Cell-type-specific transcriptomic signatures associated with Alzheimer’s disease in the ROSMAP cohort: a single-nucleus RNA-seq pseudobulk analysis.

doi:10.21203/rs.3.rs-8475138/v1

Download PDF

Research Article

Cell-type-specific transcriptomic signatures associated with Alzheimer’s disease in the ROSMAP cohort: a single-nucleus RNA-seq pseudobulk analysis.

https://doi.org/10.21203/rs.3.rs-8475138/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Alzheimer’s disease (AD) is a complex neurodegenerative disorder involving interactions between neuronal, glial, and vascular processes. While single-cell and single-nucleus RNA sequencing studies have identified cell-type-specific transcriptional alterations in AD, the extent to which these patterns are consistently associated with disease status at the donor level remains incompletely characterized.

Methods

We performed a secondary analysis of publicly available single-nucleus RNA sequencing (snRNA-seq) data from the ROSMAP cohort, comprising 188,941 nuclei from 111 post-mortem human donors. Cells were aggregated by donor and major brain cell type to generate normalized pseudobulk expression profiles. For each cell type, regularized logistic regression models were used to evaluate whether transcriptomic patterns differed between AD and cognitively normal control donors. Model performance was assessed using donor-stratified cross-validation across multiple random seeds, permutation testing, and exploratory coefficient analysis.

Results

Non-neuronal cell types showed more consistent transcriptomic differences between AD and control donors than neuronal populations. Astrocyte-derived profiles demonstrated the strongest association with disease status (maximum AUC = 0.646), followed by microglia/immune cells (mean AUC = 0.574 ± 0.088) and vascular cells (mean AUC = 0.540 ± 0.069). Excitatory and inhibitory neurons exhibited limited disease-associated transcriptomic signal. Genes contributing most strongly to these patterns were functionally enriched in pathways related to neuroinflammation, lipid and cholesterol metabolism, oxidative stress response, and blood–brain barrier integrity. Exploratory in silico perturbation analyses suggested greater model sensitivity to coordinated transcriptomic changes in microglial gene sets.

Conclusions

Cell-type-specific pseudobulk transcriptomic profiles derived from snRNA-seq data exhibit reproducible associations with Alzheimer’s disease status at the donor level, predominantly driven by glial and vascular cell populations. These findings support a multifactorial view of AD pathophysiology and provide a computational framework for prioritizing biological hypotheses for future experimental validation. This study is exploratory and hypothesis-generating in nature and does not support diagnostic, prognostic, or clinical application.

Epigenetics & Genomics

Bioinformatics

Neurobiology of Disease

Cellular & Molecular Neuroscience

Computational Neuroscience

Alzheimer's disease

snRNA-seq

pseudobulk

astrocytes

microglia

neuroinflammation

blood-brain barrier

ROSMAP

computational analysis

Alzheimer's disease (AD) is a complex neurodegenerative pathology characterized by β-amyloid deposits, neurofibrillary tangles of hyperphosphorylated tau, and progressive cognitive decline. Beyond these classical hallmarks, there is growing evidence that glial dysfunction, chronic neuroinflammation, and vascular alterations play a central role in disease progression.1,2

The development of single-cell and single-nucleus RNA sequencing technologies (scRNA-seq, snRNA-seq) has enabled high-resolution characterization of cell-type-specific transcriptomic changes in post-mortem human brains.3,4 Most existing studies have focused on identifying cellular states or differentially expressed genes within specific cell populations. However, the extent to which these cell-type-specific transcriptomic patterns are consistently associated with disease status at the donor level remains incompletely characterized.

In this study, we performed a systematic secondary analysis of publicly available snRNA-seq data from the ROSMAP cohort to evaluate whether pseudobulk transcriptomic profiles derived from major brain cell types differ reproducibly between AD donors and cognitively normal controls. We adopted a conservative analytical framework, using simple classification models, donor-level cross-validation, and multiple robustness analyses, with the objective of identifying stable transcriptomic associations rather than maximizing predictive or clinical performance.

