Cerebellar post-mortem tissue samples from 15 donors with a clinical diagnosis of A-T and 56 control donors were used in these analyses. The A-T cohort comprised five female (8–31 years old) and ten male (8–32 years old) donors with neuropathological findings consistent with the disease. Control donors (27 males aged 8–93 years and 29 females aged 17–96) were selected for having no known CNS disease and having died from a non-CNS related etiology. Post-mortem intervals ranged from 2 to 21 h for A-T donors and 5–79 h for controls donors. Donor information is summarized in Supplementary Table 1.

To better understand the cell type specific molecular changes that occur in A-T, we extended the NETSseq cell type purification method from Xu et al. (2018) to isolate additional cell types from post-mortem human cerebellar tissue (Figure 1A). Using a combination of cell type specific antibodies, we developed FANS gating strategies to isolate specific populations of nuclei and determine the relative abundance of each cell type (Figure 1B and Supplementary Table 2). Using a refined antibody panel, we were able to isolate nuclei from an additional three key cerebellar cell types – microglia, Golgi and Purkinje neurons, in addition to the five cell types isolated in the previous publication (granule, basket, astrocyte, oligodendrocyte and OPCs). To isolate the four neuronal cell types, we used antibodies against the splicing factor NeuN/RBFOX3, the ER resident membrane protein Inositol 1,4,5-Triphosphate Receptor Type 1 (ITPR1), and the transcription factor Forkhead Box K2 (FOXK2). Combining ITPR1 and FOXK2 labeling into a single channel and plotting this against NeuN in a different channel, enabled efficient detection of pure populations of nuclei from granule and basket neurons, along with those from exceedingly rare Golgi (0.1%) and Purkinje (0.01%) neurons. The complete sorting strategy for Purkinje neurons is summarized in Supplementary Figure 1D. We also isolated oligodendrocytes and OPCs based on differential positive labeling of oligodendrocyte transcription factor 2 (OLIG2), and astrocytes based on positive labeling of Glutamate Aspartate Transporter (GLAST/EAAT1/SLC1A3). Finally, we confirmed that labeling against interferon regulatory factor 8 (IRF8), which we previously used to isolate cortical microglia (Ossola et al., 2023), is also effective for isolating cerebellar microglia.

RNA-seq was used to generate nuclear gene expression profiles for each population of nuclei separated by FANS. The number of nuclei collected per donor varied by cell type from a mean of 1,096 for Purkinje neurons to 90,720 for granule neurons. From each sorted nuclei population, we subsequently generated deep RNA-seq libraries (median reads >16 million after adapter trimming, Supplementary Table 1). To validate the identity and purity of nuclei obtained from each sorting strategy, we evaluated the expression of known markers for each cell type (Supplementary Figure 1A) and performed a deconvolution analysis (Figure 1C). This involved assembly of a reference signature for each cell type; then, for each sample, we estimate the fraction of similarity against all the cell types in the assembled reference. Consistent with previous findings (Xu et al., 2018) each cell type exhibited a specific expression pattern of marker genes, indicating that negligible contamination was present between cell types.

Next, we used two unbiased methods, hierarchical clustering (HC, Supplementary Figure 1B) and principal component analysis (PCA, Figure 1D), to visualize the similarities and differences across samples. With both methods, we observed a high level of similarity for samples from the same cell type and clear separation of samples from different cell types. Using these methods, we can also observe similarities between basket and Golgi cells, which are two subtypes of GABAergic interneurons. Cell type specific expression of key cell type specific marker genes was confirmed using in situ hybridization (ISH) analysis in human donor tissue (Supplementary Figure 1C). These results indicate that NETSseq can purify and profile distinct cell types with high accuracy and low levels of cross contamination.

Single-nuclei analysis provides superior resolution and the capacity to delineate cellular heterogeneity compared to bulk analysis; however, it faces limitations in detecting genes with low expression profiles. To assess the detection thresholds and sensitivity of our NETSseq platform against single-cell technologies, we performed a comparative analysis, directly comparing our findings with those of Lai et al. (2024), which generated a snRNA-seq dataset from the cerebellum of control and A-T donors. First, we identified cell type specific genes in both datasets and categorized them based on expression quantiles (low: <33%, medium: 33%–66%, high: >66%). We generated expression distribution plots (Figures 1E–G, Supplementary Figure 3A, and Supplementary Table 3) and heatmaps (Supplementary Figures 2A–D) for these genes across all examined cell types in both datasets. In all snRNA-seq expression tertiles, computed using the specific genes generated by NETSseq, we noted the presence of granule neuron markers in the single cell data for the non-granule cell types. This may be due to the presence of ambient RNA (Caglayan et al., 2022; Gautier et al., 2023) originating from the high number of granule neurons, which comprise ∼70%–90% of total cells. This contamination was also found when the expression of snRNA-seq-derived specific genes were plotted across all cell types in the snRNA-seq study (Supplementary Figure 3A). Moreover, for the lower tertiles, the distribution of gene expression for the specific genes in each cell type in the single-cell cohort is significantly skewed toward zero, indicating limited sensitivity in detecting these genes (Figure 1G). In contrast, when we examined the specific genes identified in each expression tertile in the single-cell cohort, we observed no loss of signal in the NETSseq dataset (Supplementary Figure 2C); this suggested that NETSseq did not face such limitations and was able to detect the majority of the specific genes.

