Single-cell multiregion epigenomic rewiring in Alzheimer’s disease progression and cognitive resilience

SUMMARY

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by progressive cognitive decline, yet its epigenetic underpinnings remain elusive. Here, we generate and integrate single-cell epigenomic and transcriptomic profiles of 3.5 million cells from 384 postmortem brain samples across 6 regions in 111 AD and control individuals. We identify over 1 million candidate cis-regulatory elements (cCREs), organized into 123 regulatory modules across 67 cell subtypes. We define large-scale epigenomic compartments and single-cell epigenomic information and delineate their dynamics in AD, revealing widespread epigenome relaxation and brain-region-specific and cell-type-specific epigenomic erosion signatures during AD progression. These epigenomic stability dynamics are closely associated with cell-type proportion changes, glial cell-state transitions, and coordinated epigenomic and transcriptomic dysregulation linked to AD pathology, cognitive impairment, and cognitive resilience. This study provides critical insights into AD progression and cognitive resilience, presenting a comprehensive single-cell multiomic atlas to advance the understanding of AD.

In brief

A single-cell epigenomic and transcriptomic atlas across six brain regions in Alzheimer’s disease reveals widespread epigenomic relaxation and information loss, associated with the selective depletion of vulnerable neurons and exhaustion of activated glial cells, while preserved epigenomic stability supports cognitive function and resilience.

Graphical abstract

INTRODUCTION

Alzheimer’s disease (AD) is the most common cause of dementia, characterized by progressive cognitive decline and neurodegeneration.^1–3 A hallmark of AD is the accumulation of amyloid plaques and tau neurofibrillary tangles (NFTs), which disrupt synaptic integrity and drive neurodegeneration.^1,4–7 Notably, the progression of AD pathology is regionally selective, with the entorhinal cortex (EC) and hippocampus (HC) among the earliest and most vulnerable brain regions, followed by broader cortical and subcortical involvement.^8–10 Unraveling how region-specific patterns of vulnerability arise at molecular and cellular levels is critical for advancing our understanding of AD and guiding the development of targeted therapeutic strategies.

Single-cell transcriptomic and epigenomic technologies have advanced our understanding of brain cell types, states, and regulatory mechanisms.^11–25 Recent studies using single-cell/single-nucleus RNA sequencing (sc/snRNA-seq) and single-nucleus assay for transposase-accessible chromatin sequencing (snATAC-seq) in human AD brains and mouse models have illuminated key cell types and pathways involved in AD,^26–31 particularly through large-scale transcriptomic landscapes of the prefrontal cortex (PFC) from aged and AD individuals.^32–35 However, how epigenomic and transcriptomic alterations across brain regions and cell types contribute to AD progression, cognitive decline, and resilience remains largely unknown. Emerging evidence also suggests that epigenomic dysregulation plays a critical role in AD pathogenesis.^33,36–38 Genome-wide association studies (GWASs) have identified that AD risk variants are enriched in microglia (Mic)-specific regulatory elements, implicating epigenomic mechanisms in AD.^32,38–44 Yet, key questions remain regarding the role of epigenomic regulation in AD: how do epigenomic dynamics vary across brain regions and cell types during AD progression? Which cell populations are most susceptible to epigenomic dysregulation, and how does such instability contribute to the degeneration and exhaustion of selectively vulnerable cell types? What is the relationship between epigenomic shifts and glial cell-state transitions? Furthermore, how are these epigenomic changes linked to cognitive impairment, decline, and resilience? Finally, what regulatory networks govern the maintenance or breakdown of epigenomic integrity in the AD brain?

To address these questions, we generated single-nucleus ATAC-seq, RNA-seq, and multiome datasets from six brain regions in aged and AD individuals. We defined large-scale active and repressive compartments and quantified single-cell epigenomic information as a balance between regulatory activation and repressive region compaction. Our analyses reveal widespread compromised compartmentalization and epigenomic information loss during AD progression, particularly in selectively vulnerable excitatory neurons (Exc) from the entorhinal cortex and hippocampus, whereas cognitively resilient individuals had preserved epigenomic integrity. We found that epigenomic dynamics were linked to AD genetic risk enrichment and glial-state transitions, including activation, exhaustion, and depletion. Finally, we identified key genes, pathways, and transcription factors (TFs) associated with epigenomic stability and its loss, providing molecular insights into AD pathology, cognitive function, cognitive decline, and resilience. Overall, our findings reveal that epigenomic erosion is closely associated with AD pathology progression, cognitive impairment, and decline, while preserved epigenomic stability correlates with better cognitive function and resilience. This study provides insights into the epigenomic basis of AD progression and lays a foundation for future therapeutic strategies.

RESULTS

A comprehensive single-cell epigenomic and transcriptomic landscape of AD across six brain regions

To systematically characterize the epigenomic and transcriptomic dynamics during AD, we collected 384 postmortem samples from 111 age-matched individuals with and without AD from the Religious Orders Study (ROS) and the Rush Memory and Aging Project (MAP), two longitudinal cohort studies of aging and dementia (Figures 1A and S1A).⁴⁵ Individuals exhibited varying degrees of AD pathology based on clinical diagnoses, clustering into normal individuals without AD pathologies (nonAD; n = 57), early-stage AD (earlyAD, n = 33), and late-stage AD (lateAD, n = 21) (Figures 1A and S1A; Table S1).²⁹ To investigate brain-region-specific and cell-type-specific dynamics of AD progression, we analyzed six brain regions: EC, HC, anterior thalamus (TH), angular gyrus (AG), midtemporal cortex (MTC), and PFC. These regions exhibit varying susceptibilities to AD, with the EC and HC being the earliest and most vulnerable, while cortical regions like the PFC and AG exhibit later pathological involvement (Figure 1A). We generated 799 single-nucleus libraries from snATAC-seq, snRNA-seq, and snMultiome (paired snRNA-seq and snATAC-seq) across 6 brain regions (Figure 1A; Table S1).^38,46,47 We integrated all snATAC and snRNA data separately (Figures S1B–S1D). After quality control, we obtained 1,217,165 nuclei from snATAC and 2,263,395 nuclei from snRNA data (Figures 1B, 1C, and S1B–S1D).

Figure 1. Single-cell epigenomic and transcriptomic landscape of AD across six brain regions.

(A) Overview of study design, ROSMAP sample collection, single-cell profiling, and data analyses.

(B and C) UMAP embeddings of 1.2 million snATAC nuclei (B) and 2.3 million snRNA nuclei (C), annotated into 7 major cell classes and 67 subtypes. Cell classes and brain regions are color-coded on the left; cell-subtype annotations are shown on the right for both modalities.

(D) Heatmaps showing cell-subtype abundance across brain regions (left), gene activity/expression of 7,520 cell-subtype-specific genes across snATAC/snRNA (middle), and TF motif deviations (right).

See also Figures S1 and S2, Table S1, and Data S1, pages 1–3.

This multimodal atlas is hierarchically organized into three levels of cell-type annotations in both snATAC and snRNA datasets, including 7 major cell classes, 34 subclasses, and 67 high-resolution subtypes (Figures 1B, 1C, and S1E–S1H; Data S1, pages 1 and 2). The seven cell classes include Exc, inhibitory neurons (Inh), oligodendrocytes (Oli), oligodendrocyte precursor cells (OPCs), astrocytes (Ast), Mic/immune cells (Mic/Immune), and vascular cells (Vasc) (Figures 1B, 1C, and S1E). Among subclasses, we defined 17 excitatory neuron subclasses spanning neocortical layers and the entorhinal, hippocampal, and thalamic regions, along with 6 inhibitory neuron subclasses characterized by selective expression of individual marker genes: VIP, SST, PVALB, LAMP5, PAX6, or MEIS2 (Figure S1G). The high-resolution subtypes include 25 excitatory, 22 inhibitory, 4 astrocyte, 3 oligodendrocyte, 3 OPC, 4 microglia/immune (including 2 subtypes of microglia, CNS-associated macrophages [CAMs], and T cells), and 6 vascular subtypes (Figures 1B, 1C, and S1H).

We analyzed the brain-regional diversity and cell-type composition of 7 major cell classes and 67 subtypes in 6 brain regions in both snATAC and snRNA datasets, observing clear brain-region specificity across different cell types (Figures 1D and S1F; Data S1, pages 1A and 1B). Overall, neocortical brain regions (i.e., AG, MTC, and PFC) show greater cell fractions of excitatory neurons (Figures 1D and S1F), and we identified 12 excitatory subtypes primarily found in various neocortical brain regions and 6 excitatory subtypes enriched in the entorhinal cortex, 6 in the hippocampus, and 1 thalamic-specific subtype (Exc NXPH1 RNF220) (Figure 1D; Data S1, pages 2A–2F). The entorhinal cortex-specific subtypes were marked by genes associated with the lateral or medial entorhinal cortex, e.g., RELN, TOX3, and GPC5 (Figure 1D; Data S1, pages 2A, 2B, and 2E).^46,48,49 In the hippocampus, identified neuronal subtypes included cells from the structured hippocampal formation, such as CA1 and CA2/CA3, as well as the dentate gyrus (DG) and subiculum (Figure 1D; Data S1, pages 2A, 2B, and 2E). In contrast to the greater regional diversity of excitatory neuron subtypes, most inhibitory neuron subtypes (21 out of 22) were shared across brain regions, with the exception of Inh MEIS2 FOXP2, which was more specific to the thalamus (Figure 1D; Data S1, pages 2C, 2D, and 2F). The thalamus, compared with other regions, exhibited a lower fraction of neuronal cell types and a higher fraction of various glial and vascular cell types (Figures 1D and S1F).

Cell-subtype-specific genes and TFs of human brain

To explore the regulatory networks that govern cell-type-specific identity in the human brain, we then analyzed cell-subtype-specific marker genes across both snATAC and snRNA modalities (Figure 1D). We identified 7,520 cell-subtype-specific genes shared between snATAC and snRNA datasets across the 67 subtypes, with 3,335 genes specific to excitatory and inhibitory neuron subtypes (Figure 1D; Data S1, page 2G; Table S1). We observed substantial cross-modality concordance between cell-subtype-specific transcriptional expression and epigenomic activity (Figure 1D; Data S1, page 1C).

We next identified key TFs that define the molecular identities of brain cell types (STAR Methods). This analysis revealed distinct TF motif enrichment patterns across brain cell types (Figures 1D, S2A, and S2B). Excitatory neurons were enriched for NEUROD family TFs. HOX TFs were associated with astrocytes, oligodendrocytes, and OPCs, suggesting their role in glial lineage specification. FOX TFs (FOXO1, FOXO3, FOXD1) were enriched in oligodendrocytes, OPCs, and vascular cells.⁵⁰ Microglia-specific TFs (SPI1/PU.1, SPIB, SPIC) were associated with microglial identity and immune function (Figures 1D and S2B), consistent with previous studies.^51–53

Notably, among neuronal cell types, thalamic neuron subtypes displayed distinct TF enrichment patterns, compared with other brain regions (Figures 1D and S2C). Further analysis revealed a strong enrichment of core circadian regulators (NR1D1/REV-ERBα, NR1D2/REV-ERBβ) and retinoic acid receptors (RORs: RORA, RORB, RORC) in both thalamic excitatory and inhibitory neurons, with inhibitory thalamic neurons additionally enriched for PITX and GSC family TFs (Figures S2C and S2D; Data S1, pages 3A and 3B). To further investigate the functionality of thalamic-specific TFs and their regulatory targets, we identified TF-regulated candidate cis-regulatory element (cCRE)-gene pairs (STAR Methods). Functional enrichment analysis of NR1D1/2 and RORs-regulated genes revealed associations with neuronal projection development, synaptic signaling, and cell junction organization, as well as additional pathways involved in circadian entrainment, sleep regulation, memory, cognition, and lipid metabolism, aligning with thalamic functions of regulating sleep-wake cycles and cognitive functions (Figure S2E).^54–56 Interestingly, our analysis revealed that the RORA gene locus was highly enriched for TF binding sites of NR1D1/2 and RORs, along with their cCRE-gene regulatory interactions, suggesting a potential self-regulatory network that contributes to thalamic neuron-specific functions (Figure S2F).

cCREs, modules, pathways, and GWAS interpretation across brain regions and cell types

To elucidate the epigenomic regulatory elements that underlie the identity and function of various brain cell types, we computed cCREs in each of the 67 cell types, grouped them into regulatory modules, and examined their roles in transcriptional regulation and the interpretation of brain disease risk variants using GWAS data (Figures 2A–2D and S3A–S3E). Across cell subtypes, we identified a union set of 1,010,153 cCREs (Figure S3A). These cCREs were highly enriched in active promoter and enhancer chromatin states, validating these regions as regulatory elements (Figures S3B and S3C). To characterize the cell-type specificity of cCREs, we employed non-negative matrix factorization (NMF) to partition cCREs into 123 cis-regulatory modules across the 67 cell subtypes and 6 brain regions, based on normalized chromatin accessibility (Figure 2A; STAR Methods).^12,25,57 The first module contained non-cell-type-specific regulatory elements, and the remaining 122 modules showed highly cell-type-restricted accessibility (Figure 2A). The enrichment and cell-type specificity of cCRE modules were validated by scoring modules in uniform manifold approximation and projection (UMAP) embeddings of nuclei and by clustering cCREs within each module in the cCRE embedding space (Figure 2A).

