Blood transcriptomic signatures of vascular risk in Alzheimer's disease

Abstract

INTRODUCTION

Lifestyle and environmental factors contributing to cardiovascular risk are components of the exposome, yet molecular signatures associated with cumulative cardiovascular risk in Alzheimer's disease (AD) remain incompletely understood.

METHODS

We analyzed blood‐based transcriptomics data from two independent cohorts: Seoul National University Bundang Hospital (SNUBH; N = 148, Korean) and Alzheimer's Disease Neuroimaging Initiative (ADNI; N = 743, predominantly European ancestry). Framingham cardiovascular risk score (FRS)–associated co‐expression network modules were identified, followed by pathway‐based enrichment, hub gene–based participant stratification, and bidirectional cross‐cohort discovery–replication analyses.

RESULTS

Two co‐expression modules were associated with FRS in SNUBH, and six in ADNI. Bidirectional analyses identified two preserved modules with consistent module–FRS associations across cohorts, enriched in pathways implicated in cellular responses to exposure, including RNA–protein homeostasis, mitochondrial function, and hematopoietic processes. Hub genes in the two modules stratified participants into molecularly distinct subgroups concordant with clinical FRS categories.

DISCUSSION

FRS‐associated molecular pathways were identified and warrant further investigation as potential targets for risk modification.

Highlights

• Blood transcriptomics network analysis identifies Framingham cardiovascular risk score–associated modules in Alzheimer's disease.

• Bidirectional analyses replicate two modules across ancestrally distinct cohorts.

• Modules were enriched in pathways related to cellular responses to exposure.

• Hub genes stratify participants by vascular risk status.

1. BACKGROUND

The exposome refers to the totality of environmental exposures across the life course, encompassing physical, chemical, biological, social, and behavioral factors, and their interactions with genetic susceptibility. ¹ Accumulating evidence indicates that these cumulative exposures contribute to brain aging and individual vulnerability and resilience to Alzheimer's disease (AD) and related dementias. ¹ Within this framework, vascular health has emerged as a central biological pathway through which long‐term environmental and lifestyle exposures may converge to impact neurodegenerative processes. ²

Composite vascular risk indices, particularly the Framingham general cardiovascular risk score (FRS), operationalize the vascular exposome by integrating blood pressure, lipid levels, smoking status, glucose metabolism, and demographic factors into a quantitative measure. ³ Numerous studies have demonstrated that higher cardiovascular risk is associated with cognitive decline, white‐matter injury, and alterations in AD‐related biomarkers. ⁴ , ⁵ , ⁶ These findings underscore that vascular burden exerts biological effects relevant not only to cerebrovascular disease but also to AD pathophysiology.

Beyond these clinical and epidemiological associations, experimental and biomarker evidence indicates that vascular risk may mechanistically potentiate AD‐related pathology. Chronic cardiovascular burden has been linked to immune dysregulation, endothelial dysfunction, oxidative stress, blood–brain barrier (BBB) dysfunction, and impaired proteostasis that overlap with pathways implicated in amyloid beta (Aβ) accumulation, tau hyperphosphorylation, and neurodegeneration. ² , ⁷ , ⁸ Experimental and human tissue data suggest that vascular insults disrupt neuronal metabolic homeostasis, mitochondrial function, and synaptic maintenance, thereby lowering resilience to Aβ and tau pathology. ⁹ , ¹⁰ , ¹¹ Despite growing recognition of intersections between vascular burden and AD, the molecular responses linking cumulative vascular risk to downstream biological alterations remain incompletely understood. In particular, transcriptional changes associated with vascular risk across the AD spectrum have not been systematically investigated, as prior studies have largely focused on stroke or cardiometabolic disease. ¹² , ¹³ This gap is notable because transcriptomic responses represent dynamic, exposure‐sensitive molecular signatures that may capture how long‐term vascular exposures influence neurodegenerative susceptibility.

Another major limitation of the existing vascular–molecular literature is the predominance of Western populations, limiting their relevance for East Asian cohorts in which cardiovascular risk profiles and genetic architecture distributions differ substantially. ¹⁴ , ¹⁵ These cross‐population variations require investigation, particularly as South Korea exhibits one of the highest burdens of vascular disease and dementia in East Asia. ¹⁶ Inclusion of ancestrally diverse cohorts may broaden the biological interpretation of vascular contributions to AD and improve generalizability.

To characterize vascular risk–related transcriptomic architecture, we used a network‐based analytic framework. Weighted gene co‐expression network analysis (WGCNA) provides a systems‐level approach for identifying groups of co‐expressed genes that share functional properties and participate in coordinated biological processes. ¹⁷ , ¹⁸ By modeling gene–gene correlation structure and reducing high‐dimensional data into interpretable network modules, WGCNA has been widely applied to uncover key molecular pathways and regulatory hubs in complex disorders, including AD and cardiovascular diseases, underscoring its utility for investigating transcriptional signatures linked to vascular risk. ¹⁷ , ¹⁹ , ²⁰ This framework is advantageous for investigating vascular risk, as it enables the detection of coherent, biologically meaningful gene networks that may reveal how cumulative vascular stress contributes to neurodegenerative vulnerability.

To address these limitations, we investigated blood‐based transcriptomic correlates of vascular burden, quantified using the FRS, across two AD cohorts representing distinct ancestry backgrounds: the Seoul National University Bundang Hospital (SNUBH) cohort (N = 148) and the Alzheimer's Disease Neuroimaging Initiative (ADNI, N = 743). Using co‐expression network analysis, we identified FRS‐associated network modules and their hub genes and performed pathway‐based enrichment analyses to characterize biological processes underlying vascular risk–related molecular alterations. In addition, we evaluated whether aggregate expression profiles of FRS‐associated hub genes could stratify participants into biologically distinct subgroups using unsupervised clustering. Each cohort was analyzed as an independent discovery dataset, and a bidirectional cross‐cohort replication framework was applied to assess module preservation, eigengene‐level replication, and robustness of hub gene–based stratification across cohorts, thereby minimizing cohort‐specific bias.

2. METHODS

2.1. Study participants

Two independent cohorts were used in this study: the SNUBH and ADNI cohorts. Participants in the SNUBH cohort were identified among individuals who sought neurological evaluation for cognitive decline at Seoul National University Bundang Hospital between April 2020 and February 2023. Each participant completed a standardized diagnostic workup that included comprehensive physical and neurological examinations, routine laboratory testing, and detailed neuropsychological evaluations, ensuring the acquisition of high‐quality clinical data. Diagnostic classification followed the National Institute on Aging–Alzheimer's Association criteria for the diagnosis of probable AD group, ²¹ whereas cognitively normal (CN) individuals were defined based on Mini‐Mental State Examination (MMSE) ²² z scores > −1.0 standard deviations relative to age‐, sex‐, and education‐adjusted norms and Informant Questionnaire on Cognitive Decline in the Elderly scores ²³ < 3.6. The study received approval from the institutional review board (IRB) of SNUBH (IRB No. B‐2004‐604‐305), and all procedures adhered to the principles outlined in the Declaration of Helsinki.

RESEARCH IN CONTEXT

1. Systematic review: We conducted a literature review using PubMed and Google Scholar to identify molecular signatures associated with cumulative cardiovascular risk in Alzheimer's disease (AD). Prior studies have primarily examined single risk factors or focused on stroke or cardiometabolic disease, with limited investigation of blood transcriptomics data for association with cumulative cardiovascular risk.

