Alzheimer’s disease (AD), the most common cause of dementia, is a progressive degenerative disorder that attacks the brain’s cells resulting in loss of memory, thinking, and language skills, and changes in behavior with extracellular β-amyloid (Aβ) aggregation and neurofibrillary tangles (NFT) formed by hyperphosphorylated tau (Knopman et al., 2021). The global number of individuals with AD is estimated to be 32 million. At the same time, prodromal and preclinical AD comprised 416 million persons worldwide, or 22% of all people aged 50 and over (Gustavsson et al., 2023). AD is a complex and multifaceted disease, and much is still unknown about its causes and progression (Herrup, 2021). As a result, there are many hypotheses about the disease, but no single answer or successful cure (Alves et al., 2021).

A single gene can create numerous proteins through alternative splicing (Gabut et al., 2011). It is an essential mechanism of gene regulation in the brain and is known to be altered in aging and neurodegeneration (Chabot and Shkreta, 2016; Deschenes and Chabot, 2017). Alternative splicing is mediated by a ribonucleoprotein complex called the spliceosome, which consists of five small nuclear ribonucleoprotein subunits and several protein cofactors (Matera and Wang, 2014). Because of the short and degenerate splicing sites in higher eukaryotes, the spliceosome usually requires RNA binding proteins (RBP) called splicing factors (SF) to identify exons accurately (Biamonti et al., 2021). Generally, seven basic types of alternative splicing events (ASEs) have been identified, including alternative 3′-splice site (A3SS), alternative 5′-splice site (A5SS), mutually exclusive exons (MXE), intron retention (RI), exon skipping (SE), alternative polyadenylation (APA), and alternative promoter (Bhadra et al., 2020). Recent transcriptomic studies have found several ASEs to be disrupted in AD, such as MBP, ABCA7, APP, CLU, PICALM, and PTK2B (Raj et al., 2018; Yang et al., 2021). Also, it is observed that the U1 snRNP deposition with NFT in postmortem brains of AD patients (Bai et al., 2013; Hales et al., 2014), MAPT transgenic mice (Vanderweyde et al., 2016; Maziuk et al., 2018), and in laboratory conditions (Bishof et al., 2018). As alternative splicing is a cell type-specific process (Zhang et al., 2016), and the balance of cell types is disturbed in AD so that there are fewer neurons and more glia (Penney et al., 2020), studying ASEs at the cellular level becomes necessary. Sharing of RNA sequencing (scRNA-seq) data of AD patients and control with single cell RNA sequencing data of Allen brain atlas of Brodmann area 22 suggested that there is cell-type differential transcript usage (DTU) pattern for APP and BIN1 (Marques-Coelho et al., 2021). ScRNA-seq data analyses a showed that Sipa1l1 transcripts were changed differently between the brains of APOE-deficient transgenic mice and control mice (Weller et al., 2022).

In this study, we analyzed transcripts variations at the cell level in the postmortem prefrontal cortex of AD patients and control patients without neuropathological or neuropsychiatric disorders (Supplementary Table S1) to identify notable ASEs in AD. To do this, we took advantage of three available transcriptomics datasets (Figure 1A). GSE125050 results from bulk RNA sequencing of sorted brain cells in four groups, including astrocytes, endothelial cells, microglia, and neurons (Srinivasan et al., 2020). GSE157827 dataset consists of single nucleus RNA sequencing of neurons, glial cells, and endothelial cells (NGE). Individuals in the NGE dataset were grouped based on the Braak stage in the control group with 0, Alzheimer’s disease with mild severity (ADM) with 3 or 4, and Alzheimer’s disease with high severity (ADH) with 5 or 6 scores (Lau et al., 2020). GSE129308 dataset includes transcriptomic data of single soma RNA sequencing which is grouped based on NFT’s presence in somas (Otero-Garcia et al., 2022). Single-cell data allow the identification of ASEs in rare cell types or subpopulations that may be overlooked in bulk analysis, thereby contributing to a more comprehensive understanding of splicing patterns (Joglekar et al., 2023). Despite the challenges posed by the short reads and 3′ bias in the 10x Genomics’ scRNA-seq data (Westoby et al., 2020), a good understanding of splicing changes can still be achieved due to the presence of a large number of junctional reads in the datasets and the application of a robust statistical approach (Figure 1B). Furthermore, due to the strong correlation between splicing changes and Braak NFT stages in AD (Arizaca Maquera et al., 2023), the utilization of NFT and NGE single-cell data can provide a broader view of ASEs in PFC neuronal subtypes based on NFT pathogenesis and ASEs in different brain cell types based on Braak stages, respectively. A broad identification of ASEs can be gained by analyzing splicing alterations in various cell populations at different stages of AD progression. We also analyzed differential gene expression (DGE) and revealed ontology cell type-specific.

