Single-cell transcriptomic analysis elucidates APOE genotype specific changes across cell types in two brain regions in Alzheimer’s disease.

Alzheimer’s Disease (AD) is a complex neurodegenerative disease that gravely affects patients and imposes an immense burden on caregivers. Apolipoprotein E4 (APOE4) has been identified as the most common genetic risk factor for AD, yet the molecular mechanisms connecting APOE4 to AD are not well understood. Past transcriptomic analyses in AD have revealed APOE genotype-specific transcriptomic differences; however, these differences have not been explored at a single-cell level. Here, we leverage the first two single-nucleus RNA sequencing AD datasets from human brain samples, including nearly 55,000 cells from the prefrontal and entorhinal cortices. We observed more global transcriptomic changes in APOE4 positive AD cells and identified differences across APOE genotypes primarily in glial cell types. Our findings highlight the differential transcriptomic perturbations of APOE isoforms at a single-cell level in AD pathogenesis and have implications for precision medicine development in the diagnosis and treatment of AD. supports myelination, and maintains blood brain barrier (BBB) 4,6,9 . In regard to APOE isoforms, APOE4 has been linked to promoting A β retention by blocking its LRP1-mediated clearance 9,10 , insulin resistance through impaired insulin signaling 11 , BBB dysfunction and increased permeability 8,12 , and regulating glycogen synthase kinase 3 (GSK3), a kinase highly involved in phosphorylation of tau 6,13 . Our study aims to identify transcriptomic differences associated with APOE isoforms at a single-cell level to better understand the underlying mechanisms contributing to AD pathophysiology and their specificity to each isoform. Transcriptomics represent a valuable means of understanding molecular underpinnings in disease conditions 10,14–18 ; however, to our knowledge, in AD, APOE isoforms are yet to be investigated at a single-cell level, which can depict molecular profiles that would be otherwise masked in a bulk analysis.

Alzheimer's disease (AD) is a heterogeneous neurodegenerative disorder, which accounts for at least 60% of dementia cases 1 . Further underscoring the importance of AD research, cases of AD are projected to increase by more than 3-fold by 2050, yet there currently are no disease altering treatments 2,3 . AD is defined by pathological hallmarks of aggregated extracellular amyloid-β (Aβ) plaques, and intracellular tau neurofibrillary tangles 1,4 . As a complex disease, AD has a number of environmental risk factors. Demographic risk factors include advanced age, low education level, and female sex. AD genetic risk factors such as Aβ precursor protein (APP), presenilin 1 (PSEN1), and presenilin 2 (PSEN2) lead to dominantly inherited early-onset AD and account for < 1% of AD cases 1,4,5 .
The strongest genetic risk factor for late-onset or sporadic AD is the ε4 allele of the apolipoprotein E (APOE) gene. In humans, there are three common APOE allelic variants (ε2, ε3, and ε4), which differ based on single substitutions at amino acid residues 112 and 158, with the most common ε3 allele generally considered as a neutral form 4,6,7 . While the ε2 allele is considered protective, the ε4 allele is associated with increasing the risk of developing AD in a gene dose dependent manner 4,8 . Specifically, one copy of the ε4 allele of APOE increases the risk of developing AD by 3-to 4-fold, and two copies increases this risk by 12-to 15-fold 4,5 .
APOE is a lipid-binding protein which serves a central role in regulating lipid transport and metabolism. It is highly expressed in the liver and brain, where it is primarily expressed in astrocytes 4,6 . APOE's functionality in the central nervous system has implications for AD in both Aβ-dependent and Aβ-independent pathways. For instance, in addition to regulating Aβ clearance, APOE regulates lipoprotein metabolism, supports cell proliferation, repairs membranes, supports myelination, and maintains blood brain barrier (BBB) integrity 4,6,9 . In regard to APOE isoforms, APOE4 has been linked to promoting Aβ retention by blocking its LRP1-mediated clearance 9,10 , insulin resistance through impaired insulin signaling 11 , BBB dysfunction and increased permeability 8,12 , and regulating glycogen synthase kinase 3 (GSK3), a kinase highly involved in phosphorylation of tau 6,13 . Our study aims to identify transcriptomic differences associated with APOE isoforms at a single-cell level to better understand the underlying mechanisms contributing to AD pathophysiology and their specificity to each isoform. Transcriptomics represent a valuable means of understanding molecular underpinnings in disease conditions 10,[14][15][16][17][18] ; however, to our knowledge, in AD, APOE isoforms are yet to be investigated at a single-cell level, which can depict molecular profiles that would be otherwise masked in a bulk analysis.
In recent years, single-cell transcriptomic datasets were generated from the prefrontal 19 and entorhinal 20 cortices of human AD patients. With single-nucleus transcriptomes from the prefrontal cortex, Mathys and colleagues performed cell type specific differential expression analyses using 80,660 droplet-based nuclei across 24 individuals with varying degrees of AD pathology and 24 non-AD individuals, across six major cell types: excitatory neurons, inhibitory neurons, astrocytes, oligodendrocytes, oligodendrocyte precursor cells (OPCs), and microglia 19 .
On examining APOE expression, they report that APOE was strongly upregulated in AD-specific microglia, but downregulated in astrocytes in AD; however, the authors did not look at the effects of APOE genotype. With single-nucleus transcriptomes from the entorhinal cortex, Grubman and colleagues analyzed 13,214 droplet-based nuclei from 6 men and women with AD and 6 non-AD sex and age matched controls 20 . Similar to Mathys et al., neurons, astrocytes, oligodendrocytes, OPCs, and microglia were surveyed. APOE expression was upregulated in an AD-specific microglial subpopulation and downregulated in AD-specific OPC and astrocyte subpopulations, which is consistent with findings from Mathys et al. Again, similar to Mathys et al., the relationship between APOE genotype and gene expression was not examined.
To expand on these single-cell analyses, we explore the cell type specific transcriptomic effects of APOE genotype across two brain regions: the prefrontal and entorhinal cortices, using these two publicly available datasets. In this study, we aim to answer the following questions: 1) Which cell types are most affected at the transcriptomic level by APOE genotype in the context of AD? 2) What are the global and cell type specific transcriptomic changes with respect to APOE genotype in the context of AD? and 3) Are there any transcriptomic changes that are specific to APOE4 that better explain AD pathophysiology?

