Integrating genetically predicted transcriptomic signatures with longitudinal real-world data enables scalable drug repurposing for Alzheimer’s disease

Integrating genetically predicted transcriptomic signatures with longitudinal real-world data enables scalable drug repurposing for Alzheimer’s disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Integrating genetically predicted transcriptomic signatures with longitudinal real-world data enables scalable drug repurposing for Alzheimer’s disease Monika E. Grabowska, Rui Chen, Ying Zhou, Avi U. Vaidya, Xue Zhong, and 14 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9518587/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Drug repurposing offers a potential strategy to expand treatment options for conditions with limited therapies, but advancing repurposing candidates toward clinical implementation remains a challenge. Large-scale data, together with advanced genetic and epidemiological methods, may help address this gap. Here, we present an integrative digital medicine approach that combines genetically predicted transcriptomic signatures and perturbation screening for candidate identification with multi-cohort real-world validation for systematic evaluation of prioritized candidates. We applied this approach to Alzheimer’s disease (AD), a disease with substantial unmet clinical need and persistent difficulty in developing effective therapies. We constructed AD signatures from genetically predicted expression changes across bulk tissues and microglia, then queried Connectivity Map profiles to identify compounds predicted to oppose these signatures. Aspirin emerged as a reproducible candidate across multiple signatures and underwent further evaluation. We then examined its association with incident AD in longitudinal electronic health record data from Vanderbilt University Medical Center and the NIH All of Us Research Program, as well as national insurance claims data. Across independent cohorts, aspirin initiation before age 65 was consistently associated with lower risk of incident AD, with signals suggesting that cumulative exposure and APOE ε4 status may influence effect size. Transcriptomic analysis of human cortical organoids provided additional experimental support, showing that aspirin more strongly opposed AD-related neuronal pathway alterations in wild-type organoids than in an organoid model of AD. This integrative approach offers a scalable strategy for genetically informed drug repurposing that bridges candidate discovery and clinical evaluation. Biological sciences/Computational biology and bioinformatics Health sciences/Neurology Biological sciences/Neuroscience Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Many serious diseases still lack effective therapies despite substantial investment in drug development. Drug repurposing offers a practical complement to de novo drug development by identifying new uses for existing drugs. However, many repurposing studies stop at candidate nomination, creating a bottleneck in determining which of the many proposed candidates warrant further study 1 . Scalable approaches that integrate large-scale data, human genetics, and epidemiological methods can help move drug repurposing beyond candidate identification toward systematic evaluation, with the potential to accelerate clinical translation and improve patient outcomes. Transcriptomic signature reversal has become a widely used strategy for drug repurposing, supported by the growing availability of large-scale genetic and transcriptomic datasets. This approach queries perturbational gene expression resources, such as the Connectivity Map (CMap), to identify compounds predicted to reverse disease-associated transcriptional profiles 2 . While early implementations often derived disease signatures from RNA sequencing or microarray data, transcriptome-wide association study (TWAS) methods offer a distinct genetics-informed approach by integrating expression quantitative trait loci (eQTLs) with genome-wide association study (GWAS) data to identify genes whose genetically predicted expression is associated with disease risk 3 . Because TWAS leverages germline genetic variation fixed at conception, it is less susceptible to reverse causation and may better prioritize causal genes and therapeutic targets. Given evidence that genetically supported drug targets are approximately 2.6 times more likely to succeed in clinical development 4 , TWAS-based signature reversal has emerged as a promising strategy for drug discovery and repurposing in complex diseases and has been applied to endometrial cancer, hypertension, and hyperlipidemia 5 , 6 . After candidate identification, additional validation is needed to confirm preliminary signals and distinguish the most promising candidates. Real-world clinical data, including electronic health records (EHRs) and claims data, contain longitudinal information on medication exposures and clinical outcomes at scale and are increasingly used to generate and assess repurposing hypotheses in human populations 7 – 10 . This is especially relevant for conditions with a long preclinical phase, such as Alzheimer’s disease (AD), as candidate effects may depend on exposure timing and duration. Alongside clinical data, experimental studies in cellular models, such as human induced pluripotent stem cell (iPSC)-derived organoids, can provide complementary biological support for candidate assessment. Given the time and cost of downstream prospective clinical studies, integrating these sources of evidence is critical for directing resources toward the most compelling candidates. In this study, we developed a genetics-informed repurposing workflow that integrates TWAS-derived disease signatures, perturbational screening, and longitudinal real-world clinical data. We then applied this workflow to AD. First, we identified AD-associated genes using GWAS summary statistics and eQTL data from bulk tissues and primary microglia, and used these genes to construct cross-tissue and microglia-specific disease signatures. We then queried CMap profiles to identify compounds opposing these signatures. Aspirin emerged as a recurrent candidate across multiple signatures and was subsequently evaluated in longitudinal EHR data from Vanderbilt University Medical Center (VUMC) and the National Institutes of Health All of Us Research Program, as well as in national claims data from the MarketScan Research Databases. We further examined aspirin-induced transcriptional responses in human iPSC-derived cortical organoids with wild-type and AD-associated APP mutant genotypes. Our findings illustrate how genetically informed candidate prioritization can be coupled with validation, including longitudinal clinical evaluation and experimental investigation, to advance drug repurposing and accelerate clinical implementation. Results Genetically informed Alzheimer’s disease transcriptomic signatures An overview of the study design is shown in Fig. 1. Because the compounds returned by CMap depend directly on the disease signature used for querying, we constructed multiple TWAS-derived AD signatures spanning both bulk tissues and microglia rather than relying on a single transcriptomic profile. We performed TWAS using S-PrediXcan 11 and S-MultiXcan 12 to identify candidate AD risk genes. Using pre-trained transcriptome prediction models for 49 tissues from the Genotype-Tissue Expression (GTEx) project 13 and summary statistics from a large AD GWAS 14 , we conducted three S-MultiXcan analyses to balance tissue specificity and statistical power: (1) a brain-specific analysis combining the 13 GTEx brain tissues, (2) an AD-relevant tissue analysis with the 13 brain tissues plus four peripheral tissues previously related to AD (whole blood, spleen, and sun-exposed and unexposed skin) 15,16 , and (3) an all-tissue analysis combining all 49 GTEx tissues. AD risk genes were defined within each S-MultiXcan analysis as Bonferroni-significant TWAS associations ( P <0.05/number of tested gene associations) and are reported in Supplementary Data File 1 (sheets 1-3). We then constructed three GTEx-derived AD transcriptomic signatures (brain, AD-relevant, and all tissues) using the gene inclusion criteria described in Methods. The final brain-tissue AD transcriptomic signature contained 72 genes (37 positively associated with AD risk, 35 inversely associated), while the AD-relevant tissue signature contained 78 genes (41 positively, 37 inversely associated) and the all-tissue signature contained 79 genes (40 positively, 39 inversely associated) (Fig. 2a). A shared set of 43 genes was observed across all three GTEx signatures (21 positively associated with AD risk, 22 inversely associated). Because bulk tissues can obscure cell type-specific signals and microglia play a central role in AD, we also constructed a microglial AD signature by performing TWAS with custom transcriptome prediction models trained on microglial eQTL summary statistics from the Microglia Genomic Atlas (MiGA) 17 . This signature comprised 53 genes (25 positively associated with AD risk, 28 inversely associated) (Fig. 2a). MiGA microglial TWAS results are provided in Supplementary Data File 1, sheet 4. Across the four signatures, we identified genes previously implicated in AD, including APOE , TREM2 , and BIN1 (Fig. 2b; Supplementary Table 1). An annotated comparison of the four AD transcriptomic signatures is shown in Supplementary Fig. 1. Nine genes were shared across all four signatures (Fig. 2c), of which seven had directionally concordant associations with AD risk (Fig. 2d). Signature reversal prioritizes aspirin for downstream evaluation We queried the Connectivity Map (CMap) 18 using each TWAS-derived AD signature to identify compounds predicted to reverse disease-associated gene expression changes. Genes were assigned to up and down sets by TWAS Z -score sign. Compounds with negative connectivity scores (τ<0) were considered repurposing candidates 6,19 . Out of 2,428 small-molecule compounds in CMap, 590 (24.3%), 709 (29.2%), and 688 (28.3%) had negative connectivity to the GTEx brain-tissue, AD-relevant tissue, and all-tissue signatures, respectively (Supplementary Data File 1, sheets 5-7). Of these, 218 showed negative connectivity to all three GTEx signatures, including mycophenolate, fluticasone, sirolimus, sertraline, clozapine, and losartan, although connectivity magnitude varied widely (e.g., mycophenolate τ: brain-tissue -72.2; AD-relevant tissue -7.6; all-tissue -0.6). Additionally, 794 (32.7%) compounds had negative connectivity to the microglia-specific signature (Supplementary Data File 1, sheet 8). The top ten repurposing candidates for each signature are shown in Table 1. The number and ranking of negatively connected compounds varied substantially across signatures, underscoring the importance of cross-signature support in candidate selection. Aspirin emerged as a recurrent candidate across multiple signatures, ranking among the top ten compounds for the all-tissue GTEx AD signature (τ=-69.96), with negative connectivity also observed for the AD-relevant GTEx tissue signature (-17.89) and the MiGA microglial signature (-30.88), although no connectivity was detected for the GTEx brain-tissue signature. For downstream evaluation, we focused on compounds showing negative connectivity across more than one AD signature and practical feasibility for longitudinal clinical evaluation. Aspirin met these criteria, with support across three signatures, widespread clinical use, and a well-characterized safety profile. Table 1. Top ten AD repurposing candidates identified in CMap queries. Drugs approved by the United States Food and Drug Administration are marked with an asterisk and their clinical indications are provided in parentheses. GTEx brain tissues AD-relevant GTEx tissues All GTEx tissues Microglia mupirocin* (impetigo and other uncomplicated bacterial skin infections) CAY-10618 indinavir* (human immunodeficiency virus) anisomycin mycophenolic-acid* (prophylaxis of organ rejection, autoimmune disease) bufalin isogedunin emetine BX-795 vidarabine* (herpes simplex keratitis) deforolimus homoharringtonine* (chronic myeloid leukemia) fluticasone* (asthma, allergic rhinitis, and certain inflammatory skin conditions) PP-30 phorbol-12-myristate-13-acetate narciclasine prostratin ivermectin* (parasitic infections) U-46619 digitoxigenin PD-123319 rhodomyrtoxin-b BX-795 troxipide phenylbutyrate* (urea cycle disorders) pentylenetetrazol aspirin* (pain, fever, primary and secondary cardiovascular disease prevention, rheumatoid arthritis) pyrvinium-pamoate desoxypeganine ingenol* (actinic keratosis) pirinixic-acid roscovitine SA-63133 prostratin ON-01910 cycloheximide indinavir* (human immunodeficiency virus) salubrinal praziquantel* (parasitic infections) isoliquiritigenin Real-world clinical validation across three independent databases We evaluated the prioritized aspirin signal in three real-world datasets: (1) VUMC’s de-identified EHR database, (2) the NIH All of Us Research Program database, and (3) the MarketScan Research Databases. In EHR data from VUMC and All of Us , we used a retrospective cohort study design to compare AD incidence after age 65 between aspirin-exposed patients (≥1 year of aspirin use before age 65) and propensity score-matched unexposed patients. In MarketScan claims, shorter follow-up limited ascertainment of incident AD among individuals with medication exposures documented before age 65; therefore, we used a case-control design comparing prior aspirin exposure in AD cases versus propensity score-matched controls. Descriptive characteristics for matched EHR cohorts are shown in Table 2. The characteristics of the matched MarketScan claims-based cohort are provided in Supplementary Table 2. Information on AD outcomes in all three datasets is provided separately in Supplementary Table 3. Table 2. Description of matched patient cohorts used in EHR validation studies. Clinical characteristics VUMC ( N = 19,413) All of Us ( N = 1,995) Aspirin Exposed Unexposed Exposed Unexposed N 6,656 12,757 666 1,329 Mean age at last follow-up (s.d.) 77.8 (2.8) 77.9 (2.8) 77.3 (2.4) 77.3 (2.4) Sex (%) Female 49.8 51.8 45.6 45.7 Male 50.2 48.2 52.4 52.5 Race (%) White 92.3 92.2 77.3 77.4 Black 6.8 6.8 7.1 7.3 Other 0.9 1.0 15.6 15.3 Baseline comorbidities (%) † Cardiovascular disease 25.1 21.2 28.5 28.4 Cerebrovascular disease 6.0 5.6 8.1 7.5 Rheumatoid arthritis 3.6 4.1 4.1 4.1 † Baseline comorbidities were defined as ≥1 diagnosis code recorded at or before time zero (age 65 or first encounter after 65). Code lists are provided in Supplementary Data File 1, sheets 14-16. In VUMC, aspirin use before age 65 was associated with a significantly reduced risk of incident AD after age 65 (hazard ratio [HR]=0.77, 95% confidence interval [CI]: 0.65-0.91, P =0.003; Fig. 3a). In All of Us , the association was directionally similar but limited by low statistical power (HR=0.40, 95% CI: 0.15-1.08, P=0.07; N =24 AD events among 1,995 participants). Meta-analysis of the two EHR cohorts showed that aspirin initiation before age 65 was associated with 24% lower risk of incident AD (HR=0.76, 95% CI: 0.64-0.89, P =0.001; Fig. 3a). In MarketScan, patients diagnosed with AD were less likely to have prior aspirin exposure compared to matched controls (OR=0.32, 95% CI: 0.28-0.38, P <0.001). We performed secondary analyses in VUMC to examine whether the association varied by aspirin dose or cumulative exposure. Due to the limited number of AD events in the All of Us cohort and the small number of aspirin prescriptions in MarketScan, we were unable to evaluate these measures in those datasets. In dose-stratified analyses (high-dose ≥325mg/day; low-dose ≤81mg/day 20 ), aspirin use remained associated with lower AD risk for both high-dose (HR=0.63, 95% CI: 0.45-0.90, P =0.01) and low-dose regimens (HR=0.82, 95% CI: 0.68-0.99, P =0.04), with no significant difference between dose groups ( P =0.19). Because higher aspirin doses may reflect more severe underlying indications, we also examined cumulative exposure. Given inconsistent documentation of dosing frequency and treatment duration in the EHR, we used documented aspirin exposure rate, defined as the number of unique aspirin records divided by the years between first and last recorded exposure, as a proxy. Exposure rates above the cohort median (>5/year) were associated with lower AD risk (HR=0.58, 95% CI: 0.39-0.88, P =0.009; Fig. 3b). In addition, we conducted APOE -stratified analyses in VUMC and All of Us (MarketScan does not contain genetic data). Among APOE ε4 carriers, aspirin use before age 65 showed a suggestive inverse association with incident AD in both cohorts, although neither analysis reached statistical significance individually (VUMC HR=0.60, 95% CI: 0.33-1.10, P =0.0986; All of Us HR=0.48, 95% CI: 0.10-2.25, P =0.35). Combined meta-analysis showed a 41% decreased risk of incident AD after age 65 in APOE ε4 carriers (HR=0.59, 95% CI: 0.33-1.02, P =0.06), suggesting a potentially stronger protective association in this subgroup but limited by statistical power. Aspirin use was not significantly associated with decreased AD risk among non-carriers (VUMC HR=0.63, 95% CI: 0.26-1.55, P =0.317; All of Us HR=0.52, 95% CI: 0.15-1.79, P =0.3; meta-analysis HR=0.59, 95% CI: 0.29-1.22, P =0.16), although the power of these analyses was limited by the small number of AD cases. Transcriptomic evaluation in human iPSC-derived cortical organoids To assess the biological plausibility of the aspirin signal, we treated 90-day-old cortical organoids derived from isogenic control wild-type (WT; WTC11) and heterozygous APP mutant (KM670/671NL) iPSCs with aspirin (0.5 mM) or vehicle (PBS) for one week, followed by RNA sequencing (RNA-seq). High uniquely mapped paired-end rates (~90% across samples) indicated successful library preparation and sequencing (Supplementary Data File 1, sheet 9). Replicates showed high concordance and clustered cleanly by their assigned labels (Supplementary Fig. 2), supporting the expected data quality. Over-representation analysis of differentially expressed genes (|log2FC| ≥log2(1.1), FDR<0.05) highlighted synapse- and axon-related pathways, including glutamatergic synapse and axon guidance, in both the baseline APP mutation signature and the WT aspirin response (Supplementary Fig. 3a). Gene set enrichment analysis (GSEA) showed broad downregulation of neuronal and synaptic terms in the baseline APP signature, including synaptic signaling (NES=-2.29, FDR=4.11×10 -31 ) and axon development (NES=-2.11, FDR = 2.76×10 -15 ) (Supplementary Fig. 3b). In WT organoids, aspirin produced the opposite pattern, with strong upregulation of synaptic signaling (NES=2.52, FDR=8.16×10 -48 ) and axon development (NES=2.56, FDR=3.75×10 -39 ) (Supplementary Fig. 3c). In contrast, aspirin-treated APP mutant organoids showed attenuated enrichment of neuronal terms relative to WT, with top GSEA hits predominantly reflecting cell cycle and chromosome-associated processes (Supplementary Fig. 3d). Full GSEA results for all contrasts are provided in Supplementary Data File 1, sheets 10-12. In WT organoids, the transcriptional effects of aspirin were modestly but significantly negatively correlated with the baseline APP mutation signature (Spearman’s ρ=-0.15, P <2.2×10 -16 , N =30,848 genes), consistent with partial reversal of disease-associated changes. Rank-rank hypergeometric overlap (RRHO) analysis 21 identified a discordant hotspot comprising genes upregulated by aspirin in WT organoids and downregulated in the baseline APP signature (Fig. 4a). In APP mutant organoids, the aspirin response was positively correlated with the APP signature (ρ=0.37, P <2.2×10 -16 , N =30,848 genes), and RRHO showed a dominant concordant overlap signal (Fig. 4b), providing little evidence of global transcriptomic reversal in the APP background. Genes in the discordant hotspot in Fig. 4a were enriched for