Interpretable deep learning translation of GWAS and multi-omics findings to identify pathobiology and drug repurposing in Alzheimer’s disease

SUMMARY

Translating human genetic findings (genome-wide association studies [GWAS]) to pathobiology and therapeutic discovery remains a major challenge for Alzheimer’s disease (AD). We present a network topology-based deep learning framework to identify disease-associated genes (NETTAG). We leverage non-coding GWAS loci effects on quantitative trait loci, enhancers and CpG islands, promoter regions, open chromatin, and promoter flanking regions under the protein-protein interactome. Via NETTAG, we identified 156 AD-risk genes enriched in druggable targets. Combining network-based prediction and retrospective case-control observations with 10 million individuals, we identified that usage of four drugs (ibuprofen, gemfibrozil, cholecalciferol, and ceftriaxone) is associated with reduced likelihood of AD incidence. Gemfibrozil (an approved lipid regulator) is significantly associated with 43% reduced risk of AD compared with simvastatin using an active-comparator design (95% confidence interval 0.51–0.63, p < 0.0001). In summary, NETTAG offers a deep learning methodology that utilizes GWAS and multi-genomic findings to identify pathobiology and drug repurposing in AD.

Graphical abstract

In brief

Xu et al. develop an interpretable deep learning framework (termed NETTAG) to identify Alzheimer’s disease (AD) risk genes and drug targets through leveraging GWAS, multimodal genomics, and human protein-protein interactome network data. Combining human interactome network-based prediction and EHR-based patient data validation identifies gemfibrozil’s effect on reduced incidence of AD.

INTRODUCTION

Alzheimer’s disease (AD), first described in 1907 by Alois Alzheimer, is the most common type of dementia with gradual cognitive decline and memory loss.¹ AD and AD-related dementias (AD/ADRD) are a major global health challenge, and the number of affected individuals is expected to double by 2050,^2,3 affecting more than 150 million people worldwide.^2,4 The prevalence of AD in the United States is expected to double by 2050,^5,6 while the attrition rate for AD clinical trials (2002–2012) is estimated at 99%.⁷ High-throughput DNA/RNA sequencing technologies have rapidly led to a robust body of genetic and genomic data in multiple national AD genome projects, including the Alzheimer’s Disease Sequencing Project (ADSP)⁸ and the Alzheimer’s Disease Neuroimaging Initiative (ADNI).⁹ Genome-wide association studies (GWAS) have also identified ~100 AD susceptibility loci.^10-14 Despite this progress in understanding AD genetic risk, the complex, polygenic, and pleiotropic genetic architecture has precluded the development of therapeutics. This is partly because the growing mass of genetic datasets has not been effectively explored for AD drug discovery and development.

Recent advances in genetics and systems biology have shown that AD is governed by network-associated molecular determinants (termed disease modules) of common endotypes or endophenotypes.^15,16 Approaching AD with a simplistic single-target approach has been demonstrated as effective for developing symptomatic therapies and for monoclonal antibody development, and specifically modulating genetic risk genes may lead to disease-modifying treatments in AD.¹⁶ However, more effective means of advancing therapies for disease modification are urgently needed.¹⁵ A recent study showed that selecting genetically supported targets can double the success rate in clinical development.¹⁷ However, existing data, including genomics, transcriptomics, proteomics, and interactomics (protein-protein interactions), has not yet been fully utilized and integrated for targeted therapeutic development for AD.^18-20

Understanding AD from the point of view of human interactome perturbations is the essence of network medicine.^15,16 The central hypothesis of AD network medicine is that cellular networks disturbed by genetic variants gradually rewire throughout disease pathogenesis and progression.^15,16 Systematic characterization and identification of underlying pathobiology can provide a foundation for identifying disease-modifying targets for AD. Integration of the genome, transcriptome, proteome, and human interactome is essential for such identification. Several network-based analytic techniques have recently been developed to address the myriad different types of inputs of omics layers. Weighted gene co-expression network analysis (WGCNA)²¹ and multiscale embedded gene co-expression network analysis (MEGENA)²² are two commonly used network-based methods that are based on brain gene/protein expression profiles of AD patients. Random walk with restart (RWR)²³ has also been used to predict AD-associated genes, based on input data from literature-reported genes. However, in contrast to NETTAG, these existing network-based approaches do not leverage the integration of multi-omics profiles, such as genetics, functional genomics, transcriptomics, and proteomics, for risk gene prediction and drug target identification.

This study introduces a network topology-based deep learning framework to identify disease-associated genes (NETTAG) and drug targets from genetic and multi-omics discoveries for AD. The fundamental premise of NETTAG is that disease risk genes (1) exhibit distinct functional characteristics compared with non-risk genes and therefore can be distinguished by their aggregated genomic features, (2) converge to a limited number of pathobiological pathways captured by the human protein-protein interactome, and (3) include multiple AD pathobiological modulators and potential therapeutic targets. In addition to pinpointing well-known AD-risk genes, (e.g., APOE, BIN1, and PICALM), NETTAG identified potential AD-associated genes, such as MEF2D and CPLX2. Importantly, we demonstrated the putative risk genes identified by NETTAG offer actionable drug targets for drug repurposing, and we further validated drug findings in a large real-world patient database.

RESULTS

An interpretable deep learning framework

Here, we present NETTAG, a network-based deep learning framework to identify likely risk genes from GWAS and brain-specific multi-omics findings in AD. NETTAG integrates multi-omics data and the human protein-protein interaction (PPI) network to infer likely risk genes and potential drug targets impacted by GWAS loci. The fundamental premise of NETTAG is that disease risk genes exhibit distinct functional characteristics compared with non-risk genes and that they can be distinguished by their aggregated genomic features under the human protein interactome. Specifically, we assembled non-coding GWAS loci effects on expression quantitative trait loci (eQTLs), histone QTLs, transcription factor binding QTLs, enhancers and CpG islands, promoter regions, open chromatin, and promoter flanking regions from GTEx,^24,25 NIH RoadMap,²⁶ Ensembl Regulatory Build,²⁷ SNPnexus,²⁸ and ENCODE.²⁹ The procedure for NETTAG (Figure 1) is divided into five components. First, we utilized a deep learning model to cluster PPIs into multiple functional network modules by capturing the topological structures within the human protein-protein interactome (STAR Methods). We then characterized each functional network module by linking its nodes (genes) with protein annotation from the Gene Ontology (GO) knowledgebase.³⁰ Second, we quantified a node’s (gene’s) score by integrating its functional similarity with each gene identified with multiple brain-specific gene regulatory evidences via influencing GWAS loci. Third, we prioritized likely risk genes in AD by their aggregated gene regulatory features. Fourth, we prioritized repurposable drugs for potential treatment of AD by evaluating network proximities between NETTAG-inferred AD-risk genes (alzRGs) and known drug targets under the human protein-protein interactome network model. Fifth, we identified supportive information for drug-AD outcomes using a large-scale patient longitudinal, patient electronic health record (EHR) database (Figure 1).

Figure 1. A diagram illustrating NETTAG.

We first applied a deep learning model to capture the topological structure of the PPIs and divided it into multiple subnetwork modules (STAR Methods). Then we discovered that the divided subnetwork module could approximate protein functions annotated by the Gene Ontology (GO) knowledge portal (STAR Methods). Next, we predicted likely AD-risk genes (alzRGs), which are functionally similar to genes that have been identified by different gene regulatory elements, i.e., CpG island, CCCTC-binding factor (CTCF), enhancer, expression quantitative trait loci (eQTL), histone, open chromatin, promoter, promoter flanking region, and transcriptional factor(TF). Finally, we prioritize repurposed drugs (e.g., gemfibrozil) for potential AD treatment and identified supportive information with the large-scale patient longitudinal database (STAR Methods).

A gene regulatory landscape of GWAS loci in AD

After mapping AD loci (p < 1.0 × 10⁻⁵) from multiple brain-specific gene regulatory elements (STAR Methods), we pinpointed 23 genes with CpG islands (e.g., APOE, PVR, and STK11), 19 genes with CTCF binding sites (e.g., BIN1, JPH1, and SYK), 13 genes with enhancers (e.g., BIN1, FARP1, and MARK4), 21 genes with eQTL (e.g.,CD2AP, IL6, and PVR), 169 genes with histone modifications including H3K27ac, H3K27me3, H3K36me3, H3K4me1, H3K4me2, H3K4me3, H3K9ac, and H4K20me1 (such as APOE, CKAP5, DST, and NECTIN2), 48 genes with open chromatin (e.g., BIN1, CLU, and INPP5D), 23 genes with promoters (e.g., APOE, IL6, and STK11), 59 genes with promoter flanking regions (e.g., BIN1, CLU, and MARK4), and 20 genes (e.g., BCAM, CLU, and VSNL1) with transcriptional factor binding sites, respectively (Figure 2A and Table S1). As shown in Figure 2A, 69 genes have AD loci with multiple gene regulatory evidence, such as apolipoprotein E (APOE), BIN1, CLU, IL6, and PTK2B (Table S1). Specifically, APOE loci have regulatory influence with CpG islands (rs429358 and rs7412), histones (rs405509 and rs769449), promoters (rs769449), and promoter flanking regions (rs75627662) (Figure 2A and S1A). Bridging integrator 1 (BIN1), another risk factor of late-onset Alzheimer’s disease (LOAD),^31,32 is associated with multiple regulatory elements, including CTCF binding sites (rs12989701), enhancers (rs10207628), histones (rs6431219, rs10194375, rs72838215, rs10207628), open chromatin (rs6733839), and promoter flanking regions (rs10194375). Phosphatidylinositol binding clathrin assembly protein (PICALM) loci in AD display regulatory activity involving histones (rs867611, rs527162, rs639012, and rs17817600), promoters (rs867611), and promoter flanking regions (rs10792832). PICALM regulates AD pathology with Aβ generation and disordered lipid metabolism.³³ Inositol polyphosphate-5-phosphatase D (INPP5D), a LOAD-risk gene,¹⁴ is associated with open chromatin regulation (rs10933431). Furthermore, a mouse model (5xFAD) study found that expression of Inpp5d was elevated in microglia as the disease progressed.³⁴ We also found that spleen-associated tyrosine kinase (SYK) (rs1172922) is linked with a CTCF binding site. Activation of SYK boosts inflammation³⁵ and modulates both Aβ- and tau-induced pathologies.³⁶ Together, these multi-omics analyses highlight the crucial roles of gene regulation involving various AD GWAS loci. We therefore developed NETTAG to infer gene regulatory variants and putative risk genes in AD using network-based multi-omics evidence aggregation analyses.

