Non-muscle Invasive Bladder Cancer Molecular Subtypes Predict Differential Response to Intravesical Bacillus Calmette-Guérin

Abstract

The recommended treatment for patients with high-risk non-muscle invasive bladder cancer (HR-NMIBC) is tumor resection followed by adjuvant Bacillus Calmette-Guérin (BCG) bladder instillations. However, only 50% of patients benefit from this therapy. If progression to advanced disease occurs, patients must undergo a radical cystectomy with risks of substantial morbidity and poor clinical outcome. Identifying tumors unlikely to respond to BCG can translate into alternative treatments, such as early radical cystectomy, targeted therapies or immunotherapies. Here we conducted molecular profiling of 132 patients with BCG-naïve HR-NMIBC and 44 patients with recurrences after BCG (34 matched), which uncovered three distinct BCG response subtypes (BRS1–3). Patients with BRS3 tumors had reduced recurrence and progression-free survival compared to BRS1/2. BRS3 tumors expressed high EMT-basal markers and had an immunosuppresive profile, which was confirmed with spatial proteomics. Tumors that recurred after BCG were enriched for BRS3. BRS stratification was validated in a second cohort of 151 BCG-naïve patients HR-NMIBC and the molecular subtypes outperformed guideline-recommended risk stratification based on clinicopathological variables. For clinical application, we validated that a commercially approved assay was able to predict BRS3 tumors with an AUC of 0.87. These BCG response subtypes will allow for improved identification of patients with HR-NMIBC at highest risk of progression and might be used to select more appropriate treatments for patients unlikely to respond to BCG.

One Sentence Summary:

Molecular subtypes are predictive of response to intravesical Bacillus Calmette-Guérin immunotherapy in non-muscle invasive bladder cancer.

Introduction

Non-muscle invasive bladder cancer (NMIBC) accounts for 75% of all bladder tumors (1). The recommended treatment for high-risk NMIBC (HR-NMIBC) consists of transurethral resection of the bladder tumor (TURBT) and adjuvant intravesical Bacillus Calmette-Guérin (BCG) instillations (1). Initial treatment response to BCG is usually excellent, however the long-term efficacy is only moderate because HR-NMIBC patients have a 50% risk of developing recurrent disease within 5 years (2). Furthermore, a 20% risk of progression to advanced disease is observed, which is associated with high mortality (1, 3). Patients with HR-NMIBC and recurring or progressive disease will have been exposed to unnecessary BCG toxicity and a delay in radical treatment, resulting in reduced survival (4, 5). Therefore, it is critical to identify patients who are at risk for treatment failure before initiation of BCG therapy (6, 7). These patients might be candidates for early radical cystectomy , which although negatively affecting quality of life, has excellent long-term outcomes (8). Additional treatments are also urgently needed because of an ongoing global BCG shortage (9).

The genetic makeup of tumors can dictate therapeutic response, and it is hypothesized that molecular profiling of HR-NMIBC may reveal mechanisms of response to BCG (10, 11). Several studies from whole-transcriptome analyses of NMIBC have led to clustering-based classification systems and identification of predictive signatures of disease progression (12–16). Dyrskjot et al. prospectively validated a 12-gene expression signature that corresponded to progression in HR-NMIBC (15). However, the progression signature had limited added value compared to the standard clinicopathological risk stratification. A recent study identified five molecular subtypes in 73 HR-NMIBC patients treated with BCG, however none of the subtypes corresponded to progression-free survival (PFS) (17). A major drawback of NMIBC molecular subtyping studies has been the lack of BCG-treated patients, preventing the identification of specific molecular signatures predictive of BCG treatment response. Moreover, studies were hampered by a lack of detailed information on BCG treatment and a lack of patients who experienced progressive disease. In addition, a user-friendly tool for subtype identification has not yet been developed. Because of these caveats, the translational impact of molecular subtypes thus far remains limited.

In this study, we leveraged two large and fully annotated cohorts of patients with HR-NMBC treated with BCG and with clinical follow-up to define three molecular subtypes that associated with clinical outcomes. The molecular subtypes have unique features that correlate with differential response to BCG. We identified an aggressive epithelial-to-mesenchymal transition (EMT) basal immunosuppressive subtype [BCG response subtype 3 (BRS3)] and we showed that BRS3 was the dominant subtype in post-BCG recurrences. We demonstrated that subtyping improved guideline-recommended risk stratification and that a commercially available pathway-assay was able to accurately identify BRS3 tumors for clinical utility. Lastly, we identified druggable genes and pathways associated with resistance to immunotherapy in post-BCG recurrences as candidate targets for bladder-sparing therapies in HR-NMIBC patients.

Results

HR-NMIBC molecular subtypes with divergent clinical outcome after BCG.

To identify gene signatures predictive of clinical outcome, we performed whole-transcriptome sequencing of two HR-NMIBC patient cohorts: discovery Cohort A, n=132 with pre-BCG and n=44 with post-BCG tumors (34 matched tumors) and validation Cohort B n=151 with pre-BCG tumors. A synopsis of the study design is shown in Fig. 1A. Detailed information on resources (Table S1) and patient inclusion, can be found in the Materials and Methods section. Patient and tumor-specific information and follow-up data for both cohorts (Table S2) and an overview of BCG responders vs non-responders is provided (Table S3). No differences in PFS were observed between Cohort A and B (fig. S1A). A detailed comparison of Cohort A vs. B is in Table S4. n=38 TaG3 tumors were included in Cohort B, otherwise no differences were found between cohorts.

Fig. 1. Study design and gene signatures of patients with high-risk NMIBC according to BCG response subtypes (BRS).

A: RNA-sequencing was performed on Cohort A: n=132 pre-BCG, T1G3 NMIBC tumors (n=64 BCG-responders [R] vs. n=68 BCG non-responders [NR]). From the BCG NR, n=44 post-BCG tumors were also sequenced (n=34 matched pre- and post BCG samples + n=10 non-matched). Paired analysis showed enrichment of BRS3 in post-BCG tumors and identified candidate druggable genes. Cohort B consisted of n=151 pre-BCG, high-risk NMIBC tumors (n=88 BCG-responders vs. n=63 BCG non-responders). For both cohorts, PFS (Kaplan-Meier) is stratified according to BRSs. Last, a qPCR pathway assay had an AUC of 0.87 in identifying BRS3 patients. B: Heatmap of gene signatures and annotation of Cohort A grouped according to BRS. From top to bottom: i) Progression to MIBC; ii) BCG response; iii) carcinoma in situ signature (94); iv) 12-gene progression signature (17); v) UROMOL21 NMIBC subtypes (13); vi) T1BC subtypes (17); vii) TCGA MIBC subtypes (18); viii) Consensus MIBC subtypes (17); ix) UNC MIBC subtypes (95); x) MDA MIBC subtypes (96); xi) Lund BC subtypes (14); xii) Clinicopathological parameters associated to pre-BCG tumors; xiii) BRS signatures based on mean gene expression of selected genes (details in methods). Abbreviations: BCG = Bacillus Calmette-Guérin; (N)MIBC = (non-)muscle-invasive bladder cancer.

