A quantitative polymerase chain reaction based method for molecular subtype classification of urinary bladder cancer—Stromal gene expressions show higher prognostic values than intrinsic tumor genes

Transcriptome‐based molecular subtypes of muscle‐invasive bladder cancer (MIBC) have been shown to be both prognostic and predictive, but are not used in routine clinical practice. We aimed to develop a feasible, reverse transcription quantitative polymerase chain reaction (RT‐qPCR)‐based method for molecular subtyping. First, we defined a 68‐gene set covering tumor intrinsic (luminal, basal, squamous, neuronal, epithelial‐to‐mesenchymal, in situ carcinoma) and stromal (immune, extracellular matrix, p53‐like) signatures. Then, classifier methods with this 68‐gene panel were developed in silico and validated on public data sets with available subtype class information (MD Anderson [MDA], The Cancer Genome Atlas [TCGA], Lund, Consensus). Finally, expression of the selected 68 genes was determined in 104 frozen tissue samples of our MIBC cohort by RT‐qPCR using the TaqMan Array Card platform and samples were classified by our newly developed classifiers. The prognostic value of each subtype classification system and molecular signature scores were assessed. We found that the reduced marker set combined with the developed classifiers were able to reproduce the TCGA II, MDA, Lund and Consensus subtype classification systems with an overlap of 79%, 76%, 69% and 64%, respectively. Importantly, we could successfully classify 96% (100/104) of our MIBC samples by using RT‐qPCR. Neuronal and luminal subtypes and low stromal gene expressions were associated with poor survival. In conclusion, we developed a robust and feasible method for the molecular subtyping according to the TCGA II, MDA, Lund and Consensus classifications. Our results suggest that stromal signatures have a superior prognostic value compared to tumor intrinsic signatures and therefore underline the importance of tumor‐stroma interaction during the progression of MIBC.

that the reduced marker set combined with the developed classifiers were able to reproduce the TCGA II, MDA, Lund and Consensus subtype classification systems with an overlap of 79%, 76%, 69% and 64%, respectively. Importantly, we could successfully classify 96% (100/104) of our MIBC samples by using RT-qPCR. Neuronal and luminal subtypes and low stromal gene expressions were associated with poor survival. In conclusion, we developed a robust and feasible method for the molecular subtyping What's new?
Transcriptome-based molecular subtypes of muscle-invasive bladder cancer have been demonstrated to be both prognostic and predictive. However, due to their complexity and high costs, transcriptome-based methods are not used in routine clinical practice. Here, the authors present a feasible 68-gene panel-and RT-qPCR-based method for molecular subtyping according to the most commonly used molecular classification systems of muscle-invasive bladder cancer. The method has been validated using in silico datasets and was further tested in an institutional bladder cancer cohort. The data revealed different prognoses for some of the molecular subgroups and underlined the prognostic relevance of stroma-related gene expression signatures.

