrophysiological abnormalities on routine electrocorticography, and were histologically normal on a rapid hematoxylin and eosin stain. Fresh tissues were acquired in close collaboration with neurosurgeons trained in tissue isolation techniques to minimize tissue disruption from tools such as electrocautery. Postmortem autopsy specimens were screened for the lack of neurological involvement in the cause of death and a postmortem interval <24 hours, and tissues were deemed histologically normal on a rapid hematoxylin and eosin stain by trained neuropathologists. All tissue specimens were flash-frozen in liquid nitrogen and maintained at −80°C for long-term storage or formalin-fixed and paraffin-embedded using standard practices. All de-identified fresh and postmortem donor demographic information is provided in Table S1.

METHOD DETAILS

Individual human sc/snRNA-seq dataset processing

Winkler²⁶, Wälchli²⁸, and Yang²⁷ datasets were first processed starting from the FASTQ files (CellRanger v7.1.0, Human Genome GRCh38 Assembly) as those files were publicly accessible at the time of data acquisition. The Winkler and Wälchli FASTQs were directly obtained from the lead authors. The Yang FASTQs were obtained by following the listed data availability link. Garcia²⁹ and Siletti³⁰ datasets were obtained as processed count matrices and metadata tables by following the listed data availability links. Each individual dataset was further filtered for quality using each respective dataset’s reported quality control thresholds. In addition to these dataset-specific thresholds, we also implemented CellBender²⁴⁴ (v0.2.2) for ambient RNA removal for datasets processed starting from FASTQs (i.e., Winkler, Wälchli, and Yang) and DoubletFinder²⁴⁵ (v2.0.4) for doublet detection and removal in all datasets. For CellBender, “expected-cells” and “total-droplets-included” parameters were estimated based on the 10X’s barcode rank plot for each individual sample. For DoubletFinder, we conservatively estimated a uniform 7.5% doublet rate as the number of loaded cells/nuclei per 10X lane was not provided by all datasets.

