Single‐microglia transcriptomic transition network‐based prediction and real‐world patient data validation identifies ketorolac as a repurposable drug for Alzheimer's disease

Abstract

INTRODUCTION

High microglial heterogeneities hinder the development of microglia‐targeted treatment for Alzheimer's disease (AD).

METHODS

We integrated 0.7 million single‐nuclei RNA‐sequencing transcriptomes from human brains using a variational autoencoder. We predicted AD‐relevant microglial subtype‐specific transition networks for disease‐associated microglia (DAM), tau microglia, and neuroinflammation‐like microglia (NIM). We prioritized drugs by specifically targeting microglia‐specific transition networks and validated drugs using two independent real‐world patient databases.

RESULTS

We identified putative AD molecular drivers (e.g., SYK, CTSB, and INPP5D) in transition networks of DAM and NIM. Via specifically targeting NIM, we identified that usage of ketorolac was associated with reduced AD incidence in both MarketScan (hazard ratio [HR] = 0.89) and INSIGHT (HR = 0.83) Clinical Research Network databases, mechanistically supported by ketorolac‐treated transcriptomic data from AD patient induced pluripotent stem cell–derived microglia.

DISCUSSION

This study offers insights into the pathobiology of AD‐relevant microglial subtypes and identifies ketorolac as a potential anti‐inflammatory treatment for AD.

Highlights

• An integrative analysis of ≈ 0.7 million single‐nuclei RNA‐sequencing transcriptomes from human brains identified Alzheimer's disease (AD)–relevant microglia subtypes.

• Network‐based analysis identified putative molecular drivers (e.g., SYK, CTSB, INPP5D) of transition networks between disease‐associated microglia (DAM) and neuroinflammation‐like microglia (NIM).

• Via network‐based prediction and population‐based validation, we identified that usage of ketorolac (a US Food and Drug Administration–approved anti‐inflammatory medicine) was associated with reduced AD incidence in two independent patient databases.

• Mechanistic observation showed that ketorolac treatment downregulated the Type‐I interferon signaling in patient induced pluripotent stem cell–derived microglia, mechanistically supporting its protective effects in real‐world patient databases.

1. BACKGROUND

Alzheimer's disease (AD) affects ≈ 6.7 million Americans aged ≥ 65 in 2023, ¹ and it is estimated to affect 13.8 million Americans and 152 million people worldwide by 2050 (2023 AD facts and figures). Lecanemab (leqembi) and donanemab (kisunla), monoclonal antibodies targeting aggregated amyloid, were recently approved as the first disease‐modifying medicine for early AD patients. ² , ³ However, anti‐amyloid drugs are unlikely to act after symptom onset and potential side effects and high cost may limit their use in broad AD populations. Without effective disease‐modifying treatments, this represents an unprecedented crisis of human suffering and financial cost. Compounding this issue, the attrition rate for AD clinical trials (2002–2012) is estimated at 99%. ⁴ The underlying pathophysiology of AD is poorly understood and AD biology is influenced by a complex, polygenic, and pleiotropic genetic architecture. ⁵ Beyond amyloid plaques and tau tangles, recent studies strongly implicate crucial roles for neuroinflammation and dysregulated microglia in the pathophysiology of AD and AD and related dementias (ADRD). ⁶ However, broad anti‐inflammatory therapies have not been clinically efficacious against AD. ⁶ There is a pressing need to better understand the heterogeneity of immune cells in AD, in particular the microglia, which could translate to identification of drug targets for emerging development of disease‐modifying therapies for AD/ADRD.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literatures using various traditional sources. Microglia have been implicated as a key aspect of the pathology of Alzheimer's disease (AD). However, high microglial heterogeneities, including disease‐associated microglia (DAM), tau microglia (tau‐pathology related), and neuroinflammation‐like microglia (NIM), hinder the development of microglia‐targeted treatment. The hypothesis of microglia‐targeted therapy—via specifically targeting the molecular drivers in AD‐relevant microglia subtypes—has emerged as a promising approach for development of effective AD treatment.

2. Interpretation: Via integrative analysis of ≈ 0.7 million single‐nuclei RNA‐sequencing transcriptomes from AD patient frozen brain samples using a variational autoencoder, we identified transition networks across DAM, tau microglia, and NIM. We identified a set of potential molecular drivers (e.g., SYK, CTSB, and INPP5D) in transition networksofDAM and NIM. Via network‐based drug repurposing prediction by specifically targeting NIM subpopulations and population‐based validation, we identified that usage of ketorolac (a US Food and Drug Administration–approved anti‐inflammatory medicine) was associated with reduced AD incidence in both MarketScan and INSIGHT Clinical Research Network databases.

3. Future directions: We demonstrate that single‐microglia transcriptomic transition network‐based prediction and real‐world patient data observation framework identify ketorolac as a potential treatment for AD, which can be broadly applied to AD‐related dementias and disease‐relevant cell types as well. Future experimental and clinical investigations in ethnically diverse cohorts are essential to establish a likely causal relationship of ketorolac in potential treatment of AD.

In the context of AD, microglia are implicated in the clearance of amyloid beta (Aβ) protein, ⁷ the regulation of synapses, ⁸ and the presence of neuroinflammation. ⁹ Microglia may play neuroprotective roles in the early stage of AD as they could normally help clear Aβ plaques from the brain. However, in the late stages of AD, microglial function becomes more diverse, with some microglia acquiring a pro‐inflammatory phenotype that could worsen the disease progression. ¹⁰ Recent studies using single‐cell RNA sequencing (scRNA‐seq) technologies have revealed the heterogeneity of microglia. ¹¹ One scRNA‐seq study based on the 5XFAD mouse model discovered one microglia transcriptional subcluster called disease‐associated microglia (DAM). ¹² It divided the formation of DAM into two different stages: TREM2 signaling‐independent stage 1 and dependent stage 2. Similarly, a positron emission tomography (PET) scan–based study found that the one microglia subtype annotated as activated microglia when colocalized with Aβ and tau could enhance tau spread, which might explain why co‐occurrence of amyloid plaques, tau tangles, and activated microglia could severely worsen disease progression. ¹³ The latest human single‐nuclei RNA sequencing (snRNA‐seq) study, from Prater et al., ¹⁴ used fluorescence‐activated nuclei sorting (FANS) for PU.1 as a marker of microglial enrichment to identify 10 unique microglial clusters associated with AD versus controls. This study further related microglia subclusters to specific functional pathways, such as motile microglia, canonical inflammatory microglia, and senescent‐like microglia. These studies emphasize the importance of identifying specific microglia transcriptional profiles in AD for the development of microglia‐targeted therapies.

Understanding microglial heterogeneity in AD is necessary for the development of microglia‐targeted therapies to promote neuroprotection and improve cognitive function or slow its decline. Integrative analyses of multiple human brain snRNA‐seq datasets will provide stronger statistical power to study microglia subtypes. Here, we reported integrative analyses of ≈ 0.7 million snRNA‐seq transcriptomes derived from AD patient frozen brain samples across different brain regions and we reidentified 0.18 million microglia. We used trajectory analysis to identify the emergence of different microglia subtypes at different pseudo‐times and to pinpoint putative molecular drivers associated with microglia subtype transition. In addition, we prioritized candidate drugs by specifically targeting microglial transition network modules using the Connectivity Map (CMap) database. ¹⁵ We identified that usage of ketorolac (a US Food and Drug Administration (FDA)–approved non‐steroidal anti‐inflammatory analgesic drug) was associated with reduced incidence of AD in two independent real‐world patient datasets (including INSIGHT Clinical Research Network [CRN; 15 million patients] and MarketScan insurance claims [172 million insured individuals]). In summary, this study demonstrates a proof of concept of the translation of high‐throughput single‐cell/nucleus omics findings into microglia‐targeted drug development for AD and other microglia‐related neurological diseases if broadly applied.

2. METHODS

2.1. Resources of snRNA‐seq data

The complete snRNA‐seq datasets used in this study (Table S1 in supporting information) are available from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) database under accession numbers GSE148822, ¹⁶ GSE157827, ¹⁷ and GSE147528. ¹⁸ One snRNA‐seq dataset (GSE148822) ¹⁶ contains 18 samples including 482,472 sequenced nuclei with 10 AD samples, 5 control without amyloid plaques (CTR) samples, and 3 control plus amyloid plaques (CTR+) samples. CTR+ samples were observed with Aβ pathology, which differentiated them from CTR samples. Each sample had sequenced nuclei from two brain regions: the occipital cortex and occipitotemporal cortex. Braak stage information is also provided for each sample. Another snRNA‐seq dataset (GSE157827) ¹⁷ collected 21 samples, 12 AD samples, and 9 normal control (NC) samples, from the prefrontal cortex region. Braak stage and apolipoprotein E (APOE) genotype information are also provided for each sample, 169,496 sequenced nuclei in total. The last snRNA‐seq dataset (GSE147528) ¹⁸ gathered 10 APOE ε3/ε3 samples with Braak stages 0, II, and VI. Each sample has nuclei sequenced from two brain regions: entorhinal cortex with 42,528 nuclei, and superior frontal gyrus with 63,608 nuclei.

2.2. Data preprocessing of snRNA‐seq datasets

For dataset GSE148822, we followed the snRNA‐seq data analysis procedures based on the original study. ¹⁶ We first removed the doublets via DoubletFinder ¹⁹ with 10% estimated doublet rate for each sample separately. Then we merged each doublet‐free sample into one using the “merge” function in Seurat (4.0.6). ²⁰ Nuclei with < 500 or > 2500 genes and mitochondrial content > 5% were removed. After data preprocessing, the remaining total nuclei count is 383,233, consistent with the original study ¹⁶ (369,615: microglia = 148,606; astrocyte = 128,764; endothelial = 26,957; central nervous system–associated macrophages = 17,979; lymphocytes = 12,675; fibroblasts = 22,238; and one unidentified cluster = 12,396). For dataset GSE157827, the raw total nuclei count downloaded from the GEO database was 179,392. Based on the original study, ¹⁶ nuclei with > 20,000 unique molecular identifiers (UMIs), < 200 genes, and > 20% mitochondrial content were removed. The total nuclei count after quality control was 169,506, which was almost identical (169,496) to the original study. ¹⁷ For dataset GSE147528, we directly used the processed data (https://kampmannlab.ucsf.edu/ad‐brain), including 42,737 and 64,257 nuclei for entorhinal cortex and superior frontal gyrus brain regions, respectively. We then merged all three datasets into one Seurat object using the “merge” function. We plotted quality controls (QCs) to merged data, finding the merged data presented inconsistent ranges for several QC measurements, such as “nCount_RNA,” and “nFeature_RNA” (Figures S1A,B in supporting information). Therefore, to remove as many outliers while simultaneously preserving as many nuclei after QC, we performed a second round of QC for the merged datasets with uniform criterions, that is, nCount_RNA < 8000 (Figure S1C), nFeature_RNA < 4000 (Figure S1D), mitochondrial content < 5% (Figure S1E), and ribosomal gene content < 5% (Figure S1F).