2. Interpretation: We identified blood‐based transcriptomic network modules associated with cumulative cardiovascular risk, quantified by the Framingham cardiovascular risk score, in two ancestrally distinct AD cohorts. Preserved and replicated modules were enriched in pathways implicated in cellular responses to exposure, including RNA–protein homeostasis, mitochondrial function, and hematopoietic processes. Their hub genes stratified participants into molecularly distinct subgroups concordant with clinical cardiovascular risk status.

3. Future directions: Longitudinal studies and replication in larger independent cohorts are warranted to further investigate cardiovascular risk–associated molecular pathways as potential targets for risk modification.

The ADNI, launched in 2003, represents a large‐scale public–private initiative supported by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, the US Food and Drug Administration, numerous pharmaceutical industry partners, and non‐profit organizations. The first stage of the project (ADNI‐1) was designed to evaluate whether longitudinal neuroimaging measures, cerebrospinal fluid and plasma biomarkers, along with standardized clinical and neuropsychological testing, could serve as reliable tools for detecting and tracking the progression of AD. ²⁴ Building upon this foundation, later phases (ADNI‐GO, ADNI‐2, and ADNI‐3) expanded the initial study framework by continuing longitudinal follow‐up of existing participants and recruiting additional cohorts to enhance coverage across the disease spectrum. ²⁵ , ²⁶ , ²⁷ Detailed information on ADNI, including recent updates, inclusion criteria, and standardized clinical and neuroimaging procedures, is publicly available through the project website (https://www.adni‐info.org). Complementary resources, including access to clinical data, imaging files, biomarker datasets, multi‐omics profiles, neuropathological findings, and cognitive assessments, can be obtained through the ADNI Laboratory of Neuro Imaging (LONI) portal (http://adni.loni.usc.edu). Diagnostic classifications were based on established neuropsychological assessments, such as the Logical Memory from the Wechsler Memory Scale–Revised, ²⁸ MMSE, ²² and Clinical Dementia Rating scale. ²⁹ All participants provided written informed consent at enrollment, authorizing the use and open sharing of their data. Study procedures were reviewed and approved by the IRB of each participating institution.

Participants with cognitive impairment primarily attributable to significant cerebrovascular disease or other non‐AD neurological conditions were excluded according to the eligibility criteria of each cohort.

2.2. Blood‐based RNA expression profiling

For the SNUBH cohort, total RNA was isolated from whole blood using the PAXgene Blood RNA Kit (Qiagen). Library construction was carried out by Macrogen (South Korea) with the TruSeq Stranded Total RNA Library Prep Globin Kit (Illumina, USA), following the manufacturer's reference guidelines. RNA sequencing was performed on the NovaSeq 6000 platform at Macrogen. Initial quality assessment of raw reads was completed using FastQC (version 0.11.7; www.bioinformatics.babraham.ac.uk/projects/fastqc/), and low‐quality bases and adapters were trimmed using Trimmomatic (version 0.38). ³⁰ The filtered reads were aligned to the GRCh38 reference genome (NCBI Annotation Release 109.20200522 GTF file) with HISAT2 (version 2.1.0). ³¹ Gene and transcript expression levels were subsequently quantified from aligned BAM files using StringTie2 (version 2.1.3b). ³² A total of 46,427 genes were initially detected, including 19,672 protein‐coding genes. After removal of genes with minimal expression, 15,376 genes were retained for downstream analysis. Sample‐level quality control (QC) included exclusion of individuals showing sex discordance, assessed by t‐distributed stochastic neighbor embedding (t‐SNE) clustering based on expression of sex chromosome markers (XIST, RPS4Y1, RPS4Y2, KDM5D, DDX3Y, and USP9Y). After quality filtering, trimmed mean of M values normalization was applied prior to downstream statistical analyses.

For the ADNI cohort, total RNA was extracted from whole blood collected in PAXgene Blood RNA Tubes using the PAXgene Blood RNA Kit (Qiagen). Gene expression profiling was performed using the Affymetrix Human Genome U219 microarray platform (Affymetrix), consistent with prior ADNI analytical pipelines. Processing of the raw microarray data followed established QC procedures outlined in a previous study. ³³ Briefly, probe sets were matched to the human reference genome (hg19), and normalization procedures were applied to generate comparable expression values across samples. Sample‐level QC involved verifying concordance between the annotated sex and sex inferred from the expression of sex‐linked genes, including XIST and USP9Y. After all QC steps were completed, the final dataset included expression measurements for 21,150 probes. Probe sets were mapped to their corresponding genes, and when multiple probes mapped to the same gene, the probe with the highest variance across participants was selected, resulting in 20,092 genes for downstream analyses.

2.3. Framingham general cardiovascular risk scores

The FRS represents a well‐validated multivariable model for estimating sex‐specific absolute risk of subsequent cardiovascular events. These risk algorithms have demonstrated strong predictive performance in external cohorts, including the Whitehall II study, in which they reliably predicted incident cardiovascular outcomes. ³⁴ For the present analysis, we derived each participant's 10‐year cardiovascular risk using the non–laboratory‐based FRS. This composite metric incorporates age, sex, body mass index, systolic blood pressure, antihypertensive medication use, smoking status, and diabetes mellitus as predictors. ³ Because the distribution of FRS values was right‐skewed, scores were natural log transformed and then standardized before inclusion in downstream statistical analyses.

2.4. Assessment of AD‐related biomarkers

For the SNUBH cohort, plasma concentrations of phosphorylated tau (p‐tau)217, Aβ40, and Aβ42 were quantified using Simoa‐based assays (Quanterix). Plasma samples were measured in singlicate, while internal controls were loaded in duplicate onto 96‐well Quanterix plates. Measurements were acquired using a two‐step digital immunoassay protocol on the Simoa HD‐X Analyzer (Quanterix), with eight calibrator levels run in triplicate and two QC levels in duplicate, in accordance with the manufacturer's recommendations. To mitigate distributional skew, plasma biomarker values were log–transformed prior to downstream analyses.

For the ADNI cohort, brain magnetic resonance imaging (MRI) was acquired using 3T scanners with 3D magnetization prepared rapid gradient echo sequences according to the standardized ADNI protocol. ³⁵ Structural MRI data were processed using the FreeSurfer version 5.3 longitudinal pipeline to obtain entorhinal cortex and total intracranial volumes (ICVs). Preprocessed [^{1 8}F] florbetapir positron emission tomography (PET) data were obtained from the ADNI database, and cerebral amyloid burden was quantified using standardized uptake value ratios (SUVRs) normalized to the cerebellum as the reference region. ³⁶ Florbetapir PET SUVR and ICV were log transformed to account for skewed distributions. Composite memory performance was assessed using a standardized score derived from multiple neuropsychological measures, including the Alzheimer's Disease Assessment Scale Cognitive Subscale, Rey Auditory Verbal Learning Test, memory‐related items from the MMSE, and Logical Memory test. ³⁷

2.5. Construction of weighted gene co‐expression networks and identification of their hub genes

WGCNA was applied to transcriptomic data from all participants to delineate global patterns of gene co‐expression and to define discrete network modules. For the identification of co‐expression network modules and their associations with the FRS, all available genes within each cohort were used independently, resulting in 15,376 genes in the SNUBH cohort and 20,092 genes in the ADNI cohort. Using the WGCNA package, we constructed a weighted, scale‐free co‐expression network from normalized expression values. ¹⁷ The optimal soft‐thresholding power (β) required to approximate scale‐free topology was selected with the “pickSoftThreshold” function. Pairwise gene correlations were subsequently converted into a topological overlap matrix (TOM), which provides a robust measure of network interconnectedness. Genes exhibiting similar co‐expression profiles were grouped into modules via hierarchical clustering combined with the dynamic tree‐cut algorithm. For each module, a module eigengene (ME), defined as the first principal component summarizing the dominant expression pattern, was computed to represent the overall gene expression pattern within each module. To identify intramodular hub genes, we estimated network connectivity using the “softConnectivity” function based on the TOM. Within the FRS‐associated module, the top 10% of genes with the highest connectivity were designated as hub genes.

