Abstract
Purpose :
Aggregation of single-marker information into polygenic risk scores (PRS) is a popular glaucoma risk-prediction approach, but its implementation and utility is impeded by the variability of disease genetic architectures in ethnically diverse populations. The purpose of this study was to evaluate the factors that affect the performance of PRS built on genome-wide association study (GWAS) results from cohorts of predominantly European ancestry in populations of African and admixed genetic heritage.
Methods :
Variable selection using Elastic Net models were conducted on markers previously reported as associated with glaucoma in a large GWAS among subjects participating in the UK Biobank cohort. The predictive models built on the selected markers in ethnically homogeneous groups were tested both in UK Biobank subjects (through a 80:20 holdout cross-validation) and in a fully independent cohort of European, African-Caribbean and “other” glaucoma cases and controls (576:287, 298:194 and 124:79, respectively) genotyped on a Human Omni Express Exome 8v1-2 BeadChip (Illumina) and imputed on 1000 Genomes haplotypes.
Results :
Our optimized Elastic Net models effectively built sufficiently sparse pan-ancestral models that performed reasonably well, even in populations of non-European descent, including UK participants of African-Caribbean (Area Under the ROC Curve, AUC=0.65-0.71). These models performed better in the UK Biobank than in the smaller case-control cohorts (AUC difference of 0.05-0.10 in subjects of African and European ancestry), suggesting that the predictive performance of the PRS is in part affected by genotyping coverage and quality. Other factors likely affecting performance include similarity between training and testing datasets, levels of genetic admixture, and models shrinkage and scaling of parameters.
Conclusions :
It may be possible to adapt the current European-driven information for effective PRS-based prediction in multiethnic populations. Future predictive PRS-based models will benefit from more ethnicity-specific GWAS information, but also from finer mapping of risk-associated genetic loci.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.