Ancestry-Dependent Performance of a Multiple Sclerosis Polygenic Risk Score in the All of Us Research Program
Multiple sclerosis (MS) is a chronic, immune-mediated disorder of the central nervous system in which both environmental exposures and inherited genetic variation contribute to disease susceptibility. Over the last decade, large genome-wide association studies (GWAS) have established MS as a quintessential polygenic trait: hundreds of common variants each exert modest effects, but collectively can meaningfully shift an individual’s liability to disease. This framework has motivated the use of polygenic risk scores (PRS)—weighted sums of risk alleles across many loci—as quantitative biomarkers to stratify individuals by genetic predisposition, potentially informing prevention strategies, cohort enrichment for clinical trials, and etiologic research. The key unresolved issue is generalizability: most MS GWAS have been conducted in populations of predominantly European ancestry, raising well-known concerns that PRS derived from these studies may underperform in other ancestries due to differences in linkage disequilibrium structure, allele frequencies, genetic architecture, and gene–environment interactions.
Data Source and Study Design
Rivier and colleagues addressed this question using the All of Us Research Program, a U.S.-based initiative explicitly designed to increase diversity in biomedical research. The investigators conducted a cross-sectional analysis leveraging prospectively collected All of Us data (2018–2023), restricting the cohort to participants with both whole-genome sequencing (WGS) and electronic health record (EHR) data. Genetic ancestry was assigned centrally by All of Us using principal components analysis anchored to diverse reference panels, and the analysis focused on the three largest strata available: European, African, and Latino/admixed American (L/A). Importantly, the authors randomly sampled European and African participants to match the L/A sample size (32,428 per group), improving comparability of effect estimation across ancestry strata while keeping the analytical framework consistent.
Construction of the MS Polygenic Risk Score
The exposure of interest was a PRS constructed from 282 independent MS-associated single nucleotide polymorphisms (SNPs) meeting conventional PRS-quality criteria: common variation (minor allele frequency >1%), biallelic loci, genome-wide significant MS associations, and low inter-variant correlation (r² < 0.1) to reduce redundancy. Each participant’s score was computed as the sum of risk allele counts weighted by reported effect sizes, then normalized within ancestry stratum and categorized into quintiles (very low to very high risk). This quintile-based design enables intuitive risk partitioning and supports trend testing while avoiding overinterpretation of absolute PRS units across ancestries. The primary outcome—MS prevalence—was ascertained via EHR diagnostic coding using ICD-10 and SNOMED concept mappings, capturing both prevalent and incident diagnoses relative to baseline assessment.
Cohort Characteristics and Baseline Disease Frequency
From 413,457 All of Us participants, 173,153 met inclusion criteria (both WGS and EHR available), with a mean enrollment age of ~52 years and a female majority (60%). MS cases were uncommon overall, consistent with population prevalence expectations; across the three matched ancestry samples, the crude MS case proportions were 1.0% in Europeans (327/32,428), 0.56% in Africans (183/32,428), and 0.46% in L/A (150/32,428). Table 1 (page 11) further indicates that MS cases were disproportionately female (77.3% among MS vs 60.0% among non-MS) and showed statistically significant differences across ancestry categories, reflecting both biology and ascertainment patterns that can accompany EHR-based phenotyping. These baseline distributions are consequential because low absolute prevalence places an upper bound on the clinical utility of PRS as a stand-alone screening tool even when relative risk gradients are detectable.
Main Findings: Stratification Works in European and Latino/Admixed, Not African Ancestry
The central result is an ancestry-dependent divergence in PRS performance. In unadjusted analyses, the proportion of MS cases increased across PRS quintiles in Europeans and in L/A participants, while the African ancestry group showed a weaker and less consistent pattern (Figure 1, page 12). In multivariable logistic regression adjusting for age, sex, and the first four genetic principal components, the PRS was strongly associated with MS risk in Europeans: participants in the highest versus lowest PRS quintile had an odds ratio (OR) of 2.41 (95% CI 1.69–3.50) with a highly significant trend test. L/A participants showed a comparable gradient, with the highest versus lowest quintile yielding OR 2.56 (95% CI 1.45–4.78) and a significant trend. In contrast, the African ancestry stratum did not reach statistical significance for risk partitioning (highest vs lowest quintile OR 1.45 (95% CI 0.95–2.25); trend p = 0.10), and the confidence interval suggests uncertainty around a modest effect at best (Figure 2, page 12). Collectively, these data demonstrate that a PRS built from loci discovered largely in European GWAS can transfer reasonably to L/A participants in this dataset, but does not provide robust stratification among African ancestry participants under the authors’ modeling assumptions.
Interpretation: Why Does Transferability Fail for African Ancestry?
The authors’ discussion aligns with a well-established statistical genetics narrative: PRS portability is constrained by ancestry-related differences in genetic architecture and genomic correlation structure. African populations typically exhibit greater haplotype diversity and different linkage disequilibrium patterns, which can degrade the tagging efficiency of European-discovered variants and impair the fidelity of effect-size transfer. Additionally, allele frequencies for risk variants—and the true causal variants they tag—may differ, and some MS-relevant loci may be ancestry-specific or have heterogeneous effect sizes across populations. The paper also highlights gene–environment interaction as a plausible contributor: if environmental exposures relevant to MS (e.g., latitude-related vitamin D effects, smoking patterns, infectious history, socioeconomic-linked stressors) differ by population and modulate genetic liability, then effect estimates derived in one context may not transport. The implication is not that genetics is irrelevant in African ancestry MS risk, but rather that the current PRS representation of that genetics is incomplete or miscalibrated due to underrepresentation in discovery GWAS and limited fine-mapping across diverse ancestries.
Implications, Limitations, and Next Steps for Equitable Precision Neurology
Scientifically, this study reinforces a practical rule: PRS validity is population-contingent, and using an unvalidated score across ancestries risks attenuated performance and inequitable downstream applications. Clinically, the authors appropriately caution that MS prevalence is low, limiting the utility of PRS as a screening instrument even in groups where odds ratios are significant; PRS is more defensible as one component of a multivariable risk model incorporating environmental, behavioral, and clinical correlates. Methodologically, several limitations merit emphasis: EHR-based phenotype ascertainment can introduce misclassification; cross-sectional modeling blends incident and prevalent cases; and even within “L/A” or “African” categories there is substantial heterogeneity that may obscure substructure-specific signal. Nonetheless, the work provides an empirical benchmark using a major U.S. biobank-scale dataset and directly supports a clear agenda: expand non-European representation in MS GWAS, perform ancestry-aware fine-mapping, and develop or recalibrate PRS using multi-ancestry methods to improve portability. In this sense, the study is less an endpoint than a diagnostic assessment—showing precisely where current PRS approaches succeed, where they fail, and why inclusive genomics is not optional if polygenic prediction is to become a credible component of precision neurology.
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., Payabvash, S., Zhao, H., Hafler, D. A., Falcone, G. J., & Longbrake, E. E. (2024). Differential Results of Polygenic Risk Scoring for Multiple Sclerosis in European and African American Populations. medRxiv, 2024-06.