To determine the presence and robustness of molecular subtypes, we split Cohort A at a 3:1 ratio into a training and testing set. We applied consensus clustering of the top 2000 protein coding genes with the most varying expression in the training set to identify three subtypes with differential risk of progression (fig. S1, B and C). To prevent clustering bias caused by immune populations based on gene expression analyses, all clustering in this study was performed with removal of genes predictive of immune cell populations. Using this strategy, weight of tumor-specific markers during clustering was increased. A nearest shrunken centroid classifier was then trained and used to predict subtypes in the testing set. Based on estimated PFS and overlapping pathway activity, the presence of three molecular subtypes was confirmed in the testing set (fig. S1, D and E). Subtypes were called BCG Response Subtypes (BRS)1, BRS2, and BRS3. After cross validation to estimate the robustness of the BRS in the training and testing set, we retrained the model on the full dataset of Cohort A (fig. S1F), and set out to investigate the clinical and molecular features of the BRS. Patients with BRS3 tumors had a worse PFS (Fig. 1A) than BRS1/2, with a 2-year PFS of 61% for BRS3 versus 78% for BRS2 and 83% for BRS1 (Kaplan-Meier [KM] estimate). Differential pathways associated with the BRS in Cohort A are illustrated in Fig. 1B; the top 30 non-synonymous single nucleotide variations (SNVs) are in fig. S2 and Table S5. BRS3 patients showed increased EMT pathway activity and were enriched for mutations associated with the extracellular matrix (ECM) (for example, adipocyte enhancer-binding protein 1, AEBP1).

Next, we validated the BRSs in terms of clinical outcome and molecular features. We performed whole-transcriptome sequencing on a second HR-NMIBC patient cohort (Cohort B, n=151 pre-BCG tumors; see Table S2 for patient information) and used the BRS classifier trained on Cohort A to predict molecular subtypes in Cohort B. Fig. S3 illustrates highly similar pathway activity per BRS in patients from Cohort B vs Cohort A. Survival analysis showed that patients with BRS3 tumors had the worst PFS (2-year PFS 94% for BRS1, 87% for BRS2, and 67% for BRS3, Fig. 1A). To explore how gene expression of the BRSs would overlap with previously published NMIBC and MIBC subtypes, we applied the BRS classifier to UROMOL21 and The Cancer Genome Atlas (TCGA) cohorts (13, 18). In UROMOL21, BRS3-predicted tumors overlapped with the more aggressive Class 2A/B NMIBC subtype, whereas the less aggressive Class 1/3 tumors were more likely to be BRS1/2 (p<0.01, fig. S4A). Although UROMOL21 contained patients that were not primary HR-NMIBC or were treated with ≥5 instillations of BCG, we did observe overlapping gene signatures and a similar BRS stratification, which was specifically evident for stage T1 HR-NMIBC (fig. S4B). In the TCGA cohort, BRS1/2 tumors were predominantly luminal papillary (p<0.001). BRS3 tumors were more frequently luminal infiltrated and basal/squamous (p<0.001) (fig. S4C), and these tumors are known to be more aggressive. As a last step, we validated pathway activity in a second independent dataset with BCG-treated, HR-NMIBC patients (fig. S4D). BRS2 (MYC high) seemed to associate with T1-MYC, whereas BRS3 (basal high) did not contain any T1-Lum tumors. No association with high-grade recurrence-free survival (HG-RFS, p=0.48) or PFS (p=0.75) was observed, but only eight patients progressed (17).

Last, we investigated whether clinicopathological parameters could explain differences between subtypes in the combined dataset (Table 1), or in Cohort A or B separately (Table S6–7). We noticed a correlation between tumors with extensive tumor invasion into the lamina propria (T1e disease) and patients with BRS3 tumors, which is in line with our previous reporting on the increased risk of BCG treatment failure and progression in patients with T1e tumors (3). Due to the probability of understaging in 10% of T1 HR-NMIBC and thus BCG treatment failure, guidelines recommend that patients undergo a re-TURBT before initiation of BCG treatment (19). Although a re-TURBT itself was not associated with PFS (HR = 1.0, p=0.97), the proportion of patients who had undergone a re-TURBT was higher in BRS1 patients, who had a favorable outcome. We recognize that a difference in re-TURBTs might lead to a decreased risk of treatment failure. Therefore, we repeated clustering in only patients who received a re-TURBT (91/132, 69%). Neither risk stratification nor expression patterns differed between patients with and without re-TURBT (fig. S5), thus we concluded that a re-TURBT did not affect the prognostic capability of the BRS. In summary, we discovered three molecular subtypes (BRS) of HR-NMIBC and showed their clinical relevance and consistency in pathway activity across multiple patient cohorts.

Table 1.

Baseline patient characteristics and clinical follow-up of all n=283 patients with BCG-naive high-risk non-muscle invasive bladder cancer treated with BCG and stratified according to subtype.

Patient characteristicsa		All	BRS1	BRS2	BRS3	p
		n (%)	n (%)	n (%)	n (%)
Sex	male	56 (80)	70 (25)	90 (33)	67 (24)	0.895
	female	227 (20)	19 (7)	22 (8)	19 (5)
Age	Median (IQR)	70 (62–77)	70 (62–77)	69 (62–76)	71 (62–77)	0.826
Smoking	No	87 (31)	23 (8)	36 (13)	28 (10)	0.432
	Yes/stopped	173 (61)	61 (22)	64 (23)	64 (17)
	Missing	23 (8)	5 (2)	12 (4)	6 (2)
re-TURBT	No	93 (32)	23 (8)	37 (13)	33 (12)	0.138
	Yes	190 (67)	66 (23)	75 (27)	49 (17)
Focality	Unifocal	137 (48)	43 (15)	55 (19)	39 (14)	0.772
	Multifocal	142 (50)	46 (16)	55 (19)	41 (14)
	Missing	4 (1)	0 (0)	2 (1)	2 (1)
Size b	< 3cm	43 (15)	16 (6)	17 (6)	13 (4)	NA
	≥ 3cm	46 (16)	18 (6)	18 (6)	7 (2)
	Missing	194 (69)	55 (22)	62 (24)	13 (22)
Concomitant CIS	No	221 (78)	65 (23)	93 (33)	63 (22)	0.216
	Yes	62 (22)	24 (8)	19 (7)	19 (7)
LVI c	No	11 (16)	76 (85)	86 (77)	72 (88)	0.553
	Yes	234 (84)	4 (4)	4 (4)	3 (4)
	NA (TaG3)	38 (13)	9 (10)	22 (20)	7 (9)
Variant histology	No	237 (83)	71 (25)	104 (37)	62 (22)	0.002
	Yes	46 (16)	18 (6)	8 (3)	20 (7)
T1 substage c,d	T1 micro	63 (22)	23 (26)	33 (29)	7 (9)	<0.001
	T1 extensive	182 (64)	57 (64)	57 (51)	68 (83)
	NA (TaG3)	38 (13)	9 (10)	22 (20)	7 (9)
EAU Risk group e	High-risk	124 (44)	37 (13)	56 (19)	31 (11)	0.213
	Very high-risk	159 (56)	52 (18)	56 (19)	51 (18)
BCG failure f	No	156 (55)	59 (21)	68 (24)	29 (10)	<0.001
	Yes	127 (45)	30 (11)	44 (15)	53 (19)
HG-RFS	1 year (CI)	73 (67–78)	77 (68–86)	76 (68–84)	64 (55–76)	<0.001
Progression	No	187 (66)	68 (24)	80 (28)	39 (14)	<0.001
	Yes	96 (34)	21 (7)	32 (11)	43 (15)
PFS	2 year (CI)	80 (75–84)	88 (82–96)	83 (76–90)	64 (54–76)	<0.001
	5 year (CI)	70 (65–76)	76 (64–86)	76 (69–85)	55 (46–67)
Follow-up (months)	Median (IQR)	73 (51–104)	76 (56–103)	75 (55–104)	68 (29–100)	0.192

BCG = Bacillus Calmette-Guérin; BRS = BCG Response Subtype; CI = confidence interval; CIS = carcinoma in situ; EAU = European Association of Urology; HG-RFS = high-grade recurrence-free survival; IQR = Interquartile range; NA = not applicable; re-TURBT = repeat transurethral resection of bladder tumor before BCG-induction.

^aPercentages may not add up to 100% due to rounding;

^bp value not applicable due to missing variables;

^cp values assessed only in T1 disease (removal of NA values);

^dT1 substaging was performed (3).

