EURURO Vol. 72 No. 4

3.2.

Similarities between subtypes across different methods

As recently reported, claudin-low tumors have basal

characteristics with high expression levels of genes indica-

tive of epithelial-to-mesenchymal transition (EMT) and

immune infiltration

( Fig. 1

C–E, and Supplementary Fig. 4A

and 4C). Tumors in TCGA cluster IV identify a similar subset

of basal tumors: 21/28 cluster IV tumors were labeled

claudin-low. Cluster IV showed enrichment for EMT,

chemokine signaling, and immune infiltration compared

with cluster III tumors

( Fig. 1

C, 1D, 1F, and Supplementary

Fig. 4B and 4D). Cluster II represents luminal tumors that

are immune infiltrated

[18]

. Compared with cluster I

tumors, cluster II tumors were EMT-signature positive

and enriched for inflammatory response and chemokine

signaling

( Fig. 1 F

, and Supplementary Fig. 4B and 4E).

3.3.

Patient benefit from NAC differs between subtypes

Extravesical extension of residual primary tumor (pT3/4)

after NAC was observed more frequently in claudin-low

(49%), p53-like (

[20_TD$DIFF]

38%), and cluster II (40%) subtypes. None of

the subtypes was associated with a major response to NAC

(ypT

2N0) on final pathology (Supplementary Table 3).

Overall survival (OS) varied by molecular subtype and

between non-NAC and NAC patients

( Fig. 2

). By all

subtyping methods, non-NAC-treated patients (TCGA co-

hort) with basal tumors had worse outcomes compared

with those luminal tumors

( Fig. 2 )

. Of both luminal TCGA

clusters, patients with cluster II tumors experienced worse

OS than those with cluster I tumors. Poor OS was also

observed in patients with claudin-low tumors.

The association between subtype and prognosis differed

in our NAC cohort. The most pronounced difference was in

patients with basal or equivalent subtypes (cluster III,

urobasal [Uro] B, and squamous cell carcinoma [SCC]-like),

who experienced a dramatic improvement in OS after NAC

compared with the TCGA patients who received no NAC

( Fig. 2 )

. However, there was no such shift in survival in

patients with claudin-low and cluster IV tumors, indicating

that these patients fared poorly regardless of NAC even

though they are subsets of the basal subtype. Across the

different subtyping methods, patients with luminal or

equivalent tumors had the best OS with or without NAC,

with the exception of cluster II patients, who fared poorly in

both settings, as did patients with p53-like tumors. There

was no noteworthy difference in these trends when

the patients were analyzed according to NAC regimen

(gentacibine vs methotrexate, vinblastine, adriamycin, and

cisplatin; Supplementary Fig. 5).

In summary, the differences in OS by subtype and the

apparent impact of NAC suggest that a classification into

four subtypes would have the greatest clinical relevance.

Basal tumors warrant subclassification into tumors without

and with EMT and immune infiltration (ie, basal and

claudin-low, respectively), since NAC appeared to have the

greatest impact on noninfiltrated basal tumors. Luminal

tumors similarly warrant subclassification into tumors

without and with EMT and immune infiltration (ie, luminal

and luminal-infiltrated, respectively), since OS differed

between the two groups.

3.4.

Single-sample classifier to predict bladder cancer subtypes

Based on the biological characteristics and different impacts

on clinical outcome, we trained a single-sample genomic

subtyping classifier (GSC) to predict four classes based on the

consensus of the different classification schemes: claudin-

low, basal, luminal-infiltrated, and luminal

( Fig. 3

A and

Supplementary Table 4). Compared with previously pub-

lished methods, the single-sample GSC was more discrimi-

nate in assigning individual patients to a definitive subtype,

as seen by the number of patients who have a dominant

subtype score

( Fig. 3

B and Supplementary Fig. 6). Under 10-

fold cross validation, the overall accuracy of GSC assignment

to the four classes in the discovery cohort was 76% (

= 223).

In the validation cohort (

= 82), the accuracy was 73% and

significantly higher (

0.001) compared with the no-

information rate of 39% (defined by assignment of the

subtype based on majority class). A multinomial goodness-

of-fit test comparing the predicted probabilities from the

GSC with the consensus subtype classes using the validation

cohort found that the independent set was well predicted by

the model and not significantly different compared with the

discovery cohort (

= 0.47). Furthermore, in both the NAC

and the non-NAC cohort (

= 476), consensus subtype

classes (ie, those obtained using previously published

clustering-based approaches) were all predicted with areas

under the curve

0.85

( Fig. 3

C).

3.5.

Prediction of clinical endpoints using GSC

In both NAC datasets, we used only cases that were treated

with cisplatin-based NAC for analysis of outcomes. The

clinical significance of the predicted classes in the NAC

datasets was compared with an independent non-NAC

active p53 signature (p53

[6_TD$DIFF]

-like). The TCGA subtyping defines four clusters that are also basal (clusters III and IV) and luminal (clusters I and II). The

Lund group discovered five subtypes that can be considered basal (Uro B and SCC-like), luminal (Uro A and genetically unstable), and infiltrated. (B)

Heatmap of biologically relevant gene signatures (rows) in pre-NAC TUR samples from the discovery cohort. The column annotation across the top

provides the subtype calls from each classification system. Claudin-low and cluster IV tumors showed the highest expression of the T-cell and myeloid

cell signatures. Genes expressed in the ECM are expressed in the p53-like and infiltrated subtypes. Proliferation markers were highly expressed in

genomically unstable tumors, while the Uro A subtype expressed an FGFR3 signature. Luminal tumors across subtyping methods express urothelial

differentiation genes more highly. (C) Heatmap of two bidirectional EMT signatures (Tan et al and Kardos et al

[15]

). Claudin-low and cluster IV

tumors both were EMT-signature positive, and cluster II tumors showed more EMT than cluster I tumors. (D) Enrichment plots of the hallmark EMT

signature in claudin-low versus basal tumors (left) and cluster IV versus III (right). Both subgroups showed significant enrichment of EMT markers.

Differential expression of immune markers (E)

CXCL9

and (F)

CD8A

in the UNC subtypes (left) and the TCGA clusters (right). The claudin-low and

cluster IV tumors showed the highest expression of immune markers. Reference levels: claudin-low, cluster IV. Diff. = differentiation;

ECM = extracellular matrix; EMT = epithelial-to-mesenchymal transition; GU = genomically unstable; MDA = MD Anderson Cancer Center;

NAC = neoadjuvant chemotherapy; SCC = squamous cell carcinoma; TCGA = The Cancer Genome Atlas; TUR = transurethral resection;

TURBT = transurethral resection of bladder tumor; UNC = University of North Carolina; Uro = urobasal.

E U R O P E A N U R O L O G Y 7 2 ( 2 0 1 7 ) 5 4 4 – 5 5 4

548