Cross-disorder and disease-specific pathways in dementia revealed by single-cell genomics

SUMMARY

The development of successful therapeutics for dementias requires an understanding of their shared and distinct molecular features in the human brain. We performed single-nuclear RNA-seq and ATAC-seq in Alzheimer’s disease (AD), frontotemporal dementia (FTD), and progressive supranuclear palsy (PSP), analyzing 41 participants and ~1 million cells (RNA + ATAC) from three brain regions varying in vulnerability and pathological burden. We identify 32 shared, disease-associated cell types and 14 that are disease specific. Disease-specific cell states represent glial-immune mechanisms and selective neuronal vulnerability impacting layer 5 intratelencephalic neurons in AD, layer 2/3 intratelencephalic neurons in FTD, and layer 5/6 near-projection neurons in PSP. We identify disease-associated gene regulatory networks and cells impacted by causal genetic risk, which differ by disorder. These data illustrate the heterogeneous spectrum of glial and neuronal compositional and gene expression alterations in different dementias and identify therapeutic targets by revealing shared and disease-specific cell states.

In brief

Functional genomic analyses of multiple brain regions across three major neurodegenerative disorders involving tau pathology, AD, bvFTD, and PSP, at the single-cell level enables the identification of markers of neuronal vulnerability, glial states that vary across disease, and disorder-specific cellular differences in the expression and regulation of known-risk genes.

Graphical Abstract

INTRODUCTION

Alzheimer’s disease (AD), the behavioral variant of frontotemporal dementia (bvFTD), and progressive supranuclear palsy (PSP) are syndromes that involve different forms of tau pathology.^1,2 Combined, these three disorders affect over 28 million people worldwide.³ Disease-altering therapeutic options are lacking, and there are few ongoing clinical trials in PSP and FTD.⁴

A foundational observation in dementia is the presence of selective neuronal vulnerability, wherein neurodegeneration, tau pathology, and neuroinflammation impact specific cell types and brain regions in temporally and spatially distinct patterns.^5,6 Both AD and bvFTD have distinct patterns of cortical-layer-specific pathology.^7–9 In PSP, frontal and subcortical motor pathways are more vulnerable.^2,10 The genetic architectures of AD, PSP, and bvFTD are also distinct, highlighting differential contributions of neurons and glia to disease risk.^11–17 All three disorders manifest tau pathology that begins in vulnerable regions that differ across disorders.^2,10,18,19

Advances in single-cell sequencing have revealed candidate markers of selective vulnerability in AD, including regulators of neuronal transcription,^9,20–24 signatures of neurons containing hyperphosphorylated tau,⁷ and signatures of neural-immune activation, an early, persistent feature in dementia.^{8,21,23,25–27} Despite these advances, few studies have yet to formally compare across dementias, and much remains unknown, including the specificity of changes observed in AD compared with other disorders, their role in selective neuronal vulnerability, or glial diversity. We addressed these important gaps in knowledge through direct comparison of post-mortem brain at the single-cell level across three major disorders involving tau pathology, AD, bvFTD due to Pick’s disease, and PSP, using both single-nucleus RNA sequencing (snRNA-seq) and the single-nucleus assay for transposase-accessible chromatin (snATAC-seq)^28,29 to validate disease-associated cell states and their predicted regulatory drivers. Our design enabled identification of new markers of neuronal vulnerability, varying glial states across disease, and disorder-specific cellular differences in the expression and regulation of known-risk genes.

RESULTS

We compared three disorders collectively referred to as tauopathies that display different patterns of neuronal loss and glial activation^2,9: AD, bvFTD, and PSP, using snRNA-seq in conjunction with snATAC-seq (Figure 1A). Using a well-curated, neuropathologically characterized brain collection (STAR Methods), we selected three cortical brain regions with differential vulnerability to disease^10,19,30 and prospective, semi-quantitative ratings of pathological burden using a published grading scale³¹ (see STAR Methods). The distribution of pathology scores matched the expected distribution of pathology in each disorder based on published annotations (Figure 1A; Table S1). For example, the insular cortex (INS) showed the highest disease burden in bvFTD¹⁹ and moderate burden in PSP and AD. In all three disorders, the motor cortex (primary motor cortex [M1]; BA4) had comparable, moderate levels of tau pathology, which in PSP was the highest of the three cortical regions profiled, as expected.¹⁰ In contrast, the primary visual cortex (V1) had low levels of tau pathology in all three disorders, and we hypothesized that it would have a higher expression of resilience factors compared with more vulnerable regions.

Figure 1. Comparison of cell types, subclasses, and disease states across brain regions with variable disease vulnerability across neurodegenerative tauopathies.

(A) Schema depicting cross-disorder analysis of brain tissue from AD, bvFTD, PSP, and controls in INS, BA4, and V1 (1.4 M cells pre-QC; RNA + ATAC) to define changes in cellular and molecular composition and gene regulatory networks (GRNs), including (left) cartoon with heatmaps showing, by region and disorder, neuropathology scores averaged across subjects for neurodegeneration (blue, below) and tau (red above) (Table S1). FTD: abbreviation for bvFTD used interchangeably throughout.

(B) snRNA-seq clusters separating into 9 canonical cell types by condition (AD, bvFTD, PSP, and control; left top) and brain region (V1, INS, and BA4; left lower) (≈590,000 cells, 101 samples post-quality control [QC] snRNA-seq; Table S1).

(C) Unsupervised re-clustering of excitatory neurons from one brain region (BA4) across all disease conditions to identify disease-associated subclasses and states.

(D) Pie charts showing cell-type distribution of clusters with differential composition in multiple diagnosis groups (upper), or one diagnosis group (lower), and table (right) listing clusters with distinct compositional changes (STAR Methods; Table S4; EX, excitatory neuron; IN, inhibitory neuron; OL, oligodendrocyte; AST, astrocyte; OPCs, oligoprogenitor cells; MICs, microglia; ENDs, endothelial cells).

See also Figures S1, S2, and S3.

We generated 880,000 single-nuclear mRNA expression profiles from 120 brain samples representing 41 subjects (see STAR Methods), with three cortical regions from each (Figures 1A and S1; STAR Methods). Following stringent quality control (QC) and outlier removal (STAR Methods), ≈590,000 high-quality mRNA profiles remained (Table S1; STAR Methods). Confounding variables of age, sex, post-mortem interval, and RNA integrity (RIN) were not significantly different between control and subjects (p < 0.05, Wilcoxon; Figure S1A; Table S1). Reference-based analysis was used (STAR Methods; Figures 1B and S1B; Table S2) to establish conserved clusters with standardized nomenclature.^32,33 We identified ninecanonical cell classes (Figures 1B and S1B–D), excitatory neurons (EXs), inhibitory neurons (INs), astrocytes (ASTs), oligodendrocytes (OLs), oligoprogenitor cells (OPCs), microglia (MICs), endothelial cells (ENDs), and pericytes, as well as peripheral lymphocytes,^34–37 and 24 canonical subclasses that overlap with established reference data^32,33 (Figure S1B). We leveraged our multi-region, multi-disease design to identify both shared and distinct disease-associated cell states in a stepwise manner (Figures 1A–1C).

Molecular taxonomy of CNS cell types across brain regions in dementia

We next re-clustered the 8 major CNS cell types (STAR Methods) in each of the three brain regions, identifying 178 clusters from 4 conditions (3 disorders and controls, excluding lymphocytes for low abundance; STAR Methods; Figure 1C; Table S2). We then performed hierarchical clustering to define an unbiased nomenclature based on the condition, region, cell type, and cluster number that distinguishes each cluster (Figures S2 and S3). For example, an excitatory neuronal cluster from BA4 was labeled BA4_EX-numeric, ending with a unique numeric identifier assigned based on relative cluster size (BA4/V1/INS_EX/IN/OL/AST/OPC/MIC-0–15; Figures S2 and S3; STAR Methods). The first level of clustering was driven predominantly by cell subclass, such as parvalbumin interneuron (Pvalb) (Figure S2), whereas the next branch represented brain region; 94% of clusters were identified in multiple brain regions based on their shared gene expression patterns (Figures S2 and S3).

The 33 excitatory and 26 IN clusters represented all canonical neuronal subclasses in the reference data³² (Figures S1B and S2; Table S2). Oligodendroglia clusters belonged to the three subtypes of OL: early myelinating BCAS1+ OL³⁸ and two more mature OL expressing either high PLP1 or RBFOX1, separated into 39 clusters based on brain region and state (Figure S3A). We identified two subclasses of immature and differentiating OPC, clustered into 17 state- and region-dependent groups^39–41 (Figure S3B). The 27 AST clusters were divided across two canonical subtypes, protoplasmic and fibrous ASTs (Figure S3C). The 18 microglial clusters were grouped into six transcriptomically distinct states (Figures S3D and S3E). Reassuringly, each of these six transcriptomically distinct microglial states overlapped significantly with those previously profiled from the fresh human brain²⁶ (Table S3). Among these, five overlapped significantly with MIC clusters previously identified in the frontal cortex in AD^36,42 (Table S3). The 13 clusters of ENDs were divided into two major groups marked by either higher expression of immune signaling genes or genes involved in angiogenesis and endothelial maintenance (Figure S3F).^43,44 This systematic classification and annotation of 178 cell states represented neuronal and glial cell types and subclasses (Figures S2A, S2B, and S3A–S3G), their regional localization across three human brain regions, their distinguishing marker genes, associated biological pathways, transcriptional regulators, and differential gene expression (DGE) across disorders, summarized in Tables S2, S4, S5, and S6.

We first identified genes that were significantly differentially expressed (DE) in major cell types across subjects by disorder (STAR Methods). It was notable that disorder-specific DE genes were a minority; the vast majority were shared by more than one disorder (89% of 5,933 genes with false discovery rate [FDR] < 0.05 and abs(log₂FC) > 0.1 had p < 0.05 in multiple disorders, cross-disorder linear mixed effects [LMEs]; Table S5; STAR Methods). The majority (91%; 621 genes) of genes DE in only one disorder were DE in only a single-cell type (Table S5). Notable examples include VPS54, increased in bvFTD (BA4-MICs) and SLCO1A2, a genome-wide association study (GWAS) locus in PSP⁴⁵ that is increased in PSP (INS-MICs) (Table S5).

Patterns of shared and distinct composition of transcriptomic cell states across brain regions in dementia

We next systematically characterized the shared and distinct disorder-associated changes of 178 cell clusters across brain regions and different dementias to characterize new molecular markers and drivers of neuronal vulnerability, neuroinflammation, and resilience. We used multivariate analysis (STAR Methods) to identify changes in abundance of cells or states across disorders (Figures 1D, 2A–2C, S2, and S3; Table S4).

Figure 2. Shared and distinct neuronal and glial disease states in disorders vs. controls.

(A) Stacked barplot distribution of disease-associated clusters by cell type.

(B and C) Heatmap showing compositional changes in each diagnosis group vs. controls of selected (B) neuronal and (C) glial subclusters (red, enriched in disease; blue, depleted in disease; log₁₀(FDR) × sign(log₂ fold change [FC]) of differential composition. Significance thresholds indicated by boundary thickness corresponding to FDR< 0.05 = thick border, and <0.1 = thin border (see STAR Methods and Table S4). Listed above are marker genes with putative functions shared by closely related clusters based on hierarchical clustering (Figures S2 and S3; Table S2) and below are cluster-enriched genes (Table S2). FDR corresponds to FDR-corrected p value unless otherwise indicated (STAR Methods).

(D) Characterization of shared depleted (AST-1) and enriched (AST-0) astrocyte clusters (top: barplot of differential composition by diagnosis group vs. controls in each brain region (*FDR < 0.1 [limma; all disease vs. control]; Table S4); below left: marker genes differentially expressed (DE) in INS_AST-1 or INS_AST-0 compared with other INS-AST (Table S2); below right: scatterplot of INS_AST-0 cell proportion vs. tau pathology score per sample, colored by disorder (Spearman’s correlation for PSP) (gold, n = 11; Table S4).

(E) AD-specific microglia state from motor cortex showing differential composition in AD vs. other samples (log₂FC, **FDR < 0.01 [limma], corrected over 4 BA4-MIC clusters), and (below) protein-protein interaction (PPI) network with direct PPI enrichment p value among genes upregulated in BA4_MIC-7 in AD vs. all other conditions (Table S5) highlighting AD disease genes (large circles) and functional categories (colored circles).

(F) ITMB immunostaining in microglia in AD brains compared with bvFTD brains (left, frontal cortex; right panel boxplot, unpaired Student’s t test, ***p = 0.0009, n = 4; see Figure S5D for ITM2B+ neuron image).

(G) Enrichment of AD GWAS variants among genes upregulated (up) in BA4_MIC-7 microglia (AD case vs. control), but not in genes downregulated (down), or in other BA4 microglia clusters (MAGMA software, —log₁₀(FDR) corrected over 8 comparisons shown).

See also Figures S2, S3, S4, S5, and S7.