Individual mouse sc/snRNA-seq dataset processing

Mouse datasets^{31,37,56–59} were first processed and annotated individually from raw expression matrices provided by the original authors (see Key Resource Table for data availability links). Where possible, quality control and processing were performed in accordance with the methods reported in the original studies. Only cells from the adult mouse cerebral cortex and hippocampus were considered for analysis across all datasets where multiple regions were sampled and annotated. Due to the bias in representation towards parenchymal cells, parenchymal cells in the Langlieb⁵⁶ and Yao³⁷ datasets were consolidated into the 3000 metacells using the SEACells²⁶⁰ package (v.0.3.3). Cerebrovascular cells and associated perivascular macrophages were subsetted from each original dataset in full without metacell consolidation. DoubletFinder (v2.0.4) was implemented on all datasets to identify and remove doublets using a standard doublet rate of 7.5%.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
KCNMA1 rabbit anti-human polyclonal (1:200)	LSBio	LS-A9574; RRID:AB_1953080
STXBP6 goat anti-human polyclonal (1:100)	Invitrogen	PIPA518157; RRID:AB_10981131
SLC5A5 mouse anti-human monoclonal (1:100)	QED Bioscience	11563; RRID:N/A
NGFR mouse anti-human monoclonal (1:200)	Invitrogen	MA5-13314; RRID:AB_10982037
ABCA8 rabbit anti-human polyclonal (1:200)	Atlas Antibodies	HPA044914; RRID:AB_10964859
PROX1 rabbit anti-human monoclonal (1:100)	Cell Signaling	14963T; RRID:AB_2783562
DARC goat anti-human polyclonal (1:100)	Novus Biologicals	NB100-2421; RRID:AB_10001541
COL1A1 sheep anti-human polyclonal (1:200)	R&D Systems	AF6220; RRID:AB_10891543
Calponin mouse anti-human monoclonal (1:200)	Sigma-Aldrich	MABT1504; RRID: N/A
CD31 mouse anti-human monoclonal (1:200)	Agilent	M082301-2; RRID:AB_2114471
PDGFRB goat anti-human polyclonal (1:100)	R&D systems	AF385; RRID:AB_355339
PDGFRB rabbit anti-human monoclonal (1:100)	Cell Signaling Technology	3169S; RRID: AB_2162497
Alpha-SMA rabbit anti-human monoclonal (clone D4K9N) (1:100)	Cell Signaling Technology	19245; RRID:AB_2734735
Biotinylated Ulex Europaeus Agglutinin I (UEA I) (1:300)	Vector Laboratories	B-1065-2; RRID:AB_2336766
Isolectin GS-IB4 AlexaFluor 647 Conjugate (1:500)	ThermoFisher Scientific	I32450; RRID:SCR_014365
COLIV mouse anti-human monoclonal (1:100)	Sigma-Aldrich	C1926; RRID:AB_476828
COLIV goat anti-human polyclonal (1:100)	Sigma-Aldrich	AB769; RRID:AB_92262
CA4 goat anti-mouse polyclonal (1:100)	R&D Systems	AF2414; RRID:AB_2070332
CA4 rabbit anti-human polyclonal (1:100)	Millipore Sigma	HPA017258; RRID:AB_1845997
ADAMTS1 sheep anti-human/mouse polyclonal (1:100)	R&D Systems	AF5867; RRID:AB_2044595
Opal dye 690 (1:1000)	Akoya Biosciences	FP1497001KT
Biological samples
Human Middle Temporal Gyrus	UCSF Neurosurgery Biospecimen Repository	N/A
Human Hippocampus	UCSF Neurosurgery Biospecimen Repository	N/A
Chemicals, peptides, and recombinant proteins
Nuclear Fast Red	Abcam	ab246831
Hematoxylin Solution, Gill No. 2	Millipore Sigma	GHS232
Eosin Y Solution, Alcoholic	Millipore Sigma	HT110116
Bluing Buffer	Dako	CS70230-2
Critical commercial assays
10x Genomics Xenium consumables v1	10x Genomics	N/A
10x Genomics Xenium cell segmentation add-on	10x Genomics	N/A
10x Genomics Xenium standalone custom 101-300 gene panel	10x Genomics	N/A
RNAscope multiplex fluorescent reagent kit v2	ACD Bio-Techne	323100
Deposited data
Resource website	This paper	https://github.com/jwangbio/spatial-brain-vasc
Archival Xenium files	This paper	GEO: GSE335898
Interactive web-based UCSC Cell Browser	This paper and hosted by Speir et al.243	https://brain-vasc-spatial.cells.ucsc.edu
Winkler single-cell dataset (human)	Winkler et al.26	dbGaP: phs002624.v2.p1
Yang single-nucleus dataset (human)	Yang et al.27	GEO: GSE163577
Wälchli single-cell dataset (human)	Wälchli et al.28	GEO: GSE256493
Garcia single-nuclei dataset (human)	Garcia et al.29	GEO: GSE173731
Siletti single-nucleus dataset (human)	Siletti et al.30	https://cellxgene.cziscience.com/collections/283d65eb-dd53-496d-adb7-7570c7caa443
Vanlandewijck single-cell dataset (mouse)	Vanlandewijck et al.31	http://betsholtzlab.org/VascularSingleCells/database.html
Yao single-cell dataset (mouse)	Yao et al.37	GEO: GSE246717
Langlieb single-nuclei dataset (mouse)	Langlieb et al.56	RRID: SCR_016152
Pietilä single-cell dataset (mouse)	Pietilä et al.57	GEO: GSE227713, GSE228882, GSE233270
Ximerakis single-cell dataset (mouse)	Ximerakis et al.58	GEO: GSE129788
Schroeder single-nuclei dataset (mouse)	Schroeder et al.59	GEO: GSE309702
Experimental models: Organisms/strains
Mus musculus C57BL/6J wild type	Jackson Labs	Strain #:000664; RRID:IMSR_JAX:000664
Oligonucleotides
Hs-ADAMTS1 (targeting 1078-2026 region)	ACD Bio-Techne	524501
Hs-CA4 (targeting 2-1136 region)	ACD Bio-Techne	438561-C2
Hs-DKK2 (targeting 706-1581 region)	ACD Bio-Techne	531131
Hs-PROM1 (targeting 145-1454 region)	ACD Bio-Techne	311261
Hs-TSHZ2 (targeting 427-1435 region)	ACD Bio-Techne	890131
Software and algorithms
CellRanger	10x Genomics	https://github.com/10XGenomics/cellranger.git
CellBender	Fleming et al.244	https://github.com/broadinstitute/CellBender.git
DoubletFinder	McGinnis et al.245	https://github.com/chris-mcginnis-ucsf/DoubletFinder.git
Harmony	Korsunsky et al.246	https://github.com/immunogenomics/harmony.git
scANVI	Xu et al.40	https://github.com/scverse/scvi-tools.git
scVI	Lopez et al.39	https://github.com/scverse/scvi-tools.git
Spapros	Kuemmerle et al.247	https://github.com/theislab/spapros.git
Baysor	Petukhov et al.248	https://github.com/kharchenkolab/Baysor.git
Xeniumranger	10x Genomics	https://www.10xgenomics.com/support/software/xenium-ranger/latest
SCTransform	Butler et al.249	https://github.com/satijalab/sctransform.git
Seurat	Butler et al.249	https://github.com/satijalab/seurat.git
Gprofiler2	Kolberg et al.250	https://github.com/cran/gprofiler2.git
EnrichR	Kuleshov et al.251	https://cran.r-project.org/web/packages/enrichR/index.html
UCell	Andreatta et al.252	https://github.com/carmonalab/UCell.git
Homologene	NCBI	https://cran.r-project.org/web/packages/homologene/index.html
CellChat	Jin et al.117	https://github.com/sqjin/CellChat.git
MAGMA:CellTyping	Skene et al.253	https://github.com/neurogenomics/MAGMA_Celltyping.git
MungeSumStats	Murphy et al.254	https://github.com/Al-Murphy/MungeSumstats.git
Drug2Cell	Kanemaru et al.219	https://github.com/Teichlab/drug2cell.git
BANKSY	Singhal et al.255	https://github.com/prabhakarlab/Banksy_py.git
Slingshot	Street et al.256	https://github.com/kstreet13/slingshot.git
Monocle3	Cao et al.257	https://github.com/cole-trapnell-lab/monocle3.git
RANN	Arya et al.258	https://github.com/jefferislab/RANN.git
ImageJ	Schroeder et al.259	https://github.com/imagej/ImageJ.git
SEACells	Persad et al.260	https://github.com/dpeerlab/SEACells.git
GraphPad	Dotmatics	https://www.graphpad.com/