To determine whether the NETSseq gating strategy for each cell type accurately isolated the target cell type, we performed a topological analysis using UMAP generated from the snRNA-seq data, by comparing the overlap between the snRNA-seq assigned cell types and the snRNA-seq derived expression of the top cell type markers identified from NETSseq data (within the highest expression quantile). The analysis revealed a complete overlap between the specific markers and the regions covered by individual cell types, confirming that the correct populations were sorted, without contamination from other cell types (Supplementary Figure 3B). Furthermore, this analysis demonstrated that the sorted populations for each cell type capture the entire cell type population, without any biases for any subpopulations. The only exception is cerebellar astrocytes, which can be split into three regionally, morphologically, and molecularly distinct subpopulations i.e., specialized Bergmann glia (BG), which are intimately connected to the Purkinje cell bodies, granule layer and white matter astrocytes. snRNA-seq data from Lai et al. (2024) subdivided astrocytes into two subgroups i.e., BG and other astrocytes. Comparison of expression profiles shows that our sorted astrocytes contain both BG and other astrocytes (Figures 1E–G, Supplementary Figure 2C). Finally, the cell type–specific genes identified by NETSseq in Golgi and basket cells also delineate distinct regions in the density plot, suggesting that these markers successfully resolve previously unrecognized cellular subpopulations.

Histological studies have shown a progressive loss of Purkinje neurons and to a lesser degree, granule neurons in A-T patients (Crawford, 1998; Rothblum-Oviatt et al., 2016; Sahama et al., 2014). However, less information is available regarding the proportions of other cerebellar cell types and how these change with disease progression. To address these questions, we used various methods to characterize the typical cell type composition of the cerebellum, and to examine how cell type proportions change in disease.

First, we analyzed the flow cytometry profiles obtained during nuclei sorting to quantify the relative proportion of each cell type population (Supplementary Table 2). Consistent with previous stereology studies (Andersen et al., 1992; Korbo and Andersen, 1995), we found that granule neurons accounted for 81%–93% of the cells in cerebellum, whereas Purkinje neurons are extraordinarily rare, comprising 0.001%–0.027% of the cells. Due to the abundance of granule neurons, we found that cell types such as astrocytes and mature oligodendrocytes, which are normally abundant in other brain regions, are relatively rare in the cerebellum, comprising only 2.2% and 1.3% of cells, respectively. To complement this analysis, we estimated cell type proportions using global gene expression deconvolution (Figure 2A). Briefly, for each cerebellar sample we performed RNA-seq on unsorted nuclei, which could then be deconvoluted using reference signatures from each constituent cell type, to obtain cell type proportions of the different cell types. The deconvolution analysis yielded similar results to those from the flow analysis, showing granule neurons as the predominant cell type at 70%, while Purkinje neurons were the least common at 0.7%. Both methods have technical limitations: flow analysis, limited by the efficiency of antibody labeling, likely underestimates cell type proportions, while deconvolution, biased by nuclear size and RNA content, tends to underestimate the proportions of cell types with smaller nuclei (such as granule neurons and glia) and overestimate those with larger nuclei (Purkinje, Golgi, basket neurons). Importantly, the two methods provide relatively consistent proportions and align with cell type proportions determined using stereology (Andersen et al., 1992; Korbo and Andersen, 1995) and single nuclei RNA-seq methods (Lai et al., 2024; Figure 2A, Supplementary Figure 3C).

Next, we aimed to identify changes in cell type proportions in A-T. Given the imbalance of cell type proportions in the cerebellum, where most cell types skew toward zero, we employed a modified deconvolution approach which more accurately captures differences between two conditions. Here, rather than conducting a global deconvolution, where an unsorted sample is split into all constituent cell types and the proportion of these cell types must add up to 1, we used a comparative deconvolution approach. This method assesses a single cell type at a time, making it more useful for measuring changes across conditions rather than absolute cell type proportions.

The density distribution of the relative abundance for each cell type shows the change of cell composition between A-T versus controls (Figure 2B). We detected a significant (t-test with Benjamini-Hochberg correction for multiple comparison) loss of Purkinje neurons, with a decrease of 1.85-fold, which aligns with the typical cerebellar degeneration seen in ataxia (Xia et al., 2013). This degeneration contributes to the impaired motor coordination and balance deficits experienced by affected individuals. We also observed a significant loss of granule neurons with a reduction of 1.35-fold (Figure 2B, Supplementary Figure 3C), highlighting the extensive neuronal loss within the cerebellum. This loss of granule neurons further exacerbates the dysfunction of cerebellar circuits, leading to the pronounced clinical symptoms of ataxia. Consistent with more general neuronal loss, we also observed a decrease of Golgi and basket neuronal cell types with a reduction of 1.52-fold and 1.45-fold, respectively. Additionally, our analysis revealed a 1.35-fold increase in OPCs and a 2-fold increase in ODCs, (Figure 2B, Supplementary Figure 3C), suggesting a compensatory response to ongoing myelin damage (Tse et al., 2018). Astrocytes and microglia show a 1.37- and 1.44-fold increase, respectively, in cell numbers.

We further confirmed the decrease in granule neurons and increase in mature oligodendrocyte numbers using flow analysis. Because the fluorescence profiles for these two cell types were distinct and well separated from other populations in both control and A-T donors, we could confidently quantify their proportions across a subset of donors. This analysis showed that almost half of the granule neuron population degenerate in A-T, decreasing from around 90% in control to 55% in A-T donors. In contrast, oligodendrocytes increase in relative proportion by approximately 4-fold, from 1.5% in control donors to 6% in A-T donors (Supplementary Figure 3D and Supplementary Table 4).