Shared disease-associated cell states in dementias

We observed 49 subclusters out of 178 (27%) that showed differential composition in one or more diseases (FDR < 0.1), with 19 depleted and 30 enriched in one or more disease conditions, the majority of which were observed across multiple disorders (67%; n = 33 clusters; Figure 1D; Table S4). These shared disease-associated states include interneurons, ASTs, MICs, and OPCs that show robust changes in composition across all disorders and that have not been previously implicated in either AD, bvFTD, or PSP (Figures 2B–2D, S4A, S4B, S4E, and S4F; Tables S2 and S4). A summary of clusters changing across disorders and brain regions is presented in Figures 2A–2C and Table S4, demonstrating a wide spectrum of glial and neuronal subtypes that are similarly depleted or enriched in all disorders.

Shared changes in neuronal cell states

We were able to identify previously unknown disease-associated changes across multiple cell types and disorders (summarized in Tables S4 and S5). One such cellular change involved a population of white matter interneurons marked by MEIS2/ADAMTS19 in the human brain³² (Figures 2B and S4A). These interneurons were split into two clusters in the insula, one composed predominantly of control cells (INS_IN-6) and one enriched for patient cells (INS_IN-0) (Figure S4A). The disease-enriched MEIS2 neurons (INS_IN-0) showed transcriptional evidence of cell stress and injury response, including downregulation of DNA repair genes⁴⁶ and upregulation of genes involved in protein folding and amyloid (Figure S4A). This reflects a previously unrecognized change in state between disease and controls.

Another salient example involved APOO, previously reported as upregulated across multiple regions in AD,⁴⁷ which marks a specific population of layer 2/3 EXs marked (V1_EX-3) upregulated in the visual cortex in AD, bvFTD, and PSP (Figures 2B and S4B). The transcriptomic signature of APOO neurons from disease brain overlapped with that of induced pluripotent stem cell (iPSC)-derived neurons (iN) expressing MAPT mutations⁴⁸ (Figure S4B). We hypothesize that the APOO+ neurons represent an early, reactive neuronal disease state that is shared across disorders, given the relative mild pathological stage of neurons in the visual cortex across all disorders, the immaturity of the iN, and relative sparing of neurons in visual cortex across all disorders.^10,19,30

Shared changes in glial cell states

We also observed shared glial states that were distinct from those previously described in dementia (Figure 2C). Notable examples included a homeostatic protoplasmic AST state that was depleted across all disorders in multiple brain regions (INS_AST-1, BA4-AST-1, and V1-AST-1) and a related disease-enriched AST (INS_AST-0) expressing regulators of hypoxic response that was more abundant in the insula in all disorders and correlated with tau pathology in PSP (Figures 2D and S4C). Shared transitions involving OPCs were observed, including depletion of OPC populations marked by genes related to injury response and N-methyl-D-aspartate (NMDA)-directed migration (INS_OPC-3; Figures 2D and S4C). In addition, we observed one population of MICs, BA4_MIC-4 (Figures S4E–S4G), whose composition was correlated with higher tau pathology (TAU pathology cor = 0.57, p = 0.02) and likely represents dystrophic MICs, marked by upregulation of FLT and expression of senescence-associated genes (Figure S4E; Tables S2 and S5).

We also replicated changes in glia previously reported in AD, including increases in QDPR+ OL (INS_OL-7, BA4_OL-6, and V1_OL-4) and decreases in PDE1A+ OL (INS_OL-2), which were shared with bvFTD and PSP (Figure 2C; Tables S3 and S4). We observed reproducible downregulation of SLC1A3,^8,9,36,49 a marker of protoplasmic ASTs, across all disorders and regions (INS_AST, V1_AST, and BA4_AST) (Table S2). We observed a reproducible enrichment of SEMA3E+ OPC (INS_OPC-0) and depletion of GPC5+ OPC (INS_OPC-1) in multiple disorders³⁶ (Figures 2C and S4D; Tables S3 and S4). A population of MICs in INS was enriched in all disorders (INS_MIC-3) (Figures 2C, S3E, and S4F), which overlapped significantly with MICs described in the MS brain⁵⁰ (Figure S4H), suggesting that it is a generalized activated state. Other microglial states observed across disorders include depleted homeostatic states (Figures 2C, S3E, and S4F; Table S4). These various states represented: phosphatidylinositol 3-kinase (PI3K) loss of function (LoF) (INS_MIC-0), apoprotein and WNT LoF (INS_MIC-1), PPAR-receptor agonist with WNT and PI3K pathway dependence (INS_MIC-3), or V type ATPase LoF with WNT LoF (INS_MIC-11) (Figure S4G), replicating microglial responses previously reported in AD^21,23,25 and other conditions.⁵⁰

Disorder-specific changes in cell composition in dementias

Next, we analyzed clusters changing in a single disorder and region to leverage the differential regional vulnerability across disorders. We identified three AD, eight PSP, and six bvFTD clusters with disease-specific trends (Figure 1D; Table S4).

AD-specific compositional changes in neuronal and glia cell states

We first assessed the extent to which our findings from AD reproduced previously reported cell states in AD. Reassuringly, we identified previously reported patterns of cellular responses to AD, including changes involving EXs, interneurons, ASTs, OLs, OPCs, and MICs in both AD datasets (Figures S4H, S4I, and S5A; Table S3). We validated robust AD-associated cell states and clarified their variable expression across brain regions with differing degrees of AD pathology and across AD, bvFTD, and PSP (Figures 2B, 2C, S4I, and S5A; Table S4). This included AD-associated depletion of a cluster of layer 5 EXs marked by RORB and NEFM (BA4_EX-4),⁹ which was not observed in the other two disorders and thus was likely specific to AD (Figure S4I). Among interneurons, we observed changes in DPP10+ Pvalb INs (INS_IN-2 [significant] and INS_IN-10 [trending]), consistent with previous findings in AD³⁶ (Figures 1D and S5A; Table S4). We also observed AD-specific DE genes in MICs, including PTPRG and IL15 in BA4-MICs and INS-MICs, respectively (Table S5), consistent with published observations.⁵¹

An AD-specific amyloid-associated MICs marked by ITM2B

We observed changes in the motor cortex, a region relatively resilient in AD, which differed compared with MICs observed in regions with more advanced pathology. DE genes in MICs cluster BA4_MIC-7 (Figure 2E; Table S5) overlapped with a MICs subtype enriched in AD risk genes and associated with aging⁵² (Figure S5B) and with amyloid-plaque-associated MICs described previously⁵³ (Figure S5B). But these BA4_MIC-7 MICs lacked expression of CD163, LRRK2, FOXP1, or SPP1 seen in the plaque-associated MICs observed previously.^8,21,25,42 Enriched pathways in BA4_MIC-7 included amyloid processing, cellular buffering of oxidative stress, and chaperone-mediated autophagy (Figure 2E), consistent with their role in resilience. They also displayed upregulation of ITM2B (Figure S5C), whose expression may be beneficial, especially considering that mutations in ITM2B cause a dominantly inherited AD-like dementia.^53–55 Immunohistochemistry (IHC) validated that ITM2B was more abundant in AD than bvFTD in MICs (Figures 2F and S5D). Linkage disequilibrium (LD)-score regression of AD GWAS demonstrated that the upregulated genes in this AD-associated MICs were significantly enriched for common genetic risk variants (Figure 2G; STAR Methods). The expression of candidate resilience factors, such as ITM2B and enrichment of genetic risk for AD in genes expressed in BA4_MIC-7 in AD, supports their potential as therapeutic targets.

PSP- and bvFTD-specific compositional changes in neuronal and glia cell states

AST depletion in PSP

Cell proportions of the 8 canonical CNS cell classes were largely similar across disease conditions and matched reference data³² (Figure S1D). A notable exception was ASTs in PSP visual cortex, which were significantly depleted (Figure 3A). We validated this finding in three ways. First, to assess contamination by misclassified cells or ambient RNA, we applied a more stringent filter to remove ambiguous cells and used published markers with bootstrap analysis, which confirmed our results (STAR Methods; log₂FC = −1.29 PSP vs. control, FDR = 0.023; Table S4). Second, we reclassified cells using reference-based mapping in Azimuth, confirming AST depletion in PSP (log₂FC = −2.04, FDR = 0.07). Lastly, to ensure that this was not an artifact of snRNA-seq, we performed bulk RNA-seq and deconvoluted cell proportions using two algorithms, Bisque (Figure 3B) and CIBERSORT (Figure S5E; STAR Methods^,56,57). This AST depletion was not correlated with differences in local tau pathology, which was minimal in V1 (Table S1).

Figure 3. Disease-specific neuronal and glial states.

(A) Decreased astrocyte cell count proportions by subject in PSP for V1 astrocytes (controls n = 7, PSP n = 6, AD n = 7, FTD n = 6; FDR [limma] over 72 comparisons of 3 diagnosis groups, 3 brain regions, and 9 cell classes; Table S4).

(B) Deconvoluted cell proportions based on bulk RNA-seq (Bisque, Wilcoxon p value; controls n = 7, PSP n = 8; FDR corrected over 21 conditions).

(C) Heatmap showing DE genes in PSP V1-AST (log₂FC, linear mixed effects model [LME], Table S5).

(D) Chromatin accessibility peaks at the GFAP promoter of different AST subclusters based on snATAC-seq (C18–C25; see Figures S6A and S6B; coverage, ATAC signal range normalized by ReadsInTSS).

(E) Barplot showing differential chromatin accessibility in PSP vs. controls, but not other disorders, at astrocyte-specific hypermethylated regions (%mC peaks) in astrocyte cluster C22 and (below) at regions differentially methylated by cell type as indicated by the x axis (BA4 and INS, n = 8–11 per diagnosis group and brain region; loss of heterochromatin (red) and gain of heterochromatin (blue) interpreted as loss or gain of chromatin silencing at cell-type-specific hypermethylated sites.

(F) Heatmap of differential gene activity of REST and additional C22 marker genes compared with other AST snATAC-seq clusters across all samples (log₂FC of gene activity score, disease vs. controls, ArchR; Table S7).

(G) Chromatin accessibility peaks showing higher activity at the REST but not adjacent POL2RB promoter in C22 (coverage, ATAC signal range normalized by ReadsInTSS, across all samples).

(H) Differential composition of nuclei in INS_EX-5 from bvFTD vs. control, compared with PSP and AD (Table S4, *FDR = 0.06 adjusted [limma] for 3 disorders and 11 INS-EX clusters).

(I) DE of bvFTD/ALS or PSP risk genes in INS_EX-5 in bvFTD AD or PSP vs. controls (*Z score > 3, LME).

(J) OPTN protein staining of layer 2 excitatory neurons (TUJ1+) in the bvFTD insular cortex.

(K and L) (K) PPI network among genes upregulated in INS_EX-5 in bvFTD vs. all other conditions (Table S5) highlighting ALS/bvFTD risk genes (large circles) or enriched Gene Ontology (GO) (key). PSP-enriched neuronal state in visual cortex (V1_EX-2), showing (L) differential composition compared with control of V1_EX-2 in bvFTD, PSP, or AD (*PSP vs. all other samples, FDR < 0.05; Table S4).

(M) PPI involving genes enriched in layer 2/3/4 vs. 5/6 neurons (WNT3) and genes upregulated in V1_EX-2 neurons in PSP vs. controls (Table S5) showing PSP risk genes (large circles) and enriched GO (key).

(N) PSP risk gene expression in layer 5/6 EX compared with layer 2/3/4 EX from V1, shown as stacked barplot combining changes observed in each diagnosis (***FDR < 0.001, **FDR < 0.01, *FDR < 0.05, t-statistic, Table S5).

See also Figures S5, S6, and S7.

To consider the potential mode of PSP-specific AST depletion, we performed disorder-specific differential gene expression(DGE) analysis and found 55 genes DE in ASTs in the visual cortex from PSP cases (FDR < 0.1; Figure 3C; Table S5). REST, which suppresses neuronal gene expression in non-neuronal cells, was downregulated specifically in PSP and bvFTD ASTs in V1 (Figure 3C; Table S5). ASCL1, which is sufficient to drive non-neuronal cells⁵⁸ toward a neuronal fate, was upregulated (Figure 3B; Table S5). ASTs in PSP also upregulated genes that were relatively repressed in ASTs compared with other cell types, including neuronally enriched genes ATP2B3, ILDR2, ELOVL4, PCMT1, NSF, and CUX2,⁸ which showed increased expression trends (t-statistic > 2) in PSP ASTs (Table S5).

Consistent with the RNA-seq, analysis of snATAC-seq (STAR Methods; Figures 3D–3G) showed the highest gain of chromatin accessibility in regions typically hypermethylated (silenced) in PSP ASTs compared with other cells (Figures 3E and S6A–S6C). These changes suggested the potential at the level of chromatin for increased expression of genes normally not expressed in ASTs, which is what we observed. Of particular interest is AST ATAC cluster C22, which is depleted in PSP (t-statistic = −2.10, p-adj = 0.05; Table S7; STAR Methods) and typically marked by high gene activity at the REST promoter (Figures 3F and 3G). Together, these findings identify previously unrecognized changes in ASTs in PSP and suggest that relaxation of cell-restricted gene expression may relate to altered chromatin regulation, an observation that recently has been observed in AD.²² Because AST pathology is a defining feature of PSP, whether the leaky expression of neuronal genes and concurrent depletion of ASTs in PSP are causally related warrants further investigation.

