Protective exercise responses in the dentate gyrus of Alzheimer’s disease mouse model revealed with single-nucleus sequencing

Abstract

Exercise’s protective effects in Alzheimer’s disease (AD) are well-recognized, but cell-specific contributions to this phenomenon remain unclear. Here, we used single-nuclei RNA-sequencing (snRNA-seq) to dissect the response to exercise (free-wheel running) in the neurogenic stem cell niche in the hippocampal dentate gyrus in male APP/PS1 transgenic AD model mice. Transcriptomic responses to exercise were distinct between wild-type and AD mice, most prominent in immature neurons. Exercise restored the transcriptional profiles of a proportion of AD-dysregulated genes in a cell type-specific manner. We identified a neurovascular-associated astrocyte (NVA) subpopulation whose abundance was reduced in AD, whereas its gene expression signature was induced with exercise. Exercise also enhanced the gene expression profile of disease-associated microglia. Oligodendrocyte progenitor cells were the cell-type with the highest proportion of dysregulated genes recovered by exercise. Lastly, we validated our key findings in a human AD snRNA-seq data set. Together, these data present a comprehensive resource for understanding the molecular mediators of neuroprotection by exercise in AD.

INTRODUCTION

Neurological impairment caused by neurodegenerative diseases, such as Alzheimer’s disease (AD), is a significant and growing health burden. AD is the most common form of age-related dementia. AD pathophysiology is characterized by the accumulation of amyloid plaques, hyperphosphorylated tau, glial activation, and neuroinflammation, eventually leading to synapse loss and neuronal cell death. Excitingly, exercise has well-proven neuroprotective effects in AD. In humans, more physical activity is associated with less risk of AD¹. Exercise – especially endurance exercise – can enhance cognitive function in the elderly and lessen cognitive decline in AD². In AD mouse models, exercise improves cognitive function, decreases Aβ-plaque burden and neuroinflammation, and restores adult hippocampal neurogenesis (AHN)^3–6. To harness exercise’s neuroprotective benefits for the development of innovative treatments against neurodegenerative diseases like AD, we first require a detailed understanding of its molecular mediators at the single-cell level.

Exercise effects are most apparent in the hippocampus and its dentate gyrus (DG), a brain region that displays special neuroplasticity in exercise and selective vulnerability in AD^7,8. A special part of this neuroplasticity with exercise is enhanced AHN. The DG contains the neurogenic stem cell niche, where AHN occurs, and the niche is an important regulator of AHN⁹. The neurogenic niche includes astrocytes, microglia, vascular cells, and resident and migrating immune cells – all important cells in AD pathology¹⁰. During aging and AD, the DG and its neurogenic niche undergo several structural changes – astrogliosis, microgliosis, vascular remodeling, reduced pro-neurotrophic factors, and increased proinflammatory mediators – that collectively impair neurogenesis^5,11.

Recent technological advancements in transcriptional analyses with the advent of single-cell (snRNA-seq) and single-nuclei RNA-sequencing (snRNA-seq) have helped us understand cellular architecture and development of the hippocampus and its DG at the single-cell level^12–14. snRNA-seq has also revealed cell-specific contributions to AD pathology in mouse models and humans^15–17 and has been used to study exercise effects in non-diseased rodent models^18–20. However, we still lack a similarly refined evaluation of exercise-induced neuroprotection in AD. Important open questions include: which cell types respond the most to exercise? Is the exercise-induced transcriptional program similar in wild-type (WT) and AD mice? How much can exercise rescue the transcriptional phenotype in AD?

Here, we used snRNA-seq to systematically dissect the adaptive neuroprotective response to exercise in the DG in AD model mice. Male AD mice and WT littermates were either sedentary or exercised with running wheels, resulting in improved cognitive flexibility. Our snRNA-seq identified cell-type-specific responses to exercise that differed between AD and WT mice. We show that exercise regulates disease-associated microglia in AD mice and discovered a neurovascular-associated astrocyte (NVA) subpopulation that was reduced in AD and strengthened with exercise. Interestingly, in oligodendrocyte precursor cells, the proportion of dysregulated genes recovered by exercise was the highest. Lastly, we validated our key findings in a human AD snRNA-seq data set. Collectively, these data are a unique resource for understanding neuroprotective exercise pathways in AD, an essential step toward developing future drug targets.

RESULTS

Neurogenic niche response to exercise and AD at the single-cell level.

To investigate the adaptive response by the neurogenic stem cell niche to exercise and AD, seven-month-old male wildtype and age-matched APP/PS1 mice, an established transgenic AD mouse model, were exercised by voluntary free-wheel running for 60 days (wildtype-sedentary=WS, wildtype-running=WR, APP/PS1-sedentary=AS, APP/PS1-running=AR). Then, animals were subjected to behavioral tests, and tissues were collected for transcriptomic assessments at 10-months-of-age (Fig. 1a). We used the Morris water maze with reversal, which tests cognitive flexibility and depends on DG function and AHN²¹. Cognitive flexibility was significantly reduced in APP/PS1 mice, consistent with these animals having deficits in the DG and AHN. Conversely, running exercise significantly improved cognitive flexibility in both genotypes, indicating improvements in the DG and AHN (Fig. 1b, Extended Data Fig. 1a–c). Of note, we observed no differences in total running activity or general (loco-) motor activity in the Open field test (OPF), which could bias performance in the Morris water maze (Extended Data Fig. 1d, e). No significant differences were observed in either the Spontaneous alternation behavior, which evaluates working memory, or the contextual fear conditioning with dissimilar contexts. Neither task depends on AHN or the DG (Extended Data Fig. 1f, g)^22,23.

Fig. 1. Neurogenic niche response to exercise and AD at the single-nuclei level.

a, Overview of the experimental design. Seven-month-old male APP/PS1 and WT mice were run for 60 days. Mice were injected with BrdU during the first week. After 60 days of running, behavioral tests were performed at nine months. Afterwards, the dentate gyrus was collected for single-nuclei RNAseq and histological analysis. Created with BioRender.com. b, MWM latency to reach the target platform in reversal (WT-Sed n = 12, WT-Run n = 12, APP/PS1-Sed n = 9, APP/PS1-Run n = 9; three-way ANOVA, Exercise **P = 0.0072, Genotype *P = 0.0434, Exercise × genotype n.s. P = 0.2975). Data represent the mean ± s.e.m. of biologically independent samples. c, UMAP plots of 106,655 annotated nuclei of dentate gyrus from all four groups (WT-Sed or WS, WT-Run or WR, APP/PS1-Run or AR n= 5, APP/PS1-Sed or AS n=4). The 11 distinct cell clusters are shown in different colors. Each dot represents an individual nucleus. d, Canonical marker genes for each cell cluster with the normalized expression and fraction of cells expressing the genes. e, Number of differential expressed genes (DEGs) across the different cell-types. Left bars (darker color), corresponds to WS>WR and right bars (lighter color) corresponds to AS>AR. f, Proportion of DEGs across cell types. Purple, unique DEGs in WS>WR; yellow, unique DEGs in AS>AR; darker grey, DEGs shared between both comparisons; lighter grey, DEGs with opposite direction. g, Scatter plot representations of significant regulator genes in the Immature Neurons via GeneWalk analysis of exercise in WT and APP/PS1 mice. Each dot represents a regulator gene. h, Number of significant cell-cell interactions for all comparisons by CellChat analysis. i, Number of significant ligand-receptor pairs. For each cell type, upregulated connections are displayed in red bars, and downregulated interactions in blue bars. Cells expressing the ligand are designated as “sender cells”. Cells expressing the corresponding receptor are designated as “receiver cells”.

snRNA-seq was performed on the microdissected DG, which contains the neurogenic niche with a 10X Genomics microfluidic system using 3’ RNA capture. From 20 mice (n=5/group), we successfully sequenced the transcriptomes of 108,613 single nuclei with a mean depth and complexity of 19,772 reads and 1,692 genes per nucleus. A sufficient number of nuclei was captured from each animal (Extended Data Fig. 1h, Supplementary Data 1). Upon initial quality control analysis, one mouse from the APP/PS1 sedentary group was identified as a significant outlier using principle-component analysis of pseudobulk transcriptomes and thus excluded from all downstream analyses (Extended Data Fig. 1i). We identified eleven major brain cell types, six neuronal and five non-neuronal, using well-known marker genes matching earlier data sets (Fig. 1c, d, Extended Data Fig. 1j, Supplementary Data 2)^12,24. Importantly, in addition to mature neuronal cell types (mature granule cells (mGCs), Interneurons, and Cajal-Retzius cells), we also captured different stages of neurogenesis (Neuroblast I and II and Immature Neurons). The classification of these neurogenic cell types was more complex and will be further discussed below. No significant changes in cellular composition were identified by Two-way ANOVA, except for OPCs, where there was a significant effect of exercise (p=0.0451) (Extended Data Fig. 1k).

To determine cell type-specific transcriptional responses in exercise and/or AD, we performed differential gene expression (DE) analysis using a Wilcoxon-based approach, which facilitates testing for differentially expressed genes (DEGs) in rarer cell types and more lowly expressed genes, maximizing potential biological insight in this discovery-centered dataset (Fig. 1e, f, Supplementary Data 3). To identify more global regulatory patterns, DE outputs were then utilized for the discovery of driver genes and dysregulated pathways, whereas cell-cell communication pathways were calculated from the full transcriptional profiles (Fig. 1g–i, Extended Data Fig. 1l, 2, Supplementary Data 4–6)^25–27. While cell-type-specific transcriptional responses are described in more detail in subsequent sections, several patterns are apparent at this wider level. Globally, many genes were differentially expressed due to exercise in the two genotypes. Although the number of DEGs was comparable for most cell types, the gene program induced by exercise was distinct between the two genotypes (Fig 1e and f, Supplementary Data 3).

Interestingly, a network-based analysis using GeneWalk revealed significant driver genes, genes that are nodes of connectivity in pathway-centric analyses. We identified large numbers of putative driver genes in exercise in WT and AD mice in Immature Neurons (58 and 37, respectively) and mGCs (528 and 310, respectively) but not in other cell types (Fig 1g, Supplementary Data 4). As expected, the majority (>75%) of exercise-induced driver genes was distinct between the genotypes. Interestingly, Immature Neurons and mGCs shared many driver genes. About 50% of Immature Neurons’ driver genes were also driver genes in mGCs. The glucocorticoid receptor Nr3c1 is the top driver gene in WSvsWR. The calcium-binding protein Calmodulin1 (encoded by Calm1) is the top driver gene in ASvsAR (Extended Data Fig. 1l). Gene set enrichment analysis (GSEA) also points towards different exercise-enriched pathways in WT mice compared to AD mice (Supplementary Data 5). Interestingly, some AD-dysregulated pathways partially recover with exercise (Extended Data Fig. 2, Supplementary Data 5).

Lastly, we examined putative cell-to-cell communication by assessing the cell-specific expression of receptor-ligand pairs using CellChat (Fig. 1h–i, Supplementary Data 6)²⁸. Despite representing less than 10% of cells in our dataset, signaling to and from neurogenic lineage cells (Neuroblast I, Neuroblast II, Immature Neurons) was exceptionally prominent in this analysis. Overall, neurogenic connections were downregulated with AD. Interestingly, when exercised, the majority of neurogenic connections were upregulated, suggesting that part of the dysregulated cell-to-cell communication in AD was restored by exercise (Fig. 1i).

In summary, our data show that exercise elicits cell-type-specific responses that are distinct between WT and AD mice. Furthermore, our data indicate that exercise can rescue transcriptional dysregulation in AD mice, although the rescue was not complete. Interestingly, cells of the neurogenic lineage were an important part of this plasticity.

Remodeling of adult hippocampal neurogenesis in exercise and AD.

Enhancement of AHN is an important mechanism of neuroplasticity in exercise. To assess how exercise and AD affect AHN, the mice were injected with BrdU during the first ten days of running, and BrdU+/NeuN+ cells were quantified in the DG by immunofluorescence (IF). These cells represent surviving newborn neurons generated during these first ten days. We also quantified doublecortin+ (DCX+) cells in the DG, representing immature neurons that were 14–21 days old at tissue collection²⁹. BrdU+/NeuN+ cells and DCX+ cells were significantly reduced in AD and were rescued by running, in line with earlier reports^4,5,30 (Fig. 2a, b, Extended Data Fig. 3a, b).

Fig. 2. Remodeling of adult hippocampal neurogenesis in exercise and AD.

a, Quantification of DG’s BrdU+NeuN+ adult-born neurons and representative confocal images of the DG stained with anti-BrdU (green) and NeuN (red). The white arrows indicate double-positive cells. Scale bar, 100 μm. n = 6 per group. Two-way ANOVA, Exercise ***P = 0.0006, Genotype *P = 0.0446, Exercise × genotype n.s. P = 0.101.b, Quantification of DCX+ cells and representative confocal images of the DG stained with anti-DCX (red) and DAPI (blue). Scale bar, 50 μm. n = 6 per group. Two-way ANOVA, Exercise **P = 0.0035, Genotype ****P <0.0001, Exercise × genotype n.s. P = 0.1222. Data (a, b) represent the mean ± s.e.m of biologically independent samples. c, Schematic of adult hippocampal neurogenesis (adapted from³¹). Molecular layer (ML), granular cell layer (GCL), and subgranular zone (SGZ). Created with BioRender.com. d, Heatmap shows the normalized mean expression (z-score) of markers for each neurogenesis stage. e, Heatmap shows the normalized mean expression (z-score) per group of the top 50 marker genes for Neuroblast I, Neuroblast II, Immature Neurons, and Mature Granule Cell clusters.

