UCLA

Polygenic score (PGS) is a tool for understanding an individual's predisposition to certain diseases or complex traits based on its genetic profile. In the burgeoning era of genomic medicine, PGS has emerged as a promising tool in advancing precision healthcare, demonstrating versatile utility such as patient risk stratification, disease risk prediction, and disease subtyping. However, its real application in clinical settings is limited by its uncertainty, bias, and low portability across diverse populations. For example, an individual may receive different genetic risk reports from different providers, and the score for a non-European individual may be less accurate than for a European individual. To fully understand and partially address these limitations, I first developed a Bayesian method to quantify the uncertainty in PGS at the individual level. I find trait-specific genetic architecture such as larger polygenicity and lower heritability combined with a small training sample size will lead to large uncertainty in PGS estimate, which in turn results in unreliable patient stratification in downstream analysis. Next, I expanded this approach to encompass individuals from varied genetic ancestry backgrounds. I find that the PGS performance varied from individual to individual with genetic distance playing a key role in impacting the performance of PGS; larger genetic distance from training data correlates with higher uncertainty and lower accuracy in testing individuals. These findings highlight the necessity of integrating individual-level PGS metrics in personalized medicine and the need for increasing genetic research diversity to ensure equitable and responsible use of PGS in clinical settings.