Cancer ResearchResearch PaperOpen Access

Urine RNA Test Detects Bladder Cancer with 95% Sensitivity and Predicts Treatment Response

A Stanford liquid biopsy using urine cell-free RNA detects localized bladder cancer with 95% sensitivity and predicts who will respond to BCG therapy.

Sunday, October 4, 2026 1 view
Published in Nat Med
A clear urine sample collection cup next to a sequencing data readout on a laptop screen in a clinical laboratory setting

Summary

Bladder cancer is often watched for years with repeated invasive camera exams, and the standard urine cell test misses most low-grade tumors. Researchers built uRARE-seq, which reads fragments of RNA floating in urine and picks up signals from bladder, kidney and prostate tissue. Tested on 683 urine samples, it detected localized bladder cancer in 95% of cases at 90% specificity, far better than urine cytology (96% versus 34% in paired samples). It also found leftover cancer after treatment and separated patients cured by surgery from those cured by BCG immunotherapy. Before treatment, people who later responded to BCG had urine RNA showing a pre-existing immune attack on the tumor, while nonresponders showed fast-dividing tumor signals. A model built on this predicted BCG response with an AUC of 0.93. Prospective studies are still needed.

Detailed Summary

Bladder cancer is common, and patients need frequent surveillance with invasive cystoscopy, which is costly and burdensome. Urine cytology is the usual noninvasive option, but it misses many tumors, especially low-grade ones. Urine touches bladder tumors directly, so it is a natural liquid biopsy, yet cell-free RNA (cfRNA) in urine has mostly been studied one or a few genes at a time. This Nature Medicine study asked whether broad RNA profiling could detect bladder cancer and reveal how a tumor will behave.

The team adapted a plasma method into uRARE-seq, which uses random priming and hybrid capture to sequence the short, highly fragmented RNA in urine. cfRNA yield and gene expression were stable for up to 48 hours after collection and after one freeze-thaw cycle. Urine cfRNA was enriched for epithelial transcripts, while urinary sediment was mostly blood-cell RNA. In 48 healthy controls it carried bladder, kidney and prostate transcripts, with prostate genes appearing only in men. The authors then identified 3,834 rarely expressed genes in controls and built a 4,782-gene capture panel.

In a training cohort of 100 controls (including people with hematuria or urinary symptoms) and 151 patients (123 nonmuscle-invasive, 28 muscle-invasive), a machine-learning detection model reached an AUC of 0.97 and 95% sensitivity at 90% specificity. Sensitivity was 86% for low-grade Ta tumors and 100% for carcinoma in situ, high-grade T1 and T2 disease. The analytical limit of detection was 0.05% tumor fraction. In 93 paired samples, uRARE-seq found 95.7% of cancers versus 34.4% for cytology, and 90% of low-grade Ta tumors versus none. It also beat tumor-naive urine DNA testing (99% versus 85%). A locked-threshold validation cohort (142 patients, 140 controls) performed nearly identically, and field-effect mutations did not affect results.

Separate models predicted high grade (AUC 0.84 in training and validation) and muscle invasion (AUC 0.82 in training, 0.75 in validation). Among clinically low-grade tumors, higher grade-model scores signaled greater recurrence risk. In 36 patients with intermediate- or high-risk nonmuscle-invasive disease treated with surgery plus BCG, no patient without recurrence had detectable cfRNA after BCG, while 92% of those with high-grade recurrence did. Post-BCG positivity carried a hazard ratio of 41.17 for recurrence (P<0.0001). The assay separated molecular complete response after surgery (n=10), after BCG (n=15) and residual disease (n=11).

Pretreatment urine from eventual BCG responders was enriched for interferon, TNF, inflammatory and T cell signatures, with more T cell receptor clones (P=0.01) and greater diversity (P=0.04). Nonresponders showed mitotic spindle and G2M proliferation programs. A model built on these features, in pretreatment urine from 114 patients, predicted BCG versus chemotherapy response with an AUC of 0.93 and tracked with recurrence risk. This points to a pre-existing antitumor immune response as a prerequisite for BCG efficacy.

The findings are promising but not yet practice-changing. The authors stress that prospective studies are needed to establish clinical utility. The MRD and response-prediction cohorts were small, the muscle-invasion model dropped to an AUC of 0.75 on validation, and molecular response was defined using the same assay being evaluated.

Key Findings

  • Detected localized bladder cancer with 95% sensitivity at 90% specificity (AUC 0.97) in a training cohort of 100 controls and 151 patients, with near-identical results in an independent validation cohort of 142 patients and 140 controls
  • In 93 paired samples, uRARE-seq detected 95.7% of cancers versus 34.4% for urine cytology, and 90% (28 of 31) of low-grade Ta tumors versus none for cytology
  • Outperformed tumor-naive urine DNA testing (99% versus 85% sensitivity, P<0.001) and matched tumor-informed DNA testing (96% for both), without being affected by field-effect mutations
  • Analytical limit of detection was 0.05% tumor fraction, and performance was not linked to sex, RNA concentration or sequencing depth
  • Among 36 BCG-treated patients, 92% with high-grade recurrence still had detectable tumor cfRNA after BCG versus none without recurrence (hazard ratio 41.17, P<0.0001)
  • Pretreatment urine from BCG complete responders showed immune signatures (interferon, TNF, T cell) and greater T cell receptor clonality (P=0.01) and diversity (P=0.04), while nonresponders showed proliferation signatures
  • A pretreatment model in 114 patients predicted BCG versus chemotherapy response with an AUC of 0.93 and was strongly associated with recurrence risk

Methodology

The authors adapted the plasma RARE-Seq hybrid-capture method into uRARE-seq and tested preanalytic variables such as storage, pH, preservatives and freeze-thaw. They profiled 683 urine samples overall and built a 4,782-gene capture panel from genes rarely expressed in control urine. Elastic net machine-learning models for detection, grade and muscle invasion were trained with 10-fold cross-validation and then applied with locked thresholds to an independent validation cohort. MRD and treatment-response analyses used 36 patients with nonmuscle-invasive bladder cancer treated with surgery plus BCG, with gene set enrichment, T cell receptor repertoire analysis and Kaplan-Meier/log-rank statistics, and comparisons to cytology and urine tumor DNA assays.

Study Limitations

The authors state that prospective studies are needed to assess clinical utility, and the data are largely from retrospective, single-assay analyses. The MRD and BCG-response cohorts were small (36 patients for MRD, 114 for response prediction), and the muscle-invasion model performed more modestly in validation (AUC 0.75). Conflict-of-interest disclosures were not included in the supplied text.

Enjoyed this summary?

Get the latest longevity research delivered to your inbox every week.

Enter your email to subscribe: