Intravesical Bacillus Calmette-Guérin (BCG) with transurethral tumor resection is considered first-line organ-preserving therapy for patients with high-risk nonmuscle-invasive bladder cancer (NMIBC).¹ Despite this treatment, up to 20%-50% of patients will eventually fail to respond, and 15% of patients will progress to muscle-invasive disease.² This represents a missed opportunity to identify the most appropriate candidates for BCG therapy, while potentially allowing for lead time toward muscle-invasive disease that harbors a poorer prognosis.³

Currently, the best predictors of BCG response are clinicopathological features such as tumor grade and stage. Previously studied molecular predictors of BCG response have included genomic alterations, molecular subtyping, mutational burden, DNA methylation, neoantigen load, and inflammatory markers.⁴ An important drawback with many of these approaches is the dependence on primary tumors as the marker source, where NMIBC tissue availability may be very limited. MicroRNAs (miRNAs) are short noncoding RNAs that regulate gene expression and are dysregulated in several malignancies including bladder cancer,⁵ exerting their effects as oncogenes or tumor suppressors. miRNAs are detectable in various body fluids including urine, and are relatively stable under various storage conditions.⁶ Their assessment in urine can therefore potentially serve as a unique liquid biopsy assay in the setting of NMIBC.

Our study hypothesized that urine-based miRNA profiles from NMIBC patients are reflective of primary tumor biology, and are distinct based on BCG response. We identified 120 patients with NMIBC who were treated at MD Anderson Cancer Center between 2005 and 2010. These patients underwent transurethral bladder tumor resection and subsequently received intravesical BCG treatment. Total RNA was isolated from frozen banked pretreatment urine specimens from these patients under an Institutional Review Board-approved protocol. The amplification-free nCounter platform was used to profile miRNAs, and samples that passed quality metrics were used for downstream analysis. Patients were categorized as responders or nonresponders; the latter included those who were unresponsive after adequate BCG therapy.⁷ Median follow-up for the entire cohort was 9.1 years.

A group of 80 patients formed a discovery set (comprised of 52 responders and 28 nonresponders), and 40 patients were used as a validation set (consisting of 26 responders and 14 nonresponders). There were no statistically significant differences in demographic or clinicopathological parameters between the discovery and validation sets. Variance filtering was used to identify differentially expressed miRNAs in the discovery set. Random forests, a machine learning method for classification,⁸ were then used to generate predictive signatures for BCG response on the discovery set. This approach constructs a multitude of decision trees with internal cross-validation to generate optimal classifiers while correcting for overfitting.⁹ Each tree generates a class prediction, and the class with the most votes becomes the model’s prediction. Random forest classifiers constructed using this approach were then blindly applied to the validation set.

A 233-miRNA classifier was initially constructed on the discovery set. The sensitivity, specificity, and negative predictive value were noted at 81%, 64%, and 64%, respectively (see Figure). This classifier was then used to blindly interrogate a validation set. The corresponding performance metrics were 78%, 53%, and 64%, respectively. In an effort to generate a more concise signature, the highest-ranked 35 miRNAs were used to construct another classifier on the discovery set. Sensitivity, specificity, and negative predictive value herein were noted at 85%, 68%, and 70%, respectively. When applied to the validation set, the corresponding performance metrics were 78%, 46%, and 71%, respectively.

NMIBC patients who progress on BCG therapy often have compromised survival due to delay in curative therapy (ie, radical cystectomy or trimodal therapy). In addition to the clinical impact, recurrent or nonprogressive disease can also impose a heavy financial burden on health care systems and is associated with significant morbidity. Furthermore, in an era of BCG shortage, predicting optimal response to BCG is clinically relevant for identifying those who may benefit from early consideration of alternate bladder-sparing therapies¹⁰ or more aggressive management. Given the multifaceted mechanism of action of BCG, it is conceivable that biomarkers reflecting the oncologic and immunological bladder milieu in NMIBC patients may be poised to offer a global perspective of the complex disease state. miRNAs assayed from pretreatment urine specimens of this heterogeneous patient population can therefore serve as an informative liquid biopsy test in this setting. This study serves as a proof-of-concept that such assays may hold promise as robust predictive tools for BCG response in NMIBC.