2.3. Integration of snRNA‐seq datasets

We applied a deep generative model for snRNA‐seq data integration, termed single‐cell variational inference (scVI), ²¹ built on a variational autoencoder framework. Variational inference is used to approximate the posterior distribution of the latent variables/cell embedding. The objective is to maximize the evidence lower bound (ELBO), which is a lower bound on the log‐likelihood of the observed data (gene expression in this study). To integrate data from multiple batches or conditions, scVI ²¹ models the batch effect explicitly. It introduces a batch‐specific latent variable that helps remove batch‐specific variations while preserving biological variations. In addition to removing the batch effect/“categorical_covariate_keys” in scVI, it also removes the “continuous_covariate_keys,” such as “percent_mito” and “percent_ribo” (https://docs.scvi‐tools.org/en/stable/tutorials/index.html). To correct for the remaining differences of QC measurements among each sample after second‐round QCs, we used “nCount_RNA” (total number of RNA molecules [counts]), “nFeature_RNA” (number of unique genes [features] detected in each nuclei), “percent.mt” and “percent.ribo” together as the “continuous_covariate_keys” in scVI (Figures S1C–F). After obtaining the cell embeddings for the integrated datasets, we loaded these embeddings into Seurat, ²⁰ and applied the functions “FindNeighbors” and “FindClusters” for cell clustering. Finally, we annotated cell types with experimentally validated gene markers from the original studies. ¹⁶ , ¹⁷ , ¹⁸ The same procedure was implemented when performing microglia subclustering. To evaluate the fidelity of cell‐type labels from the dataset integration process, we compare it to annotations by Seurat ²⁰ from The Alzheimer's Cell Atlas (TACA) database. ²² We applied the “precision_score” and “recall_score” functions from the scikit‐learn Python package to evaluate the consistency.

2.4. Trajectory analysis of snRNA‐seq profiles

Trajectory analysis is a technique used in snRNA‐seq data analysis to identify and visualize cellular differentiation pathways or developmental trajectories. Trajectory analysis uses these gene expression profiles to infer the order and timing of cell‐state transitions, or trajectories, during disease progression. In this study, we chose Monocle3, ²³ which would first cluster nuclei into supergroups based on partition‐based graph abstraction (PAGA). ²⁴ PAGA converted all nuclei into a k nearest neighbor graph with the provided nuclei embedding (in our study, we used the nuclei embedding from a previous integration step). The generated k nearest neighbor graph was further partitioned into supergroups with the Louvian algorithm. A coarse‐grained graph was created, connecting only supergroups, based on a statistical model. ²⁴ This model deemed supergroups to be connected if the count of their connecting edges surpassed the expected number under random conditions by a certain fraction. ²⁴ After that, a starting point needs to be determined to infer the trajectory. We used the function “get_earliest_principal_node” in Monocle3 (https://github.com/cole‐trapnell‐lab/monocle3/blob/master/examples/c_elegans_embryo.R) to achieve this. We chose homeostatic microglia as the input. Then this function would determine the supergroup which contained the most homeostasis microglia as the starting point for the trajectory. Monocle3 ²³ calculated the length of the path from the starting points to each nucleus, following the trajectory of the graph. The pseudo‐time assigned to a nucleus corresponds to the shortest path length from that nucleus to the nearest start point on the graph.

2.5. Building the human protein–protein interactome

To build the comprehensive human interactome from the most contemporary data available, we assembled 18 commonly used protein–protein interactome (PPI) databases with experimental evidence: (1) binary PPIs tested by high‐throughput yeast‐two‐hybrid (Y2H) systems, ²⁵ (2) kinase–substrate interactions, (3) signaling networks, (4) binary PPIs from three‐dimensional protein structures, (5) protein complexes data, and (6) carefully literature‐curated PPIs. In total, 351,444 PPIs connecting 17,706 unique proteins were used in this study (see Appendix III in supporting information). More details about the building of human PPI networks have been provided in previous studies. ²⁶ , ²⁷

2.6. Transition network analysis

We conducted a transition network analysis to study how each microglia subtype transitions from a non–disease‐related status to a disease‐related status. The transition network should include the genes that contribute most to this process. We applied a tool called TENET, ²⁸ which uses transfer entropy ²⁹ (TE) to reconstruct gene regulatory network (GRN) from single cell transcriptome data. TE is a measure of information transfer between two or more stochastic processes. It measures the amount of information about the future state of one system that can be obtained from the past state of another system, above and beyond the information provided by the past states of the first system alone. TENET ²⁸ used TE to measure the causal relationships between genes at different time points (Formula 1).

$$\mathrm{T}{\mathrm{E}}_{X\to Y}=H\left(Y_t|Y_{t-1:t-L}\right)-H\left(Y_t|Y_{t-1:t-L},X_{t-1:t-L}\right)$$ (1)

Here, H(X) represents the Shannon entropy of X, and L refers to the sequence length of past events taken into account when calculating TE. TE determines how much the knowledge of X from time t‐1 to t‐L reduces the uncertainty in Y at time t. When computing TE, the original manuscript (https://github.com/neocaleb/TENET) provided two options: (1) computing TEs only between transcription factors and their target genes at different time points; and (2) computing TEs at different time points for all possible combinations of gene pairs. We used a third option as we wanted to incorporate human PPIs: We computed TEs only between gene–gene pairs with physical binding evidence (human protein interactome), restricted our analysis to gene–gene pairs exhibiting z score normalized TE values > 1 (see Section 2.5), and then constructed the final transition network. As an example, we illustrate below how to apply TENET for transition network estimation using DAM between pseudo‐time intervals T2 and T3 in the occipital cortex brain region. We collected the differentially expressed genes (DEGs) for DAM at pseudo‐time intervals T2 and T3 separately; the differentially expressed genes (using DAM at pseudo‐time interval T2 as the example) were computed by comparing genes from DAM at T2 intervals versus the rest of the nuclei. We used the union of significant DEGs (false discovery rate < 0.05 and |log₂FC| ≥ 0.1) and DAM nuclei only in AD samples between the two pseudo‐time intervals to construct the input count matrix.

2.7. Predicting putative molecular drivers within the transition network

In this study, the putative molecular drivers refer to the key network proteins/genes that are highly associated with disease pathogenesis or progression based on previous studies. ³⁰ , ³¹ , ³² We considered four different connectivity measurements (topological properties) to identify putative molecular drivers: degree centrality, closeness centrality, betweenness centrality, and eigenvalue centrality. We define proteins that have connectivity greater than the average plus one standard deviation in at least two topological properties as the putative molecular drivers. ³³ We used the NetworkX Python package to evaluate each topological property.

2.8. Gene set enrichment analysis

We assembled drug–gene signatures from the CMap database containing 6100 expression profiles relating 1309 compounds. ¹⁵ We used the gene set enrichment analysis (GSEA) algorithm to predict drugs across each transition network, ³⁴ with the enrichment score (ES) computed as follows (Formulas (2), (3), (4)).

$$\mathrm{ES}=\left\{\begin{array}{cc}\mathrm{E}{\mathrm{S}}_{\mathrm{up}}-\mathrm{E}{\mathrm{S}}_{\mathrm{down}}& \mathrm{sgn}\left(\mathrm{E}{\mathrm{S}}_{\mathrm{up}}\right)\ne \mathrm{sgn}\left(\mathrm{E}{\mathrm{S}}_{\mathrm{down}}\right)\\ 0& \mathrm{otherwise}\end{array}\right.$$ (2)

The ESs $\mathrm{E}{\mathrm{S}}_{\mathrm{up}}$ and $\mathrm{E}{\mathrm{S}}_{\mathrm{down}}$ for upregulated and downregulated genes, respectively, in the input transition network are calculated independently using the same two‐step process outlined below. We first compute intermediate parameters a and b:

$$a=\underset{1\le \mathrm{j}\le \mathrm{s}}{\max }\left(\frac{j}{s}-\frac{V\left(j\right)}{r}\right)b=\underset{1\le \mathrm{j}\le \mathrm{s}}{\max }\left(\frac{V\left(j\right)}{r}-\frac{j-1}{s}\right)$$ (3)

where j = 1, 2, …, s were the gene sets from the transitions network sorted in ascending order by their rank in the gene profiles of the drug being evaluated. The rank of gene j is denoted by V(j), where $1\le V(j)\le r$, with r being the number of genes from the drug profile. Then, the corresponding $\mathrm{E}{\mathrm{S}}_{\mathrm{up}}$ and $\mathrm{E}{\mathrm{S}}_{\mathrm{down}}$ equal:

$$\mathrm{E}{\mathrm{S}}_{\mathrm{up}}=\left\{\begin{array}{cc}{a}_{\mathrm{up}}& \mathrm{if}{a}_{\mathrm{up}}>b_{\mathrm{up}}\\ -b_{\mathrm{up}}& \mathrm{if}{b}_{\mathrm{up}}>a_{\mathrm{up}}\end{array}\right.\mathrm{E}{\mathrm{S}}_{\mathrm{down}}=\left\{\begin{array}{cc}{a}_{\mathrm{down}}& \mathrm{if}{a}_{\mathrm{down}}>b_{\mathrm{down}}\\ -b_{\mathrm{down}}& \mathrm{if}{b}_{\mathrm{down}}>a_{\mathrm{down}}\end{array}\right.$$ (4)

In the equations mentioned, $a_{\mathrm{up}/\mathrm{down}}$ and $b_{\mathrm{up}/\mathrm{down}}$ are calculated individually for the upregulated and downregulated genes within the transition network. The GSEA ES indicates the drug's potential to reverse the gene expression patterns in the given transition network. To assess the significance of the calculated ES value, permutation tests were conducted 1000 times using randomly generated gene lists. These lists contained the same number of up‐ and downregulated genes as found in the input molecular network. Consequently, drugs with a significantly large positive ES value, evidenced by a q value of ≤ 0.05, were chosen.