AHN follows well-described cellular stages with established marker gene expression (Fig. 2c)³¹. Excitingly, by combining well-established markers with markers from a recent scRNA-seq study of the DG by Hochgerner et al¹², we identified several cell types of the neurogenic lineage, namely neuroblasts (Neuroblast I and Neuroblast II), Immature Neurons, and mGCs in our data set (Fig. 2d and e, Supplementary Data 7). Of note, our Neuroblasts I and Neuroblasts II are most similar to Hochgerner’s Neuroblast 2 stage, not their Neuroblast 1 stage. While there was good overlap for the Neuroblast and mGCs marker genes between the data sets, overlap was less for the Immature Neurons, possibly due to the difference in age of the mice used in the studies: middle-age vs. young-adult (Extended Data Fig. 3c). We did not identify the much rarer earlier stages of AHN, such as radial glia-like cells (RGL) (Nes, Sox2, Gfap, Ascl1, Lpar1, Tfap2c) or neuronal intermediate progenitor cells (nIPC) (Nes, Sox2, Fabp7, Neurog2, Eomes, Top2a) (Extended Data Fig. 3d).

Next, we investigated genes that were dysregulated in AD and recovered with exercise. These genes we termed “recDEGs” (Supplementary Data 8). AD genes were considered recovered if ASvsAR adjusted p-value < 0.05. The recovery score represents the absolute value of the logFC of ASvsAR. In Neuroblast I, Neuroblast II, and Immature Neurons, there were 11, 4, and 454 genes significantly dysregulated in AD, respectively. Of those, 4, 2, and 213 genes, were recovered by exercise, respectively (Fig. 3a, Extended Data Fig. 3e, f, Supplementary Data 8). The smaller number of AD-dysregulated and recDEGs in Neuroblasts was probably due to their overall lower cell number (Extended Data Fig. 3e, f). RecDEGs of interest in the Immature Neurons are highlighted in Fig. 3b, c. For visualization purposes, we separated genes that were expressed >3% of cells in one group, but <10% of all cells, and named these genes “Rare genes” hereafter (Fig. 3c). ATP Synthase Inhibitory Factor Subunit 1 (Atpif1) and ATP/ADP translocase, solute carrier family 25 member 4 (Slc25a4) are important for mitochondrial function. ATPase H+ Transporting V0 Subunit C (Atp6v0c) is involved in lysosomal function in neuronal-like cell lines³². Altered neuronal ATPIF1 expression is observed in AD^15,33. Brk1 and Atp6v0c were also significant driver genes in the GeneWalk analysis in ASvsAR (Fig. 1g). Interestingly, most recDEGs were similarly regulated in Neuroblast I and Neuroblast II, and several were recDEGs in mGCs (Atpif1, Atp6v0c, Brk1, Slc25a4, Rabac1) (Extended Data Fig. 3g).

Fig. 3. snRNA-seq of the Immature Neurons reveals specific changes in exercise and AD.

a, Scatter plot showing the correlation between AD and exercise effects in Immature Neurons. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. b, c, Dot plots showing recDEGs (b) and ‘rare’ recDEGs (c) in Immature Neurons. In each, the hue and size of the dot represent mean expression and fraction of non-zero count nuclei, respectively. d, Embryonic neural stem and progenitor cells were transduced with LV-shRNA and maintained in proliferating media for 5 days. Cell proliferation rate was measured using EdU+ cells normalized to DAPI+ cells. n = 4 per group. One-way ANOVA followed by Dunnett’s against shCtrl, shAtp6v0c ***P = 0.0002, shAtpif1 ***P = 0.0001. e-h, Primary cortical neurons were transduced with LV-shRNA for five days and treated with 20μM recombinant Amyloid-beta 42 for the last 16h (e and f), or Abeta-enriched Tg2576 conditioned-media for the last 3h (g and h). PrestoBlue HS normalized cell viability (e and g), and ratio between dead and live cells (f and h). n = 6 per group. One-way ANOVA followed by Dunnett’s against shCtrl, ****P < 0.0001. i-j, Primary adult hippocampal neural stem and progenitor cells were transduced with LV-shRNA and maintained in proliferating media for 5 days. Representative confocal images of IF staining for EdU (red) and nestin (green) with scale bar = 100μm (i) and quantification of cell proliferation rate using EdU+ cells normalized to DAPI+ cells (j), n = 4 per group. Unpaired two-tailed t-test, ***P = 0.0006. k-m, LV-shRNA was delivered by unilateral stereotaxic injection into the dentate gyrus of WT mice. Representative confocal images of the DG stained with anti-DCX (red) and DAPI (blue) (k), quantification of DCX+ neurons in the DG (l, n = 5 animals per group, unpaired two-tailed t-test, **P = 0.004), Atpif1 gene expression in the hippocampus by QPCR (m, n = 10 animals per group, one-way ANOVA followed by Dunnett’s against shCtrl, ****P < 0.0001). Scale bar, 200μm. Data represent the mean ± s.e.m. of biologically independent samples.

To validate top recDEGs identified in the Immature Neurons, we used lentiviral shRNA-mediated knockdown in differentiating primary cortical neuronal (PCN) cultures. Successful knockdown of Atpif1 and Atp6v0c, but not Slc25a4, reduced neuronal cell viability (Extended Data Fig. 4a–d). Furthermore, knockdown of Atpif1 and Atp6v0c in primary embryonic neural stem and progenitor cultures greatly diminished cell proliferation, with Atpif1 knockdown abnormally altering morphology (Fig. 3d, Extended Data Fig. 4e). When these progenitors were differentiated into neurons, they displayed diminished expression of neuronal differentiation markers and synaptic genes, indicating a less mature neuronal phenotype (Extended Data Fig. 4f–h). Next, we tested whether Atpif1 and Atp6v0c were neuroprotective in two different Aβ-dependent cytotoxicity assays. First, we incubated PCNs with synthetic Aβ42 peptide (Fig. 3e, f, Extended Data Fig. 4i, j). In a second assay, we incubated the neurons with conditioned media from PCNs from APPswe transgenic mice, which is enriched with human Aβ40 and Aβ42 oligomers (Fig. 3g, h, Extended Data Fig. 4k, l). In both assays, loss-of- Atpif1 or Atp6v0c significantly reduced cell survival. Similarly, knock-down of Atpif1 in primary adult neural stem and progenitor cells derived from the subgranular zone of the DG resulted in reduced proliferation, reduced neuronal differentiation, and decreased cell survival (Fig. 3i and j, Extended Data Fig.4m–o).

Lastly, as a proof-of-principle that our recDEG Atpif1 is an important regulator of AHN in vivo, we used unilateral stereotaxic injections to deliver shRNA-lentivirus directly into the DG of wild-type mice, where VSVg-pseudotyped lentivirus preferentially targets neuronal progenitor cells and immature neurons (as previously described^34,35). LV-shAtpif1 significantly reduced Atpif1 expression compared to LV-shControl or the contralateral, untreated side. Furthermore, loss-of-Atpif1 severely impaired AHN as indicated by fewer DCX+ cells in the DG (Fig. 3k–m). Taken together, the data show that our snRNA-seq data is an important resource for novel exercise mediators.

Exercise regulates DAM-like microglia in AD mouse models.

Microglia, the resident immune cells in the brain, are a crucial part of the neuroinflammatory response in AD³⁶. To assess the number of microglia, hippocampal sections were stained with an anti-IBA1 antibody, which labels microglia and tissue-resident CNS macrophages. The number of IBA1+ cells was elevated in the DG of AD mice compared to WT, and lowered with exercise AD mice (Fig. 4a, b). Interestingly, the effect was slightly stronger in the dorsal than the ventral DG (Extended Data Fig. 5a, b). Although in our snRNA-seq data the percentage of Aif1+ cells was decreased in AD (Aif1 encodes for IBA1), the expression levels of Aif1 in those cells were heightened in AD and lowered with running, thus increasing or decreasing, respectively, the ability to detect these cells by an IBA1 antibody in IF (Fig. 4c). These findings strengthen the argument for using multiple techniques to interrogate microglia function. Of note, we confirmed that perivascular macrophages (Mrc1, Cd163, Cd74), monocytes (S100a4), B cells (Cd79b, Rag1), T cells (Trbc2, Cd3g), or natural killer cells (Nkg7) had no significant contributions to the microglia cell cluster (Extended Data Fig. 5c)^17,37.

Fig. 4. Exercise regulates DAM-like microglia in AD mouse models.

a, Representative confocal microscopy images of IBA1+ microglia (red), GFAP+ astrocytes (green), Aβ plaques (3D6 antibody, magenta), DAPI (blue), and composite images from WT and APP/PS1, sedentary or running, mice. Scale bar, 100 μm. b, Quantification of IBA1+ microglia per mm² in the DG (WT-Sed n = 6, WT-Run n = 6, APP/PS1-Sed n = 5, APP/PS1-Run n = 6; Two-way ANOVA, Exercise ***P = 0.0007, Genotype ****P < 0.0001, Exercise × genotype **P = 0.0026, followed by Fisher’s LSD, WT-Sed vs Run ns P = 0.6816, APP/PS1 Sed vs Run ****P < 0.0001). Data represent the mean ± s.e.m. of biologically independent samples. c, Dot plot of Aif1 mean non-zero expression in the microglia cluster. d, Scatter plot showing the correlation between AD and exercise effects in Microglia. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. e, Dot plots showing recDEGs in Microglia. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. f, Overlap between the mouse AD-dysregulated genes (mAPP/PS1) and their change with running (mAPP/PS1-Run) from the present data set with snRNA-seq data from human parietal cortex microglia of sporadic AD (hAD) or AD autosomal dominant (APP and PSEN1, hADAD) in Brase et al., 2023. LogFC, calculated in comparison to healthy age-matched controls (hAD and hADAD), sedentary WT mice (mAPP/PS1), or sedentary AD mice (mAPP/PS1-Run), are color-coded and listed for each gene and comparison. Bold logFC indicates statistically significant adjusted p values. g, UMAP projection of the microglia subclusters (subcluster 0 in blue and subcluster 1 in orange). h, Cell composition in percentage for each subcluster shown in g (WS, WR, AR n = 5/group; AS n = 4). i, Heatmap shows the normalized mean expression (z-score) of the marker genes for each microglia subcluster (subclusters color-coded as in g). The homeostatic and DAM genes that appeared in our dataset are labeled. j, Heatmap shows the normalized mean expression (z-score) per group of homeostatic and DAM microglia genes reported by Chen et al. in our dataset (subclusters color-coded as in g).

In microglia, from 156 AD-dysregulated genes, nine were recovered by exercise (Fig. 4d, Supplementary Data 8). In addition, we highlight the sub-significant targets Marchf1 (ASvsAR adj. p=0.05382) and Maf (ASvsAR adj. p=0.06186) (Fig. 4e). Expression of MARCHF1, a regulator of key immunoreceptors, was altered in border-associated macrophages in meninges in humans with AD³⁸. Interestingly, mutations in MAF and CST3 are linked to an increased AD risk^39,40. This CST3 mutation decreases Cystatin C secretion, and overexpression of Cystatin C decreases Aβ load in AD models⁴¹. Of note, Cst3 was also a recDEG in astrocytes.

To provide the crucial validation of our data in humans, we used a snRNA-Seq data set from the parietal cortex of AD and control brains⁴². This data set was of special interest to us because, in addition to patients with sporadic AD (sAD), it also contained patients with autosomal-dominant AD (ADAD), with mutations in amyloid precursor protein (APP) or presenilin 1 (PSEN1), which form the basis for our transgenic AD mouse model. Notably, this data set contained male and female subjects to examine sex-specificity of our target genes. Fifteen of our AD-dysregulated genes in microglia were also significantly dysregulated in patients with ADAD, and one gene in patients with sAD (Fig. 4f, Supplementary Data 9).

We detected two distinct microglia subclusters, the first enriched for Elmo1 and Nav2 and the other enriched for Fgf13 and Myo1e. Only AD mice, but not WT mice, contributed to the second subcluster (Fig. 4 g–i, Extended Data Fig. 5d, Supplementary Data 10, 11). Using published classifications for disease-associated microglia (DAM), activated response microglia (ARMs), and a recent sc/snRNA-seq-based classification for five distinct microglia subtypes^17,37,43, we found that our first subcluster aligned well with “homeostatic microglia” and our second subcluster aligned well with DAM microglia (Fig. 4i, j, Extended Data Fig. 5e). DAMs ameliorate AD progression by facilitating Aβ plaque uptake through TREM2^44,45. Interestingly, exercise upregulated established DAM-marker genes as well as our subcluster 1 marker genes in AD mice (Fig. 4j, Extended Data Fig. 5f, Extended Data Fig. 5g). Using QPCR of microglia isolated from 5xFAD transgenic AD mouse model, we confirmed a trend for exercise to increase this DAM-related gene signature (p=0.0616) (Extended Data Fig. 5h). In contrast, using a high-confidence gene signature for homeostatic microglia, we did not detect a significant exercise effect in our snRNA-seq or QPCR data (Extended Data Fig. 5g and h). In summary, our study demonstrates that DAMs respond to exercise.

Exercise shifts the transcriptional state of astrocytes.

Astrocytes are major homeostatic cells in the brain and play an important role in AD pathogenesis⁴⁶. Glial fibrillary acidic protein (GFAP) is a well-accepted marker of reactive astrocytes. IF staining of hippocampal sections for GFAP showed no alterations in the number of GFAP+ cells in the DG with exercise or AD, similar to our snRNA-seq data. (Fig. 4a and 5a, Extended Data Fig. 6a, b). However, Gfap mean expression was increased in AD mice (Fig. 5b).

