Multiple polygenic score approach in colorectal cancer risk prediction.

Authors

Jiang SJ, Thomas M, Rosenthal EA, Phipps AI, Sakoda LC, van Duijnhoven FJB, Pellatt AJ, Avery CL, Berndt SI, Bishop DT, Castellví-Bel S, Chan AT, Grant RC, Gignoux C, Gsur A, Gunter MJ, Haiman CA, Hoffmeister M, Jarvik GP, Jenkins MA, ...show more Keku TO, Küry S, Lee JK, Marchand LL, Moreno V, Newcomb PA, Newton CC, Ogino S, Palmer JR, Pearlman R, Qu C, Schoen RE, Um CY, Van Guelpen B, Visvanathan K, Vymetalkova V, White E, Woods MO, Platz EA, Brenner H, Corley DA, Vogelaar IL, Hsu L, Peters U

Affiliations

• University of Washington, Seattle, WA, USA.
• Public Health Sciences Division, Fred Hutchinson Cancer Center, Seattle, WA, USA.
• Department of Medicine, Division of Medical Genetics, University of Washington, Seattle, WA, USA.
• Kaiser Permanente Division of Research, Oakland, CA, USA.
• Division of Human Nutrition and Health, Wageningen University & Research, Wageningen, The Netherlands.
• Intermountain Health, Salt Lake City, UT, USA.
• Department of Epidemiology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA.
• Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institutes of Health, Bethesda, MD, USA.
• Leeds Institute of Cancer and Pathology, University of Leeds, Leeds, UK.
• Gastroenterology Department, Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas (CIBEREHD), Hospital Clínic, Institut d'Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), University of Barcelona, Barcelona, Spain.

Abstract

Recent studies have demonstrated that for various diseases, incorporating polygenic risk scores (PRSs) for other traits and diseases into the PRS-based risk prediction model may improve predictive performance - known as Multiple Polygenic Score (MPS) approach. We aimed to examine whether the MPS approach improves colorectal cancer (CRC) risk prediction. We included 2,187 non-CRC PRSs from the polygenic Score (PGS) Catalog and used machine learning (ML) models to select the most predictive non-CRC PRSs, utilizing individual-level data from 31,257 CRC cases and 33,408 controls. An independent dataset from the Genetic Epidemiology Research in Adult Health and Aging (GERA) cohort (4,852 cases and 67,939 controls) was randomly split into subsets for model estimation and validation. The model combined MPS with two existing CRC-PRSs based on known loci and genome-wide genotyping. We then assessed model performance by calculating the area under the receiver operating curve (AUC) in the validation set and performed 1,000 bootstrapped iterations to evaluate AUC improvements. The ML model selected 337 non-CRC PRSs predictive of CRC risk. Adding MPS to the CRC-PRSs significantly improved AUC by 0.017 (95% CI: 0.011-0.022, p < 0.0001) when combined with known-loci CRC-PRS, 0.005 (95% CI: 0.002-0.007, p = 0.0005) with genome-wide CRC-PRS, and 0.004 (95% CI: 0.002-0.006, p = 0.0005) with both the known loci and genome-wide CRC-PRSs. These findings demonstrate MPS's potential to refine CRC risk prediction models and highlight opportunities for further advancements in risk prediction.