A systems biology-based identification and in vivo functional screening of Alzheimer’s disease risk genes reveals modulators of memory function

Summary

Genome-wide association studies (GWAS) have uncovered over 75 genomic loci associated with risk for late-onset Alzheimer’s Disease (LOAD), but identification of the underlying causal genes remains challenging. Studies of induced pluripotent stem cell (iPSC)-derived neurons from LOAD patients have demonstrated the existence of neuronal cell-intrinsic functional defects. Here, we searched for genetic contributions to neuronal dysfunction in LOAD, using an integrative systems approach that incorporated multi-evidence-based gene-mapping and network analysis-based prioritization. A systematic perturbation screening of candidate risk genes in C. elegans revealed that neuronal knockdown of the LOAD risk gene orthologs vha-10 (ATP6V1G2), cmd-1 (CALM3), amph-1 (BIN1), ephx-1 (NGEF), and pho-5 (ACP2) alters short/intermediate-term memory function, the cognitive domain affected earliest during LOAD progression. These results highlight the impact of LOAD risk genes on evolutionarily conserved memory function, as mediated through neuronal endosomal dysfunction, and identify new targets for further mechanistic interrogation.

eTOC

Hudgins et al. use an integrative systems approach to search for genetic contributions to neuronal dysfunction in Alzheimer’s disease. Perturbation screening of candidate Alzheimer’s disease risk genes in C. elegans memory assays reveal their impact on evolutionarily conserved memory function and identify new targets for further mechanistic interrogation.

Introduction

Alzheimer’s Disease (AD), the most common cause of dementia, is an age-related neurodegenerative disorder that affects millions worldwide¹. Although our understanding of the molecular mechanisms underpinning the progression of AD has increased steadily over the past several decades^2,3, the precise etiology of the disease remains elusive, and no preventative or curative treatments currently exist^3,4. In recent years, large consortia-based genome-wide association studies (GWAS) have identified over 75 genomic loci associated with risk for sporadic late-onset AD (LOAD)^5–10, the predominant form (>90% of cases) of the disease. However, the majority of risk variants reside in non-coding regions of the genome and are enriched in cell type-specific transcriptional regulatory elements such as enhancers, suggesting that they contribute to genetic risk by altering gene expression regulatory networks^11,12. Yet, we still have a limited ability to predict how non-coding variants affect cell- and tissue-specific gene regulatory interactions that alter transcriptional outputs, confounding efforts to identify LOAD-causal variants and target genes^12–14, a critical step to fully realize the promise of GWAS for clinical applications.

Thus far, the genes and loci implicated in LOAD genetic risk have nominated multiple pathways for disease relevance, including endosomal trafficking, cholesterol regulation, mitochondrial function, myelination, vascular biology, inflammation and immunity, and APP and tau metabolism^15,16, most of which are active in multiple cell types in the brain. Due to the discovery of LOAD-associated coding variation in genes such as TREM2^17–19, PLCG2^17,20, and ABI3¹⁷, which are predominantly expressed in microglia, functional studies in both cell and animal models have increasingly been focused on the role of microglial biology with regard to genetic risk for LOAD and the pathways listed above^21–25. However, in contrast to previous work which had highlighted the importance of microglia-expressed genes to transcriptional network dysregulation in the LOAD brain²⁶, a recent co-expression network study of brain RNA-seq data from a large-scale LOAD cohort found neuron-specific co-expression modules to be the most profoundly affected by disease state²⁷. Additionally, in vitro studies of neuronal cultures derived from LOAD patient iPSCs have demonstrated several cell-intrinsic defects in neuronal function, including hyperexcitability and altered synapse formation dynamics, absent interactions with other cell types or exposure to external neurotoxic insults^28,29. The precise mechanisms through which common genetic variation contributes to neuronal cellular dysfunction and genetic risk for LOAD are understudied and remain largely unknown.

Here, we searched for genetic contributions to neuronal dysfunction in LOAD pathobiology by taking a systems biology approach. Using summary statistics data from a LOAD GWAS meta-analysis⁷ that is no longer the most recent, but that reports associations that have been replicated in the latest studies^9,10,16, we analyzed this data in the context of large-scale brain omics data, utilizing 1) multi-evidence-based gene-mapping; 2) transcriptome-wide correlation with clinical and neuropathological traits and network analysis-based prioritization; and 3) in vivo functional screening to identify high-confidence neuronal genes and pathways contributing to LOAD pathophysiology. We found that many candidate LOAD risk genes that are dysregulated in the LOAD brain and more strongly correlated with clinically-assessed cognitive function and dementia severity than with post-mortem assessment of neuropathological burden are central members of network modules involved in critical neuronal functions.

As modeling cognitive dysfunction in vitro presents considerable challenges, we chose to screen our candidate genes for LOAD-relevant effects in vivo, through the use of C. elegans associative memory assays, a well-established experimental paradigm of cognitive function assessment with evolutionarily conserved molecular underpinnings^30–32. C. elegans shares similarities with mammals in age-related physiological changes, including learning and memory decline³¹. Like mammals, memory loss is one of the earliest features of neuronal aging in C. elegans^31,33. Furthermore, conserved molecular machinery is required in C. elegans to learn and remember^30,31,34. A systematic perturbation screening of candidate risk genes in C. elegans revealed that neuronal knockdown of the LOAD risk gene orthologs vha-10 (ATP6V1G2), cmd-1 (CALM3), amph-1 (BIN1), ephx-1 (NGEF), and pho-5 (ACP2) significantly altered short/intermediate-term memory function, the cognitive domain affected earliest during LOAD progression, highlighting these genes for further in vitro and in vivo evaluation as potential therapeutic targets.

Results

Integrative multi-omics analysis for target gene identification and functional screening in vivo

To identify high-confidence target genes that underlie LOAD genetic risk and contribute to neuronal dysfunction in LOAD pathobiology, we analyzed LOAD GWAS summary statistics⁷ in the context of large-scale brain omics data, as outlined in Figure 1. Our analysis framework incorporates data from large-scale brain expression quantitative trait loci (eQTL) studies (PsychENCODE³⁵, CommonMind Consortium (CMC)³⁶, BRAINEAC³⁷, BrainSeq³⁸, ROSMAP³⁹, and GTEx⁴⁰), chromatin interaction data from various Hi-C analyses of brain and neural tissue (PsychENCODE - Dorsolateral Prefrontal Cortex (DLPFC)³⁵, Giusti- Rodríguez et al.- Adult and Fetal Cortex⁴¹, and Schmitt et al. - DLPFC, Hippocampus, and Neural Progenitor Cell⁴²), and both single-nucleus RNA-seq human brain data from the Religious Orders Study/Memory and Aging Project (ROSMAP)⁴³ and bulk RNA-seq data from a cohort of 215 brains from the Mount Sinai Brain Bank (MSBB)⁴⁴, two resources made publicly available as part of the Accelerating Medicines Partnership-Alzheimer’s Disease Consortium (AMP-AD). A key strength of our approach is the use of the C. elegans short/intermediate-term associative memory assay as an organismal level readout of the relevance of our prioritized candidate genes to neuronal circuit integrity and function.

Figure 1. Integrative systems biology approach for LOAD risk gene identification and functional screening.

(A) Candidate risk genes are identified from LOAD GWAS summary statistics, using functional genomics data from large-scale brain eQTL and chromatin interaction studies. (B) Relevance of candidate risk genes to LOAD biology is assessed by correlation of expression patterns with clinical and neuropathological traits, and connectivity within co-expression networks built from LOAD cohort brain RNA-seq data. (C) Prioritized candidate risk genes are screened for in vivo effects on memory function through the use of associative memory assays in C. elegans.

LOAD GWAS variants are enriched in open chromatin regions in neurons in addition to microglia

Previous studies have shown that LOAD SNP heritability is specifically enriched in transcriptional regulatory elements active in microglia^45–48, findings which have contributed to the recent focus on microglial biology in LOAD functional genomics studies. While clearly important to genetic risk for LOAD, this microglial enrichment does not explain the dysfunctional phenotypes observed in LOAD patient iPSC-derived neurons^28,29. To test for the presence of LOAD GWAS signal⁷ in neuronal transcriptional regulatory elements, we used single-cell open chromatin profiles generated from the human brain by Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq)⁴⁷, and a statistical enrichment methodology employed by Wang and colleagues⁴⁹ (see Methods). We found an enrichment of LOAD GWAS signal in the open chromatin of several neuronal cell types, over a wide range of statistical significance, from genome-wide significant (GWS) p-values (P<5×10⁻⁸) to a sub-GWS p-value of P=1×10⁻⁴, although the overall neuronal enrichment observed was much weaker than that seen in microglia (Figure 2A).