2.6. Pathway‐based enrichment analysis

To characterize the molecular pathways and biological functions associated with genes derived from FRS‐associated network modules, we conducted pathway‐based enrichment analysis using the Metascape online platform. ³⁸ Metascape provides an integrated analytical framework that combines functional enrichment, gene annotation, and membership search by leveraging > 40 curated knowledge bases. In the present study, enrichment analyses focused on Gene Ontology (GO), with particular attention to the biological process (BP) domain. ³⁹ GO‐BP terms with a p value < 0.01, a minimum overlap of three, and an enrichment factor > 1.5 were considered statistically significant. Significant terms were subsequently clustered based on similarity in their membership to facilitate biological interpretation.

As a complementary approach, we additionally conducted GO enrichment analyses using the Database for Annotation, Visualization, and Integrated Discovery (DAVID), a functional annotation tool for identifying biologically related gene sets and pathways. ⁴⁰ Similar to the Metascape analyses, we focused on the GO with particular focus on the BP category. Bonferroni correction was applied to account for multiple comparisons.

2.7. Participant stratification using the predicted scores derived from FRS‐associated hub genes

A predicted score was generated to capture the aggregate expression changes of the FRS‐associated hub genes. This score was derived using least absolute shrinkage and selection operator (LASSO) logistic regression applied exclusively to the FRS‐associated hub genes identified within the FRS‐associated network modules. The resulting single numeric score reflected the integrated expression of these hub genes and was subsequently used as the primary feature for participant stratification.

This predicted score was then used as the input feature for participant stratification. Specifically, k‐means clustering was performed via the kmeans function in the R stats package to group participants into data‐driven subclusters. The optimal number of clusters was selected using the elbow criterion, assessed with the fviz_nbclust function in the R factoextra package. To evaluate whether FRS‐associated hub genes could distinguish participants into biologically meaningful subgroups, we applied unsupervised clustering based solely on the predicted scores.

To examine whether these transcriptomically derived clusters reflected FRS status, participants were stratified into low‐ and high‐FRS groups based on a cohort‐specific median split. Individuals with FRS values greater than the median within each cohort were classified as high risk, whereas those with values at or below the median were classified as low risk. This comparison enabled evaluation of how well the predicted score–based clusters aligned with clinically meaningful FRS stratification.

For visualization, dimensionality reduction was conducted using uniform manifold approximation and projection (UMAP) implemented with the umap function in the R umap package, enabling projection of individuals onto a two‐dimensional plane that reflects similarity in FRS‐associated hub gene expression levels.

2.8. Network module preservation analysis for cross‐cohort replication

To ensure robust cross‐cohort validation, we implemented a bidirectional discovery–replication framework. For cross‐cohort analyses, including module preservation and replication, gene expression data were restricted to 13,425 common genes to ensure comparability across platforms. Given the substantial differences in transcriptomic platforms between cohorts, including RNA sequencing in the SNUBH cohort and microarray profiling in the ADNI cohort, this framework is more appropriately interpreted as an orthogonal validation rather than strict replication. In the primary analysis, FRS‐associated co‐expression modules were identified in the SNUBH cohort and subsequently evaluated for module preservation and eigengene‐level replication in the ADNI cohort. In a complementary analysis, the ADNI cohort was treated as an independent discovery dataset, and FRS‐associated modules identified in the ADNI cohort were assessed for preservation and replication in the SNUBH cohort. This reciprocal analytic design allowed us to systematically evaluate the reproducibility of vascular risk–associated transcriptomic network signatures across cohorts differing in ancestry background and transcriptomic measurement platforms.

If a co‐expression network module reflects a biologically meaningful network rather than a random gene aggregation, its modular structure should be reproducible in independent datasets obtained under different conditions. In this study, network module preservation analysis was performed using the WGCNA framework to assess the extent to which modules identified in one cohort were preserved in the other. ⁴¹ Preservation was quantified using the Zsummary statistic, which evaluates whether genes within a predefined module remain densely interconnected and whether inter‐gene connectivity patterns are maintained across datasets, based on permutation testing (n permutations = 200). Modules with a Zsummary values > 10 indicate strong preservation, values between 2 and 10 indicate moderate preservation, and values < 2 indicate little or no evidence of preservation across datasets. ⁴¹ In addition, median rank statistics were calculated as a complementary preservation metric, providing a rank‐based measure of module preservation that is less sensitive to module size. ⁴¹ Modules with lower median rank values were considered better preserved across cohorts than those with higher ranks.

After module preservation analyses, MEs corresponding to preserved FRS‐associated network modules were computed in the replication cohort. These MEs were derived using the same gene sets and eigengene calculation procedures as applied in the discovery cohort, thereby ensuring methodological consistency across datasets. The resulting MEs were subsequently used for downstream association analyses with the FRS within the replication cohort. These replication analyses were performed to assess the robustness and generalizability of associations between FRS and module‐level transcriptional signatures across populations with differing ancestry backgrounds and across distinct transcriptomic measurement platforms.

2.9. Statistical analysis

Linear regression models were used to evaluate the associations of MEs or hub genes with the FRS and plasma AD biomarkers, including the Aβ42/40 ratio and p‐tau217. All models were adjusted for key covariates, including age, sex, and dyslipidemia, with additional adjustment for education in analyses of cognitive function and for ICV in analyses of entorhinal cortex volume.

Additional sensitivity analyses were performed to evaluate the robustness of the primary findings. The primary regression models were further adjusted for apolipoprotein E (APOE) ε4 carrier status to assess the consistency of associations between MEs, hub genes, and the FRS. We also repeated the analyses using unadjusted regression models and performed sex‐stratified analyses to evaluate interactions between FRS and sex. Finally, analyses were repeated after excluding CN participants by restricting the SNUBH cohort to participants with AD and the ADNI cohort to participants with AD. Given the limited sample size of the AD group in ADNI, an additional analysis was conducted in participants with mild cognitive impairment (MCI) and AD. Unless otherwise specified, all analyses were performed using the same regression models and covariate adjustments as the primary analyses. 

FIGURE 1.

Schematic overview of the workflow of our analysis. This figure was created with BioRender.com. ADNI, Alzheimer's Disease Neuroimaging Initiative; FRS, Framingham general cardiovascular risk score; SNUBH, Seoul National University Bundang Hospital; UMAP, uniform manifold approximation and projection.

Multiple testing was adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure. ⁴² Analyses were conducted in R (version 4.2.2). Statistical significance was defined as p < 0.05 after appropriate corrections for multiple comparisons, unless stated otherwise. A comprehensive overview of the full analytical workflow is presented in Figure 1.

