Authors: Kevin J. Liu, William Y. Shi, Monica C. Nesselbush, Nick Phillips, Bogdan Luca, Ravi Bajpai, Isabel Jabara, Kathleen E. Mach, Christian R. Hoerner, Jordan Cheng, Rui Wang, Mohammad Esfahani, Diego Almanza, Emily G. Hamilton, Angela Hui, Vinh La, Grace Holton, Gabriela Rodriguez, Timothy J. Lee, Patrick Mullane, Kris Prado, Alice C. Fan, Eila C. Skinner, Ash A. Alizadeh, Joseph C. Liao, Maximilian Diehn
Published: 2025-04-22
DOI: 10.1158/1538-7445.am2025-3773
Source: Full article
Urine is a promising biofluid for bladder cancer (BLCA) detection because it can be collected non-invasively and is in direct contact with the tumor. Urine cell-free RNA is a potential analyte for BLCA detection and molecular characterization but has not been extensively explored. Here, we develop a novel urine cell-free RNA sequencing method for bladder cancer detection and monitoring. First, we characterize urine cell-free RNA and show that we can detect transcripts from the bladder, kidney, and prostate. Next, we show that BLCA urine cfRNA profiles highly correlate with tumor tissue gene expression (R = 0.91; p < 0.0001). To determine the performance of using urine cfRNA for BLCA detection, we analyze 414 urine samples from 213 non-malignant controls and 201 BLCA patients. The cohort consisted of 160 (80%) non-muscle invasive bladder cancer (NMIBC) and 41 (20%) muscle-invasive bladder cancer (MIBC) patients. NMIBC patients included 59% Ta, 32% T1, 9% CIS, with 28% low-grade (LG) and 72% high-grade (HG). Using machine learning, we develop a BLCA detection model that has a limit of detection of 0.05% for cancer-derived RNA and has an overall performance of 95% sensitivity at 90% specificity. We then validated the locked BLCA detection model in an independent cohort, achieving performance of 94% sensitivity and 88% specificity. The sensitivity of BLCA detection model correlates with grade and stage, ranging from 86% for low-grade Ta, 97% for HG Ta, 100% for HG CIS, and 100% for HG tumors greater than T1. To determine the performance of the BLCA detection model for minimal residual disease, we analyze 37 post-treatment urine samples from patients who received adjuvant Bacillus Calmette-Guérin (BCG) treatment after resection of NMIBC. Patients with residual BLCA cfRNA have high rates of recurrence compared to those without (HR = 42.27; log-rank P < 0.0001). Lastly, we trained two additional machine learning models to: 1) distinguish between low- and high-grade BLCA (AUC = 0.85) and 2) distinguish between NMIBC and MIBC (AUC = 0.82). Together, these results highlight the potential utility of urine cfRNA analysis for BLCA detection, minimal residual disease detection, and subtype distinction.