Systems genetic dissection of brain gene expression reveals excitotoxic mechanisms of Alzheimer’s disease

Abstract

Gene expression changes likely mediate the impact of Alzheimer’s disease (AD) neuropathology on cognition, but there are challenges to resolve the proximal causal pathways from postmortem brain transcriptome profiles which lack temporal resolution and are further confounded by mixed pathologies. Here, we functionally dissect 30 AD-associated human brain gene co-expression modules using fruit fly (Drosophila melanogaster) models. Integrating longitudinal RNA-sequencing and behavioral phenotyping, we interrogated the consequences of amyloid beta (Aβ) plaques, tau neurofibrillary tangles, and aging, highlighting hundreds of conserved, differentially expressed genes. To pinpoint causal modules and drivers, we manipulated 344 prioritized targets in vivo, identifying 141 modifiers of Aβ- or tau-induced neurodegeneration. We discovered an upregulated immune module enriched for AD risk variants that promotes neurodegeneration based on genetic manipulations in neurons. By contrast, a downregulated human brain synaptic regulatory network includes many loss-of-function suppressors of Aβ/tau and modulates glutamatergic hyperexcitation injury. Additional analyses support a biphasic model in which early AD pathology activates expression of a synaptic transcriptional signature that promotes neuronal injury, followed by a decrease that is compensatory. In sum, our cross-species strategy establishes a causal chain linking AD pathology, transcriptome perturbation, N-Methyl-D-Aspartate receptor excitotoxicity, and neurodegeneration.

Introduction

A major goal for the Alzheimer’s disease (AD) field is to map the pathophysiologic cascade from its earliest neuropathologic triggers to the downstream cognitive manifestations (Fig. 1A) [1]. Gene expression profiling from human brains is one promising approach to highlight molecular mediators. In a meta-analysis of RNA-Sequencing (RNAseq) from ~2000 postmortem brain tissue samples, the Accelerating Medicines Partnership (AMP)-AD target discovery consortium defined 30 AD-associated consensus co-expression networks [2,3,4,5]. This study highlighted robust transcriptional signatures for immune and synaptic regulatory mechanisms in AD pathogenesis, among other pathways, and similar findings have been independently reported by others [6,7,8,9,10,11]. Additional evidence from human genome-wide association studies further implicates immune-mediated mechanisms in AD [12, 13]. However, a definitive understanding of the causal roles for these and other gene expression signatures requires functional characterization in vivo.

Several major challenges hinder the comprehensive functional dissection of the causal chain in AD pathogenesis. First, aging and comorbid brain pathologies can confound interpretation of differentially-expressed genes. Second, AD pathophysiology is a dynamic process that evolves over decades, but human brain transcriptional profiles can only be examined at one timepoint following death, which usually occurs at the end stage of the disease. Early, causal changes may thus be obscured by cell loss or late-stage pathologies. Among causal gene expression perturbations that alter AD onset and progression, it is further important to differentiate changes that promote from those that compensate for AD pathophysiology (Fig. 1A). Lastly, because changes in the transcriptome comprise correlated, modular co-expression networks, it is also necessary to dissect the fine-scale structure to identify key drivers with the potential to restore regulatory homeostasis.

By contrast with expression profiling of human postmortem brain tissue, animal models enable controlled experimental manipulations that isolate specific molecular triggers (e.g., amyloid-beta (Aβ) plaques and tau neurofibrillary tangles) along with longitudinal study designs that account for the influence of aging. Mice harboring amyloid precursor protein (APP) or microtubule associated protein tau (MAPT) transgenes recapitulate gene expression signatures similar to AD human brain transcriptomes, including dysregulation of immune and synaptic processes [2, 14,15,16,17,18,19]. Importantly, genetic manipulations of immune regulators such as TREM2 or NFkB in microglia have also been demonstrated to modify neurodegeneration, supporting causal roles for these pathways [20, 21]. However, the co-expression modules derived from the AMP-AD human brain transcriptome meta-analysis range in size from 500 to 5000 genes, creating an impediment to evaluating a larger number of potential targets. High-throughput systematic genetic dissection, or systems genetics, to pinpoint causal genes and network drivers is not feasible in mouse models, requiring alternative strategies. In the fruit fly, Drosophila melanogaster, pan-neuronal expression of the secreted human Aβ peptide or MAPT/tau protein triggers misfolding and aggregation similar to amyloid plaque or neurofibrillary tangle pathology, respectively, along with age-dependent, progressive neurodegeneration [22,23,24,25]. We recently showed that pan-neuronal expression of tau in flies triggers conserved gene expression signatures strongly overlapping with differentially expressed genes from AMP-AD human brain transcriptomes [26]. Here, we employ fly models to functionally dissect AD gene expression signatures, differentiating specific and shared responses to Aβ, tau, and/or aging, and mapping the key driver genes and causal network mechanisms that mediate the impact of disease pathology on progressive central nervous system (CNS) dysfunction.

Methods

Human subjects

The AMP-AD consortium meta-analysis of AD differential expression and consensus co-expression networks was previously published [2]. In particular, our study utilized the available summary statistics and lists of all reported genes from co-expression modules. The module names incorporate abbreviations denoting the relevant brain regions contributing to RNAseq analysis: CBE, cerebellum; TCX, temporal cortex; FP, frontal pole; IFG, inferior frontal gyrus; PHG, parahippocampal gyrus; STG, superior temporal gyurs; and DLPFC, dorsolateral prefrontal cortex. Overlapping transcriptional signatures define 5 consensus clusters (A-E, Supplemental Table 1) among the 30 AD-associated AMP-AD coexpression modules. Modules within each consensus cluster recapitulate similar enrichment for cell type specific markers and biological pathways [2]. We also leveraged published clinical, pathologic, demographic, and RNAseq data from the Religious Orders Study and Rush Memory and Aging Project (ROSMAP). All ROSMAP participants enrolled without known dementia and agreed to detailed clinical evaluation and brain donation at death [27]. We also employed summary statistics from published analyses of ROSMAP data, including differential expression analyses of bulk [28] and single-nucleus RNAseq [29]. Additional single-nucleus RNAseq data was leveraged from a study of cortical biopsies taken during surgery for normal pressure hydrocephalus [30]. GWAS summary statistics were also used from published analysis of AD [31, 32], Parkinson’s disease [33], and height [34].

Drosophila stocks and husbandry

For transcriptome profiling, we used previously characterized UAS-tau (ref. [35]), UAS-Aβ42 (ref. [36]), and UAS-αSyn (ref. [37]). The UAS-tau is a chromosome III line (P{UAS-Tau.wt}7B) encoding a full-length, wildtype human 2N4R MAPT protein isoform. The UAS-Aβ42 encodes the 42-amino-acid Aβ42 peptide fused to the Argos secretion peptide (UAS-Aos:hAβ42; line M17A on chromosome III). The UAS-αSyn stock has a recombinant second chromosome with 2 transgenes codon-optimized for Drosophila, encoding full-length human alpha-synuclein. Two published [38, 39] transgenic models of Huntington’s disease expressing human huntingtin (UAS-HTT^NT231Q128 and UAS-HTT^FLQ200) were also included in the longitudinal RNAseq study, and these data contributed to derivation of Drosophila co-expression modules. For pan-neuronal expression, we used the elav^c155-GAL4 (ref. [40]) driver strain on the first chromosome, which is available from the Bloomington Drosophila Stock Center (BDSC; Bloomington, IN, USA). Experimental animals were heterozygous for the Elav-GAL4 driver and respective transgenes encoding Aβ42, tau, or alpha-synuclein. For controls, we included both Elav-GAL4/+ and w¹¹¹⁸. For locomotor testing and RNAseq, female adult flies were aged to 2, 5, 7, 10, 14, 21, 28, 42, and 57, as detailed in Fig. 1B (not all genotypes were examined at all timepoints due to reduced survival). For genetic modifier screening using locomotor behavior, we used the identical UAS-tau and UAS-Aβ42 strains, which were combined in a stock together with Elav-GAL4, a heat-shock-induced lethality mutation on the Y chromosome (P{hs-hid}) to increase the efficiency of virgin collection [41], and a balancer carrying tub-GAL80 to silence the disease associated protein, permitting maintenance of the stock. Virgin females were collected from this stock and crossed to male flies carrying potential modifier alleles. All modifier strains (RNAi and other alleles) showing genetic interactions are detailed in Supplementary Table 11; these lines were obtained from the BDSC, Vienna Drosophila Resource Center (VDRC), or the Japan National Institute of Genetics (NIG). All modifier gene manipulations were tested in heterozygosity. For secondary tests of genetic interactions using adult brain histology, we used the same UAS-Aβ42 strain, but as in prior publications [42] we used an alternate UAS-Tau^R406W strain optimized for modifier studies using the brain histology assay. This chromosome III line encodes a 0N4R mutant form of MAPT associated with familial frontotemporal dementia. For histology experiments, we used a control line (UAS-empty-VK33) with an empty UAS site at a chromosome III VK33 docking site from Dr. Hugo Bellen). We obtained the following additional stocks from BDSC or VDRC: P{UAS-VGlut.D}2 (ref. [43]); P{VGlut-GAL4.D}1(ref. [44]); Df(2L)VGlut[2] (ref. [45]); and P{GD834}v2574 (ref. [46]). The Transcriptional Reporter of Intracellular Ca2+ (TRIC) (ref. [47]) was also obtained from BDSC: w[*]; P{y[+t7.7] w[+mC]=nSyb-MKII::nlsLexADBDo}attP24, P{w[+mC]=QUAS-p65AD::CaM}2, P{y[+t7.7] w[+mC]=13XLexAop2-mCD8::GFP}attP40/CyO; P{y[+t7.7] w[+mC]=nSyb-QF2.P}attP2, P{w[+mC]=tub-QS.P}heph[9B]. For the locomotor screen for modifiers of tau, flies were raised at 23 °C. Otherwise, flies were raised on standard molasses-based Drosophila media at 25 °C. Flies were maintained in ambient room lighting unless otherwise noted.

Analysis of AMP-AD co-expression modules

We leveraged published data from the ROSMAP bulk brain RNAseq differential expression analysis [28] in order to examine overlaps with each AMP-AD co-expression module (Supplementary Table 1). Differentially expressed genes from analyses considering either Aβ/amyloid or tau/tangle pathologic burden were defined based on |log10 (p-value)| > 1.3, and we considered up- and down-regulated genes separately. For statistical analysis, we performed a hypergeometric test, evaluating the overlap of all genes in each AMP-AD module with the AD pathology-associated, differentially expressed gene sets. For the background, we used the overall total number of shared genes between the AMP-AD and ROSMAP datasets.