In recent years, advancements in molecular techniques such as GWAS and RNA-seq have facilitated the identification of potent molecules associated with the pathology of AD, which were previously overlooked (Penney et al., 2020). However, these studies were frequently confined to analyzing tissues rather than individual cells and prioritized gene expression over transcript expression, representing a more functional level than gene expression. This study addresses these limitations by examining cell-type-specific transcripts alterations in AD, achieving a more comprehensive analysis of AD transcriptomics data than previous studies. To the best of our knowledge, only two studies have attempted to identify DTUs in Alzheimer’s disease AD brain and AD model based on cell types. Marques-Coelho et al. indirectly assigned genes to unique cell types in the medial temporal gyrus region of AD brains and compared DEGs/DTUs from bulk RNAseq to unique cell populations from previously published scRNAseq data (Marques-Coelho et al., 2021). In contrast, our study directly analyzed scRNAseq reads in the PFC of the AD brain. In a separate investigation, Weller et al. identified DTUs in scRNAseq data from the hippocampus of APOE null mutant mice using Sierra (Weller et al., 2022), which uses pseudobulk analyses to detect DTUs and peak calling to detect APA (Patrick et al., 2020). Our method employed SpliZ scalar quantification to measure true exon junctions between cell types and detected APA using at least two peaks in human postmortem data. The present study highlights the cell type and subtype-specific alterations in splicing patterns of certain genes in AD, even in the absence of expression changes. Notably, an increase in ASEs is observed with disease progression. We used an annotation-free approach to scRNAseq data analysis to identify novel splicing ASEs and disruptions. Collectively, our findings offer a detailed account of cell type-specific modifications in gene expression in the AD brain and imply that alternative splicing may play a significant role in the pathological progression of the disease.

In order to cover the shortcomings of each technique in this study, the data of three different RNA sequencing approaches, including bulk RNA-seq, single nuclei RNA-seq, and single soma RNA-seq, were used to identify ACEs in AD. While bulk sequencing provides full-length coverage of transcripts in contrast to 10x single-cell sequencing, it is biased in favor of more populated subtypes. snRNA-seq is a widely used method to determine brain cell type complexity and is used to construct a comprehensive human brain cell atlas, although it relies on the nuclear mRNA pool (Kim et al., 2023; Pavan et al., 2023). While bulk RNA sequencing data shows a strong correlation between nuclei and whole-cell samples in differential expression analysis, at the single-cell or single-nucleus levels, cell-to-cell or nucleus-to-nucleus correlations decrease and replicate variations become larger than the bulk samples (Kim et al., 2023). Recent studies report that the single-nucleus RNA sequencing technique is biased towards genes with longer sequence lengths and roughly >10 exons, whereas the single-cell RNA sequencing technique captures shorter genes more efficiently (Gupta et al., 2022; Pavan et al., 2023). Moreover, Characterization of the nuclear and cytosolic transcriptomes in human brain tissue reveals that transcripts encoding nuclear-encoded mitochondrial proteins are significantly enriched in the cytosol, while lncRNAs significantly enriched in the nucleus (Zaghlool et al., 2021). Therefore, these studies suggest that examining single-cell sequencing data alongside bulk sequencing data is better to gain a more detailed insight. Our analysis showed that most of the known junctions were identified in common across all three studies, bulk, NGE, and NFT (Figure 10A). There were 51,196 intersected ASEs between NGE and NFT, most including unannotated junctions. After we applied a threshold of 0.05 on the p.adj values, the number of genes with significant splicing changes in NGE and NFT was less than the bulk data (Figure 10B). There were more intersected genes with significant splicing alteration between NFT and NGE rather than bulk with NFT or NGE. The most intersected differential annotated junctions were observed between NFT and NGE neurons (Figure 10C). In every study, most differential annotated junctions were observed in neurons rather than the other cell types. However, more significant ASEs were observed in NFT and NGE neurons data than in the bulk neurons data. This highlights the advantage of single-cell data, which ignores the splicing variation between different cell subtypes by comparing diagnoses in a given subtype. The intronic/exonic ratio, junctional reads proportion of the total reads, and novel/known junction ratio was higher in NFT than NGE and higher in NGE than bulk data (Figures 10D–F).