synaptic signaling and neurotransmission across GO Biological Process, KEGG, and Reactome databases, with top terms including regulation of trans-synaptic signaling, glutamatergic and GABAergic synapse, and transmission across chemical synapses (Fig. 4c; Supplementary Data File 1, sheet 13). Building on these findings, we performed targeted GSEA of synapse-, axon-, and neurotransmitter-related gene sets curated from MSigDB using keyword filters. The baseline disease signature (APP vehicle vs. WT vehicle) showed broad downregulation across these pathway families, consistent with neuronal dysfunction in AD. In WT organoids, aspirin robustly upregulated these pathways, counteracting the baseline AD signature (synapse and axon pathways in Fig. 5 and Supplementary Figs. 4-5; neurotransmitter pathways in Supplementary Fig. 6). By contrast, in APP mutant organoids, aspirin’s effects were attenuated and did not consistently oppose the baseline disease signature. Discussion We developed an integrative drug repurposing framework that combines genetically informed transcriptomic disease signatures, perturbational signature matching, longitudinal real-world clinical data, and experimental investigation in human cellular models to prioritize candidates for further study and accelerate clinical translation. Applying this framework to AD, we identified aspirin as a candidate supported across multiple disease signatures, three independent clinical datasets, and transcriptomic analyses in human cortical organoids. This study illustrates how genetically informed candidate prioritization can be paired with large-scale longitudinal clinical data to evaluate repurposing hypotheses in a scalable and clinically relevant manner. An important feature of this study is the use of multiple disease signatures spanning both bulk tissues and microglia rather than a single transcriptomic representation of AD. Our TWAS analyses highlighted genes at established AD risk loci, including APOE , TREM2 , and BIN1 , and corroborated previously reported TWAS associations (Supplementary Table 1), while also revealing substantial variability across tissues and cell types. Only seven AD risk genes appeared in all four signatures with concordant effect directions. Notably, BIN1 , a leading late-onset AD risk locus, showed a positive association with AD risk in microglia but inverse associations in all GTEx signatures. This variability was also reflected in the top-ranked CMap compounds identified across the four TWAS-derived AD signatures, supporting the use of cross-signature consistency as a prioritization criterion. We therefore focused on compounds with negative connectivity across more than one AD signature and practical feasibility for longitudinal clinical evaluation. Aspirin met these criteria, with support across the all-tissue GTEx, AD-relevant GTEx, and microglial signatures, making it a useful test case for this framework. Across three real-world datasets, aspirin exposure was consistently associated with lower AD risk. In meta-analysis of the two EHR cohorts, aspirin initiation before age 65 was associated with a 24% reduced risk of incident AD. The claims-based study, while constrained by shorter observation windows that precluded time-to-event analysis, showed directionally consistent results in a case-control analysis. We did not detect a significant difference between high- and low-dose regimens, but individuals with higher documented aspirin exposure rates, used here as a proxy for cumulative exposure, had lower AD risk than matched individuals with lower exposure rates. This observation is consistent with UK Biobank analyses in which the inverse association between low-dose aspirin and AD was most evident with long-term use (>10 years) 22 . Analysis of EHR-linked genetic data suggested a stronger inverse association among individuals carrying at least one APOE ε4 allele, although statistical power was limited by the relatively small number of APOE ε4 carriers and low prevalence of AD diagnoses among non-carriers. Overall, these findings support the consistency of the aspirin signal across distinct clinical data sources and illustrate how longitudinal clinical data can refine repurposing signals beyond simple exposed-versus-unexposed comparisons. Transcriptomic analysis of human iPSC-derived cortical organoids provided complementary experimental support for the aspirin signal. In wild-type organoids, aspirin-induced transcriptional changes were directionally opposite to the APP mutation signature and were enriched for synaptic, axonal, and neurotransmission-related pathways suppressed in the baseline disease state. In contrast, aspirin responses in APP mutant organoids were attenuated and did not show the same degree of opposition to the baseline disease signature. These findings provide biological plausibility for the prioritized signal and suggest that aspirin’s effects may depend on underlying disease context. Randomized trials have not demonstrated a clear population-level benefit of aspirin in AD, although most initiated treatment relatively late in life 23–26 , after AD-related pathology may already have been established 27,28 . Our findings do not contradict those trials; rather, they raise the possibility that any benefit of aspirin, if present, may depend on exposure timing and persistence during the long preclinical interval preceding diagnosis, when relevant neuronal pathways may still be modifiable. This interpretation is consistent with the genetically anchored, life-course nature of TWAS-derived signatures, and with our real-world analyses, which specifically emphasized aspirin initiation before age 65. It is also broadly concordant with the organoid results, where aspirin more strongly opposed disease-associated neuronal pathway alterations in wild-type than APP mutant backgrounds, suggesting reduced efficacy in a disease context. This study has several strengths. It integrates human genetics, perturbational screening, and multi-cohort real-world clinical data within a single repurposing workflow and evaluates the leading candidate across independent EHR and claims datasets, as well as through transcriptomic analysis in human iPSC-derived organoids. Using TWAS, we mapped AD genetic risk to genetically predicted gene expression across multiple tissues and in microglia, generating disease-relevant signatures for candidate identification. Longitudinal EHR data enabled evaluation of aspirin initiation over intervals that are difficult to study in conventional trials and supported additional analyses by exposure pattern and APOE ε4 status. Importantly, this framework is transferable to other complex diseases with available GWAS summary statistics and longitudinal EHR or claims data, where genetics-informed prioritization can be integrated with real-world data to support more systematic assessment of repurposing candidates. This study also has several limitations. Repurposing hypotheses were restricted to compounds profiled in CMap, which is not exhaustive and is derived largely from non-brain cell lines that may not fully reflect AD-relevant biology. TWAS was constrained by the ancestral composition of available AD GWAS and eQTL reference datasets, and we lacked power for ancestry-stratified EHR analyses despite known differences in AD incidence among populations, which may limit generalizability. AD case definitions relied on structured diagnosis codes and may incompletely capture clinical heterogeneity and potentially introduce outcome misclassification. Aspirin exposure was also likely under-ascertained because over-the-counter use is not represented in pharmacy claims and is inconsistently recorded in EHR medication lists. These outcome and exposure measurement limitations would both be expected to attenuate associations toward the null. Although we addressed measured confounding through propensity score matching, residual confounding remains possible. Finally, organoids lack systemic vascular and immune context, precluding assessment of aspirin’s antithrombotic and peripheral anti-inflammatory effects that are likely relevant in vivo . The workflow described here provides a scalable strategy for moving from computationally prioritized repurposing signals to clinical evaluation in longitudinal real-world data, particularly relevant for diseases in which conventional drug development is slow, costly, or poorly aligned with disease stage. Given the flexibility of this workflow, future updates could incorporate additional identification or validation steps, such as AI-facilitated screening 10 . In AD, our findings support further study of aspirin with earlier initiation rather than broad late-life use. Future observational studies incorporating deeper phenotyping, including neuroimaging, cerebrospinal fluid biomarkers, polygenic risk scores, and cardiovascular and metabolic risk measures, may help define the subgroups most likely to benefit. Overall, this approach demonstrates a scalable next step for advancing drug repurposing toward eventual clinical implementation. Methods This study was conducted with approval from the VUMC Institutional Review Board and the NIH All of Us Research Program. All EHR data from VUMC and All of Us are de-identified; use of these data is considered non-human subjects research. Construction of microglia transcriptome prediction models We downloaded full nominal eQTL summary statistics for 255 primary human microglia samples across four brain regions (medial frontal gyrus, superior temporal gyrus, subventricular zone, thalamus) from 100 human subjects from the Microglia Genomic Atlas (MiGA) 17 . Details on genotyping, RNA-seq generation and processing, and eQTL mapping can be found in the MiGA flagship paper 29 . Using the multivariate adaptive shrinkage approach to eQTL analysis introduced by Urbut et al . 30 , we trained region-specific Multivariate Adaptive Shrinkage in R (MASHR) transcriptome prediction models. For each brain region, we selected the top five eQTLs with the lowest local false sign rate (a measure analogous to false discovery rate) 30 per gene and included these eQTLs in the final models. In cases where multiple eQTLs were tied for lowest local false sign rate, we retained all eQTLs. An initial top-eQTL-per-gene specification was evaluated but not used in downstream analyses due to limited overlap with the AD GWAS variant set. Imputation of transcriptomic signatures for Alzheimer’s disease We applied two statistical methods, S-PrediXcan 11 and S-MultiXcan 12 , to publicly available Alzheimer’s GWAS summary statistics to compute transcriptomic signatures for AD. S-PrediXcan and S-MultiXcan predict gene expression for a trait or disease of interest using prediction models trained on reference transcriptome datasets. S-PrediXcan computes single-tissue gene-level association results from GWAS summary statistics. S-MultiXcan then aggregates the single-tissue S-PrediXcan results across multiple tissues, thereby increasing the statistical power to detect associations. We used GWAS summary statistics for a total of 762,917 individuals (86,531 AD cases and 676,386 controls, excluding participants from 23andMe, which were not available due to data access restrictions) 14 to generate all AD transcriptomic signatures used in this study. To improve GWAS-QTL integration, we harmonized the AD GWAS summary statistics to GTEx v8 variants and imputed summary statistics for missing variants 31 . We first used S-PrediXcan to impute single-tissue gene expression levels in 49 available GTEx tissues, using pre-developed MASHR expression prediction models 11,32,33 . We then ran S-MultiXcan on the S-PrediXcan results to predict gene expression across multiple tissues. Although AD predominantly affects the brain, studies have suggested that peripheral tissues such as the skin and vascular tissues can also capture genetic effects on gene expression in AD and may even represent potential pathogenic tissues in AD 15 . Thus, we conducted three separate S-MultiXcan analyses to predict AD-associated changes in gene expression combining brain and non-brain tissues, including: (1) a brain-specific analysis combining predictions from the 13 GTEx brain tissues, (2) an AD-relevant tissue analysis combining the 13 brain tissues with four other tissues previously related to AD (whole blood, spleen, and two skin tissues) 15,16 , and (3) a non-specific analysis combining predictions from all 49 GTEx tissues. We then constructed a separate AD virtual transcriptomic signature for each of the brain-restricted, AD-relevant, and all-tissue S-MultiXcan analyses using the AD risk genes with ≥2/3 tissue concordance (i.e., showing the same directionality of gene expression changes in at least two-thirds of the tissues with non-NA results). We required this tissue-wide consensus because several genes with statistically significant S-PrediXcan results exhibited opposing effects across tissues, producing near-zero effect estimates when averaged in S-MultiXcan analysis (e.g., APOE in the all-tissue analysis, Supplementary Fig. 7). The tissue-specific S-PrediXcan results for the AD risk genes, their consensus direction of effect, and mean Z score computed across the tissues with the consensus effect direction are available in Supplementary Data File 1, sheets 1-3. We defined AD risk genes using Bonferroni correction ( P 0) or inversely associated with AD risk (mean Z score<0). We also used S-PrediXcan and S-MultiXcan to calculate a microglia-specific AD gene expression signature using the microglia MASHR transcriptome prediction models we trained with MiGA data. We then applied S-PrediXcan to predict microglial gene expression in the medial frontal gyrus, superior temporal gyrus, subventricular zone, and thalamus. We initially used S-MultiXcan to construct a single AD gene expression signature integrating the S-PrediXcan results for microglia in the four different brain regions; however, the high correlations between the S-PrediXcan associations across the brain regions prevented computation of S-MultiXcan association statistics for >80% (7,852/9,687) of the genes tested. We thus used a generalized Berk-Jones (GBJ) test to combine the S-PrediXcan Z scores from the four brain regions 34 . Again, AD risk genes were defined using a Bonferroni-corrected significance threshold (GBJ P <0.05/18,222 genes tested). As GBJ does not provide directional effects (GBJ statistic is always positive), we constructed the microglial signature including only the AD risk genes with concordant effect directions across all regions with available results. NA values did not influence the concordance assessment. To capture inverse associations with AD risk, we negated the GBJ statistic for genes with consistently negative S-PrediXcan Z scores across the brain regions. Identification of drug repurposing candidates We queried the Connectivity Map (CMap) to identify drugs with perturbation profiles opposing TWAS-derived AD signatures. CMap contains over 1.5 million gene expression profiles (including 978 directly measured landmark genes) capturing the effects of over 8,000 small-molecule compounds and genetic reagents across multiple human cell lines 35 . Each AD signature was represented as two gene sets split by TWAS direction ( Z >0 “up”; Z <0 “down”) and submitted to the CMap CLUE Query tool, which returns a connectivity score (τ) quantifying similarity between the queried signature and each perturbagen signature. Small-molecule compounds with τ<0 were considered potential repurposing candidates, consistent with prior CMap-based studies 6,19 . For downstream evaluation, we prioritized compounds with negative connectivity across multiple AD signatures, feasible exposure ascertainment in real-world data, and sufficient expected numbers of exposed individuals for longitudinal analysis. Clinical evaluation using real-world data To evaluate the prioritized aspirin signal in longitudinal clinical data, we investigated the association between aspirin exposure and AD using EHR and claims data. In the two EHR databases, we performed a retrospective cohort study evaluating the risk of new AD diagnosis after age 65 in individuals previously exposed to aspirin compared to individuals without documented exposure. In the claims database, we performed a case-control study comparing the odds of prior aspirin use in AD cases and controls. Data sources We performed clinical validation studies using diagnosis and medication data from three independent sources: (1) VUMC’s Synthetic Derivative (SD) database, (2) the NIH All of Us Research Program database, and (3) the MarketScan Commercial Claims and Encounters (CCAE) and Medicare Supplemental (MDCR) databases. The VUMC SD contains decades of longitudinal clinical data, including diagnosis and procedure codes, laboratory test results, and medications, extracted from de-identified EHRs for over 4 million unique patients 36 . The SD is linked to VUMC’s biobank, BioVU, which contained over 300,000 unique DNA samples as of January 2023, allowing for the integration of EHR and genetic data. At the time of the study, the All of Us Research Program database contained data for over 633,540 participants, with genomic sequencing data for over 414,840 participants 37 . The MarketScan CCAE and MDCR databases collectively contain detailed insurance claims data, including diagnosis codes and pharmacy claims, for over 200 million unique patients 38 . Cohort formation and outcome assessment: VUMC and All of Us To capture endpoints relevant to AD and ensure adequate EHR follow-up time was available for all patients in the study population, we restricted our analysis to patients over 65 years of age with at least one visit at ≥75 years. We excluded individuals with AD diagnosed at ≤65 years, as well as individuals with diagnoses of non-AD dementia (occurring at any time). Information on patient medications is available as structured data in the VUMC and All of Us EHR databases (within a DRUG_EXPOSURE table, in accordance with the Observational Medical Outcomes Partnership Common Data Model 39 ). Drug exposures documented in this table may originate from various sources in the EHR, including clinical notes (e.g., medication lists appearing in history and physical [H&P] reports, progress notes, and discharge summaries), inpatient and outpatient medication orders (e.g., prescriptions), and historical medication records. We identified aspirin-exposed patients by mapping patient medications to their ingredients using the RxNorm standardized terminology for clinical drugs and filtering for all medications containing aspirin (RxCUI = 1191). We used aspirin-exposed patients with at least one year of documented aspirin use prior to age 65 to form the aspirin-exposed group; patients without any documented aspirin use were considered unexposed controls. Patients with first aspirin use occurring after age 65 were excluded. We then identified two groups within the aspirin-exposed cohort for additional comparisons based on the two most common doses at VUMC: high-dose (≥325mg/day) and low-dose (≤81mg/day) aspirin users, excluding individuals exposed to intermediate aspirin doses and individuals switching between low-dose and high-dose aspirin use. Aspirin is used in the treatment of multiple medical conditions, including in doses of 81 and 325 mg for primary and secondary prevention of atherosclerotic cardiovascular disease 40 , and historically, in much higher doses for pain relief in inflammatory diseases such as rheumatoid arthritis 41 , which may also impact AD risk. To minimize the effects of confounding by indication and other sources of confounding, we matched patients in the aspirin-exposed cohort to control patients with no recorded aspirin use in a 1:2 ratio, using a propensity score based on sex, race, EHR time after age 65 (i.e., length of time between age 65 and last EHR record), and presence of clinical indications for aspirin use (cardiovascular disease, cerebrovascular