Sample classification and analytic workflow
We classified samples based on tau tangle and Aβ plaque burdens, using Braak clinical staging and Consortium to Establish a Registry for Alzheimer's Disease (CERAD) scores, respectively (AD: Braak stage ≥ IV, CERAD score ≤ 2; Control: Braak stage I-III, CERAD score ≥ 3) (Fig.   1). Following, from the prefrontal cortex cohort (Table 1), we analyzed single nucleus RNA-Seq (snRNA-seq) data containing 43,831 cells (Supplementary Table 1) and 17,593 genes, and from the entorhinal cortex cohort (Table 2), we analyzed snRNA-seq data containing 9,587 cells (Supplementary Table 2) and 10,850 genes. Both datasets were acquired from different sets of individuals. To examine cell type specific differences in gene expression in APOE3/3 (homozygous for allele ε3) and APOE3/4 (heterozygous ε3/ε4) cells, we performed an APOE genotype stratified differential gene expression (DGE) analysis comparing age-matched AD cases to controls, with sex as a covariate, in excitatory (Ex), and inhibitory (In) neurons for the prefrontal cortex specifically, undistinguished neurons (Neu) for the entorhinal cortex, and astrocytes (Ast), microglia (Mic), oligodendrocytes (Oli), and oligodendrocyte progenitor cells (OPCs) for both cohorts ( Supplementary Fig. 1). Differentially expressed genes (DEGs) were selected using cutoffs of a Benjamini-Hochberg (BH) adjusted p-value < 0.05 and log2 fold change > 0.25. DEGs were further passed as inputs to identify pathways for subsequent network analysis. We examined gene expression and network changes in AD compared to non-AD samples to identify cell type specific and universal changes based on APOE genotype (Fig. 1).
Due to sample limitations for relatively rare APOE genotypes, we focused our analysis on comparisons between AD and non-AD groups with an APOE3/3 or APOE3/4 genotype in this study. Interestingly, DEGs were primarily downregulated in APOE3/4 astrocytes, oligodendrocytes and OPCs, while they were primarily upregulated in both APOE3/3 and APOE3/4 neurons (Fig. 2a).
Altogether, across all cell types we identified 278 unique DEGs (Supplementary Table 3). Of the 278 DEGs, 8 were specific to APOE3/3 and 135 were specific to APOE3/4. We observed DEGs previously linked to AD (CLU 5,21 , CCK [22][23][24] , NRGN 25,26 , DHFR 27,28 , ERBB4 29,30 , NRXN1 31 ), which were shared by APOE3/3 and APOE3/4 cells. In most cases, expression differences in these genes were in the same direction across genotypes, but with greater fold changes in APOE3/4 as compared to APOE3/3 cells (Fig. 2b). Across cell types, while the majority of DEGs were shared and in consistent direction across APOE3/3 and APOE3/4 cells (Fig. 2c, yellow color; Supplementary Fig. 2), there were a few shared DEGs with opposite directionality of expression changes, such as DOCK4 in microglia, SPARCL1 in neurons, and FRYL in oligodendrocytes Notably, some DEGs in AD patients relative to controls were shared across multiple cell types ( Fig. 3a). For example, APP binding family B member 1 interacting protein (APBB1IP), and DOCK8, a protein highly involved in brain development and immune response 32 , were differentially expressed in most APOE3/4 cell types and in APOE3/3 neurons. Interestingly, for both APBB1IP and DOCK8, we observed cell-type specific effects. Both genes were downregulated in astrocytes, oligodendrocytes and OPCs and upregulated in microglia and neurons from APOE3/4 AD patients versus APOE3/4 controls. In APOE3/3 individuals, both genes were only significantly upregulated in neurons in AD patients versus controls. APOE itself was also differentially expressed in AD patients versus controls, with an increase in both APOE3/4 and APOE3/3 neurons and in APOE3/3 microglia as well as a decrease in APOE3/4 astrocytes, oligodendrocytes, and OPCs. MTRNR2L12 expression, which encodes a humanin isoform necessary for neuroprotection and anti-apoptotic function suggested to have utility as a blood marker for cognitive disability and early dementia for adults with Down Syndrome 33,34 , was very similar to APOE expression. Hierarchical clustering of samples using AD compared to control pseudobulk cell type gene expression (Fig. 2d) showed clustering of samples dominated by APOE genotype before cell type identity for all cell types except neurons, which instead cluster by cell type. Generally, we observed more similarities in DEGs of AD versus control across APOE genotypes in neuronal populations (both excitatory and inhibitory neurons), and differences primarily in non-neuronal cells (astrocytes, oligodendrocytes, and OPCs). In addition to identifying shared DEGs across cell types and APOE genotypes, we also observed a larger  Table 4). Of these DEGs, 18 were specific to APOE3/3 cells and 128 to APOE3/4 cells. Similar to the previous analysis, we observed more differences in perturbed gene profiles across APOE genotypes in astrocytes, oligodendrocytes, and OPCs, where DEGs were primarily downregulated in APOE3/4 cells.
Additionally, clustering of DEGs by log2 fold change also showed a stronger clustering by APOE genotype than cell type identity ( Supplementary Fig. 3).