Fig. 5. Exercise shifts the transcriptional state of astrocytes.

a, Quantification of GFAP+ astrocytes per mm² in the DG (WT-Sed n = 6, WT-Run n = 6, APP/PS1-Sed n = 5, APP/PS1-Run n = 6; Two-way ANOVA, Exercise n.s. P = 0.7639, Genotype n.s. P = 0.7652, Exercise × genotype n.s. P = 0.6449). b, Gfap mean non-zero expression in the astrocyte cluster across different experimental groups. c, Scatter plot showing the correlation between AD and exercise effects in Astrocytes. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. d, e, Dot plots showing recDEGs (d) and ‘rare’ recDEGs (e) in Astrocytes. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. f, Overlap between the mouse AD-dysregulated genes (mAPP/PS1) and their change with running (mAPP/PS1-Run) from the present data set with snRNA-seq data from human parietal cortex astrocytes of sporadic AD (hAD) or AD autosomal dominant (APP and PSEN1, hADAD) in Brase et al., 2023. LogFC, calculated in comparison to healthy age-matched controls (hAD and hADAD), sedentary WT mice (mAPP/PS1), or sedentary AD mice (mAPP/PS1-Run), are color-coded and listed for each gene and comparison. Bold logFC indicates statistically significant adjusted p values. g, UMAP projection of the astrocytes subclusters (subcluster 0 in blue and subcluster 1 in orange). h, Cell composition in percentage for each subcluster shown in g (WS, WR, AR n = 5/group; AS n = 4). i-l, Validation of the presence of cadherin-4^high astrocytes in the DG, by IF. Representative confocal images of cadherin-4 (CDH4, magenta), GFAP (orange), and vascular endothelium (ICAM2+ cells, green). White arrows mark CDH4^high astrocytes. Scale bar, 100μm. On the far right, examples of CDH4^−/low and CDH4^high astrocytes in the DG. Scale bar, 20μm (i). Quantification of CDH4^high astrocytes in APP/PS1 (j), CDH4^high astrocytic domain size (l, CDH4^high = 61 and CDH4^−/low = 73 cells; two-tailed Mann-Whitney test ****P < 0.0001), and CDH4^high astrocytes distance to blood vessel (k, CDH4^high=81 and CDH4^−/low=82 cells; two-tailed Mann-Whitney test ***P < 0.001). n = 5 animals, 3 dorsal sections per animal. m, n, Heatmap shows the normalized mean expression (z-score) per group of the marker genes for astrocyte subclusters 0 (m) and 1 (n) for each subcluster (color-coded as in g). Data represent the mean ± s.e.m (a and j), violin plots show the median and quartiles (k-l) of biologically independent samples.

In astrocytes, from the 193 AD-dysregulated genes, 48 were restored by exercise (Fig 5c, Supplementary Data 8). In Fig. 5d, we present interesting recDEGs (Grin2b, Nme7, Atp6v0a1 and St7) and Thra, a sub-significant target of interest, which has been implicated in human AD, showing a very strong recovery in exercise, and was a recDEG in mGCs⁴⁷. Although these genes are highly expressed in neuronal cell types, they are also expressed in 15–35% of astrocytes. Astrocytic glutamate ionotropic receptor NMDA type subunit 2B (encoded by Grin2b) confers neuroprotection against Aβ in vitro⁴⁸. Serine incorporator 1, Serinc1, a “rare” recDEGs in astrocytes, shows the highest fold-change with exercise in AD (Fig. 5e). Eighteen of our AD-dysregulated genes were also significantly altered in patients with ADAD, four of those were significantly recovered with exercise in our data set (Fig. 5f, Supplementary Data 9).

To functionally validate recDEGs identified in the astrocytes, we used lentiviral-shRNA mediated knockdown in primary astrocyte cultures. Interestingly, successful knock-down of Nme7 and Thra, increased the expression of reactive astrocyte markers (Gfap, C3, Serpin3n, or Hspb1) (Extended Data Fig. 7a–d). Knock-down of Nme7 and Thra also increased expression of reactive astrocyte markers when astrocytes were stimulated with either an established astrocyte activating cytokine cocktail (IL1α, TNF, and C1q) or with synthetic Amyloidβ42 (Extended Data Fig.7e–l). Taken together, these data suggest that altering recDEGs expression regulates astrocyte reactivity.

We noted two distinct astrocyte subclusters: subcluster 0 (high expression of Cdh20 and Rorb) and subcluster 1 (high expression of Cdh4 and Csmd1) (Fig. 5g–h, Extended Data Fig. 6c, Supplementary Data 10, 11). All four groups (WS, WR, AS, and AR) contributed to both clusters, but in AD subcluster 1 was significantly reduced. To test whether RGLs were potentially hidden in the astrocyte subclusters, we used the high-confidence signatures Lpar1+/Tfab2c+/Aqp4- and Ascl1+/Tfap2c+/Aqp4-¹². We only identified two Lpar1+/Tfab2c+/Aqp4- cells and one Ascl1+/Tfap2c/Aqp4- cell in subcluster 1, and none in the astrocyte subcluster 0 (Extended Data Fig. 6d).

Next, we investigated whether subcluster 1 astrocytes are disease-associated astrocytes (DAA), as described by Habib et al. in 5xFAD mice using snRNA-seq⁴⁹. However, only 1/3 of their 24 high-confidence DAA-makers were increased in our subcluster 1 (Extended Data Fig. 6e). Furthermore, since DAAs are increased in AD, and our subcluster 1 astrocytes were decreased in AD, we conclude that subcluster 1 astrocytes are not DAAs (Fig. 5h). A recent consensus paper on astrocyte nomenclature recommended the term “reactive astrocytes” and supplied markers of the reactive state in AD or other neurodegenerative diseases⁴⁶. Interestingly, only 2/3 of these “reactive markers” were enriched in subcluster 0 and only 1/3 were enriched in subcluster 1 (Extended Data Fig. 6f). Taken together, these results suggest that subcluster 1 astrocytes do not represent classical “DAAs” or classical “reactive astrocytes” but rather a distinct transcriptional cell state.

To better understand the functional role of the astrocyte subcluster 1, we first performed IF staining for the subcluster marker Cadherin-4 (CDH4, or R-cadherin) on hippocampal sections. CDH4 is a calcium-dependent cell-cell adhesion protein that promotes angiogenesis in the developing retina^50,51. We confirmed the existence of an astrocyte subpopulation expressing high levels of CDH4 protein (CDH4^high astrocytes) (Fig. 5i, j). In APP/PS1 mice, this population constituted ~2%, comparable to our snRNA-seq data. Since changes in morphology correlate well with functional changes, we then assessed the morphology of the CDH4^high and CDH4^−/low astrocytes (subcluster 0). CDH4^high astrocytes displayed significantly larger astrocytic domains than the CDH4^−/low astrocytes, further supporting their neuroprotective role (Fig. 5k). An astrocytic domain, also referred to as territory, is defined as the area formed by connecting the tips of their processes⁵². Furthermore, CDH4^high astrocytes were significantly closer to the blood-brain barrier than CDH4^−/low astrocytes (Fig. 5l), suggesting a potential role in neurovascular coupling, which is in line with a neuroprotective function^50,51. Lastly, functional enrichment analysis highlights neuroprotective functions and functions related to neuronal support in this CDH4^high astrocyte subcluster (Extended Data Fig. 6g, Supplementary Data 12). Hence, we termed them neurovascular astrocytes (NVA).

We also evaluated the expression of CDH4 in astrocyte subpopulations in the human data set and identified a distinct human CDH4^high astrocyte subcluster (human subcluster 1) (Extended Data Fig. 6h), which was decreased in ADAD compared to controls⁴². Interestingly, this human CDH4^high astrocyte subcluster was increased in carriers of rs1582763-A, an intergenic allele associated with decreased risk for AD and higher sTREM2 CSF levels^42,53.

Interestingly, exercise upregulated a high-confidence marker gene signature for astrocyte subcluster 1 (Fig. 5m, n, Extended Data Fig. 6i). This exercise effect was validated by QPCR of astrocytes isolated from 5xFAD mice (Extended Data Fig. 6j). Finally, knockdown of the recDEGs Nme7, St7, and Thra reduced expression of Cdh4 and Csmd1 (markers for subcluster 1), while increasing Cdh20 and Rorb expression (markers for subcluster 0) (Extended Data Fig. 7). Taken together these data further validate a neuroprotective role of our novel neurovascular CDH4^high astrocyte subcluster.

Exercise remodels AD-dysregulated pathways in mGCs.

Multiple lines of evidence have shown that physical exercise and environmental enrichment reduce amyloid burden in various AD mouse models^3,54,55. In line with this, we observed smaller Aβ plaques in the dorsal DG of exercised AD mice (Fig. 6a, Extended Data Fig. 8a). Therefore, we investigated APP processing and Aβ degradation pathways at the single-cell level (Extended Data Fig. 8b). Because our AD mice overexpress human mutated APP and PSEN1 predominantly in CNS neurons (Extended Data Fig. 8c) and because intracellular processing of transgenic APP is the major contributor to amyloid plaques, we investigated the APP processing pathway only in neuronal cells. Since transgenic APP and PSEN1 can influence endogenous levels, we excluded the App and Psen1 genes from our analysis. Adam10 and Bace1 showed a general pattern of up-regulation with exercise, especially in neuronal cells (Extended Data Fig. 8d). The same effect of exercise is visible in some genes of the γ-secretase complex (Psenen and Ncstn). Interestingly, Ide but not Mme is increased with exercise in most neuronal cells (Extended Data Fig. 8e). Overall, the expression changes in these genes were subtle and sub-significant.

Fig. 6. Exercise remodels AD-dysregulated pathways in mGCs.

a, Average size of amyloid-beta plaques (3D6 staining) in dorsal DG sections (APP/PS1-Sed n = 5, APP/PS1-Run n = 6, 3 section per animal; Two-tailed unpaired t-test, *P = 0.0188). b, Scatter plot showing the correlation between AD and exercise effects in mGCs. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. c, d, Dot plots showing recDEGs (c) and ‘rare’ recDEGs (d) in mGCs. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. e, f, Representative RNA images of DG showing Chgb puncta (magenta, e) and Scg2 puncta (red, f) in mGCs in WT-Sed, APP/PS1-Sed and APP/PS1-Run. Scale bar, 20μm. g, h, Quantification of Chgb punta (g) and Scg2 punta (h) in the different groups. n = 2 biological independent animals. Kruskal-Wallis followed by the Dunn’s test, Chgb: WSvsAS ****P < 0.0001, ASvsAR ***P = 0.0002, WSvsAR ****P < 0.0001, Scg2: WSvsAS ****P < 0.0001, ASvsAR ****P < 0.0001, WSvsAR ****P < 0.0001. i, j, Overlap between the mouse AD-dysregulated genes (mAPP/PS1) and their change with running (mAPP/PS1-Run) from the present data set with snRNA-seq data from human parietal cortex excitatory neurons of sporadic AD (hAD) or AD autosomal dominant (APP and PSEN1, hADAD) in Brase et al., 2023. LogFC, calculated in comparison to healthy age-matched controls (hAD and hADAD), sedentary WT mice (mAPP/PS1), or sedentary AD mice (mAPP/PS1-Run), are color-coded and listed for each gene and comparison. Bold logFC indicates statistically significant adjusted p values. k, Immediate early genes mean expression and the fraction of cells expressing the genes in mGCs. Violin plots show the median (middle bold line) and upper and lower quartiles (lighter dotted lines) of biologically independent samples.

In mGCs, from 6610 AD-dysregulated genes, 1638 are recovered by exercise (Fig. 6b, Supplementary Data 8). Five top-ranked recDEGs (Fig 6c) were previously reported to be associated with AD: Pkm, Scg2, Chgb, Mdh1, and Hat1^56–60. For example, iPSC neurons derived from AD patients show increased expression of a pathologic isoform of the glycolytic enzyme PKM leading to neuronal cell death. On the contrary, inhibition of this isoform restored neuronal metabolic profile and survival⁶⁰. Using RNAScope on hippocampal sections, we validated the gene expression changes in Scg2 and Chgb in mGCs with exercise and AD (Fig 6e–h).

Four of the 10 top-ranked “rare” recDEGs (Fig 6d) play a role in cytoskeleton-related functions: Dynll2, Tubb2a, Tpt1, Nefl^61–64. Neurofilament Light Chain (Nefl) is downregulated in both human and mouse AD brains and is being evaluated as an AD biomarker⁶². Interestingly, overexpression of Parkinsonism Associated Deglycase (Park7) reduced Aβ pathology and improved the cognitive function of AD transgenic mice⁶⁵. Pkm and Nefl were also driver genes in the GeneWalk analysis in ASvsAR (Fig. 1g, Supplementary Data 4). 346 of our AD-dysregulated genes were also significantly altered in excitatory neurons in patients with ADAD, 52 of which were recDEGs (Fig 6i, j Supplementary Data 9).

As a proxy for neuronal activity around the time of tissue collection, we analyzed the expression of immediate early genes in mGCs (Fig 6k). There was a trend towards increased Bdnf expression in exercised AD mice, which is in line with earlier reports. In addition, the Bdnf downstream target Nrn1 was a recDEG in mGCs (Fig 6i, Extended Data Fig. 8f). Fos, Arc, Egr1 were expressed at much lower levels and in much fewer cells (<1%) than Bdnf or Nrn1 (~20%), and showed a trend towards reduced gene expression with AD and upregulated with exercise.

OPCs and oligodendrocytes are highly plastic in exercise and AD.

Cells of the oligodendrocyte lineage are emergent cell types in AD pathology, including the discovery of “disease-associated oligodendrocytes” (DAO)^16,66. Exercise has been known to enhance oligodendrogenesis⁶⁷. Interestingly, in OPCs AD-dysregulated transcription was the most rescued by exercise, namely 64%. In OPCs, from 39 AD-dysregulated genes, 25 are recovered by exercise (Fig.7a, Supplementary Data 8). Stearoyl-CoA desaturase 2 (Scd2) stood out as a recDEG in OPCs, because exercise almost completely restored its expression to WT levels (Fig. 7b). SCD2 is the rate-limiting enzyme that catalyzes the de novo synthesis of monounsaturated fatty acids (MUFA), which are needed for the sphingomyelin in the myelin sheath⁶⁸. SCD2 is the predominant SCD isoform in oligodendrocytes, and its levels rise with the maturation of OPCs into oligodendrocytes^68,69, which we also observed in our data (Extended Data Fig. 9a). Interestingly, in oligodendrocytes, Scd2 was not a recDEG (Extended Data Fig. 9b).