Integration of the human and mouse cerebrovasculature datasets

Following best practices for atlas-level single-cell integration⁴² and a tiered integration strategy inspired by JOINTLY⁴¹, we used the same three-tier integration framework to assemble the mouse and human atlases. In the first tier, we performed dataset-level integration using Harmony²⁴⁶ to first integrate across sample batches within the same dataset, with care placed on ensuring that across datasets only the same annotation vocabulary is used. Clusters with indeterminate identity were labeled as ‘cryptic’ across all datasets. In the second tier, we performed semi-supervised integration to merge all five human and six mouse datasets using scANVI framework⁴⁰ with the tier one annotation labels used as the labels_key, ‘author’ set to the batch_key, and the ‘cryptic’ category set to the unlabeled_category parameter. ‘Author’ was selected as the batch_key due to the strong transcriptional covariation explained by this metadata variable in principal component analyses. The underlying scVI model for scANVI was trained with 2 layers, 30 latent variables, and a negative binomial distribution. Leiden clustering was performed on the resulting consensus atlas and manually annotated based on the top 25 Wilcoxon-nominated markers. Low-quality clusters including the cryptic clusters are then dropped at this step from the final consensus atlas. Finally, in the third tier, we subset the consensus atlas into five daughter datasets: endothelial cells, smooth muscle cells, pericytes, fibroblasts, and myeloid cells, and re-embedded cells using the unsupervised scVI framework³⁹. Here, the batch_key was set to ‘sample’. This scVI model was also trained with 2 layers, 30 latent variables, and a negative binomial distribution. Following scVI embedding, iterative Leiden clustering was performed on each dataset. Annotations on ensuing subclusters were performed in a similar manner to the canonical-level annotations, principally based on the top 25 Wilcoxon-nominated markers. To assess integration quality, we computed the integration local inverse Simpson’s index (iLISI), as in prior benchmarking studies,^246,261 to confirm improved dataset and batch intermixing in the integrated dataset relative to the harmonization-naïve dataset.

Xenium gene panel curation

To curate genes for our 300-gene Xenium panels, we adopted a mixed approach leveraging both data-driven approaches such as the spapros package²⁴⁷ and manual literature curation. Candidate genes were assessed for compatibility using the 10X Genomics Xenium Panel Designer software (accessed February 27, 2024, for Xenium v1 chemistry kit) to safeguard against optical crowding prior to the finalization of the 300-gene panel. Included genes are found in Table S4.

Tissue placement and processing for spatial transcriptomics

Tissues were prepared for tissue placement on the Xenium slide according to the manufacturer’s instructions. For flash-frozen specimens, tissues were embedded in OCT, equilibrated for cryosectioning, and mounted onto the cryosection chuck at −20°C. Test sections were initially cut at 10μm, stained with Nuclear Fast Red (Abcam, ab246831), and assessed by light microscopy to confirm tissue integrity and the presence of vascular structures. Once suitable blocks were identified, sections were cut at the manufacturer-recommended thickness using an anti-roll plate and mounted onto Xenium slides while avoiding fiducial markers. When multiple adjacent sections were collected from the same specimen, sections were separated by approximately 10-30 μm depending on section quality. After section placement, all Xenium slides were placed in slide mailers, sealed with parafilm to minimize desiccation, and stored at −80°C for up to four weeks before processing on the 10X Genomics Xenium Analyzer. All Xenium Slides were processed using the same manufacturer-recommended workflow, including the Xenium Multi-Tissue Stain add-on protocol. Briefly, slides were incubated overnight with the resuspended custom gene expression probe buffer. After removal of unbound probes, ligation was performed to circularize hybridized probes, followed by rolling circle amplification to generate gene-specific barcodes. Cell boundaries and interiors were then labeled using the add-on cell segmentation antibodies with overnight incubation. Autofluorescence quenching and DAPI staining were subsequently performed according to the manufacturer’s protocol before imaging and transcript detection on the Xenium Analyzer. Xenium onboard instrument software (v2.0.1.0) was used for transcript decoding, cell segmentation, and quality control metric generation, and outputs were exported for downstream bioinformatic analysis.

