Development and Validation of an 18-Gene Urine Test for High-Grade Prostate Cancer

Abstract

Importance: Benefits of prostate cancer (PCa) screening with prostate-specific antigen (PSA) alone are largely offset by excess negative biopsies and overdetection of indolent cancers resulting from the poor specificity of PSA for high-grade PCa (ie, grade group [GG] 2 or greater).

Objective: To develop a multiplex urinary panel for high-grade PCa and validate its external performance relative to current guideline-endorsed biomarkers.

Design, setting, and participants: RNA sequencing analysis of 58 724 genes identified 54 markers of PCa, including 17 markers uniquely overexpressed by high-grade cancers. Gene expression and clinical factors were modeled in a new urinary test for high-grade PCa (MyProstateScore 2.0 [MPS2]). Optimal models were developed in parallel without prostate volume (MPS2) and with prostate volume (MPS2+). The locked models underwent blinded external validation in a prospective National Cancer Institute trial cohort. Data were collected from January 2008 to December 2020, and data were analyzed from November 2022 to November 2023.

Exposure: Protocolized blood and urine collection and transrectal ultrasound-guided systematic prostate biopsy.

Main outcomes and measures: Multiple biomarker tests were assessed in the validation cohort, including serum PSA alone, the Prostate Cancer Prevention Trial risk calculator, and the Prostate Health Index (PHI) as well as derived multiplex 2-gene and 3-gene models, the original 2-gene MPS test, and the 18-gene MPS2 models. Under a testing approach with 95% sensitivity for PCa of GG 2 or greater, measures of diagnostic accuracy and clinical consequences of testing were calculated. Cancers of GG 3 or greater were assessed secondarily.

Results: Of 761 men included in the development cohort, the median (IQR) age was 63 (58-68) years, and the median (IQR) PSA level was 5.6 (4.6-7.2) ng/mL; of 743 men included in the validation cohort, the median (IQR) age was 62 (57-68) years, and the median (IQR) PSA level was 5.6 (4.1-8.0) ng/mL. In the validation cohort, 151 (20.3%) had high-grade PCa on biopsy. Area under the receiver operating characteristic curve values were 0.60 using PSA alone, 0.66 using the risk calculator, 0.77 using PHI, 0.76 using the derived multiplex 2-gene model, 0.72 using the derived multiplex 3-gene model, and 0.74 using the original MPS model compared with 0.81 using the MPS2 model and 0.82 using the MPS2+ model. At 95% sensitivity, the MPS2 model would have reduced unnecessary biopsies performed in the initial biopsy population (range for other tests, 15% to 30%; range for MPS2, 35% to 42%) and repeat biopsy population (range for other tests, 9% to 21%; range for MPS2, 46% to 51%). Across pertinent subgroups, the MPS2 models had negative predictive values of 95% to 99% for cancers of GG 2 or greater and of 99% for cancers of GG 3 or greater.

Conclusions and relevance: In this study, a new 18-gene PCa test had higher diagnostic accuracy for high-grade PCa relative to existing biomarker tests. Clinically, use of this test would have meaningfully reduced unnecessary biopsies performed while maintaining highly sensitive detection of high-grade cancers. These data support use of this new PCa biomarker test in patients with elevated PSA levels to reduce the potential harms of PCa screening while preserving its long-term benefits.

Figure 1.. Biomarker Discovery and Development of the MyProstateScore 2.0 Urinary Test for High-Grade Prostate Cancer

A, Discovery and nomination of candidate biomarkers for the multiplex urinary panel. Biomarker discovery was performed using RNA sequencing data from 220 benign prostates, 71 with cancers of grade group 1, and 484 with cancers of grade group 2 or greater available through the Cancer Genome Atlas, the Genotype-Tissue Expression portal, and the University of Michigan. A total of 72 markers met predefined criteria. Of these, quantitative polymerase chain reaction probes could not be successfully designed for 19, and 9 genes were highly cross-correlated, resulting in exclusion from the candidate panel. The remaining 44 transcripts meeting nomination criteria were supplemented with 10 curated genes to yield a 54-gene candidate panel. B, Development of the optimal 18-gene model for high-grade cancer. To avoid multicollinearity in regression models, highly correlated variables were identified and removed with a stepwise procedure. We assessed 3 model-building approaches: (1) logistic regression with stepwise feature selection, (2) logistic regression with recursive feature elimination, and (3) regularized logistic regression with elastic net. Performance of each model-building approach was quantified as the area under the receiver operating characteristic curve on repeated cross-validation (10-fold cross-validation repeated 3 times) with upsampling of the minor class to yield balanced classes. Elastic net modeling yielded the highest median area under the curve and was used for development. Using an ensemble approach, the development set was randomly divided into 4 partitions, and the model yielding the highest area under the curve was identified for each partition. This approach was repeated 10 times with different random seeds, yielding 40 elastic net models in total. For each candidate gene, the frequency of model inclusion and importance to high-grade prostate cancer detection was tabulated across models. Based on analysis of optimal feature size and technical features of the OpenArray platform (Thermo Fisher Scientific), the 17 biomarkers providing optimal discriminative accuracy for prostate cancer of grade group 2 or greater were included with standard clinical variables and the normalization gene KLK3 in the MyProstateScore 2.0 model (without prostate volume) and MyProstateScore 2.0 Plus model (with prostate volume). Models were calibrated and internally cross-validated prior to external validation.

Figure 2.. Calibration Curves for High-Grade Prostate Cancer for MyProstateScore 2.0 (MPS2) and MPS2 Plus Prostate Volume (MPS2+) in the External Validation Cohort

Figure 3.. Decision Curve Analysis for High-Grade Prostate Cancer in the External Validation Cohort

A, Decision curve analysis plots for net clinical benefit of prebiopsy testing with prostate-specific antigen (PSA) alone, the Prostate Cancer Prevention Trial risk calculator, Prostate Health Index (PHI), derived multiplex 2-gene model, derived multiplex 3-gene model, MyProstateScore (MPS), MPS2, and MPS2+ compared with baseline approaches of biopsy all or biopsy none. The threshold probability (x-axis) reflects how the patient and clinician value potential clinical outcomes. For example, a threshold probability of 5% applies to patients that would choose to pursue biopsy if their risk of high-grade cancer is 5% or higher. For high-grade prostate cancer, a 5% threshold probability represents a risk-averse population, such as younger men with a long life expectancy. At a practice level, this implies that the clinician would be willing to perform as many as 20 biopsies to detect an additional high-grade cancer. At the other end of the spectrum, a threshold probability of 20% applies to patients that would choose to pursue biopsy only if their risk of high-grade cancer was 20% or greater. Such a population strongly values avoiding biopsy and is willing to accept a higher risk of delayed detection of high-grade cancer. The unit of net benefit (y-axis) is true positives. A net benefit of 0.15 is equivalent to an approach in which 15 patients per 100 are directed to biopsy based on use of the test, and all 15 patients are found to have high-grade cancer. As illustrated in the figure, the MPS2 and MPS2+ models provided the highest net benefit across the range of clinically pertinent threshold probabilities (5% to 20%). B, Decision curve analysis plots illustrating the net reduction in biopsies performed per 100 patients without missing a single diagnosis of cancer of grade group 2 or greater based on prebiopsy testing with PSA alone, the Prostate Cancer Prevention Trial risk calculator, PHI, the derived multiplex 2-gene model, the derived multiplex 3-gene model, MPS, MPS2, and MPS2+ compared with a baseline approach of biopsying all patients. The MPS2 and MPS2+ models provided the largest net reduction in biopsies performed across clinically pertinent threshold probabilities.