EX clusters specifically enriched in PSP and bvFTD

We identified several clusters of EXs with significant disorder-specific enrichments, one in bvFTD (INS_EX-5; Figures 3H–3K) and three others in PSP, one of which we highlight (V1_EX-2; Figures 3L–3N). In bvFTD, we observed enrichment in INS_EX-5, comprising a layer 2/3 excitatory neuronal cluster in the insular cortex (Figure 3H; Table S4). INS_EX_5 also uniquely manifested increased expression of multiple bvFTD/ALS risk genes and optineurin, validated by immunocytochemistry (LME, t-statistic > 3; Figures 3I and 3J) and genes that function in associated pathways, including macro-autophagy and stress response (Figure 3K). This is notable because layer 2/3 is the most vulnerable cortical layer in this region, having the highest pathological burden.⁵⁹

The PSP-enriched neuronal cluster was marked by GRIA4/EPHA6/IL1RAPL1 localized to layer 2/3 EXs in the calcarine cortex (V1_EX-2) (Figure 3L; Table S4), a relatively spared region. This cluster overlapped with PSP-associated changes previously identified in bulk tissue from PSP⁶⁰ (Figure S7A), which we localized to a distinct PSP-enriched neuronal cluster. Pathway analysis of genes upregulated in layer 2/3 EX identified cholesterol biosynthesis, WNT signaling, and synaptic vesicle, including NSF, a PSP risk gene^14,45,61,62 (Figures 3M and 3N). These neurons in a relatively spared region, the visual cortex, appeared to be upregulating risk genes potentially related to the resilience of these cells.

Disorder-enriched glial states

We also observed several distinct glial states that are specifically enriched in one disorder, including AD-enriched MICs (described above, BA4_MIC-7), bvFTD-enriched OLs (INS_OL-14) and ASTs (BA4_AST-8), and PSP-enriched OLs (BA4_OL-11) (Figures 1D, 2C, and S7B–S7I; Table S4). One MICs population was most enriched in bvFTD in the motor cortex and to a lesser extent in PSP, but not AD (BA4_MIC-1; Figures S4F–S4I). These ATP2C1- and SORL1-expressing MICs uniquely upregulated pro-inflammatory signals, such as NAIP, STK3, and SERPINL1 in bvFTD cases^63,64 (Figure S7G; Table S5) and were abundant in BA4 and INS, vulnerable regions in bvFTD (Table S4).

Characterization of selectively depleted projection neurons in AD, bvFTD, and PSP

We next sought to understand factors driving the differential vulnerability of specific neuronal subtypes. We observed notable layer-specific depletion of disorder-specific neuronal subclusters in brain regions with moderate to high neuropathology scores (AD: BA4_EX-4; bvFTD: INS_EX-2; PSP: INS_EX-13; Figures 4A and S8A; Table S4). The depleted populations we identified matched known expected patterns of vulnerability based on prior neuropathological data,^9,59,65 but they have not been identified in PSP and have not previously been characterized at the molecular level. In bvFTD, the selectively depleted layer 2/3 insular neurons (INS_EX-2; Figures 4A, 4B, and S8A) expressed layer-specific markers CBLN2, CUX2, and RASGFR2 (Figure S8B).⁵⁹ In AD, the selectively depleted layer 5 intratelencephalic (IT) EX neuronal subpopulation (BA4_EX-4; Figures 4A, 4B, and S8A) expressed layer-specific marker genes TSHZ2, FOXP2, and IL1RAPL2 together with previously reported markers of AD vulnerability, including RORB and NEFM^9,66 (Figure S8B). We performed bulk RNA-seq and deconvolution across seven regions in the 40 subjects (237 samples post-QC; Table S1; STAR Methods) to assess changes in neuronal cell composition, validating a significant proportional loss of L2/3 IT neurons in FTD insular cortex (Figure S8C).

Figure 4. Cross-disorder comparisons of selectively depleted neuronal clusters identify RORB as shared repressor of disease-associated genes.

(A) Differential cluster composition by diagnosis group across all excitatory neuronal clusters (black, insula; gray, BA4; and white, V1). Clusters grouped hierarchically based on overlapping marker genes (Figure S2). Below each cluster, colored by disorder (AD in red, bvFTD in blue, PSP in yellow), is the differential composition score (STAR Methods; −log₁₀(p value) × sign (log₂FC) of each disease vs. all other samples [limma];Table S4).

(B) Left panel bar graphs: differential composition per disease (*FDR < 0.1, **FDR < 0.01, ***FDR < 0.001; corrected for two comparisons shown); right panel: volcano plots showing DE genes, comparing each disorder-specific depleted cluster with the non-depleted “matched” cluster sharing the most similar marker genes (see Figure S4C and Table S5).

(C) Overlap of DE genes in selectively depleted vs. matched clusters for each disease (log₂FC > 0.20, FDR < 0.05).

(D) IHC (left) and quantification (right) showing the depletion of KCNH7+ neurons in layer 5 (RORB high) neurons in motor cortex used for snRNA-seq (n = 7 per diagnosis) and in independent frontal cortex (n = 3 per condition) (p = 0.043 AD vs. control, p = 0.22 motor vs. frontal cortex, two-way ANOVA).

(E) Differential expression of RORB in INS_EX-2 (layer 2/3 IT neurons; *p adjusted < 0.05, FDR corrected over 3 disorders).

(F) High RORB immunostaining in layer 2/3 cortical neurons in INS in bvFTD.

(G) Differential RORB binding in layer 2/3 EX (ATAC INS_EX C2 subcluster C8) based on footprinting.

(H) Model of RORB repressing gene expression of NPTX2 in selectively depleted neurons.

(I) DGE of RORB relative to NPTX2 in INS_EX-2 neurons in disease vs. control (t-statistic, LME; Table S5).

(J) Chromatin accessibility peaks (Peaks) at the RORB binding site proximal to the NPTX2 promoter in INS_EX (ATAC cluster C2; coverage, ATAC signal range normalized by ReadsInTSS; Chr7: 98611427–98621427).

(K) PPI plots and p value (STRING) of bvFTD-specific RORB target genes, based on footprinting, also downregulated in bvFTD in INS_EX-2 compared with INS_EX-5 neurons (t < −2, LME, Table S5), highlighting enriched GO terms (key).

See also Figures S8 and S9.

In PSP, where cortical laminar vulnerability patterns have not been fully established,⁶⁷ we observed a selectively depleted population of layer 5/6 near-projecting (NP) neurons (INS_EX-13; Figures 4A, 4B, S8A, and S8B; Table S4). These NP neurons may innervate nearby subcortical structures with high tau pathology burden in PSP,¹⁰ including the striatum and globus pallidus. We also identified known PSP risk genes enriched in these depleted neurons, including RUNX2, MOPB, STX6, and EIFAK3^45,61 (STAR Methods; Figure S8D). Orthogonal analysis of candidate risk genes at GWAS loci (Figure S8E), including genes identified by high-throughput genomics¹⁴ (STAR Methods), further confirmed PSP risk gene enrichment in deep-layer neurons compared with superficial layer neurons, including RUNX2, KANSL1, ARL17B, MAPT, ASAP1, LINC02210-CRHR1, and SP1 (13/15 risk genes expressed significant at FDR < 0.05, LME, Table S5). These analyses support a causal link between PSP genetic risk and this vulnerable cell population (INS_EX-13) identified by cross-disorder differential cell composition analysis (Figures 4A, 4B, S8A, and S8B; Table S4).

Markers of depleted/vulnerable neurons

We next compared markers of selectively depleted neurons across disorders and found examples of both disorder-shared and disorder-distinct genes (Figure 4B). GRM8⁶⁸ and GPC6 were shared markers of PSP- and bvFTD-depleted neurons, which we validated in bvFTD by immunocytochemistry (Figure S9A). KCNH7, OPCML, PDE1C, and NLGN1 were notable as shared markers of selectively vulnerable neurons in bvFTD (INS_EX-2, layer 2/3 IT), AD (INS_EX-4, layer 5 IT), and PSP (INS_EX-13, layer 5/6 NP) (Figure 4C). PDE1C and NLGN1 are known drivers of neuronal vulnerability to toxic stress^69,70 and OPCML and KCNH7 (Kv11.3) correlate with resilience in mouse.⁷¹ We extended these observations to human brain and further validated that KCNH7 marked vulnerable, depleted neurons in an independent single-cell dataset in AD⁹ and by immunohistochemistry in our cohort (Figures 4D and S9B), nominating KCNH7 as a potential therapeutic target. In contrast to the layer 5 IT neurons depleted from AD samples (BA4_EX-4), which expressed high levels of RORB at baseline, layer 2/3 neurons depleted in bvFTD do not typically express high levels of RORB. However, in bvFTD samples, the depleted, vulnerable layer 2/3 IT neurons upregulated RORB (Figures 4E and 4F).

Next, we conducted pseudotime analysis an alternative method to delineate gene expression trajectories that might reflect differential vulnerability among layer 2/3 IT EXs in insular cortex (STAR Methods). Depleted neurons from bvFTD (INS_EX-2) populated the trajectory’s start, whereas neurons enriched in disease samples (INS_EX-5) occupied the end (Figure S9C). Correlations between pseudotime and pathology scores in bvFTD cases showed a positive correlation trend with tau pathology (cor = 0.44, p = 0.17; Figure S9D), but not with neurodegeneration score. This aligns with the accumulation of tau pathology over disease progression and the precedent of distinct pools of tau-enriched and -depleted neurons observed in AD.⁷ Genes positively correlated with pseudotime included modifiers of protein pathology and/or neuronal survival (e.g., FYN, PDE1A, MAP3K5, VCP, TMEM106B, OPTN, CD47, BDNF, and NPTX2) (Figure S9E; Table S5). Conversely, negatively correlated genes were markers of depleted neurons (KCNH7, RORB, GRM8, and GPC6) (Figure S9E; Table S5); RORB was the most highly correlated with initiation of disease trajectory (Figures S9E and S9F). We further validated that RORB, KCNH7, PDE1C, NLGN1, and OPCML expression marked differentially vulnerable neurons using independent data from the SEA-AD gene expression trajectory viewer.⁸ These markers for vulnerable neurons across three tauopathies warrant further investigation as modifiers of resilience.

Finally, as a proof of principle to explore the utility of these data characterizing cell-type-specific vulnerability, we identified compounds predicted to reverse the bvFTD-disease signature of L2/3 IT neurons⁷² (STAR Methods). This analysis revealed that 14 of the top 35 compounds predicted to reverse signatures of L2/3 vulnerable, depleted, neurons involved mechanisms recognized for ameliorating cognitive deficits or disease pathology in cell or mouse models of AD or FTD, such as memantine (Table S5).

Transcriptomic drivers of shared and distinct disease-associated cell states

We next experimentally confirmed that RORB TF binding activity was increased in bvFTD samples (Figure 4G), by performing snATAC-seq in INS followed by TF footprinting in vulnerable layer 2/3 EXs from insula in bvFTD (STAR Methods; Figure 4G). RORB gene promoter occupancy was higher in bvFTD cases, but not in AD or PSP (Figure 4G). Of particular note was NPTX2, a candidate prognostic biomarker in AD²⁰ and bvFTD^73,74 (Figure 4H). NPTX2 expression was anti-correlated with RORB expression (Figure 4I) along with significantly reduced chromatin accessibility at its RORB binding motif (Figure 4J). The majority of RORB targets in bvFTD were downregulated in selectively vulnerable neurons (STAR Methods; 120 downregulated, 14 upregulated, FDR < 0.05; INS_EX-2 relative to INS_EX-5 in bvFTD). RORB-targeted genes formed a significant PPI network reflecting a coordinated stress response (enrichment p = 1.3E—5; STAR Methods; Figure 4K). Its downregulation here suggests a relative dampening of neuroprotective pathways (Figure 4K).

The identification of RORB as a potential driver of neuronal vulnerability in both AD and bvFTD suggested that analysis of additional TF-mediated drivers of disease-associated cellular states would be informative.³³ We used single-cell regulatory-network inference and clustering analysis (SCENIC)⁷⁵ (Figure 5A; STAR Methods) to identify cell-type-specific gene regulatory network (GRN) regulons for EXs, ASTs, OLs, and MICs and define the disease specificity of cell-type-specific regulons (Figures 5A and S9G–S9K; Table S6). We validated snRNA-seq-based SCENIC predictions of cell-type- and disease-specific regulon activity by snATAC-seq, assessing TF binding site enrichment (Figure 5A) and TF footprinting (STAR Methods), focusing our analysis on EXs (1,146,872 nuclei) and MICs (15,715 nuclei). We empirically defined 1,332 cell-context-specific regulons driven by 250 TFs, 65% of which were active in only one cell type, including TFs well known to be cell type specific, such as TBR1 and SCRT1 in EX (16 total unique), SOX9 and NFAT in AST (82 unique), and PRMD1, SPI1, TAL1, and IRF8 in MIC (57 unique) (Table S6). We highlight the 10 most specifically active TF regulons in each disease (Figures 5B–5D; Table S6). We confirmed increased chromatin accessibility at TF binding sites with high regulon specificity scores (Figures 5E, S9J, and S9K) and observed significant correlation across disorders between ranked differences in regulon specificity and chromatin accessibility (Figure 5E).

Figure 5. Transcription factor network inference identifies regulons active across cell types, disorders, and brain regions.

(A) Schema of strategy for cross-disorder comparison of TF activity within cell type and brain region and workflow for validation of SCENIC predictions using snATAC-seq and footprinting.