DGE analysis in the entorhinal cortex identifies distinct AD-related changes in microglia and oligodendrocytes in APOE3/4 versus APOE3/3 cell type specific disease signatures
Leveraging data from Grubman et al., we identified DEGs in all cell type and APOE genotype pairings when comparing AD to control tissue from 9,587 cells and 10,850 genes, where DEGs were primarily downregulated in APOE3/3 AD versus control and upregulated in APOE3/4 AD versus control (Fig. 4a). Altogether, across all cell types we identified 232 unique DEGs (Supplementary Table 5). Of the DEGs, 29 were specific to the APOE3/4 AD, and none were specific to the APOE3/3 AD. In each cell type, we observed more DEGs in the APOE3/4 comparison, some of which were shared with APOE3/3 analysis, though often with consistent opposite directionality ( Overall, in the APOE3/4 case control comparisons, 87 DEGs were shared in all cell types, with 64 consistently upregulated in AD tissue and 23 with mixed directionality across cell types when comparing AD to control tissue ( Supplementary Fig. 4). Of these shared DEGs, a few with higher absolute log2 fold changes between AD and controls include MBP, a gene important for myelination 40,41 that was upregulated in all APOE3/4 cell types in AD except oligodendrocytes, and LINGO1, which was upregulated in all APOE3/4 cell types as well as APOE3/3 astrocytes and OPCs in AD. Interestingly the average log2 fold change for LINGO1 in APOE3/4 AD samples (3.52) was much higher than that of the APOE3/3 AD samples (0.451). Additionally, protein folding HSPA1A, the neuroprotective chaperone and apoptosis regulator CRYAB 42 , and quinoid dihydropteridine reductase (QDPR) were upregulated in all APOE3/4 cell types in AD.
However, HSPA1A was also downregulated in APOE3/3 microglia, oligodendrocytes, and OPCs, CRYAB was downregulated in APOE3/3 oligodendrocytes, and QDPR was downregulated in APOE3/3 microglia in AD. The latter two genes have previously been observed to be upregulated in oligodendrocytes and OPCs of pathologically confirmed AD individuals 19 , most of them are usually APOE4 carriers. We also observed a larger range of log2 fold change in APOE3/4 cells (-2.918, 3.839; median= 0.688) in AD compared to APOE3/3 cells (-2.385, 2.227; median= -0.436) in AD, which we visualized in a number of shared DEGs such as LINGO1, NRXN1, FTL, and ADGRL3 ( Supplementary Fig. 4). Largely, when comparing AD to non-AD cells in the entorhinal cortex, while we observed changes relevant to AD pathophysiology across APOE3/3 and APOE3/4 genotypes, we observed a number of flipped expression profiles across both APOE genotypes primarily in non-neuronal cells, and more universal transcriptional changes and changes of higher amplitude in APOE3/4 AD versus control comparison as compared to APOE3/3 AD versus control comparison.