Figure 2. cCREs, cCRE modules, gene activity, expression, and GWAS risk variant interpretation across human brain cell types.

(A) Overview of 1,010,153 cCREs organized into 123 distinct modules. Top: cell embeddings based on cCRE module activity, showing representative modules enriched in specific cell types. Middle: cCRE embeddings illustrating the distribution of cCREs within each representative module. Bottom: heatmap showing the activity of the 123 cCRE modules (columns) across various cell types and brain regions (rows). Cell types associated with each cCRE module are color-coded according to their cell subtype and subclass at the bottom.

(B) snATAC gene activity for genes associated with the corresponding 123 cCRE modules from (A).

(C) snRNA gene expression for genes associated with the corresponding 123 cCRE modules from (A).

(D) LDSC enrichment analysis of GWAS risk variants across 67 brain cell subtypes for 90 neurological, psychiatric, neurodegenerative, and other complex traits. Statistical significance of GWAS enrichment is represented as −log₁₀(p value).

See also Figure S3.

To explore how cCREs regulate gene expression and contribute to cell identity, we integrated snATAC and snRNA datasets using unified cell-subtype annotations to link RNA expression to 123 regulatory modules for each cell type (Figures 2B and 2C). Correlating snATAC gene activity and snRNA expression with cCRE module activity revealed strong concordance, supporting the quality of cell-type annotations and module definitions (Figures 2A–2C). Functional enrichment reveals that the cell-type-specific modules are enriched for distinct functional pathways and Gene Ontology (GO) terms (Figure S3D). Modules specific to excitatory and inhibitory neurons are enriched for pathways related to synapse structure, neuronal signal transduction, and neuron migration (Figure S3D). Microglia/immune cell-associated modules are enriched for immune response and cell activation pathways, while vascular cell-associated modules are highly enriched for blood vessel morphogenesis and vasculogenesis (Figure S3D).

We next leveraged our single-cell epigenomic landscape and regulatory modules to interpret non-coding genetic risk variants associated with complex traits and diseases (Figures 2D and S3E).^42–44 Using linkage disequilibrium score (LDSC) regression, we assessed SNP heritability from 90 GWASs by evaluating the enrichment of trait-associated variants across cCREs in 67 cell subtypes and 123 regulatory modules in the human brain.⁵⁸ We found that bipolar disorder, intelligence, schizophrenia, and depression-related variants are enriched in the cCRE modules from excitatory and inhibitory neuron subtypes (Figures 2D and S3E). Immune disorders (including inflammatory bowel disease and multiple sclerosis) and AD were enriched in microglia or immune cells, consistent with prior studies (Figures 2D and S3E).^42–44 Vascular cells were enriched for traits associated with blood pressure and hypertension (Figures 2D and S3E). These results highlight the crucial role of cell-type-specific regulatory elements in shaping the genetic basis of brain-related complex traits and diseases.

Characterizing cell-type-specific epigenomic compartments

The human genome is organized into two distinct compartments with specific spatial localization and epigenetic signatures.^59,60 Active compartments, located in the interior of the nucleus, have more accessible chromatin, while repressive compartments are less accessible and tend to be located near the nuclear periphery and nuclear lamina, characterized by a higher A/T content in their sequence.^59,60 To systematically characterize large-scale epigenomic dynamics during AD progression, we inferred active and repressive compartments in major cell classes across 6 brain regions using hidden Markov model and ATAC signals from 27,993 high-quality genome-wide 100-kb bins and further grouped them into 25 epigenomic compartment groups (ECGs), based on their active or repressive status, using K-means clustering (Figure 3A; Data S1, pages 4A–4F; STAR Methods). We validated the inferred epigenomic compartments using multiple genetic and epigenetic features, including CpG density, A/T content, enrichment in nuclear domains (active speckle-associated domains [SPADs] and repressive lamina-associated domains [LADs]), and EpiMap chromatin states (Figures 3A and S4A–S4D; Data S1, pages 4A–4E; STAR Methods).^61,62

Figure 3. Large-scale epigenomic compartment dynamics in AD across brain regions and cell types.

(A) Heatmap of 27,993 100-kb genome-wide bins grouped into 25 ECGs, showing active (pink) and repressive (blue) compartments across 7 cell classes, 6 brain regions, and 3 AD stages. Highlighted are cell-type-specific active compartments and compartment switches in AD, including activation of repressive chromatin and repression of active regions. Bar plots (top) summarize compartment activity by brain region, cell class, and pathology. Bottom annotations include CpG density, A/T content, LAD/SPAD enrichment, and chromatin states.

(B) Integrative analysis of LAD proportion, A/T content, epigenomic compartment states, and AD-associated dynamics across chromosomes. Left to right: bar plot of nuclear lamina interactions (ranked by LAD proportion), violin plot of A/T content (higher in repressive chromosomes), dot plots of active/repressive compartment proportions, and dynamic scores indicating compartment switching during AD. Positive values indicate repressive-to-active transitions; negative values reflect the reverse.

(C) Chromatin accessibility increases in LADs and decreases in SPADs in lateAD across cell classes (Wilcoxon test).

(D) Genome-wide view of epigenomic changes in lateAD. Top-left: whole-genome view of log₂(fold change) in snATAC signals (lateAD versus nonAD), with blue indicating repression and red indicating activation. Bottom-left: zoom-in views highlight global activation of repressive chr18 and repression of active chr19. Right: snATAC changes across brain regions and glial cell types, highlighting vulnerable brain regions with epigenomic dynamics.

(E) Schematic model of compromised compartmentalization and epigenomic relaxation in AD.

See also Figure S4 and Data S1, pages 4 and 5.

Characterization of active and repressive compartments across brain cell types revealed distinct cell-type-specific patterns (Figure 3A). Microglia and oligodendrocytes exhibited a higher proportion of repressive compartments, whereas excitatory and inhibitory neurons showed fewer repressive regions (Data S1, page 4D). Further analysis revealed that some compartments (e.g., ECG-1, -24, -25) are broadly shared across cell classes, whereas others display cell-type-specific activity (Figure 3A). For instance, ECG-18 and -19 are selectively active in inhibitory and excitatory neurons, while ECG-20 to -23 are enriched in astrocytes, OPCs, oligodendrocytes, and microglia, respectively (Figure 3A). These patterns coincide with the activation of cell-type-specific gene expression programs (Data S1, page 4G).

Globally compromised compartmentalization and decreased transcriptomic fidelity in AD

We then investigated large-scale epigenomic compartment alterations in AD (Figures 3A–3D and S4E–S4H). By clustering compartments, we found cell classes grouped by AD pathology groups (Figure 3A). LateAD samples exhibited more frequent transitions between repressive and active compartments (Figures 3A, S4E, and S4F). Notably, oligodendrocytes, microglia/immune, and vascular cells in the lateAD entorhinal cortex showed the most pronounced compartment transitions, highlighting the entorhinal cortex as particularly vulnerable to epigenomic dynamics (Figures 3A and S4G).

Previous studies have demonstrated that chromosomes have distinct chromosome territory and nuclear radial positioning, with higher A/T content chromosomes (e.g., chr18) typically located at the nuclear periphery interacting with the nuclear lamina, whereas lower A/T content chromosomes (e.g., chr19) are centrally positioned and exhibit active epigenetic states (Figure 3B).^63–65 We investigated whether the compartment dynamics in AD are linked to the intrinsic localization or epigenomic signatures of each chromosome. Using nuclear lamina and chromatin interaction data from the human brain,⁶¹ we sorted chromosomes by their interactions with the nuclear lamina (the proportion of LADs for each chromosome), indicating their radial and spatial positioning from the nuclear periphery to the interior and their overall epigenetic states from repressive to active (Figure 3B). Indeed, we found that chromosomes with greater lamina interactions tend to have higher A/T content and a greater proportion of inferred repressive compartments across cell classes (Figure 3B).

When examining the dynamics of chromatin compartments across cell classes for chromosomes with distinct epigenomic signatures in AD, we found that repressive chromosomes tended to be activated (repressive-to-active compartment transitions) and active chromosomes tended to be repressed (active-to-repressive compartment transitions) in lateAD (Figure 3B). To validate this globally compromised compartmentalization, we directly surveyed snATAC signals within previously defined compartment domains (i.e., SPADs and LADs) in the human brain (Figures 3C and S4H). Consistent with the compartment transition analysis, a genome-wide overview of epigenomic alterations between lateAD and nonAD demonstrated that repressive chromatin marked by greater nuclear lamina interactions, higher A/T content, and weaker association with nuclear speckles was activated, while active chromatin characterized by lower nuclear lamina interactions, lower A/T content, and stronger association with nuclear speckles was repressed (Figures 3B–3D and S4H). Notably, typically repressive chromosomes with more nuclear lamina interactions, such as chr18, chr4, and chr13, were widely activated in AD (Figures 3B, 3D, and S4H). Conversely, the active chromosomes, including chr19 and chr22, tended to be globally repressed in AD (Figures 3B, 3D, and S4H). The entorhinal cortex and hippocampus, two regions most vulnerable to AD pathology, exhibited the strongest epigenomic alterations, particularly in oligodendrocytes, OPCs, and microglia/immune cells (Figure 3D).

To further validate the compromised epigenomic identity in AD, we analyzed external AD epigenomic datasets, including those profiling active (H3K27ac and H3K4me3) and repressive (H3K27me3) histone modifications.^40,66 Consistent with our snATAC analyses, active histone modifications (H3K27ac and H3K4me3) exhibited increased signals in repressive compartments and decreased signals in active compartments (Figure S4I). Conversely, the repressive histone modification H3K27me3 exhibited elevated signals in active compartments and reduced signals in repressive compartments, further validating the compromised compartmentalization in AD (Figure S4I).

We next examined whether global epigenomic dynamics are associated with overall gene expression changes in AD by calculating the average expression score for genes located in repressive or active compartments. We observed that genes in repressive compartments showed overall increased expression in lateAD, while genes in active compartments exhibited decreased expression (Figure S4J; Data S1, page 5A). Motivated by the coordinated epigenomic and transcriptomic changes in AD, we then analyzed the relationship between gene expression in repressive and active compartments and various variables of AD pathology and cognitive measures (Figure S4K; Data S1, page 5B). We found that expression level of genes in repressive compartments was positively associated with AD pathology progression but negatively correlated with cognitive function (Figure S4K; Data S1, page 5B). Conversely, expression level of genes in active compartments was positively associated with cognitive function but negatively correlated with AD pathology, with the higher correlations observed in inhibitory neurons and astrocytes (Figure S4K; Data S1, page 5B).

Taken together, we identified widespread compromised compartmentalization and epigenomic erosion in AD, characterized by activation of repressive chromatin regions and repression of active regions, particularly in the vulnerable entorhinal cortex and hippocampus (Figure 3E). These epigenomic alterations are linked to a global decline in transcriptomic fidelity, providing insights into regulatory dysfunction underlying AD.

Global loss of epigenomic information in AD

Motivated by our findings of large-scale compromised epigenomic identity during AD, we next sought to dive into these epigenomic dynamics at single-cell resolution across cell types and brain regions to identify the most vulnerable populations. From the information and erosion theory of aging, epigenetic information increases with cellular differentiation, specialization, and development and is lost during aging.^38,67–69 We quantified the epigenomic information at the single-cell level based on the relative distribution of reads in cCREs (preferentially located in active enhancer and promoter regions) versus genome-wide backgrounds (STAR Methods). This single-cell resolution epigenomic information was further validated based on the independent external references, including Human Brain Atlas (HBA) snATAC cCREs and H3K27ac cCREs derived from sorted human brain cell types, as well as EpiMap chromatin states across multiple brain regions,^25,41,62,70 revealing a high consistency of the epigenomic information derived from all alternative references with high statistically significant correlations (p < 2.2 × 10⁻¹⁶ and Spearman correlation > 0.97) (Figure S5A).