(B) Heatmap of relative regulon specificity score (RSS) ranks among top 25 TF regulons ranked for each disorder, cell type and brain region, highlighting distinct and shared GRNs.

(C and D) PPI and associated pathways enriched among GRN with disorder-specific differences in activity in INS-EX, including (C) YY1 and (D) DBP (p values from STRING; see Figure S9K and Table S6).

(E) Scatterplot of correlation between relative RSS determined from SCENIC (gain in rank vs. control) and differential accessibility of 11 TFs with greater activity in disease than control (Pearson’s correlation = 0.60, p = 0.00025; Table S6).

See also Figures S9 and S10.

Next, we identified combinations of regulons that reflected the majority of the gene expression difference associated with distinct disorder-enriched clusters (Figures S10A–S10D). We confirmed disorder-specific increase in chromatin accessibility at binding sites of RUNX1 in bvFTD and NR3C1 in AD (Figure S10E). Orthogonal experimental data from gene knockout studies⁷² allowed us to independently confirm that both RUNX1 and IKZF1 were functional drivers of the bvFTD-enriched ATP2C1 MICs signature (BA4_MIC-1; Table S6). Furthermore, validated IKZF1 regulon target genes were upregulated in bvFTD MICs relative to controls (Figure S10F). These data support a role for RUNX1 and IKZF1 in bvFTD-specific reactive microglial states observed in a brain region with moderate bvFTD pathology (BA4_MIC-1).

AD disease genes engage combinatorial TF programs regulating AD-specific MICs

Distinct clusters of TFs co-varied by diagnosis and brain region in microglia (Figure 6A), suggesting that distinct TF networks contribute to the diverse microglial transcriptomic states observed across disorders (Figures S3D, S3E, S7C, and S7D). One example was an AD-specific microglial TF network comprising SPI1, NR3C1, MXI1, and USF2, whose combined targets were responsible for a remarkable 60% of upregulated genes in the amyloid-associated, AD-specific microglial cluster, BA4_MIC-7 (OR 3.4–6.8, FDR < 0.05, Figures 6B and 6C; Table S6). Examining the combinatorial regulation of disease-specific microglial states both at the RNA and chromatin level (Figures 6B and 6C) revealed that SPI1 drives expression of a gene set involved in lysosome and phagocytosis, including multiple AD risk genes and the AD biomarker, IL15^76,77 (Figure 6B). We validated the AD-specific increase in IL15 promoter accessibility at the SPI1 binding site by footprinting (Figure S10G; Table S7). In contrast, genes regulated by another transcription factor, NR3C1, participate in amyloid processing⁷⁸ (Figure 6C), including ITM2B, the marker of BA4_MIC-7 in AD (Figure 2F), which we validated by footprinting (Table S7). This demonstrates that the effects of NR3C1 and SPI1 direct discrete pathways acting combinatorially to modulate the microglial neuroinflammatory state in AD.

Figure 6. Distinct GRNs drive disorder-specific microglial states in AD and PSP.

(A) Multidimensional scaling (MDS) plots showing distinct and shared microglial GRNs by brain region (STAR Methods).

(B) Diffferential gene expression (DGE) (disease vs. controls) of select SPI1 network genes in BA4 MIC (Table S5). Black boxes below indicate genes indicated to be bound by SPI1 in AD MIC based on footprinting (BA4 C7; Table S7).

(C) Combined GRN of TFs with increased activity in AD BA4_MIC-7 (USF2, NR3C1, MX11, and SPI1) based on gene overlap (Figures S10A, S10B, and S10D) with edge length proportion to (1-GRN) score, node color proportion to DGE in BA4_MIC-7 from AD vs. non-AD, and node border color indicating GO (key).

(D) MDS plots of AST GRNs in V1 (STAR Methods).

(E) Combined PPI (gray edges) and GRN (blue edges) plot of TFs with increased activity in PSP V1-AST (turquoise; Figures S10A and S10B; Table S6) with enriched GO terms. Enrichment p value from STRING.

(F) Immunohistochemistry and quantification of CUX1 staining in V1 AST of PSP vs. control (unpaired t test, *p = 0.023, n = 5). Arrows indicate astrocyte staining for CUX1. Barplot showing mean ± SEM. See Figures S11A–S11C.

See also Figures S10 and S11.

ASTs in PSP ectopically express neuronal TFs and disease-related pathways

As described above, a prominent pathological feature of PSP is astrocytic tau inclusions.² We observed that V1 ASTs in PSP downregulate REST (Figure 3C), and four neuronally-enriched TFs were uniquely active in V1 ASTs in PSP: CUX1, CUX2, ZMAT4, and FOXP1 (Figure 6D). The V1_AST-3 cluster, which was initially removed because of ambiguous marker gene expression (Table S1; STAR Methods), was of interest because it contained genes upregulated in PSP that significantly overlapped with these 4 regulons (Figure S11B). V1_AST_3 also showed PSP-specific upregulation of ZMAT4 (log₂FC = 0.24, FDR = 0.023, Table S5) and a broad trend of increased expression of the 4 TF targets (Figure S11C), including known Parkinson’s disease (PD) risk genes (PARK7 and SNCA)⁷⁹ and regulators of protein homeostasis, such as ATP6V0B and RHEB (Figure 6E). We confirmed that CUX1 predominantly stained neurons in control brain tissue, but robustly stains S100b-positive ASTs in PSP samples (Figures 6F and S11A).

MAFG-NFEL2L1 regulates PSP disease genes and neuroprotective pathways in EXs that are selectively dampened in PSP vulnerable neurons

We next interrogated GRN data for candidate TF regulators of differential neuronal resilience to tau pathology. Notably, this GRN analysis identified NFE2L1, a known buffer of cellular stress⁸⁰ (Figures 7A, S11D, and S11E). In the INS_EX-4 neurons, NFE2L1 was co-expressed with its cofactor MAFG, a combination that is a potent driver of proteostasis.⁸⁰ Genes downstream of NFE2L1 and MAFG regulons included genes with known neuroprotective effects, including SQSTM1 (Figures 7B and S11E; Table S6) and VCP, another ALS/bvFTD risk gene.⁸¹ Other components include the PD risk gene SCNA, and genes involved in the promotion of selective autophagy, proteostasis, RNA stabilization, and synaptic vesicle exocytosis (Figure 7B). We confirmed that VCP expression was positively correlated with MAFG and NFE2L1 expression across EX in the insular cortex (Figure 7C) and validated several MAFG targets by footprinting (STAR Methods), including UBE2N, KIF5C, RARB, and VCP (Figure 7B; Table S7). These findings suggested a model whereby the MAFG/NFE2L1 regulon protects neurons from tau aggregation (Figure 7D), which was supported by our observation that tissue containing more neurons with high regulon activity had significantly lower tau pathology scores (cor −0.49 p = 0.04; STAR Methods; Figure 7E). Moreover, neurons that were selectively depleted in one disease also lost MAFG /NFE2L1 regulon activity (Figure 7F), consistent with their roles in regulating resilience.

Figure 7. MAFG/NFE2L1 drives a resilience program.

(A) Cross-cluster comparison between INS-EX of MAFG GRN scaled activity score (SCENIC, Table S6), and expression of MAFG and NFE2L1 and their target gene, VCP, in all samples (*t-statistic > 2, LME; Table S2).

(B) PPI plot of the MAFG GRN from INS-EX (top 250 genes) including MAFG targets validated by TF footprinting (large circles) (Table S7) and enriched GO.

(C) Scatterplot showing differences across clusters in the relative expression of MAFG (top) and NFE2L1 (bottom) compared with VCP (% of cells where gene is detected; each dot represents one INS_EX cluster per disease group; Pearson’s correlation, p = 2.9e–13, n = 42).

(D) Model showing that MAFG/NFE2L1 drives a neuroprotective program in cells otherwise highly vulnerable to neurodegeneration, including VCP and PSP risk genes. Large circle = target validated by footprinting.

(E) Scatterplot showing inverse correlations between tau pathology score and proportion of MAFG/NFE2L1-high neurons (INS_EX-4; Pearson’s correlation and p value across all disorders shown at trendline, and within each disorder shown at key).

(F) Barplots comparing layer 5/6 NP neurons (INS_EX-13) in AD (spared) vs. PSP (depleted), showing the percent of cells expressing MAFG/NFE2L1, NFE2L2, VCP, or PSAP (left) and MAFG/NFE2L1 target genes (right), and also showing their widespread reduced expression in PSP.

See Figures S11D and S11E and Tables S6 and S7.

DISCUSSION

Our analyses strongly support the value of cross-disorder comparative analysis. Several studies have illustrated the power of single-cell RNA-seq to elucidate pathways dysregulated in AD.^7,8,20–25 But, whether these changes were specific to AD was not known. We demonstrate that several findings in AD are observed across disorders, identifying targets for therapeutic development. We annotated shared and distinct disease states comprehensively to provide an extensive resource, which we used to understand the molecular features and transcriptional drivers of selective vulnerability and resilience, including the role of disease-specific risk genes. By coupling snRNA-seq with ATAC-seq, we experimentally validated the bio-informatically predicted regulatory drivers of these altered cell states. We provide these data in the Synapse database and Cell 3 Gene application hosted by Chan Zuckerberg Initiative (CZI) to facilitate their browsing.

The molecular basis of selective neuronal vulnerability in neurodegeneration is relatively uncharacterized. We identified four genes that were reproducibly enriched in depleted EX neurons of multiple classes across all three disorders (bvFTD, layer 2/3 IT neurons [INS_EX-2]; AD, layer 5 IT neurons [BA4_EX-4]; PSP, layer 5/6 NP neurons [INS_EX-13]), including OPCML and KCNH7, which have also been identified experimentally in mouse models of AD.^71,73,74 To reduce the confound of survival bias, we compared molecular and cellular phenotypes within disease individuals, rather than via comparisons of disease and control. For example, comparing depleted neurons (INS_EX-2) vs. disease-enriched neurons (INS_EX-5) within bvFTD samples to define markers.

Other components of neuronal resilience across disorders included upregulation of MAFG/NFE2L1, whose relationship to vulnerability and pathological tau burden in dementias was not previously known. The relative upregulation of MAFG/NFE2L1 regulons in spared populations of EXs illustrates how cellular differences in injury response may determine vulnerability. This includes differentially affected neuronal subtypes that match expected patterns of selective vulnerability, such as those harboring gene regulatory programs driven by RORB in AD- and bvFTD-depleted neurons, and GRM8 in PSP- and bvFTD-depleted neurons. We were able to identify the specific molecular programs underlying superficial projection neuron vulnerability in bvFTD. We also identified an additional class of projection neurons, INS_EX-13, manifesting reduced MAFG/NFE2L1 activity, that was selectively depleted in PSP. These pathways that vary across spared and depleted cell populations become potential therapeutic targets, a model supported by our connectivity map analysis of gene expression changes in FTD L2/3 neurons.

In this regard, our cross-region, cross-disorder design complements previous studies characterizing cellular-molecular correlates with cognition and neuropathology across individuals within disorder.^8,20,65,71 We identified NFE2L1 proteostasis networks as a potential cross-disorder resilience factor. Similarly, we identified specific neuronal clusters associated with vulnerability, such as the depleted cluster of layer 2/3 IT neurons (INS_EX-2) in bvFTD insula, a particularly vulnerable region. Similarly, in PSP, the depleted neuronal cluster occurred among layer 5/6 NP neurons (INS_EX-13) projecting into highly affected subcortical regions. In AD, we found changes involving layer 5 IT neurons expressing RORB. Recent studies implicate IT neurons in residual cognitive changes that extend beyond what can be explained by AD neuropathology alone.⁶⁵ Here, we observe decreased KCNH7 and OPCML expression in layer 5 IT neurons in AD, consistent with their correlation with cognitive resilience across individuals.⁷¹

Glial subtype heterogeneity in the brain is increasingly appreciated, but how it corresponds to specific neurodegenerative conditions is not understood. We replicated a dystrophic microglial state in AD^26,42 and observed additional heterogeneity, including a unique amyloid-associated MICs state in AD with resilient features (BA4_MIC-7), a bvFTD-enriched state with upregulation of genes involved in sterile inflammation⁶⁴ (BA4_MIC-1), and a cross-disorder state expressing high levels of WNT signaling genes (INS_MIC-3). We also uncovered glial diversity that correlated with degree of tau pathology, such as BA4_MIC-4 and INS_AST-0, findings that depended on profiling brain regions with intermediate pathology and relative resilience. Furthermore, we defined a combinatorial TF GRN underlying the AD-specific microglial state, BA4_MIC-7, which includes the risk gene SPI1, which was recently demonstrated to repress AD risk genes in late-stage AD²⁴ and whose reduced expression has been associated with delayed AD onset.¹⁵ We speculate that these shared and distinct neural-immune pathways represent candidate therapeutic targets.

PSP is pathologically defined by its unique pattern of neuronal and astrocytic tau pathology,^2,82 but its molecular correlates have not been well defined. Here, we find that depleted ASTs (V1_AST) downregulate REST, which represses neuronal gene expression, whereas they upregulate transcription factors typically observed in neurons (Table S5). We hypothesize that a relative weakening of factors involved in AST self-maintenance may be related to their reduction in PSP, which warrants further study.