Candidate driver genes were prioritized from the AMP-AD co-expression modules using 3 distinct criteria: i) magnitude of differential expression (extreme), ii) average network degree (hub), and iii) correlation with the module first principal component (PC1, average). Each criterion was used to generate a rank-list from each of the 30 AMP-AD modules. “Extreme” genes were ranked using the true effect (TE) differential expression absolute values from the published AMP-AD RNAseq random effects meta-analysis (data available at syn11914606) [2]. “Hub” gene average degree was computed using the igraph R package [48, 49] (data available at syn10208070). “Average” genes were ranked based on absolute PC1 correlation (data available at syn11940434). Genes were selected from the top 1% of genes ranked by each criterion from each module, with a maximum list size of 30 ranked genes and a minimum list size of 10 ranked genes for each criterion. A maximum of 90 genes and minimum of 30 genes could therefore be prioritized from each module based on the 3 independent criteria. As detailed in the AMP-AD consortium publication [2], since consensus co-expression networks were derived independently from multiple brain regions, they can be organized into overlapping clusters. We adjusted our final selections of n = 400 gene drivers based on both cluster and module size. Cluster selection sizes were weighted using the formula:

$${Cluster}\,{selection}\,{size}=400* \frac{{No}.\,{of}\,{genes}\,{in}\,{cluster}}{{No}.\,{of}\,{genes}\,{in}\,{all}\,{clusters}}$$

Module selection sizes were then weighted within each cluster using the formula:

$${Module}\,{selection}\,{size}={Cluster}\,{selection}\,{size}* \frac{{No}.\,{of}\,{genes}\,{in}\,{module}}{{No}.\,{of}\,{genes}\,{in}\,{cluster}}$$

To generate putative causal subnetworks (Supplementary Fig. 9), we first obtained published network adjacency matrices for the AMP-AD modules for each of the 7 brain regions from Synapse (cerebellum/CBE, syn8281722; dorsolateral prefrontal cortex/DLPFC, syn8268669; frontal pole/FP, syn8340017; inferior frontal gyrus/IFG, syn8349785; parahippocampal gyrus/PHG, syn8345109; superior temporal gyrus/STG, syn8343704; temporal cortex/TCX, syn8276546). Each brain region’s adjacency matrix was used to produce network graphs with the network_from_adjacency_matrix function in the igraph R package. For each module, prioritized driver gene candidates for which Drosophila homologs were experimentally validated as modifiers were used as “network seeds” plotting along with first-order neighbors from the network adjacency matrix, using the neighbors function in the igraph R package. “First-order neighbor” refers to any human gene from the coexpression module that shares an edge with one of the prioritized candidate driver genes that are experimentally validated as modifiers; edges are based on pairwise correlation of coexpression. Subnetworks containing the modifier genes plus their directly coexpressed first-order neighbors were produced for each of the 30 AMP-AD modules using the induce.subgraph command, and the resultant undirected subgraph was used for downstream analyses.

The MAGMA tool [50] was used to conduct gene and gene set enrichment analyses for GWAS signal in the 30 AMP-AD modules and the respective subnetworks, derived using either (i) all prioritized candidate driver genes or (ii) experimentally validated driver gene modifiers (causal subnetworks). GWAS summary statistics were downloaded from studies of AD [31, 32] (NIAGADS: NG00075; https://ctg.cncr.nl/software/summary_statistics), Parkinson’s disease [33] (https://pdgenetics.org/resources), or height [34] (https://portals.broadinstitute.org/collaboration/giant/). The org.Hs.eg.db R package [49, 51] was used to map gene symbols to Entrez IDs for each gene in the AMP-AD modules or subnetworks prior to running MAGMA. All analyses with AD GWAS data were repeated without summary statistics from chromosome 19 in order exclude potential confounding due to APOE.

Causal mediation analysis examining the relation of PHGbrown and STGblue with AD pathologic burden and cognitive decline were conducted using the mediation package in R [49, 52]. These analyses employed clinical, pathologic, and RNAseq (DLPFC bulk brain) data from ROSMAP. For the outcome variable, we use a global cognitive function summary measure (cogn_global). For the predictor variable, we used a summary measure of global AD pathology (globalpath). We computed mean gene expression for all genes in each AMP-AD module as a mediator variable. The age at death (age_death), post-mortem interval (pmi), education (edu) and sex (sex) were included as covariates. For statistical analysis, 1000 simulations were completed using the mediate() function.

Gene expression analysis in Drosophila models of aging and AD/ADRD

For longitudinal RNAseq profiles of Drosophila heads from the Elav > Aβ42, Elav>tau, and Elav>alpha-synuclein models and controls (Elav-Gal4 and w¹¹¹⁸), we prepared 3 biological replicates for each genotype and aging time point. For the VGlut overexpression model and VGlut-Gal4 controls, 4 biological replicates were generated. Each replicate consisted of 100 heads collected from flash frozen, age-matched virgin female flies. mRNA was extracted using TRIzol (#15596026, Invitrogen) followed by DNAse treatment. A minimum of 500 nanograms of total DNase-treated extract RNA were used per replicate. Preparation of sequencing libraries, RNA-sequencing, and alignment was performed by the New York Genome Center (longitudinal analysis of Drosophila AD/ADRD models) and the Baylor College of Medicine Genomic and RNA Profiling Core (VGlut overexpression model). Briefly, samples were prepared using the Illumina TruSeq Stranded mRNA Library Prep Kit. Samples were then sequenced on an Illumina NovaSeq 6000 with 100 bp paired-end reads. Raw reads were aligned to Drosophila reference genome dm6 r6.06 and quantified. Genes with an average read count < 50 across all samples were excluded. Gene quantification was performed using the featureCounts function from the Rsubread package.

Differential expression analysis was performed using DESeq2 (1.32.0) in R (4.1.0) [49, 53] (Supplementary Table 2). Raw transcript counts were normalized for library depth and composition using the DESeq2 median of ratios normalization approach. To identify gene expression changes during normal aging within control flies (w1118), we first modeled age as categorical variable and performed the likelihood-ratio test to determine the significance of the age term (reduced model: expression ~ 1). To identify gene expression changes triggered by each AD/ADRD pathologic trigger, we employed the following linear model with age as a categorical variable (expression ~ genotype + age). We tested the significance of the genotype term coefficient using a likelihood-ratio test comparing with a reduced model (expression ~ age). Functional pathway enrichment analysis was completed using the G:GOSt function from the R package gProfiler2 (0.2.1) [49, 54, 55]. The Gene Ontology (GO), and KEGG databases were used for querying genes. We used the hypergeometric overlap test (phyper function in R) to evaluate the overlap/enrichment of each AMP-AD human coexpression module with the differential expressed gene sets defined from the tau and Aβ, as well as the overlap between aging-induced differential expression signatures and the perturbed gene sets defined from tau, Aβ, and αSyn models (Fig. 1E, F, Supplementary Fig. 1B and D, and Supplementary Table 5). The genes within each AMP-AD human co-expression module were mapped to Drosophila using DIOPT [56] with a DIOPT score filter >= 5 (Supplementary Table 4). When more than one fly homolog had a DIOPT score >= 5, all were included. For overlap testing, the total number of shared genes between the Drosophila RNAseq and AMP-AD datasets (after DIOPT conversion) was used as the background.

Analysis of Drosophila co-expression networks

Weighted gene co-expression network analysis (WGCNA) was implemented as previously described [26] using the WGCNA package in R [49, 57]. Expression counts from 147 unique Drosophila samples (13,192 transcripts) were included in the analysis after normalization in DESeq2 (median-of-ratios depth normalization). The soft power value was set as 4 (scale-free R² = 0.9). To achieve maximum sensitivity, deepSplit was set as 4 and cutHeight was set as 0.1. Minimum module size was set as 23. Closely related modules were merged based on module eigengenes at a distance threshold of MEDissThres = 0.1 (Supplementary Fig. 3 for the cluster dendrogram). To indicate expression behavior of the modules, module “eigengenes” were calculated with the moduleEigengenes() function in the WGCNA package, and we also computed mean module expression by averaging counts of all gene in each module. We examined cell type signatures for each module using markers from a published Drosophila brain single cell sequencing dataset [58] (Supplementary Fig. 3), employing the hypergeometric overlap test (phyper function in R. Functional enrichment analysis of Drosophila gene modules was done using the “WebGestaltR” package in R [49, 59]. The enrichMethod parameter was set to “ORA”. The organism parameter was set to “dmelanogaster”. The enrichDatabase was set to ‘geneontology_Biological_Process’. In order to determine module conservation and cross-species overlaps, Drosophila genes were mapped to their human orthologs using DIOPT [56] with a filter of DIOPT score >= 5. When more than one human homolog had a DIOPT score >= 5, all were considered. The hypergeometric test (phyper) was next used to compute overlaps between each Drosophila co-expression module and each AMP-AD human co-expression module. For the background, we used the total number of shared genes between complete Drosophila RNAseq dataset (following DIOPT conversion) and the AMP-AD dataset. Results were plotted using the pheatmap package in R [49, 60].

In order to examine for differential expression of co-expression modules in relation to age or AD/ADRD pathologic triggers, we used linear regression with mean expression of all genes in each module as the outcome variable (Fig. 2B, Supplementary Figs. 5, 6, and Supplementary Table 8). We considered RNAseq data from each AD/ADRD model (Elav > Aβ42, Elav>tau, or Elav>alpha-synuclein) along with the Elav-Gal4 driver control. We compared the full model (module expression ~ age + genotype) to a reduced model (module expression ~ age). Age was coded as a categorical, factor trait to account for the possibility of non-linear effects. The genotype term coefficient was used for significance testing using likelihood ratio test. To determine whether fly-human overlapping modules also shared expression changes in the same direction, we computed a concordance score for each significantly overlapping module pair, adapting a method from our prior work [2]. Specifically, the concordance score (ConcScore) is the percentage of genes in the concordant direction, which is computed as follows:

$${ConcScore}={ConcUp}-{ConcDown}$$

whereby

$${ConcUp}=\frac{{{\#}}{overlapped}\,{genes}\,{up}\,{in}\,{both}\,{fly}\,{and}\,{human}}{{{\rm{\#}}}{overlapped}\,{genes}\,{up}\,{in}\,{human}},$$

$${ConcDown}=\frac{{{\#}}{overlapped}\,{genes}\,{down}\,{in}\,{both}\,{fly}\,{and}\,{human}}{{{\rm{\#}}}{overlapped}\,{genes}\,{down}\,{in}\,{human}}$$

To assess the relationship between module expression and locomotor behavior (Fig. 2C and Supplementary Fig. 7), we first tested the Pearson correlation between module expression and climbing speed using the cor.test() function in R. For fM1/fM11, we further implemented linear regression using a base model with locomotor behavior (climbing speed) as the outcome variable, and both age and genotype as predictor variables. Mean fM1 or fM11 module expression was then included as an additional predictor variable. To assess for potential mediation, we computed the percentage change in the genotype term β-coefficient (absolute value) with or without inclusion of the term for module expression in the model. The β-coefficient of the genotype term from the full model, including fM1/fM11 expression was considered as the direct effect.