3. RESULTS

3.1. Demographics and clinical characteristics

The SNUBH cohort included a total of 148 participants, comprising 17 CN individuals and 131 individuals with AD, as detailed in Table 1. The median age of participants was 76 years, and 64.2% were female. No statistically significant differences were observed between the CN and AD groups with respect to age, sex distribution, or other cardiovascular and metabolic variables, including systolic blood pressure, diabetes, smoking status, antihypertensive medication use, body mass index, and dyslipidemia. FRS values did not differ significantly between the two groups as well.

TABLE 1.

Demographic information of participants from the SNUBH and ADNI cohorts.

	SNUBH
	CN (N = 17)	AD (N = 131)	Total (N = 148)	p value b
Female (%)	76.5	62.6	64.2	0.262
Age, years a	74 (71–77)	76 (68–82)	76 (68–81)	0.601
FRS	19.8 (17.5–23.9)	25.3 (14.8–42.3)	24.3 (15.1–41.3)	0.508
Systolic blood pressure, mmHg a	139 (124–144)	132 (120–149)	133 (121–149)	0.801
Diabetes (%)	23.5	23.7	23.6	0.990
Current smoker (%)	0	6.11	5.41	0.295
Treatment with anti‐hypertensive medication (%)	58.8	44.3	45.9	0.257
Body mass index a	24.6 (22.4–25.8)	23.5 (22.0–25.4)	23.6 (22.0–25.4)	0.413
Dyslipidemia (%)	47.1	41.2	41.9	0.646

	ADNI
	CN (N = 230)	MCI (N = 291)	AD (N = 222)	Total (N = 743)	p value c
Female (%)	53.9	42.3	39.6	45.1	0.004
Age, years a	75 (70–79)	74 (68–80)	77 (71–81)	75 (69–80)	0.003
FRS	31.4 (19.6–50.5)	34.8 (23.2–55.1)	39.0 (22.0–57.9)	35.4 (21.7–54.6)	0.012
Systolic blood pressure, mmHg a	132 (121–143)	135 (123–146)	132 (120–145)	133 (121–145)	0.312
Diabetes (%)	6.52	9.28	10.81	8.88	0.265
Current smoker (%)	39.1	39.9	40.1	39.7	0.976
Treatment with anti‐hypertensive medication (%)	42.2	44.7	40.5	42.7	0.634
Body mass index a	26.8 (24.3–29.7)	26.5 (24.3–29.3)	25.9 (23.6–28.4)	26.3 (24.1–29.2)	0.026
Dyslipidemia (%)	37.8	40.2	44.6	40.8	0.332

Abbreviations: AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; CN, cognitively normal; FRS, Framingham general cardiovascular risk score; MCI, mild cognitive impairment; SNUBH, Seoul National University Bundang Hospital.

^a Data are presented as median followed by (interquartile range).

^b The Mann–Whitney U test or chi‐squared test was used to determine the p value for comparisons between CN and AD groups, as appropriate.

^c The Kruskal–Wallis H test or chi‐squared test was used to determine the p value for comparisons among CN, MCI, and AD groups, as appropriate.

The ADNI dataset consisted of 743 participants, including 230 CN individuals, 291 individuals with MCI, and 222 individuals with AD (Table 1). The median age was 75 years, with 45.1% of the cohort comprising females. Significant group differences were identified in sex distribution, age, FRS, and body mass index. However, systolic blood pressure, diabetes and dyslipidemia prevalence, current smoking status, and antihypertensive medication use did not differ significantly across diagnostic groups.

3.2. Discovery in the SNUBH cohort and replication in the ADNI cohort

3.2.1. Identification of co‐expression network modules and their associations with the FRS in the SNUBH cohort

A scale‐free weighted gene co‐expression network was constructed using transcriptomic data from 148 participants (Figure S1 in supporting information). A soft‐thresholding power (β = 18) was selected to achieve a scale‐free topology index > 0.8 while substantially reducing mean connectivity. Average linkage hierarchical clustering based on the topological overlap matrix derived from pairwise gene expression correlations identified 21 distinct co‐expression modules, with a minimum module size of 30 genes. Genes that did not cluster into any module were assigned to the M0 module and were excluded from downstream analyses. 

Higher ME values of the M7, M15, and M16 modules showed significant associations with higher FRS values (β for linear regression coefficients = 1.503, 1.386, and 1.430 for the M7, M15, and M16, respectively; FDR‐corrected p = 0.047 for all three modules; Table 2). In contrast, the remaining modules did not exhibit significant associations with the FRS.

TABLE 2.

Association of module eigengenes with the Framingham general cardiovascular risk score (FRS) in the SNUBH cohort.

Module	β a	Standard error	Uncorrected P	FDR‐corrected P
M1	1.045	0.515	0.044	0.135
M2	0.662	0.510	0.196	0.375
M3	0.115	0.517	0.825	0.860
M4	1.203	0.508	0.019	0.101
M5	−0.588	0.509	0.250	0.414
M6	−1.023	0.506	0.045	0.135
M7	1.503	0.508	0.004	0.047 *
M8	0.499	0.524	0.343	0.480
M9	0.830	0.519	0.112	0.234
M10	0.192	0.517	0.711	0.793
M11	0.443	0.554	0.425	0.557
M12	1.009	0.546	0.067	0.175
M13	0.876	0.507	0.086	0.202
M14	1.065	0.503	0.036	0.135
M15	1.386	0.503	0.007	0.047 *
M16	1.430	0.501	0.005	0.047 *
M17	0.276	0.512	0.590	0.729
M18	0.535	0.511	0.297	0.445
M19	0.588	0.516	0.256	0.414
M20	0.185	0.512	0.718	0.793
M21	0.091	0.513	0.860	0.860

Abbreviations: FDR, false discovery rate; M, module; SNUBH, Seoul National University Bundang Hospital.

^a Linear regression coefficients of association between module eigengenes and the Framingham general cardiovascular risk score.

Notes: Statistical significance is denoted by asterisks based on FDR‐corrected p values.

*FDR‐corrected p value < 0.05.

To assess the robustness of the primary findings, several additional analyses were performed. Further adjustment for APOE ε4 carrier status resulted in modest attenuation of the associations between FRS and the three FRS‐associated M7, M15, and M16 modules, with FDR‐corrected p values increasing from 0.047 to 0.056 (Table S1 in supporting information). In unadjusted models without age, sex, or dyslipidemia adjustment, the M7 and M15 modules remained significantly associated with FRS after FDR correction, whereas the M16 module retained a similar effect size and direction of association with borderline significance (Table S2 in supporting information). Analyses restricted to participants with AD demonstrated that the three FRS‐associated modules remained significantly associated with FRS in the AD group (Table S3 in supporting information). Finally, sex‐stratified analyses showed no significant interactions between FRS and sex for any of the three primary FRS‐associated modules (Table S4 in supporting information).

3.2.2. Module preservation and replication analysis of SNUBH‐derived FRS‐associated modules in the ADNI cohort

To assess the reproducibility of the FRS‐associated modules identified in the SNUBH cohort, we conducted network module preservation analysis by comparing co‐expressed modules derived from the SNUBH dataset with those in an independent transcriptomic dataset from the ADNI cohort, with particular emphasis on the preservation of the FRS‐associated M7, M15, and M16 modules. Module preservation analysis based on the overlapping gene sets revealed that 13 of the 21 modules identified in the SNUBH dataset were strongly preserved in the ADNI dataset with Zsummary statistics > 10, while seven modules showed moderate preservation with Zsummary statistics between 2 and 10 (Figure 2A). Among the FRS‐associated modules, M7 and M16 demonstrated moderate preservation based on relatively high Zsummary statistics and low median ranks, whereas the M15 module exhibited little or no evidence of preservation.