disease, and rheumatoid arthritis) at baseline (defined as the time zero of age 65 or, if no data available at age 65, the age of the first EHR visit after 65). We used the MatchIt R 42 package to perform nearest-neighbor propensity score matching. The diagnoses used to define the comorbidities for matching are provided in Supplementary Data File 1, sheets 14-16. We defined AD cases using a requirement of at least one AD diagnosis code (ICD-9-CM 331.0; ICD-10-CM G30.1, G30.8, G30.9). We excluded patients with a first recorded AD code before age 65 and those with codes for non-AD dementias (Supplementary Data File 1, sheet 17). Statistical analysis: VUMC and All of Us We used Cox proportional hazards regression models to investigate incident AD risk after age 65 in aspirin-exposed and unexposed individuals. Age 65 served as time zero; individuals were followed until first recorded AD diagnosis or otherwise right censored at last recorded EHR observation. We first compared AD risk between the aspirin-exposed cohort and the propensity score-matched unexposed cohort. We then performed subgroup analyses based on aspirin dose (high-dose versus low-dose versus no aspirin), documented aspirin exposure rate, and APOE ε4 genotype. We used the metafor R package for meta-analysis of hazard ratios 43 . Heterogeneity was assessed using Cochran’s Q and I 2 . Based on these metrics, all meta-analyses were conducted under a fixed-effects model. Documented aspirin exposure rate The inconsistent recording of medication end dates, dosing frequency (e.g., daily versus as needed), and therapy duration within the EHR, along with multiple sources of medication documentation in patient charts (including medication lists in clinical notes as well as prescriptions), hindered precise quantification of total aspirin exposure using EHR data. To address this limitation, we developed a proxy measure: the documented aspirin exposure rate, defined as the total number of unique aspirin records divided by the time (in years) between the first and last recorded aspirin exposures. This measure was intended to capture the frequency and duration of aspirin use documented in the EHR, with a higher rate reflecting more consistent and sustained aspirin exposure. We calculated the documented aspirin exposure rate for all individuals in the aspirin-exposed cohort and then classified aspirin users into high- and low-exposure groups based on the median documented exposure rate of 5 aspirin records per year. To ensure balanced comparisons, we matched individuals with high documented aspirin exposure rates to those with low rates in a 1:1 ratio using propensity score matching. The variables used in matching were sex, race, baseline comorbidities (cardiovascular disease, cerebrovascular disease, rheumatoid arthritis), EHR time after age 65, total number of EHR visits, and aspirin duration. The final matched cohort comprised 3,690 individuals ( N = 1,845 per group). A Cox proportional hazards regression model was used to investigate the risk of AD after age 65 in the high-exposure group relative to the low-exposure group. APOE genotyping APOE genotype was determined using the combination of alleles at SNPs rs429358 and rs7412. The APOE ε4 variant was defined as the presence of a C allele at both SNPs. Genetic data was available for 1,856 patients in the VUMC cohort (including 41 APOE ε4 homozygotes and 430 heterozygotes) and 1,450 patients in the All of Us cohort (with 23 APOE ε4 homozygotes and 271 heterozygotes). Information on APOE genotype was not available in the MarketScan dataset. MarketScan validation Given the shorter observation time in the MarketScan Research Databases, which prevented us from capturing AD-relevant timepoints in patients exposed to aspirin before age 65, we performed a case-control study to investigate the association between aspirin use and AD. We first identified AD cases using ICD-9-CM code 331.0 and ICD-10-CM codes G30.1, G30.8, and G30.9. We matched AD cases to comparable controls in a 1:2 ratio based on propensity score, using sex, comorbidities (cardiovascular disease, cerebrovascular disease, and rheumatoid arthritis, diagnosed at any age), and claims follow-up time (difference in years between first and last claims records) as covariates. We did not match on race as this is not reported in MarketScan. Aspirin prescriptions were identified using National Drug Codes. We then calculated the odds ratio for aspirin exposure among the AD cases compared to their matched controls. Human iPSC culture and cortical organoid generation Isogenic control wild-type (WT) and heterozygous APP mutant (KM670/671NL) iPSCs were maintained in mTeSR1 medium and passaged every 6-7 days. Human cortical organoids were generated following a previously established differentiation protocol, with modifications 44,45 . Briefly, iPSC colonies were enzymatically detached using 1 mg/mL collagenase IV for 1 hour. The iPSC colonies were collected and cultured as embryoid bodies (EBs) in DMEM/F12 medium (Invitrogen) supplemented with 20% Knockout Serum Replacement (KSR, Invitrogen), 1× GlutaMAX (Invitrogen), 1×MEM Non-Essential Amino Acids (Invitrogen), 0.1 mM beta-mercaptoethanol (Invitrogen), 2 µM Dorsomorphin (PeproTech), and 2 µM A-83 (PeproTech). EBs were maintained in 10 cm low-attachment dishes for 5 days, allowing uniform spheroid formation. From day 6 to day 16, EBs were changed into a neural medium (NM) comprising Neurobasal medium (Invitrogen), 1× B-27 supplement (minus vitamin A, Invitrogen), 1× GlutaMAX (Invitrogen), and 100 U/mL penicillin-streptomycin (Invitrogen), further supplemented with 20 ng/mL bFGF (Pepro Tech) and 20 ng/mL EGF (Pepro Tech). The medium was replaced daily to support robust neuroectodermal patterning. Between day 17 and day 24, cultures were maintained in the same medium with changes every other day. On day 25, the neural medium was supplemented with 20 ng/mL BDNF, with medium renewal continued every other day. By day 43, the medium was shifted to differentiation medium with growth factor-free neural medium, consisting of Neurobasal, 1× B-27 supplement (minus vitamin A), 1× GlutaMAX, and 100 U/mL penicillin-streptomycin, refreshed every four days. From day 70 onward, organoids were cultured in NM supplemented with 1× B-27 containing vitamin A, with medium changes every three days, promoting long-term maturation. At day 90, mature human cortical organoids were treated with acetylsalicylic acid (aspirin; Sigma A5376; 0.5 mM) or vehicle (PBS) for one week. RNA extraction and RNA-seq Human cortical organoids were homogenized in TRIzol Reagent (Invitrogen, 15596018), and total RNA was extracted using the Direct-zol TM RNA Miniprep Kit (Zymo, R2052) according to the manufacturer’s instructions. Total RNA was quantified using the Qubit 2.0 Fluorometer (ThermoFisher Scientific, Waltham, MA, USA) and assessed for integrity with the 4200 TapeStation (Agilent Technologies, Palo Alto, CA, USA). Strand-specific libraries were prepared using the NEBNext Ultra II Directional RNA Library Prep Kit for Illumina (NEB, Ipswich, MA, USA), following the manufacturer’s instructions. RNA was fragmented at 94 °C for 8 minutes, and first- and second-strand cDNA synthesis was performed, with dUTP incorporated during second-strand synthesis to maintain strand specificity. After 3’ end adenylation, adapter ligation, and limited-cycle PCR amplification, libraries were validated using the Agilent TapeStation and quantified by Qubit 2.0 (ThermoFisher Scientific) and qPCR (KAPA Biosystems, Wilmington, MA, USA). Libraries were multiplexed, clustered onto a flowcell, and sequenced on the Illumina NovaSeq 6000 system using a 2 × 150 bp paired-end configuration, according to the manufacturer’s protocol. Image analysis and base calling were performed using the NovaSeq Control Software (Illumina, San Diego, CA, USA), and raw BCL files were converted to FASTQ format and demultiplexed with bcl2fastq v2.20 (Illumina), allowing one mismatch for index recognition. RNA-seq analysis RNA-seq data were processed following a previously described workflow. Raw reads were quality-checked to confirm that library preparation and sequencing met requirements for downstream analyses. Adapters were removed with Trimmomatic 46 . Cleaned reads were aligned to the human hg38 reference genome using HISAT2 47 , and read counts were generated with featureCounts 48 . Differential expression was assessed with DESeq2 49 , controlling the FDR at 0.05 with lfcThreshold = log2(1.1). We analyzed three contrasts: (1) a baseline AD signature (APP vehicle vs WT vehicle), (2) aspirin in a non-AD background (WT aspirin vs WT vehicle), and (3) aspirin in an AD background (APP aspirin vs APP vehicle). Functional enrichment analyses were performed in R using clusterProfiler 50 and fgsea 51 . Over-representation analysis (ORA) was conducted on differentially expressed genes (|log2FC| ≥log2(1.1), FDR<0.05) for each contrast, with enrichment significance evaluated by hypergeometric test using all genes tested in the differential expression analysis as background. GSEA was performed on preranked gene lists for each contrast using GO, KEGG, and Reactome gene set collections, with genes ranked by Wald statistic (log2 fold change/lfcSE). For targeted neuronal analyses, human MSigDB gene sets were curated by keyword filtering of pathway names (“synapse”, “axon”, “neurotransmitter”); genes were again ranked by Wald statistic. For all enrichment analyses, multiple testing correction was performed using the Benjamini-Hochberg method, and significance was defined as FDR<0.05. RRHO was performed using RRHO2 52 to compare expression patterns between contrasts. Genes in each contrast were ranked by signed significance [-log 10 ( P ) × sign(log2 fold change)], and the significance of overlap between two ranked lists was assessed using hypergeometric tests across rank thresholds (step size=100). ORA was performed on discordant overlap genes (i.e., genes with significant opposite direction changes between contrasts), as described above. Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki. Analyses of Vanderbilt University Medical Center de-identified electronic health record and genetic data were determined to be non-human subjects research by the Vanderbilt University Medical Center Institutional Review Board under IRB #211489. Analyses using NIH All of Us Research Program data and MarketScan Research Databases involved de-identified data and complied with applicable access and data use requirements. All of Us participants provided informed consent for participation in the All of Us Research Program. No additional informed consent to participate in the study was required because analyses did not involve participant recruitment, participant contact, or access to identifiable private information. Acknowledgements This work was supported by the National Institute on Aging under grants F30AG080885 (MEG), R01AG069900 (BL, WQW, ZW), and R01AG084550 (WQW, QF); by the National Institute of General Medical Sciences under grant T32GM007347 (MEG); and by the National Heart, Lung, and Blood Institute under grants R01HL163854 (QF), R01HL133786 (WQW), and R01HL171809 (WQW, QF). The primary datasets were obtained from Vanderbilt University Medical Center’s BioVU, which is supported by institutional funding, 1S10RR025141-01, and CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. Additional support was provided by the National Institutes of Health through grants P50GM115305 and U19HL065962. We acknowledge the expert technical support of the VANTAGE and VANGARD core facilities, supported in part by the Vanderbilt-Ingram Cancer Center (P30CA068485) and Vanderbilt Vision Center (P30EY08126). Validation datasets were obtained from the All of Us Research Program. We thank All of Us participants, whose contributions made this research possible. The All of Us Research Program is supported by the National Institutes of Health, Office of the Director: Regional Medical Centers: 1 OT2 OD026549; 1 OT2 OD026554; 1 OT2 OD026557; 1 OT2 OD026556; 1 OT2 OD026550; 1 OT2 OD026552; 1 OT2 OD026553; 1 OT2 OD026548; 1 OT2 OD026551; 1 OT2 OD026555; IAA #: AOD 16037; Federally Qualified Health Centers: HHSN 263201600085U; Data and Research Center: 5 U2C OD023196; Biobank: 1 U24 OD023121; The Participant Center: U24 OD023176; Participant Technology Systems Center: 1 U24 OD023163; Communications and Engagement: 3 OT2 OD023205; 3 OT2 OD023206; and Community Partners: 1 OT2 OD025277; 3 OT2 OD025315; 1 OT2 OD025337; 1 OT2 OD025276. The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. Author contributions MEG, WQW, BL, ZW, and QF conceived the study. MEG, WQW, BL, ZW, QF, RC, YZ, XZ, and ALD contributed to the study design. MEG performed the primary analyses. YZ performed the organoid experiments, and RC contributed to downstream analysis of the organoid data. MEG drafted the initial manuscript. All authors contributed to data interpretation, critically reviewed and revised the manuscript, and approved the final manuscript. Competing interests All authors declare no financial or non-financial competing interests. Code availability The microglial MASHR models trained and used in this study are available on Zenodo: https://doi.org/10.5281/zenodo.18156902. Data availability All data are available in the main text or the supplementary materials. The AD GWAS summary statistics used in this study are publicly available at https://cncr.nl/research/summary_statistics/. The microglia eQTL summary statistics from the Microglia Genomic Atlas used in this study can be downloaded from the NIAGADS Data Sharing Service using accession number NG00105.v3. The MASHR GTEx v8 transcriptome prediction models can be downloaded from PredictDB (https://predictdb.org/categories/downloads/). Access to VUMC’s EHR database requires institutional approval and compliance with a data use agreement. Data from the All of Us Research Program can be accessed through the Researcher Workbench (https://workbench.researchallofus.org). The MarketScan claims data used in this study can be requested from Merative®. References Grabowska, M. E., Huang, A., Wen, Z., Li, B. & Wei, W.-Q. Drug repurposing for Alzheimer’s disease from 2012–2022—a 10-year literature review. Frontiers in Pharmacology 14 , (2023). Musa, A. et al. A review of connectivity map and computational approaches in pharmacogenomics. Brief Bioinform 19 , 506–523 (2017). Li, B. & Ritchie, M. D. From GWAS to Gene: Transcriptome-Wide Association Studies and Other Methods to Functionally Understand GWAS Discoveries. Front Genet 12 , 713230 (2021). Minikel, E. V., Painter, J. L., Dong, C. C. & Nelson, M. R. Refining the impact of genetic evidence on clinical success. Nature 629 , 624–629 (2024). Kho, P. F. et al. Multi-tissue transcriptome-wide association study identifies eight candidate genes and tissue-specific gene expression underlying endometrial cancer susceptibility. Commun Biol 4 , 1211 (2021). Wu, P. et al. Integrating gene expression and clinical data to identify drug repurposing candidates for hyperlipidemia and hypertension. Nat Commun 13 , 46 (2022). Zong, N. et al. Computational drug repurposing based on electronic health records: a scoping review. NPJ Digit Med 5 , 77 (2022). Tan, G. S. Q., Sloan, E. K., Lambert, P., Kirkpatrick, C. M. J. & Ilomäki, J. Drug repurposing using real-world data. Drug Discov Today 28 , 103422 (2023). Zang, C. et al. High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data. Nat Commun 14 , 8180 (2023). Yan, C. et al. Leveraging generative AI to prioritize drug repurposing candidates for Alzheimer’s disease with real-world clinical validation. npj Digit. Med. 7 , 1–6 (2024). Barbeira, A. N. et al. Exploring the phenotypic consequences of tissue specific gene expression variation inferred from GWAS summary statistics. Nat Commun 9 , 1825 (2018). Barbeira, A. N. et al. Integrating predicted transcriptome from multiple tissues improves association detection. PLOS Genetics 15 , e1007889 (2019). Lonsdale, J. et al. The Genotype-Tissue Expression (GTEx) project. Nat Genet 45 , 580–585 (2013). Wightman, D. P. et al. A genome-wide association study with 1,126,563 individuals identifies new risk loci for Alzheimer’s disease. Nat Genet 53 , 1276–1282 (2021). Gerring, Z. F., Lupton, M. K., Edey, D., Gamazon, E. R. & Derks, E. M. An analysis of genetically regulated gene expression across multiple tissues implicates novel gene candidates in Alzheimer’s disease. Alzheimer’s Research & Therapy 12 , 43 (2020). Ongen, H. et al. Estimating the causal tissues for complex traits and diseases. Nat Genet 49 , 1676–1683 (2017). NG00105 - MiGA – Microglia Genomic Atlas. DSS NIAGADS https://dss.niagads.org/datasets/ng00105/. Zhao, Y., Chen, X., Chen, J. & Qi, X. Decoding Connectivity Map-based drug repurposing for oncotherapy. Briefings in Bioinformatics 24 , bbad142 (2023). Taubes, A. et al. Experimental and real-world evidence supporting the computational repurposing of bumetanide for APOE4-related Alzheimer’s disease. Nat Aging 1 , 932–947 (2021). Narcisse, D. I. et al. Comparative Effectiveness of Aspirin Dosing in Cardiovascular Disease and Diabetes Mellitus: A Subgroup Analysis of the ADAPTABLE Trial. Diabetes Care 47 , 81–88 (2024). Plaisier, S. B., Taschereau, R., Wong, J. A. & Graeber, T. G. Rank–rank hypergeometric overlap: identification of statistically significant overlap between gene-expression signatures. Nucleic Acids Res 38 , e169 (2010). Nguyen, T. N. M. et al. Long-term low-dose acetylsalicylic use shows protective potential for the development of both vascular dementia and Alzheimer’s disease in patients with coronary heart disease but not in other individuals from the general population: results from two large cohort studies. Alzheimer’s Research & Therapy 14 , 75 (2022). Ryan, J. et al. Randomized placebo-controlled trial of the effects of aspirin on dementia and cognitive decline. Neurology 95 , e320–e331 (2020). AD2000 Collaborative Group et al. Aspirin in Alzheimer’s disease (AD2000): a randomised open-label trial. Lancet Neurol 7 , 41–49 (2008). Parish, S. et al. Effects of aspirin on dementia and cognitive function in diabetic patients: the ASCEND trial. Eur Heart J 43 , 2010–2019 (2022). Kang, J. H., Cook, N., Manson, J., Buring, J. E. & Grodstein, F. Low dose aspirin and cognitive function in the women’s health study cognitive cohort. BMJ 334 , 987 (2007). Bateman, R. J. et al. Clinical and Biomarker Changes in Dominantly Inherited Alzheimer’s Disease. New England Journal of Medicine 367 , 795–804 (2012). Li, Y. et al. Timing of Biomarker Changes in Sporadic Alzheimer’s Disease in Estimated Years from Symptom Onset. Ann Neurol 95 , 951–965 (2024). Lopes, K. de P. et al. Genetic analysis of the human microglial transcriptome across brain regions, aging and disease pathologies. Nat Genet 54 , 4–17 (2022). Urbut, S. M., Wang, G., Carbonetto, P. & Stephens, M. Flexible statistical methods for estimating and testing effects in genomic studies with multiple conditions. Nat Genet 51 , 187–195 (2019). Best practices for integrating GWAS and GTEX v8 transcriptome prediction models. GitHub https://github.com/hakyimlab/MetaXcan/wiki/Best-practices-for-integrating-GWAS-and-GTEX-v8-transcriptome-prediction-models. Gamazon, E. R. et al. A gene-based association method for mapping traits using reference transcriptome data. Nat Genet 47 , 1091–1098 (2015). Barbeira, A. N. et al. Exploiting the GTEx resources to decipher the mechanisms at GWAS loci. Genome Biology 22 , 49 (2021). Sun, R. & Lin, X. Genetic Variant Set-Based Tests Using the Generalized Berk-Jones Statistic with Application to a Genome-Wide Association Study of Breast Cancer. Journal of the American Statistical Association 115 , 1079 (2019). Subramanian, A. et al. A Next Generation Connectivity Map: L1000 Platform and the First 1,000,000 Profiles. Cell 171 , 1437-1452.e17 (2017). Roden, D. M. et al. Development of a large-scale de-identified DNA biobank to enable personalized medicine. Clin Pharmacol Ther 84 , 362–369 (2008). All of Us Research Program Investigators et al. The ‘All of Us’ Research Program. N Engl J Med 381 , 668–676 (2019). Merative MarketScan Research Databases. https://www.merative.com/documents/merative-marketscan-research-databases. OMOP Common Data Model. https://ohdsi.github.io/CommonDataModel/index.html. Mainous, A. G., Tanner, R. J., Shorr, R. I. & Limacher, M. C. Use of Aspirin for Primary and Secondary Cardiovascular Disease Prevention in the United States, 2011–2012. Journal of the American Heart Association 3 , e000989 (2014). Solomon, D. H. et al. The potential benefits of aspirin for primary cardiovascular prevention in rheumatoid arthritis: a secondary analysis of the PRECISION Trial. Rheumatology (Oxford) 57 , 1364–1369 (2018). Ho, D., Imai, K., King, G. & Stuart, E. A. MatchIt: Nonparametric Preprocessing for Parametric Causal Inference. Journal of Statistical Software 42 , 1–28 (2011). Viechtbauer, W. Conducting Meta-Analyses in R with the metafor Package. Journal of Statistical Software 36 , 1–48 (2010). Kang, Y. et al. A human forebrain organoid model of fragile X syndrome exhibits altered neurogenesis and highlights new treatment strategies. Nat Neurosci 24 , 1377–1391 (2021). Kuehner, J. N. et al. 5-hydroxymethylcytosine is dynamically regulated during forebrain organoid development and aberrantly altered in Alzheimer’s disease. Cell Rep 35 , 109042 (2021). Bolger, A. M., Lohse, M. & Usadel, B. Trimmomatic: a flexible trimmer for Illumina sequence data. Bioinformatics 30 , 2114–2120 (2014). Kim, D., Paggi, J. M., Park, C., Bennett, C. & Salzberg, S. L. Graph-based genome alignment and genotyping with HISAT2 and HISAT-genotype. Nat Biotechnol 37 , 907–915 (2019). Liao, Y., Smyth, G. K. & Shi, W. featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics 30 , 923–930 (2014). Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15 , 550 (2014). Yu, G., Wang, L.-G., Han, Y. & He, Q.-Y. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. OMICS 16 , 284–287 (2012). Korotkevich, G. et al. Fast gene set enrichment analysis. 060012 Preprint at https://doi.org/10.1101/060012 (2021). Cahill, K. M., Huo, Z., Tseng, G. C., Logan, R. W. & Seney, M. L. Improved identification of concordant and discordant gene expression signatures using an updated rank-rank hypergeometric overlap approach. Sci Rep 8 , 9588 (2018). Additional Declarations No competing interests reported. Supplementary Files npjSupplementaryMaterials.pdf SupplementaryDataFile1.xlsx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 06 May, 2026 Reviewers agreed at journal 06 May, 2026 Reviewers agreed at journal 06 May, 2026 Reviewers agreed at journal 06 May, 2026 Reviewers invited by journal 06 May, 2026 Editor assigned by journal 29 Apr, 2026 Submission checks completed at journal 29 Apr, 2026 First submitted to journal 24 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9518587","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":636685460,"identity":"27d133df-27d5-4795-a7cc-8004dc5a05e6","order_by":0,"name":"Monika E. Grabowska","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Monika","middleName":"E.","lastName":"Grabowska","suffix":""},{"id":636685462,"identity":"b452bd96-9b70-4947-8535-dd4a4e1059a9","order_by":1,"name":"Rui Chen","email":"","orcid":"","institution":"Vanderbilt University","correspondingAuthor":false,"prefix":"","firstName":"Rui","middleName":"","lastName":"Chen","suffix":""},{"id":636685464,"identity":"21b2214f-c8ea-4dfc-9fc8-0b7c21e29d16","order_by":2,"name":"Ying Zhou","email":"","orcid":"","institution":"Emory University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Zhou","suffix":""},{"id":636685466,"identity":"e821259d-5409-4223-998f-2ba2cb5613e9","order_by":3,"name":"Avi U. Vaidya","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Avi","middleName":"U.","lastName":"Vaidya","suffix":""},{"id":636685468,"identity":"b1f8a54c-fbcf-4a1e-8264-71bfd26746c9","order_by":4,"name":"Xue Zhong","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Xue","middleName":"","lastName":"Zhong","suffix":""},{"id":636685469,"identity":"1f0386c8-9b3f-4222-b796-808d1fe4383d","order_by":5,"name":"Chris Guardo","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Chris","middleName":"","lastName":"Guardo","suffix":""},{"id":636685470,"identity":"1066309f-90ef-438e-b021-8a0ead8fa5d7","order_by":6,"name":"Alyson L. Dickson","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Alyson","middleName":"L.","lastName":"Dickson","suffix":""},{"id":636685472,"identity":"fd09cab3-3138-4634-9d83-17470747b283","order_by":7,"name":"Mojgan Babanejad","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Mojgan","middleName":"","lastName":"Babanejad","suffix":""},{"id":636685473,"identity":"c4715a39-2614-4cc0-8678-1835170fd547","order_by":8,"name":"Chao Yan","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Chao","middleName":"","lastName":"Yan","suffix":""},{"id":636685474,"identity":"1d921feb-2ec6-4739-b9e1-d99b64a4b3ab","order_by":9,"name":"Yi Xin","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Xin","suffix":""},{"id":636685475,"identity":"c4fb2a3d-8513-4b71-87cb-0899614e1d3f","order_by":10,"name":"Sergio Mundo","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Sergio","middleName":"","lastName":"Mundo","suffix":""},{"id":636685476,"identity":"1df873e2-ca43-4039-a6a5-0e66266396dc","order_by":11,"name":"Josh F. Peterson","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Josh","middleName":"F.","lastName":"Peterson","suffix":""},{"id":636685477,"identity":"6d42b779-de5b-416e-b0aa-789f7e3c774a","order_by":12,"name":"Lang Li","email":"","orcid":"","institution":"Ohio State University","correspondingAuthor":false,"prefix":"","firstName":"Lang","middleName":"","lastName":"Li","suffix":""},{"id":636685479,"identity":"6f635017-f0f5-445d-bb84-938092174dc2","order_by":13,"name":"Peter Embí","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Peter","middleName":"","lastName":"Embí","suffix":""},{"id":636685480,"identity":"2ef140d5-8252-467e-b6ca-a2bbe30bc9f8","order_by":14,"name":"QiPing Feng","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"QiPing","middleName":"","lastName":"Feng","suffix":""},{"id":636685483,"identity":"4f100a43-a085-465d-ad65-bc74d42f71f4","order_by":15,"name":"James Eaton","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"James","middleName":"","lastName":"Eaton","suffix":""},{"id":636685484,"identity":"f191a4d4-40cc-4075-9a24-cabc86df089b","order_by":16,"name":"Zhexing Wen","email":"","orcid":"","institution":"Emory University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Zhexing","middleName":"","lastName":"Wen","suffix":""},{"id":636685486,"identity":"af00d66f-48fb-42d4-b8ab-abf4e7caacde","order_by":17,"name":"Bingshan Li","email":"","orcid":"","institution":"Vanderbilt University","correspondingAuthor":false,"prefix":"","firstName":"Bingshan","middleName":"","lastName":"Li","suffix":""},{"id":636685487,"identity":"a4200236-69ea-4446-b789-8750f5aa8e03","order_by":18,"name":"Wei-Qi Wei","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0UlEQVRIiWNgGAWjYFACxgYGBgMgzcx8gGQtbAkkW8djQJw6g+PNDQxvChgS+9l5vknz5tgx8Esfv4Bfy5mDDYxzDBgSZzbzbpPm3ZbMINmXU4BXi9mNxPbfQCflbjjMu9mYd9sBoCE8Cfi13H/YwAzSsv8wz2MitdxghGjZwMzD+Biihf0AXi32ZxJBfpGon3GYzfDh3G3JPJI9PHh1MEi2H3/A8OaPjTF//+EHB95us5Pj52F/gF8PCPAwSCCxiYogNJcQY8soGAWjYBSMJAAA4G0/s6u6rKcAAAAASUVORK5CYII=","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Wei-Qi","middleName":"","lastName":"Wei","suffix":""}],"badges":[],"createdAt":"2026-04-24 14:53:58","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9518587/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9518587/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109328854,"identity":"74505ea8-a424-448a-8e6f-cebd6b79d09e","added_by":"auto","created_at":"2026-05-15 15:31:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":293481,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIntegrative drug repurposing framework for Alzheimer’s disease.\u003c/strong\u003e 1) Signature development: genetically informed AD signatures across 49 GTEx tissues and primary human microglia were derived from GWAS summary statistics using the S-PrediXcan and S-MultiXcan transcriptome-wide association study methods. 2) Candidate identification and prioritization: these AD signatures were queried against Connectivity Map (CMap) perturbational profiles to identify compounds predicted to reverse AD-associated expression changes. 3) Clinical validation: the leading candidate, aspirin, was evaluated using pharmacoepidemiologic analyses in longitudinal real-world clinical data, including electronic health record (EHR) data from Vanderbilt University Medical Center (VUMC), EHR data from the \u003cem\u003eAll of Us\u003c/em\u003e Research Program, and healthcare claims data from the MarketScan Research Databases.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/e5f9ca822c2811e68d453aa3.png"},{"id":109405524,"identity":"5a12f90c-40ce-42a5-9867-50b740ef7112","added_by":"auto","created_at":"2026-05-17 13:18:42","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":209189,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGenetically informed Alzheimer’s disease signatures across tissues and microglia. a)\u003c/strong\u003e Genes included in the TWAS-derived AD signatures from all GTEx tissues, AD-relevant GTEx tissues, brain GTEx tissues, and microglia. Each vertical bar represents an individual gene, with the direction indicating a positive (upward) or negative (downward) association with AD risk. \u003cstrong\u003eb)\u003c/strong\u003e Associations between genetically predicted expression and AD risk for selected AD-relevant genes. \u003cstrong\u003ec)\u003c/strong\u003e Overlap of genes across the GTEx- and microglia-derived AD signatures irrespective of direction of association. \u003cstrong\u003ed)\u003c/strong\u003e Overlap of genes shared across signatures with concordant direction of association with AD risk.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/5760b45fdf03116472b06bb3.png"},{"id":109405853,"identity":"32a56e71-621d-43b8-a1e1-cc87887ca98d","added_by":"auto","created_at":"2026-05-17 13:20:40","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":106788,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eLongitudinal real-world clinical evaluation of aspirin. a)\u003c/strong\u003e Associations between aspirin initiation before age 65 and incident AD after age 65 in electronic health record (EHR) data from Vanderbilt University Medical Center (VUMC) and the \u003cem\u003eAll of Us\u003c/em\u003e Research Program, including pooled estimates from meta-analysis. Forest plots show hazard ratios (HRs) with 95% confidence intervals (CIs); the dashed vertical line indicates the null (HR=1). \u003cstrong\u003eb)\u003c/strong\u003e Dose- and exposure-stratified analyses within the VUMC cohort. Comparisons evaluate high-dose (≥325mg/day) and low-dose (≤81mg/day) aspirin use versus no aspirin use, and high versus low documented aspirin exposure rate. Documented aspirin exposure rate was used as a proxy for cumulative aspirin exposure, with high exposure defined as above the cohort median of 5 records/year.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/9d347cca14a8c34574b2614c.png"},{"id":109328856,"identity":"77323592-058e-4b14-a2b2-aab56bf3cf91","added_by":"auto","created_at":"2026-05-15 15:31:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":252007,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRank-rank hypergeometric overlap between aspirin-induced transcriptional changes and the baseline APP mutation signature. a)\u003c/strong\u003e Rank-rank hypergeometric overlap (RRHO) heatmap comparing the aspirin response in wild-type organoids (WT aspirin vs. vehicle) with the baseline APP mutation signature (APP vehicle vs. WT vehicle). The strongest discordant overlap is observed among genes downregulated in APP and upregulated by aspirin in WT (upper left quadrant). \u003cstrong\u003eb) \u003c/strong\u003eRRHO heatmap comparing the aspirin response in APP mutant organoids (APP aspirin vs. vehicle) with the baseline APP mutation signature, showing predominantly concordant overlap (down-down in the upper right quadrant, up-up in the lower left quadrant). For RRHO, genes were ranked from most upregulated to most downregulated using -log\u003csub\u003e10\u003c/sub\u003e(\u003cem\u003eP\u003c/em\u003e) × sign(log2 fold change). Heatmap intensity reflects overlap significance across rank thresholds, shown as -log\u003csub\u003e10\u003c/sub\u003e(\u003cem\u003eP\u003c/em\u003e) from the hypergeometric test, with warmer colors indicating stronger overlap. \u003cstrong\u003ec)\u003c/strong\u003e Functional enrichment of the discordant genes from panel a (up in WT aspirin, down in APP). The top five enriched gene sets per database (GO BP, KEGG, Reactome) are shown; bars denote -log\u003csub\u003e10\u003c/sub\u003e(FDR). Gene sets related to neuronal signaling and organization are bolded and outlined in red.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/e0c310d2e09cc0818322972a.png"},{"id":109328859,"identity":"e15381bb-6f4e-4bdd-8f58-d0cd2c9b69d7","added_by":"auto","created_at":"2026-05-15 15:31:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":186858,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSynapse- and axon-related pathway enrichment across aspirin-treated wild-type and APP mutant organoids. a)\u003c/strong\u003e Gene set enrichment results for synapse-related MSigDB gene sets across three contrasts: wild-type (WT) aspirin versus vehicle, APP aspirin versus vehicle, and APP vehicle versus WT vehicle. \u003cstrong\u003eb)\u003c/strong\u003e Gene set enrichment results for axon-related MSigDB gene sets across the same contrasts. Only pathways with false discovery rate (FDR) \u0026lt;0.05 are shown. Bubbles are colored by normalized enrichment score (NES), with red indicating positive enrichment and blue indicating negative enrichment. Bubble size reflects statistical significance, shown as -log\u003csub\u003e10\u003c/sub\u003eFDR. Across both pathway families, aspirin induced stronger reversal of the baseline disease-associated pattern in WT than in APP mutant organoids.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/1592a169733e4077229540a5.png"},{"id":109406151,"identity":"44c4a9c8-d0f6-499b-b5d9-ef14204c1d36","added_by":"auto","created_at":"2026-05-17 13:25:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1381248,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/396d5f27-6c7f-4de8-8507-b43313e1eeaf.pdf"},{"id":109328853,"identity":"ca3bdd81-f134-43de-a2d0-d447f65efb9d","added_by":"auto","created_at":"2026-05-15 15:31:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":663644,"visible":true,"origin":"","legend":"","description":"","filename":"npjSupplementaryMaterials.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/750dfc51ae8f600049b5a1b5.pdf"},{"id":109405519,"identity":"5bc4860d-e04b-499a-ae7b-08f4e9dede2e","added_by":"auto","created_at":"2026-05-17 13:18:39","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1802317,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryDataFile1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9518587/v1/d36d089541122707b656e0e8.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating genetically predicted transcriptomic signatures with longitudinal real-world data enables scalable drug repurposing for Alzheimer’s disease","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMany serious diseases still lack effective therapies despite substantial investment in drug development. Drug repurposing offers a practical complement to de novo drug development by identifying new uses for existing drugs. However, many repurposing studies stop at candidate nomination, creating a bottleneck in determining which of the many proposed candidates warrant further study\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Scalable approaches that integrate large-scale data, human genetics, and epidemiological methods can help move drug repurposing beyond candidate identification toward systematic evaluation, with the potential to accelerate clinical translation and improve patient outcomes.\u003c/p\u003e \u003cp\u003eTranscriptomic signature reversal has become a widely used strategy for drug repurposing, supported by the growing availability of large-scale genetic and transcriptomic datasets. This approach queries perturbational gene expression resources, such as the Connectivity Map (CMap), to identify compounds predicted to reverse disease-associated transcriptional profiles\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. While early implementations often derived disease signatures from RNA sequencing or microarray data, transcriptome-wide association study (TWAS) methods offer a distinct genetics-informed approach by integrating expression quantitative trait loci (eQTLs) with genome-wide association study (GWAS) data to identify genes whose genetically predicted expression is associated with disease risk\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Because TWAS leverages germline genetic variation fixed at conception, it is less susceptible to reverse causation and may better prioritize causal genes and therapeutic targets. Given evidence that genetically supported drug targets are approximately 2.6 times more likely to succeed in clinical development\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, TWAS-based signature reversal has emerged as a promising strategy for drug discovery and repurposing in complex diseases and has been applied to endometrial cancer, hypertension, and hyperlipidemia\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAfter candidate identification, additional validation is needed to confirm preliminary signals and distinguish the most promising candidates. Real-world clinical data, including electronic health records (EHRs) and claims data, contain longitudinal information on medication exposures and clinical outcomes at scale and are increasingly used to generate and assess repurposing hypotheses in human populations\u003csup\u003e\u003cspan additionalcitationids=\"CR8 CR9\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. This is especially relevant for conditions with a long preclinical phase, such as Alzheimer\u0026rsquo;s disease (AD), as candidate effects may depend on exposure timing and duration. Alongside clinical data, experimental studies in cellular models, such as human induced pluripotent stem cell (iPSC)-derived organoids, can provide complementary biological support for candidate assessment. Given the time and cost of downstream prospective clinical studies, integrating these sources of evidence is critical for directing resources toward the most compelling candidates.\u003c/p\u003e \u003cp\u003eIn this study, we developed a genetics-informed repurposing workflow that integrates TWAS-derived disease signatures, perturbational screening, and longitudinal real-world clinical data. We then applied this workflow to AD. First, we identified AD-associated genes using GWAS summary statistics and eQTL data from bulk tissues and primary microglia, and used these genes to construct cross-tissue and microglia-specific disease signatures. We then queried CMap profiles to identify compounds opposing these signatures. Aspirin emerged as a recurrent candidate across multiple signatures and was subsequently evaluated in longitudinal EHR data from Vanderbilt University Medical Center (VUMC) and the National Institutes of Health \u003cem\u003eAll of Us\u003c/em\u003e Research Program, as well as in national claims data from the MarketScan Research Databases. We further examined aspirin-induced transcriptional responses in human iPSC-derived cortical organoids with wild-type and AD-associated APP mutant genotypes. Our findings illustrate how genetically informed candidate prioritization can be coupled with validation, including longitudinal clinical evaluation and experimental investigation, to advance drug repurposing and accelerate clinical implementation.