A pericyte-to-myofibroblast transition links APOE4 to cerebrovascular degeneration

Summary

Cerebrovascular disease is a major but poorly understood feature of Alzheimer’s disease (AD). The strongest genetic AD risk factor, APOE4, is associated with cerebrovascular degeneration, including vascular amyloid deposition and fibrosis. To uncover how APOE4 promotes cerebrovascular pathology, we assembled a single-cell transcriptomic atlas of human brain vasculature. In APOE4 carriers, pericyte abundance was significantly reduced and accompanied by the emergence of a myofibroblast-like cell population co-expressing contraction and extracellular matrix genes. Immunostaining confirmed non-vascular myofibroblasts in APOE4 human and mouse brains. We show that APOE4 pericytes transition into myofibroblasts that secrete fibronectin, which promotes vascular amyloid accumulation. Computational and experimental analyses identified elevated TGF-β signaling as the driver of this pericyte-to-myofibroblast transition. Inhibition of TGF-β restored pericyte coverage and reduced vascular fibrosis and amyloid to APOE3 levels, revealing a targetable mechanism linking APOE4 to cerebrovascular pathology in AD.

In Brief (“Aim to communicate the interest and excitement of your work to a general audience in a sentence or two”):

APOE4 drives a pericyte-to-myofibroblast transition that promotes cerebrovascular fibrosis and perivascular amyloid accumulation. TGF-β inhibition reverses this pathological remodeling, identifying a therapeutic strategy to restore cerebrovascular homeostasis in these high-risk individuals.

Graphical Abstract

Introduction

Alzheimer’s disease (AD) is a devastating neurodegenerative disorder characterized by progressive cognitive decline and brain atrophy. Most cases are late-onset and arise through interactions between aging and genetics¹. APOE4, the strongest genetic risk factor for late-onset AD, profoundly affects the cerebrovasculature ^2–6. Perivascular amyloid accumulation, blood-brain barrier (BBB) breakdown, and microvascular degeneration emerge early in AD and increasingly appear to contribute causally to neuronal dysfunction and degeneration^7–9. Even in the absence of AD pathology, APOE4 carriers show increased cerebrovascular dysfunction, suggesting that the vascular effects of APOE4 may be an upstream driver of AD pathogenesis^10–14. Although anti-amyloid monoclonal antibodies represent a major therapeutic advance, their use in APOE4 carriers is complicated by increased risk of amyloid-related imaging abnormalities (ARIA), including cerebral hemorrhage and edema^15,16. This heightened vulnerability underscores the need for therapies that prevent or reverse APOE4-associated cerebrovascular dysfunction.

The development of such therapies is hindered by the absence of a pathogenic mechanism linking APOE4 to cerebrovascular pathology. In endothelial cells, APOE4 promotes inflammatory and prothrombotic signaling while impairing tight junctions and compromising BBB integrity^17,18. Pericytes are key contributors to cerebral amyloid angiopathy (CAA) and BBB breakdown, with APOE4 expression in pericytes directly driving these pathologies^10,19–25. Beyond these canonical vascular cell types, single-cell studies have revealed an unexpected diversity of fibroblasts within brain barrier structures, suggesting emerging roles in cerebrovascular homeostasis, even in AD^26–31. Supporting this, recent studies have identified protective genetic variants in vascular basement membrane components, such as fibronectin, that reduce AD risk in APOE4 carriers^32,33. Moreover, the profibrotic cytokine TGF-β has been linked to vascular degeneration, further implicating dysregulation of the extracellular matrix in APOE4-associated cerebrovascular pathology^34–36. Despite these findings across cell types and compartments of the APOE4 cerebrovasculature, the precise mechanisms through which APOE4 expression in vascular cells leads to cerebrovascular dysfunction and AD pathogenesis remain poorly defined.

To address this gap, we assembled a high-resolution transcriptomics atlas of the human cerebrovasculature by integrating previously published single-nuclei RNA-sequencing (snRNA-seq) datasets from APOE4 carriers and non-carriers. We complement this computational approach with postmortem immunostainings, isogenic iPSC models of the human brain, and humanized APOE knock-in mice. Through this comprehensive three-pronged approach (human postmortem, human isogenic iPSC models, and APOE-TR mice), we discover that APOE4 promotes a pericyte-to-myofibroblast transition in the brain. We establish that elevated TGF-β signaling in the APOE4 brain drives this transition, resulting in a loss of pericyte coverage and excessive fibronectin deposition along the vasculature. We show that fibronectin secretion from APOE4 myofibroblasts directly promotes perivascular amyloid accumulation and correlates with an earlier age of AD diagnosis and more severe cerebral pathology. Importantly, we demonstrate that TGF-β inhibition in APOE4 iPSC-derived human brain tissue and APOE4 mice reverses the pericyte-to-myofibroblast transition and reduces cerebrovascular fibrosis and amyloid accumulation, thus providing a direct mechanistic link between APOE4, TGF-β, vascular fibrosis, and AD pathogenesis.

Results

Harmonized single-nucleus atlas of the human APOE4 cerebrovasculature

To assemble a single-nuclei transcriptomics atlas of the human cerebrovasculature from APOE3/3 and APOE4 carriers, we integrated three previously published snRNA-seq datasets of the postmortem human brain^37–39 (Figure 1A). After within-dataset integration and vascular cell annotation, propensity score matching was performed independently within each dataset to generate age-, sex-, AD status-, and vascular cell yield-matched APOE3/3 and APOE4 cohorts (Figure S1A–C). Matched vascular cells were then integrated into a unified cerebrovascular atlas (Figure 1B, S1D–G), comprising approximately 66,000 nuclei from 220 individuals. Integration quality was confirmed by pseudobulk correlation analysis and variance partitioning (Figure S1E–G). Canonical marker genes identified astrocytes, endothelial cells, pericytes, and smooth muscle cells (Figure 1B, S1D; Table S1).

Figure 1. A myofibroblast-like mural cell population is enriched in the APOE4 brain.

A. Schematic of cerebrovascular atlas generation from single-nucleus RNA sequencing of post-mortem human brains from APOE4 carriers and non-carriers.

B. UMAP of harmonized blood-brain barrier cell types.

C. Differentially abundant (DA) cell neighborhoods by cell type in APOE4 carriers versus non-carriers identified using miloR (spatial FDR < 0.05; positive logFC indicates enrichment in APOE4 carriers). Data points represent individual cell neighborhoods. Solid lines indicate mean logFC. One-sample Wilcoxon signed-rank tests with Benjamini-Hochberg correction.

D. Representative images of post-mortem human hippocampus stained for NG2, αSMA, and VE-CAD. Scale bar, 50 μm. Arrows indicate vascular pericytes (top) and nonvascular αSMA+ cells (bottom). Pericyte coverage was quantified as the NG2-VE-CAD overlap normalized to VE-CAD. αSMA-positive area was normalized to nuclei. Both were expressed relative to APOE3/3. Data represent individuals (NG2: n = 13 APOE3/3, n = 7 APOE4; αSMA: n = 11 APOE3/3, n = 9 APOE4). Bars indicate mean ± SEM. Unpaired two-tailed Student’s t-test.

E. UMAP of mural cell subclusters.

F. Top upregulated and downregulated REACTOME pathways in APOE4 SMC_2 cells relative to APOE3/3 identified by GSEA (FDR < 0.05). Redundant pathways were merged for visualization.

G. Top CellMarker annotations of APOE4 SMC_2 cells based on pseudobulk differential expression (Figure S2D). Redundant annotations were merged for visualization.

H. DAseq differential abundance analysis of mural cell subclusters. Red and blue indicate enrichment and depletion in APOE4 carriers, respectively. The APOE4-enriched SMC_2 region was annotated as myofibroblasts based on ECM and contraction signatures (Figure S2H–J).

I. Model-estimated subject-level myofibroblast abundance from the covariate-adjusted binomial generalized linear model (Figure S2M). Data represent individuals (n = 53 APOE3/3, n = 50 APOE4). Values were normalized to APOE3/3. Bars indicate model-estimated marginal means ± SEM. P-values represent the APOE genotype effect.

J. Representative images of corpus callosum microvessels (<6 μm) from age- and sex-matched APOE3/3 and APOE4/4-TR mice stained for lectin, αSma, and Ng2. Arrows indicate nonvascular αSma+ myofibroblasts. Scale bar, 10 μm. αSma-positive area was normalized to nuclei and pericyte coverage was quantified as Ng2-lectin overlap, both expressed relative to APOE3/3. Data represent individual mice (6 images/mouse; n = 3 mice/genotype). Bars indicate mean ± SEM. Unpaired two-tailed Student’s t-test.

K. Schematic of the miBrain model.

L. UMAP of miBrain mural cells.

M. Mural cell cluster proportions by genotype. Bars indicate mean ± SEM (n =2 miBrains/genotype). Welch’s unpaired Student’s t-test.

N. Representative marker genes of mural cluster 2.

O. Top enriched GO pathways in mural cluster 2.

P. Pericyte and myofibroblast module scores across mural clusters.

Q. Representative images of isogenic APOE3/3 and APOE4/4 miBrains stained for CD144, αSMA, and NG2. Scale bar, 75 μm. Three-dimensional pericyte coverage was quantified as NG2-CD144 overlap normalized to total CD144 volume. αSMA-positive area was normalized to nuclei. Values were expressed relative to APOE3/3. Data represent individual miBrains (6 images/miBrain; n = 6 miBrains/genotype). Bars indicate mean ± SEM. Unpaired two-tailed Student’s t-test.

See also Figures S1, S2, and S3.

Pericyte loss in APOE4 carriers coincides with the emergence of atypical non-vascular SMCs

Previous transcriptomic studies of the APOE4 vasculature had limited statistical power to detect changes in cell-type abundance⁴⁰. By harmonizing multiple datasets into a single, integrated atlas, overcame these limitations enabling robust assessment of APOE genotype cerebrovascular cell-type composition. We performed differential abundance analysis across all vascular cell types using MiloR, a k-nearest-neighbor graph-based approach designed to detect changes in non-discrete cell populations along differentiation or developmental trajectories⁴¹. MiloR identified 250 cell neighborhoods with significant differential abundance between APOE4 carriers and APOE3/3 individuals (spatial FDR < 0.05) (Table S2).

To examine the directionality of APOE4-associated changes in vascular cell composition, we plotted the fold changes of significantly altered neighborhoods for each cell type (Figure 1C). APOE4 carriers showed a significant reduction of pericytes and a significant increase of SMCs compared to APOE3/3 individuals. In contrast, astrocytes and endothelial cells exhibited no significant net changes in abundance. We confirmed these results with Dirichlet regression, controlling for age, sex, AD status, and dataset of origin, leading us to focus subsequent analyses on mural cells (Figure S1H).

To directly assess APOE4-associated pericyte loss, we performed NG2 immunostaining on postmortem human hippocampal sections. APOE4 carriers exhibited a significant reduction in NG2-positive pericyte coverage along hippocampal vasculature compared to age-matched APOE3/3 individuals (Figure 1D). In parallel, human postmortem staining revealed significantly increased αSMA immunoreactivity in the APOE4 hippocampus compared to APOE3/3 controls (Figure 1D). However, while αSMA positivity was primarily confined to large vessels in APOE3/3 hippocampi, APOE4 carriers displayed abundant αSMA immunoreactivity throughout the parenchyma (Figure 1D). Together, these findings indicated that APOE4 carriers exhibit both pericyte loss and the emergence of atypical non-vascular αSMA-positive cells, suggesting that APOE4 may alter mural cell identity.

A myofibroblast-like mural cell population is enriched in the APOE4 brain

To investigate APOE4-associated transcriptional changes in mural cells, we subclustered the original pericyte and SMC populations. Consistent with previous studies^38,39, we identified two pericyte subtypes (Pericyte_1 and Pericyte_1) and two SMC subtypes (SMC_1 and SMC_2) (Figure 1E, S2A; Table S3, S4). We then performed differential expression and gene set enrichment analyses across mural cell subtypes between APOE4 carriers and non-carriers (Figure S2B, 2C; Table S5, S6). The Pericyte_1 subcluster had the most differentially expressed genes (1210 genes), followed by SMC_2 (260 genes), SMC_1 (184 genes), and Pericyte_2 (98 genes) (Figure S2B). However, for the GSEA results, SMC_2 showed the greatest number of dysregulated pathways with 61, followed by Pericyte_1 (9 pathways), SMC_1 (7 pathways), and Pericyte_2 (6 pathways) (FDR < 0.05; Figure S2C; Table S5, S6).

Due to the large number of SMC_2 dysregulated pathways, we hypothesized that this population may underly the nonvascular αSMA staining observed in APOE4 postmortem brains. The most upregulated pathways in APOE4 SMC_2 cells were related to cellular contraction and extracellular matrix (ECM) production (Figure 1F), which was confirmed by pseudobulk analysis (Figure S2D, Table S7, S8). Although SMCs typically adopt either a contractile or ECM-synthetic state ^42–44, both contraction and ECM signature scores were increased in APOE4 SMC_2 cells, with more cells co-expressing both programs (Figure S2E,F), suggesting a hybrid phenotype.

Co-expression of contractile and ECM programs is a defining feature of myofibroblasts known to promote tissue fibrosis^45–50. Although best characterized in peripheral organs^51–54, brain myofibroblasts appear following traumatic injury or ischemic events^55,56, where some are thought to arise from pericytes that detach from the vasculature⁵⁵. To determine whether the APOE4-enriched SMC_2 population resembles myofibroblasts, we performed unbiased cell-type annotation using EnrichR with the CellMarker database^57–59. Myofibroblasts were identified as the top-matching cell type for the APOE4 SMC_2 gene signature (Figure 1G, Table S9). Consistent with this annotation, myofibroblast gene activity scores were significantly higher in SMC_2 cells from APOE4 compared with APOE3/3 carriers (Figure S2G).

To determine whether myofibroblasts are selectively enriched in APOE4 carriers, we performed differential abundance analysis of mural cell clusters using DAseq. This identified a distinct APOE4-enriched subregion within SMC_2 that spatially colocalized with the peak of ECM and contractile gene co-expression (Figure 1H). Differential expression analysis demonstrated upregulation of genes involved in ECM organization, cell-substrate junctions, and contractile fibers, whereas the remaining SMC_2 cells preferentially expressed genes associated with vascular homeostasis (Figure S2H,I; Table S10,S11). Unbiased annotation of this subregion again identified myofibroblasts as the top-matching cell type (Figure S2J; Table S12), and enrichment of this population in APOE4 carriers was independently confirmed by Dirichlet regression and MiloR (Figure S2K,L). Together, the transcriptional profile of APOE4 SMC_2 cells, unbiased cell-type annotation, differential abundance analyses, and the presence of non-vascular parenchymal αSMA staining in postmortem brain tissue strongly support that the APOE4-enriched SMC_2 subpopulation represents myofibroblasts.