Figure 2. Data from eQTL and chromatin interaction studies implicates potential causal genes in LOAD GWAS loci.

(A) Enrichment signal for sub-threshold LOAD GWAS SNPs in neuronal open chromatin becomes evident following the removal of GWS loci and nearby SNPs (+/− 1 Mb), becoming similar in magnitude to that of microglia. Each point on the curves represents the difference in fold of the proportion of SNPs with a p-value below the cutoff in the ATAC-seq peaks versus all SNPs present in the GWAS summary statistics. (B) Numbers of candidate risk genes unique to, and shared by, the two gene-mapping methods. (C) Distribution of candidate risk genes by gene type and significance threshold. (D-E) Heatmap of expression patterns of candidate risk genes from (D) GWS LOAD-associated loci and (E) sub-GWS LOAD-associated loci across cell types in the human frontal cortex. Color scale represents relative expression across cell types (red = higher, blue = lower). Abbreviations: Endo, endothelial; Micro, microglia; Oligo, oligodendrocytes; Exc, excitatory neurons; Inh, inhibitory neurons; Astro, astrocytes; OPC, oligodendrocyte precursor cells. (F) Example LOAD GWAS locus (CELF1/SPI1), highlighting challenges in the identification of causal genes. Top to bottom – Manhattan plot of -log10(p-value) association statistics from Jansen et al., with the top SNP rs10437655 highlighted in purple and remaining variants colored according to LD (r²) with the lead SNP; Genome browser track showing all coding genes present in the locus. Gene names colored in green or blue are candidate risk genes nominated by QTL evidence or SNP-promoter interaction evidence, respectively. Gene names colored in red are candidate risk genes nominated by both kinds of evidence; Track showing the location of significant GWAS SNPs (P<1×10⁻⁵), and SNPs in LD (r²>0.6); Tracks indicating the positions of enhancer elements identified in different human brain cell types; Track illustrating the significant chromatin interactions between LOAD GWAS SNPs and gene promoters in the locus; Track illustrating the significant eQTL associations between LOAD GWAS SNPs and genes in the locus. See also Figure S1 and Tables S1–S3.

Since any enrichment of signal in the sub-GWS range could potentially be explained by linkage disequilibrium (LD) with above-threshold LOAD GWAS variants¹¹, we performed an additional enrichment analysis after removing all variants within 1 Mb of the GWS loci. Surprisingly, the enrichment of sub-GWS signal in neuronal open chromatin regions was significantly strengthened, a result observed for all neuronal subtypes in the scATAC-seq dataset, including both excitatory and inhibitory neurons (Figure 2A). In comparison, the same analysis using scATAC-seq data from the human lung⁵⁰ did not show enrichment in the open chromatin of lung cell types (Figure S1A). This result indicates that sub-threshold LOAD GWAS loci likely harbor causal non-coding risk variants in transcriptional regulatory elements active in neurons, which may lead to the dysregulation of causal risk genes underlying the dysfunction observed in LOAD patient iPSC-derived neurons. Thus, for our integrative approach outlined in Figure 1 we chose to include loci which reached a suggestive significance threshold of P<1×10⁻⁵, in addition to GWS loci, as this approach has been successfully used for post-GWAS gene mapping^49,51–53.

eQTL and chromatin interaction data nominate potential causal genes in LOAD GWAS loci

Increasing evidence suggests that the gene nearest to the most significant variant in a GWAS loci is often not the causal gene^54–57. To identify and prioritize candidate causal LOAD risk genes, we incorporated functional genomics data with summary statistics from the recent LOAD GWAS meta-analysis conducted by Jansen and colleagues⁷ using the web-based platform Functional Mapping and Annotation (FUMA)⁵⁸ (see Methods). We selected genes which were nominated by eQTL^35–40 or chromatin interaction data^35,41,42 (Figure 1) and disregarded genes that were only implicated through positional mapping. These datasets expand upon those used for gene-mapping in the original Jansen et al. study⁷, incorporating two more eQTL studies (PsychENCODE³⁵ and BrainSeq³⁸), and two additional Hi-C studies (PsychENCODE³⁵ and Giusti- Rodríguez et al.- Adult and Fetal cortex⁴¹), although unfortunately the most recent LOAD-relevant brain eQTL study⁵⁹ was not available through FUMA and thus was not included. In addition, we included those genes that contained protein-coding variants in LD (r² > 0.8) with the tag variant at each LOAD GWAS locus. This strategy nominated 1,630 coding and non-coding genes, in 29 GWS and 71 sub-GWS loci, as candidate causal LOAD risk genes (Figure 2C; Table S1). The majority of mapped genes were protein-coding, with lncRNAs and pseudogenes making up the next two largest categories, in roughly equal proportions in both GWS and sub-GWS loci (Figure 2C).

More candidate risk genes were mapped by variant-promoter chromatin interactions (n=1,353) than by eQTL evidence (n=542). In total, 282 genes (17%) were supported by both eQTL and chromatin interaction evidence (Figure 2B). The PsychENCODE and the Adult and Fetal cortex Hi-C data provided most of the chromatin interaction-implicated genes (Figure S1B; Table S2), and the majority of cis-eQTL-associated genes came from the PsychENCODE, GTEx, and CMC datasets (Figure S1B; Table S3). Using publicly available human frontal cortex single-nucleus RNA-seq (snRNA-seq) data from individuals from the Religious Orders Study/Memory and Aging Project (ROSMAP)⁴³, the most recent and largest AMP-AD snRNA-seq dataset available at the time of our analysis, we examined the expression patterns of the protein-coding candidates (n=957) and found that most of the candidate risk genes have some level of ubiquitous expression across all the major cell types, although most also exhibit higher levels of expression in one or two specific cell types, regardless of whether they came from GWS or sub-GWS loci, with no one cell type dominating (Figures 2D–E). Taken together with the findings of the scATAC-seq enrichment analysis, this result suggests that enrichment of LOAD genetic risk in cell type-specific transcriptional regulatory elements does not directly correspond to obvious cell type dominance in transcriptional signatures of functional genomics-nominated risk genes.

Of particular interest is the CELF1/SPI1 locus (11p11.2), which did not reach GWS in the Jansen et al. study⁷ but has been found as GWS in several previous studies^5,6,8, as well as the most recent LOAD GWAS⁹ (Figure 2F). Previous work has implicated SPI1 as the causal gene in the locus⁶⁰ based on the data from microglia. However, whether or not SPI1 is the only causal gene in the locus, and microglia are the only causal cell type, remains unclear. Overlaying the locus with recently generated brain cell type-specific epigenomic annotation data⁴⁶, we found that the region of LOAD association is rich with regulatory elements that are active in several cell types, including dense clusters of neuronal enhancers (Figure 2F). Our integrated analysis including eQTL and chromatin interaction data implicates almost all (n=29) of the protein-coding genes in the CELF1/SPI1 locus as candidate causal LOAD risk genes (Figure 2F), highlighting the challenges in identifying the true causal genes and relevant cell types underlying GWAS associations and the need for further prioritization and functional screening.

Neuronal co-expression network modules are enriched for candidate LOAD risk genes and are associated with dementia severity in LOAD.

Functional genomics-informed post-GWAS gene mapping cannot identify causal genetic risk genes alone, so we sought additional lines of evidence to support our candidate risk gene nominations. To determine the potential relevance of our candidate risk genes to the transcriptional alterations occurring in the LOAD brain, we performed weighted gene co-expression network analysis (WGCNA)⁶¹ on bulk RNA-seq data of the parahippocampal gyrus from a cohort of 215 human brains from the Mount Sinai Brain Bank (MSBB)⁴⁴. This robust and large-scale dataset has been previously successfully utilized for the identification of LOAD-relevant transcriptional co-expression relationships²⁷, and samples from this cohort span a wide spectrum of LOAD-related neuropathological and cognitive disease severities. From this analysis we identified 32 distinct co-expression modules (Figure 3A).

Figure 3. Neuronal co-expression network modules are enriched for candidate LOAD risk genes and are associated with dementia severity in LOAD.