2.1 Data and cohort

We used publicly available snRNA-seq data from the ROSMAP cohort (Religious Orders Study and Memory and Aging Project), accessed through the AD Knowledge Portal.⁵ The dataset comprises 188,941 nuclei from post-mortem prefrontal cortex of 111 donors: 57 cognitively normal controls (nonAD, 101,711 nuclei, 53.8%), 32 early AD (earlyAD, 55,094 nuclei, 29.2%), and 22 late AD (lateAD, 32,136 nuclei, 17.0%). For classification purposes, earlyAD and lateAD were combined into a single AD category (n = 54 donors). Available metadata included age at death, sex, and neuropathological diagnosis.

2.2 Preprocessing and quality control

Data were processed using standard single-cell transcriptomics analysis tools. Quality filters were applied at both cell and gene levels, and the top 2,000 globally highly variable genes were selected. Cells were annotated into six major cell types: excitatory neurons (Exc), inhibitory neurons (Inh), astrocytes (Ast), microglia/immune cells (Mic_Immune), oligodendrocytes (Oli), and vascular cells (Vasc).

2.3 Pseudobulk profile generation

For each donor and cell type, a pseudobulk profile was constructed by calculating the mean normalized expression of each gene across all corresponding cells.⁶ This approach reduces technical variability at the cell level, mitigates pseudoreplication,⁷ and enables individual-level analysis, aligned with the clinical study design.

2.4 Classification models and validation

Independent L2-regularized logistic regression models were trained for each cell type. Evaluation was performed using donor-stratified cross-validation (80/20 split), repeated with 10 random seeds (seeds: 13, 23, 37, 41, 59, 73, 101, 131, 151, 181). Performance was quantified using area under the ROC curve (AUC), with confidence intervals calculated as standard deviation across seeds. Seeds resulting in degenerate splits were excluded from averaging.

2.5 Exploratory differential expression analysis

Differential expression analysis between AD and controls was performed for each cell type using Welch's two-sample t-test on pseudobulk profiles. Log₂ fold change was calculated. Results were interpreted descriptively and prioritized for functional analysis, explicitly acknowledging their exploratory, non-confirmatory nature.

2.6 In silico perturbation

A transcriptomic perturbation was simulated by normalizing the top 50 genes with the highest absolute weights in each model toward control levels (0.25 standard deviation shift), evaluating the resulting change in predicted AD probability (Δprob). This analysis was used exclusively as an exploratory tool for gene prioritization, without implying causality.

3.1 Cohort characteristics

The analyzed dataset included 111 unique donors: 57 controls (51.4%) and 54 AD cases (48.6%). All six major cell types were represented in both diagnostic groups, with per-cell-type donor counts ranging from 107 (vascular cells) to 111 (oligodendrocytes). Balanced representation enabled robust classifier training (Fig. 1).

3.2 Classification performance by cell type

Classification performance varied substantially across cell types, with non-neuronal populations demonstrating superior discriminative capacity (Fig. 2, Table 1). Astrocytes achieved the highest AUC in a single-seed configuration (AUC = 0.646), followed by microglia/immune cells with the most robust multi-seed performance (AUC = 0.574 ± 0.088). Vascular cells showed moderate performance (AUC = 0.540 ± 0.069). Notably, both excitatory neurons (AUC = 0.480 ± 0.079) and inhibitory neurons (AUC = 0.459 ± 0.123) exhibited limited discriminative capacity, with inhibitory neurons performing near chance levels. Although modest, these values were significantly superior to those obtained in permutation tests, indicating reproducible biological signal.

Table 1

Classification performance by cell type and analytical configuration.
Cell type	Config.1	Config.2	Config.3	Average	Rank
Astrocytes	0.583	0.646	0.550 ± 0.067	0.593	1
Microglia/Immune	0.590	0.571	0.574 ± 0.088	0.578	2
Vascular cells	0.600	0.525	0.540 ± 0.069	0.555	3
Oligodendrocytes	0.463	0.524	0.529 ± 0.113	0.505	4
Excitatory neurons	0.516	0.559	0.480 ± 0.079	0.518	5
Inhibitory neurons	0.523	0.482	0.459 ± 0.123	0.488	6

Config.1: tun5 (single split with Naive Bayes comparison); Config.2: tun8 (single split, extended analysis); Config.3: tun9 (cross-validation with 10 seeds, values expressed as mean ± SD). Bold values indicate best performance per configuration.

