Methodological Developments for Advancing Genetic Risk Prediction Across Populations

Abstract

This thesis develops and evaluates statistical methods to improve genetic association and risk prediction across diverse populations and variant types. In Chapter 1, I investigate optimal strategies for multi-ancestry genome-wide association studies (GWAS), comparing pooled analysis, fixed-effect meta-analysis, and MR-MEGA under both fixed- and mixed-effects frameworks. Using extensive simulations across five ancestry groups, varying sample size configurations, admixture levels, and both continuous and binary traits, as well as analyses of 13 traits in the All of Us Research Program and UK Biobank, I show that pooled analysis consistently yields higher power than meta-analytic approaches while maintaining appropriate type I error in realistic scenarios. A theoretical framework links these gains to allele frequency differences between ancestry groups, supporting pooled analysis as a robust and scalable choice for multi-ancestry GWAS. Chapter 2 extends Mendelian risk prediction models to allow for variant-specific penetrance. Focusing on BRCA1/2, I relax the common assumption of homogeneous gene-level risk by incorporating penetrances defined for pathogenic sequence variants grouped into breast cancer clustering regions (BCCR), ovarian cancer clustering regions (OCCR), and other regions, yielding the Fam3PRO-variant model. Simulation studies demonstrate good calibration, high discrimination, and improved region-specific carrier probability estimates relative to gene-level models, even under realistic patterns of underreported family history. Validation in two large iii clinical cohorts shows that Fam3PRO-variant maintains gene-level performance comparable to the original Fam3PRO while providing more granular and clinically informative region-specific risk estimates. Chapter 3 examines polygenic risk scores (PRS) for four common cancers across ten large cohorts, totaling >1,000,000 individuals from multiple ancestries and age ranges, including

300,000 non-European participants. Comparing four PRS construction methods, I quantify the impacts of genetic distance, sample overlap with discovery GWAS, case-control imbalance, and ancestry on PRS transportability and effect size estimates. I show that the bias induced by validating PRS in a typical overlapping cohort is modest relative to population-specific attenuation, suggesting that carefully characterized overlapping cohorts can be judiciously leveraged to increase total validation sample size. The results highlight substantial attenuation of PRS effects in African American cohorts relative to European populations and demonstrate that more sophisticated genome-wide methods can enhance predictive performance but are more sensitive to study design biases. Collectively, these chapters provide methodological guidance for designing multi-ancestry GWAS and implementing gene- and genome-based risk prediction models that are more accurate and equitable across diverse populations.