(A) Co-expression network analysis of RNA-seq data from the parahippocampal gyrus identifies 32 distinct co-expression modules. Correlations between module eigengenes and traits are shown in the heatmap, with Pearson’s correlation and FDR-corrected P-values indicated for each significant association (FDR < 0.05). Colored bars indicate significant enrichment for cell type gene expression signatures (Fisher’s exact test, FDR < 0.05) for each module. (B) Overlap of protein-coding candidate risk genes whose expression is significantly associated with one or more clinical and neuropathological traits. (Pearson’s correlation, FDR < 0.05) (C) Significance of enrichment of protein-coding LOAD candidate risk genes in each module. Bar color corresponds to cell type gene expression signature enrichment, using the same color schema shown in (A). Grey bars = modules with no cell type enrichment. Bars extending above the line represent FDR < 0.05. (D-F) Violin plots of module scores across cell types in human brain snRNA-seq for the three candidate risk gene-enriched modules that were also enriched for neuronal gene expression signatures. Enrichment of cell type expression was assessed by Fisher’s exact test. ^#FDR < 0.05. (G-I). Bar plots of the top 10 enriched Gene Ontology biological process terms are shown for the candidate risk gene-enriched neuronal modules M9 (G), M10 (H), and M16 (I). The dotted vertical line in (G) corresponds to the significance threshold of FDR < 0.05 which was not met by any of the enriched terms for module M9. (J-L) Expression of the module eigengene decreases significantly with increased dementia severity for both module M9 (J) and module M16 (L), but not for module M10 (K). Pearson’s correlation and FDR-corrected P-values are indicated. Differences in the expression of the module eigengene at each CDR score with respect to cognitive baseline (CDR=0) was also assessed by t test. *P < 0.05, **P < 0.01, ***P < 0.001. (M) Gene expression correlation with CDR is significantly correlated with network connectivity as measured by kME. Pearson’s correlation and FDR-corrected P-value are indicated. Core network candidate risk genes, according to max kME, are shown in teal. The top 20 high-priority risk gene candidates, as determined by correlation with CDR and network centrality, are highlighted in orange. (N) Bar plot of significantly enriched (FDR < 0.05) Gene Ontology biological process terms are shown for the core network genes. From the top 30 significantly enriched terms, 10 non-redundant terms are shown. See also Figures S2 and S3 and Tables S4–S7.

We utilized the spectrum of neuropathology and cognitive function present across the MSBB dataset to identify significant associations between gene expression and disease severity. We assessed the correlations between both the expression of individual genes, and the expression of module eigengenes, and neuropathological category (CERAD), neurofibrillary tangle burden (Braak), and clinically-assessed cognitive function (CDR). On the level of individual genes, 14,392 genes were associated with at least one trait (FDR<0.05, Table S4), indicating large-scale rewiring of transcriptional activity in the LOAD brain. With regard to module-trait correlations, we found that 26 of the 32 modules were significantly associated with at least one trait (FDR<0.05, Figure 3A, Table S5). To assess for enrichment of brain cell type transcriptional signatures, we again utilized the ROSMAP snRNA-seq data⁴³, and assigned module scores (see Methods) to all nuclei in the dataset for each of the 32 co-expression modules. Using a module score threshold of > 0.1, enrichment of cell type expression for each module was assessed by overrepresentation testing, and all modules except for module M29 showed enrichment for a least one cell type (FDR<0.05), Figures 3A and S2). Co-expression modules with significant enrichment for neuronal transcriptional signatures were generally negatively correlated with CERAD category, and Braak and CDR scores, while glial and endothelial modules generally showed positive correlation with these traits (Figure 3A; Table S5). We then assessed if expression of our candidate LOAD risk genes were associated with any of the clinical and neuropathological traits and whether any of the modules were enriched for our candidate LOAD risk genes, considering only protein-coding genes. Out of the 825 protein-coding candidate risk genes that were present in the MSBB RNA-seq data, 520 (63%) were significantly associated with at least one trait (Figure 3B). In total 5 modules showed enrichment for our candidate risk genes, including 3 modules enriched for neuronal signatures (M9, M10, and M16; Figures 3C and 3D–F), the module with the strongest enrichment for microglia signature, module M28 (Figures 3C and S2), whose members include the well-supported LOAD risk genes CD33^62,63, TREM2^17,19, and PLCG2^17,20, and module M29, which was the only module not enriched for any cell type gene expression signature (Figures 3C and S2). Gene ontology (Biological Process) enrichment analysis of the 3 modules enriched for both candidate risk genes and neuronal signatures (Figures 3D–F) found only nominal enrichment for proteostasis and mitochondrial function terms for module M9 (Figure 3G; Table S6), significant enrichment of synaptic transmission and neuron morphogenesis terms for module M10 (Figure 3H; Table S6), and significant enrichment of ribonucleoprotein and mitochondrial function terms for module M16 (Figure 3I; Table S6). For the modules enriched for candidate risk genes and neuronal signatures the associations with dementia status were more pronounced than the associations with neuropathological burden, with module M9 showing moderate negative correlation with CDR score and weaker negative correlation with Braak score and CERAD category (Figures 3A and 3J; Table S5), module M10 showing no significant correlation with any trait (Figures 3A and 3K; Table S5), and module M16, the neuronal and endothelial signature-enriched module with the strongest enrichment for candidate risk genes (Figure 3C) showing robust negative correlation with all three traits, with the strongest association with CDR score (Figures 3A and 3L; Table S5).

Since our primary focus was to identify neuronal contributions to LOAD genetic risk, we wanted to determine if the relationships we observed between expression of the neuronal signature-enriched modules and the clinical and neuropathological traits merely reflected neuronal loss or were an actual readout of lower neuronal expression. To explore this, we investigated whether module scores for the neuronal co-expression modules differed between control and LOAD patient neurons in the ROSMAP snRNA-seq data. We found significant differences in module score between case and control neurons in the snRNA-seq data for every neuronal module that was significantly associated with the LOAD clinical and neuropathological traits in the bulk RNA-seq data (Figures 3A and S3). This included significantly decreased module scores in LOAD patients, in comparison to controls, for module M2, the neuronal module with the strongest negative correlation with the clinical and neuropathological traits, in both excitatory and inhibitory neurons (Figure S3A), as well as significantly decreased module scores in LOAD patients, in comparison to controls, for the candidate risk gene and neuronal signature-enriched modules M9 and M16 in inhibitory, but not excitatory, neurons (Figures S3C and S3D). These findings support the conclusion that transcripts of genes in neuronal signature modules decrease during LOAD progression, and that the correlations seen between these modules and the clinical and neuropathological traits present in the LOAD cohort data do not merely or only reflect neuronal loss and are thus relevant to our understanding of neuronal gene expression dynamics during LOAD pathobiology.

Next, we wanted to use the results of the co-expression network analysis to identify the most promising candidate risk genes for functional testing. Since candidate risk genes with higher centrality in the network are more likely to have disease-relevant effects if perturbed^64–66, we focused on candidates which occupied centrally-connected nodes within the overall co-expression network. We identified “core” genes as those positioned in the top 10% of the network, as determined by the eigengene-based connectivity measure kME (see Methods). Interestingly, overall connectivity in the co-expression network displayed a strong correlation with gene-trait association (Figure 3M) so that genes with the highest absolute correlation with CDR score were more likely to have high network centrality. Furthermore, core genes were enriched for biological process terms involved in neuronal functions, including synaptic plasticity, synaptic vesicle cycle, and synapse organization, as well as mitochondrial biology-related terms such as ATP metabolic processes, and protein localization to the membrane (Figure 3N), all terms that were also shared by the neuronal candidate risk gene-enriched modules (Figures 3G–I; Table S6). Since our neuronal signature modules were more strongly associated with CDR score than with neuropathological burden (Figures 3A and 3J–L; Table S5), we ranked these core network candidate risk genes by absolute correlation with dementia status and prioritized the top 20 as high-priority targets for functional validation (Table S7). Notably, eighteen of these top 20 candidates were either not the genes usually nominated from their respective loci or were genes that came from sub-GWS loci^6,7. The two exceptions were the well-replicated LOAD GWAS gene PTK2B, and the familial AD gene APP (Table S7), which resides in a locus that reached the suggestive association threshold (P<1×10⁻⁵)⁷. As a result of the prioritization process, all 20 of the high-priority candidates were members of neuronal signature-enriched modules (M2, M16 and M32, Figures 3A, 3F, and S2), with four candidates coming from M16, the module with the strongest enrichment for candidate LOAD risk genes (Figure 3C).