Cell segmentation

Guided by best practices for analysis of imaging-based spatial transcriptomics platforms, such as 10X Genomics’ Xenium In Situ, cell segmentation was performed utilizing Baysor²⁴⁸ (v0.6.2). Xenium Onboard Analysis was first performed by leveraging the add-on cell segmentation kit to generate preliminary transcripts to cell assignments. The output transcript file was then filtered to remove negative control transcripts and low-quality transcripts (Phred-scaled quality score < 20) before passing into the Baysor algorithm. The -m flag or the minimum transcript detection threshold was selected depending on the distribution of transcripts per cell found for each slide, with the empirical values of 15 to 120 chosen such that lower-quality segmentation calls are excluded. The -n-cluster flag was set to 14 reflecting the number of major cell types expected from the integrated human cerebrovascular cell atlas. The --prior-segmentation-confidence flag was tested across 0.25, 0.5, and 0.75, with the final value of 0.25 selected due to its ability to best call segmentations based on manual inspection. The --save-polygons flag was enabled to save geoJSON files from each slide and imported into Xeniumranger (v2.0.1.2) software for visualization of the Baysor cell segmentation calls.

Quality control of spatial transcriptomics dataset

To ensure robust interpretation of the cell segmentation calls following Baysor, a stringent 7-point QC was followed, modeled on the implementation of MERFISH²⁶². The points are as follows: (1) Baysor-assigned average transcript assignment confidence > 80%, (2) nCount thresholds empirically tuned per slide with the lower bound at 10-20% and the upper bound at 99%, (3) DoubletFinder set to 10% rate, (4) counts normalized by volume area in order to account for truncated cells then further normalized using SCTransform with ‘Section_ID’ set as the regression variable, (5) volume selection relaxed to a lower bound of 20 μm³ in order to capture smaller vascular and mural cells and an upper bound set to the 99th percentile, or 452 μm³, (6) unassigned codewords percentage in any given cell must not exceed 5%, and (7) any placement artifacts such as curling in the tissue are grossly excluded in silico.

Tiered hierarchical annotation of cellular subsets

Consistent with prior atlases of the brain^37,69,262, we used a tiered hierarchical strategy to annotate cellular subsets. At the global tier, cell clusters were assigned canonical identities using Seurat (v5.1.0) ‘FindAllMarkers’ with the Wilcoxon rank-sum test. This initial annotation resolved the major vascular and vascular-associated cell classes (endothelial cells, pericytes, smooth muscle cells, fibroblasts, and perivascular macrophages), parenchymal cell populations (neurons, astrocytes, oligodendrocytes, and oligodendrocyte progenitor cells), and broad immune cell classes (lymphoid and myeloid cells). To further resolve subtype-level heterogeneity within the cerebrovasculature, each major vascular and vascular-associated population was then subsetted, re-embedded, re-clustered, and re-analyzed with ‘FindAllMarkers’. This second tier enabled finer discrimination of subtypes within each major cell class after canonical identity had been established, using subtype-enriched markers that distinguish closely related populations.

Differential gene expression analysis

Differential gene expression between vascular cells, segmental cell communities, and brain regions was performed using Seurat’s (v5.1.0) ‘FindMarkers’ function as previously reported^27,85. For the purposes of differential gene expression analysis, segmental communities were defined as the unique vascular cell types that were found between the groups of comparison. Differentially expressed genes (DEGs) were filtered to exclude mitochondrial and ribosomal genes and to retain only genes with an adjusted p-value < 0.01. Average log₂-transformed fold change thresholds ranging from 1 to 40 were utilized and the top 55 to 90 genes ranked by either the fold change or adjusted p-value were drawn for input into gprofiler2^250,263 using only the ‘Biological Process’ Gene Ontology (GO) source database in order to best maximize interpretability of the results.

Gene signature scoring

Gene signatures are calculated using the UCell²⁵² (v2.8.0) package using the ‘AddModuleScore_UCell’ function. For the calculation of the fibromyocyte (FBMC) signatures and venous smooth muscle cell (vSMC) signatures, gene markers were obtained from both the Winkler et al.²⁶ and Muhl et al.³⁶ papers. For the calculation of contractility, migration, extracellular matrix organization, and synaptic signaling signatures, the GO terms for each gene signature as specified in Table S6 were used. Gene lists were then extracted from each set of GO terms and then intersected with the top 100 significant DEGs ranked by False Discovery Rate for both comparison groups (i.e. arterial/arteriolar SMC vs venous SMC or arterial SMC vs arteriolar SMC). The gene intersection was then taken as the final feature list for input into UCell.

Human and mouse species comparisons

To examine molecular patterns of conservation and divergence between human and mouse cerebrovascular cells, conserved vascular subsets were obtained from both human and mouse datasets. Log-transformed and normalized counts were used, and mouse genes were converted to their orthologous equivalents in human using the “homologene” R package (v1.4.68.19.3.27). Pearson correlations were calculated for each of the vascular subset comparisons. To classify a gene as species-specific, we extended a previously established criterion²⁶⁴ such that the gene had to (1) rank in the top 15th percentile of gene expression for the species and (2) be expressed at a > 2-fold higher level than in the other species. To classify a gene as species-shared, it had to (1) rank in the top 15th percentile of gene expression for both species and (2) fall within a 1.25-fold window around the equivalence line (i.e., equal expression between human and mouse). Species-specific genes are catalogued in Table S5. GO was completed using the “enrichR” R package (v3.2) on the species-specific and shared gene sets as defined above, and only five per gene set were shown.