Drosophila genetic modifier screen

For genetic screening, we mapped homology for each prioritized candidate AMP-AD driver using DIOPT [56]. In most cases, we selected the single best ortholog (DIOPT score of 5 or greater) for genetic screening. For a minority of genes (n = 16), in which multiple fly homologs were very strongly or equally well-conserved (DIOPT score of 10 or greater), we included up to 3 homologs. The negative geotaxis climbing assay was performed using a custom robotic system (SRI International, available in the High Throughput Behavioral Screening Core at the Duncan Neurological Research Institute) as previously described [38]. The robotic instrumentation elicited negative geotaxis by ‘tapping’ Drosophila housed in 96-vial arrays. After three taps, video cameras recorded and tracked the movement of animals at a rate of 30 frames per second for 7.5 s. For each genotype, we collected 4 replicates of 10 females to be tested in parallel (biological replicates). Each trial was repeated five times (technical replicates) per trial day. Flies were transferred into new food every day. The automated, high-throughput system is capable of assaying 16 arrays (1536 total vials) in ~3.5 hr. To transform video recordings into quantifiable data, individual Drosophila were treated as an ellipse, and the software deconvoluted the movement of individuals by calculating the angle and distance that each ellipse moves between frames. Replicates were randomly assigned to positions throughout the plate and were blinded to users throughout the duration of experiments. The results of this analysis were used to compute the average speed of the 10 females in each vial/time point. Software required to run and configure the automation and image/track the videos include Adept desktop, Video Savant, MatLab with Image Processing Toolkit and Statistics Toolkit, RSLogix (Rockwell Automation), and Ultraware (Rockwell Automation). Additional custom-designed software includes Assay Control – SRI graphical user interface for controlling the assay machine; Analysis software bundles: FastPhenoTrack (Vision Processing Software), TrackingServer (Data Management Software), ScoringServer (Behavior Scoring Software), and Trackviewer (Visual Tracking Viewing Software).

The modifier screen used 1,069 alleles targeting 350 Drosophila genes that were homologous to 344 human genes. The female progeny from each cross were collected within a 12-hour period. We assessed climbing from 5 to 18 days post-eclosion in our initial locomotor screen. On average 6 timepoints were assessed for the tau model (as it shows an earlier phenotype) and 9 time points for Aβ. As in prior published work [38, 61], we assessed locomotion in Drosophila as the average speed at which animals moved in each vial as a function of age and genotype using a nonlinear random mixed effects regression model. ANOVA was used to assess statistical significance followed by Bonferroni-Holm correction to determine adjusted p-values. Specifically, we looked at differences in regression between genotypes, genotypes with time (additive effect, represented by a shift in the curve) or the interaction of genotype and time (interactive effect, represented by a change in the slope of the curve). We estimated the expected statistical power to detect differences by each of our models using a threshold for statistical significance (alpha = 0.05). All plots were reviewed to confirm that statistically significant results from the nonlinear mixed effects models represented meaningful differences on visual inspection. Statistical analysis and graphical visualizations were generated in R [49]. For those strains that enhanced both tau- and Aβ- induced locomotor behavior, we performed an additional test to evaluate the consequences of gene manipulation independent of tau/Aβ, crossing with the elav > GAL4 driver line. Female progeny were then aged to determine whether the modifier can be classified as additive or synergistic, based on its locomotor toxicity profile in the absence of tau/Aβ (Supplementary Table 11). All modifiers from PHGbrown were retested and locomotor behavior was assayed longitudinally from 3 to 28 days (Fig. 3D and Supplementary Fig. 8).

Annotation of human coexpression subnetworks with causal inferences

As noted above, in order to consider a prioritized candidate driver as an experimentally validated modifier gene, we required consistent results from at least 2 independent allele strains (either consistent, significant enhancement or suppression for a given direction of manipulation, either loss- or gain-of-function). For all such validated modifier genes, we attempted to generate “causal predictions” (amplifying vs. compensatory) by integrating the results from the Drosophila locomotor tests (either loss- or gain-of-function enhancer vs. suppressor) with the human differential gene expression (up- versus down-regulated in AD) (Supplementary Fig. 9). First, for each modifier gene, we first determined a consensus for the observed interactions based on the results for alleles tested and considering tau versus Aβ separately. Second, we used the consensus results for loss- or gain-of-function of the fly gene to infer the potential impact of differential expression in human brains. For example, if a given human gene is down-regulated in brains affected by AD, and loss-of-function of the fly homolog led to suppression of Aβ/tau neurotoxcity, the gene would be annotated as “compensatory”; whereas, if loss-of-function led to enhancement of Aβ/tau neurotoxcity, the gene would be annotated as “amplifying”. When the direction of perturbation is opposite to genetic manipulation in Drosophila, we infer the consequences of the observed human gene expression change. For genes in which the consensus result had opposite interactions in the tau versus Aβ models, we did not assign any causal predictions (gray nodes in the subnetwork plots; Fig. 4A and Supplementary Fig. 10).

Drosophila brain histology

Drosophila heads were fixed using 8% glutaraldehyde (Electron Microscopy Sciences) at 4 °C for 7–10 days, followed by paraffin embedding and microtome sectioning, as previously described [36]. Serial 5 μm-thick frontal sections (Leica microtome) were prepared from Drosophila heads and mounted on slides, followed by hematoxylin and eosin staining to examine brain morphology. We quantified vacuoles greater than 5 μm in diameter in the central brain in all frontal brain sections (~20 per animal) and computed the average vacuole count per section. All light microscopic images were acquired using a Leica DM 6000 B system. Brain vacuole counts for individual Drosophila were then used to conduct two-sample unpaired t-tests between genotypes for significance using GraphPad Prism software. For multiple comparisons, we performed one-way ANOVA with Dunnett’s post-hoc test.

Drosophila calcium imaging

Immunofluorescence imaging was performed to quantify the TRIC signal. Initiation of TRIC expression via the QUAS system was performed as previously described [62]. Flies carrying TRIC elements were placed in food vials containing 300ul of 250 mg/ml quinic acid solution 3 days before they reach the target age. The immunostaining procedure was adapted from published protocols [63]. Flies were anesthetized with CO2 and brains were dissected with forceps in PBS solution (GeneDEPOT, P2101-050) and fixed in 3.7% formaldehyde (J.T.Baker, 2106-01) for 20 min in room temperature. Formaldehyde was removed and brains were washed using PBS with 0.3% Triton-X (Sigma-Aldrich, T8787) (PBST) for 15 min by 3 times. Primary antibodies were diluted in 0.3% PBST and samples were incubated in primary antibody at 4 °C, rocking for 2 days. Primary antibodies were removed, and brains were washed 3 times using 0.3% PBST for 15 min each. Secondary antibodies were diluted in 0.3% PBST, and samples were incubated with secondary antibody at 4 °C, rocking for 2 days. Secondary antibodies were removed, and brains were again washed 3 times using 0.3% PBST for 15 min each. Whole brains were mounted in Vectashield antifade mounting medium (Vector Laboratories, H-1000-10) and stored in the dark at 4 °C prior to imaging. Samples were imaged on a Leica Microsystems SP8X confocal microscope. Z-stacks of 1 micron covered the entirety of whole-mount brains. We used the following antibodies and dilutions: Rabbit anti GFP (Invitrogen A11122, 1:500), Mouse anti bruchpilot (DSHB nc18, 1:100), Alexa Fluor® 488 AffiniPure™ Donkey Anti-Rabbit IgG (H + L) (Jackson ImmunoResearch 711-545-152, 1:500), Cy™3 AffiniPure™ Goat Anti-Mouse IgG (H + L) (Jackson ImmunoResearch 115-165-003, 1:500). Quantification of calcium signal was performed using ImageJ software on z-stacks of imaged sections. The entirety of brain region (ROI) of the z-stacked image was selected and regions not containing brain tissue were selected as background. Mean pixel intensity was measured and intensity from the background regions was subtracted from that of ROI for each fluorescence channel. Two sample t-tests were conducted between genotypes for significance using the GraphPad Prism software.

Single cell RNA-sequencing

A comprehensive description and analysis of the Drosophila single cell RNAseq dataset will be published elsewhere. The overall data generation, quality-control, and analysis workflow was adapted from our prior publication [64] and is also available with the full dataset on synapse (syn34767207). Three biological replicates per genotype were generated, and all animals were aged to 10 days post-eclosion (21 samples total). Each replicate consisted of 16–18 intact, dissected brains from female flies, and were enzymatically dissociated into a single cell suspension as described in Davie et al., 2018. Single cell libraries were prepared per manufacturer’s protocol for the Chromium Single Cell Gene Expression 3’ v3.1 kit (10x Genomics) by the BCM Single Cell Genomics Core. Completed libraries were sequenced by the Baylor Genomic and RNA Profiling Core using the Illumina NovaSeq 6000 platform with a minimum depth of 300,000,000 reads per sample. Illumina BCL files were demultiplexed into FASTQ files by calling the Cell Ranger 4.0.0 mkfastq function. FASTQ files were aligned to the Drosophila reference genome (BDGP6.22.98) and gene counts were quantified using the Cell Ranger 4.0.0 count pipeline. Given the 10x recovery rate estimations, the cell calling algorithm in Cell Ranger was applied by setting the --expect-cells parameter in count to 10,000 per library.