3.3 Discriminative genes and associated biological processes

Model coefficients identified genes whose expression contributed most strongly to classification. Top contributing genes clustered functionally into the following processes (Fig. 3, Table 2):

Neuroinflammation and immune activation (especially in microglia): IKZF3, HLA-DRA, CCL4, CXCL1, CXCR4, VAV1, LAT2
Lipid metabolism and cholesterol transport: CH25H, CETP, SCD, PIPOX
Oxidative stress response and detoxification: MT1E, MT1X, NXN, PARP14
Vascular integrity and blood-brain barrier: F3, CLDN5, FBLN1, ABCB1, AQP1

These patterns were consistent across seeds and functionally related cell types.

Table 2

Top discriminative genes with biological function.
Gene	Cell type	Coef.	Biological function
MCC	Microglia	-0.444	Cell cycle regulator; tumor suppressor
IKZF3	Microglia	+ 0.375	Ikaros family transcription factor; lymphocyte development
HLA-DRA	Astrocytes	-0.110	MHC class II antigen presentation; immune response
BCL2A1	Astrocytes	+ 0.094	BCL-2 family anti-apoptotic protein
CH25H	Vascular	+ 0.062	Cholesterol 25-hydroxylase; oxysterol synthesis, inflammation
MT1E	Vascular	+ 0.060	Metallothionein 1E; oxidative stress, heavy metal binding
CETP	Microglia	+ 0.224	Cholesteryl ester transfer protein; lipid metabolism
TFPI	Astrocytes	-0.082	Tissue factor pathway inhibitor; coagulation regulation

Coef.: logistic regression model coefficient. Positive values indicate overexpression in AD; negative values indicate underexpression in AD.

3.4 Differential expression analysis

Differential expression analysis identified genes with significant changes (p < 0.05) between AD and controls (Fig. 4). In astrocytes: ITGA1 (log₂FC = + 1.11, p = 0.008), TLR5 (log₂FC = + 1.33, p = 0.012), TYMP (log₂FC=-0.98, p = 0.012), IGFBP6 (log₂FC=-1.73, p = 0.023), TFPI (log₂FC=-1.43, p = 0.027). In vascular cells, MT1E showed the most statistically significant change (log₂FC = + 0.16, p = 0.00026), representing a 113% increase in AD relative to controls.

3.5 Convergent pathological axes

Integration of discriminative genes across cell types revealed convergence on four main biological axes (Fig. 5): (1) Neuroinflammatory axis: multiple chemokines (CCL4, CXCL1, CXCR4) and immune signaling molecules (VAV1, LAT2, IKZF3), suggesting chronic immune activation and peripheral cell recruitment. (2) Cholesterol/lipid metabolism axis: CH25H (produces 25-hydroxycholesterol with immunomodulatory properties), CETP, and SCD, implicating altered brain cholesterol homeostasis. (3) Oxidative stress response axis: upregulated metallothioneins (MT1E, MT1X), indicating activation of antioxidant defense mechanisms. (4) Vascular/blood-brain barrier axis: F3 (tissue factor), CLDN5 (claudin-5), FBLN1 (fibulin-1), and ABCB1 (P-glycoprotein), suggesting compromised BBB integrity.

3.6 In silico perturbation

Simulated normalization of the top 50 genes with the highest absolute weights (0.25 standard deviation shift toward control levels) resulted in marked changes in model output probabilities (Fig. 6). In microglia, the simulated perturbation shifted the mean predicted AD probability from 0.436 to 0.082 (Δprob = − 0.354), indicating high model sensitivity to coordinated transcriptomic variation within this cell type. In astrocytes, the effect was more modest (Δprob = − 0.019), potentially reflecting greater redundancy or complexity in astrocytic transcriptional programs.

This analysis does not establish causality or imply biological reversibility of disease processes. Instead, it serves as an exploratory computational approach to prioritize genes and pathways contributing disproportionately to the observed transcriptomic associations.