On a subset of cerebellar cortical tissue samples, we performed histopathological analysis to evaluate gross neuropathological features in control and A-T donors (Figures 2C–H). In this analysis, we used antibodies against ITPR1 and FOXK2 to label Purkinje neurons (Figures 2C, D), RBFOX3 to label granule neurons (Figure 2E), IBA1 for microglia (Figure 2F), GFAP for astrocytes (Figure 2G), and OLIG2 for oligodendrocytes and OPCs (Figure 2H). In control Purkinje neurons, FOXK2 is normally localized to the nucleus, while ITPR1 labels the cell body, including dendritic arborizations. Qualitative analysis in A-T donors showed a clear reduction in Purkinje neuron number (Figures 2C, D), while surviving Purkinje neurons have features characteristic of neurodegeneration, including disturbed morphology and ectopic appearance of dendrites in the molecular layer (Figure 2C). Importantly, although Purkinje neurons in A-T donors are reduced in numbers, they can still be reliably labeled using antibodies against both ITPR1 and FOXK2. Qualitative analysis of glial cell markers in A-T donors showed more abundant IBA1 microglia labeling (Figure 2F, Supplementary Figure 3E), more abundant and more intense GFAP labeling in astrocytes (Figure 2G, Supplementary Figure 3E) and more abundant OLIG2 oligodendrocyte lineage labeling, particularly in the white matter (Figure 2H).

Finally, for granule neurons, we performed quantitative analysis using tissue sections from three control and four A-T donors. We observed a 70% decrease in granule neuron numbers in A-T donors, which is consistent with the comparative deconvolution analysis and FACS result (Supplementary Figure 3E). Overall, the histology results support the finding that neuronal cell types decrease, while glial cell types increase in the disease. Furthermore, the increased expression of GFAP suggests increased astrocyte activation in A-T donors.

We conducted a comparative analysis of gene expression values to assess the similarities between our cohort and the single-cell cohort, by computing the pairwise correlation across each cell type (Figure 3A). Interestingly, we observed a strong overall correlation between the gene expression signatures of corresponding cell types across the two cohorts (Figure 3A); however, we detected granule neuron contamination in all cell types in the snRNA-seq dataset (Figure 3B). Moreover, our expressed genes demonstrate greater specificity to individual cell types compared to those in the single-cell cohort (Figure 3B).

We next performed differential gene expression analysis comparing A-T and control samples across all cell types (Supplementary Figure 4 and Supplementary Table 5). The presence of technical variance was accounted for and corrected in the DESeq model using surrogate variables (Supplementary Figure 4). Moreover, we evaluated if any correlation existed between the number of nuclei collected and the transcriptomic signal and found no significant correlation between these two variables (Supplementary Figure 4). To validate our findings, we compared our results with those reported in the snRNA-seq study by Lai et al. (2024) (Supplementary Table 6). Intriguingly, in our cohort, we observed that a higher proportion of differentially expressed genes are cell type specific (Figure 3C). In contrast, this trend is not uniformly observed in the study of Lai et al. (2024), particularly in granule neurons where differentially expressed genes overlap more extensively with those from other cell types (Figure 3D). These results highlight the ability of NETSseq to resolve these cell type specific, disease-associated expression changes, particularly for granule cells (Figure 3C), where the differences in the snRNA-seq data are likely obscured by ambient RNA contamination across all cell types (Figure 3D).

In the cerebellar cortex, mossy fibers (MF) initiate a cascade where they stimulate granule neurons, which subsequently activate Purkinje neurons. Together, these signals are essential for motor coordination (Lee et al., 2023), and disruptions in Purkinje neurons are known to lead to motor impairments such as ataxia. In their study, Lee et al. (2023) emphasized the critical role of granule neurons in motor function by showing that the integrity of MF-induced signaling and consistent firing of Purkinje neurons relies on granule neuron function. Lai et al. (2024) observed a significant loss of granule neurons in A-T; however, challenges in single-cell technology prevented the detection of differentially expressed genes. Our NETSseq platform can overcome those limitations and better elucidate the role of granule neurons in the cerebellum of A-T donors. GSEA analysis performed with the upregulated genes in granule neurons in A-T vs. control donors, highlighted the gene set “Release of cytochrome C from mitochondria.” This process is crucial for apoptosis regulation and is influenced by interferon gamma (IFN-γ), a “response to interferon gamma” gene set was also significantly enriched (Figure 3E). We also noted a significant downregulation of BCL2, an anti-apoptotic protein, in A-T donors, suggesting compromised regulation of apoptosis (Supplementary Table 5). Downregulation of synaptic vesicle transport in cerebellar granule neurons may disrupt neurotransmitter release, impairing excitability of Purkinje neurons and interneurons (Van Der Heijden et al., 2024). Consistent with this observation, processes associated with neurotransmitter release and vesicle docking were markedly enriched among downregulated genes in A-T granule neurons, indicating potential disruptions in synaptic function associated with motor coordination mechanisms (Figure 3F). Metabolic flux analysis computed using NETSseq gene expression values aggregated across entire pathways, suggested that granule neurons in A-T donors may have lower glutamate levels (Figure 3G), likely impairing synaptic transmission within the cerebellum and contributing to functional deficits. Specifically, we noticed a significant downregulation of the glutamate transporter SLC1A6 (Figure 3H), which is crucial for the clearance of glutamate from the synaptic cleft following synaptic release. By taking up glutamate into the cells, SLC1A6 helps maintain low extracellular glutamate concentrations, preventing excitotoxicity and ensuring precise synaptic signaling. Pathways associated with glutamate receptor signaling are also downregulated in Purkinje neurons of A-T donors, indicating a broader disruption of glutamatergic transmission throughout the cerebellum. The reduced glutamate receptor signaling in Purkinje neurons likely impairs neuronal function and exacerbates motor coordination deficits since these neurons play a central role in controlling cerebellar output (Supplementary Figure 4). Taken together, the changes in glutamate levels and the dysregulated synaptic transmission processes may indicate a broader disruption in the balance / control between excitatory and inhibitory signaling. This dysregulation, together with the loss of granule neurons could impair the integrated timing and coordination of motor commands. These findings suggest that disruptions in glutamate neurotransmission contribute to the motor dysfunction observed in A-T, highlighting potential therapeutic targets for future studies.