Cell-cell communication analysis

Cell-cell communication analysis was performed using CellChat¹¹⁷ (v2.1.2) using default tutorial settings and with the inclusion of all cell subsets. In instances where the CellChat analysis focused on specific arteriovenous segments, the CellChat object was subset to the endothelial, mural, and perivascular cell types that were spatially mapped to the queried arteriovenous segment. Rare immune subsets, such as monocytes and neutrophils, were combined into a single category, “circulating myeloids”. Neurons were further sub-annotated as excitatory or inhibitory for the purposes of the CellChat analysis. For examination of neurovascular signaling across neuronal subtypes and microvascular ensembles, we extracted and summed the cell-cell communication probabilities across all source and target cell group pairs to obtain a single aggregate information flow value per pathway per neurovascular comparison group (i.e., Excitatory by Arteriole, Excitatory by Capillary, Excitatory by Venule, Inhibitory by Arteriole, Inhibitory by Capillary, Inhibitory by Venule). The top 100 pathways ranked by the total information flow across comparisons were then retained, log1p transformed, and standardized row-wise by z-score normalization before being visualized in a hierarchical heatmap.

MAGMA enrichment

We employed MAGMA:CellTyping^185,253 (v2.0.8) to integrate genome-wide association study (GWAS) summary statistics data with our cell atlas for 18 polygenic traits spanning cerebrovascular disease, neuroimmune disease, neurodegenerative disease, headaches, and miscellaneous categories (Table S7). GWAS summary statistics were formatted using MungeSumStats (v1.15.1)²⁵⁴; MAGMA analyses were then performed using default settings. Linkage disequilibrium (LD) was accounted for using the 1000 Genomes Phase 3 European reference²⁶⁵ to match the ancestry of the GWAS summary statistics that were tested. Analyses were limited to European-ancestry GWAS because robust LD reference panels were most readily available for these studies, and the human vascular cell atlas is also composed predominantly of donors of European ancestry. For each trait, common risk variants reaching genome-wide significance were identified and mapped to their respective genes using the default MAGMA parameters. Adjusted F statistics were then computed using a multiple linear principal components regression model to test whether gene-level association statistics are enriched among genes with greater expression specificity in each cell subset. The gene-specificity score was defined continuously as the mean expression of a gene in one cell subset divided by its mean expression across the full dataset, which is also known as a “Cell Type Dataset” (CTD) by MAGMA. Our CTD was derived using the expression matrix of our single-cell Seurat object (genes as rows and cells as columns) and cell annotations listed in the metadata. Significance values were then stringently Bonferroni-corrected row-wise for each phenotypic trait to account for multiple cell type testing. For the vascular cell ensemble analysis, vascular cells were subset from the cell atlas with respect to their arteriovenous segmental identity. Cell types that were found over multiple segments were compiled multiple times, once for each vascular cell ensemble they localize to. Dural border fibroblasts, arachnoid barrier fibroblasts, inner arachnoid fibroblasts, parenchymal cells, and circulating immune cells were excluded from this analysis as they are not exclusively enriched perivascularly. We then performed MAGMA using these segmental annotations, using the same methodology described above to identify genetic risk associations between each polygenic trait and arteriovenous segment.

Monogenic stroke enrichment

Stroke Gene Panel 1 was obtained from the supplementary materials of Ilinca et al. 2023²²³, totaling 168 curated monogenic causes of stroke. Cluster markers were calculated for the global cell atlas and stringently filtered for genes with an average log₂ fold change > 2 and an adjusted p-value less than 0.01. For each cell cluster, a logistic regression model in which each gene was annotated by stroke-gene status and cluster-marker status was fitted as previously described^266–268, testing whether marker genes for that cluster were more likely to belong to a given monogenic stroke category. The background gene set for this analysis comprised all unique genes that passed the marker-gene filtering criteria in the cell atlas. False discovery rate correction was applied to coefficient p-values calculated across clusters within each monogenic stroke category. To assess the robustness of the enrichment analysis to threshold selection, we progressively increased the thresholds. This revealed increased non-specific enrichments, particularly for systemic monogenic stroke etiologies such as cardioembolic or metabolic stroke, which are not expected to be enriched in cerebrovascular cells. We therefore retained stringent thresholds, as defined above, to minimize these non-specific enrichments.