We extracted the mean expression for all gene members from fM1 or PHGbrown (mapped fly orthologs) for each cell. Cluster-wise Log2 fold-change for the gene set was evaluated comparing Elav> model with driver control was calculated by averaging module expression of all cells in the cluster. Significance of expression change was determined using Wilcoxon signed-rank test. Cluster-wise log2 fold change was used to generate UMAP in Fig. 6D using the FeaturePlot() function in the Seurat package in R. For the fly-human cross-species analyses, we used published human postmortem brain single-cell RNAseq data [29] as well as complementary data from living cortical biopsies taken during neurosurgery for normal pressure hydrocephalus [30]. For the Mathys et al. postmortem brain dataset, differential expressed genes from analyses of amyloid or neuritic plaque burden (plaq_n) for each cell cluster were tested for overlap with PHGbrown module genes and its subnetworks. For Gazestani et al. cortical biopsy dataset, differentially expressed genes from excitatory neurons (ExN) analyses, including either (i) amyloid-beta positive (tau negative) or (ii) amyloid-beta and tau positive samples were similarly evaluated for overlap with the PHGbrown module geneset. The hypergeometric overlap test (phyper in R) was applied to examine overlaps between the PHGbrown module gene set and differentially expressed genes among excitatory neuron cell-types, considering either all genes or up- vs. down-regulated genes separately.

Statistical analysis

Statistical tests are described with relevant methods, above, and in all figure legends along with sample sizes, consistent with prior publications. We randomized sample processing order and performed at least 2 technical replicates. Investigators were blinded to genotypes and experimental conditions when possible. We used two-tailed tests and evaluated sample variance with F tests. For RNAseq analyses, p-values were adjusted using the Benjamini-Hochberg false discovery rate (FDR), with significance set at p_FDR < 0.05.

Results

Cross-species dissection of causal pathologic triggers

Our previously published AMP-AD meta-analysis of postmortem brain RNAseq defined 3,774 AD differentially expressed genes and implicated 30 human brain gene co-expression network modules, relying on an integrated clinical-pathologic diagnosis [2]. The modules were independently derived from seven brain regions and are organized into five overlapping consensus clusters (Supplementary Table 1). Leveraging complementary differential gene expression analyses [28] based on Aβ or tau neuropathology, we confirmed that all 30 modules were associated with AD pathology (Supplementary Table 1). Moreover, since amyloid and tau pathology co-occur and are highly correlated in AD, most modules showed consistent enrichment with differentially expressed genes from both analyses.

To determine which transcriptional changes might be directly or indirectly caused by AD pathology and to differentiate Aβ versus tau as potential triggers (Fig. 1A), we turned to two Drosophila transgenic models. As introduced above, pan-neuronal expression of either the secreted human 42-amino-acid Aβ peptide or the tau protein (wild-type 2N4R isoform) recapitulates age-dependent neuronal loss and progressive CNS dysfunction, including locomotor impairment (Supplementary Fig. 1A) [25]. We used previously published Drosophila strains [36, 65] harboring human transgenes responsive to the yeast Upstream Activating Sequence (UAS), whereby expression is directed selectively to neurons by the Elav-GAL4 driver [66]. As Elav > Aβ, Elav>tau, and controls (Elav-GAL4 /+ or w¹¹¹⁸) were aged, we generated longitudinal transcriptome profiles from fly heads at six to nine timepoints across the lifespan (Fig. 1B). We used multivariate linear regression to model the influence of each AD pathologic trigger on gene expression, including age as a covariate. Aβ and tau induced 2746 and 3830 differentially expressed genes, respectively (Fig. 1C, Supplementary Fig. 1B, and Supplementary Table 2). Based on enrichment for gene ontology terms, genes related to lipid and oxidative metabolism were perturbed in both models (Fig. 1D and Supplementary Table 3). Conversely, regulators of synaptic transmission and calcium related processes were specifically enriched among tau differentially-expressed genes, whereas glycerophospholipid metabolism was specifically highlighted in the Aβ transgenic model. Similar functional enrichments characterized several of the AD-associated human co-expression modules from the AMP-AD meta-analysis (Table 1). For example, module PHGbrown, which is down-regulated in AD, is characterized by abundant neuronal and synaptic genes [2].

Fig. 1: Cross-species dissection of Alzheimer’s disease (AD) pathologic triggers.

A Hypothetical causal chain linking AD triggers, either directly or indirectly, to gene expression perturbations, and Alzheimer’s disease pathophysiology. Perturbations in gene expression that are causal modifiers of disease pathophysiology may be either amplifying or protective. Amplifying changes are pathogenic, promoting AD pathogenesis; whereas protective changes are compensatory, attenuating AD progression. B Control and AD transgenic models (n = 3 for each genotype) were profiled for locomotor behavior and gene expression using RNA-sequencing from Drosophila heads, including up to 9 timepoints between 2 and 57 days. AD model genotypes were as follows: Aβ (elav-Gal4/+; UAS-Aβ/+) and tau (elav-Gal4/+; UAS-tau/+). Both wildtype (w¹¹¹⁸) and driver (elav-GAL4/+) controls were evaluated. See also Supplementary Fig. 1. C Venn diagram highlighting shared and unique differentially expressed genes following pan-neuronal expression of Aβ or tau. Linear regression was performed comparing elav > Aβ or elav>tau with elav-GAL4 driver controls, including longitudinal data (days 2–28 for Aβ and days 2–42 for tau) and adjusting for age. Statistical analysis was based on a Wald test (false discovery rate (FDR) < 0.5). See also Supplementary Fig. 1B and Supplementary Table 2. D Plots highlight shared and unique pathways perturbed in Aβ and tau transgenic flies, based on gene ontology term enrichment. Statistical analysis based on the hypergeometric overlap test (FDR < 0.5). E Cross-species analysis highlights genes within human AD-associated coexpression modules, for which expression of conserved Drosophila homologs are triggered by Aβ (green), tau (purple), or both (yellow). Each bar indicates the total module size, based on the total number of conserved genes. Purple asterisks indicate significant enrichment for tau differentially-expressed genes; black asterisks indicate enrichment for both tau and Aβ DEGs. See also Supplementary Table 5. F Cross-species analysis highlights genes within human coexpression modules, for which Drosophila homologs are differentially expressed in response to AD and related dementia pathologic triggers (blue: Aβ, tau, or alpha-synuclein), aging (orange), or both (yellow). See also Supplementary Fig. 1C, D. (E-F) Module names incorporate abbreviations denoting the relevant brain regions: CBE, cerebellum; TCX, temporal cortex; FP, frontal pole; IFG, inferior frontal gyrus; PHG, parahippocampal gyrus; STG, superior temporal gyurs; and DLPFC, dorsolateral prefrontal cortex.

To integrate the AMP-AD human consensus co-expression modules with Drosophila differential expression data, we mapped gene homologs across species (Supplementary Table 4). Most human co-expression modules (25 out of 30) were significantly enriched for fly gene homologs showing differential expression in response to tau (n = 25) and/or Aβ (n = 12) (Fig. 1E and Supplementary Table 5). Based on cross-species overlaps, we infer that most modules are at least partially activated by both AD triggers. We therefore partitioned each module into discrete Aβ- and/or tau-responsive gene sets, along with a component for changes shared by both the Aβ and tau transgenic flies (Fig. 1E). Overall, among all conserved genes within each module, the proportion of genes responsive to AD pathologic triggers varied (range_Aβ = 14–22%; range_tau = 18–43%). For example, among 887 conserved genes in CBEturquoise, which is related to phagocytosis and enriched for microglial markers, 279 (31.5%) or 455 (51.3%) genes, respectively, significantly overlapped with Aβ- or tau-induced differentially expressed genes from transgenic flies, and these overlaps were highly unlikely to have occurred by chance (p_Aß = 4.1 × 10⁻⁶; p_tau = 1.3 × 10⁻¹⁶). In other cases, such as for the PHGbrown module, we note a more selective enrichment for differentially expressed genes from the Drosophila tau model (p = 1.2 × 10⁻⁷) rather than the Aβ flies (p = 0.74). We further considered cross-species overlaps using a more stringent requirement of concordant up- versus down-regulation). Among all AD differentially expressed genes overlapping between human brains and fly models, the concordant gene overlaps per module range from 43–78% or 41–77% for the Aβ or tau transgenic models, respectively (Supplementary Fig. 2).

In human brains, amyloid plaques and tau tangles frequently co-occur with other pathologies, such as alpha-synuclein (αSyn) Lewy bodies, which can increase vulnerability for cognitive impairment in AD and also cause Lewy body dementia [67]. Pan-neuronal expression of human αSyn (Elav > αSyn) in Drosophila causes Lewy-body like protein aggregation, dopaminergic and other neuron loss, and progressive locomotor impairment [68,69,70]. We therefore obtained complementary transcriptome profiles from αSyn transgenic flies using identical aging timepoints and experimental conditions, facilitating comparisons with the complementary results from Elav>tau and Elav > Aβ models. Overall, there was substantial overlap in brain gene expression changes triggered by distinct disease models, with 875 common differentially-expressed genes among the three genotypes and 38–49% of gene perturbations shared among pairwise comparisons of Aβ, tau, or αSyn flies (Supplementary Fig. 1B and Supplementary Table 2). Since aging influences the brain independently of disease pathologies, we also defined aging-induced gene expression changes from control flies (w¹¹¹⁸). Consistent with our prior work [26], aging had a profound effect on the transcriptome, significantly altering expression of 10,717 genes. Moreover, ~80% of expression changes observed in the neurodegenerative models were also perturbed in aging control animals, and these overlaps were highly unlikely to occur by chance (Supplementary Fig. 1B). We therefore extended the cross-species approach to pinpoint putative αSyn- and aging-induced differentially expressed genes among the human AD-associated co-expression modules (Fig. 1F and Supplementary Fig. 1D). In flies, αSyn altered the expression for 24–39% of conserved genes from each human co-expression module, while aging had an even greater effect (86–95% genes). Overall, the combined effects of aging along with Aβ, tau, and αSyn pathologies appear to explain the majority of changes in the human AD brain transcriptome (Fig. 1F).

Cross-species dissection of brain expression networks and behavior

The preceding analyses revealed extensive conservation between the transcriptional changes observed in human brain and Drosophila models of AD, Lewy body dementia, and aging. We next examined if this conservation extends to the network level. We independently performed weighted gene co-expression network analysis (WGCNA) [57] using transcriptome profiles from the Drosophila adult brain. For optimal power, we made use of all available Drosophila transcriptome data, including longitudinal profiles obtained from control, Aβ, tau, or αSyn flies. The combined dataset comprises 147 total samples, including bulk RNAseq in triplicate from fly heads of seven distinct genotypes and up to nine timepoints between 2 and 57 days of aging. Importantly, the network detection algorithm is implemented independent of sample age and genotype in order to identify gene sets that covary across a range of conditions. WGCNA identified 19 modules ranging from 32 to 4277 genes (Supplementary Fig. 3 and Supplementary Table 2). All 30 AMP-AD human co-expression modules significantly overlap with at least one Drosophila co-expression module (Fig. 2A and Supplementary Table 6). Several overlapping fly and human modules showed similar enrichment for cell-type specific markers and biological pathways (Supplementary Fig. 3 and Supplementary Table 7). For example, the Drosophila module fM1 is strongly enriched for neuronal markers and genes that mediate synaptic signaling (GO:0099536, p < 1.0 × 10⁻¹⁵) and overlaps significantly with PHGbrown (p = 1.7 × 10⁻⁵⁰). Overall, these results support conservation of gene regulatory networks in the adult Drosophila brain, including models of AD and related neurodegenerative disorders.