Fig. 7. OPCs and oligodendrocytes are highly plastic in exercise and AD.

a, Scatter plot showing the correlation between AD and exercise effects in OPCs. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. b, Dot plots showing recDEGs in OPCs. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. c, Number of significant ligand-receptor pairs from CellChat analysis. For each cell type, upregulated connections are displayed in red bars, and downregulated interactions in blue bars. Cells expressing the ligand are designated as “sender cells”. Cells expressing the corresponding receptor are designated as “receiver cells”. d-f, OPCs subcluster UMAP representation (d), mean proportion (e), and marker genes for each subcluster (f). g, Scatter plot showing the correlation between AD and exercise effects in oligodendrocytes. The legend details are the same as in a. h, i, Dot plots showing recDEGs (h) and ‘rare’ recDEGs (i) in oligodendrocytes. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. j, Overlap between the mouse AD-dysregulated genes (mAPP/PS1) and their change with running (mAPP/PS1-Run) from the present data set with snRNA-seq data from human parietal cortex oligodendrocytes of sporadic AD (hAD) or AD autosomal dominant (APP and PSEN1, hADAD) in Brase et al., 2023. LogFC, calculated in comparison to healthy age-matched controls (hAD and hADAD), sedentary WT mice (mAPP/PS1), or sedentary AD mice (mAPP/PS1-Run), are color-coded and listed for each gene and comparison. Bold logFC indicates statistically significant adjusted p values. k-m, Vascular cells subcluster UMAP representation (k), marker genes for each subcluster (l), and mean proportion (m).

CellChat analysis further highlighted the plasticity of OPCs, showing a higher number of downregulated connections with AD and of upregulated connections with exercise. Of note, their communication with neuronal cell types was most pronounced (Fig. 7c). Similarly, in the GSEA, AD downregulated neuronal support functions, such as synaptic transmission and plasticity, whereas exercise enhanced pathways associated with the regulation of membrane potential, such as anion and cation transmembrane transporter activity (Extended Data Fig. 2b). Lastly, we identify a small subcluster of more developed types of OPCs, also known as committed OPCs, marked by high expression of Fyn, Tns3, Plp1, Mpzl1, Gpr17, Bmp4 (Fig. 7d–f, Supplementary Data 10–11)^70,71. However, neither exercise nor AD significantly changed the number of committed OPCs (Fig. 7e, Extended Data Fig. 9c).

In oligodendrocytes, from 152 AD-dysregulated genes, 61 are recovered by exercise (Fig.7g, Supplementary Data 8). Upregulation of Reticulon 3 (Rtn3) decreased the production of amyloid-beta (Aβ)⁷². The Ferritin Heavy Chain 1 (encoded by Fth1), which is secreted by oligodendrocytes through extracellular vesicles, promotes the antioxidant defense system to protect neurons from iron-mediated cytotoxicity⁷³. Neurocalcin Delta (encoded by Ncald), a calcium-binding protein, is downregulated in the hippocampus of patients with AD⁷⁴. Interestingly, all three genes, Rtn3, Fth1, and Ncald, were downregulated in DAOs in AD⁷⁰ (Fig. 7h). Hsp90ab1 and Hsp90aa1, which encode heat shock proteins, were “rare” recDEGs (Fig. 7i), a finding in line with the “cellular response to heat stress” being downregulated in AD and upregulated by exercise in the GSEA (Extended Data Fig. 2b). Twelve of our AD-dysregulated genes were also significantly altered in patients with ADAD, eight of which were recDEGs in our data set (Fig. 7j, Supplementary Data 9).

Buckley et al. (2023) recently used scRNA-seq to evaluate exercise responses in oligodendrocytes and OPCs in the subventricular zone in WT mice throughout the lifespan²⁰. When we compared our findings in WT mice to those in WT mice of similar age in Buckley et al. (2023), we found that ~50% of our WT exercise DEGs were regulated in the same direction as in Buckley et al. (2023) (Extended Data Fig. 9d). Semaphorin-3C (encoded by Sema3c) is a secreted protein associated with axon guidance and cell migration and linked with AD pathology^75,76. Sema3c was downregulated in both datasets in both oligodendrocytes and OPCs, highlighting a potential exercise-related marker gene in these cells across distinct neurogenic regions of the brain. Interestingly, while Sema3c was downregulated with exercise in WT mice, it was upregulated with exercise in APP/PS1 mice and identified as a recDEG across several cell types. In summary, our results affirm an important role for the plasticity of the oligodendrocyte lineage cells in exercise and AD^16,20,66,67.

Exercise and AD responses in interneurons and vascular cells.

To ascertain whether exercise- and/or AD-specific cell states exist in the remaining cell types in our data set, we subclustered the GABAergic interneurons and the vascular cells. Our interneuron cell cluster contained the following important subtypes: parvalbumin (Pvalb), somatostatin (Sst), vasoactive intestinal peptide (Vip), cholecystokinin (Cck), and reelin (Reln) interneurons (Extended Data Fig. 9e–g). Of note, Lamp5 emerged as a more precise marker gene for reelin interneurons than Reln⁷⁷.

In Interneurons, of 70 AD-dysregulated genes, 32 were recovered by exercise (Extended Data Fig. 9h, Supplementary Data 8). Protein Disulfide Isomerase Family A Member 3 (Pdia3), Transmembrane Protein 106B (Tmem106b), and DnaJ heat shock protein family (Hsp40) member A1 (Dnaja1) have reported roles in AD^78–80 (Extended Data Fig. 9i).

Subclustering of the vascular cells revealed the following cell types using established markers: fibroblasts (Lama1), endothelial cells (Flt1), and mural cells (Pdgfrb), containing smooth muscle cells and pericytes (Fig. 7k–m, Supplementary Data 10–11)^81,82. Exercise has well-documented effects on the vasculature, including boosting de novo angiogenesis^83,84. In line with these effects, exercise significantly increased the percentage of endothelial cells (Fig. 7m, Extended Data Fig. 9j). However, we only sequenced 441 nuclei (~23 nuclei per animal), since vascular cells require special enrichment methods for optimal representation in snRNA-seq^82,85. For Cajal-Retzius cells, which are small, non-GABAergic neurons located in the hippocampal molecular layer, we only sequenced 602 nuclei (~32 nuclei per animal). We did not identify any subclusters in this cell type. Due to the low numbers of nuclei in both cell types, we decided not to perform in-depth analyses.

Lastly, we wanted to discuss several genes that were recDEGs in many different cell types (Extended Data Fig. 9k, Supplementary Data 8). The most striking example was the above-mentioned Sema3c, which was a recDEG in eight cell types: interneurons, mGCs, neuroblast I, neuroblast II, OPCs, oligodendrocytes, microglia, and astrocytes. These findings confirm that exercise not only has cell-specific effects but also has pleiotropic effects.

DISCUSSION

Understanding exercise’s neuroprotective effects on the molecular level is an essential step toward developing disease-modifying therapies for neurodegenerative diseases like AD. Here, we used snRNA-seq of the DG to dissect, at the single-cell level, the adaptive neuroprotective response to exercise in a transgenic AD mouse model. We demonstrated that the exercise-induced transcriptional program is distinct on the single-cell level for male WT and AD mice. Furthermore, at the single-cell level, we identified several critical AD-deregulated genes that exercise restores in AD mice. These recDEGs provide an excellent starting point for future drug targets for AD therapy. Lastly, we discovered astrocytes and microglia-specific cell populations that were dysregulated by AD and restored with exercise.

AHN is an essential mechanism of neuroplasticity in exercise. Enhancing AHN in aging or AD increases cognitive function^4,5,86,87. Recent work has highlighted the importance of a healthy neurogenic stem niche to support AHN^5,9,88. Although the relevance of AHN in humans has been periodically called into question, a recent snRNA-seq study confirmed fewer immature neurons in human AD¹³. We confirmed by IF that AHN is reduced in AD and increased by exercise^4,5,30. We outline how the transcriptional states of several neurogenic lineage cells, namely Neuroblast I, Neuroblast II, and Immature Neurons, change during AD and improve with exercise. We identified several significant recDEGs in Immature Neurons that we validated in vivo and in vitro. We uncovered the exercise-remodeled cellular signaling pathways to and from neurogenic cells. Lastly, we showed that many important neurogenic niche cells, such as astrocytes, microglia, and vascular cells, improved with exercise in AD.

Astrocytes and microglia are key players in the pathology of AD^16,17,37,49. Here, we identify neurovascular astrocytes (NVA), with a potential role in neurovascular coupling and neuroprotection. NVAs were reduced in number in AD. Exercise not only restored the expression of individual recDEGs but also shifts the transcriptional state of this NVA subpopulation. Exercise shifts microglia more towards a DAM-like phenotype. Since DAM ameliorate AD progression by facilitating Aβ plaque uptake through TREM2^17,43–45, it is possible that our observed reduction in Aβ plaques with exercise is due to this shift towards a DAM-like phenotype; future studies are required to explore this mechanism in more detail. Interestingly, the oligodendrocyte lineage cells show remarkable plasticity, in line with recent results in AD and exercise^16,20,66. Surprisingly, we did not find DAO⁷⁰, possibly due to the AD mouse models used, the brain region analyzed, or the time point selected.

Lastly, to increase the translational value of our study, we validated our AD-dysregulated and recDEGs in a snRNA-seq data set from the parietal cortex of subjects with familial AD and sporadic AD⁴². Validation of certain recDEGs in a data set from a different brain region suggests that the importance of these recDEGs is not limited to AD pathology in the DG. Of note, the human data set contained male and female subjects, thereby addressing the potential limitation of using only male mice. In summary, these analyses show that our findings are relevant to the human condition.

To capture the entire gestalt of the neuroprotective response to exercise, we used whole, microdissected tissue. Since mGCs comprised ~70% of nuclei, rarer cell types, such as vascular cells or DAM, were not covered in the same depth. We were also limited to a single time point, therefore, we might not have captured all the transcriptional dynamics occurring over time. Because the APP/PS1 mouse model at this age does not display tau pathology or neurodegeneration, we could not investigate them in the present study. Lastly, running exercise affects a multitude of aspects, among others food intake, metabolism, myokine secretion, or neuronal activity (Extended Data Fig. 9l)^89,90. Our current study was designed to dissect exercise’s effect on the brain, but it cannot determine the causality of all these physiological changes.

Critical next steps in deciphering exercise’s neuroprotection will include simultaneous transcriptional and epigenomic profiling, spatial transcriptomics as well as efforts to elucidate the therapeutic potential of our targets. In conclusion, our study provides a unique resource for investigating exercise effects in AD and provides insight into the underlying mechanisms of neuroprotective effects in different cell types. Taken together, the data show that our snRNA-seq data is an important treasure trove to discover exercise mediators.

METHODS

Animal procedures and ethics statement

All animal procedures were approved by the Institutional Animal Care and Use Committee of the Massachusetts General Hospital (MGH). APP/PS1 mice, their wild-type littermates on a C57BL6 background (JAX: 34832). The 5xFAD (B6SJL-Tg (APPSwFlLon,PSEN1*M146L*L286V)6799Vas/Mmjax, RRID:MMRRC_034840-JAX), was obtained from the Mutant Mouse Resource and Research Center (MMRRC) at The Jackson Laboratory, an NIH-funded strain repository, and was donated to the MMRRC by Robert Vassar, Ph.D., Northwestern University. C57BL/6J (000664, JAX) were obtained from JAX. All experimental animals were housed in the specific pathogen-free environment animal facility at MGH with a regular 12h light and 12h dark cycle from 07:00 to 19:00, at 20–22°C and 30–70% humidity. Water and standard chow diet (Prolab^® IsoPro^® RMH 3000, Irradiated) were provided ad libitum. Mice were group-housed. Seven-month-old male wild-type sedentary (Wildtype Sed, n = 12), wild-type running (Wildtype Run, n = 12), APP/PS1 sedentary (APP/PS1 Sed, n = 9), and APP/PS1 running (APP/PS1 Run, n = 9) were placed individually in cages with (running) or without (sedentary controls) stainless steel running wheels. During the first 10 days, mice were injected with BrdU (50mg/kg i.p). After 60 days of running, behavioral testing was performed. At ten months of age, mice were euthanized for tissue collection. Running activity was tracked using a revolution counter that collected data every hour (VitalView Animal Activity v1.4, Starr Life Science). Separate cohorts of male APP/PS1 of the same age was used for tissue collection for RNAcope (n = 2/group) and IF of CDH4/GFAP/ICAM2 (n = 5). Two-month-old male 5xFAD mice were housed in either cages with a running wheel 5xFAD-run (n = 8) or without a running wheel 5xFAD sedentary (n = 8). Tissues and isolated cells were collected after two months of running. Nine-week-old male wildtype mice were injected stereotaxically with lentivirus (LV) (n = 10/group), injected with BrdU after seven days daily for five days, and tissues were collected 21 days later. The mice were housed individually after surgery to prevent injury.

Stereotaxic Surgery

Mice were injected unilaterally using stereotaxic surgery to deliver one μl of LV-shScramble or LV-shAtpif1 (titer 1E8/ml) to the dorsal and ventral dentate gyrus (DG) using a Model 1900 Stereotaxic Alignment Instrument (Kopf Instruments, Tujunga, CA). The following spatial coordinates relative to bregma were used: Dorsal DG, anterior-posterior (AP) = −2.10 mm; medial-lateral (ML) = 1.9 mm; dorsoventral (DV) = −2.20 mm, and ventral DG, AP = −3.10 mm; ML = 2.88 mm; DV = −3.20 mm⁹¹.