^eSee methods for characteristics of very high-risk disease;

^fBCG failure as specified by major urology guidelines, which includes patients with tumors who developed muscle-invasive recurrences, persistent T1HG after BCG-induction and HG recurrences after adequate BCG therapy.

BRS3 is an aggressive molecular subtype of HR-NMIBC.

BRS3 delineated a group of patients with a high risk of progression. Unraveling the underlying molecular characteristics that drive BRS3 could explain differences in clinical outcome and possibly lead to the identification of candidate druggable targets for targeted treatments. Because of the overlapping gene expression patterns and PFS, we combined Cohort A and B (n=283) for further analysis. Differential gene expression (DGE) analysis showed that BRS3 tumors had overexpression of basal and EMT markers (cluster of differentiation [CD]44, keratin 5 and 6 [KRT5/6], discoidin domain receptor tyrosine kinase 2 [DDR2], vimentin [VIM], snail family transcriptional repressor [SNAI]1, zinc finger e-box binding homeobox 1/2 [ZEB1/2], and twist-related protein [TWIST]1). Overexpression of EMT markers could be the consequence of deep tumor invasion into the stromal area, which may explain the correlation of BRS3 with T1 extensive disease. Furthermore, BRS3 tumors were characterized by overexpression of immune suppressive genes, including immune-checkpoint genes (programmed death (ligand) 1) [PD-1/PD-L1] and cytotoxic T-lymphocyte-associated protein 4 [CTLA-4]), T regulatory cell markers (T regs; CD4/CD25/forkhead box P3 [FOXP3]), myeloid derived suppressor cell markers (MDSC; interleukin 6/10 [IL6/IL10], CD11b, CD14), colony stimulating factors (CSFs) and chemokines (CXCLs and CXCRs), which have been implicated in BCG treatment failure (20). A full list of differentially expressed genes is reported in Table S8.

We then analyzed pathways associated with distinct subtypes. Individual genes upon which the BRS pathway activity is based are listed in fig. S6A. In BRS3 tumors, gene set enrichment analysis (GSEA) identified pathways known to be associated with cancer progression, such as EMT, Notch, mitogen-activated protein kinase (MAPK), and angiogenesis (Fig. 2, A and B). Moreover, immune-related pathways were strongly enriched, for instance complement, IL6-janus kinase (JAK)-signal transducer and activator of transcription (STAT) 3 pathway and IL2-STAT5 (Fig. 2, A and B). Complete results are listed in Table S9.

Fig. 2. Gene set enrichment analysis and regulon analysis grouped according to theBRS.

A: Heatmap of all 50 GSEA hallmarks in n=283 pre-BCG, HR-NMIBC patients. Clustering on gene signatures (rows) indicates the existence of different molecular subtypes in HR-NMIBC. B: Boxplots of selected gene signatures in n=283 pre-BCG HR-NMIBC patients grouped according to BRS; p-values are Kruskal-Wallis tests. C: Heatmap of the top 200 most varying regulons using single-sample VIPER analysis on n=283 pre-BCG tumors grouped by BRS (columns). Hierarchical clustering between samples confirms three distinct molecular subtypes; key regulators for which small molecule inhibitors exist are highlighted. Abbreviations: BCG = Bacillus Calmette-Guérin; BRS = BCG response subtypes; GSEA = gene set enrichment analysis; HR-NMIBC = high-risk non-muscle invasive bladder cancer; VIPER = Virtual Inference of Protein Activity by Enriched Regulon Analysis.

Regulon analysis using VIPER was performed to computationally infer protein activity based on downstream target gene expression patterns in each subtype (21). Regulons comprising the top 150 most varying transcriptional regulators for which small molecular inhibitors are available, as provided by the druggable genome, are shown for the three HR-NMIBC subtypes (Fig. 2C) (22). These genes may prove to be attractive candidates for subtype-specific drug targeting. Several regulatory genes have previously been identified as important genes driving bladder cancer (BC) (18). These genes have been grouped according to BRS in fig. S6B We observed low retinoblastoma 1 (RB1) and tumor protein (TP)63 in BRS3 tumors, which is often seen in aggressive BC (23). We confirmed the activity of EMT-associated regulators (ZEB2, TWIST2, and SNAI3), and enrichment of regulators related to Myeloid derived suppressor cells (MDSCs) and B cells (CD14, CD86), T regs (Basic Leucine Zipper ATF-Like Transcription Factor 3 [BATF3], IL10, IL16, FOXP3) and T cell polarization [STAT1/STAT4 and T-Box Transcription Factor 21 (TBX21)] (Table S9). Taken together, BRS3 tumors were enriched for regulators associated with EMT, a basal phenotype, and immune checkpoint proteins, and these results associated biological features of BRS3 tumors with the corresponding poor clinical outcome.

BRS1 and BRS2 HR-NMIBC have luminal characteristics.

We then described relevant molecular features associated with BRS1/2. DGE analysis showed that BRS1/2 tumors overexpressed the luminal marker peroxisome proliferator activated receptor gamma (PPARG) (fig. S6A, Table S8). A typical luminal marker is fibroblast growth factor receptor 3 (FGFR3), but FGFR3 expression was high in BRS2 tumors only. Pathway analysis revealed BRS1 tumors had upregulated cell cycle and metabolic processes involved in mycobacterial (BCG) processing, such as enrichment of autophagy, ubiquitination and proteasome degradation and protein secretion (Fig. 2B, Table S9). These findings provided mechanistic insight into the more favorable outcome associated with BRS1 patients after BCG treatment. GSEA indicated that BRS2 was the least enriched for immune-related pathways (Fig. 2A) (24). MYC pathway activity (Fig. 2B), the absence of a carcinoma in situ (CIS) profile (Fig. 1C and fig. S3), enriched luminal signatures (fig. S6C) and a high proportion of pure urothelial cell carcinoma (Table 1) were associated with BRS2. Hence, BRS2 tumors align with features seen in papillary BC. Subtype-specific and potentially druggable regulatory proteins are listed in Fig. 2C. The expression of luminal regulators, forkhead box M1 (FOXM1), forkhead box A1 (FOXA1), GATA Binding Protein 3 (GATA3) and Erb-B2 Receptor Tyrosine Kinase 3 (ERBB3) was high in both BRS1 and BRS2 tumors, but not in BRS3 tumors. Consistent with luminal papillary MIBC, BRS2 tumors retained sonic hedgehog (SHH) expression (fig. S6B) (18). Taken together, characterization of BRS1 and BRS2 tumors shows these are molecularly distinct, but both share luminal features, which may explain the better outcomes after BCG.

The tumor microenvironment of BRS3 displays immune suppressive features.

BCG immunotherapy is dependent on interactions with the tumor microenvironment (TME). Because findings pointed towards high immunological activity in BRS3 with a poor outcome after BCG treatment, we analysed the TME using transcriptome deconvolution and spatial proteomics. First, we assessed the overall immune infiltration using the ESTIMATE algorithm and confirmed that BRS3 tumors had the highest immune and lowest tumor purity scores as compared to BRS1/2 tumors (Fig. 3A) (25). To gain further insight into immune cell composition of BRSs, we applied the quanTIseq immune deconvolution method, which allowed for between tumor and between immune cell population comparisons (26). BRS3 showed the highest infiltration of B cells, tumor-associated macrophages (TAMs) M1/M2, CD8+ T cells and T regs (Fig. 3B). Findings were verified using CIBERSORT and EPIC algorithms (Table S10) (27, 28).

Fig 3. Immune deconvolution of pre-BCG tumors from Cohort A+B and spatial proteomics of the tumor microenvironment grouped by BRS from Cohort A.