Figure 2. Gene regulatory landscape of AD GWAS loci.

(A) Overview of AD GWAS loci across different chromosomes after considering nine gene regulatory elements: GpG island, CCCTC-binding factor (CTCF), enhancer, expression quantitative trait loci (eQTL), histone, open chromatin, promoter, promoter flanking region, and transcriptional factor (Table S1).

(B) Proteins’ cluster numbers are positively correlated with their gene ontology (GO) annotation. We divide proteins into 10groups according to their GO terms. For example, G1 group include the proteins that have at least one, but less than ten GO annotation (Table S2). Error bars denote 1,000 randomly replicated experiments.

(C) Receiver operating characteristic (ROC) analyses of NETTAG based on four collected AD-association gene sets, i.e., AlzGene, DistiLD, DISEASES (knowledge), and TIGA (STAR Methods; Table S2).

NETTAG-based prediction of likely risk genes in AD

Through NETTAG, we first clustered PPIs into functional subnetwork modules using a topology-based self-supervised learning framework. We found that these subnetwork modules could reflect biological relationships (Figure 2B). Specifically, proteins with more GO terms (not used in training) tend to have more network clusters in the human PPI interactome (Figure 2B and Table S2). We found that proteins in the same subnetwork module tend to have more shared GO annotation (Wilcoxon signed rank test, p < 2.2 × 10⁻¹⁶, STAR Methods), indicating that network-based fingerprints of the PPI module overlay among genes and can characterize functional modularities and similarities.³⁷ We inferred likely risk genes by integrating PPI-derived network modules and multimodal analyses of nine types of gene regulatory elements implicated by AD GWAS loci. Taking CpG islands as an example, the predicted score for one particular gene regarding CpG islands could estimate its functional overlap (Spearman correlation, r = 0.44, p = 1.86 × 10⁻²⁴, Figure S1B and STAR Methods) with all 23 AD CpG island-linked genes (Table S1). Finally, we inferred likely risk genes by integrating (summing) all nine types of gene regulatory elements (STAR Methods). The area under the curve (AUC) values for receiver operating characteristic curve (ROC) using reported AD-associated genes collected from AlzGene,³⁸ DistiLD,³⁹ TIGA,⁴⁰ and DISEASES⁴¹ (Figures 2C and S2A) are 0.81, 0.80, 0.72, and 0.78 respectively, suggesting reasonable accuracy of NETTAG. We further compared NETTAG with three classical approaches: RWR,²³ spectral clustering,⁴² and k-means clustering.⁴³ We found that NETTAG predicted more genes non-overlapped with the mapped input genes associated with different regulatory elements (80%) compared with RWR (less than 50%, Figure S3A). Furthermore, genes predicted by NETTAG were more significantly enriched in AD-relevant functional pathways compared with those identified by RWR (Figure S3B). In addition, we found that the overlapping clustering method implemented in NETTAG outperformed traditional spectral clustering and k-means clustering (Figures S3C and S3D).

Via NETTAG (Figure 1), we identified 156 likely alzRGs, such as APOE, APP, BIN1, FYN, and STK11. Products (proteins) of 139 alzRGs out of 156 formed the largest connected component within 294 PPIs (Figure 3A and Table S3) in the human protein-protein interactome. Via gene and functional enrichment analyses, we found that the NETTAG-predicted alzRGs are significantly enriched by gene regulatory elements (Figures 3B and S4A), compared with the same number of randomly selected genes with a similar degree distribution in the human interactome network. We assembled AD-associated genes from the GWAS Catalog,⁴⁴ UK Biobank GWAS⁴⁵ and DisGeNET with published experimental evidence from animal models and human studies.⁴⁶ We found that alzRGs were significantly enriched in all three AD-associated gene sets: GWAS Catalog (adjusted p value [q] = 2.25 × 10⁻⁷), UK Biobank GWAS (q = 8.59 × 10⁻³), DisGeNET (q = 1.19 × 10⁻⁸, Fisher’s exact test, Table S3). Pathway enrichment analyses⁴⁷ showed that alzRGs are significantly enriched in multiple immune pathways (Table S3 and Figure S4B), including B cell (q = 5.32 × 10⁻⁴), T cell receptor (q = 1.17 × 10⁻²), and cytokine signaling pathways (IL-2: q = 6.99 × 10⁻³, IL-7: q = 1.43 × 10⁻², IL-18: q = 1.65 × 10⁻²). In summary, NETTAG achieved a high accuracy (AUC = 0.81) in predicting likely AD-risk genes with diverse functional pathways, including key immune pathways. We next performed multi-omics validation for NETTAG-predicted alzRGs, including single-cell/nuclei transcriptomics in disease-associated microglia (DAM) and astrocytes (DAA) isolated from transgenic mouse and human brains with well-established AD neuropathology.

Figure 3. Observations of 156 prioritized AD-risk genes (alzRGs) by NETTAG.

(A) Network-based visualization of 156 predicted alzRGs. 139 alzRGs are non-isolated and form a subnetwork with 294 protein-protein interactions (PPIs). Prioritized alzRGs are colored with various evidence. Green genes are the ones identified by GWAS Catalog but with no gene regulatory element evidence. Blue genes are the ones identified by GWAS Catalog and simultaneously with single or multiple gene regulatory element evidence. Yellow genes are the predicted genes with other types of evidence, e.g., multi-omics or literature evidence. Gray genes are the rest of predicted alzRGs (Table S3).

(B) Cumulative distributions of predicted scores with alzRGs and the same amount of random non-alzRGs with similar degree distribution (the human protein-protein interactome network) for expression quantitative trait loci (eQTLs), histones, and promoters, respectively.

NETTAG-predicted genes are differentially expressed in AD brains

We found 95 alzRGs (p = 2.67 × 10⁻⁷, Fisher’s exact test) that are differentially expressed regarding at least one type of transcriptomic study in AD brains. Specifically, 29 (p = 0.0185), 67 (p = 2.36 × 10⁻³), and 39 (p = 2.96 × 10⁻⁷) alzRGs are differentially expressed genes (DEGs) according to microarray (human AD patients and controls), bulk RNA sequencing (human AD patients and controls), and single-cell/nucleus RNA sequencing (AD transgenic mouse model and human postmortem brain samples) analyses, respectively (STAR Methods). Nine genes (ACTL6B, ATP2B1, EPB41L3, ABCA1, CPLX2, P2RX7, PDE1A, SLC38A2, and VSNL1) are DEGs based on all three types of differential expression evidence (Figure S5A). Additionally, the purinergic receptor P2X7 (P2RX7) inhibitor was found to reduce tau accumulation in P301Stau transgenic mice,⁴⁸ and P2X7R antagonists were supported as potential therapeutic AD options.⁴⁹ Also, visinin-like 1 (VSNL1) is co-expressed with multiple genes involved in molecular mechanisms of AD, including amyloid beta precursor protein (APP).⁵⁰ Moreover, nineteen genes (ABCA1, APOE, BCL3, BIN1, CKAP5, CLU, FARP1, HSPG2, MADD, MAPK7, MARK4, NCS1, PICALM, PTK2B, SPRED2, TGFB2, TOMM40, TOP1, and VSNL1) have been identified by gene regulatory elements and AD GWAS studies⁴⁴ (Figure S5B). More explicitly, microtubule affinity regulating kinase 4 (MARK4) is the protein linked with most gene regulatory elements, including CpG island (rs28469095), CTCF (rs12463049), enhancer (rs536518226), eQTL (rs8100183), histone (rs9653111 and rs10421247), open chromatin (rs138137383), promoter flanking region (rs10421247 and rs138137383), and TF-binding site (rs12463049) (Table S1 and Figure S5B). Furthermore, MARK4 has been suggested as a potential target for AD via binding with acetylcholinesterase (AChE) inhibitors, such as donepezil (an AChE inhibitor used for patients with AD).⁵¹ In addition to APOE, BIN1, and PICALM, CLU loci are linked with four types of regulatory elements, including histones (rs9331896 and rs1532278), open chromatin (rs2279590), promoter flanking regions (rs9331888), and TF-binding sites (rs1532278) (Table S1 and Figure S5B). Clusterin (CLU) is another risk gene for LOAD⁵² by multiple pathologies, including neuroinflammation, Aβ accumulation, and lipid metabolism.⁵³

Among the 67 alzRGs that are differentially expressed based on bulk RNA-seq studies, 42 alzRGs form a subnetwork within the human interactome network (Figure 4A). Within the 23 top differentially expressed alzRGs such as HOMER2, ABCA1, and HSPG2 (∣log₂FC∣ > 1, q < 0.05, Table S4), homer scaffold protein 2 (HOMER2) (downregulated log₂FC = −1.65, q = 2.39 × 10⁻³, Table S4) was found to inhibit APP production and secretion of Aβ peptide together with HOMER3.⁵⁴ ATP binding cassette sub-family A member 1 (ABCA1) was among the most upregulated genes according to bulk RNA-seq studies (log₂FC = 1.21, q = 1.37 × 10⁻³, Table S4), and its mutation has been associated with an elevated risk of AD.⁵⁵ Heparan sulfate proteoglycan 2 (HSPG2) was another highly upregulated gene according to bulk RNA-seq studies (Table S4). Correlation studies suggested that people with an APOEϵ4 allele have increased risk of AD if also carrying an HSPG2 A allele.⁵⁶

Figure 4. Transcriptomics-based observation of 156 prioritized AD-risk genes (alzRGs) by NETTAG.

