Molecular classification of urothelial carcinoma: global mRNA classification versus tumour‐cell phenotype classification

1Global mRNA expression analysis is an efficient method for phenotypic profiling of tumours, and has been used to define molecular subtypes for almost every major tumour type 1, 2, 3. However, a key limitation is that tumours, in most cases, are communities of both tumour and non‐tumour cells. Accordingly, the proportions of cells in such communities may severely affect the global gene expression profile. This problem is particularly pertinent for analysis of advanced invasive tumours, which are known to induce major changes and responses in both the tumour and surrounding tissue. We have benefited from our previous analyses of non‐muscle‐invasive (NMI) tumours in defining molecular subtypes of urothelial carcinoma. NMI tumours grow into the bladder cavity with limited influence from surrounding tissue, making it more likely that global gene expression analyses will capture the nature of the tumour‐cell phenotypes. Extensive analyses of NMI tumour samples led us to the definition of three major molecular subtypes of urothelial carcinoma (UC), i.e. urothelial‐like (Uro) (previously termed urobasal 4), genomically unstable (GU), and basal/squamous cell carcinoma (SCC)‐like, with well‐defined molecular phenotypes and genotypes 5, 6, 7, 8. These subtypes show differential activity of established tumour markers and key regulatory genes: Uro tumours express FGFR3 and CCND1, and frequently show 9p21 (CDKN2A) loss; GU tumours express FOXM1 but not KRT5, and frequently show RB1 loss; and basal/SCC tumours express KRT5 and KRT14, but not FOXA1 and GATA3. Apart from our efforts, UC gene expression subtypes have been described for muscle‐invasive (MI) cases 9, 10, 11, 12 and for NMI cases 13. For MI tumours, the UNC classification 10 limits classification to two subtypes, luminal and basal, and points to a similarity with breast cancer luminal and basal‐like tumours. The MDA classification 9 also defines a luminal and a basal subtype, but introduces a third group, named TP53‐like, distinguished by a TP53‐related gene expression signature. This classification also points to the clinical utility of molecular classification. The TCGA consortium 11 identifies four groups of tumours defined by consensus clustering, but, apart from comparing the clusters with subtypes defined by others, TCGA limits the efforts to name the clusters I, II, III, and IV. With four classification systems at hand, all based on genome‐wide gene expression analyses, Aine et al 14 set out to compare the systems by applying all of them to the TCGA dataset. They concluded that the four systems showed extensive overlap and could be organized in a hierarchical order, with the UNC classification at the top with the lowest resolution, followed by the MDA, TCGA and Lund classifications with the highest resolution. However, owing to the complexity of MI carcinomas, the Lund taxonomy had to be extended to include six categories. In the present investigation, we expand previous studies by analysing advanced UC by both global mRNA profiling and by extensive immunohistochemistry (IHC) investigations to define tumour‐cell phenotypes. The results have major implications for how UC should be understood from a molecular perspective.

2Formalin‐fixed paraffin‐embedded preoperative TUR‐B specimens reviewed by a uropathologist were collected from consecutive patients who underwent radical cystectomy in four hospitals in southern Sweden from 2006 to 2011. None of the included tumours received any treatment prior to sample‐taking (Trans uretral resection of the bladder (TUR‐B)). The histological variants at pathological review are shown in supplementary material, Table S1. The pathological stage and grade based on TUR‐B specimens are given in supplementary material, Table S2. Sufficient tissue for embedding of dual cores (1.0 mm) in tissue microarrays (TMAs) and for extraction of RNA was obtained from 307 TUR‐B specimens. For three cases, single cores were embedded. Positioning of the cores within the TUR‐B specimen was performed by selecting areas with >80% tumour cells morphologically and histologically representative of the total tumour. RNA extraction was performed on 4–10 10‐µm sections from macrodissected tissue areas located as close as possible to the positions of the TMA cores. Sampling and RNA extraction are described in detail Supplementary materials and methods. The included samples had an average total spectrometric RNA yield of 2.5 µg (range 0.4–9.3 µg). Informed consent was obtained from all patients, and the study was approved by the Local Ethical Committee of Lund University, in accordance with the Helsinki Declaration.