| INTRODUCTION
Bladder cancer (BC) is a common malignancy with approximately 550 000 new cases each year worldwide. 1 Urothelial carcinoma is the most frequently diagnosed histological type of BC. About 30% of cases are muscle-invasive at first presentation or will ultimately progress to muscle-invasive bladder cancer (MIBC). The standard care of MIBC is radiochemotherapy or radical cystectomy (RC) with perioperative platinum-based chemotherapy; however, the 5-year survival rate of these patients is less than 50%. Currently, checkpoint and fibroblast growth factor receptor (FGFR) inhibitors as well as Nectin-4 antibody conjugates have become available for platinum-resistant and/or -ineligible patients. Only few routinely available predictive biomarkers, such as Programmed death-ligand 1 expression by immunohistochemistry or FGFR3 mutational or fusion-status, are available. [2][3][4] MIBC patients may show remarkable differences regarding their response to therapies. Therefore, a more detailed characterization of MIBC is required to decipher this clinical heterogeneity.
In the last years, several studies demonstrated that MIBCs with similar histological patterns may have distinct molecular properties.
Transcriptome analyses of MIBC samples revealed distinct molecular subtypes with different prognosis. One of the earliest molecular classifications was developed by a research group from Lund and distinguished five subtypes with diverse gene expression patterns and clinical outcome. 5 Then the "University of North Carolina (UNC) classification" defined luminal and basal subtypes similarly to the determined subtypes in breast cancer and revealed that luminal tumors have a significantly better prognosis compared to basal cases. 6 Subsequent studies confirmed the presence of luminal and basal subtypes in independent MIBC cohorts. The "MD Anderson (MDA) classification" described a p53-like subtype in addition to luminal and basal subtypes, which was associated with an improved survival compared to basal tumors; however, it showed resistance to chemotherapy. 7 In addition, Seiler et al found that only patients classified as basal subtype benefited from a platinum-based neoadjuvant chemotherapy (NAC). 8 In the first The Cancer Genome Atlas (TCGA) study in 2014, 129 MIBC samples were analyzed by transcriptome sequencing, 9 which in 2017 has been extended to 412 samples. This "TCGA II" study distinguished three luminal (luminal-infiltrated, luminal-papillary and luminal), a basal/squamous and a neuronal subtype. 10 Of these molecular subgroups, neuronal tumors had the worst while the luminal-papillary tumors exhibited the most favorable prognosis.
Recently, an international consensus classification with six molecular subtypes has been suggested based on a reanalysis of 1750 formerly published MIBC transcriptome profiles. 11 This study included an own nomenclature with significant overlap with previously suggested classification systems. The authors confirmed a more favorable prognosis for luminal papillary subtype, as well as for the luminal nonspecified and stroma-rich subtypes, while luminal unstable, basal/squamous and neuronal subtypes had a poor prognosis. However, no significant differences could be observed in NAC-treated patients regarding the outcomes between various consensus subtypes.
All these subtype classification systems are based on transcriptome data and consider the expression of thousands of genes, which is hardly compatible with daily clinical routine. Therefore, we aimed to develop a simple and applicable system for potential inclusion in daily clinical routine with the final aim to translate the molecular findings to clinical application. To achieve this, we utilized a reverse transcription quantitative polymerase chain reaction (RT-qPCR)-based gene expression analysis method with a reduced marker set and a respective evaluation method in order to recapitulate the TCGA II, MDA, Lund and 2 | MATERIALS AND METHODS 2.1 | Development of a classifier method for molecular subtyping using a reduced marker set First, we selected the markers as the genes with the highest discriminating effect between distinct molecular subtypes in the TCGA II study.
Then, identified marker set were further reduced and those markers also used in other classification systems were preferred. Based on a comparison of respective subtype classification studies, a panel of 68 genes was defined in order to distinguish molecular subtypes according to the TCGA II, MDA, Lund and Consensus systems. 7,10-13 These markers covered six tumor cell-specific and three stroma-related gene signatures ( Table 1).
The tumor intrinsic signatures were as follows: luminal, basal, squamous, neuronal, epithelial-to-mesenchymal transition (EMT), carcinoma in situ (CIS), while stroma-specific signatures included p53, extracellular matrix (ECM)/smooth muscle and immune cell-specific genes. 10 Second, we used respective publicly available data sets for the in silico development and validation of subtype classification rule sets for the TCGA II, MDA, Lund and Consensus classification systems. For each classification system, two data sets with available transcriptome-based subtype class information were used for the elaboration (training set) and validation (validation set) of our classifier method (Figures 1, S1A, S3A and S5A).

| Patients' characteristics
The main characteristics of our institutional (Essen cohort) and the other data set cohorts (TCGA II, MDA, Lund, CIT, Riester, Seiler) are summarized in

| Molecular subtype classification of our institutional cohort and its correlations with clinicopathological parameters
Gene expression of the reduced marker set was determined by RT-qPCR and applied for molecular subtype classification according to our above-described TCGA II classifier rule set. During RT-qPCR data evaluation, four samples were excluded because their low housekeeping gene expression level (Ct >33), leaving 100 samples for the final evaluation ( Figure 2). The subtype distribution within our institutional cohort proved to be remarkably similar to that of the TCGA cohort (  (Figures 2 and S7).
Pearson's chi-square test was used to examine the association between "summa luminal" (luminal-papillary, luminal-infiltrated, luminal), basal/squamous subtypes and main clinicopathological parameters (Table S2). Similar to the findings made in the TCGA cohort, basal subtype tended to associate with female sex and was more frequent in high-stage tumors; however, these correlations proved not to be statistically significant.

| Survival analysis of clinicopathological parameters and signature scores
Univariable analysis of our institutional cohort revealed lymph node metastasis as a significant risk factor for survival (OS and CSS: P < .001). Patients' age and sex had no significant impact on OS and CSS, while tumor stage significantly correlated with CSS (P = .014) (Table 3A). Basal, squamous, luminal, CIS and EMT signatures had no significant effect on OS and CSS (Table 3A) (Table 3B).
In order to exclude the possible influencing effect of chemotherapy use on results, we performed survival analyses also after the exclusion of 15 patients who received postoperative platinum therapy. In this cohort, univariable analysis of clinicopathological parameters, signature scores and survivals revealed the same associations as for the whole cohort; lymph node metastasis, neuronal, ECM, immune and p53 score correlated with OS (P = .002, P < .001, P = .001, P = .001 and P = .015, respectively) and CSS (P < .001, P < .001, P = .004, P = .002 and P = .025) (Table S4A).
Furthermore, multivariable analysis showed that the presence of lymph node metastases and high neuronal scores were independently associated with poor OS and CSS (P = .001, P < .001 and  (Table S4B). In addition, low ECM score was correlated with worse OS (P = .039), while low immune score tended to associate also with poor OS and significantly correlated with shorter CSS (P = .071 and P = .040).

