A comparative transcriptomic analysis of mouse demyelination models and multiple sclerosis lesions Authors: Erin L Aboelnour, Veronica R Vanoverbeke, Elizabeth A Maupin, Madelyn M Hatfield, Katrina L Adams Published: 2026 ABSTRACT AbstractDemyelinating diseases, including multiple sclerosis (MS), are characterized by loss of myelin and progressive neurodegeneration. It remains unclear if demyelination mouse models, such as cuprizone (CPZ) and lysophosphatidylcholine (LPC) elicit distinct responses or are comparable to human disease. Here, we integrate new and published single-cell transcriptomic datasets from CPZ- and LPC-induced demyelination and compare them with human MS data. We find that CPZ induces a distinct, stressed oligodendrocyte (OL) state, marked by Cdkn1a and Nupr1, that resembles phenotypes in MS lesions. The models converge on an immune responsive OL state expressing Socs3, B2m, and interferon-response genes during remyelination. Mouse microglia share a conserved activation program, although LPC drives a stronger, prolonged response. However, neither model captures the oligodendrocyte progenitor and microglial heterogeneity observed in MS. These results provide a cross-model, cross-species atlas of glial states and offer a framework to strategically leverage mouse models to study myelin injury and repair.Subject terms: Oligodendrocyte, Microglia, Multiple sclerosis Introduction Demyelination, the loss of myelin sheaths around neuronal axons, is a major driver of neurodegenerative disease. In the young, healthy central nervous system (CNS), demyelination is spontaneously repaired and sufficient to provide neuroprotection and slow disease progression1,2. In multiple sclerosis (MS), however, evidence of myelin repair is highly variable and often incomplete, and there are no approved therapies that directly enhance remyelination. Understanding the cellular and molecular mechanisms underlying repair is therefore a pressing need. Toxin-induced demyelination models are widely used to study these mechanisms because they provide a temporally controlled and reproducible injury. The most commonly used models are LPC injection and CPZ dietary administration3,4. LPC is a detergent that creates a focal lesion that undergoes repair through reproducible phases of demyelination, increased reactive gliosis, and differentiation of oligodendrocyte progenitor cells (OPCs) into OLs within approximately 4 weeks. CPZ is a copper chelator that, when fed to mice for 4–6 weeks, induces OL cell death throughout the grey and white matter (GM and WM, respectively) of the brain. After CPZ withdrawal, the brain undergoes spontaneous OPC differentiation and remyelination within 2–3 weeks5,6. These two toxicity models are often used interchangeably by researchers, with a particular method selected because of experimental design constraints rather than scientific rationale. However, it remains unclear whether LPC and CPZ induce comparable molecular changes and cellular states. The advancement of single-cell (sc) and single-nucleus (sn) RNA sequencing (sc/sn-RNA-seq) and supporting bioinformatics tools has enabled unbiased characterization of the molecular responses across neuronal and glial populations from a multitude of CNS pathologies7. Previous scRNA-seq studies of LPC and CPZ models have described distinct disease-associated oligodendrocyte (DAO) states, including similar transcriptional changes to those in Alzheimer’s disease models, but have not directly compared LPC and CPZ8–10. Similarly, sequencing analyses of postmortem human brain samples have provided valuable insight into the cellular and molecular pathology of MS11–14. MS lesions, particularly the rims of chronic active lesions, contain stressed OLs, reactive astrocytes, and activated phagocytosing microglia/macrophages11,12,15. While mouse models have been used to identify altered signaling pathways conserved in MS lesions, a direct comparison of mouse toxicity-induced demyelination models to human MS lesion transcriptomics across multiple cell types has not been performed9,16. A greater understanding of the similarities and differences between mouse models and MS lesions is critical for effectively using these models to identify and test new therapeutic targets. In this study, we directly compare LPC- and CPZ-induced demyelination and remyelination in subcortical WM by integrating newly generated and published sc/snRNA-seq datasets from the corpus callosum8,9. We also reanalyze an available snRNA-seq dataset of the largest cohort of MS lesions to date. Within the OL lineage, CPZ uniquely induces a strongly dysregulated DAO state, marked by the expression of Cdkn1a and Nupr1, while both models converge on a similar remyelination state characterized by immune gene upregulation, including Socs3 and B2m. Notably, CDKN1A and SOCS3 are also upregulated in human DAOs, and we show that CPZ better recapitulates the human DAO transcriptional state, as evidenced by shared stress pathway activation and downregulation of myelin maintenance genes. Although OPCs in both mouse models adopt stress- and immune-associated programs that partially overlap with DAO signatures, human OPC responses are highly heterogeneous across lesion types, indicating only partial cross-species conservation. In the innate immune cells, we find that microglia and macrophages from both models exhibit a largely conserved core activation program; however, LPC elicits a more robust and sustained response. A conserved disease-associated microglia/macrophage (DAM) signature is observed across species, but we find that microglia in MS display distinct transcriptional states across lesion types. DAMs in remyelinated lesions notably enriched for hypoxia and glycolysis pathways that are broadly induced in mouse microglia after demyelination. Together, these findings reveal conserved and divergent lineage-specific responses between toxin models and MS, providing rationale for strategically using these models to dissect mechanisms of myelin injury and repair. Results Integrated single-cell and single-nucleus transcriptomic datasets from mouse demyelination models and MS patient samples To compare the effects of LPC- and CPZ-induced demyelination on different WM cells, we performed snRNA-seq of the adult mouse corpus callosum, and integrated our data with two additional published scRNA-seq studies of the corpus callosum following LPC- or CPZ-induced demyelination8,9. Together, these datasets span a time course from baseline (homeostatic conditions) through demyelination and subsequent remyelination (Fig. 1a and Supplementary Data 1). Notably, the LPC baseline controls in the snRNA-seq samples were saline-injected, whereas in the scRNA-seq dataset they were uninjected. We obtained 112,017 high-quality cells and nuclei after filtering, and the fully integrated UMAP displayed strong overlap between datasets and transcriptome types while maintaining clear cell type separation (Fig. 1b and Supplementary Fig. 1a–c). Using expression of known marker genes, we annotated clusters corresponding to OL lineage cells (OPCs, committed oligodendrocyte precursors (COPs), and mature OLs (MOLs)), microglia, perivascular macrophages (PVMs), astrocytes, neural progenitor cells (NPCs), ependymal cells, endothelial cells, vascular cells, pericytes, T cells, and three clusters undergoing cell division (OPCs, microglia, and NPCs) (Fig. 1c and Supplementary Fig. 1d–e). MOLs and microglia constituted the majority of captured cells (Fig. 1c). Single-cell and single-nucleus RNA-seq datasets from mouse demyelination models and MS tissue. a Schematic depicting mouse datasets included in this study. Adult corpus callosum was microdissected from the LPC- or CPZ-induced demyelinated and remyelinated tissue, and controls. Single-cell suspensions were generated by enzymatic dissociation followed by FACS. Replicate numbers for each condition and timepoint are indicated. Datasets were integrated using sysVI. Created in BioRender. Aboelnour, E. (2026) https://BioRender.com/hrx906t. b UMAP embedding of the integrated mouse dataset, annotated by cell type. Insets show UMAPs colored by dataset origin, and labeled with demyelination treatment and library type. c Bar chart (left) showing the cell-type composition and dot plot (right) of canonical marker gene expression. d UMAPs showing cell-type distribution across treatments (LPC or CPZ) and experimental stages. e Overview of the human snRNA-seq dataset. White matter control samples (n = 13) and multiple sclerosis lesions classified by histopathology (79 lesions from 34 patients) were integrated using scVI. Created in BioRender. Aboelnour, E. (2026) https://BioRender.com/hrx906t. f UMAP embedding of the integrated human dataset, annotated by major cell type. Inset shows concordance of the cell-type annotations with the original publication. g Bar chart (left) showing the cell-type composition and dot plot (right) of canonical marker gene expression. h UMAPs colored by cell-type annotation for each lesion type. sc single cell, sn single nucleus, LPC lysophosphatidylcholine, CPZ cuprizone, UMAP uniform manifold approximation and projection, FACS fluorescent activated cell sorting, OPCs oligodendrocyte progenitor cells, COPs committed oligodendrocyte precursors, MOLs mature oligodendrocytes, PVMs perivascular macrophages, NPCs neural precursor cells, VSMCs vascular smooth muscle cells, VLMCs vascular and leptomeningeal cells, Endo + Peri endothelial cells and pericytes, WM white matter, NAWM normal-appearing white matter, AL active lesion, CAL chronic active lesion, CIL chronic inactive lesion, RL remyelinated. Visualization of UMAPs stratified by condition and timepoint revealed shifts in cell type abundance and transcriptional states in LPC and CPZ models. Both paradigms exhibited a strong reduction in MOLs and an increase in microglia at demyelination, as expected (Fig. 1d and Supplementary Fig. 1f–g). Although astrocytes are known to respond robustly in both demyelination models17,18, astrocyte representation was insufficient in these datasets for robust downstream analyses. Notably, PVMs were specifically enriched in the LPC model, showing distinctions at the cellular level between LPC- and CPZ-induced responses. Using scCODA to statistically test for compositional changes, we found that LPC demyelination was characterized by increased PVM and microglia abundance and MOL depletion (Supplementary Fig. 1g). At the LPC remyelination timepoint, microglia remained elevated, accompanied by increases in T cells and MOLs and a modest decrease in Dcx+ NPCs19. In the CPZ model, microglia also significantly increased during both demyelination and remyelination, whereas MOL depletion was restricted to demyelination. To generate a comparable human cohort, we also reanalyzed a recently published snRNA-seq MS dataset, extracting only the WM samples from their atlas14. In total, the human data includes 13 control WM samples from non-MS patients, 17 normal-appearing white matter (NAWM) samples, 21 active lesions (ALs), 17 chronic active lesions (CALs), 13 chronic inactive lesions (CILs), and 11 remyelinated lesions (RLs) from a total of 47 individuals (Fig. 1e and Supplementary Data 2). There were no significant differences in the age, sex (25 female, 22 male), postmortem interval (PMI), or counts per nucleus (Supplementary Fig. 2a–c). Multiple lesions were collected from many individuals, totaling 92 samples, and the majority of the lesion samples were derived from patients with progressive forms of the disease–secondary progressive MS (SPMS) and primary progressive MS (PPMS) (Supplementary Fig. 2d and Supplementary Fig. 2e–g). After quality filtering, we integrated 321,565 nuclei and annotated the human cell types using canonical markers, recapitulating the major cell classes reported in the original study (Fig. 1f–g and Supplementary Fig. 3a–c). As in the mouse data, the human WM atlas is composed primarily of glial cells, particularly MOLs, microglia, and astrocytes. UMAPs stratified by lesion type revealed shifts in microglia, astrocyte, and MOL transcriptional states (Fig. 1h), and scCODA analysis showed a significant decrease in MOL abundance in ALs and CALs (Supplementary Fig. 3d, e). Peripheral immune cells, including T and B cells, were predominantly detected in MS lesion samples and were largely absent from control WM, consistent with their known pathological roles in MS. Together, these integrated mouse demyelination and human MS datasets recapitulate expected cellular and compositional changes and provide a harmonized cross-species framework for comparative transcriptomic analyses of demyelination and remyelination. Cuprizone demyelination induces a distinct DAO state not present in LPC demyelination To define how OL lineage cells respond to LPC versus CPZ demyelination, we subclustered the OL lineage cells from the integrated mouse sn/scRNA-seq dataset (Fig. 1b) and reannotated them using established markers for OPCs, COPs, newly formed OLs (NFOLs), myelin-forming OLs (MFOLs), and MOLs (Fig. 2a, b and Supplementary Fig. 4a–c). PAGA connectivity recapitulated the known developmental progression of progenitors into mature states, inferring connectivity from MFOL to MOL_A. We defined four homeostatic mouse MOL states (MOL_A-D), which all expressed mature markers (Plp1, Mog, Mag, Opalin), along with cluster-specific markers (MOL_A: Il33low; MOL_B: Il33high and Ptgdshigh; MOL_C: