Predictive performance of polygenic scores for coronary artery disease in multiple genetic ancestries

Abstract

Abstract Background Polygenic scores (PGS) for coronary artery disease (CAD) measure a person's genetic liability for CAD. PGSs for CAD are generally considered sufficiently predictive in individuals of European ancestry, but their performance is attenuated in non-European ancestry groups (AGs). We collated publicly available CAD PGS and benchmarked their performance across several AGs to evaluate their utility for personalized prevention. Methods We queried the “PGS Catalog” to extract standardized odds/hazard ratios (OR/HR) for published CAD PGSs across multiple AGs (European-EUR, African-AFR, Hispanic-HIS, South-Asian-SAS, East-Asian-EAS). We restricted our analysis to PGSs specifically developed for prevalent and/or incident CAD and identified the five best-performing PGSs in each AG. We then computed these PGSs in a sample of 504,096 individuals of EUR (94.5%), AFR (2.6%), SAS (1.9%), EAS (0.7%) and HIS (0.3%) AG and evaluated their performance through logistic regression (covariates: age, sex, family history of CAD, principal components of AG). Results PGS000018 performed best in SAS (1.69[1.56-1.83]), EUR (1.62[1.60-1.64]) and EAS (1.45[1.11-1.89]). PGS001780 was the best performing PGS in AFR (1.18[1.04-1.33], on par with PGS000749) and ranked second in EUR (1.59[1.57-1.61]) and SAS (1.64[1.53-1.76]). Finally, PGS002262 was best in HIS (1.39[1.07-1.81], almost on par with PGS000337) and second in EASs (1.41[1.13-1.75]). Overall, the highest standardized ORs were found for the SAS and the EUR AGs, while the AFR AG had the worst performances. PGSs performed differently than previously reported both in an absolute and relative way, corroborating the need of a standardized comparison. Conclusions There is currently no gold-standard trans-ancestry CAD PGS, but the employment of different AG-specific CAD PGSs can achieve the best possible predictive performance in every individual, an essential prerequisite towards implementing PGSs in CAD personalized prevention protocols. Key messages • No gold standard PGS for CAD exists that can be applied with equal predictive performance across all AGs, which can be partially counterbalance by using the best performing PGS in each AG. • In our analysis, the absolute performance of all PGSs was worse than expected and the best performing PGS for each AG was different from the one reported in the “PGS Catalog” across all AGs.