We also found several associations between cell composition and qualitative tau pathology scores, including the expected positive correlation between tau hyperphosphorylation and dystrophic MICs expressing senescent-associated genes⁴² (Figure S4E). In PSP ASTs, tau pathology scores were correlated with increased numbers of protoplasmic ASTs expressing genes involved in synapse phagocytosis,^83,84 inhibition of hypoxic response, and inhibition of WNT signaling (Figure 2C). Another interesting correlation with tau score in PSP ASTs was the upregulation of PKFYVE, a drug target in development for ALS (Figure S4C).⁸⁵ Among FTD layer 2/3 IT neurons, we found an inverse relationship between tau score and neuronal expression of vulnerability markers, such as RORB and KCNH7, suggesting greater vulnerability in samples with higher tau pathology (Figures S9D–S9F).

Tying gene expression to causal factors is informed by understanding its relationship to genetic risk. In this regard, we provide several examples of AD, bvFTD, or PSP risk genes that are enriched in specific cell states observed in that disease, suggesting a causal link. This includes AD risk genes, such as ITM2B, APOE, BIN1, and SPI1 in MICs that are enriched in AD brain; bvFTD/ALS risk genes, including OPTN, CTSF, TPRM7, and TMEM106B, enriched among layer 2/3 neurons that are specifically vulnerable in bvFTD; and PSP-associated genes, such as WNT3, ASAP1, and RUNX2, which were DE in layer 2/3 neurons (e.g., V1_EX-2) that are more abundant in PSP, compared with layer 5/6 neurons that are depleted in PSP (e.g., INS_EX-13). These data show that known-risk genes act in specific neuronal and glial states or cell types that differ across related disorders. Moreover, causally associated disease states may be limited to specific cell types and regions. This underscores the importance of examining multiple brain regions to understand causal disease pathways at the cellular level, which we show as providing a clearer picture of shared and disease-specific aspects of resilience and vulnerability to inform the therapeutic roadmap.

Limitations of the study

Although this work provides a proof of principle as the first multi-region, multi-disorder single-cell analysis, it has several limitations. Large-scale replication in independent cohorts is crucial for generalizability. We replicated previous results where possible and addressed technical challenges in post-mortem tissue by measuring group-level trends and implementing conservative filtering to eliminate poorly represented changes. Limited sample sizes hinder comprehensive gene expression and cell composition analysis by snRNA-seq, particularly for low-abundance cells. Independent datasets and orthogonal marker gene validations are necessary, as we demonstrated by selected key examples. To mitigate error in marker gene detection stemming from cell non-independence, we validated our findings using reference-based cell type mapping³² and by aggregating subclusters across regions. Further experimental confirmation of candidate resilience factors and TF drivers identified here in model systems is necessary to validate their causal impact on disease states. Additionally, exploring other non-tau dementias, such as Lewy body dementia, PD, and FTD with TDP-43 pathology, will elucidate whether resilience and vulnerability factors extend beyond tau-related dementia. Finally, spatial transcriptomic analysis will be critical to further support findings, such as the identity of MEIS⁺ interneurons as white matter interstitial cells and to define the spatial relationships between different glial and neuronal states and cell-cell interactions that are essential to understanding disease mechanisms.

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Daniel Geschwind (dhg@mednet.ucla.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• RNA-seq and ATAC-seq data have been deposited at Synapse (syn52074156) and are available as of the date of publication. Accession numbers are listed in the key resources table.

• All original code has been deposited at Zenodo and is publicly available as of the date of publication. DOIs are listed in the key resources table.

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

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit KCNH7	Elabscience	E-AB-53444
Goat IBA1	Abcam	ab5076; RRID:AB_2224402
Rabbit GPC6	Bioss	bs-2177R; RRID:AB_11053347
Rabbit GPC5	Abcam	ab124886; RRID:AB_10971204
Rabbit ITM2B	Invitrogen	PA5–31441; RRID:AB_2548915
Mouse AT8	Invitrogen	MN1020; RRID:AD_223647
Chicken TUJ1	NovusBio	NB100–1612;RRID:AB_10000548
Rabbit S100	Abcam	ab41548;RRID:AB_956280
Rabbit RORB	Sigma-Aldrich	HPA008393;RRID:AB_1079830
Rabbit KCNH7	Atlas antibodies	HPA018039;RRID:AB_1852125
Guineapig NeuN	Synaptic Systems	266004;RRID:AB_2619988
Goat GFAP	Novus biologicals	NB100–53809;RRID:AB_829022
Rabbit Anti-Optineurin Antibody (C-2)	Fisher	PA5–28249;RRID:AB_2545725
Mouse Anti-Protein CASP antibody (CUX1 Antibody)	Abcam	Ab54583;RRID:AB_941209
Rabbit Cux2	Bioss	50–198-1706
Donkey anti-Rabbit IgG (H+L) Highly	Invitrogen™	A-21206;RRID:AB_2535792
Cross-Adsorbed Secondary Antibody, Alexa Fluor™ 488
Donkey anti-mouse 488	Invitrogen™	A-21202;RRID:AB_141607
Goat anti-guinea pig 647	Invitrogen™	A21450;RRID:AB_2535867
Donkey anti-Rabbit 555	Invitrogen™	A-31572;RRID:AB_162543
Alexa Fluor® 647 AffiniPure™ Donkey	Jackson ImmunoResearch Labs	703-605-155;RRID:AB_2340379
Anti-Chicken IgY (IgG) (H+L)
Biological samples
Human Postmortem Tissue Frozen Control	UCSF Neurodegenerative	See Table S1
(minimal pathology, no clinical diagnosis of	Disease Brain Bank and
cognitive impairment or dementia) Pick’s	UPENN Neurodegenerative
disease (bvFTD), PSP or AD.	Disease Research Brain Bank
Chemicals, peptides, and recombinant proteins
Trueblack ® lipofuscin autofluorescence quencher	Biotium	23007
Citrate Antigen Retrieval buffer, 10x	Sigma	c9999
NeuroTrace™ 640/660 Deep-Red	Invitrogen™	N21483
Fluorescent Nissl Stain (Nissl)
Donkey Serum	Fisher	NC1697010
Alexa Fluor™ 555 Tyramide SuperBoost Kit, goat anti-rabbit IgG	Fisher	B40923
VECTASHIELD ® PLUS Antifade Mounting Medium	Vectorlabs	H-1900-10
VECTASHIELD® Antifade Mounting Medium with DAPI	Vectorlabs	H-1800-10
Critical commercial assays
Chromium™ Chip B Single Cell Kit, Chromium™	10x genomics	1000073, 1000075, 1000076
Single Cell 3‘ GEM, Library & Gel Bead Kit v3,
Chromium Single Cell 3’ Gel Bead Kit v3,
Chromium™ Single Cell 3’ Library & Gel Bead	10x genomics	120267, 120264,120265
Kit v2, Chromium™ Single Cell 3’ Library Kit v2,
Chromium™ Single Cell 3’ Gel Bead Kit v2
Chromium™ Next GEM Single Cell ATAC Library &	10x genomics	1000175, 1000159, 1000163
Gel Bead Kit v1.1, Chromium Next GEM Single Cell
ATAC Gel Bead Kit v1.1, Chromium Next GEM
Single Cell ATAC Library Kit v1.1
Deposited data
Bulk and single nuclear RNAseq, ATACseq; “Raw Data”	This publication	Synapse: syn52074156; syn52369053
Single nuclear RNAseq, ATACseq’ “Processed Data”	This publication	Synapse: syn52074156; syn52082747
Analysis Code RNAseq	This publication	https://doi.org/10.5281/zenodo.12734741
Analysis Code ATACseq	This publication	https://doi.org/10.5281/zenodo.12734743
Human postmortem single cell data – Motor cortex (Azimuth)	Bakken86	https://doi.org/10.5281/zenodo.4546932
GRCh38.p12	NCBI Homo sapiens Updated	https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.39/
Annotation Release 109.20211119
ENSEMBL 93	EMBL-EBI	https://ftp.ensembl.org/pub/release-93/gtf/homo_sapiens/
CisTarget hgnc19 references	Stein Aerts lab	https://resources.aertslab.org/cistarget/databases/old/homo_sapiens/hg19/refseq_r45/mc9nr/gene_based/
TRANSFAC 7.0 Public motifs	generegulation.com	http://gene-regulation.com/pub/databases.html
Software and algorithms
10X Genomics Cell Ranger 3.0	https://www.10xgenomics.com/support/software/cell-ranger/latest	N/A
LIGER 0.5.0.9000	Liu et al.87	https://github.com/welch-lab/liger
SCENIC 1.1.2	Aibar88	https://scenic.aertslab.org/
Seurat 4.1.0	Hao et al.89	https://satijalab.org/seurat/
WGCNA 1.71	Langfelder and Horvath90	http://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/
lme4 1.1–33	Bates91	https://github.com/lme4/lme4/
limma 3.50.3	Ritchie92	https://bioinf.wehi.edu.au/limma/
Azimuth 0.3.1	Hao et al.89	https://azimuth.hubmapconsortium.org/
ComplexHeatmap 2.10.0	Gu93	https://jokergoo.github.io/ComplexHeatmap-reference/book/
BisqueRNA 1.0.5	Jew and Alvarez94	https://github.com/cozygene/bisque
ArchR 1.0.2	Granja et al.95	https://github.com/GreenleafLab/ArchR
MAGMA 1.08bb	de Leeuw et al.96	https://cncr.nl/research/magma
CIBERSORTx 1.0	Newman et al.97	https://cibersortx.stanford.edu/
ALLCools 1.0.5	Liu et al.98	https://lhqing.github.io/ALLCools/intro.html
ArchR 1.0.3	Granja et al.95	https://www.archrproject.com/index.html

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Human tissue samples

Freshly frozen human brain tissue (BA4: precentral gyrus, V1: calcarine cortex, INS: insular cortex) were obtained from the UCSF Neurodegenerative Disease Brain Bank and University of Pennsylvania Center for Neurodegenerative Disease Research Brain Bank. We obtained samples from 41 total individuals, including 10 subjects with clinical diagnosis of bvFTD and neuropathological diagnosis of Pick’s disease (FTLD-tau), 10 subjects with clinical diagnoses of AD-type dementia and a neuropathological diagnosis of Alzheimer’s disease, and 11 subjects with a clinical diagnosis of PSP-RS and a neuropathological diagnosis of PSP (FTLD-tau), and 10 non-demented controls, sex-matched to the patients (Table S1). All procedures involving the use of postmortem human brain were conducted after obtaining the written informed consent, and after approval by the Committee on Human Research at the University of California, San Francisco and University of Pennsylvania. IRB exemption was obtained from the UCLA IRB to authorize use of de-identified human postmortem brain single nuclear sequencing data in this study. Neuropathological diagnoses were made prospectively at the contributing brain banks following standard criteria.^99,100 We added an alternative available PSP case to complete n = 10 for insula, as one PSP case did not have insula tissue available for snRNA-sequencing. Of 120 samples used as input for snRNA-seq, 118 passed nuclear isolation of which 12 failed library synthesis and 5 were removed as sample outliers (Table S1) leaving 101 final samples. Additional sequencing data was generated for V1 libraries 4–5, following the completion of our analyses. This additional data has been included for public data sharing due to its potential value. For snATAC-seq, 80 samples yielded 78 final libraries post quality control, filtering, and outlier removal (Table S1).

METHOD DETAILS

Neuropathological scoring and brain region selection

Patients autopsied at UCSF underwent a standardized, semi-quantitative scoring of various pathological features (Table S1). These assessments were carried out prospectively, at the time of autopsy. Frozen tissue blocks were taken from the same regions that underwent scoring; these regions were taken from an adjacent frozen slab whose tissue face fell across the plane of the dissection blade from the scored region. Scoring for neurodegenerative features and tau burden were carried out using methods previously described.³¹ Briefly, morphological and immunohistochemical analyses of glial and neuronal tau pathomorphologies were performed in 40 regions of 13 cases with progressive supranuclear palsy (PSP), 5 with Pick’s disease (bvFTD with tau pathology), and 20 with Alzheimer’s disease. Throughout, we use the terms FTD and bvFTD interchangably, using FTD as a shorthand notation. Nonspecific features of neurodegeneration were scored based on the hematoxylin and eosin stain and included microvacuolation, astrogliosis, and neuronal loss, each graded on a 0 to 3 scale (absent, mild, moderate, severe). Tau aggregates were visualized using a monoclonal anti-phospho-tau (pS202) antibody CP13. Pathomorphologies were assessed using the same 0–3 scale and included neurofibrillary tangles, Pick’s bodies, neuronal cytoplasmic inclusions, globose tangles, astrocytic plaques, tuft-shaped astrocytes, thornshaped astrocytes, tau-positive threads and grains in the gray and white matter, and glial cytoplasmic inclusions. To analyze the pattern of tau-related pathomorphologies of the three patient groups, we calculated a composite score by adding tau and neurodegeneration scores. The composite score was used to prioritize brain regions from mild to severe stage of tau inclusions in each patient group of tauopathies. Overall, three cortical brain regions were selected across three patient groups to reflect selective regional vulnerability of tauopathies for this study, including middle insula (INS), precentral gyrus (BA4), and calcarine cortex (V1).