Fig. 2: Cross-species dissection of brain gene expression networks and behavior.

A Overlap between Drosophila WGCNA modules (rows) and AMP-AD human gene coexpression modules (columns). Black shading indicates significant overlaps based on the hypergeometric overlap test [false discovery rate (FDR) < 0.05]. The heatmap indicates the direction and magnitude of mean differential expression for module genes in either Aβ (elav-Gal4/+; UAS-Aβ/+) or tau (elav-Gal4/+; UAS-tau/+) models compared with controls (elav-GAL4/+), using linear regression including longitudinal data (days 2-28 for Aβ and days 2-42 for tau) and adjusting for age. Statistics are based on the likelihood ratio test with significance-level adjusted based on FDR (*, p < 0.05, **, p < 0.01). See also Supplementary Figs. 3, 4 and Supplementary Tables 6, 8. B Drosophila modules fM11 and fM1 are significantly up- and down-regulated, respectively, in response to Aβ and tau. Mean expression for module genes is shown for AD models (red, elav > Aβ or elav>tau) or elav-GAL4 controls (black). See also Supplementary Figs. 5, 6 and Supplementary Table 8. C Scatter plots demonstrate correlations between fM11/fM1 expression and climbing speed, based on Pearson correlation (significance adjusted based on FDR). The plots include data from 44 experimental conditions, including 7 genotypes at 7 timepoints between 2 and 28 days. The color key indicates the relevant AD models (red) and controls (black), with other conditions in grey, and age denoted by size of the points. Gray shading denotes the 95% confidence interval (standard error of the mean) for the best fitting line. See also Supplementary Fig. 7. D Expression of fM11 and fM1 partially mediate the impact Aβ and tau, respectively, on progressive CNS dysfunction (locomotor behavior). Mediation analysis considered the association between Aβ or tau and locomotor performance phenotype, decomposing into the component explained by module expression (mediated effect), and the unexplained component (direct effect). Proportion mediated indicates the fraction of the mediated effect to the total effect.

We hypothesize a causal chain in which disease triggers, such as Aβ, tau, and aging, modulate brain gene expression networks, which subsequently affect CNS function (Fig. 1A). In nested linear regression models considering both AD pathologies and aging, Aβ significantly perturbed six co-expression modules, whereas tau altered expression of one module (Fig. 2A, B, Supplementary Figs. 5, 6, and Supplementary Tables 6, 8). To model the link between brain transcriptional networks and CNS function/behavior, we incorporated locomotor performance data (negative geotaxis) generated from the same fly cohort used for RNA-sequencing. Overall, mean expression of 6 out of 19 transcriptional modules correlated significantly with locomotor performance (Fig. 2C and Supplementary Fig. 7), including networks enriched for genes that regulate the immune response (fM11), synaptic transmission (fM1), and endocytosis/vesicle trafficking (fM8) (Supplementary Table 7). Lastly, using multivariate mediation models [71], we asked whether expression levels of these modules explain the influence of Aβ and/or tau on progressive CNS dysfunction (Fig. 2D). Among our results, Aβ triggered up-regulation of module fM11, which in turn strongly mediated the locomotor impairment caused by Aβ (model beta reduced by 16%) (Fig. 2D, left). Interestingly, fM11 overlapped significantly with STGblue and other AMP-AD microglial modules and is similarly enriched for immune pathways, and these transcriptional networks are consistently up-regulated in human brains with AD pathology and the fly Aβ transgenic model (Supplementary Figure 4). Complementary analyses indicate more modest attenuation of tau-induced CNS dysfunction. For example, tau causes down-regulation of module fM1, concordant with PHGbrown, and in turn, partially mediates the locomotor impairment in tau flies (model beta reduced by 2%) (Fig. 2D, right). In sum, these cross-species results—integrating Drosophila AD models, behavioral assays, and brain transcriptome profiles—highlight conserved gene regulatory networks that respond to neurodegenerative pathology and may mediate progressive behavioral impairment.

Prioritizing network drivers

We hypothesized that AD-associated co-expression modules possess a fine-scale, hierarchical structure, in which selected gene drivers regulate the broader network [72]. We therefore sought to systematically prioritize candidate driver genes within each regulatory network. We considered three distinct but complementary bioinformatic criteria for ranking genes within each module (Fig. 3A). Since the 30 human AMP-AD modules were identified based on their enrichment for AD differentially expressed genes, we first considered the magnitude of differential expression, prioritizing genes within networks with the most extreme expression differences between AD cases and controls. Second, we ranked genes based on correlation with the first principal component of expression for each module, highlighting network “eigengenes” that best approximate the average module transcriptional activity. Lastly, we identified hub genes based on network connectivity, or degree, reflecting the number of edges due to pairwise co-expression of module gene members. Overall, the three selected parameters (extreme, average, and hub) define largely independent and non-overlapping criteria for ranking candidate key drivers within each module (Fig. 3A and Supplementary Table 9). Further, the nominated driver genes recapitulate the cell-type specific and functional signatures of each module (Supplementary Table 10). For example, the 614 prioritized drivers from PHGbrown, representing the top 10% of genes ranked using each criterion, were strongly enriched for synaptic signaling gene ontology terms (GO:0099536, p = 5 × 10⁻⁸⁵) similar to the entire module (n = 2123 total genes). Consistent results were also obtained when separately considering PHGbrown drivers nominated based on each of the distinct prioritization criteria (p_average=2.5 × 10⁻³³; p_extreme = 1.3 × 10⁻⁶; p_hub = 1.2 × 10⁻⁵⁴).

Fig. 3: Prioritization and screening for causal drivers from AD brain expression networks.

A Schematic network plot for the PHGbrown coexpression module, showing 3 independent and non-overlapping criteria for selection of candidate drivers for experimental manipulations. The top-ranked 10% (black) or 1% (yellow) genes are highlighted for each prioritization criterion (extreme, average, hub). See also Supplementary Table 9. B Plot shows summary statistics for results of all 1069 alleles tested, denoting the number of recovered enhancers (red) and suppressors (green) of Aβ- and/or tau-mediated locomotor behavior impairment. An enhancer or suppressor is defined as any allele that worsens or improves the locomotor impairment phenotype, respectively. Each modifier test included at least n > 4 replicates of 10 animals each examined at 6 or more timepoints over 28 days of aging. Statistical comparisons based on one-way ANOVA considering three nested models (genotype, genotype + time, and genotype*time) and applying a significance threshold of p < 0.05, following Holm-Bonferroni adjustment. See also Supplementary Table 11. C All three prioritization criteria identify causal modifiers of Aβ- and/or tau-mediated CNS dysfunction. Driver genes were based on consistent, statistically significant modifier effects from at least 2 independent allele strains. D Representative results from the screen of prioritized PHGbrown candidate drivers in the Aβ (elav-Gal4/+; UAS-Aβ42/+) and tau (elav-Gal4/+; UAS-tau/+) transgenic fly models. Plots show the results of longitudinal locomotor behavior assays (climbing speed during startle-induced negative geotaxis). The Drosophila gene homolog tested is noted above each plot, along with human gene in parentheses. The following RNAi knockdown or loss-of-function alleles were tested in heterozygosity: Nmdar2(NIG14794R-3III); Syt12(f00785); pcx(v29893); and CG43980(v107170). Natural spline curve with 3 degrees of freedom was fit to data, with shading to denote the 95% confidence band. See also Supplementary Fig. 8 and Supplementary Table 11.

Screening for causal drivers of Aβ- and tau-mediated neurodegeneration

Our preceding analyses identify conserved transcriptional modules as potential mediators of Aβ- or tau-triggered neurodegeneration, along with promising candidate key drivers within these networks. Nevertheless, without directed experimental manipulations, it is challenging to establish whether differentially expressed genes are upstream causal factors or downstream changes in response to pathology. We therefore next applied our cross-species strategy to determine whether prioritized driver genes and modules are causal (Fig. 1A), screening for in vivo modifiers of Aβ- and/or tau-induced progressive neuronal dysfunction. We focused on gene candidates representing the top 1% ranked by each driver criterion, calibrating our selections based on module size and overlaps (see Methods). Genes prioritized based on extremes of differential expression were less conserved from humans to Drosophila (55.7%) when compared with those selected based on average expression (66.1%) or hub-like connectivity (68.3%). Overall, among 357 prioritized, conserved, candidate drivers, Drosophila strains were available to evaluate 344 targets, including 87 extreme, 175 average, and 93 hub-like genes (Supplementary Table 9). Our screen thus considered an average of 27 (range=13–57) conserved candidate driver genes per module. Because of overlaps, a larger number of genes were in fact tested per module (mean=65, range=22–114), whereby predicted drivers from one module may be members of a related module from a different brain region. In most cases, we selected the single best fly ortholog for genetic screening based on the Drosophila RNAi Screening Center Integrative Ortholog Prediction Tool (DIOPT) [56], except for a few cases with multiple strongly conserved homologs. Thus, for the 344 human candidate genes, our screen evaluated a total of 350 Drosophila homologs. We obtained 1,069 genetic strains, including RNA-interference (RNAi) transgenic lines or other loss-of-function alleles (n = 955), as well as gain-of-function lines predicted to activate target gene expression (n = 114). We thus screened an average of 2.9 independent lines per target gene (range=1–11).

All lines were tested by crossing with Drosophila strains in which secreted Aβ or wild-type 2N4R tau are expressed throughout the nervous system, using the identical Elav > Aβ and Elav>tau models characterized above. We utilized an automated locomotor behavioral assay which is amenable for high-throughput genetic screening and detection of dose-sensitive interactions from genetic manipulations in either direction [38]. The climbing speed of adult flies was evaluated longitudinally between 1 and 3 weeks of age, scoring for enhancement or suppression of the locomotor phenotypes caused by Aβ- or tau-induced neuronal dysfunction. An enhancer or suppressor is defined as any experimental manipulation (e.g., RNAi or overexpression allele) that worsens or improves the Aβ- or tau-induced locomotor impairment phenotype, respectively. RNAi knockdown was activated pan-neuronally via the Elav-GAL4 driver, which also directs expression of Aβ/tau transgenes. All alleles were tested in heterozygosity. While Drosophila RNAi lines are designed for optimal specificity [73, 74], we considered genes as modifiers only when they were supported by consistent evidence from at least two independent RNAi strains or other alleles to minimize the possibility of off-target or genetic background effects.