Previous studies have shown that cerebrovascular degeneration in APOE4 carriers occurs independently of amyloid and tau pathology¹⁰. To assess whether APOE4-associated myofibroblasts are similarly independent of AD pathology we modeled subject-level myofibroblast abundance using a covariate-adjusted binomial generalized linear model including APOE genotype, AD status, age, sex, dataset, and APOE×AD status and APOE×sex interaction terms (Figure S2M). APOE4 emerged as the strongest predictor of myofibroblast abundance (Figure S2M), with significantly higher model-predicted myofibroblast proportions than APOE3/3 when averaged across all covariates (Figure 1I). This association was consistent across sex and AD status, with neither factor significantly modifying myofibroblast abundance within APOE4 carriers (Figure S2N). Finally, leave-one-dataset-out sensitivity analyses confirmed that the APOE4 effect was not driven by any individual dataset (Figure S2O). Together, these findings indicate that APOE4 independently promotes myofibroblast accumulation.

To further test whether APOE4 promotes myofibroblast formation in the absence of AD or other cerebrovascular pathology, we analyzed APOE targeted replacement (TR) mice expressing human APOE3/3 or APOE4/4. Bulk cortical RNA-seq data⁶⁰ demonstrated a positive association between APOE4 gene dosage and myofibroblast gene activity (Figure S3A). Consistent with these findings, aged APOE4/4 mice exhibited reduced pericyte coverage of small vessels (<6 μm) and increased αSMA immunoreactivity in small vessels and the surrounding parenchyma, whereas large cerebral vessels showed comparable αSMA expression between genotypes (Figure 1J, S3B). We further validated this phenotype using another myofibroblast contractile protein, Tagln, once again finding an increase in small vessel and parenchymal immunoreactivity in aged APOE4/4 mice (Figure S3C). These results indicate that APOE4 promotes the induction of non-vascular myofibroblasts in aged mice in the absence of other AD genetic factors.

Finally, we assessed whether APOE4 is sufficient to induce myofibroblasts in a human context using miBrains, a 3D human brain tissue derived from isogenic APOE3/3 or APOE4/4 iPSCs differentiated into the six major brain cell types, including vasculature^61,62 (Figure 1K). Analysis of our previously generated miBrain snRNA-seq dataset⁶² identified two mural cell subclusters, with APOE3/3 cells enriched in cluster 1 and APOE4/4 cells enriched in cluster 2 (Figure 1L,M; S3D,E). Mural cluster 2 exhibited enrichment of ECM and contractile pathways that closely resembled the APOE4 human cerebrovascular myofibroblast program, which was confirmed by pseudobulk analysis (Figure 1N,O; S3F; Table S13–S16). Module scoring further demonstrated a shift from a pericyte identity in cluster 1 to a myofibroblast identity in cluster 2 (Figure 1P). Consistent with these transcriptional findings, APOE4/4 miBrains exhibited reduced pericyte coverage and increased nonvascular αSMA immunoreactivity, which was independently validated by increased TAGLN immunoreactivity (Figure 1Q, S3G). Together, these results demonstrate that APOE4 is sufficient to promote pericyte loss and the emergence of non-vascular myofibroblasts, recapitulating the transcriptional and morphological features observed in human and mouse APOE4 brains.

APOE4 expression in mural cells is necessary and sufficient for myofibroblast accumulation

Having identified an APOE4-associated myofibroblast phenotype across three model systems, we next asked whether it arises intrinsically within mural cells or required interactions with other APOE4-expressing cell types. To test this, we performed a genetic mixing experiments in which isogenic APOE3/3 mural cells (iMCs) were selectively integrated into APOE4/4 miBrains (Figure 2A). As expected, APOE4/4 miBrains exhibited significantly reduced pericyte coverage and increased αSMA immunoreactivity compared to isogenic APOE3/3 controls (Figure 2B). Strikingly, replacing APOE4/4 iMCs with APOE3/3 iMCs restored pericyte coverage and reduced αSMA immunoreactivity (Figure 2B), indicating that APOE4 expression in mural cells is necessary for myofibroblast formation. Similar results were found using two additional independently derived isogenic iPSC lines with reciprocal APOE gene editing (Figure S3H–K).

Figure 2. APOE4 expression in mural cells is necessary and sufficient for myofibroblast accumulation.

A. Schematic of the hybrid genetic experiment in which APOE3/3 iPSC-derived mural cells (iMCs) were integrated into APOE4/4 miBrains.

B. Representative images of isogenic APOE3/3, APOE4/4, and APOE4/4 miBrains integrated with APOE3/3 iMCs stained for PECAM, αSMA, and NG2. Scale bar, 50 μm.

C. Schematic of the reciprocal hybrid genetic experiment in which APOE4/4 iMCs were integrated into APOE3/3 miBrains.

D. Representative images of isogenic APOE3/3, APOE4/4, and APOE3/3 miBrains integrated with APOE4/4 iMCs stained for PECAM, αSMA, and NG2. Scale bar, 50 μm.

E. Schematic of a previously published mouse single-cell RNA-sequencing dataset with conditional human APOE expression in mural cells.

F. UMAP of mural cells from this dataset.

G. DAseq differential abundance analysis of mural cell clusters. Red and blue indicate enrichment and depletion in aged iE4/Cre+ mice, respectively. The enriched SMC subregion was annotated.

H. Top enriched GO pathways of the annotated iE4/Cre+ region.

I. Representative marker genes of the annotated iE4/Cre+ region.

J. Top Cell-Marker annotations of the iE4/Cre+ enriched region.

For B and D, three-dimensional pericyte coverage was quantified as NG2-CD144 overlap normalized to total CD144 volume. αSMA-psositive area was normalized to nuclei. Values were expressed relative to APOE3/3. Data represent individual miBrains (6 images/miBrain; n = 5-6 miBrains/genotype). Bars indicate mean ± SEM. One-way ANOVA with Dunnett’s multiple-comparisons test.

See also Figures S3 and S4.

To assess whether APOE4 mural cells are sufficient to induce myofibroblasts, we performed the opposite genetic mixing experiment, integrating APOE4/4 iMCs into an APOE3/3 miBrain (Figure 2C). This single manipulation reduced pericyte coverage and increased αSMA immunoreactivity to levels observed in APOE4/4 miBrains (Figure 2D), demonstrating that APOE4 expression in mural cells is sufficient to induce the phenotype. Consistent with these findings, APOE4/4 iMC monocultures exhibited increased αSMA and fibronectin expression across all donor lines (Figure S4A–C). Finally, analysis of a published single-nucleus RNA-seq dataset in which human APOE4 was conditionally expressed in mural cells of Apoe-KO mice identified an APOE4-enriched SMC subpopulation with myofibroblast-associated transcriptional signatures that was independently annotated as myofibroblasts (Figure 2E–J; Table S17–S19), confirming that APOE4 expression in mural cells is sufficient to drive this fibrotic phenotype.

Reduced vascular pathology following APOE depletion suggests APOE4 acts through a toxic gain-of-function mechanism^25,63. To determine whether a similar mechanism underlies myofibroblast accumulation, we reanalyzed the dataset from Figure 2E comparing iE4/Cre+ and Apoe-KO mice. Consistent with a toxic APOE4 gain of function, the myofibroblast population was significantly reduced in Apoe-KO mice (Figure S4D). Similarly, recombinant APOE4 induced αSMA and fibronectin expression in APOE-KO iPSC-derived iMCs (Figure S4E). Together, these findings support a causal role for APOE4 in mural cell myofibroblast induction.

APOE4 promotes a pericyte-to-myofibroblast transition

Pericytes differentiate into myofibroblasts following injury in multiple organ systems, including the brain^51,55. A pericyte-to-myofibroblast transition (PMT) would explain the observed correlation between myofibroblast expansion and pericyte loss (Figure 1D, J, Q). Pseudotime analysis of the human cerebrovascular atlas identified two distinct pericyte-derived trajectories: one terminating in the APOE4-enriched myofibroblast population (PMT lineage) and the other in pericytes (pericyte lineage) (Figure 3A). Gene signature scoring confirmed progressive acquisition of myofibroblast programs with loss of pericyte identity along the PMT lineage (Figure 3B). Curve weight analysis further showed that APOE4 mural cells were enriched along the PMT lineage, whereas APOE3/3 cells preferentially aligned with the pericyte lineage (Figure 3C).

Figure 3. APOE4 promotes a pericyte-to-myofibroblast transition.

A. Pseudotime analysis of the mural cell subclusters identified in Figure 1.

B. Myofibroblast and pericyte signature trajectories across pseudotime.

C. Lineage bias between APOE4 carriers and APOE3/3 individuals. A value of 0 indicates exclusive assignment to the pericyte-to-pericyte lineage, whereas a value of 1 indicates exclusive assignment to the pericyte-to-myofibroblast lineage.

D. Schematic of ACTA2 and CSPG4 expression during the pericyte-to-myofibroblast transition.

E. Density plot of mural cells stratified by ACTA2 and CSPG4 expression. Subject-level predicted mural cell-state proportions were generated using the Dirichlet regression model (Figure S5A).

F. Representative images of the pericyte-to-myofibroblast transition in APOE3/3 and APOE4/4 miBrains and APOE-TR mice stained for CD144/lectin, αSMA, and NG2.

G. Schematic of cell-type deconvolution of the mouse bulk RNA-sequencing dataset.

H. Pericyte and myofibroblast scores across ages and APOE-TR genotypes.

I. Representative images of APOE4/4 iPSC-derived mural cells from two donors stained for αSMA and NG2 over serial passage. Scale bar, 50 μm.

For E, data represent individuals (n = 100 APOE3/3, n = 102 APOE4 carriers). Values were normalized to APOE3/3. Boxes indicate median and IQR. Benjamini-Hochberg-adjusted P values are shown.

For F and I, cells were classified as pericytes (vascular NG2+/αSMA−), intermediate cells (NG2+/αSMA+), or myofibroblasts (nonvascular NG2−/αSMA+). Cell-state proportions were normalized to the total number of classified cells. Data represent individual miBrains or mice (F; 6 images/miBrain or mouse, n = 6 miBrains and n = 3 mice/genotype) or wells (I; 6 images/well, n = 6 wells per donor line and passage). Statistical analyses were performed using two-way repeated-measures ANOVA with Geisser-Greenhouse correction and Tukey’s multiple-comparisons test (F) or binomial generalized linear models with Holm correction (I).

For H, data represent individual mice (n = 16 mice per age group and genotype). Boxes indicate median and IQR. Statistical analysis is described in the Methods.

See also Figure S5.

The pericyte-to-myofibroblast transition occurs as a continuum, with cells passing through an intermediate CSPG4⁺/ACTA2⁺ state before acquiring a mature myofibroblast identity (Figure 3D)^51,64. Accordingly, we found APOE4 carriers displayed a higher abundance of CSPG4⁺/ACTA2⁺ intermediate cells than APOE3/3 individuals (Figure 3E). Covariate-adjusted Dirichlet regression confirmed an APOE4-associated shift in mural cell composition, characterized by reduced CSPG4⁺/ACTA2⁻ pericytes and increased CSPG4⁺/ACTA2⁺ intermediate and CSPG4⁻/ACTA2⁺ myofibroblast populations (Figure 3E). Notably, the intermediate CSPG4⁺/ACTA2⁺ population was more abundant in APOE4 individuals without AD than in those with AD (Figure S5A), suggesting that APOE4-associated vascular remodeling may occur prior to overt neurodegenerative pathology.

To validate PMT in our model systems, we classified mural cells in miBrains and APOE-TR mice as pericytes (vascular NG2+/αSMA−), intermediate cells (NG2+/αSMA+), or myofibroblasts (nonvascular NG2−/αSMA+) while restricting analyses to small vessels (<6 μm) to exclude SMCs. Consistent with our human atlas, APOE4/4 miBrains and APOE4/4-TR mice exhibited reduced pericyte abundance with increased intermediate and myofibroblast populations (Figure 3F). Vascular distance also progressively increased from pericytes to intermediate cells to myofibroblasts in both model systems (Figure S5B), consistent with pericyte detachment from the vasculature during PMT.

Given our finding that APOE4 promotes myofibroblast characteristics in iMC monocultures (Figure S4A–C), we also examined the APOE4-induced PMT in this context. Using CIBERSORTx⁶⁵ deconvolution of a bulk RNA-seq dataset²⁵, we mapped iMCs onto mural cell subclusters from our cerebrovascular atlas (Figure S5C). Overall, iMCs most closely aligned with pericytes (Figure S5D). However, APOE4/4 iMCs exhibited reduced pericyte identity and increased myofibroblast identity compared with APOE3/3 iMCs (Figure S5E). Differential expression analysis further demonstrated upregulation of canonical myofibroblast pathways in APOE4/4 iMCs (Figure S5F; Table S20 and S21).

If myofibroblasts arise from pericytes rather than representing a pre-existing population, we would expect a progressive shift in mural cell states over time. Consistent with this model, analysis of the conditional APOE mouse dataset (Figure 2E) revealed an age-dependent decrease in pericytes accompanied by an increase in myofibroblasts in iE4/Cre+ mice (Figure S5G). We next performed CIBERSORTx deconvolution of a longitudinal bulk RNA-seq dataset from APOE-TR mice⁶⁶ using pericyte and myofibroblast marker genes derived from the conditional APOE mouse dataset (Figure 3G). APOE3-TR and APOE4-TR mice showed similar mural cell compositions at 3 months of age, whereas APOE4-TR mice progressively lost pericyte signatures and gained myofibroblast signatures by 12 and 24 months, resulting in significant genotype-by-age interactions for both populations (Figure 3H).

We next assessed differentiation potential within the miBrain snRNA-seq dataset. CytoTRACE2 predicted progressively decreasing developmental potential from the pericyte cluster toward the myofibroblast cluster, consistent with a pericyte-to-myofibroblast transition (Figure S5H). Integrating these scores with the k-nearest neighbor graph using CellRank2 further inferred a differentiation trajectory from pericytes toward myofibroblasts (Figure S5I).

