Multi‐omics consensus ensemble refines the classification of muscle‐invasive bladder cancer with stratified prognosis, tumour microenvironment and distinct sensitivity to frontline therapies

Dear Editor, The molecular classification of muscle-invasive bladder cancer (MIBC) based on transcriptomic signatures has been extensively studied.1–4 The complementary nature of information provided by different molecular profiles motivated us to refineMIBC classification by aggregatingmultiomics data. The entire workflow is outlined in Figure 1A, information on the eight datasets is summarized in Tables S1-S2, and technical details are listed in the Supporting Materials and Methods. A consensus ensemble was generated through 10 multi-omics integrative clustering approaches on five omics datasets of 396 MIBCs from the TCGA database.5 We identified four robust integrative consensus subtypes (iCSs; Figure S1A,B), which showed distinctive molecular patterns and were significantly associated with clinicopathological features and previously identified molecular classifications (Figure 1B); iCS1 and iCS4 dramatically overlapped with basal-like subtypes, whereas iCS2 and iCS3 were enriched for luminal-like subtypes (p < 0.001; Figure 1C, Table S3). We relabelled iCS1 to iCS4 as iBS1, iLS2, iLS3, and iBS4, respectively. Our classification system was tightly associated with prognosis (Figure 1D,E), showing superior performance than PAM, oneNN, and Lund classifications, but inferior than TCGA and consensus molecular subtype (CMS) classifications regarding overall stratification (Figure S2); iBS4with theworst outcome that was refined from basal-like subtypes caught our attention. Regulon analysis strongly manifested the biological pertinency of a four-classification where differentiated activity of cancerous chromatin remodelling regulons highlighted other possible differential regulatory patterns among iCSs (Figure 2A), indicating that epigenetically driven transcriptional networksmight be important differentiators of these subtypes.


L E T T E R T O E D I T O R
Multi-omics consensus ensemble refines the classification of muscle-invasive bladder cancer with stratified prognosis, tumour microenvironment and distinct sensitivity to frontline therapies Dear Editor, The molecular classification of muscle-invasive bladder cancer (MIBC) based on transcriptomic signatures has been extensively studied. [1][2][3][4] The complementary nature of information provided by different molecular profiles motivated us to refine MIBC classification by aggregating multiomics data. The entire workflow is outlined in Figure 1A, information on the eight datasets is summarized in Tables S1-S2, and technical details are listed in the Supporting Materials and Methods.
A consensus ensemble was generated through 10 multi-omics integrative clustering approaches on five omics datasets of 396 MIBCs from the TCGA database. 5 We identified four robust integrative consensus subtypes (iCSs; Figure S1A,B), which showed distinctive molecular patterns and were significantly associated with clinicopathological features and previously identified molecular classifications ( Figure 1B); iCS1 and iCS4 dramatically overlapped with basal-like subtypes, whereas iCS2 and iCS3 were enriched for luminal-like subtypes (p < 0.001; Figure 1C, Table S3). We relabelled iCS1 to iCS4 as iBS1, iLS2, iLS3, and iBS4, respectively. Our classification system was tightly associated with prognosis ( Figure 1D,E), showing superior performance than PAM, oneNN, and Lund classifications, but inferior than TCGA and consensus molecular subtype (CMS) classifications regarding overall stratification ( Figure S2); iBS4 with the worst outcome that was refined from basal-like subtypes caught our attention.
Regulon analysis strongly manifested the biological pertinency of a four-classification where differentiated activity of cancerous chromatin remodelling regulons highlighted other possible differential regulatory patterns among iCSs (Figure 2A), indicating that epigenetically driven transcriptional networks might be important differentiators of these subtypes.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. We observed that iLS2 showed a higher tumour mutation burden (TMB, p = 0.002; Figure 2B,C) and harboured more mutations in APOBEC-relevant signatures (p = 0.007; Figure 2D). Of the frequently mutated genes (>5%), iBS1 harboured more mutations of TP53, RB1 and KMT2A, while iBS4 was enriched for mutations in NFE2L2 (also known as NRF2) and TRANK1; KIAA0947, MED12, COL6A6, and ARID2 were mutated more frequently in iLS2, whereas iLS3 was enriched for mutations in FGFR3, STAG2 and SPEN (Table S4). iBS1 had better chromosomal stability with lower copy number alterations (CNAs) (p < 0.001; Figure 2E). Ch9p21.3 was susceptible to inactivation in cell immortalization and diseases; within this region, interferon-alpha (IFNA) genes, MTAP and CDKN2A/B were differentially altered in iBS1 compared to others, which may contribute to shaping different basal-like subtypes ( Figure 2B).
