Abstract
OBJECTIVE: To evaluate the benefit of combining polygenic risk scores with the QCancer-10 (colorectal cancer) prediction model for non-genetic risk to identify people at highest risk of colorectal cancer. DESIGN: Population based cohort study. SETTING: Data from the UK Biobank study, collected between March 2006 and July 2010. PARTICIPANTS: 434 587 individuals with complete data for genetics and QCancer-10 predictions were included in the QCancer-10 plus polygenic risk score modelling and validation cohorts. MAIN OUTCOME MEASURES: Prediction of colorectal cancer diagnosis by genetic, non-genetic, and combined risk models. Using data from UK Biobank, six different polygenic risk scores for colorectal cancer were developed using LDpred2 polygenic risk score software, clumping, and thresholding approaches, and a model based on genome-wide significant polymorphisms. The top performing genome-wide polygenic risk score and the score containing genome-wide significant polymorphisms were combined with QCancer-10 and performance was compared with QCancer-10 alone. Case-control (logistic regression) and time-to-event (Cox proportional hazards) analyses were used to evaluate risk model performance in men and women. RESULTS: Polygenic risk scores derived using the LDpred2 program performed best, with an odds ratio per standard deviation of 1.584 (95% confidence interval 1.536 to 1.633), and top age and sex adjusted C statistic of 0.733 (95% confidence interval 0.710 to 0.753) in logistic regression models in the validation cohort. Integrated QCancer-10 plus polygenic risk score models out-performed QCancer-10 alone. In men, the integrated LDpred2 model produced a C statistic of 0.730 (0.720 to 0.741) and explained variation of 28.2% (26.3 to 30.1), compared with 0.693 (0.682 to 0.704) and 21.0% (18.9 to 23.1) for QCancer-10 alone. In women, the C statistic for the integrated LDpred2 model was 0.687 (0.673 to 0.702) and explained variation was 21.0% (18.7 to 23.7), compared with 0.645 (0.631 to 0.659) and 12.4% (10.3 to 14.6) for QCancer-10 alone. In the top 20% of individuals at highest absolute risk, the sensitivity and specificity of the integrated LDpred2 models for predicting colorectal cancer diagnosis was 47.8% and 80.3% respectively in men, and 42.7% and 80.1% respectively in women, with increases in absolute risk in the top 5% of risk in men of 3.47-fold and in women of 2.77-fold compared with the median. Illustrative decision curve analysis indicated a small incremental improvement in net benefit with QCancer-10 plus polygenic risk score models compared with QCancer-10 alone. CONCLUSIONS: Integrating polygenic risk scores with QCancer-10 modestly improves risk prediction over use of QCancer-10 alone. Given that QCancer-10 data can be obtained relatively easily from health records, use of polygenic risk score in risk stratified population screening for colorectal cancer currently has no clear justification. The added benefit, cost effectiveness, and acceptability of polygenic risk scores should be carefully evaluated in a real life screening setting before implementation in the general population.