Behavioral Assays

Open field test (OPF)

For the open field test (OPF)⁹², mice were placed individually in an open field arena (27.3 × 27.3 cm, height 20.3 cm) housed within a sound-attenuating cubicle (Med Associates, Inc., St. Albans, VT) and permitted to move freely. Trials lasted 30 or 60 minutes with a five-minute time bin as specified in each experiment. Animal motion and cumulative path length were automatically tracked via 2X and 1 Z 8-beam IR arrays and recorded by Activity Monitor software (Version 4.0, Med Associates, Inc., St. Albans, VT). Mean substitution of distance traveled was used for incomplete measurements when mice escaped the open field box.

Morris water maze (MWM)

In the Morris water maze (MWM), mice were trained to find the hidden platform (15 cm in diameter) within the maze (1.4 m in diameter)^93,94. The water was made opaque using nontoxic white paint (00011–1009, Blick Art Materials) and the temperature was maintained at 25°C. Mice were trained for seven consecutive days with four trials per day. Starting points were changed daily for each trial. The animals were allowed a maximum of 60 s to locate the platform with a ten-minute inter-trial interval. A 60 s probe trial was performed at the indicated time points after the training. For reversal, the platform was moved to the opposite quadrant, and mice were trained to find the hidden platform following the same trial protocol as the acquisition described above. Probe trials were performed at indicated time points after the training. On probe trials, the hidden platform was removed, and time spent in each quadrant was measured. Data collection and analysis were performed using the ANY-maze tracking system (v6.0, Stoelting Co. Wood Dale, IL). Latency, time spent in the correct quadrant, swim distance, and swim speed were calculated.

Contextual fear conditioning (CFC)

Contextual fear conditioning (CFC) was used to assess memory and is based on the murine tendency to show a fear response (freezing) when re-exposed to the context where they received an aversive stimulus (in this case, foot shock). Mice were placed into a conditioning chamber (17 cm long × 17 cm wide × 25 cm high, Maze Engineers, Stokie, IL) with Plexiglas sidewalls and a floor consisting of steel bars. Mice were allowed to explore the chamber for 2 min and were given 2 presentations of a 2 s foot shock (0.5 mA) separated by 2 min. Mice were removed from the chamber 1 min after the last foot shock. Twenty-four hours after the training session, mice were placed back into the conditioning chamber (context A) for 3 min (no electric shock was delivered during this session) and freezing was videotaped and scored with ANY-maze software v6.0 (Stoelting Co. Wood Dale, IL). The freezing response was used as a surrogate marker of memory performance because the memory of receiving the shock, during the training session in context A on day 1, is expected to induce significant freezing episodes during the day 2 session. After testing in the context A, mice were placed into the alternate context chamber (no grid and new patterned walls) (context B) for 3 min as before. Mice that did not freeze after receiving the shocks on day 1 were excluded from that analysis as we could not use freezing as a proxy for learning or memory⁹⁵.

Spontaneous alternation behavior (SAB)

Spontaneous alternation behavior (SAB) tests were conducted in a Y-maze. Each arm of the Y-maze was 35 cm long, 5 cm wide, and 10 cm high. To reduce anxiety in the animals, the testing area was dimmed to 50 lux. Mice were handled for three days before testing. This test consisted of a single 5 min trial, in which the mouse was allowed to explore all three arms of the Y-maze. If a mouse climbed on the maze walls, it was immediately returned to the abandoned arm. The start arm was varied between animals to avoid placement bias. Spontaneous Alternation [%] was defined as consecutive entries in 3 different arms (ABC), divided by the number of possible alternations (total arm entries minus 2) and was scored with ANY-maze software v6.0 (Stoelting Co. Wood Dale, IL). Re-entries into the same arm were rated as separate entries. Mice with fewer than eight arm entries during the 5-min trial were excluded from the analysis as significantly lower exploratory activity biases the analysis⁹⁶.

Antibodies

All primary and secondary antibodies, with source, dilutions, and validations, are detailed in Supplementary Table 1.

BrdU labeling

Mice were injected daily intraperitoneally (i.p.) with BrdU (50 mg/kg, (Sigma, dissolved in 0.9% saline) for the indicated timepoints. Solutions were filtered at 0.22 μm. At the end of the experiment, the mice were anesthetized with isoflurane and perfused transcardially with ice-cold PBS before tissues were collected.

Immunofluorescence (IF)

IF stainings were done following methods described previously⁵. Briefly, PFA-fixed and cryoprotected brains were cryosectioned to 35μm thick. Staining was performed on 35 μm coronal free-floating sections (one-in-six series) from embedded tissue. IF antibodies for BrdU, DCX, 3D6, GFAP, IBA1, nestin, Cdh4, ICAM2 were used. Nuclei were counterstained using Hoechst 33342 (Thermo Fisher Scientific). For antigen retrieval of BrdU in IF, brain sections were incubated for two hours in 50% formamide/2X SSC (0.3 M NaCl and 0.03 M sodium citrate dihydrate) at 65°C followed by two 10-min washes in 2X SSC buffer. After washing, the sections were incubated in 2N HCl at 37°C for 30 minutes, followed by neutralization with 0.1 M boric acid (pH 8.5) and several washes in TBS (pH 7.5). The usual IF protocol was resumed after these steps.

Immunofluorescence Analysis

Unless otherwise noted, all images were captured in Z-stack mode using a confocal microscope with a 20x/0.8 air objective (LSM 900, Zeiss). 16-bit Z-stack images were taken 2μm apart (1μm for GFAP/IBA1 images) for a total Z-stack of 18–30 μm. Maximum projected images were created in ZEN software (ZEN Blue 2, v3.3.89.0000). Images were analyzed with ImageJ software in a blinded manner (v2.9.0/1.53t). Counting of positive cells followed the principle for design-based stereology⁹⁷.

BrdU+/NeuN+ and DCX+ quantification

To quantify BrdU+/NeuN+ and DCX+ cells, for each animal, every 6^th hippocampal section from bregma −1.34mm to −3.88 mm was analyzed⁹⁸. Subregions were defined as follows: dorsal DG: −1.34 mm to −2.06 mm, intermediate DG: −2.06 mm to −2.54 mm, and ventral DG: −2.54 mm to −3.88 mm from bregma⁹⁸. Only cells located within the granular cell layer (GCL) or the subgranular zone (SGZ were counted²⁹. Counting of positive cells followed the principle of design-based stereology for AHN⁹⁷. For missing sections, a mean substitution was performed. Imaging of the brains of LV-shScramble or LV-shAtpif1-injected mice were acquired using a Nikon Ti2 Inverted Microscope with Yokogawa W1 and SoRa Module(W1) and NIS-Elements AR 6.10.01 (Nikon Instruments Inc., Melville, NY).

GFAP+ and IBA1+ quantification

To quantify GFAP+ and IBA+, three sections for dorsal and three sections for ventral DG from each mouse were analyzed. Two regions of interest (ROIs) measuring 200μm × 200μm × 20μm were outlined in the suprapyramidal and infrapyramidal blades of the DG, respectively. A cell was considered positive if it showed a signal for DAPI and GFAP or DAPI and IBA1 and was counted using the Cell Counter ImageJ plugin or QuPath 0.5.0⁹⁹.

CDH4^high astrocytes quantification, morphology and localization assessment

To quantify CDH4^high and CDH4^low astrocytes, three sections of dorsal DG from each mouse were analyzed. Positive cells were detected using the positive cell detection tool from QuPath 0.5.0⁹⁹. Astrocytes were identified using the GFAP signal and CDH4^high astrocytes were identified using the GFAP signal and the CDH4 signal. CDH4^high and CDH4^−/low astrocytes were individually inspected for the presence of a DAPI signal to ensure only single astrocytes were included in the analysis. For the distance to ICAM2+ blood vessel cells and astrocytic domain size, CDH4^−/low cells were selected randomly, using a random number generator. An astrocytic domain, also referred to as territory, is defined as the area formed by connecting the tips of their processes⁵². The astrocytic domains were analyzed using QuPath positive cell detection feature based on the GFAP signal, and the area of each astrocyte as delimited by the GFAP staining was quantified. The threshold for the GFAP channel was kept the same throughout all the images.

Amyloid plaque quantification

To quantify amyloid plaques (staining with 3D6 antibody), three sections for dorsal and three sections for ventral DG from each mouse were imaged with 20X magnification using a slide scanner (Zeiss Axio Scan.Z1). The DG area on each section was outlined as ROI. The respective fluorescent channel for 3D6 antibody staining was isolated. A background subtraction using the rolling ball algorithm (rolling ball=130) and a gaussian blur filtering (σ=2.5) was performed. A binary image was created with an otsu dark thresholding algorithm, and the Analyze Particles plugin in ImageJ obtained the area of every particle >20μm^{2 100}.

RNA in situ hybridization

RNA in situ hybridization was performed on fixed-frozen brain tissue. After perfusion with ice-cold PBS, the brains were immediately dissected and fixed in 4%PFA-PBS for 24h. After the sucrose gradient steps, the brains were embedded in optimal cutting temperature (OCT; Tissue Tek) on dry ice, and processed into 14-μm cryostat sections directly mounted on Superfrost plus slides. RNA in situ hybridizations were carried out using the RNAscope^® 4-Plex Ancillary kit (323120, Advanced Cell Diagnostics-ACD) and the RNAscope Multiplex Fluorescent Reagent Kit v2 (323100, ACD) for 4 plex fluorescent in situ hybridization per the manufacturer’s instructions for fixed-frozen tissue. Commercially available and validated probes for Prox1 (488591, ACD), Scg2 (477691-C2, ACD), and Chgb (1057371-C3, ACD) were used. Whole-DG images were acquired with a Leica Stellaris Confocal Microscope, LAS X Life Science Microscope Software (Leica Micrisystems Inc., Deerfield, IL). For each animal, an unstained tissue was imaged as a negative control to assess levels of background fluorescence. Analysis was done using QuPath 0.5.0⁹⁹. Three ROIs were outlined in the granule cell layer, and Scg2 and Chgb positive puncta in the nuclear space of Prox1+ neurons were identified and quantified.

Microglia and Astrocyte cells isolations

5xFAD mice, either sedentary or runners, were perfused with ice-cold PBS and the brains were immediately dissected to ice-cold HBSS 1x. Cortex and HC halves were transferred to gentleMACS C-tubes (130-093-237, Miltenyi Biotec) and digested using the Neural Tissue Dissociation Kit Postnatal Neurons (130-094-802, Miltenyi Biotec) for 30 min at 37°C on the gentleMACS Octo Dissociator with Heaters (#130-096-427, Miltenyi Biotec) using the 37C_ABDK_01 program. The following transcription and translation inhibitors were included during the digestion step: Actinomycin D (5 μg/ml, A1410, MilliporeSigma), Triptolide (10 μm, T3652, MilliporeSigma), Anisomycin (10 μg/ml, A9789, MilliporeSigma)¹⁰¹. CD11b magnetic labeling and separation was performed as in Bordt et al. 2020, using CD11b (Microglia) MicroBeads (#130-093-636, Miltenyi Biotec) LS Columns (130-042-401) and QuadroMACS Separator (130-090-976) after debris/myelin removal using the Debris Removal solution (#130-109-398, Miltenyi Biotec)¹⁰². The CD11b negative fraction was further processed using the mouse Anti-ACSA-2 MicroBead Kit (#130-097-678, Miltenyi Biotec) for the isolation of astrocytes.

Plasmids and Lentiviral Vector Production

The following lentiviral p.LKO1-puro vectors were procured from Sigma: MISSION^® Non-Target shRNA Control (#SHC016), ATP6V0C-TRCN0000101611; ATPIF1-TRCN0000190288; SLC25A4-TRCN0000069153; NME7-TRCN0000024704; THRA-TRCN0000027109; ST7-TRCN0000179803. Lentiviruses were produced in 293T HEK cells using the pMD2.G and psPAX2 for packing and the Fugene6 transfection reagent as per the manufacturer’s protocol. The lentivirus was purified by ultracentrifugation, following established protocols¹⁰³, and was resuspended in PBS. Small aliquots were stored at −80°C. Viruses were tittered using p24 antigen using QuickTiter^™ Lentivirus Titer Kit (VPK-107). psPAX2 was a gift from Didier Trono (Addgene plasmid # 12260; http://n2t.net/addgene:12260; RRID:Addgene_12260). pMD2.G was a gift from Didier Trono (Addgene plasmid # 12259; http://n2t.net/addgene:12259; RRID:Addgene_12259).

In Vitro embryonic neurosphere and monolayer culture

Embryonic cortices were collected from time-pregnant C57BL/6J wildtype mice on embryonic day (E) 12–13 as previously described¹⁰⁴. After removing the meninges, the cortices were passed through a 1 ml pipette tip. The obtained single cell suspension was cultured in DMEM/F12 (11330–032, Gibco) supplemented with B27 without Vitamin A (12587010, Gibco), EGF (E-1257, Sigma), bFGF (F-0291, Sigma), heparin (H-3149, Sigma) and Penicillin/Streptomycin (30–001-CI, Corning) to encourage neurosphere formation¹⁰⁵. Cells from primary neurospheres were dissociated into single cells using a 1 ml pipette and then treated with indicated lentivirus at a multiplicity of infection (MOI) 4 and seeded in Matrigel-coated 8 well-chamber slides in proliferating media for five days or transferred to differentiation media DMEM/F12 (11330–032, Gibco) supplemented with B27 (17504–044, Gibco), N2, 2% Fetal Bovine Serum (FBS) and Penicillin/Streptomycin (P/S) (30–001-CI, Corning) for five days.