Finally, we examined iMC monocultures over serial passage to determine whether APOE4 pericytes progressively acquire myofibroblast features. At early passages (P1–P2), nearly all classified APOE4/4 iMCs were NG2+/αSMA−, consistent with a pericyte identity. By passage 5, APOE4/4 iMCs exhibited a marked shift toward NG2+/αSMA+ intermediate and NG2−/αSMA+ myofibroblast states. This temporal transition was reproduced across two independent donor iPSC lines (Figure 3I).

Collectively, these results across human tissue, iPSC-derived models, and APOE-TRmice demonstrate that APOE4 expression in mural cells drives a pericyte-to-myofibroblast transition.

Myofibroblast-derived fibronectin increases amyloid accumulation in APOE4 models

Myofibroblasts are the principal mediators of fibrosis through secretion of ECM proteins, including collagens and fibronectin (FN1)⁶⁷. Because genetic variants in FN1 are associated with cognitive resilience in APOE4 carriers³², we asked whether APOE4-associated myofibroblasts promote cerebrovascular fibronectin deposition. APOE4/4-TR mice exhibited increased fibronectin immunoreactivity along small vessels compared with age- and sex-matched APOE3/3 controls, coinciding with elevated αSMA staining (Figure 4A). Similarly, APOE4/4 miBrains showed increased vascular fibronectin deposition relative to isogenic APOE3/3 miBrains (Figure 4B). Increased deposition of collagen I and collagen III in both APOE4/4-TR mice and APOE4/4 miBrains further confirmed cerebrovascular fibrosis (Figure S6A,B).

Figure 4. Myofibroblast-derived fibronectin drives amyloid accumulation in APOE4 models.

A. Representative images of corpus callosum microvessels from APOE3/3-TR and APOE4/4-TR mice stained for lectin, αSma, and fibronectin. Scale bar, 10 μm.

B. Representative images of APOE3/3 and APOE4/4 miBrains stained for PECAM and fibronectin. Scale bar, 50 μm.

C. Validation of cerebrovascular fibrosis in additional isogenic iPSC-derived mural cell (iMC) pairs using the experimental paradigm shown in Figure S3I. ADRC scale bar, 25 μm; sAD scale bar, 50 μm.

D. Representative images of isogenic APOE3/3 and APOE4/4 miBrains stained with three amyloid-β antibodies recognizing distinct epitopes. Scale bar, 25 μm.

E. Representative images of isogenic APOE3/3, APOE4/4, and APOE4/4 miBrains integrated with APOE3/3 iMCs stained for PECAM, fibronectin, and amyloid (12F4). Scale bar, 25 μm.

F. Schematic of the FN1 knockdown experiment.

G. Representative images of APOE4/4 miBrains integrated with scramble shRNA- or FN1 shRNA-transduced iMCs stained for fibronectin and amyloid (12F4). Scale bar, 25 μm.

H. Experimental paradigm for treating encapsulated APOE3/3 iPSC-derived endothelial cells (iECs) with amyloid-β (1-40/1-42) alone or in combination with fibronectin for 96 h, followed by D54D2 immunostaining. Scale bar, 50 μm.

For all panels, αSMA-, fibronectin-, and amyloid-positive area were normalized to nuclei and expressed relative to APOE3/3 (A–D), APOE4/4 (E and G), or amyloid-treated iECs (H). Data represent individual miBrains (6 images/miBrain; n = 4–6 miBrains/condition), mice (6 images/mouse; n = 4 mice/genotype), or iEC cultures (6 images/culture; n = 6 cultures/condition). Bars indicate mean ± SEM. Unpaired two-tailed Student’s t-tests (A–D and G) or one-way ANOVA with Dunnett’s multiple-comparisons test (E and H).

See also Figure S6.

Myofibroblasts were the largest contributor of FN1 expression among individuals from our cerebrovascular atlas (Figure S6C). Furthermore, most of the fibronectin deposition in APOE4/4-TR mice occurred near αSMA+ regions (Figure S6D), implicating myofibroblasts as a causal source of vascular fibrosis. To directly test this, we swapped APOE4/4 mural cells to APOE3/3 in an APOE4/4 miBrain, an approach that previously reversed APOE4-associated myofibroblast phenotypes (Figure 2B). Remarkably, this single swap led to a significant reduction in fibronectin protein across multiple donor lines (Figure 4C), suggesting that APOE4 mural cell-derived myofibroblasts are the main contributor to fibronectin accumulation.

In the central nervous system, perivascular fibroblasts are also a major source of fibronectin and contribute to fibrotic scar formation in response to brain injury^68,69. To rule out the involvement of this cell type, we analyzed perivascular fibroblasts from the original snRNA-seq datasets corresponding to the same subjects included in our cerebrovascular atlas. We found that the myofibroblasts identified in this study exhibited significantly higher FN1 expression compared to fibroblasts (Figure S6E), and that APOE genotype had no significant effect on ECM-related pathways in fibroblasts (Table S22, S23), further highlighting mural cell-derived myofibroblasts as the primary cell type responsible for APOE4-associated vascular ECM remodeling.

To determine whether mural cell FN1 expression is associated with clinical features of AD, we analyzed AD individuals within the Haney et al. dataset³⁷, which includes detailed clinical annotations. Consistent with our earlier findings, APOE4 carriers exhibited higher mural cell FN1 expression than APOE3/3 individuals (Figure S6F). Moreover, mural cell FN1 expression was inversely correlated with age at AD diagnosis (Figure S6G). Individuals with the highest global neuropathological burden also exhibited significantly higher mural cell FN1 expression (Figure S6H). Consistent with an APOE4-mediated effect, APOE4/4 miBrains immunostained with three independent amyloid-β antibodies recognizing distinct Aβ epitopes showed significantly increased vascular amyloid immunoreactivity in APOE4/4 miBrains compared with isogenic APOE3/3 controls, suggesting that myofibroblast secreted fibronectin may promote amyloid pathology (Figure 4D).

If fibronectin causally contributes to amyloid-β pathology, then reducing myofibroblast-derived fibronectin should also lower amyloid accumulation. Indeed, APOE4/4 miBrains containing APOE3/3 iMCs exhibited significantly reduced vascular amyloid deposition compared with APOE4/4 controls, a finding confirmed with two independent amyloid antibodies (Figure 4E, S6I). Conversely, introducing APOE4/4 iMCs into APOE3/3 miBrains increased both fibronectin and amyloid deposition (Figure S6J), demonstrating that APOE4 mural cells are both necessary and sufficient to drive vascular fibrosis and amyloid accumulation. APOE4/4 miBrains also exhibited increased vascular APOE deposition with enhanced APOE colocalization with both fibronectin and amyloid, suggesting that APOE may associate with the fibrotic extracellular matrix to promote amyloid accumulation (Figure S6K,L).

Fibronectin directly binds to amyloid and may promote its aggregation⁷⁰. To determine whether the reduction in amyloid pathology observed when APOE3/3 iMCs replaced APOE4/4 iMCs in the miBrain was mediated by decreased fibronectin, APOE4/4 iMCs were transduced with FN1 shRNA or a scramble control before incorporation into APOE4/4 miBrains (Figure 4F). Efficient FN1 knockdown was confirmed by qRT-PCR and immunostaining (Figure 4G, S6M). Specific knockdown of FN1 in APOE4 iMCs significantly reduced vascular amyloid deposition compared with scramble controls suggesting that fibronectin promotes vascular amyloid retention (Figure 4G). Independently validating this mechanism, APOE3/3 iPSC-derived endothelial cells treated with an amyloid-fibronectin mixture exhibited significantly increased amyloid accumulation compared with amyloid alone (Figure 4H). Together, these findings demonstrate that myofibroblast-derived fibronectin is required for APOE4 vascular amyloid accumulation, establishing a mechanistic link between APOE4-induced myofibroblasts and AD pathology.

Increased TGF-β signaling in the APOE4 brain promotes the pericyte-to-myofibroblast transition

Having identified a key role for APOE4-derived myofibroblasts in AD pathology, we next investigated the molecular mechanisms driving the PMT. NicheNet⁷¹ analysis of APOE4 myofibroblast marker genes identified TGF-β as the top predicted upstream regulator (Figure 5A,B; Table S24). Consistent with its established profibrotic role, TGF-β was predicted to regulate multiple myofibroblast-associated genes, including FN1 and ACTA2 (Figure 5C,D). Within the human cerebrovascular atlas, TGFβ1 was predominantly expressed by myofibroblasts and TGFβ2 by astrocytes, with only modest differences in ligand expression between APOE4 and APOE3/3 individuals (Figure 5C,E,F). Despite these small expression differences, TGF-β pathway activity progressively increased along the inferred pericyte-to-myofibroblast trajectory in human mural cells and the CytoTRACE-derived trajectory in miBrain mural cells (Figure 5G; S5H,I; S7A). Mural cells from both APOE4 carriers and APOE4/4 miBrains exhibited increased expression of key TGF-genes compared with APOE3/3 controls (Figure 5H; Figure S7B). Immunostaining for nuclear pSMAD2 confirmed increased TGF-β pathway activation around the cerebrovasculature of both APOE4/4 miBrains and APOE4/4-TR mice (Figure 5I).

Figure 5. Increased TGF-β signaling in the APOE4 brain drives the pericyte-to-myofibroblast transition.

A. Schematic of NicheNet analysis using APOE4 myofibroblast marker genes.

B. Top predicted ligands driving APOE4 myofibroblast marker expression.

C. Circos plot of the top predicted ligands and their target genes.

D. Predicted target genes of the top ligands.

E. Expression of the top ligands across cerebrovascular cell types.

F. Differential ligand expression in APOE4 carriers relative to APOE3/3 by cell type.

G. TGF-β signaling signature trajectories across pseudotime.

H. TGF-β signaling scores in mural cells from the cerebrovascular atlas and miBrain single-nucleus RNA-sequencing datasets.

I. Representative images of APOE3/3 and APOE4/4 miBrains and APOE-TR mice stained for PECAM/lectin and pSMAD2. Scale bar, 25 μm.

J. Upregulated TGF-β signaling GO pathways in APOE4/4 iPSC-derived mural cells.

K. TGF-β receptor expression in isogenic iPSC-derived mural cells from two donor lines.

L. Representative images of APOE4/4 iPSC-derived mural cells from two donors stained for pSMAD2 and PDGFRβ over serial passage. Scale bar, 50 μm.

For H, data represent individuals (n = 106 APOE3/3, n = 106 APOE4 carriers) or miBrains (n = 2/genotype). Values are expressed relative to APOE3/3. Bars indicate mean ± SEM. Welch’s unpaired Student’s t-test.

For I and L, pSMAD2-positive nuclei were normalized to the total number of vascular nuclei (I) or total nuclei (L). Data represent individual miBrains or mice (I; 6 images/miBrain or mouse, n = 6 miBrains and n = 4 mice/genotype) or wells (L; 6 images/well, n = 6 wells per donor line and passage). Statistical analyses were performed using unpaired two-tailed Student’s t-tests (I) or binomial generalized linear models with Holm correction (L).

For K, MSSM3 expression was measured by bulk RNA sequencing (n = 3 biological replicates/genotype) and sAD expression by qRT-PCR (3 technical replicates/well; n = 4 wells/genotype). Values are expressed relative to APOE3/3. Bars indicate mean ± SEM. Limma-adjusted P values (MSSM3) and unpaired two-tailed Student’s t-tests (sAD) are shown.

See also Figure S7.

Given that APOE4 mural cells are both necessary and sufficient for myofibroblast formation, we next asked whether elevated TGF-β signaling is intrinsic to APOE4 mural cells. APOE4/4 iMCs exhibited significantly higher nuclear pSMAD2 levels than isogenic APOE3/3 controls, whereas treatment of APOE3/3 iMCs with recombinant TGFβ1 increased pSMAD2 to levels comparable to baseline APOE4/4 cells (Figure S7C). Bulk RNA-seq of iMCs further demonstrated increased TGF-β pathway activity together with upregulation of TGFβR1 and TGFβR2 in APOE4/4 iMCs, findings that were independently validated by qRT-PCR (Figure 5J,K). Inhibition of TGF-β signaling with SB431542 reduced αSMA and fibronectin expression (Figure S7D), supporting a functional role for TGF-β signaling in myofibroblast induction. Consistent with the temporal emergence of myofibroblasts (Figure 3I), aged APOE4/4 iMC cultures (P5) also exhibited increased pSMAD2-positive cells compared with early passages (P1–P2) (Figure 5L).

Taken together, these findings indicate that APOE4 mural cells have increased TGF-β signaling that promotes a pericyte-to-myofibroblast transition.

TGF-β inhibition reverses APOE4-driven myofibroblast accumulation, cerebrovascular fibrosis, and amyloid deposition

Having identified TGF-β signaling as a driver of the pericyte-to-myofibroblast transition, we next tested whether pharmacologic inhibition could reverse the APOE4 phenotype. APOE4/4 miBrains were treated with two chemically distinct ALK5 (TGFβR1) inhibitors, SB431542 and Galunisertib (Figure 6A). Both inhibitors restored pericyte coverage while reducing non-vascular αSMA immunoreactivity, indicating reversal of the pericyte-to-myofibroblast transition (Figure 6B; Figure S7E). Consistent with this, both inhibitors also reduced fibronectin deposition and vascular amyloid accumulation to levels comparable to APOE3/3 controls, findings reproduced with two independent amyloid antibodies (Figure 6C; S7F,G). Both compounds additionally reduced pSMAD2-positive vascular nuclei, confirming effective suppression of canonical TGF-β signaling (Figure S7H).

Figure 6. TGF-β inhibition reverses APOE4-driven myofibroblast accumulation, cerebrovascular fibrosis, and amyloid accumulation.

A. Timeline of TGF-β inhibition in miBrains.

B. Representative images of myofibroblast phenotypes in APOE3/3 and APOE4/4 miBrains treated with DMSO vehicle or APOE4/4 miBrains treated with 10 μM SB431542 for 2 weeks. miBrains were stained for PECAM, αSMA, and NG2. Scale bar, 25 μm.