Our results indicate that cell-type-specific transcriptomic alterations, particularly in glial and vascular populations, exhibit reproducible associations with Alzheimer’s disease status at the donor level, albeit with moderate effect size. This observation is consistent with the multifactorial nature of AD and the substantial clinical and pathological heterogeneity observed across affected individuals. Rather than supporting individual-level classification, these findings highlight population-level transcriptomic patterns relevant to disease-associated biological processes.

4.1 Primacy of glial signatures

The superior discriminative capacity of astrocytes (AUC = 0.646) and microglia (AUC = 0.574) compared to excitatory neurons (AUC = 0.480) and inhibitory neurons (AUC = 0.459) challenges the traditional neurocentric view of AD. Although neuronal loss is the ultimate cause of cognitive decline, our results suggest that glial transcriptomic changes are more consistent across individuals and potentially precede or accompany neuronal dysfunction. These findings align with GWAS studies implicating microglial genes (TREM2, CD33, MS4A cluster) as AD risk loci,⁸ and with evidence of neurotoxic reactive astrocytes.⁹

4.2 Convergence of pathological mechanisms

The four identified pathological axes—neuroinflammation, cholesterol metabolism, oxidative stress, and BBB dysfunction—are not independent but interconnected. 25-hydroxycholesterol produced by CH25H modulates both inflammatory signaling and lipid metabolism.¹⁰ BBB dysfunction allows peripheral immune cell infiltration, amplifying neuroinflammation.¹¹ The convergence of these axes in vascular cells supports the vascular hypothesis of AD.¹²

4.3 Therapeutic implications

The in silico perturbation analyses performed in this study suggest that coordinated modulation of multiple genes highlights biologically coherent transcriptomic programs associated with Alzheimer’s disease, particularly within microglial populations. These findings should be interpreted as a computational prioritization of biological hypotheses rather than evidence of therapeutic efficacy.

The genes and pathways identified are proposed as candidates for future mechanistic investigation and experimental validation, potentially informing downstream functional studies in cellular or animal models. No conclusions regarding treatment strategies, clinical intervention, or therapeutic benefit can be drawn from these analyses. All results presented here are exploratory and hypothesis-generating in nature.

4.4 Limitations

This study has several important limitations. First, AUC values in the 0.55–0.65 range indicate moderate discriminative capacity, reflecting the biological heterogeneity of AD and the challenge of distinguishing disease states from transcriptomic data alone; these models are hypothesis-generating tools, not diagnostic classifiers. Second, our analysis aggregated early and late AD stages; transcriptomic signatures likely differ between stages. Third, covariates such as age, sex, post-mortem interval, and APOE genotype were not explicitly modeled. Fourth, the ROSMAP cohort represents specific demographics (predominantly Caucasian, highly educated); generalization requires validation in diverse cohorts. Fifth, the differential expression analysis is exploratory and non-confirmatory in nature. Nevertheless, this work provides a robust methodological framework for future studies.

This study demonstrates that pseudobulk transcriptomic profiles derived from single-nucleus RNA sequencing data exhibit reproducible associations with Alzheimer’s disease status at the donor level, predominantly driven by glial and vascular cell populations. Non-neuronal cell types, particularly astrocytes and microglia, display more consistent transcriptomic alterations than neuronal populations.

The identified signatures converge on four major biological processes—neuroinflammation, cholesterol and lipid metabolism, oxidative stress response, and blood–brain barrier dysfunction—highlighting interconnected mechanisms relevant to Alzheimer’s disease pathophysiology. Although these transcriptomic patterns do not constitute diagnostic or prognostic biomarkers, they provide a robust computational framework for biological hypothesis generation, mechanistic studies, and future experimental validation.

Data availability

All data used in this study are publicly available through the ROSMAP consortium via the AD Knowledge Portal (https://adknowledgeportal.synapse.org/) under controlled access. Only de-identified, aggregated data were analyzed. No individual-level identifiable data were accessed or generated. The code and analysis scripts will be made publicly available in a dedicated repository upon manuscript acceptance (https://github.com/jnadvid/AlzheimerAdvance).