Immunofluorescence

IHC-P human brain slides were deparaffinized by being placed in a Clarity^™ oven for 15–30 minutes then submerged in Citrasolv (Cat# c9999, Sigma) followed by graded ethanol washes. Primary antibodies were visualized either with standard Alexa Fluor^® fluorescent secondaries or antibodies from the Alexa Fluor^™ 555 Tyramide SuperBoost^™ Kit (B40923), as indicated. Sections prepared for Alexa Fluor^™ antibodies were heated in 1x Citrate Antigen retrieval (Cat #c9999, Millipore Sigma) for 15 min, before blocking (5% Donkey serum cat#NC1697010, Fisher) for 1 hour at room temperature. Primary antibodies were as follows: KCNH7–1:500, Cat# E-AB-53444 Elab; GPC6–1:100 Cat#bs-2177R Bioss; CUX1–1:150, Cat #ab54583 abcam; S100-beta 1:2000 Cat #ab41548 abcam; NeuN-1:100, Cat #266004; ITM2B-1:200 Cat# PA5–31441 Invitrogen; IBA1–1:200, Cat# ab5076 abcam; TUJ1–1:100 Cat#NB100–1612 Novusbio; OPTN-1:100, Cat #PA5–28249 Fisher; RORB-1:100, Cat #HPA008393 Sigma-Aldrich). ITM2B antibody specificity was validated by IHC and Western blot. The same ITM2B antibody was used in the following prior studies.¹⁰¹ The GPC6 antibody was used in prior studies.^102,103 Because the KCNH7 antibodies were not well established, we verified reproducible staining pattern in our samples with a second KCNH7 antibody from the Human Protein Atlas which underwent antigen-purification followed by validation of antigen specificity using protein array (Atlas Antibodies #HPA018039, 1:650). We verified similar patterns across antibodies by IHC and Western blot. Primary antibodies were incubated at 4°C overnight. Secondary antibodies (Donkey anti-mouse 488, Cat#A-21202, Invitrogen; Donkey anti-Rabbit 555, Cat#A31572; Goat anti-guinea pig 647, Cat#A21450) were diluted 1:500 and incubated for 1 hour. For KCNH7 and GPC6 staining, to prepare for tyramide amplification, sections were incubated in 3% hydrogen peroxide for 1 hour at room temperature and washed (1X PBS Cat#10010049, Fisher) before blocking (C2-kit) for 1 hour at room temperature. KCNH7 (1:500 Cat# E-AB-53444, Elab) or GPC6 (1:100 Cat#bs-2177R, Bioss) primary incubation was followed by an HRP secondary incubation for one hour and a tyramide amplification step (555). Sections were then stripped and reheated for 17 minutes in 1x Citrate Antigen retrieval. After rinsing with 1X PBS, a blocking step (5% Dk Serum) for 1 hour followed and then a 1-hour incubation of primary antibodies or counterstain (RORB; TUJ1; Nissl N21483 −1:50, Invitrogen). Secondary antibodies (Donkey anti-rabbit 488, Cat#A-21206, Invitrogen; Donkey anti-chicken Cat#703–605-155, Jacksonimmuno) were diluted 1:500 and incubated for 1 hour. To use Neurotrace Nissl, slides were first rehydrated in 0.1 M PBS (pH 7.2) for 40 minutes, then permeabilized for 10 minutes with.1% Triton X-100. Nissl was added at dilution 1:20 for 20 minutes and then left overnight in PBS at 4°C. All slides were treated with 0.3% Sudan Black in 70% EtOH for 3 min to reduce autofluorescence before imaging.

Image Quantification

All imaging and data capture and analysis was completed in a blinded fashion. For CUX1, images were captured with an upright Zeiss Axioplan II microscope with Zeiss software. For ITM2B, GPC6 and KCNH7 quantification analysis, images were scanned and digitized using a Vectra^® Polaris^™ microscope to scan slides, and samples were visualized using Qupath (v0.2.0-m5). For CUX1 analysis, we captured 20X images including 5 randomly selected, representative regions in gray matter from PSP and control cases (n=5 per condition, visual cortex, cases from snRNAseq study cohort). The number of CUX1-positive and S100beta-positive cells were then counted in a blinded fashion fashion. For ITM2B, GPC6 and KCNH7 quantification and analysis, we used the Vectra^® Polaris^™ microscope to scan slides and Qupath (v0.2.0-m5) to visualize data. Settings were thresholded according to optimal signal to background minimization and uniformly applied to all images. For ITM2B cell quantification, images were captured at 40x magnification, selecting 5 randomly selected regions in gray matter from AD and bvFTD with tau pathology (Pick’s disease) cases. In total, we assessed for ITM2B staining in 331 microglia in AD and 357 in FTD samples. Following immunostaining, slides were counterstained with Nissl (NeuroTrace^™ 640/660). To compare enrichment of ITM2B in microglia with other cells, we similarly counted ITM2B positivity in cells with large nuclei characteristic of pyramidal neurons morphology (verified by NeuN), in the same tissue samples. In total, we assessed for ITM2B staining in 189 total neurons in AD and 228 in FTD samples. To quantify GPC6 positive neurons in outer layers, we applied the Nissl staining to visualize these layers and counterstained with TUJ1 to positively label neurons. We first verified assignment of cortical layer into outer (layer 2–3) and deep (layer 5) by staining 3 slides from controls with CUX1 for outer layer and RORB for deeper layer neurons, verifying these stains specified neurons with the same average distance from cortical edge relative to our assignment of layers based on Nissl. We counted the number of GPC6 + TUJ1 + positive neurons in bvFTD (n = 4) and control cases (n=4) For KCNH7, we stained cases from the snRNAseq cohort as well an independent series of samples from the UCLA brain bank (n = 7 per condition for motor cortex from snRNAseq cohort, n = 3 per condition for independent frontal cortex from UCLA cohort). We used Nissl staining to identify layer 5 neurons which we validated based on co-staining with RORB as described above. Then, we counted the number of KCNH7 stained neurons in layer 5, using Nissl morphology as the criteria for neurons. In total, we assessed for KCNH7 staining across 656 neurons in AD and 506 neurons in control samples.

Single nucleus isolation

Nuclei were prepared from 60–70mg of frozen brain tissue per sample, with all procedures carried out on ice or at 4°C with RNase-free reagents. Briefly, postmortem frozen brain tissue was gentle lysed in 3mL homogenization buffer (250mM sucrose, 150mM KCl, 30mM MgCl₂, 60mM Tris, 0.01% v/v Triton X-100, 0.001% v/v Digitonin, 0.01% v/v NP40, 1mM DTT, supplemented with 0.2U/mL RNase Inhibitor (NEB, M0314), Complete protease inhibitor cocktail (Roche, 11697498001) using a Wheaton Dounce Tissue Grinder (30 strokes with pestle B). The lysate was filtered through a 40μm cell strainer and centrifuged at 1000×g for 8 minutes to obtain a nuclear pellet. To remove debris, the nuclear pellet was resuspended in 350μL homogenization buffer and 1:1 with an equal volume of 50% iodixanol buffer (Iodixanol 60% v/v combined with buffer of 250 mM sucrose, 150mM KCl, 3mM MgCl₂, 60mM Tris), then layered over 600μL of 29% iodixanol buffer (Iodixanol 29% v/v combined with buffer of 250mM sucrose, 150mM KCl, 3mM MgCl₂, 60mM Tris) and centrifuged at 13500×g for 20 minutes. The supernatant was discarded, and nuclei gently resuspended and washed in 1mL of 1% BSA/PBS. The nuclei were visually inspected to confirm complete lysis and nuclear integrity. Nuclei were manually counted and diluted to a concentration of 1000 nuclei/μL in 1% BSA/PBS. For single-nucleus RNA sequencing (snRNA-seq), libraries were prepared using the Chromium Single Cell 3’ Reagent Kits (v2 for BA4 and V1, v3 for INS) according to the manufacturer’s protocol (10X Genomics). For snATAC-seq, libraries were prepared using the Chromium Single Cell Next GEM Single Cell ATAC kit (v1.1). RNA-seq libraries were sequenced on a Novaseq S2 or S4 sequencer with paired end reads (read 1: 26 bp, read 2: 96 bp) targeting over 50,000 paired reads per nucleus. ATAC-seq libraries were sequenced on a Novaseq S4 with paired end reads (2×50 bp) targeting 25,000 paired reads per nucleus.

Single nucleus RNA-seq Alignment and Filtering

Raw single-nuclei RNA-seq data was processed using the 10X Genomics Cell Ranger (v3.0) pipeline. Reads were aligned to the Ensembl release 93 Homo sapiens genome. Cells were selected for downstream analysis using the cell barcodes associated with the most UMIs. We estimated the number of cells expected to be captured based on input nuclei concentration and retained this many cell barcodes for downstream analysis. Cells with < 200 unique genes detected were removed (gene detection: > 1 count). Cells with > 8% of their counts mapping to MT genes were removed. Genes detected in < 3 cells were removed. Normalization was performed using Seurat (v3.1).¹⁰⁴ Briefly, raw counts are read depth normalized by dividing by the total number of UMIs per cell, then multi-plying by 10,000, adding a value of 1, and log transforming (ln (transcripts-per-10,000 + 1)). Raw UMI counts data were assessed for the effects from biological covariates (clinical diagnosis, anatomical region, donor, age, sex), and technical covariates (RIN, PMI, library batch, number of UMI, number of genes detected, percentage MT). The effects of the number of UMI (sequencing depth) were removed from the read depth normalized expression values using a linear model. Outlier samples were identified based on abnormal frequencies of major cell types and divergent gene expression patterns and were removed from the analysis (Table S1). Doublet cells were removed using Doublet Finder (doubletFinder_v3).¹⁰⁵ To achieve this, for each brain region we ran independent scaling, principal component analysis, and clustering through Seurat. pK was estimated using paramSweep_v3 on the first 40 principal components. The homotypic doublet proportion was initially estimated at 7.6% and refined through modelHomotypic. This doublet rate is within estimates expected based on cell number. High confidence doublets were identified and removed by region and disorder. The final cell count was 590, 224 after removal of outliers and low-quality cells (Table S1).

Annotation of major cell types, subtypes and states

All samples (BA4, INS, V1; all diagnoses) were jointly clustered in Seurat (v3.1). Counts were log-normalized, and the 2000 most variable genes were selected for dimensional reduction and clustering using FindVariableGenes / vst. log₁₀(UMI) was regressed in ScaleData. PCA was performed to reduce the dataset to 200 principal components, and the first 100 were used as input to UMAP (see Data S1). Clusters were identified using the Leiden method in Seurat::FindClusters with resolution = 0.1. Each cluster was then annotated as a major cell type using mean expression of groups of cell type marker genes. Canonical genes were selected based on mouse and human studies as well as published reference atlas enriched genes (Astrocyte GFAP/SLC1A2/AQP4/SLC14A1, Endothelia VWF/CLDN5/FLT1, Inhibitory neurons GAD1/GAD2, Excitatory neurons SATB2/SLC17A7/NRGN/SNAP25, Oligodendrocyte MOBP/MOG/TF, Microglia CD47/CSF1R/C3, OPC VCAN/PDGFRA/CSPG4, Pericyte ACTA2/RGS5/PDGFRB, Lymphocyte CD247 taken from Hodge et al.,³⁴ Kelley et al.,³⁵ Mathys et al.,³⁶ and Sweeney et al.³⁷ (see Data S1). We also confirmed all major cell type classifications used human motor cortex reference dataset using Azimuth,³² a web-based portal from the Allen Brain Atlas (https://azimuth.hubmapconsortium.org). Extratelencephalic (ET) neurons contribute cells to BA4_EX3 and BA4_EX-1 as shown in Figure S1B but do not populate a distinct cluster. IN, OL and AST subclasses were labeled based on reference-based classifiers (STAR Methods; for example OLIGO L3-L6 ENPP6³² (Table S2). We include examples such as ENPP6, a marker of putative newly formed OL,¹⁰⁶ RBFOX1, PLP1 indicating myelinating OL, and BCAS1 marking early pre-myelinating and/or disease associated OL^38,106,107 (See Tables S2 and S4 for additional markers with literature references).

Where indicated in the text for stringent filtering, an additional round of filtering was conducted to confirm that findings were not influenced by variations in sample quality. To prevent ambiguously annotated cells from affecting reported changes in composition or gene expression differences related to astrocytes in PSP, we implemented a filter to exclude cells whose Seurat cluster type annotation did not match that of the single cell. Results from this analysis were reported separately from the main analysis and are available in Table S5.

After identification and clustering of 9 main cell types, we sub-clustered each of the 8 main CNS cell types independently to identify distinct cell states within each cell-class. To maximize the influence of disorder-specific effects on subclustering, we performed this analysis on each brain region separately, generating 178 total clusters (post-filtering) (Table S2). Specifically, we combined nuclei from the same major cell type and different diagnosis groups, and performed batch correction, data integration and subsampling using Liger (v0.4.2; with kappa = 20 and lambda = 0.5).^87,108 We removed clusters that were non-representative across multiple subjects (with less than three libraries contributing >10 nuclei from at least one diagnostic group), low-quality clusters based on significant association with multiple sample quality metrics (FDR <0.05, limma, number of gene detected per cell, sample post-mortem interval, percent_mitochondrial genes detected per cell), or ambiguous cell type with 30% or more nuclei in a given cell-type enriched cluster bearing markers more suggestive of a disparate cell type (see Table S1 for full list of filtered clusters with justification). Cluster V1_AST-3 was originally removed during the course of our standard QC due to ambiguous marker expression, including expression of neuronal TFs. It was reintroduced because further investigation (e.g., overlap with GRNs) revealed that rather than being an artifact, it potentially represented cell state changes in PSP related to the observed astrocyte depletion in V1.