A: Boxplots predicting tumor purity and immune infiltration in n=283 pre-BCG tumors grouped by BRS (25); p-values are Kruskal-Wallis tests. B: Immune cell deconvolution from RNA-sequencing: a single column is the sum of ten immune cell subpopulations and non-characterized cells grouped by BRS; uncharacterized cells are a mixture of malignant and normal cells; p-values are Wilcoxon tests for BRS1/2 vs BRS3; **p.adj<10⁻³; ***p.adj<10⁻⁴. C: Immunohistochemistry of CD8⁺ cytotoxic T cell infiltration in n=36 BRS1, n=48 BRS2 tumors and n=35 BRS3 tumors from Cohort A; p-values are Wilcoxon tests for BRS1/2 vs BRS3. D: Selection of regions for spatial proteomics; on average 20 regions of pre-BCG tumors surrounding the macrodissected areas used for RNA-sequencing were selected. Mean intratumoral protein expression was used for analyses. E: Boxplots of multiplex immunofluorescence results for n=38 BRS1, n=50 BRS2 tumors and n=38 BRS3 tumors; p-values are Wilcoxon tests for BRS1/2 vs BRS3 or Kruskal-Wallis tests (overall). F: Boxplot for immunofluorescence results for only BRS3 patients. BRS3 tumors with progression had higher T regs and macrophages than tumors that did not progress; p-value is a Wilcoxon tests for BRS3-R vs BRS3-P. Abbreviations: BCG = Bacillus Calmette-Guérin; BRS = BCG response subtypes; BRS3-R = BRS3 responder; BRS-P = BRS3 progressor.

Previous in vivo studies demonstrated a critical role for CD8⁺ T cells in BCG-mediated anti-tumor immunity (29, 30). Therefore, high infiltration of CD8⁺ T cells in treatment-naive BRS3 tumors seemed a paradoxical finding, as most patients with BRS3 did not respond to BCG treatment. Thus, we tested if BRS3 tumors showed high CD8⁺ T cells. Immunohistochemistry was successfully performed on 119 tumors from Cohort A with whole slides available for staining with a clinically approved antibody for CD8⁺ T cells. Consistent with RNA expression data, we found that the BRS3 vs BRS1/2 subtype was significantly (p<0.001) more infiltrated with CD8+ T cells (Fig. 3C). We also investigated protein differences between progressors (n=19) and non-progressors within BRS3 (n=16), but found no significant (p=0.716) differences for CD8⁺ T cell infiltration (Wilcoxon test). The luminal FGFR3 expressing BRS2 subtype had the lowest infiltration with CD8⁺ T cells (Fig. 3C).

From Cohort A, we successfully performed multiplex proteomics for CD4⁺, T regs, B cells and macrophages (M1/M2), EMT (VIM) and cytokeratin (CK). Areas surrounding the macrodissected tumor punches used for RNA isolation were selected (Fig. 3D). We confirmed an increased number of tumor-infiltrating immune cells (CD4⁺ [p=0.021], T regs, B cells and macrophages) and higher intratumoral VIM protein expression in BRS3 patients as compared to BRS1/2 using automated quantification on tumoral vs stromal tissue (Fig. 3E). We then tested if immune cells could serve as markers for a poor prognosis in all patients that developed progressive disease vs all patients that did not have progressive disease. None of the markers showed differences between progressors and non-progressors, which further highlights the importance of subtyping in these patients. Within BRS3 tumors, patients with progressive disease had more macrophages and T regs, suggesting that this might serve as a marker for treatment failure in a subset of BRS3 tumors (Fig. 3F).

Taken together, a trend towards higher immune cell infiltration was seen in RNA-expression, and these results were confirmed at proteomics. BRS3 tumors were more infiltrated with T regs, macrophages and B cells, which are associated with immune suppression, and these findings have previously been linked to a poor clinical outcome after BCG (31, 32).

BRS improves current clinical risk stratification of HR-NMIBC.

We investigated how the BRSs impact BCG response and clinical outcome in Cohort A+B. According to international recommendations, BCG response is defined as being free of high-grade (HG) disease after 6 months of adequate BCG exposure (33). Using this endpoint, a differential response to BCG treatment was observed between subtypes (BRS1 85%, BRS2 82%, BRS3 68%; p=0.017, KM method). In addition to PFS (Fig. 4A), patients with BRS3 tumors had a poorer HG recurrence-free survival (Fig. 4B). In multivariable analyses, BRS3 vs. BRS1/2 was an independent predictor of PFS (HR 2.7, p<0.001, Fig. 4C).

Fig. 4. Kaplan-Meier estimates of survival according to the BRS in pre-BCG HR-NMIBC patients (Cohort A+B) and assay performance to predict BRS3 in Cohort A.

A: Progression-free survival (PFS) stratified according to BRS. B: Recurrence-free survival (RFS) for high-grade tumor recurrences stratified according to BRS. C: Forest plot of multivariable Cox regression analysis; BRS3 vs BRS1 + BRS2. D: PFS stratified for the current EAU guideline recommendations (high-risk [HR] patients vs subgroup of very high-risk [VHR] patients). E: PFS stratified for current EAU risk-stratification and the BRSs. F: Subtype prediction for BRS3 vs BRS1/2 based on the assay results in Cohort A. Figure depicts the ROC with the area under the curve based on a logistic regression model, with corresponding test sensitivity and specificity reported at a threshold of 0.385. Abbreviations: AUC = area under the curve; BCG = Bacillus Calmette-Guérin; BRS = BCG response subtypes; EAU = European Association of Urology; (V)HR-NMIBC = (very) high-risk non-muscle invasive bladder cancer; Progression-free survival (PFS).

Next, we analyzed the added value of the BRS classifier to the clinicopathological risk stratification of HR-NMIBC recommended by the EAU guidelines (1). This risk stratification is based on the presence of stage T1, grade 3 or high-grade, concurrent CIS, multifocality, large tumors (≥3cm), aggressive forms of variant histology and lymphovascular invasion which are associated with a high risk (HR) or a very high risk of progression (VHR) (1). Because VHR patients have a higher risk of progression than HR patients, guidelines recommend consideration of an early radical cystectomy (instead of BCG). In our study, VHR patients had a 2-year PFS of 75% vs 85% for HR patients (p=0.01, Fig. 4D). Combining the BRS with the EAU HR-NMIBC risk model improved risk stratification by identification of a subgroup of patients with the highest risk of progression (VHR3) (Fig. 4E). VHR3 patients had a 2-year PFS of only 54%. Importantly, VHR1 vs VHR2 patients, normally also considered for early radical cystectomy, now had a similar 2-year PFS as HR patients. We then developed a preliminary nomogram for the research community to estimate the risk of progression, which should be validated prospectively (fig. S7). Lastly, we assessed whether previously reported signatures were able to predict response to BCG treatment in our total Cohort A+B. A significant (p<0.001) association was seen in BRS1 (lowest), BRS2 (intermediate) to BRS3 (highest) for the presence of the 12-gene progression signature (ANOVA test) (15, 17). Though a trend was seen for the 12 gene signature, none of the published MIBC or NMIBC subtypes could identify a clinically relevant subset of patients with the highest risk of progression (fig. S8, A to H).

BRS3 patients can be identified with a commercially available test.

To investigate the clinical utility of the BRSs, we assessed whether a commercially approved assay (OncoSignal^®) was able to identify BRS3 versus BRS1/2 tumors. The assay is qPCR-based and measures seven signal transduction pathways (details in Materials and methods) (34). The same RNA used for transcriptomic sequencing of Cohort A was selected for OncoSignal^® analysis. The assay was able to distinguish BRS3 from BRS1/2 tumors with BRS3 having high signal transduction of MAPK, estrogen receptor (ER), and Notch pathways (fig. S9). Slight differences between BRS3 and BRS1/2 were found based on hedgehog, Phosphoinositide 3-kinases (PI3K) and transforming growth factor beta (TGF-β) pathways. From a clinical perspective, it is important to distinguish aggressive BRS3 tumors from BRS1/2 tumors, because BRS3 tumors might benefit from a different therapeutic strategy than standard-of-care. Thus, based on the pathway activity scores, we developed a logistic regression model to predict if a tumor was BRS3. The model resulted in an AUC of 0.87, with optimized thresholds resulting in a sensitivity of 78% and specificity of 86% (Fig. 4F). Although findings need validation in an independent cohort, we have established for the first time that a commercially available assay can be used to identify a clinically relevant aggressive type of HR-NMIBC.

