Immune Contexture and Differentiation Features Predict Outcome in Bladder Cancer

Background: An improved risk assessment of patients with bladder cancer (BC) is important to optimize clinical management. Objective: To identify whether immune cell subpopulations and cancer cell-intrinsic features are associated with outcome and response to ﬁrst-line chemotherapy in BC. Design, setting, and participants: Primary tumor tissue from 785 patients with BC (stage Ta-T4b) were stained using multiplex immunoﬂuorescence (CD3, CD8, FOXP3, CD20, CD68, CD163, and MHC-I) and immunohistochemistry (pancytokeratin, CK5/6, GATA3, programmed death 1 [PD-1], and programmed death ligand 1 [PD-L1]). A digital image analysis quantiﬁed staining results within the carcinoma cell and stromal part of the tumor. Outcome measurements and statistical analysis: Primary endpoints were progression-free survival, recurrence-free survival, and response to ﬁrst-line chemotherapy. Optimal cutoff values for investigated markers were estimated using maximally selected rank statistics and receiver operating characteristic for each primary endpoint. Time-to-event analyses were


Introduction
Bladder cancer (BC) is a highly molecularly heterogeneous disease [1], and the phenotype of each cancer cell is influenced by a multitude of cancer cell-intrinsic and cellextrinsic features, which may drive disease development and treatment resistance [2].
BC is characterized by a high tumor mutational burden (TMB), and a higher TMB has been associated with an increased T-cell influx in different cancer types [3,4].However, tumors can escape immune surveillance through alterations in neoantigen processing or presentation, or through upregulation of programmed death ligand 1 (PD-L1) [5].High PD-L1 expression has been associated with a poor prognosis in both non-muscle-invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC) [6][7][8][9].Recent transcriptomic profiling of BC has provided additional insights, revealing that immune infiltration is greatly associated with the different molecular subtypes: high immune infiltration has been associated with class 2b in NMIBC and the basal/squamous (Ba/Sq) subtype in MIBC [10,11].
The prognostic value of immune system biomarkers in BC has been addressed in several studies [12,13].A high level of cytotoxic T lymphocyte (CTL) infiltration has generally been associated with a favorable prognosis in BC [7,14,15]; however, it has also been associated with an increased risk of recurrence in NMIBC [16].High infiltration of macrophages and regulatory T cells (Tregs) have primarily been linked to tumor progression [17,18], but high levels of Tregs have also been associated with a positive prognosis in BC [19].Consequently, larger studies are needed to obtain consistent reliable results in order to determine optimal predictors of tumor recurrence, progression, and treatment response.Here, we present multiplex immunofluorescence (mIF) and immunohistochemical analyses of tumors from 785 patients with BC.We use a digital image analysis to investigate the spatial dynamics of cancer cell-intrinsic and cell-extrinsic features during disease development and their impact on clinical outcomes.

Patient details and samples
Biological specimens (n = 785) from transurethral resection of bladder tumor or cystectomy collected between 1992 and 2014 were retrieved from 12 Danish hospitals.Tissue samples were placed on tissue microarrays (TMAs).Further details regarding samples, procedures, and clinical followup are listed in the studies of Lindskrog et al [11], Taber et al [20], and Jensen et al [21], and in the Supplementary material.Informed written consent to take part in future research projects was obtained from all patients, and the study was approved by the National Committee on Health Research Ethics (#1706291 and #1300174).

Endpoints
Recurrence-free survival and progression-free survival were measured in months from the time of primary surgery to the event or end of follow-up.Evaluation of first-line chemotherapy response was based on pre-and posttreatment imaging according to RECIST 1.1.Survival data were updated in June 2020.

Automated quantification of scanned images
A digital image analysis was carried out using the Visiopharm image analysis software version 2018.9.5.5952 (Visiopharm A/S, Hørsholm, Denmark), as described in the Supplementary material.Identification and quantification of specific cell types were based on the colocalization of selected markers or positive brown staining in near proximity to a nucleus, as shown in Figure 1.MHC-I staining was quantified using an H score (1 Â [% area low intensity] + 2 Â [% area moderate intensity] + 3 Â [% area high intensity]).

Quantification and statistical analysis
Statistical comparisons between groups were performed using Fisher's exact test, Kruskal-Wallis test, or Wilcoxon rank sum test.The log-rank test was used to compare survival curves.Maximally selected rank statistics and receiver operating characteristic (ROC) statistics were used to estimate the optimal cutoff for time-to-event endpoints and treatment response, respectively.The prognostic potential of individual markers was analyzed by univariate and multivariable Cox regression analyses.For multivariable testing, all clinicopathological parameters significant in the univariate analysis were included.We performed ROC and time-dependent ROC statistics, and compared areas under the curve (AUCs) using DeLong test and inverse probability of censoring weighting, respectively.Significance levels were adjusted for multiple testing using the Bonferroni method (p.adjust*).All p values below 0.05 were considered significant.Data analysis was performed using R version 3.6.1 (RStudio Team, 2019).