2.9. Enrichment analysis

All pathway enrichment analyses were conducted using either Kyoto Encyclopedia of Genes and Genomes (KEGG) Human 2021 or Gene Ontology (GO) biological process 2021 from Enrichr, ³⁵ respectively. The 22 immune pathways were selected from the KEGG pathway database (https://www.genome.jp/kegg/pathway.html) module 5.1 Immune system. We selected synaptic‐related biological processes from the GO database if it was significantly enriched according to one set of DEGs. In total, we selected 20 synapse‐related biological processes.

2.10. Estimating treatment effects using real‐world evidence (MarketScan)

We now describe the methods used to estimate comparative effectiveness of ketorolac for preventing or delaying onset of AD in the general population, based on analysis of health insurance claims records.

2.10.1. Study design, setting, and population

We performed a cross‐sectional cohort study using data from the MarketScan insurance claims database, which contains records on almost 172 million insured individuals across the United States. Individuals were enrolled with either Medicare, Medicaid, or other commercial health insurers from January 2011 to December 2020.

2.10.2. Eligibility criteria

Eligible participants had to be over the age of 70 with at least 1 year of history prior to the index date. We excluded individuals with a prior diagnosis of AD in their record history and those with previous prescriptions of ketorolac prior to their eligibility. Additionally, individuals with missing sex and age (no date of birth) were excluded from the analysis.

2.10.3. Index date/time zero

Index date/time zero—the time from which outcomes were measured and until which baseline characteristics were extracted—was defined for the treatment group as the first prescription of ketorolac after eligibility was met. Then, ³⁶ the index date for the control group for each of the drugs was defined from a pool of individuals who had a prescription of any drug from the Anatomical Therapeutic Chemical (ATC) class M01, which is the second‐level ATC class of ketorolac. The exposure of individuals in the control group was then the first prescription of a randomly selected drug for each individual from the drugs they took from the corresponding ATC level.

2.10.4. Follow‐up period

Eligible participants were followed up from their index date until their first AD diagnosis, their last clinical record (either a visit or a prescription), December 2020 (end of data), or 3 years (end of administrative censoring), whichever occurred the earliest.

2.10.5. Variables

The exposure of ketorolac was defined as a prescription record of RXCUI codes 35827. The primary outcome, time to AD onset, was defined as the first post index date record of Clinical Classifications Software (CCS) International Classification of Disease (ICD)‐9 code 331.0 and ICD‐10 codes G30 and F00 (and their descendant codes). Additionally, because participants self‐assign into treatment in a non‐randomized way, we controlled for multiple potential confounders recorded in the 1‐year period prior to index date. These included age at index date and biological sex; 18 Charlson Comorbidity Indices ³⁷ ; 277 CCS diagnosis codes indications; indications of any prescription to any of the 86 ATC‐2 drugs; and 18 indications to cognitive‐related events, such as diagnoses of dementia, persistent mental disorder, memory loss, and so forth.

As variables indicate existence of diagnoses or prescriptions in claim records, non‐existence of any given code was interpreted as non‐existence of the corresponding diagnosis or prescription. Consequently, Charlson indices, calculated from the different diagnostic codes, do not contain missing values.

2.10.6. Covariate overlap/trimming

To ensure the analytic samples did not violate the positivity assumption required to make causal inferences, ³⁸ the dataset was first filtered to exclude extreme non‐overlapping observations between groups prior to the main time‐to‐event analysis. There are multiple ways to assess covariate overlap, ³⁹ but because covariate distributions are multi‐dimensional objects, many methods use the propensity score to summarize them and reduced the problem to assessing whether two one‐dimensional distributions overlap. A propensity model, using a gradient boosting trees estimator, was fitted and observations with extreme scores passing a dynamic optimal threshold were excluded ⁴⁰ (see Appendix I in supporting information).

2.10.7. Statistical analysis

Statistical analysis was applied to ketorolac after trimming on the remaining (non‐excluded) samples. As the primary outcome was time to first AD diagnosis, time‐to‐event modeling was applied. Cumulative incidence curves for treatment and controls groups were estimated using weighted Kaplan–Meier estimators. The weights were derived from an inverse probability weighting (IPW) model, regressing the treatment indicator against the variables described above using logistic regression and transforming its predicted probabilities into balancing weights. The weights were then used to estimate covariate‐adjusted hazard ratios (HRs) using weighted Cox regression and cumulative incidence curves per group using a Kaplan–Meier estimator. ⁴¹ Covariate‐adjusted differences in cumulative incidence were estimated by taking the difference in cumulative incidence between the treatment groups at year 3. As the treatment is modeled as a point intervention for which we emulate randomization at baseline, the causal contrast of the average treatment effect is the intention to treat (Figures S2 and S3 and Table S2 in supporting information).

Covariate balance after weighting was evaluated in two ways. First, for each covariate separately—calculated as the standardized mean difference between treatment and control groups, with an absolute difference not exceeding 0.1 being considered acceptable. ⁴² , ⁴³ Second, balancing was assessed for all covariates jointly, using the area under the receiver operating characteristic curve (AuROC) for the propensity model weighted by the IP‐weights, ⁴⁴ with an accepted score being within 0.01 of the chance AuROC of 0.5 (Figure S4 in supporting information).

To obtain valid confidence intervals (CI) for the curves, the entire procedure—chaining the overlap filtering and the subsequent survival model—was bootstrapped 500 times. Bias‐corrected P values ⁴⁵ for the difference in cumulative incidence were calculated by applying the entire procedure to 500 random permutations of the treatment assignment to estimate a null distribution of treatment effects under the entire modeling procedure. CIs for HR were not bootstrapped but calculated analytically with robust standard errors.

For the sex‐stratified results, we applied the same procedure to generate the cumulative incidence curves using weighted Kaplan–Meier estimators for female and male samples separately. To measure how the effect of treatment differs between sexes, we conducted a difference in differences analysis estimating the sex differences in cumulative incidence differences, and fit a separate weighted Cox regression on treatment, sex, and their interaction term. In the interaction analysis, female is coded as 1 and male is coded as 0 or reference.

2.11. Estimating treatment effects using real‐world evidence (INSIGHT CRN)

2.11.1. Patient data

In addition to insurance claims data, this study also used de‐identified electronic health record (EHR) datasets (real‐world data) from the INSIGHT CRN, which contained longitudinal clinical data from ≈ 15 million patients in the New York City metropolitan area. The use of the INSIGHT data was approved by the institutional review board of Weill Cornell Medicine under protocol 21‐07023759. The number of patients with mild cognitive impairment (MCI) from January 2000 to September 2022 is ≈ 40,000 (38,896), which contains 7270 AD‐diagnosed patients (Figure S5 in supporting information).

2.11.2. Eligibility criteria

We consider the following criteria in our design. (1) Patients should have at least one MCI diagnosis according to ICD‐9/10 codes (Table S3 in supporting information). (2) The patient's age was ≥ 50 years old at MCI diagnosis. (3) There was no history of AD or AD‐related dementia diagnoses (Table S3) before the baseline. (4) The patient's first MCI diagnosis date should be before the baseline. (5) The patient should have at least 1 year of records before the baseline. (6) The baseline was defined as the first date of the trial drug prescription, and at that time point, all of the above criteria must be met.

2.11.3. Treatment strategies

For each drug trial, there are two strategies: the initiation of the trial drug at baseline (treated group) and the initiation of an alternative drug at baseline (comparison group). The treatment initiation date of the drug was defined as the first prescription date of the drug. We required at least two consecutive drug prescriptions over 30 days since the first prescription date in our database as a valid drug initiation.

2.11.4. Treatment assignment procedures

Patients were divided into different drug groups according to their baseline eligibility criteria and their treatment strategies. We assumed the treated group and comparison group were exchangeable at baseline conditional on baseline covariates. We mainly considered three types of covariates at baseline: comorbidities, medications, and other variables. For comorbidities, there are 64 covariates including comorbidities from Chronic Conditions Data Warehouse and AD risk factors that were selected by experts. ⁴⁶ The comorbidities were defined by a set of ICD‐9/10 codes. We grouped medications based on their major active ingredients that were coded in RXNorm as defined in the Unified Medical Language System. In this study, we considered the 200 most prevalent prescribed drug ingredients for building covariates for medication history. For other variables, we considered age, sex, and the time from the MCI diagnosis date to the drug initiation date (time of MCIToDrug). In total, there were 267 covariates, including two continuous variables (age and time of MCIToDrug) and 265 binary variables.

2.11.5. Follow‐up

Each patient was followed from his/her baseline until the day of the first AD diagnosis, loss to follow‐up (censoring), or 2 years after baseline, whichever happened first.

2.11.6. Outcome

The outcome was the incident AD diagnosis that was defined based on ICD‐9/10 codes within the follow‐up period.

2.11.7. Causal associations of interest

The observational analogy of the intention‐to‐treat effect of being assigned to trial drug initiation versus comparison drug initiation at baseline.

2.11.8. Emulation process

We emulated trials for all drugs. We selected drugs with at least 500 eligible patients in the treated groups for our analyses. During emulation of each trial, patients in the treated group initiated the trial drug, and patients in the comparison group initiated alternative drugs. Two methods were used to build the comparison group: (1) Choose patients who were administered a random drug other than the target trial drug. (2) Select patients who were administered a similar drug from the same ATC‐2 code as the target trial drug. The patients would be excluded if they were also in the trial drug group or prescribed the trial drug before baseline. For each targeted drug, to maintain robust validation, we performed 100 emulated trials: 50 emulated trials using random controls and 50 emulated trials using ATC‐2 controls as introduced above.