3RNA samples were amplified and labelled by use of the SensationPlus kit (Affymetrix, Santa Clara, CA, USA), and raw data were generated with the human Gene ST 1.0 platform (Affymetrix). Batch effects adjustments were performed with the COMBAT algorithm 15. A complete technical description of data processing is included in supplementary material, Supplementary materials and methods. The resulting dataset included 14 062 genes and 307 samples. Raw and normalized data have been deposited in Gene Expression Omnibus under GSE83586. Sample clusters were identified by sequential two‐way splits with the ConsensusClusterPlus R package on the top 50% genes (n = 7031) with respect to variance.

4Antibodies against CCNB1, CCND1, CDH1, CDH3, CDKN2A (p16), CHGA, E2F3, EPCAM, FGFR3, FOXA1, GATA3, KRT5, KRT14, KRT20, NCAM1, PPARG, RB1, RXRA, SYP, TP63, TUBB2B, UPK3, VIM and ZEB2 and antibodies for eight additional supportive markers were used (Supplementary materials and methods). The full dataset (IHC evaluations) is available as supplementary material, Table S6. Staining was also performed with haematoxylin and eosin and with antibodies against CD3, CD68 and smooth muscle actin (ACTA2), to facilitate the identification of immune cell infiltration and stroma. Staining procedures have been described previously 8. Slides were scanned (AxioScan Z1; Zeiss, Oberkochen, Germany), and image files were evaluated with the digital pathology platform PathXL (PathXL, Belfast, UK).

5A genomic circuit score 5 was calculated from log2 mRNA expression data and from IHC data as FGFR3 + CCND1 + RB1 – E2F3 of either log2 values (mRNA) or labelling intensity (IHC). Similarly, consensus basal/SCC‐like marker scores (Ba/Sq scores) 4 were calculated as FOXA1 + GATA3 – KRT5 – KRT14 of either log2 values (mRNA) or labelling intensity (IHC). The proposed definitions of tumour‐cell phenotypes were translated to IHC phenotype scores as described in Supplementary materials and methods. Cases with missing data for the defining markers were not classified.

6Statistical tests were performed with the R software environment for statistical computing. For group comparisons of IHC data, two‐sided non‐parametric test with significance threshold α = 0.05 were used unless otherwise indicated. For group comparisons of gene expression data, false discovery rates were controlled at a maximum of q = 0.01.

7We performed stepwise unsupervised hierarchical clustering using whole genome mRNA expression for 307 advanced bladder cancers, resulting in six consensus clusters of approximately equal sizes. To compare the clustering results with known molecular subtypes of UC, we used six gene expression signatures with subtype‐specific expression (supplementary material, Figure S1). On the basis of these signatures, we concluded that the structure of the present data conforms to, and reproduces, clusters obtained in the TCGA RNA‐Seq dataset (N = 238) reported by Aine et al 14, and assigned the six consensus clusters the following labels: urothelial‐like (Uro, n = 41), genomically unstable (GU, n = 66), epithelial‐infiltrated (Epi‐Inf, n = 51), Squamous cell carcinoma-like/mesenchymal-infiltrated (SCCL/Mes‐Inf, n = 56), SCCL/UroB (n = 46), and small‐cell/neuroendocrine‐like (Sc/NE, n = 47) (Figure 1A). Throughout the analyses, we will use subtype abbreviations in italics to refer to these gene expression clusters (e.g. Epi‐Inf), whereas we will use non‐italicized abbreviations for tumour‐cell phenotypes determined by IHC (e.g. basal/SCC‐like).