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eGenetically informed Alzheimer\u0026rsquo;s disease transcriptomic signatures\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn overview of the study design is shown in Fig. 1. Because the compounds returned by CMap depend directly on the disease signature used for querying, we constructed multiple TWAS-derived AD signatures spanning both bulk tissues and microglia rather than relying on a single transcriptomic profile.\u003c/p\u003e\n\u003cp\u003eWe performed TWAS using S-PrediXcan\u003csup\u003e11\u003c/sup\u003e and S-MultiXcan\u003csup\u003e12\u003c/sup\u003e to identify candidate AD risk genes. Using pre-trained transcriptome prediction models for 49 tissues from the Genotype-Tissue Expression (GTEx) project\u003csup\u003e13\u003c/sup\u003e and summary statistics from a large AD GWAS\u003csup\u003e14\u003c/sup\u003e, we conducted three S-MultiXcan analyses to balance tissue specificity and statistical power: (1) a brain-specific analysis combining the 13 GTEx brain tissues, (2) an AD-relevant tissue analysis with the 13 brain tissues plus four peripheral tissues previously related to AD (whole blood, spleen, and sun-exposed and unexposed skin)\u003csup\u003e15,16\u003c/sup\u003e, and (3) an all-tissue analysis combining all 49 GTEx tissues. AD risk genes were defined within each S-MultiXcan analysis as Bonferroni-significant TWAS associations (\u003cem\u003eP\u003c/em\u003e\u0026lt;0.05/number of tested gene associations) and are reported in Supplementary Data File 1 (sheets 1-3).\u003c/p\u003e\n\u003cp\u003eWe then constructed three GTEx-derived AD transcriptomic signatures (brain, AD-relevant, and all tissues) using the gene inclusion criteria described in Methods. The final brain-tissue AD transcriptomic signature contained 72 genes (37 positively associated with AD risk, 35 inversely associated), while the AD-relevant tissue signature contained 78 genes (41 positively, 37 inversely associated) and the all-tissue signature contained 79 genes (40 positively, 39 inversely associated) (Fig. 2a). A shared set of 43 genes was observed across all three GTEx signatures (21 positively associated with AD risk, 22 inversely associated).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBecause bulk tissues can obscure cell type-specific signals and microglia play a central role in AD, we also constructed a microglial AD signature by performing TWAS with custom transcriptome prediction models trained on microglial eQTL summary statistics from the Microglia Genomic Atlas (MiGA)\u003csup\u003e17\u003c/sup\u003e. This signature comprised 53 genes (25 positively associated with AD risk, 28 inversely associated) (Fig. 2a). MiGA microglial TWAS results are provided in Supplementary Data File 1, sheet 4.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAcross the four signatures, we identified genes previously implicated in AD, including \u003cem\u003eAPOE\u003c/em\u003e, \u003cem\u003eTREM2\u003c/em\u003e, and \u003cem\u003eBIN1\u003c/em\u003e (Fig. 2b; Supplementary Table 1). An annotated comparison of the four AD transcriptomic signatures is shown in Supplementary Fig. 1. Nine genes were shared across all four signatures (Fig. 2c), of which seven had directionally concordant associations with AD risk (Fig. 2d).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSignature reversal prioritizes aspirin for downstream evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe queried the Connectivity Map (CMap)\u003csup\u003e18\u003c/sup\u003e using each TWAS-derived AD signature to identify compounds predicted to reverse disease-associated gene expression changes. Genes were assigned to up and down sets by TWAS \u003cem\u003eZ\u003c/em\u003e-score sign. Compounds with negative connectivity scores (\u0026tau;\u0026lt;0)\u0026nbsp;were considered repurposing candidates\u003csup\u003e6,19\u003c/sup\u003e. Out of 2,428 small-molecule compounds in CMap, 590 (24.3%), 709 (29.2%), and 688 (28.3%) had negative connectivity to the GTEx brain-tissue, AD-relevant tissue, and all-tissue signatures, respectively (Supplementary Data File 1, sheets 5-7). Of these, 218 showed negative connectivity to all three GTEx signatures, including mycophenolate, fluticasone, sirolimus, sertraline, clozapine, and losartan, although connectivity magnitude varied widely (e.g., mycophenolate\u0026nbsp;\u0026tau;: brain-tissue -72.2; AD-relevant tissue -7.6; all-tissue -0.6). Additionally, 794 (32.7%) compounds had negative connectivity to the microglia-specific signature (Supplementary Data File 1, sheet 8). The top ten repurposing candidates for each signature are shown in Table 1.\u003c/p\u003e\n\u003cp\u003eThe number and ranking of negatively connected compounds varied substantially across signatures, underscoring the importance of cross-signature support in candidate selection. Aspirin emerged as a recurrent candidate across multiple signatures, ranking among the top ten compounds for the all-tissue GTEx AD signature (\u0026tau;=-69.96), with negative connectivity also observed for the AD-relevant GTEx tissue signature (-17.89) and the MiGA microglial signature (-30.88), although no connectivity was detected for the GTEx brain-tissue signature. For downstream evaluation, we focused on compounds showing negative connectivity across more than one AD signature and practical feasibility for longitudinal clinical evaluation. Aspirin met these criteria, with support across three signatures, widespread clinical use, and a well-characterized safety profile.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1.\u003c/strong\u003e \u003cstrong\u003eTop ten AD repurposing candidates identified in CMap queries.\u003c/strong\u003e Drugs approved by the United States Food and Drug Administration are marked with an asterisk and their clinical indications are provided in parentheses.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGTEx brain tissues\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAD-relevant GTEx tissues\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAll GTEx tissues\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMicroglia\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003emupirocin*\u003c/p\u003e\n \u003cp\u003e(impetigo and other uncomplicated bacterial skin infections)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eCAY-10618\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eindinavir*\u003c/p\u003e\n \u003cp\u003e(human immunodeficiency virus)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eanisomycin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003emycophenolic-acid*\u003c/p\u003e\n \u003cp\u003e(prophylaxis of organ rejection, autoimmune disease)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ebufalin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eisogedunin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eemetine\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eBX-795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003evidarabine*\u003c/p\u003e\n \u003cp\u003e(herpes simplex keratitis)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003edeforolimus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ehomoharringtonine*\u003c/p\u003e\n \u003cp\u003e(chronic myeloid leukemia)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003efluticasone*\u003c/p\u003e\n \u003cp\u003e(asthma, allergic rhinitis, and certain inflammatory skin conditions)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ePP-30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ephorbol-12-myristate-13-acetate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003enarciclasine\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eprostratin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eivermectin*\u003c/p\u003e\n \u003cp\u003e(parasitic infections)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eU-46619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003edigitoxigenin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ePD-123319\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003erhodomyrtoxin-b\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eBX-795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003etroxipide\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ephenylbutyrate*\u003c/p\u003e\n \u003cp\u003e(urea cycle disorders)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003epentylenetetrazol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003easpirin*\u003c/p\u003e\n \u003cp\u003e(pain, fever, primary and secondary cardiovascular disease prevention, rheumatoid arthritis)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003epyrvinium-pamoate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003edesoxypeganine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eingenol*\u003c/p\u003e\n \u003cp\u003e(actinic keratosis)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003epirinixic-acid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eroscovitine\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eSA-63133\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eprostratin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eON-01910\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003ecycloheximide\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eindinavir*\u003c/p\u003e\n \u003cp\u003e(human immunodeficiency virus)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003esalubrinal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003epraziquantel*\u003c/p\u003e\n \u003cp\u003e(parasitic infections)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eisoliquiritigenin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003eReal-world clinical validation across three independent databases\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe evaluated the prioritized aspirin signal in three real-world datasets: (1) VUMC\u0026rsquo;s de-identified EHR database, (2) the NIH \u003cem\u003eAll of Us\u003c/em\u003e Research Program database, and (3) the MarketScan Research Databases. In EHR data from VUMC and \u003cem\u003eAll of Us\u003c/em\u003e, we used a retrospective cohort study design to compare AD incidence after age 65 between aspirin-exposed patients (\u0026ge;1 year of aspirin use before age 65) and propensity score-matched unexposed patients. In MarketScan claims, shorter follow-up limited ascertainment of incident AD among individuals with medication exposures documented before age 65; therefore, we used a case-control design comparing prior aspirin exposure in AD cases versus propensity score-matched controls. Descriptive characteristics for matched EHR cohorts are shown in Table 2. The characteristics of the matched MarketScan claims-based cohort are provided in Supplementary Table 2. Information on AD outcomes in all three datasets is provided separately in Supplementary Table 3.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u003c/strong\u003e Description of matched patient cohorts used in EHR validation studies.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical characteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 24px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVUMC\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e(\u003cem\u003eN\u003c/em\u003e = 19,413)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 25px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAll of Us\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e(\u003cem\u003eN\u003c/em\u003e = 1,995)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAspirin\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eExposed\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnexposed\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eExposed\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnexposed\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003eN\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e6,656\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e12,757\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e666\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e1,329\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean age at last follow-up (s.d.)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e77.8\u003c/p\u003e\n \u003cp\u003e(2.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e77.9\u003c/p\u003e\n \u003cp\u003e(2.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e77.3\u003c/p\u003e\n \u003cp\u003e(2.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e77.3\u003c/p\u003e\n \u003cp\u003e(2.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e49.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e51.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e45.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e45.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e50.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e48.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e52.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e52.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRace (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eWhite\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e92.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e92.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e77.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e77.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eBlack\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e6.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e6.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e7.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e7.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eOther\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e15.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e15.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBaseline comorbidities (%)\u003c/strong\u003e\u003cstrong\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eCardiovascular disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e25.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e21.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e28.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e28.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eCerebrovascular disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e6.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e8.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e7.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 50px;\"\u003e\n \u003cp\u003eRheumatoid arthritis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e3.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e4.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e4.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e4.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003e Baseline comorbidities were defined as \u0026ge;1 diagnosis code recorded at or before time zero (age 65 or first encounter after 65). Code lists are provided in Supplementary Data File 1, sheets 14-16.\u003c/p\u003e\n\u003cp\u003eIn VUMC, aspirin use before age 65 was associated with a significantly reduced risk of incident AD after age 65 (hazard ratio [HR]=0.77, 95% confidence interval [CI]: 0.65-0.91, \u003cem\u003eP\u003c/em\u003e=0.003; Fig. 3a). In \u003cem\u003eAll of Us\u003c/em\u003e, the association was directionally similar but limited by low statistical power (HR=0.40, 95% CI: 0.15-1.08, P=0.07; \u003cem\u003eN\u003c/em\u003e=24 AD events among 1,995 participants). Meta-analysis of the two EHR cohorts showed that aspirin initiation before age 65 was associated with 24% lower risk of incident AD (HR=0.76, 95% CI: 0.64-0.89, \u003cem\u003eP\u003c/em\u003e=0.001; Fig. 3a). In MarketScan, patients diagnosed with AD were less likely to have prior aspirin exposure compared to matched controls (OR=0.32, 95% CI: 0.28-0.38, \u003cem\u003eP\u003c/em\u003e\u0026lt;0.001).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe performed secondary analyses in VUMC to examine whether the association varied by aspirin dose or cumulative exposure. Due to the limited number of AD events in the \u003cem\u003eAll of Us\u003c/em\u003e cohort and the small number of aspirin prescriptions in MarketScan, we were unable to evaluate these measures in those datasets. In dose-stratified analyses (high-dose \u0026ge;325mg/day; low-dose \u0026le;81mg/day\u003csup\u003e20\u003c/sup\u003e), aspirin use remained associated with lower AD risk for both high-dose (HR=0.63, 95% CI: 0.45-0.90, \u003cem\u003eP\u003c/em\u003e=0.01) and low-dose regimens (HR=0.82, 95% CI: 0.68-0.99, \u003cem\u003eP\u003c/em\u003e=0.04), with no significant difference between dose groups (\u003cem\u003eP\u003c/em\u003e=0.19). Because higher aspirin doses may reflect more severe underlying indications, we also examined cumulative exposure. Given inconsistent documentation of dosing frequency and treatment duration in the EHR, we used documented aspirin exposure rate, defined as the number of unique aspirin records divided by the years between first and last recorded exposure, as a proxy. Exposure rates above the cohort median (\u0026gt;5/year) were associated with lower AD risk (HR=0.58, 95% CI: 0.39-0.88, \u003cem\u003eP\u003c/em\u003e=0.009; Fig. 3b).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn addition, we conducted \u003cem\u003eAPOE\u003c/em\u003e-stratified analyses in VUMC and \u003cem\u003eAll of Us\u003c/em\u003e (MarketScan does not contain genetic data). Among \u003cem\u003eAPOE\u0026nbsp;\u003c/em\u003e\u0026epsilon;4 carriers, aspirin use before age 65 showed a suggestive inverse association with incident AD in both cohorts, although neither analysis reached statistical significance individually (VUMC HR=0.60, 95% CI: 0.33-1.10, \u003cem\u003eP\u003c/em\u003e=0.0986; \u003cem\u003eAll of Us\u003c/em\u003e HR=0.48, 95% CI: 0.10-2.25, \u003cem\u003eP\u003c/em\u003e=0.35). Combined meta-analysis showed a 41% decreased risk of incident AD after age 65 in \u003cem\u003eAPOE\u0026nbsp;\u003c/em\u003e\u0026epsilon;4 carriers (HR=0.59, 95% CI: 0.33-1.02, \u003cem\u003eP\u003c/em\u003e=0.06), suggesting a potentially stronger protective association in this subgroup but limited by statistical power. Aspirin use was not significantly associated with decreased AD risk among non-carriers (VUMC HR=0.63, 95% CI: 0.26-1.55, \u003cem\u003eP\u003c/em\u003e=0.317; \u003cem\u003eAll of Us\u003c/em\u003e HR=0.52, 95% CI: 0.15-1.79, \u003cem\u003eP\u003c/em\u003e=0.3; meta-analysis HR=0.59, 95% CI: 0.29-1.22, \u003cem\u003eP\u003c/em\u003e=0.16), although the power of these analyses was limited by the small number of AD cases.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTranscriptomic evaluation in human iPSC-derived cortical organoids\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess the biological plausibility of the aspirin signal, we treated 90-day-old cortical organoids derived from isogenic control wild-type (WT; WTC11) and heterozygous APP mutant (KM670/671NL) iPSCs with aspirin (0.5 mM) or vehicle (PBS) for one week, followed by RNA sequencing (RNA-seq). High uniquely mapped paired-end rates (~90% across samples) indicated successful library preparation and sequencing (Supplementary Data File 1, sheet 9). Replicates showed high concordance and clustered cleanly by their assigned labels (Supplementary Fig. 2), supporting the expected data quality.