2.11.9. Propensity score matching and statistical analysis

To control confounding factors, we performed a propensity score matching (PSM) ⁴⁷ analysis. The propensity score for each individual was estimated using logistic regression with ridge penalty, with the 267 covariates identified as described in Treatment assignment procedures, after which 1:1 nearest‐neighbor matching was performed. ⁴⁸ The covariate balance after matching was assessed by absolute standardized mean difference (SMD). ⁴⁹ For each covariate, it was considered balanced if its SMD value was no more than 0.1, and the treated and control group were balanced if the number of unbalanced covariates was no more than 2% of the total number of covariates. ⁵⁰ To ensure robustness, we required that there were at least 10 successfully balanced trials after matching regarding a specific drug among 100 emulated trials. ⁴⁶ Based on all balanced emulated trials of one drug, we computed HR modeled by the Cox proportional hazard model ⁵¹ for each balanced emulated trial and reported median HR with 95% CI obtained by bootstrapping. ⁵² The successful drug candidate for AD drug repurposing was selected if its HR was < 1. All analyses were implemented using Python 3.9 and packages psmpy 0.3.13 ⁵³ for PSM and lifelines‐0.27 ⁵⁴ for Cox proportional hazard model.

2.12. Mechanistic observations of ketorolac using patient induced pluripotent stem cell–derived microglia

2.12.1. Cell culture

Patient induced pluripotent stem cell (iPSC) line (IUGB267.2) was obtained from the National Centralized Repository for Alzheimer's Disease (NCRAD) at Indiana University. The iPSCs were maintained and expanded in complete mTeSR medium (Stemcell Technologies) containing recombinant human basic fibroblast growth factor and recombinant human transforming growth factor β. iPSC‐derived microglia were achieved following the protocol from Stemcell Technologies. Briefly, iPSCs were first differentiated into hematopoietic progenitor cells (HPCs) using STEMdiff Hematopoietic kit (Stemcell Technologies) for a total of 12 days. HPCs were collected from the culture supernatant and further differentiated and maturated into microglia using STEMdiff Microglia Differentiation Kit and STEMdiff Microglia Maturation Kit (Stemcell Technologies) for a total of 30 days. Cell surface markers of CD45, CD11b, and TREM2 were stained and analyzed by flow cytometry to confirm the microglia population. Patient iPSC‐derived microglia were treated with ketorolac (50μ m, Sigma‐Aldrich) or untreated (vehicle control) for 24 hours.

2.12.2. RNA isolation and sequencing

After ketorolac treatment, total RNA from microglia was extracted using Quick‐RNA Miniprep Kit (Zymo Research). RNA quality was analyzed by the Agilent Tapestation and bulk RNA sequencing was performed using Illumina NovaSeq platform at Genomics Core at Cleveland Clinic Lerner Research Institute.

2.12.3. QC, alignment, and differential expression analysis for RNA sequencing

We used “fastp,” ⁵⁵ which is a widely used tool for QC and preprocessing of sequencing data (in terms of FASTQ files). Next, we aligned sequencing reads to the reference genome using the “Rsubread” ⁵⁶ R package. The output BAM files were used as the input for function “featureCounts” within “Rsubread” for count matrix generation. Finally, we conducted differential expression analyses for ketorolac treated versus untreated cells using the “edgeR” R package. ⁵⁷

2.13. Data and code availability

Key custom codes used in this work are available at the following GitHub repository: https://github.com/ChengF‐Lab/microglia‐NIM

3. RESULTS

3.1. Network‐based integration of human AD brain snRNA‐seq data

We describe an analytical framework that integrates multiple human post mortem brain snRNA‐seq data and human protein interactome data to inspect the dynamic evolution of microglia subtypes, including DAM, tau microglia, and neuroinflammation‐like microglia (NIM). Specifically, we integrated three snRNA‐seq datasets from human brain tissues covering five different regions: prefrontal cortex, entorhinal cortex, superfrontal gyrus, occipital cortex, and occipitotemporal cortex. The integrated datasets contain 768,858 nuclei, of which 175,075 are microglia nuclei. We applied Monocle3 ²³ to extract the trajectories of all microglia with the assumption that different microglial subtypes might emerge at different stages. We combined human PPI data with an entropy‐based approach ²⁸ to identify transition networks across different microglial subpopulations. Via unique integration of microglial transition network–based drug prediction and population‐based validation from two independent, large real‐world patient databases, we prioritized and validated repurposable drugs for AD using large real‐world patient databases (see Sections 2.10 and 2.11). The entire framework is summarized in Figure 1.

FIGURE 1.

A diagram illustrating the network‐based deep learning pipeline for single‐microglia‐based target identification and drug repurposing in AD. Three datasets covering five different brain regions, including the occipital cortex, occipitotemporal cortex, entorhinal cortex, super frontal gyrus, and prefrontal cortex, were evaluated. A variational autoencoder (namely scVI) model was used to learn the embeddings for cells from the integrated datasets after feeding all the snRNA‐seq data (see Sections 2.2 and 2.3). The trajectory plots were used to define the occurrences of different microglia subtypes. The transition networks were further used to facilitate the microglia subtypes changing from non‐disease to disease‐related status. Finally, candidate drugs (e.g., ketorolac) were prioritized via specifically targeting the microglia‐specific transition network and drug‐AD outcomes were further evaluated using two independent large‐scale patient longitudinal databases (see Sections 2.10 and 2.11). AD, Alzheimer's disease; DAM, disease‐associated microglia; iPSC, induced pluripotent stem cell; MHC‐II, major histocompatibility complex class II; NIM, neuroinflammation‐like microglia; scVI, single‐cell variational inference; RWD, real‐world data; snRNA‐seq, single‐nuclei RNA sequencing.

3.2. A draft microglial atlas in AD

We used scVI, ²¹ a deep generative model, to integrate 768,858 nuclei from 77 brain samples (Table S1). The integrated datasets covered 127,571 (21.1%) astrocytes, 174,682 (28.9%) microglia, 112,769 (18.6%) oligodendrocytes, 61,296 (10.1%) excitatory neurons, 25,827 (4.3%) inhibitory neurons, and other cell types (102,734, 17.0%; Figures S1G–I, S6A in supporting information, and Table S1). To improve the quality of cell‐type identification with the integrated dataset, we compared it to that analyzed with Seurat ²⁰ from the TACA database. ²² We used precision, recall scores, and confusion matrix (see Section 2.3) to evaluate the cell type identification consistency. Common cell types shared by all three datasets presented high precision (astrocytes: 0.994; endothelial: 0.989; microglia: 0.984; occipital cortex: 0.957; and oligodendrocytes: 0.975) and recall (astrocytes: 0.994; endothelial: 0.944; microglia: 0.982; occipital cortex: 0.988; and oligodendrocytes: 0.991) scores (Table S1).

We further subclustered the integrated 174,682 microglia nuclei and re‐clustered each microglial subtype using well‐characterized marker genes ¹⁶ : (1) TMEM163, CX3CR1, SOX5, P2RY12 for homeostatic microglia; (2) ITGAX, SPP1, MSR1, MYO1E for DAM; (3) GRID2, ADGRB3, CX3CR1, DPP10 for tau microglia; and (4) GPNMB, IL1B, CD83, NFKB1 for NIM (Figures 2B and S6B). For the largest microglial population (GSE148822) among the three datasets for both occipital and occipitotemporal cortex, ¹⁶ DAM in both brain regions were characterized by elevated expression of ITGAX, SPP1, MYO1E, and MSR1 (Table S1, Figure S6C). Tau microglia highly expressed GRID2, ADGRB3, and CX3CR1 in each of the two brain regions, and significantly expressed DPP10 in the occipitotemporal cortex (Table S1, Figure S6C). NIM was marked by genes associated with immune responses, that is, GPNMB, IL1B, CD83, and NFKB1 (Table S1, Figure S6C) in both the occipital and occipitotemporal cortex. For dataset GSE157827 (see Section 2.1), the original study didn't identify any microglia subtypes across DAM, tau microglia, and NIM. ¹⁷ After integrative analysis using a deep generative model (see Section 2.3, Figure S6D), we identified a NIM subcluster for this dataset (IL1B, log₂FC = 0.48, q = 1.98 × 10⁻⁸, CD83, log₂FC = 0.78, q = 1.17 × 10⁻²⁰, Table S1), indicating the power of our deep learning–based data integration framework. For GSE147528, we also identified tau microglia in the entorhinal cortex brain regions (entorhinal cortex TAU: GRID2, log₂FC = 2.53, q < 1 × 10⁻³⁰⁰; CX3CR1, log₂FC = 0.32, q = 5.06 × 10⁻⁵; superior frontal gyrus TAU: GRID2, log₂FC = 2.61, q = 1.38 × 10⁻²⁷⁶; ADGRB3, log₂FC = 0.28, q = 0.06; Figure S6E, Table S1). These observations demonstrate that integrating multiple snRNA‐seq data will identify rare, disease‐relevant microglial subpopulations (including NIM and tau microglia) by overcoming underpowered issues from original small cell population studies.

FIGURE 2.

Deep microglial subtyping and trajectory analysis of the assembled snRNA‐seq data. A, UMAP for the integrated datasets consisting of twelve different identified cell types (e.g., microglia, astrocyte, oligodendrocyte, and endothelial cell) using well‐known marker genes (see Section 2.3). B, UMAP for the integrated microglial clustering. Using well‐established markers, we identified six different microglia subtypes: homeostasis, tau, DAM, inflammation, MHC‐II, and proliferation microglia. C, Trajectory plots (with Monocle3) for samples from occipital cortex brain regions across different groups (i.e., AD, CTR+, and CTR). D, Trajectory plots (with Monocle3) for samples from occipitotemporal cortex brain regions across different biological groups (i.e., AD, CTR+, and CTR). CTR+ group represented donors with no diagnosis of AD but with observed amyloid plaque. The CTR group represents donors with no diagnosis of AD. The number “1” in the circle denoted the starting point of each trajectory. The width of the trajectory curves coded the pseudo‐time. AD, Alzheimer's disease; DAM, disease‐associated microglia; MHC‐II, major histocompatibility complex class II; NIM, neuroinflammation‐like microglia; snRNA‐seq, single‐nuclei RNA sequencing; UMAP, Uniform Manifold Approximation Plot.

