PublishedOctober 2, 2026
JournalNature Medicine
TrialuRARE-seq (non-randomized biomarker study)

Urine Cell-Free RNA Test Detects Bladder Cancer at 95% Sensitivity and 90% Specificity, but Residual-Disease and BCG-Response Findings Rest on Small Cohorts

How this article was produced

How this article was produced

This article was written about a single named primary publication, with AI assistance for literature discovery, drafting and editing. AI tools are not the author and are not responsible for the conclusions.

How the draft was prepared

  • Discovery: PubMed is searched for oncology publications from the previous 14 days in a fixed list of major journals — JAMA Oncology, The Lancet Oncology, Journal of Clinical Oncology, Annals of Oncology, Nature Reviews Clinical Oncology, New England Journal of Medicine, Nature Medicine, Cancer Discovery and Clinical Cancer Research — falling back to a named set of open-access journals only if that list yields nothing. Candidates are ranked by how completely they report quantitative results, and studies already covered are excluded.
  • Source: the article is drafted from the open full text in PubMed Central where one exists. Where only the abstract is available, it is supplemented with cited external sources — the trial registry record, conference abstracts, sponsor releases — and facts drawn from those sources are labelled as such.
  • Numbers are reported exactly as the source gives them, with confidence intervals and p-values, and are not rounded. Data the source does not report is stated as not specified rather than estimated.
  • Subgroup statistics are reported only where the source states them explicitly, and quotations are reproduced verbatim.
  • Benchmarks from other studies are labelled as cross-study comparisons and never presented as results of the trial itself.
  • Validation: every factual claim in the draft is checked against the source, and key claims are spot-checked against independent sources such as the trial registry. Fixable discrepancies are corrected before the draft is handed on for publication.

These steps are instructions followed by an AI system, not checks enforced by software. The validation step is itself performed by AI rather than a person, and its verdict does not by itself stop an article from being published. This describes the current workflow; some earlier articles predate parts of it.

Checks completed before publication

  • The primary publication is recorded in structured form — journal, publication date, DOI, PubMed ID and PMC ID where they exist — rather than only as prose in a reference list.
  • Every cited DOI is confirmed to resolve.
  • Every cited PubMed ID is confirmed to exist in PubMed.
  • The cited DOI is cross-checked against the DOI that PubMed records for that PubMed ID, so a citation cannot point at a different paper from the one the article describes.
  • Trial registry identifiers are format-checked and linked to their ClinicalTrials.gov record.
  • Every reference must carry a resolvable identifier or link, unless it is explicitly labelled as a funding note, press release or background statistic.
  • The article must contain a limitations discussion specific to the study — what it could not establish — not a generic disclaimer.
  • Absolutist claims such as “proves”, “guarantees”, “will cure” or “completely effective” are rejected.
  • HTML markup and entities that would render as literal text — such as an unsupported tag or " " — are rejected outside the reference list.
  • Structural checks confirm the article has a unique URL and complete metadata.

These run automatically on every article. If any of them fails, the build fails and the article is held rather than published.

What these checks do not cover

These checks confirm that the sources are real and correctly cited. They do not confirm that the analysis drawn from them is sound. This article has not been read against its primary source by a human editor, and should be read as a signpost to the primary publication rather than a substitute for it.

Evidence note

This article summarizes research findings and does not provide medical advice or treatment recommendations. Investigational findings, preliminary results, and preclinical evidence should not be interpreted as established clinical care. Always consult the cited primary source, and a qualified clinician, before acting on anything described here.

Corrections

Report a factual error or a sourcing concern through our corrections policy. Substantive corrections are dated and described on the article page. Our editorial policy explains how these articles are sourced, verified, and corrected.