8We compared the obtained clusters with published molecular subtype classification algorithms, clinical and pathological data, and expression of Uro‐diff. The combined data indicate that high Uro‐diff expression is a hallmark of the ‘luminal‐type’ mRNA clusters Uro, GU, and Epi‐Inf, whereas lack of signature expression is characteristic for SCCL/Mes-Inf, SCCL/UroB, and Sc/NE clusters (Figure 1A). Next, we compared global (mRNA) and tumour‐cell specific (IHC) phenotypes of the Uro‐Diff‐positive mRNA clusters Uro, GU, and Epi‐Inf. We used mRNA and protein expression data for FGFR3, CCND1, E2F3, and RB1, which make up a genomic circuit that separates the Uro subtype from the GU subtype (Figure 1B) 5. The calculated genomic circuit scores were, as expected, high for tumours in the Uro cluster, i.e. FGFR3+, CCND1+, RB1+, and E2F3−, both at the mRNA level and the tumour‐cell protein level (Figure 1B). This was true regardless of stage, as indicated by the T1 tumours included in the study (Figure 1A). In the GU consensus cluster, approximately half showed the anticipated strong decrease in circuit score, i.e. FGFR3−, CCND1−, RB1−, and E2F3+, whereas the other half showed scores similar to those of the Uro cluster. This is in line with results obtained by the Lund classification algorithm, which identified several potential Uro cases as a part of the GU cluster, and thus GU‐Uro cases (Figure 1A). We confirmed this finding by analysis of CDKN2A (p16) protein expression, which was also low in both Uro and GU‐Uro cases, but high in GU‐GU cases (Figure 1B, C), consistent with CDKN2A deletions/mutations being frequent in progressed Uro cases 5 and with GU showing frequent overexpression of p16 8. The GU‐Uro tumours differed from Uro tumours by increased proliferation, immune and extracellular matrix (ECM) mRNA signatures, but protein expression levels of the canonical Uro genes FGFR3, CCND1 and TP63 were not different in the Uro versus GU‐Uro groups of tumours (supplementary material, Figure S2). We therefore conclude that tumours of both Uro and GU tumour‐cell phenotypes are observed in the GU consensus mRNA cluster. Analysis of the Epi‐Inf consensus cluster at the mRNA level is compromised by high levels of infiltrating non‐tumour cells (supplementary material, Figure S1). However, of 47 Epi‐Inf cases subjected to IHC analysis, 30 were considered to have Uro and 10 GU tumour phenotypes, whereas seven had indecisive results at this stage (Figure 1B, C). Thus, the Epi‐Inf cluster is mainly composed of tumours with Uro or GU tumour‐cell phenotypes.

9Next, we set out to dissect tumour‐cell phenotypes in the Uro‐diff‐negative subtypes. We used the consensus definition of basal/SCC‐like tumours, KRT5/KRT14‐high and FOXA1/GATA3‐low 4. This set of markers clearly identified the SCCL/Mes‐Inf and the SCCL/UroB clusters as being composed mainly of basal/SCC‐like cases, with scores based on either mRNA expression or tumour‐cell protein expression (Figure 2A). The typical basal/SCC‐like cases also showed a shift from high EPCAM and CDH1 and low CDH3 expression in Uro and GU cases to lower EPCAM and CDH1 and high CDH3 expression (Figure 2B). One portion of the SCCL/Mes‐Inf cluster was negative for KRT5/KRT14 and FOXA1/GATA3, as well as for CDH3 expression, making it distinct from basal/SCC‐like tumours (Figure 2A, B). The most upregulated mRNAs in this group, as compared with the basal/SCC‐like cases in the same cluster, were ZEB2 and VIM (supplementary material, Table S1), identifying this group as the Mes‐Inf mRNA cluster 14. The Mes‐Inf tumours were negative for a large number of basal cell‐related and SCC‐related cytokeratins, but positive for tumour‐cell expression (IHC) of both ZEB2 and VIM (Figure 3A, B). It is of note that, at the mRNA level, a large proportion of the basal/SCC‐like tumours in the same consensus cluster expressed VIM, but the protein was expressed in infiltrating mesenchymal cells and not by the tumour cells in these cases (Figures 2B and 3A, B). This makes the Mes‐Inf tumour‐cell phenotype distinct from the basal/SCC‐like cases and more similar to a mesenchymal than to a basal epithelial phenotype, even though they belong to the same global mRNA‐based tumour cluster.

