Risk factors affecting polygenic score performance across diverse cohorts

Area of Science:

• Genetics and Genomics
• Biostatistics
• Personalized Medicine

Background:

• Polygenic scores (PGS) predict complex traits but their performance varies across populations and individuals.
• Factors beyond genetics, such as personal and environmental covariates, can influence PGS accuracy and effect size.

Approach:

• Analyzed body mass index PGS (PGS_BMI) performance across European and African ancestry cohorts, stratifying by 62 covariates.
• Investigated PGS_BMI-covariate interactions and utilized quantile regression to understand effect modification.
• Explored machine learning (neural networks) and genome-wide association study (GWAS) interaction effects to enhance PGS_BMI models.

Key Points:

• 18 out of 62 covariates showed significant, replicable differences in PGS_BMI R², with age, sex, lipids, physical activity, and alcohol being prominent.
• 28 covariates exhibited significant PGS_BMI-covariate interactions, modifying effects by up to 20% per standard deviation.
• Covariates strongly associated with BMI demonstrated the largest R² differences and interaction effects.

Conclusions:

• Covariate effects significantly impact PGS_BMI performance and interpretation across diverse ancestries.
• Machine learning models improved PGS_BMI R² by a mean of 23%, and GxAge GWAS effects improved R² by 7.8%.
• Accounting for non-linear covariate effects and interactions is crucial for improving PGS models and their clinical utility.