When MS Genetics Meets scRNA-seq
Every result in a genetics paper carries a sign. The allele raises risk or lowers it, the gene is higher in patients or lower, the correlation is positive or negative. Signs are also the first thing to go when a finding is carried from one layer of evidence to the next, and three recent multiple sclerosis (MS) papers show this unusually clearly, because all three are built the same way: public datasets, a statistical pipeline, a short gene list at the end, and single-cell or single-nucleus RNA sequencing somewhere in the middle. In each one a direction established at one layer reverses at another, and in each one the authors report the reversal rather than smoothing it away. Reading them for the sign changes is a good way to see what the single-cell step is measuring.
A Chain of Causal Estimates With One Step Reversed
Huang and colleagues assembled a three-link causal chain entirely from summary data. Exposures came from genome-wide studies of 1,091 metabolites and 309 metabolite ratios in 8,299 people; the MS outcome came from the International MS Genetics Consortium, 47,429 cases and 68,374 controls of European ancestry. Eight SNPs survived harmonisation for serotonin and six for adenosine deaminase (ADA), all with F statistics above 10. Inverse variance weighted estimates put serotonin on the risk side (beta 0.350, p 3.63E-05, FDR 0.049), ADA on the protective side (beta minus 0.395, p 2.73E-04, FDR 0.012), and serotonin negatively on ADA (beta minus 0.089, p 0.0087), with no reverse causation detected. Three signs, one coherent chain. Their own methods then supply the caveat that bears on the title: progression was assessed indirectly through cross-sectional case-control data, because longitudinal data were not available, and the authors state that risk and progression are distinct aspects of the disease.
Where the Single-Cell Step Put the Enzyme
The single-cell layer came from one GEO dataset, filtered at 200 to 4,000 genes per cell and under 10 percent mitochondrial content, reduced by principal components and then UMAP, and annotated with MonacoImmuneData. It gave 18 clusters and six major immune types, with ADA highest in T-NK cells and lowest in platelets. The second reversal arrives here: ADA expression was higher in the group with low immune receptor signalling, not the high one. Three machine learning methods were run on the differentially expressed genes and their overlap taken as the answer, returning IK, UBA52 and CCDC25; a nomogram built on those three reached an area under the curve of 1.000 in training and 0.976 in validation. The authors set out their own constraints: predominantly European genomic data, technical artefacts in high-throughput omics, MR assumptions, and a need for prospective clinical studies.
One Tissue, Two Genes, Opposite Directions
Qiao and colleagues worked in MS brain tissue rather than blood. After batch correction and merging, weighted gene co-expression network analysis at a soft threshold of 13 picked the module most strongly associated with MS; intersecting 1,932 differentially expressed genes with that module left 324, and intersecting those with a published endoplasmic reticulum stress gene set left 13. LASSO and Boruta together narrowed the 13 to four, of which three cleared an area under the curve of 0.7 in both the merged and an external dataset and were higher in MS than in controls. Two of them, GPX1 and RCN1, were then confirmed raised in the brains of mice with experimental autoimmune encephalomyelitis by quantitative PCR. So far the direction is simple: both genes up in disease, in human tissue and in the animal model alike.
What Resolution Did to That Direction
Single-nucleus sequencing of 15 MS and 5 control samples, clustered into 12 cell types, took that simple direction apart. Inside the same tissue the two genes moved opposite ways across cell populations: one fell in pericytes while rising in microglia and fibroblasts, the other fell in oligodendrocyte precursor cells and endothelial cells while rising in pericytes. Module scoring put microglia and pericytes highest, and cell-to-cell signalling analysis found the APP-CD74 pathway strengthened in MS, with pericytes as senders and microglia as receivers; the same appeared in a mouse dataset, where staining showed APP raised around brain vessels alongside PDGFRB and CD74 raised in cells of microglial shape alongside IBA-1. Then a third reversal: the Mendelian randomisation step returned a protective odds ratio of 0.876 for one of the two genes that the expression data had shown raised in MS, and nothing significant for the other. The authors note that their APP immunohistochemistry reference came from inflamed tissue that was not MS tissue.
One Gene, Three Diseases, Three Signs
Wang and colleagues asked what rheumatoid arthritis, MS and type 1 diabetes hold in common. Eight genes were differentially expressed in whole blood across all three, generally lower in the first two but trending higher in MS. LASSO run per disease left two candidates, two and three respectively, and ROMO1 was the only gene chosen in all three. Its correlation with monocyte fraction was positive in rheumatoid arthritis (0.20) and MS (0.41) and negative in type 1 diabetes (minus 0.43). The single-cell step, on 8,564 cells for rheumatoid arthritis, 2,046 for MS, 3,011 for type 1 diabetes and 2,961 for controls, found ROMO1 highest in CD14 and CD16 HLA-DR positive monocytes in every group, and reported MS samples no higher than controls across all cells, with significant falls in CD4 T cells, CD14 HLA-DR positive monocytes and NK cells. Mendelian randomisation then gave odds ratios of 0.52 in rheumatoid arthritis, 3.42 in MS and a non-significant 0.87 in type 1 diabetes, on three instruments the authors call too few for firm conclusions.
Which Layer the Sign Belongs To
Read together, the reversals are the most informative part of these three papers, because each marks a boundary between layers rather than a contradiction inside one. A gene raised in bulk tissue can be falling in most of its cell types while rising steeply in a few, which is what the single-nucleus step in the second paper showed. An allele that raises expression can lower risk, because expression and liability are different quantities estimated in different people. A gene can behave one way in one autoimmune disease and the opposite way in another, and the genetic correlations between those diseases are themselves of mixed sign. All three sets of authors say the same thing about what is still missing: these are computational findings from public data, and the functional work has not been done. In two of the three, the single-cell step is where one number first became several.
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:
Huang, L., Shi, J., Li, H., & Lin, Q. (2025). Mendelian randomisation and single-cell transcriptomic analyses reveal serotonin promotes multiple sclerosis progression by suppressing adenosine deaminase activity. BMJ Open, 15(9), e102876. https://doi.org/10.1136/bmjopen-2025-102876
Qiao, Z., Wang, Y., Ma, X., et al., Li, H., & Liu, Y. (2025). GPX1 and RCN1 as new endoplasmic reticulum stress-related biomarkers in multiple sclerosis brain tissue and their involvement in the APP-CD74 pathway: an integrated study combining machine learning and multi-omics. International Journal of Molecular Sciences, 26(13), 6286. https://doi.org/10.3390/ijms26136286
Wang, T., He, Q., & Chan, K. H. K. (2025). A multi-omics approach to identify and validate shared genetic architecture in rheumatoid arthritis, multiple sclerosis, and type 1 diabetes: integrating GWAS, GEO, MSigDB, and scRNA-seq data. Functional & Integrative Genomics, 25(1), 91. https://doi.org/10.1007/s10142-025-01598-x