(A) Visualization of 67 predicted alzRGs that are also DEGs according to human bulk RNA-seq studies with both late-stage AD (LAD) and control donors (STAR Methods).

(B) Violin plots show alzRGs are more likely differentially expressed in disease-associated microglia (DAM) according to mouse single-nucleus (GEO: GSE140511; three mice for each group; 5xFAD, wild-type, Trem2 knockout 5xFAD, and Trem2 knockout wild type) RNA-seq datasets (unpaired t test, t test statistic = 34.65, p = 2.59 × 10⁻²⁰⁶, numbers of replicates = 1,000).

(C) Violin plot shows alzRGs are more likely differentially expressed in disease-associated astrocyte (DAA) according to human prefrontal cortex single-nucleus (GEO: GSE157827; nine normal control and 12 AD human postmortem brain samples) RNA-seq dataset (unpaired t test, t test statistic = 52.01, p = 0.00, numbers of replicates = 1,000).

(D) Violin plot shows alzRGs are more likely differentially expressed in DAA according to human entorhinal cortex (EC) single-nucleus (GEO: GSE138852; six control and six AD human postmortem brain samples) RNA-seq dataset (unpaired t test, t test statistic = 15.18, p = 2.33 × 10⁻⁴⁹, numbers of replicates = 1,000).

(E) Visualization of 32 predicted alzRGs that are also differentially expressed genes (DEGs) according to single-cell/nucleus RNA-seq studies collected from both mouse models and human postmortem brain tissues in two disease-associated immune subtypes, i.e., DAM and DAA (STAR Methods).

We collected and analyzed six sets of proteomics data from transgenic mouse AD models (STAR Methods). Among the 156 predicted AD-associated genes, protein products of 29 genes (e.g., HSPA5, ENO1, FERMT2, and VAV1) are differentially expressed (p = 3.98 × 10⁻⁷, Fisher’s exact test, Table S4). Heat shock protein family A (Hsp70) member 5 (HSPA5) is important in tau phosphorylation and has been proposed as a potential target for AD treatment.⁵⁷ Another study showed that oxidative inactivation of enolase 1 (ENO1) can significantly accelerate AD progression from mild cognitive impairment (MCI).⁵⁸ Additionally, FERM domain containing kindlin 2 (FERMT2) was identified as an AD-risk gene by GWAS studies.¹⁴ It is observed that FERMT2 can modulate APP metabolism and Aβ formation, therefore linking its mechanism with AD pathology.⁵⁹ Another mouse model study found that targeting vav guanine nucleotide exchange factor 1 (VAV1) could rescue neuronal death by inhibiting JNK signaling pathway.⁶⁰

alzRGs are differentially expressed in AD-associated microglia and astrocytes

Neuroinflammation plays a crucial role in pathogenesis and progression of AD.⁶¹ We found that predicted alzRGs are significantly enriched by multiple immune pathways, including B cell receptor, T cell receptor, and cytokine (IL-2, IL-7, and IL18) signaling pathways (Table S3 and Figure S4B). We then investigated how neuroinflammatory pathways were impacted by predicted alzRGs using DAM and disease-associated astrocytes (DAA) as two examples. We found that 14 alzRGs were differentially expressed (∣log₂FC∣ > 0.25, q < 0.05) in DAM from a 5xFAD mouse-model-derived single-cell RNA-seq dataset (one-sided t test: statistic = 33.85, p = 9.76 × 10⁻¹⁸³, Figure S5C) and a 5xFAD mouse-model-derived single-nucleus RNA-seq dataset (one-sided t test: statistic = 34.65, p = 2.59 × 10⁻²⁰⁶, Figure 4B). For DAA, 25 alzRGs are differentially expressed (∣log₂FC∣ > 0.25, q < 0.05) across three single-nucleus RNA-seq datasets from human postmortem brains with several brain regions, including prefrontal cortex (p < 1.0 × 10⁻³, Figure 4C), entorhinal cortex (p = 2.33 × 10⁻⁴⁹, Figure 4D), and super frontal gyrus (p = 8.30 × 10⁻²⁷) (Figure S5D, Table S4, and STAR Methods). Among 39 differentially expressed alzRGs in DAM or DAA, 28 alzRG-coding proteins form a subnetwork within the human interactome network (Figure 4E). Activating transcriptional factor 3 (ATF3) overexpressed in DAM (one top DEGs log₂FC = 0.62, q = 9.73 × 10⁻¹⁶ with 5xFAD mouse models, Table S4), was observed to have elevated expression in another mouse model study.⁶² An elevated expression level of ATF3 is positively correlated with Aβ accumulation.⁶² Microtubule-associated protein RP/EB family member 2 (MAPRE2) overexpressed in DAA (one top DEG with log₂FC = 0.51, q = 7.03 × 10⁻²⁰in human postmortem brain tissues, Table S4) was identified as an AD-associated gene based on GWAS from the ADNI cohort.⁶³ In summary, NETTAG-predicted alzRGs are differentially expressed in AD-associated microglia and astrocytes.

NETTAG-based discovery of potential risk genes in AD

We further identified potential risk genes in AD through combining multiple types of evidence: (1) top predicted genes by NETTAG (156 alzRGs), (2) NETTAG-predicted genes that are supported with at least three types of multi-omics evidence (Table S4), and (3) NETTAG-predicted genes that have not previously been identified by the GWAS Catalog.⁴⁴ In total, we have identified 25 potential AD-risk genes, e.g., CPLX2, FYN, MAPKAPK2, MEF2D, KLF4, P2RX7, BACE1, and HK2. (Figure 5 and Table S4). Myocyte enhancer factor 2D (MEF2D), an alzRG with currently non-existing AD-associated DNA regulatory or GWAS evidence, has the highest NETTAG-predicted score. MEF2D is differently expressed in both mouse single-nucleus RNA-seq and human microarray data analyses (Figures 5 and S5E and Table S4). A previous study showed that protocatechuic acid rescued a cell model from okadaic acid-induced cytotoxicity (tau hyperphosphorylation) by modulating the Akt/GSK-3β/MEF2D pathway, and it also exhibited neuroprotective effects for AD.⁶⁴ Complexin 2 (CPLX2) is the second top predicted alzRG with currently non-existing AD-associated DNA regulatory or GWAS evidence. CPLX2 was identified as AD associated with five types of multi-omics evidence and was differentially expressed according to both transcriptome (DAA) and proteome studies (Figures 5 and S5F and Table S4). Experimental data with hippocampus from 3x-Tg AD mice showed abnormally lower proteomic levels of CPLX2,⁶⁵ which is consistent with the observed downregulation in DAA. Another independent study based on the memory and aging project (MAP) cohort found that lower levels of CPLX2 were significantly associated with cognitive decline.⁶⁶ Together, these potential risk genes by NETTAG identify disease-associated genes and potential drug targets in AD.

Figure 5. Multi-omics observations of 156 prioritized AD-risk genes (alzRGs) by NETTAG.

Summary of multi-omics validations for all 156 predicted alzRGs (Table S3 and S4). The genes are sorted in predicted score decreasing order (clockwise direction). We have collected seven types of evidence, including drug target, differentially expressed genes (DEG) by microarray studies, DEG by bulk RNA-seq studies, DEG in disease-associated microglia (DAM), DEG in disease-associated astrocyte (DAA), DEG by proteome studies, and literature evidence. There are 126 predicted alzRGs that could be proved as associated with AD with at least one type of evidence.

Discovery of repurposable drugs via targeting NETTAG-predicted genes

Among the 156 predicted alzRGs, 38 proteins (gene products of alzRGs) have been identified as known drug targets with FDA-approved or clinically investigational medicines (p = 8.78 × 10⁻⁴, Fisher’s exact test, Table S4 and Figure 5). In total, nine targets (e.g., BACE1, CDK5, FYN, GSK3B, MARK4, MKL1, and PTK2B; Table S4) have been widely investigated as therapeutic approaches for treating AD. FYN proto-oncogene, Src family tyrosine kinase (FYN), which contributes to Aβ production and tau phosphorylation,⁶⁷ has been suggested as one potential target for AD. Glycogen synthase kinase 3 beta (GSK3B) has been found to have a role in hyperphosphorylation of tau and Aβ production.⁶⁸ A previous study showed that thiadiazolidinone (a GSK3B inhibitor) decreased tau phosphorylation and improved neuronal survival.⁶⁸ Beta-secretase 1 (BACE1), a β-secretase enzyme involved in Aβ peptide generation, has also been demonstrated to be a promising target in AD.⁶⁹ We next turned to identify repurposable drugs by specifically targeting protein products of alzRGs (Figure 3A).