Post-BCG tumor recurrences are enriched for BRS3.

To gain further insight into molecular changes that occur after BCG treatment failure, we applied the BRS classifier training and prediction pipeline to now include the 44 post-BCG tumors (34 pre/post-treatment matched tumors). BRS3-predicted tumors were enriched in post-BCG 28/44 (64%) recurrences vs pre-BCG tumors 82/283 (28%) (Fig. 5A). From the 34 matched tumors, 26/34 patients had BRS1/2 tumors pre-BCG of which 14/26 tumors switched to BRS3 post-BCG.

Fig. 5. Pre- vs post-BCG transcriptomic and spatial proteomic comparisons.

A: Frequency of BRS in pre-BGC vs post-BCG tumors in Cohort A+B. Alluvial plot details 34 patients with matched pre- and post-BCG tumor samples. B: Volcano plot of differentially expressed mRNAs in matched post-BCG and pre-BCG tumors from n=34 patients. C: Top regulons with small molecule inhibitors that are up- or down-regulated in pre-BCG tumors and post-BCG tumors depicted. In red: regulators enriched in post-BCG recurrences. In blue: regulators enriched in pre-BCG tumors. VIPER uses a Student’s t-test with 2-tailed p values. D: Selected gene set enrichment hallmarks comparing n=34 paired pre- and post-BCG tumors samples. Dashed lines represent pathway activity scores between paired samples of a single patient; p-values are paired Wilcoxon tests. E: Immune cell deconvolution based on RNA-sequencing: a single column is the sum of ten different immune cell populations and non-characterized cells found in the tumor microenvironment grouped by pre-BCG tumors or post-BCG recurrences; uncharacterized cells are a mixture of malignant and normal cells; p-values are Wilcoxon tests; *p.adj<0.05. F: Heatmap with spatial proteomics grouped by subtype and sorted by z-score. G: Proteomic results of n=30 post-BCG recurrences grouped by post-BCG BRS; p-values are Wilcoxon tests. Abbreviations: BCG = Bacillus Calmette-Guérin; BRS = BCG response subtypes.

Overexpressed genes in post-BCG samples can indicate mechanisms involved in BCG treatment failure and might therefore provide candidate target genes for potential bladder-sparing treatments. Paired DEG analysis revealed that in post-BCG tumors, Cytochrome P450 superfamily CYP1A1 and CYP1B1 were highly overexpressed, both key genes involved in phase-1 drug metabolism (modification), as well as genes essential for the attachment of BCG to urothelial cells (fibronectin 1 [FN1], Fibronectin Type III Domain Containing 1 [FNDC1], Integrin Subunit Alpha 5 [ITGA5]). These results in tumors that do not respond to treatment, indicated that overexpression could be a consequence of ineffective mycobacterial antigen exposure to the urothelium. EMT-related genes (ZEBs, TGFβ1/3), fibroblast and hepatocyte growth factors (FGF2, FGF7, HGF), and genes associated with basal-like disease and cancer progression (CD44, KRT6/16, galectin 7 [LGALS7]) were also overexpressed (Fig. 5B) (35–37). Collagen receptor DDR2 and ECM remodeling genes (matrix metalloproteinases [MMPs], a disintegrin and metalloproteinase with thrombospondin motifs [ADAMTSs], versican [VCAN], fibroblast activastion protein alpha [FAP]) were also highly overexpressed after treatment. Although these genes are involved in wound healing and tumor scarring, extensive preclinical work has shown that they promote bladder cancer metastases and are predictive of a poor outcome (38–40). Additionally, post-BCG tumors showed increased expression of the WNT antagonists secreted frizzled-related proteins [SFRPs], which is consistent with the WNT pathway activity found in pre-BCG BRS3 tumors and might indicate WNT signaling involvement in BCG-resistant tumors (41). Overall, of the multiple genes differentially overexpressed in tumors displaying the pre-BCG BRS3 gene signature, most were overexpressed in tumors that failed treatment and were described in the literature to be associated with cancer progression (Table S11). Many of these genes can be targeted by available small molecules, providing a starting point for the development of targeted treatments in these BCG-resistant patients (22).

Next, we investigated consistently expressed gene fusions, important regulatory networks, pathways, and non-synonymous SNVs to investigate which potentially actionable mechanisms underlie recurring disease post-BCG. We identified gene fusions that were present before and after treatment within the same patient. Although infrequent, gene fusions such as FGFR3 with transforming acidic coiled-coil containing protein 3 [TACC3] were detected in three patients (fig. S10A, Table S12), which may provide some opportunity for anti-FGFR treatment. Transcriptional regulators that were enriched in post-BCG tumors include DDR2, TGFβs, AEBP1, glucocorticoid receptor NR3C1, and FGF7 (Fig. 5C, Table S11). By performing a within-patient pathway analysis, we noted enrichment of EMT, complement, IL6-JAK-STAT3, and angiogenesis in post-BCG recurring tumors (Fig. 5D; Table S11), supporting the hypothesis that recurrences have more aggressive biology. Signatures that contributed to the pre-BCG BRS3 were also the dominant signatures present in recurring tumors (fig. S10B, Table S11). Variant calling identified mutational differences between pre- and post-BCG tumors. Twelve of these genes belonged to ECM organization, which is interesting as ECM signatures were overexpressed in treatment failures as well (fig. S10C; Table S11). To summarize, post-BCG tumors displayed enrichment of pathways that are involved in cancer progression. In addition, several promising actionable targets in post-BCG recurrences were detected.

Lastly, we investigated the TME of post-BCG recurrences. Immune deconvolution on matched samples showed more T regs, macrophages, and B cells in post-BCG tumors (Fig. 5E). Intratumoral multiplex spatial proteomics of n=85 samples showed an increased presence of immune subpopulations in post-BCG BRS3 (Fig. 5F). From the post-BCG recurrences, only BRS3 tumors showed a trend towards increased numbers of regulatory T cells and macrophages (M2) as compared to BRS1/2 tumors (Fig. 5G). Protein analysis pointed towards a phenotype that is associated with immune suppression in tumors that fail BCG treatment.

Altogether, these data support the hypothesis that BRS3 tumors have a clinically and biologically aggressive phenotype. In addition, paired analyses of tumors that did not respond to BCG provide targetable gene candidates for preclinical research into alternative treatments for BCG non-responsive tumors.

Discussion

BCG treatment failure is a major risk factor for progressive disease and death from BC. Thus, much effort has been made to identify patients that do not benefit from treatment and to search for alternatives to radical surgery. Previous RNA-based molecular subtyping studies were not associated with clinical outcome after BCG treatment, possibly because these studies lacked sufficient numbers of patients or were difficult to interpret in the clinic due to the high number of subtypes (14). UROMOL21 did identify three T1HG subtypes, but these subtypes were not able to predict failure of BCG treatment (13). The T1BC classifier was built for recurring disease, while a limited number of included patients developed progression (17). Here, we leveraged a unique cohort of BCG responsive and non-responsive tumors to identify three molecular subtypes with clinical implications. Individuals with BRS1 and BRS2 subtypes and very high-risk disease who normally are candidates for early radical cystectomy have a lower risk of progressive disease, allowing for the possibility to preserve the bladder. In contrast, patients with BRS3 tumors showed a poorer PFS, which suggests that these patients should strongly be considered for early radical cystectomy. To pave the way for the clinical translation of our findings, we used a commercially approved qPCR-based assay and show that this test can accurately identify patients with BRS3 HR-NMIBC. As a next step, we have initiated a follow-up study in a general HR-NMIBC cohort to validate the predictive value of the BRS classifier and OncoSignal (IMPASSE study).