Drug2Cell

To investigate the relationship between arteriovenous localization and drug specificity and to identify drugs used in cSVD that target matrix pericytes, we utilized Drug2Cell (v0.1.2)²¹⁹ on default settings. Putative drugs of interest and their gene targets were extracted from the CHEMBL database²⁶⁹. Clinically approved drugs with the Anatomical Therapeutic Chemical (ATC) classifications of ACE inhibitors and immunomodulatory biologics were used. Drug-to-cell scores were derived based on their target gene expression, generating a drug-to-cell matrix. Composite drug class scores were calculated as the mean of individual drug reactivity scores comprising the drug class.

Spatial cluster annotation

BANKSY²⁵⁵ was applied to two settings in order to accomplish cluster annotations on the Xenium dataset. In the first setting, a lower coherence value of λ=0.2 was set to perform spatially weighted cell-typing at both the canonical and subclustering levels. In the second setting, a higher coherence value of λ=0.95 was set to perform cortical depth and hippocampal transverse ordering of cells. Due to the high level of variability in cortical geometries across sections, we sought to leverage a data-driven approach to assign cortical depth measurements and hippocampal subfield location to each cell. At high coherence levels, BANKSY strongly skews the cell embeddings based on spatial location over gene expression profiles, favoring an embedding that sorts cells along the cortical depth and hippocampal transverse length in practice. In all cases, BANKSY clustering was performed on SCT-transformed values using ‘Section_ID’ as the grouping variable and k_geom set to 10. Principal component analysis (PCA), Uniform Manifold Approximation Projections (UMAPs), and multilevel refinement Louvain clustering were performed on the top 30 PCs. To find enriched marker genes per cluster, Seurat’s ‘FindAllMarkers’ function was employed using the Wilcoxon rank sum test.

Spatial trajectory analysis

Trajectory analysis was applied on BANKSY embeddings for two scenarios: (1) leveraging low-coherence BANKSY embeddings to perform mapping of endothelial cells along the arteriovenous axis and (2) ordering of all cell types at high-coherence BANKSY embeddings to perform cortical depth and transverse hippocampal mapping. For arteriovenous mapping, a trajectory was calculated using Slingshot²⁵⁶ with the start clusters set as arteries and the end cluster set as veins in both cortex and hippocampus. For the cortical depth mapping, we removed admixed clusters (i.e. cells that were annotated to a canonical cell type but harbored ambient number of transcripts from a neighboring canonical cell type), non-specific clusters (i.e., cells of indeterminate identity), and clusters containing less than 100 cells at a multilevel refinement Louvain clustering resolution of 0.25. We found that this QC step substantially improved the ability to perform trajectory mapping. Next, we isolated clusters corresponding to the gray matter and overlying sulcus, subarachnoid space, and arachnoid mater, and performed trajectory analysis using Monocle3²⁵⁷. Cells were clustered with a partition qval threshold of 0.05 and random seed of 67. Graph learning was performed with a minimum branch length of 5. We found that the most faithful cortical depth mappings came from tethering the two poles of the trajectory with markers expected at either end of the cortical span during the order_cells step. In this case, we set the root cells to those with the greatest expression of CD22 (oligodendrocyte-enriched gene, most expressed in the white matter) and COL1A1 (fibroblast marker, most expressed in the leptomeninges). Following trajectory analysis of the cortical depth, we segmented the cortical depth into distinct regions to correspond with known cortical laminar demarcations. This was done by evaluating the distribution of cells expressing known laminar and regional markers and then defining trajectory cut points that cleanly demarcate each lamina and regions from each other. Because laminar marker expression alone is insufficient for unambiguously defining excitatory neuron identity by laminae, we also leveraged our trajectory-informed laminar demarcations to refine excitatory neuron annotations in a layer-specific manner. Additionally, domains with low representation (< 5% of segmented cells) for any given section were omitted from further analysis. As with cortical depth mapping, we applied BANKSY to high-coherence data to embed cells of the hippocampal pyramidal cell layer and sort cells from the subiculum to the cornu ammonis 4 (CA4) subfields. We found in our testing that Slingshot best recapitulated the transverse hippocampal axis when the start cluster is set as CA4 and the end cluster as the subiculum and adopted this approach for the hippocampus. As with cortical depth, we rigorously benchmarked our subfield demarcations against the distribution of cells expressing known subfield markers and then defined trajectory cut points that cleanly demarcate each subfield from the others. In both cortical and hippocampal contexts, inter-donor reproducibility was further assessed by plotting the scaled expression of lamina/subfield-specific genes in excitatory neurons from each donor’s annotated laminae/subfields.

Neighborhood analysis

A 2-dimensional ranged nearest neighbors search was performed using the ‘nn2’ function in RANN (v2.6.1) package. Except for neighborhood complexity analyses, for which k = 100 was used, we used k = 10 as an upper bound on the number of neighbors within the search radius to prevent excessive compute runtimes, as over 99% of cells had 10 or fewer neighboring cells. To assess for the enrichment of any given cell type within the search radius of a query cell type, a hypergeometric test was performed using the ‘phyper’ function in the stats (v4.4.1) package with lower.tail set to false. Search radii relative to endothelial cells are adjusted to 10μm for mural cells and 20μm for perivascular fibroblasts and macrophages, owing to their differing radial position within the neurovascular unit. For computation of the number of perivascular macrophages between the arterial and venous ends of the arteriovenous axis, a search radius of 100μm was used to search for the identity of the nearest non-capillary endothelial cell (i.e., arterial, arteriolar, venular, or venous), which was taken to identify the arteriovenous identity of the perivascular macrophage. For the calculation of vascular ratios relative to the endothelial cell across cortical domains and hippocampal subfields, a search radius of 100μm was used. Neighborhood complexities were calculated as previously described^67–69 by quantifying the number of distinct neighboring cell states found within a fixed search radius of the origin cell type.