Article Summary
  • uRARE-seq is a tumor-naive urine cell-free RNA (cfRNA) sequencing approach, applied to 683 urine samples from patients with cancer and controls. In the training cohort (100 controls, 151 bladder cancer samples), the detection model reached an AUC of 0.97, with a sensitivity of 95% at 90% specificity.
  • Among 93 paired samples, uRARE-seq detected 89 cases (95.7%) and urine cytology detected 32 (34.4%). In 36 patients with non-muscle-invasive bladder cancer (NMIBC), post-BCG cfRNA positivity was associated with worse high-grade recurrence-free survival (HR = 41.17; log-rank P < 0.0001).
  • This was a non-randomized, non-blinded biomarker study. The residual-disease and treatment-response cohorts were small, the BCG-response signature was derived from 26 patients, and the chemotherapy validation cohort contained only 5 patients with recurrence. The authors state that larger cohorts and prospective clinical trials are needed to define clinical utility.

Why This Study Matters

The authors note that improved noninvasive biomarkers for genitourinary malignancies are needed, "especially in bladder cancer (BLCA), given its rising prevalence and resource-intensive surveillance requirements". Urine is in direct contact with bladder tumors and can be collected noninvasively. Prior urine cfRNA studies mainly focused on a small number of genes; this study profiles thousands of transcripts by sequencing. Urine cytology, the commonly used clinical test, has modest sensitivity, especially for low-grade tumors, per the authors.

1
In the training cohort (100 controls, 151 BLCA samples), the detection model reached an AUC of 0.97, with a sensitivity of 95% at 90% specificity.
2
Among 93 paired samples, uRARE-seq detected 89 cases (95.7%) and urine cytology detected 32 (34.4%), an absolute difference of 61.3%.
3
In 36 patients with NMIBC, post-BCG cfRNA positivity was associated with worse high-grade recurrence-free survival (HR = 41.17; log-rank P < 0.0001).
4
A cfRNA model predicting BCG response achieved an AUC of 0.93 by leave-one-out cross-validation and was associated with recurrence in an independent BCG cohort (n = 57; P = 0.0005).

Study Design

uRARE-seq (urine random priming and affinity capture of cfRNA fragments for enrichment analysis by sequencing) adapts RARE-Seq, a hybrid capture-based sequencing approach originally developed for plasma cfRNA, to urine supernatant. The team applied it to 683 urine samples from patients with cancer and controls. Using 38 noncancer controls they identified 3,834 urine rare abundance genes (uRAGs) and designed a custom capture panel; with supplemental genes the total reached 4,782 genes. A machine-learning (elastic net) detection model was trained on 1,066 preselected genes, of which the model selected 381. The authors state the experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment. ClinicalTrials.gov ID: not specified in the source (non-randomized biomarker study). Funding included the NIH, Stanford ACED Canary Fellowship, Cui Scholar Graduate Fellowship, TRDRP, the Virginia and D.K. Ludwig Fund for Cancer Research and a philanthropic gift.

Patient Population

The training cohort comprised 100 noncancer controls (48 asymptomatic and 52 with hematuria or lower urinary symptoms) and 151 pretreatment BLCA samples (NMIBC = 123; MIBC = 28). An independent validation cohort included 142 patients with BLCA and 140 noncancer controls. Separate cohorts were used for minimal residual disease (MRD; 36 patients with intermediate- and high-risk NMIBC undergoing surgery plus adjuvant BCG induction), a BCG validation cohort (n = 57) and an intravesical chemotherapy cohort (n = 38). Number of study sites: not specified in the source; all listed affiliations are Stanford University units, Resero Bio, and the Veterans Affairs Palo Alto Health Care System.

Detection Results

The study does not designate a formal primary endpoint; detection performance is its lead result. In the training cohort (251 urine samples) the model achieved an AUC of 0.97 with 95% sensitivity at 90% specificity. Applying the locked model to the independent validation cohort "yielded nearly identical performance as in the training cohort" per the text. The analytical 95% limit of detection (LOD95) was 0.05%. Matched urine and tumor RNA were highly correlated (R = 0.91; P < 0.0001).