Comparative analysis across brain regions shows more AD-related transcriptomic changes in the entorhinal cortex compared to the prefrontal cortex, with consistent APOE genotype specific disease signatures
We observed a higher number of DEGs and larger log2 fold change magnitudes across cell types in the entorhinal cortex than in the prefrontal cortex in AD. The number of shared DEGs within cell types across APOE genotype groups was highest in the entorhinal cortex in AD, while the number of shared DEGs within cell types across brain regions was highest in APOE3/4 cells in AD (Fig. 5a). With hierarchical clustering of per-cell and genotype group pseudobulk expression, we observed strongest clustering by brain region, followed by APOE genotype (Fig.   5b).

Pathway and network analysis reveal APOE genotype specific perturbed biological processes primarily in glial cells across brain regions
Pathway enrichment was performed using gprofiler 43  In astrocytes from the prefrontal cortex, we identified six enriched functional modules in both APOE3/3 and APOE3/4 AD relative to controls (Fig. 6a). We found five of these common modules (glutamatergic NMDA receptor transmission of glutamate; NLGN1 and NRXN1 clustering of NMDA receptor, periphery intrinsic plasma membrane function, trans synaptic anterograde signaling, transporter transmembrane activity) to be downregulated in AD, and one, the LINGO1-TROY-NgR complex, which was previously suggested to be important for modulating glial-neuronal interactions in demyelinating lesions, upregulated in AD 44 . In APOE3/3 astrocytes, Ion and acid transport (receptor ion channel activity, regulator ion channel activity, acid sodium organic symporter, ferrous ferric ion binding, intracellular ferritin iron sequestering), glutamate receptor activity (mGLUR2, mGLUR3, mGLUR4, mGLUR7, mGLUR8), metabolic (aspartate uptake, astrocytic metabolism) as well as autolysosome activities (scavenging class receptors, secondary lysosome, autolysosome) were downregulated in AD, and myelin maintenance (PRNP, ASAH1), cell adhesion (FLRT3, LPHN3, UNC5B, UNC5D), and Vascular endothelial growth factor (VEGF) induced heat shock protein 90 (hsp90) complex were upregulated in AD, indicating perturbation in processes important for autophagy and stress response which are known to accompany disease progression 4,5 . APOE3/4 astrocytes uniquely showed upregulation in pathways related to post-synaptic scaffold proteins (e.g., DLGAP1, DLG4, DLC1 and SHANK3) and actin assembly at cell junctions, but downregulation of synaptic membrane and neurotransmitter pathways, neurogenesis and nervous system development in AD.
In the entorhinal cortex, common astrocyte modules for AD cells relative to controls showed perturbation for processes such as aspartate update metabolism, astrocytic affinity, synaptic membrane organization, macromolecular cellular localization and transmembrane transport ( Supplementary Fig. 6). In APOE3/3 astrocytes of the entorhinal cortex, we observed a downregulation of ion and neurotransmitter transport related pathways (intracellular ion and ferritin iron sequestering) in AD. APOE3/4 astrocytes in the entorhinal cortex had mostly upregulated pathway enrichment modules in AD, in contrast to what was observed in prefrontal cortex. Many of these pathways governing cellular homeostasis, such as ATP synthesis, transmembrane cation transport, amyloid fibril formation and exosome regulation, and macromolecule and protein plasma membrane localization.