Previous findings from network-based analysis of LOAD transcriptomic data have highlighted the disease-relevance of modules representing aspects of microglia²⁶ and oligodendrocyte⁶⁷ biology. Most recently, a co-expression network analysis of the same RNA-seq dataset we analyzed here, utilizing a different methodology, determined that neuronal modules were the most significantly affected by LOAD pathobiology²⁷. A key difference and advantage of our study is the use of genetic association as the fundamental basis of our prioritization schema, upon which we leverage network approaches to derive new insights from LOAD brain transcriptome data. It has been recognized that genetically-supported drug targets have a much greater chance of success in clinical trials⁶⁸. By using genetics as a foundation, we increase confidence in our prioritized risk genes while also increasing the probability of successful therapeutic development. Indeed, our analysis confirmed that the significant disease-relevant neuronal modules contain well-supported AD risk genes, including the familial AD gene APP and the APP processing pathway member SORL1^69,70 in module M2, which was the most strongly associated with dementia status (CDR) (Figure 3A; Table S4).

Identification of high-priority candidate causal risk genes for functional screening in vivo

The network analysis identified important relationships between genes and modules involved in synapse and mitochondrial biology, critical components of healthy neuronal function, and LOAD. Since our candidate LOAD risk genes were enriched in these important neuronal function co-expression modules, which were more closely correlated with dementia status than with neuropathological burden, we chose to functionally screen our candidates for effects on memory, in a non-amyloidosis model, in neurons in vivo. The C. elegans short/intermediate-term associative memory (S/ITAM) assay was chosen as the ideal experimental paradigm due to the highly evolutionarily conserved molecular biology which underpins memory function from worms to mammals^30–32, as well as the practicality and efficiency the model affords, allowing for the testing of large numbers of candidates.

We aimed for both un-biased and targeted screening of our candidate risk genes and included 4 categories of criteria for candidate selection (Table 1. Candidate LOAD risk genes screened for memory function and functional screen results.). 1) High-priority candidate risk genes. This category is comprised of the top 20 highest-priority candidate risk genes that were prioritized as a combined result of the post-GWAS functional genomics nomination and the LOAD cohort bulk brain RNA-seq co-expression network analysis (Table S7). 2) CELF1/SPI1 locus candidates. A major challenge for post-GWAS gene mapping is the identification of target genes in gene-rich loci that may contain multiple potentially causal genes. Since the expression of 18 of the 29 eQTL- and chromatin interaction-implicated genes in the CELF1/SPI1 locus (Figure 2E) were significantly correlated with CDR score (Tables 1 and S4), including one of the top high-priority candidates, CKAP5, we selected multiple additional members (n=4) of this locus to screen for potential memory effects. 3) Well-studied LOAD GWAS genes. We selected two of the best studied GWS LOAD GWAS genes, BIN1 and PICALM, based on recent fine-mapping analyses using cell type-specific approaches^46,47 and their known role in synaptic function and memory^71–73, as strong candidates which could act as surrogate positive controls, as we would expect perturbation of their expression to affect memory function. 4) Candidates unsupported by prioritization schema. Since any effects on memory function we observed in our screen could conceivably occur due to perturbation of important neuronal genes that coincidentally exist in LOAD GWAS loci but have no actual relevance to LOAD genetic risk, we included genes from LOAD GWAS loci that were not prioritized by our analysis (RAPSN and GRIN3B (not present in MSBB RNA-seq data); GNB2 and TRPM7 (not correlated with CDR score); and CHL1, GDE1, and RORA (previously identified sub-GWS LOAD GWAS loci⁵ which did not meet significance criteria⁷)), to act as surrogate negative controls. To identify the appropriate targets for our 33 prioritized candidate LOAD risk genes (Tables 1 and S7), we used the web-based comparative genomics tool OrthoList 2⁷⁴ to identify the closest C. elegans orthologs for our perturbation screen. Keeping only those genes with orthology predictions supported by more than one database, we found 27 well-supported orthologs for 24 of our candidate risk genes (Tables 1, S8, and S9). As a final layer of prioritization, we selected orthologs which had been shown to be expressed in C. elegans neurons, as determined by our previous work characterizing the C. elegans neuronal transcriptome⁷⁵ (Tables 1 and S8), leading to a total of 27 worm orthologs of 24 LOAD GWAS candidate risk genes in 17 loci for in vivo functional screening.

Table 1.

Candidate LOAD risk genes screened for memory function and functional screen results

Gene	Chr	GWAS locus	eQTL evidence	Chromatin interaction evidence	*Cell type expression	Module	CDR correlation	CDR correlation FDR p-value	Braak correlation	Braak correlation FDR p-value	CERAD correlation	CERAD correlation FDR p-value	Max kME	C. elegans ortholog(s)	Orthology databases	**Effect on memory function
High-priority candidate LOAD risk genes
ATP6V1G2	6p21.32	GWS	✓		Below threshold	M2	−0.47	2.75E-10	−0.46	2.06E-09	−0.44	1.10E-08	0.92	vha-10	3	↓
MAPK9	5q35.3	sub-GWS		✓	Ex, In	M2	−0.47	3.27E-10	−0.45	4.31E-09	−0.44	8.05E-09	0.94	jnk-1	5	—
CISD1	10q21.2	sub-GWS		✓	Ex, In	M2	−0.45	1.76E-09	−0.42	1.97E-08	−0.40	1.20E-07	0.96	cisd-1	5	a
AP2S1	19q13.32	GWS		✓	Ex	M16	−0.43	9.92E-09	−0.41	4.72E-08	−0.39	2.20E-07	0.90	aps-2	6	b
CKAP5	11p11.2	sub-GWS		✓	Ex, In	M2	−0.42	1.24E-08	−0.39	1.72E-07	−0.34	7.80E-06	0.89	zyg-9; F54A3.2	4;2	—; b
CALM3	19q13.32	GWS		✓	Ex, In, Oli, OPC, Ast, Mic, End	M16	−0.42	1.41E-08	−0.36	1.34E-06	−0.39	3.42E-07	0.84	cmd-1	5	↓
PTK2B	8p21.1	GWS		✓	Ex	M32	−0.41	2.45E-08	−0.38	4.72E-07	−0.37	7.88E-07	0.85	kin-32	4	a
APP	21q21.3	sub-GWS		✓	Ex, In, Oli, OPC, End	M2	−0.41	2.69E-08	−0.43	1.61E-08	−0.40	1.64E-07	0.83	apl-1	5	—
POP7	7q22.1	GWS		✓	Below threshold	M16	−0.41	3.60E-08	−0.40	1.02E-07	−0.39	2.83E-07	0.88	Y62E10A.2	5	—
NGEF	2q37.1	GWS	✓		Ex	M32	−0.41	3.94E-08	−0.37	8.57E-07	−0.38	4.98E-07	0.88	ephx-1	2	↑
SLC25A11	17p13.2	GWS		✓	Below threshold	M16	−0.40	5.11E-08	−0.35	2.64E-06	−0.40	1.78E-07	0.89	misc-1	6	—
CELF1/SPI1 locus candidates
ACP2	11p11.2	sub-GWS		✓	Below threshold	M16	−0.26	6.54E-04	−0.20	0.01	−0.18	0.02	0.54	pho-5; pho-14	5;3	↑; —
CKAP5	See above
MTCH2	11p11.2	sub-GWS	✓	✓	Ex	M2	−0.30	6.87E-05	−0.25	1.40E-03	−0.25	1.36E-03	0.64	mtch-1	6
NDUFS3	11p11.2	sub-GWS	✓	✓	Ex, In	M16	−0.43	5.85E-09	−0.40	1.36E-07	−0.37	1.32E-06	0.82	nuo-2	5	—
PSMC3	11p11.2	sub-GWS		✓	Below threshold	M16	−0.35	2.65E-06	−0.35	3.27E-06	−0.32	3.33E-05	0.82	rpt-5	6	c
Well-studied LOAD GWAS genes
BIN1	2q14.3	GWS	✓	✓	Ex, In, Oli, Mic	M29	0.20	9.39E-03	0.16	0.04	0.15	0.06	0.65	amph-1	5	↑
PICALM	11q14.2	GWS	✓	✓	Ex, In, Oli, OPC, Ast, Mic, End	M1	0.07	0.42	−0.01	0.94	0.06	0.45	0.88	unc-11	4	c
Candidates unsupported by prioritization schema
RAPSN	11p11.2	sub-GWS	✓	✓	—	—	—	—	—	—	—	—	—	rpy-1	6	—
GRIN3B	19p13.3	GWS	✓		Below threshold	—	—	—	—	—	—	—	—	nmr-2	2	—
GNB2	7q22.1	GWS	✓	✓	Ex, End	M29	0.03	0.77	0.03	0.74	0.02	0.80	0.54	gpb-1	5	c
TRPM7	15q21.2	sub-GWS	✓		Ex, In, Ast	M1	0.05	0.57	0.07	0.43	0.10	0.21	0.75	gtl-2	4	—
CHL1	3p26.3	#	—	—	Ex, In, OPC, Ast	M2	−0.47	4.11E-10	−0.41	3.55E-08	−0.41	5.43E-08	0.95	lad-2	3	—
GDE1	16p12.3	#	—	—	Below threshold	M2	−0.30	9.73E-05	−0.27	5.05E-04	−0.23	3.31E-03	0.69	T09B9.3	2	—
RORA	15q22.2	#	—	—	Ex, In, OPC, Ast, End	M6	−0.08	0.37	−0.04	0.66	−0.02	0.83	0.65	nhr-23	3	—

