Can We Predict Multiple Sclerosis Before It Starts? What Genetic and Lifestyle Risk Scores Really Tell Us
Multiple sclerosis (MS) prevention sounds straightforward in theory: identify people at high risk, test preventive strategies early, and reduce future disease burden. In practice, the low incidence of MS makes prevention trials extremely difficult to run efficiently—either you need enormous sample sizes, or you need a way to “enrich” the trial population with people who are meaningfully more likely to develop MS. Hone and colleagues frame risk prediction primarily as a research tool for smarter trial design, while cautioning that “predicting MS” for a specific individual is a much tougher (and possibly unrealistic) clinical goal.
MS risk is a mosaic: many small genetic effects plus key exposures
The review reinforces a core point about MS biology: susceptibility is complex, shaped by both genetics and environment. On the genetic side, large studies have identified hundreds of associated loci (including a prominent effect in the HLA region), but no reproducible single-gene “MS cause” that would make prediction simple. On the environmental side, the most consistently replicated associations include smoking, childhood/adolescent obesity, infectious mononucleosis/EBV-related measures, and lower vitamin D status—factors that are plausible, measurable (to varying degrees), and potentially modifiable.
What a polygenic risk score really is (and what it is not)
A polygenic risk score (PRS) is essentially a weighted tally: across many genetic variants, you count how many risk-increasing alleles a person carries, and weight each by its estimated effect from genome-wide association studies. Environmental risk scores (ERS) follow a similar logic—combine exposures (e.g., smoking history, prior infectious mononucleosis) using effect estimates from epidemiologic studies—while more advanced models may also try to account for gene–environment interactions (for example, situations where a given exposure matters more in people carrying specific HLA risk alleles). This is conceptually elegant: a single number that summarizes a complicated risk landscape. But elegance does not guarantee clinical usefulness.
What the data show so far: discrimination improves, individual prediction does not
Across published cohorts and methods, MS PRS (and PRS+ERS hybrids) can often separate groups of people with MS from controls “better than chance,” with reported AUCs spanning roughly the low 0.5s up to around 0.8 depending on design and features included. Adding environmental variables can yield modest improvements. Yet the review’s key message is that these gains are not the same as being able to tell a particular person, in clinic, whether they will develop MS. The literature is also heterogeneous—different variant selection strategies, different covariate handling (age, sex, ancestry principal components), different exposure definitions—which makes results difficult to compare cleanly.
Theoretical ceilings: heritability and “missing pieces” cap what PRS can ever do
The authors emphasize that heritability sets an upper bound on prediction accuracy from genetics alone. For MS, SNP heritability estimates from large GWAS place a firm ceiling on what common-variant PRS can capture, and even “best practice” methods cannot exceed what the underlying genetic architecture allows. Performance is also degraded when PRS include non-causal proxy variants (imperfect linkage disequilibrium with the true causal alleles), when rare high-impact variants are missed, and when interaction effects exist but are not modeled. In other words, some limitations are mathematical, not merely technical.
Real-world friction: measuring exposures, transferring across ancestries, and interpreting metrics honestly
Environmental factors are dynamic and time-dependent—“ever smoked” or “ever obese” is a crude summary of biology that likely depends on timing and dose, and may be inconsistently recorded across cohorts. Some exposures may not be truly causal (or may be confounded), so including them can add noise rather than signal. PRS also port poorly across ancestries because linkage disequilibrium patterns and allele frequencies differ, meaning a score trained in largely European datasets may misestimate risk elsewhere. Finally, the paper highlights a common interpretability trap: metrics like AUC can look impressive while the absolute risk difference between “low” and “high” score groups remains small for a low-prevalence disease. The authors’ examples (including UK Biobank-style analyses) illustrate why positive predictive value can remain low even when discrimination seems strong—an issue that matters directly for any real-world screening concept.
The realistic opportunity: risk stratification for prevention research, not a crystal ball
Where does this leave the field? The review takes a pragmatic stance: MS risk scores may be valuable for enriching prevention trials—finding subsets of the population with substantially higher incidence than baseline, where interventions (for example, targeting EBV-related pathways) can be tested more efficiently. The authors also outline clear technical directions that could improve research-grade prediction: better causal variant mapping, inclusion of rare variation as evidence matures, explicit modeling of gene–environment interplay, and larger non-European datasets to improve portability. Just as importantly, they stress the communication challenge—working with people with MS and stakeholders to convey both the promise and the significant caveats—because an imperfect risk score can mislead if it is treated like a definitive forecast rather than a probabilistic tool.
Disclaimer: This blog post is based on the provided 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.
References:
Hone, L., Giovannoni, G., Dobson, R., & Jacobs, B. M. (2022). Predicting multiple sclerosis: challenges and opportunities. Frontiers in Neurology, 12, 761973.