FIGURE 2.

Preservation and replication analysis in the ADNI cohort for FRS‐associated modules identified in the SNUBH cohort. A, Left panel shows module preservation statistics (Zsummary, y axis) relative to module size (x axis). Each circle represents an individual module, with colors corresponding to module identities. Two dashed red horizontal lines indicate conventional Zsummary thresholds at 2 and 10, corresponding to no (Zsummary < 2), moderate (2–10), and strong (Zsummary > 10) evidence of module preservation. Right panel shows the median preservation rank (y axis) in relation to the module size (x axis). Each circle represents a module labeled in different colors, with the rank scores. Modules with higher Zsummary values and lower median rank are interpreted as demonstrating stronger preservation relative to other modules. Modules that were both preserved and significantly associated with FRS in the ADNI cohort are annotated based on pathway enrichment analyses, including M7 associated with immune response and cellular stress pathways and M16 associated with RNA processing and mitochondrial function. B, Linear regression analysis evaluating the associations of preserved FRS‐associated modules with FRS values in the ADNI cohort. C, Association between the M16 module eigengene and plasma Aβ42/40 ratio. The scatter plot represents a correlation of the M16 module eigengene with plasma Aβ42/40 ratio. The solid line represents the fitted linear regression line, and the gray zone represents the 95% confidence interval. Statistical significance is denoted by asterisks based on p values. * p value < 0.05, ** p value < 0.01. ^aLinear regression coefficients of association between module eigengenes and the FRS. Aβ, amyloid beta; ADNI, Alzheimer's Disease Neuroimaging Initiative; FRS, Framingham general cardiovascular risk score; M, module; MMSE, Mini‐Mental State Examination; p‐tau217, phosphorylated tau217; SNUBH, Seoul National University Bundang Hospital.

After the module preservation analyses, MEs corresponding to the two preserved FRS‐associated network modules were recalculated in the ADNI cohort. These MEs were subsequently evaluated for their associations with the FRS within the ADNI cohort. Association analysis revealed that higher ME values of the M16 module were significantly associated with higher FRS values (β coefficient = 1.049; p = 0.018), consistent in direction with the SNUBH cohort (Figure 2B). In contrast, no significant association was detected between ME values of the M7 module and the FRS.

To further investigate the relevance of the preserved and replicated FRS‐associated M16 module to AD‐related phenotypes, we examined its associations with plasma AD biomarkers (Figure 2C and Table S5 in supporting information). Higher ME values of the M16 modules were significantly associated with lower plasma Aβ42/40 ratios.

3.2.3. Pathway‐based enrichment analysis in the SNUBH cohort

We performed pathway‐based enrichment analyses using Metascape for the preserved and replicated FRS‐associated M16 module to delineate biological processes. GO‐BP pathway–based enrichment analysis revealed that 89 genes in the M16 module were enriched in pathways associated with ribonucleoprotein complex biogenesis, protein folding, messenger RNA (mRNA) metabolic process, aerobic respiration, ribosomal RNA (rRNA) metabolic process, RNA localization to Cajal body, positive regulation of mitophagy, amino acid metabolic process, regulation of translation, mitochondrial membrane organization, negative regulation of apoptotic signaling, carboxylic acid metabolic process, and regulation of reactive oxygen species metabolic process (Figure 3A and Table S6 in supporting information).

FIGURE 3.

Gene Ontology enrichment analysis and stratification of participants in the SNUBH cohort. A, Enrichment network and gene ontology biological process terms for the preserved and replicated FRS‐associated M16 module are shown, as visualized using Metascape. Each Gene Ontology term is represented as a circle node, with node color indicating cluster membership such that terms sharing the same color belong to the same cluster. Edges connect pairs of terms with a similarity score > 0.3, and edge thickness corresponds to the degree of similarity. Gene‐set interaction networks for the preserved and replicated FRS‐associated M16 module are also displayed after applying a degree filter of at least five to highlight densely connected functional relationships among the top 20 enriched terms ranked by p value. B, UMAP plots represent stratification of participants based on the FRS‐associated hub gene expression profiles in the SNUBH cohort. Upper panel shows participant subgroups (Cluster 1–4) identified by unsupervised k‐means clustering, while lower panel displays stratification according to phenotype groups (FRS low risk and high risk). Colors indicate subgroup cluster in the upper panel and FRS risk categories in the lower panel. FRS, Framingham general cardiovascular risk score; M, module; mRNA, messenger RNA; rRNA, ribosomal RNA; SNUBH, Seoul National University Bundang Hospital; UMAP, uniform manifold approximation and projection.

We additionally performed pathway‐based enrichment analyses of the preserved but non‐replicated FRS‐associated M7 module. Seventy‐seven genes in the M7 module were significantly enriched in pathways related to positive regulation of innate immune response, positive regulation of endocytosis, autophagy, pentose–phosphate shunt, membrane invagination, regulation of type I interferon (IFN‐1) production, negative regulation of RNA splicing, actin filament organization, neuron apoptotic process, response to nutrient levels, carbohydrate metabolic process, positive regulation of cell–cell adhesion, response to ketone, and carboxylic acid metabolic process (Figure S2 and Table S6 in supporting information).

Complementary pathway analyses using DAVID demonstrated highly consistent enrichment patterns, further supporting the biological interpretations of the M7 and M16 modules (Table S7 in supporting information).

3.2.4. Identification of FRS‐associated hub genes and participant stratification based on hub gene–derived scores in the SNUBH cohort

We identified the top 10% hub genes with the highest TOM‐based intramodular connectivity within each of the preserved FRS‐associated M7 and M16 modules (Table S8 in supporting information). Association analysis of the hub genes revealed that higher expression levels of HCLS1, NBR1, NEU1, ATP6V0D1, IST1, and RTF2 in the M7 module and DDOST, ENO1, XRCC6, HDAC1, IDH3B, SARS1, and EDF1 in the M16 module were significantly associated with higher FRS values.

To determine whether the FRS‐associated hub genes could delineate biologically meaningful subgroups, we first computed a predicted score summarizing the aggregate expression of these genes using LASSO logistic regression. Using the predicted score, participants were stratified through unsupervised k‐means clustering, which identified distinct data‐driven subgroups. The elongated UMAP structures reflect continuous gradients in transcriptomic variation, consistent with the use of predicted scores as a continuous variable. Participants in the SNUBH cohort were classified into four subgroups, designated C1, C2, C3, and C4 (Figure 3B). The stratification of participants across subgroups revealed that C1 predominantly consisted of individuals with FRS low‐risk status, while C3 was primarily comprised of individuals with FRS high‐risk status. The predicted scores used for clustering differed markedly across the four subgroups, aligning with their respective FRS profiles.