increased expression of Selenop, Spock3, Anln; and MOL_D: Klk6, S100b, Anxa5), which have been defined in previous studies (Fig. 2b)20,21. MOL_B and MOL_C constituted the majority of adult mouse corpus callosum OLs, accounting for 75–80% of MOLs at baseline, consistent with a previous report of low Klk6 expression in the adult mouse corpus callosum22. Additionally, we identified three MOL subclusters that express genes associated with previously described DAO states, including Serpina3n, C4b, and H2-D1 (MOL_E and F) and Cdkn1a, Ddit3, and Nupr1 (MOL_G) (Fig. 2b and Supplementary Fig. 4c)9,10,16,21. CPZ induces a distinct transcriptional state in oligodendrocytes. a UMAP embedding of reclustered mouse oligodendrocyte lineage cells. Inset shows PAGA connectivity with edge weights >0.4. b Dot plot of lineage- and cluster-specific marker genes. c Differentially abundant cell neighborhoods overlaid on the UMAP embedding, where each node represents a nhood colored by the logFC (spatial FDR < 0.1). d Bar chart showing the proportion of increased, decreased or unchanged nhoods annotated by cell type. Significance annotations: *p < 0.05, **p < 0.01 (two-sided Fisher’s exact or chi-squared test, results in Supplementary Table 3). e Dot plot showing Hallmark pathway enrichment derived from differential expression testing of pseudobulk MOLs. Dot color is the NES and dot size indicates –log10(padj). f Venn diagram illustrating shared and condition-specific significantly upregulated genes (logFC ≥1.5, padj < 0.05). g Heatmaps showing cytokine-associated gene expression (left) and MYC V1, p53 pathway, and UPR-associated genes (right), colored by the average log₂-transformed CPM expression. Black tiles in annotation panels indicate pathway membership, blue asterisks indicate significantly upregulated genes (fold change ≥1.5, padj < 0.05), and red asterisks indicate significantly downregulated genes (fold change ≤ –1.5, padj < 0.05), (*padj < 0.05, **padj < 0.01, ***padj < 0.001). Representative RNAscope images of mouse corpus callosum showing Cdkn1a (h), Nupr1 (i) or both Cdkn1a and Nupr1 (j) expression. LPC timepoints: baseline = saline at 7 dpl, demyelination = LPC at 7 dpl, remyelination = LPC at 28 dpl. CPZ timepoints: baseline = normal diet, demyelination = 5 weeks CPZ, remyelination = 5 weeks CPZ + 3 weeks recovery (h, i) and saline or LPC-inject at 2 dpl (j). Yellow arrows indicate positive oligodendrocytes. Scale bars = 10 μm. Bar charts showing: the percentage of Cdkn1a+ OLs (**p = 0.0011, ****p < 0.0001) for n = 4 (7 dpl, CPZ) and n = 5 (28 dpl) animals (k); Nupr1+ OLs (****p < 0.0001) for n = 4 (7 dpl, 28 dpl, CPZ_5wks, CPZ_5+3wks) and n = 5 (CPZ_baseline) animals (l); and the percentage of either Cdkn1a+ or Nupr1+ Olig2+ cells at 2 dpl from n = 5 animals (m). Bars represent group means, dots indicate individual animals, error bars are the SEM, and statistical significance was determined by one-way ANOVAs with Tukey’s multiple comparisons test (k, l) and unpaired t-tests with Welch’s correction (m). OPCs oligodendrocyte progenitor cells, COPs committed oligodendrocyte precursors, NFOLs newly formed oligodendrocytes, MFOLs myelin-forming oligodendrocytes, DAOs disease-associated oligodendrocytes, PAGA partition-based graph abstraction, DA differential abundance, logFC log fold-change, demyel demyelination, remyel remyelination, nhood neighborhood, GSEA gene set enrichment analysis, NES normalized enrichment score, padj adjusted p value, DEGs differentially expressed genes, TNFα tumor necrosis factor alpha, IFNγ interferon gamma, IFNα interferon alpha, UPR unfolded protein response, CPM counts per million, dpl days post lesion, wks weeks, SEM standard error of the mean. We hypothesized that DAO states would be selectively enriched during demyelination in a model-specific manner. To quantify differential abundance (DA), we applied Milo, which models partially overlapping local cellular neighborhoods (nhoods) on a k-nearest neighbor graph23. Both LPC and CPZ induced widespread MOL depletion and a moderate increase in progenitor cells during demyelination (Fig. 2c and Supplementary Data 3). To directly compare CPZ and LPC cellular composition, we performed a separate statistical test comparing the distribution of nhoods. We aggregated counts of upregulated, downregulated, and non-significant nhoods by cell type and condition, and performed statistical testing separately for demyelination and remyelination timepoints (Supplementary Data 3). We found that LPC-induced demyelination preferentially enriched DAO states MOL_E and F, whereas CPZ-induced demyelination resulted in an accumulation of the DAO MOL_G state (Fig. 2d). Interestingly, LPC samples at remyelination showed near-complete recovery of homeostatic MOL populations, whereas CPZ samples exhibited persistent MOL depletion, and both remyelination conditions exhibited a sustained increase in MOL_E DAO cells. To define the transcriptional programs underlying LPC- and CPZ-induced demyelination MOL states, we performed pseudobulk differential expression analysis comparing all MOLs at demyelination and remyelination to their respective baselines (Supplementary Data 4). CPZ remyelination samples exhibited high inter-sample heterogeneity and were excluded from this analysis (Supplementary Data 4). Similar numbers of DEGs were induced by LPC across de- and remyelination timepoints (902 and 1044 genes, respectively), whereas CPZ demyelination induced a substantially larger transcriptional shift (4230 DEGs) indicating a more profound reprogramming of MOL identity (Supplementary Fig. 4d). Gene set enrichment analysis (GSEA) using a ranking metric on all genes tested in the DEG analysis identified TNFα signaling via NFκB and interferon (IFN) pathways as shared responses across conditions, with shared upregulation of genes Gadd45a/b, Stat1/3, B2m, H2-D1, and Socs3 (Fig. 2e–g ad Supplementary Fig. 4e). LPC-induced MOL DEGs were particularly enriched for IFNγ and IFNα signaling, marked by upregulation of Irf1, Ifi27, Ifi35, and Stat2. In contrast, CPZ demyelination uniquely enriched Myc target genes (Eif4e, Eif4a1, Odc1), DNA damage and repair genes (Tp53, Ddb1, Pcna), and unfolded protein response (UPR) signatures (Atf4, Hspa5) (Fig. 2e–g). These signatures indicate increased protein synthesis and dysregulated genotoxic and ER stress responses, consistent with oxidative injury and metabolic toxicity previously described in CPZ-treated OLs24,25. To validate the CPZ-specific DAO signature in situ, we performed RNAscope for two highly upregulated MOL_G genes Cdkn1a and Nupr1. Cdkn1a encodes p21, a p53-dependent kinase inhibitor that is induced in response to DNA damage, and Nupr1 encodes a transcriptional regulator involved in the oxidative stress response, DNA damage, and ferroptosis. Both transcripts were robustly induced in OLs after 5 weeks of CPZ demyelination (5 weeks) and returned to baseline following 3 weeks of recovery (5+3 weeks) (Fig. 2h–k). In contrast, neither Cdkn1a nor Nupr1 expression was significantly altered in OLs at 7 days post-lesion (dpl) or 28 dpl (demyelination and remyelination, respectively) following LPC injection, compared to saline-injected controls (Fig. 2h–k). Given that CPZ induces a more continuous and protracted OL loss over weeks, we examined whether an early stress response might occur in LPC-induced lesions. However, analysis of LPC-injected brains at 2 dpl revealed no detectable expression of Cdkn1a or Nupr1 (Fig. 2l, m). Together, these data demonstrate that CPZ demyelination induces a distinct, stress-associated DAO state characterized by DNA damage and UPR activation that is not observed in the inflammatory LPC model. At remyelination stages, LPC and CPZ converge onto a shared DAO state that upregulates immune-related genes Our analysis of MOL state dynamics across demyelination and repair revealed that the MOL_E DAO cluster remains elevated in both LPC and CPZ models. To validate the presence of this shared DAO state, we performed RNAscope for two genes highly enriched in MOL_E: Socs3 and B2m. Socs3 encodes a suppressor of cytokine signaling (SOCS) that negatively regulates immune responses and inflammation, and B2m encodes the invariant light chain of the class I major histocompatibility complex (MHC-I). We observed significant upregulation of both Socs3 and B2m in OLs of LPC-injected and CPZ-fed mice relative to their respective controls (Fig. 3a–d). The proportion of Socs3+ and B2m+ OLs in LPC remained similar from demyelination to remyelination, whereas the percentages gradually increased from CPZ demyelination to remyelination (Fig. 3b, d). While B2m+ OLs did not significantly differ between 5 weeks and 5 + 3 weeks (p = 0.08), these trends suggest progressive emergence of immune-associated OL programs during recovery. We also examined the spatial distribution of DAOs, particularly in the CPZ model, which induces diffuse demyelination. Cdkn1a+ MOLs were strongly induced at 5 weeks of CPZ exposure and resolved by 5 + 3 weeks, localizing predominantly to the corpus callosum with minimal presence in deep grey matter and striatum (Supplementary Fig. 5a). In contrast, Socs3+ DAOs were sparse at demyelination but increased during remyelination, distributing broadly throughout the white matter in parallel with widespread Mbp loss and Iba1+ microglial accumulation (Supplementary Fig. 5b). No significant medial–lateral differences in DAO abundance were detected in CPZ lesions (Supplementary Fig. 5c). In the focal LPC model, Socs3+ DAOs were confined to the lesion core, appearing at low frequency at 7 dpl and increasing by 28 dpl (Supplementary Fig. 5d). A convergent oligodendrocyte state expressing immune-related genes emerges during remyelination in both CPZ and LPC models. Representative RNAscope images of mouse corpus callosum showing Socs3 (a) or B2m (b) expression. LPC timepoints: baseline = saline at 7 dpl, demyelination = LPC at 7 dpl, remyelination = LPC at 28 dpl. CPZ timepoints: baseline = normal diet, demyelination = 5 weeks CPZ, remyelination = 5 weeks CPZ + 3 weeks recovery. Yellow arrows indicate positive OLs. Scale bar = 10 μm. Bar charts showing the percentage of Socs3+ OLs (*p = 0.0125 for 28 dpl, *p = 0.0273 for CPZ) for n = 4 (7 dpl, CPZ_baseline), n = 6 (28 dpl), and n = 5 (CPZ_5wks, CPZ_5+3 weeks) animals (c); and B2m+ OLs (28 dpl: **p = 0.0037, CPZ: **p = 0.0057) for n = 4 (7 dpl, CPZ_5 weeks, CPZ_5+3 weeks), n = 6 (28 dpl) and n = 5 (CPZ_baseline) animals (d). Bars show group means and error bars represent the SEM. Statistical significance was assessed by one-way ANOVA with Tukey’s multiple comparisons test. e Schematic showing the timeline used for the EdU labeling experiment. Created in BioRender. Aboelnour, Erin. (2026) https://BioRender.com/hf51lrm. f Bar chart showing the percentage of EdU+ cells that are OPCs in the corpus callosum of LPC-injected animals at 28 dpl (n = 3 animals) and in CPZ-fed animals at 5 + 3 weeks (n = 5). g Bar chart showing the percentage of EdU+ cells that express Olig2 in the corpus callosum of LPC-injected and CPZ-fed mice (n = 3 animals for 28 dpl, n = 4 for 42 dpl, n = 5 for CPZ_5+3 weeks, n = 4 for CPZ_5+5 weeks). h Representative RNAscope images of mouse corpus callosum showing Socs3 expression in EdU+Olig2+ cells. Yellow arrows indicate triple-positive cells. Scale bar = 10 μm. i Bar chart showing the percentage of newly generated OLs that express Socs3 in the corpus callosum of LPC-injected and CPZ-fed mice (n = 3 animals for 28 dpl, n = 4 for 42 dpl, n = 5 for CPZ_5+3 weeks, n = 4 for CPZ_5+5 weeks). Statistical significance was assessed by unpaired t-tests with Welch’s correction (*p = 0.004121) (f, g, i). OLs oligodendrocytes, OPCs oligodendrocyte progenitor cells, LPC lysophosphatidylcholine, CPZ cuprizone, dpl days post lesion, EdU 5-ethynyl-2′-deoxyuridine, SEM standard error of the mean, IP intraperitoneal. As DAO populations are absent under homeostatic conditions, we reasoned that they arise either from stressed mature OLs or from newly generated OLs adopting a DAO-like transcriptional program. Supporting the latter possibility, MFOLs displayed direct connectivity to MOL_E in PAGA analysis (Fig. 2a), suggesting that demyelination alters OL maturation trajectories rather than inducing an irreversible damaged fate. To determine if immune-associated DAO markers are expressed by newly generated OLs during remyelination, we performed EdU labeling during demyelination and analyzed brains at two remyelination stages for each model (Fig. 3e). The majority of EdU-labeled cells expressed the OL lineage marker Olig2, with minimal expression of the OPC marker Pdgfra, confirming successful labeling of newly generated OLs (Fig. 3f–h). Strikingly, the majority of Olig2+EdU+ cells expressed Socs3 in both models (Fig. 3i). Interestingly, Socs3 expression persisted in newly generated OLs up to 42 days post LPC injection (5 weeks after our demyelination timepoint), but declined significantly 5 weeks after CPZ withdrawal (Fig. 3i), indicating subtle differences in the temporal resolution of immune-associated OL states between models. Across both LPC and CPZ models, MOLs upregulate a conserved interferon- and stress-associated gene signature, including MHC-I antigen presentation genes (H2-D1, H2-K1, B2m, Tap2, Nlrc5), cytokine signaling mediators (Stat1/2/3, Irf1/3/9), and interferon-stimulated genes (Ifi27, Rtp4, Zc3hav1). This state is accompanied by upregulation