\u003c/p\u003e\n\u003cp\u003eOver-representation analysis of differentially expressed genes (|log2FC|\u0026nbsp;\u0026ge;log2(1.1), FDR\u0026lt;0.05) highlighted synapse- and axon-related pathways, including glutamatergic synapse and axon guidance, in both the baseline APP mutation signature and the WT aspirin response (Supplementary Fig. 3a). Gene set enrichment analysis (GSEA) showed broad downregulation of neuronal and synaptic terms in the baseline APP signature, including synaptic signaling (NES=-2.29, FDR=4.11\u0026times;10\u003csup\u003e-31\u003c/sup\u003e) and axon development (NES=-2.11, FDR = 2.76\u0026times;10\u003csup\u003e-15\u003c/sup\u003e) (Supplementary Fig. 3b). In WT organoids, aspirin produced the opposite pattern, with strong upregulation of synaptic signaling (NES=2.52, FDR=8.16\u0026times;10\u003csup\u003e-48\u003c/sup\u003e) and axon development (NES=2.56, FDR=3.75\u0026times;10\u003csup\u003e-39\u003c/sup\u003e) (Supplementary Fig. 3c). In contrast, aspirin-treated APP mutant organoids showed attenuated enrichment of neuronal terms relative to WT, with top GSEA hits predominantly reflecting cell cycle and chromosome-associated processes (Supplementary Fig. 3d). Full GSEA results for all contrasts are provided in Supplementary Data File 1, sheets 10-12.\u003c/p\u003e\n\u003cp\u003eIn WT organoids, the transcriptional effects of aspirin were modestly but significantly negatively correlated with the baseline APP mutation signature (Spearman\u0026rsquo;s \u0026rho;=-0.15, \u003cem\u003eP\u003c/em\u003e\u0026lt;2.2\u0026times;10\u003csup\u003e-16\u003c/sup\u003e, \u003cem\u003eN\u003c/em\u003e=30,848 genes), consistent with partial reversal of disease-associated changes. Rank-rank hypergeometric overlap (RRHO) analysis\u003csup\u003e21\u003c/sup\u003e identified a discordant hotspot comprising genes upregulated by aspirin in WT organoids and downregulated in the baseline APP signature (Fig. 4a). In APP mutant organoids, the aspirin response was positively correlated with the APP signature (\u0026rho;=0.37, \u003cem\u003eP\u003c/em\u003e\u0026lt;2.2\u0026times;10\u003csup\u003e-16\u003c/sup\u003e, \u003cem\u003eN\u003c/em\u003e=30,848 genes), and RRHO showed a dominant concordant overlap signal (Fig. 4b), providing little evidence of global transcriptomic reversal in the APP background. Genes in the discordant hotspot in Fig. 4a were enriched for synaptic signaling and neurotransmission across GO Biological Process, KEGG, and Reactome databases, with top terms including regulation of trans-synaptic signaling, glutamatergic and GABAergic synapse, and transmission across chemical synapses (Fig. 4c; Supplementary Data File 1, sheet 13).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Building on these findings, we performed targeted GSEA of synapse-, axon-, and neurotransmitter-related gene sets curated from MSigDB using keyword filters. The baseline disease signature (APP vehicle vs. WT vehicle) showed broad downregulation across these pathway families, consistent with neuronal dysfunction in AD. In WT organoids, aspirin robustly upregulated these pathways, counteracting the baseline AD signature (synapse and axon pathways in Fig. 5 and Supplementary Figs. 4-5; neurotransmitter pathways in Supplementary Fig. 6). By contrast, in APP mutant organoids, aspirin\u0026rsquo;s effects were attenuated and did not consistently oppose the baseline disease signature.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe developed an integrative drug repurposing framework that combines genetically informed transcriptomic disease signatures, perturbational signature matching, longitudinal real-world clinical data, and experimental investigation in human cellular models to prioritize candidates for further study and accelerate clinical translation. Applying this framework to AD, we identified aspirin as a candidate supported across multiple disease signatures, three independent clinical datasets, and transcriptomic analyses in human cortical organoids. This study illustrates how genetically informed candidate prioritization can be paired with large-scale longitudinal clinical data to evaluate repurposing hypotheses in a scalable and clinically relevant manner.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAn important feature of this study is the use of multiple disease signatures spanning both bulk tissues and microglia rather than a single transcriptomic representation of AD. Our TWAS analyses highlighted genes at established AD risk loci, including \u003cem\u003eAPOE\u003c/em\u003e, \u003cem\u003eTREM2\u003c/em\u003e, and \u003cem\u003eBIN1\u003c/em\u003e, and corroborated previously reported TWAS associations (Supplementary Table 1), while also revealing substantial variability across tissues and cell types. Only seven AD risk genes appeared in all four signatures with concordant effect directions. Notably, \u003cem\u003eBIN1\u003c/em\u003e, a leading late-onset AD risk locus, showed a positive association with AD risk in microglia but inverse associations in all GTEx signatures. This variability was also reflected in the top-ranked CMap compounds identified across the four TWAS-derived AD signatures, supporting the use of cross-signature consistency as a prioritization criterion. We therefore focused on compounds with negative\u0026nbsp;connectivity across more than one AD signature and practical feasibility for longitudinal clinical evaluation. Aspirin met these criteria, with support across the all-tissue GTEx, AD-relevant GTEx, and microglial signatures, making it a useful test case for this framework.\u003c/p\u003e\n\u003cp\u003eAcross three real-world datasets, aspirin exposure was consistently associated with lower AD risk. In meta-analysis of the two EHR cohorts, aspirin initiation before age 65 was associated with a 24% reduced risk of incident AD. The claims-based study, while constrained by shorter observation windows that precluded time-to-event analysis, showed directionally consistent results in a case-control analysis. We did not detect a significant difference between high- and low-dose regimens, but individuals with higher documented aspirin exposure rates, used here as a proxy for cumulative exposure, had lower AD risk than matched individuals with lower exposure rates. This observation is consistent with UK Biobank analyses in which the inverse association between low-dose aspirin and AD was most evident with long-term use (\u0026gt;10 years)\u003csup\u003e22\u003c/sup\u003e.\u0026nbsp;Analysis of EHR-linked genetic data suggested a stronger inverse association among individuals carrying at least one \u003cem\u003eAPOE\u0026nbsp;\u003c/em\u003eε4 allele, although statistical power was limited by the relatively small number of \u003cem\u003eAPOE\u0026nbsp;\u003c/em\u003eε4 carriers and low prevalence of AD diagnoses among non-carriers. Overall, these findings support the consistency of the aspirin signal across distinct clinical data sources and illustrate how longitudinal clinical data can refine repurposing signals beyond simple exposed-versus-unexposed comparisons.\u003c/p\u003e\n\u003cp\u003eTranscriptomic analysis of human iPSC-derived\u0026nbsp;cortical organoids provided complementary experimental support for the aspirin signal. In wild-type organoids, aspirin-induced transcriptional changes were directionally opposite to the APP mutation signature and were enriched for synaptic, axonal, and neurotransmission-related pathways suppressed in the baseline disease state. In contrast, aspirin responses in APP mutant organoids were attenuated and did not show the same degree of opposition to the baseline disease signature. These findings provide biological plausibility for the prioritized signal and suggest that aspirin’s effects may depend on underlying disease context.\u003c/p\u003e\n\u003cp\u003eRandomized trials have not demonstrated a clear population-level benefit of aspirin in AD, although most initiated treatment relatively late in life\u003csup\u003e23–26\u003c/sup\u003e, after AD-related pathology may already have been established\u003csup\u003e27,28\u003c/sup\u003e. Our findings do not contradict those trials; rather, they raise the possibility that any benefit of aspirin, if present, may depend on exposure timing and persistence during the long preclinical interval preceding diagnosis, when relevant neuronal pathways may still be modifiable. This interpretation is consistent with the genetically anchored, life-course nature of TWAS-derived signatures, and with our real-world analyses, which specifically emphasized aspirin initiation before age 65. It is also broadly concordant with the organoid results, where aspirin more strongly opposed disease-associated neuronal pathway alterations in wild-type than APP mutant backgrounds, suggesting reduced efficacy in a disease context.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study has several strengths. It integrates human genetics, perturbational screening, and multi-cohort real-world clinical data within a single repurposing workflow and evaluates the leading candidate across independent EHR and claims datasets, as well as through transcriptomic analysis in human iPSC-derived organoids. Using TWAS, we mapped AD genetic risk to genetically predicted gene expression across multiple tissues and in microglia, generating disease-relevant signatures for candidate identification. Longitudinal EHR data enabled evaluation of aspirin initiation over intervals that are difficult to study in conventional trials and supported additional analyses by exposure pattern and \u003cem\u003eAPOE\u0026nbsp;\u003c/em\u003eε4 status. Importantly, this framework is transferable to other complex diseases with available GWAS summary statistics and longitudinal EHR or claims data, where genetics-informed prioritization can be integrated with real-world data to support more systematic assessment of repurposing candidates.\u003c/p\u003e\n\u003cp\u003eThis study also has several limitations. Repurposing hypotheses were restricted to compounds profiled in CMap, which is not exhaustive and is derived largely from non-brain cell lines that may not fully reflect AD-relevant biology. TWAS was constrained by the ancestral composition of\u0026nbsp;available AD GWAS and eQTL reference datasets, and we lacked power for ancestry-stratified EHR analyses despite known differences in AD incidence among populations, which may limit generalizability. AD case definitions relied on structured diagnosis codes and may incompletely capture clinical heterogeneity and potentially introduce outcome misclassification.\u0026nbsp;Aspirin exposure was also likely under-ascertained because over-the-counter use is not represented in pharmacy claims and is inconsistently recorded in EHR medication lists. These outcome and exposure measurement limitations would both be expected to attenuate associations toward the null. Although we addressed measured confounding through propensity score matching, residual confounding remains possible. Finally, organoids lack systemic vascular and immune context, precluding assessment of aspirin’s antithrombotic and peripheral anti-inflammatory effects that are likely relevant \u003cem\u003ein vivo\u003c/em\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe workflow described here provides a scalable strategy for moving from computationally prioritized repurposing signals to clinical evaluation in longitudinal real-world data, particularly relevant for diseases in which conventional drug development is slow, costly, or poorly aligned with disease stage. Given the flexibility of this workflow, future updates could incorporate additional identification or validation steps, such as AI-facilitated screening\u003csup\u003e10\u003c/sup\u003e. In AD, our findings support further study of aspirin with earlier initiation rather than broad late-life use. Future observational studies incorporating deeper phenotyping, including neuroimaging, cerebrospinal fluid biomarkers, polygenic risk scores, and cardiovascular and metabolic risk measures, may help define the subgroups most likely to benefit. Overall, this approach demonstrates a scalable next step for advancing drug repurposing toward eventual clinical implementation.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis study was conducted with approval from the VUMC Institutional Review Board and the NIH \u003cem\u003eAll of Us\u003c/em\u003e Research Program. All EHR data from VUMC and \u003cem\u003eAll of Us\u003c/em\u003e are de-identified; use of these data is considered non-human subjects research.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConstruction of microglia transcriptome prediction models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe downloaded full nominal eQTL summary statistics for 255 primary human microglia samples across four brain regions (medial frontal gyrus, superior temporal gyrus, subventricular zone, thalamus) from 100 human subjects from the Microglia Genomic Atlas (MiGA)\u003csup\u003e17\u003c/sup\u003e. Details on genotyping, RNA-seq generation and processing, and eQTL mapping can be found in the MiGA flagship paper\u003csup\u003e29\u003c/sup\u003e. Using the multivariate adaptive shrinkage approach to eQTL analysis introduced by Urbut \u003cem\u003eet al\u003c/em\u003e.\u003csup\u003e30\u003c/sup\u003e, we trained region-specific Multivariate Adaptive Shrinkage in R (MASHR) transcriptome prediction models. For each brain region, we selected the top five eQTLs with the lowest local false sign rate (a measure analogous to false discovery rate)\u003csup\u003e30\u003c/sup\u003e per gene and included these eQTLs in the final models. In cases where multiple eQTLs were tied for lowest local false sign rate, we retained all eQTLs. An initial top-eQTL-per-gene specification was evaluated but not used in downstream analyses due to limited overlap with the AD GWAS variant set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImputation of transcriptomic signatures for Alzheimer\u0026rsquo;s disease\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe applied two statistical methods, S-PrediXcan\u003csup\u003e11\u003c/sup\u003e and S-MultiXcan\u003csup\u003e12\u003c/sup\u003e, to publicly available Alzheimer\u0026rsquo;s GWAS summary statistics to compute transcriptomic signatures for AD. S-PrediXcan and S-MultiXcan predict gene expression for a trait or disease of interest using prediction models trained on reference transcriptome datasets. S-PrediXcan computes single-tissue gene-level association results from GWAS summary statistics. S-MultiXcan then aggregates the single-tissue S-PrediXcan results across multiple tissues, thereby increasing the statistical power to detect associations. We used GWAS summary statistics for a total of 762,917 individuals (86,531 AD cases and 676,386 controls, excluding participants from 23andMe, which were not available due to data access restrictions)\u003csup\u003e14\u003c/sup\u003e to generate all AD transcriptomic signatures used in this study. To improve GWAS-QTL integration, we harmonized the AD GWAS summary statistics to GTEx v8 variants and imputed summary statistics for missing variants\u003csup\u003e31\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eWe first used S-PrediXcan to impute single-tissue gene expression levels in 49 available GTEx tissues, using pre-developed MASHR expression prediction models\u003csup\u003e11,32,33\u003c/sup\u003e. We then ran S-MultiXcan on the S-PrediXcan results to predict gene expression across multiple tissues. Although AD predominantly affects the brain, studies have suggested that peripheral tissues such as the skin and vascular tissues can also capture genetic effects on gene expression in AD and may even represent potential pathogenic tissues in AD\u003csup\u003e15\u003c/sup\u003e. Thus, we conducted three separate S-MultiXcan analyses to predict AD-associated changes in gene expression combining brain and non-brain tissues, including: (1) a brain-specific analysis combining predictions from the 13 GTEx brain tissues, (2) an AD-relevant tissue analysis combining the 13 brain tissues with four other tissues previously related to AD (whole blood, spleen, and two skin tissues)\u003csup\u003e15,16\u003c/sup\u003e, and (3) a non-specific analysis combining predictions from all 49 GTEx tissues. We then constructed a separate AD virtual transcriptomic signature for each of the brain-restricted, AD-relevant, and all-tissue S-MultiXcan analyses using the AD risk genes with \u0026ge;2/3 tissue concordance (i.e., showing the same directionality of gene expression changes in at least two-thirds of the tissues with non-NA results). We required this tissue-wide consensus because several genes with statistically significant S-PrediXcan results exhibited opposing effects across tissues, producing near-zero effect estimates when averaged in S-MultiXcan analysis (e.g., \u003cem\u003eAPOE\u003c/em\u003e in the all-tissue analysis, Supplementary Fig. 7). The tissue-specific S-PrediXcan results for the AD risk genes, their consensus direction of effect, and mean \u003cem\u003eZ\u003c/em\u003e score computed across the tissues with the consensus effect direction are available in Supplementary Data File 1, sheets 1-3. We defined AD risk genes using Bonferroni correction (\u003cem\u003eP\u003c/em\u003e\u0026lt;0.05/number of S-MultiXcan gene associations). We used the mean \u003cem\u003eZ\u003c/em\u003e score among the concordant tissues to classify the risk genes as positively associated with AD risk (mean \u003cem\u003eZ\u003c/em\u003e score\u0026gt;0) or inversely associated with AD risk (mean \u003cem\u003eZ\u003c/em\u003e score\u0026lt;0).\u003c/p\u003e\n\u003cp\u003eWe also used S-PrediXcan and S-MultiXcan to calculate a microglia-specific AD gene expression signature using the microglia MASHR transcriptome prediction models we trained with MiGA data. We then applied S-PrediXcan to predict microglial gene expression in the medial frontal gyrus, superior temporal gyrus, subventricular zone, and thalamus. We initially used S-MultiXcan to construct a single AD gene expression signature integrating the S-PrediXcan results for microglia in the four different brain regions; however, the high correlations between the S-PrediXcan associations across the brain regions prevented computation of S-MultiXcan association statistics for \u0026gt;80% (7,852/9,687) of the genes tested. We thus used a generalized Berk-Jones (GBJ) test to combine the S-PrediXcan \u003cem\u003eZ\u003c/em\u003e scores from the four brain regions\u003csup\u003e34\u003c/sup\u003e. Again, AD risk genes were defined using a Bonferroni-corrected significance threshold (GBJ \u003cem\u003eP\u003c/em\u003e\u0026lt;0.05/18,222 genes tested). As GBJ does not provide directional effects (GBJ statistic is always positive), we constructed the microglial signature including only the AD risk genes with concordant effect directions across all regions with available results. NA values did not influence the concordance assessment. To capture inverse associations with AD risk, we negated the GBJ statistic for genes with consistently negative S-PrediXcan \u003cem\u003eZ\u003c/em\u003e scores across the brain regions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIdentification of drug repurposing candidates\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe queried the Connectivity Map (CMap) to identify drugs with perturbation profiles opposing TWAS-derived AD signatures. CMap contains over 1.5 million gene expression profiles (including 978 directly measured landmark genes) capturing the effects of over 8,000 small-molecule compounds and genetic reagents across multiple human cell lines\u003csup\u003e35\u003c/sup\u003e. Each AD signature was represented as two gene sets split by TWAS direction (\u003cem\u003eZ\u003c/em\u003e\u0026gt;0 \u0026ldquo;up\u0026rdquo;; \u003cem\u003eZ\u003c/em\u003e\u0026lt;0 \u0026ldquo;down\u0026rdquo;) and submitted to the CMap CLUE Query tool, which returns a connectivity score (\u0026tau;) quantifying similarity between the queried signature and each perturbagen signature. Small-molecule compounds with \u0026tau;\u0026lt;0 were considered potential repurposing candidates, consistent with prior CMap-based studies\u003csup\u003e6,19\u003c/sup\u003e. For downstream evaluation, we prioritized compounds with negative connectivity across multiple AD signatures, feasible exposure ascertainment in real-world data, and sufficient expected numbers of exposed individuals for longitudinal analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical evaluation using real-world data\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the prioritized aspirin signal in longitudinal clinical data, we investigated the association between aspirin exposure and AD using EHR and claims data. In the two EHR databases, we performed a retrospective cohort study evaluating the risk of new AD diagnosis after age 65 in individuals previously exposed to aspirin compared to individuals without documented exposure. In the claims database, we performed a case-control study comparing the odds of prior aspirin use in AD cases and controls.