ComparisonuRARE-seqComparator
Urine cytology (n = 93 paired)89 detected (95.7%)32 detected (34.4%); absolute difference 61.3%
Low-grade Ta tumors, cytology comparison90% (28 of 31)Cytology detected none (P < 0.0001)
Tumor-naive utDNA (n = 85)99%85% (P < 0.001)
Tumor-informed utDNA (n = 28)96%96%

Grade and invasiveness models performed more modestly: the BLCA grade model achieved an AUC of 0.84 in training (10-fold cross-validation) and 0.84 in validation, while the muscle invasion model achieved 0.82 and 0.75, respectively.

Subgroup Analyses

Training-cohort sensitivity at 90% specificity by tumor grade and stage, as reported in the text:

Stage / gradeSensitivity
Low-grade (LG) Ta86% (31 of 36)
High-grade (HG) Ta95% (38 of 40)
Carcinoma in situ100% (13 of 13)
HG T1100% (34 of 34)
HG T2100% (28 of 28)

Detection performance was similar regardless of the presence of field-effect mutations (55 controls and 28 BLCA samples evaluated). Subgroup hazard ratios and confidence intervals beyond those shown are not specified in the source.

Minimal Residual Disease and Treatment Response

Among 36 patients with NMIBC (13 with high-grade recurrence, 23 without), all patients without recurrence had detectable BLCA cfRNA presurgery, 58% pre-BCG and none post-BCG. Among patients with HG recurrence, all had detectable cfRNA presurgery and pre-BCG, and 92% had it after BCG. cfRNA-positive cases had significantly worse HG recurrence-free survival at the pre-BCG (HR = 4.30; log-rank P = 0.02) and post-BCG (HR = 41.17; log-rank P < 0.0001) time points.

Molecular response groupn
Complete molecular response after surgery (mCR)10
Residual disease after surgery (mRD)26
mCR after BCG15
mRD after BCG11

Pretreatment urine from BCG responders was enriched for immune-related pathways (TNF signaling, interferon response, inflammatory response) and T cell signatures, while nonresponders showed increased proliferation-related pathways such as mitotic spindle and G2M checkpoint. TCR clones per million reads (P = 0.01) and TCR diversity (P = 0.04) were higher in responders (n = 7) than nonresponders (n = 10). An "intravesical therapy response prediction model" built from these pathways reached an AUC of 0.93 by leave-one-out cross-validation; in an independent BCG cohort (n = 57) scores were associated with recurrence (P = 0.0005), and in an independent chemotherapy cohort (n = 38) patients without recurrence had low scores and those with recurrence had high scores (P < 0.0001).

Safety Profile

Not applicable: this is a diagnostic and prognostic biomarker study with no therapeutic intervention assigned by the investigators; adverse-event data are not specified in the source.

Interpretation and Broader Context

The authors conclude that urine cfRNA analysis is a promising biomarker approach for bladder cancer and potentially other urologic malignancies. They emphasize that single-sample, tumor-naive classification avoids the need for matched tumor tissue. The authors also compared against published performance of other assays across studies; this is a cross-study comparison, not a head-to-head trial.

While these results are encouraging, larger cohorts and prospective clinical trials will be required to define fully the performance characteristics and clinical utility of cfRNA analysis in these settings.
— Liu et al., Nature Medicine, Discussion

Limitations

The experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment. The MRD and response-prediction cohorts are small: the BCG-response signature was derived from 26 patients with three landmark samples, and the chemotherapy validation cohort contained only 5 patients with recurrence. The comparison with other assays is cross-study rather than head-to-head. The number of study sites is not specified in the source. The authors state that larger cohorts and prospective clinical trials will be required to define the performance characteristics and clinical utility of cfRNA analysis. Disclosures as reported: K.J.L., W.Y.S., M.C.N. and E.G.H. report ownership interest or stock options in Resero Bio; M.C.N. and E.G.H. are employees of Resero Bio.

Primary source: DOI 10.1038/s41591-026-04673-3 · PubMed 42827132

References

  1. Liu KJ, Shi WY, Nesselbush MC, et al. Urine cell-free RNA for bladder cancer detection and treatment response prediction. Nature Medicine. 2026. DOI: 10.1038/s41591-026-04673-3. PMID: 42827132.