of complement and glial stress markers (Serpina3n, C4b, Gadd45a/b, Socs3), and cytoskeletal remodeling factors (Vim, Anxa2, Cd63), suggesting an active, adaptive response of MOLs to inflammatory cues in a manner that resembles that observed in experimental autoimmune encephalomyelitis (EAE). Notably, previous studies have described upregulation of immune-related genes within OL lineage cells in both the EAE mouse model and postmortem MS brain tissue11,16,26. To directly compare these responses, we examined the overlap between our DEG analysis in LPC- and CPZ-induced demyelination with a previously published DEG list generated from scRNA-seq of OL lineage cells isolated from the spinal cords of EAE mice (Supplementary Fig. 6a–c and Supplementary Data 4)26. We found that 27% and 22% of upregulated EAE DEGs overlapped with LPC and CPZ MOLs, respectively (Supplementary Fig. 6d, e). In contrast, CPZ and EAE shared a substantially higher number of downregulated DEGs (23%) compared to LPC (7.5%). Both EAE and CPZ MOLs downregulated metabolic and membrane maintenance genes (Slc1a3, Abhd12, Gprc5b, Stard5) and structural or adhesion factors (Ncam1, Dclk1, Tspan15), suggesting destabilization of the mature OL phenotype and reduced metabolic resilience in CPZ and EAE. Interestingly, both EAE and CPZ MOLs also downregulated expression of epigenetic modifiers (Dnmt3a, Hdac11, Tet1), potentially contributing to the broader transcriptional remodeling observed in EAE and CPZ relative to LPC. Collectively, these data indicate that while LPC and CPZ induce demyelination through distinct mechanisms, both models converge during remyelination onto a shared immune-associated DAO state characterized by interferon signaling and MHC-I expression. Importantly, CPZ initially induces a unique stress-associated DAO program driven by metabolic and genotoxic injury, yet ultimately resolves to produce the same immune-like MOL fate observed in LPC. These findings demonstrate that immune-reactive OLs arise in toxin-induced demyelination, providing opportunities to explore how demyelination induces immune responses in OLs in the mammalian brain. CPZ demyelination more faithfully recapitulates the DAO state observed in MS lesions than LPC The identification of distinct DAO states in mouse models prompted us to investigate whether analogous states are present in MS lesions. We subclustered and reannotated the human OL lineage cells from the snRNA-seq MS WM dataset in Fig. 1 (Fig. 4a and Supplementary Fig. 7a). Using established marker expression, we identified human OPCs, COPs, and 4 MOL populations (MOL_A–D) (see “Methods”, Fig. 4b and Supplementary Fig. 7b). Human MOL states were defined by cluster-specific markers (MOL_A: RBFOX1high and SLC5A11high; MOL_B: expresses MOL_A markers, CTNND2low and PLXDC2low; MOL_C: loses RBFOX1 and SLC5A11, expresses CTNND2high and PLXDC2high and gains OPALIN and LAMA2 expression; and MOL_D: similar to MOL_C but increased FTLhigh, CSPRP1high). Additionally, we identified clusters that express canonical DAO-associated markers (human MOL_E–G). Specifically, human MOL_E expresses high levels of HSPH1, FOS, and JUN, while human MOL_F is marked by CDKN1A, NEDD4L, and GADD45B expression, and human MOL_G lowly expresses GPC5 and CD63. We validated our DAO annotation using a gene signature from a previous MS study, showing consistent enrichment of DAO genes in human MOL_E–F and OPC clusters (Supplementary Fig. 7c)9. PAGA connectivity analysis revealed a primary differentiation trajectory from OPCs to COPs, and sequential progression from MOL_A to B, C, and D, with additional connectivity linking the DAO states, consistent with stress-induced divergence from the homeostatic lineage path (Supplementary Fig. 7d). Mouse demyelination models share transcriptional phenotypes with disease-associated oligodendrocytes in human MS lesions. a UMAP embedding of reclustered human oligodendrocyte lineage cells. b Dot plot of lineage- and cluster-specific marker genes. c Differentially abundant cell neighborhoods overlaid on the UMAP embedding (top), where each node represents a nhood colored by the logFC (spatial FDR < 0.1). Bar charts (bottom) showing the proportion of increased, decreased or unchanged nhoods annotated by cell type, compared to NAWM. Significance annotations: *p < 0.05, **p < 0.01 (two-sided Fisher’s exact or chi-squared test, results in Supplementary Table 5). d Dot plot showing enrichment of Hallmark pathways calculated on differential expression testing of pseudobulk MOLs per lesion type. Dot color is the NES and dot size reflects –log10(padj). e Bar charts showing overlaps of upregulated (fold change ≥1.5, padj < 0.05) and downregulated (fold change ≤–1.5, padj < 0.05) DEGs in LPC and CPZ at demyelination compared with MS lesion types (fold change 1.5, FDR < 0.05). Hashed regions represent genes shared between LPC, CPZ and the indicated lesion type. Genes in side panels occur in the overlap of at least one MS lesion type with either LPC (orange box) or CPZ (magenta box). The green box marks genes upregulated across LPC, CPZ, and at least one MS lesion type. f UMAP plots showing mouse and human OL lineage cells, and gene scores derived from CPZ-associated and remyelination-associated mouse DEGs. OLs oligodendrocytes, OPCs oligodendrocyte progenitor cells, MOLs mature oligodendrocytes, DA differential abundance, UMAP uniform manifold approximation and projection, NES normalized enrichment score, DEGs differentially expressed genes, WM white matter, NAWM normal-appearing white matter, LPC lysophosphatidylcholine, CPZ cuprizone, MS multiple sclerosis, padj adjusted p-value. To quantify changes in OL lineage composition across MS conditions, we performed DA testing using Milo (Supplementary Data 5). Homeostatic MOL_A–D populations were largely unchanged in NAWM compared to control WM. In contrast, MS lesions exhibited variable but significant alterations in homeostatic MOL states, with the fewest changes observed in RLs (Fig. 4c). All MS lesion types were strongly enriched for human DAO states MOL_F and G, though MOL_F is notably less abundant than MOL_G (mean of 33 cells per sample for MOL_F, compared to mean of 245 for MOL_G) (Fig. 4c and Supplementary Fig. 7e, f). Interestingly, MOL_E, a comparatively small cluster, was highly enriched in two control WM samples (44% of cells), both from 85-year-old individuals (Supplementary Fig. 7g). Human MOL_E shares gene expression changes with the MS-enriched DAOs, suggesting that physiological aging may induce a stressed state in MOLs resembling demyelination-associated transcriptional changes. To assess transcriptional alterations across MS lesion types, we performed pseudobulked differential gene expression analysis comparing MOLs from each MS lesion type to control WM (Supplementary Data 6). CALs displayed the greatest transcriptional perturbation (1629 genes), whereas NAWM exhibited the fewest changes (328 genes), with substantial overlap across lesion types (Supplementary Fig. 8a, b). Pathway enrichment analysis revealed consistent upregulation of TNFα signaling (NFKB1A, BIRC3, CEBPD), p53 pathway (IER5, CDKN2A, CDKN2B) and IFNα response (IRF7) across multiple MS lesion types (Fig. 4d). These inflammatory and stress-response pathways mirrored those observed in mouse demyelination models. CIL and RL lesions shared additional pathways, including TGFβ and PI3K/AKT/mTOR signaling, suggesting acquisition of lesion-stage-specific remodeling programs. To directly compare the mouse and human OL lineage cells, we applied MetaNeighbor analysis. Cell types clustered into two major cross-species groups corresponding to progenitors and MOLs (Supplementary Fig. 8c). Within the MOL branch, DAOs formed a distinct shared subcluster separate from homeostatic MOLs, indicating conserved transcriptional identities across species. To better understand which genes drive the similarity between the DAOs, we next quantified shared DEGs between mouse toxicity models and human MS lesions. LPC demyelination MOLs shared 22–47 upregulated DEGs with MS MOLs, whereas CPZ demyelination MOLs shared 25–60 (Fig. 4e and Supplementary Fig. 8d). In particular, multiple mouse DAO genes were upregulated in CIL and/or RL lesions, including SOCS3, CDKN1A, FOS, JUNB, and GADD45A/B (Supplementary Fig. 8e). NUPR1 was also upregulated in CALs, but did not reach our significance threshold (log2FC = 1.48, FDR = 0.0703). Interestingly, B2M was not significantly upregulated by MOLs in any MS lesion type, however it has been shown to be upregulated in other MS datasets11–13. Intersection of all three datasets identified 37 genes robustly upregulated regardless of demyelination mechanism, which were enriched for cytoskeletal remodeling and stress-response programs. There were 41 genes uniquely overlapping at least 1 MS lesion type and LPC MOLs at demyelination, including genes with roles in neuroprotection (GDNF) and lipid metabolism and cholesterol trafficking (APOE), glutamatergic and axon-glial ephrin signaling genes (GRM8, EPHA4), interferon signaling and MHC-I regulation (IRF1, NLRC5) and ECM and cell adhesion molecules (TNR, CNTN3, ITGA8). The CPZ-specific overlap included 106 genes and was dominated by stress-response transcription factors (KLF15, KLF9, KLF10, MAFK, CEBPG, JUNB and FOS). The CPZ set contained several genes involved in OL fate decisions and differentiation arrest (ID4, NGFR, EGR1, SESN2, RUNX2), and the anti-apoptotic factor BCL2. When we examined the number of overlapping downregulated genes, we found that MOLs in CPZ demyelination share substantially greater overlap (137 genes) with MS lesions than LPC (7 genes) and only 12 in common across all 3 states (Fig. 4e). The CPZ-specific overlapping downregulated genes include key mediators of OL homeostasis and myelin integrity, such as MOG, MAG, OMG, PLLP, and GJB1. Additional MS/CPZ-specific downregulated genes further support broad impairment of OL function, including the neuregulin co-receptor ERBB3 (OL survival and differentiation), the citrullinating enzyme PADI2 (myelin basic protein post-translational modification), lipid biosynthesis genes ST3GAL5 and LPCAT2, the pro-survival receptor LPAR1, and the differentiation-associated transcription factor E2F3. Downregulation of myelin structural and metabolic genes suggests that both CPZ-induced and MS-associated DAOs are functionally compromised in maintaining stable myelin. Finally, to examine whether distinct mouse DAO states correspond to human DAOs, we generated gene signatures representing the CPZ-specific DAO state (mouse MOL_G) and the remyelination DAO state (mouse MOL_E). The CPZ signature was strongly enriched in human DAO clusters, supporting the conclusion that CPZ demyelination closely models the transcriptional state of lesion-associated DAOs (Fig. 4f). The CPZ-associated gene signature includes stress response genes (ATF4, DDIT3, NUPR1), hypoxia and injury response genes (VEGFA, SERPINE1), and pro-inflammatory mediators (CD44, LGALS3). The mouse remyelination signature was also enriched in human DAO clusters, albeit more modestly, and comprised negative regulators of immune signaling (SOCS3, BCL3, TNFRSF1A, GADD45B/G) and membrane/ECM remodeling genes (CD63, ANXA2, SNX10). Heatmaps of the gene scores showed the expected enrichment of the CPZ signature in MOL_G and the remyelination signature in MOL_E, respectively, and enrichment in human DAOs, especially in MOL_F (Supplementary Fig. 8f). Collectively, these analyses demonstrate that both mouse demyelination models recapitulate key features of DAOs in MS lesions, including increased TNFα and interferon signaling pathways. However, CPZ demyelination more faithfully recapitulates the human DAO transcriptional program, particularly the downregulation of myelin and homeostatic genes. Importantly, these findings establish SOCS3 and CDKN1A as conserved elements of the DAO state across species, warranting further mechanistic investigation into their roles in OL pathology. Mouse OPCs express DAO genes after demyelination but do not capture the heterogeneous OPC response across MS lesions Given the central role of OPCs in remyelination, we next examined their transcriptional response to demyelination. In addition to differentiating into new OLs, OPCs induce immune activation and participate in debris phagocytosis26–28. We analyzed gene expression changes in mouse OPCs at the LPC-induced demyelination timepoint, but due to insufficient cell numbers in CPZ demyelination (fewer than 30 cells per replicate) a robust pseudobulk analysis could not be performed for CPZ (Supplementary Data 4). At LPC demyelination, OPCs exhibited substantial transcriptional remodeling, with 553 upregulated and 1012 downregulated genes (Supplementary Fig. 9a). Many altered genes and enriched pathways overlapped with DAO programs from both LPC and CPZ models (Supplementary Fig. 9b). Upregulated genes included immune-related and interferon response factors (Ifit1, Stat2, Socs3, and C4b), as well as mediators of cell cycle arrest and DNA damage response (Gadd45b, Gadd45g). Interestingly, one of the most strongly downregulated genes we observed was Mertk, which could indicate that OPCs are suppressing their phagocytic function in favor of differentiation. As we previously observed enrichment of our DAO gene signatures in OPC clusters (Fig. 4f), we next examined the expression of DAO-associated genes in OPCs. Using RNAscope, we quantified