3.3. Trajectory analyses identified disease‐relevant microglial subtypes

Trajectory analysis is a computational method used to reconstruct the temporal order of biological events in a snRNA‐seq dataset. ⁵⁸ Trajectory analysis is based on the assumption that cells in a particular developmental or differentiation pathway follow a reproducible sequential order of gene expression changes. ⁵⁸ Therefore, we applied trajectory analyses to identify the emergence of microglial subtypes using Monocle3 ²³ (see Section 2.4). Our analyses revealed that the trajectories of different brain regions were different across all three datasets for donors with different phenotypes (e.g., healthy control donors, AD donors with amyloid plaque only, AD donors with both amyloid plaque and tau tangles), especially for two datasets (GSE148822 and GSE147528) having donors with paired regions (Figures 2, and S7A–C in supporting information). These observations indicate the contribution of microglia to AD pathology in a brain region–specific manner. Therefore, we hypothesized that each microglia subtype might emerge at different stages across different human brain regions as well. To test this hypothesis, we began by dividing pseudo‐time into equal time intervals (i.e., T1 = 3 > pseudo‐time ≥ 0; T2: 6 > pseudo‐time ≥ 3; T3: 9 > pseudo‐time ≥ 6; and T4: pseudo‐time ≥ 9). We found that for samples with sequencing data from paired brain regions (i.e., dataset IDs: GSE147528 and GSE148822), microglia subtypes presented distinct population distributions in different brain regions (Figures 3, 4, and Figures S8A–C and  Table S4 in supporting information). For example, the nuclei distribution of NIM significantly increased from T2 to T3 pseudo‐time interval in occipital cortex, while its distribution surged from T3 to T4 pseudo‐time interval in occipitotemporal cortex (Figures 3, 4, and Table S4). The occurrence of different microglia subtypes at different pseudo‐time intervals indicated the extent of expression profile change regarding the homeostasis microglia. To be more specific, the trajectory analyses showed that in the AD donors, the formations of tau microglia (T2 marker genes: e.g., DPP10: log₂FC = 0.22, q = 3.43 × 10⁻⁸, ADGRB3: log₂FC = 0.28, q = 5.66 × 10⁻¹⁹, Table S5 in supporting information) were “earlier” than DAM (T4 marker genes: e.g., SPP1: log₂FC = 1.11, q = 0; MSR1: log₂FC = 0.72, q = 1.96 × 10⁻⁷⁹, Table S5) and NIM (T4 marker genes: e.g., IL1B: log₂FC = 0.47, q = 1.48 × 10⁻⁴⁴; CD83: log₂FC = 2.02, q = 0, Table S5) in the occipitotemporal cortex brain region. This indicated that in the occipitotemporal cortex, the transcriptional profiles of tau microglia were more similar with homeostasis microglia compared to DAM and NIM for AD donors (GSE148822). In the occipital cortex, among AD donors, we could detect the formation of tau microglia at T2 pseudo‐time interval as well (T2 marker genes: e.g., DPP10: log₂FC = 0.23, q = 3.66 × 10⁻¹², ADGRB3: log₂FC = 0.46, q = 1.32 × 10⁻²⁶, Table S5). Most DAM (73% of total DAM nuclei in occipital cortex, Table S4) and NIM (91% of total NIM nuclei in occipital cortex, Table S4) were observed in T3 (DAM marker genes: e.g., MSR1: log₂FC = 0.28, q = 5.89 × 10⁻¹¹; NIM marker genes: e.g. IL1B: log₂FC = 0.47, q = 1.75 × 10⁻³⁷, Table S5) and T4 (DAM marker genes: e.g., MSR1: log₂FC = 1.06, q = 4.91 × 10⁻¹⁹²; NIM marker genes: CD83: log₂FC = 0.85, q = 4.45 × 10⁻⁷, Table S5) pseudo‐time intervals.

FIGURE 3.

Network‐based pathobiological analysis of microglia from AD donors with occipital cortex. A, Microglia nuclei distribution across different groups (i.e., AD, CTR+, and CTR) and different pseudo‐time intervals. The CTR+ group represents donors with no diagnosis of AD but with observed amyloid plaque. The CTR group represents donors with no diagnosis of AD. B, Enrichment analyses with KEGG immune pathways for significantly upregulated genes (log2FC ≥ 0.5 and q < 0.05) in three microglia subtypes (DAM, NIM, and tau microglia) across pseudo‐time intervals. C, Transition network from non‐DAM (in T2 pseudo‐time interval) to DAM (in T3 pseudo‐time interval; see Section 2.6). This transition network included 27 nodes (proteins) and 30 protein–protein interactions (edges). Green nodes are upregulated genes and gray nodes are downregulated genes. D, Transition network from non‐NIM (in T2 pseudo‐time interval) to NIM (in T3 pseudo‐time interval; see Section 2.6). The transition network included 89 nodes (proteins) and 106 edges. Yellow nodes are upregulated genes and blue nodes are downregulated genes. Diamond nodes represented putative network drivers (see Section 2.7). AD, Alzheimer's disease; DAM, disease‐associated microglia; KEGG; Kyoto Encyclopedia of Genes and Genomes; MHC‐II, major histocompatibility complex class II; NIM, inflammation‐like microglia.

FIGURE 4.

Network‐based pathobiological analysis of microglia from AD donors with occipitotemporal cortex. A, Microglia nuclei distribution across different groups (i.e., AD, CTR+, and CTR) and different pseudo‐time intervals. The CTR+ group represents donors with no diagnosis of AD but with observed amyloid plaque. The CTR group represents donors with no diagnosis of AD. B, Enrichment analyses with KEGG immune pathways for significantly upregulated genes (log2FC ≥ 0.5 and q < 0.05) cross three microglial subtypes (DAM, NIM, and tau) across pseudo‐time intervals. C, Transition network from non‐DAM (in T3 pseudo‐time interval) to DAM (in T4 pseudo‐time interval) (see Section 2.6). The transition network included 81 nodes (proteins) and 87 edges (PPIs). Green nodes denote upregulated genes and gray nodes denote downregulated genes. D, Transition network from non‐NIM (in T3 pseudo‐time interval) to NIM (in T4 pseudo‐time interval; see Section 2.6). The transition network included 89 nodes (proteins) and 99 edges (PPIs). Yellow nodes denote upregulated genes and blue nodes denote downregulated genes. Diamond nodes denote inferred KNDs (see Section 2.7). AD, Alzheimer's disease; DAM, disease‐associated microglia; KEGG; Kyoto Encyclopedia of Genes and Genomes; KNDs, key network drivers; MHC‐II, major histocompatibility complex class II; NIM, inflammation‐like microglia; PPI, protein–protein interactome.

FIGURE 5.

Network‐based discovery and real‐world patient data‐based validation of repurposable drugs for AD via specifically targeting the transition from NIM to non‐NIM (see Sections 2.6, 2.7, 2.8). A, Top prioritized drugs that specifically target (reverse the transcriptomic patterns of the transition network measured by GSEA, see Section 2.8) two transition networks from NIM to non‐NIM in occipital and occipitotemporal cortex. “+” sign indicates the adjusted P value is < 0.05. B, Hazard ratio plots for ketorolac (a US Food and Drug Administration–approved anti‐inflammatory medicine) across two independent patient databases (INSIGHT Clinical Research Network / INSIGHT CRN and MarketScan, see Sections 2.10 and 2.11). The comparison drug cohort was selected by randomly choosing patients who received a random drug different from the target trial (see Sections 2.10 and 2.11). For the ATC‐based matched drug cohort design (Figure S4 in supporting information), the comparison drug cohort was composed of patients who were prescribed a drug similar to the target trial drug with the same ATC‐2 category. C, Cumulative incidence plot of ketorolac and its second level ATC comparator group from the MarketScan database. The x axis represents the followed‐up time and y axis denotes the cumulative incidence of having an AD diagnosis. D, Drug‐target network analysis for ketorolac based on RNA sequencing results from the ketorolac treated and untreated iPSC‐derived microglia. E, Volcano plot visualization for differentially expressed gene between ketorolac treated versus untreated iPSC‐derived microglia. AD, Alzheimer's disease; ATC, Anatomical Therapeutic Chemical; CRN, Clinical Research Network; GSEA, gene set enrichment analysis; iPSC, induced pluripotent stem cell; NIM, inflammation‐like microglia.

We next turned to investigate microglia subtypes among CTR+ and CTR donors at each pseudo‐time interval and identified DAM at T3 (marker genes: e.g., SPP1: log₂FC = 1.50, q = 2.63 × 10⁻¹⁸⁴; MYO1E: log₂FC = 0.84, q = 4.97 × 10⁻²⁷) and T4 (marker genes: e.g., SPP1: log₂FC = 1.28, q = 3.40 × 10⁻²³; MYO1E: log₂FC = 0.94, q = 5.75 × 10⁻⁷) pseudo‐time intervals only in CTR+ donors with occipital cortex brain regions (Table S5). The CTR donors represent healthy controls with neither observed amyloid plaque nor tau tangles, ¹⁶ while CTR+ donors represent healthy controls with observed amyloid plaque only. ¹⁶ However, we were unable to identify either NIM or tau microglia subpopulations in the CTR and CTR+ donors in either occipitotemporal cortex or occipital cortex at any pseudo‐time interval.