Previous work showed that the presence of TAMs and FOXP3-positive T regs in the TME are associated with worse outcome in BCG-treated BC (42). Kates et al. found that BC progression occurred in BCG-treated immunocompetent rats with an increased presence of FOXP3-positive T regs and rats having a decreased adaptive immunity (31). BRS3 tumors also showed signs of CD8⁺ T cell infiltration, Th1 polarization, and interferon-γ pathway activity, vital for anti-tumor immunity and BCG effectiveness (43, 44). This is a seemingly paradoxical finding as BRS3 tumors have the worst clinical outcome, but strong Th1 predisposition should be counterbalanced to avoid auto-immunity, which may explain the T regs, B cells, and macrophages in BRS3 tumors (45). Because of the high CD8⁺ T cell activity in BRS3 tumors and the non-response to BCG immunotherapy, results are suggestive of T cell exhaustion, which may lead to an ineffective cascade of additional anti-tumor immunity (46, 47).

Recently, low durable response rates were observed to anti-PD-1 in Keynote-057, where pembrolizumab was given to patients with BCG unresponsive disease (48). The investigation of matching pre- and post-BCG treatment samples in this study showed that synergistic therapy regimens in BRS3 tumors with checkpoint inhibitors may harness potential (49). Previous work showed that TGF-β attenuates anti-PD-L1 therapy outcomes by exclusion of cytotoxic T cells from the TME (50). In our study, high expression of TGFβ1/3 was demonstrated in post-BCG tumors (mostly BRS3), suggesting that TGF-β is a key player to overcome BCG resistance and should be investigated as a druggable target in combination with anti-PD-L1. Anti-PD-1 combined with anti-CTLA-4 showed promising response rates to neoadjuvant therapy in MIBC, and our data indicates this strategy has potential as well (51). In addition, in vivo studies showed positive synergistic effects on tumor load and increased CD8⁺ T cell influx when dasatinib (anti-DDR2) was combined with PD-L1 blockade (52). In line with these findings, we demonstrated an increased expression of DDR2 in post-BCG tumors, thus supporting the potential of combination treatment with anti-PD-L1 and anti-DDR2. Finally, recent work from Chamie et al. showed that a combination of BCG with NAI (Nogapendekin alfa-inbakicept, an IL-15 superagonist) had a 55.4% disease-free survival rate at 12 months in n=72 patients with BCG-unresponsive high-grade Ta/T1 NMIBC (53). IL-15 stimulates natural killer cells and CD8⁺ memory T cells, both essential for the effectiveness of BCG. We speculate that BRS3 patients with an immune suppressive profile may benefit from treatment with NAI.

BRS1 tumors had the most favorable PFS and showed molecular similarities to Genomic Subtype 2, previously identified by Hurst et al. in stage Ta/low-grade BC (12). BRS1 tumors had increased expression of genes related to intracellular protein trafficking and antigen presentation, which is vital for mycobacterial processing, and major histocompatibility complex 1 (MHC1)-mediated cytotoxic T cell killing of cancer cells (54–58). Enhanced autophagy has been found to be important in the processing of BCG and improved antigen presentation in vivo (59, 60). SNPs in autophagy related 5 and 2B (ATG5/ATG2B) negatively influence trained immunity after restimulation with BCG and inhibition of autophagy directly attenuated bacterial processing and antigen presentation (61). These results led us to hypothesize that enhanced autophagy during BCG treatment might improve the immune response.

BRS2 tumors were luminal, with high FGFR3 expression, activation of MYC, but substantial within-subtype heterogeneity. Luminal heterogeneity is a consistent finding in BC and complicates the clinical implementation of subtyping (62). The relationship of MYC to FGFR3 activity was elucidated in vitro, in which Mahe et al. discovered that FGFR3 expression led to downstream MYC accumulation and increased FGFR3 activity in a positive feedback loop (63). Given the high prevalence of FGFR alterations in NMIBC, high FGFR3 expression in BRS2 tumors and recent effective anti-FGFR treatment in metastatic BC, therapeutic inhibition of FGFR in BRS2 tumors is an option that has potential therapeutic translation (64).

This study is not without limitations. First, tumors from BCG non-responders were overrepresented as compared to a real-world situation (65). For the discovery cohort, we specifically selected true BCG responders and BCG non-responders, because the objective was to find molecular characteristics associated with response to BCG. The inclusion of a similar number of T1G3 responders and non-responders may also explain why progression rates are on the high-end. Hence, the IMPASSE study has been setup to determine the clinical applicability and validity of the BRS in a general HR-NMIBC cohort. Second, even after taking all necessary precautions, understaging in a small percentage of patients may occur, which might also explain why BRS3 tumors are associated with more extensive stromal infiltration. None of the tumors showed muscle invasion, yet assessment of clinicopathological parameters by a uropathologist is currently the only available tool for urologists to decide which patients should receive BCG. Third, there could be a treatment-bias, because approximately half of the patients from Cohort B were also treated in the hospitals from Cohort A. Fourth, the OncoSignal assay was run on the same RNA as we performed RNA-sequencing on. Although comparison between these assays is a necessary first step, independent validation must be done. Last, future studies should address the issue of RNA-based spatial heterogeneity, because false positive and false negative subtypes might be a result of sample collection bias or inherent BC subtype heterogeneity (66, 67). Assessment of histological and molecular heterogeneity will likely help to explain BCG resistance due to selection pressure, which may have prognostic and therapeutic implications (68, 69). We found that the proportion of BSR3 increased in post-BCG recurrences. Therefore, we speculate that progression from BRS1/2 to BRS3 after treatment failure is due to tumor plasticity and clonal expansion of aggressive tumor cells.

In conclusion, our findings demonstrated the prognostic relevance of molecular subtypes in HR-NMIBC patients. We translated our findings for potential clinical use via a consumer-ready test. In addition, the data presented provided unique pre- and post-treatment evidence for development of targeted therapies targeting tumors that fail BCG treatment. Identification of BRS3 tumors may be a critical step for implementation of more aggressive therapeutic regimens such as an early radical cystectomy or recruitment into clinical trials.

Materials and Methods

Study design

The aim of this study was to identify molecular differences that may explain differences in response to BCG treatment in patients with BCG-naïve and BCG-resistant tumors using whole-transcriptome sequencing and (spatial) proteomics. To this end, tumor samples from primary HR-NMIBC patients who had received ≥5/6 BCG-induction instillations between 2000–2018 at four different Dutch hospitals (Erasmus University Medical Center Rotterdam, Franciscus Gasthuis and Vlietland Rotterdam, Amphia Breda and Reinier de Graaf Gasthuis, Delft) and one Norwegian hospital (Stavanger University Hospital) were compiled. IRB approval was obtained from the Erasmus MC Medical Research Committee (MEC-2018–1097). For all hospitals, the regimen of BCG instillations was according to the Southwest Oncology Group (SWOG) protocol; clinical follow-up was according to the European Association of Urology (EAU) guidelines (1). Formalin-fixed paraffin-embedded (FFPE) material, including TURBTs, re-TURBTs, tumor recurrences, random biopsies, radical cystectomies (RC), pelvic lymph node dissections, and distant metastases were reviewed by an expert uropathologist (R.F.H.) in accordance with WHO standards for classification of the urinary system and was previously published (3, 70). Patients were classified as high-risk or as very high-risk subgroup according to the EAU guidelines on NMIBC. Very high-risk criteria included: multiple or large tumors (≥3cm), concomitant CIS, lymphovascular invasion (LVI), or certain aggressive forms of variant histology (micropapillary, sarcamatoid, nested and neuroendocrine) (1).

