Bayesian polygenic risk score methods outperformed pruning-and-thresholding for predicting body mass index (overall R² 9.0% vs. 6.6%), with performance varying by race and clinical context.
Observational (n=501,247)
Yes
Do Bayesian Polygenic Risk Scores improve body mass index prediction compared to pruning-and-thresholding methods across diverse populations?
Bayesian polygenic risk scores improve BMI prediction compared to standard methods, but their accuracy remains significantly lower in non-Hispanic Black individuals and specific clinical subgroups, highlighting the need for context-specific evaluation.
Absolute Event Rate: 9% vs 6.6%
Introduction and Objective: Polygenic risk scores (PRS) for obesity are a promising tool for obesity risk prediction, yet their performance may vary across populations and individual-level contexts. While racial/ethnic differences in PRS performance are well documented, less is known about how demographic, lifestyle, and cardiometabolic factors influence PRS performance within and across populations. Methods: We constructed PRS for body mass index (BMI) using genome-wide association study (GWAS) summary statistics from the Genetic Investigation of Anthropometric Traits (GIANT) consortium (N≈1.95M; independent of target samples). PRS were constructed using pruning-and-thresholding (P+T) and Bayesian approaches (PRS-CS and PRS-CSx). Prediction performance was evaluated in the Population Architecture using Genomics and Epidemiology (PAGE) study and eight additional cohorts and biobanks (Max N=501,247). Context-specific performance was examined across strata defined by race/ethnicity, age, sex, smoking status, physical activity, type 2 diabetes (T2D), and hypertension. Results: Across all racial/ethnic groups in PAGE, Bayesian PRS methods outperformed P+T, with PRS-CS showing the most consistent performance (overall R²=9.0% vs. 6.6%). PRS performance varied substantially by race/ethnicity, with higher R² in non-Hispanic White participants (14.0%) compared with non-Hispanic Black participants (7.1%). Beyond race/ethnicity, significant heterogeneity in PRS performance was observed across demographic, lifestyle, and clinical contexts. Prediction accuracy was generally lower among males, older individuals, current smokers, and participants with T2D, with consistent patterns across cohorts. Conclusion: Bayesian PRS methods substantially improve obesity risk prediction; however, PRS performance varies across populations and individual-level contexts. These findings underscore the importance of context-specific evaluation to support equitable and clinically appropriate application of obesity PRS. Disclosure D. Kim: None. Funding AHA (90385)
KIM et al. (Fri,) conducted a observational in Obesity (n=501,247). Bayesian polygenic risk scores (PRS-CS) vs. Pruning-and-thresholding (P+T) polygenic risk scores was evaluated on Prediction performance (overall R²). Bayesian polygenic risk score methods outperformed pruning-and-thresholding for predicting body mass index (overall R² 9.0% vs. 6.6%), with performance varying by race and clinical context.