Using network proximity approaches⁷⁰ to evaluate the closest distance between disease module and a drug’s targets within the human protein-protein interactome network (STAR Methods), we computationally identified 118 candidate drugs using Z scores (Z) < −2 and false discovery rate (q) < 0.05 from the 2,938 US FDA-approved or clinically investigational drugs (Table S5 and STAR Methods). As shown in Figure 6A, we grouped these top predicted 118 candidate drugs into 14 pharmacological categories based on the first level of the Anatomical Therapeutic Chemical (ATC) code. Gemfibrozil, a lipid regulator used to treat hyperlipidemias, is one of these top predictions (Table S5). Gemfibrozil reduced the amyloid plaque burden in a mouse model of AD.⁷¹ Mechanistically, gemfibrozil’s targets (e.g., LPL, PPARA, and SLCO1B1) have physical interactions with proteins encoded by several predicted alzRGs (e.g., APOE, APP, HSPG2, and TOMM40) (Figure 6B). Cholecalciferol (vitamin D3) is another top predicted candidate (Table S5). Multiple cohort studies found that vitamin D deficiency was associated with elevated risk of AD.^72,73 The drug target network analysis showed that cholecalciferol’s targets (e.g., CDC25A, VDR, and GLRA1) have physical interactions with proteins encoded by several predicted alzRGs (e.g., GSK3B and MAPKAPK2) (Figure 6C). Choline, a nutrient found in many vitamins, is our fifth ranked predicted drug (Table S5). Experiments with APP/PS1 mouse models showed that dietary choline reduced Aβ production and improved spatial memory by suppressing overactivation of DAM.⁷⁴ While choline does not directly target any alzRGs, its targets interact with multiple protein products of predicted alzRGs, including BIN1, CDK5, and FYN (Figure S6A). Ibudilast, an anti-inflammatory drug used to attenuate multiple sclerosis, is another top predicted drug (Table S5). Ibudilast inhibited pro-inflammatory cytokine production and blocked neuroinflammation to prevent Aβ-induced cognitive impairment.⁷⁵ Mechanistically, ibudilast’s targets (e.g., PDE3A, PDE4B, and PDE4D) have physical interactions with proteins encoded by several predicted alzRGs (e.g., BIN1, FYN, and GSK3B) (Figure S6B). Ceftriaxone (Table S5), an antibiotic, was found to be neuroprotective by inhibiting amyloid deposition and neuroinflammation in an AD mouse model.⁷⁶ Ceftriaxone’s targets interact with multiple protein products of predicted alzRGs, such as CSNK2A1 and TOP1 (Figure S6C). Ibuprofen, one commonly used non-steroidal anti-inflammatory drug (Table S5), was capable of suppressing interleukin 1 beta (IL1B) protein level and amyloid deposition in an AD transgenic mouse model.⁷⁷ Ibuprofen’s targets interact with multiple protein products of predicted alzRGs, such as MARK4 and FYN (Figure S6D). In summary, risk genes identified by NETTAG offer potential drug targets for AD therapeutic discovery, including drug repurposing (such as gemfibrozil and cholecalciferol). We next turned to test AD drug outcomes from large-scale patient EHR data using well-established pharmacoepidemiologic approaches.^70,78

Figure 6. Network-based discovery of repurposable drug candidates for AD.

Drug-AD associations were evaluated by the network proximity between predicted alzRGs and drug-target networks.

(A) 118 prioritized drugs for AD treatment (Table S5). Drugs are grouped by fourteen different classes (e.g., immunological, respiratory, neurological, cardiovascular, and cancer) defined by the first level of the Anatomical Therapeutic Chemical (ATC) codes.

(B) Proposed mechanism of actions (MOAs) for gemfibrozil by drug-target network analysis.

(C) Proposed MOAs for cholecalciferol by drug-target network analysis.

Validating possible causal AD associations of candidate drugs in patient data

We used subject matter expertise based on a combination of prespecified factors to select candidate drugs: (1) strength of network proximity measures via quantifying relationships between drug targets and proteins encoded by NETTAG-predicted alzRGs; (2) novelty of the predicted associations through exclusion of drugs currently being tested in AD clinical trials; (3) availability of sufficient patient data for meaningful evaluation (exclusion of infrequently used medications); (4) published positive in vitro or in vivo experimental data in AD models; and (5) ideal brain penetration (STAR Methods). Applying these criteria resulted in identifying four top candidate drugs, including gemfibrozil, cholecalciferol, ceftriaxone, and ibuprofen (Table S5), for patient data analysis. We identified a total of 3,180,111 valid subject records from the Northwestern Medicine Enterprise Data Warehouse (NMEDW) database between 2011 and 2021 after screening 10 million patient records. We then adjusted various confounding factors (including age, race, sex, ethnicity, and disease comorbidities) based on previous studies.^70,78 Overall, we found that use of four drugs was significantly associated with reduced incidence of AD compared with propensity score (PS)-matched non-drug exposure (STAR Methods, Table S6A): (1) ibuprofen vs. PS-matched non-ibuprofen users (odds ratio [OR] = 0.80, 95% confidence interval [CI] 0.74–0.87, p < 0.001); (2) gemfibrozil vs. matched non-gemfibrozil users (OR = 0.76, 95% CI 0.70–0.84, p < 0.001); (3) cholecalciferol vs. matched non-cholecalciferol users (OR = 0.87, 95% CI 0.82–0.91, p < 0.001); and (4) ceftriaxone vs. matched non-ceftriaxone users (OR = 0.86, 95% CI 0.78–0.95, p = 0.003).

We next performed additional active-comparator analyses based on an active-comparator design approach allowing comparison of the agent of interest to an agent used for the same medical indication but not identified as a candidate by NETTAG-predicted alzRGs (STAR Methods): (1) 712,103 patients in the ibuprofen cohort vs. 474,110 patients in an aspirin cohort; (2) 72,691 patients in the gemfibrozil cohort vs. 119,949 patients in a simvastatin cohort; (3) 447,846 patients in the cholecalciferol cohort vs. 235,993 patients in an ergocalciferol cohort; and (4) 91,192 patients in the ceftriaxone cohort vs. 248,724 patients in a ciprofloxacin cohort. Table S6A summarizes clinical characteristics of the patients across each drug cohort. More details for each drug cohort design are provided in Table S6D. Among four active-comparator design analyses, (1) gemfibrozil use was significantly associated with a reduced risk of AD compared with simvastatin (an approved anti-lipid medicine under the phase II AD trials [ClinicalTrials.gov identifier: NCT00053599 and NCT00939822]) users (OR = 0.57, 95% CI 0.51–0.63, p< 0.001); (2) ibuprofen usage was significantly associated with a reduced risk of AD compared with aspirin (a drug failed in a large phase II AD trial⁷⁹) users (OR = 0.78, 95% CI 0.66–0.91, p = 0.002); (3) cholecalciferol was slightly associated with a reduced risk of AD compared with ergocalciferol users (OR = 0.85, 95% CI 0.73–0.99, p = 0.039); and (4) ceftriaxone use was not associated with a reduced risk of AD compared with ciprofloxacin (OR = 1.08, 95% CI 0.94–1.25, p = 0.263).

We further performed sex- and race-specific patient subgroup analyses (Figure 7), and we found that ibuprofen usage was significantly associated with a reduced incidence of AD compared with aspirin users in only male individuals; and ibuprofen usage was slightly associated with a reduced incidence of AD compared with PS-matched non-ibuprofen users in White Americans (OR = 0.82, 95% CI 0.68–0.97, p = 0.023) but not for Black Americans (OR = 0.90, 95% CI 0.56–1.42, p = 0.639). Gemfibrozil usage was significantly associated with a reduced risk of AD compared with simvastatin users and PS-matched non-gemfibrozil users, across all four sex- and race-specific subgroups (OR ranging from 0.49 to 0.58, p ≤ 0.002 for simvastatin users and OR ranging from 0.71 to 0.82, p ≤ 0.04 for PS-matched non-gemfibrozil users). Cholecalciferol usage was associated with reduced AD incidence compared with ergocalciferol users for females (OR = 0.79, 95% CI 0.65–0.94, p = 0.01) and White Americans (OR = 0.80, 95% CI 0.67–0.95, p = 0.01) but not for males (OR = 1.07, 95% CI 0.82–1.39, p = 0.636) and Black Americans (OR = 0.75, 95% CI 0.49–1.16, p = 0.196). Ceftriaxone use was not associated with reduced AD incidence compared with ciprofloxacin users across all four PS-matching subgroups.

Figure 7. Longitudinal patient data observations reveal that usage of four candidate drugs is associated with reduced incidence of Alzheimer’s disease (AD).

Odds ratios (ORs) and 95% confidence intervals (CIs) for eight drug cohort comparator studies are presented. Within each study, besides the overall group analysis (all), we also conducted four additional subgroup analyses by considering sex (female and male) and race (Black and White). Propensity score (PS)-stratified Cox proportional hazards models were used for statistical inference (Fisher’s exact test) of the ORs. For overall group analysis, the patient numbers in each group are ibuprofen (n = 712,103) vs. non-ibuprofen (n = 2,468,008); ibuprofen (n = 712,103) vs. aspirin (n = 474,110); gemfibrozil (n = 72,691) vs. non-gemfibrozil (n = 3,107,420); gemfibrozil (n = 72,691) vs. simvastatin (n = 119,949); cholecalciferol (n = 447,846) vs. non-cholecalciferol (n = 2,732,265); cholecalciferol (n = 447,846) vs. ergocalciferol (n = 235,993); ceftriaxone (n = 91,192) vs. non-ceftriaxone (n = 3,088,919); ceftriaxone (n = 91,192) vs. ciprofloxacin (n = 248,724); Table S6D summarizes more detailed patient counts and event counts in each drug cohort design. Table S6A summarizes detailed clinical characteristics of patients used for each subgroup comparison.