10A subset of tumours in the SCCL/UroB cluster showed simultaneous expression of the Uro‐diff, FGFR3 and SCC signatures; although the latter did not reach the levels typically seen for basal/SCC‐like cases (supplementary material, Figure S1). In addition, the same cases were classified as UroA or UroB by the Lund classifier (Figure 2A). Tumours in this subset expressed FGFR3 and CCND1 as determined by IHC at the same levels as the Uro tumours, indicating a Uro‐related phenotype. Hence, this portion of the consensus cluster corresponds to UroB tumours. The basal/SCC‐like cases in the same consensus cluster showed almost no expression of FGFR3 and CCND1, whereas the reverse was seen for KRT5 and CDH3 tumour‐cell expression (Figure 3C, D). A significance analysis of microarrays (SAM) on mRNA data clearly distinguished the UroB portion from the basal/SCC‐like portion in the same cluster, with the canonical Uro‐diff genes GATA3, FOXA1 and PPARG being among the top upregulated genes (supplementary material, Table S3). Conversely, the phenotypically basal/SCC‐like cases were distinguished by expression of SCC/desmosome markers, e.g. KRT6B and DSC2. Taken together, these findings indicate that the UroB tumour‐cell phenotype is distinct and more similar to a Uro phenotype than to a basal/SCC‐like phenotype, even though they are grouped within the same cluster by global mRNA clustering.

11The above results define two possible basal/SCC‐like categories, one with strong ECM and T‐cell signatures that segregates with the Mes‐Inf subtype, and one with weaker ECM and T‐cell signatures that segregates with the UroB subtype, on global mRNA clustering. However, these two basal/SCC‐like tumour categories do not differ in their KRT5, KRT14/FOXA1, GATA3 ratios, their defining characteristics, or their shifts to high CDH3 and lower CDH1 expression. Hence, the difference between these two categories, and the reason for them to be part of two different global mRNA clusters, is determined by signatures related to infiltrating non‐tumour cells.

12Cases in the Sc/NE consensus cluster showed coordinated overexpression of E2F3, CDKAL1, SOX4 and MBOAT1, which are genes located in the core region of the 6p22 amplicon (Figure 2C). The Sc/NE tumours share this feature with those showing a true GU IHC phenotype in the GU cluster. To resolve the Sc/NE cluster further, we used mRNA levels for CHGA, SYP and ENO2 and applied IHC for CHGA, SYP, and NCAM1 (CD56), all of which are markers for either a small cell or a neuroendocrine phenotype 16 (Figure 2C). Only one‐half of the Sc/NE consensus cluster showed enriched expression of the markers. SAM revealed that several tubulin genes were among the top genes co‐expressed with markers for a neuroendocrine phenotype (supplementary material, Table S3). The top upregulated genes were normally restricted to cells of neuroendocrine origin (supplementary material, Figure S3). Expression of TUBB2B (tubulin β2B), the second most significant upregulated gene in this analysis, coincided with expression of the small‐cell/neuroendocrine differentiation genes (Figure 2C), and IHC showed that it was expressed by tumour cells (Figure 4A). This identifies a group of urothelial carcinomas with a tumour‐cell phenotype reminiscent of neuroendocrine cells. We named this phenotype small‐cell/neuroendocrine‐like (Sc/NE‐like). However, these features accounted for only one half of the Sc/NE consensus cluster. The second half showed significantly higher mRNA expression levels of the Uro‐diff gene signature (Figure 4B; supplementary material, Figure S4), as well as of the key transcription factors FOXA1 and GATA3 (Figure 4C; supplementary material, Figure S4), and appeared to be indistinguishable from GU cases at the tumour‐cell level. Expression of canonical neuroendocrine markers was not detected in this half of the cluster (Figure 4D, E). SAM confirmed that the Uro‐diff signature drives the separation of these subclusters, as the canonical genes GATA3, UPK2 and PPARG were among the top 10 genes as compared with Sc/NE‐like cases (supplementary material, Table S3). Hence, this group has several features in common with the Uro‐diff‐positive half of the cohort, and with a GU tumour‐cell phenotype in particular.