“Reactive astrocytes” refers to astrocytes that are undergoing morphological, molecular and functional remodeling in response to adverse microenvironmental changes and pathological stimuli. Astrocyte activation is implicated in the pathogenesis of multiple neurodegenerative diseases. Extensive changes were observed in our NETSseq derived astrocyte A-T datasets. Three mouse astrocyte activation gene expression signatures have previously been reported, corresponding to a common “PAN” activation profile, an “A1” neurotoxic activation profile and an “A2” neuroprotective profile (Liddelow et al., 2017). Expression analysis indicated that the “A1” neurotoxic and common “PAN” gene signatures are significantly upregulated in A-T donors, compared to controls (p-value < 0.005) (Figure 4A), while the A2 signature was not significantly different. In addition to the genes specified in the A1 gene set, release of complement components is also characteristic of the A1 activation state and is postulated to drive synapse degeneration (Liddelow et al., 2017). Our data shows that several activators of complement components are significantly upregulated in A-T donors compared to healthy controls (Figure 4B). Conversely, most complement inhibitors have higher expression in controls, further supporting the hypothesis of an overall activation (Supplementary Figure 4). Interestingly, CFI, a complement inhibitor, is significantly upregulated in A-T donors, possibly as a compensatory mechanism in response to the excessive complement activation (Davies and Spires-Jones, 2018). While there are nuances and complexities related to astrocyte subset identification, function, and nomenclature of activation (Patani et al., 2023; Qian et al., 2023), the findings mentioned above suggest an overall increased activation state in astrocytes in A-T. This heightened state may lead to an elevated neuroinflammatory response and burden, potentially contributing to the ongoing pathophysiology of the disease.

An important function of astrocytes in the healthy brain is to support synapse function, including uptake of neurotransmitters and release of trophic factors essential for neuronal survival and communication. As astrocytes become more reactive, they lose their capacity to support synapses and normal neuronal function (Brandebura et al., 2023). Given that A-T is a neurodegenerative disease, we were particularly interested in exploring how genes involved in glial support of synaptic function were altered. To address this, we performed GSEA on the genes that were either upregulated or downregulated in astrocytes from donors with A-T. We then conducted a network analysis to identify and group together similar biological processes. Our findings revealed that upregulated genes in A-T astrocytes were predominantly associated with inflammatory processes, consistent with our earlier observation of an increased neurotoxic and activated glial signature. This suggests that in A-T, astrocytes may be shifting toward a more reactive and potentially harmful state (Figure 4C and Supplementary Table 7). Downregulated genes were involved in synaptic function and neuronal interaction (Figure 4D and Supplementary Table 7) indicating a very extensive disease driven remodeling and synaptic reorganization in the cerebella of donors with A-T. In addition, genes critical for maintaining the structural integrity of the astrocyte-synapse interaction, as well as those involved in monitoring the activity of adjacent synapses (such as neurotransmitter receptors, ion channels and transporters) were significantly downregulated. This downregulation extends to receptors for the primary neurotransmitters in the cerebellum, i.e., glutamate and GABA, both of which play essential roles in maintaining normal cerebellar function. The marked reduction in the expression of these receptors (Supplementary Figure 5) indicates a significant impact of A-T on synaptic communication. The typical profile for our cerebellar astrocytes suggests a considerable similarity to that of Bergmann glia. These are highly specialized cerebellar astrocyte populations which have a close and highly interconnected interaction with Purkinje neurons, controlling and supporting the synaptic function of these rare neurons (De Zeeuw and Hoogland, 2015). Thus, the observed deficits seen in their function could be much more significant in the cerebellum and long term have a greater impact on the disease progression.

Astrocyte specific expression profiles from post-mortem tissue from patients with Alzheimer’s disease (AD) have previously been published. Mathys et al. (2019), used snRNA-seq to examine prefrontal cortex cell types including astrocytes. Although changes were observed, no explicit connection was made to A1 activation. We employed the “Gene Set Score” method of Srinivasan et al. (2020) to compare astrocyte changes in one published AD data set, an ALS SOD1 mouse model (Liu et al., 2020) and a human Parkinson’s disease (PD) data set (Smajić et al., 2022) with those changes observed in A-T. These results (Figure 4E) show a consistent directional correlation of astrocyte genes seen in A-T with the astrocyte gene sets derived from the Mathys AD studies. Additionally, the PD study showed a correlation between its upregulated genes and the changes seen in A-T.

