Differential performance of polygenic prediction across traits and populations depending on genotype discovery approach

Abstract

Polygenic scores (PGS) are widely used to estimate genetic predisposition to complex traits by aggregating the effects of common variants into a single measure. They hold promise in identifying individuals at increased risk for diseases, allowing earlier screening and interventions. Genotyping arrays, commonly used for PGS computation, are affordable and computationally efficient, while whole-genome sequencing (WGS) offers a more comprehensive view of genetic variation. In this study, we compared PGS derived from arrays and WGS across multiple traits to evaluate differences in predictive performance, portability across populations, and computational efficiency. We computed PGS for 10 traits, representing a range of heritability and polygenicity, in the three largest genetic ancestry groups in All of Us (European, African American, Admixed American), trained on multi-ancestry meta-analyses from the Pan-UK Biobank. Using the clumping and thresholding (C+T) method, we found that WGS-based PGS outperformed array-based PRS for highly polygenic traits but showed differentially reduced accuracy for sparse traits in certain populations. With the LD-informed PRS-CS method, we observed overall improved prediction performance compared to C+T, with WGS outperforming arrays across most non-cancer traits. The results obtained using PRS-CS closely align with those derived from pre-trained models in the PGS Catalog, with prediction achieving better performance using WGS than array genotypes for non-sparse traits. To further investigate factors influencing differential prediction performance between array and WGS, we ran simulations varying the proportions of causal SNPs directly captured by the technologies. These demonstrated that the proportion of causal variants genotyped dramatically affects prediction accuracy. Fine-mapping of empirical data supported this concept but also highlighted the importance of reducing non-informative variants for optimal prediction accuracy. In conclusion, while WGS-based PGS generally offer superior predictive power with PRS-CS, the advantage over arrays is context-dependent, varying by trait, population, and the PGS method. The ability to capture causal variants through these technologies largely drives the prediction accuracy. This study provides insights into the complexities and potential advantages of using different genotype discovery approaches for polygenic predictions across populations and informs on strategies to enhance accuracy.