3.3. Discovery in the ADNI cohort and replication in the SNUBH cohort

3.3.1. Identification of co‐expression network modules and their associations with the FRS in the ADNI cohort

WGCNA was applied to gene expression profiles from 743 participants (Figure S3 in supporting information). A soft‐thresholding power of β = 10 was chosen to approximate scale‐free topology. Average linkage hierarchical clustering of the TOM yielded 17 co‐expression modules that met the minimum module size of 30 genes. The M0 module comprised genes not assigned to any of the defined modules and was removed from further evaluation.

Higher ME values of the M4, M15, M16, and M17 modules and lower ME values of the M1 and M2 modules were significantly associated with higher FRS values (β coefficients = −1.569, −1.341, 1.484, 1.115, 1.287, and 1.397 for the M1, M2, M4, M15, M16, and M17, respectively; corresponding FDR‐corrected p = 0.005, 0.012, 0.006, 0.030, 0.012, and 0.008; Table 3). In contrast, the other modules did not show significant associations with the FRS.

TABLE 3.

Association of module eigengenes with the Framingham general cardiovascular risk score (FRS) in the ADNI cohort.

Module	β a	Standard error	Uncorrected P	FDR‐corrected P
M1	−1.569	0.432	3.07 x 10−4	0.005 **
M2	−1.341	0.453	0.003	0.012 *
M3	−0.916	0.450	0.042	0.098
M4	1.484	0.435	0.001	0.006 **
M5	0.127	0.457	0.781	0.802
M6	0.231	0.447	0.606	0.686
M7	−0.112	0.447	0.802	0.802
M8	0.379	0.441	0.390	0.510
M9	0.563	0.460	0.221	0.376
M10	−0.407	0.449	0.366	0.510
M11	−0.394	0.441	0.372	0.510
M12	−0.697	0.459	0.129	0.243
M13	0.289	0.442	0.513	0.622
M14	0.909	0.454	0.046	0.098
M15	1.115	0.435	0.011	0.030 *
M16	1.287	0.442	0.004	0.012 *
M17	1.397	0.435	0.001	0.008 **

Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; FDR, false discovery rate; M, module.

^a Linear regression coefficients of association between module eigengenes and the Framingham general cardiovascular risk score.

Notes: Statistical significance is denoted by asterisks based on FDR‐corrected p values.

*FDR‐corrected p value < 0.05.

**FDR‐corrected p value < 0.01.

To assess the robustness of the primary findings, several additional analyses were performed. Further adjustment for APOE ε4 carrier status demonstrated that the associations between FRS and the six FRS‐associated M1, M2, M4, M15, M16, and M17 modules remained significant after FDR correction (Table S9 in supporting information). In unadjusted models without age, sex, or dyslipidemia adjustment, four of the six primary FRS‐associated modules remained significantly associated with FRS after FDR correction (Table S10 in supporting information). Analyses restricted to participants with AD showed that only M4 module remained significantly associated with FRS, whereas the remaining modules showed attenuation of statistical significance (Table S11 in supporting information). An additional analysis including participants with MCI and AD demonstrated that both M1 and M4 modules remained significantly associated with FRS (Table S12 in supporting information). Finally, sex‐stratified analyses identified significant interactions between FRS and sex for M2 and M16. For both modules, the associations were stronger in women than in men, reaching statistical significance only in women, whereas no significant associations were observed in men (Table S13 in supporting information).

3.3.2. Module preservation and replication analysis of ADNI‐derived FRS‐associated modules in the SNUBH cohort

To assess the reproducibility of the FRS‐associated modules identified in the ADNI cohort, we performed network module preservation analysis in the SNUBH cohort, with particular focus on the preservation of the FRS‐associated M1, M2, M4, M15, M16, and M17 modules. Module preservation analysis revealed that 13 of the 17 modules identified in the ADNI dataset were strongly preserved in the SNUBH dataset with Zsummary statistics > 10, while three modules showed moderate preservation with Zsummary statistics between 2 and 10 (Figure 4A). Among the FRS‐associated modules, M1, M4, M15, M16, and M17 modules demonstrated strong preservation based on relatively high Zsummary statistics and low median rankings, whereas the M2 module exhibited moderate preservation.

FIGURE 4.

Preservation and replication analysis in the SNUBH cohort for FRS‐associated modules identified in the ADNI cohort. A, Left panel shows module preservation statistics (Zsummary, y axis) relative to module size (x axis). Each circle represents an individual module, with colors corresponding to module identities. Two dashed red horizontal lines indicate conventional Zsummary thresholds at 2 and 10, corresponding to no (Zsummary < 2), moderate (2–10), and strong (Zsummary > 10) evidence of module preservation. Right panel shows the median preservation rank (y axis) in relation to the module size (x axis). Each circle represents a module labeled in different colors, with the rank scores. Modules with higher Zsummary values and lower median rank are interpreted as demonstrating stronger preservation relative to other modules. The module that was both preserved and significantly associated with FRS in the SNUBH cohort is annotated based on pathway enrichment analyses, with M2 representing RNA processing and translational regulation with hematopoietic signatures. B, Linear regression analysis evaluating the associations of preserved FRS‐associated modules with FRS values in the SNUBH cohort. C, Association between the M2 module eigengene and composite score for memory. The scatter plot represents a correlation of the M2 module eigengene with composite score for memory. The solid line represents the fitted linear regression line, and the gray zone represents the 95% confidence interval. Statistical significance is denoted by asterisks based on p values. * p value < 0.05. ^aLinear regression coefficients of association between module eigengenes and the FRS. ADNI, Alzheimer's Disease Neuroimaging Initiative; FRS, Framingham general cardiovascular risk score; M, module; SE, standard error; SNUBH, Seoul National University Bundang Hospital.

After the module preservation analyses, MEs corresponding to the six preserved FRS‐associated network modules were computed in the SNUBH cohort. These MEs were subsequently evaluated for their associations with the FRS within the SNUBH cohort. Association analysis revealed that higher ME values of the M2 module were significantly associated with lower FRS values (β coefficient = −1.116; p = 0.028), with the direction of effect consistent with that observed in the ADNI cohort (Figure 4B). However, no significant associations were observed between FRS and the ME values of the remaining five preserved FRS‐associated modules.

To further investigate the relevance of the preserved and replicated FRS‐associated M2 module to AD‐related phenotypes, we examined its associations with amyloid PET burden, entorhinal cortex volume, and composite memory performance (Figure 4C and Table S14 in supporting information). Higher ME values of the M2 module were significantly associated with better composite memory performance.

3.3.3. Pathway‐based enrichment analysis in the ADNI cohort

We performed pathway‐based enrichment analyses using Metascape for the preserved and replicated FRS‐associated M2 module to delineate biological processes. GO‐BP pathway–based enrichment analysis revealed that 168 genes in the M2 module were enriched in pathways including mRNA processing, translation, primary microRNA (miRNA) processing, regulation of mRNA metabolic process, ribonucleoprotein complex biogenesis, nuclear receptor‐mediated signaling, hemopoiesis, morphogenesis of a polarized epithelium, aerobic respiration, RNA 3'‐end processing, regulation of translational initiation, lymphocyte migration, positive regulation of epithelial to mesenchymal transition, nuclear export, DNA metabolic process, and brain development (Figure 5A and Table S15 in supporting information). We additionally performed pathway‐based enrichment analyses of the five preserved but non‐replicated FRS‐associated M1, M4, M15, M16, and M17 modules (Figure S4 and Table S15 in supporting information).

FIGURE 5.