Adult hippocampal neurosphere culture and differentiation

Whole hippocampus was collected from 7–8 weeks old C57BL/6J mice. The hippocampal slices were digested in the MACS Neuronal Dissociation Kit-Postnatal Neurons (130-094-802, Miltenyi) as per the manufacturer’s protocol. Obtained single cells were allowed to form neurospheres in DMEM/F12 (11330–032, Gibco) supplemented with B27 without Vitamin A (12587–010, Gibco), EGF (E-1257, Sigma), bFGF (F-0291, Sigma), heparin (H-3149, Sigma) and Penicillin/Streptomycin (30–001-CI, Corning). After the second passage, dissociated single-cell suspensions were treated with indicated lentivirus at 15 MOI and seeded in Matrigel-coated well-plates in proliferating media for five days. To differentiate into neurons, the cultures were switched after two days of expansion media to differentiation media DMEM/F12 (11330–032, Gibco) supplemented with Glutamax (35050–061, Gibco), B27 (17504–044, Gibco), N2 (17502–048, Gibco), and Penicillin/Streptomycin (P/S) (30–001-CI, Corning) for three days.

Cortical neuron culture

Primary neuronal cultures were prepared from cerebral cortices from E14–15 embryos from CD-1 IGS wildtype mice (Charles Rivers Laboratories). Cortices were dissociated using Papain dissociation system (Worthington Biochemical Corporation, Lakewood, NJ, USA). Dissociated cells were maintained in PDL-coated 24 well plates for 9–10 days in vitro (DIV) in Neurobasal medium containing 2% B27 supplement (Gibco), 1% P/S (Gibco) and 1% glutamax (Gibco) in a humidified 37 °C incubator with 5% CO2 with 50% media change every third day. Listed lentiviruses were transduced on DIV 4, and cultures were maintained for five days post-viral transduction.Before cell collection, cells were incubated with PrestoBlue HS cell viability reagent 20x diluted in the culture medium (P50200, Invitrogen, Thermo Fisher Scientific). One hour later, 100μL of the culture media was transferred to a black-walled 96-well plate (3340, Corning) and the fluorescence signal was read using FLUOstar Omega (BMG Labtech).

Preparation and treatment with TgCM media

A heterozygous Tg2576 (B6;SJL-Tg(APPSWE)2576Kha, #1349, Taconic Farms Inc.) female mouse was bred with a heterozygous Tg2576 male mouse. The pregnant female was euthanized, and the embryos were extracted at E14.5. Primary neuronal cultures were prepared after the isolation of embryonic cortices as described above. The limb tissue from each embryo was used to verify the genotypes. At 14 DIV, transgenic conditioned media (TgCM) enriched with human Abeta was collected and pooled. Media was centrifuged at 1258 g, 10 min at 4⁰C, and stored as aliquots at −80⁰C. Abeta40 was measured using Aβ1–40 human/rat sandwich ELISA kit (Wako Pure Chemicals Industries, Japan #294–64701), according to the manufacturer’s instructions. The average Abeta concentration was 10nM.

Primary cortical neurons from embryonic CD1 mice were plated in a black-walled 96-well plate and were maintained and transduced with lentivirus as described above. 50% culture media was replaced by TgCM at the endpoint for 2.5h and viability was measured with Prestoblue HS or Live/Dead (L3224, Invitrogen, Thermo Fisher Scientific) assays as per the manufacturer’s protocol. Fluorescence signal was read using FLUOstar Omega.

Cortical glial cultures

Primary cortical glial cells were isolated from neonatal C57BL/6J wild type mice mice (P1). Cortices were dissociated using 0.25% Trypsin (T4549, Sigma-Aldrich) and 1mg/mL DnaseI (10104159001, Roche) for 15 min. at 37C with gentle shaking. Isolated cells were plated on poly-l-lysine (P4707, Sigma-Aldrich) tissue culture plates and maintained in DMEM/HAM F-12, GlutaMAX (10565018, Gibco) supplemented with 10% FBS 2mM GlutaMaxI and P/S (30–001-CI, Corning) in a humidified 37 °C/5% CO2 incubator. To enrich for astrocytes, at the time of passage, other low-adherent glia cells were washed away after strong shaking. For experiments, cells were plated on 24-well plates coated with poly-l-lysine at a density of 100K cells/well. Listed lentivirus were transduced when cultures reached 80% confluency for 5 days. Separate sets of astrocytes received the following AD-related insults for additional 24h: 1.) reactive astrocyte activating cocktail of IL1α (3 ng ml−1, I5396, Sigma-Aldrich), TNF (30 ng ml−1, 8902SF, Cell Signaling Technology) and C1q (400 ng ml−1, MBS14730, MyBioSource)¹⁰⁶; 2.) 20 uM Amyloid-β 42 peptide (AS-21793, Anaspec Inc). The efficiency/effect of the insults was confirmed by an increase of astrocyte reactive marker genes compared with non-treated cells.

EdU pulse label, staining in vitro

After 5 days in proliferative condition, cells were treated with 20 uM EdU (Invitrogen) for 30 minutes. Then, the cells were washed with cold PBS and then fixed with 4% PFA for 20 minutes. Cells were washed 3x for 10 min with PBS. The fixed cells were blocked with 5% BSA in PBS and incubated with primary antibodies overnight at 4⁰C. A secondary fluorophore antibody with DAPI was incubated for 1hr at room temperature followed by 3×10mins with PBS. Following the staining, EdU labelling was performed using Click-it EdU Cell Proliferation Kit (Invitrogen C10340). After mounting, the cells were imaged at 20x magnification in LSM 900 confocal microscope.

RNA preparation and expression analysis

Total RNA was isolated from cells, either culture or fresh isolated, using RLT lysis buffer supplemented with 1% β-mercaptoethanol and the RNeasy micro kit (Qiagen) according to manufacturer’s instructions. For RNA isolations from PFA-fixed HC sections, the RNeasy FFPE kit (Qiagen) was used starting from the Proteinase K digestion step, according to the manufacturer’s instructions. For QPCR analysis, first-strand cDNA was generated using equal amounts of RNA and the Superscript IV first-strand synthesis system (18091050, Invitrogen, Thermo Fisher Scientific). QPCR was performed using Power SYBR Green PCR Master Mix (4367660, Thermo Fisher Scientific) in a QuantStudio5 Real-Time PCR system (Applied Biosystems). Relative quantification of gene expression normalized to Rps18, an average of Rps18 and Tbp, or an average Rsp18 and Slc13a, as appropriate, was determined by the comparative Ct method. Primer sequences are listed in Supplementary Table 1.

Single Nuclei RNA Sequencing and Data Analysis

Single nuclei isolation and sequencing

We used five animals for each group. Animals representative of the group average for the MWM results were sent to single-nuclei RNA-sequencing. In addition, selection criteria for the animals in the running groups included a minimal average running activity of 2.5 km/day. After the mice were perfused with ice-cold PBS, the brains were immediately transferred to ice-cold HBSS, and cut across the midline. The hemibrains were sliced coronally into 400 um slices using a McIlwain Tissue Chopper. Slices were transferred to a petri dish with ice-cold HBSS. Under a microdissection microscope, the slices containing the hippocampus were selected, and the DG was microdissected with the help of a 23G needle. While holding the slice in the intersection of the DG and CA3 with forceps, the needle was used to perform the cut. All DG slices of an animal were collected into a 2 mL tube, briefly spun down to remove HBSS, and snap-frozen in liquid nitrogen¹⁰⁷. The nuclei were isolated from the snap-frozen dentate gyrus tissue following the previously described protocol¹⁰⁸. The 10x Genomics Chromium with v3.1 NextGEM Single Cell 3’ Kit was used to generate the single nuclei barcoded emulsions, targeting 8000 nuclei per device. Barcoded libraries were constructed according to the manufacturer’s protocols, with few modifications, and multiplexed for sequencing on Illumina’s HiSeq2500 or NovaSeqS1 flow cells with a total of 35,000 nuclei per condition at an average depth of 50,000 reads per nucleus. All data processing was performed in the Terra Platform or on a local machine with the same computational environment settings. Scanpy v1.8.2 in Python v.3.8.12 was used to postprocess the data¹⁰⁹.

Preprocessing, quality control, and normalization

All data processing was performed in the Terra Platform or on a local machine with the same computational environment settings unless otherwise noted. After sequencing, the FASTQ files were generated using the CellRanger v6.0.0 on Terra.bio¹¹⁰. The reads were further trimmed using Cutadapt v2.8¹¹¹ with default parameters to trim off the homopolymers (A30, T30, G30, and C30), the templated switch oligo sequence (CCCATGTACTCTGCGTTGATACCACTGCTT), and its complementary sequence (AAGCAGTGGTATCAACGCAGA GTACATGGG). The trimmed reads were aligned to the GRCm38 mus musculus reference genome. Lastly, CellRanger Count was used to generate the final count matrices.

To check read quality, the number of mapped reads per nucleus, the percentage of mapped mitochondrial reads, and the decay curve of the unique molecular identifier (UMI) were inspected, and no abnormal samples were found or removed at this point. CellBender v2.0 was used to remove ambient RNA byproducts of nuclear isolation and resulted in a total of 148,757 nuclei at this stage.

Further, Scrinvex v13 was used to calculate the proportion of reads mapped to exonic regions to the total mapped reads for each droplet. We removed the droplets with an exon ratio greater than the 75^th percentile and IQR range for downstream analysis. Scrublet was used to identify doublets, and the threshold for doublets was manually adjusted for each sample based on the distribution of predicted doublet scores. The doublets were removed. Moreover, nuclei with fewer than 200 genes mapped and more than 5% mitochondrial reads were filtered out. These steps resulted in a total of 113,625 nuclei for further analysis. The reads were normalized and logarithmized, highly variable genes were identified, and the data were scaled.

Final UMAP clustering

We calculated the principal components from the highly variable genes and used the first 50 PCs to calculate the near neighbors and generate the final Uniform Manifold Approximation and Projection (UMAP) graph. The cluster assignment was computed using the Leiden clustering with a resolution of 0.225¹¹². Note that nuclei in two clusters stemmed from only one mouse and have high expression of IEGs (Fos and Npas4). Therefore, all the nuclei in those clusters were removed from further analysis, resulting in a total of 108,613 nuclei.

Marker genes and cluster identification

The marker genes for each identified cluster were calculated using a Wilcoxon rank-sum test compared to all other clusters. The log fold change, percentage of nuclei expressing each gene, and the area under the receiver operator curve scores were calculated for all genes with functions from Scanpy v1.8.2 and scikit-learn v1.0.2. Genes with an AUC score greater than 0.7 and log2FC > 0.6 were considered marker genes for a cluster. We referred the marker genes to Hochgerner et al. 2018 paper and Saunders et al., 2018 to assign our cluster’s identity^12,24. For the subcluster marker genes, we applied the same method and threshold, except that the AUC threshold was set to 0.6. The heatmap visualizations used pheatmap v.1.0.12 library in R.

Compositional Analyses

The absolute number of nuclei was used to calculate the percentage of each cell types and subcluster. Statistical analysis was performed by Two-way ANOVA.

Differential Gene Expression Testing

We used the Wilcoxon rank-sum testing to identify the differentially expressed (DE) genes in cluster and subcluster using four different comparisons: (1) WT Sed vs. WT Run, (2) WT Sed vs. APP/PS1 Sed, (3) WT Run vs. APP/PS1 Run, and (4) APP/PS1 Sed vs. APP/PS1 Run. Briefly, the gene matrix was filtered for genes with at least 3% of cells containing non-zero expression in at least one condition. Differentially expressed genes were calculated between pairs of conditions using a Wilcoxon rank-sum test (rank.genes.groups function in Scanpy). To account for multiple testing, adjusted P-values were calculated using Benjamini-Hochberg FDR adjustment.

To identify genes that were dysregulated in AD and recovered with exercise, recDEGs, we first selected the genes dysregulated with AD (WSvsAS FDR-adjusted p-value <0.1) and tested again this subset list considering the gene recovered if ASvsAR FDR-adjusted p-value < 0.05. The plotted recovery score is equal to the absolute value of the ASvsAR logFold-Change. For visualization purposes, we separated lowly expressed genes, typically being absent in one condition but induced in the other (expressed >3% of cells in one group, but <10% of all cells), and further named “Rare” recDEGs.

Subclustering

We ran subclustering analyses on all cell types. Our subcluster identification approach was similar to that used to determine the global clusters. We extracted one cell type at a time, recalculated the highly variable genes (with a mean range from 0.0125 to 3), and ran PCA with the first 50 PCs. Next, we reset the local neighborhood size with 10 near neighbors to recalculate the final UMAP in the subclusters. After regenerating the UMAP for the subclusters, Leiden clustering was used to cluster the nuclei at a resolution of 0.225. The UMAP visualizations used Scanpy built-in functions. We compared our marker genes in the microglia subclusters or astrocyte subclusters with the published disease-associated microglia (DAM) and activated response microglia (ARM) or disease-associated astrocyte (DAA) genes with adjusted p-values below 0.05 and logFC > |0.6|^17,49,113.

Pathway Enrichment Analysis

GSEA pre-ranked analysis was performed for each cell type to identify the cellular pathways and biological processes associated with AD and exercise¹¹⁴. All analyses used the GSEA v4.3.3 (Broad Institute), following a previously published protocol¹¹⁴. Briefly, using the DEGs output we ranked genes from each cell type by the −log10 of the adj p-value transforming the metric to a negative value for the genes that were downregulated across conditions. These pre-ranked gene lists are the input for GSEA. Pathway enrichment was calculated based on the m5.go.mf/bp.v2022 and m2.cp.reactome.v2022 databases. We used 15 and 500 as the min and max size, respectively, of the gene set for a given term, and “classics” enrichment statistic. A pathway was considered enriched with a cell type with an FDR threshold of 0.25.