For reference-based assignment of clusters to cell classes and subclasses, cells were mapped to an external human motor cortex dataset³² using the reference-based mapping workflow described.⁸⁹ We prefiltered our cells as recommended by the web interface: UMI count in the range [212, 33185], gene count in the range [201, 6486], and proportion of mitochondrial genes <= 2%. For visualization Sankey plots (Figure S1B) were made between our subclusters and the predicted subclusters. High gene count (nFeature_RNA >= 10000) and cells with low mapping scores (mapping.score <= 0.9) were filtered out. Subcluster/predicted subcluster links were restricted to those in the same cell type. Clusters that did not map to reference cell type were discarded from further analysis (Table S1). In this way, each final cluster was assigned a reference cell class and subclass, as shown (Figures 1, S1, and S2; Table S2). Representative reference cell assignments for each cluster ( >5% cells mapped) were visualized as Sankey plots for control cells. Finally, to organize subclusters into related cell types, subtypes and states across all brain regions and diagnoses groups, we performed hierarchical clustering based on their marker genes, and then jointly annotated based on marker gene expression and reference cell type assignments. First, for each cluster, we calculated cluster-specific marker gene expression by performing differential gene expression (DGE) analysis to compare each cluster with all other clusters of the same cell type and brain region. Genes were filtered using a minimum proportion, keeping only genes detected in 10% or more of cells per cell type x brain region group. DGE was calculated based on mixed effects model used lmerTest::lmer¹⁰⁹ with formula expression ~ clinical_dx + pmi + age + sex + number_ umi + percent_mito + (1 | library_id). To complete hierarchical clustering across brain regions, we then filtered DGE lists to genes detected in each region. We then used the estimate terms (beta) for the top 100 genes per cluster based on DGE results for all regions, sorted by the cross-region variance. The top 100 most variable genes were used to compute a Euclidian distance matrix (stats::dist) and complete-linkage hierarchical clustering (stats::hclust) using default parameters. Groups of clusters were annotated based on shared marker genes. Reference-based cluster assignment results from Azimuth were overlaid manually to the hierarchical clustering framework to assign cluster groups to reference cell type and subtypes based on high confidence matches where >70% of cells of a cluster assigned to the same reference cell type (Figures 1G, S1, and S2; Table S2).

Cell type composition analysis

Cluster composition was defined as the proportion of cells in a given cluster relative to the total number of cells of that major cell class and brain region, per subject. To measure cell type composition across subjects and diagnosis groups, cluster composition percentages were used as pseudobulk counts, forming a cluster by subject count matrix for each major cell type and brain region. The matrix was normalized using TMM¹¹⁰ and Limma-voom,¹¹¹ then fit with model formula ~0 + dx + pmi + age + sex + mean_percent_mito + median_genes. T-statistics were calculated using eBayes.¹¹² As an initial screen to identify shared and distinct compositional changes by diagnosis, we measured compositional changes within each disease group vs. control samples (Table S4, tab = “Diagnosis_vs_Control_screen_ALL”) applying FDR corrected across all final clusters within a given cell type and brain region (post-QC), but not corrected for comparisons across multiple disorders (designated FDR “one dx”, Table S4, tab = “Diagnosis_vs_Control_screen_ALL”), to achieve thresholds comparable with single disorder studies. We then applied an additional FDR correction for all three diagnosis groups compared to control (designated FDR “three dx”; Table S5, tab = “Diagnosis_vs_Control_screen_ALL”), and these were reported were indicated (Figures 2B, 2C, 3H, 3L, S4A, S4B, and S7F). To designate and count clusters with “shared” compositional changes, we identified clusters with similar compositional trends in each disease vs control comparison, measured their combined “all disease vs control” log2FC, p-value, and adjusted p-value, and applied the threshold FDR <0.05 to designate “shared” clusters (Table S4, tab = “Shared_Summary”, Figures 1D and 2C (INS_MIC-3, INS_OL-7, V1_END-1, INS-OPC-3, BA4_OL-7, BA4_INS-7, INS_AST-0, INS_END-1), Figures S2 and S4). P-values were FDR adjusted for multiple testing across final clusters (post-QC) from each corresponding cell type and brain region (as shown in Table S4). Importantly, these shared combined-disorder compositional changes were distributed across each disease group, such that for each disease vs. control abs(log2FC) >1.0 (Table S4). To identify clusters with “distinct” trends, we used three criteria. First, we compared samples from one disorder group with all remaining samples (other disease groups and controls), calculated log2FC, p-value (as shown in Figure 4A) and adjusted p-values that were corrected across clusters from a given cell type and brain region within the disorder group tested. We used the statistical cut off of p-adjusted <0.1 to count distinct clusters (Table S4, tabs = “One Diagnosis vs Others”, “Distinct_Summary”, “Category 1”; Figures 1D, S2, and S3; 10 total clusters). In addition, we included five clusters as distinct based on the criteria that only one diagnosis group showed significant enrichment or depletion vs control samples at the threshold of FDR < 0.1 (Table S4 tabs = “Diagnosis_vs_Control_screen_ALL”, “Distinct_Summary”, listed under “Category 2”). We chose to include these clusters based on the supportive evidence that they also differentially upregulated transcription factor (TF) gene regulatory networks (GRN) with distinctly high activity in the same cell type and disorder based on SCENIC (ranked among top 10 regulons) (Figures S10A–S10D; Tables S6 and S5). For neuronal subclasses with trends indicative of regionally specific vulnerability in one disease, we measured differential composition of the depleted neuronal subclass in the corresponding disease group vs all other samples, and then correction for observations made over 3 brain regions (Figures 1D, 2A purple circles, and 4B; Table S4 tab = “Vulnerable Classes”).

For selected clusters of interest, to correlate cluster counts with variation in neuropathology and tau scores across subjects (Table S4), we measured within each diagnosis group and brain region the biweight midcorrelation and Pearson’s p-value between cluster count proportion and neurodegeneration and tau score per sample. P-values were adjusted for FDR applied over all clusters of a given cell type and brain region (for example 12 INS-AST clusters) and for testing across the 3 disease groups. Control samples were excluded because their tau and neuropathology scores were either zero or unmeasured.

Bulk RNA seq sample generation, quality control and pre-processing for deconvolution analysis

For bulk RNAseq library generation, 1μg of total RNA was used for rRNA depletion with the RiboZero Gold kit (Illumina). Remaining RNA was size selected using AMPure XP beads (Beckman Coulter) and standard libraries were prepared following Illumina’s TruSeq protocols for 50bp paired end reads. Libraries were sequenced at an average read-depth of 50–60 million reads per sample on a HiSeq2500 instrument using rapid mode. All samples were aligned with Salmon¹¹³ to GRCh38 and annotated with Gencode v25.¹¹⁴ Expression data from three batches of library preparation and sequencing were concatenated together to form the raw counts matrix. We dropped one sample (ID:19–2) pre-QC due to incomplete sample metadata (see Table S1). For the remaining samples, we log-transformed the counts, then quantile filtered the genes (kept genes with 80% of expression above 5 reads). We dropped 11 samples based on poor metrics (more than 3 standard deviations in connectivity as calculated from WGCNA::fundamentalNetworkConcepts). Prior to cell type deconvolution, we additionally regressed outtechnical covariates by fitting a linear model (expression ~ Primary.Neuropath.Dx + Batch + RIN + PMI + Sex + Age + Source + seqpc1 + seqpc2 + seqpc3 + seqpc4 + seqpc5) and subtracting the predicted effects of all terms except for Primary.Neuropath.Dx.

CIBERSORT deconvolution analysis

For inferring the cellular composition in the log-transformed bulk RNAseq data (post-QC and outlier removal described above), we applied CIBERSORTx (https://cibersortx.stanford.edu/)⁹⁷ without any modification to the developer’s instruction. Although less sensitive for compositional analysis, particularly of rarer cell types, cell composition analysis based on deconvolution lacks many of the potential confounders in single cell data.¹¹⁵ Docker images were downloaded on 8/10/2023. Signature matrix of human motor cortex was generated according to the developers’ instruction using human motor cortex reference snRNAseq data, reported by,³² downloaded on 12/10/2023, with the subclass level annotation. The normalized bulk brain RNAseq data was then deconvoluted with their fraction function according to their instructions. The deconvoluted estimates of the cell proportions are visualized with R (4.3.0). Statistical testing was performed with Wilcoxon test, and multiple testing correction was applied with Benjamin-Hochberg method across all cell types and diagnostic conditions. For astrocytes shown in Figure S5E, FDR correction was applied across astrocytes from each disease group and brain region tested.

BISQUE deconvolution analysis

Bisque deconvolution was performed using Bisque v1.0.3 performed on log transformed bulk RNAseq data (post-QC and outlier removal described above). We use a reference-based decomposition using the frontal or visual cortex expression matrices from Lake,¹¹⁶ as indicated in figure legends and Table S5. The deconvoluted estimates of the cell proportions were visualized with R (4.3.0). Statistical testing was performed with Wilcoxon test, and multiple testing correction was applied using the Benjamin-Hochberg method across astrocytes from each diagnostic group and brain region.

Cross disorder differential gene expression

Differential gene expression (DGE) for each diagnosis group was calculated using a linear mixed effects model. Counts for all cell type within a brain region were derived from the Seurat counts matrix (STAR Methods, Single nucleus RNA-seq Alignment and Filtering). To assess DGE by diagnosis within a given cluster, we subset the normalized counts to cells from that cluster to generate a per-cluster cell by gene counts matrix. Genes were then filtered using a minimum proportion, keeping only genes expressed in at least 10% of cells within any condition, calculated independently per cluster. We then calculate DGE using a linear mixed effects model (LME) using lmerTest::lmer¹⁰⁹ with model formula ~ clinical_dx + pmi + age + sex + number_umi + percent_mito + (1 | library_id). Resulting p-values were then FDR-adjusted across the number of genes measure in the DGE analysis. For the analysis of shared vs distinct DGE at major cell type level (Table S5), we first identified genes with DGE in any disorder vs control, by using the cut offs of FDR <0.1 and abs(log₂FC)>0.1 (5,933 genes). To distinguish genes with disorder-specific DGE trends, we filtered for genes meeting this DGE cut off in only one disorder, and then we further applied a second stringent filter to remove genes with p-value <0.05 in two or more disorder in order to remove genes with closely overlapping trends across multiple disorders, resulting in 682 genes (Table S5). For analyses of DGE within clusters, different contrasts were used to measure DGE by disease depending on the experiment, as specified. In some cases, we contrasted one disease group, such as AD, with samples from all other diagnosis groups (such as PSP, bvFTD and control). In other cases, we generated one combined linear model to measure the effect on gene expression of each diagnosis group, contrasting each diagnosis with control (expression ~ clinical_dx+ pmi + age + sex + number_umi + percent_mito + (1 | library_id)).

In several instances, as indicated in the text and figures (Figures 3N and 4B; Table S5) we used an LME model to measure DGE between two specific clusters, within only samples of one disease samples (for example, comparing gene expression of INS_EX-2 vs INS_EX-5 in bvFTD cases, or between layer 2–4 vs layer 5–6 excitatory neurons). We subset the normalized counts matrix to only those cells assigned to the cluster or cell subtypes being compared (such as INS_EX-2 and INS_EX-5) and to only the libraries of the disease group of interest (such as bvFTD). As before, genes were then filtered using a minimum proportion, keeping only genes expressed in at least 10% of cells. The cells were subset to form two levels of a contrast labeled custom_split. We calculate DGE on the counts matrix using a linear mixed effects model (LME) using lmerTest::lmer¹⁰⁹ with formula expression ~ custom_split + pmi + age + sex + number_umi + percent_mito + (1 | library_id). Resulting p-values were then FDR-adjusted across the number of genes measure in the DGE analysis.

For astrocytes in the visual cortex, to ensure changes in gene expression observed in PSP samples was not due to low quality cells, we performed additional stringent cell filtering to retain only astrocytes with high confidence annotations (each cell assigned astrocyte as cell type based on marker gene expression or Azimuth), and these gene expression changes were maintained (Table S5).

Pseudotime analysis with SlingShot

Slingshot 2.2.0.¹¹⁷ was performed on a subset of the Liger subclusters (INS_EX-2, INS_EX-5, INS_EX-6) for a total of 12218 cells across 37 samples. A UMAP-based dimensional reduction was calculated on the first 30 principal components of the scaled expression data. UMAP data and subcluster labels were supplied as input to Slingshot 2.2.0, generating an imputed lineage and pseudotime values per cell. Pseudotime was correlated with gene expression through Pearson correlation. P-values were adjusted with FDR correction across the number of genes.