Cohort A consisted of a similar number of BCG responders (n=63) and BCG non-responders (n=69). No prior sample size calculation was performed. Investigators were aware of all patient characteristics, no blinding techniques were used. Response to BCG was defined as the absence of a high-grade recurrence after at least 5 out of to 6 BCG induction instillations and ≥9 BCG maintenance instillations (1-year schedule with at least 3 cycles of BCG maintenance). Cohort A had a median of 18 instillations. Non-response to BCG was defined as the development of one of the following: i) biopsy-proven muscle-invasive bladder cancer, ii) persistent T1HG NMIBC after BCG induction, iii) high-grade NMIBC after adequate BCG therapy, which was defined as 5 out of 6 BCG induction instillations plus 2 out of 3 BCG maintenance instillations. Only patients were included in whom pathology review showed invasion of Grade 3 / high-grade (HG) urothelial cancer cells into the lamina propria (T1 G3/HG NMIBC) with the absence of cancer cells in the detrusor muscle of the primary TURBT or re-TURBT. Pre-BCG tumor samples were selected for whole transcriptome analysis with a matching number of BCG responders and non-responders, and from the latter group, we additionally included samples that were the first HG tumor recurrences during BCG therapy, which could be TaT1HG or T2. Ten post-BCG, HG recurrences were included for which the pre-BCG matching FFPE blocks missing or of insufficient quality to isolate RNA. For Cohort B, additional patients with HR-NMIBC were included. Cohort B resembled Cohort A, with the aim to have a similar number of tumors that responded to BCG (n=88) and that did not respond to BCG (n=63). To increase the number of patients in Cohort B, we included patients with Ta (n=38) tumors and with G2 (HG) disease (n=16) in this group. In Cohort B, patients received ≥5/6 BCG induction instillations, with a median of 13 instillations. Half of the patients from Cohort B originated from two sites that were not included in Cohort A: Reinier de Graaf (the Netherlands) and Stavanger University Hospital (Norway).

Clustering analysis pipeline and molecular clustering

Details on whole-transcriptome sequencing and tissue selection are in the supplementary materials and methods. The R package cluster was used to randomly split (3:1) Cohort A based on two 2 variables (BCG-failure [yes/no] and Progression [yes/no]) (71). The training cohort (n=99) was filtered for protein coding transcripts (R package AnnotationDbi) (72). We also removed all immune related genes (union of the signature genes from quanTIseq, EPIC and CIBERSORT algorithms, see following section) from our dataset (26–28). This strategy prevented predominant clustering on immune-related genes and enabled between-cluster comparisons using RNA-sequencing-based immune cell deconvolution, which makes use of the filtered signature genes. Then, we performed unsupervised clustering (R package ConsensusClusterPlus) with top variant genes (500–6000) with steps of 250 genes ranked based off of the calculated row variance (73). Arguments for ConsensusClusterPlus included: partitioning around medoids (PAM), Pearson correlations and iterating 2500 times with 95% samples. We determined that clustering with the 2000 top variant mRNAs resulted in the most robust sample allocation for three clusters (Fig. S1B) as determined by intercluster variation defined by the CDF plot, consensus matrix, and leading to clear differences in PFS (Fig. S1C). After analyzing survival and interpretation of GSEA results between clusters, we deemed three gene clusters as clinically informative and useful. Clustering labels from the training cohort were used to produce a multiclass predictive gene signature with the shrunken nearest centroid method (R package pamr) (74). The resulting pamr object was able to predict three classes in the testing cohort (n=33). Based on estimated survival (Fig. S1D) and overlapping GSEA hallmarks between identical subtypes from both training and testing cohorts (Fig. S1E), we confirmed the discovery of three distinct molecular subtypes in Cohort A.

The next step in establishing the algorithm parameters was clustering of all pre-BCG Cohort A patients using the established input parameters. To this end, Cohort B was quantile normalized and matched to Cohort A (the top 2000 most varying genes between samples). Results of ConsensusClusterPlus were used to investigate three, four and five gene clusters (Fig. S1F). We determined that the 3-cluster solution was again the best fit for the data based on intercluster variation through internal validation, differences in GSEA pathway analyses between clusters, and survival analysis. We generated a pamr-based nearest-centroid gene classifier based on the PAM model from Cohort A. Using cross validation we determined an optimal threshold value, minimizing the misclassification rate. We applied the pre-BCG trained model to validate the BRS in Cohort B and external datasets. For the post-BCG analysis within Cohort A, pre- and post-BCG data were normalized and scaled together using the same pamr pipeline.

Subtype validation and calculating molecular signatures

The BRS classifier was used to predict the BRSs for Cohort B. Similarly, the BRS classifier was also applied to the BCG samples from UROMOL21, to the T1BC dataset (GSE154261) and the TGCA dataset (13). The TCGA BLCA dataset was downloaded using the Genomic Data Commons (GDC) portal (via https://portal.gdc.cancer.gov/projects/TCGA-BLCA). For both datasets, raw counts were VST normalized using DESeq2 and then used for downstream analyses. Using the consensusMIBC R package, we generated all MIBC subtypes from our expression matrices (75). A 68-gene CIS signature score was also built based on a previously reported study from Dyrskjot et al. (76). The luminal BC signature (Fig S6C) included the following genes: FGFR3, PPARG, FOXA1, GATA3, ERBB2 ERBB3, DDR1, UPK1A, UPK2, UPK3A, UPK3B, KRT18, KRT19, KRT20, CDH1, FABP4, CD24, XBP1 and CYP2J2. For T1 classification, the classifyT1BC R package was used (17). We used the instructions reported in Lindskrog et al to calculate the 12-gene progression signature (13). Results are depicted with Kaplan-Meier estimates, boxplots, and heatmaps. All generated heatmaps were ordered according to the following signatures: i) BRS, ii) Consensus MIBC, iii) TCGA, iv) Lund, and v) MDA subtypes.

Code availability

Coefficients of the BRS classifier and datasets are found in the R package on GitHub (https://github.com/CostelloLab/BRSpred) and supplied in the supplementary data. Gene expression data matrices can be supplied to the BRS classifier. If missing genes are detected, mean values are imputed.

Immunohistochemistry

Sections were obtained from the same FFPE blocks as used for RNA isolation. Samples were removed from the dataset if insufficient whole tissue slides could be obtained for analysis. No samples were intentionally removed from the dataset. Immunohistochemistry was performed according to the BenchMark ULTRA (Ventana) protocol. Sections were deparaffinated and washed (EZ Prep, Ventana). Heating and incubation were performed according to the Ultra-CC1 condition (Cell Conditioning 1, Ventana, 32 minutes at 100 °C) with Reaction Buffer (10X, Tris-based buffer solution (pH 7.6 +/− 0.2) washing steps. Next, the primary diagnostics-approved antibody (anti-CD8, clone SP57, Ventana) was added and incubated (32 minutes at 36 °C). Antibody detection was done with the OptiView DAB IHC Detection Kit according to standard diagnostics protocol. For analyses, two independent researchers selected up to six regions surrounding the macrodissected areas used for RNA sequencing and counted the CD8⁺ positive T cells within a single microscopic high-powered field (40x). Boxplots were generated from the mean number of CD8⁺ positive T cells of up to twelve regions (six regions times two investigators). Wilcoxon testing was done, with statistical significance set at p < 0.05.