This paper reports several cases of SE closely related to AD pathology, while the effects of others remain unclear. One such case is SLC27A1, expressed explicitly in astrocytic clusters four and six and involved in transporting fatty acids from the BBB to the brain (Ochiai et al., 2017). SLC27A1 plays a vital role in brain health by absorbing endogenous neuroprotective factors such as docosahexaenoic acid and biotin from the blood. However, an in vitro study shows that Aβ inhibits the uptake of docosahexaenoic acid by SLC27A1 from the environment (Ochiai et al., 2019). Moreover, the absorption of substances by SLC27A1 is related to the presence of insulin in the environment (Ochiai et al., 2017). Our findings suggest an SE in the coding region of the cytoplasmic regulatory subunit of AMP-binding protein, which may be related to the insulin signaling pathway. Another case is DROSHA, an RNase enzyme that plays a crucial role in the processing of microRNAs. Our results indicate that transcripts of DROSHA are significantly SE in exons two, six, and seven in microglia, resulting in a reduction in the length of the 5′ UTR region in AD and an increase in the presence of exon seven in the structure of DROSHA transcripts in AD microglia. Recent studies have reported different types of DROSHA transcripts that show subcellular localization differences, although they do not exhibit apparent functional differences in the processing of microRNAs. Interestingly, this process appears to be cell type-specific, and the presence of exon seven in DROSHA mRNA is essential for its cytoplasmic localization (Dai et al., 2016; Link et al., 2016). In addition, new evidence suggests that increased Aβ in postmortem brains and rat brains reduces the presence of DROSHA in the nucleus and increases its presence in the cytoplasm. This phenomenon is associated with the phosphorylation of DROSHA by p38 MAPK in neurons, and the DROSHA phosphorylation site is located near the junction of exon eight to seven (Xu et al., 2021). Another example is PLEKHA5 transcripts, which show SE that includes an additional exon between exon nine and ten. This SE increases the mRNA structure in EN-L3-5 with NFT, and this region encodes the Pleckstrin homology domain, which regulates plasma membrane and trans-Golgi membrane traffic activities (Sluysmans et al., 2021). Finally, we observed that RI is another type of ASE, significantly increased in astrocytes. For instance, NPAS2, which is involved in the circadian rhythm, has a splicing disorder in its transcripts that may be related to sleep disorder in AD (Sharma et al., 2021).