the percentage of Pdgfra+Olig2+ OPCs expressing Cdkn1a, Nupr1, Socs3, or B2m following LPC and CPZ demyelination. We found that Cdkn1a was significantly upregulated in OPCs at demyelination timepoints in both models and returned to baseline during remyelination (Supplementary Fig. 9c, d). Conversely, Nupr1 expression did not significantly change in OPCs, although it did slightly increase at CPZ demyelination (Supplementary Fig. 9e). In contrast, both Socs3 and B2m expression increased in OPCs across both models and remained elevated at remyelination timepoints (Supplementary Fig. 9f–h). Together, these results indicate that OPCs adopt a stress- and immune-associated transcriptional state that partially overlaps with DAO phenotypes in both mouse models, although the functional consequences of this state for OPC differentiation and survival remain to be determined. We next examined differential gene expression in human OPCs across MS lesions (Supplementary Data 7). OPCs in CALs displayed the greatest number of DEGs (150 upregulated and 231 downregulated genes), whereas OPCs in NAWM and ALs showed comparatively fewer changes (Supplementary Fig. 10a). Interestingly, pathway analysis revealed marked heterogeneity in the OPCs across lesion types. OPCs in CALs exhibited only downregulation of pathways, including cholesterol homeostasis, interferon signaling, and fatty acid metabolism, whereas OPCs in RLs showed upregulation of hypoxia, glycolysis, and TNFα signaling pathways (Supplementary Fig. 10b, c). These results indicate distinct metabolic and inflammatory states of OPCs depending on lesion stage. This heterogeneity was further reflected in limited overlap of DEGs across lesion types. Only 8 upregulated and 10 downregulated genes were shared between OPCs in CALs and RLs (Supplementary Fig. 10d). The greatest DEG overlap occurred between CIL and RLs (38 upregulated and 16 downregulated genes), suggesting closer transcriptional similarity between these states. Interestingly, human OPCs in CILs and/or RLs upregulated several genes also induced in mouse OPCs following demyelination, including SOCS3, TGFB2, GADD45B, and SNX10 (Supplementary Fig. 10e). Consistent with this overlap, pathway enrichment in mouse OPCs and human OPCs in RLs showed concordant activation of Myc targets, IL6/JAK/STAT signaling, and TNFα pathways. In contrast, OPCs in ALs and CALs exhibited distinct or even opposing pathway regulation. Collectively, these findings demonstrate that OPC responses in MS are strongly lesion-stage-dependent. Importantly, human OPCs in CIL and RL samples share transcriptional features with demyelination-induced mouse OPCs, including activation of stress and inflammatory programs, suggesting partial conservation of injury-associated OPC states across species. LPC and cuprizone demyelination induce transcriptionally similar DAM states that persist through remyelination In addition to OL lineage alterations, demyelination induced dramatic changes in microglia abundance and state (Fig. 1d). We therefore asked whether LPC and CPZ demyelination elicit distinct microglia responses. Reclustering of mouse microglia and macrophages identified microglia subclusters (Mg_A–E), PVMs, and a cycling population (Fig. 5a and Supplementary Fig. 11a). Using a previously defined innate immunity gene set, we calculated a DAM score, which was enriched in the highly related Mg_C–E clusters, PVMs, and cycling cells (Fig. 5a, Supplementary Fig. 11b)29,30. These clusters expressed canonical DAM markers (Apoe, Cd63, Lyz2) and downregulated homeostatic microglia genes (Tmem119 and P2ry12) (Fig. 5b and Supplementary Fig. 11c). PVMs were distinguished by expression of lineage markers C1qa and Trem2, and lack of microglia-specific genes such as Tmem119 and Cx3cr1. Notably, PVMs shared expression of DAM markers and were enriched for immune signaling genes Igqap1, Ms4a7, and Msrb1. Microglia share conserved gene expression changes in mouse demyelination models, but LPC leads to increased inflammation. a UMAP embedding of reclustered mouse macrophage/microglia lineage cells. Inset shows a gene score for innate immunity-associated genes. b Dot plot of lineage- and cluster-specific marker genes across cell types. c Differentially abundant cell neighborhoods overlaid on the UMAP embedding, where each node represents a nhood colored by the logFC (spatial FDR < 0.1). d Bar chart showing the proportion of increased, decreased or unchanged nhoods annotated by cell type. Significance annotations: *p < 0.05, **p < 0.01 (two-sided Fisher’s exact or chi-squared test, results in Supplementary Table 8). e Representative IHC images (left) showing Cd31 and Cd206 expression in the mouse corpus callosum. Scale bar = 10 μm. Bar chart (right) showing the number of Cd206+ PVMs per Cd31+ vascular volume from n = 7 (7 dpl), n = 4 (CPZ_baseline) and n = 5 (CPZ_5 weeks) animals. Bars show group means and error bars indicate the SEM. Statistical significance was assessed by one-way ANOVA with Tukey’s multiple comparisons test (*p = 0.0229). f Representative IHC images (left) showing Cd11b and Apoe expression in the mouse corpus callosum. Scale bar = 10 μm. Bar chart (right) showing the percentage of Apoe+ microglia from n = 4 (7 dpl, 42 dpl), n = 3 (CPZ_baseline), and n = 5 (CPZ_5 weeks, CPZ_5+5 weeks) animals. Statistical significance was assessed by one-way ANOVA with Tukey’s multiple comparisons test (****p < 0.0001, *p = 0.0174). g Dot plot showing Hallmark pathway enrichment derived from differential expression testing of pseudobulk MG in LPC or CPZ. Dot color is the NES and dot size indicates –log10(padj). h Venn diagram illustrating unique and overlapping significantly upregulated (fold change ≥1.5, padj < 0.05, left) and downregulated (fold change ≤–1.5, padj < 0.05, right) microglial DEGs at the indicated timepoints. Colored boxes highlight DEGs associated with the indicated overlapping gene sets. i Heatmaps showing LPC remyelination-enriched pathways, colored by the average log₂-transformed CPM expression. Black tiles in annotation panels indicate pathway membership, blue asterisks indicate significantly upregulated genes, and red asterisks indicate significantly downregulated genes as indicated in (h), (*padj < 0.05, **padj < 0.01, ***padj < 0.001). Mg microglia, DAM damage-associated microglia, PVMs perivascular macrophages, DA differential abundance, logFC log fold-change, NES normalized enrichment score, DEGs differentially expressed genes, CPM counts per million, IHC immunohistochemistry, dpl days post lesio, SEM standard error of the mean, LPC lysophosphatidylcholine, CPZ cuprizone, padj adjusted p-value. To compare the distribution of these microglia states between LPC and CPZ demyelination, we performed DA testing using Milo (Supplementary Data 8). LPC demyelination had significantly decreased proportions of homeostatic microglia (Mg_A and B) compared to CPZ demyelination (Fig. 5c, d), suggesting stronger microglial activation following LPC injection. Both LPC and CPZ demyelination increased the proportion of DAMs (Mg_C–E) by similar levels, but LPC induced a significantly higher proportion of cycling microglia (Fig. 5d). In addition, PVMs were only enriched in LPC lesions, which we validated by quantifying Cd206high perivascular cells in healthy and demyelinated WM (Fig. 5e). Cycling microglia and PVMs were restricted to demyelination timepoints and resolved during remyelination. In contrast, DAM clusters persisted during remyelination in both models, but remained significantly elevated in LPC compared to CPZ (Fig. 5d). To determine if DAMs persist long term, we quantified ApoE-expressing Cd11b+ microglia at demyelination and late remyelination stages (42 days post LPC injection, 5 weeks after CPZ withdrawal). ApoE⁺ microglia remained abundant following LPC, whereas expression returned to baseline after CPZ, indicating delayed resolution of microglial activation in the LPC model (Fig. 5f). Importantly, we confirmed reproducible enrichment of DAO and DAM states in LPC and CPZ models, independent of dataset origin or sequencing platform, assessed by integration of each dataset with its respective cohort (Supplementary Fig. 12). We next compared microglia gene expression changes using pseudobulk analysis, comparing de- and remyelination microglia to their respective baseline controls (Supplementary Data 9). PVMs were compared to LPC homeostatic microglia due to their absence in baseline samples. All comparisons yielded extensive transcriptional remodeling, with the largest DEG numbers observed in LPC demyelination microglia and PVMs (Supplementary Fig. 13a). Gene set enrichment analysis identified activation of several pathways in common between LPC and CPZ DAMs/PVMs, including cholesterol homeostasis, epithelial-mesenchymal transition, and reactive oxygen species signaling (Fig. 5g and Supplementary Fig. 13b). LPC remyelination and CPZ demyelination DAMs additionally enriched immune response pathways, including IFNα and IL6 signaling (Fig. 5g). Comparison revealed 136 genes consistently upregulated in DAMs across both models and timepoints, including well-established DAM markers such as Apoe, Lgals3, and Spp1 (Fig. 5h). Approximately 67% of upregulated genes in CPZ DAMs were also induced in LPC DAMs, whereas only 22–27% of LPC-upregulated genes overlapped with CPZ, highlighting a broader LPC-specific inflammatory response. LPC-specific genes included immune-associated mediators such as Gpnmb, Ms4a7, and Mmp12. Downregulated genes were even more conserved, showing 941 genes reduced across all conditions, including core homeostatic microglia genes (P2ry12, Cx3cr1) and inhibitors of phagocytosis (Cd33, Gpr34), indicating a highly shared loss of homeostatic identity across models (Fig. 5h, Supplementary Fig. 13c). Interestingly, LPC remyelination DAMs remained enriched for inflammatory genes (Il1b, Cxcl9, Cd69), complement components (C1qa, C3), and hypoxia/glycolysis-associated genes (Mt1, Slc2a1), suggesting a continued pro-inflammatory microglial state during LPC remyelination stages (Fig. 5i)13,31–36. These findings suggest that, although both models generate a conserved DAM program, LPC lesions maintain a metabolically active and pro-inflammatory microglial state during remyelination. To determine whether microglial responses to toxic demyelination resemble those observed in EAE, we compared gene expression changes in our analysis with a previously published scRNA-seq dataset of microglia isolated from the brains of mice at peak EAE (Supplementary Fig. 13d)37. We subset the data to include only cells without the CD83 knockout, and performed pseudobulk gene expression testing comparing EAE vs control microglia. Pathway analysis of the EAE differential expression results showed strong induction of interferon signaling, Myc targets and metabolic shifts (Supplementary Fig. 13e). Across EAE, LPC, and CPZ conditions, microglia upregulated a conserved inflammatory and metabolic activation program characterized by induction of cytokine and chemokine mediators (Il1b, Cxcl10), MHC-I–associated antigen processing components (Ifi30, Psmb8, Psme1/2, H2-T22), oxidative stress regulators (Cybb, Txn1), and glycolytic enzymes (Ldha, Pkm) (Supplementary Fig. 13f). This shared response was accompanied by downregulation of homeostatic receptors and metabolic regulators, including P2ry13, Gpr34, Adrb2, and Lpin1. In contrast, LPC and CPZ elicited a substantially broader and more pronounced transcriptional remodeling than observed in EAE brain microglia. Toxin demyelination uniquely amplified lipid handling and cholesterol metabolism programs (Apoe, Lpl, Abca1, Npc2), phagolysosomal and proteolytic machinery (Gpnmb, Cd68, Ctsb, Ctsd), and innate immune signaling networks (Tyrobp, Axl, Cd14, Tlr2), while further suppressing canonical microglial identity genes (P2ry12, Tmem119, Cx3cr1, Sall1). Thus, although LPC and CPZ clearly recapitulate the core activation program present in EAE, they drive a quantitatively stronger and qualitatively expanded state marked by enhanced lipid processing, phagocytic engagement, and metabolic reprogramming. This amplified transcriptional response is consistent with the presence of overt demyelination in toxin models and underscores that brain microglia in EAE, at the timepoint examined, exhibit a comparatively restrained activation profile. Altogether, we find that LPC induces a more acute and sustained microglial response characterized by PVM recruitment, increased proliferation, persistent DAM enrichment, and broader inflammatory gene activation. In contrast, CPZ elicits a slower and more transient activation profile, with substantial resolution by remyelination. Despite these kinetic and quantitative differences, both models converge on a conserved DAM transcriptional core defined by upregulation of canonical DAM genes and coordinated downregulation of homeostatic microglial programs. Microglia enriched in MS exhibit heterogeneous states across lesion types To compare the toxin-induced microglia response with human microglia in MS pathology, we reclustered the human microglia and identified seven subclusters (Mg_A–F and PVMs) (Fig. 6a, Supplementary Fig. 14a). Scoring of pro-inflammatory gene sets revealed no enrichment in Mg_A, modest enrichment in Mg_E, and strong enrichment in Mg_F, suggesting heterogeneous activation states across microglia subtypes found in MS (Supplementary Fig. 14b). Human Mg_A expressed