^*Cain et al., Nat. Neurosci. 2022; Expression threshold: expressed in >10% of nuclei in a given cell type, Ex - excitatory neuron, In - inhibitory neuron, Oli - oligodendrocyte, OPC - oligodendrocyte precursor, Ast - astrocyte, Mic - microglia, End - endothelial

^**↓, decrease; —, no difference; ↑, increase; a, No RNAi Clone; b, Chemotaxis defect; c, Motor defect

^#Lambert 2013

Neuron-specific knockdown of LOAD risk gene orthologs alters memory function in C. elegans.

Since the expression of all our candidate genes were negatively correlated with LOAD severity, with the exception of BIN1 and PICALM (Table 1), we knocked-down the candidate genes using RNAi to mimic the directional impact of association. To generate neuronal-specific knockdown of candidate genes, we used a neuronal RNAi-sensitive strain (LC108) of C. elegans, which can otherwise be refractory to RNA interference. We knocked-down each of the LOAD candidate risk gene orthologs from egg stage and tested for effects on short/intermediate-term memory (at 1 hour and 2 hours post-conditioning) at day 1 (young adulthood). Knockdown of most of the candidate genes had no effect on naive chemotaxis (Figures S4A–G), suggesting that they did not alter normal neuronal development or function, with the exceptions of F54A3.2 (CKAP5; decreased naive chemotaxis), and aps-2 (AP2S1; increased naive chemotaxis), which unfortunately prevented robust assessment of any potential memory effects for these two high-priority genes. In addition, rpt-5 (PSMC3), unc-11 (PICALM), and gpb-1 (GNB2) could not be assayed for memory effects due to motor deficits resulting from knockdown. Finally, two of the high-priority candidates, kin-32 (PTK2B) and cisd-1 (CISD1), could not be tested for memory effects due to a lack of available RNAi clones in the Ahringer and Vidal libraries. However, their presence in our list of top candidates gave us further confidence in our prioritization schema because of previous findings from functional studies of these genes. In mice, PTK2B, a well-known LOAD GWAS gene, has been shown to have important roles in hippocampal-dependent memory, synaptic plasticity, and dendritic spine structure⁷⁶, and deficiency of CISD1, a gene involved in mitochondrial function, has been shown to elicit Parkinsonian phenotypes⁷⁷.

Among the high-priority candidates that could be tested in the memory assays (Table 1), knockdown of vha-10 (ATP6V1G2), cmd-1 (CALM3), and ephx-1 (NGEF) caused significant impacts on memory function (Figures 4A, 4B, and 4D), while knockdown of jnk-1 (MAPK9), apl-1 (APP), Y62E10A.2 (POP7), misc-1 (SLC25A11), and the other CKAP5 ortholog, zyg-9, had no effect (Figures 4A–C). Knockdown of our top candidate vha-10 (ATP6V1G2) resulted in a robust memory deficit at 1hr post-conditioning (Figure 4A). ATP6V1G2 encodes a neuronal-specific subunit of the vacuolar-type ATPase (V-ATPase), a proton translocating pump that plays critical roles in the acidification of endosomal compartments including lysosomes⁷⁸ and the loading and release of synaptic vesicles⁷⁹. Similarly, knockdown of cmd-1 (CALM3), encoding the calcium-binding protein calmodulin, resulted in a significant memory deficit at 1hr and 2hr post-conditioning (Figure 4B). In the brain calmodulin has diverse functions, including the regulation of synaptic signaling, endocytosis, cholesterol metabolism, and ion channel function⁸⁰. Interestingly, knockdown of the high-priority risk candidate ephx-1 (NGEF) and the well-known LOAD GWAS risk gene amph-1 (BIN1) had no effect on short/intermediate-term memory function, but instead resulted in increased memory retention at 2hr post-conditioning (Figure 4D). These results indicate that neuronal loss of expression of these genes impacts processes of active forgetting that are mediated through RAC1/CDC42^81,82. Indeed, NGEF encodes a neuronal guanine nucleotide exchange factor (GEF) that regulates the activity of GTPases such as RAC1, RHOA, and CDC42⁸³, and recent work has implicated BIN1 in RAC1-mediated synaptic remodeling⁸⁴.

Figure 4. Neuronal knockdown of LOAD risk gene orthologs alters memory function in C. elegans.

(A-G) 1 hour and 2 hour post-conditioning learning indices of worms treated with whole-life RNAi for LOAD candidate risk gene orthologs. Grouping of the tested orthologs was random and does not represent candidate prioritization. n ≥ 4 (n: technical replicates). Statistical significance determined by One-way ANOVA, with Dunnett’s post hoc test. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. See also Figure S4 and Tables S8 and S9.

Among the 5 gene orthologs from the gene dense CELF1/SPI1 locus (Table 1) that did not meet the criteria for inclusion in the list of high-priority candidates, mtch-1 (MTCH2), nuo-2 (NDUFS3), rpt-5 (PSMC3), and two orthologs of ACP2, pho-5 and pho-14, knockdown of mtch-1, nuo-2, and pho-14 showed no significant memory effects (Figures 4A, 4E, and 4G). However, knockdown of pho-5, the closest ortholog of the lysosomal acid phosphatase gene ACP2, resulted in a memory retention effect at 2hr post-conditioning, similar to what we observed for ephx-1 and amph-1 (Figure 4F). This result suggests that, like NGEF and BIN1, ACP2 is also involved in the process of active forgetting, possibly through the local turnover of synaptic proteins during dendritic spine remodeling, a process recently found to involve neuronal-activity dependent lysosome trafficking⁸⁵.

In total, out of 22 LOAD GWAS candidate risk genes (25 worm orthologs) in 15 loci tested, we identified 5 genes in 5 loci as in vivo modulators of memory function (Table 1). Taken together, the results of our systematic perturbation screen indicate that LOAD genetic risk impacts neuronal function, particularly with respect to memory, through two primary avenues – the synapse (ATP6V1G2, CALM3, BIN1, NGEF), and the lysosome (ACP2, ATP6V1G2, CALM3). The common point of interaction between these two fundamental components of neuronal biology is the endosomal trafficking system. Pathway and gene set analyses have previously found a significant enrichment of LOAD-associated genetic variation in genes involved in endolysosomal function⁸⁶. However, the involvement of the endolysosomal system in LOAD pathobiology has typically been conceptualized in the context of amyloid and tau biology^87–89. Our findings indicate that genetic contributions to neuronal dysfunction in LOAD pathobiology can affect the endolysosomal system through mechanisms which do not involve amyloid and tau, but instead directly impact the evolutionarily conserved pathways of learning and memory.

Discussion

Studies of the genetic underpinnings of LOAD continue to uncover new genomic loci of interest but identifying the responsible genes and translating genetic discoveries into druggable targets remains a major challenge for the field. In this study we searched for the genetic contributions which underlie neuronal dysfunction in LOAD pathobiology, using an integrative systems approach that incorporated multi-evidence-based gene-mapping and network analysis-based prioritization, with the C. elegans short/intermediate-term associative memory assay as an organismal level readout of the impact of our prioritized candidate risk genes on neuronal circuit integrity and function.