The findings pertaining to astrocytes indicate that alternate splicing in this cell type is closely associated with Alzheimer’s disease (AD). Moreover, certain sub-types exhibit distinct patterns of alternative splicing and expression linked to specific AD pathologies. Notably, astrocytes exhibit significantly impaired splicing of CLU transcripts, a gene highly expressed in astrocytes. CLU is an extracellular chaperone that interacts and binds to Aβ, reducing its aggregation and promoting its clearance, suggesting a potential neuroprotective role. However, some studies have suggested that CLU may contribute to the spread of Aβ in the brain and serve as a key mediator in Aβ-induced neurotoxicity. This interaction may depend on the distribution ratio of CLU to Aβ (Foster et al., 2019). Most of the astrocyte-related CLU splicing disruptions observed in this study are related to the alpha domain. Some of the abnormal junctions miss the glycosylation and phosphorylation residues of CLU, which disruption of these structures can lead to a decrease in the polarity of the protein and a suitable substrate for the deposition of insoluble compounds such as Aβ and NFT or Its secretion is prevented. The CLU gene is known as the third risk factor for late-onset AD. Studies of allele-specific quantitative loci indicate that CLU splicing changes are related to rs9331896 and rs7982 polymorphisms in the intronic region of the CLU gene (Raj et al., 2018; He et al., 2022). Another crucial gene in astrocytes is APOE, the first risk factor for late-onset AD, which undergoes APA and reduced expression in astrocytes. In humans, there are three common types, ApoE3, ApoE2 and ApoE4, which differ by only one amino acid at position 112 or 158. APOE plays a vital role in Aβ clearance, which ApoE2 has the strongest effect on Aβ clearance, followed by ApoE3 and least of all ApoE4 (Huynh et al., 2017). Studies of allele-specific quantitative expression loci reveal an association between APOE ASE and rs429358 polymorphism, coding ApoE4, in the coding region of the 3′ end of the APOE gene (He et al., 2022). Although the PCR data related to this polymorphism was available for the samples used in this study, the correlation between the allele and the reduction of the length UTR3’ and the N-terminal coding of APOE was not statistically significant, possibly due to the smaller sample size compared to the previous study. A network of genes involved in metal homeostasis, oxidation, and metabolism, such as FTL, MT1E, MT1G, and MT3, is evident in astrocyte clusters related to the nutritional support of neurons and synapses. Disruption in the transcripts related to these genes primarily leads to autophagy disorder, one of the pathological features of AD (Koh and Lee, 2020). Furthermore, a significant decrease in the expression of DDX5, an RNA helicase that regulates spliceosome structure and prevents the formation of unwanted RNA loops, was observed in astrocytes. This decrease correlated with an increase in the dispersion of scZ scores of CLU transcripts. Additionally, studies suggest that DDX5 increases the access of U1snRNP to the 5′ binding site of MAPT exon ten by structurally changing the stem-loop structure, which prevents the formation of incomplete tau proteins (Kar et al., 2011). Additionally, studies suggest that DDX5 increases the access of U1snRNP to the 5′ binding site of MAPT exon ten by structurally changing the stem-loop structure, preventing incomplete tau proteins (Li et al., 2021). PTGDS is a highly abundant protein found in cerebrospinal fluid that has been suggested to act as an endogenous chaperone for Aβ. The expression level of PTGDS is related to dementia, with its expression being regulated both directly and indirectly by estradiol (Unno et al., 2015). It has been reported that plasma levels of PTGDS in AD patients are associated with increased levels of inflammatory cytokines and reactive oxygen species (Dharshini et al., 2019). Our study revealed that the A5SS event of PTGDS transcripts endothelial and astrocyte cells. UniProt data suggests this event occurs within the coding region of turn between beta sheets with disulfide bonds near the enzyme’s active site, potentially impacting the enzyme’s regulatory activities. A recent NMR study has shown that PTGDS can inhibit primary and secondary nucleation of Aβ40 by interacting with both monomers and the surface of fibrils, reducing the final fibril content (Kannaian et al., 2019). The proposed binding region of PTGDS with Aβ is adjacent to the turn affected by the A5SS event in PTGDS.

Our analysis revealed that genes contributed to RNA splicing and the steroid hormone signaling pathways in microglia undergoing DTU and DGE. Studies suggest that the loss of ovarian hormones during menopause may increase the risk of AD in women, as androgens play a crucial role in various brain functions, such as neurotransmission, neurodevelopment, survival, protection against oxidative stress, reduction of Aβ peptide levels, and reduced tau hyperphosphorylation (Breijyeh and Karaman, 2020). Furthermore, our findings showed a correlation between the splicing changes of FTL transcripts and DDX5 expression. In microglia, FTL has increased expression and alternative splicing of exon 3, despite only one FTL transcript in the reference genome and the skipping of exon three observed in ADH being considered abnormal. Exon three encodes FTL alpha helices, and mutations in this region have been linked to Parkinsonian symptoms, ataxia, and mild non-progressive cognitive impairment (Maciel et al., 2005). Our analysis also revealed several alternative APAs in microglia. Moreover, we observed expression changes in factors involved in shortening the poly A tail of mRNA molecules synthesized in the nucleus of microglia. While APOE splicing changes were also observed in microglia, we did not statistically confirm its relationship with rs429358. HSP90AA1, an intracellular chaperone that corrects misfolded proteins and prevents their disruptive aggregation, underwent 3′ end elongation in ADH. Studies suggest that this difference in elongation may be due to alternative promoter recognition by different transcription factors in combination with RNA polymerase rather than APA. An increase in this transcript’s 3′ end length is associated with factors such as viral infection, inflammation, cell death, and increased glucose concentration(Zuehlke et al., 2015; Bohush et al., 2019). Increased expression of HSP90AA1 was also observed in microglia and EN in other studies (Bohush et al., 2019).