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData sources\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe performed clinical validation studies using diagnosis and medication data from three independent sources: (1) VUMC\u0026rsquo;s Synthetic Derivative (SD) database, (2) the NIH \u003cem\u003eAll of Us\u003c/em\u003e Research Program database, and (3) the MarketScan Commercial Claims and Encounters (CCAE) and Medicare Supplemental (MDCR) databases. The VUMC SD contains decades of longitudinal clinical data, including diagnosis and procedure codes, laboratory test results, and medications, extracted from de-identified EHRs for over 4 million unique patients\u003csup\u003e36\u003c/sup\u003e. The SD is linked to VUMC\u0026rsquo;s biobank, BioVU, which contained over 300,000 unique DNA samples as of January 2023, allowing for the integration of EHR and genetic data. At the time of the study, the \u003cem\u003eAll of Us\u003c/em\u003e Research Program database contained data for over 633,540 participants, with genomic sequencing data for over 414,840 participants\u003csup\u003e37\u003c/sup\u003e. The MarketScan CCAE and MDCR databases collectively contain detailed insurance claims data, including diagnosis codes and pharmacy claims, for over 200 million unique patients\u003csup\u003e38\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eCohort formation and outcome assessment: VUMC and All of Us\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo capture endpoints relevant to AD and ensure adequate EHR follow-up time was available for all patients in the study population, we restricted our analysis to patients over 65 years of age with at least one visit at \u0026ge;75 years. We excluded individuals with AD diagnosed at \u0026le;65 years, as well as individuals with diagnoses of non-AD dementia (occurring at any time).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInformation on patient medications is available as structured data in the VUMC and \u003cem\u003eAll of Us\u003c/em\u003e EHR databases (within a DRUG_EXPOSURE table, in accordance with the Observational Medical Outcomes Partnership Common Data Model\u003csup\u003e39\u003c/sup\u003e). Drug exposures documented in this table may originate from various sources in the EHR, including clinical notes (e.g., medication lists appearing in history and physical [H\u0026amp;P] reports, progress notes, and discharge summaries), inpatient and outpatient medication orders (e.g., prescriptions), and historical medication records.\u003c/p\u003e\n\u003cp\u003eWe identified aspirin-exposed patients by mapping patient medications to their ingredients using the RxNorm standardized terminology for clinical drugs and filtering for all medications containing aspirin (RxCUI = 1191). We used aspirin-exposed patients with at least one year of documented aspirin use prior to age 65 to form the aspirin-exposed group; patients without any documented aspirin use were considered unexposed controls. Patients with first aspirin use occurring after age 65 were excluded. We then identified two groups within the aspirin-exposed cohort for additional comparisons based on the two most common doses at VUMC: high-dose (\u0026ge;325mg/day) and low-dose (\u0026le;81mg/day) aspirin users, excluding individuals exposed to intermediate aspirin doses and individuals switching between low-dose and high-dose aspirin use.\u003c/p\u003e\n\u003cp\u003eAspirin is used in the treatment of multiple medical conditions, including in doses of 81 and 325 mg for primary and secondary prevention of atherosclerotic cardiovascular disease\u003csup\u003e40\u003c/sup\u003e, and historically, in much higher doses for pain relief in inflammatory diseases such as rheumatoid arthritis\u003csup\u003e41\u003c/sup\u003e, which may also impact AD risk. To minimize the effects of confounding by indication and other sources of confounding, we matched patients in the aspirin-exposed cohort to control patients with no recorded aspirin use in a 1:2 ratio, using a propensity score based on sex, race, EHR time after age 65 (i.e., length of time between age 65 and last EHR record), and presence of clinical indications for aspirin use (cardiovascular disease, cerebrovascular disease, and rheumatoid arthritis) at baseline\u0026nbsp;(defined as the time zero of age 65 or, if no data available at age 65, the age of the first EHR visit after 65). We used the MatchIt R\u003csup\u003e42\u003c/sup\u003e package to perform nearest-neighbor propensity score matching. The diagnoses used to define the comorbidities for matching are provided in Supplementary Data File 1, sheets 14-16.\u003c/p\u003e\n\u003cp\u003eWe defined AD cases using a requirement of at least one AD diagnosis code (ICD-9-CM 331.0; ICD-10-CM G30.1, G30.8, G30.9). We excluded patients with a first recorded AD code before age 65 and those with codes for non-AD dementias (Supplementary Data File 1, sheet 17).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStatistical analysis: VUMC and All of Us\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used Cox proportional hazards regression models to investigate incident AD risk after age 65 in aspirin-exposed and unexposed individuals. Age 65 served as time zero; individuals were followed until first recorded AD diagnosis or otherwise right censored at last recorded EHR observation. We first compared AD risk between the aspirin-exposed cohort and the propensity score-matched unexposed cohort. We then performed subgroup analyses based on aspirin dose (high-dose versus low-dose versus no aspirin), documented aspirin exposure rate, and \u003cem\u003eAPOE\u003c/em\u003e \u0026epsilon;4 genotype.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe used the metafor R package for meta-analysis of hazard ratios\u003csup\u003e43\u003c/sup\u003e. Heterogeneity was assessed using Cochran\u0026rsquo;s Q and I\u003csup\u003e2\u003c/sup\u003e. Based on these metrics, all meta-analyses were conducted under a fixed-effects model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDocumented aspirin exposure rate\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe inconsistent recording of medication end dates, dosing frequency (e.g., daily versus as needed), and therapy duration within the EHR, along with multiple sources of medication documentation in patient charts (including medication lists in clinical notes as well as prescriptions), hindered precise quantification of total aspirin exposure using EHR data. To address this limitation, we developed a proxy measure: the documented aspirin exposure rate, defined as the total number of unique aspirin records divided by the time (in years) between the first and last recorded aspirin exposures. This measure was intended to capture the frequency and duration of aspirin use documented in the EHR, with a higher rate reflecting more consistent and sustained aspirin exposure. We calculated the documented aspirin exposure rate for all individuals in the aspirin-exposed cohort and then classified aspirin users into high- and low-exposure groups based on the median documented exposure rate of\u0026nbsp;5 aspirin records per year. To ensure balanced comparisons, we matched individuals with high documented aspirin exposure rates to those with low rates in a 1:1 ratio using propensity score matching. The variables used in matching were\u0026nbsp;sex, race, baseline comorbidities (cardiovascular disease, cerebrovascular disease, rheumatoid arthritis), EHR time after age 65, total number of EHR visits, and aspirin duration. The final matched cohort comprised 3,690 individuals (\u003cem\u003eN\u0026nbsp;\u003c/em\u003e= 1,845 per group). A Cox proportional hazards regression model was used to investigate the risk of AD after age 65 in the high-exposure group relative to the low-exposure group.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAPOE genotyping\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAPOE\u003c/em\u003e genotype was determined using the combination of alleles at SNPs rs429358 and rs7412. The \u003cem\u003eAPOE\u003c/em\u003e \u0026epsilon;4 variant was defined as the presence of a C allele at both SNPs. Genetic data was available for 1,856 patients in the VUMC cohort (including 41 APOE \u0026epsilon;4 homozygotes and 430 heterozygotes) and 1,450 patients in the \u003cem\u003eAll of Us\u0026nbsp;\u003c/em\u003ecohort (with 23 APOE \u0026epsilon;4 homozygotes and 271 heterozygotes). Information on \u003cem\u003eAPOE\u003c/em\u003e genotype was not available in the MarketScan dataset.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMarketScan validation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the shorter observation time in the MarketScan Research Databases, which prevented us from capturing AD-relevant timepoints in patients exposed to aspirin before age 65, we performed a case-control study to investigate the association between aspirin use and AD. We first identified AD cases using ICD-9-CM code 331.0 and ICD-10-CM codes G30.1, G30.8, and G30.9. We matched AD cases to comparable controls in a 1:2 ratio based on propensity score, using sex, comorbidities (cardiovascular disease, cerebrovascular disease, and rheumatoid arthritis, diagnosed at any age), and claims follow-up time (difference in years between first and last claims records) as covariates. We did not match on race as this is not reported in MarketScan. Aspirin prescriptions were identified using National Drug Codes. We then calculated the odds ratio for aspirin exposure among the AD cases compared to their matched controls.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHuman iPSC culture and cortical organoid generation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIsogenic control wild-type (WT) and heterozygous APP mutant (KM670/671NL) iPSCs were maintained in mTeSR1 medium and passaged every 6-7 days. Human cortical organoids were generated following a previously established differentiation protocol, with modifications\u003csup\u003e44,45\u003c/sup\u003e. Briefly, iPSC colonies were enzymatically detached using 1 mg/mL collagenase IV for 1 hour. The iPSC colonies were collected and cultured as embryoid bodies (EBs) in DMEM/F12 medium (Invitrogen) supplemented with 20% Knockout Serum Replacement (KSR, Invitrogen), 1\u0026times; GlutaMAX (Invitrogen), 1\u0026times;MEM Non-Essential Amino Acids (Invitrogen), 0.1 mM beta-mercaptoethanol (Invitrogen), 2 \u0026micro;M Dorsomorphin (PeproTech), and 2 \u0026micro;M A-83 (PeproTech). EBs were maintained in 10 cm low-attachment dishes for 5 days, allowing uniform spheroid formation. From day 6 to day 16, EBs were changed into a neural medium (NM) comprising Neurobasal medium (Invitrogen), 1\u0026times; B-27 supplement (minus vitamin A, Invitrogen), 1\u0026times; GlutaMAX (Invitrogen), and 100 U/mL penicillin-streptomycin (Invitrogen), further supplemented with 20 ng/mL bFGF (Pepro Tech) and 20 ng/mL EGF (Pepro Tech). The medium was replaced daily to support robust neuroectodermal patterning. Between day 17 and day 24, cultures were maintained in the same medium with changes every other day. On day 25, the neural medium was supplemented with 20 ng/mL BDNF, with medium renewal continued every other day. By day 43, the medium was shifted to differentiation medium with growth factor-free neural medium, consisting of Neurobasal, 1\u0026times; B-27 supplement (minus vitamin A), 1\u0026times; GlutaMAX, and 100 U/mL penicillin-streptomycin, refreshed every four days. From day 70 onward, organoids were cultured in NM supplemented with 1\u0026times; B-27 containing vitamin A, with medium changes every three days, promoting long-term maturation. At day 90, mature human cortical organoids were treated with acetylsalicylic acid (aspirin; Sigma A5376; 0.5 mM) or vehicle (PBS) for one week.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRNA extraction and RNA-seq\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHuman cortical organoids were homogenized in TRIzol Reagent (Invitrogen, 15596018), and total RNA was extracted using the Direct-zol\u003csup\u003eTM\u003c/sup\u003e RNA Miniprep Kit (Zymo, R2052) according to the manufacturer\u0026rsquo;s instructions. Total RNA was quantified using the Qubit 2.0 Fluorometer (ThermoFisher Scientific, Waltham, MA, USA) and assessed for integrity with the 4200 TapeStation (Agilent Technologies, Palo Alto, CA, USA). Strand-specific libraries were prepared using the NEBNext Ultra II Directional RNA Library Prep Kit for Illumina (NEB, Ipswich, MA, USA), following the manufacturer\u0026rsquo;s instructions. RNA was fragmented at 94 \u0026deg;C for 8 minutes, and first- and second-strand cDNA synthesis was performed, with dUTP incorporated during second-strand synthesis to maintain strand specificity. After 3\u0026rsquo; end adenylation, adapter ligation, and limited-cycle PCR amplification, libraries were validated using the Agilent TapeStation and quantified by Qubit 2.0 (ThermoFisher Scientific) and qPCR (KAPA Biosystems, Wilmington, MA, USA). Libraries were multiplexed, clustered onto a flowcell, and sequenced on the Illumina NovaSeq 6000 system using a 2 \u0026times; 150 bp paired-end configuration, according to the manufacturer\u0026rsquo;s protocol. Image analysis and base calling were performed using the NovaSeq Control Software (Illumina, San Diego, CA, USA), and raw BCL files were converted to FASTQ format and demultiplexed with bcl2fastq v2.20 (Illumina), allowing one mismatch for index recognition.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRNA-seq analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRNA-seq data were processed following a previously described workflow. Raw reads were quality-checked to confirm that library preparation and sequencing met requirements for downstream analyses. Adapters were removed with Trimmomatic\u003csup\u003e46\u003c/sup\u003e. Cleaned reads were aligned to the human hg38 reference genome using HISAT2\u003csup\u003e47\u003c/sup\u003e, and read counts were generated with featureCounts\u003csup\u003e48\u003c/sup\u003e. Differential expression was assessed with DESeq2\u003csup\u003e49\u003c/sup\u003e, controlling the FDR at 0.05 with lfcThreshold = log2(1.1). We analyzed three contrasts: (1) a baseline AD signature (APP vehicle vs WT vehicle), (2) aspirin in a non-AD background (WT aspirin vs WT vehicle), and (3) aspirin in an AD background (APP aspirin vs APP vehicle).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFunctional enrichment analyses were performed in R using clusterProfiler\u003csup\u003e50\u003c/sup\u003e and fgsea\u003csup\u003e51\u003c/sup\u003e. Over-representation analysis (ORA) was conducted on differentially expressed genes (|log2FC| \u0026ge;log2(1.1), FDR\u0026lt;0.05) for each contrast, with enrichment significance evaluated by hypergeometric test using all genes tested in the differential expression analysis as background. GSEA was performed on preranked gene lists for each contrast using GO, KEGG, and Reactome gene set collections, with genes ranked by Wald statistic (log2 fold change/lfcSE). For targeted neuronal analyses, human MSigDB gene sets were curated by keyword filtering of pathway names (\u0026ldquo;synapse\u0026rdquo;, \u0026ldquo;axon\u0026rdquo;, \u0026ldquo;neurotransmitter\u0026rdquo;); genes were again ranked by Wald statistic. For all enrichment analyses, multiple testing correction was performed using the Benjamini-Hochberg method, and significance was defined as FDR\u0026lt;0.05.\u003c/p\u003e\n\u003cp\u003eRRHO was performed using RRHO2\u003csup\u003e52\u003c/sup\u003e to compare expression patterns between contrasts. Genes in each contrast were ranked by signed significance [-log\u003csub\u003e10\u003c/sub\u003e(\u003cem\u003eP\u003c/em\u003e) \u0026times; sign(log2 fold change)], and the significance of overlap between two ranked lists was assessed using hypergeometric tests across rank thresholds (step size=100). ORA was performed on discordant overlap genes (i.e., genes with significant opposite direction changes between contrasts), as described above.\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted in accordance with the Declaration of Helsinki. Analyses of Vanderbilt University Medical Center de-identified electronic health record and genetic data were determined to be non-human subjects research by the Vanderbilt University Medical Center Institutional Review Board under IRB #211489. Analyses using NIH \u003cem\u003eAll of Us\u003c/em\u003e Research Program data and MarketScan Research Databases involved de-identified data and complied with applicable access and data use requirements. \u003cem\u003eAll of Us\u003c/em\u003e participants provided informed consent for participation in the \u003cem\u003eAll of Us\u003c/em\u003e Research Program. No additional informed consent to participate in the study was required because analyses did not involve participant recruitment, participant contact, or access to identifiable private information.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Institute on Aging under grants\u0026nbsp;F30AG080885 (MEG), R01AG069900 (BL, WQW, ZW),\u0026nbsp;and\u0026nbsp;R01AG084550 (WQW, QF);\u0026nbsp;by the National Institute of General Medical Sciences under grant\u0026nbsp;T32GM007347 (MEG); and by the National Heart, Lung, and Blood Institute under grants R01HL163854 (QF), R01HL133786 (WQW),\u0026nbsp;and R01HL171809 (WQW, QF). The primary datasets were obtained from Vanderbilt University Medical Center\u0026rsquo;s BioVU, which is supported by institutional funding, 1S10RR025141-01, and CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. Additional support was provided by the National Institutes of Health through grants P50GM115305 and U19HL065962. We acknowledge the expert technical support of the VANTAGE and VANGARD core facilities, supported in part by the Vanderbilt-Ingram Cancer Center (P30CA068485) and Vanderbilt Vision Center (P30EY08126). Validation datasets were obtained from the \u003cem\u003eAll of Us\u003c/em\u003e Research Program. We thank \u003cem\u003eAll of Us\u003c/em\u003e participants, whose contributions made this research possible.\u0026nbsp;The \u003cem\u003eAll of Us\u003c/em\u003e Research Program is supported by the National Institutes of Health, Office of the Director: Regional Medical Centers: 1 OT2 OD026549; 1 OT2 OD026554; 1 OT2 OD026557; 1 OT2 OD026556; 1 OT2 OD026550; 1 OT2 OD026552; 1 OT2 OD026553; 1 OT2 OD026548; 1 OT2 OD026551; 1 OT2 OD026555; IAA #: AOD 16037; Federally Qualified Health Centers: HHSN 263201600085U; Data and Research Center: 5 U2C OD023196; Biobank: 1 U24 OD023121; The Participant Center: U24 OD023176; Participant Technology Systems Center: 1 U24 OD023163; Communications and Engagement: 3 OT2 OD023205; 3 OT2 OD023206; and Community Partners: 1 OT2 OD025277; 3 OT2 OD025315; 1 OT2 OD025337; 1 OT2 OD025276.