Metabolism Off Reused Single-Cell Data in MS

Public single-cell and single-nucleus datasets from multiple sclerosis (MS) tissue have become infrastructure. A metabolic claim built on top of one of them inherits whatever the original study measured, and the original study measured transcripts. That is not a flaw in reuse, which is often the only way to get human MS tissue at this resolution, but it does mean the metabolic part of any such claim arrives second-hand twice over: inferred from transcripts, and from transcripts somebody else collected for a different question. Three recent papers all read metabolism off single-cell data they largely did not generate, and what separates them is how much independent evidence each one stacks on top of the borrowed atlas.

Patient Cells, a Methylome, and a Senolytic
Park and colleagues brought their own material first. They reprogrammed skin fibroblasts from people with progressive MS directly into induced neural stem cells, using a protocol that keeps the donor epigenome rather than resetting it, then ran whole-genome bisulfite sequencing on both the parental fibroblasts and the matched stem cell lines across 28 million CpG sites. Genes governing lipid metabolism, inflammation and interferon production came back hypomethylated in the patient lines, and the pattern held in both the fibroblasts and the cells derived from them, so the epigenetic difference predates the reprogramming. Single-cell and single-nucleus sequencing with chromatin accessibility profiling then split those lines into clusters, and the patient-enriched ones carried cholesterol biosynthesis, oxidative phosphorylation and neurodegeneration markers alongside senescence and interferon signalling.

Checking the Finding Against Public Atlases
The reuse came at the validation step rather than the discovery step, which is the reverse of the usual order. Having named a senescent, interferon-responsive radial glia-like population in their own cultures, they went looking for it in published post-mortem single-nucleus and spatial datasets and found it: a roughly 2.3-fold higher proportion of senescent radial glia-like cells than senescent non-radial-glia cells within chronic active lesions, under 5 per cent in control tissue, and about twice the fraction at the chronic active lesion edge compared with the lesion core. Pseudotime followed the population back to mature astrocytes. Conditioned medium from patient stem cells pushed oxidative stress, lipid metabolism and senescence transcripts up in control cells, and the senolytic ABT-263 abolished that transfer. The metabolic claim is still transcriptional, but a drug now sits behind it.

A Screen of Datasets Rather Than Samples
Soler-Sáez and colleagues did the opposite and made the reuse the entire design. They screened GEO, ArrayExpress and the UCSC Cell Browser in September 2025 for human MS single-cell and single-nucleus studies, requiring cases and controls, sex recorded as a variable, at least three individuals per sex per condition, and an available expression matrix. Three datasets survived, with grey matter neurons from a fourth kept for validation under the original authors' annotation. Each was processed separately through quality control, normalisation, variable gene selection, clustering and cell type annotation, then compared for expression changes, functional profiles, pathway activation and ligand-receptor interactions in females, males and the difference between them. The pipeline was written in R, with package versions recorded, and the output is a browsable atlas rather than a mechanism. No new tissue was collected at any point, and the inclusion rules are what carry the weight: by demanding sex as a recorded variable and three donors per sex per condition, the screen threw away most of what is published.

Named Genes, Stated as Potential
What that design buys is gene-level specificity across cell types. In relapsing-remitting disease, male adaptive immune cells carried higher expression of electron transport chain genes at every step: NDUFA13 and NDUFA4 in CD4 T cells for the NADH to ubiquinone stage, UQCR10 for ubiquinol to cytochrome c, COX6B1, COX6C and COX7C for the final step, with CD8 T cells adding UQCRB, UQCRH and ATP5MG. Female CD8 T cells instead raised oxidative phosphorylation and mitochondrial maintenance genes in a homeostatic cluster, while male CD8 T cells leaned cytolytic through granzyme and perforin. In progressive forms the male signature included glycolytic genes LDHA and TPI1. The authors write that these genes could be driving the chain harder in males, with a potential rise in reactive oxygen species, and that further work is needed to see whether the transcriptional changes translate into measurable effects.

A Chain Built Entirely From Existing Data
The third design keeps only the computation. Li and colleagues combined network pharmacology, machine learning and single-cell transcriptome analysis to look for MS genes that gut microbial metabolites might act on, scored feature contributions with SHapley Additive exPlanations, tested causal direction with Mendelian randomisation, and finished with molecular docking plus predicted drug-likeness and toxicity. Caspase-3 came out as a core target interacting with several named metabolites, among them L-isoleucine, aromatic lactic acid derivatives, 3-hydroxyphenethyl alcohol and D-xylose, with the proposed routes running through TNF, MAPK, interleukin-17 and galectin signalling, and particular microbial taxa tied to those metabolites. The metabolites here are named rather than measured, and the authors close by stating that experimental validation is needed to confirm the mechanisms and any translational potential. Mendelian randomisation is the one step in that chain that tests direction rather than association, and it was run on summary data rather than on the cells.

What a Reused Atlas Settles and What It Leaves
The three sit at three depths. Park generates new patient cells and a new methylome, uses the public atlases to ask whether the population exists in real brains, and lands a perturbation on top. Soler-Sáez generates nothing and returns gene-level sex differences across cell types and disease courses, shipped as a browsable resource, with the authors' own hedging kept in the text. Li generates nothing and returns a target plus four candidate metabolites, with the request for experiments written into the abstract. For a geneticist the common output is a list of named genes, from NDUFA13 through LDHA to CASP3, which is what makes any of this testable. None of the three measured a metabolite in MS tissue.

Disclaimer: This blog post is based on the cited studies and is intended for informational purposes only. It is not intended to provide medical advice. Please consult with a healthcare professional for any health concerns.

References:
Park, B., Nicaise, A. M., Tsitsipatis, D., et al., Pluchino, S., & Beerman, I. (2025). Integrated omics reveals disease-associated radial glia-like cells with epigenetically dysregulated interferon response in multiple sclerosis. Neuron, 113(24), 4158-4177.e10. https://doi.org/10.1016/j.neuron.2025.09.022
Soler-Sáez, I., Gómez-Cabañes, B., Grillo-Risco, R., et al., Hidalgo, M. R., & García-García, F. (2026). Single cell landscape of sex differences in the different courses of multiple sclerosis. Biology of Sex Differences, 17(1). https://doi.org/10.1186/s13293-026-00957-y
Li Li, Liu, H., Ye, Z., & Xu, W. (2026). Gut microbiota-derived metabolites regulate CASP3 and neuroimmune pathways in multiple sclerosis: an integrative multiomics study. Mediators of Inflammation, 2026, 3572399. https://doi.org/10.1155/mi/3572399