Ethics statement

This study is a secondary analysis of de-identified, publicly available human data from the ROSMAP consortium.

All original data collection procedures were approved by the relevant institutional review boards, and all participants provided informed consent.

No new human subjects were recruited, and no identifiable personal data were accessed or generated in this study.

Conflict of interest

The author declares no conflicts of interest.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The author welcomes inquiries regarding collaboration or funding opportunities for experimental validation of these findings.

Acknowledgements

We thank the participants of the Religious Orders Study and the Memory and Aging Project for their invaluable contribution to Alzheimer's disease research.

Selkoe DJ, Hardy J (2016) The amyloid hypothesis of Alzheimer's disease at 25 years. EMBO Mol Med 8(6):595–608. 10.15252/emmm.201606210
Long JM, Holtzman DM (2019) Alzheimer disease: An update on pathobiology and treatment strategies. Cell 179(2):312–339. 10.1016/j.cell.2019.09.001
Mathys H, Davila-Velderrain J, Peng Z et al (2019) Single-cell transcriptomic analysis of Alzheimer's disease. Nature 570(7761):332–337. 10.1038/s41586-019-1195-2
Grubman A, Chew G, Ouyang JF et al (2019) A single-cell atlas of entorhinal cortex from individuals with Alzheimer's disease reveals cell-type-specific gene expression regulation. Nat Neurosci 22(12):2087–2097. 10.1038/s41593-019-0539-4
Bennett DA, Buchman AS, Boyle PA et al (2018) Religious Orders Study and Rush Memory and Aging Project. J Alzheimers Dis 64(s1):S161–S189. 10.3233/JAD-179939
Crowell HL, Soneson C, Germain PL et al (2020) muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data. Nat Commun 11(1):6077. 10.1038/s41467-020-19894-4
Squair JW, Gautier M, Kathe C et al (2021) Confronting false discoveries in single-cell differential expression. Nat Commun 12(1):5692. 10.1038/s41467-021-25960-2
Kunkle BW, Grenier-Boley B, Sims R et al (2019) Genetic meta-analysis of diagnosed Alzheimer's disease identifies new risk loci and implicates Aβ, tau, immunity and lipid processing. Nat Genet 51(3):414–430. 10.1038/s41588-019-0358-2
Liddelow SA, Guttenplan KA, Clarke LE et al (2017) Neurotoxic reactive astrocytes are induced by activated microglia. Nature 541(7638):481–487. 10.1038/nature21029
Cyster JG, Dang EV, Reboldi A, Yi T (2014) 25-Hydroxycholesterols in innate and adaptive immunity. Nat Rev Immunol 14(11):731–743. 10.1038/nri3755
Sweeney MD, Sagare AP, Zlokovic BV (2018) Blood-brain barrier breakdown in Alzheimer disease and other neurodegenerative disorders. Nat Rev Neurol 14(3):133–150. 10.1038/nrneurol.2017.188
Zlokovic BV (2011) Neurovascular pathways to neurodegeneration in Alzheimer's disease and other disorders. Nat Rev Neurosci 12(12):723–738. 10.1038/nrn3114

The authors declare no competing interests.

Download PDF

Version 1

posted

You are reading this latest preprint version

Cell-type-specific transcriptomic signatures associated with Alzheimer’s disease in the ROSMAP cohort: a single-nucleus RNA-seq pseudobulk analysis.

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1. Introduction

2. Materials and Methods

2.1 Data and cohort

2.2 Preprocessing and quality control

2.3 Pseudobulk profile generation

2.4 Classification models and validation

2.5 Exploratory differential expression analysis

2.6 In silico perturbation

3. Results

3.1 Cohort characteristics

3.2 Classification performance by cell type

3.3 Discriminative genes and associated biological processes

3.4 Differential expression analysis

3.5 Convergent pathological axes

3.6 In silico perturbation

4. Discussion

4.1 Primacy of glial signatures

4.2 Convergence of pathological mechanisms

4.3 Therapeutic implications

4.4 Limitations

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1