13We then inspected the expression of the transcription factor genes RXRA, PPARG, FOXA1, GATA3, and ELF3, which are known to be central in differentiation of the normal urothelium 17, 18, 19, 20, 21. These factors showed coordinated high expression in both the Uro and GU and in the Epi‐Inf consensus clusters, as determined by both mRNA analysis and IHC (ELF3 was not included in the latter analysis) (Figure 5A). Coordinated downregulation of these genes was observed in the SCCL/Mes‐Inf and SCCL/UroB clusters. In the latter group, UroB cases showed expression of these factors to a varying extent. The Sc/NE consensus cluster was split into two groups, one associated with the Sc/NE‐like cases with an absence of factor expression, and one in which the factors and signature were expressed at varying levels, further strengthening the conclusion that the Sc/Ne consensus cluster consists of tumours of two different tumour‐cell phenotypes.

14Next, we investigated the relationship between relative mRNA levels and the tumour‐cell expression of the terminal differentiation markers KRT20 and UPK3. We hypothesized that not only marker levels but also subcellular expression patterns may reveal where and if urothelial differentiation takes place. KRT20 and UPK3 staining was remarkably inconsistent; a multitude of aberrant and heterogeneous expression patterns were observed for each marker, and, unlike that of the regulatory factors, KRT20 and UPK3 expression was hardly ever homogeneous in a tumour area 1 mm in diameter. Among the most conspicuous of these observations was the stochastic appearance of tumour‐cell KRT20 staining, and, even more, the aberrant subcellular localization of these markers (Figure 5B, C). Moreover, KRT20 and UPK3 were observed in each of the three types of invasive growth pattern described for UC, i.e. infiltrative, nodular, and trabecular 22 (supplementary material, Figure S5), as well as in cells with low differentiation levels, loss of cellular adhesion, and highly atypical and pleomorphic nuclei (Figure 5B, C; supplementary material, Figures S5 and S6). Taken together, these results show that regulatory factors accurately identify Uro‐diff‐positive tumours at both the mRNA level and the protein level, but, whereas KRT20 and UPK3 mRNA expression overlaps with the Uro‐diff signature, only subsets of the tumour cells are positive at the IHC level, and expression patterns are inconsistent with the proteins' structural functions in normal differentiation.

15On the basis of the extensive IHC analyses, we propose working definitions for five major tumour‐cell phenotypes in advanced bladder cancer: urothelial‐like (Uro), genomically unstable (GU), basal/SCC‐like, mesenchymal‐like (Mes‐like), and small‐cell/neuroendocrine‐like (Sc/NE‐like), as described in Figure 6. As there is no reliable way to distinguish UroA from UroB, or from GU‐Uro, these tumours are grouped under the ‘Uro’ heading. The defining protein markers represent key genes in tumour biology central to the respective tumour cell phenotypes. We then assigned each tumour an mRNA phenotype based on the deconstructed consensus clusters, and grouped Uro‐Uro with GU‐Uro and SCCL/UroB‐UroB in one Uro class, combined GU‐GU and Sc/NE‐GU into one GU class, and combined all basal/SCC‐like cases into one class, and tested to what extent the IHC tumour phenotype definitions could recreate the data (supplementary material, Figures S7 and S8). The overall accuracy was 75%, being mainly reduced by the low sensitivity for the two small Mes‐like and Sc/NE‐like classes of tumour. The sensitivity was, however, high for Uro and GU tumours (0.89 and 0.79, respectively), and the specificity for the basal/SCC‐like tumour‐cell phenotype definition was excellent (0.93) (supplementary material, Tables S4 and S5). Hence, we see the presented tumour phenotype definitions as starting points for efficient classification of advanced bladder cancer into relevant and distinct molecular tumour‐cell phenotypes.

