When Genetic Risk Scores Don’t Fit Everyone: What Multiple Sclerosis Teaches Us About Ancestry
Genetic risk scores (GRS) are often described as a kind of “credit score” for disease risk: they compress information from hundreds of genetic variants into a single number that helps estimate how likely someone is to develop a condition. For multiple sclerosis (MS), large genome-wide association studies in Europeans have mapped 232 genetic variants that, together, capture a substantial part of inherited susceptibility. Rivier and colleagues asked a critical question: if we build an MS GRS from primarily European data, does it work equally well for people of African and Latino/admixed ancestry in the United States? Using the highly diverse All of Us Research Program, they show that the answer is “not yet”—and that this has direct implications for fairness and reliability in genetic prediction.
A diverse cohort and a shared genetic score
The team leveraged whole-genome sequencing and electronic health records from 173,153 All of Us participants with linked data between 2018 and 2022. Genetic ancestry was assigned using principal components and the Rye software, which estimates each person’s proportion of African, European, Latino/admixed American, East Asian, and Middle Eastern ancestry; the violin plots on page 12 show that most people in each group had ancestry very close to their assigned category, with only a minority displaying substantial admixture from other ancestries. The authors focused on three major groups: European (EUR), African (AFR), and Latino/admixed American (AMR). To make performance directly comparable, they randomly selected 32,428 individuals from each group and applied the same MS GRS, built from 232 independent single nucleotide polymorphisms identified by the International Multiple Sclerosis Genetics Consortium. Each person’s score was the sum of risk alleles across these variants, weighted by the effect sizes reported in the large European GWAS.
How well did the European-derived score work across ancestries?
First, the investigators divided each ancestry group’s GRS into quintiles—very low, low, intermediate, high, and very high genetic risk—and asked a simple question: as you move from the lowest to the highest quintile, does the proportion of people with MS rise? In individuals of European ancestry, the answer was clearly yes: MS prevalence climbed from 0.66% in the lowest GRS quintile to 1.59% in the highest, and after adjustment for age, sex, smoking, BMI, principal components, and admixture, those in the top quintile had more than twice the odds of MS compared with those in the bottom quintile (odds ratio [OR] 2.30, 95% CI 1.60–3.36). The score also performed well in the Latino/admixed American group, where MS prevalence rose from 0.23% to 0.63% across quintiles and the adjusted odds in the top quintile were 2.53 times higher than in the bottom quintile. The forest plots on page 14 visually underscore these trends, with a clear stepwise increase in risk across quintiles for both EUR and AMR groups.
When a “good” score fails: the African ancestry gap
In stark contrast, the same GRS did not meaningfully stratify MS risk in individuals of African ancestry. In the AFR group, MS prevalence was 0.45% in the intermediate quintile and 0.82% in the top quintile, but the trend across all quintiles was weak and not statistically significant (p-trend = 0.17). The adjusted OR comparing top to bottom quintile was only 1.30 (95% CI 0.88–1.95), with wide, overlapping confidence intervals in the bottom-left panel of the forest plots on page 14. This is not simply a power problem—the authors downsampled the European and African groups to match the Latino/admixed sample size, and sensitivity analyses without downsampling and with Firth’s penalized logistic regression gave similar results. Rather, it highlights that a GRS tuned on European genomes can lose much of its predictive power when applied to African-ancestry genomes, even when the same variants are present.
Can trans-ancestry methods rescue portability?
To probe whether better statistical modeling could recover performance, the authors turned to JointPRS, a data-adaptive framework designed to improve GRS portability by leveraging genetic correlations between populations. In practical terms, they combined the large European MS GWAS with a smaller African-ancestry GWAS from the Million Veteran Program and asked JointPRS to re-weight variants for predicting MS in African-ancestry individuals. This secondary analysis markedly improved stratification: after applying JointPRS, MS prevalence in the AFR group rose more consistently across quintiles, and the trend became statistically significant (p = 0.004 in unadjusted analyses). In multivariable models, the odds of MS in the top vs bottom quintile increased to 3.02 (95% CI 1.00–8.95), as shown in the bottom-right panel of the forest plot on page 14. Although overall discrimination still lagged behind that seen in Europeans, this result demonstrates that explicitly integrating ancestry-specific GWAS data—even when modest in size—can substantially enhance risk prediction in underrepresented groups.
Why prediction is harder in African ancestry populations
The performance gap observed here is not unique to MS; it reflects structural issues in human genetics. Existing GWAS heavily over-represent European populations, so effect sizes and even the lead variants at a locus may not translate cleanly to African genomes, where patterns of linkage disequilibrium and allele frequencies differ. For admixed African Americans, local ancestry adds another layer: key MS risk alleles in the HLA region appear to reside disproportionately on European ancestry segments, and a standard GRS that assigns one effect size per variant cannot capture ancestry-specific effects at each locus. On top of this, African populations carry greater overall genetic diversity, including variants that are rare or absent in Europeans, and gene–environment interactions may differ across populations. The authors note that even within European-ancestry participants, GRS performance in All of Us was somewhat attenuated compared with more homogeneous cohorts such as the UK Biobank, likely reflecting the broader sociodemographic and environmental diversity in All of Us. Together, these factors argue that portability is not a simple technical fix; it requires both richer data and ancestry-aware modeling.
What this means for equitable precision medicine
From a translational perspective, this study is both a warning and a roadmap. It cautions against the uncritical use of European-derived GRS in diverse clinical populations: while such scores can meaningfully stratify risk in European and Latino/admixed individuals, they may provide little benefit—and potentially misleading reassurance or alarm—for people of African ancestry. At the same time, the JointPRS results in this paper show that even relatively small non-European GWAS can substantially improve prediction when combined wisely with large European datasets. The path forward is clear but demanding: expand genetic studies in underrepresented populations, routinely validate GRS across ancestries and cohorts, and adopt tools that model ancestry and local ancestry explicitly. For MS and other complex diseases, this is not only a technical challenge but an equity imperative: the promise of personalized medicine will only be realized when risk prediction tools work reliably for everyone, not just for those whose genomes have historically been most studied.
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:
Rivier, C. A., Xu, L., Clocchiatti-Tuozzo, S., Zhao, H., Ohno-Machado, L., Hafler, D. A., ... & Longbrake, E. E. (2025). Differential results of genetic risk scoring for multiple sclerosis in European and African American populations. Multiple Sclerosis Journal, 31(11), 1304-1313.