Gene regulatory network analysis

We used a modified version of the SCENIC (Single-Cell rEgulatory Network Inference) approach⁷⁵ for constructing GRNs from single-cell RNAseq data. Briefly, SCENIC contains three steps: (1) identify co-expression modules between TF and the potential target genes; (2) for each co-expression module, infer direct target genes based on those potential targets for which the motif of the corresponding TF is significantly enriched. Each regulon is then defined as a TF and its direct target genes; (3) the regulon activity score (RAS) in each single cell is calculated through the area under the recovery curve. We performed this analysis on four cell types (excitatory neurons, oligodendrocytes, astrocytes, microglia) from each brain region (INS, BA4, V1), using final cell clusters post-quality control, outlier removal, regression and normalization in Seurat.¹⁰⁴ On disease and control cells combined, and for each cell type and brain region, we ran SCENIC. We first filtered genes to those expressed in at least 5% of cells. We then subsampled from each library to draw a similar number of cells from each library and disorder for the analysis. To achieve this, we sampled at the 15% (decile range) of counts across samples, and then performed stratified random sampling to keep this decile range of cells from each sample. Stratified random sampling starts off by dividing a population into groups with similar attributes. This method was used to ensure that different segments in a population were equally represented. We then applied the SCENIC package algorithm version 1.1.2 in R (https://github.com/aertslab/SCENIC) to generate gene regulatory networks and RAS in each cell type. To compare regulon activity across disorders, we used Shannon entropy algorithm described in¹¹⁸ to quantify regulon specificity scores (RSS) for each regulon for each disease condition. Briefly, this analysis quantifies the differences between two probability distributions, scored from 0 to 1. The essential regulators are predicted to be those with the highest specific scores. This application has been used previously to define cell-type-specific regulons. Therefore, we first confirmed that our entropy-based scores reproduced known cell-type-specific TFs, by comparing microglia to all other cell types and observing top ranked microglia specific TFs to include RUNX1 and SP11 (Figures S9G and S9H), which are two known core microglial transcriptional regulators. We then modified this application to assign a specificity score for a regulon for each disorder (Figures 5 and 6). We plotted multidimensional scaling plots (MDS) using Sammon projections to show the relatedness of TF specific score ranks across diagnosis within brain region for MIC or AST (Figures 6A and 6D) to highlight distinct and shared TFs across disorder based on ranked differential entropy score. Findings related to TF GRN with differential RSS by disorder were further validated against disorder-enriched clusters and DGE, and with snATAC-seq, and reported (Figures 5C, 5D, 5E, 6B, 7B, S9J, S9K, and S10).

snATAC sequencing and data analysis

Nuclei were extracted as above simultaneously (BA4, INS) with 78 samples subjected to snRNA-seq balanced by age, PMI and cause of death (n = 8 – 11 per diagnosis and region pair as shown in Table S1). Libraries were aligned to reference genome with 10X Genomics software, Cell Ranger 3.0 (cellranger-atac count; https://support.10xgenomics.com/single-cell-atac/software) then all libraries were aligned together using the cellranger-atac aggr function. We used ArchR version 1.0.2⁹⁵ to filter out low quality cells through nucleosome banding score < 4, TSS < 2, minimum fragments < 1000, and blacklist region ratio > 0.1. We additionally dropped low-quality samples P1_7at1_7, I3_6_at, and I1_7 (Table S1), leaving 141038 AD, 136732 PSP, 145106 bvFTD and 152369 control (total = 575,245) high quality nuclei from each dataset (Table S1).

We normalized with ArchR’s iterative Latent Semantic Indexing and Harmony batch correction on the sequencing and preparation batches. Clustering was performed using the Seurat 4.1.0 clustering algorithm with resolution = 0.8 to generate 25 total clusters. Two clusters (C17, C20) were removed because of low cell count (<1,000) resulting in 23 clusters labeled C1-C25 (Data S1). To annotate the clusters we first created a gene activity score matrix where gene expression levels are roughly computed from fragment counts within gene body elements. We then used the snATAC-seq gene activity matrix to integrate with our fully annotated snRNAseq clusters through Seurat’s ‘FindTransferAnchors’ function (Data S1). We then used spearman correlation to compare cluster-specific marker gene profiles from the snATAC-seq clusters (from the gene activity matrix) with our snRNAseq cluster marker gene profiles to identify overlapping clusters (Figure S9I; Data S1). This identified snATAC-seq clusters corresponding to excitatory neurons (C2), microglia (C7), and astrocytes (C18-C25) which we used for further analysis. For differential composition analysis, we fit a linear model on disease group vs. astrocyte subcluster proportion. C2 was subclustered further to identify subclusters C7 and C8, which were both relatively enriched in markers of L2/3 IT neurons, and were therefore combined and used for RORB footprinting.

Peak calling and prediction of TF activity in snATAC-seq data

We merged reads from individual cells of snATAC clusters by disease condition and brain regions. Pseudo-bulk replicates were further created in ArchR using customized parameters to ensure balanced number of cells for all pseudo-replicates (ArchR:: addGroupCoverage, minCell = 950, minRep = 4). Peak calling was performed using macs2 (https://github.com/macs3-project/MACS) in ArchR,⁹⁵ resulting in a reproducible, non-overlapping peakset (ArchR::addReproduciblePeakSet). Differential accessible peak regions were determined by pair-wise comparisons of chromatin accessibility between disease and control cells. We identified significant TFs sourced from our SCENIC analysis and JASPAR via ArchR wrappers (addBgdPeaks, addDeviationsMatrix) for chrom-VAR¹¹⁹ 2017, selecting TF motifs that were correlated with significant deviations in chromatin accessibility. To infer transcription factor activity, we performed TF footprinting analysis on peak regions differentially accessible in each disease using TOBIAS,¹²⁰ as described in our previous study.¹²¹ This method starts with Tn5 bias correction using the TOBIAS ATACorrect module, subtracting the background Tn5 insertion cuts highlighting the effect of protein binding. To match footprints to potential TF binding sites, and to estimate TF binding activity on its target loci, we applied TOBIAS BINDetect module to the corrected ATAC-seq signals within peaks, with TF motif PWMs used in TRASFAC Pro. Many transcription factors are represented by more than one motif. To avoid motif redundancy, we clustered the motifs based on their sequence similarity using TOBIAS ClusterMotifs, and chose one motif per TF that is most similar to others in the same cluster TOBIAS BINDetect compares the positions and activities of TF footprinted sites in disease or control per clusters. Each footprint site was assigned a log2FC (fold change) between two conditions, representing whether the binding site has larger/smaller TF footprint scores in comparison. To calculate statistics, a background distribution of footprint scores is built by randomly subsetting peak regions at ~200bp intervals, and these scores were used to calculate a distribution of background log2FCs for each comparison of two conditions. The global distribution of log2FC’s per TF was compared to the background distributions to calculate a differential TF binding score, which represents differential TF activity between two conditions. A P-value is calculated by subsampling 100 log2FCs from the background and calculating the significance of the observed change. By comparing the observed log2FC distribution to the background log2FC, the effects of any global differences due to sequencing depth, noise etc. are controlled. To visualize, we used soGGI to plot TF footprints, and bar graphs to show the global TF footprint activity changes comparing disease vs. control. Distribution of footprints were further annotated with ChIPseeker¹²² and the footprints located within ± 10 kb of a gene’s TSS were linked to that gene. Genes that are footprinted by the TF are considered putative targets of each TF and were used for further analysis.

Analysis of chromatin accessibility at cell-specific methylated regions

To assess for evidence of altered chromatin accessibility at cell-type specific hypermethylated sites in diseased astrocytes, we focused on the 7 ATAC-seq derived astrocyte clusters, namely C18-C25, and categorize cells into four conditions: one control and three diseases. We called peaks for each cluster per condition using macs2 in ArchR.⁹⁵ We integrated with the public single-nucleus methyl-3C sequencing data¹²³ to quantify the methylation levels within ATAC peak regions. We detected significantly methylated regions utilizing ALLCools,⁹⁸ using ‘all-to-region-count’ to compute methylation levels (reads supporting methylation/read coverage), and ‘generate-dataset’ to calculate the significant hyper-methylation score. ATAC peaks were annotated as methylated if they met the criteria of a hyper-methylation score >= 0.9 and a methylation level > 0. A gain in chromatin accessibility in disease vs. control samples at locations with cell type-specific hypermethylation was interpreted as a loss of heterochromatin and vice versa. The significance was determined using a permutation test with n = 10,000 iterations. Empirical p values were then adjusted for multiple testing across 210 comparisons (*** FDR ≤ 0.001, ** FDR ≤ 0.01, * 0.01 < FDR ≤ 0.05,).

Protein-Protein Interaction and Gene Ontology Analysis

To assess and visualize protein-protein interactions among cluster markers, DGE and GRN, to emphasize groups of genes with functional associations, we used STRING¹²⁴ (version 11.5) with the following setting (organism: Homo sapiens for human data; meaning of network edges: confidence; active interaction sources: experiments and databases; minimal required interaction score: high confidence, max number of interactors to show: none; we reported either full STRING networks indicating edges representing both functional and physical protein associations, or physical STRING networks indicating only physical protein interactions based on databases, as indicated in Figures and Figures Legends). Direct protein-protein interaction enrichment p-values were reported as generated by STRING. Gene ontology enrichment represented within protein-protein interaction networks were prioritized and displayed along with their geneset enrichment FDR corrected p-values as generated by STRING, using whole genome as the default background. Data was exported and visualized either directly or by using the Cytoscape software.¹²⁵

Connectivity Map (CMAP) Analysis

For a given module, the top 100 module genes (ranked by t-statistic, LME) were used as input for the QUERY app in the Broad’s CMAP database, version CLUE.⁷² This signature was used to query 7,494 gene overexpression or knockdown experiments carried out across 9 cell lines for similar (positive connectivity score) or opposite (negative connectivity score) effects on gene expression signatures, incorporating Kolmogorov-Smirnov statistics (a nonparametric, rank-based pattern-matching strategy) as described.⁷² Per the CMAP website (https://clue.io), for each module-perturbagen pair, the connectivity score (tau) is a standardized percentile score that compares the similarity of the query geneset to the perturbagen compared to all other reference genesets in CMAP; such that 95 indicates that 5% of reference genesets show stronger connectivity to the perturbagen than the query dataset. For our analysis, we used the mean “connectivity scores” which was calculated from the combining data generated independently in 9 cell lines. For the vulnerability analysis, the signatures of “depleted” neurons and “disease-enriched neurons” were generated by taking the top 100 gene differentially up-regulated, or down-regulated, respectively, in INS_EX-2 vs. INS_EX-5 measured among bvFTD cases (Table S5). CMAP analysis was conducted individually to each signature, and then a combined “delta scores” was calculated for each by subtracting the CMAP score for the “depleted” signature from the “disease-enriched” signature.

GWAS risk variant enrichment

Summary statistics for genome-wide association studies for AD,¹²⁶ FTD¹²⁷ or PSP¹²⁸ were used as an input for MAGMA (v1.08bb)⁹⁶ for gene annotation to map SNPs onto genes (with annotation window = 20) and the competitive gene set analysis was performed to test module associations with GWAS variants (permutations = 100,000). Genesets analyzed included marker genes of differentially disease enriched clusters (LME, t-statistic>2); or as indicated in the text or accompanying figure legends. FDR correction was applied across competitive p-value outputs from MAGMA for all genesets tested across comparisons as shown in each figure.

Geneset enrichment

We implemented a Fisher’s exact test for geneset enrichment analysis. We compared cluster-specific and DGE by disorder (log₂FC > 0.2 in one disease vs. all others, LME, see above section) with genes assigned to TF regulons by SCENIC (see Table S6). To compare cluster-specific marker gene expression (T-statistic >2, LME, see above section) to published disease-associated cell types and markers, we used published genesets and markers shown in Table S3.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical methods for each experiment can be found in the corresponding section for each analysis. Additional experimental details including sample sizes, meaning of n, and comparisons for multiple testing corrections can be found in the STAR Methods section for each analysis type, as well as in figure legends and/or corresponding supplementary tables. Boxplots for single nuclear RNA or ATAC-seq results throughout show median values at center with upper and lower quantiles. For immunohistochemistry, boxplots show center at median with minimum and maximum values, and barplots show SEM. Methods were used to determine whether the data met assumptions of the statistical approaches applied in this study. This resulted in limiting differential gene expression analysis applied to single nuclear RNA sequencing data to genes detected in >10% of cells from each cell type, calculated by region, to avoid inflation of type 1 error due to drop out of genes at low capture. For calculating cell-type abundance using deconvolution analysis, we applied the cell proportion threshold of > 2%, as recommended.¹²⁹ Unless otherwise stated all statistical test were run using R (4.1.0). P-values were reported with post-hoc FDR using Benjamin-Hochberg (BH) method for multiple testing correction. Throughout the Results, figures, and legends, we used the term “FDR” in reference to BH-adjusted p-values to save space.

ADDITIONAL RESOURCES

Final processed data has been prepared for public browsing at CZI CellxGene: https://cellxgene.cziscience.com/collections/c53573b2-eff4-4c5e-9ad0-b24d422dfd9b.

Highlights.

• Perform comparative genomic analysis of AD, bvFTD, and PSP at the single-cell level

• Pinpoint markers and candidate drivers of selective neuronal vulnerability in dementia

• Identify disorder-specific microglia, astrocyte, and oligodendrocyte glial-immune states

• Causal genetic risk impacts disorder-specific gene regulatory networks and cells