16Recent advances in the field of UC classification have put increasing weight on urothelial differentiation signatures to subdivide UCs. Such a signature is the driving force behind both the UNC and the MDA classes termed ‘luminal’ 9, 10. With the nomenclature used in the present investigation, their ‘luminal’ categories consist of our Uro, GU and Epi-Inf clusters of gene expression. Indeed, our urothelial differentiation mRNA signature (Uro‐diff) includes both KRT20 and uroplakin genes normally expressed in luminal cells. Furthermore, the luminal category of tumours frequently express the transcription factors RXRA, PPARG, FOXA1, and GATA3, which are known to have a crucial role in urothelial differentiation. However, the full complement of factors is not always expressed in tumour cells (Figure 5A) 23. We compared gene expression data and IHC data for the terminal differentiation markers KRT20/UPK3, and for RXRA, PPARG, FOXA1, and GATA3, to establish to what extent urothelial differentiation occurs in the Uro‐diff signature subtypes. According to the mRNA data, a large proportion of the tumours expressed both KRT20 and UPK3. Protein expression was, however, typically absent or patterns were far from normal. This is in contrast to NMI cases, which may retain KRT20 and UPK expression in the most apical cells of the tumour papillae 24, 25. We consequently hypothesize that the urothelial differentiation programme becomes corrupted during progression, that KRT20 and uroplakins lose their relevance as luminal differentiation markers, and that pseudo‐differentiation occurs in advanced urothelial tumours.

17By exploring gene expression signatures, tumour clustering, and data on tumour‐cell expression of a large panel of IHC markers, we arrived at definitions of five tumour‐cell phenotypes: Uro, GU, basal/SCC‐like, Mes‐like, and Sc/NE‐like. We see these as working definitions that have to be validated in independent series of tumours for which both global mRNA and IHC data are available. The definition of the Uro category of tumours does not distinguish between the predominantly NMI UroA, and UroB. Furthermore, a large proportion of the invasive cases with a Uro tumour‐cell phenotype clustered with GU cases in the GU consensus cluster, making them different from both UroA and UroB. We tentatively name this group of Uro tumours UroC. This group differs from early‐stage UroA by being invariably high‐grade, by having lost urothelial‐like stratification, and by showing frequent CDKN2A genomic losses 14. UroC shares this last feature with UroB, but, whereas UroB cases start to attain SCC‐like features and cluster together with frank basal/SCC‐like cases, UroC cases instead become increasingly similar to, and finally converge with, frank GU cases at the global gene expression level. As UroA, UroB and UroC are very similar in fundamental gene regulatory features, and thus in ‘type’, designing a simple IHC marker system to distinguish them as separate entities is challenging. Information on global gene expression profiles (cluster assignment) would, in this case, be useful. Nevertheless, this makes Uro the most heterogeneous of the molecular phenotypes and with the most prominent signs of biological progression routes.