In summary, among the four NETTAG-predicted drugs, these comprehensive pharmacoepidemiologic observations revealed that gemfibrozil was a strong candidate drug for potential prevention and treatment of AD. The association between gemfibrozil use and decreased incidence of AD will require a randomized controlled trial with diverse population to establish the causality.

DISCUSSION

In this study, we presented the development of a deep learning framework known as NETTAG that integrates multi-omics data to infer putative risk genes in AD. To avoid “black-box” deep learning models, we utilized the human protein-protein interactome network to make NETTAG more transparent and interpretable in inferring AD-risk genes from GWAS and multi-omics findings. Specifically, there are 16,720 proteins in the human protein-protein interactome network according to the GO knowledgebase³⁰ (Table S2). We found that 97% (16,214) of annotated proteins in the human interactome have multiple GO terms, ranging from two to around 200 (Table S2 and Figure S7A), and our NETTAG-predicted subnetwork modules were highly correlated with GO-annotated protein function (Figure 2B). We performed additional spearman (r) correlation analyses to evaluate the relationship between NETTAG-predicted scores (cumulative overlays of divided subnetwork modules) and cumulative overlays of protein functions by considering each gene regulatory element and found that deep-learning-predicted scores showed significant correlations across all gene regulatory elements (Figure S1B). For example, CpG islands (r = 0.44, p= 1.86 × 10⁻²⁴), enhancer (r = 0.34, p = 9.07 × 10⁻¹³), histones (r= 0.50, p = 6.49 × 10⁻³⁷), and TF-binding sites (r = 0.36, p = 3.50 × 10⁻¹⁷) show strong correlations, whereas promoter flanking regions (r = 0.26, p = 2.60 × 10⁻¹²), open chromatin (r = 0.22, p = 2.91 × 10⁻¹⁰), promoters (r = 0.20, p = 2.42 × 10⁻¹⁵), CTCFs (r = 0.09, p = 1.45 × 10⁻²), and eQTLs (r = 0.14, p = 8.62 × 10⁻¹⁰) show weak to moderate correlations. In addition, we scored genes according to their clustering similarities with genes associated with each regulatory element as shown in heatmaps (Figure S8). We found that top genes ranked by NETTAG can be attributed to their protein functional overlap with the input proteins associated with each regulatory element. Consistent with correlation analyses (Figure S1), CpG islands, enhancers, histone QTLs, and TF-binding sites play more important gene regulatory roles of GWAS loci in AD compared with promoter flanking regions, open chromatin, promoters, CTCF binding, and eQTLs. We also found that the AUC values for NETTAG-predicted gene scores, generated by integrating a total of nine genome regulatory elements, are 23% (AlzGene), 23% (DistiLD), 18% (DISEASES-knowledge), and 17% (TIGA) higher than the average AUC value when considering each single gene regulatory element alone (Figures S2B-S2F). Taken together, NETTAG offers an interpretable deep learning framework to identify AD-risk genes from GWAS loci by leveraging multi-omics findings.

To evaluate effects of modifications implemented in NETTAG (STAR Methods), we compared NETTAG with three additional experiments: (1) without any modifications, (2) subsampled topological similar subgraphs (Modification 2 in STAR Methods), and (3) coupling second-order term (Modification 1 in STAR Methods). Each group contained 10 random experiments with fixed seeds. We evaluated the performance of four experiments using the GWAS Catalog for validation, and we found superior efficacy of NETTAG with either modification, compared with the other groups (Figure S7C). In detail, we compared NETTAG with existing network-based approaches, including RWR, spectral clustering, and k-means (STAR Methods). We found that the percentage of predicted genes (not overlapped with input genes) by NETTAG was much higher compared with RWR (Figure S3A and Table S3). In addition, clustering approaches implemented in NETTAG outperformed classical spectral and k-means clustering methods (Figures S3C and S3D) as well. This indicates that overlapping clustering-based approaches could better coincide with the multiple biologic functions of any given protein.

We found that NETTAG-predicted drugs could guarantee consistency with forecasts from other methodologies. For example, pioglitazone was proposed as one potential AD treatment from another artificial intelligence-based approach.⁸⁰ Sildenafil was suggested as promising AD treatment based on Aβ-tau synergistic endophenotype study.⁷⁸ Our approach also stably highlights both drugs as top-ranked AD candidate options (Table S5). Combining network-based prediction and large-scale patient data observations, we identified four candidate drugs for treating AD: gemfibrozil, cholecalciferol, ceftriaxone, and ibuprofen. The beneficial effect of ceftriaxone with AD was demonstrated with both Aβ (APP1)⁸¹ and tau (3xTg-AD)⁸² mouse models as well. In both APP1⁸¹ and 3xTg-AD mouse models,⁸² ceftriaxone restored glutamate-transporter-1 (GLT-1) expression levels and ameliorated cognitive decline. Ibuprofen’s treatment effects with AD were consistent with a mouse model study that ibuprofen could inhibit neuroinflammation and amyloid deposition.⁷⁷

Gemfibrozil is a lipid regulator used to treat hyperlipidemia.⁷¹ Recent mouse (5xFAD) model studies have shown that oral gemfibrozil treatment reduces amyloid plaque pathology.⁷¹ The same team found that gemfibrozil reduced neuroinflammation by inhibiting microgliosis and astrogliosis.⁷¹ We selected gemfibrozil to test its possible association with AD using a large-scale patient longitudinal database (STAR Methods). With PS-matching, we found that gemfibrozil was significantly associated with a decreased risk of AD in the general population group after adjusting various cofounding factors, including age, gender, race, ethnicity, MCI, hypertension (HT), type-2 diabetes (T2D), and coronary artery disease (CAD) diagnoses, based on our sizable efforts (Figure 7). Compared with non-gemfibrozil users, gemfibrozil was significantly associated with reduced AD across all four subgroup (i.e., sex and race) analyses as well (Figure 7). The active-comparator design analyses showed that gemfibrozil was associated with greater risk reduction for AD compared with simvastatin in both general and all subgroup studies (Figure 7). Simvastatin is an HMG-CoA reductase inhibitor used for cardiovascular-related disease treatment, which is under several phase II AD trials (ClinicalTrials.gov identifier: NCT01439555 and NCT00486044). Altogether, gemfibrozil offers a promising candidate drug to be tested in future randomized controlled trial in AD patients with diverse populations to establish the causality.

In summary, we established a deep learning framework (NETTAG) that incorporates multi-genomic information along with human protein interactome to infer AD-associated genes. We showed that predicted genes are enriched with drug targets, differently expressed in DAM and astrocytes, and most importantly significantly associated with AD. We demonstrated that the predicted genes offer potential targets for drug repurposing, and we validated one NETTAG-predicted drug, gemfibrozil, in reducing risk of AD using large-scale, longitudinal patient data. We believe that the NETTAG presented here, if broadly applied, could significantly catalyze innovation in drug discovery for AD and other neurodegenerative diseases.

Limitations of the study

We acknowledge several potential limitations in this study. For example, incompleteness of the human protein-protein interactome and GWAS loci by limited population size may influence the model performance. More brain-specific functional genomics data should be integrated in the future studies,⁸³ and the latest GWAS data (Bellenguez et al., 2022)¹⁰ should be integrated in the future. While considering all protein pairs sharing at least one common GO annotation, we found that shortest path-based distances among those protein pairs are 2 (53.97%), 3 (39.05%), and 4 (5.05%) in decreasing order (Figure S7B). Therefore, in NETTAG, we have two hidden layers to indirectly aggregate features from second-order neighbors. We postulate that reformulating the model to sample second- and third-order neighbors directly may improve the model performance further.

The EHR-based observational studies may have limitations. First, since true drug administration data were not available, for validation of the NMEDW data, we considered the patients that had been prescribed the drugs. Although patients are likely to adhere to their prescriptions, this is not always the case. Second, pharmacoepidemiologic studies could be biased due to unavailable confounding factors. Although we adjusted for age, sex, and disease comorbidities, including CAD, HT, T2D, and MCI, other factors not included in the database may also be associated with risk of AD, such as education level, socioeconomic status, and genotyping information (such as APOE).^84-87 Finally, EHR-based observational studies cannot build causal relationships between drug use and beneficial clinical response of AD.^78,88 Our results therefore warrant rigorous clinical trial testing of the treatment efficacy in patients with AD, inclusive of both sexes and controlled by placebo.

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and software should be directed to and will be fulfilled by the lead contact, Feixiong Cheng (chengf@ccf.org).

Materials availability

This study did not generate new materials.

Data and code availability

• This paper analyzes existing, publicly available data. The accession numbers for the datasets are listed in the key resources table.

• The Python implementation of NETTAG is publicly available at https://github.com/ChengF-Lab/NETTAG and archived on Zenodo (https://doi.org/10.5281/zenodo.7159120). The Github repository contains the regulatory element data and NETTAG default parameters used in the manuscript.

