Bladder cancer is a major global burden. With an estimated 430,000 new cases and 165,000 deaths annually, bladder cancer is one of the most prevalent and difficult-to-treat malignancies worldwide. It divides into two broad clinical categories: non-muscle-invasive bladder cancer (NMIBC, stages Ta and T1) and muscle-invasive bladder cancer (MIBC, stages T2-T4), each with distinct treatment approaches and survival rates.
Progression from NMIBC to MIBC is a critical clinical problem. Most bladder cancers are initially non-muscle-invasive and treated with transurethral resection followed by intravesical BCG therapy. However, many patients, particularly those with T1 high-grade tumors, progress to invasive disease. MIBC carries a five-year survival below 50% with high rates of metastatic relapse after radical cystectomy.
Immune checkpoint inhibitors help only a minority of patients. Anti-PD-1 and anti-PD-L1 immunotherapy has emerged as an important treatment option for advanced bladder cancer and BCG-unresponsive NMIBC. However, only a fraction of patients respond, and PD-L1 protein expression alone has shown inconsistent predictive value, leaving clinicians without a reliable way to identify which patients will benefit.
Molecular subtypes could unlock better patient selection. While multiple molecular subtype classifications have been developed for MIBC, none prior to this study had successfully classified progressive NMIBC or identified subtypes specifically predictive of immune checkpoint inhibitor response in early-stage bladder cancer, creating a critical gap in precision oncology for this disease.
Large multi-cohort analysis of 1934 samples. The study integrated gene expression data from seven patient cohorts spanning both NMIBC and MIBC cases. The discovery cohort came from Chungbuk National University Hospital (165 patients), with validation performed in six independent cohorts including TCGA (408 MIBC), UROMOL (460 NMIBC and 16 MIBC), and IMvigor210 (298 MIBC treated with atezolizumab).
Gene selection focused on consistently variable genes. Starting with genes that showed at least two-fold expression differences in more than 15% of samples, the analysis identified 1627 genes that were commonly variable across three microarray cohorts. These genes were used for initial clustering to ensure the resulting subtypes would generalize across different gene expression platforms.
Unsupervised hierarchical clustering defined four subtypes. Using centroid linkage hierarchical clustering with median gene centering, four distinct bladder cancer subtypes were identified in the discovery cohort. Silhouette width analysis confirmed the quality of cluster assignments, and only samples with positive silhouette values were retained for classifier training.
A 786-gene classifier was built for prospective application. Significance analysis of microarrays (SAM) identified subtype-specific gene signatures, and prediction analysis of microarrays (PAM) with 10-fold cross-validation then selected an optimal 786-gene genomic subtype predictor (GSP786) that could classify new samples with the lowest prediction error rate.
Class 1 contains low-risk, slow-growing tumors. Class 1 was predominantly composed of low-grade NMIBC tumors. These tumors showed decreased expression of cell proliferation genes, reflecting less aggressive clinical behavior. In the RNA-seq validation cohort, class 1 tumors were enriched for the luminal-papillary TCGA subtype, which has a favorable prognosis, and many high-grade tumors in this class were histologically papillary, which also carries better outcomes.
Class 2 shows immune evasion features. Class 2 included both low-grade NMIBC and a small proportion of MIBC tumors. The defining molecular feature was downregulation of immune response pathways, particularly antigen processing, presentation, and T cell receptor signaling. HLA gene expression was specifically inhibited in class 2, a pattern associated with poor prognosis. Class 2 also showed upregulation of FGFR3 and CCND1, suggesting these tumors evade immune detection while retaining FGFR3-driven oncogenic signaling.
Class 3 captures aggressive disease across stages. Class 3 showed a mixture of high-grade NMIBC and MIBC, with 69% of T1 high-grade tumors falling into this class. Molecular hallmarks included activation of cell cycle genes including E2F1, FOXM1, CCNB1, and CCNE1, inhibition of the Notch tumor suppressor pathway, high somatic mutation burden, and enrichment of DNA damage response and repair gene alterations including BRCA1 and TP53 mutations. These features were consistent across both NMIBC and MIBC subgroups within this class.
Class 4 reflects EMT-driven invasive disease. Class 4 contained the highest proportion of MIBC cases with prominent upregulation of genes involved in extracellular matrix organization, epithelial-mesenchymal transition, and myofibroblast markers. Most class 4 MIBC samples corresponded to the basal-squamous TCGA subtype or the luminal-infiltrated subtype. Class 4 also showed strong activation of immune response pathways, resembling the infiltrated subtype from prior classification systems.
Subtype classification predicted survival in all cohorts tested. Kaplan-Meier analysis across the CNUH, YUSH, UHL, and SSH cohorts consistently showed statistically significant differences in cancer-specific survival among the four classes. Class 1 patients had the best outcomes, class 2 showed intermediate survival, and classes 3 and 4 showed reduced survival compared with other subtypes.
Class 3 specifically predicted NMIBC progression. When analyzing progression-free survival in NMIBC patients across three independent cohorts, class 3 showed significantly higher rates of progression to muscle-invasive disease compared with other classes. This finding validates that the molecular features of class 3 capture biologically aggressive NMIBC that is at high risk for stage advancement regardless of its initial non-invasive classification.
GSP786 classifies across different staging systems. The four subtypes were consistently identified regardless of whether samples were classified by conventional stage or grade, suggesting that the molecular subtype reflects underlying biology that crosses traditional histopathological boundaries. In particular, class 3 identified a subset of T1 tumors with MIBC-like molecular behavior.
GSP786 resolved ambiguities in prior classification systems. Previous subtyping systems had difficulty distinguishing the genomically unstable (GU) from the infiltrated subtypes within the UROMOL class 2 category. GSP786 successfully separated these into class 3 (GU features) and class 4 (infiltrated features), demonstrating added resolution over existing classification frameworks.