We then studied epigenomic information dynamics in AD across 7 cell classes and 67 subtypes spanning 6 brain regions and observed a global reduction in epigenomic information across all cell classes in lateAD (Figure 4A). Detailed analysis at cell class and subtype resolution across brain regions revealed region-specific and cell-type-specific patterns of epigenomic deterioration (Figures 4A and S5B). Consistent with the compartment dynamic analysis, the entorhinal cortex and hippocampus showed a more pronounced decline of the epigenomic information in lateAD, compared with neocortical regions, aligning with the region-specific progression of AD pathology (Figures 4A and S5B). Among major cell classes, glial cells, including oligodendrocytes, OPCs, and microglia, exhibited a notable loss of epigenomic information in lateAD (Figure 4A). At the cell-subtype level, consistent with the high vulnerability of the entorhinal cortex and hippocampus, excitatory neuron subtypes from the entorhinal cortex (e.g., Exc RELN COL5A2 and TOX3 TTC6) and hippocampus (e.g., Exc CA1 pyramidal cells, GRIK1 CTXND1 [Subiculum], and DG granule cells) displayed a marked decrease in epigenomic information in lateAD, compared with other excitatory neuron subtypes (Figure 4A). Among neocortical excitatory neurons, the upper-layer subtypes (e.g., Exc L2–3 CBLN2 LINC02306, L3–4 RORB CUX2) experienced greater epigenomic information loss in lateAD, compared with deeper-layer neurons (Figure 4A). For inhibitory neurons, the Inh L3–5 SST MAFB subtype was identified as particularly vulnerable, exhibiting the most pronounced loss of epigenomic information in lateAD (Figure 4A). Our findings are consistent with previous studies identifying upper-layer excitatory neurons and SST-expressing inhibitory neurons as selectively vulnerable populations in AD.^47,71

Figure 4. Epigenomic information changes during AD progression across brain regions and cell types.

(A) Heatmap of epigenomic information changes in 7 cell classes and 67 subtypes across 6 brain regions in AD. Bar plot on the right summarizes epigenomic information differences between lateAD and nonAD. Cell types with pronounced information loss are marked in blue; vulnerable brain regions are highlighted with black boxes.

(B) UMAP and ridge plots showing glial subtypes and their activation/inflammation scores.

(C) Density plots of epigenomic information in glial subtypes across AD stages.

(D) Epigenomic information changes along the microglial activation trajectory across AD stages.

(E) AD-related GWAS enrichment across the microglial activation trajectory across AD stages.

(F) Modified epigenetic Waddington’s landscape model depicting glial-state transitions during AD progression—from homeostatic to reactive/activated and eventually to exhausted states. Homeostatic glial cells initially gain epigenomic information and identity upon activation, supporting reactivity function. However, with sustained activation, they progressively lose this information, culminating in an exhausted state characterized by diminished epigenomic identity in lateAD.

See also Figure S5.

Interplay between epigenomic information dynamics, glial cell-state transitions, and AD genetic risk enrichments

Glial cells play indispensable and complex roles in neurodegenerative disorders by maintaining homeostasis and brain plasticity, providing support and protection for neurons, and regulating immune responses.^4,72 In AD, microglia and astrocytes become reactive or activated to protect neuronal functions through glia-neuron crosstalk.^73–75 To study the epigenomic changes across glial cell types and cell states, we examined the activation and inflammatory signatures in glial subtypes (Figures 4B and S5C–S5F). Among the four annotated astrocyte subtypes, we observed distinct activation states, with Ast DCLK1 showing a relatively higher activation score (Figures 4B and S5C).^76–79 In oligodendrocytes, Oli OPALIN and Oli RAFGRF1 are more inflammatory with higher activity for inflammation or disease-associated oligodendrocyte genes (Figures 4B and S5D).^76,80 For microglia, we annotated classic homeostatic (Mic P2RY12) and activated (Mic TPT1) subtypes, the latter showing higher activity of canonical activation genes including CD9, B2M, APOE, and C1QA (Figures 4B and S5E).^{32,76,81–83} Next, we investigated epigenomic information dynamics across glial subtypes during AD and observed a remarkable decrease in epigenomic information in activated or inflammatory glial cells in lateAD, including Ast DCLK1, Oli OPALIN, Oli RAFGRF1, and Mic TPT1 (Figures 4A and 4C).

To further study epigenomic deterioration across microglial activation states, we inferred the microglial activation trajectory based on the activity of the canonical microglial activation genes (Figure S5F; STAR Methods).^{32,76,81–83} Integrating microglial activation trajectories with cell states from previously reported large-scale datasets, including ROSMAP and Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD) datasets,^32,35,71 revealed a progressive transition trajectory from early surveillance and homeostatic phases, through intermediate reactive states, to terminally activated microglia characterized by lipid metabolism, inflammatory, ribosome biogenesis, and disease-associated microglia (DAM) signatures (Figure S5G). We further explored the changes in epigenomic information along the trajectory of microglial activation, demonstrating that microglia exhibit an initial increase of epigenomic information during activation, followed by a substantial decline of the epigenomic information in the terminally activated microglia in lateAD (Figure 4D). This indicates epigenomic deterioration and premature degeneration of the activated microglia in lateAD individuals. Consistent trends were observed across all six brain regions, with the entorhinal cortex and hippocampus showing the most pronounced decline in epigenomic information in activated microglia (Figure S5H).

Microglia are highly enriched for risk variants associated with AD, indicating a significant link between microglia function and the AD progression.^44,84 To investigate the relationships between AD progression, epigenomic dynamics, microglial activation, and AD-GWAS enrichment, we calculated the AD-GWAS enrichment score at single-cell resolution using SCAVENGE⁸⁵ and integrated them with microglial activation trajectories (Figure 4E; STAR Methods). The AD-GWAS enrichment score increased during early microglial activation but declined in terminally activated microglia, peaking in activated microglia with the highest epigenomic information, indicating a strong connection between genetic risk, epigenomic instability, and microglial exhaustion in AD (Figures 4D and 4E). Analysis across APOE genotypes revealed that APOE4 homozygotes exhibited the most severe epigenomic erosion, followed by heterozygotes, compared with non-carriers (Figure S5I). A similar pattern was observed in astrocytes, with early activation showing increased epigenomic information and AD-GWAS enrichment, followed by declines in terminally activated astrocytes (Figure S5J). These findings suggest a shared mechanism linking genetic susceptibility and epigenomic instability driving glial dysfunction in AD.

Together, these findings reveal the interplay between glial cell-state transitions, epigenomic dynamics, and AD genetic risk variants. While homeostatic glial cells gain epigenomic identity upon activation, terminally activated or inflammatory glial cells experience epigenomic erosion, cellular degeneration, and exhaustion. This epigenomic deterioration is particularly pronounced in individuals carrying AD-related risk variants (Figure 4F).

Preserved epigenomic information in cognitive resilience

Cognitive resilience (CR) is defined as the ability to preserve cognitive function despite the degree of AD pathology.^46,47,86 In addition to examining epigenomic information dynamics in AD progression, we explored how epigenomic dynamics are associated with CR.^46,47,86 We calculated continuous CR scores based on the discrepancy between observed cognitive function (from clinical diagnosis) and that expected from global AD pathology (Figure S6A; STAR Methods). Individuals were categorized into three quantile-based groups: cognitively vulnerable (low CR scores), cognitively resilient (high CR scores), and an intermediate group (intermediate CR scores). Individual-level association analyses across various AD-related pathological, phenotypic, and cognitive variables showed that CR is positively associated with higher cognitive function and slower decline across multiple cognitive function domains while being negatively associated with AD pathology burden (Figures S6B and S6C).

We then investigated the dynamics of epigenomic information across individuals with varying degrees of CR (Figure 5A). Contrary to the global decrease in epigenomic information observed across cell classes in lateAD (Figure 4A), we discovered a global increase in epigenomic information in cognitively resilient individuals across major cell classes and brain regions, compared with the cognitively vulnerable individuals (Figure 5A). Among major cell classes, glial cells, including oligodendrocytes, OPCs, and microglia, exhibited a remarkable increase in epigenomic information in the cognitively resilient group (Figure 5A).

Figure 5. Epigenomic information dynamics link to CR and cell-type composition changes in AD.

(A) Heatmap showing higher epigenomic information in cognitively resilient individuals compared to cognitively vulnerable individuals, across brain regions, cell classes, and top-ranked vulnerable subtypes.

(B) Negative correlation between epigenomic information changes in AD progression and CR. Vulnerable subtypes (top left) exhibit epigenomic information loss in AD but preservation in resilience.

(C) Cell-type enrichment or depletion patterns linked to epigenomic information dynamics, CR, and AD pathology progression, shown as log₂(odds ratio) heatmaps.

(D) Pearson correlations between cell-type fraction changes linked to epigenomic information dynamics, CR, and AD pathology progression.

(E) Positive association between epigenomic information loss and cell-type depletion in AD based on snATAC data, highlighting vulnerable subtypes.

(F) Schematic model illustrating AD-associated cellular community changes, including epigenomic information loss, selective neuronal depletion in vulnerable brain regions, and exhaustion of activated glial cells.

See also Figure S6.

Subtype-resolved analysis revealed that epigenomic information varied across CR groups in a cell-subtype-specific manner (Figure 5A). Vulnerable subtypes that lost epigenomic information in AD showed notable increases in the cognitively resilient group (Figures 4A, 5A, and 5B). These included excitatory neurons in the entorhinal cortex (e.g., Exc RELN COL5A2, RELN GPC5, and TOX3 TTC6) and hippocampus (e.g., DG granule cells and GRIK1 CTXND1), as well as neocortical upper-layer excitatory neurons (e.g., Exc L2–3 CBLN2 LINC02306 and L3–4 RORB CUX2) (Figures 4A, 5A, and 5B). Notably, Inh L3–5 SST MAFB and CUX2 MSR1, previously identified as selectively vulnerable in AD, also showed increased information in the resilient group, highlighting their roles in both AD progression and resilience.⁴⁷ Additionally, inflammatory oligodendrocytes (Oli OPALIN, RASGRF1) and OPC TPST1 exhibited higher epigenomic information in resilience (Figures 5A and 5B). We further examined epigenomic information along microglial and astrocyte activation trajectories in the context of CR, and we found that unlike the loss observed in AD, cognitively resilient individuals preserved epigenomic information in activated glial states, compared with cognitively vulnerable individuals (Figure S6D).

Collectively, these results reveal a general decline in epigenomic information in AD and its preservation in CR, particularly within selectively vulnerable cell types.

Epigenomic dynamics underpin cell-type preservation in CR and depletion in AD

We next investigated whether epigenomic stability dynamics are linked to changes in cell-type composition during AD progression and CR. For each cell class and subtype, we calculated individual-level epigenomic information scores and assessed cell-type fraction changes in relation to epigenomic information dynamics, AD pathology progression, and CR (Figure 5C; STAR Methods). Overall, cell-type composition changes linked to epigenomic information dynamics were positively correlated with those in CR but negatively correlated with changes during AD progression (p < 0.01), consistently across both snATAC and snRNA datasets (Figure 5D). At the major cell-class level, the abundance of excitatory and inhibitory neurons tended to be higher in individuals with higher levels of epigenomic information and CR but depleted in individuals with higher global AD pathological levels (Figure 5C). Consistent with the vulnerable subtypes identified in the analysis of epigenomic information dynamics during AD and CR (Figures 4A, 5A, and 5B), vulnerable excitatory neuron subtypes in the entorhinal cortex (e.g., Exc RELN COL5A2, RELN GPC5), hippocampus (e.g., DG granule cells), and neocortical upper-layer (e.g., L2–3 CBLN2 LINC02306, L3–4 RORB CUX2) were more abundant in individuals with higher epigenomic information and CR but depleted in those with a higher global AD pathology burden (Figures 5C and S6E). Direct comparisons between epigenomic information changes and cell-type fraction dynamics across cell subtypes during AD progression further demonstrated that a reduction in epigenomic information in vulnerable excitatory neuron subtypes (e.g., Exc RELN COL5A2, RELN GPC5, and DG granule cells) from vulnerable brain regions (EC and HC) was closely associated with their depletion in AD progression (Figure 5E).

We further explored the relationship between glial cell-state transitions, epigenomic information dynamics, and cell fraction changes in microglia and astrocytes in relation to AD pathology progression and CR. By investigating cell-type fraction changes across states along the microglial and astrocyte activation trajectories, we assessed their relationship with AD pathological variables, cognitive functions, cognitive decline slopes, and CR scores (Figure S6F). Our analysis revealed that the loss of epigenomic information was associated with a depletion of cell proportions in late-stage activated microglia and astrocytes, which correlated with an advanced AD pathology burden (Figures 4D, S5J, and S6F). In contrast, individuals with preserved cognitive function, slower cognitive decline, and higher CR exhibited preserved epigenomic information and cell fractions in activated microglia and astrocytes (Figures S6D and S6F). These results suggest that maintaining epigenomic stability in these activated glial cells may protect them from exhaustion, degeneration, and depletion.