Our NETSseq data revealed a striking down-regulation of synaptic transmission genes in A-T associated astrocytes. GSEA analysis showed that this downregulation is also observed for key genes in the Mathys et al. (2019) AD data set (Supplementary Figure 5). This suggests that down-regulation of synaptic transmission genes is a common feature of activated astrocytes regardless of brain region or disease. Ongoing astrocytic dysfunction in regulating synaptic function may therefore be a phenomenon found across neurodegenerative diseases rather than being specific to A-T.

Given the size and spread of age across the current cohort of control donors is relatively large, we were able to investigate the impact of normal aging on cerebellar gene changes and compare these findings with those from the A-T donors. PCA analysis of the most differentially expressed genes revealed that PC1 is associated with age in control donors. Interestingly, the subjects with A-T align with much older control donors, suggesting an accelerated aging phenomenon (Figure 4F). To better understand how A-T donors compare with age-matched controls, we analyzed the expression of A1-associated genes in younger and older control groups and A-T donors (Figure 4G). Interestingly, some markers showed similar expression levels between the two control groups but were upregulated in A-T donors, suggesting that neurotoxicity is more pronounced in the disease than in normal aging. In contrast, PAN markers such as GFAP, CD44, and CP were highly expressed in older controls compared to younger donors but showed an even greater elevation in expression in A-T donors (Figure 4H), indicating an enhanced or accelerated activation signature of PAN-associated markers in A-T.

To further explore the changes occurring in donors with A-T and with aging, we performed enrichment analysis on the upregulated differentially expressed genes in both conditions (Supplementary Tables 5, 8). A consistent and significant enrichment of gene sets associated with changes in AD or associated with cognitive decline was observed (e.g., Mathys et al., 2019 - Astrocyte AD vs. non-AD (up); Mostafavi et al., 2018 - ROSMAP-Clin AD enriched; Mostafavi et al., 2018- ROSMAP- Cognitive decline (UP)) in both aging and A-T. However, specific enrichment for gene sets associated with inflammation and astrocytic activation was seen in A-T only (Figure 4I and Supplementary Table 9), further suggesting that the conversion to a more toxic astrocytic phenotype may be part of the normal aging process, but which is significantly accelerated as a consequence of the ATM mutated protein in A-T.

Lai et al. (2024) showed that activated microglia in the A-T cerebellum exhibit transcriptomic signatures resembling those seen in aging and neurodegeneration-associated microglia. However, they performed their analysis comparatively. Our microglia cohort consists of control donors with age ranges between 8 and 93 years old. Therefore, an assessment of the gene expression patterns associated with age were compared within the A-T dataset. PCA, based on the most variable genes identified through age regression analysis, demonstrated a significant correlation with age progression. Interestingly, donors with A-T are positioned at a similar level to older control donors (Figure 5A) suggesting that the microglial gene expression profiles seen in the A-T cerebellum mirrors the aging process but in much younger donors with A-T. This may suggest that the ATM mutation is accelerating the aging phenotype of microglia in A-T, contributing to an accelerated neurodegenerative disease process.

Sun et al. (2023) identified 12 microglia clusters associated with distinct biological functions. Using gene set scoring to analyze our A-T microglia samples, we see that the homeostatic and surveillance functions, MG00 and MG01, are impaired in A-T (p-values of 0.093 and 0.019, respectively); among the inflammation pathways, MG02 is the most elevated (Figure 5B). The homeostatic microglial cluster is crucial for maintaining normal brain function, while the surveillance cluster monitors and responds to the brain microenvironment for signs of cellular damage or infection. The downregulation of these clusters in A-T suggests a compromised ability of microglia to perform their normal essential homeostatic roles, potentially as a consequence of chronic inflammation and ongoing cellular stress which are hallmark features of A-T and other neurodegenerative conditions, such as AD. Interestingly the MG10 inflammation cluster was previously associated with changes in early AD, while the MG02 inflammation cluster was linked to late AD changes (Sun et al., 2023). Overall, our A-T microglia signature reflects changes seen in late AD indicating a more advanced inflammatory state in A-T brains. Taking these microglial observations, together with the astrocyte inflammatory signaling changes, these findings reflect the importance of glial cell dysfunction in the pathophysiology of A-T, pointing toward disrupted immune surveillance and altered inflammatory responses across cell types as contributing factors to A-T disease progression.

A-T is characterized by defects in the ATM protein kinase (Lai et al., 2024), which plays a central role in DNA damage response pathways. ATM is expressed at significantly higher levels in control donor microglia compared to the other cerebellar cell types examined here suggesting that microglia may be more affected by deficient ATM than other cell types (Figure 5C). This is likely attributable, at least in part, to the ability of adult microglia to proliferate as part of their normal homeostatic physiological condition and exacerbated during pathology-induced stimulation (Askew et al., 2017).

Consistent with this hypothesis, the most upregulated pathway, as determined by the most overexpressed genes in A-T, is the Reactome gene set “Activation of ATR in response to replication stress” (Figure 5D and Supplementary Table 7). Within this group, genes involved in cell cycle checkpoint and DNA replication are consistently upregulated, including CHEK1 (Figure 5E), a positive regulator of cell cycle arrest. This potential adaptive response aligns with the importance of ATR signaling in preserving genomic integrity under conditions of increased replication stress. This process is likely to be specific to glial cell types and particularly microglia during periods of elevated inflammation and cellular damage. Expression analysis of the leading-edge genes identified in the GSEA results associated with ATR response show a microglia-specific upregulation of these genes relative to other cell types (Figure 5F).