Class 3 has the highest somatic mutation burden. Analysis of TCGA somatic variant data confirmed that class 3 tumors had significantly higher total mutation rates compared with other subtypes. TP53 mutations were present in 70% of class 3 tumors and RB1 mutations were also enriched, consistent with the aggressive biology of class 3. By contrast, FGFR3 mutations were significantly more frequent in classes 1 and 2, corresponding to their less invasive characteristics.
Class 3 patients responded to anti-PD-L1 treatment in IMvigor210. Applying GSP786 to 298 metastatic bladder cancer patients treated with atezolizumab in the IMvigor210 clinical cohort, class 3 patients showed a significantly higher response rate than patients in other subtypes (Fisher's exact test p-value less than 0.001). Class 3 patients also showed better survival on immunotherapy compared with other subtypes (log-rank p = 0.028), a reversal of the poor prognosis class 3 shows on standard treatment.
TGFb pathway inhibition in class 3 supports immunotherapy sensitivity. Gene set enrichment analysis revealed that class 3 was negatively associated with TGFb signaling, which is known to attenuate T cell infiltration and response to anti-PD-L1 therapy. Decreased TGFb activity in class 3 tumors may therefore help explain their superior immunotherapy responsiveness by allowing greater T cell access to the tumor microenvironment.
Subtype classification was an independent predictor of ICI response. In a multivariate logistic regression model including PD-L1 expression on tumor cells, PD-L1 expression on immune cells, and tumor mutation burden, the bladder cancer subtype retained a statistically significant independent association with immunotherapy response (p = 0.00287). This confirms that GSP786 subtype classification adds predictive value beyond existing biomarkers.
Class 3 changes how early-stage bladder cancer should be treated. By identifying a subgroup of NMIBC patients with MIBC-like molecular characteristics and worse progression-free survival, GSP786 suggests that a subset of T1 high-grade patients may need more aggressive upfront treatment rather than standard BCG therapy. This provides a molecular basis for treatment escalation decisions in early-stage disease.
Classes 1 and 2 may be candidates for targeted therapy. The enrichment of FGFR3 mutations and overexpression in classes 1 and 2 identifies these as potential candidates for FGFR3-targeted therapies, an area of active clinical investigation. The distinct immune evasion pattern in class 2, characterized by reduced HLA gene expression, represents another potential therapeutic vulnerability worth exploring.
GSP786 outperforms the GU subtype as an immunotherapy biomarker. While the previously described genomically unstable subtype from the Lund taxonomy was associated with ICI response in prior work, it did not show a higher complete response rate than other subtypes. GSP786 class 3 provided better discrimination of immunotherapy responders because it captures the full biological program associated with ICI sensitivity, not just one feature of the GU classification.
TMB alone is insufficient for patient selection. Tumor mutation burden showed significant association with ICI response in the multivariate model, but class 3 subtype assignment added independent predictive value. Since whole-exome sequencing for TMB quantification is expensive and lacks validated disease-specific thresholds for bladder cancer, a gene expression-based classifier like GSP786 may offer a more practical and comprehensive biomarker approach.
A single classifier applicable across disease stages. Unlike prior molecular subtyping systems that focused exclusively on MIBC or NMIBC, GSP786 was developed and validated across both stages in 1934 patients from seven cohorts. This breadth means a single classifier could potentially guide treatment decisions throughout the entire clinical trajectory of bladder cancer, from initial presentation through progression.
Addresses an unmet need in early bladder cancer management. The absence of validated molecular tools for identifying high-risk NMIBC patients who will progress despite standard therapy has left clinicians relying on clinical risk scores that miss important biological heterogeneity. GSP786 class 3 classification provides a biologically grounded approach to this unmet clinical need.
Supports rational combination therapy development. Understanding the molecular biology of each subclass points toward logical therapeutic combinations. For class 3, checkpoint inhibitors combined with cell cycle inhibitors targeting the activated CDK-CCNB1-E2F1 axis could be explored. For class 2, immune activation strategies combined with HLA pathway restoration may help restore immunosurveillance.
Prospective validation is needed before clinical adoption. The authors appropriately emphasize that retrospective analysis in the IMvigor210 cohort is hypothesis-generating but not sufficient for clinical implementation. Prospective clinical trials stratifying patients by GSP786 class 3 status and testing checkpoint inhibitor therapy in this selected group are needed to confirm the predictive value observed in this study.
Four molecular subtypes with distinct biology and prognosis. Using a 786-gene classifier validated in 1934 patients across seven cohorts, this study established four bladder cancer molecular subtypes with reproducibly distinct biological features, prognostic significance, and differential responses to treatment, providing a comprehensive molecular framework applicable across disease stages.
Class 3 as the key immunotherapy-responsive subtype. Class 3, characterized by cell cycle activation, high tumor mutation burden, DNA damage response gene alterations, and inhibited TGFb signaling, was identified as the subtype most likely to respond to PD-L1 checkpoint inhibitor therapy, as confirmed in the IMvigor210 atezolizumab clinical cohort.
Molecular subtype adds to existing biomarkers. In multivariate analysis including PD-L1 expression and tumor mutation burden, GSP786 subtype classification remained an independent predictor of immunotherapy response, confirming that it captures biologically relevant information not encoded by currently used biomarkers.
Foundation for precision therapy in bladder cancer. By connecting molecular subtype to both prognosis and immunotherapy response across NMIBC and MIBC, this work provides a foundation for precision oncology in bladder cancer, with potential clinical utility ranging from treatment escalation decisions in early-stage disease to immunotherapy patient selection in advanced disease.