Gene Ontology enrichment analysis and stratification of participants in the ADNI cohort. A, Enrichment network and gene ontology biological process terms for the preserved and replicated FRS‐associated M2 module are shown, as visualized using Metascape. Each Gene Ontology term is represented as a circle node, with node color indicating cluster membership such that terms sharing the same color belong to the same cluster. Edges connect pairs of terms with a similarity score > 0.3, and edge thickness corresponds to the degree of similarity. Gene‐set interaction networks for the preserved and replicated FRS‐associated M2 module are also displayed after applying a degree filter of at least five to highlight densely connected functional relationships among the top 20 enriched terms ranked by p value. B, UMAP plots represent stratification of participants based on the FRS‐associated hub gene expression profiles in the ADNI cohort. Upper panel shows participant subgroups (Cluster 1–3) identified by unsupervised k‐means clustering, while lower panel displays stratification according to phenotype groups (FRS low risk and high risk). Colors indicate subgroup cluster in the upper panel and FRS risk categories in the lower panel. ADNI, Alzheimer's Disease Neuroimaging Initiative; FRS, Framingham general cardiovascular risk score; M, module; mRNA, messenger RNA; miRNA, microRNA; UMAP, uniform manifold approximation and projection.

Complementary pathway analyses using DAVID demonstrated highly consistent enrichment patterns, further supporting the biological interpretations of the preserved and replicated M2 module (Table S16 in supporting information).

3.3.4. Identification of FRS‐associated hub genes and participant stratification based on hub gene–derived scores in the ADNI cohort

We identified the top 10% hub genes within the preserved FRS‐associated M1, M2, M4, M15, M16, and M17 modules (Table S17 in supporting information). Association analysis showed that increasing FRS values were associated with significantly higher expression levels of 34 hub genes in M4, 3 hub genes in M15, 87 hub genes in M16, and 8 hub genes in M17. In contrast, increasing FRS values were associated with significantly lower expression levels of five hub genes in M1 and nine hub genes in M2. Additional analyses further adjusting for APOE ε4 carrier status demonstrated that 145 of 146 hub genes remained significantly associated with FRS, while maintaining consistent effect directions (Table S18 in supporting information).

To evaluate whether FRS‐associated hub genes could define biologically meaningful participant subgroups, we calculated a predicted score summarizing their aggregate expression through LASSO logistic regression. Participants in the ADNI cohort were classified based on the predicted score using unsupervised k‐means clustering, resulting in three subgroups, designated C1, C2, and C3 (Figure 5B). C1 predominantly consisted of individuals with FRS low‐risk status, while C3 was primarily comprised of individuals with FRS high‐risk status. The predicted‐score distributions differed markedly across clusters, consistent with their corresponding FRS profiles, indicating that the clustering structure also reflected vascular risk status in an independent cohort.

4. DISCUSSION

This study provides novel network‐based insights into how vascular burden, quantified using the FRS, is associated with variation in blood‐based peripheral transcriptomic signatures in participants enrolled in AD‐spectrum cohorts. By integrating blood RNA sequencing from the Korean‐ancestry SNUBH cohort with genome‐wide microarray profiling from the predominantly European‐ancestry ADNI cohort, we identified co‐expression network modules reproducibly associated with vascular risk across distinct ancestry backgrounds. Importantly, these modules represent vascular risk–associated transcriptomic responses identified within the AD‐spectrum cohorts rather than general cardiovascular populations. Beyond their association with vascular burden, the preserved and replicated modules were also associated with AD‐related phenotypes, including plasma Aβ42/40 and cognitive performance, suggesting that vascular risk–related transcriptomic alterations may extend beyond cardiovascular biology and be relevant to AD‐related processes. Furthermore, additional analyses restricted to participants with AD demonstrated that a subset of these FRS–module associations remained detectable within the AD spectrum, supporting their relevance beyond differences across diagnostic groups. Our study represents the first cross‐cohort transcriptomic investigation of FRS in AD cohorts, providing a foundational molecular framework linking vascular risk to reproducible blood‐based transcriptomic network signatures.

Several biologically relevant signatures identified in the FRS‐associated network modules may reflect molecular processes linking vascular burden to AD‐related molecular alterations. Importantly, some biological processes were reproduced across the Korean‐ancestry and predominantly European‐ancestry cohorts, whereas others appeared cohort specific, suggesting both trans‐ancestry and cohort‐specific transcriptomic responses to vascular risk. Module preservation analysis further revealed that a substantial proportion of FRS–associated modules exhibited moderate to strong preservation across cohorts, indicating that these network modules reflect conserved biological responses rather than cohort‐specific effects. In particular, SNUBH‐derived M16 and ADNI‐derived M2 modules showed both preservation and consistent associations with the FRS across cohorts, providing convergent evidence for transcriptomic signatures of vascular burden that are reproducible across cohorts with distinct ancestry backgrounds. Specifically, higher ME values of the SNUBH‐derived M16 were associated with higher FRS, whereas higher ME values of the ADNI‐derived M2 were associated with lower FRS across both cohorts. M16 was also associated with plasma Aβ42/40 in the SNUBH cohort, whereas higher M2 ME values were associated with better memory performance in ADNI.

With a particular focus on biological pathways relevant to the exposome, functional annotation of both modules converged on pathways related to ribonucleoprotein complex biogenesis, protein folding, mitochondrial bioenergetics, primary miRNA processing, and hemopoiesis. Ribonucleoprotein complex biogenesis and ribosomal maturation constitute stress‐responsive molecular systems sensitive to cumulative metabolic, oxidative, and inflammatory exposures. ⁴³ Disruption of ribonucleoprotein complex biogenesis is observed in AD, where aberrant U1 small nuclear ribonucleoprotein (snRNP) assembly co‐occurs with pathological tau neurofibrillary tangles. ⁴⁴ Moreover, altered snRNP function and RNA splicing have been shown to influence amyloid precursor protein (APP) processing, linking stress‐related RNA dysregulation to Aβ‐associated pathogenic mechanisms in AD. Protein folding and proteostasis constitute exposure‐responsive molecular systems that are highly sensitive to cumulative metabolic, oxidative, and inflammatory stressors within the exposome. Recent evidence demonstrates that disruption of protein homeostasis and activation of endoplasmic reticulum stress responses occur in AD‐like neuropathology, indicating that impaired proteostasis drives chronic neuroinflammation, protein aggregation, and progressive neurodegeneration in AD. ⁴⁵ Mitochondrial bioenergetics represent a central exposure‐responsive system, as cumulative metabolic, vascular, and inflammatory stressors progressively impair oxidative phosphorylation efficiency and increase reactive oxygen species (ROS) production, thereby promoting vascular and neuronal injury that contributes to AD. ⁴⁶ MiRNA biogenesis and processing provide an exposure‐responsive epigenetic interface through which cumulative vascular and lifestyle‐related exposures are translated into molecular changes relevant to disease pathophysiology. ⁴⁷ Altered primary miRNA processing contributes to neurodegenerative vulnerability by disrupting miRNA‐mediated regulation of APP expression, tau‐related signaling, and neuronal stress responses. ⁴⁸ Hematopoiesis represents an exposure‐responsive biological system, as cumulative metabolic, inflammatory, and environmental stressors progressively reshape hematopoietic stem and progenitor cell function. ⁴⁹ In AD, altered hematopoiesis contributes to neurodegenerative vulnerability by sustaining pro‐inflammatory myeloid signaling that interacts with oxidative stress to exacerbate Aβ accumulation, tau pathology, and cognitive decline. ⁵⁰