Microglia, the resident brain macrophage, contributes to neuroinflammation in AD and produces APOE upon activation in the brain 1,4 . Differentially enriched pathways were predominantly upregulated in APOE3/4 microglia in AD patients in both prefrontal ( Supplementary Fig. 5) and entorhinal cortices (Fig. 6b). In APOE3/3 microglia, however, most significantly enriched pathways were downregulated in AD. Within the entorhinal cortex, changes in gliogenesis, myelination, cation transmembrane transport, cellular projection, synaptic spine development, and synaptic junction assembly pathway network modules were shared in APOE3/3 AD and APOE3/4 AD but perturbed in opposite directions, downregulated in APOE3/3 and upregulated in APOE3/4 microglia (Fig. 6b). The ITGAV-ITGB-SPP1 complex, not previously linked to AD to our knowledge, was significantly upregulated in both brain regions in APOE3/3 microglia in AD, but only in the prefrontal cortex in APOE3/4 microglia in AD ( Fig. 6b and Supplementary   Fig. 5). The downregulation of iron homeostasis and ferritin complex, a protein that binds to iron and reflects the level of iron storage in the body, was observed in APOE3/3 microglia and astrocytes of both prefrontal and entorhinal cortex in AD (Fig. 6a, Fig. 6b, and Supplementary Overall, network analysis comparing neurons from two brain regions yielded many similar perturbed biological processes within each APOE genotype in AD (Fig. 6c). In APOE3/3 neurons, shared differentially perturbed processes between brain regions were mostly related to regulation of membrane homeostasis (intrinsic integral membrane component, transporter transmembrane activity regulation), neuron projection (positive neurogenesis and differentiation regulation) and synaptic development (presynapse organization assembly, synapse assembly structure regulation, postsynaptic specialization assembly density). Pathway networks in APOE3/3 neurons specific to the prefrontal cortex pertain to cell structure development (actomyosin actin-based structure, extension growth development, anchoring junction, adherens cell), while the entorhinal cortex showed unique modules relevant to cellular energy production (oxidative respirasome synthesis, metabolic ATP nucleotide process). From APOE3/4 neurons, we observed a more diverse population of shared network modules between the two brain regions, including functional processes related to protein trafficking vesicles, myelination, membrane assembly, and voltage gated channel and neurotransmitter receptor regulation.
Amyloid fibril formation was uniquely differentially regulated in APOE3/4 neurons and observed in both brain regions in AD, while an amyloid beta precursor formation module was only observed APOE3/4 neurons in prefrontal cortex in AD.
In oligodendrocytes, which provide myelination, we observed common upregulation in proteasomal degradation (e.g., PSMA1) and the LINGO1-TROY-NgR complex, and downregulation in regulating ion activity in the prefrontal cortex ( Supplementary Fig. 5 Fig. 6). As in prefrontal cortex, for APOE3/3 oligodendrocytes, we found an upregulation for the ITGAV-ITGB-SPP1 complex and a downregulation for ion transport activity, protein refolding, and regulation of MAP kinase signaling activity (e.g., positive regulation of Erk1 and Erk2) in AD.
Lastly, in APOE3/4 oligodendrocytes, we observed postsynaptic structural specialization to be uniquely downregulated For OPCs in the prefrontal cortex, there were no common network modules across APOE genotypes. In APOE3/3 AD, we identified downregulation for brain cell development processes (AHI1-NPHP1-HAP1) ( Supplementary Fig. 5). In APOE3/4 OPCs, we observed upregulated modules for the ferritin, GAIT and LINGO1-TROY-NgR complexes, and downregulation for glutamatergic synaptic activity, plasma membrane and cell organization, and lipoprotein density in AD, which may have implications for neuronal integrity and lipid transport and metabolism.
In the entorhinal cortex of AD, OPCs across APOE3/3 and APOE3/4 genotypes shared modules for cell motility, neurogenesis and ion regulation synapse assembly, and neurotransmitter transport ( Supplementary Fig. 6). We also observed upregulation of the LINGO1-TROY-NgR, and downregulation of glutamatergic signaling in APOE3/3 OPCs in AD. Specific to APOE3/4 OPCs in AD, we identified upregulation of processes related to aerobic metabolic processes, stress response, autophagy, amyloid fibril regulation, demyelination, and immune response.