We compiled and employed a large array of functional genomics data to identify candidate risk genes from LOAD GWAS loci. Examined in the transcriptional context of the LOAD brain, we found significant associations between many candidate risk genes and phenotypic measures of cognitive dysfunction and LOAD neuropathology. Network analysis identified several neuronal signature-enriched co-expression modules that were the most significantly associated with LOAD-associated cognitive dysfunction. We prioritized candidate risk genes by using genetic association and functional genomics evidence, focusing on core genes in the co-expression modules, particularly those that were highest ranked for dementia association. All of the prioritized genes came from either pure neuronal signature modules (M16, M32) or modules that were enriched for multiple cell type signatures, including neuronal (M16). Some caveats to our study include the use of LOAD GWAS data that is no longer the most recent⁹, and the lack of inclusion of the largest LOAD-relevant brain eQTL study⁵⁹ due to its absence from the list of data available to FUMA⁵⁸. These caveats noted, a limitation of functional genomics-enabled post-GWAS gene mapping is the possibility of false positive gene nominations due to such factors as non-causal overlap between QTL and GWAS associations and non-disease-relevant promiscuous chromatin interactions between GWAS variants and gene promoter regions. This limitation persists regardless of the quality and comprehensiveness of the tools and datasets used. Because of this fact, functional follow-up is critical to gaining confidence in a set of GWAS-implicated genes. Thus, we conducted functional studies of the prioritized candidate neuronal risk genes, as well as low-priority and non-prioritized genes, for effects on in vivo memory function in C. elegans. Testing 25 worm orthologs out of 22 LOAD GWAS candidate risk genes in 15 loci, this study is to our knowledge the first comprehensive functional screen of its kind. The most notable finding of this study is the identification of 5 LOAD risk genes, ATP6V1G2, CALM3, BIN1, NGEF, and ACP2, in 5 loci, as in vivo modulators of evolutionarily conserved memory function.

ATP6V1G2 encodes a neuronal-specific subunit of the large V-ATPase complex. Our analysis prioritized ATP6V1G2 as our top candidate risk gene, both due to its membership within the core network of genes, as well as being the candidate most significantly associated with cognitive function. ATP6V1G2 has not been previously nominated as a LOAD risk gene, most likely due to the fact that it resides in the 6p21.32 major histocompatibility (MHC) locus, a region well-known for having an extremely complex LD structure that makes the identification of causal variants and genes in the locus particularly difficult. Multiple members of the V-ATPase complex are associated with neurological disorders and neurodegenerative conditions arising due to defective lysosomal acidification⁹⁰. Additionally, V-ATPase function has also been shown to be important for the maintenance of neural stem cell renewal capacity⁹¹, and the age-related loss of this capacity is also implicated in impaired cognitive function⁹². In support of our findings, a network-based study found ATP6V1A (3q13.2), another member subunit of V-ATPase, to be one of the top drivers of neuronal function that is dysregulated in the LOAD brain²⁷. Furthermore, testing for LOAD relevance in a Drosophila model of Aβ pathology, the authors of the study²⁷ found that Vha68–1 (ATP6V1A) deficiency negatively affected neuronal activity and exacerbated Aβ-mediated neuronal toxicity. These findings complement our observation that vha-10 (ATP6V1G2) deficiency causes deficits in short/intermediate-term memory function in C. elegans, and further highlights evolutionarily conserved V-ATPase function as an attractive target for LOAD therapeutic development.

CALM3 is one of the three identical isoforms of the calmodulin gene that is encoded in the human genome. A calcium-binding factor, calmodulin is ubiquitously expressed, and has central roles in a wide variety of processes critical to cellular health and function. Calmodulin function has been tied to LOAD pathobiology for some time, leading some to postulate a “calmodulin hypothesis” for AD pathogenesis⁹³, as an extension of the already-established “calcium hypothesis”⁹⁴. With respect to CALM3 in particular, it has been difficult to definitively tie alterations in CALM3 expression in the LOAD brain to genetic risk because CALM3 resides within the greater APOE locus. Due to the powerful LOAD association of APOE, along with the strong LD relationships in this locus, identification of additional signals beyond the well-studied APOE coding variants^95–97 has been challenging. We identified and prioritized CALM3 by our analyses for in vivo testing. In contrast to mammals, the C. elegans genome contains only one ortholog of calmodulin, cmd-1. We found that neuron-specific knockdown of this critical gene cmd-1 (CALM3) resulted in a significant memory deficit at 1hr post-conditioning without causing significant motor or chemotaxis defects. This interesting phenotype likely involves differential regulation of calmodulin-dependent kinase II (CaMKII) activity, given its well-known roles in learning, memory, and forgetting^98,99.

The BIN1 (Bridging Integrator 1) locus has the second strongest LOAD GWAS association behind APOE. Recent variant fine-mapping studies have indicated that transcriptional regulatory elements specific to microglia might be the mediators of LOAD genetic risk in the region, resulting in altered microglial BIN1 expression^46,47. However, different cell types in the brain express different isoforms of BIN1, and while global transcription of BIN1 is increased in the LOAD brain, the transcription of neuron- and astrocyte-specific isoforms are downregulated and are associated with tau pathology¹⁰⁰. Additionally, recent work has shown that neuron-specific conditional knockout of BIN1 results in reduced synapse density, decreased presynaptic vesicle release, and learning and memory deficits in mice⁷¹. Such studies prompted the inclusion of BIN1 as a surrogate positive control for our assay, but importantly BIN1 was not prioritized for testing according to our criteria due to its membership in module M29, which was not significantly associated with any LOAD traits, did not contain genes in the core network, and was not enriched for any cell type transcriptional signatures. Interestingly, however, we found that neuronal-specific knockdown of the C. elegans ortholog amph-1 (BIN1) resulted in a decreased ability to “forget” an associated memory, in line with its role in RAC1-mediated synaptic remodeling⁸⁴, an important component in the process of active forgetting^81,82. These results suggest complex roles for neuronal BIN1 function that may have isoform-dependent phenotypes upon perturbation.

NGEF, or ephexin-1, is a neuronal guanine nucleotide exchange factor (GEF) for GTPases such as RAC1, RHOA, and CDC42. Besides central functions in axon guidance⁸³ it also has major roles in dendritic spine morphogenesis, post-synaptic organization, and pre-synaptic vesicle release through its interactions with Eph receptors like EphA4^101,102. We found that, similar to amph-1 (BIN1), neuronal knockdown of ephx-1 (NGEF) resulted in a persistence of associative memory. In the LOAD GWAS locus that includes NGEF, INPP5D is the gene usually nominated as causal. However, a recent fine-mapping study identified neuron-specific chromatin interactions between LOAD risk variants and the NGEF promoter⁴⁶, nominating NGEF as one of the top candidate neuronal causal genes. These results indicate that there might be multiple, cell type-specific, causal genes in this locus, in contrast to the prevailing view that LOAD genetic risk is conferred by dysregulation of INPP5D primarily in microglia¹⁰³.

The CELF1/SPI1 LOAD GWAS locus (11p11.2), which we screened extensively for functional effects on memory, is a gene dense locus that did not reach GWS in the Jansen et al. 2019 GWAS/X meta-analysis⁷, but has been found as GWS in previous studies^5,6,8 and the most recent LOAD GWAS⁹. SPI1 has been found to be a likely causal gene with regard to the relevance of this locus for microglial function in AD⁶⁰, but LD relationships in this locus are complex and other lines of evidence^46,47,104 as well as the results presented here indicate that this locus harbors additional causal genes, including ACP2. ACP2 encodes lysosomal acid phosphatase 2, a phosphatase present in the lysosomal membrane which assists in the maturation of lysosomal enzymes and helps maintain the optimal pH for proper lysosomal function¹⁰⁵. In humans ACP2 is broadly expressed in all tissues, with particularly strong expression in pyramidal neurons of the cortex and cerebellar Purkinje cells¹⁰⁶. ACP2 deficiency in mice causes lysosomal storage defects, seizures, skin, cerebellum, and vertebral malformations, and ataxia^107,108. Intriguingly, a recent LOAD whole exome sequencing (WES) study identified a rare missense variant in ACP2 (D353E) to be enriched in controls compared to LOAD patients¹⁰⁹, suggesting a protective role of ACP2 in LOAD. Correspondingly, we found that neuron-specific knockdown of pho-5 (ACP2) results in extended associative memory in C. elegans, even up to 3 hours post-conditioning, an interesting result which agrees directionally with the finding from the WES analysis. While complete loss of ACP2 function results in severe neurological phenotypes^107,108, these results suggest that reduced ACP2 function could be protective with respect to LOAD-associated cognitive impairment.

