This study compared single-ancestry (European), multi-ancestry (MAMA), and cross-ancestry Bayesian (PRS-CSx) polygenic risk scores for Alzheimer's disease (AD) across individuals of African, Amerindian, and European ancestry in two cohorts: ADSP (n=19,398; 7,111 AD cases, 12,287 controls) and HABS-HD (n=2,559). Models were evaluated for AD risk prediction, cognitive function, CSF/plasma biomarkers, and neuropathological burden within the A/T/N framework.
The cross-ancestry Bayesian PRS-CSx model showed the strongest performance in non-European groups: OR 1.71 (95% CI 1.52–1.93) in African ancestry and OR 1.29 (95% CI 1.21–1.37) in Amerindian ancestry, with significantly higher R² than single-ancestry models (ΔR² p=4.6×10⁻⁷ vs. EUR PRS in African ancestry). Single-ancestry EUR PRS showed no significant association in African ancestry participants (OR 1.08, p=0.22). PRS-CSx also linked to lower CSF Aβ42, poorer cognition across all three domains, and the most severe CERAD/Thal/Braak neuropathology categories. An AD latent variable (pTau181 + Aβ PET) was significantly associated with PRS-CSx in the full HABS-HD cohort (SEM CFI=0.983, RMSEA=0.037).
- Non-European GWAS training datasets were substantially smaller than European datasets, limiting power for ancestry-specific models. - South and East Asian ancestry groups were too small for inclusion, restricting generalizability beyond African, Amerindian, and European groups. - Genotype–environment interactions were not modeled, and all cohorts were US-based, limiting global transportability.
European-only AD polygenic risk scores should not be used for risk stratification in patients of African or Amerindian ancestry, as they fail to predict AD risk in these groups. Clinicians and researchers working with diverse populations should use cross-ancestry Bayesian PRS models (e.g., PRS-CSx) combined with ancestry normalization to improve equitable risk prediction.
Explore related topics