3.4. Identifying pathobiological pathways for AD‐relevant microglial subtypes

We next turned to inspect pathobiological pathways across different microglial subtypes. We included DEGs that were upregulated with log₂FC ≥ 0.5 and q < 0.05 in each microglia subtype within each pseudo‐time interval. We then conducted enrichment analysis with 22 immune pathways from the KEGG database ⁵⁹ and 20 synapse‐related biological processes from the GO database ⁶⁰ (see Section 2.9). Compared to DAM and tau microglia, upregulated DEGs in NIM were enriched with more immune pathways across all datasets and brain regions in AD or Braak stage VI samples (Figures 3, 4, and Figures S9A,C and S10A in supporting information). Compared to DAM and NIM, tau microglia presented consistent synapse‐enriched patterns across most pseudo‐time intervals and datasets (Figures S8D,E, S9B,D, S10B and Table S6 in supporting information).

We focused on one immune response pathway, Fc gamma R‐mediated phagocytosis, because it was enriched by upregulated genes from NIM across all five brain regions. In total, there are 11 upregulated genes (i.e., LYN, HCK, FCGR3A, SCIN, PTPRC, FCGR2A, LIMK2, PRKCA, SCIN, BIN1, and FCGR1A) associated with Fc gamma R‐mediated phagocytosis (Table S7 in supporting information). LYN proto‐oncogene, Src family tyrosine kinase (LYN) is critical for inducing neurotoxicity caused by Aβ and tau hyperphosphorylation. Knocking down LYN inhibited neuronal cell death that was induced by Aβ1‐42. ⁶¹ Bridging integrator 1 (BIN1), one of the most well‐known AD risk genes, has been demonstrated to be the key modulator of proinflammatory activity of microglia in a primary microglia culture study. ⁶² It has been found that BIN1 could positively regulate the behavior of type I interferon (IFN) and immune pathways, as well as antigen presentation through major histocompatibility complex class I (MHC‐I). ⁶² One in vitro study found that in J20 mice lacking HCK proto‐oncogene, Src family tyrosine kinase (HCK), increased Aβ plaque load while diminished microglial coverage of plaques was observed. ⁶³ This study indicated that the Hck pathway significantly influences microglial neuroprotective activity in the early phases of AD progression. ⁶³ In summary, the Fc gamma R‐mediated phagocytosis pathway plays an important role in regulating neuroinflammation activities of microglia.

3.5. Discovery of putative molecular drivers for DAM

We next turned to pinpoint putative molecular drivers between non‐DAM and DAM. We conducted transition network analysis for each microglia subtype with a tool called TENET ²⁸ (see Section 2.6). We identified putative key network drivers (KNDs) based on topology properties (including degree, closeness, betweenness, and eigenvector centralities [see Section 2.7]) based on a previous study. ³³ The putative molecular drivers refer to the key network proteins/genes in which they are highly associated with disease pathogenesis or progression based on previous studies. ³⁰ , ³¹ , ³² For AD donor samples from the occipital cortex brain region, the transition network from non‐DAM (pseudo‐time interval T2) to DAM (pseudo‐time interval T3) contained 27 nodes (proteins) and 30 edges (PPIs; Figure 3 and Table S8 in supporting information). We identified several putative molecular drivers from the non‐DAM to DAM transition network of the occipital cortex brain region, such as autophagy related 7 (ATG7), SLIT‐ROBO Rho GTPase activating protein 3 (SRGAP3), inositol polyphosphate‐5‐phosphatase D (INPP5D), protein kinase C alpha (PRKCA), and spleen associated tyrosine kinase (SYK). Deficiency of ATG7 was associated with both accumulated Aβ ⁶⁴ and phosphorylated tau ⁶⁵ deposition. Drosophila and mouse tau models found that SRGAP3 regulated the binding between tau and presynaptic vesicles and reducing SRGAP3 could prevent the tau‐induced defects. ⁶⁶ The transition network also suggests multiple pathological regulators that did not present strong network connectivity properties. One noteworthy target is INPP5D, one of the most common risk genes for AD (rs35349669). ⁶⁷ One recent study found that INPP5D expression is decreased in the brains of a mouse model of AD and this decrease is associated with increased plaque formation and glial reactivity, which are hallmark features of AD pathology (Castranio et al. ⁶⁸ ). PRKCA has been suggested as a promising therapeutic target for AD. ⁶⁹ Bryostatin 1 has been tested in an AD clinical trial. ⁷⁰ Activation of SYK was found to be crucial for DAM‐like microglia in restraining Aβ pathology. ⁷¹ , ⁷²

In AD occipitotemporal cortex, the transition network from non‐DAM (pseudo‐time interval T3) to DAM (pseudo‐time interval T4) contained 81 nodes (proteins) and 87 edges (PPIs; Figure 4 and Table S8). Our analysis revealed multiple putative molecular drivers involved in the transition network, including ferritin light chain (FTL), cytokinesis 4 (DOCK4), B cell linker (BLNK), and activated leukocyte cell adhesion molecule (ALCAM). One histological study focusing on relations between iron levels and AD found a subset of microglia ⁷³ with elevated expression of FTL and Iba1, also decreased expression of transmembrane protein 119 (TMEM119) and purinergic receptor P2Y12 (P2RY12), which were consistent with those in DAM. Other proteins from the transition network, which were not identified as putative molecular drivers according to topological properties, have also been demonstrated as having an association with AD. For example, a multi‐trait association study identified the dedicator of DOCK4 as the likely causal gene responsible for the shared genetic influence on diastolic blood pressure and AD at the genomic region where DOCK4 is located. ⁷⁴ BLNK was identified as an AD genome‐wide association study (GWAS) gene ⁷⁵ and another single‐cell RNA‐seq analysis study confirmed that BLNK is one of the genes related to Aβ pathology. ⁷⁶ One cognitive impairment study (including 44 patients with MCI, 71 patients with AD, and 18 patients with other dementia) found that participants’ plasma level of ALCAM was positively correlated with degree of atrophy in the medial temporal lobe structure. ⁷⁷ The same study found that using regression analyses, ALCAM, together with APOE ε4, education, age, and vascular cell adhesion molecule 1, could predict AD with high precision (area under the curve = 0.891) even without considering the imaging diagnosis. ⁷⁷

3.6. Discovery of molecular drivers for NIM

The transition network from non‐NIM to NIM microglia in the occipital cortex (from pseudo‐time T2 to T3) contained 89 nodes (proteins) and 106 edges (PPIs; Figure 3 and Table S8). We discovered several putative molecular drivers (e.g., cathepsin B [CTSB], RAN binding protein 9 [RANBP9], phosphatidylinositol binding clathrin assembly protein [PICALM], and ADAM metallopeptidase domain 10 [ADAM10]) that may play a role in the transition process from non‐NIM to NIM networks. CTSB was suggested as a potential drug target involved in lysosomal leakage of AD pathology. ⁷⁸ Several mouse model studies ⁷⁹ , ⁸⁰ have showed that RANBP9 exacerbated synaptic damage, elevated production of microglial neurotoxic cytokines, specifically tumor necrosis factor‐alpha (TNF‐α) and interleukin‐1 beta (IL‐1β). ⁸¹ Multiple genetic studies have confirmed the significant overlap between PICALM expression quantitative trait loci and AD GWAS signals in microglia. ⁸² , ⁸³ , ⁸⁴ , ⁸⁵ The role of ADAM10 in the non‐amyloidogenic processing of amyloid precursor protein is well recognized as a defensive mechanism against the advancement of AD. ⁸⁶ Additionally, pharmacological data points to ADAM10 as a candidate target for therapeutic intervention in AD. ⁸⁶

The transition network from non‐NIM (pseudo‐time interval T3) to NIM (pseudo‐time interval T4) with respect to the occipitotemporal cortex contained 89 nodes (proteins) and 99 edges (PPIs; Figure 4 and Table S8). We pinpointed several putative molecular drivers (e.g., protein kinase C beta [PRKCB], sortilin related receptor 1 [SORL1], RUNX family transcription factor 2 [RUNX2], and alpha‐2‐macroglobulin [A2M]) that facilitate the transition from non‐NIM to NIM networks. PRKCB has been identified as a potential biomarker for AD onset, and a decrease in PRKCB expression has been proposed as a possible causal factor of AD. ⁸⁷ Loss of SORL1 altered the lysosomal activity in human microglia, such as lowered lysosomal enzyme activity and decreased lysosomal exocytosis. ⁸⁸ A study with AD patient iPSCs revealed that RUNX2‐related pathway's activation inhibited neurogenesis. ⁸⁹ Expression of the A2M has the capability to forecast the onset of MCI in men. ⁹⁰ In summary, the transition network covers multiple AD risk genes and pathobiology modulators which might help explain the non‐neuroprotective role of microglia underlying AD pathogenesis.

3.7. Microglia transition network‐based discovery of repurposable drugs

We next turned to identify drug candidates by specifically blocking transition networks of NIM in both occipital and occipitotemporal cortex regions. As shown in Figure 1, we assembled drug–gene signatures in human cell lines from the CMap database. ¹⁵ We posited that if a drug significantly reverses dysregulated gene expression identified by the transition network between NIM and non‐NIM, this drug may have potential for prevention or even treatment effects of AD. For GSEA, we used ES _{z_score} > 2 and q < 0.05 as a cutoff to prioritize drug candidates. For 1309 drugs from the CMap, ¹⁵ we obtained 35 (occipital cortex) and 14 (occipitotemporal cortex) candidate FDA‐approved drugs (ES _{z _score} > 2 and q < 0.05) that are significantly enriched in the transcriptomic transition network of NIM derived from occipital cortex and occipitotemporal cortex, respectively (see Section 2.8, Figure 5, and Table S9 in supporting information).

We found that NIM to non‐NIM transition network‐predicted drugs were grouped into seven pharmacological categories (Figure 5): immunosuppressive, antihypertensive, antineoplastic, hormone antagonist, antifungal, lipid modifying, and antibacterial. Diflunisal (ES _{z _score} = 2.67, q < 0.001, Figure 5), a non‐steroidal anti‐inflammatory drug (NSAID) used for mild to moderate pain, was found to be associated with lower likelihood of AD. ⁹¹ A rat model study found that administration of bezafibrate (a pan‐peroxisome proliferator‐activated receptor [PPAR] agonist, ES _{z _score} = 2.20, q < 0.001, Figure 5), a lipid‐lowering fibrate used for primary and secondary hyperlipidemia, attenuated abnormalities with neuronal loss, tau pathology, and microgliosis in cortex and hippocampus. ⁹² Isradipine (ES _{z _score} = 2.26, q < 0.001, Figure 5), a dihydropyridine (DHP) class of calcium channel blockers (CCBs) used for treatment of hypertension, presented neuroprotective effects in transgenic mouse models of AD. ⁹³ Altogether, these network‐predicted drugs (Table S9) offer potential candidate agents via specifically blocking transition network between NIM and non‐NIM to be further tested in various experimental models.