Clinicopathological data and staining results
A total of 785 patients were included in this retrospective study: 220 with NMIBC (pTa-T1), 441 with localized MIBC (pT2-T4a), and 124 with advanced BC (T4b, N+, or M+).Summarized clinical and histopathological information is available in Supplementary Table 2.An overview of the included cohort, study design, and representative staining results is shown in Figure 1.

Overall immune infiltration association with histopathological and cancer cell-intrinsic features
We first investigated the overall degree of immune cell infiltration and cancer cell-intrinsic markers in 541 patients with available staining results from both immune panels.Overall, immune infiltration increased significantly with tumor stage (both regions, p< 0.001; carcinoma cell region, CROI, p < 0.001; Fig. 2A and 2B) and grade (NMIBC only, p = 0.025;Fig.2C, and Supplementary Fig. 1A and 1B).
For NMIBC and MIBC, overall high (>median) immune infiltration was associated with high MHC-I and PD-L1 carcinoma cell expression (MHC-I: p < 0.001; PD-L1: p < 0.001; Fig. 2D and 2E, and Supplementary Fig. 2A-H).Furthermore, RNA-seq-derived TMB data were available for a subset of NMIBC (n = 135) cases, and we found that tumors with high immune infiltration within the carcinoma cell part had a significantly higher TMB (p = 0.018; Fig. 2F, and Supplementary Fig. 2I and 2J).We also investigated differentiation features in NMIBC and MIBC defined by the most dominant GK5/6 and GATA3 staining patterns (Fig. 2G and the Supplementary material).NMIBC cases were almost exclusively luminal (72%) or double positive (24%), whereas 50% of MIBC cases were luminal, 17% basal, 16% double positive, and 16% double negative.As described in the study of Lindskrog et al [11], the double positive differentiation feature in NMIBC was associated with transcriptomic UROMOL 21 class 3 (Fig. 2H).Furthermore, we observed a partial overlap between the four differentiation features in MIBC and the consensus molecular subtypes (74% luminal/double positive tumors were classified as luminal papillary or luminal unstable, and 80% of basal tumors were classified as Ba/ Sq) [10].We observed no significant difference in overall immune cell infiltration across differentiation features (NMIBC: p = 0.26; MIBC: p = 0.21, Fig. 2I and 2J, and Supplementary Fig. 2K-N).

Specific immune subsets among different stages and differentiation features
Next, we explored immune cell subpopulations across tumor stage and differentiation features (Fig. 3A).Generally, we observed that Ta and T1 tumors were characterized by lower infiltration of the different subsets of immune cells and low PD-L1, PD-1, and MHC expression compared with higher-stage disease (Fig. 3B).Only infiltration with CTLs were found to be significantly higher in Ta tumors than in higher-disease stages.
In NMIBC, the two major differentiation features (luminal and double positive) showed similar immune composi-       tion (Supplementary Fig. 3A and 3B).In MIBC, basal tumors were characterized by high carcinoma cell infiltration of CTLs and high expression of PD-L1 and MHC-I compared with the other differentiation features (Fig. 3C, and Supplementary Fig. 3C and 3D).

3.4.
Prognostic role of immune subtypes in NMIBC and MIBC For a subset of MIBC patients, we have previously reported that immune infiltrated (high infiltration into the CROI) and immune excluded (immune cells restricted to the S ROI ) tumors were associated with a better response to first-line chemotherapy than immune desert (few immune cells present in both regions) tumors [20].No natural clustering into these immune subtypes was observed in either NMIBC or MIBC in this study (Supplementary Fig. 4A and 4B).Immune subtypes were not prognostic in NMIBC (Supplementary Fig. 4C and 4D).However, recurrence-free survival was significantly improved in MIBC patients with infiltrated tumors compared with the other subtypes (p = 0.019; Supplementary Fig. 4E).
We also identified subtypes based on hierarchical clustering of immune cell subsets and immune evasion markers.The resulting clusters had no prognostic value in NMIBC (p = 0.24-0.093;Supplementary Fig. 5).However, for MIBC, Kaplan-Meier survival analysis indicated improved survival for patients with clusters 2 and 4, enriched with immune cells and high PD-L1 expression (p = 0.003; Supplementary Fig. 6A-C).Furthermore, we observed the lowest response rate to first-line chemotherapy in cluster 3, characterized by tumors with low immune infiltration (overall difference, p = 0.3; Supplementary Fig. 6D).