We generated a 120-gene signature to predict iCSs in external cohorts with 403 MIBCs ( Figures S3A,B and S4A-C, Table S5). The signature-predicted iCSs highly overlapped with the CMS but further refined the basal-like classification (p < 0.001; Figure S4D,E). Consistently, iBS4 showed the most unfavourable prognosis ( Figure S4F-G).
Since cancer immunity plays a critical role in tumour progression, tumour microenvironment deconvolution suggested that immunocyte infiltration was dramatically higher in both iBS1 and iLS2 than in the other subtypes ( Figure 3A, Table S6); iBS1 had relatively higher expression of genes that represent potential targets for immunotherapy. The immune landscape was validated ( Figure S5A , and potential regulators that are relevant with chromatin remodelling (middle panel); Similar patterns were shared by iBS1 and iBS4, but iBS1 differed with high activity of GATA6, FGFR1 and ESR1 for regulon activity, while iBS4 was distinctly associated with high activity of TP63. A total of 296 unique differentially methylated promoters derived from each iCS versus adjacent normal samples (bottom panel); the iLS3 (265 vs. 45 in iLS2) and iBS4 (191 vs. 26 in iBS1) had more hypermethylated promoters (296 unique loci) than the 21 normal samples had. (B) Genomic alteration landscape according to iCS. Samples are sorted in descending order according to the cumulative contribution of APOBEC-relevant mutational signatures (i.e., SBS2 + SBS13) within each iCS. TMB, relative contribution of four mutational signatures, selected differentially mutated genes (>5%) and broad-level CNAs (>20%), and selected genes located within Chr9p21.3 are shown from the top to the bottom panels. Of note, iBS1 harboured more mutations of TP53 (74%; p < 0.001), RB1 (39%; p < 0.001) and KMT2A (18%; p = 0.033) than others, while iBS4 was enriched in mutations of NFE2L2 (16%; p = 0.001; p = 0.005 compared to iBS1 [3%]) and TRANK1 (10%; p = 0.06; p = 0.001 compared to iBS1 [0%]); KIAA0947 (11%; p = 0.005), MED12 (11%; p = 0.005), COL6A6 (13%; p = 0.008), and ARID2 (14%; p = 0.01) were mutated more frequently in iLS2, whereas iLS3 was enriched in mutations of FGFR3 (34%; p < 0.001), STAG2 (22%; p = 0.006), and SPEN (11%; p = 0.05). The proportion of iCS in each alteration is presented in the right bar charts. The distributions of TMB and APOBEC contributions are shown in (C) and (D), respectively. (E) Distribution of fraction genome altered (FGA) and fraction genome gain/loss (FGA/FGG). Bar charts are presented as the mean ± standard error of the mean. Statistical p values were calculated by Kruskal-Wallis rank sum test for multiple comparison Figure 3C). The predicted iBS4 showed poorer outcome than immune-hot iCSs (i.e., iBS1 and iLS2; p < 0.05 at 6 months, p < 0.1 at 12 months) and had a segregated survival curve compared to immune-cold iLS3 ( Figure S6A). Due to the potentially delayed clinical effect of immunotherapy, we also compared the long-term overall survival (OS) rates after three months of treatment, leading to the observation that iBS4 was associated with poorer long-term OS than iBS1 (p = 0.046), iLS2 (p = 0.027) and iLS3 (p = 0.095), suggesting its high malignancy and potential resistance to immune checkpoint inhibitors (ICIs) ( Figure S6B).  6 In this context, we renamed iBS1 'basal-inflamed', iBS4 'basal-noninflamed', iLS2 'luminal-excluded', and iLS3 'luminal-desert' according to the phenotype proportions ( Figure 3D,E). SubMap revealed that only the basal-inflamed subtype showed a high likelihood of responding to ICIs (FDR < 0.01; Figures 3F and S7), which indicated that the current classification may be useful to identify ideal candidates for immunotherapy, especially for basal-like MIBC.