\u0026nbsp;The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMEG, WQW, BL, ZW, and QF conceived the study. MEG, WQW, BL, ZW, QF, RC, YZ, XZ, and ALD contributed to the study design. MEG performed the primary analyses. YZ performed the organoid experiments, and RC contributed to downstream analysis of the organoid data. MEG drafted the initial manuscript. All authors contributed to data interpretation, critically reviewed and revised the manuscript, and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors declare no financial or non-financial competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe microglial MASHR models trained and used in this study are available on Zenodo: https://doi.org/10.5281/zenodo.18156902.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data are available in the main text or the supplementary materials. The AD GWAS summary statistics used in this study are publicly available at https://cncr.nl/research/summary_statistics/. The microglia eQTL summary statistics from the Microglia Genomic Atlas used in this study can be downloaded from the NIAGADS Data Sharing Service using accession number NG00105.v3. The MASHR GTEx v8 transcriptome prediction models can be downloaded from PredictDB (https://predictdb.org/categories/downloads/). Access to VUMC\u0026rsquo;s EHR database requires institutional approval and compliance with a data use agreement. Data from the \u003cem\u003eAll of Us\u003c/em\u003e Research Program can be accessed through the Researcher Workbench (https://workbench.researchallofus.org). The MarketScan claims data used in this study can be requested from Merative\u0026reg;.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGrabowska, M. E., Huang, A., Wen, Z., Li, B. \u0026amp; Wei, W.-Q. Drug repurposing for Alzheimer\u0026rsquo;s disease from 2012\u0026ndash;2022\u0026mdash;a 10-year literature review. \u003cem\u003eFrontiers in Pharmacology\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, (2023).\u003c/li\u003e\n \u003cli\u003eMusa, A. \u003cem\u003eet al.\u003c/em\u003e A review of connectivity map and computational approaches in pharmacogenomics. \u003cem\u003eBrief Bioinform\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 506\u0026ndash;523 (2017).\u003c/li\u003e\n \u003cli\u003eLi, B. \u0026amp; Ritchie, M. D. From GWAS to Gene: Transcriptome-Wide Association Studies and Other Methods to Functionally Understand GWAS Discoveries. \u003cem\u003eFront Genet\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 713230 (2021).\u003c/li\u003e\n \u003cli\u003eMinikel, E. V., Painter, J. L., Dong, C. C. \u0026amp; Nelson, M. R. Refining the impact of genetic evidence on clinical success. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e629\u003c/strong\u003e, 624\u0026ndash;629 (2024).\u003c/li\u003e\n \u003cli\u003eKho, P. F. \u003cem\u003eet al.\u003c/em\u003e Multi-tissue transcriptome-wide association study identifies eight candidate genes and tissue-specific gene expression underlying endometrial cancer susceptibility. \u003cem\u003eCommun Biol\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 1211 (2021).\u003c/li\u003e\n \u003cli\u003eWu, P. \u003cem\u003eet al.\u003c/em\u003e Integrating gene expression and clinical data to identify drug repurposing candidates for hyperlipidemia and hypertension. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 46 (2022).\u003c/li\u003e\n \u003cli\u003eZong, N. \u003cem\u003eet al.\u003c/em\u003e Computational drug repurposing based on electronic health records: a scoping review. \u003cem\u003eNPJ Digit Med\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 77 (2022).\u003c/li\u003e\n \u003cli\u003eTan, G. S. Q., Sloan, E. K., Lambert, P., Kirkpatrick, C. M. J. \u0026amp; Ilom\u0026auml;ki, J. Drug repurposing using real-world data. \u003cem\u003eDrug Discov Today\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 103422 (2023).\u003c/li\u003e\n \u003cli\u003eZang, C. \u003cem\u003eet al.\u003c/em\u003e High-throughput target trial emulation for Alzheimer\u0026rsquo;s disease drug repurposing with real-world data. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 8180 (2023).\u003c/li\u003e\n \u003cli\u003eYan, C. \u003cem\u003eet al.\u003c/em\u003e Leveraging generative AI to prioritize drug repurposing candidates for Alzheimer\u0026rsquo;s disease with real-world clinical validation. \u003cem\u003enpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 1\u0026ndash;6 (2024).\u003c/li\u003e\n \u003cli\u003eBarbeira, A. N. \u003cem\u003eet al.\u003c/em\u003e Exploring the phenotypic consequences of tissue specific gene expression variation inferred from GWAS summary statistics. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 1825 (2018).\u003c/li\u003e\n \u003cli\u003eBarbeira, A. N. \u003cem\u003eet al.\u003c/em\u003e Integrating predicted transcriptome from multiple tissues improves association detection. \u003cem\u003ePLOS Genetics\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, e1007889 (2019).\u003c/li\u003e\n \u003cli\u003eLonsdale, J. \u003cem\u003eet al.\u003c/em\u003e The Genotype-Tissue Expression (GTEx) project. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 580\u0026ndash;585 (2013).\u003c/li\u003e\n \u003cli\u003eWightman, D. P. \u003cem\u003eet al.\u003c/em\u003e A genome-wide association study with 1,126,563 individuals identifies new risk loci for Alzheimer\u0026rsquo;s disease. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e, 1276\u0026ndash;1282 (2021).\u003c/li\u003e\n \u003cli\u003eGerring, Z. F., Lupton, M. K., Edey, D., Gamazon, E. R. \u0026amp; Derks, E. M. An analysis of genetically regulated gene expression across multiple tissues implicates novel gene candidates in Alzheimer\u0026rsquo;s disease. \u003cem\u003eAlzheimer\u0026rsquo;s Research \u0026amp; Therapy\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 43 (2020).\u003c/li\u003e\n \u003cli\u003eOngen, H. \u003cem\u003eet al.\u003c/em\u003e Estimating the causal tissues for complex traits and diseases. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, 1676\u0026ndash;1683 (2017).\u003c/li\u003e\n \u003cli\u003eNG00105 - MiGA \u0026ndash; Microglia Genomic Atlas. \u003cem\u003eDSS NIAGADS\u003c/em\u003e https://dss.niagads.org/datasets/ng00105/.\u003c/li\u003e\n \u003cli\u003eZhao, Y., Chen, X., Chen, J. \u0026amp; Qi, X. Decoding Connectivity Map-based drug repurposing for oncotherapy. \u003cem\u003eBriefings in Bioinformatics\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, bbad142 (2023).\u003c/li\u003e\n \u003cli\u003eTaubes, A. \u003cem\u003eet al.\u003c/em\u003e Experimental and real-world evidence supporting the computational repurposing of bumetanide for APOE4-related Alzheimer\u0026rsquo;s disease. \u003cem\u003eNat Aging\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 932\u0026ndash;947 (2021).\u003c/li\u003e\n \u003cli\u003eNarcisse, D. I. \u003cem\u003eet al.\u003c/em\u003e Comparative Effectiveness of Aspirin Dosing in Cardiovascular Disease and Diabetes Mellitus: A Subgroup Analysis of the ADAPTABLE Trial. \u003cem\u003eDiabetes Care\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, 81\u0026ndash;88 (2024).\u003c/li\u003e\n \u003cli\u003ePlaisier, S. B., Taschereau, R., Wong, J. A. \u0026amp; Graeber, T. G. Rank\u0026ndash;rank hypergeometric overlap: identification of statistically significant overlap between gene-expression signatures. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e38\u003c/strong\u003e, e169 (2010).\u003c/li\u003e\n \u003cli\u003eNguyen, T. N. M. \u003cem\u003eet al.\u003c/em\u003e Long-term low-dose acetylsalicylic use shows protective potential for the development of both vascular dementia and Alzheimer\u0026rsquo;s disease in patients with coronary heart disease but not in other individuals from the general population: results from two large cohort studies. \u003cem\u003eAlzheimer\u0026rsquo;s Research \u0026amp; Therapy\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 75 (2022).\u003c/li\u003e\n \u003cli\u003eRyan, J. \u003cem\u003eet al.\u003c/em\u003e Randomized placebo-controlled trial of the effects of aspirin on dementia and cognitive decline. \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e95\u003c/strong\u003e, e320\u0026ndash;e331 (2020).\u003c/li\u003e\n \u003cli\u003eAD2000 Collaborative Group \u003cem\u003eet al.\u003c/em\u003e Aspirin in Alzheimer\u0026rsquo;s disease (AD2000): a randomised open-label trial. \u003cem\u003eLancet Neurol\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 41\u0026ndash;49 (2008).\u003c/li\u003e\n \u003cli\u003eParish, S. \u003cem\u003eet al.\u003c/em\u003e Effects of aspirin on dementia and cognitive function in diabetic patients: the ASCEND trial. \u003cem\u003eEur Heart J\u003c/em\u003e \u003cstrong\u003e43\u003c/strong\u003e, 2010\u0026ndash;2019 (2022).\u003c/li\u003e\n \u003cli\u003eKang, J. H., Cook, N., Manson, J., Buring, J. E. \u0026amp; Grodstein, F. Low dose aspirin and cognitive function in the women\u0026rsquo;s health study cognitive cohort. \u003cem\u003eBMJ\u003c/em\u003e \u003cstrong\u003e334\u003c/strong\u003e, 987 (2007).\u003c/li\u003e\n \u003cli\u003eBateman, R. J. \u003cem\u003eet al.\u003c/em\u003e Clinical and Biomarker Changes in Dominantly Inherited Alzheimer\u0026rsquo;s Disease. \u003cem\u003eNew England Journal of Medicine\u003c/em\u003e \u003cstrong\u003e367\u003c/strong\u003e, 795\u0026ndash;804 (2012).\u003c/li\u003e\n \u003cli\u003eLi, Y. \u003cem\u003eet al.\u003c/em\u003e Timing of Biomarker Changes in Sporadic Alzheimer\u0026rsquo;s Disease in Estimated Years from Symptom Onset. \u003cem\u003eAnn Neurol\u003c/em\u003e \u003cstrong\u003e95\u003c/strong\u003e, 951\u0026ndash;965 (2024).\u003c/li\u003e\n \u003cli\u003eLopes, K. de P. \u003cem\u003eet al.\u003c/em\u003e Genetic analysis of the human microglial transcriptome across brain regions, aging and disease pathologies. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e54\u003c/strong\u003e, 4\u0026ndash;17 (2022).\u003c/li\u003e\n \u003cli\u003eUrbut, S. M., Wang, G., Carbonetto, P. \u0026amp; Stephens, M. Flexible statistical methods for estimating and testing effects in genomic studies with multiple conditions. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, 187\u0026ndash;195 (2019).\u003c/li\u003e\n \u003cli\u003eBest practices for integrating GWAS and GTEX v8 transcriptome prediction models. \u003cem\u003eGitHub\u003c/em\u003e https://github.com/hakyimlab/MetaXcan/wiki/Best-practices-for-integrating-GWAS-and-GTEX-v8-transcriptome-prediction-models.\u003c/li\u003e\n \u003cli\u003eGamazon, E. R. \u003cem\u003eet al.\u003c/em\u003e A gene-based association method for mapping traits using reference transcriptome data. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, 1091\u0026ndash;1098 (2015).\u003c/li\u003e\n \u003cli\u003eBarbeira, A. N. \u003cem\u003eet al.\u003c/em\u003e Exploiting the GTEx resources to decipher the mechanisms at GWAS loci. \u003cem\u003eGenome Biology\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 49 (2021).\u003c/li\u003e\n \u003cli\u003eSun, R. \u0026amp; Lin, X. Genetic Variant Set-Based Tests Using the Generalized Berk-Jones Statistic with Application to a Genome-Wide Association Study of Breast Cancer. \u003cem\u003eJournal of the American Statistical Association\u003c/em\u003e \u003cstrong\u003e115\u003c/strong\u003e, 1079 (2019).\u003c/li\u003e\n \u003cli\u003eSubramanian, A. \u003cem\u003eet al.\u003c/em\u003e A Next Generation Connectivity Map: L1000 Platform and the First 1,000,000 Profiles. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e171\u003c/strong\u003e, 1437-1452.e17 (2017).\u003c/li\u003e\n \u003cli\u003eRoden, D. M. \u003cem\u003eet al.\u003c/em\u003e Development of a large-scale de-identified DNA biobank to enable personalized medicine. \u003cem\u003eClin Pharmacol Ther\u003c/em\u003e \u003cstrong\u003e84\u003c/strong\u003e, 362\u0026ndash;369 (2008).\u003c/li\u003e\n \u003cli\u003eAll of Us Research Program Investigators \u003cem\u003eet al.\u003c/em\u003e The \u0026lsquo;All of Us\u0026rsquo; Research Program. \u003cem\u003eN Engl J Med\u003c/em\u003e \u003cstrong\u003e381\u003c/strong\u003e, 668\u0026ndash;676 (2019).\u003c/li\u003e\n \u003cli\u003eMerative MarketScan Research Databases. https://www.merative.com/documents/merative-marketscan-research-databases.\u003c/li\u003e\n \u003cli\u003eOMOP Common Data Model. https://ohdsi.github.io/CommonDataModel/index.html.\u003c/li\u003e\n \u003cli\u003eMainous, A. G., Tanner, R. J., Shorr, R. I. \u0026amp; Limacher, M. C. Use of Aspirin for Primary and Secondary Cardiovascular Disease Prevention in the United States, 2011\u0026ndash;2012. \u003cem\u003eJournal of the American Heart Association\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, e000989 (2014).\u003c/li\u003e\n \u003cli\u003eSolomon, D. H. \u003cem\u003eet al.\u003c/em\u003e The potential benefits of aspirin for primary cardiovascular prevention in rheumatoid arthritis: a secondary analysis of the PRECISION Trial. \u003cem\u003eRheumatology (Oxford)\u003c/em\u003e \u003cstrong\u003e57\u003c/strong\u003e, 1364\u0026ndash;1369 (2018).\u003c/li\u003e\n \u003cli\u003eHo, D., Imai, K., King, G. \u0026amp; Stuart, E. A. MatchIt: Nonparametric Preprocessing for Parametric Causal Inference. \u003cem\u003eJournal of Statistical Software\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 1\u0026ndash;28 (2011).\u003c/li\u003e\n \u003cli\u003eViechtbauer, W. Conducting Meta-Analyses in R with the metafor Package. \u003cem\u003eJournal of Statistical Software\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, 1\u0026ndash;48 (2010).\u003c/li\u003e\n \u003cli\u003eKang, Y. \u003cem\u003eet al.\u003c/em\u003e A human forebrain organoid model of fragile X syndrome exhibits altered neurogenesis and highlights new treatment strategies. \u003cem\u003eNat Neurosci\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 1377\u0026ndash;1391 (2021).\u003c/li\u003e\n \u003cli\u003eKuehner, J. N. \u003cem\u003eet al.\u003c/em\u003e 5-hydroxymethylcytosine is dynamically regulated during forebrain organoid development and aberrantly altered in Alzheimer\u0026rsquo;s disease. \u003cem\u003eCell Rep\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 109042 (2021).\u003c/li\u003e\n \u003cli\u003eBolger, A. M., Lohse, M. \u0026amp; Usadel, B. Trimmomatic: a flexible trimmer for Illumina sequence data. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 2114\u0026ndash;2120 (2014).\u003c/li\u003e\n \u003cli\u003eKim, D., Paggi, J. M., Park, C., Bennett, C. \u0026amp; Salzberg, S. L. Graph-based genome alignment and genotyping with HISAT2 and HISAT-genotype. \u003cem\u003eNat Biotechnol\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 907\u0026ndash;915 (2019).\u003c/li\u003e\n \u003cli\u003eLiao, Y., Smyth, G. K. \u0026amp; Shi, W. featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 923\u0026ndash;930 (2014).\u003c/li\u003e\n \u003cli\u003eLove, M. I., Huber, W. \u0026amp; Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. \u003cem\u003eGenome Biol\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 550 (2014).\u003c/li\u003e\n \u003cli\u003eYu, G., Wang, L.-G., Han, Y. \u0026amp; He, Q.-Y. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. \u003cem\u003eOMICS\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 284\u0026ndash;287 (2012).\u003c/li\u003e\n \u003cli\u003eKorotkevich, G. \u003cem\u003eet al.\u003c/em\u003e Fast gene set enrichment analysis. 060012 Preprint at https://doi.org/10.1101/060012 (2021).\u003c/li\u003e\n \u003cli\u003eCahill, K. M., Huo, Z., Tseng, G. C., Logan, R. W. \u0026amp; Seney, M. L. Improved identification of concordant and discordant gene expression signatures using an updated rank-rank hypergeometric overlap approach. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 9588 (2018).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9518587/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9518587/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDrug repurposing offers a potential strategy to expand treatment options for conditions with limited therapies, but advancing repurposing candidates toward clinical implementation remains a challenge. Large-scale data, together with advanced genetic and epidemiological methods, may help address this gap. Here, we present an integrative digital medicine approach that combines genetically predicted transcriptomic signatures and perturbation screening for candidate identification with multi-cohort real-world validation for systematic evaluation of prioritized candidates. We applied this approach to Alzheimer\u0026rsquo;s disease (AD), a disease with substantial unmet clinical need and persistent difficulty in developing effective therapies. We constructed AD signatures from genetically predicted expression changes across bulk tissues and microglia, then queried Connectivity Map profiles to identify compounds predicted to oppose these signatures. Aspirin emerged as a reproducible candidate across multiple signatures and underwent further evaluation. We then examined its association with incident AD in longitudinal electronic health record data from Vanderbilt University Medical Center and the NIH \u003cem\u003eAll of Us\u003c/em\u003e Research Program, as well as national insurance claims data. Across independent cohorts, aspirin initiation before age 65 was consistently associated with lower risk of incident AD, with signals suggesting that cumulative exposure and \u003cem\u003eAPOE\u003c/em\u003e ε4 status may influence effect size. Transcriptomic analysis of human cortical organoids provided additional experimental support, showing that aspirin more strongly opposed AD-related neuronal pathway alterations in wild-type organoids than in an organoid model of AD. This integrative approach offers a scalable strategy for genetically informed drug repurposing that bridges candidate discovery and clinical evaluation.\u003c/p\u003e","manuscriptTitle":"Integrating genetically predicted transcriptomic signatures with longitudinal real-world data enables scalable drug repurposing for Alzheimer’s disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-15 15:31:19","doi":"10.21203/rs.3.rs-9518587/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"118787013116575439350177813903665002868","date":"2026-05-06T20:34:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"234155489385315912008879836947051702658","date":"2026-05-06T09:03:25+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"110027295319481950557292345104369911444","date":"2026-05-06T07:29:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"311245031897827219361773920163047547034","date":"2026-05-06T07:22:12+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-06T07:12:30+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-29T23:04:10+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-29T04:56:32+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2026-04-24T14:46:08+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"58350911-1ea3-4ccf-b783-15c17a6b5915","owner":[],"postedDate":"May 15th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"118787013116575439350177813903665002868","date":"2026-05-06T20:34:55+00:00","index":30,"fulltext":""},{"type":"reviewerAgreed","content":"234155489385315912008879836947051702658","date":"2026-05-06T09:03:25+00:00","index":29,"fulltext":""},{"type":"reviewerAgreed","content":"110027295319481950557292345104369911444","date":"2026-05-06T07:29:41+00:00","index":27,"fulltext":""},{"type":"reviewerAgreed","content":"311245031897827219361773920163047547034","date":"2026-05-06T07:22:12+00:00","index":26,"fulltext":""},{"type":"reviewersInvited","content":"15","date":"2026-05-06T07:12:30+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-29T23:04:10+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67765426,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":67765427,"name":"Health sciences/Neurology"},{"id":67765428,"name":"Biological sciences/Neuroscience"}],"tags":[],"updatedAt":"2026-05-15T15:31:20+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-15 15:31:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9518587","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9518587","identity":"rs-9518587","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}