high levels of lineage-defining genes and low inflammatory markers and was classified as homeostatic microglia (Fig. 6b and Supplementary Fig. 14c). In contrast, human Mg_B and C retain expression of the homeostatic marker P2RY12 and show modest induction of DAM-associated genes, suggesting they are transitional states. Human Mg_D strongly expressed genes previously used to describe ‘foamy’ microglia, such as lipid storage genes LPL and PPARG and the lysosome gene CTSL, supporting a role for these cells in clearance of myelin debris13. Human Mg_E shared features with Mg_D, but additionally expressed NUPR1, a stress-response gene implicated in ferroptosis38,39. Human Mg_F was transcriptionally distinct, marked by high levels of VEGFA and FOSL2. Finally, the human PVM cluster was defined by specific expression of macrophage markers (F13A1, MRC1, CR1, and LYVE1). Together, these data indicate greater state heterogeneity in MS lesions than we observed in mouse demyelination models. MS lesions enrich distinct microglial states that share core transcriptional features with mice. a UMAP embedding of reclustered human microglia/macrophage lineage cells. b Dot plot of lineage- and cluster-specific marker genes across cell types. c Differentially abundant cell neighborhoods overlaid on the UMAP embedding (top), where each node represents a nhood colored by the logFC (spatial FDR < 0.1). Bar charts (bottom) showing the proportion of increased, decreased or unchanged nhoods annotated by cell type, compared to NAWM. Significance annotations: *p < 0.05, **p < 0.01 (two-sided Fisher’s exact or chi-squared test, results in Supplementary Table 10). d Dot plot showing enrichment of Hallmark pathways calculated on differential expression testing of pseudobulk MG per lesion type. Dot color is the NES and dot size reflects –log10(padj). e Bar charts showing overlaps of upregulated (fold change ≥1.5, padj < 0.05) and downregulated (fold change ≤–1.5, padj < 0.05) DEGs in LPC and CPZ at demyelination compared with MS lesion types (fold change 1.5, FDR < 0.05). Hashed regions represent genes shared between LPC, CPZ and the indicated lesion type. Genes in side panels occur in the overlap of at least one MS lesion type with either LPC (orange box) or CPZ (magenta box). The green box marks genes upregulated across LPC, CPZ, and at least one MS lesion type. f UMAP plots showing mouse and human MG cells, and gene scores derived from hypoxia- and inflammatory response-associated genes. MG microglia, DA differential abundance, UMAP uniform manifold approximation and projection, NES normalized enrichment score, DEGs differentially expressed genes, WM white matter, NAWM normal-appearing white matter, LPC lysophosphatidylcholine, CPZ cuprizone, MS multiple sclerosis, padj adjusted p-value. DA testing using Milo demonstrated widespread deviation from homeostasis in MS lesions (Supplementary Data 10). Although NAWM microglia exhibited state shifts, consistent with previous reports showing reactive microglia throughout the CNS in MS patients, changes were modest compared to the lesion tissues (Fig. 6c)38. All lesion types except for RLs showed significant depletion of homeostatic microglia (Fig. 6c). ALs and CALs were strongly enriched for DAMs (Mg_D and Mg_E), consistent with the accumulation of foamy microglia in regions of high myelin debris. In contrast, CILs and RLs were preferentially enriched for DAM cluster Mg_F, with reduced enrichment of Mg_D and E. Mg_F expresses VEGFA, which may promote angiogenesis and support OPC survival and differentiation, raising the possibility that this state is associated with repair40,41. PVM enrichment was modest and largely restricted to ALs and CALs. Pseudobulked differential gene expression testing comparing MS lesions to control WM revealed extensive transcriptional remodeling in microglia (Supplementary Data 11). NAWM exhibited the fewest changes (369 total DEGs), whereas CAL microglia showed the highest (955 total DEGs) (Supplementary Fig. 15a). Only 27 upregulated and 20 downregulated genes were shared across all lesion types, underscoring lesion type specificity in microglia. ALs and CALs shared an additional 32 upregulated and 62 downregulated genes, whereas CILs and RLs shared 61 upregulated and 17 downregulated genes, suggesting two major activation trajectories corresponding to inflammatory and chronic/repair-associated lesions (Supplementary Fig. 15b). Across multiple lesion types, enriched pathways included UPR (BAG3, DNAJA4, TSPYL2), MTORC1 signaling (HSPD1, HSPA4, SERPINH1), and epithelial-mesenchymal transition (TGFBI, FERMT2, FLNA) (Fig. 6d). Notably, CIL and RL microglia additionally enriched for glycolysis (PKM, PYGL) and hypoxia-related genes (VEGFA, PLIN2, FOSL2, NDRG1), mirroring the metabolic profile observed in LPC remyelination DAMs. Mouse demyelination models recapitulate core microglia activation states observed in MS lesions To directly compare mouse and human microglia, we performed cross-species analyses. MetaNeighbor analysis showed cells segregated into two principal clusters, corresponding to homeostatic microglia and DAMs/PVMs (Supplementary Fig. 15c). Within the activated branch, two subclusters emerged, one containing mouse and human Mg_E/F and PVMs, and another containing mouse and human Mg_D and cycling microglia. Mouse DAMs shared a conserved upregulation of 40 genes with MS-enriched microglia, including canonical DAM-associated genes such as APOE, LGALS1/3, VIM, CD44, and CXCL16 (Fig. 6e), indicating cross-species conservation of fundamental activation programs. The triple-overlap signature prominently enriched for lipid metabolism genes (LPL, FABP5, PLIN2, ABCA1, COLEC12), reflecting the central role of DAMs in supporting lipid handling following demyelination. LPC DAMs shared more upregulated genes with MS lesions (55 uniquely overlapping DEGs at demyelination) than CPZ DAMs (17 unique DEGs), suggesting closer similarity of the LPC model to the MS inflammatory state. The LPC/MS unique genes were enriched for lysosomal biogenesis and lipid degradation programs (GPNMB, NPC2, GALC, CTSZ, TPP1, IGF2R), consistent with active phagocytosis of myelin debris, whereas the CPZ-specific genes suggest a chronic stress and metabolic reprogramming response (CRYAB, CPT1A, FGR, CDKN2B, GPX3). Downregulated gene overlap was comparable across models (106 shared across LPC/CPZ/MS, 58 specific to LPC/MS and 48 specific to CPZ/MS), though notably, CAL microglia showed the most DEGs and the greatest similarity to mouse DAM downregulation patterns (Fig. 6e and Supplementary Fig. 14a, d). The signature across MS and both demyelination models constituted strong suppression of the homeostatic microglial transcriptional program, including downregulation of P2RY12, CX3CR1, ENTPD1 (CD39), BACH2, MEF2A and SALL1. Together, this cross-species comparison reveals conserved remodeling in MS and toxin demyelination models, including upregulation of core DAM signature genes and suppression of homeostatic microglial identity, though LPC notably reflects a shift toward lysosomal myelin debris clearance, whereas CPZ reflects chronic stress and metabolic adaptation. Finally, gene signature analysis further showed that hypoxia- and inflammation-associated programs were most enriched in the human Mg_F state present in RLs, though the inflammatory signature was also modestly enriched in human Mg_D/E and PVMs (Fig. 6f and Supplementary Fig. 15e). In contrast, both signatures were broadly distributed across mouse DAM populations (Mg_C–E, PVMs and cycling cells), rather than restricted to a subtype (Supplementary Fig. 15f). Collectively, these findings indicate that toxin-induced demyelination models capture a conserved microglial activation core characterized by loss of homeostatic identity and induction of shared inflammatory programs. However, MS lesions show greater microglial heterogeneity and activation of pathways not captured in the mouse models. This likely reflects the chronic, spatially heterogeneous nature of human disease compared to acute toxin-induced demyelination. Alternatively, this may suggest that mouse microglia perform multiple functions required to clear myelin debris and create a permissive environment for remyelination, but these functions may be diversified in humans and restricted to specific subtypes, consistent with previous reports of evolutionary divergence in human microglia42. Discussion The use of animal models to study demyelination and remyelination has been critical in advancing our understanding of OL susceptibility in neurodegenerative disease, and has been a valuable tool to address the urgent, unmet need for therapies promoting myelin protection and repair. Toxicity models, including CPZ and LPC, offer precise temporal control over damage and a predictable timeline for repair, making them particularly well-suited for assessing therapies for halting disease progression and promoting remyelination. Numerous preclinical therapies have been identified using these models, with a 2019 report stating that ~23% have entered clinical trials and 5 received approval (Fingolimod, Fumaric acid ester, Glatiramer acetate, Methotrexate, Methylprednisolone)43. However, whether distinct cellular responses are induced by different demyelination paradigms remains unexplored, and these models have not been systematically compared to human patient samples across multiple cell types. Therefore, we analyzed single-cell/nucleus datasets to profile transcriptional changes in LPC- and CPZ-induced demyelination and remyelination, comparing them to the largest MS dataset to date. We found that CPZ demyelination induces a distinct DAO phenotype characterized by a pronounced transcriptional shift not observed following LPC injection. This CPZ-specific state exhibited substantially greater gene overlap with DAOs identified in both EAE and MS lesions compared to LPC-induced changes. Stress-associated pathways enriched in CPZ DAOs—including the unfolded protein response (UPR), MYC targets, and mTORC1 signaling—were similarly elevated in human DAOs, whereas these programs were comparatively weak in LPC-induced cells. In addition, core myelin structural genes were selectively downregulated in CPZ DAOs and MS lesions, reinforcing the disease relevance of this state. Notably, the CPZ-induced DAO program was marked by robust upregulation of Cdkn1a and Nupr1. Cdkn1a (p21) is a mediator of cell cycle arrest following DNA damage and its upregulation is linked to cellular senescence, suggesting that CPZ injury may push OLs towards a senescent-like state44. While the function of Nupr1 in OLs has not been examined, it is known to protect against ferroptosis in cancer cells39,45,46. Markers of ferroptosis have been found upregulated in MS lesions, mostly within microglia42,47–50, but our results suggest a role for Nupr1 within OLs as well. Interestingly, Nupr1 has also been linked with cellular senescence in MS lesions51. Collectively, these data reveal a stress-adapted, senescence-like DAO state specific to CPZ-induced demyelination that recapitulates key transcriptional features of OLs enriched in MS lesions. Targeting senescence-associated or stress-response pathways may therefore represent a strategy to preserve OL function in demyelinating disease. In addition to the CPZ specific programs, our cross-model analysis revealed that both LPC and CPZ converge during remyelination onto a shared state that expresses canonical DAO genes such as Serpina3n and C4b. We found that this remyelination DAO state additionally expresses immune-related genes, including Socs3 and B2m, and emerges in newly generated OLs as a sustained stress-adapted transcriptional state during repair. Notably, many immune-associated genes induced in LPC and CPZ overlap with those identified in EAE-associated OLs, such as H2-D1, B2m, Stat1/2/3, and Gadd45a/b. This convergence across all three mouse models indicates that interferon-responsive and antigen-presentation programs in the OL lineage are not unique to autoimmune demyelination but instead represent a broader injury-adaptive response. Consistent with this, many genes enriched in the remyelination signature overlapped with MS lesions (Supplementary Fig. 8d), including immediate early and interferon-associated genes such as JUNB, FOS, SOCS3, APOE, IRF family members, and GADD45A/B, supporting immune-associated OL states are a conserved feature in human disease. Among the conserved enriched pathways, TNFα signaling emerges as a recurrent inflammatory axis. TNFα activity is elevated in MS lesions, and blockade of soluble TNF improves outcomes in EAE, supporting a potentially detrimental role of excessive TNF signaling during inflammatory demyelination52. However, blocking TNF signaling increases MS risk and can exacerbate disease, underscoring the context-dependent and cell-type-specific complexity of this pathway53,54. These observations highlight the need to resolve how TNF-responsive programs within OLs contribute to either adaptive stress responses or persistent dysfunction during remyelination. The functional consequences of both the de- and remyelination MOL states remain unclear. For example, it is not known if immune-OLs produce functional myelin that matures over time. A recent study demonstrated that even incomplete remyelination following CPZ-induced demyelination is sufficient to restore neuronal function, though the transcriptional state of these cells