Recently, Hao et al. suggested that Chr4 loss could induce an unfavourable 'immune-cold' phenotype. 7 We demonstrated that only the basal-noninflamed subtype showed significant Chr4p and 4q deletions (FDR < 0.25; Table S7), and Chr4p loss was tightly linked to unfavourable OS (HR = 1.56, 95% CI: 1.001-2.437, p = 0.0496). Chr4 harbours several immune regulatory genes and genes that encode chemokines which are crucial for T cell recruitment. Additionally, Chr4 deletion is genetically linked with immune deficiency syndromes. 8 Integrative analysis of RNA-Seq and CNA indicated that Enrichment analysis revealed that the cell cycle oncogenic pathway was activated in basalinflamed/noninflamed MIBC (p < 0.001; Figures 4A,B  and S8A,B). Activation of the cell cycle pathway induces cell cycle checkpoint regulatory proteins (e.g., ataxia telengiectasia and Rad3-related [ATR] and WEE1) involved in replication stress, which are associated with DNA damage responses that contribute to cisplatin resistance; such an association mirrored a potentially higher sensitivity of responding to ATR and WEE1 inhibitors in basal-inflamed/noninflamed MIBC (Figures 4D,E and S8C,D). A recent study reported that NRF2 enables an immune-cold microenvironment by inducing COX2/PGE2 and inhibiting the DNA-sensing innate immune response 9 ; consistently, the luminal-desert/basalnoninflamed subtypes with activated NRF2 oncogenic pathway had dramatically higher COX2 expression levels than others in both the TCGA and IMvigor210 (Figures 4F and S9A-C) cohorts, suggesting that dysfunction of oncogenic pathways might drive the low immune infiltration of these MIBCs. Unfortunately, it is not feasible to combine the currently unavailable NRF2 inhibitors with ICIs. Nevertheless, preclinical models have demonstrated that COX inhibitors and anti-PD-1 immunotherapy have a synergistic effect 10 ; therefore, targeting NRF2 downstream markers provides a therapeutic opportunity for immune-cold MIBCs. Additionally, the different biological characteristics of the four subtypes may indicate the need for different targeted therapeutic interventions (Table S8).
Given the distinct molecular and prognostic characteristics of basal-inflamed/noninflamed MIBC compared with traditional basal-like classifications, we developed a random forest predictor, which is superior than decision tree, to refine basal-like MIBC (Table S9-10, Figure S10). The final basal classifier contained five immune-related genes, C3AR1, CCL8, FCGR3A, LILRB2 and PDCD1LG2, and the predictor achieved superior performance ( Figure 4G) and showed the capability of prognostic stratification in refining different basal-like classifications ( Figure 4H-L). Strikingly, almost all kinds of epithelial tumours could be refined into immune-hot or immune-cold basal-like phenotypes ( Figure 4M), suggesting a global immunological divergence across basal-like carcinomas.
We acknowledge the limitations of our study, where signals quantified by bulk RNA-Seq and microarray profiles were confounded from mixed cell populations; thus, incorporating these findings with multiplex immunohistochemistry assays to investigate intrinsic tumour cell variation and their crosstalk with the tumour microenvironment, which interferes with the therapeutic response is warranted.
In summary, we identified four MIBC subtypes with distinct landscapes using a multi-omics approach that stratifies prognosis, tumour microenvironment characteristics and distinct sensitivity to frontline therapies. Additionally, we offered the R package "'refineMIBC' as a translational research tool to refine MIBC classification from a singlesample perspective in retrospective or prospective studies.

A C K N O W L E D G E M E N T S
We greatly appreciate the patients and investigators who participated in the corresponding medical project for providing data. This work was supported by the Active Com-

C O N F L I C T O F I N T E R E S T
The authors have no conflict of interest. database) that are involved in DNA maintenance as well as activated cell cycle regulation in DNA damage response and replication stress. Two replication stress (RS) subtypes were identified for the MIBC-TCGA cohort. Subtypes with (D) high replication stress (RS-High) or (E) basal-like MIBC (iBS and iBS4) were inferred to be much more sensitive to both ATR (i.e., VE-822, AZD6739 and VE-821) and WEE1 (i.e., Wee1 inhibitor and MK-1775) inhibitors by applying a ridge regression model using 727 human cancer cell lines. Drug sensitivity was measured as ln(IC 50 ), and the lower the value was, the more sensitive the patient would be to the treatment. (F) Distribution of COX2 expression between immune-hot (i.e., iBS1 and iLS2) and immune-cold (i.e., iLS2 and iBS4) phenotypes of MIBC-TCGA. We developed the R package, "refineMIBC", which is documented and freely available at https://github.com/ xlucpu/refineMIBC. This package implements a 120-gene template that assigns subtype labels according to the multi-omics consensus ensemble of muscle-invasive bladder cancer (MIBC) using nearest template prediction. The consensus ensemble identifies 4 integrative consensus subtypes: basal-inflamed, basal-noninflamed, luminalexcluded, and luminal-desert. This package also deploys a 5-immune-gene classifier to refine each basal-like MIBC as either basal-inflamed or basal-noninflamed by a random forest classifier if basal-like classification has already been identified by other approaches (e.g., CMS, PAM, oneNN, Lund, etc.).