Characterizing polygenic risk scores for cancer in prospective cohorts: Considering PRS method construction, age of diagnosis, genetic ancestry, and sample overlap

Well-established prospective cancer cohorts provide a unique opportunity to investigate the combined use of polygenic risk scores (PRSs) and non-genetic factors for personalized disease risk prediction. However, their application can vary depending on cohort-specific characteristics. We compared four PRS weighting strategies: PRS-CSx, Lassosum, a multi-ethnic joint analysis of marginal summary statistics (mJAM) forward selection procedure, and standard genome-wide association study (GWAS) derived weights, across eight large, prospective cohorts, totaling over a million subjects from the Multiethnic Cohort (MEC, n=73,139), the Genetic Epidemiology Research on Aging (GERA, n=103,358), the Women's Health Initiative (WHI, n=46,794), the Nurses' Health Studies I (NHS, n=20,195) and II (NHS2, n=16,082), the Health Professionals Follow-up Study (HPFS, n=12,649), the UK Biobank (UKB, n=474,775) and All of Us (AoU, n=317,956). Together, these cohorts represent a uniquely diverse study population spanning multiple racial and ethnic groups, broad age distributions, and both overlapping and non-overlapping samples with discovery GWAS populations. We evaluated PRS performance for four major cancers (breast, prostate, colorectal, and lung) across all weighting methods and contexts. Our findings show that both the choice of PRS construction method and the characteristics of the target cohort substantially influence the characterization of disease risk. By integrating four distinct PRS approaches across eight large and diverse prospective cohorts, this study provides one of the most comprehensive evaluations to date of context-dependent variation and provides guidance for ongoing development of risk prediction models that combine genetic and non-genetic factors.