In contrast to the bulk sequencing data, which showed limited cases of ASEs and DTU and lacked any DGE in neuronal types, single-cell sequencing data revealed multiple cases of expression and transcriptome changes. This may be due to the wide transcriptome difference between EN and IN or may be due to an error in the neuron collection protocol. The events identified in neurons using single-cell sequencing data appear to be repeated in the bulk sequencing data but are not statistically significant. Increasing the number of donors for neuron lineages may improve this issue. Nonetheless, important genes such as TSPAN14 are observed in the DTU of the bulk data. Studies on quantitative allele-specific expression loci methylation support the hypothesis that increased brain expression of TSPAN14 is linked to an increased risk of AD (Bellenguez et al., 2022). CLU is an example of widespread dysregulation of ENs transcripts, and this disruption seems to enhance with disease severity. This irregularity increased correlated with the decrease of CELF2 and DDX5 expression. CELF2, also called ETR3, has been identified as a novel risk factor associated with AD, particularly in individuals carrying the high-risk APOE ε4 allele, and it has also shown significant association with AD independent of APOE ε4, indicating its potential as a valuable therapeutic target for addressing underlying genetic causes of the disease (Tran, 2023). Prior research has demonstrated that the alternative splicing of exon three in TREM2, a genetic risk factor for AD, is regulated by two paralogous RNA-binding proteins, CELF1 and CELF2, with CELF2 being implicated in the reduction of full-length TREM2 protein expression through exon three skipping and nonsense-mediated mRNA decay, which effects on microglial responses to the Aβ aggregation (Yanaizu et al., 2020). Also, CELF2 is involved in influencing various transcripts, such as different exons of MAPT and NMDAR1 exon five, in the mis-splicing event in myotonic dystrophy type 1, particularly in reducing the inclusion of MAPT exon 10, indicating its role in alternative splicing regulation and potential nuclear and cytoplasmic functions in the brain (Liu et al., 2021). To understand the specific relationship between CELF2 and tau splicing in AD, further studies are needed to investigate the interactions between CELF2 and the tau gene in AD-affected brain tissues.

In addition to protein-coding genes, our study identified changes in transcripts of lncRNAs such as MEG3 and MALAT1 in neurons. MEG3 displayed one of the most distinct cell type-dependent splicing patterns among ENs, INs, and OPCs, as well as among different AD pathological conditions. Previous studies have reported the potential therapeutic effects of MEG3 in AD, as overexpression of MEG3 was shown to improve cognitive disorders, reduce nerve damage, and inhibit astrocyte activation in the hippocampal tissues by inhibiting the PI3K/Akt signaling pathway in rats (Yi et al., 2019). However, another study using human neurons transplanted into mouse brains exposed to Aβ found specific NFT pathology and cell death in human neurons, while mouse neurons showed only mild pathology (Balusu et al., 2022). That study further demonstrated that the induction of MEG3 in ex vivo conditions led to cell necroptosis, unlike in mice, which showed no effect (Balusu et al., 2022). Also, unannotated events of SYT1 in the 5′ region are actually part of the SYT1 alternative promoter and are not splicing disruption. This is due to some of the SYT1 variants had not been annotated in the version of the reference genome that we used, GENCODE V39 (Ensemble 105), while they were annotated in the recent version of the reference genome, GENCODE V42 (Ensemble 108). Our study identified an undocumented splicing event, including an unusual junction in position chr17:29,515,203-29,546,866 of the TAOK1 gene, which is present in all individuals in control, ADM, and ADH groups. Interestingly, the expression of this unknown transcript is increased in ADH ENs compared to the control and ADM ENs. The region lacked pseudogenes, nonsense RNA, and DNase signals, and the deleted region corresponded to the regulatory region of TAOK1 through ATM-mediated phosphorylation (Raman et al., 2007). This unknown transcript was increased in neurons with NFT in all cortical layers, but the increase was not statistically significant. TAOK1 plays a crucial role in regulating microtubule dynamics and organization, and it can induce apoptotic changes through the activation of JNK, MAPK, and caspases. Furthermore, TAOK1 regulates MAPKs and stimulates the JNK and p38 MAPK signaling pathways. Our findings also show that TAOK1 expression is increased in neurons with NFT, consistent with its role in regulating tau phosphorylation. Two phosphorylated tau residues (T123 and T427) are identified in the AD brain that appears to be specifically targeted by TAOK1. Notably, a TAOK1 inhibitor reduced tau phosphorylation in cortical neurons without affecting synapse markers and neuronal health (Giacomini et al., 2018). These results suggest that TAOK1 may be a potential therapeutic target for tauopathy in AD.