3.8. Real‐world patient analyses pinpoint an association between ketorolac use and reduced AD incidence

We next selected drug candidates using subject matter expertise based on a combination of factors: (1) strength of the predicted associations, (2) novelty of the predicted associations with established mechanisms‐of‐action on blocking transition network between NIM and non‐NIM, (3) literature‐based evidence in support of prediction, and (4) availability of sufficient patient data for meaningful evaluation (exclusion of infrequently used medications). Applying these criteria resulted in identifying ketorolac, an approved NSAID in the treatment of moderate to severe acute onset pain, as a candidate drug. As shown in Figure 5, anti‐inflammatory agents are the biggest network‐predicted drug class, including ketorolac. We thus evaluated ketorolac as a potential candidate treatment for AD by analyzing two independent real‐world patient databases: (1) 114,593 individuals over the age of 70 who were prescribed ketorolac or other drugs from its second ATC level M01 from January 2011 to December 2020 across the United States (MarketScan) and (2) 38,896 individuals with MCI from January 2000 to September 2022 (INSIGHT CRN). We conducted cohort analysis to evaluate ketorolac–AD outcomes using state‐of‐the‐art pharmacoepidemiologic analysis (see Sections 2.10 and 2.11). Using the MarketScan database, we found that individuals taking ketorolac are associated with 11% reduced incidence of AD (HR = 0.89, 95% CI: 0.78–1.01, P = 0.065, Figure 5). The survival analysis further suggested lower cumulative risk of developing AD for those who took ketorolac compared to non‐ketorolac users after we adjusted for various confounding factors (P = 0.002, Figure 5). The additional sex‐stratified interaction analysis suggested that the Ketorolac may have a better efficacy in women (HR = 0.802, 95% CI: 0.619–1.039, P = 0.094) than men (HR = 1.015, 95% CI: 0.839‐1.229, P = 0.876), consistent with sex‐specific survival analysis (Figures 6).

FIGURE 6.

Real‐world patient data‐based stratified validation of repurposable drugs for AD via specifically targeting the transition from NIM to non‐NIM (see Sections 2.6, 2.7, 2.8). Cumulative incidence plot of ketorolac and its second level ATC comparator group from the MarketScan database for female (A) and male (B) groups. C, HR plots for ketorolac for females and males from the MarketScan database. CI was calculated by fixing the trimming (overlap) model fitted once and bootstrapping the subsequent IPW model. HR plots for three subgroup (sex: D, age: E, and race: F) analyses compared to ketorolac with both random drugs and the same level ATC‐2 drugs from Insight CRN database. AD, Alzheimer's disease; ATC, Anatomical Therapeutic Chemical; CI, confidence interval; CRN, Clinical Research Network; HR, hazard ratio; IPW, inverse probability weighting; NIM, inflammation‐like microglia.

Using the INSIGHT CRN database, propensity score stratification confirmed that usage of ketorolac was significantly associated with reduced risk of AD compared to both (1) non‐ketorolac users (HR = 0.67, 95% CI: 0.58–0.79, P = 0.002, Figure 5) and ATC‐2 code‐matched non‐ketorolac users (individuals taking drugs of the same ATC second level classification [ATC‐2]; HR = 0.83, 95% CI: 0.77–0.92, P = 0.004, Figure 5) after adjusting various confounding factors (see Methods Section). In addition, we conducted three subgroup analyses based on sex, age, and race. For sex‐stratified results, consistent with the results from the MarketScan database, compared to both random drugs and the same level ATC‐2 drugs, ketorolac shows greater beneficial effects in women ([1] non‐ketorolac users: HR = 0.64, 95% CI: 0.56–0.72, P = 0.002; and ² ATC‐2 code‐matched non‐ketorolac users: HR = 0.79, 95% CI: 0.72–0.88, P = 0.02) compared to men ([1] non‐ketorolac users: HR = 0.74, 95% CI: 0.68–0.80, P < 0.001; and ² ATC‐2 code‐matched non‐ketorolac users: HR = 0.87, 95% CI: 0.80–0.93, P = 0.002, Figure 6D and Table S10 in supporting information). The age‐stratified analysis was conducted by splitting the entire population into: (1) older than 70 cohort (individuals older than 70) and (2) younger than 70 cohort (younger than 70). We found that ketorolac presented slightly better beneficial effect in the older than 70 cohort (Figure 6E and Table S10). The race‐stratified analysis was conducted by dividing the population into two groups: (1) White and (2) Black or African American. We found that usage of ketorolac is significantly associated with reduced incidence of AD compared to both non‐ketorolac White American users (HR = 0.64, 95% CI: 0.60–0.68, P < 0.001, Figure 6F) and ATC‐2 code‐matched non‐ketorolac White American users (HR = 0.81, 95% CI: 0.79–0.83, P < 0.001, Figure 6F and Table S10). We didn't obtain enough Black individuals (only 82 individuals in this group) in the INSIGHT CRN database and there is lack of race‐specific information in the MarketScan database. We thus didn't conduct race‐specific analysis in Black individuals.

3.9. Mechanistic observation of ketorolac in patient iPSC‐derived microglia

Ketorolac is an FDA‐approved cyclooxygenase 1 (COX1 encoded by PTGS1) and 2 (COX2 encoded by PTGS2) enzyme inhibitor for moderate to severe pain treatment. Another independent cohort study based on the Korean National Health Insurance Service‐National Sample Cohort (2003–2013) ⁹⁴ showed that ketorolac is associated with decreased risk of all‐cause dementia (HR = 0.69, 95% CI: 0.53–0.91) as well. To better understand the mechanism of action of ketorolac in potential treatment of AD, we performed RNA‐seq analysis for both ketorolac‐treated and untreated microglia derived from AD patient iPSC line (see Section 2.12). We identified 56 DEG (adjusted P value [q] < 0.05, Figures 5D,E) between ketorolac‐treated versus untreated iPSC‐derived microglia. Microglia with elevated FTL expression levels show loss of homeostasis and a more activated status. ⁷³ We found upregulated FTL expression in NIM derived from both the occipital cortex (log₂FC = 1.29, q = 2.21 × 10⁻¹⁰, Table S11) and the occipitotemporal cortex (log₂FC = 0.67, q = 7.24 × 10⁻³, Table S11 in supporting information) of AD donors. We found that ketorolac treatment downregulated expression of FTL (log₂FC = –0.21, q = 4.16 × 10⁻³, Table S11). A preclinical study ⁹⁵ showed that exposure to Aβ1‐42 increases the production of bioactive type I IFN by primary microglia and upregulates expression of type I IFN genes (IFIT1 and IFIT2). Both IFIT1 and IFIT2 were upregulated in NIM derived from both occipital cortex (IFIT1: log₂FC = 1.17; IFIT2: log₂FC = 1.91, Table S11) and occipitotemporal cortex (IFIT1: log₂FC = 2.16; IFIT2: log₂FC = 0.96, Table S11) of AD donors. Our RNA‐seq analysis showed that both IFIT1 (log₂FC = ‐0.69, q = 2.29 × 10⁻³, Table S11) and IFIT2 (log₂FC = –1.31, q = 4.19 × 10⁻⁵, Table S11) were significantly downregulated after ketorolac treatment in AD patient iPSC‐derived microglia. Type I IFN signaling has been shown to directly impact cognitive function and play a role in modulating the long‐term neuroinflammatory and neurotoxicity in AD and AD‐associated microgliosis. ⁹⁶ , ⁹⁷ In summary, these preliminary mechanistic observations suggest that ketorolac may target type I IFN signaling in AD‐relevant microglia, in supporting beneficial effects in potential prevention and treatment of AD from real‐world patient databases. Further functional and mechanistic observations of ketorolac in AD microglia and other cell types are highly warranted.

4. DISCUSSION

Studies in animal models and AD human‐oriented samples have demonstrated the microglia heterogeneity in AD. ¹¹ The scRNA‐seq of 5xFAD mice revealed that DAM appeared at 8 months of age. ¹² Another ontogeny study which integrated multiple mouse scRNA‐seq data concluded that DAM was embryonically derived while disease inflammatory macrophages emerged with aging and were enhanced in neurodegenerative diseases. ⁹⁸ To uncover human microglial heterogeneity in AD, we present a network‐based methodology that integrates multiple snRNA‐seq datasets from human AD patient brains, human PPIs, and drug target networks, along with large‐scale patient‐level data observation. The transition networks from non–disease‐related microglia subtypes to disease‐related microglia subtypes showed that DAM in the late disease stages was not neuroprotective and NIM are more neurotoxic, as they are involved with proinflammatory processes which worsen disease progression. We demonstrate that blocking the transition network from non‐NIM to NIM could offer potential targets for drug repurposing. We validated one such network‐predicted drug, ketorolac, that could reduce the risk of AD using two independent large‐scale, longitudinal patient datasets.