Cross-atlas cluster matching analysis

We first pseudobulked single-cell transcriptomes into cluster-level transcriptomes for both the query and reference datasets. We then extracted the expression profiles for markers that are expressed in both the query and reference clusters, and manually and/or bioinformatically prioritized highly variable gene markers that exhibited a strong ability to resolve cluster-based differences in the reference dataset. If prioritized bioinformatically, genes were prioritized by ranking them in descending order of the adjusted p-value significance following marker calculation using the Wilcoxon rank sum test found in Seurat’s ‘FindAllMarkers’ function. Pearson’s correlation values were then calculated for each query-reference cluster pair, and the best matching cluster was made based on the highest correlation value found for each query cluster.

Bimodality Testing of Mural Cells at the Arteriole-Capillary Transition

Following prior work aimed at distinguishing continuous from discrete cellular states¹¹⁰, we evaluated the distribution of mural cell states along a transcriptomic variational trajectory. First, we used Slingshot to infer a trajectory from UMAP embeddings of subsetted mural cells, with the starting cluster defined as arterial smooth muscle cells and the ending cluster as transport pericytes. After fitting the Slingshot curve, trajectory scores for each mural cell were extracted and passed into Hartigans’ Dip Test for Unimodality. Under this framework, the null hypothesis is that the mural cell distribution is unimodal; therefore, a non-significant p-value indicates that the data are consistent with a unimodal, continuous distribution, whereas a significant p-value suggests deviation from unimodality and is consistent with multimodal or more discretized mural transcriptomic structure.

Immunofluorescent staining

Formalin-fixed paraffin-embedded (FFPE) blocks were sectioned into 5-μm-thick sections on an automated microtome (Leica RM2255) and placed onto a glass slide (Fisher Scientific Superfrost Microscope Slides). Following overnight drying, slides were de-paraffinized with 3-minute incubation in the following order: 2x Xylene (Fisher Scientific HistoPrep), 2x 100% Ethanol (EtOH), 95% EtOH, 90% EtOH, 70% EtOH, 50% EtOH. Slides were then incubated with 1X low pH Antigen Retrieval Solution (Invitrogen eBioscience) overnight at 58° C and washed thoroughly afterward. Flash-frozen OCT-embedded non-pathologic human cortical and hippocampal blocks were mounted within a cryostat (Microm HM 525) set to an operating temperature of −20°C and sectioned at 10–14 μm and mounted onto a glass slide (Fisher Scientific Superfrost Microscope Slides). Prepared slides were then stored at −80°C until use. Slides of either preparation were then retrieved and then blocked using 10% Donkey Serum (Sigma Aldrich Biosciences) for 1 hour. Primary antibodies (listed in the Key Resource Table) were prepared in 2% Donkey Serum (Sigma Aldrich Biosciences) and then added to the sections and left in a dark humidified chamber overnight. Primary antibodies were washed off using 1X Tris Buffered Saline with Tween 20 (TBST, Cell Signaling Technology) for a total of 3x washes at 10 minutes each. Alexa Fluor-conjugated secondary antibodies (1:100) were incubated for 1 hour, and DAPI (1:500, EN62248, Invitrogen) was then added for a 10-minute incubation, with 3x 1X TBST washes being conducted following both incubations. To quench autofluorescence, sections were incubated with 0.1% Sudan Black (Fisher Scientific BioReagents) for 30 minutes and then washed >3x with 1X PBS (Fisher Scientific). Slides were then cover slipped using mounting media (H-5700, Vector Laboratories) and allowed to fully set prior to confocal imaging. Slides were imaged using a Nikon Eclipse Ti2-E inverted microscope, and post-processing of images was done using ImageJ.

RNAscope staining

RNAscope in situ hybridization was performed on fresh-frozen sections of adult human brain tissue using the RNAscope Multiplex Fluorescent Reagent Kit v2 (ACD Bio-Techne), according to the manufacturer’s instructions. Target detection was carried out for ADAMTS1, CA4, DKK2, PROM1, and TSHZ2 using probes from ACD Bio-Techne (see Key Resource Table). Opal fluorophores (Akoya Biosciences) were used for signal detection. Following RNAscope, immunohistochemistry was performed as described above, omitting the target retrieval step and using primary antibodies at twice the manufacturer’s listed concentration. Slides were imaged using a Nikon Eclipse Ti2-E inverted microscope, and post-processing of images was done using ImageJ.