C. Representative images of cerebrovascular fibrosis and amyloid deposition in APOE3/3 and APOE4/4 miBrains treated with DMSO vehicle or APOE4/4 miBrains treated with 10 μM SB431542 for 4 weeks. miBrains were stained for PECAM, fibronectin, and amyloid (D54D2). Scale bar, 25 μm.

For panels B and C, three-dimensional pericyte coverage was quantified as overlapping NG2 and PECAM volume normalized to total PECAM volume. αSMA−, fibronectin-, and amyloid-positive area were normalized to nuclei and expressed relative to APOE4/4 vehicle. Data represent individual miBrains (6 images/miBrain; n = 6 miBrains/condition). Bars indicate mean ± SEM. One-way ANOVA with Dunnett’s multiple-comparisons test.

D. Timeline of TGF-β inhibition in aged APOE4/4-TR mice.

E. Representative cortical images of aged APOE4/4-TR mice treated with DMSO vehicle or 10 mg/kg GW788388 daily for 4 weeks and stained for lectin, αSMA, and NG2. Scale bar, 75 μm.

F. Representative cortical images of aged APOE4/4-TR mice treated with DMSO vehicle or 10 mg/kg GW788388 daily for 4 weeks and stained for lectin, fibronectin, and amyloid-β. Scale bar, 75 μm.

For panels E and F, pericyte coverage was quantified as overlapping Ng2 and lectin area and expressed relative to vehicle. αSMA−, fibronectin-, and amyloid-positive area were normalized to nuclei and expressed relative to vehicle. Data represent individual mice (6 images/mouse; n = 6 mice/condition). Bars indicate mean ± SEM. Unpaired two-tailed Student’s t-test.

G. Proposed model of the TGF-β-mediated pericyte-to-myofibroblast transition in APOE4 carriers.

See also Figure S7.

Finally, we asked whether TGF-β inhibition could reverse the pericyte-to-myofibroblast transition, cerebrovascular fibrosis, and amyloid accumulation in vivo. To test this, we treated 12-month-old APOE4/4-TR mice with either a DMSO vehicle or the TGF-β inhibitor GW788388 for four weeks via daily intraperitoneal injections (Figure 6D). Following GW788388 treatment, we observed a significant reduction in pSMAD2+ nuclei, confirming suppression of canonical TGF-β signaling (Figure S7I). Consistent with our miBrain results, TGF-β inhibition in aged APOE4/4-TR mice significantly increased pericyte coverage and reduced αSMA immunoreactivity relative to control (Figure 6E), indicating a reversal of the pericyte-to-myofibroblast transition in vivo. Strikingly, this reversal was accompanied by a significant decrease in fibronectin deposition and amyloid-β immunoreactivity in drug-treated mice (Figure 6F). Together, this demonstrates that TGF-β inhibition can reverse APOE4 cerebrovascular pathologies in human brain tissue in vitro and in aged APOE4 mice.

Discussion

Overall, the results presented here reveal that elevated TGF-β signaling drives a pericyte-to-myofibroblast transition in the APOE4 brain. This leads to a loss of pericyte coverage and fibrosis of the APOE4 vasculature, with fibronectin accumulation directly promoting perivascular amyloid deposition. Pharmacologically inhibiting TGF-β signaling in APOE4 in vitro and in vivo models reverses the pericyte-to-myofibroblast transition and attenuates cerebrovascular pathology. These results establish a causal link between TGF-β and cerebrovascular dysfunction in APOE4 carriers, which may influence AD progression in this large patient population. This study thus identifies a pathological mechanism in AD and reveals therapeutic opportunities for APOE4-driven cerebrovascular pathology.

AD has long been characterized by the presence of amyloid plaques and neurofibrillary tangles in the brain parenchyma⁷². However, cerebrovascular pathology is often one of the earliest manifestations of AD, sometimes occurring even before symptoms arise⁹. As part of disease progression, pericytes are lost along the vasculature leading to microvessel degeneration and the leakage of intravascular components into the brain parenchyma^23,73–76. In addition to forming plaques, amyloid often accumulates along the vasculature in AD patients, leading to cerebral amyloid angiopathy⁷⁷. CAA can occur in both large vessels and capillaries (Type 2 and Type 1 CAA, respectively), with APOE4 being particularly associated with capillary involvement⁷⁸. Perivascular ECM accumulation is another feature of aging and AD, and several cerebrovascular ECM proteins, particularly fibronectin, are positively correlated with amyloid pathology^79–97. Strikingly, TGF-β overexpression in mice leads to thickening of the cerebrovascular basement membrane that precedes amyloid accumulation and vascular dysfunction³⁵. These findings suggest a link between TGF-β, perivascular fibrosis, and AD pathogenesis.

This study demonstrates that APOE4 directly increases TGF-β signaling along the cerebrovasculature, leading to vascular fibrosis and amyloid deposition via pericyte differentiation into myofibroblasts. Cerebrovascular degeneration is a well-characterized phenotype in APOE4 carriers that is linked to accelerated cognitive decline¹⁰. Independent of AD pathology, pericyte coverage is significantly reduced in APOE4 individuals and carriers of this allele are more susceptible to cerebral hemorrhages^10,98,99. Additionally, perivascular amyloid accumulation is far more common in APOE4 carriers^98,100. More recently, ECM dysregulation has also been implicated in APOE4-associated AD. APOE4 astrocytes and microglia upregulate matrisome pathways^101,102 and solutes from the brain parenchyma, including amyloid, drain along the vasculature^103–106. This suggests that perivascular accumulation of pro-aggregatory ECM proteins may drive impaired amyloid clearance^79,84. Indeed, perivascular amyloid drainage is decreased in human APOE4-targeted replacement mice¹⁰⁷. Strong evidence for a fibrotic phenotype in APOE4-mediated AD pathogenesis is a recent study that identified genetic factors protecting carriers of this allele from developing the disease³². They found that many protective variants are in ECM genes and that a particular loss-of-function FN1 variant reduces the risk of developing AD by approximately 71%. Coinciding with these results, the brains of APOE4 carriers have increased perivascular fibronectin deposition compared to non-carriers^32,33. To date, these APOE4-associated vascular pathologies – pericyte loss, amyloid deposition, and fibrosis – have all been investigated separately. However, this study identifies a TGF-β-dependent mechanism that links these phenotypes together into a single, targetable pathway. We show that inhibiting TGF-β can reverse cerebrovascular pathology in APOE4 which is, at least in part, due to reduced fibronectin accumulation. Therefore, the work presented here simultaneously explains a mechanism of AD risk and resilience and offers a therapeutic opportunity for APOE4 individuals.

While our data support a primarily cell-autonomous role for APOE4 in driving the pericyte-to-myofibroblast transition, several important questions remain. First, increased TGF-β ligand expression has been reported across multiple APOE4 cell types, including astrocytes and microglia, which raises the possibility that non–cell-autonomous signaling may contribute to mural cell reprogramming in APOE4 brains^108,109. Defining the contributions of autocrine versus paracrine TGF-β signaling in vivo will be an important area for future study. Second, our findings that APOE expression is required for myofibroblast accumulation (Figure S4D–E) suggest that therapeutic strategies aimed at lowering APOE levels may represent an alternative approach to targeting downstream TGF-β signaling. Finally, APOE4 has been widely associated with lipid dysregulation across multiple CNS cell types, and cholesterol-lowering therapies have been shown to reduce myofibroblast accumulation and fibrosis in other organ systems^110–113. Although we do not define the precise mechanism linking APOE4 to increased TGF-β signaling, these observations raise the possibility that altered lipid homeostasis may act upstream of, or in parallel with, TGF-β pathway activation in driving the pericyte-to-myofibroblast transition.

Overall, our study utilizes an orthogonal approach integrating single-nuclei transcriptomic and immunocytochemical analysis of the postmortem human brain with isogenic iPSC brain models and humanized APOE knock-in mice. From this rigorous approach, we discover a causal link between APOE4, TGF-β signaling, cerebrovascular fibrosis, and AD pathology, opening a path to therapeutic opportunities in the treatment and prevention of AD in high-risk individuals.

Limitations of the Study:

This study has several limitations. The miBrain single-nucleus RNA-seq (n = 2 per genotype) and iPSC-derived mural cell (iMC) RNA-seq (n = 3 per genotype) datasets were relatively small, potentially limiting statistical power. The study was also not designed or powered to evaluate sex-specific effects, and the influence of sex on the observed phenotypes could therefore not be determined. In addition, the miBrain snRNA-seq dataset was generated using a previously established model containing A53T mutant neurons. However, because our data support a mural cell-autonomous mechanism underlying the pericyte-to-myofibroblast transition, neuronal genotype is unlikely to explain the observed phenotype. Furthermore, pharmacological TGF-β inhibition affects multiple cell types in the miBrain model, precluding definitive attribution of the rescue to mural cells. Finally, although TGF-β inhibition mitigated APOE4-associated vascular pathology in our experimental models, the broad physiological functions of TGF-β signaling may limit the safety of systemic inhibition.

Resource availability

1. Lead Contact

Further information and requests for resources and reagents should be directed to the lead contact, Joel Blanchard (joel.blanchard@mssm.edu).

2. Materials availability

This study did not generate new unique reagents.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Human iPSC cell lines

Isogenic APOE3/3 and APOE4/4 iPSC pairs were previously established using CRISPR/Cas9 genome editing. iPSCs were cultured on pre-coated Geltrex^™ Matrix plates in StemFlex^™ medium and maintained until they reached 60–70% confluency. For routine maintenance, iPSC colonies were passaged using 0.5 mM EDTA. For differentiation, iPSCs were dissociated into single cells using Accutase^™ for 5–10 minutes at 37°C and used according to the differentiation protocols described below. Prior to experimentation, the APOE genotype of each cell line was confirmed by Sanger sequencing. All iPSC lines used in this study, including donor sex and parental APOE genotype, are listed in the Key Resources Table.