Together, our findings reveal that higher epigenomic information is linked to cell-type preservation in cognitive resilience, while epigenomic information loss is associated with cell-type depletion during AD pathology progression. Our analysis highlights the importance of maintaining epigenomic integrity for cognitive function and resilience, offering insights into protective mechanisms against AD progression (Figure 5F).

Transcriptomic alterations associated with changes in epigenomic stability

Next, we investigated the regulatory networks underpinning epigenomic dynamics across brain regions and cell types in AD. We focused on differentially expressed genes (DEGs) and AD-GWAS risk genes, alongside the enrichment of functional pathways, TFs, and histone modifications at the transcriptomic level, as well as on differentially accessible genomic regions (DARs) and chromatin-state enrichments at the epigenomic level (Figure 6A).

Figure 6. Regulatory networks linking epigenomic information dynamics to AD pathology progression, cognitive function, and resilience.

(A) Overview of DEGs, DARs, AD-GWAS risk genes, TFs, histone modifications, and chromatin states associated with epigenomic information dynamics and their links to AD pathology progression, cognitive function, and resilience.

(B) GO term and pathway enrichment for DEGs associated with higher (red) or lower (blue) epigenomic information across brain regions and cell types. (C) Network of TFs and histone modifications enriched in DEGs with lower epigenomic information. Circles represent brain regions by cell class; diamonds indicate enriched factors. Edge thickness reflects enrichment significance −log₁₀(p value).

(D) Representative differential expression of AD-GWAS genes associated with epigenomic information dynamics across brain regions and cell classes.

(E) Chromatin-state enrichment of DARs associated with higher (top) or lower (bottom) epigenomic information.

(F) Correlation between sample-level epigenomic information and AD pathology, cognitive function, and resilience.

(G) Heatmap showing the Spearman correlations of log₂(fold change) for DARs and DEGs linked to epigenomic information dynamics, AD pathology progression, and cognitive measures.

See also Figures S7, S8, S9, and S10 and Data S1, page 6.

To interpret the DEGs associated with epigenomic dynamics, we leveraged our snMultiome datasets (including a total of 288,480 nuclei with paired high-quality snATAC and snRNA data) to map snATAC-derived single-cell epigenomic information to the matched snRNA-derived gene expression. We then identified DEGs associated with the epigenomic information changes and categorized the DEGs into two groups: (1) genes whose expression levels positively correlate with epigenomic information (i.e., higher expression in cells with greater epigenomic information), indicating association with higher epigenomic maintenance and stability; and (2) genes whose expression levels negatively correlate with epigenomic information (i.e., higher expression in cells with lower epigenomic information), reflecting association with epigenomic information loss or erosion (Figure 6A). Among the top DEGs with associations to epigenomic information dynamics across seven cell classes and six brain regions, several notable genes tied to higher epigenomic information include DSCAM, ASTN2, VCAN, and RORA (Figures S7A and S7B).^87–89 Among these genes, ASTN2 plays a crucial role in neuronal morphogenesis and cognitive functions, and its disruption is a risk factor for various neurological disorders.^90,91 Interestingly, we identified several AD-GWAS risk genes among these genes, including DENND4A and ENSA (Figures S7A and S7B).^42,44 Conversely, genes associated with lower epigenomic information (i.e., higher expression levels in epigenomically eroded cells) include GLUL, CST3, EGLN3, MT3, and VIM (Figures S7A and S7B). CST3 has been implicated in cerebral amyloid angiopathy and AD.⁹² EGLN3 and MT3 are associated with cellular stress responses.^93,94 Notably, we also found that APOE, one major genetic risk factor of AD, is negatively associated with epigenomic information, particularly in astrocytes, across all brain regions (Figures S7A and S7B).^42,44 APOE drives inflammation in human astrocytes, and its overexpression induces various AD-related pathologies, suggesting a link between AD-GWAS gene dysregulation and epigenomic instability.^95,96

DEGs positively associated with higher epigenomic information are enriched in functional pathways or GO terms related to nuclear and chromatin structure organization (including the nucleus, nuclear envelope, nucleolus, chromatin organization, and telomere maintenance) and to protein and organelle localization and organization (i.e., maintenance of protein location in the nucleus and organelle assembly) (Figure 6B). These findings underscore the indispensable role of global nuclear and chromatin structure organization and maintenance in preserving epigenomic identity and information, serving as protective mechanisms against AD. Additionally, basic biological processes such as mRNA processing and regulation of translation and some specific functional signaling pathways, including the circadian rhythm signaling pathway, were also enriched among DEGs associated with higher epigenomic information (Figure 6B).

Conversely, DEGs associated with lower epigenomic information are enriched in pathways linked to neurodegeneration, inflammation, and cellular stress responses, such as apoptosis, ferroptosis, cell death, and oxidative stress (Figure 6B). The Ras signaling pathway, which stimulates apoptosis and is associated with oxidative stress, neuroinflammation, and neuronal dysfunction during neurodegeneration, is also enriched in cells with lower epigenomic information.⁹⁷ Notably, we found that neurodegenerative disorders (including AD, Parkinson’s disease, and Huntington’s disease) and related pathologies (e.g., β-amyloid binding and Tau protein binding) are highly enriched in astrocytes, indicating that epigenome dysregulation in astrocytes is strongly associated with various neurodegenerative disorders and related pathologies (Figure 6B). This suggests the loss of epigenomic stability as a key mechanism underlying various neurodegenerative disorders.

Upstream regulators and AD genetic risk genes associated with epigenomic information dynamics

We then analyzed the upstream regulators and histone modifications enriched in DEGs associated with epigenomic stability dynamics (Figure 6C). Our findings show that DEGs linked to lower epigenomic information, with higher expression levels in epigenomically eroded cells, are enriched for key components of the Polycomb repressive complex 1 and 2 (PRC1/2), such as EZH2, EED, SUZ12, and H3K27me3 histone modification (Figure 6C). To directly examine the coordinated epigenomic and transcriptomic changes at DEGs in Polycomb-repressed regions, termed Polycomb region Activation Genes in Erosion (PAGEs), we analyzed their epigenomic activity dynamics in epigenomic erosion, AD progression, and CR. Correlation analyses revealed that activation of PAGEs was positively correlated with AD pathology progression and negatively associated with cognitive function and resilience, indicating a protective role of Polycomb repression in preserving cognitive function and mitigating AD progression (Figures S7C and S7D). Interestingly, we also identified an enrichment of various heterochromatin stabilizers and transcriptional silencers, including TRIM28/KAP1, SETDB1, and SMRT (Figure 6C). Moreover, consistent with the functional pathway enrichment analysis, which highlighted inflammatory pathways among DEGs with lower epigenomic information (Figure 6B), we also observed that those genes are enriched for immune regulators like RELA, MAF, and IRF8 (Figure 6C). We also observed enrichment in components of the cohesin complex (e.g., SA1, SMC3) (Figure 6C).

Given that several top-ranked DEGs associated with epigenomic information dynamics are linked to AD-GWAS risk variants (Figures S7A and S7B), we investigated how dysregulation of AD-GWAS risk genes correlates with epigenomic information dynamics (Figure 6D). We obtained the AD risk genes (with p < 5 × 10⁻⁸) from two recent studies and identified the differentially expressed AD risk genes by overlapping DEGs correlated with epigenomic information dynamics with AD-GWAS risk genes (Figure 6D).^42,44 Besides APOE, we identified more AD risk genes positively associated with epigenomic erosion, including CLU, BCAM, CTSB, and JAZF1 (Figure 6D). Conversely, some AD risk genes showed higher expression levels in cells with greater epigenomic information, including NDUFAF6, ANKH, and YWHAQ (Figure 6D).

We further investigated DARs associated with epigenomic information dynamics. Consistent with global epigenomic relaxation and compromised compartmentalization (Figures 3B–3E), genomic regions with increased accessibility during epigenomic erosion were enriched in heterochromatin, quiescent, and Polycomb-repressed states, whereas regulatory elements such as active promoters and enhancers were more accessible in cells with higher epigenomic information (Figure 6E).

Epigenomic stability correlates with AD pathology progression, cognitive function, and resilience

Next, we explored the relationship between epigenomic stability and the progression of various variables of AD pathology, cognitive function, cognitive decline, and CR. To achieve this, we performed sample-level embedding, based solely on epigenomic information, and examined the correlations with these pathological and cognitive variables (Figures 6F, S7E, and S7F; STAR Methods). Notably, we found that sample-level epigenomic information is positively correlated with higher cognitive function, slower cognitive decline, and CR (Figures 6F and S7F). Conversely, epigenomic information is negatively associated with a range of clinical and neuropathological features of AD, including Braak stage, global AD pathology, tangle density, NFT burden, amyloid level, and diffuse/neuritic plaque burden (Figures 6F and S7F).

To examine molecular associations between epigenomic stability, AD pathology, and cognition, we analyzed DEGs and DARs linked to epigenomic information dynamics, AD pathologies (e.g., global pathology, tangle density, plaque burden), cognitive functions, cognitive decline slopes, and CR across cell subtypes (Figures 6G, S8A, and S8B). Consistent with sample-level trends (Figure 6F), fold changes of DARs and DEGs linked to epigenomic information dynamics were positively correlated with cognitive function and resilience but negatively correlated with the progression of various AD pathologies (Figures 6G, S8A, and S8B). Additionally, we generated genome-wide visualizations of epigenomic dynamics and performed chromatin-state enrichment analysis of the DARs (Figures S8C, S8D, S9A–S9C, and S10A–S10D). We found that, during the progression of AD pathologies, repressive genomic regions become more accessible, while active chromatin regions (i.e., promoters and enhancers) are repressed (Figures S8C and S9A–S9C; Data S1, page 6A). This compromised epigenomic identity emerges early in AD, predominantly affecting the most vulnerable entorhinal cortex and hippocampus (Figures S10B and S10C). With advanced AD progression, these epigenomic erosion signatures intensify in lateAD and extend into other cortical areas, highlighting the progressive and region-specific epigenomic deterioration in AD (Figures S10B–S10D). Conversely, repressive heterochromatin and quiescent regions are more repressed, while cis-regulatory regions are more active in individuals with higher cognitive function and resilience (Figures S8D, S9A, and S9B; Data S1, page 6B).

In summary, we identified that the loss of epigenomic information is associated with the progression of AD pathology, while maintaining epigenomic information and identity is crucial for preserving cognitive function and resilience. Our findings highlight the importance of maintaining epigenomic stability to support cognitive function and to combat the progression of AD.

DISCUSSION

Single-cell studies of the human brain and AD mouse models have uncovered cell-type-specific transcriptomic changes and key pathways in AD.^{4,26–30,35,83,98–104} However, the dynamics of the epigenome and transcriptome across brain regions and cell types in relation to AD pathology, cognitive function, and resilience remain elusive. By analyzing the single-cell epigenomic and transcriptomic landscape from aged and AD individuals with varying levels of pathology across larger cohorts and six brain regions, our study reveals compromised chromatin compartmentalization and loss of epigenomic identity in AD. We identify vulnerable brain regions and cell types with coordinated epigenomic dynamics and cell fraction changes in AD and demonstrate that preserving epigenomic stability is crucial for cognitive function and resilience, while epigenomic erosion is associated with AD pathology progression. We delineate regulatory networks that govern epigenomic stability, providing a mechanistic foundation for therapeutic targeting in AD.

The epigenome comprises DNA methylation, histone modifications, chromatin accessibility, and 3D genome organization, all of which regulate gene expression and maintain cellular identity.^36,105–109 Based on the epigenetic information and erosion theory of aging, epigenomic information increases during development, cell differentiation, and cell fate commitment.^{67,69,110–112} In contrast, aging and aging-related diseases are marked by epigenomic erosion—the loss of this regulatory information.^{38,67–69,113} This is reflected by decreased repressive histone modifications (e.g., H3K9me3, H3K27me3) in heterochromatin regions and diminished active histone marks and chromatin accessibility at active regulatory elements.^{38,67,114–116} Collectively, these changes represent a loss of epigenomic fidelity and identity.^{38,67–69,113} We analyzed large-scale compartment and single-cell epigenomic information changes across brain regions and cell types in AD. Our analyses revealed global epigenomic relaxation and an overall decline in epigenomic information with AD progression but a preserved epigenomic stability in cognitively resilient individuals. Correlation analyses further revealed that epigenomic information was positively associated with cognitive function and resilience and negatively associated with the progression of various AD-related pathologies. These findings align with emerging evidence that chromatin plasticity enhances memory by promoting neuronal excitability and synaptic remodeling.¹¹⁷ Our results, together with prior work, underscore the role of epigenomic rewiring in the formation, maintenance, and decline of memory and cognitive function in development and neurodegenerative disease.