18The definition of basal/SCC‐like urothelial tumours is established and conforms well to definitions of similar subtypes in other tumour types, i.e. basal‐like breast cancer and SCC of the lung. Nevertheless, frank basal/SCC‐like UC was grouped into two different global gene expression clusters. In this case, divergence is most likely attributable to the profiles of tumour‐infiltrating non‐tumour cells. One cluster, SCCL/Mes‐Inf, showed a strong ECM signature, indicating the presence of a large proportion of mesenchymal cells in these tumours. On application of the tumour‐cell phenotype approach, it became evident that a portion of these tumours did not conform to the basal/SCC‐like definition. Instead, the tumour cells were themselves mesenchymal‐like and expressed the typical mesenchymal genes ZEB2 and VIM, indicating a tumour type that has undergone epithelial–mesenchymal transition (EMT). The basal/SCC‐like tumour samples also expressed ZEB2 and VIM, but only in surrounding stromal cells. Thus, the Mes‐like subtype defines a new entity of bladder cancer showing a tumour‐cell phenotype that is in stark contrast to previously defined subtypes, and cases of which are biologically very different from the basal/SCC‐like cases that they cluster together with.

19The consensus cluster Sc/NE‐like turned out to harbour two very distinct tumour‐cell phenotypes. One‐half of these tumours expressed markers that are typical for neuroendocrine differentiation, e.g. SYP and ENO2, and, more robustly, TUBB2B. This and other tubulin isoforms have normal expression that is highly specific for tissues of neuroendocrine origin. This part of the Sc/NE consensus cluster also showed an absence of PPARG, FOXA1 and GATA3 expression, as well as of uroplakin and KRT20 expression. We name this tumour‐cell phenotype small‐cell/neuroendocrine like (Sc/NE‐like). The second half of the Sc/NE consensus cluster expressed PPARG, FOXA1, GATA3, and ELF3, as well as the Uro‐diff signature. This, in combination with the absence of FGFR3 and CCND1 protein expression, suggests that these tumours have a GU tumour‐cell phenotype. The forces driving Sc/NE‐like and GU tumours into one consensus cluster are most likely a high proliferation rate and frequent E2F3, CDKAL1 and SOX4 genomic amplification at 6p22, which are characteristic of both subtypes.

20The main finding of the present study is that global mRNA clustering and tumour‐cell phenotype analyses lead to different groupings of bladder cancer samples. We observed several instances where global mRNA clusters did not show the expected tumour‐cell phenotype: (1) approximately half of the tumours in the GU consensus cluster were of the Uro tumour‐cell phenotype; (2) basal/SCC‐like tumours were allocated to two different global mRNA tumour clusters; (3) the two small but distinct tumour subtypes Mes‐like and UroB grouped with typical basal/SCC‐like cases; (4) the Epi‐Inf consensus cluster was found to consist of cases with Uro and GU tumour‐cell phenotypes; and (5) the Sc/NE consensus cluster consisted of tumours with typical neuroendocrine and GU tumour‐cell phenotypes. We interpret these observations as signs of convergence/divergence at the global gene expression level that become evident when tumour phenotypes are measured both at the tumour‐cell level and by global mRNA profiling, as shown schematically in Figure 6. Hence, a complex relationship is apparent between global mRNA clusters and tumour‐cell phenotypes in UC. It is worth noting that this effect is only observed in an MI setting, and not in NMI‐based datasets (e.g. 7, 26), suggesting that broad global commonalities, perhaps related to the invasive process itself, exist between MI tumours regardless of subtype. This also makes the denotation of tumour subtype labels problematic; is the label intended to denote a specific tumour‐cell type, or a tumour type with a given composite organization? This discrepancy suggests that a bi‐nominal classification system consisting of both tumour‐cell phenotype and gene expression cluster (context) would be more appropriate. Such a bi‐nominal system may also bridge the gap between genome‐wide expression profiling and traditional (molecular) pathology. Future clinical studies/trials will show when, and under what circumstances, such a high‐resolution classification system is needed, and when a low‐resolution classification is fully adequate.

21GS: performed IHC, including selection of antibodies, scoring of staining, and biological interpretation; PE: performed bioinformatic analyses, including hierarchical clustering and specific signature analyses; FL: responsible for data collection, including pathological re‐evaluation, clinical follow‐up, and ethical issues; MH: conceived the study, was responsible for final interpretations and conclusions, and was responsible for the final draft of the manuscript. All authors contributed to the drafting of the manuscript.