Spatial proteomics

Multispectral imaging was done using the Akoya Vectra Polaris instrument at the Human Immune Monitoring Shared Resource (HIMSR) at the University of Colorado Anschutz Medical Campus. Tissue sections from the same tumor blocks as used for RNA-isolation were stained consecutively with specific primary antibodies according to standard protocols provided by Akoya and performed routinely by the HIMSR. Briefly, the slides were deparaffinized, heat treated in antigen retrieval buffer, blocked, and incubated with primary antibodies for CD19, CD4, CD11c, FOXP3, CSFR1 (CD115), CD14, Vimentin, and pan-CK followed by horseradish peroxidase (HRP)-conjugated secondary antibody polymer, and HRP-reactive OPAL fluorescent reagents. To prevent further deposition of fluorescent dyes in subsequent staining steps, the slides were stripped in between each stain with heat treatment in antigen retrieval buffer. Whole slide scans were collected using the 10x objective with a 1 micron resolution. Regions of interest surrounding the macrodissected areas used for RNA sequencing were selected for multispectral imaging and quantification of tumor areas (3 – 28 fields per tissue, depending on tissue size). Multispectral imaging was performed using the 20x objective with a 0.5 micron resolution. The 9 color images were analyzed with inForm software version 2.5.1 (Akoya Biosciences) to unmix adjacent fluorochromes, subtract autofluorescence, segment tumor and stroma regions of the tissue, segment cellular compartments, and phenotype infiltrating cells according to cell marker expression. Trained phenotyping algorithms developed in inForm for each marker were applied across the entire image set and data were compiled and summarized using Phenoptr Reports software. Summarized findings for intratumoral regions were visualized with boxplots and heatmaps. Wilcoxon tests were used to compare groups, with a two-sided statistical significance threshold of p<0.05.

Quantative signaling pathway assay

All pre-BCG samples for Cohort A were used for the RT-qPCR quantitative signal transduction assay (OncoSignal^®). Five out of 132 samples failed due to depleted RNA. Samples were checked for RNA integrity and possible PCR disturbing contaminants via performing control PCRs on a number of reference genes. Samples that passed incoming Quality Control were processed further and the activity of ER, AR, MAPK, PI3K, HH, Notch and TGFb pathways were measured. In short: 90μl Purified and DNase treated RNA of tumor tissue is mixed with One-step RT-qPCR reagents (SuperScript III Platinum One-Step qRT-PCR Kit (Thermo Fisher Scientific, cat. no. 11732088) and Nuclease-free water according to the manufacturer’s instructions. 25 μl of this mix was added to each well of a 96 wells OncoSignal Testing Plate after which the plate is sealed. Each well contains dried-down primers and probes to specifically amplify and detect one of the target genes. The PCR cycling and detection was performed in a Bio-Rad CFX96 (Touch) Real-Time PCR Detection System according to the following protocol: 30’ 50°C for reverse transcription followed by initial denaturation at 95°C for 5’ and 45 cycles of denaturation (15” 95°) and annealing (30” 60°C). The raw fluorescence data were exported from the Bio-Rad device and uploaded into the Pathway Activity software to calculate the 7 pathway activity scores for every sample. Finally, samples were labeled according to the generated molecular subtypes and visualized with the use of boxplots for all pathways. P-values are based on Wilcoxon testing. A logistic regression model was generated to identify BRS3 vs BRS1/2 patients. Sensitivity, specificity, and the area under the curve were calculated using the pROC package (77).

Statistical analyses

The primary end point was progression-free survival (PFS), defined as the time from HR-NMIBC diagnosis until the development of MIBC (≥pT2), lymph node (LN), or distant metastatic (M+) disease. The secondary endpoint was the HG-RFS, defined as the time from HR-NMIBC diagnosis until a biopsy-proven HG recurrence had occurred. Disease-specific survival (DSS) was defined as the time from HR-NMIBC diagnosis until death from BC. Patients who were lost to follow-up were censored at the last date of follow-up or death. For the between subtype comparisons, a Fisher’s Exact Test for categorical data or a Kruskal-Wallis test for non-parametric continuous data was used. KM estimation coupled with the log-rank statistical test was used to model survival over time with KM plots truncated at 8 years. A multivariable Cox-proportional hazard model was generated to determine whether BRS was associated with PFS. Model variables included age, sex, re-TURBT, tumor focality, CIS, LVI, and BRSs (BRS1 + BRS2 vs BRS3). Tumor size was excluded due to missing variables. T1 substaging, using micro vs extensive invasion as done previously, was excluded due to patients from Cohort B having Ta disease, which cannot be substaged (3). To prevent understaging of primary T1 disease and to exclude bias from the absence of a re-TURBT, additional analyses were performed on patients only with a re-TURBT available. For all between survival analyses, results were determined to be significant at p<0.05 (two sided).

Supplementary Material

Data File S2

Data file S2. Raw counts data for Cohort A + B.

Data File S1

Data file S1. Combined expression matrix for n=283 BCG-naïve high-risk non-muscle invasive bladder cancer patients.

Data File S3

Data file S3. pamr coefficients of the BCG response subtype classifier.

Table S1

Table S1. Materials and resources.

Table S2

Table S2. Complete clinical data with patient and tumor characteristics for Cohort A and Cohort B.

Table S3

Table S3. Cohort A + B. Baseline patient characteristics and clinical follow-up of n=283 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according BCG treatment success.

Table S4

Table S4. Cohort A vs Cohort B. Baseline patient characteristics and clinical follow-up of n=283 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to cohort.

Table S5

Table S5. Cohort A. Overview of non-synonymous single nucleotide exonic variants in n=132 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to BCG response subtype.

Table S6

Table S6. Cohort A. Baseline patient characteristics and clinical follow-up of n=132 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to BCG response subtype.

Table S7

Table S7. Cohort B. Baseline patient characteristics and clinical follow-up of n=151 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to BCG response subtype.

Table S8

Table S8. Cohort A + B. Differentially expressed genes between in n=283 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to BCG response subtype.

Table S9

Table S9. Cohort A + B. Pathway and regulon analyses in n=283 BCG-naïve high-risk non-muscle invasive bladder cancer patients and stratified according to BCG response subtype.

Table S10

Table S10. Cohort A + B. Immune deconvolution analyses comparing BCG-naïve high-risk non-muscle invasive bladder cancer patients with BCG response subtype 3 vs BCG response subtypes 1 and 2.

Table S11

Table S11. Cohort A. Differential gene expression, regulon and pathway analyses in n=34 paired analyses pre- and post-BCG tumor samples.

Table S12

Table S12. Cohort A. All discovered gene fusions, top expressed gene fusions and gene fusions found in both pre- and post-BCG samples of the same patient.

Supplementary Methods

Materials and methods

Supplementary Figures

fig. S1. Consensus clustering results, Kaplan-Meier estimates of survival and gene signatures associated with BCG response subtypes (BRS) in the training and testing set in Cohort A.

fig. S2. Single nucleotide variant analysis for BCG response subtypes in Cohort A.

fig. S3. Heatmap of predicted BCG response subtypes in Cohort B.

fig. S4. Signatures and progression-free survival of BCG response subtypes in external cohorts.

fig. S5. BCG response subtype analyses in patients who had undergone a re-TURBT in Cohort A.

fig. S6. Heatmap of signature genes for BCG response subtypes, luminal gene set enrichment in BRS2 and regulon analysis. Result are based on the combination of Cohort A+B.

fig. S7. Nomogram to improve risk stratification in HR-NMIBC..

fig. S8. Kaplan-Meier estimates of Progression-Free Survival (PFS) based on currently published (N)MIBC molecular subtypes (Cohort A+B combined).

fig. S9. OncoSignals^® pathway results in HR-NMIBC (Cohort A)fig. S10. Gene fusions, gene set enrichment analysis, and single-nucleotide variants in n=34 paired pre- and post-BCG tumors from n=34 HR-NMIBC patients.

Data and materials availability:

All data associated with this study are in the paper or supplementary materials. A detailed overview of all resources including assays, commercial systems, and companies, software, critical R packages including versions and associated references used in this manuscript are in Table S1. The clinical data table is included in Table S2. A material and transfer agreement (MTA) between Stavanger and Erasmus MC is in place for tumor tissue from Norwegian patients. Raw sequencing data has been deposited for public use in the European Genome-Phenome Archive under accession number: EGAS00001006879. Code has been deposited at DOI: 10.5281/zenodo.788351. Coefficients, the BRS classifier and gene expression matrices are also found in the R package on GitHub (https://github.com/CostelloLab/BRSpred).