Examining the Performance of Polygenic Risk Scores for Alzheimer Disease Within and Across Populations Using -Fold Cross-Validation.

Background And Objectives: Alzheimer disease (AD) has a complex etiology with a strong genetic component. Despite mounting evidence that genetic risk effect sizes vary by population, most research on the genetics of AD has examined only data sets of individuals with European ancestry. In this study, we investigate the variable performance and transferability of polygenic risk scores (PRSs) by deriving a PRS from analyses of AD for various race and ethnic categories and applying this across groups using a -fold cross-validation approach.

Methods: After quality control and application of inclusion criteria, we analyzed 11,254 individuals from the Alzheimer's Disease Sequencing Project Release 3 with 3 predominant self-identified race/ethnicity groups: Hispanic individuals (n = 2,207), non-Hispanic Black individuals (n = 2,437), and non-Hispanic White individuals (n = 6,610). For each group, a 5-fold cross-validation approach was used to perform a genome-wide association study (training) and construct PRS estimates (test). This approach allowed for consideration of within-group and across-group PRS performance. Area under the curve (AUC) was calculated for each race/ethnicity-specific PRS within each cross-validation fold. Various PRS criteria were considered, including rare single-nucleotide polymorphisms and the region in calculation, sex-specific PRSs, and different sample sizes of training and test groups, and their effects were examined through goodness-of-fit metrics in reduced and full models.

Results: Across all participant groups, the PRS trained in the same race/ethnicity category as the test group nearly exclusively outperformed the other PRSs with considerable benefit in the Hispanic group. The range of AUCs across iterations was also greater when applying PRSs from other groups. These findings were consistent after inclusion of , sex, and age covariates and through different PRS generation criteria.

Discussion: In this study, we demonstrated the variable performance of AD PRSs within race/ethnicity groups and the weakness in transferability across groups, accounting for different criteria in PRS calculation. This work provided a framework for improving PRS application for AD and highlighted the importance of recruiting diverse populations in genetic studies.