In addition to endolysosomal biology, mitochondrial function was also enriched in the top LOAD-associated neuronal modules, and both have been implicated in the etiology of other neurodegenerative diseases, including Parkinson’s disease (PD)¹¹⁰. Interestingly, several PD risk genes are members of these top LOAD-associated neuronal signature-enriched modules, including GBA and PINK1 (mitochondrial function and candidate risk gene-enriched module M16), and SNCA and PRKN (top CDR-associated neuronal module M2). Additionally, gene ontology analysis of the LOAD-downregulated module M16 found a significant enrichment for genes involved in antigen presentation (Table S6), and recent studies have drawn links between mitochondrial antigen presentation and immune responses in PD¹¹¹, suggesting potentially common mechanisms of pathogenesis between the two diseases, centered on mitochondrial biology. Notably, a previous co-expression network study found two modules which were conserved between normal aging and LOAD, one representing mitochondrial processes, and the other representing synaptic function, and identified ATP6V1G2 as a top hub gene common to both LOAD and aging¹¹². Since modules and genes that we identified through our work have also been found to be relevant to the normal aging process, this suggests that perhaps LOAD genetic risk factors which affect neuronal function are the earliest contributors to disease pathophysiology, as aging is the greatest risk factor for neurodegenerative disease, including LOAD.

In summary, our integrative analysis and in vivo screening revealed genetic contributions to neuronal dysfunction in LOAD pathobiology and identified evolutionarily conserved key neuronal genes and pathways involved in this process. When combined with the growing publicly available human genomic data, simple model organism systems, such as the C. elegans behavioral paradigm used here, have great potential to advance the functional genetic understanding of the complex etiology of LOAD.

RESOURCE AVAILABILITY

Lead contact

For further information and resource/reagent requests, please direct all inquiries to the Lead Contact, Yousin Suh (ys3214@cumc.columbia.edu).

Materials availability

This study did not generate any unique reagents.

Data and code availability

• All bulk and single-nucleus RNA-seq are publicly available on Synapse (https://www.synapse.org). Accession numbers are listed in the key resources table.

• This study does not report original code. All analyses can be replicated using the standard WGCNA (https://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/) and Seurat (https://satijalab.org/seurat/index.html/) workflows and the specific parameters provided in the Method Details.

• Any additional information required to reanalyze the data reported in this study is available from the lead contact upon request.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Bacterial and virus strains
E. Coli: OP50	Caenorhabditis Genetics Center	OP50
E. Coli: HT115	Caenorhabditis Genetics Center	HT115
Biological samples
Chemicals, peptides, and recombinant proteins
2-Butanone, 99+%, extra pure	Acros Organics	Cat# 149670250
Critical commercial assays
Deposited data
Human brain bulk RNA sequencing data	(Wang et al., 2018)	Synapse: syn3159438; https://www.synapse.org/#!Synapse:syn3159438
Human brain single-nucleus RNA sequencing data	(Cain et al., 2023)	Synapse: syn16780177; https://www.synapse.org/#!Synapse:syn16780177
Experimental models: Cell lines
Experimental models: Organisms/strains
C. elegans strain LC108: vIs69 [pCFJ90(Pmyo-2::mCherry + Punc-119::sid-1)]	Caenorhabditis Genetics Center	LC108
Oligonucleotides
Recombinant DNA
Plasmid: pL4440 RNAi control	Ahringer RNAi library	N/A
Plasmid: pL4440-vha-10 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-jnk-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-aps-2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-zyg-9 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-F54A3.2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-cmd-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-apl-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-Y62E10A.2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-ephx-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-misc-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-pho-5 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-pho-14 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-mtch-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-nuo-2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-rpt-5 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-amph-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-unc-11 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-rpy-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-nmr-2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-gpb-1 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-gtl-2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-lad-2 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-T09B9.3 RNAi	Ahringer RNAi library	N/A
Plasmid: pL4440-nhr-23 RNAi	Ahringer RNAi library	N/A
Software and algorithms
WGCNA v1.70.3	(Langfelder and Horvath, 2008)	https://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/
R version 4.0.5	R Core Team	https://www.R-project.org/
Rstudio v1.4	R Studio	https://rstudio.com/products/rstudio/
Seurat v4.1.0	(Hao et al., 2021)	https://satijalab.org/seurat/index.html/
Prism 8	Graphpad Prism	https://www.graphpad.com/scientific-software/prism/
Other

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Worm cultivation

All animals were maintained at 20°C on plates made from high growth medium (HGM: 3 g/L NaCl, 20 g/L Bacto-peptone, 30 g/L Bacto-agar in distilled water, 4 mL/L cholesterol (5 mg/mL in ethanol), 1 mL/L 1M CaCl2, 1 mL/L 1M MgSO4, and 25 mL/L 1M potassium phosphate buffer (pH 6.0) added to molten agar after autoclaving (Brenner, 1974) with OP50 E.coli as the food source. For RNAi treatment, the standard HGM was supplemented with 1 mL/L 1M IPTG (isopropyl b-d-1-thiogalactopyranoside) and 1 mL/L 100 mg/mL carbenicillin, and plates were seeded with HT115 E. coli for ad libitum feeding. Worms were synchronized by collecting eggs from hermaphrodites via exposure to an alkaline-bleach solution (80 mL water, 5 mL 5N KOH, 15 mL sodium hypochlorite); collected eggs were repeatedly washed in M9 buffer (6 g/L Na2HPO4, 3 g/L KH2PO4, 5 g/L NaCl and 1 mL/L 1M MgSO4 in distilled water; Brenner, 1974).

C. elegans strains

LC108 (vIs69 [pCFJ90(Pmyo-2::mCherry + Punc-119::sid-1)])

METHOD DETAILS

Data sources

Alzheimer’s disease GWAS summary statistics from the Jansen et al. meta-analysis⁷ were retrieved from https://ctg.cncr.nl/software/summary_statistics. The MSBB LOAD RNA-seq data are available through the AD Knowledge Portal (https://adknowledgeportal.synapse.org) under the synapse ID# syn3159438. Processed single-nucleus RNA-seq data from the human frontal cortex⁴³ are available through the AD Knowledge Portal (https://adknowledgeportal.synapse.org) under the synapse ID# syn16780177. Human brain cell type-specific enhancer tracks from Nott et al.⁴⁶ are available through the UCSC Genome Brower (https://genome.ucsc.edu/s/nottalexi/glass_lab_BrainCellTypes_hg19).

LOAD patient cohorts

The Mount Sinai Brain Bank (MSBB) LOAD cohort consists of 364 postmortem control and LOAD patient brains, each accompanied by robust clinical and neuropathological phenotype metadata, with various sample subsets used for the generation of genome-, transcriptome- and proteome-scale molecular datasets, as has been described in detail previously⁴⁴. For our co-expression network analyses, we utilized bulk RNA-seq data that had been generated from the Brodmann area 36 parahippocampal gyrus region of a subset of the greater cohort (n=215). Each individual had full neuropathological assessments according to the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) protocol¹¹³, a Braak staging score for neurofibrillary neuropathology burden¹¹⁴, and a Clinical Dementia Rating (CDR) scale score¹¹⁵ based on premortem dementia and cognitive function assessment. The human frontal cortex single-nucleus RNA-seq data we utilized for characterization of the expression patterns of our post-GWAS nominated candidate risk genes and for the cell type transcriptional signature module enrichment analyses came from subjects enrolled in two cohort studies of aging and dementia, the Religious Orders Study (ROS) or the Rush Memory and Aging Project (MAP), typically referred to collectively as ROSMAP and previously described⁴³. All brains from these subjects have undergone quantitative neuropathological assessment, and the associated participants had longitudinal cognitive measures and clinical diagnoses of dementia status at death, as has been extensively previously described^116–118. For our analyses we used single nucleus RNA-seq data from 24 individuals from the ROSMAP cohort previously created and analyzed in the Cain et al. study. We demarcated these 24 individuals into 12 LOAD cases and 12 cognitively normal controls, based upon clinical diagnoses of dementia status at death, and regardless of individual neuropathological burden.