22Supplementary materials and methods Supplementary figure legends Figure S1. Consensus clustering of global mRNA expression from cystectomized patients (TUR‐B samples) Figure S2. Cases with Uro phenotype that cluster in the GU consensus cluster retain their Uro molecular profile, but proliferation, T‐cell and ECM signals are affected Figure S3. Top 10 upregulated genes in Sc/NE–Sc/NE group (compared to Sc/NE–GU) frequently have a normal tissue expression pattern restricted to neuronal/endocrine tissues Figure S4. Cases with GU phenotype that cluster in the Sc/NE consensus cluster retain expression of the Uro‐Diff signature (Mean mRNA) and key factors (IHC) Figure S5. Protein expression pattern of KRT20 in advanced urothelial carcinoma is de‐regulated and inconsistent with normal urothelial differentiation Figure S6. Protein expression of UPK3 in advanced urothelial carcinoma is de‐regulated and inconsistent with normal urothelial differentiation Figure S7. IHC data from 29 markers showing differential expression in tumour cells were considered in establishing tumour‐cell phenotype definitions Figure S8. Concordance of tumour‐cell phenotype with consensus clusters and deconstructed subtypes Table S1. Histological variants included in the consecutive series of 307 tumours according to the 2016 WHO classification of invasive bladder tumours Table S2. The pathological stage and grade based on uro‐pathologist re‐evaluation of TUR‐B specimens for the consecutive series of 307 tumours Table S3. Differentially expressed genes between Mes‐Inf and Basal/SCC‐like, between UroB and Basal/SCC‐like, and between Sc/NE and GU Table S4. Comparison of tumour‐cell phenotypes (applied IHC definitions) to global mRNA consensus clusters (top) Table S5. Comparison of tumour‐cell phenotypes (applied IHC definitions) to deconstructed mRNA clusters (as described in text) Table S6. The complete IHC‐evaluation data set

23Supplementary materials and methods

24Supplementary figure legends

25Figure S1. Consensus clustering of global mRNA expression from cystectomized patients (TUR‐B samples)

26Figure S2. Cases with Uro phenotype that cluster in the GU consensus cluster retain their Uro molecular profile, but proliferation, T‐cell and ECM signals are affected

27Figure S3. Top 10 upregulated genes in Sc/NE–Sc/NE group (compared to Sc/NE–GU) frequently have a normal tissue expression pattern restricted to neuronal/endocrine tissues

28Figure S4. Cases with GU phenotype that cluster in the Sc/NE consensus cluster retain expression of the Uro‐Diff signature (Mean mRNA) and key factors (IHC)

29Figure S5. Protein expression pattern of KRT20 in advanced urothelial carcinoma is de‐regulated and inconsistent with normal urothelial differentiation

30Figure S6. Protein expression of UPK3 in advanced urothelial carcinoma is de‐regulated and inconsistent with normal urothelial differentiation

31Figure S7. IHC data from 29 markers showing differential expression in tumour cells were considered in establishing tumour‐cell phenotype definitions

32Figure S8. Concordance of tumour‐cell phenotype with consensus clusters and deconstructed subtypes

33Table S1. Histological variants included in the consecutive series of 307 tumours according to the 2016 WHO classification of invasive bladder tumours

34Table S2. The pathological stage and grade based on uro‐pathologist re‐evaluation of TUR‐B specimens for the consecutive series of 307 tumours

35Table S3. Differentially expressed genes between Mes‐Inf and Basal/SCC‐like, between UroB and Basal/SCC‐like, and between Sc/NE and GU

36Table S4. Comparison of tumour‐cell phenotypes (applied IHC definitions) to global mRNA consensus clusters (top)

37Table S5. Comparison of tumour‐cell phenotypes (applied IHC definitions) to deconstructed mRNA clusters (as described in text)

38Table S6. The complete IHC‐evaluation data set