Adult mouse brain harvest and fixation

Adult mice were anesthetized with isoflurane and transcardially perfused with 1X PBS (Fisher Scientific) followed by 4% paraformaldehyde (prepared in PBS). Brains were post-fixed in 4% PFA at 4 °C for 24 hours, rinsed in 1X PBS at 4 °C for another 24 hours. Mouse brains were then incubated in 30% sucrose prepared in PBS until they had fully settled. Excess sucrose was removed with Kimwipes before being embedded in 50:50 (v/v) 30% sucrose/OCT solution at −80° C for hardening. Mouse brains were then sectioned into 10μm thick sections in a conventional cryostat as a tiled array for downstream staining experiments.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical analysis of Xenium spatial distributions

For spatial analyses of cellular distributions, we fit linear mixed-effects models with cell-type abundances or ratios as the outcome, anatomical subregion as a fixed effect and donor as a random intercept to account for the non-independence of multiple sections obtained from the same donor, consistent with repeated-measures RNA-seq analysis frameworks that model donor as a random effect to avoid pseudoreplication^270,271. When multiple pairwise comparisons were performed, contrasts among estimated marginal means were tested with Tukey adjustment. Multiple-testing correction with the Benjamini-Hochberg technique was applied where applicable. Enrichment analyses, including tests for enrichment of cell populations across cortical laminae, hippocampal subfields, or arteriovenous segments, were performed using hypergeometric tests across all segmented cells passing quality control.

For manual quantitative analysis of Xenium images for endothelial arteriovenous zonation markers, morphology images were opened using the Xenium Explorer software (v4.1.1) at 0.5x zoom. At least 5 randomly selected image fields were obtained per section, and no tissues were excluded. A cell was considered positive if 3 or more RNA spots for the queried gene were within the cell boundary stain. Cell counting was performed by a blinded investigator. Statistical analyses were performed in R (v4.4.1, “Race for Your Life”) using a two-sided Student’s t-test. All data plots show individual data points.

Quantitative immunofluorescence analyses

For all quantitative immunostaining imaging experiments, at least 5 randomly selected image fields were acquired from 3 non-adjacent tissue sections per individual separated by at least 100 μm. For each field, a 10- to 12-μm z-stack was acquired. Maximum projection images were generated with NIH ImageJ software. No tissues were excluded from any analysis. At least 3 donors per region were used for each experiment. Quantification was performed by a blinded investigator.

For quantification of endothelial venous enrichment in cortex (gray versus white matter) and hippocampus (gray versus white matter), DARC (encoded by ACKR1, venous endothelial marker) and CD31 (pan-endothelial cell marker) signals were thresholded separately. Signal area was measured using the NIH ImageJ Area measurement tool. Venous endothelial enrichment was calculated as DARC-positive area divided by total CD31-positive endothelial area. Statistical analyses were performed in GraphPad Prism (v11.0.0, Dotmatics, Boston, MA) using two-sided Student’s t-test. All plots show individual data points.

For quantification of transport and matrix pericytes in cortex and hippocampus, the number of CA4⁺ PDGFRβ⁺ cells (transport pericytes) and ADAMTS1⁺ PDGFRβ⁺ cells (matrix pericytes) were counted using the NIH ImageJ multipoint tool. Pericyte number was then expressed as the number of cells per mm² of lectin-positive endothelial surface area as previously described^103,272. The same approach was used to quantify transport pericytes between cortical gray and white matter. Statistical analyses were performed in GraphPad Prism (Dotmatics, Boston, MA) using two-sided Student’s t-test. All plots show individual data points.

Supplementary Material

Table S1

Table S1: Donor demographics of cell and spatial atlas samples, related to Figures 1–2.

Table S3

Table S3: Cell census from cell and spatial atlases, related to Figures 1–2.

Table S4

Table S4: Xenium spatial transcriptomics targeted gene panel and robust arteriovenous markers, related to Figure 2.

Table S5

Table S5: Comparison of human and mouse gene expression across vascular and vascular-associated cell types, related to Figures 5, 6, S4N, S7F, S7J, S8L, S8S.

Table S6

Table S6: Gene ontology terms used in smooth muscle cell UCell score, related to Figures 5E–F.

Table S7

Table S7: GWAS summary statistics used for MAGMA analysis and expected cell associations, related to Figures 7A–D.

Table S2

Table S2: Cell type marker genes, related to Figure 1.

Data and code availability

• Archival Xenium files are deposited on the NCBI Gene Expression Omnibus (GEO: GSE335898) and are publicly available as of the date of publication. No novel scRNA-seq data were generated in this study. Cell and spatial atlases are available to explore via the interactive web-based UCSC cell browser (https://brain-vasc-spatial.cells.ucsc.edu) and are publicly available as of the date of publication.

• Microscopy data reported in this paper will be shared by the lead contact upon request.

• All original code has been deposited at GitHub at https://github.com/jwangbio/spatial-brain-vasc and is publicly available as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.