Key Resources Table

REAGENT OR RESOURCE	SOURCE	IDENTIFIER
Antibodies
α-smooth muscle actin	R&D Systems	Cat# MAB1420
α-smooth muscle actin	Abcam	Cat# Ab321866
Amyloid-β (12F4)	BioLegend	Cat# 805501 Recognizes the C-terminus of amyloid-β and is specific for Aβ42
Amyloid-β (6E10)	BioLegend	Cat# 803001 Recognizes amino acids 1–16 of Aβ, detecting full-length amyloid-β peptides
Amyloid-β (D54D2)	Cell Signaling Technology	Cat# 8243S Recognizes residues near the N-terminus of amyloid-β and detects multiple Aβ species, including Aβ37, Aβ38, Aβ39, Aβ40, and Aβ42
Amyloid-β	Rockland	Cat# 600-401-253 Recognizes amino terminus (amino acids 1-14) of Aβ
APOE	Abcam	Cat# Ab52607
CD31/PECAM-1	R&D Systems	Cat# AF806
CD31/PECAM-1	R&D Systems	Cat# BBA7
CD31/PECAM-1	Thermo Scientific	Cat# MA5-29475
Collagen I	Abcam	Cat# Ab34710
Collagen III	Thermo Scientific	Cat# PA5-92066
Fibronectin	R&D Systems	Cat# AF1918
NG2	Abcam	Cat# Ab255811
PDGFRB	Abcam	Cat# Ab69506
pSmad2	Millipore-Sigma	Cat# AB3849-I
Smad2	Cell Signaling Technology	Cat# 5339S
Tagln	Abcam	Cat# Ab14106
VECAD/CD144	R&D Systems	Cat# AF938
Biological samples
Human postmortem brain tissue	Blanchard et al.25	N/A
Chemicals, peptides, and recombinant proteins
Accutase	Stemcell	Cat# 07920
Activin A	Peprotech	Cat# 120-14P
Recombinant APOE4	Peprotech	Cat# 350-04
Astrocyte Growth Supplement	ScienCell	Cat#1852
Astrocyte Medium (AM)	ScienCell	Cat# 1801
B27	Gibco	Cat# 17504044
B27 without Vitamin A	Gibco	Cat# 12587010
BDNF	Peprotech	Cat# 450-02
Biotin	Sigma	Cat#B4639
Blasticidin	Gibco	Cat # A1113903
BMP4	Peprotech	Cat# 120-05ET
CD Lipids	Gibco	Cat# 11905031
CHIR99021	Tocris	Cat# 4423
CNTF	Peprotech	Cat# 450-13
DAPT	Cayman	Cat# 13197
Dibutyryl cAMP	Biogems	Cat# 1698950
DMEM	Gibco	Cat# 11965092
DMEM/F12 with GlutaMAX	Gibco	Cat# 10565018
Doxycyline	Millipore	Cat# D3072
FGF-basic	Peprotech	Cat# 100-18B
Fibronectin	R&D Systems	Cat# 4305-FNB-200
Forkskolin	R&D Systems	Cat# 1099/10
Galunisertib	Tocris	Cat# 6956
GDNF	Peprotech	Cat# 450-10
Gelxtrex	Gibco	Cat# A1413201
GlutaMAX	Gibco	Cat# 35050061
GW788388	MedChem Express	Cat# HY-10326
HGF	Peprotech	Cat# 100-39H
Hoechst33342	Thermo Scientific	Cat# 62249
Human Endothelial SFM	Gibco	Cat# 11111044
IGF-1	Peprotech	Cat# 100-11
IL-34	Peprotech	Cat# 200-34
Insulin	Sigma	Cat# I9278
Laminin	Gibco	Cat# 23017015
L-Ascorbic Acid	Fisher Scientific	Cat# BP351
LDN193189	Tocris	Cat# 6053
Lipofectamine Stem Transfection Reagent	Invitrogen	Cat# STEM00001
Lycopersicon Esculentum (Tomato) Lectin (LEL, TL), DyLight 488	Thermo Scientific	Cat# L32470
MCSF	Peprotech	Cat# 300-25
MEM-Non-Essential Amino Acids	Gibco	Cat#111400050
N2	Gibco	Cat# 17502048
Neurobasal	Gibco	Cat# 21103049
NT3	Peprotech	Cat# 450-03
PDGF-AA	Peprotech	Cat# 100-13A
PDGF-BB	Peprotech	Cat# AF-100-14B
PEG300	MedChem Express	Cat# HY-Y0873
Penicillin-Streptomycin	Gibco	Cat# 15140122
Puromycin	Gibco	Cat# A1113803
Retinoic Acid	MIllipore	Cat# R2625
SAG	Cayman	Cat# 11914
SB431542	Stemgent	Cat# 04-0010
StemFlex	Gibco	Cat# A3349401
TGFB1	Peprotech	Cat# 100-21
TrypLE Select	Gibco	Cat# 12563011
Tween-80	MedChem Express	Cat# HY-Y1891
VEGF-A	Peprotech	Cat# 100-20
Y-27632	Tocris	Cat# 1254
Deposited data
Human brain snRNA-seq dataset	Yang et al.38	GEO: GSE163577
Human brain snRNA-seq dataset	Sun et al.39	Synapse: syn51015750
Human brain snRNA-seq dataset	Haney et al.37	GEO: GSE254205
miBrain snRNA-seq dataset	Mesentier-Louro et al.62	GEO: GSE342463
Conditional APOE mouse brain snRNA-seq dataset	Yamazaki et al.137	Synapse: syn22313650
APOE-TR mouse brain bulk RNA-seq dataset	Zhao et al.66	Synapse: syn20808171
APOE-TR mouse brain bulk RNA-seq dataset	Foley et al.60	Synapse: syn26561824
iPSC-derived mural cell bulk RNA-seq dataset	Blanchard et al.25	GEO: GSE125869
Experimental models: Cell lines
AG09173 (MSSM3) iPSCs (female, APOE3/3 parental)	Massachusetts Institute of Technology	RRID: CVCL_4L66
ADRC5 iPSCs (male, APOE3/3 parental)	UCI	N/A
sAD iPSCs (female, APOE4/4 parental)	Massachusetts Institute of Technology	N/A
Experimental models: Organism/strains
B6.Cg-Apoeem2(APOE*)Adiu j/J (APOE*3 KI)	Jackson Labs	RRID: IMSR_JAX:029018
B6(SJL)- Apoetm1.1(APOE*4)A diuj/J (APOE*4 KI)	Jackson Labs	RRID: IMSR_JAX:027894
Recombinant DNA
FN1 shRNA plasmid for lentiviral production	Thermo Scientific	TRCN0000293840
piggyBac-rtTA (4th_Gen)-NGN2–2A-PURO-IRES-SNAP	Addgene	Cat# 209077
PB_iETV2_P2A_GFP_Puro	Addgene	Cat# 168805
Scramble shRNA plasmid for lentiviral production	Millipore-Sigma	Cat# SHC016
Software and algorithms
CellProfiler (v4.2.8)	Stirling et al.114	https://cellprofiler.org/
CellRank2 (v2.1.0)	Weiler et al.115	https://pypi.org/project/cellrank/
CIBERSORTx	Newman et al.65	https://cibersortx.stanford.edu/
ClusterProfiler (v4.10.1)	Yu et al.116	https://guangchuangyu.github.io/software/clusterProfiler/
Condiments (v1.10.0)	Roux de Bézieux et al.117	https://github.com/HectorRDB/condiments
CytoTRACE2 (v1.1.0)	Kang et al.118	https://github.com/digitalcytometry/cytotrace2
DAseq (v1.0.0)	Zhao et al.119	https://github.com/KlugerLab/DAseq
DirichletReg (v0.7.2)	Maiermarco 120	https://github.com/maiermarco/DirichletReg
DoubletFinder (v2.0.4)	McGinnis et al. 121	https://github.com/chris-mcginnis-ucsf/DoubletFinder
enrichR (v3.4)	Chen et al.57	https://github.com/guokai8/EnrichR
Fgsea (v1.28.0)	Korotkevich et al.122	https://github.com/alserglab/fgsea
Fiji (v2.3.0)	Schindelin et al.123	imagej.net/software/fiji/
GraphPad Prism (v10.4.1)	GraphPad Software	graphpad.com
GSVA (v1.50.5)	Hänzelman n et al.124	https://github.com/rcastelo/GSVA
Harmony (v1.2.3)	Korsunsky et al.125	https://github.com/immunogenomics/harmony
Limma (v3.58.1)	Ritchie et al.126	https://bioconductor.org/packages/release/bioc/html/limma.html
MAST (v1.28.0)	Finak et al.127	https://github.com/RGLab/MAST
Mgcv (v1.9.3)	Wood128	https://cran.r-project.org/web/packages/mgcv/index.html
miloR (v1.10.0)	Dann et al.41	https://github.com/MarioniLab/miloR
Msigdbr (v10.0.2)	Dolgalev129	https://igordot.github.io/msigdbr/
NicheNet (v2.1.5)	Broweays et al.71	https://github.com/saeyslab/nichenetr
Nikon NIS Elements software	Nikon	microscope.healthcare.nikon.com
Python (v3.13.5)	Python	https://www.python.org
R (v4.3.2)	R Project	https://www.r-project.org/
R Studio (v2024.04.2)	Posit	https://posit.co/download/rstudio-desktop/
Scanpy (v1.11.4)	Wolf et al.130	https://pypi.org/project/scanpy/
Seurat (v5.2.1)	Hao et al.131	https://satijalab.org/seurat/
Slingshot (v2.10.0)	Street et al.132	https://github.com/kstreet13/slingshot
Other
96-well plate with cover glass thickness polystyrene bottom	Greiner Bio-One	Cat# 655096
Ti2E Ax R Confocal Microscope	Nikon	microscope.healthcare.nikon.com

Mouse models

All experiments were performed in accordance with the Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at the Icahn School of Medicine at Mount Sinai (ISMMS). For all mouse experiments, C57BL/6J background mice with their endogenous murine Apoe replaced with human APOE3/3 or APOE4/4 were used. Aged, retired breeder APOE-TR mice were purchased from Jackson Laboratories (APOE3/3-TR strain number: 029018; APOE4/4-TR strain number: 027894). All mice used in APOE3/3- and APOE4/4-TR comparisons were female, and ages ranged from 8-10 months. Mice used in the TGF-β inhibition experiment had equal numbers of males and females in both treatment and control groups (n = 6 per sex) and ages ranged from 12-13 months. All mice were healthy at the time of experimentation and had not undergone prior experimental treatments or procedures other than breeding where applicable. Mice were housed in the ISMMS vivarium under standard husbandry conditions with ad libitum access to food and water until experimentation.

Human postmortem brain tissue

Human postmortem hippocampal tissue samples for immunofluorescence analyses were previously described²⁵. Donors were of European ancestry with a mean age of 88.7 ± 3.2 years. Samples were grouped for analysis according to APOE genotype. No statistical methods were used to predetermine sample size. Donor sex information was not available for these samples.

METHOD DETAILS

Cerebrovascular atlas assembly and analysis

Processing and subject selection of individual snRNA-seq studies

Datasets from the following studies were downloaded and processed following the original methods as a guide. These studies were selected for their inclusion of APOE genotype information and complementary strategies to enrich vascular cell populations that are typically underrepresented in single-cell transcriptomic datasets. All analysis was performed in R with the Seurat package¹³¹ unless otherwise specified.

Yang et al., 2022:

Raw data from hippocampal samples were downloaded from GSE163577. To pass quality control, nuclei had to meet the following criteria: (1) more than 200 features, (2) less than 5000 features, and (3) less than 5% mitochondrial RNAs. Doublets were filtered out using DoubletFinder¹²¹. The standard Seurat pipeline of NormalizeData(), FindVariableFeatures() (nfeatures = 2000), and ScaleData() was performed. The top 30 principal components (PCs) were used for UMAP generation at a resolution of 0.2. Harmony was used for batch correction¹²⁵. Clusters were annotated using the marker genes defined in the source publication. Clusters annotated as astrocytes, endothelial cells, and mural cells were then extracted for generation of the harmonized cerebrovascular atlas.

Sun et al., 2023:

Processed and clustered data was downloaded from the source publication. Only individuals with APOE3/3, APOE3/4, and APOE4/4 genotypes were included. Nuclei labeled as endothelial cells, smooth muscle cells, and pericytes were extracted for construction of the cerebrovascular atlas.

Haney et al., 2024:

Raw data of AD samples was downloaded from GSE254205. Nuclei that passed the following quality control parameters were included for further analysis: (1) more than 500 features, (2) more than 1000 reads, (3) less than 10% mitochondrial reads, and (4) less than 10% ribosomal reads. Doublets were removed using DoubletFinder. The standard Seurat pipeline of NormalizeData(), FindVariableFeatures() (nfeatures = 2500), and ScaleData() was performed. The top 20 principal components (PCs) were used for UMAP generation at a resolution of 0.2. Harmony was used for batch correction. Clusters were annotated using the marker genes defined in the source publication. Clusters annotated as astrocytes and vascular cells were extracted.

After extracting the cerebrovascular cell types from each dataset, propensity matching for APOE4 carriers and non-carriers was performed within each dataset to minimize confounding variables and equalize the number of individuals with each genotype. Specifically, the R package MatchIt¹³³ was employed to match individuals using the following parameters:

$$\mathit{\text{APOE}}~\mathit{\text{age}}+\mathit{\text{sex}}+\mathit{\text{AD status}}+\mathit{\text{total cell count}},\mathit{\text{method}}=‘\mathit{\text{nearest}}’,\mathit{\text{ratio}}=1$$

Cerebrovascular atlas integration and clustering

After propensity matching, data from the selected individuals were merged. The standard Seurat pipeline of NormalizeData(), FindVariableFeatures() (nfeatures = 2000), and ScaleData() was performed. Harmony was used for integration of the datasets with the following parameters:

$$RunHarmony\left(c\left(“subject”,“dataset”\right),theta=c\left(2,4\right)\right)$$

The top 15 principal components (PCs) were used for UMAP generation at a resolution of 0.5. Clusters with high expression of markers of two or more cell types were excluded from further analysis. After excluding a cluster, the Seurat pipeline and harmony integration was repeated. Ultimately, three rounds of clustering were performed and the final integrated UMAP was generated at a resolution of 0.2. Clusters were manually annotated using canonical marker genes for astrocytes, endothelial cells, pericytes, and smooth muscle cells (SMCs).

For mural cell subclustering, pericytes and smooth muscle cell clusters were extracted. FindNeighbors(), FindClusters(), and RunUMAP() were re-run with the top 15 PCs at a resolution of 0.2. Subclusters were annotated using previously described marker genes^38,39. Specifically, the marker genes used were Pericyte 1 (SLC1A3, SLC6A1), Pericyte 2 (ADAMTS1, ADAMTS9), SMC 1 (SLIT3, RGS6), and SMC 2 (TAGLN, MYH11).

Differential abundance analysis

miloR:

The miloR package was implemented in R⁴¹. For the entire cerebrovascular atlas, a k-nearest-neighbor (KNN) graph was generated using the buildGraph() function with parameters k = 50 and d = 15 Neighborhoods were generated with the makeNhoods() function with parameters prop = 0.1, k = 50, and d = 15 followed by calcNhoodDistance() with d = 15. Differential abundance between APOE4 carriers and non-carriers was assessed using the testNhoods() function with default parameters and the following design:

$$\mathit{\text{Design}}=~age+\mathit{\text{sex}}+AD\mathit{\text{status}}+\mathit{\text{dataset}}+\mathit{\text{apoe}}$$

For miloR differential abundance analysis of the mural cell subclusters, the same pipeline was used as above. For all above functions, Harmony embeddings were used as the dimensionality reduction input.

Dirichlet regression on all vascular cell types:

For each individual, vascular cell types were aggregated and the number of nuclei assigned to each cell type was counted. The proportion of each cell type was then calculated relative to the total number of vascular cells within that subject. The resulting subject-level composition matrix (cell-type proportions that sum to one for each individual) was modeled using the DirichletReg package in R¹²⁰. Cell-type proportions were used as the Dirichlet response variable and modeled as a function of APOE genotype, Alzheimer’s disease diagnosis, sex, age (modeled using a natural spline with 3 degrees of freedom), and dataset of origin. To account for differences in sampling depth across donors, models were weighted by the total number of cells contributed by each subject.

Dirichlet regression on mural cell subclusters:

For each individual, nuclei belonging to mural subclusters were counted and converted to proportions by dividing each subcluster count by the total number of mural cells for that subject. The resulting subject-by–mural subcluster proportion matrix (where proportions across mural subclusters sum to one per subject) was analyzed using Dirichlet regression with the DirichletReg package in R¹²⁰. Mural subcluster proportions were used as the multivariate Dirichlet response variable and modeled as a function of APOE genotype, Alzheimer’s disease diagnosis, sex, standardized age (z-scored), and dataset of origin. To account for differences in the number of mural cells profiled per donor, models were weighted by the total number of mural cells contributed by each subject.

Dirichlet regression on mural cell states:

Mural cells were classified into three transcriptional states based on expression of the canonical pericyte marker CSPG4 and the smooth muscle marker ACTA2. CSPG4 and ACTA2 expression values were extracted from the Seurat object metadata, and expression cutoffs were defined using the 1st percentile of non-zero expression for each gene across cells. Cells were then assigned to one of three mutually exclusive states:CSPG4_high/ACTA2_low, CSPG4_high/ACTA2_high, or CSPG4_low/ACTA2_high. Cells not meeting any of these criteria were excluded from downstream compositional analyses. For each individual, we aggregated cells by state to obtain subject-level counts and proportions, with proportions calculated relative to the total number of classified state-assigned mural cells per subject. Subject-level state compositions (three proportions summing to one) were modeled using Dirichlet regression¹²⁰, with state composition as the multivariate response and APOE genotype, AD diagnosis, and an APOE×AD interaction as primary predictors, adjusting for standardized age, sex, and dataset of origin. To account for unequal sampling depth across individuals, models were weighted by the total number of classified state-assigned mural cells contributed by each subject.

DASeq:

The DAseq package was implemented in R¹¹⁹. To compute APOE4 carrier vs.non-carrier differential abundance in mural cell subclusters, Harmony embeddings were used as input to DAseq along with APOE genotype labels for each subject. A range of neighborhood sizes (k.vector = seq(50, 500, 50)) was used to compute differential abundance (DA). Cells with DA scores above the 90^th percentile or below the 10^th percentile were classified as differentially abundant and visualized on UMAP coordinates. DA regions were identified using the getDAregion() function at a resolution of 0.5.