On the other hand, it is important to note that not all ASEs of certain genes can be detected in single-cell data due to poly A tail capture. Although almost all transcripts significantly altered in AD have a stop codon in the ORF or 3’ UTR, except for some CLU and TAOK1 transcripts, which are likely to be eliminated by non-stop decay mechanisms, it is still unclear whether these transcripts decayed by other mRNA surveillance mechanisms such as peptide dependent translation arrest, 18S nonfunctional rRNA decay, and mRNAs containing a strong secondary structure within the ORF (Wolin and Maquat, 2019). it is not completely clear which sequence length increase or decrease can affect RNA turnover (Andrzejewska et al., 2020). Furthermore, the question remains whether these ASEs contribute to the development of AD pathology or are a result of it or whether a positive feedback loop is involved. Lastly, it is unclear whether these altered transcripts are translated into functional proteins and, if so, what impact these changes have on protein function in some ASEs. Further investigation is required to address these questions.

In the future, the VASA-Seq technology could be a valuable tool for identifying all relevant ASEs in AD postmortem brain tissues due to its ability to provide full-length reads in over 10,000 cells (Salmen et al., 2022). However, to validate and further investigate the spread of these ASEs as the first steps, more precise methods such as in situ hybridization and immunohistochemistry should be used. There is also currently less data available in the databases for CELF2 and DDX5 than for other RBPs. Crosslinking and immunoprecipitation methods could also be employed to explore the interaction between splicing factors and ASEs (Sternburg and Karginov, 2020). The recently developed single-cell long-read method can provide more details of the final RNA sequence and may predict more translational regulation by the asset of the machine learning approach than the current studies (Hardwick et al., 2022; Joglekar et al., 2023). However, single-cell long read transcriptomics has been limited in capturing a wide range of isoform diversity due to the sequencing depth constraints inherent in its protocols. As a result, datasets typically exhibit low redundancy levels between cells of the same cell type (Arzalluz-Luque et al., 2022). Also, polysome profiling provides information about the translational control consequences of ASEs (Reixachs-Solé and Eyras, 2022). Also, after researching these cases, it is possible to integrate RNA sequencing data with other omics data, such as proteomics and epigenomics, and identify potential therapeutic targets. These methods may provide a more complete understanding of the relationship between ASEs and AD pathology, which could lead to the development of novel therapeutic strategies. In summary, our study sheds light on the importance of splicing in the AD pathology that might be considered AD as a spliceopathy during the disease progression. By applying a cell-level approach, we have identified several novel ASEs in the PFC cells of AD. Further research on splicing in AD may lead to the development of novel diagnostic and therapeutic strategies for neurodegenerative diseases.

Publicly available datasets, including GSE125050, GSE157827, and GSE129308, available at https://www.ncbi.nlm.nih.gov/, were used in this study. Processed post SICILIAN data in this study, including all true junctions used for the SpliZ pipeline, are openly available in the compressed tsv file for NGE at https://figshare.com/articles/dataset/NGE_prespliz/23522391 and in the compressed tsv file for NFT at https://figshare.com/articles/dataset/NFT_prespilz/23530641. The code used in this study is available through the following repository: https://github.com/MEFarhadieh/SCADSplice.

The Brain and Body Donation Program has been supported by the National Institute of Neurological Disorders and Stroke (U24 NS072026, National Brain and Tissue Resource for Parkinson’s Disease and Related Disorders), the National Institute on Aging (P30 AG19610, Arizona Alzheimer’s Disease Core Center), the Arizona Department of Health Services (contract 211002, Arizona Alzheimer’s Research Center), the Arizona Biomedical Research Commission (contracts 4001, 0011, 05-901, and 1001 to the Arizona Parkinson’s Disease Consortium), and the Michael J. Fox Foundation for Parkinson’s Research for GSE125050 dataset. The SWDBB for providing brain tissues for this study. The SWDBB is part of the Brains for Dementia Research program, which is jointly funded by Alzheimer’s Research United Kingdom and the Alzheimer’s Society and is supported by Bristol Research into Alzheimer’s and Care of the Elderly and the Medical Research Council for GSE157827 dataset. Human tissue was obtained from UCLA-Easton Center, NIH Neurobiobank (Sepulveda repository, Los Angeles, CA, and Mount Sinai, New York, NY), and the Stanford Alzheimer Disease Research Center (NIH/NIA P30 AG066515) for GSE129308 dataset. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

M-EF and KG conceptualized and designed study. M-EF collected and analyzed data and wrote manuscript. KG supervised study and reviewed manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.