Among the 12 microglial clusters, MG11 was the most significant and differentially regulated between donors with A-T and control donors (Figure 5B). This gene set was associated with antiviral response, thus suggesting similarities in activation mechanisms needed for viral-induced responses and A-T microglia activation. It is known that cytosolic viral DNA induces type 1 interferon as a defense mechanism (Nakad and Schumacher, 2016). It has been argued previously that DNA damage or inhibited DNA repair can lead to nuclear DNA being transported into the cytoplasm leading to a “sterile inflammation response” and the induction of interferons (Nakad and Schumacher, 2016). In support of this, genes in the KEGG pathway “Cytosolic DNA Sensing,” which involves the detection of cytosolic DNA by immune receptors and triggers an antiviral response through the production of type I interferons and cytokines, are specifically expressed in microglia compared to other cerebellar cell types (Supplementary Figure 5) (Kanehisa et al., 2025).

Collectively, these findings indicate that microglial pathways are significantly changed in A-T, and these disruptions may play a crucial role in driving aberrant neuroinflammation and progressive neurodegeneration (Lai et al., 2024; Song et al., 2019).

All human tissue donations used in this study were fully consented for research. Tissue was stored and data managed in compliance with the UK Human Tissue Act, with local Research Ethics Committee approval obtained from the Health Research Authority and in accordance with the World Medical Associations Declaration of Helsinki for medical research. Tissue was obtained from the NIH Neurobiobank at the University of Maryland, Baltimore, MD, Queen Square Brain Bank (The Queen Square Brain Bank is supported by the Reta Lila Weston Institute of Neurological Studies, UCL Queen Square Institute of Neurology), The Netherlands Brain Bank (Netherlands Institute for Neuroscience, Amsterdam); all material was collected from donors for or from whom a written informed consent for a brain autopsy and the use of the material and clinical information for research purposes had been obtained.

Snap frozen cerebellar tissue samples were dissected on a brass plate, over dry ice. Tissue blocks for FFPE processing were first fixed in formalin at room temperature, the excess formalin drained and washed in water. Using an automated tissue processor (TP1020, Leica) the tissue was passed through a rising concentration of ethanol (70%, 80%, 95% and 100% twice for 1 h), then HistoChoice (twice for 1 h), before being immersed in Paraplast Plus (twice for 1 h). Tissues were placed in molds, embedded in paraffin using a Paraffin Embedding Station (EG1150H, Leica), left to set on a cool plate (EG1150C, Leica), and were removed from the mold and stored at room temperature until required for sectioning. FFPE sections were cut from blocks at a thickness of 4 μm on a rotary microtome (Leica RM2255; 535 blades). Sections were floated in a water bath to flatten, before mounting on microscope slides. Slides were dried on a hotplate at 40°C for 4 h.

Frozen blocks were sectioned at a thickness of 10 μm in a freezing cryostat (Leica CM1850; MX35 blades), before mounting on microscope slides and stored at −80°C.

Sections were fixed in 4% paraformaldehyde/phosphate buffered saline (PBS) for 10 min, washed, blocked for non-specific protein binding and endogenous peroxidase activity and then incubated with antibodies for 20 min at room temperature. Following primary incubation, sections were washed and incubated with horseradish peroxidase (HRP)-conjugated secondary antibodies (20 min, room temperature), washed again and then incubated with chromogen (10 min, room temperature). Sections were counterstained using hematoxylin, dehydrated through an ascending ethanol series, cleared and cover-slipped under Xylene.

Sections were evaluated using brightfield microscopy (Zeiss Axiophot) and IHC staining patterns were compared with negative control incubations (non-immune rabbit immunoglobulin G) performed under identical conditions on adjacent sections.

RNAscope^® 2.5 HD Assay (ACDBio) was carried out according to the manufacturer’s instructions, against the probes listed in Supplementary Table 10, on cerebellum tissue from healthy donors. A GFAP probe was included as a positive assay control. Frozen sections were fixed with pre-chilled 4% paraformaldehyde for 15 min at room temperature and passed through a series of rising ethanol gradients (50%, 70% and 100% twice for 5 min each), then left at room temperature for 5 min to fully dehydrate. Sections were treated with hydrogen peroxide for 10 min, before washing in diethylpyrocarbonate-treated water for 5 min. Pre-warmed probes were applied for 2 h at 40°C, within a humidity-controlled chamber (HybEZ oven, ACD Bio). Following the incubation with probes, slides were washed twice in wash buffer (WB, ACDBio) for 2 min, at room temperature.

Signal amplification and detection reagents (RNAscope^® 2.5 HD Detection Reagents-RED, ACDBio) were applied sequentially. Sections were incubated in AMP 1, AMP 2, AMP 3, and AMP 4 for 30, 15, 30 and 15 min, respectively, at 40°C, with 2 × 2 min WB washes in between. AMP 5, and AMP 6 reagents were applied for 30 and 15 min, respectively, at room temperature, with 2 × 2 min WB washes in between. Chromagen FAST RED was then applied for 10 min, at room temperature, before washing in WB twice for 2 min. The sections were counterstained with 50% Gill’s hematoxylin I (Pioneer Research Chemicals) for 30 s at room temperature, rinsed, and dried for 15 min at 60°C, before applying a coverslip with VectaMount (Vector Laboratories).

Sections immunohistochemically stained using a polyclonal antibody to RBFOX3 were digitally scanned (Hamamatsu Nanozoomer). Three neuroanatomically-matched fields of view were selected from three normal donors and four donors with A-T. Image processing and analysis was carried out using Image J (Ver.1.51 National Institutes of Health). Positively stained nuclei were accurately segmented according to both color and staining intensity and automatically counted. Adjustments based on the average size of positive nuclei were made for aggregates.

