Prostate cancer risk-stratification using proteomic estimation of KLK3 improves accuracy in screening applications: a population-based cohort study in the UK Biobank

Background The prostate specific antigen (PSA) test is the most used clinical tool for prostate cancer risk stratification. PSA-based screening remains controversial due to modest predictive power (ROC AUC<0.75), high false-positive rates and racial disparities. Here, we evaluated the prostate cancer risk-stratification potential for the KLK3 protein, measured by Olink Explore-3072, in the UK Biobank. Methods 19,364 men in the UK Biobank's proteomics dataset were cancer-free at assessment centre visit. We used logistic regression to evaluate potential of KLK3 to predict prostate cancer within 2, 5, and 10 years of recording, as an independent predictor and with age and genetic risk score. Prostate cancer cases were classified by severity based on clinical action taken post-diagnosis. All predictive models were performed under 5-fold cross validation, and diagnostic accuracy statistics reported for the test set. Findings KLK3 was strongly associated with prostate cancer incidence (HR per SD: 3.00 (2.87 - 3.13), p<2e-16). ROC AUC for a 2-year prediction horizon was 0.918 (0.906 - 0.93), reducing to 0.854 (0.848 - 0.86) over 10 years. 10-year ROC AUC was stronger in individuals of European ancestry (0.857 (0.851 - 0.863)) than African (0.801 (0.762 - 0.841)) or South Asian (0.795 (0.727 - 0.862)) ancestry. Power to predict highly aggressive cancer cases within 2 years of recording was strong (ROC AUC 0.928 (0.919 - 0.937)) but weaker for a 10-year period (0.87 (0.865 - 0.876)). Inclusion of age and genetic risk score provided small improvements in individuals of primarily genetic European ancestry, but no improvement in African or South Asian ancestry. Interpretation The ROCAUC values reported are superior to those seen previously for the PSA test. We demonstrate high predictive accuracy across 2-, 5- and 10-year windows. Findings were consistent for low and high-risk cases. These findings suggest that proteomic PSA measurements may be helpful in prostate cancer risk stratification, while highlighting the need for improved predictive models across diverse ancestral groups.