Abstract
Non–muscle invasive bladder cancer (NMIBC) is a biologically and clinically heterogeneous disease, accounting for approximately 75% of bladder cancer cases. Over the past decade, multiple RNA-based molecular classification systems for NMIBC have been proposed, demonstrating associations with tumor stage, grade, prognosis, and treatment response. However, unlike muscle-invasive bladder cancer, a unified consensus classification for NMIBC is lacking, limiting clinical translation and cross-study comparability. To address this gap, the first NMIBC Consensus Classification Meeting was convened in November 2024 at Erasmus University Medical Center, bringing together 13 international experts from seven countries representing major NMIBC molecular classification efforts. This report summarizes the discussions, comparative evaluation of existing classification systems, and the agreed strategy towards developing a consensus molecular model. The meeting focused on harmonizing heterogeneous transcriptomic datasets, integrating RNA sequencing and microarray platforms, and assessing the applicability of classification models across tumor stages (Ta/T1) and grades. Key features, strengths, and limitations of the Lund Taxonomy, Leeds, Birmingham, UROMOL, and Rotterdam (BCG response) subtypes were critically examined. Building on prior experience from the muscle-invasive bladder cancer consensus, participants agreed on a centroid-based clustering framework using rigorously defined core samples to derive stable consensus subtypes and a single-sample classifier with associated probability scores. A large, curated dataset encompassing approximately 2500 NMIBC samples from published cohorts will facilitate model development and validation. The resulting consensus classification aims to improve biological understanding, facilitate reproducibility across studies, and enable future evaluation of prognostic and therapeutic relevance, including response to intravesical BCG therapy.
Introduction
Non-muscle invasive bladder cancer (NMIBC) is a heterogeneous disease at both the clinical and molecular level - unraveling the molecular biology of these tumors has the potential to improve risk stratification and treatment. As a result, extensive investigational efforts have led to multiple molecular classifications of NMIBC, which correlate with tumor stage, grade, and patient prognosis and response to treatment. Whereas a consensus has been reached for molecular classification 1 of muscle-invasive bladder cancer (MIBC), a consensus among molecular classifications for NMIBC is lacking. With these premises and considering that NMIBC represents 75% of bladder cancer patients, we felt there were compelling reasons to convene a meeting dedicated to reaching a consensus on the molecular classification of NMIBC and to stimulate collaboration between involved groups. In November 2024, the first NMIBC Consensus Classification Meeting was held at Erasmus University Medical Center in Rotterdam, The Netherlands. An international group of 13 healthcare professionals and researchers from 7 countries that have proposed RNA-based NMIBC molecular subtypes in recent years2–7 gathered to participate in the two-day conference. This report provides an overview of the proceedings and scientific exchange from this meeting.
Meeting discussion report
The main objectives of the consensus meeting were to:
Provide an overview of available datasets and methods used for molecular classification Discuss the strategy on how to integrate all datasets for analysis, irrespective of data type (RNA microarray vs. RNA-sequencing [RNA-seq] data). Discuss the possibility of a classification model that is applicable irrespective of the clinical stage (cTa/cT1) and/or pathological tumor grade (G1-3 or LG/HG).
Secondly, for each NMIBC classification, properties and specifications were discussed with the participating representatives, considering potential pitfalls and generalizability while applying individual classification models to different cohorts during the development of a consensus classification model.
Lund taxonomy
In December 2024, the Lund group reported a versatile and updated version of the Lund Taxonomical (LundTax) classification algorithm, applicable to both NMIBC and MIBC. 7 At the meeting, Dr Pontus Eriksson and Dr Gottfrid Sjödahl pointed out that, according to the Lund Taxonomy, the majority of NMIBC tumors are classified as Urothelial-like (Uro). This aligns with the consensus MIBC subtype system that classifies >90% of NMIBC as luminal papillary. 2 Urothelial-like is considered as one subtype, although it may be subdivided into UroA, UroB, and UroC. By their definition, a subtype is a stable differentiation state of the cancer cells, defined by immunohistochemistry using cancer cell-directed antibodies rather than clustering of bulk biopsy expression. Consequently, LundTax classification is not driven by the broad signatures originating from proliferation or the presence of non-cancer cells present in the tumor microenvironment. Instead, these signatures are treated as separate secondary variables derived directly from the gene expression data.
Leeds and Birmingham classifications
In December 2021, the Leeds group (led by prof. Maraget Knowles) reported their stage-stratified molecular profiling of NMIBC and showed that separate subtyping of Ta and T1 tumors enhanced biological, clinical, and therapeutic insight. 3 In June 2022, the Birmingham group (led by prof. Rik Bryan) reported on associations between genomic changes, expression subtypes, and clinical outcomes by clustering NMIBCs into three expression subtypes (Class A/B/C). 4 While the research groups from Leeds and Birmingham proposed NMIBC classifications, neither provided classifiers for independent external application of their classifications. During the meeting, Dr Alberto Nakauma-Gonzalez presented his efforts in developing classifiers for Leeds and Birmingham to enable representation of these schemes across the external samples to be used for development of a consensus classification model. These classification models will be rigorously evaluated by their groups Prof. Margaret Knowles (Leeds) and Dr Anshita Goel (Birmingham), before inclusion in the NMIBC consensus model.