• Any additional information required for this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
GWAS Catalog	Buniello et al.44	https://www.ebi.ac.uk/gwas/
SNPnexus	Oscanoa et al.28	https://www.snp-nexus.org/v4/
GTEx	Lonsdale et al.24; GTEx Consortium25	https://gtexportal.org/home/datasets
Ensembl Regulatory Build	Zerbino et al.27	https://grch37.ensembl.org/index.html
ENCODE	Kundaje et al.29	https://www.encodeproject.org/
NIAGADS	NA	https://www.niagads.org
The Alzheimer’s disease Knowledge portal	NA	https://adknowledgeportal.synapse.org/
Human protein-protein interactome	Zhou et al.89	https://alzgps.lerner.ccf.org
Gene Ontology	Mi et al.30	http://geneontology.org
AlzGene	Bertram et al.38	http://www.alzgene.org
DistiLD	Palleja et al.39	http://distild.jensenlab.org
TIGA	Yang et al.40	https://unmtid-shinyapps.net/shiny/tiga/
DISEASES	Pletscher-Frankild et al.41	https://diseases.jensenlab.org/Search
Human microarray: 31 late-stage AD and 32 controls	Miller et al.90; Zhou et al.89	GEO: GSE29378; https://alzgps.lerner.ccf.org
Human microarray: 42 late-stage AD and 173 controls	Berchtold et al.91; Zhou et al.89	GEO: GSE48350; https://alzgps.lerner.ccf.org
Human microarray: 328 late-stage AD and 214 controls	Wang et al.92; Zhou et al.89	GEO: GSE84422; https://alzgps.lerner.ccf.org
Human bulk-RNA seq: 4 late-stage AD and 4 controls	Zhou et al.89; Magistri et al.93	https://alzgps.lerner.ccf.org
Human bulk-RNA seq: 6 late-stage AD and 6 controls	Zhou et al.89; Annese et al.94	https://alzgps.lerner.ccf.org
Human bulk-RNA seq: 20 late-stage AD and 10 controls	Zhou et al.89; van Rooij, et al.95	https://alzgps.lerner.ccf.org
Mouse single-cell RNA seq: 6-month-old 16 5xFAD and16 WT mouses	Keren-Shaul et al.96	GEO: GSE98696; http://taca.lerner.ccf.org/
Mouse single-nucleus RNA seq: 7-month-old 3 TREM2-KO WT, 3 TREM2-KO 5xFAD, 3 5xFAD, and 3 WT mouses	Zhou et al.97	GEO: GSE140511; http://taca.lerner.ccf.org/
Human single-nucleus RNA seq: 3 Braak stage 0, 4 Braak stage and 3 Braak stage brain samples	Leng et al.98	GEO: GSE147528; http://taca.lerner.ccf.org/
Human single-nucleus RNA seq: 6 AD and 6 controls	Grubman et al.99	GEO: GSE138852; http://taca.lerner.ccf.org/
Human single-nucleus RNA seq: 12 AD and 9 controls	Lau et al.100	GEO: GSE157827; http://taca.lerner.ccf.org/
Mouse proteomic data (7- and 10-month-old ADLP mouse models [JNPL3 mouse model cross with 5xFAD mouse model])	Zhou et al.89; Kim et al.101	https://alzgps.lerner.ccf.org
Mouse proteomic data (7- and 10-month-old 5xFAD mouse model)	Zhou et al.89; Kim et al.101	https://alzgps.lerner.ccf.org
Mouse proteomic data (12-month-old 5xFAD mouse model)	Zhou et al.89; Savas et al.102	https://alzgps.lerner.ccf.org
Mouse proteomic data (12-month-old hAPP mouse model)	Zhou et al.89; Savas et al.102	https://alzgps.lerner.ccf.org
Software and algorithms
NETTAG	This paper	https://github.com/ChengF-Lab/NETTAG
diffusr R package	Valdeolivas et al.23	https://github.com/dirmeier/diffusr
Seurat R package	Butler et al.103	https://github.com/satijalab/seurat
MAST R package	Finak et al.104	https://github.com/RGLab/MAST
scikit-learn Python package: spectral clustering, k-means, SVD	Pedregosa et al.105	https://scikit-learn.org/stable/index.html
Enrichment analysis	Chen et al.106; Kuleshov et al.107	https://maayanlab.doud/Enrichr/

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Patient electronic health records (EHR) database

The de-identified patient EHR database is from Northwestern Medicine Enterprise Data Warehouse (NMEDW) which covers clinical and medical data from approximately 10 million individuals. The more detailed information is in the method details - pharmacoepidemiologic methods section.

METHOD DETAILS

Construction of multi-genomic features

In this study, we collected 1,047,489 SNPs across multiple genetic traits from GWAS catalog,⁴⁴ such as Alzheimer’s disease, cerebral amyloid deposition measurement. Next, we performed web server SNPnexus²⁸ to annotate all SNPs in human genome (GRCh38) and collected the regulatory elements information from five databases, including CpG Islands,²⁸ Ensembl Regulatory Build,²⁷ ENCODE,²⁹ the Genotype-Tissue Expression (GTEx) portal^24,25 and Roadmap.²⁶ Finally, nine regulatory elements (histone, open chromatin, CpG Island, TF, CTCF, eQTL, enhancer, promoter and promoter flanking region) were used as features to evaluate AD disease-associated genes. To be more specific, step 1: for SNPs with respect to each regulatory elements, e.g., CpG island, we merge them with SNPs curated by GWAS Catalog⁴⁴ with AD as the mapped traits. Step 2: For each AD related SNPs, the corresponding genes were identified by the “MAPPED GENE(S)” column as provided by GWAS Catalog.⁴⁴ For SNPs with no mapped genes, if there were any reported genes (REPORTED GENE(S) column in GWAS Catalog⁴⁴) associated with this SNP, we then map the SNP to its reported genes. For ENCODE²⁹ and Ensembl Regulatory Build²⁷ databases, we only consider epigenomes from brain and neuron tissues and normal karyotype. For eQTLs from GTEx,^24,25 we only consider those from human brain tissues. When mapping eQTLs with GWAS Catalog, we only counted significant eSNP (q < 0.05, LD r² < 0.1) associated eGenes as input features for downstream analysis.^24,25 To enhance the statistical analysis from eQTL feature, all eGenes were counted as eQTL features when an eSNP associated with multiple genes. However, we only counted one most significant eQTL feature when an eGene associated with multiple SNPs. The specific epigenomes with brain and neurons for each database are presented in Table S1, separately. The final mapped genes for each regulatory element are provided in Table S1.

Building the human protein-protein interactome

To build the comprehensive human interactome from the most contemporary data available, we assembled commonly used PPI databases with experimental evidence and in-house systematic human PPIs: (i) binary PPIs assessed by high-throughput yeast-two-hybrid (Y2H) experiments¹⁸; (ii) kinase-substrate interactions via literature-derived low-throughput and high-throughput experiments from Human Protein Resource Database (HPRD),¹⁰⁸ DbPTM 3.0¹⁰⁹, Phospho.ELM,¹¹⁰ KinomeNetworkX,¹¹¹ PhosphoNetworks,¹¹² and PhosphositePlus¹¹³; (iii) binary PPIs from 3D protein structures from Instruct¹¹⁴; (iv) signaling networks by literature-derived low-throughput experiments from the SignaLink2.0¹¹⁵; (v) protein complex data (~56,000 candidate interactions) identified by a robust affinity purification-mass spectrometry collected from BioPlexV2.0¹¹⁶; and (vi) literature-curated PPIs identified by affinity purification followed by mass spectrometry from HPRD,¹¹⁷ PINA,¹¹⁸ MINT,¹¹⁹ InnateDB,¹²⁰ IntAct,¹²¹ and BioGRID.¹²² In total, 351,444 PPIs connecting 17,706 distinctive proteins are now freely available at https://alzgps.lerner.ccf.org. In this study, we consider only the largest connected components of this dataset, which includes 17,456 proteins and 336,549 PPIs.

Description of NETTAG

NETTAG involves 3 steps. Step 1: we build up a graph neural network (GNN) model to capture PPI’s topology structure and establish the appropriate overlapping clustering. The GNN model is motivated by NOCD-G¹²³ which is one GNN-based overlapping community detection framework. And in NETTAG, we made 2 modifications with respect to NOCD-G.¹²³

Modification 1: The model architecture used in NETTAG is defined below (Equation 1)

$$F:=ReLU\left[GC{N}_{{\Theta }_2}^K\left(A,ReLU\left[GC{N}_{{\Theta }_1}^2(A,X)\right]\right)\right],\text{with}GC{N}_{{\Theta }_l}^2=\sum _{k=0}^2{\widehat{A}}^{{}^k}X{W}^{k,l}$$ (Equation 1)

Here: ${\widehat{\mathit{A}}}^{{}^0}=\mathit{I}$, ${\widehat{\mathit{A}}}^{{}^1}=\widehat{\mathit{A}}={\mathit{D}}^{{}^{-1∕2}}\mathit{A}{\mathit{D}}^{{}^{-1∕2}}$, ${\widehat{\mathit{A}}}^{{}^2}$ is the elementwise square of $\widehat{\mathit{A}}$ and the sub/super-script l is the layer index. A is the adjacency matrix of the PPI, D is the corresponding diagonal degree matrix. X in generate denotes the node feature matrix, and here we set X = A as implemented by NOCD-G.¹²³ In classical graph convolution network (GCN) models,¹²⁴ nodes with various degrees share the same weight matrix, the normalized matrix $\widehat{\mathit{A}}$ could alleviate this. In NETTAG, coupling ${\widehat{\mathit{A}}}^{{}^2}$ into the model is to enhance this, i.e., to strengthen different weight matrices for nodes with high and low degrees. Finally, the dimension of the final output layer interprets the clustering number.

