Evaluation of functional annotation-informed and ancestry-specific polygenic risk scores for ischemic stroke

Ischemic stroke (IS) is a leading cause of morbidity and mortality, and predicting events remains challenging. Polygenic risk scores (PRSs) aggregate genetic variants, but performance varies across ancestries and remains modest, particularly in non-European populations.

We compared Bayesian PRS methods incorporating functional annotations and expanded linkage disequilibrium reference panels for IS prediction. PRSs were generated using PRS-CS with HapMap3 and TagIt panels, and SBayesRC integrating functional priors. Scores were optimized in 1,403 European ancestry (EA) and 6,342 African ancestry (AA) participants from the Reasons for Geographic and Racial Differences in Stroke (REGARDS) study and independently validated in All of Us participants (78,155 EA and 29,594 AA participants), a REGARDS holdout cohort (11,459 EA participants), and the Genetics of Hypertension Associated Treatments (GenHAT) study (6,908 AA participants). Predictive performance was evaluated using logistic regression adjusted for age, sex, the top 10 genetic principal components, and antihypertensive treatment assignment where applicable. Performance was assessed using odds ratios per standard deviation (OR/SD), liability-scale R², and area under the curve (AUC).

Among EA participants, SBayesRC demonstrated the largest effect size and liability R² in the REGARDS holdout validation cohort (OR/SD = 1.24, 95% CI 1.12-1.36, R² = 2.42%). PRS-CS using the expanded TagIt panel (OR/SD = 1.11, 95% CI 1.00-1.22) showed numerically stronger performance than the HapMap3 panel (OR/SD = 1.09, 95% CI 0.98-1.20). Performance was consistent across the All of Us EA cohort, with OR/SD ranging from 1.10 to 1.16 across methods. Among AA participants, associations and predictive metrics were attenuated, with no significant associations observed in All of Us. A modest association was observed for PRS-CS-TagIt in GenHAT (for an all-stroke outcome). PRSs modestly improved discrimination beyond covariate models among EA participants (AUCs up to 66.6%), but showed little to no improvement among AA participants (AUCs 62.7%-66.7%). High PRS percentile groups (e.g., top 2% versus 98%) showed modest stroke enrichment with high specificity and limited sensitivity.

Functional annotation-informed PRSs and expanded reference panels modestly improved prediction, particularly among EA participants. Limited performance among AA participants underscores the need for larger, more diverse genomic datasets and integrative models incorporating genetic, clinical, lifestyle, and social determinants of health.