UROMOL classification
In April 2021, the UROMOL consortium (led by Prof. Lars Dyrskjot) reported an integrated multi-omics analysis identifying prognostic molecular subtypes of NMIBC, proposing four transcriptomic classes (1, 2a, 2b and 3). 2 During the consensus meeting, Dr Sia Viborg Lindskrog discussed that their classifier is based on whole-transcriptome RNA-seq data and thus care should thus be taken when applying the classifier on microarray data because depending on the platform used, these may not include non-coding genes. Nevertheless, application of the UROMOL classifier on microarray data within the most recent UROMOL publication did not reveal any issues. The UROMOL cohort includes the entire disease spectrum of NMIBC, and the classifier can, therefore, be applied to both Ta and T1 tumors. However, the UROMOL cohort included more low-risk tumors compared to other cohorts and the classifier was not trained to evaluate response to Bacillus Calmette-Guérin (BCG) treatment.
Rotterdam subtypes
In May 2023, the Rotterdam group (led by dr. Tahlita Zuiverloon) proposed three distinct BCG response subtypes (BRS1-3) using whole-transcriptome RNA-seq on 283 high-risk NMIBC patients treated with intravesical BCG immunotherapy. 6 Analyses revealed a worse prognosis for patients with BRS3 tumors, a subtype characterized by high epithelial-to-mesenchymal transition and basal markers, and an immunosuppressive profile (confirmed by spatial proteomics). During the consensus meeting, dr. Christiaan de Jong acknowledged that the BRS (Rotterdam) classifier uses an additional step to re-scale the gene expression from an external cohort to their development data set. As such, the BRS classifier may solely be applicable to similar cohorts containing T1 sample tumors.
Consensus model development plan
As preparation for this meeting, all RNA-seq data were processed using a single pipeline and GRCh38 as the reference genome. Similarly, all cohorts of RNA microarrays were processed in a standard manner. An overview of all cohorts and included patients is given in Table 1. The RNA counts were then used for subtype classification of the various NMIBC classification systems. These preliminary results are being reviewed by all groups that have developed an NMIBC classifier. Of note, the NMIBC Consensus Subtype Consortium plans to undertake an approach whereby subtypes are labeled as missing when applying classifiers developed specifically for certain tumor stage (e.g., the BRS, Leeds).
Cohorts and transcriptomic subtypes included for building a consensus NMIBC subtype system.
In analogy to the MIBC consensus method,
1
the following has been proposed regarding consensus model methodology:
The method of defining the consensus subtypes will be based on a centroid model, and all samples with the identified subtype for each classification system will be clustered using the BuildConsensus R package (developed by the MIBC Consensus Consortium) with silhouette score threshold >0.95 to be used to define the cluster number. Core samples belonging to a specific cluster will be defined based on high probability of belonging to a specific cluster (threshold and method to be defined). Subtype assignment of core samples should be based on a less stringent p-value (since the subtypes are not as distinct as in MIBC) or using multiple corrections (use cutoff points dynamically). Those core samples will be used to build a single-sample mRNA classifier for the consensus classes, and then used for consensus label assignment to all training cohort samples. To build a consensus classifier, the two largest cohorts containing Ta and T1 core samples are selected: Leeds (Affymetrix) and UROMOL (whole-transcriptome RNA-seq). Several strategies will be used to test the classifiers, e.g., cross-validation or by dividing training and testing datasets. The classifier will be further tested in the rest of the core cohort samples. The stability of the subtypes will be checked, and next to the subtype label, a probability score for each consensus subtype will be generated per sample in the final classifier to include an approximation of the level of intra-tumor subtype heterogeneity.
Future perspectives
While the NMIBC Consensus Subtype Consortium acknowledges that multiple validation cohorts are presently being processed, the development of the NMIBC consensus model will be based only on published literature. When the model development phase is completed, future available cohorts may serve as independent NMIBC consensus model validation studies. We are targeting a total cohort size of 2500 NMIBC samples, facilitating sufficient power to characterize the biology of the foreseen consensus subtypes of NMIBC. Moreover, the aggregated dataset may facilitate identification of druggable targets across biological subgroups of NMIBC, supporting future therapeutic development. Whereas NMIBC classification models have been correlated with prognosis after BCG, previous studies are underpowered to draw conclusions on BCG response. 13 Clinical data curated by each participating study center will be centralized in a database. In this context, the consensus NMIBC will be tested for its ability to identify high-risk NMIBC with a need for treatment intensification (or de-intensification).
Footnotes
Acknowledgments
JJ acknowledges the Erasmus MC Young Investigator Grant
Ethical approval number/ IRB statements
NA
Consent statement
NA
Author contribution for each author
JJ wrote the initial manuscript under supervision of TZ, SR, and AN and all authors reviewed, adjusted and provided suggestions for the manuscript.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.