The output matrix F is then feed into Bernoulli-Poisson model¹²⁵ (Equation 2.1) to learn PPI’s topology (Equation 2.2). The output matrix F has N (number of total nodes in PPI) rows and C (clustering numbers) columns. Locally minimal communities¹²⁶ estimate the lower bound of the clustering number and we tested different clustering number with increment 100. The final clustering number was determined as the one with lowest akaike information criterion (AIC) value (Equation 2.3, Figure S7F). Each specific row (vector) denotes the node’s weights for being assigned to each cluster. Therefore, we can interpret the loss as follows: if two nodes have multiple commonly shared clusters ($\overrightarrow{F_u}\cdot \overrightarrow{F_v^T}$ is large), then there should exist an edge connecting each other ($1-\text{exp}(\overrightarrow{F_u}\cdot \overrightarrow{F_v^T})$ is close to 1) and vice versa.

$$P_{uv}(c)=1-exp(-F_{uc}\cdot F_{vc})$$ (2.1)

$$(Loss∕-Loglikelihood)L=-\sum _{(u,v)\in E}\log \left(1-\text{exp}\left(\overrightarrow{F_u}\cdot \overrightarrow{F_v^T}\right)\right)+\sum _{(u,v)\notin E}\overrightarrow{F_u}\cdot \overrightarrow{F_v^T}$$ (2.2)

$$AIC=\frac{L}{N\times (N-1)∕2}+\frac{C}{N}$$ (2.3)

Modification 2: The PPI network is a highly sparse network which implies number of connected edges is far less than that of non-connected edges (For our PPI, the ratio between number of connected edges over number of non-connected edges = 1/451). To address this imbalanced training problem, the authors in¹²³ first uniformly subsampled certain amounts of connected edges and then subsampled the equal amounts of non-connected edges for training. Instead of uniformly sampling edges directly, we first group nodes according to their degrees into multiple bins. And in each training iteration, we first uniformly subsample same amounts of nodes from each bin, then we extract an adjacency matrix A_sub which compromising only selected nodes. In this way, we can keep the topology similarities among sampled subgraphs from different iterations. Next, we use the connected and non-connected edges in A_sub to compute the train loss, and the rest connected and non-connected edges in A A_sub for test loss evaluation. With this graph subsampling scheme, we find that we are capable to maintain similar connected and non-connected training edge percentages (Figure S7G).

After learning the clustering affinity matrix F, we used the threshold defined in¹²⁷ as the cutoff for F to determine the node (gene) clustering membership (Equation 3.1).

$$F_{ij}^{\ast }=\{\begin{array}{cc}1\hfill & \text{if}{F}_{ij}\ge \sqrt{-\log (1-1∕N)}\hfill \\ 0\hfill & \text{Otherwise}\hfill \end{array}\phantom{\}}$$ (Equation 3.1)

The rationality of the threshold can be explained from the Bernoulli-Poisson model (Equation 2.1). In detail, we have:

$$\phantom{\{}\begin{array}{c}\hfill F_{uc}\ge \sqrt{-\log (1-1∕N)}\hfill \\ \hfill F_{vc}\ge \sqrt{-\log (1-1∕N)}\hfill \end{array}\}\Rightarrow P_{uv}(c)=1-exp(-F_{uc}\cdot F_{vc})\ge \frac{1}{N}$$ (Equation 3.2)

This calculation indicates that a given node could be designated as belonging to a given cluster because it was connected with one or more nodes in the same cluster, rather than being randomly selected alone.

Step 2 of NETTAG: After clustering the entire PPI into overlapping communities, we next scored each node with respect to each gene regulatory element. The detailed steps are described below: Step 2.1. Compute node (gene, e.g., ‘APP’) score regarding one particular regulatory element $S_{feat}^i=\sum _g\frac{\mid C^i\cap C_{feat}^g\mid }{\mid C_{feat}^g\mid }$, here Cⁱ denotes the cluster assignments of node i, $C_{feat}^g$ denotes the cluster assignments of gene g identified with this regulatory element. The node score equals the sum of clustering overlaps with all genes identified regarding this regulatory element. ∣feat∣ denotes the numbers of genes associated with this regulatory element.

Step 2.2 Construct background distribution to evaluate the significance for the score computed in Step 2.1. For node (gene, e.g., ‘APP’), we run 1000 random experiment, and in each random experiment, we do two things as listed below:

1. Select ∣feat∣ of genes randomly from Gene_PPI {feat∪i}

2. $$\text{Compute}{S}_{rand}^i=\sum _g\frac{\mid C^i\cap C_{rand}^g\mid }{\mid C_{rand}^g\mid }$$ (Equation 4)

Compute the Z score for $S_{feat}^i$ as $Z_{feat}^i=\frac{S_{feat}^i-{\mu }_i}{{\sigma }_i}$, with ${\mu }_i=\text{mean}\{S_{rand}^i\}$ and ${\sigma }_i=\text{std}\{S_{rand}^i\}$, and the corresponding p value ($P_{feat}^i$) with respect to a standard normal distribution. The final node (gene, e.g., ‘APP’) score regarding the current regulatory element after considering statistical significance is:

$${\widehat{S}}_{feat}^{{}^i}=\{\begin{array}{cc}{S}_{feat}^i\hfill & \text{if}{P}_{feat}^i<0.05\hfill \\ 0\hfill & \text{Otherwise}\hfill \end{array}\phantom{\}}$$ (Equation 5)

Here, it means we consider the gene score to be meaningful only if its clustering overlap with input genes is statistically significant.

Step 2.3. integrating multiple regulatory elements

The final node (gene, e.g., ‘APP’) score after integrating all regulatory elements is:

$$S_i=\sum _{feat}{\widehat{S}}_{feat}^{{}^i}$$ (Equation 6)

Step 3of NETTAG: After inferring the nodes’ score, we extracted the disease module by considering only genes inferred as significantly associated with AD. We collected all positive gene scores {S_i > 0}, and computed the mean μ, standard deviation σ and the Z score. We considered genes with p value ≤0.01 as the predicted AD associated genes and mapped those genes to the background PPI network to generate disease module.

Other existing network-based approaches for risk genes prediction

Random walk with restart (RWR)

We used R package ‘diffusr’ to implement risk gene prediction based on RWR.²³ We used genes associated with each regulatory element as the starting distribution for RWR, and gene scores regarding each specific regulatory element equaled to their stationary distribution. The integrated gene scores, same as NETTAG, equaled to summation after considering scores with all gene regulatory elements. We prioritized genes predicted by RWR with the same Z score cutoff as with NETTAG. We conducted disease enrichment analyses using GWAS Catalog⁴⁴ and DisGeNET⁴⁶ (Enrichment analysis).

Spectral clustering.

We used the function ‘SpectralClustering’ from Python package ‘scikit-learn’ to implement risk gene prediction based on spectral clustering.⁴² We used the adjacency matrix of our PPI as the precomputed affinity matrix, and we set the clustering number to be 1200 because this was the same as we set with NETTAG. The subsequent steps of gene scoring were the same as NETTAG (STAR Methods - Step 2 of NETTAG).

K-means.

We used function ‘KMeans’ from Python package ‘scikit-learn’ to implement risk gene prediction based on k-means.⁴³ Considering the curse of dimensionality, we performed singular value decomposition (SVD)¹²⁸ with function ‘TruncatedSVD’ from Python package ‘scikit-learn’ to the adjacency matrix of our PPI first. The selection of desired dimensionality (parameter ‘n_components’) was described in Table S3. The reduced version of adjacency matrix was then used as the input for k-means clustering. We set the clustering number to be 1200 because this was the same as we set with NETTAG. The subsequent steps of gene scoring were the same as NETTAG (STAR Methods - Step 2 of NETTAG).

Network proximity for drug prediction

We assembled drugs from the DrugBank database relating 2,938 FDA-approved drugs or clinically investigated molecules.¹²⁹ To predict drugs using the extracted disease module from NETTAG, we adopted the closest-based network proximity measure⁷⁰ shown below.

$${\mathrm{d}}_{closest}(\mathrm{X},\mathrm{Y})=\frac{1}{\Vert \mathrm{X}\Vert +\Vert \mathrm{Y}\Vert }\left(\sum _{x\in X}\underset{y\in Y}{\text{min}}d(x,y)+\sum _{y\in Y}\underset{x\in X}{\text{min}}d(x,y)\right)$$ (Equation 7)

where d(x,y) is the shortest path length between protein x and y from protein sets X and Y, respectively. In our work, X denotes the disease module from NETTAG, and Y denotes the drug targets (protein/gene set) for each drug. To evaluate whether such proximity was significant, the computed network proximity is transferred into Z score form as shown below:

$$Z_{d_{closest}}=\frac{d_{closest}-{\mu }_d}{{\sigma }_d}$$ (Equation 8)

Here, μ_d and σ_d are the mean and standard deviations of permutation tests with 1,000 random experiments. In each random experiment, two random subnetworks X_r and Y_r are constructed with the same numbers of nodes and degree distribution as the given 2 subnetworks X and Y, separately, in the PPI network. This reduces any literature bias associated with well-studied proteins.

Pharmacoepidemiologic methods

Dataset source

The data for validation is from Northwestern Medicine Enterprise Data Warehouse (NMEDW),¹³⁰ which contains medical and clinical data from approximately 10 million patients across 11 hospitals in Illinois.¹³¹ We extracted the data from NMEDW between 2011 and 2021 for analysis.

Study design

We evaluated the effect of four target drugs on AD, i.e., Ibuprofen, gemfibrozil, cholecalciferol and ceftriaxone. For each drug, to minimize confounding factors, we conducted two types of drug cohort-based observational studies: (1) users versus non-users of a certain drug, denoted as drug vs. non-drug, such as Ibuprofen vs. non-Ibuprofen; (2) active-comparator design¹³² that constituted users of a drug versus users of a comparator drug with similar U.S. FDA-approved indications. The target drugs and comparator drugs we used are listed in Table S6B. The primary outcome variable was incidence of AD defined by the International Classification of Diseases (ICD-9/ICD-10) codes (Table S6C).