Discussion
As APOE4 is the greatest known genetic risk factor for AD, we analyzed recently available single-cell transcriptomic datasets from two brain regions to better understand how APOE genotype plays into transcriptional profiles of AD in a cell type specific manner. We aimed to understand whether transcriptional differences exist, and if so, how they might be represented in different cell types across brain regions; which cell types were most affected by APOE genotype; what changes were shared or dissimilar across cell types; and whether such findings are consistent across brain regions. Our differential gene expression analysis involved a comparison between AD and control within each cell type, stratified by APOE genotype. Due to the limitations of the datasets, we restricted analysis to compare AD patients versus control individuals with an APOE3/3 or APOE3/4 genotype. In both the prefrontal and entorhinal cortices, we observed shared and unique gene signatures across these APOE genotypes that were often cell type specific, but sometimes spanned many cell types (Fig. 2, Fig. 3, Fig. 4, and Supplementary Fig. 4). Additionally, we observed more DEGs unique to APOE3/4 cells in AD versus control when compared to DEGs for APOE3/3 cells in AD versus control and more DEG overlaps across cell types in APOE3/4 AD, suggesting a stronger and more global AD-related molecular response when one copy of the APOE4 allele is present.
In the prefrontal cortex, most DEGs that are common across cell types tend to be more strongly differentially expressed in APOE3/4 AD as compared to those in APOE3/3 AD. Additionally, we observed most of the APOE genotype specific changes in APOE3/4 astrocytes, oligodendrocytes and OPCs, where these genes are predominantly downregulated in AD as compared to controls. Neurons, on the other hand, tended to exhibit DEGs of AD versus control that were common across APOE genotypes ( Fig. 2a and Supplementary Fig. 2). Through hierarchical clustering of samples using AD compared to control pseudobulk cell type gene expression (Fig. 2c), we observed clustering by APOE genotype in all cell types except neurons.
In the entorhinal cortex, microglia and oligodendrocytes had the highest proportion of DEGs of AD versus control that were shared across APOE genotypes. Interestingly, these DEGs frequently exhibited opposite log fold-change direction between APOE3/3 AD cells and APOE3/4 AD cells. Additionally, through hierarchical clustering of samples using AD compared to control pseudobulk cell type gene expression, we observed that samples clustered primarily by APOE genotype (Fig. 4c). Compared to the prefrontal cortex, the entorhinal cortex, which is implicated in early stages of AD where tau begins to accumulate and the occurrence of synaptic and neuronal loss is associated with the onset of cognitive impairment 1,4,45 , had a larger number of DEGs of AD versus control in each cell type and larger log2 fold change range, implying a greater magnitude of molecular changes in this region in AD, as compared to prefrontal cortex.
Through pathway and network analysis, we identified biological processes potentially involved in AD pathogenesis that were uniquely modified by APOE genotype (Fig. 6, Supplementary Fig.   5, and Supplementary Fig. 6). While many essential cellular processes were differentially regulated in APOE3/3 neurons in AD, most were related to energy production, membrane regulation, and cellular signaling through synapse. APOE3/4 neurons in AD, on the other hand, demonstrated a perturbation of enriched pathways linked to myelination and protein trafficking vesicle regulation (both endocytosis and exosome), which are important cellular processes that protect the integrity of neurons by providing insulation and filtering toxic elements from these cells. This evidence suggests that APOE, a known lipid metabolizing protein, may play differential roles in maintaining essential metabolic processes for neuronal myelination and vesicle trafficking based on its isoform. Glial cells from APOE3/3 and APOE3/4 AD had many uniquely versus common altered biological processes, identified by the APOE genotype specific pathway modules. This suggests that APOE genotype modifies glial cell biology in different ways compared to its effects on neuronal cell biology during AD progression. Further study on AD pathogenesis focusing on glial cell modification by the APOE genotype might facilitate personalized therapeutic development for AD patients with different APOE genotypes.
While we were able to examine APOE genotype specific changes across cell types in both brain regions, some limitations exist. For one, as we did not have any APOE4/4 controls, we designed our analysis to examine only APOE3/3 and APOE3/4 samples. Each dataset contained only one APOE3/4 control, which was a male sample in both cases. We performed a sensitivity analysis in males of the prefrontal cortex cohort ( Supplementary Fig. 3), where we also observed more differences in perturbed gene profiles across APOE genotypes in astrocytes, oligodendrocytes, and OPCs, and a stronger clustering by APOE genotype than cell type identity. We also observed an imbalance in the entorhinal cohort, where all female samples were APOE3/3 cells, and male APOE3/3 cases were not present. However, here we were able to mitigate this caveat by including sex as a covariate in our model for differential expression to account for the confounding variables introduced by the experimental design.
Secondly, due to the design of the Grubman et al. study, we observed a batch effect, where cases were sequenced in separate batches from controls, and each batch contained only one sex. To mitigate this limitation, we used Seurat's integration workflow and dimensionality reduction to confirm appropriate batch correction and included sex as a covariate in our model ( Supplementary Fig. 1). Comparing across brain regions, we recognize that reasons for the differences we observed include the variability in acquiring each cohort, which is sourced from different sets of individuals and studies. The expected variability inherent in human studies also raises the concern that a limited number of single nucleus transcriptomes (n = 30 for the prefrontal cortex, n = 9 for the entorhinal cortex) may not be reflective of the greater human population. Although this concern is mitigated by the consistent observation that variance in gene expression is explained more by APOE genotype than cell type specificity, it underscores the need to extend these findings to a larger and more diverse population of human subjects in the future. The nature of our analysis only allows for association of transcriptomic changes with APOE genotype, so links to causality might be hypothesized, but additional followup are needed to prove any such potential links.
Moreover, we hope that future studies will generate more data from diverse sets of individuals, across different ages, racial and ethnic backgrounds, with a wider variety of APOE genotypes, and more brain regions, thus allowing for even more extensive insights. With more diverse genomic data, researchers may be able to 1) examine AD relative to control expression differences based on the number of copies of the AD risk variant, 2) stratify by age groups to determine whether changes are age-dependent, and 3) identify more concrete changes that are unique to and shared by different regions of the brain. Ultimately, we identified key AD-related genes and pathways that are specific to APOE genotypes and cell types, especially glial cells, as well as certain consistently affected pathways. These results will inform how glial cells are potentially primary sites of AD-related transcriptional differences based on APOE genotype, suggesting possible mechanisms and vulnerable cell subpopulations relevant to AD pathogenesis, and thus can help to facilitate precision medicine diagnostic and drug discovery efforts.