We identified vulnerable brain regions and cell types exhibiting epigenomic alterations and fractional changes linked to AD progression and CR. Epigenomic information loss correlated with cell depletion in AD, while higher epigenomic information levels were associated with cell-type preservation and resilience, underscoring the importance of the maintenance of epigenomic stability. The entorhinal cortex and hippocampus, early sites of AD pathology, showed pronounced epigenomic deterioration, particularly in RELN/TOX3-expressing excitatory neurons, DG granule cells, and pyramidal cells. The DG, a key region for adult neurogenesis, was notably affected, suggesting that the maintenance of epigenomic integrity plays a critical role in supporting neural plasticity.^118,119

Epigenomic information dynamics during AD are tightly linked to glial-state transitions. Epigenomic information increases during microglial and astrocyte activation, suggesting a reprogramming toward reactivity and repair—paralleling patterns seen in cell differentiation.^69,120 However, terminally activated or inflammatory glial cells are prone to subsequent epigenomic erosion—resembling the identity loss observed with cellular senescence.^67,69,121 In oligodendrocytes, such erosion may impair myelination and axonal support, worsening neurodegeneration. These findings highlight the importance of maintaining epigenomic integrity to prevent glial exhaustion and to mitigate AD progression.

Our integrative analyses revealed genes, functional pathways, and upstream regulators linked to epigenomic stability. Genes associated with nuclear and chromatin structure maintenance are highly enriched in cells with higher epigenomic stability, indicating that nuclear structure integrity is essential for the maintenance and preservation of epigenomic identity. Consistently, we observed a global activation of repressive genomic regions associated with LADs and repression of active nuclear speckle regions in AD, suggesting a decline in epigenomic identity and transcriptomic fidelity. Previous studies highlighted the importance of nuclear lamina proteins and heterochromatin scaffolds in maintaining epigenome stability and functional integrity in aging and neurodegenerative diseases.^113,122,123 Our findings highlight a mechanistic axis in which nuclear disorganization and loss of epigenomic stability undermine transcriptional fidelity, contributing to cell-type vulnerability, degeneration, and exhaustion during AD progression and cognitive decline. Conversely, preservation of nuclear architecture and epigenomic stability may help maintain transcriptional accuracy and support cognitive resilience in the aging brain.

We found that the loss of epigenomic information is linked to the activation of genes residing in heterochromatic regions, regulated by heterochromatic organizers and transcriptional repressors such as TRIM28/KAP1, SETDB1, PRC1/2, and the repressive histone modification H3K27me3.^{115,124–127} Consistent with these transcriptomic changes, cells exhibiting greater epigenomic erosion or advanced AD progression displayed increased chromatin accessibility in heterochromatin and Polycomb-repressed regions. Erosion and activation of heterochromatin regions have been implicated in aging.^128–132 These findings suggest that the maintenance of heterochromatin integrity and Polycomb-mediated repression is associated with preserving cellular function and identity in AD. Genes associated with epigenomic erosion were also enriched for the cohesin complex, a key regulator of 3D genome organization and double-strand break repair.^133–135 Cohesin overexpression in PFC with AD pathology has been linked to increased DNA damage.^33,46

In summary, our single-cell epigenomic and transcriptomic atlas across brain regions and cell types offers a valuable resource for understanding epigenetic regulation in AD. We reveal region-specific and cell-type-specific epigenomic deterioration in AD and preservation in CR, highlighting opportunities for epigenetic-based therapeutic development.

Limitations of the study

While we provide a detailed view of the epigenomic and transcriptomic dynamics in AD, several limitations warrant further exploration. Our study is based on postmortem cross-sectional snRNA-seq and snATAC-seq data, which may introduce biases related to tissue quality and data sparsity in single-cell analyses. Future integration with large-scale ROSMAP and SEA-AD datasets, along with the inclusion of single-cell histone modification profiling, Hi-C, DNA methylation, and spatial epigenomic and transcriptomic analyses could provide a more nuanced understanding of regulatory dynamics in AD. While we identified key molecular factors underlying epigenomic and transcriptomic dysregulation across cell types in AD, further investigation is needed to validate erosion mechanisms and to identify therapeutic targets.

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to the lead contact, Manolis Kellis (manoli@mit.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• All snATAC-seq and snMultiome data generated in this study are available through the AD Knowledge Portal on Synapse at the following URL: https://www.synapse.org/Synapse:syn66271521 and https://www.synapse.org/Synapse:syn66271522. snATAC-seq and snMultiome of PFC data were reported in our previous study³⁸ and are available at https://www.synapse.org/Synapse:syn52293417. snRNA-seq data across 6 brain regions from 48 individuals reported in our previous study are available at https://www.synapse.org/Synapse:syn52293442.⁴⁶ The data are available under controlled use conditions set by human privacy regulations. To access the data, a data use agreement is needed. This registration is in place solely to ensure the anonymity of the ROSMAP study participants. A data use agreement can be agreed with either Rush University Medical Center (RUMC) or with SAGE, which maintains synapse, and can be downloaded from their websites (https://adknowledgeportal.synapse.org/). Additional processed data for snATAC and snRNA datasets are available at https://compbio.mit.edu/AD_Multiomic_MultiRegion.

• All original code has been deposited on GitHub (https://github.com/ZunpengLiu/Multi-region_AD).

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

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Human subjects

We selected 111 aged and AD individuals based on the modified NIA-Reagan criteria for AD diagnosis and Braak stage scores from the Religious Orders Study and the Rush Memory and Aging Project (ROSMAP), ongoing longitudinal clinical-pathologic cohort studies of aging and dementia.⁴⁵ The cohort includes 57 controls (32 female, 25 male), 33 individuals with early-stage AD (17 female, 16 male), and 21 individuals with late-stage AD (14 female, 7 male) samples, exhibiting a range of AD-related pathological and cognitive measurements (Table S1). Among the 111 study participants, 110 identified as White, and 1 identified as African American. Participants were balanced by age at death across groups, with a median age of 87 years. The AD groupings were determined using k-means clustering based on seven pathological measures, including a global AD pathology, molecularly specific amyloid-β and tangles, and global cognition proximate to death, as previously described.³⁸ The Religious Orders Study and Rush Memory and Aging Project were approved by the Institutional Review Board (IRB) of Rush University Medical Center. Informed and repository consents were obtained from each participant, and participants also signed an Anatomic Gift Act.

METHOD DETAILS

Clinical, phenotypic, and pathological characterization of ROSMAP participants

Details of the clinical and pathological data collection methods have been previously reported.^3,6,7,45,153 The key pathology traits for ROSMAP individuals analyzed in our study include several measures of AD.^154–159 First, counts of neuritic plaques, diffuse plaques, and neurofibrillary tangles were done on five sectioned stained with modified Bielschowsky which were used to generate Braak stage (a semiquantitative measure of the severity and distribution of neurofibrillary tangle pathology) and a semiquantitative measure of neuritic plaques as recommended by the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD); The global AD pathology burden is a quantitative measure based on standardized and averaged counts of three hallmark features of AD: neuritic plaques, diffuse plaques, and neurofibrillary tangles. These three AD pathologies were assessed through microscopic examination of silver-stained slides from five brain regions: midfrontal cortex, midtemporal cortex, inferior parietal cortex, entorhinal cortex, and hippocampus. Regional counts for each pathology type were standardized by dividing their respective standard deviations. The five scaled regional measures were then averaged to generate summary scores for neuritic plaques, diffuse plaques, and neurofibrillary tangles. Finally, these three summary scores were averaged to calculate the global AD pathology burden score.^153,157 The global AD pathology score has been used in prior studies to investigate the transcriptomic dynamics in AD.^46,47

Each participant underwent an assessment with 21 cognitive tests, 11 of which informed clinical diagnosis of AD, mild cognitive impairment (MCI), and AD dementia each year.^6,160,161 Further, 19 cognitive tests were Z-scored and averaged to make a global measure of cognitive function and separate measures of five cognitive domains including episodic memory, semantic memory, working memory, perceptual speed, and visuospatial ability.^7,162 To assess cognitive decline, the individual-specific slopes of cognitive function decline over time were estimated using a linear mixed-effects model, with annual cognitive test scores as the longitudinal outcome, adjusting for age, sex, and years of education.¹⁶³ A steeper cognitive decline slope indicates a more rapid decline in cognitive function, corresponding to a lower slope value, whereas a flatter cognitive decline slope represents a slower decline, reflected by a higher slope value. At the time of death, select clinical data are reviewed without reference to neuropathologic data to make a final clinical diagnosis. Final cognitive consensus diagnosis, clinical consensus diagnosis of cognitive status at time of death.¹⁶⁴ Postmortem interval is the time between death and tissue preservation. APOE is genotyped.¹⁶⁵

Tissue dissection criteria, preparation, and nuclei isolation

Multiregion brain dissections were performed using anatomical criteria consistent with those described in a previous study.⁴⁶ All tissue dissections were done on a bed of dry ice using either a fine-toothed razor saw (for cortical regions) or a jewelers saw with diamond wire (for subcortical regions). Region-specific descriptions are as follows. (1) EC: full thickness cortex from the EC (Brodmann area; BA 28); take at the level of the amygdala. Avoid amygdala. Minimize white matter. (2) HC: take from the last slab containing HC. If the last slab has less than 5 mm of HC, take from the next slab anterior. Collect a full cross section. (3) TH: take from the first slab with thalamus. Include the most medial aspect. (4) AG: full thickness cortex from the AG (BA 39/40); take from the first or second slab posterior to the end of the HC. Minimize white matter. (5) MTC: full thickness cortex from the middle temporal gyrus (BA 22); take as close to the level of the anterior commissure as possible. Minimize white matter. (6) PFC: full thickness cortex from the frontal pole (BA 10); take from the lateral side of the first or second slab. Minimize white matter.

To isolate nuclei from frozen post-mortem brain tissue, we followed the procedure described in previous studies.^38,47 Specifically, we homogenized brain tissue in homogenization buffer (700 μL), filtered homogenate through a 40 μm cell strainer (Corning, NY), and added working solution (450 μL). We placed the resulting solution on top of a 30%–40% OptiPrep density gradient (750 μL 30% OptiPrep solution, 300 μL 40% OptiPrep solution) as a 25% OptiPrep (Sigma-Aldrich, St-Louis, MO) solution and centrifuged on a tabletop centrifuge at 4°C for 5 minutes at 10,000 × g. We then collected and transferred the resulting nuclei pellets to a new tube, washed twice with 1 mL ice-cold PBS/0.04% BSA (centrifuged 3 minutes, 300 × g, 4 °C), and resuspended pellets in 100 μL PBS/0.04% BSA.

Single nucleus ATAC library construction and sequencing

For droplet-based snATAC, we spun down the nuclei, resuspended them in 1x Diluted Nuclei buffer, and adjusted their concentration to 2,500 nuclei per μL. We then prepared libraries targeting 5,000 nuclei per brain region and individual using the Chromium Single-Cell ATAC Reagent Kit v1(10x Genomics, Pleasanton, CA) according to the manufacturer’s instruction and sequenced the pooled libraries using NovaSeq 6000 S2 flowcell (100 cycles, Illumina).

Single nucleus multiome library construction and sequencing

We used the Chromium Next GEM Single Cell Multiome ATAC and Gene Expression Reagent kit A (10x Genomics, Pleasanton, CA), which allows for simultaneous transcriptomic and chromatin accessibility profiling from the same nuclei. For this purpose, we isolated nuclei as described above. After the final wash, we resuspended nuclei pellets in 1x Diluted Nuclei buffer/1mM DTT according to the manufacturer’s protocol at a concentration of 2,500–5,000 nuclei per μL. We transposed 7,000 nuclei in bulk and loaded them on a Chromium controller for GEM generation. We then splitted the pre-amplified sample and used it as input for the snATAC and Gene expression (GEX) library construction. We generated separate pools of snATAC and snGEX libraries and sequenced them on the NovaSeq 6000 S4 flowcell (150 cycles, Illumina) according to the suggested sequencing depth and read cycles for each library type.

QUANTIFICATION AND STATISTICAL ANALYSIS

Preprocessing of snRNA-seq, snATAC-seq and snMultiome data

For snRNA-seq data, gene counts matrices were obtained by aligning snRNA-seq reads to the GRCh38 (hg38) genome using Cell Ranger software (v.7.2.0) (10x Genomics) with all default parameters adding “–include-introns”.¹⁶⁶ We used the Scanpy (v.1.9.3) for the quality control, data integration, normalization, dimensionality reduction, clustering, and visualization.¹³⁶

For snATAC-seq data, BCL files were demultiplexed into raw Fastq format sequencing data using cellranger-atac mkfastq (v.1.1.0). We mapped snATAC-seq reads to the human GRCh38 reference genome and generated bam and fragment files with the identified transposase cut sites for each sample using the cellranger-atac count function.