Candidate gene mapping

To map LOAD GWAS loci to genes we used the web-based tool Functional Mapping and Annotation (FUMA, v1.3.6a)⁵⁸. Using the summary statistics from the Jansen et al. meta-analysis, a sub-genome-wide significance threshold of P<1×10⁻⁵ was used to identify all independent (r²<0.1, EUR population, 1000 Genomes) loci. Within each identified locus, all SNPs that met the significance threshold were used for mapping, as well as SNPs in strong linkage disequilibrium (r²>0.6, EUR population, 1000 Genomes) with the index variant of each locus. Gene mapping was conducted using two strategies: 1) Selecting genes with significant cis-eQTL associations (FDR < 0.05) with the LOAD GWAS SNPs (i.e., expression of the gene is associated with allelic variation at the SNP). Six large-scale brain eQTL studies were utilized for this purpose – PsychENCODE³⁵, CommonMind Consortium³⁶, BRAINEAC³⁷, BrainSeq³⁸, ROSMAP³⁹, and GTEx v8 (Brain and Nerve tissue only)⁴⁰; and 2) Selecting genes by identifying significant chromatin interactions (FDR ≤ 1e-6) between gene promoter regions (250 bp up- and 500 bp downstream of the transcription start site) and the LOAD GWAS SNPs, as identified by Hi-C data. Data from three Hi-C studies of brain and neural tissue were utilized – DLPFC from PsychENCODE³⁵, Adult and Fetal cortex data from the study of Giusti-Rodríguez et al.⁴¹, and DLPFC, Hippocampus, and Neural Progenitor Cell data from the study of Schmitt et al.⁴².

Cell type expression specificity of candidate risk genes

We investigated the cell type-specific expression of our candidate risk genes by examining their expression patterns in publicly available human single-nucleus RNA-seq data from the frontal cortex⁴³. Preprocessed and annotated single-nucleus expression matrices were downloaded as described above and analyzed in R using Seurat v4¹¹⁹. All nuclei (n=162,562) were assigned the primary cell identity classifications determined by Cain et al. and specified in the corresponding metadata file retrieved from Synapse. Due to low representation in the data, nuclei annotated as “Pericytes” were removed, leaving a total of 161,399 nuclei remaining. Expression data for all protein-coding candidate risk genes were averaged across each annotated cell type using the AverageExpression function in Seurat, scaled and log-normalized and displayed as a heatmap by using the R package ‘pheatmap’.

Co-expression network analysis

The R package ‘WGCNA’⁶¹ was used to construct a co-expression network from the MSBB LOAD brain RNA-seq data⁴⁴. For the creation of the network we utilized the publicly available preprocessed expression matrix (see Data sources) which had already been normalized and adjusted for sex, race, age, RNA integrity, post-mortem interval, and batch effect. A weighted co-expression network was built using the preprocessed expression values and the blockwiseModules WGCNA function with the following parameters: soft-thresholding power = 8, TOMType = “signed”, deepSplit = 2, minimum module size of 15, merge cut height of 0.25, signed hybrid network with pamRespectsDendro = FALSE. This resulted in the identification of 32 modules of co-expressed genes, from which we calculated module eigengenes (MEs). Correlations between clinical and neuropathological traits and individual gene expression or MEs were computed as Pearson’s correlations and were corrected for multiple testing according to the FDR (Benjamini-Hochberg) method. Significance was determined using an adjusted P-value cutoff of 0.05.

Enrichment analyses

Enrichment of LOAD GWAS signal⁷ in open chromatin regions of human brain cell types was calculated according to the methodology of Wang et al.⁴⁹, using called peaks from scATAC-seq of the human brain⁴⁷. At each p-value significance cut-off, using a sliding -log(p-value) threshold from 0 to 10 in steps of 0.1, the proportion of SNPs in ATAC-seq peaks with p-values more significant that the cut-off, the foreground, was calculated against the proportion of SNPs present in the summary statistics (~13 m). Co-expression modules were tested for significant enrichment in cell type expression signatures of seven primary brain cell types (excitatory neurons, inhibitory neurons, microglia, astrocytes, oligodendrocytes, oligodendrocyte precursor cells (OPC), and endothelial cells), using the preprocessed human frontal cortex single-nucleus transcriptome data⁴³ we downloaded and annotated as described above. Using Seurat v4¹¹⁹, we followed the exact procedures specified by Cain et al. for normalization (NormalizeData), identification of variable features (FindVariableFeatures), data scaling (ScaleData), dimensionality reduction (RunPCA), and visualization (RunUMAP). Using the AddModuleScore function in Seurat we assigned module scores to all nuclei in the dataset for expression of each of the 32 modules identified in the co-expression network analysis. Briefly, this approach assesses the difference between the average expression levels of each module gene set and random control genes. Positive scoring of nuclei indicates that the module of genes is expressed in those nuclei more highly than would be expected, given the average expression of that module across the population of nuclei¹²⁰. Using a module score threshold of > 0.1, overrepresentation of expression of a co-expression module within a given cell type in the single-nucleus transcriptome data was assessed by one tailed Fisher’s exact test and corrected for multiple comparisons by the FDR (Benjamini-Hochberg) method. Significance was determined using an adjusted P-value cutoff of 0.05. Functional enrichment of biological pathways within the co-expression modules was assessed by over-representation test, using the R package ‘clusterProfiler’¹²¹, considering all genes present in the MSBB RNA-seq dataset as the set of background genes. The Gene Ontology (Biological Process) gene sets used for the enrichment analysis came from the Molecular Signatures Database (MSigDB) v7.0^122,123. Multiple testing correction was performed according to the FDR (Benjamini-Hochberg) method and significance was determined using an adjusted P-value cutoff of 0.05. Significantly enriched terms were visualized as bar plots indicating -log10FDR P-values.

Identification of core network genes

To identify genes with high trait-relevance that also reside in centrally-located positions within the co-expression network, we took the approach used by Chateigner et al.¹²⁴. By utilizing the module membership measure kME (the correlation between the expression of a gene and the module eigengene), it can be appreciated that the genes with the highest kME in a given module are also the most correlated to the traits that are most closely associated with the module eigengene. This relationship demonstrates the utility of employing kME as a centrality score when prioritizing genes with both relevance to the trait of interest and high network connectivity. Using kME to define the topological positions of all the genes in the co-expression network, the max kME was identified for every gene (i.e., the score with respect to the module to which the gene was assigned), and “core” network genes were then defined as the top 10% of genes with the highest global absolute scores.

Short/intermediate-term associative memory training

Worms were tested for short/intermediate-term memory as previously described (Kauffman et al., 2010). Briefly, synchronized day 1 adult hermaphrodites were washed from HGM plates with M9 buffer for 3 times. Then the animals were starved for 1 hr in M9 buffer. For training, worms were transferred to 10 cm NGM conditioning plates seeded with OP50 E. coli bacteria and with 18 ul 10% 2-butanone (Acros Organics) in ethanol on the lid for 1 hr. After conditioning, the trained worms were tested for chemotaxis towards 10% butanone vs. an ethanol control either immediately (0 hr) or after being transferred to 10 cm NGM plates with fresh OP50 for specified intervals before testing (30 min-2 hr). Chemotaxis indices (CI) were calculated as follow: (#wormsButanone – #wormsEthanol)/(Total #worms). Learning indices (LI) are: LI_trained=CI_trained-CI_naive.

Supplementary Material

2

Table S1. Candidate risk genes nominated by eQTL- and chromatin interaction-based gene-mapping, related to Figure 2.

3

Table S2. Chromatin interaction evidence nominating candidate risk genes, related to Figure 2.

4

Table S3. eQTL evidence nominating candidate risk genes, related to Figure 2.

5

Table S4. Co-expression module memberships, trait correlations, max kME, and module membership scores across all modules, for all genes in the BM36 RNA-seq data, related to Figure 3.

6

Table S5. Module eigengene-trait corrleations, related to Figure 3.

7

Table S6. Gene Ontology Biological Process term enrichment for all co-expression modules, related to Figure 3.

8

Table S7. Top 20 high-priority candidate risk genes, related to Figure 3.

9

Table S8. C. elegans orthologs of candidate LOAD risk genes, related to Figure 4.

10

Table S9. Amino acid similarity of screened C. elegans orthologs of candidate LOAD risk genes, related to Figure 4.

Highlights

Genetic contributions to neuronal dysfunction in AD are explored

Candidate risk genes dysregulated in AD are core members of neuronal network modules

Perturbation screening of candidate risk genes in C. elegans reveals impacts on memory

Evolutionarily conserved neuronal endosomal function genes are identified as new targets