| Survival analysis by the molecular subtype classifications
The original, transcriptome-based subtype classification of the TCGA II cohort proved to be prognostic (P = .001); patients with luminalpapillary subtypes showed the most favorable prognosis, while neuronal and luminal subtypes had the worst OS ( Figure 1B). Similarly, when classifying the TCGA cohort by our selected marker and rule set, distinct subtypes had significantly different survival rates (P = .038) ( Figure 1C). Subtype classification of our institutional cohort using the RT-qPCR-based gene expression data also proved to be prognostic (P = .003). Similar to the findings of the TCGA study, we found the worst prognosis for the neuronal and luminal subtypes. The other three subtypes did not show a significant difference regarding their OS rates. (Figure 1D). According to the univariable analysis, only neuronal subtypes associated with survival, and was a significant risk factor for OS and CSS (P = .002 and P = .001) (Table S5).
Survival analysis using our MDA classifier rule set resulted in a similar significant risk stratification compared to the transcriptome-based classification on the reference TCGA cohort (P = .002) ( Figure S1B,C).
The same method on our institutional cohort using our RT-qPCR generated gene expression data revealed the p53-like subtype to have a more favorable prognosis, which is in line with the findings of the MDA study ( Figure S1D). 7 The classifier according to Lund subtyping had no significant impact on OS (P = .203) ( Figure S3B).  were classified into luminal, basal and p53-like subtypes. Interestingly, they found luminal subtype to be an independent risk factor for patients' survival. 18 Kardos et al transposed the BASE47 mRNA expression classifier 6 to a NanoString-based platform and divided samples into luminal and basal subtypes but could not observe a prognostic difference between these two groups. 19 (Table S6).

T A B L E 3 Cox univariable (A) and multivariable (B) survival analyses with dichotomized signature scores of our institutional cohort
Overall, IHC analyses have the advantage that they can be easily which was similar to the findings of the TCGA II study. 10 Neuronal subtype represents $5% of all MIBCs but is associated with a devastating prognosis and poor response to NAC. 10,25 On the other hand, tumors assigned to this subtype may well respond to immune checkpoint inhibitor therapy. 24 Therefore, distinguishing the neuronal subgroup is of clinical importance, but former RT-qPCR or NanoString platform-based studies were not able to identify this subtype.
On the other hand, we could not confirm the favorable prognostic value of luminal-papillary subtype, described in the TCGA II study. 10 However, in a recent transcriptome-based study with 283 MIBC patients, the luminal-papillary subtype was also not associated with favorable survival. 26 When classifying our institutional cohort according to the MDA systems, we found the p53-like subgroup to have the most favorable prognosis, while the luminal subtype had a slightly better prognosis compared to the basal group, which is similar to that of was found in the MDA study. 7 When classifying our institutional cohort according to the Lund system, the mesenchymallike subtype proved to be associated with poor survival, which is similar to the findings of the Lund group on their cohort. 26 Figure S7). 27 In accordance, Ikeda et al revealed that high tumorassociated immune cell status associated with significantly better CSS in MIBC patients. 22 In a further study, a newly defined stromal immunotype of MIBCs with high CD8 + cytotoxic T cells and natural killer cells was significantly associated with better overall and recurrence-free survival as well. 28 Limitations of our work are that our patient cohort may be less representative for some reason.

| CONCLUSIONS
In this study, we present a robust and feasible RT-qPCR-based molecular subtype analysis method, which is able to recapitulate the TCGA II and the MDA classifications with an estimated overlap of $80% and the Lund and Consensus classifications with 69% and 64%. Applying this method, we were able to classify 96% (100/104) of our institutional MIBC cohort and found a significant prognostic value for the TCGA II and MDA subtype classification systems. Furthermore, our method was able to identify smaller but clinically relevant subgroups (such as neuronal and luminal-infiltrated) which was formerly only possible by transcriptome-based analysis. Moreover, high stroma-related signature scores (ECM, p53, immune) were associated with favorable OS. In contrast, tumor intrinsic signature scores (basal, squamous, luminal) were not prognostic, except of neuronal score which-similar to immune score-proved to be an independent prognostic factor in MIBC. These results underline the importance of the molecular features of stromal component and suggesting that a dichotomous classification into basal and luminal subtypes may be of limited clinical value.
The final aim is to translate the findings of large transcriptome studies to a marker system that can be easily applied in the clinical routine using common platforms and robust assays. Our method needs to be transferred to formalin-fixed paraffin-embedded samples and be prospectively validated in independent patient cohorts also with regard to its predictive value for chemotherapy and/or immune therapy effectiveness prediction, which is the planned next step.

ACKNOWLEDGMENTS
We thank Dr Mikl os Sárvári and Dr Attila Pat ocs for providing us the equipment for real-time PCR analyses and Dr Tibor Füle for his help in primary evaluation of real-time qPCR data. We are grateful to Dr Woonyoung Choi for providing subtype calls to publicly available MDA data set. Open Access funding enabled and organized by Projekt DEAL.

CONFLICT OF INTEREST
The authors declare no conflicts of interest.

DATA AVAILABILITY STATEMENT
Data sources and handling of the publicly available data sets used in this study are described in the Materials and Methods section and in Table 1. Further details and other data that support the findings of this study are available from the corresponding author upon request.

ETHICS STATEMENT
The study was performed according to the Declaration of Helsinki and the institutional ethics committee approved the study protocol (08-3942-BO/15-6400-BO).