MS Genetic Risk, Healthy Brains, and an Effect Measured in Milliseconds

Does genetic liability to multiple sclerosis (MS) leave a mark on the brain of someone who does not have MS? Feng and colleagues tested that in 35,952 neurologically healthy UK Biobank participants aged 45 to 83, all with a polygenic risk score for MS, diffusion MRI, and cognitive testing, after excluding 6,279 people carrying any ICD-10 record of demyelinating disease, neurodegenerative disease, dementia, brain tumour, head trauma, or cerebrovascular disease. The score matters here because MS risk is not carried by one variant. HLA-DRB1*15:01 is the strongest known allele and confers roughly a threefold increase, while hundreds of other variants each contribute modest effects that only a polygenic score can sum. One methodological point deserves credit before the results: the score was built from GWAS of 10,820 cases and 234,517 controls that excluded UK Biobank participants entirely, so score and outcome come from non-overlapping samples and cannot be inflated by that particular kind of overfitting.

The Primary Result Is a Null, and the Paper Says So
In the full sample, the polygenic score was not associated with any of 11 cognitive measures once corrected for multiple comparisons. The strongest nominal signal was paired associate learning at β = −0.012 with p = 0.038, which became 0.418 after false discovery rate correction. Reaction time sat at p = 0.212, corrected to 0.732. Across all 11 measures the corrected values ran from 0.418 to 0.971, meaning not one came close. That is the headline result of a large and well-powered study, and the authors report it as such: at the level of the general population, MS genetic risk does not have a direct and widespread effect on cognitive function. The analysable sample also varied by test, from 23,593 for picture vocabulary to 33,806 for prospective memory, since not every participant completed every task.

What Turned Up in the Top Quartile, Converted Into Milliseconds
Participants were then split into quartiles by their score and the analysis rerun. In the highest quartile, a higher score was associated with longer reaction time at β = 0.048 with a corrected p of 0.047, just inside the threshold. No other cognitive measure reached significance there, and the lowest quartile produced nothing at all. Because every variable was standardised, that coefficient can be converted into something a reader can hold. Reaction time in this sample averaged 595.5 ms with a standard deviation of 109.3 ms, so 0.048 standard deviations works out to roughly five milliseconds per standard deviation of polygenic risk. That is the magnitude of the finding. Reporting it in milliseconds rather than standard deviations does not diminish it, but it does make clear what scale of effect is being discussed.

One Imaging Metric Out of Seven, in One Tract Out of Eight
Because reaction time was the only cognitive measure to survive in the top quartile, the imaging follow-up was confined to eight white matter regions previously linked to reaction time, testing four diffusion tensor metrics and three neurite orientation dispersion and density metrics in each. One association survived: mean diffusivity in the splenium of the corpus callosum, β = 0.021, corrected p = 0.024. Fractional anisotropy, radial diffusivity, axial diffusivity, and all three of the neurite metrics were null. The authors read those nulls carefully rather than conveniently. The pattern that usually accompanies overt demyelination or marked axonal injury involves fractional anisotropy and radial diffusivity, and its absence here means the result should not be taken as evidence of irreversible white matter destruction. The null on isotropic volume fraction likewise argues against free water explaining the change. Mean diffusivity, as they note, is sensitive but nonspecific and cannot separate demyelination from axonal damage or other processes.

A Mediation Effect of Three Percent
Mediation analysis in the top quartile put splenium mean diffusivity between the score and reaction time. The total effect was 0.048 (p = 0.004), the direct effect 0.047 (p = 0.006), and the indirect effect 0.0014 with a 95% confidence interval of 0.0002 to 0.0032. The mediated proportion is 3.0%. Read plainly, 97% of an association already measured in single-digit milliseconds travels through something other than the splenium, and the confidence interval on the mediated part sits close to zero at its lower bound. The authors describe the mediating effect as unlikely to have immediate clinical relevance and better read as a modest population-level association, which is the accurate characterisation. They also point out that because predictor, mediator, and outcome were all measured at the same moment, no temporal precedence is established and reverse causality cannot be ruled out.

The Framing Is the Part Worth Copying
Several statements in this paper draw lines that similar studies often blur. The quartile analysis is described not as confirmatory but as hypothesis-generating exploratory analysis. The authors note that multiplicity was not controlled across the whole analytical framework of 56 tests, since correction was applied within each conceptual unit rather than across all of them, so false positives cannot be fully excluded. They reject outright the interpretation a reader might reach for, writing that these alterations should be conceptualised not as a clinical biomarker for detecting preclinical MS but as a neural endophenotype reflecting the subtle, continuous impact of genetic susceptibility on brain morphology in the general population. And they check whether undiagnosed preclinical cases could be driving the result, estimating from a radiologically isolated syndrome prevalence of 0.1% to 0.5% that at most 35 to 175 such individuals would sit in a cohort this size, under 0.5% of the sample.

What Is Missing From the Model
The covariate list is where the residual doubt lives, and the authors name it themselves. Models adjusted for age, age squared, sex, ten genetic principal components, body mass index, and intracranial volume. They did not adjust for years of schooling, hypertension, diabetes, smoking history, or white matter hyperintensity burden. Part of the reason is practical: hyperintensity measures derived from FLAIR imaging were unavailable for all participants, and restricting the sample to those who had them would have cost statistical power. Part is conceptual, since some of those variables may lie on the causal path rather than confound it. Either way, every one of them influences white matter microstructure in the splenium specifically and also influences cognition, so residual confounding cannot be excluded on the one imaging finding the paper rests on. Imaging also came from multiple sites without harmonisation, which the authors recommend for future work using this resource. The study that would settle the question follows directly from those gaps: the same analysis with vascular burden and hyperintensity load included as covariates, harmonised across sites, in participants imaged more than once, so that reaction time and splenium diffusivity can be watched changing rather than compared at a single moment.

Disclaimer: This blog post is based on the cited research article 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.

Reference:
Feng, C., Kikuta, J., Hagiwara, A., Zou, R., Ma, S., Guo, S., Kitagawa, T., Mizuta, K., Takabayashi, K., Wada, A., Takai, I., Hoshino, Y., Tomizawa, Y., Hatano, T., & Kamagata, K. (2026). Associating a polygenic risk score for multiple sclerosis with brain MRI metrics and cognitive performance in healthy UK Biobank participants. Magnetic Resonance in Medical Sciences, 25(4), 2026-0094. https://doi.org/10.2463/mrms.mp.2026-0094