When performing the snRNA‐seq data integration, we did not choose a supervised ⁹⁹ or semi‐supervised ¹⁰⁰ approach for cell type identification. Both supervised and semi‐supervised methods depend on existing knowledge about cell types/subtypes and we have limited knowledge about the right cell type abundances because of heterogeneities of brain cells in AD. In the case of highly heterogeneous cell types such as microglia, we posited that an unsupervised approach could be more suitable as it does not require prior information and can thus potentially identify the natural structure of brain cell abundances accurately. We further validated our unsupervised models using nuclei label from a dataset of GSE15728 (Appendix II and Table S12 in supporting information). In addition to homeostasis microglia, we identified another five subtypes of microglia, including DAM, tau microglia, NIM, MHC‐II microglia, and proliferation microglia. In addition to well‐known homeostatic microglia, the other five microglial subtypes identified in this study are consistent with previous studies. ¹⁴ , ⁹⁶ , ⁹⁷ , ⁹⁸ DAM was first discovered in an AD mouse model. ¹² One recent large‐scale human snRNA‐seq study (a total of 443 human samples) identified 12 microglial subtypes, ¹⁰¹ including proinflammatory microglia, DAM, and proliferative microglia. Another snRNA‐seq study from human cerebral cortex samples ¹⁰² identified a total of nine microglial subclusters, including DAM and proliferating microglia. Taken together, our current findings and previous studies highlight the heterogeneity of microglial subtypes in AD pathobiology. Creating effective computating algorithms to identify disease‐relevant microglial subtypes will offer powerful tools to develop microglia‐targeted therapies for AD and other brain disorders if broadly applied.

In particular, considering dataset GSE148822, which contained AD, CTR, and CTR+ donors, we could identify DAM in both AD and CTR+ donors (Tables S1 and S5), which could indicate DAM is amyloid plaque specific. Microglia are also known to play key roles in forming neuronal circuits. ¹⁰³ During the process of neurodevelopment, the formation of synapses in the cortex is induced by microglial contact. ¹⁰⁴ Tau microglia are a newly reported microglia subtype that is positively correlated with tau pathology load in AD according to snRNA sequencing data. ¹⁶ Tau microglia shared some highly expressed genes with homeostasis microglia, such as CX3CR1 and P2RY12. At the same time, tau microglia also expressed neuron‐associated genes, including GRID2, ADGRB3, and DPP10. Tau microglia presented microglia‐like morphology and the same study found that they also localized with tau pathology. ¹⁶ To further test the association between tau microglia with synapse dysfunction or synapse damage, we have collected 14 proteins (i.e., C1QA, C1QB, C1QC, C3, CX3CR1, MERTK, GPR56, SIRPA, TREM2, C3AR1, TARDBP, TNF, IL6, and MFGE8) from four manuscripts ¹⁰⁵ , ¹⁰⁶ , ¹⁰⁷ , ¹⁰⁸ studying the role of microglia in regulating synapses. The differential expression gene analyses of these 14 proteins in tau microglia did not show association with synapse loss in either AD or control groups (Table S5). However, we did find elevated expression levels of C1q complex and C3 in NIM in AD donors only in the occipital cortex brain region (pseudo‐time interval T3, C1QA: log₂FC = 0.13, q = 3.09 × 10⁻⁵; C1QB: log₂FC = 0.31, q = 3.65 × 10⁻²⁴; C1QC: log₂FC = 0.26, q = 2.59 × 10⁻¹⁵; C3: log₂FC = 0.93, q = 1.21 × 10⁻²⁵⁹; Table S5). They presented consistently elevated expression in pseudo‐time interval T4 as well (Table S5). One study ¹⁰⁶ using a mouse model discovered that inhibiting C1q and C3 could decrease the number of phagocytic microglia and reduce the extent of early synapse loss, suggesting TAU microglia may be associated with synaptic loss/damage. However, more functional observations of tau microglia are highly warranted in future studies using various transgenic mouse models. Additionally, for NIM, except the Fc gamma R‐mediated phagocytosis pathway, which was consistently enriched by AD donors across all datasets (Figures 3, 4, S9A,C, and S10A), several other immune pathways were enriched by multiple datasets as well, such as toll‐like receptor, IL17, and Th17 cell differentiation signaling pathways (Figures 3, 4, S9A,C). Additionally, our sex‐difference analysis showed that the toll‐like receptor signaling pathway was enriched only in the male donors in both brain regions (T3 and T4 pseudo‐time intervals for occipital cortex and T4 pseudo‐time interval for occipitotemporal cortex; Table S7). In contrast, we found that there were fewer enriched immune pathways in control donors across different brain regions (Figures S11A,B,D, and S12 in supporting information) except those from the prefrontal cortex (Figure S10C) and occipital cortex (Figure S11C) brain regions.

Furthermore, we identified several potential molecular drivers associated with AD pathology, such as CTSB, a key member of the lysosomal cysteine protease family. Several studies have identified functional associations between lysosomal dysfunction and CTSB in AD. For example, the Czech Brain Aging Study reported that levels of serum transcription factor EB (TFEB), the master gene for lysosomal biogenesis, were reduced in AD samples, leading to decreased executive function and language abilities. ¹⁰⁹ In a mouse model, elevated CTSB from exercise was found to boost brain‐derived neurotrophic factor (BDNF) and doublecortin (DCX) in hippocampal cells, which are essential for neurogenesis and cognitive function. Besides lysosomal dysfunction, impaired mitophagy is also involved in AD pathology; for example, an accumulation of damaged mitochondria in the hippocampus of AD patients and models. ¹¹⁰ , ¹¹¹ Stimulating mitophagy reduces insoluble Aβ1‐42 and Aβ1‐40 levels and prevents cognitive impairment in an APP/PS1 mouse model by promoting microglial phagocytosis of extracellular Aβ plaques and suppressing neuroinflammation. ¹⁰⁶ Consistently, one potential anti‐AD molecule, urolithin A, was found to be capable of reducing memory loss in AD animal models by inducing mitophagy and optimizing lysosomal function. ¹¹¹

We acknowledge several potential limitations in the current study. The current study lacked a pre‐planned strategy to harmonize or standardize the datasets; we, therefore, followed pre‐processing procedures from the original publications of each dataset. We will design more standard harmonization process in the future when integrating more sc‐/snRNA seq datasets. Also, although the overall nuclei distribution was overall consistent among the deep learning (e.g., scVI ²¹ ) and machine learning (e.g., Seurat ²⁰ )‐based tools, we found discrepancies regarding microglia subpopulations identified from multiple sources (Table S1). The main potential reason that we consider here is when using variational autoencoder–based tools for data integration, the explicit form of probability distribution involved in its design might cause limitations when in reality microglia subpopulations from different datasets followed different distributions. Furthermore, for very heterogeneous cell types during disease progression, like microglia, sequencing enough samples that could cover as many microglia subtypes as possible is required. The EHR study could also have some limitations. ²⁶ , ²⁷ , ¹¹² Pharmacoepidemiologic studies may be biased due to the absence of certain confounding factors. Despite adjusting for factors such as age, sex, and comorbidities, other clinical variables not available in our database, including education level, socioeconomic status, and specific genotypes (such as APOE and TREM2), could be linked to AD risk. Future studies should aim to gather and analyze these factors. Ultimately, observational studies with population‐based samples are unable to establish causal links between drug use and positive clinical outcomes in AD.

The pseudo‐time analyses in this study were performed with Monocle3, ²³ which constructed the trajectory purely based on the gene expression pattern. We cannot ensure that the pseudo‐time computed by Monocle3 perfectly aligns with the true biological processes. Based on trajectory analytical results, we found the existence of DAM and NIM in the same pseudo‐time interval (e.g., T3 in occipital cortex brain region for GSE148822 AD donors) highlights heterogeneities of microglial biology in AD. This does not necessarily mean that in actual disease progression DAM and NIM are formed at the same or similar disease stages. Additionally, the trajectory analyses presented various patterns for donors with the same endophenotype (e.g., AD), and even for the same donor with different brain regions. These phenomena were consistently observed across multiple datasets (Figures 2, and S7) which might indicate that AD pathology could be individualized. However, for rare population cell types like microglia, which usually account for < 10% of the total sequencing population, the sequenced population per donor could be too small for sufficient computational power. Therefore, patient stratification that could refine AD samples according to microglia expression should be one promising future research direction as it could help reduce the microglia heterogeneity, increase computational power (considering multiple patients from the same group), and consequently provide more precise interpretation for distinct AD pathology patterns and therapeutic design.

Finally, we recognize that considering snRNA‐seq alone even after integration is not adequate for studying roles of microglia in AD pathology. For example, as suggested by one PET‐based study, microglia activation and tau pathology are spatially related, where microglia activation could promote tau migration. ¹³ Another study which integrated both spatial RNA‐seq and snRNA‐seq demonstrated that microglia and astrocytes are the cell types that are, respectively, primarily and secondarily closest to Aβ plaques, while oligodendrocytes locate near the tau tangle. ¹¹³ This study suggested that characterizing the cross‐talks between microglia, astrocytes, and oligodendrocytes could better help understand AD pathology.

In summary, we believe that the network‐based methodology presented here, if broadly applied, would significantly catalyze innovation in AD drug discovery by using the large‐scale single‐cell/nucleus omics data and real‐world patient databases.

CONFLICT OF INTEREST STATEMENT

Dr. Cummings has provided consultation to AB Science, Acadia, Alkahest, AlphaCognition, ALZPathFinder, Annovis, AriBio, Artery, Avanir, Biogen, Biosplice, Cassava, Cerevel, Clinilabs, Cortexyme, Diadem, EIP Pharma, Eisai, GatehouseBio, GemVax, Genentech, Green Valley, Grifols, Janssen, Karuna, Lexeo, Lilly, Lundbeck, LSP, Merck, NervGen, Novo Nordisk, Oligomerix, Ono, Otsuka, PharmacotrophiX, PRODEO, Prothena, ReMYND, Renew, Resverlogix, Roche, Signant Health, Suven, Unlearn AI, Vaxxinity, VigilNeuro pharmaceutical, assessment, and investment companies. Dr. Leverenz has received consulting fees from Vaxxinity, grant support from GE Healthcare, and serves on a data safety monitoring board for Eisai. Any analysis, interpretation, or conclusion based on these data is solely that of the authors and not of Merative L.P. and its subsidiaries. M.D., E.K., Y.S., J.H., and M.R.Z. are employees of IBM Research. The other authors have no competing interests. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All human subjects used in this study have undergone adequately informed consent by the previous studies. Consent was not necessary for this study.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Xu J, Song W, Xu Z, et al. Single‐microglia transcriptomic transition network‐based prediction and real‐world patient data validation identifies ketorolac as a repurposable drug for Alzheimer's disease. Alzheimer's Dement. 2025;21:e14373. 10.1002/alz.14373