Binomial generalized linear model to estimate subject-level myofibroblast proportions:

For myofibroblast proportions, the number of cells in the SMC_2 DAseq region defined as myofibroblasts (Figure 1H) were counted for each individual. Myofibroblast proportion per individual was defined as the number of myofibroblast cells divided by the total number of SMC_2 cells. We modeled these subject-level proportions using a binomial generalized linear model (GLM) with a logit link in R (glm, family = binomial). Counts were represented in the response as a two-column matrix of successes and failures. The primary model tested whether the effect of APOE genotype differed by sex or AD diagnosis by including interaction terms for APOE×sex and APOE×AD, while adjusting for age and dataset of origin:

$$glm(cbind(myofibroblastcount,totalSMC2count-myofibroblastcount)~apoe*sex+apoe*ad+scale(age)+dataset,family=binomial())$$

Age was standardized prior to modeling (z-scored) so that effect sizes corresponded to a change per standard-deviation increase in age. Model diagnostics were assessed with the DHARMA package¹³⁴. Model coefficients for each for the variables were extracted and expressed as odds ratios with 95% confidence intervals. To report interpretable effect sizes, estimated marginal means were computed with the emmeans package and contrasts were performed on the logit scale, then exponentiated to obtain odds ratios (ORs) with 95% confidence intervals. In addition to reporting marginal main effects averaged over covariates (e.g., APOE4 vs APOE3), we quantified APOE-stratified contrasts within sex and within AD strata using Tukey-adjusted pairwise comparisons from emmeans for the APOE×sex and APOE×AD interaction terms.

Binomial GLM sensitivity analysis:

To assess robustness to potential dataset-specific effects, we performed a leave-one-dataset-out sensitivity analysis by refitting the model after excluding each dataset in turn (e.g., Sun+Yang only, Haney+Yang only, Haney+Sun only). Because the dataset variable is undefined once a dataset is removed and sample sizes are reduced, a simplified binomial GLM was used for the sensitivity analysis that omitted the dataset term and interaction terms while retaining the primary covariates (APOE genotype, sex, AD diagnosis, and standardized age). Individuals contributing fewer than 5 SMC_2 cells were excluded to improve model stability. For each dataset combination, the APOE4 versus APOE3 effect was summarized as an odds ratio with 95% confidence intervals derived from the fitted model.

Differential gene expression (DGE) analysis

Single-cell DGE analysis of mural cell subclusters:

Single-cell differential gene expression was performed with MAST¹²⁷ by using the Seurat FindMarkers() function. Pairwise comparisons between APOE4 carriers and non-carriers were performed for each mural cell subcluster. Age, sex, AD status, and dataset were included as covariates. A differentially expressed gene was defined as |logFC| > 0.25 and an adjusted p-value < 0.05.

Pseudobulk DGE analysis of SMC_2 cluster:

For the mural cell subclusters, gene expression counts were aggregated across cells for each subject using the AggregateExpression() function in Seurat, grouping by subject, APOE genotype, and mural cell type. The data was filtered to only include SMC_2 cells from individuals contributing at least 5 cells. A voom-limma¹²⁶ pipeline was then employed to model differential gene expression between APOE4 carriers and non-carriers. The following linear model was fit for each gene:

$$Y_{ij}={\beta }_0+{\beta }_1\cdot APOE{4}_j+{\beta }_2\cdot {\mathit{\text{age}}}_j+{\beta }_3\cdot {\mathit{\text{sex}}}_j+{\beta }_4\cdot {\mathit{\text{dataset}}}_j+{\beta }_5\cdot {\mathit{\text{ADstatus}}}_j+{\epsilon }_{ij}$$

In this model, Y_ij represents the voom-transformed expression value of gene i in subject j; APOE4_j is an indicator variable for APOE4 genotype (with APOE3 as the reference); age_j, sex_j, dataset_j, and ADstatus_j represent the subject-level covariates for age, sex, dataset, and AD status, respectively; β₀ is the model intercept; β₁–β₅ are the estimated coefficients for each predictor; and ε_ij is the residual error term. For each gene, statistics were calculated using the eBayes() function.

Myofibroblast marker genes:

DGE was performed with MAST by using the Seurat FindMarkers() function. Within the SMC_2 subcluster, a pairwise comparison was performed between the cells in the APOE4-enriched DAseq region (Figure 1H) and the remaining SMC_2 cells. A myofibroblast marker gene was defined as logFC > 0.25 and an adjusted p-value < 0.05.

Pathway analysis

Gene set enrichment analysis (GSEA) was performed using the fgsea R package¹²². The REACTOME database was input via the msigdbr R package¹²⁹. Following APOE4 carrier vs. non-carrier differential gene expression analysis via MAST or voom-limma (see differential gene expression analysis methods section), genes were ranked according to the following formula:

$$\mathit{\text{ranking score}}=sign\left({\mathit{\text{avg\_log}}}_{2}FC\right)\times -{log}_{\mathrm{10}}(p-\mathit{\text{value}})$$

Genes were then sorted in decreasing order and input into the fgsea package. Enrichment scores were calculated using a running-sum statistic across the ranked gene list. The normalized enrichment score (NES) represents the enrichment score scaled to a null distribution generated by gene-set permutations, correcting for gene-set size and enabling comparison across pathways. The enrichment score, normalized enrichment score, p-values, and FDR-adjusted p-values were computed for each pathway. Significant pathways were defined as an adjusted p-value < 0.05. Redundant pathways were merged to enhance readability.

For Gene Ontology (GO) enrichment analysis, differentially expressed genes were input into the ClusterProfiler R package¹¹⁶ using the enrichGO() function with ont = “ALL”. Redundant pathways were merged with the simplify() function. Significant pathways were defined as an adjusted p-value < 0.05.

Unbiased cell type annotation

Cell type annotation was performed with the CellMarker_Augmented_2021 database via the enrichR package^57,135. APOE4-enriched genes from the pseudobulk SMC_2 DGE analysis (logFC > 0.25, p < 0.05, see DGE analysis methods) or the myofibroblast marker genes (see DGE analysis methods) were input into enrichR with default parameters. Only cell types with a defined tissue of origin and at least 10 marker genes and 3 hits were included. Redundant cell types were merged to enhance readability.

Seurat module scores

Cellular contraction, extracellular matrix (ECM), and co-expression scores in mural cells:

Module scores were calculated for each cell using the AddModuleScore() function in Seurat. The following genes were used for each module:

Contraction: ACTA2, TAGLN, ITGA1, ITGA2, ITGA8, CDH11, PALLD, SORBS1, LMOD2, TPM2

ECM: COL4A1, COL6A2, COL1A1, COL3A1, COL1A2, COL5A2, COL14A1, FN1, FBN1, DCN

Co-expression: Minimum of the two scores for each cell

For comparison of module scores in APOE4 carriers vs. non-carriers, the data was filtered to only include SMC_2 cells from individuals contributing at least 5 cells. The module scores were then averaged across all cells per individual for subject-level comparisons between APOE genotypes.

TGF-β signaling in mural cells:

Mural cell subclusters were extracted and a module score was calculated for each cell using the AddModuleScore() function in Seurat. The following genes were used for the module:

TGF-β signaling: TGFβR1, TGFβR2, TGFβR3, TGFβ1, TGFβ2, TGFβ3, SMAD2, SMAD3, SMAD4

For comparison of TGF-β gene expression in APOE4 carriers vs. non-carriers, the module score was then averaged across all cells per individual for subject-level comparisons between APOE genotypes.

Myofibroblast gene activity scores

Gene set variation analysis was performed using the GSVA R package¹²⁴. The data was filtered to only include SMC_2 cells from individuals contributing at least 5 cells. Gene expression counts were aggregated across cells for each subject using the AggregateExpression() function in Seurat, grouping by subject and APOE genotype. Myofibroblast marker genes were downloaded from a previous publication (cluster 11)¹³⁶ and GSVA scores were assigned to each subject using the top 10 highly expressed myofibroblast marker genes in our dataset.

Pseudotime analysis

Lineage assignment:

Lineages were generated using the slingshot R package¹³². The UMAP of mural cell subclusters was input into slingshot with the Pericyte_2 cluster assigned as the start cluster.

Lineage bias:

For the lineages inferred by slingshot, each cell was assigned a pseudotime value along each lineage, as well as a corresponding curve weight. Curve weights represent the probability that a given cell belongs to a particular lineage when multiple branching trajectories are present. In this framework, cells located near a branch point may have fractional weights across lineages, whereas cells committed to a single trajectory have weights approaching 1 for that lineage and 0 for others.

After weight calculation, the R package condiments¹¹⁷ was then used to test for differential lineage assignment between APOE genotypes. Specifically, the differentationTest() function was performed for this anlaysis.

Generalized additive models (GAMs):

Three sets of module scores were assigned to mural cells using the AddModuleScore() function in Seurat with the following genes:

Pericyte signature score: PDGFRB, ANPEP, CSPG4, KCNJ8, SLC20A2, SLC6A1

Myofibroblast signature score: the co-expression score defined in the Seurat module scores methods section

TGF-β signaling signature score: the hallmark TGF-β signaling gene set

After assigning scores, separate GAMs with slingshot-derived lineage weights were fit for each module score with a smoothing parameter of k = 7 using the mgcv R package¹²⁸. To compare scores between lineages, the differences in areas under the curve (Δ-AUC: pericyte-to-myofibroblast – pericyte-to-pericyte lineage) was calculated for each GAM. Permutation testing (n = 1000) was then used to generate a null distribution of Δ-AUC values. A two-sided p-value was calculated as the proportion of permuted Δ-AUC values exceeding the observed Δ-AUC in magnitude.

NicheNet

The NicheNet R package was utilized for this analysis⁷¹. A sender-focused approach was performed. The DAseq APOE4-enriched myofibroblast region (Figure 1H) was assigned as the receiver cell type. The astrocyte, endothelial, pericyte_1, pericyte_2, SMC_1, SMC_2, and myofibroblast clusters were assigned as the sender cell types. Genes expressed in at least 10% of receivers or senders were used in the analysis, and the differential genes input into NicheNet for ligand prediction were the myofibroblast marker genes defined in the DGE analysis methods section.

Association of mural cell FN1 expression with AD phenotypes

Mural cells from the Haney (2024) dataset³⁷ were extracted from the cerebrovascular atlas and FN1 expression was averaged per individual. Mural cell FN1 expression was then correlated with age of AD diagnosis. A global cerebral pathology score was assigned to each individual by averaging the following metadata columns reported in the original dataset: sum_lb_density, PlaqueTotal, TangleTotal, infarct_cerebral_total_volume. Scores were then dichotomized into high (top 25%) and low (bottom 75%) pathology groups to compare mural cell FN1 expression between them.

FN1 gene expression in fibroblasts versus myofibroblasts

The original Yang and Sun snRNA-seq datasets each contained annotated fibroblast clusters. To enable direct comparison, we returned to these datasets and extracted fibroblast cells from the same subjects represented in our integrated cerebrovascular atlas, while myofibroblasts were defined as previously described and mapped back to the original objects by cell barcode. The datasets were harmonized and merged, and FN1 expression was pseudobulked by averaging log-normalized expression per subject within each cell type. Only subjects containing both fibroblasts and myofibroblasts were retained for analysis.

miBrain single-nuclei RNA-seq analysis

Dataset processing and clustering

Raw data was downloaded from a previous dataset generated in our laboratory. To pass quality control, nuclei had to meet the following criteria: (1) more than 500 features, (2) less than 6500 features, and (3) less than 10% mitochondrial RNAs. Doublets were filtered out using DoubletFinder¹²¹. The standard Seurat pipeline of NormalizeData(), FindVariableFeatures() (nfeatures = 2000), and ScaleData() was performed. The top 15 principal components (PCs) were used for UMAP generation at a resolution of 0.2. Harmony was used for batch correction¹²⁵. Clusters were annotated using canonical cell type marker genes. Clusters with high expression of markers of two or more cell types were excluded from further analysis. After excluding a cluster, the Seurat pipeline and harmony integration was repeated. Clusters annotated as astrocytes, endothelial cells, and mural cells were then extracted for downstream vascular cell analysis. For mural cell subclustering, FindNeighbors(), FindClusters(), and RunUMAP() were re-run with the top 15 PCs at a resolution of 0.05. Subclusters were annotated using previously described marker genes^38,39.

Differential gene expression (DGE) analysis

Single-cell DGE analysis of mural cell markers:

Single-cell differential gene expression was performed with MAST¹²⁷ by using the Seurat FindMarkers() function. A pairwise comparison between APOE4/4-enriched mural cluster 2 and APOE3/3-enriched mural cluster 1 was performed. A differentially expressed gene was defined as |logFC| > 0.75 and an adjusted p-value < 0.05.

Pseudobulk DGE analysis of mural cells:

Mural cell gene expression counts were aggregated across cells for each miBrain using the AggregateExpression() function in Seurat, grouping by sample and APOE genotype. A voomlimma¹²⁶ pipeline was then employed to model differential gene expression between APOE4/4 and APOE3/3 miBrains. The following linear model was fit for each gene:

$$Y_{ij}={\beta }_o+{\beta }_1\cdot APOE{4}_j+{\epsilon }_{ij}$$

In this model, Y_ij represents the voom-transformed expression value of gene i in subject j; APOE4_j is an indicator variable for APOE4 genotype (with APOE3 as the reference); β₀ is the model intercept; β₁ is the estimated coefficients; and ε_ij is the residual error term. For each gene, statistics were calculated using the eBayes() function.

Pathway analysis

See section under ‘Cerebrovascular atlas assembly and analysis.’

Seurat module scores

Pericyte and myofibroblast scores in mural cells:

Module scores were calculated for each cell using the AddModuleScore() function in Seurat. The following genes were used for each module:

Pericyte: Pericyte_1 marker genes from the human cerebrovascular atlas Myofibroblast: Myofibroblast marker genes from the human cerebrovascular atlas

TGF-β signaling in mural cells:

A a module score was calculated for each cell using the AddModuleScore() function in Seurat. The following genes were used for the module:

TGF-β signaling: TGFβR1, TGFβR2, TGFβR3, TGFβ1, TGFβ2, TGFβ3, SMAD2, SMAD3, SMAD4

For comparison of TGF-β gene expression in APOE4/4 vs. APOE3/3 miBrains, the module score was then averaged across all cells per miBrain for sample-level comparisons between APOE genotypes.