was not measured2. The mechanisms governing DAO resolution, the contribution of cell–cell crosstalk to its maintenance, whether DAO states actively promote or constrain remyelination, and the precise timing of their resolution– particularly in LPC where the immune-state persisted at 42 dpl– remain important open questions. Furthermore, whether restoring proper MOL heterogeneity is essential for functional tissue regeneration remains unknown, requiring further investigation into how transcriptional phenotypes translate to MOL structure and function. Addressing these questions will be essential for determining whether modulating injury-induced OL stress programs can enhance myelin repair. Interestingly, Socs3 deficiency reduces CPZ-induced demyelination, supporting that modulation of this axis may directly influence OL survival and remyelination efficiency53,54. Sustained Socs3 expression during late LPC remyelination paralleled the prolonged DAM activation we observed (Fig. 5f). Whether this prolonged inflammatory response is beneficial for remyelination remains unclear, but microglia have been previously shown to promote remyelination independent of debris clearance55–57. The upregulation of SOCS3 in human MS lesions further underscores its potential as a regulatory node linking inflammatory signaling to OL fate. As OL survival has been linked as a critical step of remyelination54, determining if ablation of Socs3 after demyelination increases OL survival is an important future step. A key strength of the human MS dataset used in our analysis is the inclusion of multiple lesion types, whereas prior studies have largely focused on CALs. The original report primarily emphasized inter-patient variability rather than systematic comparisons of transcriptional states across lesion classes, and leveraging this broader sampling framework enabled us to examine how oligodendrocyte lineage states differ across lesion contexts14. Although defining mechanistic differences among DAO states will require further study, we observed that DAOs are comparatively conserved across MS lesion types. In contrast to the relative conservation of DAO states, both OPCs (Supplementary Fig. 10b–d) and DAMs (Fig. 6d and Supplementary Fig. 14b) exhibit a greater transcriptional heterogeneity. In ALs and CALs, OPCs downregulate pathways linked to cholesterol homeostasis, peroxisome function, Myc, mTORC1, oxidative phosphorylation, and IFNγ signaling, suggesting a reduced metabolic and biosynthetic state that may reflect quiescence or a protective adaptation to inflammation. In comparison, OPCs in CILs and RLs upregulate inflammatory IL-6/JAK/STAT and TNF-α pathways, and RL OPCs uniquely activate hypoxia and glycolysis pathways, indicative of metabolic reprogramming and emergence of a stress-adapted state. Inflammatory gene induction in OPCs from active MS lesions has previously been linked to differentiation arrest and chronic inflammation, suggesting strategies to ameliorate these detrimental changes may be clinically relevant11,12,27. However, we observe limited transcriptional overlap between mouse and human OPC responses to demyelination, indicating that commonly used toxin models incompletely recapitulate human OPC pathology. One potential explanation is the more advanced age of human OPCs in postmortem tissue. As mouse OPCs undergo transcriptional and functional changes with aging58, it would be interesting to compare OPCs following demyelination in older mice. Across the OL lineage, we therefore conclude that, while DAO states show relative conservation between mouse and human, OPC responses diverge significantly, revealing cell-type-specific differences in the models and human disease. In contrast to these discordant OPC phenotypes, we identify a conserved core microglial response to demyelination characterized by suppression of homeostatic identity and induction of canonical DAM-associated genes such as APOE, LPL, LGALS3, CD44, and CXCL16. This shared signature prominently enriches lipid handling and lysosomal programs, consistent with the need for myelin debris clearance during lesion repair. Thus, toxin-induced demyelination models robustly recapitulate the core transcriptional remodeling observed in MS lesions. However, MS-enriched microglia exhibit substantially greater state heterogeneity than observed in either LPC or CPZ models. Human lesions segregate into distinct activation states associated with inflammatory (AL/CAL) and chronic/repair-associated (CIL/RL) environments. In particular, the Mg_F population enriched in CILs and RLs displays coordinated activation of hypoxia, glycolysis and angiogenic programs, suggesting metabolic adaptation to sustained tissue stress. In mice, however, hypoxia- and inflammation-associated signatures are broadly distributed across DAM populations rather than restricted to discrete subtypes, indicating reduced functional specialization. This compression of states likely reflects the acute and temporally constrained nature of toxin-induced demyelination, in which remyelination occurs within weeks. In contrast, the prolonged and spatially heterogeneous inflammatory milieu of MS may drive progressive diversification of microglial states over years, allowing specialization into metabolically, angiogenically, or stress-adapted phenotypes. Model-specific biases are also evident. LPC-induced DAMs more closely resemble inflammatory MS lesions, with enrichment for lysosomal and lipid degradation pathways consistent with active phagocytosis. CPZ microglia, by contrast, preferentially engage stress-response and metabolic adaptation programs, partially mirroring features of chronic lesions. Furthermore, human microglia do not show evidence of proliferation like we observed in mouse models, aligning with previous findings that microglial populations are largely stable in humans, with an estimated lifespan of ~4 years59. Notably, LPC demyelination induces a PVM response that we do not observe in CPZ, and would thus be the preferred model for distinguishing microglia versus macrophage function across the span of injury and repair. While it remains difficult to definitively identify CNS macrophages versus microglia in postmortem MS tissue, previous work has found evidence of increased macrophage infiltration in active MS lesions, where they are hypothesized to contribute to chronic inflammation60. Thus, while both toxin models capture essential elements of demyelination-associated activation, neither fully recapitulates the spectrum of microglial heterogeneity present in human disease. Together, these findings suggest that murine demyelination models are well-suited to interrogate conserved mechanisms of myelin clearance and core inflammatory remodeling. While no current MS therapies directly target microglia, some drugs with off-target microglial effects show regenerative benefits, suggesting that targeting microglia directly could enhance remyelination59. However, they incompletely model the chronic metabolic specialization and lesion-type-specific diversification that characterize MS microglia. In longer-term neurodegenerative models, such as AD, distinct temporal microglial activation states emerge61. Extending demyelination paradigms to longer-term, repetitive, or age-associated models may better approximate the state complexity observed in human lesions and refine translational interpretation. Altogether, our study provides a systematic cross-model and cross-species framework for interpreting demyelination paradigms in the context of human MS. We demonstrate that LPC and CPZ models do not induce interchangeable cellular states, but instead bias distinct aspects of OL and microglial biology. CPZ more faithfully recapitulates the DAO stress program observed in MS, whereas LPC more closely mirrors inflammatory lesion environments. Both toxin models converge on conserved injury-responsive programs shared with MS, including suppression of homeostatic microglial identity and induction of interferon-responsive and antigen-presentation pathways within the OL lineage. These shared signatures likely represent fundamental mechanisms of myelin injury and debris clearance. However, as neither model fully captures the degree of cellular heterogeneity observed in human lesions, particularly within microglia, model selection should be guided by the specific biological question: CPZ for interrogating OL stress and senescence-like dysfunction, LPC for studying inflammatory dynamics, microglial proliferation, and macrophage contributions. The compressed timeline of toxin-induced injury likely limits state diversification compared to the prolonged and spatially heterogeneous environment of MS. Extending demyelination paradigms to aged, chronic, or repetitive injury models may better approximate the complexity of human disease. Moreover, expanding cross-cell-type analyses—particularly incorporating astrocyte states and cell–cell interaction networks—will be essential to understand how coordinated tissue responses govern successful remyelination versus chronic pathology. Altogether, by defining both the convergent and divergent features of mouse demyelination models relative to MS, this work provides a roadmap for strategically leveraging experimental systems to dissect conserved mechanisms of myelin repair while acknowledging their limitations. Such precision in model selection will be critical for translating mechanistic insights into effective remyelination therapies. Methods All methods were performed in accordance with the guidelines and regulations of Children’s National Hospital and the University of Notre Dame. For generation of the CPZ and LPC snRNAseq datasets, animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of the Children’s National Hospital (protocol #30578). For all other experiments, all animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Notre Dame (protocol #23-01-7578). Experimental procedures Animals Wild-type C57BL/6 mice (JAX strain #000664) were maintained in the animal facility at the Children’s National Hospital and the University of Notre Dame according to the Institutional Animal Care and Use Committee and the National Institutes of Health guidelines. Both male and female mice were analyzed for all experiments. Mice were maintained in the animal facility under a 12 h (h) dark–light cycle, and constant temperature (20–26 °C) and humidity (40–60%). All studies complied with all relevant animal use guidelines and ethical regulations. Demyelination protocols For LPC demyelination, adult (8–12 weeks) C57BL/6 mice were anesthetized with ketamine/xylazine, the skull exposed, and a small hole drilled 1 mm rostral and 1 mm lateral of bregma. The animal was placed in a stereotactic rig and injected with 1 µL of 1% LPC into the corpus callosum of one brain hemisphere at 2 mm depth over 5 min. Saline was injected on the contralateral side using the same steps as above. Mice were sacrificed at 7 dpl, as described below, for snRNA-seq and RNAscope/IHC for the demyelination timepoint, at 28 dpl for RNAscope/IHC at the remyelination timepoint, and at 42 dpl for the late recovery timepoint. For snRNA-seq, CPZ demyelination was conducted by feeding adult (8–12 weeks) C57BL/6 mice 0.2% CPZ pellets (Envigo TD.140803) for 5 weeks. For RNAscope/IHC validation, mice were fed 0.3% CPZ pellets (Envigo TD.140805) for 5 consecutive weeks, and then either sacrificed as described below at the end of 5 weeks for the demyelination timepoints, or returned to normal chow and collected 3 or 5 weeks later for the remyelination and late recovery timepoints. During the study, the lab switched from 0.2% to 0.3% CPZ pellets to improve the efficiency of demyelination. Control mice were fed normal chow and processed as described below. Dissection and single-nucleus preparation For LPC demyelination snRNA-seq collection at 7 dpl, mice were intraperitoneally (IP) injected with neural red dye 2–3 h (h) prior to tissue collection to label the lesion within the corpus callosum. Mice were then anesthetized with isoflurane, perfused with cold PBS for 5 min, and the brain was removed and placed into cold PBS on ice. Coronal slices of the brain were made with a fresh razor blade to identify the lesion, and the region of the corpus callosum containing the lesion was dissected. For LPC controls, the corresponding contralateral saline-injected side was also collected. For CPZ-fed mice, the whole corpus callosum was dissected and collected as above. The dissected tissues were placed in Eppendorf tubes, excess liquid was carefully removed, and the tubes were flash frozen in liquid nitrogen. Frozen tissue was stored at –70 °C. For each condition, nuclei were pooled from 5 female and 5 male mice, with two replicates per condition (LPC: baseline = saline injected, demyelination = 7 dpl, CPZ: baseline = normal chow, demyelination = 5 weeks of 0.2% CPZ). Pooled tissue was transferred to chilled glass dounce with 1 mL of lysis buffer + 1% TritonX100 + 1 μM DTT (Sigma Nuclei Pure Prep Kit, Cat #NUC-201). Samples were dounced on ice 10 times with pestle A, then 10× with pestle B. The suspension was passed through a 40 μm FlowMi cell strainer. Dounces were rinsed with 1 mL of low sucrose buffer and passed through the cell strainer as well. Samples were centrifuged at 300 × g at 4 °C for 5 min in a fixed-angle centrifuge. Supernatant was carefully decanted, and the pelleted nuclei were washed with 1.5 mL of low sucrose