Our primary screen identified a total of 134 Drosophila genetic modifiers, homologous to 141 predicted human gene drivers (Fig. 3B and Table 1, Supplementary Tables 9, 11). The overall modifier hit rate of 38% (134 out of 350) was significantly higher than that seen in previously published, unbiased modifier screens in Aβ or tau transgenic fly models (1–23%), albeit using somewhat different assays [65, 75,76,77,78]. Among 164 prioritized, down-regulated genes from AMP-AD human brain RNAseq data, we identified 63 candidate causal drivers based on concordant loss-of-function experimental genetic manipulations. Conversely, among 173 prioritized up-regulated genes from human brains, we identified 75 candidate causal drivers, with 31 genes having support from concordant gain-of-function genetic manipulations based on available activating strains and the rest based on loss-of-function experimental genetic manipulations. In 87 out of 350 fly genes evaluated, the availability of both loss- and gain-of-function alleles permitted reciprocal tests of the consequences for both down- and up-regulation of candidate drivers; 26 out of 134 modifier genes had evidence of opposing, bidirectional dose-sensitive interactions.

All three prioritization criteria led to the successful identification of causal modifiers of Aβ- and tau-induced progressive neuronal dysfunction. Proportionally, candidate drivers prioritized based on either extreme or average differential expression were somewhat more likely to be validated as modifiers (43.7 and 42.3%, respectively) than the hub-like genes (36.6%) (Fig. 3C). We identified similar numbers of Aβ (n = 93) and tau (n = 86) modifiers, including 38 genes that interact with both AD models following knockdown and/or activation. The screen highlights an average of 11 modifiers per module (range = 4 to 22 modifiers), pinpointing those regulatory systems comparatively enriched (or depleted) for candidate drivers with causal evidence (Table 1). PHGbrown had the highest overall modifier hit rate (53.8%, Fig. 3D), whereas TCXblue, enriched for genes with roles in developmental morphogenesis, had the lowest number of modifiers recovered from our screen (23.5%). While most modules identified modifiers of both fly models, others suggested specificity. For examples, DLPFCturquoise is characterized by disproportionate numbers of Aβ modifiers (n = 10) compared to tau (n = 2). Table 1 summarizes tallies of all recovered modifiers from each module, notable representative driver genes, and implicated pathways.

Table 1 AMP-AD coexpression modules and cross-species screen results.

Strains that enhanced both Aβ and tau were retested to examine the consequences for locomotor behavior in the absence of either AD pathologic trigger (e.g., elav > RNAi / +). Based on these results, we further classified enhancer alleles as either additive or synergistic modifiers (Supplementary Table 11). Additive modifiers influence locomotor behavior independent of Aβ or tau, making them excellent candidates to mediate alternate biological pathways important for neuronal maintenance with aging. We identified 14 Drosophila genes with modifier alleles that consistently enhanced Aβ or tau additively. Interestingly, module STGyellow, which is characterized by oligodendrocyte marker genes, was particularly enriched in causal drivers for which fly homologs showed CNS requirements independent of Aβ or tau (5 out of 25 genes evaluated), including MYRF, SCRIB, KIF13B, ST18, and PLPP1.

Defining and validating human AD causal subnetworks

AD-associated, consensus co-expression modules are large, ranging from 480 to 4519 gene members; therefore, our genetic screen examined a systematically prioritized subset of candidate driver genes. To define putative causal networks, partition subnetworks, and further define potential mechanisms, we next integrated our Drosophila genetic screen results with human network structure (Fig. 4A and Supplementary Figs. 9,10). Starting with the experimentally validated driver genes as seeds, we plotted all first degree neighbors (i.e., connected genes sharing edges) from the consensus co-expression network and also determined whether each driver and other member genes are up- or down-regulated in human postmortem brain tissue, based on the published AD case/control differential expression meta-analysis [2]. We next summarized whether the AD-associated perturbation of each validated driver gene (e.g., increased or decreased differential expression in human brains) is predicted to either promote or protect against disease pathogenesis (Fig. 1A), based on the corresponding results from our genetic manipulations of the Drosophila gene homolog in Aβ/tau models (Supplementary Table 12). Gene expression changes inferred to aggravate Aβ- and/or tau-mediated neurotoxicity were classified as disease amplifying/pathogenic gene expression changes. By contrast, differentially expressed genes inferred to mitigate Aβ- and/or tau-induced brain injury were classified as compensatory/protective genes. Together, these results define a collection of 30 AD causal subnetworks based on robust meta-analyses conducted in large human autopsy studies and integrated with results from systematic genetic manipulations to validate likely causal drivers (Fig. 4, Supplementary Fig. 10 and Supplementary Table 12).

Fig. 4: Alzheimer’s disease (AD) causal gene expression networks.

A Differential AD-associated gene expression and causal driver predictions are shown for the STGblue and PHGbrown networks. To derive subnetworks, we plotted human homologs of modifier genes (large oval nodes) along with first-order neighbors (small circle nodes). Edges indicate coexpression between genes. Left: Significant up- or down-regulated differential expression is indicated using red or blue, respectively, based on the AMP-AD case/control brain transcriptome meta-analysis. Right: Integrating results from the Drosophila modifier screen (enhancers or suppressors of Aβ/tau-induced locomotor impairment), we inferred whether drivers are amplifying or compensatory (red or blue, respectively). Amplifying gene expression perturbations promote disease pathophysiology, whereas compensatory changes are protective. Genes localized to established AD risk loci are highlighted in yellow. Genetic manipulations of fly homologs of ITPKB from STGblue, and SMAD4 and RAB6B from PHGbrown were compatible with bi-directional effects (colored gray). See also Supplementary Fig. 9,10 and Supplementary Table 12. B Mean expressions of STGblue and PHGbrown genes partially mediate the impact of AD pathology on cognition, based on clinical-pathologic data from the Religious Orders Study and Rush Memory and Aging Project (n = 636, mean age at death=88.7, age range=67-108, female sex=64%, AD = 40.1% non-demented=58%). Mediation analysis considered the association between global AD pathology (quantitative assessments of amyloid plaques and tau neurofibrillary tangles) and global cognitive performance in evaluations proximate to death (summary measure from 19 cognitive tests), decomposing into the component explained by module expression (mediated effect), and the unexplained component (direct effect). Proportion mediated indicates the fraction of the mediated effect to the total effect. Linear regression models relating pathology, cognition, and gene expression included covariates for age at death, postmortem interval, and sex. Significance testing based on 1000 simulations with the Benjamini-Hochberg false discovery rate (p_STGblue = 0.048; p_PHGbrown = 0.036).

We next systematically examined whether any of the causal subnetworks identify genes associated with AD risk (Supplementary Table 13). Leveraging publicly available summary statistics from two AD genome-wide association studies (GWAS) [31, 32], we deployed the multi-marker analysis of genomic annotation (MAGMA) tool [50], to examine whether each subnetwork is enriched for AD susceptibility variants. MAGMA computes an overall gene-set test statistic, including adjustments for gene size and regional linkage disequilibrium. Two subnetworks, STGblue and PHGturquoise, were consistently enriched for AD susceptibility variants (p < 0.05). Similar results were not observed in analyses of the full module, nor in subnetworks generated from all prioritized genes (Supplementary Table 13); therefore, integration of the Drosophila experimental results increases power to detect causal networks. Moreover, we did not detect similar enrichment when using GWAS summary statistics from either Parkinson’s disease [33] or height [34], consistent with specificity for AD risk.

The STGblue and PHGturquoise causal subnetworks are similarly enriched for microglial expression signatures (p_STGblue=1.4 × 10⁻⁴; p_PHGturquoise = 6.4 × 10⁻⁶) and genes that regulate innate immune mechanisms (GO:0006955, p_STGblue = 2.2 × 10⁻³²; p_PHGturquoise = 8.7 × 10⁻²⁹) (Supplementary Table 14). Indeed, AD GWAS meta-analysis results are similarly enriched for microglial and immune pathways [13] and these subnetworks are notable for including genes from AD susceptibility loci. For example, the STGblue subnetwork includes TREM2, HLA-DRB1, and SPI1, which all have roles in immune modulation (marked yellow in Fig. 4A). Interestingly, although these genes are poorly conserved in Drosophila, they are co-expressed with conserved, prioritized driver genes that were validated as modifiers in our screen. Genes in the STGblue subnetwork are predominantly up-regulated in AD [2]. Among the 8 STGblue modifier genes, we can infer 5 amplifying and 3 compensatory genes. For example, SASH3 is an STGblue hub driver candidate, encoding an immune signaling adapter protein [79]. Loss-of-function of the fly gene homolog SKIP, either through RNAi-mediated knockdown or heterozygosity for an insertional allele, suppressed Aβ and tau-mediated locomotor impairment (Supplementary Table 11). We therefore infer that up-regulation of SASH3 along with 4 other STGblue drivers promotes AD pathogenesis. Since most of our tested alleles were RNAi transgenes under the control of the pan-neuronal elav-GAL4 driver, our results are highly suggestive of cell autonomous, causal roles within neurons, besides the well-established microglial role of these genes in innate immunity. Indeed, prior studies in flies, mice, and human brains support the expression of immune regulators and effectors in many neuronal subtypes [64].

Among our screen results, PHGbrown had the highest hit-rate, with 14 out of 26 candidate driver genes modifying Aβ (n = 9) or tau (n = 10) neurotoxicity, with manipulations of 5 genes modifying both AD triggers (Figs. 3D, 4 and Supplementary Table 12). The resulting 167-gene PHGbrown causal subnetwork is significantly enriched for genes that regulate synaptic function (GO:0099536, p = 4.0 × 10⁻¹¹) and glutamatergic neurotransmission (KEGG:04724, p = 2.5 × 10⁻⁴) and is a predominantly down-regulated module in AD bulk brain tissue (Fig. 4A, Supplementary Fig. 11, and Supplementary Table 14). While the PHGbrown subnetwork was not directly associated with AD risk from our MAGMA analyses, many other endocytic genes have been identified at AD susceptibility loci, including several regulators of synaptic vesicle recycling [32]. Integrating PHGbrown differential expression with the results from screening, we inferred 13 compensatory and 3 amplifying gene expression changes. For example, GRIN2B encoding a subunit of the N-Methyl-D-Aspartate (NMDA) glutamate receptor, was prioritized for evaluation based on its ranking as a top, hub-like driver gene candidate. RNAi knockdown of Nmdar2, the conserved, Drosophila ortholog of GRIN2B, suppresses Aβ-mediated locomotor impairment, consistent with a possible compensatory role. Eleven other modifier genes showed similar relationships, several of which are also implicated in neurotransmission (e.g., FRMPD4/CG42788, RBFOX1/Rbfox1, ATP6V1C1/Vha44) [80, 81]. We confirmed that many of these genes show consistent interactions with Aβ or tau using an independent, histologic assay for age-dependent neurodegeneration in the adult Drosophila brain (Fig. 5 and Supplementary Fig. 12).