CytoTRACE2 and CellRank2

To estimate differentiation potential in miBrain mural cells, CytoTRACE2¹¹⁸ was implemented in R using default parameters. For trajectory inference, the miBrain mural cell object was converted to an AnnData object for downstream analysis in Python. A k-nearest neighbor graph was generated in Scanpy¹³⁰ from the Harmony embedding (X_harmony) using n_neighbors = 30 and the Gaussian kernel method (method = “gauss”). CytoTRACE2 scores were then inverted to represent pseudotime progression and incorporated into CellRank2¹¹⁵ using the PseudotimeKernel. Transition probabilities were computed using the soft thresholding scheme with nu = 0.3 and visualized by projection onto the UMAP embedding. The TGF-β signaling signature score for correlation analysis with cytoTRACE was assigned using the hallmark TGF-β signaling gene set as described in the human methods section ‘Pseudotime Analysis.’

Mouse single-cell RNA-seq analysis

Dataset processing and clustering

Raw data was downloaded from a previous study¹³⁷. To pass quality control, nuclei had to meet the following criteria: (1) more than 200 features, (2) less than 4000 features, and (3) less than 25% mitochondrial RNAs. Doublets were filtered out using DoubletFinder¹²¹. The standard Seurat pipeline of NormalizeData(), FindVariableFeatures() (nfeatures = 2000), and ScaleData() was performed. The top 20 principal components (PCs) were used for UMAP generation at a resolution of 0.2. Harmony was used for batch correction¹²⁵. Clusters were annotated using canonical cell type marker genes. Clusters with high expression of markers of two or more cell types were excluded from further analysis. After excluding a cluster, the Seurat pipeline and harmony integration was repeated. Clusters annotated as astrocytes, endothelial cells, and mural cells were then extracted for downstream vascular cell analysis.

For mural cell subclustering, FindNeighbors(), FindClusters(), and RunUMAP() were re-run with the top 15 PCs at a resolution of 0.2. Subclusters were annotated using previously described marker genes^38,39.

Differential abundance analysis

DAseq:

The DAseq package was implemented in R¹¹⁹. Differential abundance analysis was performed for aged iE4/Cre+ (>250 days old) versus all others. Harmony embeddings were used as input to DAseq along with APOE genotype labels for each mouse. A range of neighborhood sizes (k.vector = seq(50, 250, 50)) was used to compute differential abundance (DA). DA regions were identified using the getDAregion() function at a resolution of 0.05.

Temporal analysis of pericyte and myofibroblast abundance:

Cells in DA region 2 were labeled as myofibroblasts and cells in DA region 5 were labeled as pericytes. For those DA regions, the following binomial logistic mixed-effects model was fit separately:

$$logit(P(myofibroblast/pericyte))=genotype*age+sex+(1/mouse)$$

Differential gene expression (DGE) analysis

DGE was performed with MAST by using the Seurat FindMarkers() function. A pairwise comparison was performed between the cells in the aged iE4/Cre+-enriched DAseq region (DA region 2 in the source code, subsequently labeled myofibroblast) and all other mural cells. A myofibroblast marker gene was defined as logFC > 0.75 and an adjusted p-value < 0.05.

Pathway analysis

For Gene Ontology (GO) enrichment analysis, differentially expressed genes were input into the ClusterProfiler R package¹¹⁶ using the enrichGO() function with ont = “ALL”. Redundant pathways were merged with the simplify() function. Significant pathways were defined as an adjusted p-value < 0.05.

Unbiased cell type annotation

See section under ‘Cerebrovascular atlas assembly and analysis.’

Age analysis of APOE-TR mice

The mouse bulk RNA-seq data was downloaded from a previous study⁶⁶. For CIBERSORTx⁶⁵ deconvolution, the web-based application was used. Cells in DAseq region 2 and DAseq region 5 from the ‘Mouse single-cell RNA-seq analysis’ section were used as the reference. A signature matrix was generated using standard parameters, with a sampling ratio of 0.99 and inclusion of all genes. Absolute cell type scores were then estimated from the bulk RNA-seq gene expression data using default CIBERSORTx settings.

For each absolute score (pericyte or myofibroblast), age and genotype differences were tested using a fixed-effects ANOVA of the form:

$$Score~sex+genotype*agegroup$$

Post hoc comparisons between APOE3/3- vs APOE4/4-TR within each age group were performed using estimated marginal means (emmeans) with Tukey adjustment for multiple testing.

Myofibroblast gene activity in APOE-TR mice

Gene set variation analysis was performed using the GSVA R package¹²⁴. The processed and normalized 4-month-old mouse bulk RNA-seq data was downloaded from Synapse (syn26561824)⁶⁰. GSVA scores were calculated for each sample using the myofibroblast marker genes from our cerebrovascular atlas (see myofibroblast marker genes in the DGE analysis methods section, all genes with adjusted p-value < 0.05 were used). The human genes were converted to their mouse orthologs prior to GSVA input.

RNA-seq analysis of iPSC-derived mural cells

FPKM values were analyzed from a previous study²⁵. After log-transforming the FPKM values, differential expression analysis between APOE3/3 and APOE4/4 iPSC-derived mural cells (iMCs) was performed using the standard limma pipeline. Array weights were estimated using the arrayWeights() function and statistics were calculated for each gene using the eBayes() function with trend = TRUE and robust = TRUE. A differentially expressed gene was defined as |logFC| > 0.25 and an adjusted p-value < 0.05. Differentially expressed genes were input into ClusterProfiler for GO pathway analysis. Significant pathways were defined as an adjusted p-value < 0.05.

For CIBERSORTx⁶⁵ deconvolution, the web-based application was used. Mural cell subclusters from the cerebrovascular atlas were used as the reference. Specifically, 2,000 mural cells were randomly downsampled from the atlas, and their gene counts were extracted. A signature matrix was generated using standard parameters, with a sampling ratio of 0.5 and inclusion of genes expressed in at least 10% of cells. Relative cell type proportions were then estimated from the bulk RNA-seq gene expression data using default CIBERSORTx settings. Pericyte (Pericyte 1 and Pericyte 2) and SMC (SMC 1 and SMC 2) subcluster relative proportions were added to obtain overall pericyte and SMC proportions. For downstream analysis, cell type proportions in APOE4/4 samples were expressed as fold changes relative to the mean proportions in APOE3/3 samples.

Differentiation of iPSCs

Differentiation of human iPSCs into neurons

Protocol for neuron differentiation was adapted from Zhang et al.¹³⁸ and Lam et al.¹³⁹. In brief, Lipofectamine^™ Stem Transfection Reagent was used to transfect iPSCs with PiggyBac plasmids to confer doxycycline-inducible expression of the Neurogenin-2 gene (NGN2, Addgene Plasmid #209077). In short, dissociated iPSCs were seeded on day zero at ~104,000 cells/cm² onto plates coated with Geltrex^™. The seeded cells then received StemFlex^™ media containing 10μM Y27632 and 5μg/mL doxycycline supplements. On day 1, culture medium was replaced with Neurobasal N2B27 medium (Neurobasal, 1x B-27, 1x N-2, 1x MEM-NEAA, 1x GlutaMax, 1% penicillin-streptomycin) containing 10μM SB431542, 100nM LDN, 5μg/mL doxycycline, and 5μg/mL puromycin supplements. On days 3-6, daily medium changes were performed with Neurobasal N2B27 media containing 1μg/mL puromycin and 5μg/mL doxycycline supplements. On day 6, cells were treated with 0.5 μM Ara-C. On day 7, cells were lifted with accutase for integration into miBrains.

Differentiation of human iPSCs into astrocytes

Previously published iPSC-derived NPC (Chambers et al.)¹⁴⁰ and astrocyte (TCW et al.)¹⁰¹ differentiation protocols were followed to generate astrocytes. In summary, dissociated iPSCs were seeded onto Geltrex ^™ - coated plates at a cell density of 100,000 cells/cm² in pre-warmed StemFlex ^™ media containing 10μM Y27632 supplement. Cells received StemFlex ^™ every other day until cells reached 95% confluence. Once confluent, cells received NPC medium (1:1 DMEM/F12: Neurobasal Medium, 1x N-2 Supplement, 1x B-27 Serum-Free supplement, 1x GlutaMAX Supplement, 1x MEM-NEAA, 1% penicillin-streptomycin) containing 10μM SB43152 and 100nM LDN193189 supplements (day 0). From days 1 to 9, cells received daily media changes with the same media as day 0. On day 10, cells were split with Accutase and seeded onto plates coated with Geltrex ^™ and were grown in NPC media supplemented with 10 μM Y27632 and 20ng/mL bFGF. From days 11 through 13, cells received NPC media supplemented 20 ng/mL bFGF. On day 14, cells were split with Accutase and replated onto Geltrex ^™ -coated plates and received NPC media supplemented with 10μM Y27632 and 20ng/mL bFGF. From day 15 onward, cells received fresh Astrocyte Medium (AM, ScienCell) every 2-3 days and were passaged using Accutase upon reaching 90% confluence. From this point on, NPCs were differentiated into astrocytes over a 30-day time course. NPCs and differentiated astrocytes were then cryopreserved in freezing medium, which consisted of 90% knockout serum replacement (KSR) and 10% dimethyl sulfoxide (DMSO).

Differentiation of human iPSC into brain microvascular endothelial cells

Brain endothelial cells were differentiated according to adapted protocols outlined in Blanchard et al.²⁵, Qian et al.¹⁴¹, and Wang et al.¹⁴². In summary, doxycycline-inducible expression of the ETS variant transcription factor 2 (ETV2, Addgene Plasmid #168805) was conferred in iPSCs by transfecting iPSCs with a PiggyBac plasmid. Inducible ETV2-iPSCs were cultured until 60-70% confluence before they were dissociated with Accutase and seeded at 20,800 cells/cm² onto plates coated in Geltrex ^™ that contained StemFlex^™ supplemented with 10μM Y27632 on day 0. On day 1, culture medium was replaced with DeSR1 medium (DMEM/F12 with GlutaMAX, 1x MEM-NEAA, 1x penicillin-streptomycin) containing 10 ng/mL BMP4, 6μM CHIR99021, and 5μg/mL doxycycline supplements. On days 5 and 7, cell media was changed and replaced with hECSR medium (Human Endothelial Serum-free Medium, Gibco; 1x MEM-NEAA; 1 x B-27, 1% penicillin-streptomycin) that was supplemented with 50 ng/mL VEGF-A, 2μM Forskolin, and 5μg/mL doxycycline. On day 8, cells were lifted with Accutase and re-plated onto fresh plates coated with Geltrex ^™. Cells received hECSR media that was supplemented with 50 ng/mL VEGF-A and 5 μg/mL doxycycline. In the week following day 8, the supplemented day 8 medium was given to cells every 2-3 days to maintain cells until they were ready for tissue assembly in miBrains.

Differentiation of human iPSCs into mural cells

Mural cell differentiation protocol was based on published protocol from Patsch et al.¹⁴³. Dissociated iPSCs were seeded at 37,000 to 40,000 cells/cm² onto plates coated with Geltrex ^™ and grown in StemFlex ^™ containing 10μM Y27632 supplement on day 0. On day 1, StemFlex ^™ medium was replaced with N2B27 medium (1:1 DMEM/F12, Neurobasal media, 1x B-27, 1x N-2, 1x MEM-NEAA, 1x GlutaMAX, and 1% penicillin-streptomycin) containing 25 ng/mL BMP4 and 8μM CHIR99021 supplements. On days 3 and 4, cultures received a media change with N2B27 media supplemented with 10 ng/mL Activin A and 10 ng/mL PDGF-BB. On day 5, pericytes were lifted with accutase, plated at 35,000 cells/cm² onto plates coated in 0.1% gelatin, and grown in N2B27 containing 10 ng/mL PDGF-BB supplement. Cells received media changes with day 5 media every 2-3 days until confluent, at which point they were either frozen in freezing medium (90% KSR/10%DMSO) in liquid nitrogen tanks or split onto 0.1% coated gelatin plates in N2B27 media and maintained until ready for miBrain tissue assembly. For monoculture experiments, cells were plated onto 96-well μClear plastic-bottom plates (Greiner) coated with 0.1% gelatin at a density of 2500-5000 cells per well, and fixed 96 h post-plating unless specified otherwise.

Differentiation of human iPSCs into oligodendrocyte progenitor cells

At nearly 100% confluence, iPSCs in mTeSR1 with 10 μM Y27632 were plated on Matrigel (day -1). Every day until day 7, 2 mL of neural induction media (DMEM/F12 with 1x NEAA, GlutaMAX, 2-mercaptoethanol, and penicillin-streptomycin), with 10 μM SB431542, 250 nM LDN193189, and 100 nM RA (freshly added per use) was added. From day 8 through day 11, media was replaced daily with N2 medium (DMEM/F12 with 1x NEAA, GlutaMAX, 2-mercaptoethanol, and penicillin-streptomycin), along with 1x N2 and 100 nM RA and 1uM SAG (added fresh daily). On day 12, a cell scraper lifted the cells, which were gently transferred to a 6-well ultra-low attachment plate with 3 mL of OPC-N2B27 medium (DMEM/F12 with 1x NEAA, GlutaMAX, 2-mercaptoethanol, penicillin-streptomycin, 1x N2, 1x B27 without Vit A, 25 μg/=L insulin, and freshly added 100 nM RA and 1 μM SAG). About two-thirds of the OPC-NB27 media was changed every two days until day 20, without disturbance of the cell aggregates. From day 21 until day 30, PDGF medium (DMEM/F12 with 1x NEAA, GlutaMAX, 2-mercaptoethanol and penicillin-streptomycin, 1x N2, 1X B27 without Vit A, 25 ug/ml insulin, 100 ng/ml biotin, 1 μM cAMP, 5 ng/ml HGF, 10 ng/ml IGF-1, 10ng/ml NT3, and 10 ng/ml PDGFaa) was added every other day. On day 30, aggregates measuring 300-800 μm in size were picked with a p200 pipette and plated on PLO/laminin-coated 6-well adherent plates. Cells were plated at a density of 20 cells per well in PDGF medium. Until day 75, two-thirds of the PDGF media was changed every other day. PDGF media was then changed every 2-3 days, and the differentiated OPCs were enriched through FACS with a lineage-specific marker. OPC differentiation followed the protocol from Douvaras et al.¹⁴⁴.

Differentiation of human iPSCs into microglia

To generate microglia, iPSCs were first differentiated into hematopoietic progenitor cells (HPCs) using STEMdiff^™ Hematopoietic Kit (Catalog #05310). On days 12, 14, and 16 of the differentiation, floating HPCs were collected from media before freezing in CellBanker (AMSBIO). Following week 2 of miBrain assembly, HPCs were seeded into the existing tissue (25000 / 10 μL miBrain). miBrains were then maintained in miBrain media supplemented with 100ng/mL IL34 and 25ng/mL M-CSF. Microglia differentiation protocol was adapted from previous studies^145,146.