Raw data generated using the 10x Multiome platform were processed using the 10x cellranger-arc count toolkit (v.2.0.2) with manufacturer’s instructions using the GRCh38 human reference genome. This process generated fragment files with identified transposase cut sites from 10x Multiome snATAC data, along with gene expression matrices containing filtered and aligned snRNA-seq reads from 10x Multiome snRNA data.

For downstream analyses of snATAC and snRNA data, including dimensionality reduction, clustering, peak calling and annotation, identification of cell-type-specific genes, and TF analysis, we utilized the Scanpy, Muon, and ArchR toolkits.^136,,139,141

snATAC data processing and integration

We used ArchR (v.1.0.2) to process the fragment files generated from the single-omic snATAC-seq data and the snATAC-seq data derived from snMultiome and generate the arrow files and ArchR project for the downstream analysis.¹⁴¹ We used ArchR to check the quality of snATAC nuclei, and filter doublets using filterDoublets function. We filtered the cells to include those with TSS enrichments greater than 2 and number of fragments greater than 1,000. To analyze and visualize the fragment size distribution in all snATAC libraries, we used the ArchR plotFragmentSizes function. We also used Muon (v.0.1.3) to process the fragment files and generate the AnnData .h5ad files.¹³⁹

For clustering analysis of the integrated snATAC data from single-omic snATAC-seq data and snATAC-seq data derived from snMultiome, we utilized multiple rounds of clustering analysis based on the tile matrix or peak matrix. The first round of clustering analysis was performed on a 5-kb tile matrix. We divided the genome into 5-kb consecutive windows and then we determined the count of accessible fragments within each window in each cell using the Muon functions muon.atac.tl.locate_fragments and muon.atac.tl.count_fragments_features.¹³⁹ Next, we normalized the counts of the 5-kb tile matrix by term frequency-inverse document frequency (TF-IDF) and performed latent semantic indexing (LSI) on the scaled matrix to obtain latent components. According to the instruction, the first LSI component was removed as it is typically associated with quality control measures, such as the number of counts/fragments per cell. To remove batch effects between the library generation type (i.e., single-omic snATAC-seq data and the snATAC-seq data derived from snMultiome), we used Harmony for batch correction with the library construction type (i.e., single-omic snATAC-seq data and the snATAC-seq data derived from snMultiome) as the key.^136,167 We then computed a cell neighborhood graph. To visualize all snATAC nuclei in a two-dimensional embedding, UMAP was created using the Scanpy (v.1.9.3) function scanpy.tl.umap with the LSI latent components corrected by Harmony.¹³⁶

After identifying over 1 million cCREs across 67 subtypes, we performed a second round of clustering using the same approach as with the 5-kb tile matrix, applied to the cCRE matrix across all snATAC data, to enhance the visualization of fine-grained brain subtypes. To achieve high-quality cell type clustering and annotation for the snATAC data, we also performed a second round of quality control and doublet removal at the major cell class level. We then subsetted the cells from each major cell class, repeated the TF-IDF normalization, dimensionality, batch correction, and UMAP embedding in each cell class. We then discarded the Leiden clusters that showed abnormal counts of fragments, reads in blacklists, ambiguous cell type marker signals. UMAP visualizations of AD pathology, cell type clustering, and annotations were generated using the scanpy.pl.umap function applied to the AnnData .h5ad object. To calculate the gene-wise activity scores in snATAC data, we utilized the default ArchR distance-weighted gene score model that aggregates local chromatin accessibility signals around gene regions.¹⁴¹ To visualize the gene score for marker genes, we used the MAGIC algorithm implemented in ArchR to impute gene scores by smoothing signals across nearby cells to reduce sparsity noise in snATAC data.^141,168

snRNA data processing and integration

We used Scanpy (v.1.9.3) to process and integrate the filtered expression matrices generated from the single-omic snRNA-seq data and snRNA-seq data derived from snMultiome, encompassing data integration, quality control, normalization, batch correction, dimensionality reduction, and clustering.¹³⁶ For quality control, we detected and filtered doublets using the scrublet package (v.0.2.1) implemented in Scanpy.^136,137 We also filtered out nuclei with fewer than 200 gene counts, over 20% mitochondrial content, or over 5% ribosomal content. For normalization and batch correction across all snRNA libraries, we followed the Scanpy workflow. This involved normalizing read counts, log-transforming the data matrix, and performing principal component analysis (PCA) using scanpy.pp.pca function. Batch effects from different library construction methods were corrected using the scanpy.external.pp.harmony_integrate function, with the snRNA library type designated as the batch key. Subsequently, we computed the neighbor graph and UMAP for dimensionality reduction, and identified cell type clusters using the Leiden algorithm by scanpy.tl.leiden function in Scanpy.^136,169 To achieve high-quality cell type clustering and annotation, we conducted a second round of quality control and doublet removal at the major cell class level. We subsetted cells from each major cell class and repeated the processes of normalization, dimensionality reduction, batch correction, and UMAP embedding. We discarded clusters that exhibited abnormal read counts, an unusual number of expressed genes, or ambiguous cell type marker signals.

Annotation of cell classes, subclasses, and subtypes

Annotation of cell classes, subclasses, and subtypes in snRNA: Multiple approaches were taken to annotate the snRNA data. For the annotation of the major cell type classes, we examined the gene expression of known canonical marker genes for each major cell class. Marker genes included SLC17A7 for excitatory neurons, GAD2 for inhibitory neurons, GFAP for astrocytes, MOBP for oligodendrocytes, VCAN for OPCs, and P2RY12 for microglia. We annotated cell subtypes using previously published marker genes and sn/scRNA-seq data.^{29,38,47,170–172} The detailed annotation procedure for the annotations of cell subtypes based on previously published scRNA-seq data (Allen Institute’s cell types database; https://portal.brain-map.org/atlases-and-data/rnaseq/human-multiple-cortical-areas-smart-seq) as described in previous studies.^46,47 Considering some of the snRNA data of PFC samples are obtained from previous data,^38,47 the cell type annotations were also mapped to our snRNA datasets for the verification. To further verify the cell type annotations, we also trained the predictive models for the annotations of major cell classes based on previous datasets using CellTypist,^{47,170,173,174} and then predicted the cell type annotations across cell class and subtype level in our snRNA datasets for comparison and validation. The high-resolution cell subtype annotations were further grouped into the cell subclasses. The inhibitory neurons show some major clusters, including VIP⁺, SST⁺, PVALB⁺, LAMP5⁺, PAX6⁺, and MEIS2⁺ subclasses, as indicated in previous studies.^15,16,170 The excitatory neurons were also grouped into different cell subclasses based on their organizations. For the cortical excitatory neurons, they are grouped into different layers (i.e., Exc L2–3 IT, L3–4 IT, L3–5 IT, L4–5 IT-1, L4–5 IT-2, L5–6 IT, L5/6 IT Car3, Exc L6 IT, L5/6 NP, L5 ET, L6 CT, and L6b), as indicated in previous studies.^{15,16,170,175} Other excitatory neurons, specific to a certain brain region were labeled (i.e. Exc EC, Exc CA pyramidal cells, Exc DG granule cells, Exc HC, and Exc TH). Additionally, we identified 6 vascular cell types: pericyte (Per), endothelial (End), smooth muscle cells (SMC), fibroblast (Fib), choroid plexus ependymal (CPEC), and ependymal (Epd) subpopulations.

Annotation of cell classes, subclasses, and subtypes in snATAC: For the annotation of major cell classes in snATAC data, we examined gene activity scores of known marker genes using ArchR. To achieve a uniform annotation in snRNA and snATAC, for the cell subtype annotation in snATAC data, we subsetted cells for each major cell class, and repeated the processes of TF-IDF normalization, LSI, batch correction, and UMAP embedding in each cell class, cell subtype labels were transferred directly from snRNA to snATAC by the matched barcodes of the nucleus profiled from snMultiome data. Then we adopted the majority voting strategy as described in CellTypist, a method for inferring the cell type annotations based on the high-resolution Leiden clusters and known cell type annotations from a subset in each Leiden cluster, to annotate all nucleus for snATAC data.^173,174 Considering some of the snATAC data of PFC samples are obtained from previous data,³⁸ the cell type annotations were also mapped to our snATAC datasets for the verification of the cell type annotation.

Cross-modality validation of cell type annotations: The uniform and shared cell type annotations in both snRNA and snATAC data were validated using many different approaches: (1) We visualized the canonical marker genes of major cell classes by the snRNA gene expression and snATAC gene activity; (2) We calculated the cell-subtype-specific genes across 67 cell subtypes, and identified more than 7,000 shared cell-subtype-specific genes across two modalities; (3) We computed RNA gene set scores for snATAC-derived cell-subtype-specific genes and observed strong enrichment in the matching subtypes within the snRNA data; (4) We visualized the canonical marker genes and newly identified cell-subtype-specific genes in excitatory and inhibitory neurons; (5) The cell subtype annotations were also verified by the motif deviation enrichment analysis; (6) We confirmed that snRNA gene expression and snATAC gene activity showed agreement in the cCRE module analysis across subtypes.

Identification of cell-subtype-specific genes

Cell-subtype-specific genes across 67 cell subtypes in snRNA datasets were analyzed using the Wilcoxon rank-sum test as statistical method in the scanpy.tl.rank_genes_groups function with the cutoff of FDR-adjusted p value < 0.05 and log₂(fold change) > 1. The cell-subtype-specific genes in snATAC datasets were identified using the ArchR function getMarkerFeatures with FDR-adjusted p value < 0.05 and log₂(fold change) > 1. Shared cell-subtype-specific genes across 67 cell subtypes in snRNA and snATAC modalities were identified by the direct overlap of the cell-subtype-specific genes in each cell subtype. The gene lists for cell-subtype-specific genes for each cell type are listed in Table S1.

Cell-type composition differences among brain regions

For comparing the relative abundance of the defined 67 high-resolution cell subtypes across 6 brain regions in both snATAC and snRNA, the fraction of cells of a cell subtype was computed relative to all the cells from a brain region. Then the fraction of cells in each cell subtype was Z-scored across 6 brain regions as the relative abundance for the visualization.

Identification of cCREs and cCRE-gene pairs across cell types

To identify cCREs across 67 cell subtypes, we utilized ArchR (v.1.0.2) functions addGroupCoverages and addReproduciblePeakSet to call accessible chromatin peaks using MACS2 (v.2.2.6).¹⁴² We identified 501-bp fixed-width reproducible peaks in each cell subtype by checking the reproducibility of each peak across pseudo-bulk replicates in ArchR. Given our large snATAC datasets, we adjusted parameters for the peak calling, e.g., we used “minCells = 40, maxCells = 10000, minReplicates = 2, maxReplicates = 40” for addGroupCoverages function and “maxPeaks = 500000, cutOff = 0.01” for addReproduciblePeakSet function. We finally identified a union set of 1,010,153 cCREs across 67 cell subtypes. cCREs-gene links were identified using ArchR addPeak2GeneLinks function with parameters “corCutOff = 0.5, FDR = 1e-4”. Thalamic-specific TF-regulated cCRE-gene pairs were identified by overlapping TF binding sites of thalamic-specific TFs with cCRE-gene pairs in thalamic neurons.

Motif deviation, TF enrichment network, and TF footprinting analysis

We computed single-nucleus TF motif enrichment using the chromVAR algorithm implemented in ArchR, based on motifs annotated in the CIS-BP database.^176,177 This analysis generates single-cell resolution TF deviation scores across over 1.2 million snATAC nuclei, covering 870 TF motifs.

For TF enrichment analysis, the TF deviation matrix was transposed into a TF-by-cell format, followed by PCA using Scanpy function scanpy.pp.pca for dimensionality reduction. We then applied UMAP embedding using scanpy.pp.umap to project the PCA-transformed TF activity space into a lower-dimensional representation. A neighborhood graph was constructed using scanpy.pp. neighbors, which computes pairwise distances and connectivity among TFs. The resulting TF enrichment network was visualized in Cytoscape, revealing functional TF clusters with coordinated activity. To determine cell-type-specific TFs, single-cell TF deviation scores were aggregated into a pseudo-bulk matrix across cell types (7 major cell classes or 67 cell subtypes) and 356 snATAC-seq libraries. A linear mixed-effects model (TF enrichment ~ Cell-type-specific TF deviation score + Age + Sex + AD pathology groups + Random effect of samples) was applied to estimate TF enrichment coefficients. The FDR-adjusted p values and coefficient values were integrated into the TF enrichment network, generating a weighted enrichment score (coefficient × −log₁₀(FDR)).