Materials Availability
This study did not generate new unique reagents.

Data and Code Availability
Single nuclei RNA-Seq data and metadata were accessed from their respective repositories: the prefrontal cortex from the Accelerating Medicines Partnership Alzheimer's Disease Project

Study Cohort Identification
We acquired publicly available single nuclei RNA datasets from repositories specified by the first two single-cell transcriptomic AD studies 19,20 . Samples were classified based on tau neurofibrillary tangles, and amyloid β (Aβ) plaque burden, using Braak clinical staging and Consortium to Establish a Registry for Alzheimer's Disease (CERAD) scores. Cases were identified as individuals with severe tau deposition (Braak stage ≥ IV), and high Aβ load (CERAD score ≤ 2), while non-AD Controls were identified as individuals with low tau (Braak stage I-III) and low Aβ load (CERAD score ≥ 3). The prefrontal cortex dataset consisted of age and sex matched samples from 48 individuals with varying degrees of AD pathology. For APOE genotype stratified analysis, we focused on APOE3/3 and APOE3/4, which consisted of 14 APOE3/3 controls, 1 APOE3/4 control, 9 APOE3/3 cases and 8 APOE3/4 cases ( Table 1). The entorhinal cortex dataset initially consisted of age and sex matched samples from 6 AD and 6 control subjects, which were classified based on pathological analysis of amyloid β plaques, Braak clinical staging, and cognitive impairment records from corresponding treating general practitioners. All cases in this cohort have numerous diffuse and neuritic amyloid beta plaques, and a Braak staging score of VI. For our APOE-stratified analysis, we focused on APOE3/3 and APOE3/4 cells, which restricted the dataset to 4 cases, and 5 controls ( Table 2). Three of the cases were from APOE3/4 individuals, while one was from an APOE3/3 individual, and of the controls, four were from APOE3/3 individuals and the one was from an APOE3/4 individual.

Prefrontal Cortex
We downloaded a filtered raw expression matrix of 17,296 genes and 70,634 cells from the prefrontal cortex from the AMP-AD Knowledge Portal and used Seurat's Read10x function to generate a count data matrix using the raw count matrix, cell names, and barcodes files provided.
A Seurat object was created with the count data matrix and metadata, keeping genes present in at least 3 cells, and cells meeting cohort selection criteria with at least 200 genes. Additionally, we selected samples from APOE3/3 and APOE3/4 individuals (Table 1), which resulted in a dataset with 43,831 cells (Supplementary Table 1) and 17,593 genes. Log normalization was performed with a scale.factor of 10,000, and FindVariableFeatures was run using 3,188 features, as specified in the original paper. The data matrix was then scaled with "nCount_RNA" regressed out, and dimensionality reduction was performed with the appropriate dimensions selected based on the corresponding Principal component analysis (PCA) elbow plot. Dimensionality reduction confirmed that there were no batch effects present ( Supplementary Fig. 1). As we found the original paper's cell type identification to be comprehensive, we kept the cell type labels for the further analysis (Supplementary Table 1). Due to low cell counts, we did not analyze pericytes and endothelial cells.  (Table 2), and a Seurat object was created to consist of genes in at least 3 cells, and cells with at least 200 genes. Normalization was performed using Seurat's SCTransform method, and Seurat's integration workflow was performed to correct the confounded batches introduced by the experimental design. In this dataset, as shown in Table 2, control samples were processed separately from cases, male samples were processed separately from female samples, and all but one batch contained one APOE genotype. Dimensionality reduction was performed using values from the integrated assay to assess successful batch correction ( Supplementary Fig. 5).

