Abstract
Background
Alzheimer’s disease (AD), the most common type of dementia, has a complex etiology with a strong genetic component. Many genetic risk variants for AD have been identified including APOE, the largest known genetic risk factor. However, most of this research has examined only broad populations of individuals with European ancestry and there is mounting evidence that effect sizes vary by population. Here, we investigate the transferability of polygenic risk scores (PRSs), including a network‐specific PRS, in the midwestern Amish and across diverse ancestries.
Method
Data from 1,091 Amish adults with AD diagnosis by consensus were considered for analysis from the Collaborative Amish Aging & Memory Project. Genotype data (Illumina GSA and MEGAEX) were imputed using the Haplotype Reference Consortium panel. We also analyzed 15,745 individuals from the Alzheimer’s Disease Sequencing Project (ADSP) r3 with three predominant race/ethnicity groups: African American (AA; n=2,937), Hispanic (n=3,047), and non‐Hispanic White (NHW; n=9,708). An AD network‐specific PRS was constructed by including variants from AD‐implicated molecular networks in the Kyoto Encyclopedia of Genes and Genomes. A comprehensive pruning and thresholding PRS considering all variants was calculated for comparison. Effect estimates from Kunkle et al. (2019) were used. PRS‐only models, sex and age covariate‐only, and full models were constructed for each group and the full ADSP data.
Result
We observed that PRS‐only predictive ability was similar between the comprehensive PRS (AUC=0.56) and the network‐specific PRS (AUC=0.55) in the full ADSP r3 data. Network‐specific PRS performance was similar in the Amish (AUC=0.55). However, we observed better predictive ability with the comprehensive PRS compared to the network‐specific PRS in each of the subgroups (Amish AUC: 0.61 vs. 0.55; NHW: 0.70 vs. 0.53; AA: 0.56 vs. 0.53, Hispanic: 0.61 vs. 0.56). These trends were consistent after inclusion of sex and age covariates.
Conclusion
We demonstrated that a network‐specific PRS performed similarly to a comprehensive PRS in a combined diverse population and the Amish but confers less predictive ability when investigating individual subgroups. Thus, the network‐specific PRS has potential as a component of a risk model in diverse populations because it may successfully account for effects that are consistent across subgroups.