Astrocyte subcluster 1 gene set enrichment analysis was performed using the list of subcluster marker genes with AUC>0.6 using the Enrichr¹¹⁵ tool libraries Reactome 2022, KEGG 2021 Human, GO Biological Process and Molecular Function 2023 (accessed on 09 March 2024). Only enrichment terms with adjusted p-value <0.05 were considered.

Cell-Cell Communication Analysis

Communication between cell types was determined using CellChat (v1.6.0) with default parameters²⁸. Briefly, normalized counts for each group were used with cell labels, and compared to the default mouse ligand-receptor interaction database to look for overexpressed ligand-receptor pairs within and between cell types. Cells expressing the ligand are designated as “sender cells”. Cells expressing the corresponding receptor are designated as “receiver cells”. The genes were filtered to remove any genes with reads in fewer than 10 cells. Significant communication pathways (p-value<0.05) were displayed using standard CellChat plotting functions.

Regulator Genes

GeneWalk v1.5.3 was applied to identify regulator genes that drive gene networks within a biological context¹¹⁶. Briefly, GeneWalk builds biologically relevant networks from the provided gene lists by connecting genes and relevant GO terms and comparing the network to randomized networks. The cosine similarity value is calculated between each gene and GO term and compared to cosine similarity values of the randomized networks to determine significance. P-values are then adjusted using Benjamini-Hochberg multiple testing correction. Regulator genes are defined as those which contain a greater than 50% of connected GO terms identified as significant. This analysis employed default parameters. This analysis was performed using Harvard Medical School O2 high performance cluster.

Data integration with sc-RNAseq data from Buckley et al. (2023)

We assessed whether the transcriptional responses we found in oligodendrocytes and OPCs in the dentate gyrus following exercise are comparable to those reported in other neurogenic niches in the adult brain. We compared our findings in WT mice to those in WT mice of comparable age in Buckley et al. (2023) who reported transcriptional responses to exercise in oligodendrocytes and OPCs in the subventricular zone. Genes common to both projects, and those that we reported differentially expressed in exercise vs. sedentary conditions with an FDR-adjusted p value < 0.05 were selected. Two-log-fold changes in normalized, log-transformed gene counts in exercise vs. sedentary conditions were plotted across projects to assess convergence in directionality of effects.

Comparison with human data

For this comparison, we used the human snRNA-seq data from Brase et al., 2023, which includes parietal cortex samples from 31 sporadic late-onset AD (hAD), 16 autosomal dominant AD (APP or PSEN1, hADAD), and nine healthy age-matched controls (hCtrl). Their demographics are described in the Supplementary Table 1. First, we mapped our AD-dysregulated genes to the corresponding human gene. Then, we plotted the logFC for hAD (hCtrl vs. hAD), hADAD (hCtrl vs. hADAD), mAPP/PS1 (WSvsAS), and mAPP/PS1-Run (ASvsAR) for genes that were significant in either hAD and/or hADAD. The full list of the comparisons can be found in Supplementary Data 10. Differences in CDH4 counts between astrocytes subclusters were calculated using a linear mixed effect model with sex and cluster as covariates, and samples as a random effect using the tool nebula. We have used the subclusters as originally defined by Brase et al. 2023.

Gene signature analysis

We performed specific gene signature analysis as reported by Buckley et al. 2023²⁰, using either snRNA-seq data or QPCR data from isolated cells. For the snRNA-seq data, we summed all normalized expression values of genes in each gene set by cell within the indicated cell subcluster. For the microglia subcluster 1, we analyzed the DAM signature using five well-known marker genes (Irf8, Trem2, Igf1, Axl, and Csf1), and the Homeostatic signature (P2ry12, Cd33, Tmem119, Csf1r, and Cx3cr1). For the astrocyte subcluster 1 signature we used seven subcluster 1 marker genes 1 (Mfge8, Plxna2, Grin2b, Bmper, Dab1, Pde1c, and Cdh4). The differences between the two groups, sedentary versus runners, were tested using Mann Whitney test (data non-normally distributed). Microglia or astrocytes isolated from 5xFAD were processed for QPCR, and the normalized expression values of the genes in each gene set described above were summed by animal. The differences between the two groups, sedentary versus runners, were tested using two-tailed unpaired t-test (data normally distributed). For the gene signature analysis by QPCR, only animals that had enough RNA to assess the expression of all marker genes defined in the respective gene signature could be included in the analysis.

Statistical analyses.

The complex bioinformatic analyses of the snRNA-seq data are described above in detail. Statistical analysis of the remaining data was performed using GraphPad Prism 9.1 (216). Significance was assigned to differences with a p-value less than 0.05 unless stated otherwise. Data are expressed as mean ± s.e.m. unless stated otherwise. Data distribution was assumed to be normal but not formally tested for most data, except for Fig.5k, 5l, 6a, 6g, and 6h, and EDFig. 5g, 5h, 6i, 6j, and 8b where normality was formally tested (D’Agostino-Pearson and Shapiro-Wilk tests). Two-tailed Student’s t-tests were used to compare two groups, or Mann Whitney (when no normal distribution). For comparisons among multiple groups, either one-way (or Kruskal-Wallis test when no normal distribution), two-way or repeated-measures ANOVA followed by the indicated post-hoc test was used. To be able to perform repeated-measures ANOVA, missing values were substituted using a mean substitution. No statistical methods were used to pre-determine sample sizes. Sample size was based on previously published studies using similar methodologies^5,91,117. Outliers were identified using the Grubbs’ test with a significance level of p<0.05 (two-sided). For behavioral tests, exclusion criteria were pre-established following published guidance^96,118,119. In the SAB, mice with fewer than eight arm entries during the 5-min trial were excluded from the analysis as significantly lower exploratory activity biases the analysis. In the CFC, mice that did not freeze after receiving the shocks on day 1 were excluded from that analysis as we cannot use freezing as a proxy for learning or memory. For all experiments, all stated replicates are biological replicates. Operators were blinded to the true experimental groups during data collection and image analysis by de-identifying all samples with generic unique IDs. Assignment to treatment groups was done randomly.

Extended Data

Extended Data Fig. 1. Neurogenic niche response to exercise and AD at the single-nuclei level.

a, MWM latency to reach the escape platform in acquisition (Three-way ANOVA, Exercise n.s. P = 0.0964, Genotype ****P <0.0001, Exercise × genotype n.s. P = 0.2664, Exercise × genotype × time *P = 0.0263). b, Acquisition 24h probe trial (Two-way ANOVA, Exercise **P = 0.0074, Genotype n.s. P = 0.2919, Exercise × genotype n.s. P = 0.8698), and c, 24h probe trial in reversal (Two-way ANOVA, Exercise **P = 0.0067, Genotype n.s. P = 0.3812, Exercise × genotype n.s. P = 0.904). d, Daily running activity (Two-way repeated measures ANOVA, Time ****P < 0.0001, Genotype n.s. P = 0.2706, Time × genotype n.s. P = 0.3457). e, Open field (OPF) (Two-way ANOVA, Exercise n.s. P = 0.8558, Genotype n.s. P = 0.2011, Exercise × genotype n.s. P = 0.6738), f, Spontaneous alternation behavior (SAB) (Two-way ANOVA, Exercise n.s. P = 0.1004, Genotype n.s. P = 0.1187, Exercise × genotype n.s. P = 0.2082), and g, Contextual fear conditioning (CFC) test in all mice (Two-way repeated measures ANOVA, Context ****P < 0.0001, Group n.s. P = 0.438, Context × group n.s. P = 0.4465, followed by Sidak’s multiple comparisons WT-Sed AvsB ***P = 0.0002, WT-Run AvsB ***P = 0.0004, APP/PS1-Sed AvsB n.s. P = 0.0519, APP/PS1-Run AvsB n.s. P = 0.1497). h, Number of cells per cell cluster within each group. i, PCA analysis of pseudobulk data from all samples. j, UMAP representation of marker genes expression in different clusters. Color represents expression level according to the scale bar on the right. k, Percentage of cells per cell cluster within each group. l, The scatter plot shows regulator genes based on GeneWalk analysis observed in WSvsAS. Each dot represents a regulator gene, and the color represents the cell cluster. For all behavior experiments WT-Sed n = 12, WT-Run n = 12, APP/PS1-Sed n = 9, APP/PS1-Run n = 9 (a-g), for snRNAseq WT-Sed, WT-Run, APP/PS1-Run n = 5, APP/PS1-Sed n = 4 (h and k). Data represent the mean ± s.e.m. of biologically independent samples.

Extended Data Fig. 2. Representative enriched pathways across different cell types.

GSEA pre-ranked on gene ontology results for all cell types between the ASvsAR and WSvsAS in neuronal cells (a) and glia cells (b). The representative enriched pathways were selected based on FDR < 0.25.

Extended Data Fig. 3. Remodeling of adult hippocampal neurogenesis in exercise and AD.

a, Quantification of BrdU+NeuN+ adult-born neurons in the dorsal and ventral DG and representative higher magnification confocal images with anti-BrdU (green) and NeuN (red) from WT and APP/PS1, sedentary or running, mice. Scale bar, 50μm. n = 6 per group. Two-way ANOVA, Dorsal DG: Exercise ****P < 0.0001, Genotype *P = 0.0269, Exercise × genotype n.s. P = 0.058, Ventral DG: Exercise *P < 0.0101, Genotype n.s. P = 0.1629, Exercise × genotype n.s. P = 0.305. b, Quantification of DCX+ cells in the dorsal and ventral DG from WT and APP/PS1, sedentary or running, mice. n = 6 per group. Two-way ANOVA, Dorsal DG: Exercise n.s. P = 0.0914, Genotype **P = 0.0013, Exercise × genotype n.s. P = 0.1552, Ventral DG: Exercise **P = 0.0069, Genotype ****P < 0.0001, Exercise × genotype n.s. P = 0.1426. c, Heatmap shows the normalized mean expression (z-score) of neurogenesis-related genes reported by Hochgerner et al. in our dataset. d, UMAP representations of early neuronal marker genes expression. Color represents expression level according to the scale bar on the right. e, f, Scatter plots showing the correlation between AD and exercise effects in Neuroblast I (e) and II (f). Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. g, Dot plots showing Immature Neurons’ recDEGs for all neurogenic cell types. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. Data represent the mean ± s.e.m. (a and b) of biologically independent samples.

Extended Data Fig. 4. Immature Neurons recDEG Atpif1 knock-down disrupts neuronal proliferation and differentiation in vitro.

a-d, Primary cortical neurons were transduced with LV-shRNA for five days. PrestoBlue HS normalized cell viability (a, one-way ANOVA followed by Dunnett’s against shCtrl: shSlc25a4 n.s. P = 0.9967, shAtp6v0c n.s. P = 0.0573, shAtpif1 *P = 0.013), confirmation of the gene knock-down (KD) by qPCR (b, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl ****P < 0.0001), and gene expression of neuronal markers in response to Atp6v0c (c, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl: Neurod1 ****P < 0.0001, Dcx **P = 0.0017, Tubb3 n.s. P = 0.8694, Map2 ***P = 0.0003, Dlg4 **P = 0.0071, Syn1 **P = 0.0037, Bax *P = 0.028) and Atpif1 (d, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl: Neurod1, Dcx, Map2, Dlg4, and Syn1 ****P < 0.0001, Tubb3 **P = 0.0046, Bax **P = 0.0022) knock-down. shAtp6v0c and shSlc25a4 n = 6, shAtpif1 n = 4 (a), shAtp6v0c and shAtpif1 n = 6, shSlc25a4 n = 5 (b-d). e, Representative confocal images of EdU (red) and Nestin (green) staining of embryonic neural stem and progenitor cells transduced with LV-shRNA and maintained in proliferating media for 5 days. Scale bar, 50μm. f-h, Neurospheres were transduced with LV-shRNA and maintained in differentiation media for five days. Confirmation of the gene knock-down by qPCR (f, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl ****P < 0.0001), and gene expression of neuronal markers in response to Atp6v0c (g, two-way ANOVA, KD n.s. P = 0.0877, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl: Neurod1 and Tubb3 ****P < 0.0001, Dcx n.s. P = 0.3704, Map2 n.s. P = 0.0826, Dlg4 n.s. P = 0.1893, Syn1 n.s. P = 0.1901) and Atpif1 (h, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl: Neurod1, Dcx, Tubb3, and Map2 ****P <0.0001, Dlg4 **P = 0.0015, Syn1***P = 0.0004) knock-down. n = 11 per group. i-l, Primary cortical neurons were transduced with LV-shRNA for five days and treated with 20μM recombinant Amyloid-beta 42 for the last 16h (i and j), or Abeta-enriched Tg2576 conditioned-media for the last 3h (k and l). Normalized calcein fluorescent signal indicative of live cells after 16h Amyloid-beta 42 treatment (i, Welch’s ANOVA followed by Dunnett’s T3 against shCtrl ****P < 0.0001), normalized EthD1 fluorescent signal indicative of dead cells after 16h Amyloid-beta 42 (j, Welch’s ANOVA followed by Dunnett’s T3 against shCtrl, shAtp6v0c **P = 0.0055, shAtpif1 ***P = 0.0002), normalized calcein fluorescent signal after 3h Tg2576 conditioned-media (k, Welch’s ANOVA followed by Dunnett’s T3 against shCtrl ****P < 0.0001), and normalized EthD1 fluorescent signal after 3h Tg2576 conditioned-media (l, one-way ANOVA followed by Dunnett’s against shCtrl, shAtp6v0c n.s. P = 0.0726, shAtpif1 n.s. P = 0.0754). n=6 per group. m-o, Adult hippocampus derived neurospheres were transduced with LV-shRNA and maintained in differentiation media for three days. Confirmation of the gene knock-down by qPCR (m, two-tailed unpaired t-test ****P < 0.0001), PrestoBlue HS normalized cell viability (n, two-tailed unpaired t-test **P = 0.0065), and gene expression of neuronal markers in response to Atpif1 knock-down (o, two-way ANOVA, KD ****P < 0.0001, KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl: Neurod1, Dcx, and Map2 ****P < 0.0001, Tubb3 n.s. P = 0.681, Dlg4 *P = 0.0152, Syn1 n.s. P = 0.7519, Bax n.s. P = 0.9609). n=4 (n) and 12 per group (m, o). Data represent the mean ± s.e.m. of biologically independent samples.