Fig. 5: PHGbrown causal drivers modulate adult brain degeneration in Drosophila models.

Compared with controls (elav-Gal4/+), pan-neuronal expression of Aβ (elav-Gal4/+; UAS-Aβ42/+) or tau (elav-Gal4/+; UAS-tau/+), causes age-dependent, progressive vacuolar degenerative changes (arrowheads). Manipulating homologs of PHGbrown causal drivers consistently modify Aβ- or tau-induced neurodegeneration based on adult brain histopathology. The following RNAi knockdown or loss-of-function alleles were tested in heterozygosity: Nmdar2(NIG14794R-3III); Lgr3(v330603); rdgC(306); dsd(v1106); Rab6(08323); CG42788(v45034); Rbfox1(NIG32062Ra-3); Med(v19688); Vha44(MI02871); and pcx(v29893). Vacuoles were quantified in at least n = 8 animals at 10 or 15 days of age for tau and Aβ, respectively (the negative control, at left, is taken from a 10-day old animal). Statistical analysis is based on one-way ANOVA followed by Dunnett’s test for multiple hypothesis correction. Each data point represents an individual biological replicate sample. Error bars denote the standard error of the mean. **, p < 0.01; ***, p < 0.001 **** p < 0.0001. See also Supplementary Fig. 12.

We next took advantage of clinical, pathologic, and RNAseq data from human postmortem brain tissue (n = 636 autopsies) in order to model the hypothetical causal chain between AD pathology, transcriptional networks, and downstream cognitive manifestations (Fig. 4B) [27, 52]. We focused these analyses on STGblue and PHGbrown, for which our cross-species strategy highlighted potential disease-amplifying versus protective roles, respectively, in AD pathogenesis. Our models incorporated a quantitative summary measure of AD neuropathologic burden as an upstream trigger along with global cognitive performance from assessments proximate to death as the outcome trait. Consistent with our mediation analysis, above, considering the Drosophila fM1 and fM11 modules (Fig. 2C), mean expression of either the homologous human PHGbrown or STGblue explained up to 9% of the variance in cognitive impairment attributable to AD pathologic triggers. These results are consistent with a model in which AD pathology alters brain transcriptional programs that in turn influence downstream cognitive manifestations of disease.

PHGbrown modulates neuronal hyperexcitability and resulting degeneration

A preponderance of evidence suggests that Aβ peptides boost glutamatergic neurotransmission, leading to aberrant Ca²⁺ influx, CNS hyperexcitation, tau hyperphosphorylation, and ultimately neurodegeneration [82,83,84,85,86]. Based on our findings, we reasoned that PHGbrown may represent a conserved transcriptional response triggered by hyperexcitability that modulates brain hyperactivation injury in AD. To test whether the Elav > Aβ transgenic flies manifest central nervous system hyperexcitation, similar to APP transgenic mouse models [87], we used the Transcriptional Reporter of Intracellular Ca2+ (TRIC) system [46], in which activity-dependent calcium influx drives expression of a green fluorescent protein (GFP) reporter. We detected a striking increase in GFP signal in the brains of 10-day-old Elav > Aβ animals (Fig. 6A and Supplementary Fig. 14). Next, in order to assess whether excitability mediates Aβ-induced neuronal injury, we experimentally manipulated VGlut, which encodes the vesicular glutamate transporter that is required for loading glutamate into synaptic vesicles. Introducing one copy of a strong VGlut loss-of-function allele or RNAi-mediated knockdown reduced vacuolar degeneration in Elav > Aβ flies (Fig. 6B and Supplementary Fig. 13A).

Fig. 6: PHGbrown modulates glutamate-induced hyperexcitability.

A Elevated Ca²⁺ (anti-GFP, white) is detected in Drosophila brains following pan-neuronal Aβ expression (elav-Gal4/+; TRIC-GFP/+; UAS-Aβ42 /+) versus control (elav-Gal4 /+; TRIC-GFP / +), based on whole-mount immunofluorescence of 10-day-old adult brains (Z-projection with maximum intensity). The VGlut overexpression model (VGlut^OE: VGlut-Gal4 /+; UAS-VGlut / TRIC-GFP) also shows elevated Ca²⁺ (4-day-old animals), compared with controls (VGlut-Gal4 / + ; TRIC-GFP / +). Full genotype of the TRIC-GFP reporter system: TRIC-GFP: LexAop2-mCD8::GFP, nSyb-MKII::nlsLexADBD, QUAS-p65AD::CaM / +; nSyb-QF2 /+. GFP signal intensity was quantified in at least n = 13 replicates, with statistics based on two sample t-tests. Data points represent individual biological replicates. Error bars denote the standard error of the mean (SEM). **, p < 0.01; **** p < 0.0001. See also Supplementary Fig. 14. B Introducing one copy of a VGlut loss-of-function allele dominantly suppressed Aβ-induced degeneration in adult Drosophila brains. Vacuoles (arrowheads) were quantified in at least n = 8 animals at day 15. Statistics is based on two sample t-test. Data points represent individual biological replicates. Error bars denote the SEM. **, p < 0.01. C VGlut overexpression (VGlut^OE: VGlut-GAL4 /+; UAS-VGlut / +) induces activity-dependent, vacuolar degeneration, and is suppressed by reducing PHGbrown driver gene homologs’ expression. The following alleles were tested in heterozygosity: Nmdar2(NIG14794R-3III); Lgr3(v330603); Rab6(08323); and Rbfox1(NIG32062Ra-3). Vacuoles were quantified in at least n = 7 animals at day 2. Statistical analysis is based on one-way ANOVA followed by Dunnett’s test for multiple hypothesis correction. Data points represent individual biological replicates. Error bars denote the SEM. **, p < 0.01; ***, p < 0.001; **** p < 0.0001. See also Supplementary Fig. 13B. D Differential expression of Drosophila fM1 module varies by cell-type and genotype. Uniform Manifold Approximation and Projection (UMAP) plots showing mean fM1 gene differential expression (log₂ fold change) across brain cell clusters, based on single nucleus RNA-sequencing in 10-day-old Aβ (elav-Gal4 /+; UAS-Aβ42 / +) or tau (elav-Gal4 / + ; UAS-tau / +) transgenic flies versus controls (elav-Gal4 / +). Clusters with significant expression changes (Wilcoxon rank sum test p < 0.05) are labeled, with glutamatergic neurons in bold. See also Supplementary Fig. 16 and Supplementary Table 17.

Since Aβ has been implicated in myriad deleterious effects in the brain, we next sought a more restricted, direct model of hyperexcitation brain injury. We adopted a previously published strain in which a UAS-VGlut transgene is activated within the endogenous VGlut expression domain via a VGlut-GAL4 driver [43, 44, 88]. These flies manifest increased glutamatergic excitatory neurotransmission, owing to excess glutamate packaging into synaptic vesicles, causing elevated neuronal activity, progressive neurodegeneration, and reduced survival [88]. We confirmed hyperexcitation and neurodegeneration in the VGlut overexpression model using TRIC imaging and brain histology, respectively (Fig. 6A and Supplementary Fig. 13B). In the brains of 2-day-old adult flies, we noted confluent vacuolar pathology in the outer medulla, which participates in visual processing and receives strong glutamatergic input from pathways originating with photoreceptors in the Drosophila retina. This degenerative pattern was strongly suppressed when flies were raised in complete darkness, consistent with activity-dependent excitotoxic brain injury (Fig. 6C). We found that loss-of-function manipulations in several PHGbrown driver genes, including GRIN2B/Nmdar2, RAB6/Rab6, RBFOX1/Rbfox1, and RXFP1/Lgr3, also significantly suppressed vacuolar degeneration following VGlut-induced excitotoxicity (Fig. 6C). Our results are consistent with a model in which decreased expression of the PHGbrown module attenuates hyperactivation neuronal injury in AD.

To determine if hyperexcitability can directly mobilize a transcriptional response similar to PHGbrown independent of AD pathology, we performed RNAseq profiling in the VGlut overexpression model at the same 2-day-old time point. Our analyses identified 2,864 significantly differentially expressed genes, including enrichment for regulators of synaptic signaling (GO:0099536, p = 2.2 × 10⁻¹⁴, Supplementary Fig. 15 and Supplementary Tables 15, 16), and this transcriptional signature significantly overlapped with both human PHGbrown (p = 7.5 × 10⁻⁶) and the homologous Drosophila fM1 co-expression network (p = 1.9 × 10⁻¹⁴). Unexpectedly, the direction of differential expression was reversed: VGlut-induced hyperexcitability increased expression of most synaptic genes, including GRIN2B/Nmdar2 and several other driver gene modifiers from PHGbrown (Supplementary Fig. 15C-D and Supplementary Table 16), but the PHGbrown co-expression module is predominantly down-regulated in AD postmortem bulk brain tissue (Fig. 4 and Supplementary Fig. 11). In our complementary analyses of the Drosophila fM1 synaptic co-expression network (above), we similarly found that gene expression was significantly decreased during aging and this was further reduced in Elav>tau flies (Fig. 2B and Supplementary Table 8). Although we observed a similar overall trend in Elav > Aβ animals (p = 0.059), our longitudinal data hint at a more complex pattern (Supplementary Fig. 5). Since AD has distinct effects across heterogeneous neuronal subtypes, we reasoned that many transcriptional perturbations might be masked or attenuated in bulk RNAseq data. Therefore, to confirm and define cell-type specificity, we analyzed single nucleus RNAseq data generated from 10-day-old Elav > Aβ flies, when elevated intracellular calcium levels suggest hyperexcitability. Interestingly, Aβ triggered a strong increase in the fM1 co-expression module expression in numerous cell types, including several excitatory glutamatergic and cholinergic neuron clusters, and similar changes were seen for PHGbrown gene homologs (Fig. 6D and Supplementary Fig. 16A). By contrast, tau predominantly induced down-regulation for these genes among most of the same cell types, although selected excitatory neuron clusters showed the opposite pattern similar to Aβ. Therefore, while AD pathology and hyperexcitability have overlapping gene expression signatures, our cross-species analyses reveal that these transcriptional programs may be transformed by aging and disease progression in a cell-type specific manner.

