Predicting Multiple Sclerosis Risk: Scientific Constraints, Emerging Methods, and Translational Opportunities
The mini-review “Predicting Multiple Sclerosis: Challenges and Opportunities” synthesizes the current state of risk prediction research in Multiple Sclerosis, with a specific emphasis on polygenic risk scores (PRS) and environmental risk scores (ERS) as tools to identify individuals at elevated susceptibility. Frontiers in Neurology is used as the publication venue to frame a pragmatic argument: although genetic and environmental data can stratify risk distributions in populations, these approaches are unlikely to provide clinically meaningful prediction for a given individual because of theoretical ceilings (heritability), measurement limitations, and—critically—the low disease prevalence that constrains positive predictive value.
Genetic Architecture of Susceptibility and the Limits It Imposes
A key premise is that MS is archetypally polygenic: susceptibility arises from the aggregate of many loci with small individual effects rather than any reproducible single-gene cause. The review highlights that common variants explain a bounded fraction of liability (reported SNP-heritability ~19.2%), with the major histocompatibility complex—particularly HLA-DRB1*15:01—remaining the strongest established signal. This matters for prediction because PRS performance is mathematically constrained by the proportion of variance attributable to the genetic component being modeled; if common variants only explain a modest fraction of liability, even an optimally constructed PRS cannot achieve “diagnostic-grade” accuracy. The article uses this framework to motivate why apparently respectable discrimination statistics can still translate into poor clinical utility.
Principles of PRS/ERS Construction and Empirical Performance to Date
Methodologically, the review describes PRS as a weighted sum of risk alleles, where weights are typically derived from genome-wide association studies and the genotype is counted per locus; ERS is conceptually similar but depends on quantifying exposure histories (e.g., smoking, infectious mononucleosis proxies, vitamin D). Across multiple studies summarized in a comparative table, PRS and combined PRS+ERS models typically yield AUCs ranging roughly from low-to-moderate performance, with improvements when environmental variables are added, but none reaching thresholds expected for routine clinical prediction at the individual level. Importantly, the authors critique heterogeneity in reporting practices—such as inconsistent calibration analyses and limited use of absolute risk estimates across score strata—which impairs interpretability and cross-study comparability.
Technical and Biological Barriers to Improving Genetic Prediction
The review enumerates several reasons PRS may underperform even before clinical translation is considered. First, causal-variant uncertainty at associated loci means many PRS include proxy variants whose predictive value can degrade with subtle differences in linkage disequilibrium structure, motivating approaches that explicitly model local LD (e.g., Bayesian shrinkage methods). Second, restricting scores to common variants omits potentially informative rare variation; although rare alleles can matter greatly for an individual carrier, their low population frequency limits their impact on global predictive metrics unless very large sequencing datasets become available. Third, interaction structure is incompletely captured by additive models: gene–gene interactions appear limited outside the MHC, but gene–environment interactions (e.g., differential effects conditioned on high-risk HLA status) provide a plausible route to incremental gains, including via non-linear or interaction-aware modeling.
Portability, Equity, and the Cross-Ancestry Generalization Problem
A crucial translational issue is cross-ancestry portability: PRS derived primarily from European-ancestry GWAS often lose accuracy when applied to other ancestral groups due to allele-frequency differences, distinct LD patterns, and potentially ancestry-specific architectures. The review notes preliminary evidence that MS genetic architecture is not identical across some non-European populations and points to multi-ancestry methods or functionally informed weighting strategies as partial mitigations. This section implicitly reframes prediction not merely as a technical optimization problem but as a generalizability problem: a score that is only valid in a subset of ancestries risks exacerbating disparities in downstream trial recruitment or prevention strategies if deployed without careful validation.
Environmental Risk Modeling: Causality, Measurement, and Time Dependence
While adding environmental factors can improve discrimination, the article emphasizes three constraints: (i) not all associated exposures are necessarily causal (confounding and bias can inflate apparent effects), (ii) exposures are difficult to measure consistently across cohorts and cultures, and (iii) environmental effects are time-dependent in ways that simplistic binary encoding fails to capture (e.g., adolescence appears to be a critical window for obesity-related risk). The review treats Mendelian randomization as an important triangulation tool to strengthen causal inference for candidate exposures, while also warning that including non-causal variables can inject noise and degrade predictive performance. Collectively, these points underscore that ERS construction is often limited less by modeling strategy and more by phenotype fidelity and temporal resolution.
Clinical Utility, Preventive Trials, and a Realistic Path Forward
The most consequential argument is interpretive: discrimination metrics like AUC can appear impressive yet still yield low positive predictive value when applied to a low-prevalence disease such as MS. The authors illustrate that even near-perfect separation of cases and controls can label large numbers of healthy individuals as “high risk,” making individual-level prediction clinically unattractive and potentially harmful if it drives anxiety, unnecessary surveillance, or inappropriate interventions. However, the review remains constructive: risk scores may be valuable as trial-enrichment tools—identifying subgroups with substantially higher incidence than the background population to make prevention studies more feasible (for example, in the context of preventive strategies targeting Epstein–Barr virus biology). The forward-looking agenda emphasizes improved PRS methods, explicit modeling of interactions, expansion of diverse-ancestry GWAS, prospective validation, and careful stakeholder communication about what risk scores can—and cannot—deliver.
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:
Jacobs, B. M., Noyce, A., Bestwick, J., Belete, D., Giovannoni, G., & Dobson, R. (2020). Gene-environment interactions in Multiple Sclerosis: a UK Biobank study. bioRxiv, 2020-03.