buffer, and centrifuged again at 300 × g, at 4 °C, for 5 min. Pelleted nuclei were then resuspended in 500 μL Nuclei wash buffer (PBS + 1% UltraPure BSA + 0.2 U/μL Protector RNase inhibitors) + 900 μL of 1.8 M sucrose cushion buffer, and then carefully layered on top of 500 μL 1.8 M sucrose cushion buffer in a 2 mL low-bind tube. Samples were centrifuged at 13,000 × g, at 4 °C, for 45 min. Supernatant was removed, and nuclei were resuspended in 1 mL of Nuclei wash buffer, passed through a 40 μm FlowMi cell strainer and counted using a Cell Countess II. Approximately 6,000 nuclei were loaded per sample onto the 10x Chromium Controller per the manufacturer’s instructions. Single-nucleus libraries were generated using the 10 × 3’ reagents and protocol (v3.1). Libraries were sequenced on an Illumina NovaSeq 6000, with an average number of 80,000 reads per cell per sample. EdU injection and detection LPC-injected or CPZ-fed animals received IP injection of EdU (Sigma, 900584-50MG) at 50 mg/kg. For LPC animals, we injected EdU at 6, 9 and 12 dpl. For CPZ animals, we injected EdU at 2, 5 and 8 days after returning them to a normal diet. The RNAscope Multiplex Fluorescent Reagent Kit V2 (Advanced Cell Diagnostics #323100) was used to perform fluorescent labeling of the indicated RNA molecules according to the kit manual. RNAscope probes used were: Olig2 (447091-C3) and Socs3 (437741). Following RNAscope, EdU was detected using the Click-iT EdU Cell Proliferation Kit (Thermo, C10637) following the manufacturer’s instructions. Sections were then coverslipped with mounting media (Southern Biotech). Sections were imaged using the Leica Stellaris 8 DIVE confocal microscope and cell counts were analyzed for at least 3 sections per animal. Cells with at least two fluorescent puncta were counted as positive for that probe. RNAscope and IHC of OL lineage Mice were anesthetized using isoflurane and transcardially perfused with cold PBS followed by 4% PFA for 5 min each. The brain was dissected out into cold 4% PFA and post-fixed overnight. The tissue was then cryoprotected in 30% sucrose in PBS. Brains were embedded in OCT and frozen on dry ice. Sections were collected onto slides (12 µm) and stored at −70 °C, or as floating sections (40 µm) and stored in PBS at 4 °C. IHC was performed by thawing the slides (all steps were performed at RT unless otherwise indicated) in a humidifying tray for 5 min, permeabilized for 15 min with 1× PBS + 0.1% T×ween (PBST), and blocked in blocking buffer (PBST + 1% normal donkey serum (NDS)) for 30 min. Primary antibody (rabbit anti-Olig2, Millipore AB9610, 1:300 or goat anti-Olig2, R&D Systems AF2418, 1:100; mouse anti-MBP, Covance SMI-99P, 1:200; rabbit anti-Iba1, Wako 019-19741, 1:200 or chicken anti-Iba1, Aves Labs IBA1-0020, 1:250) was incubated in blocking buffer overnight at 4 °C. Slides were washed 3× for 5 min each with PBST, and incubated with the indicated fluorophore-conjugated secondary antibodies (AlexaFluor) in blocking buffer for 2 h. For tissue with combined IHC and RNAscope, the IHC-stained slides were washed again 3× with PBST, and then post-fixed with 4% PFA for 10 min. Next, the RNAscope Multiplex Fluorescent Reagent Kit V2 (Advanced Cell Diagnostics #323100) was used to perform fluorescent labeling of the indicated RNA molecules according to the kit manual, beginning with protease treatment. RNAscope probes used were: Cdkn1a (408551-C2), Socs3 (437741), Pdgfra (480661-C3), Nupr1 (434811), and B2m (415191-C2). Following RNAscope, sections were coverslipped with mounting media (Southern Biotech). Sections were imaged using the Leica Stellaris 8 DIVE confocal microscope and at least 3 sections per animal were analyzed in FIJI for each stain. Positive cells for each probe were determined by the presence of at least two fluorescent puncta within the Olig2+ DAPI+ nucleus, representing expression above background levels. Olig2+ cells positive for Pdgfra were counted as OPCs, and Olig2+ cells negative for Pdgfra were counted as MOLs. Approximately 400–900 cells were counted per animal, with 3–6 animals analyzed per marker combination. Outliers were identified using the ROUT (robust regression and outlier removal) method with a Q value of 1%. To analyze the spatial distribution of Cdkn1a+ and Socs3+ DAOs, tiled images were obtained at 40× to visualize the corpus callosum and portions of the cortex, septum, and striatum. Positive cells were then labeled with colored dots using FIJI. IHC analysis of microglia/macrophages Mouse brain tissue was collected as described above (see RNAscope and IHC of OL lineage). For analysis of PVMs, 40 µm floating sections were permeabilized for 15 min with 1× PBST, and blocked in blocking buffer for 1 h. Primary antibodies (goat anti-CD206, R&D Systems AF2535, 1:400 and rat anti-Cd31, BD Pharmingen 550274, 1:100) were incubated in blocking buffer overnight at 4 °C. Tissue sections were washed 3× for 5 min each with PBST, and incubated with the indicated fluorophore-conjugated secondary antibodies (AlexaFluor) in blocking buffer for 2 h. Sections were washed again 3× with PBST, and then mounted on slides with Southern Biotech mounting media. For PVMs, the total number of CD206high cells adjacent to CD31+ blood vessels within the demyelinated WM was counted and normalized to the volume of CD31 staining, measured using the surface tool in Imaris software. For analysis of DAMs, 12 µm brain sections on slides were permeabilized for 15 min with 1× PBST, and blocked in blocking buffer for 30 min. Primary antibodies (rabbit anti-ApoE, Cell Signaling 49285 T, 1:100 and rat anti-Cd11b, Biorad MCA74G, 1:250) were incubated in blocking buffer overnight at 4 °C. Tissue sections were washed 3× for 5 min each with PBST and incubated with the indicated fluorophore-conjugated secondary antibodies (AlexaFluor) in blocking buffer for 1 h. Sections were washed again 3× with PBST, and then coverslipped with Southern Biotech mounting media. Sections were imaged using the Leica Stellaris 8 DIVE confocal microscope and cell counts were analyzed for at least 3 sections per animal. For DAMs, the number of ApoE+ Cd11b+ cells was counted in FIJI and plotted as a percentage of total Cd11b+ cells. Statistics Statistical tests and graphs for experimental procedures were generated using Graphpad (version 10.3.0). Differences between values were examined using either a one-way ANOVA with Tukey’s multiple comparisons test or unpaired t-tests with Welch’s correction. The specific test used for each analysis is stated within the figure legends. Data analysis Single-cell and single-nucleus raw data processing Mouse LPC/CPZ Demultiplexed FASTQ files were aligned to mm10 using Cell Ranger (v6.0.1), and published datasets were downloaded from GEO and mapped using Cell Ranger (v7.2.0)8,9. Mapped h5 files were run through CellBender to remove ambient RNA62. Corrected count matrices for each sample were processed using Scanpy tools following recommended best practices for retaining high-quality cells63. Doublet detection and removal were performed using Solo, by first training a scVI model on each batch and then removing predicted doublets64. After doublet removal, data were merged across its publication of origin (i.e., our single-nucleus data, Pandey et al.9, or Shen et al.8). Each of these three datasets was then subset based on quality metrics for that dataset—cells or nuclei were removed if they had fewer than 200 genes identified, more than 20% ribosomal content, or more than 5% mitochondrial reads. Additionally, we filtered the data to remove outliers using the following criteria: the single-nucleus data was filtered to remove samples with more than 15,000 total reads or more than 7000 genes. Pandey et al.9 single-cell data was subset to remove cells with more than 8000 genes per cell, or 50,000 reads per cell. Shen et al.8 was subset to remove cells with more than 6000 genes identified, or more than 20,000 reads per cell. Human MS The MS data quality-control counts matrix was downloaded from Zenodo (DOI: 10.5281/zenodo.8338963), subset to the white matter samples, and converted to an h5ad file14. A major advantage of this dataset is that it contains carefully staged lesions. In brief, the subset data comprises 92 samples collected from 47 individuals, including 13 control (Ctr) and 34 MS patients, categorized by disease progression as relapsing-remitting (RR), secondary progressive (SP), or primary progressive (PP) MS (Supplementary Fig. 2D). The WM-only subset contains 13 controls (WM, mean age at death: 64.8 years, median: 70 years), 17 normal-appearing white matter samples (NAWM, mean: 61.3 years, median: 66 years), 21 active lesions (AL, mean: 51.5 years, median: 47 years), 17 chronic active lesions (CAL, mean: 54.7 years, median: 48 years), 13 chronic inactive lesions (CIL, mean: 65 years, median: 69 years), and 11 remyelinated lesions (RL, mean: 66.2 years, median: 69 years). Samples were further controlled by removing nuclei with greater than 10% mitochondrial counts, more than 10,000 genes per cell, and more than 50,000 total reads. Dataset integration and annotation Mouse LPC/CPZ Each dataset was integrated within its own experimental group (i.e., each data source separately) to obtain harmonized cell type labels using scVI65. Briefly, we subset the data to the top 3000 highly variable genes (HVGs), with the batch key set to each individual batch, and the percent mitochondrial and ribosomal counts as continuous covariates. The latent representation from the trained scVI model was used to calculate the neighborhood graph and then the UMAP embedding. Cell types were labeled by visual inspection of known marker gene expression, using markers calculated with Scanpy’s rank_genes_groups, as well as custom cell type scoring using the marker gene sets detailed in Koupourtidou et al. For full integration of the mouse single-cell and single-nucleus data we used sysVI, which was developed to handle challenging integration tasks, such as between single-cell and single-nuclei data or between species66. For the mouse only integration we used the top 3000 HVGs, using the Scanpy highly_variable_genes function set to “cell_ranger”, batch key set to the system variable (single-cell or single-nuclei), and each of the individual samples as a categorical covariate. We verified marker gene expression was enriched in the cell type labels and that each cell type contained single cells and nuclei. Human MS To match our mouse corpus callosum datasets, we integrated the WM samples using scVI and the top 3000 HVGs, and produced a similar embedding and annotation as the published dataset. For the scVI model, the batch key was set to the sample_id (representing each individual sample), and RNA count and percent mitochondrial count were given as continuous covariates. For training the model we used n_layers = 2, n_latent = 30, and gene_likelihood = “nb”. Mouse EAE (microglia) The mapped files were downloaded from GEO (barcodes/features/matrix), as well as the UMAP and metadata files. Raw gene expression matrices were processed using Scanpy. The four pools were concatenated into a single file, and cell-level metadata and precomputed UMAP coordinates were loaded into the object. Downstream analysis was restricted to the WT cells, removing cells with more than 20% ribosomal counts (47 cells) and removing genes expressed in fewer than 3 cells. Differential abundance analysis Compositional changes in the total UMAP for mouse LPC/CPZ and human MS datasets were quantified using scCODA19. For the mouse integration, the samples were split into LPC or CPZ groups and the respective baseline samples used as the control. Pericytes were used as the reference cell type. For the human data, the WM samples were used as the control, and the cluster “Endo + Peri” was used as the reference cell type. The default FDR of 0.05 was used. Milopy, the Python implementation of the Milo toolset, was used to calculate the DA between neighboring cells in the cell type-specific embeddings23. For the mouse LPC/CPZ OL lineage, neighborhoods (nhoods) were recalculated using the X_sysVI representation and n_neighbors set to 100; for mouse microglia, 300; for human oligodendrocytes, 200; and for human microglia, 100. The make_nhoods function was applied with proportion set to 10% to create nhoods, and the median was inspected to ensure that there were, on average, at least 5 cells per sample per nhood. Pairwise comparisons were performed to calculate DA, and build an nhood_graph for each condition compared to the relevant baseline for mouse samples or WM for human MS lesion samples. The nhoods were annotated by the cell type labels by calculating the fraction of cells in a nhood, if it was more than 60% of a single cell type. For nhoods that were less than 60% a single cell type, the annotation was set to ‘Mixed’. The final calculation and annotation was saved from the milo object, and used for statistical analysis. To compare the distribution of nhoods across conditions, we aggregated counts of upregulated, downregulated, and non-significant nhoods by cell type and condition. For each cell type and significance category, we constructed contingency tables of observed counts and totals, and applied either Fisher’s Exact Test when any expected frequency in the contingency table was <5; otherwise Pearson’s chi-squared test was used. Both tests were two-sided. Resulting p-values were corrected for multiple testing