Nuclei isolation protocols described fully in previous publications (Xu et al., 2018), were miniaturized to enable a more efficient and higher throughput method to process larger donor numbers.

To isolate nuclei, cerebellar frozen tissues were thawed on ice for a minimum of 30 min and transferred to 1 mL of homogenization medium (0.25 M sucrose, 150 mM KCl, 5 mM MgCl₂, 20 mM Tricine pH 7.8, 0.15 mM spermine, 0.5 mM spermidine, EDTA-free protease inhibitor cocktail, 1 mM DTT, 20 U/mL Superase-In RNase inhibitor, 40 U/mL RNasin ribonuclease inhibitor). Using a 2 mL glass Dounce tissue grinder set, samples were homogenized by 30 strokes of a large clearance pestle A (0.003–0.005 in.) followed by 30 strokes of small clearance pestle B (0.005 – 0.0025 in.). Homogenate was adjusted to 1.92 mL of homogenization medium and supplemented with 1.78 mL of a 50% iodixanol solution (50% Iodixanol/Optiprep, 150 mM KCl, 5 mM MgCl₂, 20 mM Tricine pH 7.8, 0.15 mM spermine, 0.5 mM spermidine, EDTA-free protease inhibitor cocktail, 1 mM DTT, 20 U/mL Superase-In RNase inhibitor, 40 U/mL RNasin ribonuclease inhibitor), and laid on a 27% iodixanol cushion. Nuclei were pelleted by centrifugation for 25 min, 10,000 rcf, 4°C in an Eppendorf 5427 R centrifuge, FA-45-12-17 rotor. The nuclear pellet was resuspended in homogenization buffer.

After isolation, resuspended nuclei were fixed with 1% formaldehyde for 8 min at room temperature, and then quenched with 0.125 M glycine for 5 min. Nuclei were pelleted at 1000 rcf, 4 min, 4°C in a FA-45-30-11 rotor, and then washed once with homogenization buffer and once with Wash Buffer (PBS, 0.05% TritonX-100, 50 ng/mL BSA, 1 mM DTT, 10 U/μL Superase-In RNase Inhibitor). Nuclei were blocked with Block Buffer (Wash buffer with an additional 50 ng/mL BSA) for 30 min, incubated with primary/fluorescently conjugated antibody for 1 h, and then washed three times with Wash Buffer with spins in between washes as described above. Nuclei were then incubated in secondary antibody for 30 min and washed three times with Wash Buffer. All incubation steps were performed at room temperature. Primary, secondary and conjugated antibodies were diluted in Block Buffer.

Prior to flow cytometry, nuclei were co-stained with DAPI to 0.01 mg/mL final concentration. Nuclei were analyzed and sorted using a BD FACSAria Fusion (BD Biosciences, San Jose, CA, USA) flow cytometer using the 355 nm, 488 nm, 561 nm, and 640 nm lasers. All samples were first gated using an FSC/SSC gate to exclude debris, followed by a single nuclei DAPI gate to exclude aggregated nuclei. Analysis was performed using FACSDiva (BD) or FlowJo software. To reduce RNA degradation all nuclei samples were sorted at 4°C and stored at −80°C, before being processed for RNA extraction.

Two different antibody panels were used to isolate cell types. RBFOX3, FOXK2, ITPR1, OLIG2 and IRF8 were used to sort Purkinje cells, basket cells, granule cells, Golgi cells, oligodendrocytes, oligodendrocyte precursor cells and microglia. RBFOX3, SLC1A3, OLIG2 and IRF8 were used to sort oligodendrocytes, oligodendrocyte precursor cells, microglia and astrocytes. For an exhaustive list of antibodies used refer to Supplementary Table 10. The Mann–Whitney U test was used to check for significant differences in percentage cell-type composition between control and AT samples for astrocytes, oligodendrocytes and oligodendrocyte precursors. Linear regression was used to rule out age as an explanatory variable.

RNA extractions from sorted nuclei were carried out with FFPE RNA Purification kits (Norgen: Product # 25300) using the manufacturer’s protocol from Step 2 (Lysate Preparation) with the following modifications: The On-column DNA removal Protocol was omitted. RNA binding and Column Wash centrifuge steps were carried out at 1,789 × g for 2 min, followed by 1,789 × g for 5 min to dry the resin. Elution was carried out in two steps, with 25 μL Elution Solution A added to each column followed by a 1 min room temperature incubation and centrifugation at 200 × g for 2 min, then a further 15 μL Elution Solution A was added to each column with centrifugation at 1,789 × g for 5 min. For samples containing > 2,000 nuclei, the final 30 μL elution was split into two 15 μL aliquots. Purified RNA was vacuum dried in a CentriVap Benchtop Vacuum Concentrator (Labconco) for 40 min (or until no liquid was visible) at 4°C and either used directly for library preparation or stored at −20°C.