TF footprinting analysis in thalamic neurons was performed by identifying motif positions using the getPosition function in ArchR.¹⁴¹ TF footprints were computed using getFootprint after coverage normalization, and footprint profiles were visualized with plotFootprints function in ArchR.

Identification of the cCRE modules

We employed NMF to classify over one million cCREs across 67 cell subtypes into distinct cCRE modules.⁵⁷ Given the sparse nature of snATAC data, we first aggregated chromatin accessibility signals across the cCREs and cell subtypes from six brain regions, creating a matrix (denoted as $V$) with dimensions $N\times \mathrm{M}$. Here, $N$ represents the cCREs, and $M$ represents the combined dimensions of the 67 cell subtypes and 6 brain regions.

We then applied NMF, utilizing the sklearn.decomposition package in Python, to factorize the non-negative matrix $V$ into two smaller non-negative matrices: the basis matrix $W$ (with dimensions $N\times R$) and the coefficient matrix $H$ (with dimensions $R\times M$), following methodologies used in prior research.^12,25 This factorization follows:

$$V\approx W\times H$$

The basis matrix $W(N\times R)$ characterizes module-related accessible cCREs, where $N$ corresponds to all identified 1 million cCREs, and each column in $W\left(R\right)$ represents a module, defining the participation of each cCRE in that module. The coefficient matrix $H(R\times M)$ defines the contribution of each cell subtype and brain region to these modules, with each row $R$ corresponding to a module and each column $M$ representing the cell subtype/brain region combination.

To assign the identified cCREs to specific modules and relate these modules to the respective cell subtypes, we utilized the basis matrix $W$ and the coefficient matrix $H$. The entries in basis matrix $W$ are non-negative and represent the degree to which each original feature (cCRE) is associated with each latent feature (module). We calculated a “feature score” for each cCRE based on the basis matrix $W$, reflecting its specificity to a module, using the “Kim” method as described previously.^12,25,57 The feature score ranges from 0 to 1, where a higher score indicates greater specificity. We considered a cCRE as highly specific to one module if it met two criteria: (1) a feature score above the median plus one standard deviation, and (2) the maximum contribution to a module exceeding the mean contribution across all modules. cCREs with feature scores below the median minus one standard deviation were aggregated into the first module (M1) as a non-specific module. cCREs with feature scores between the median minus one standard deviation and the median plus one standard deviation were considered shared across different modules.

Furthermore, we used the coefficient matrix $H$ to associate modules with distinct cell subtypes across brain regions. The values in $H$ indicate the weights of cell subtypes in each module, and we used the highest coefficient value to determine the primary association of each cell subtype with a particular module, thereby linking each cell subtype most strongly with the corresponding module.

Verification of cCRE modules by the cCRE module score and cCRE embedding

We verified the cell-type-specific enrichment of cCRE modules by computing module scores using Scanpy function scanpy.tl. score_genes based on cCREs in each module, and visualized the scores in cell embeddings using UMAP.¹³⁶ We also computed the cCRE embeddings to cluster, verify, and visualize the cCREs within the UMAP embedding. First, we transposed the cell × cCRE count matrix into a cCRE × cell matrix and applied the same processes used for cell clustering, including TF-IDF normalization, LSI, and UMAP dimensionality reduction. To correct chromosome effects, we employed the BBKNN method in Scanpy, using chromosomes as the key.¹³⁸ The corrected neighborhood graph was then used for UMAP analysis. In this cCRE embedding, cCREs from the same modules, as defined by the NMF method, clustered together.

Linking gene activity, expression, and functional pathways to cCRE modules

With unified cell subtype annotations across 67 subtypes and 6 brain regions in both snATAC and snRNA datasets, we linked gene activity (snATAC) and gene expression (snRNA) to the computed 123 cCRE modules. First, we obtained the snATAC gene activity scores and snRNA gene expression levels at the pseudo-bulk level for all genes aggregating across 67 cell subtypes and six brain regions. For each cCRE, the nearest gene was identified based on union peak annotations from ArchR. To compute gene activity per cCRE module, we averaged activity scores of genes with transcription start sites (TSS) within 3 kb of cCREs. Likewise, mean expression values of nearest genes were used to represent each module’s gene expression. Functional enrichment of each cCRE module was assessed using GREAT (v1.26.0) against a background of all peaks, using GO biological processes and visualized by log₁₀-transformed hypergeometric p values.¹⁴³

Interpretation of GWAS-risk variants with cCREs and modules

We used the liftOver tool with default parameters to convert the cCREs identified across 67 subtypes and cCREs within each cCRE module from hg38 to hg19 genomic coordinates.¹⁷⁸ GWAS heritability enrichment analysis was carried out using linkage disequilibrium (LD) score regression (LDSC, v.1.0.1) following the tutorial guidelines.⁵⁸ Enrichment was defined as the proportion of heritability normalized by the proportion of SNPs covered, with standard errors used for p value estimation. GWAS summary statistics for neuropsychiatric, neurodegenerative, and complex traits were obtained from LDSC (https://alkesgroup.broadinstitute.org/) and CATlas.^58,179 Additional AD GWAS summary statistics from multiple studies were included.^42–44

Epigenomic compartment inference and dynamics in AD

We inferred global active and repressive chromatin compartments from chromatin accessibility profiles using snATAC data at 100-kb resolution, applying a two-state Hidden Markov Model (HMM). Pseudo-bulk ATAC signals were computed by aggregating fragments from seven major cell classes across six brain regions using the ArchR function getGroupBW, with parameters tileSize = 100000 and normMethod = “nFrags”, followed by Z-score normalization. To ensure data quality, we excluded 100-kb bins overlapping with the blacklist and retained only those with at least 90% annotation of chromatin states as defined in EpiMap (https://compbio.mit.edu/epimap/), resulting in 27,993 high-quality 100-kb bins for downstream analysis. Overlap analysis for these bins was conducted using BEDTools (v.2.30.0).¹⁵¹ For compartment calculation, we used the HMMt R package (v.0.1) (https://github.com/gui11aume/HMMt). To enable comparative analyses across cell types, brain regions, and AD pathology groups, we concatenated pseudo-bulk ATAC signal profiles across all conditions and jointly inferred compartments, analogous to the concatenated modeling approach in chromHMM.¹⁸⁰ The resulting binary state assignments (active or repressive compartments) across the 27,993 100-kb bins were grouped into 25 epigenomic compartment groups using K-means clustering (R kmeans function), and visualized using ComplexHeatmap (v.2.14.0) in R (v.4.2.2).¹⁴⁵

To quantify epigenomic compartment dynamics between nonAD and AD across chromosomes and cell types, we computed compartment dynamic scores based on 100-kb genomic bin transitions. For each chromosome, we counted the number of bins switching from active to repressive (nAtR) and from repressive to active (nRtA). The compartment dynamic score, defined as (nRtA − nAtR) / (nRtA + nAtR), captures the net directional shift: positive values indicate global chromatin activation, whereas negative values reflect repression.

Validation and annotation of epigenomic compartments

We validated the inferred active and repressive chromatin compartments using complementary genetic and epigenetic approaches:

1. Dimensionality reduction. We applied LSI followed by UMAP to the transposed 100-kb genomic bin-by-cell matrix of snATAC signals across all cells or within each major cell class. For global embedding, LSI components were used to construct the neighborhood graph for UMAP visualization, revealing chromosome-specific patterns among 100-kb bins. For cell-class-specific analyses, the bin-by-cell matrix was subsetted, transposed, and analyzed using LSI followed by BBKNN (with chromosomes as the batch key) to correct for chromosome effects prior to UMAP. Clear separations of active and repressive compartments emerged across cell classes.

2. CpG density and A/T content analysis. Using the annotatePeaks function in Homer (v.4.11.1),¹⁵² we calculated the CpG density and A/T content for each 100-kb bin and mapped them to the inferred compartments. Differences in A/T content between active and repressive compartments were statistically assessed using a Wilcoxon rank-sum test implemented in ggpubr R package (v.0.6.0) (https://github.com/kassambara/ggpubr).¹⁴⁶

3. Chromatin state enrichment. We integrated EpiMap chromatin states from multiple brain regions (e.g., frontal cortex, hippocampus, middle frontal area, angular gyrus; biosample IDs BSS00369, BSS00371, etc.).⁶² For each bin, we calculated log₂ fold enrichment of 18 chromatin states, averaged across replicates, and mapped them to inferred compartments.

4. Nuclear compartment domain enrichment. We computed relative enrichment of LADs and SPADs across 100-kb bins using the reference domain BED files, confirming preferential localization of repressive compartments to LADs and active compartments to SPADs.⁶¹

Genome browser visualization of chromatin accessibility and cCRE-gene links

To visualize genome-wide chromatin accessibility across brain regions and cell classes, we generated BigWig files at 100-kb resolution using the ArchR function getGroupBW, normalizing by the total fragment count with parameters “tileSize = 100000” and “normMethod = “nFrags”. To assess accessibility differences between lateAD and nonAD, we computed log₂(fold change) values using bigwigCompare from the deepTools suite (v.3.5.1), comparing lateAD to nonAD profiles within each cell class.¹⁸¹ Resulting BigWig files were visualized in the IGV Genome Browser (v2.15.4).¹⁴⁷ cCRE-gene regulatory links were identified using ArchR’s getPeak2GeneLinks function and visualized in IGV.

Estimation and validation of epigenomic information at single-cell resolution

Given that cCREs are highly enriched in promoters and enhancers, key cis-regulatory regions in the human genome, we hypothesized that chromatin accessibility within these elements reflects meaningful regulatory information underlying cellular identity and epigenomic stability. To quantify epigenomic information at single-cell resolution and study its dynamics across brain regions, cell types, AD progression, and cognitive resilience, we first identified cCREs in nonAD control samples using ArchR. We then computed the epigenomic information score for each nucleus as the fraction of reads located within these nonAD-defined cCREs, reflecting the proportion of biologically relevant regulatory signals versus background noise.

To validate our epigenomic information scores, we employed multiple independent external references, including: (1) human brain atlas (HBA) snATAC-seq data, (2) H3K27ac ChIP-seq data from sorted human brain cell types, and (3) chromHMM-based chromatin state annotations from various brain regions.^25,41,62,70 For HBA and H3K27ac references, we calculated the fraction of reads overlapping HBA snATAC or H3K27ac cCREs at the single-cell level across all snATAC nuclei. For chromHMM-based epigenomic information validation, we adapted a previously described approach to analyze the distribution of reads across active and repressive chromatin states using the following steps: (1) We retrieved chromatin state annotations (hg38) from the EpiMap portal using biosample IDs from the frontal cortex (BSS00369, BSS00371), hippocampus (BSS01124–BSS01126), middle frontal area (BSS01271, BSS01272), and angular gyrus (BSS00077, BSS00078). (2) For each nucleus, we computed the fraction of fragments mapping to each of 18 chromatin states, yielding a 1.2 million nuclei × 18 chromatin state matrix. (3) We applied Z-score normalization across nuclei for each chromatin state. (4) We summed the normalized values per cell, assigning positive weights to active states (Enhancer, TSS, Tx) and negative weights to repressive states (Quiescent, Repressed, Heterochromatin). (5) We then averaged values across the four brain regions and applied a final Z-score transformation to generate the composite epigenomic information score per nucleus.³⁸ A higher positive score indicates greater accessibility in active chromatin regions and repression of repressive genomic regions, suggesting a more stable epigenome in the corresponding cell.

Correlation analysis across all validation methods demonstrated strong concordance, with Spearman correlation coefficients > 0.97 and p values < 2.2 × 10⁻¹⁶, confirmed the consistency and biological validity of our epigenomic information measure.

Glial cell activation scoring, trajectory analysis, and integration with predefined glial states

To investigate glial cell activation states, we curated gene sets associated with activation and inflammation for microglia, astrocytes, and oligodendrocytes from prior studies. For microglia, activation-associated genes included TPT1, DUSP1, B2M, TREM2, CCL2, CCL3, APOE, AXL, ITGAX, CD9, C1QA, C1QC, CTSS, CSF3R, CX3CR1, SLC2A5, TMEM119, and CD68.^{32,76,81–83} Astrocyte activation genes include C4B, GFAP, C3, SERPINA3, VIM, AQP4, HIF3A, S100B, CD44, SLC1A2, and NES.^76–79 Oligodendrocyte inflammation genes include C4B, IRF7, S100B, TNFRSF12A, IL18, NEAT1, and SERPINA3.^76,80 Glial activation or inflammation scores were calculated by using the AddModuleScore function in ArchR, based on these gene sets for each glial cell type.

To study the activation trajectories in microglia and astrocytes, we applied ArchR’s addTrajectory function, using deciles of activation scores as reference groups. Gene activity dynamics of the canonical glial activation genes along the trajectory were visualized using the plotTrajectory function.

