Abstract
BACKGROUND: Models with polygenic risk scores and clinical factors to predict risk of different cancers have been developed, but these models have been limited by the polygenic risk score-derivation methods and the incomplete selection of clinical variables. METHODS: We used UK Biobank to train the best polygenic risk scores for 8 cancers (bladder, breast, colorectal, kidney, lung, ovarian, pancreatic, and prostate cancers) and select relevant clinical variables from 733 baseline traits through extreme gradient boosting (XGBoost). Combining polygenic risk scores and clinical variables, we developed Cox proportional hazards models for risk prediction in these cancers. RESULTS: Our models achieved high prediction accuracy for 8 cancers, with areas under the curve ranging from 0.618 (95% confidence interval = 0.581 to 0.655) for ovarian cancer to 0.831 (95% confidence interval = 0.817 to 0.845) for lung cancer. Additionally, our models could identify individuals at a high risk for developing cancer. For example, the risk of breast cancer for individuals in the top 5% score quantile was nearly 13 times greater than for individuals in the lowest 10%. Furthermore, we observed a higher proportion of individuals with high polygenic risk scores in the early-onset group but a higher proportion of individuals at high clinical risk in the late-onset group. CONCLUSION: Our models demonstrated the potential to predict cancer risk and identify high-risk individuals with great generalizability to different cancers. Our findings suggested that the polygenic risk score model is more predictive for the cancer risk of early-onset patients than for late-onset patients, while the clinical risk model is more predictive for late-onset patients. Meanwhile, combining polygenic risk scores and clinical risk factors has overall better predictive performance than using polygenic risk scores or clinical risk factors alone.