Sequencing library generation was carried out using the Trio RNA-Seq, Human rRNA Library Preparation Kit (Tecan: Product # 0506-A01) using the manufacturer’s protocol with the following modifications: Lyophilized RNA was resuspended in 12.5 μL DNAse mix and incubated at 37°C for 30 min followed by 60°C for 5 min and 4°C hold. Library amplification I was carried out as 72°C 2 min/95°C 2 min/2 × (95°C 30 s/ 60°C 60 s)/6 × (95°C 30 s/ 65°C 60 s)/65°C 5 min/4°C hold. Library Amplification II was carried out as 95°C 2 min/2 × (95°C 30 s/ 60°C 60 s)/6 × (95°C 30 s/ 65°C 60 s)/65°C 5 min/4°C hold. Final libraries eluted in 30 μL DNA Resuspension Buffer (DR1) were checked on Agilent Bioanalyzer DNA 1000 chips, then measured on a Nanodrop 8000. 32 sample pools (normalized to 500 ng/sample then diluted to 5 ng/μL) were sent for sequencing on an Illumina NextSeq High Output platform.

FastQC¹ was used for initial quality control. Adapter trimming was performed with Cutadapt v1.12 (Martin, 2011). Reads were aligned to the hg38 human genome build (NCBI primary analysis set) using STAR aligner (Dobin et al., 2013) v2.5.2b with splice junction files generated from Gencode release 27 and RefSeq GRCh38.p10. Post-alignment quality control was performed by examination of the STAR alignment metrics, Picard tools v2.8.2 (CollectRNASeqMetrics)² and RSeQC v2.6.4 (tin.py) (Wang et al., 2012). Expression quantitation was performed with htseq-count v0.6.1p1 (Anders et al., 2015) using a custom annotation file based on RefSeq where both exonic and intronic reads are counted for each gene.

The read counts table from each sample was imported into RStudio (R version 4.0.5) running under Red Hat Enterprise Linux 8. Using edgeR (Robinson et al., 2010), the read counts were normalized into Counts Per Million (TPM), log transformed, and batch corrected. For each individual cell type, we removed technical variance associated with batch effect using the sva R package (Leek et al., 2012). Briefly, the sva R package provides detection of unwanted variation in high-dimensional data. This unwanted variation, often caused by batch effects, technical artifacts, or unknown confounding factors, can obscure the true biological signals of interest. We incorporated relevant surrogate variables in the differential analysis models; the models aim to assess differences due to the disease condition (or any variable of interest such as age), while accounting for covariates or potential confounders. Differentially expressed genes were identified using the DESseq R package (Love et al., 2014).

Hierarchal clustering analysis was performed using the pheatmap R package (Kolde, 2018). Principal component analysis was performed using the pcaMethods R package (Stacklies et al., 2007). Metabolic flux analysis was computed using the scFEA R package (Alghamdi et al., 2021). Briefly, scFEA is a tool designed to infer metabolic activity based on gene expression data. It works by modeling the flux of metabolites through predefined metabolic modules, which represent groups of reactions within metabolic pathways. Using transcriptomic data as input, scFEA predicts the activity of these modules by integrating gene expression levels with a flux balance analysis framework. This allows for the estimation of pathway activity across the samples, providing insights into cellular metabolism and its variation across different conditions. Cell type specific genes were identified using the specificity index algorithm previously described in Xu et al. (2018) and Dougherty et al. (2010). log₂(TPM) values were used as gene expression input and ranks were averaged across 1,000 iterations. The sequencing data generated in this study is publicly available at the GEO repository under accession number GSE282079.

To quantify cell type proportions in the unsorted populations we performed a deconvolution analysis using the deconRNAseq R package (Gong and Szustakowski, 2013). deconRNAseq works by leveraging reference expression profiles, which are known RNA expression levels of different cell types, to analyze the mixed tissue sample’s RNA sequencing data. It uses these profiles to construct a mathematical model that deconvolutes the mixture, identifying the contribution of each cell type. The algorithm optimizes the fit between the observed mixed sample data and a linear combination of the reference profiles, thereby estimating the proportion of each cell type present in the sample. We assembled the reference dataset using our sorted, cell type specific, RNA-seq expression profiles. This method enabled us to estimate the relative abundances of each cell type in the unsorted samples. For the comparative deconvolution analysis, we first created a reference panel for each cell type. Each reference panel consisted of two expression signatures: one representing the target cell population and the second representing the average gene expression of all other cell types. This resulted in nine distinct signature datasets, each corresponding to a different cell type in comparison to the rest. Next, we compared the gene expression profiles of the unsorted samples against each of these reference signatures. For each comparison, we quantified the fraction of similarity between the unsorted sample and a given target cell type, relative to the expression of all other cell types. This approach allows us to assess the proportion of similarity between the unsorted samples and individual cell types, providing a more accurate means of identifying changes in the proportion of specific cell types between experimental conditions.

Gene Set Enrichment Analysis (GSEA) was performed using the method described by Subramanian et al. (2005) on gene sets generated from the Molecular Signatures Database (MSigDB) v6.2 against a ranked list of genes derived from the DESeq2 analysis results. Pre-ranked gene lists (.rnk files) were created from the DESeq2 Wald statistic, low expressed genes with baseMean less than 10 were filtered out to reduce noise.

Microglial state marker genes were harvested from Sun et al. (2023) as is. For each state, labeled MG0∼MG12, genes significantly upregulated (adjusted P-value < 0.05) and with a minimum fraction of cells expressing the gene of 0.25 were used to build the gene sets for GSEA.

Similarly, we harvested disease-related and cell type specific gene sets from several relevant publications. These gene sets represent up- or down-regulation with respect to key Alzheimer’s disease traits. We extracted significantly differentially expressed genes (adjusted P-value < 0.05) and subdivided them based on the direction of their fold changes.

To compare changes between gene expression profiles we generated with other changes seen in other neurodegenerative diseases or models we used the gene set score method described in Srinivasan et al. (2020).