Entorhinal Cortex
To identify cell types, we adopted techniques from the original paper. Briefly, Grubman et al used Seurat's AddModuleScore function to calculate association scores using lists of brain cell type markers of an unspecified number from the BRETIGEA 51 package. They labeled cells based on which set of markers they had the highest score for, identified hybrids as cells where the highest and second highest score were within 20% of each other, and relabeled unidentified cells  Table 2).

Cell type specific APOE-stratified Differential Expression Analysis
To generate transcriptomic disease signatures relative to APOE genotype in each cell type, we used Limma-Voom 52,53 .We included sex as a covariate in our design formula, as sex, instead of batch, accounted for the confounding relationships introduced by the original study design, allowed for an appropriate model fit, and avoided the collinearity limitation observed with including batch in the design. Additionally, as samples were age matched, we also did not include age in our design formula. A dge list object was then created from a matrix of counts extracted from the corresponding Seurat objects. To improve the accuracy of mean-variance trend modeling and lower the severity of multiple testing correction, lowly expressed genes were filtered out using edgeR's FilterByExpr with default parameters. Normalization was performed with Trimmed Mean of M-values with singleton pairing (TMMwsp), followed by voom, model fitting with a contrast matrix of each case-control comparison for each cell type-APOE group, and Empirical Bayes fitting of standard errors. We performed a cell type specific AD versus control gene expression comparison in each APOE variant group separately in our defined prefrontal cortex cohort, entorhinal cohort, and male-only prefrontal cortex cohort, in which we excluded sex as a covariate. Differentially expressed genes were selected using a Benjamini-Hochberg (BH) corrected p-value less than 0.05, and an absolute log base 2-fold change greater than 0.25.

Functional Enrichment Analysis and Network Visualization
We performed an overrepresentation analysis of DEGs from the cell type-specific APOEstratified analysis of cells from the prefrontal and entorhinal cortex using gprofiler 43 , a web tool for functional enrichment using an input gene list. We queried differentially expressed genes comparison split by upregulated and downregulated expression to identify enriched pathways. In addition to Gene Ontology cellular components, biological processes, and molecular functions, our enrichment analysis also provided pathways from the Human Protein Atlas, Human Phenotype Ontology, KEGG, Reactome, and Wiki pathways. We followed a previously established protocol 54 for network enrichment analysis on pathway results derived from our cell type specific DEGs. Briefly, pathway results were imported into the Cytoscape visualization application, EnrichmentMap. We collapsed redundant and related pathways into single biological themes and further filtered significant pathways using a BH adjusted p-value <0.01. Individual biological themes were defined and summarized using the AutoAnnotate Cytoscape application.

Acknowledgements
We   Overview of cohort sample definition and workflow for APOE genotype stratified cell type specific differential gene expression and functional enrichment. AD and non-AD cells were determined based on tau tangle (Braak) and amyloid β plaque (CERAD) burden. Cell types were identified, and AD versus non-AD differential expression and pathway network enrichment analyses were performed separately for APOE3/3 and APOE3/4 cells of each cell type.    : APOE genotype stratified cell type specific disease signatures across brain regions. a. Upset plots indicating intersections of AD versus non-AD differentially expressed genes (DEGs) (Benjamini Hochberg (BH) adjusted p-value < 0.05 and absolute log2 fold change (FC) > 0.25) within cell types across brain region and APOE genotype. Rows correspond to brain region and APOE genotype pairings. The bar chart shows the number of single and common sets of DEGs across brain regions and APOE genotype pairings. Single filled dots represent a unique set of DEGs for the corresponding brain region and APOE genotype pairing. Multiple filled black dots connected by vertical lines represent common sets of DEGs across brain region and APOE genotype pairings. Bar chart colors correspond to whether DEGs are shared between brain regions or APOE genotype using the bottom right key, b. log2 FC scores of all genes in the DE analysis of both brain regions clustered by cell type, brain region, and APOE genotype.  Tables   Table 1: Prefrontal cortex cohort  Table 2: Entorhinal cortex cohort