Although the SNUBH‐derived FRS‐associated M7 module was not replicated in the ADNI cohort, this module showed preservation across cohorts and was enriched for several exposure‐responsive biological processes, including IFN‐1 production, the pentose–phosphate pathway (PPP), actin filament organization, and response to nutrient levels. Elevated IFN‐1 production and downstream interferon‐stimulated gene expression are detected in the AD brain, where sustained type I interferon signaling amplifies microglial activation and chronic neuroinflammation in response to Aβ and tau aggregates. ⁵¹ The PPP represents an exposure‐responsive metabolic system, as sustained redox imbalance and increased antioxidant demand drive cellular reliance to maintain redox homeostasis. ⁵² Elevated PPP activity has been observed in response to oxidative stress in AD, where disruptions in PPP‐mediated metabolism are associated with Aβ–induced metabolic impairment and oxidative damage, and emerging experimental evidence suggests that PPP‐dependent redox regulation may also modulate tau‐related neurodegenerative vulnerability. ⁵² , ⁵³ Actin filament organization constitutes a stress‐responsive cytoskeletal system that dynamically integrates metabolic and oxidative stress signals, reflecting close reciprocal interactions between cytoskeletal dynamics and cellular metabolic state. ⁵⁴ In AD, tau‐induced overstabilization of actin filaments are observed in neurons, with pathological actin filament changes associated with tau pathology and neurodegenerative processes. ⁵⁵ Chronic dysregulation of cellular responses to nutrient levels, encompassing both excess and deprivation, may arise from long‐term metabolic and dietary exposures. Such dysregulation perturbs nutrient‐sensing pathways, thereby promoting amyloidogenic processing and tau phosphorylation. ⁵⁶

Similarly, the ADNI‐derived FRS‐associated M1, M4, M15, M16, and M17 modules were not replicated in the SNUBH cohort; these modules showed preservation across cohorts and were enriched for several exposure‐responsive biological processes, including cellular response to cytokine stimulus, small GTPase‐mediated signal transduction, autophagy, and cellular response to starvation. Cellular responses to cytokine stimulation are closely linked to AD‐related neuroinflammation, as sustained cytokine signaling promotes microglial activation and accelerates amyloid pathology. ⁵⁷ Small GTPase‐mediated signaling regulates intracellular trafficking and synaptic homeostasis, and its dysregulation has been implicated in amyloid processing and neurodegeneration in AD. ⁵⁸ Autophagy is essential for maintaining proteostasis, and defective autophagy facilitates Aβ and tau accumulation while exacerbating chronic neuroinflammatory responses in AD. ⁵⁹ Cellular responses to starvation engage AMP‐activated protein kinase–mammalian target‐of‐rapamycin–dependent nutrient‐sensing pathways that regulate autophagy and metabolic adaptation, processes increasingly recognized as modifiers of AD progression. ⁵⁶ Accumulating transcriptomic studies of cardiovascular disease and stroke have consistently implicated inflammatory and immune‐related pathways. ⁶⁰ , ⁶¹ , ⁶² Consistent with these findings, the inflammatory processes identified in our FRS‐associated modules may reflect shared vascular risk‐related mechanisms, while their associations with AD‐related phenotypes further support their relevance within the AD spectrum.

Beyond module‐level associations, we demonstrated that unsupervised clustering based on aggregate expression patterns of FRS‐associated hub genes stratified participants into biologically meaningful subgroups that showed strong concordance with clinical FRS status across both cohorts. Importantly, this stratification was achieved using blood‐based gene expression profiles alone, independent of traditional clinical cardiovascular risk factors, highlighting the potential of transcriptomic signatures to complement conventional clinical measures by capturing latent biological vulnerability. Such molecular stratification may have important implications for precision medicine, as individuals with similar clinical FRS values may exhibit divergent transcriptional responses to vascular stress, contributing to heterogeneity in cognitive trajectories and therapeutic response. Interventions targeting vascular risk reduction may help evaluate whether these transcriptomic signatures capture modifiable biological responses. Furthermore, integration of broader exposome dimensions, including social and environmental factors, may further refine the interpretation of vascular risk–related transcriptomic signatures in AD.

Our study has several limitations. First, the cross‐sectional design of this study precludes causal inference regarding the temporal relationships among vascular risk and transcriptomic alterations. Second, methodological and clinical heterogeneity between the two cohorts, including differences in transcriptomic platforms, disease stage distributions, and FRS levels, may have introduced variability. However, associations between FRS and modules were evaluated independently within each cohort, minimizing the potential influence of these differences on the observed associations. Third, the modest sample size of the SNUBH cohort, particularly the limited number of CN individuals, may have reduced statistical power and constrained gene‐level analyses. Fourth, plasma Aβ measurements using immunoassay‐based methods may be susceptible to analytical variability, warranting cautious interpretation of the observed associations with plasma Aβ42/40. ⁶³ Fifth, experimental validation of the identified hub genes and pathways was not performed, and further functional studies will be necessary to elucidate their biological roles. Finally, although peripheral blood transcriptomic profiles provide valuable insights into systemic biological responses, ⁶⁴ they may not fully capture brain‐specific molecular processes.

Moreover, lifestyle or pharmacologic intervention studies targeting vascular risk reduction could leverage these transcriptional signatures to assess biological response to treatment. Future work will incorporate broader exposome dimensions, including social and environmental factors that are being collected in the SNUBH cohort, to further refine the interpretation of vascular risk–related transcriptomic signatures in AD.

In conclusion, vascular burden quantified by the FRS is captured by reproducible blood‐based transcriptomic signatures across two AD cohorts of distinct ancestry. These exposure‐responsive molecular profiles are associated with plasma AD biomarkers and enable biologically informed risk stratification that is complementary to conventional clinical measures, highlighting their potential utility for precision prevention by identifying individuals who may benefit most from targeted vascular risk modification strategies in AD.

AUTHOR CONTRIBUTIONS

Sang‐Won Han contributed to the concept and design of the work, data acquisition, interpretation of data, drafting of the manuscript, statistical analysis, and critical revision of the manuscript for important intellectual content. Young Ho Park and Kwangsik Nho contributed to the concept and design of the work, data acquisition, interpretation of data, drafting of the manuscript, statistical analysis, critical revision of the manuscript for important intellectual content, and supervision. Jeong Su Park, Jiyun Hwang, Jeong Su Park, and SangYun Kim contributed to the interpretation of data and manuscript reviewing. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

All the authors report that they have no disclosures or conflicts of interest relevant to the manuscript. Author disclosures are available in the Supporting Information.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

The protocol for the SNUBH study has been approved by the IRB of the SNUBH for the SNUBH cohort (IRB No. B‐2004‐604‐305) and conforms to the provisions of the Declaration of Helsinki. All data used from the ADNI cohort are publicly available and de‐identified. Therefore, no additional ethical approval or informed consent was required for the present study. Ethics approval for ADNI was obtained at all participating sites in accordance with the standards of their respective local ethics committees and IRBs.

Supporting information

Supporting Information

Supporting Information

Supporting Information

DATA AVAILABILITY STATEMENT

The data generated and/or analyzed in the SNUBH cohort are available from the corresponding authors upon reasonable request. Data from the ADNI cohort are available through the ADNI LONI data repository (http://adni.loni.usc.edu).