For each comparison, we executed 3 steps to generate the final cohort, i.e., data extraction, filtering and matching. For drug vs. non-drug comparisons, in the data extraction step, we extracted all the patients who had a drug order record from NMEDW for the target drug. The earliest drug order date was recorded as the index date based on the new-user design.¹³² All other patients in the NMEDW database were extracted from the non-drug group. In the filtering step, we excluded patients who had been diagnosed with AD before the index date for the target drug group. Patients in both groups whose age was missing were also excluded. Finally, in the matching step, we performed a propensity score matching algorithm¹³³ to mitigate potential confounding variables, including age, gender, race, and ethnicity, as well as diagnosis of mild cognitive impartment (MCI), hypertension (HT), type 2 diabetes (T2D), and coronary artery disease (CAD). All diagnoses were defined by ICD9/10 codes, which are listed in Table S6C. In propensity score matching,^134-136 the propensity score was estimated by fitting a logit model in which the outcome variable was a binary variable indicating whether the target drug was used, and the predictors were the potential confounding variables of age, gender, race, ethnicity, and diagnosis of MCI, HT, T2D, and CAD. By propensity score matching, each patient in the target drug group was paired with a patient from the non-drug group with a similar propensity score.

For drug vs. comparator drug comparisons, in the data extraction step we extracted all patients who had been prescribed the corresponding drug. According to the new-user design (Ho et al., 2011,2007; Stuart, 2010), we indicated the earliest drug order date as the index date for both the target group and the comparator. In the filtering step, besides excluding patients diagnosed with AD before the index date and patients with missing age, we also excluded patients who had been prescribed both the target drug and the comparator drug. Finally, in the matching step, we applied a propensity score matching algorithm with confounding variables, including age, gender, race, ethnicity, and diagnosis of MCI, HT, T2D, and CAD. For comparisons in which the target drug group had more patients than the corresponding comparator group, we matched each patient in the comparator group to a patient with similar confounding factors from the target group. In our dataset, for the comparison between ibuprofen and its comparator drug aspirin, we identified more younger patients and fewer older patients, causing failure of the propensity score matching with age. Thus, we excluded young patients with age<65 in the filtering step.

Collections of known AD-associated genes

We collected AD-associated genes from four databases, including AlzGene, DistiLD, TIGA and DISEASES. AlzGene³⁸ collected AD-associated genes via genetic association studies. Thirty-two genes supported by genetic evidence are collected from AlzGene. DistiLD made the existing GWAS studies easier for accessing disease-associated SNPs and genes.³⁹ We collected 19 genes (p < 5.0 × 10⁻⁸) with AD GWAS from DistiLD. TIGA⁴⁰ was a web application that helps drug discovery scientists to prioritize targets by leveraging gene-trait association with multiple studies and evidence. Fifty-one genes with scores greater than 80 defined by TIGA specifically for AD (EFO_0000249) were collected. DISEASES⁴¹ was a text-mining founded database that collected disease-gene associations with manual curations, cancer mutation data and existing databases. We picked up twenty-seven genes associated with AD (DOID:10,652) which were curated from multiple sources, including MedlinePlus, AmyCo, UniProtKB-KW. The complete lists of genes we used for ROC analyses are presented in Table S2.

Gene Ontology

All proteins’ gene ontology annotation (human, gaf-version 2.1) are extracted from The Gene Ontology (GO) Knowledgebase.³⁰ For our PPI with 17,706 proteins, there are 16,736 proteins with total 268,241 GO annotation.

QUANTIFICATION AND STATISTICAL ANALYSIS

Alignments between clustered subnetwork modules and protein functions

With respect to each clustered subnetwork modules, we considered all possible protein-protein pairs by counting the protein-protein pairs that shared at least one common GO annotation, and the protein-protein pairs that shared no common GO annotation. We then conducted a paired statistical test with null hypothesis of proteins in the same clustered subnetwork modules having no protein functional similarities and the alternative hypothesis that the proteins in the same clustered subnetwork module possessed common protein functions. We used Wilcoxon signed rank test with R (4.2.0). For each cluster and each possible protein pair we checked whether they shared common GO annotation. Therefore, for each cluster, we could have 2 values: a) numbers of protein pairs that share at least one common GO annotation, and b) numbers of protein pairs that share no common GO annotation. Two groups of data (Table S2) were generated after considering all clusters, and we performed the paired Wilcoxon signed rank test.

Differential expression analyses of brain transcriptomic data

Transcriptome analyses were performed based on microarray, bulk RNA-seq, and single-cell/nucleus (sc/sn) RNA-seq datasets. We utilized three sets of human brain microarray transcriptome data collected from late-stage AD and control donors. The original data are available from Gene Expression Omnibus database,¹³⁷ i.e., (GEO:GSE29378 with 31 late-stage AD and 32 controls),⁹⁰ (GEO:GSE48350 with 42 late-stage AD and 173 controls),⁹¹ and (GEO:GSE84422 with 328 late-stage AD and 214 controls),⁹² and the results are available from our recently developed AD knowledgebase AlzGPS.⁸⁹ We also included human brain bulk RNA-seq transcriptome data collected from the hippocampus of late-stage AD and control donors with three studies, including 4 late-stage AD versus 4 controls,⁹³ 6 late-stage AD versus 6 controls,⁹⁴ and 20 late-stage AD versus 10 controls.⁹⁵ The complete sc/sn RNA-seq datasets used for differentially expressed genes (DEGs) analyses in this study are available from the Gene Expression Omnibus database¹³⁷ under accession numbers: GSE98969,⁹⁶ GSE140511,⁹⁷ GSE147528,¹³⁸ GSE138 8 52⁹⁹ and GSE157827.¹⁰⁰ The DEGs between DAM and non-DAM were based on mouse scRNA-seq datasets GEO:GSE98969 and mouse snRNA-seq dataset GEO:GSE140511. The DEGs between DAA and non-DAA were based on the other three human snRNA-seq datasets GEO:GSE147528, GEO:GSE138852 and GEO:GSE157827. We used DEG analytical pipeline in our previous study⁸⁸ for datasets GEO:GSE98969, GEO:GSE140511, GEO:GSE147528 and GEO:GSE138852. For the additionally added dataset GEO:GSE157827 which contained 21 prefrontal cortex tissue postmortem samples with 12 diagnosed as AD and 9 controls, we performed bioinformatics analyses according to the parameters used in the original manuscript¹⁰⁰ with Seurat (4.0.6)¹⁰³. Nuclei with ≤200 genes and ≥20,000 unique molecular identifiers and ≥20% mitochondrial genes were removed. Then, the raw count was log-normalized and the top 1000 most variable genes were detected with the function FindVariableFeatures with selection.method = ‘Vst’. Next, all 21 samples were integrated by functions FindIntegrationAnchors and InegrateData with parameter dims = 1:20. Principal component analysis (PCA) was performed with parameter npcs = 50, and clustering was performed with the first 20 pcs and resolution 1. After identifying astrocytes with marker genes (ADGRV1, GPC5 and RYR3) provided by the original manuscript,¹⁰⁰ we continued with the subcluster analysis. We ran the subcluster with the first 20 pcs and resolution 0.3. Next, we identified disease associated astrocytes (DAA) and homeostasis astrocytes by established marker genes.¹³⁸ DEGs were calculated between DAA and HAA with MAST R package.¹⁰⁴ For all DEGs generated from sc/sn RNA-seq datasets, we applied uniform criterion with q < 0.05 and ∣log₂FC∣ ≥ 0.25. We have also included the complete DEG results based on microarray, bulk RNA-seq, sc/sn-RNA-seq aforementioned in Table S4.

Differential expression analyses from brain proteomic data

Differential expression analyses of brain proteomic data were conducted based on six mouse model datasets: (1) 7-month old ADLP mouse models (JNPL3 mouse model cross with 5xFAD mouse model),¹⁰¹ (2) 10-month old ADLP mouse models (JNPL3 mouse model cross with 5xFAD mouse model),¹⁰¹ (3) 7-month old 5xFAD mouse models,¹⁰¹ (4) 10-month old 5xFAD mouse models,¹⁰¹ (5) 12-month old 5xFAD mouse models,¹⁰² and (6) 12 month old hAPP mouse models.¹⁰² We obtained the differentially expressed proteins (DEPs) for each mouse brain proteomic dataset from the supplemental tables of the original manuscripts for the first four datasets¹⁰¹ with p value < 0.05 regarding the Student’s t-test, and last two datasets¹⁰² with FDR <0.05. All lists of DEPs across different mouse brain datasets are available from AlzGPS.⁸⁹ The complete lists of DEPs are provided in Table S4.

Enrichment analysis

All pathway and disease enrichment analyses were conducted using WikiPathways,¹³⁹ GWAS Catalog 2019,⁴⁴ UK Biobank GWAS v1⁴⁵ and DisGeNET⁴⁶ from Enrichr,^106,107 respectively. The combined score defined in Enrichr¹⁰⁶ equaled the product of log of p value from the Fisher’s test and Z score which characterized the departure from the expected rank. The GWAS Catalog⁴⁴ contributes as a freely and easily assessable database of SNP-trait associations which were identified by literatures. The DisGeNET⁴⁶ platform integrates information about disease-associated genes. Specifically, it assembles disease-associated genes from several sources: expert curated repositories, animal models, and the scientific literatures. Our prediction was enriched with 43 AD-associated genes from DisGeNET (Alzheimer’s disease [C0002395]).

Highlights.

• NETTAG identifies Alzheimer’s disease (AD) risk genes from multi-omics and network data

• NETTAG-predicted risk genes are differentially expressed in AD brains and microglia

• NETTAG-predicted AD-risk genes are enriched in known druggable targets

• Large-scale patient health record data support gemfibrozil use for reduced AD risk