Lastly, in order to confirm and translate these findings, we interrogated recently-published snRNAseq data from human brain tissue [29]. In a study of 427 human postmortem brains, we find heterogeneous responses across diverse brain cell types similar to Drosophila, including significant upregulation of PHGbrown genes in many excitatory neuron subtypes in association with AD clinicopathologic traits (Supplementary Fig. 17A and Supplementary Table 17). In another pertinent study, cortical biopsies obtained during neurosurgery for normal pressure hydrocephalus were interrogated for AD neuropathology, electrophysiology, and single-cell transcriptomes [30]. Strikingly, we found a significant overlap between the PHGbrown gene set (p = 9.5 × 10⁻⁷), and predominantly up-regulated, differentially expressed genes mediating glutamatergic neurotransmission in excitatory neuron subtypes from cortical biopsies with early-stage Aβ pathology (absent tau tangles) and a hyperexcitable electrophysiologic signature (Supplementary Fig. 17B).

Discussion

Our cross-species, systems genetics approach has delineated the causal relationships between AD pathology, global transcriptional changes, and progressive CNS dysfunction. Among 30 consensus AD co-expression modules, we dissected the conserved transcriptional responses to Aβ, tau, alpha-synuclein, and/or aging—together, these triggers account for the majority of AD-associated gene regulatory changes in human brains. Further, through systematic genetic manipulations and behavioral screening, we pinpoint 141 driver genes with Drosophila homologs that can exacerbate or mitigate Aβ- or tau-mediated neurotoxicity. The STGblue module, implicated in innate immunity, is strongly up-regulated in AD, promotes neurodegeneration based on the results of fly genetic manipulations, and is significantly enriched for AD risk variants. On the other hand, the PHGbrown module, including regulators of synaptic transmission, is down-regulated in AD postmortem human bulk brain tissue and enriched for loss-of-function suppressors of Aβ/tau. Our findings suggest that changes in these synaptic regulatory genes respond to CNS hyperexcitability and modulate downstream AD brain injury. Overall, our strategy highlights how analyses of human gene expression data combined with genetic manipulations in more tractable, experimental models can pinpoint promising targets that aggravate or protect against disease.

Based on their common transcriptional signatures in humans and Drosophila models, aging and AD pathologic triggers appear to have broadly conserved impact on the brain. Whereas the human brain can only be profiled cross-sectionally following death, Drosophila models enable analysis of the transcriptome and behavior longitudinally to generate dynamic, multi-scale models of brain gene expression and function, in the presence or absence of Aβ or tau. The majority of shared, differentially expressed genes and coexpression modules show concordant directions in fly models and human brains with AD pathology, highlighting targets that may be directly responsive to Aβ or tau. Overlapping but discordant gene expression changes may also provide important clues to identify Aβ/tau-responsive transcriptional signatures that are transformed through feedback mechanisms impacted by aging, disease progression, or other confounders like co-pathology (below). We have thus produced a powerful longitudinal, cross-species molecular atlas for gene expression in AD and related dementias, which has been made broadly accessible through a web-based tool for further exploration and analyses (http://flynda.nrihub.org).

The high prevalence of mixed pathologies presents another major challenge for gene expression analysis in human postmortem brain tissue. Based on prospective, autopsy series, more than half of clinically-diagnosed AD cases are characterized by the presence of alpha-synuclein Lewy bodies, which may modify AD clinical manifestations [89]. In combination, we find that Aβ, tau, and alpha-synuclein together account for up to 62% of AD-associated gene expression changes, with the proportion of changes uniquely responsive to each pathology varying across the AMP-AD modules. Aging was an even stronger driver of brain gene expression changes, consistent with prior work [2, 26, 90, 91]. Together, aging and pathologic triggers for AD and Lewy body dementia collectively account for up to 86% of gene expression changes within AMP-AD modules. The residual unexplained gene expression signatures may relate to other pathologies that cause dementia and commonly co-occur with AD, such as TDP-43 and cerebrovascular lesions [89].

Synaptic dysfunction and loss in AD are strongly linked with clinical manifestations and disease progression [92,93,94]. In addition, a growing body of work suggests that AD is associated with hyperexcitability due to synapse and circuit level pathology, likely contributing to cognitive impairment and risk for seizures [95,96,97,98]. Gene expression profiles from human postmortem bulk brain tissue reveal prominent signatures of reduced synaptic and neuronal markers [2, 99]. PHGbrown was the top-ranked causal module based on its strong enrichment for modifiers of Aβ and tau-mediated neurotoxicity in Drosophila. This result strongly suggests that rather than simply a marker of cell death or synaptic loss, transcriptional perturbations affecting regulators of synaptic transmission may modulate AD pathophysiology.

Several additional experimental results and analyses reveal a causal chain linking AD pathology, hyperexcitability, dynamic gene expression changes, and neurodegeneration (Fig. 7). First, we establish that Aβ causes elevated calcium influx, consistent with neuronal hyperexcitability in the adult Drosophila brain. Second, we find that either expression of Aβ or more direct induction of glutamate excitotoxicity via VGlut overexpression triggers a transcriptional signature that strongly overlaps with PHGbrown. Third, experimental manipulations in Drosophila that reduce the expression of PHGbrown driver genes also suppress excitotoxic degenerative pathology in the brain. These gene expression changes appear to be dynamic, with either up- or down-regulation depending on the model (Aβ vs. tau), cell type, and/or time course of disease progression. Since Aβ aggregation into amyloid plaques precedes the formation of tau neurofibrillary tangles [100, 101], it is possible that the direction of PHGbrown module perturbation could correspond, in part, to the early Aβ and later tau phases of disease, along with pathologic spread from susceptible to more resilient cell types. Indeed, the aging brain likely comprises a complex transcriptional mosaic, with signatures from distinct cell types corresponding in part to different disease stages. Single-cell transcriptomes from Drosophila Aβ transgenic models or human cortical biopsies with early Aβ pathology (absent tau tangles) (Fig. S17) [30], reveal prominent activation of transcriptional signatures related to glutamatergic neurotransmission overlapping with PHGbrown. Complementary analyses of mouse AD models, including APP transgenics, are potentially consistent with an early up-regulation in mutliple AMP-AD co-expression modules enriched for neuronal/synaptic markers [2, 14]. Conversely, the prominent down-regulation of PHGbrown in the AMP-AD bulk RNAseq meta-analysis may reflect the dominant neuronal signatures present at later AD stages [2], similar to bulk and single-cell transcriptomes from Drosophila tau transgenic models. In sum, we propose a biphasic model in which AD pathology causes an early activation of PHGbrown that promotes neuronal injury followed by a decrease in expression that is compensatory, but may also reflect a contribution from synaptic and neuronal loss (Fig. 7).

Fig. 7: Schematic model for causal chain in Alzheimer’s disease.

In our biphasic model, early Aβ pathology causes an activation of PHGbrown that promotes neuronal injury (left), followed by a later tau-triggered decrease in expression that may be compensatory (right). Besides AD pathologic progression, the PHGbrown transcriptional program may be dynamically transformed by aging and cell-type specific mechanisms underlying vulnerability. Although the recovery of loss-of-function suppressors supports a coordinated, protective response to tau-mediated toxicity, it is possible that expression of PHGbrown may be further attenuated by neuronal dysfunction and death.

Toxic Aβ oligomers have been shown to directly increase glutamate release [85, 87, 102] and/or block reuptake [103], leading to CNS hyperexcitability in AD. Signaling by NMDA-type glutamate receptors appear to be a central mediator [86], and our cross-species screening strategy validated GRIN2B as a PHGbrown causal driver. Genetic manipulations reducing excitatory neurotransmission, including loss-of-function in either the Drosophila GRIN2B homolog, Nmdar2, or VGlut, which loads glutamate into synaptic vesicles, similarly suppressed Aβ-induced neurodegeneration. Notably, the NMDA antagonist, memantine, which directly binds and targets GRIN2B, is an FDA-approved AD therapy developed to counteract excitotoxic neuronal injury [104, 105]. While memantine has demonstrated benefit for AD in large clinical trials [106,107,108], the overall symptomatic benefit was modest and short-lived. It is possible that targeting glutamatergic neurotransmission would be more effective at earlier, pre-symptomatic AD stages. Alternatively, memantine efficacy may be transient and limited because GRIN2B is already down-regulated in the brain as part of a compensatory transcriptional response. Regardless, since GRIN2B is only one of more than 2,000 genes in the PHGbrown module, more robust neuroprotection may require simultaneous engagement of multiple drivers. In this manner, results from our screen could serve to guide the development of more effective, network-based AD therapeutics. In further support of potential translation, several PHGbrown causal drivers implicated by our cross-species studies in Drosophila have been independently identified as modifiers in large-scale screens conducted in human neurons (Supplemental Table 18) [109,110,111,112].

We note several important study limitations. Although we start from a robust set of 30 AMP-AD consensus co-expression modules derived from more than 2,000 human postmortem tissue samples [2], we acknowledge that bulk brain RNAseq may obscure important cell-type specific transcriptional changes, as exemplified for PHGbrown. In addition, we limited our screen to prioritized candidate drivers in order to accomplish a systematic survey of all AD-associated co-expression modules. In the future, it may be informative to probe selected networks, like PHGbrown, in greater depth, or to directly incorporate model organism gene expression data into the prioritization criteria. In most cases, our genetic screening was restricted to the single best predicted Drosophila gene ortholog, and we were only able to test for reciprocal, bidirectional experimental manipulations in a subset due to the more limited availability of gain-of-function lines that activate gene expression. Therefore, despite testing 350 fly genes and 1,069 strains, we may have still missed potentially interesting candidates. A related caveat comes from potential false negative screening results due to ineffective gene knockdown using available alleles. Given the importance of neuroglial interactions in AD pathophysiology, it will also be important to consider systematic manipulations of candidate drivers in other cell types besides neurons. Lastly, while well-controlled experimental manipulations in genetic models can powerfully inform inferences of causal relations between gene expression changes and neurodegenerative outcomes, caution is warranted in the interpretation of these data when translating back to complex systems, such as human pathophysiology. Nevertheless, our cross-species, systems genetic strategy begins to functionally dissect the complex transcriptional landscape in aging human brains, revealing mechanisms of AD excitotoxic brain injury and pinpointing other causal networks and drivers for potential therapeutic targeting.