Extended Data Fig. 5. Exercise regulates DAM-like microglia in AD mouse models.

a, b, Quantification of Iba-1+ microglia per mm² in the dorsal DG (a, two-way ANOVA, Exercise **P = 0.0021, Genotype ****P < 0.0001, Exercise × genotype **P = 0.0031, followed by Fisher’s LSD, WT-Sed vs Run ns P = 0.8979, APP/PS1 Sed vs Run ***P = 0.0001) and ventral DG (b, two-way ANOVA, Exercise *P = 0.0495, Genotype ****P <0.0001, Exercise × genotype n.s. P = 0.1564) (WT-Sed n = 6, WT-Run n = 6, APP/PS1-Sed n = 5, APP/PS1-Run n = 6). c, UMAP representation of marker genes for perivascular macrophages (Mrc1, Cd163, Cd74), monocytes (S100a4), B cells (Cd79b, Rag1), T cells (Trbc2, Cd3g), and natural killer cells (Nkg7). Color represents expression level according to the scale bar on the right. d, Cell composition in percentage for each subcluster shown in Figure 4g–i. Two-way ANOVA, Exercise n.s. P = 0.9956, Genotype ****P < 0.0001, Exercise × genotype n.s. P = 0.8881. e, Heatmap shows the normalized mean expression (z-score) group of IFN, MHC, and Cyc-M microglia genes reported by Chen et al. in our dataset. f, Dot plots showing microglia subcluster 1 markers. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. g, Violin plots of gene signatures for DAM (Irf8, Trem2, Igf1, Axl, and Csf1) and Homeostatic (P2ry12, Cd33, Tmem119, Csf1r, Cx3cr1) microglia using snRNAseq data of microglia subcluster 1 from AD mice DG. Gene signature = sum of normalized gene expression for all genes of the gene signature per cell (n = 150 and 229 cells for ‘Sed’ and ‘Run’, respectively). Two-tailed Mann Whitney, DAM P = 0.0195 and Homeostatic P = 0.9858. h, Bar plots of gene signatures for DAM (Irf8, Trem2, Igf1, Axl, and Csf1) and Homeostatic microglia (P2ry12, Cd33, Tmem119, Csf1r, Cx3cr1) in isolated CD11b+ cells (microglia) from the cortex and hippocampus of the 5xFAD mouse model using QPCR. Gene signature = sum of normalized gene expression for all genes of the gene signature per animal (n=8 and 7 for ‘Sed’ and ‘Run’, respectively). Two-tailed unpaired t-test, DAM P = 0.0616 and Homeostatic P = 0.9462. Data represented by the mean ± s.e.m. (a, b, d, h) or by the median (middle bold line) and upper and lower quartiles (lighter dotted lines) (g) of biologically independent samples.

Extended Data Fig. 6. Exercise shifts the transcriptional state of astrocytes.

a, b, Quantification of GFAP+ astrocytes per mm² in the dorsal DG (a, two-way ANOVA, Exercise n.s. P = 0.9007, Genotype n.s. P = 0.5594, Exercise × genotype n.s. P = 0.3558) and ventral DG (b, two-way ANOVA, Exercise n.s. P = 0.6633, Genotype n.s. P = 0.9575, Exercise × genotype n.s. P = 0.9111) (WT-Sed n = 6, WT-Run n = 6, APP/PS1-Sed n = 5, APP/PS1-Run n = 6). c, Heatmap shows the normalized mean expression (z-score) of the marker genes for each astrocyte subcluster (subcluster 0 in blue and subcluster 1 in orange). d, UMAP representation of the expression of the high-confidence markers for Radial Glia-like cells in astrocytes. Yellow dots are all nuclei in the astrocyte cluster; grey and green dots represent potentially Radial Glia-like cells based on the expression of listed markers. e, Heatmap shows the normalized mean expression (z-score) per group of previously identified disease-associated astrocytes (DAA) markers in our dataset (subcluster 0 in blue and subcluster 1 in orange). f, Heatmap shows the normalized mean expression (z-score) per group of previously identified reactive astrocytes markers in our dataset (subcluster 0 in blue and subcluster 1 in orange). g, Bar chart of the relevant enriched terms for the astrocyte subcluster 1 marker genes from Enrichr. Enriched terms displayed presented an adjusted p-value <0.05 determined by Fisher exact test with the Benjamini-Hochberg correction for multiple hypotheses. h, CDH4 counts in astrocytes subclusters from human parietal cortex snRNA-seq in Brase et al., 2023⁴⁸. The subclusters presented are the originally described ones from⁴⁸. Linear mixed effect model (covariates, sex and cluster; random effect, sample). i, Violin plots of gene signatures in the astrocyte subcluster 1 (Mfge8, Plxna2, Grin2b, Bmper, Dab1, Pde1c, and Cdh4) using snRNAseq data of astrocytes subcluster 1 from AD mice DG. Gene signature = sum of normalized gene expression for all genes of the gene signature per cell (n = 36 and 60 cells for ‘Sed’ and ‘Run’, respectively). Two-tailed Mann Whitney P = 0.0028. j, Bar plots of gene signatures for astrocyte subcluster 1 (Mfge8, Plxna2, Grin2b, Bmper, Dab1, Pde1c, and Cdh4) in isolated ACSA2+ cells (astrocytes) from the cortex and hippocampus of the 5xFAD mouse model using qPCR. Gene signature = sum of normalized gene expression for all genes of the gene signature per animal (n=7 for ‘Sed’ and ‘Run’). Two-tailed unpaired t-test P = 0.0240. Data represented by the mean ± s.e.m. (a, b, j) or by the median (middle bold line) and upper and lower quartiles (lighter dotted lines) (i) of biologically independent samples.

Extended Data Fig. 7. Astrocytes recDEGs knock-down alters astrocytes states in vitro.

Primary cortical mixed glia cultures were transduced with LV-shRNA for five days in standard growth media (a-d) or treated in the last 24h with a reactive astrocyte activating cocktail of cytokines (e-h) or Amyloid-β 42 peptide (i-l). Confirmation of the gene knock-down by QPCR (a, e, i), and gene expression of reactive astrocyte markers and markers of our subcluster 0 and 1 in response to Nme7 (b, f, j), St7 (c, g, k), and Thra (d, h, l), knock-down. n = 3 for shCtrl in e-l, n = 4 for all other groups. Two-way ANOVA followed by Fisher’s LSD. *p<0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, n.s., not significant. Data represent the mean ± s.e.m. of biologically independent samples. a: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl ****P < 0.0001; b: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, Cdh20, Rorb, and Cdh4 ****P < .0001, C3 n.s. P = 0.3171, Serpina3n ***P = 0.0009, Hspb1 **P = 0.004, Csmd1 **P = 0.0068; c: two-way ANOVA, KD n.s. P = 0.2011, KD × gene ****P <0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, C3, Serpina3n, Cdh20, and Cdh4 ****P < 0.0001, Hspb1 *P = 0.0277, Rorb n.s. P = 0.1921, Csmd1 n.s. P = 0.9147; d: two-way ANOVA, KD and KD × gene ****P < .0001, followed by Fisher’s LSD compared to shCtrl Gfap, Serpina3n, Rorb, Cdh4, and Csmd1****P < 0.0001, C3 ***P = 0.0001, Hspb1 n.s. P = 0.283, Cdh20 *P = 0.0174; e: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl ****P < .0001; f: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, Serpina3n, Cdh20, and Rorb ****P < 0.0001, C3 n.s. P = 0.7271, Hspb1 n.s. P = 0.107, Cdh4 n.s. P = 0.0856, Csmd1 n.s. P = 0.1875; g: two-way ANOVA, KD and KD × gene ****P < .0001, followed by Fisher’s LSD compared to shCtrl Gfap **P = 0.0092, C3, Cdh20, Cdh4, and Csmd1 ****P < .000, Serpina3n **P = 0.0012, Hspb1 ***P = 0.001, Rorb n.s. P = 0.159; h: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, Serpina3n, Cdh20, Rorb, and Cdh4****P < 0.0001, C3 n.s. P = 0.8352, Hspb1 n.s. P = 0.9755, Csmd1 n.s. P = 0.1847; i: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl ****P < 0.0001; j: two-way ANOVA, KD and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap ***P = 0.0002, C3 n.s. P = 0.2851, Serpina3n **P = 0.003, Hspb1n.s. P = 0.5116, Cdh20, Rorb, and Cdh4 ****P < 0.0001, Csmd1 **P = 0.0093; k: two-way ANOVA, KD n.s. P = 0.7698 and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, ***P = 0.0001, C3, Cdh20, and Cdh4 ****P < 0.0001, Serpina3n n.s. P = 0.137, Hspb1 n.s. P = 0.0896, Rorb n.s. P = 0.6144, Csmd1 n.s. P = 0.6805. l: two-way ANOVA, KD n.s. P = 0.4382 and KD × gene ****P < 0.0001, followed by Fisher’s LSD compared to shCtrl Gfap, Serpina3n, Hspb1, Cdh20, Rorb, and Cdh4 ****P < 0.0001, C3 n.s. P = 0.9321, Csmd1 * P = 0.0236.

Extended Data Fig. 8. Exercise remodels AD-dysregulated pathways in mGCs.

a, Average size of amyloid-beta plaques (3D6 staining) in ventral DG sections (APP/PS1-Sed n = 5, APP/PS1-Run n = 6, three section per animal; Two-tailed unpaired t-test P = 0.6679). Data is represented by the median (middle bold line) and upper and lower quartiles (lighter dotted lines). b, Schematic representation of Amyloid precursor protein (APP) processing and amyloid beta degradation pathways. Adapted from the KEGG pathway database. Created with BioRender.com. c, Expression of chimeric mouse/human APPswe and the human PS1-dE9 transgene characteristic of the APP/PS1 mice, and the fraction of cells expressing the gene. d, Expression of alpha-secretase (Adam10), beta-secretase (Bace1), and gamma-secretase (Psenen, Ncstn, Aph1a) coding- genes in all neuronal clusters and the fraction of cells expressing the gene. e, Expression of the Aβ-degrading enzyme coding-gene Ide and Mme in all cell clusters and the fraction of cells expressing the gene. f, Immediate early gene expression in the different neuronal cell types by group. Data represented by biologically independent samples.

Extended Data Fig. 9. Exercise and AD responses in interneurons and vascular cells.

a, Scd2 average expression of all groups in Oligodendrocyte progenitor cells (OPCs) and oligodendrocytes. b, Scd2 expression in oligodendrocytes in different groups. c, Cell compositional analysis for the OPCs subclusters. Two-way ANOVA, Exercise n.s. P = 0.1966, Genotype n.s. P = 0.3081, Exercise × genotype n.s. P = 0.2483. d, Comparison of our exercise effects in DG oligodendrocytes and OPCs with reported exercise effects in the subventricular zone in Buckley et al., 2023. Genes common to both projects, and those we found differentially expressed in exercise vs. sedentary conditions with an FDR-adjusted p value < 0.05 are displayed. Genes with > 0.5 Log2FC change in both projects are labeled. e-g, Interneuron subcluster UMAP representation (e), mean proportion (f), and marker genes for each subcluster (g). h, Scatter plot showing the correlation between AD and exercise effects in Interneurons. Each dot represents a statistically significant DEG in AD (WSvsAS). Dots with black borders represent statistically significant DEGs with exercise in AD mice (ASvsAR). The color gradient illustrates the recovery score (|logFC ASvsAR|). The dot size represents the fraction of non-zero count nuclei in the AR group. i, Dot plots showing recDEGs in Interneurons. In each, the hue and size of the dot represent the mean expression and fraction of non-zero count nuclei, respectively. j, Cell compositional analysis for the Vascular cell subclusters. Two-way ANOVA, subcluster 0: Exercise *P = 0.0266, Genotype n.s. P = 0.9986, Exercise × genotype n.s. P = 0.2262, subcluster 1: Exercise *P = 0.0113, Genotype n.s. P = 0.378, Exercise × genotype n.s. P = 0.5948, subcluster 2: Exercise n.s. P = 0.9024, Genotype n.s. P = 0.492, Exercise × genotype n.s. P = 0.4417.k, Sema3c was a significant recDEGs shared by different cell types. l, Body weights at the start and end of the experiment. Two-way repeated measures ANOVA, Group n.s. P = 0.0603, Time *P = 0.0232, Group × time **P = 0.0018. WT-Sed n = 5, WT-Run n = 5, APP/PS1-Sed n = 4, APP/PS1-Run n = 5 (c and j). WT-Sed n = 12, WT-Run n = 12, APP/PS1-Sed n = 9, APP/PS1-Run n = 9 (l). Data represent the mean ± s.e.m. of biologically independent samples (c, j, l).

Data Availability Statement

snRNA-seq data are available at the Sequence Read Archive (SRA) with SubmissionID: SUB13470069 and BioProject ID: PRJNA976296 and the Broad Institute Single Cell Portal with the Accession number: SCP2229, URL: https://singlecell.broadinstitute.org/single_cell/reviewer_access/8e407278-1324-4cc4-99f2-1948003d3e8d, and PIN: 3JBO34XYYU. The human single nucleus data from the Knight ADRC accessed in this study are found in the National Institute on Aging Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS) with accession number NG00108 [https://www.niagads.org/datasets/ng00108]. All snRNA-seq analysis results are provided as Supplementary Data in the Additional Supplementary Files. For additional data, all numerical source data and statistical details are provided as Source Data Files.

Code Availability Statement

The relevant code is part of Additional Supplementary Files (Supplementary Code 1).