using the Benjamini–Hochberg procedure. For visualization, we generated stacked bar plots showing the proportions of upregulated, downregulated, and non-significant neighborhoods per cell type in each condition. Significant differences between conditions were denoted with asterisks on the bar with higher proportions of significant nhoods that corresponded to adjusted p-value thresholds. Differential expression analysis Mouse LPC/CPZ Differential gene expression was calculated using a pseudobulking approach, summing raw gene expression counts in each replicate for each broad cell type with at least 30 cells. Pseudobulk counts were then subset to test individual pairwise comparisons. Each test was performed separately because not all samples contained single-cell and single-nucleus samples, so batch correction was performed differently. For each pairwise comparison, count matrices were trimmed using the edgeR filterByExpr function with the design matrix as input, and keep.lib.sizes set to FALSE. Normalization factors were calculated using the trimmed mean of M-values (TMM) method. For mixed samples (with cells and nuclei) we used voom with quality weights and duplicate correlation twice, blocked on the library type67. For samples with only single-cell data, we used voom with quality weights and no blocking. Using lmFit, a linear model was fitted that included blocking on the batch key and correlation set to the duplicateCorrelation output. Empirical Bayes (eBays) smoothing was applied, and the results table extracted using topTable with n=Inf to include all genes in the output. The full differential expression output for each cell type tested is included as a supplementary table. Differential testing in the CPZ remyelination timepoint produced poor results due to low cell number (2407 across 3 replicates), small library sizes, and high heterogeneity between replicates. Unless otherwise noted in the figure legend, genes are considered significantly differentially expressed if the adjusted p-value is less than or equal to 0.05, and the absolute logFC is greater than or equal to a 1.5 fold change (corresponding to |logFC| ≥ 0.585). Venn diagrams were made using the matplotlib-venn library, and upset plots were made using UpSetPlot in Python68. Heatmaps were made using ggplot, using the normalized and transformed counts per million (log₂[CPM + 1]) that were averaged across each replicate within a group. Heatmap gene annotations were assigned based on inclusion in the mouse orthology-mapped GSEA Hallmark gene sets. Human MS Pseudobulked samples were prepared as above for the mouse data. Briefly, expression counts in each replicate and each broad cell type with at least 15 cells were pseudobulked, then individual comparisons were subset for downstream analysis. We removed outlier samples with a log library size 3× median absolute deviation below the median, as library size seemed to be one of the strongest drivers of variability. Next, for each comparison, the counts matrices were trimmed using the edgeR filterByExpr function and a design matrix with donor age (as categories of under 50, 50–70 and above 70), sex, and postmortem interval (PMI, as categories of less than 6 h, 6–12 h or more than 12 h), and keep.lib.sizes set to FALSE. The TMM method was used to calculate normalization factors. Differential expression was assessed using generalized linear mixed models (GLMMs) implemented in the glmmTMB R package, assuming a negative binomial distribution (nbinom2) to account for overdispersion. For genes with zero counts in all samples (very few), a count of 1 was added to the sample with the largest library size to enable model fitting. Each model included group (e.g., white matter vs. lesion), age_cat, sex, and PMI as fixed effects, and a random intercept per donor to capture inter-individual variability. For each gene, the model output included the estimated log-fold change for the group of interest, standard error, p-value, convergence status, positive-definite Hessian flag (pdHess), log-likelihood, and AIC. Genes for which the model failed or the group coefficient was not estimable were flagged as errors. P-values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) procedure. A post hoc quality flag was created (status = “OK” for models with convergence = 0 and pdHess = TRUE), and downstream analyses were restricted to these high-confidence models. The final output is a ranked list of genes with robust evidence of differential expression between conditions. Mouse EAE (microglia) Pseudobulk differential expression testing was performed on WT microglia cells as described above. Briefly, transcriptional clusters were simplified such that all non-cycling microglia were included in the pseudobulk per replicate (Gpr34+Siglech+, Ccnd3+MHCII+, Cd83++Ccl4+Lpl+, Cxcl10+, Apoe+MHCII+, and Stmn1+), while cycling cells (Mki67+) were excluded. Only replicates containing at least 30 cells were retained to ensure robust representation, yielding three biological replicates per condition (control vs EAE). Lowly expressed genes were filtered using filterByExpr() from the edgeR package, and counts were normalized using the TMM method. Differential expression between EAE and control samples was assessed after voom transformation, linear modeling with lmFit, eBayes, and statistics were extracted using topTable. Gene set enrichment analysis Fast gene set enrichment analysis (fgsea) was implemented to score pathway enrichment69. For the mouse LPC/CPZ datasets and the EAE microglia, the ranking score was set to the sign of the logFC multiplied by −log10(p), assigning positive values to upregulated genes and negative values to downregulated genes, with larger magnitudes reflecting greater statistical significance. For the mouse EAE oligodendrocyte dataset, the ranking score was set to the logFC, as this is the only available data for the DEG list provided from the original manuscript. For the human MS dataset, p-values of exactly zero were replaced with a minimal value to prevent numerical errors. Then, a Z-score was computed from the unadjusted p-value using a two-tailed test conversion [Z = qnorm(p/2, lower.tail=FALSE) × sign(logFC)]. This calculation incorporates both the significance and direction of change, similar to the mouse dataset. Msigdbr was used to retrieve the mouse or human hallmark gene set as the input pathway70. The output was filtered using minSize and maxSize to include only pathways with more than 15 genes but less than 500 to avoid noise, and for terms with an adjusted p-value less than 0.1. Gene scores The following gene lists were used to calculate gene scores using Scanpy functions9,30,71. For DAO signatures, genes were selected if they were upregulated ≥4-fold in the condition of interest, and visually assessed for their specificity and expression level on UMAPs across conditions and timepoints. Genes were retained only if they had a human orthologue that was expressed in our dataset. Mouse/human nomenclature was converted by mapping the gene orthologues using GRCh38.p14 and GRCm39. For microglial signatures, genes from the Hallmark Hypoxia or Hallmark Inflammatory Response gene sets were subset and only those with both human and mouse orthologues were retained. Next we filtered the lists for genes upregulated ≥2-fold in at least one human lesion condition (AL, CAL, CIL, or RL) or in mouse DAM populations. The logFC and padj/FDR values are shown for each gene score in Supplementary data tables for each relevant DEG analysis. Cell cycle scoring We downloaded the Regev lab cell cycle genes, as used in Seurat and the scverse tutorials and converted the genes to mouse nomenclature using BioMart (https://github.com/scverse/scanpy_usage/tree/master/180209_cell_cycle). Pandey GADD45B, CAMK2D, CNTN1, VIM, NUPR1, SLC3A2, FOS, JUNB, JUN, DDIT3, KLK6, CD63, HMGB2, DYNLL1, CSF1, B2M, FBL, KLF6 IRAS1 CCL2, CCL20, CSF3, CXCL1, CXCL2, CXCL3, CXCL5, FTH1, PDCD5, PDLIM4, PITX3, PSME2, SERPINE1, SOD2, TNFAIP6 IRAS2 B2M, CALR, CARHSP1, CCL2, CD47, FTH1, ICAM1, LAP3, NFKBIA, NNMT, PDCD5, PFN1, PKM, PRELID1, PSME2, SOD2, TMSB10, VCAM1 Common innate immunity IFI47, IIGP1, MDFIC, TOP2A, DYNLT1A, LGALS1, CXCL10, OASL1, ZBP1, DHX58, TRIM25, RSAD2, SLFN9, OAS1A, IFITM2, BCL3, IFI27L2A, OASL2, IRF7, IFIT1, BST2, ISG15, IFIT3B, IFIT3, IFITM3, RTP4 CPZ DAO signature SPRY2, SYT4, CHCHD10, DTHD1, TRIB3, EPHA5, NUPR1, ATF4, CYB5R2, DDIT3, GDF15, CD44, CEBPG, VEGFA, CDH6, SLC7A5, S100A6, CDKN1A, RGS20, LMNA, VIM, EGR1 Remyelination DAO signature COL6A1, GADD45G, COL27A1, COL6A2, NAV3, TMEM176A, COL13A1, FGFRL1, SNX10, PARP3, CXCL14, NRTN, SOCS3, ANXA2, TNFRSF1A, BCL3, B2M, PLVAP, GADD45B, SEMA4F Hypoxia signature ANXA2, BHLHE40, CHST2, ENO2, FOSL2, GCNT2, GLRX, HMOX1, ISG20, MIF, MXI1, NDRG1, NFIL3, NOCT, P4HA1, PAM, PGK1, PGM1, PIM1, PLAC8, PLAUR, PLIN2, PRDX5, SDC2, SDC3, TGFBI, TPD52, TPI1, VEGFA Inflammatory response signature ABCA1, AQP9, AXL, BST2, CCL5, CD14, CD48, CHST2, CLEC5A, HIF1A, HRH1, ICAM1, IL1B, IL10, IL15, IL18R1, IL1R1, MARCO, MSR1, MXD1, NAMPT, NOD2, PLAUR, PTGER4, RNF144B, SLC31A2, STAB1, TLR2 Meta-neighborhood analysis To determine cluster similarity in our cross-species analysis, we used MetaNeighbor72. We merged the human-mouse datasets and calculated the top 2000 HVGs for the oligodendrocyte or microglia lineage separately. The data was then passed to MetaNeighborUS using the fast version, with organism as study_id, and the broad cell type annotation as the ct_col variable. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary data 1 Supplementary data 2 Supplementary data 3 Supplementary data 4 Supplementary data 5 Supplementary data 6 Supplementary data 7 Supplementary data 8 Supplementary data 9 Supplementary data 10 Supplementary data 11 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements We would like to thank Dr. Vittorio Gallo at Seattle Children’s Hospital for his mentorship of K.L.A. during the early stages of this project and Dr. Cody Smith at the University of Notre Dame for critically reading this manuscript. We thank Dr. Bharat Mishra for analysis and statistical advice. We thank the Freimann Life Science Center at the University of Notre Dame for mouse breeding and husbandry, and Dr. Sara Cole from the Notre Dame Integrated Imaging Facility for her microscopy support. Images presented in this manuscript were generated using the instruments at the NDIIF/Optical Microscopy Core. We thank Karuna Panchapakesan, Dr. Susan Knoblach, and Dr. Surajit Bhattacharya at Children’s National Hospital for their assistance with single-nuclei RNA-seq library generation and analysis. Finally, we thank members of the Adams lab for their discussion. This work was supported by a career transition award from the National Multiple Sclerosis Society (K.L.A.). Author contributions Conceptualization (E.L.A., K.L.A.), Data curation (E.L.A.), Formal analysis (E.L.A., V.R.V., E.A.M.), Funding acquisition (K.L.A.), Investigation (E.L.A., V.R.V, E.A.M., M.M.H, K.L.A.), Methodology (E.L.A.), Project administration (K.L.A.), Supervision (K.L.A.), Visualization (E.L.A.), Writing – original draft (E.L.A.), and Writing – review and editing (E.L.A., V.R.V., K.L.A.). Data availability The single-nucleus sequencing data generated during this study have been deposited in the Gene Expression Omnibus (GEO) database under accession number GSE293850. The additional data used in this study are available in the GEO database under accession codes: GSE182846; GSE148676; and GSE230753; and on Zenodo under DOI 10.5281/zenodo.8338963 (https://zenodo.org/records/8338963). The human MS data were deposited to European Genome-Phenome Archive (EGA) as dataset EGA: EGAD00001009169. Source data are provided with this paper. Code availability Details of the analyses completed for this study are described in the Methods sections. Jupyter notebooks used for the data processing and analysis are available on Github (https://github.com/AdamsLabNotreDame/Aboelnour_et_al) and archived on Zenodo (https://zenodo.org/records/19075392). Additional information required to reproduce the analyses is available from the lead contact upon request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-72383-y. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary data 1 Supplementary data 2 Supplementary data 3 Supplementary data 4 Supplementary data 5 Supplementary data 6 Supplementary data 7 Supplementary data 8 Supplementary data 9 Supplementary data 10 Supplementary data 11 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The single-nucleus sequencing data generated during this study have been deposited in the Gene Expression Omnibus (GEO) database under accession number GSE293850. The additional data used in this study are available in the GEO database under accession codes: GSE182846; GSE148676; and GSE230753; and on Zenodo under DOI 10.5281/zenodo.8338963 (https://zenodo.org/records/8338963). The human MS data were deposited to European Genome-Phenome Archive (EGA) as dataset EGA: EGAD00001009169. Source data are provided with this paper. Details of the analyses completed for this study are described in the Methods sections. Jupyter notebooks used for the data processing and analysis are available on Github (https://github.com/AdamsLabNotreDame/Aboelnour_et_al) and archived on Zenodo (https://zenodo.org/records/19075392). Additional information required to reproduce the analyses is available from the lead contact upon request.