Assembly of 3D tissues

3D Tissue Assembly for miBrains

Neurons, endothelial cells, mural cells, and OPCs were lifted with Accutase, and astrocytes in TrypLE Select. Cells were then resuspended in their respective media, counted, and resuspended at 1 x 10⁶ cells/ mL. To assemble miBrains, a 15 mL Falcon tube for each miBrain condition was prepared to receive approximately 5 x 10⁶ neurons, 5 x 10⁶ endothelial cells, 1 x 10⁶ astrocytes, 1 x 10⁶ OPCs, and 1 x 10⁶ pericytes per 1 mL. Each tube of pooled cells was spun down for five minutes at RT at 200 x g. After carefully aspirating as not to disturb the cell pellet, each resulting pellet was placed on ice before resuspension in 1 mL Geltrex^™ ,10 μM Y27632 and 5 μg/mL doxycycline. Resuspension technique took care to avoid air bubbles, and cell pellets were kept on ice to prevent premature Geltrex polymerization and ensure successful seeding. 10 μL of encapsulated cell suspensions were then plated per well of a 96-well μClear plastic-bottom plate (Greiner). To allow the Geltrex^™ to polymerize, seeded plates were transferred into a 37 °C 95%/5% Air/CO2 incubator for 30 minutes. After polymerization, miBrain cultures were completely submerged in miBrain week-1 medium (Human Endothelial Serum-free Medium, 1x Pen/Strep, 1X MEM-NEAA, 1X CD Lipids, 1x Astrocyte Growth Supplement (ScienCell), 1x B27 Supplement, 10ug/mL Insulin, 1μM cAMP-dibutyl, 50μg/mL Ascorbic acid, 10ng/mL NT3, 10ng/mL IGF, 100ng/mL Biotin, 60 ng/mL T3, 50 ng/mL VEGF, 1μM SAG, 5 μg/mL doxycycline). Each well received 200 μL of media. Every 2-3 days, a half-media change was performed until day 8, before changing the media to miBrain week-2 medium (Human Endothelial Serum-free Medium, 1x Pen/Strep, 1X MEM-NEAA, 1X CD Lipids, 1x Astrocyte Growth Supplement (ScienCell), 1x B27 Supplement, 10ug/mL Insulin, 1μM cAMP-dibutyl, 50μg/mL Ascorbic acid, 10ng/mL NT3, 10ng/mL IGF, 100ng/mL Biotin, 60 ng/mL T3, 5 μg/mL doxycycline). After 4 weeks, miBrain cultures were considered ready for downstream assays.

3D tissue assembly for iPSC-derived endothelial cell monocultures

This protocol is the same as the miBrain protocol with the following differences. Each 10 μL Geltrex droplet contained 120,000 of only endothelial cells. Cultures were maintained in hECSR media supplemented with 50 ng/mL VEGF and 5 μg/mL doxycycline for two weeks until fixation (see immunofluorescence methods section).

Transduction of lentiviral shRNAs

FN1 and scramble MISSION lentiviral shRNAs were purchased from Sigma Aldrich. Lentivirus was generated in HEK293T cells following standard protocols. iPSC-derived mural cells were then transduced at a 1:40 dilution in mural cell culture media. Media was replaced 24 hours after transduction, and puromycin selection was initiated 72 hours post-transduction. Puromycin selection was maintained until cells were integrated into miBrains or collected for qRT-PCR validation of FN1 knockdown.

Amyloid and fibronectin treatments

3D iPSC-derived endothelial cell monocultures were generated as previously described. For the amyloid treatment group, 20 nM of both amyloid-β 1-40 and 1-42 were incubated in cell culture media for 1 hour at RT. For the amyloid and fibronectin treatment group, 20 nM of both amyloid-β 1-40 and 1-42 were incubated with 50 ng/mL fibronectin for 1 h in cell culture media at RT. Cells were treated with these mixtures for 96 h before fixing (see immunofluorescence methods section).

TGF-β ligand treatments

iPSC-derived mural cells were seeded onto 96-well plates as previously described. 24 h post-plating, cells were treated with 50 ng/mL TGFβ1 for 96 hours followed by fixation (see immunofluorescence methods section).

TGF-β inhibitor treatments

For 2D treatments, iPSC-derived mural cells were seeded onto 96-well plates as previously described. 24 hours post-plating, cells were treated with 10 μM SB431542 for 96 hours followed by fixation. For miBrains, treatments were either initiated at 3-4 days post-plating and maintained for 4 weeks until fixation or initiated at 2 weeks post-plating and maintained for the remaining 2 weeks until fixation, as specified in the figure legends. The concentrations used for miBrains were 10 μM SB431542 and 10 μM Galunisertib.

For TGF-β inhibition in APOE-TR mice, 12-month-old APOE4/4-TR mice were treated with the compound GW788388 for 4 weeks. Specifically, the mice were treated with 10 mg/kg of drug via daily intraperitoneal injections. The drug was prepared in a solution of 10% DMSO, 40% PEG300, 5% Tween-80, and 45% saline. This solution was also used as the DMSO vehicle for the control group. Following the treatment period, mice underwent cardiac perfusion as described in the ‘Immunofluorescence of mouse brain tissue’ section.

Immunofluorescence of cell cultures

2D cultures were fixed in 4% paraformaldehyde for 15 minutes at room temperature, rinsed with PBS and permeabilized in 0.1% Triton-X100 in PBS for 10 minutes. Cultures were then blocked in 5% normal donkey serum in PBS for 2 h. Primary antibodies diluted 1:500 in blocking solution were added to cultures and incubated overnight at 4°C, before 3 consecutive rinses with PBS for 10, 20, and 30 minutes. Cultures were then incubated for 2h at room temperature with secondary antibodies diluted 1:1000 in blocking solution. After rinsing once with PBS for 10 min, cultures were incubated for 20 min in Hoechst 33258 (Sigma-Aldrich, Cat. no. 94403) diluted at 1:1000 PBS. Cultures were then rinsed once more with PBS for 30 min. The PBS was then replaced prior to imaging. See corresponding image acquisition and quantification section.

3D cultures and miBrains were incubated overnight at 4°C in blocking solution (0.3% Triton-X100, 5% normal donkey serum, in PBS). Primary antibodies, diluted 1:500 in blocking solution, were added the following day and left for 2-3 nights at 4°C. Cultures received 5 x 30 min rinses with 0.3% Triton-X100 PBS before adding secondary antibodies and Hoechst. Both were diluted at 1:1000 in blocking solution and cultures were left to incubate at 4°C for 1-3 nights. Cultures were rinsed for 5 x 30 min with 0.3% Triton-X100 PBS, before rinsing once more and leaving cells in regular PBS for image acquisition. See corresponding image acquisition and quantification section.

Immunofluorescence of mouse brain tissue

Mice were first anaesthetized via gaseous isoflurane exposure, followed by cardiac perfusion with ice-cold PBS and subsequently 4% paraformaldehyde (PFA). The brains were dissected out and post-fixed in 4% PFA at 4C overnight. Fixed brains were placed in 30% sucrose in PBS until the tissue had sunk and were then embedded in OCT. Slices were cut at a thickness of 40 μm using a Leica cryostat and stored at −20C in cryoprotectant solution until staining.

For free-floating immunofluorescence, sections were carefully transferred into a 24-well plate with fresh PBS and rinsed two more times. Slices were then blocked in a buffer containing 0.3% Triton X-100 and 5% normal donkey serum in PBS overnight at 4° C. Primary antibodies diluted in blocking buffer were added the following day for 2-3 nights at 4° C. Sections were washed 5 x 30 min in PBS, and secondary antibodies diluted in blocking buffer were added for 2 hours at RT. Slices were then washed 5 x 30 min in PBS and mounted on glass slides for imaging. For slices stained with lectin, lectin was diluted 1:200 in PBS as part of the first wash and left for 1 h. Hoechst was added 1:1000 in PBS as part of the second wash and left for 30 min. All images were acquired in the corpus callosum or overlying cortex and in approximately the same locations of each slice to minimize biasing the results due to differences in anatomical regions (see image acquisition and quantification methods section).

Immunofluorescence of human postmortem brain tissue

Formalin-fixed human postmortem brain tissue was sectioned at a thickness of 40 μm using a Leica vibratome. Sections were incubated overnight at 4°C with primary antibodies in PBS containing 0.1% Triton X-100 and 10% bovine serum albumin. Sections were then washed four times with PBS-T and incubated with Alexa Fluor-conjugated secondary antibodies (1:1,000). Following secondary antibody incubation, sections were washed four times with PBS. Autofluorescence was quenched using TrueBlack (Sigma-Aldrich) according to the manufacturer’s instructions. Sections were mounted onto slides and imaged by confocal microscopy using identical imaging parameters across experimental groups (see image acquisition and quantification methods section).

QUANTIFICATION AND STATISTICAL ANALYSIS

Images were acquired using a Nikon AX R confocal microscope. The same parameters were used for each image within an experiment, and the experimenter was blinded to the channels of interest or experimental condition during imaging. Where applicable, images underwent deconvolution using the Nikon AX R built-in software as part of a standardized acquisition workflow, with identical parameters applied across all images within an experiment.

Images were analyzed using scripts from the Nikon AX R built-in quantification software, CellProfiler¹¹⁴, or FIJI/ImageJ¹²³. Prior to quantification, images underwent rolling-ball background subtraction. The background radius was selected based on the imaging parameters for each experiment and applied uniformly to all images within that experiment. No visualization-only image processing was performed prior to quantitative analysis.

Representative images were additionally processed for visualization only, where appropriate, using linear brightness and contrast adjustment, cropping, and/or median filtering. These visualization-only processing steps were applied identically across all conditions within each experiment and were not used for quantitative measurements. A complete summary of image acquisition, quantitative analysis, and visualization workflows for each figure is provided in Table S25.

Statistical analysis was performed with GraphPad Prism or R and the tests used are specified in each figure caption.

Supplementary Material

1

aa. Data S1. Source data, related to Figures 1–6 and S2–S7

3

b. Table S1. Cluster markers of the major cerebrovascular cell types, related to Figure 1.

4

k. Table S10. Differential gene expression analysis in APOE4-enriched DAseq subregion vs. remaining SMC_2 cells via MAST, related to Figure 1.

5

l. Table S11. ClusterProfiler pathway enrichment analysis of differentially expressed genes in APOE4-enriched DAseq subregion vs. remaining SMC_2 cells from Table S10, related to Figure 1.

6

m. Table S12. Cell type annotation of APOE4-enriched DAseq subregion via enrichR, related to Figure 1.

7

n. Table S13. Cluster markers of the miBrain mural cell subtypes, related to Figure 1.

8

o. Table S14. ClusterProfiler pathway enrichment analysis of miBrain mural cluster 2 marker genes from Table S13, related to Figure 1.

9

p. Table S15. Pseudobulked differential gene expression analysis of APOE4/4 vs. APOE3/3 miBrain mural cells, related to Figure 1.

10

q. Table S16. Gene set enrichment analysis of APOE4/4 vs. APOE3/3 miBrain mural cell limma results from Table S15, related to Figure 1.

11

r. Table S17. Cluster markers of the aged iE4/Cre+ enriched SMC region, related to Figure 2.

12

s. Table S18. ClusterProfiler pathway enrichment analysis of aged iE4/Cre+ enriched region from Table S17, related to Figure 2.

13

t. Table S19. Cell type annotation of aged iE4/Cre+ enriched region via enrichR, related to Figure 2.

14

c. Table S2. Significant differentially abundant cell neighborhoods in APOE4 carriers vs. non-carriers for the major cerebrovascular cell types via miloR analysis, related to Figure 1.

15

u. Table S20. Differential gene expression analysis of APOE4/4 iPSC-derived mural cells (iMCs) vs. APOE3/3 iMCs via limma, related to Figure 3.

16

v. Table S21. ClusterProfiler pathway enrichment analysis of differentially expressed genes in APOE4/4 iMCs vs. APOE3/3 iMCs from S20, related to Figure 3.

17

w. Table S22. Differential gene expression analysis in fibroblasts of APOE4 carriers vs. non-carriers via MAST, related to Figure 4.

18

x. Table S23. Gene set enrichment analysis of APOE4 carrier vs. non-carrier fibroblast MAST results from Table S22, related to Figure 4.

19

y. Table S24. NicheNet ligand activities in APOE4 myofibroblasts, related to Figure 5.

20

z. Table S25. Image processing workflow for quantification and representative images, related to Figures 1–6 and S3, S4, S6, and S7.

21

d. Table S3. Cluster markers of the mural cell subtypes, related to Figure 1.

22

e. Table S4. ClusterProfiler pathway enrichment analysis of mural subcluster marker genes from Table S3, related to Figure 1.

23

f. Table S5. Differential gene expression analysis in mural cell subclusters of APOE4 carriers vs. non-carriers via MAST, related to Figure 1.

24

g. Table S6. Gene set enrichment analysis of APOE4 carrier vs. non-carrier MAST results from Table S5, related to Figure 1.

25

h. Table S7. Pseudobulked differential gene expression analysis in SMC_2 cluster of APOE4 carriers vs. non-carriers via limma, related to Figure 1.

26

i. Table S8. Gene set enrichment analysis of APOE4 carrier vs. non-carrier SMC_2 limma results from Table S7, related to Figure 1.

27

j. Table S9. Cell type annotation of APOE4 SMC_2 gene signature via enrichR, related to Figure 1.

Highlights:

• Single-nucleus analysis identifies an APOE4-enriched brain myofibroblast population

• APOE4 myofibroblasts emerge from pericytes

• APOE4 myofibroblasts drive cerebrovascular fibrosis and amyloid deposition

• TGF-β inhibition reverses the pericyte-to-myofibroblast transition

3. Data and code availability

This paper analyzes existing, publicly available data, accessible at GEO: GSE163577, GSE254205, GSE342463, GSE125869, and Synapse: syn51015750, syn22313650, syn20808171, syn26561824. Microscopy and qRT-PCR data reported in this paper will be shared by the lead contact upon request.

All original code has been deposited at Zenodo and is publicly available at DOI: 10.5281/zenodo.22102870 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.