To integrate the microglial and astrocyte activation trajectories defined in this study with previously characterized microglial and astrocyte cell states, we first obtained cell state annotations from large-scale datasets, ROSMAP and SEA-AD.^32,35,71 Canonical correlation analysis (CCA) was performed using Seurat (v4.4.0) to align our snATAC-derived gene activity profiles with these reference states.¹⁴⁰ Each ATAC nucleus was assigned a “predicted cell state” based on the highest prediction score, allowing us to map our activation trajectories onto established microglial and astrocyte cell states.

AD GWAS enrichment at single-cell resolution in microglia and astrocytes

To explore the association between cell states and genetic risk for AD, we applied SCANVAGE (v.1.0.2),⁸⁵ a method that quantifies GWAS enrichment at single-cell resolution. Fine-mapped AD-associated single-nucleotide polymorphisms (SNPs) were obtained from CAUSALdb,¹⁸² which applies FINEMAP (v1.4) to AD GWAS summary statistics.^43,148 For each cell type, we supplied SCAVENGE with the cCRE-by-cell matrix alongside the fine-mapped SNP list. The resulting Trait Relevance Score (TRS) reflects the enrichment of AD genetic risk variants at the single-cell level.

Cognitive resilience score

To study the associations between epigenomic information, AD pathology, and cognitive resilience, we computed the individual-level cognitive resilience score (CR score) by fitting a linear regression model between global cognitive function and global AD pathology.^46,47 In brief, a linear regression was fit between global cognitive function and the global AD pathology using the R (v.4.2.2) stats function $lm$. For each individual, CR score is the difference (residual, $\mathrm{y}-\stackrel{\mathrm{ˆ}}{\mathrm{y}}$) between the actual value of the observed global cognitive function ($\mathrm{y}$) and the predicted value of the cognition function ($\stackrel{\mathrm{ˆ}}{\mathrm{y}}$) by the model based on their level of global AD pathology ($\mathrm{x}$). To have more power and samples for the linear regression analysis, we combined the ROSMAP individuals used in our study and the large-scale snRNA dataset of PFC.⁴⁷

Sample-level epigenomic information and its associations with AD pathology, cognitive function, and resilience

To investigate the relationship between epigenomic information dynamics and AD pathology progression, we generated a UMAP embedding based solely on sample-level epigenomic information computed for each cell using snATAC data from 1,217,165 nuclei across 67 cell subtypes. To ensure robust representation across samples, we implemented a bootstrapping strategy, resampling cells from each subtype 100 times across 356 snATAC libraries. For each bootstrap iteration, epigenomic information was calculated per subtypes, resulting in a sample-level epigenomic information matrix comprising 356 libraries × 67 subtypes × 100 bootstraps. Dimensionality reduction via UMAP was applied to visualize the associations between epigenomic information and AD pathology. We then mapped individual-level AD pathological estimates defined from the ROSMAP metadata, including global AD pathology, neurofibrillary tangle burden, tangle density, neuritic plaque burden, and amyloid levels, onto the sample embedding. To further investigate brain region-specific vulnerabilities, we also mapped regional NFT burden estimates onto the sample embedding.

Separately, to quantitatively evaluate associations with AD-related traits, we computed sample-level epigenomic information by aggregating single-cell scores within each major cell class and across all cells per snATAC library. Median values were calculated for each major cell class and the total nuclei, and Spearman correlation analysis was performed to assess associations with AD pathology measures (e.g., global pathology burden, amyloid level), cognitive traits (e.g., global cognition, episodic memory), cognitive resilience scores.

Cell-type composition changes associated with epigenomic information dynamics, AD pathology progression, and cognitive resilience

To investigate cell-type composition changes across 7 major cell classes and 67 high-resolution subtypes in 6 brain regions, we calculated cell-type fractions at the cell class and subtype levels for each sample (individual-by-region) in both snATAC and snRNA datasets. We then applied a generalized quasi-binomial linear model in R (v.4.2.2) to examine how cell-type composition varies with epigenomic information dynamics, AD pathology progression, and cognitive resilience. For cell-type fraction changes associated with epigenomic information dynamics, we modeled the fraction of each cell class or subtype in individual-by-region samples as a function of sample-level epigenomic information scores. The coefficient p values and log₂(odds ratio) were obtained from the generalized quasi-binomial linear model and visualized using the pheatmap R package (v.1.0.12).¹⁴⁹ A positive log₂(odds ratio) value indicates a greater enrichment of the specific cell type associated with higher epigenomic information, while a negative value indicates depletion. For AD pathology and cognitive resilience, we modeled the cell-type fractions against individual-level continuous global AD pathology scores and cognitive resilience scores. A positive log₂(odds ratio) with the global AD pathology score indicates enrichment of a given cell type in individuals with higher AD pathology burden, while a negative value indicates depletion. Similarly, positive or negative log₂(odds ratio) values with the cognitive resilience score indicate relative enrichment or depletion of cell types in cognitively resilient individuals.

DEGs associated with epigenomic information dynamics, AD pathology progression, cognitive function, cognitive decline, and cognitive resilience

To identify DEGs associated with epigenomic information dynamics and compare them to DEGs linked to AD pathology progression, cognitive function, cognitive decline, and cognitive resilience, we selected nuclei with paired high-quality epigenomic and transcriptomic data from our snATAC and snRNA datasets derived from snMultiome data. We then mapped snATAC-derived epigenomic information at single-cell resolution to matched snRNA-derived gene expression via barcode matching. Given the single-cell resolution of the epigenomic information scores, differential expression analysis was performed using the single-cell method Nebula (v.1.5.3).¹⁴⁴

To ensure consistency across differential expression analyses, we used the same set of snMultiome profiles to assess DEGs associated with epigenomic information dynamics, AD pathology, cognitive traits, and cognitive resilience. For AD pathology-associated DEGs, we incorporated individual-level quantitative measures, including global AD pathology, overall neurofibrillary tangle burden, tangle density, amyloid level, neuritic plaque burden, and diffuse plaque burden. For cognitive function and resilience-associated DEGs, we used various cognitive function domains, domain-specific cognitive decline slopes, and cognitive resilience scores. All models included age, sex, postmortem interval (PMI), total read count, and total gene count per cell as covariates. DEGs were identified within each cell type using a FDR-adjusted p value threshold of < 0.05.

GO term, functional pathway, upstream regulator, and histone modification enrichment analyses

GO term, functional pathway, upstream regulator, and histone modification enrichment analyses were performed using R package enrichR (v.3.2).¹⁵⁰ Upregulated and downregulated DEGs associated with the epigenomic information dynamics were analyzed separately. For GO term and pathway enrichment, we queried multiple databases available through the Enrichr platform (https://maayanlab.cloud/Enrichr/#libraries), including GO_Molecular_Function_2023, GO_Biological_Process_2023, GO_Cellular_Component_2023, WikiPathway_2023_Human, KEGG_2021_Human, and Reactome_2022. Upstream regulator enrichment was performed using the databases, including TRANSFAC_and_JASPAR_PWMs, ChEA_2022, ENCODE_TF_ChIP-seq_2015, and TRRUST_Transcription_Factors_2019. Enrichment for histone modifications was assessed using the Epigenomics_Roadmap_HM_ChIP-seq database.

DARs associated with epigenomic information dynamics, AD pathology progression, cognitive function, cognitive decline, and cognitive resilience

We identified DARs associated with epigenomic information dynamics, AD pathology progression, cognitive function, cognitive decline, and cognitive resilience, using a similar strategy to DEG analyses. The genome was partitioned into consecutive 100-kb bins, and the number of accessible fragments within each bin was quantified per cell using muon.atac.tl.locate_fragments and muon.atac.tl.count_fragments_features functions.¹³⁹ Differential accessibility analysis was performed using Nebula (v1.5.3) on the resulting 100-kb tile matrix. For epigenomic information-associated DARs, chromatin accessibility was modeled as a function of single-cell-level epigenomic information scores. For AD- and cognition-related DARs, we employed the same individual-level variables used in DEG analyses, including measures of AD pathology, cognitive function, slopes of cognitive decline, and cognitive resilience. All models included age, sex, PMI, single-cell library type (single-omic snATAC-seq or snATAC-seq derived from snMultiome), and total fragment count as covariates. Significant DARs were identified per cell type using a FDR-adjusted p value threshold of < 0.05.

Chromatin state enrichment analysis of cCREs, compartments, and DARs

To study the chromatin state enrichment in identified cCREs, inferred active and repressive compartments, and the computed DARs associated with the epigenomic information dynamics, various AD pathologies, cognitive resilience. We obtained chromatin state data (hg38 version) for different brain regions from the EpiMap portal using biosample IDs: BSS00134 for the embryonic brain, BSS00369 and BSS00371 for the frontal cortex, BSS01125 and BSS01126 for the hippocampus, BSS01271 and BSS01272 for the middle frontal area, and BSS00077 and BSS00078 for the angular gyrus.⁶² We computed the relative enrichment of 18 distinct chromatin states in cCREs, compartments, and DARs using ChromHMM software with OverlapEnrichment function.¹⁸⁰

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.cell.2025.06.031.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Frozen adult postmortem human brain tissues	RUSH, ROS/MAP	N/A
Chemicals, peptides, and recombinant proteins
Recombinant RNase inhibitor	Takara Bio	Cat# 2313B
OptiPrep™ Density Gradient Medium	Sigma-Aldrich	D1556-250ML
Phosphate-Buffered Saline (PBS)	Fisher Scientific	SH3025601
DTT	Sigma-Aldrich	Cat# 43816
BSA Fraction V, Protease-free	Sigma-Aldrich	03117332001
Critical commercial assays
Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Kit	10x Genomics	Cat# 1000283
Chromium Single Cell ATAC Library & Gel Bead Kit	10x Genomics	Cat# 1000110
Deposited data
Processed data in this study	This study	https://compbio.mit.edu/AD_Multiomic_MultiRegion
Protected individual raw data in this study	This study	Synapse: https://www.synapse.org/Synapse:syn66271521 and https://www.synapse.org/Synapse:syn66271522
Multiregion snRNA-seq	Mathys et al.46	Synapse: https://www.synapse.org/Synapse:syn52293442
PFC snRNA-seq, snATAC-seq, and snMultiome	Mathys et al.47 ; Xiong et al.38	Synapse: https://www.synapse.org/Synapse:syn52293417
Software and algorithms
cellranger-atac	10x Genomics	https://support.10xgenomics.com/single-cell-atac/software
Cell Ranger	10x Genomics	https://www.10xgenomics.com/support/software/cell-ranger/latest
Scanpy (v.1.9.3)	Wolf et al.136	https://scanpy.readthedocs.io/en/stable/
scrublet (v.0.2.1)	Wolock et al.137	https://github.com/swolock/scrublet
BBKNN	Polanski et al.138	https://github.com/Teichlab/bbknn
Muon (v.0.1.3)	Bredikhin et al.139	https://muon.readthedocs.io/en/latest/index.html
Seurat (v.4.4.0)	Hao et al.140	https://satijalab.org/seurat/
LDSC (v.1.0.1)	Bulik-Sullivan et al.58	https://github.com/bulik/ldsc
ArchR (v.1.0.2)	Granja et al.141	https://www.archrproject.com/
MACS2 (v.2.2.6)	Zhang et al.142	https://pypi.org/project/MACS2/
GREAT (v.1.26.0)	McLean et al.143	http://great.stanford.edu/
Nebula (v.1.5.3)	He et al.144	https://github.com/lhe17/nebula
ComplexHeatmap (v.2.14.0)	Gu et al.145	https://jokergoo.github.io/ComplexHeatmap-reference/book/
ggpubr (v.0.6.0)	Kassambara et al.146	https://github.com/kassambara/ggpubr
IGV Genome Browser (v.2.15.4)	Robinson et al.147	https://igv.org/
FINEMAP (v.1.4)	Benner et al.148	http://www.christianbenner.com/
SCANVAGE (v.1.0.2)	Yu et al.85	https://github.com/sankaranlab/SCAVENGE
pheatmap (v.1.0.12)	Kolde et al.149	https://github.com/raivokolde/pheatmap
enrichR (v.3.2)	Kuleshov et al.150	https://maayanlab.cloud/Enrichr/
BEDTools (v.2.30.0)	Quinlan et al.151	https://bedtools.readthedocs.io/en/latest/
Homer (v.4.11.1)	Heinz et al.152	http://homer.ucsd.edu/homer/

Highlights.

• Single-cell epigenomic and transcriptomic dissection of AD across six brain regions

• Widespread epigenomic relaxation and information loss during AD progression

• Epigenomic information dynamics link to glial activation and exhaustion

• Epigenomic stability correlates positively with cognitive resilience and declines with AD pathology