3.5.
Immune subpopulations and cell-intrinsic features predicting clinical outcomes in NMIBC First, we assessed whether the overall immune cell infiltration could predict clinical outcomes for patients with NMIBC.However, overall immune infiltration had no significant prognostic value in NMIBC (Fig. 4A, and Supplementary Tables 3 and 4).We then asked whether the different subsets of immune cells and cell-intrinsic features have prognostic value in NMIBC.For that, we estimated optimal cutoff values for predicting recurrence and progression (Supplementary Tables 3 and 4), and performed a univariate and a multivariable Cox regression analysis for each candidate biomarker.For multivariable testing, we adjusted for the currently applied risk assessment tool by the European Association of Urology (EAU) and BCG treatment either before or after the analyzed tumor (Fig. 4A).
High infiltration of the different subsets of T lymphocytes (T helper cells, CTLs, and Tregs) was independently associated with a low risk of recurrence compared with low infiltration.This was most evident for CTLs (CROI: p.adjust* = 0.015; SROI: p.adjust* = 0.0007; Fig. 4A, and Supplementary Tables 3 and 4).High infiltration of CTLs within the carcinoma cell region was also associated with a low risk of progression (CROI: p.adjust* = 0.036; SROI: p.adjust* = 0.28).
To test whether the EAU risk score combined with our most promising predictors of recurrence and progression could improve the prediction compared with EAU risk score alone, we performed time-dependent ROC analyses.Neither stromal PD-1 expression nor stromal CTL infiltration improved the prediction accuracy for recurrence (PD1: p = 0.22, CTLs: p = 0.26; Fig. 4B and Supplementary Fig. 7).For progression, addition of PD-1 or PD-L1 expression in the carcinoma cell region improved the prediction accuracy (PD-1: p = 0.020; PD-L1: p = 0.021; Fig. 4C and Supplementary Fig. 7) compared with the EAU risk score alone.However, a model containing both PD-1 and PD-L1 did not significantly improve the prediction further (p = 0.21; Fig. 4C and Supplementary Fig. 7).

Immune subpopulations and cell-intrinsic features predicting clinical outcomes in MIBC
In a similar manner as for NMIBC, we analyzed the clinical value of the overall immune cell infiltration, immune cell subsets, and cancer cell-intrinsic features in MIBC (Supplementary Tables 5 and 6).Here, we assessed whether these have any prognostic or predictive value independent of clinicopathological factors (Fig. 5A).We also adjusted for sample age, as we observed a significant difference in the overall degree of immune cell infiltration in MIBC in relation to the age of the formalin-fixed paraffin-embedded (FFPE) material (p < 0.001; Supplementary Fig. 8).
We observed that high overall immune cell infiltration within the carcinoma cell region was independently associated with a lower risk of recurrence than low infiltration (CROI: p.adjust* = 0.038 SROI: p.univariate = 0.6; Fig. 5A, and Supplementary Tables 5 and 6), whereas high stromal immune cell infiltration was independently associated with a better response to first-line chemotherapy (CROI: p.univariate = 0.77; SROI: p.adjust* = 0.0044).
A high proportion of T helper cells was independently associated with a lower risk of recurrence (CROI: p.adjust* = 0.0098; SROI: p.adjust* = 0.018).High infiltration of CTLs within the carcinoma cell region was also associated with a low risk of recurrence (CROI: p.adjust* = 0.015; SROI: p.adjust* = 0.17).On the contrary, high infiltration of stromal M2 was independently associated with an increased risk of recurrence (CROI: p.univariate = 0.13; SROI: p.adjust* = 0.005).Interestingly, we observed that high stromal infiltration of different immune cell subsets (T helper cells, CTLs, M1, and M2) was associated with a better response to first-line chemotherapy (p.adjust* = 0.024-0.0096).
In accordance with our previous study assessing gene expression consensus subtypes [20], we observed that basal tumors have a poor response to first-line chemotherapy compared with the other differentiation features (p.adjust = 0.0075; Fig. 5A).
An analysis of the two strongest independent predictors of recurrence, PD-1 and PD-L1 expression within the carcinoma cell region, showed that high PD-L1 expression improved the predictive accuracy compared with N-stage alone (PD-L1: p = 0.0023; PD-1: p = 0.26; Fig. 5B and Supplementary Fig. 10).Inclusion of both variables in the model did not provide additional value (p = 0.082).Finally, we assessed whether basal differentiation compared with the strongest immune marker for predicting response to firstline chemotherapy, PD-1 expression in the carcinoma cell region, could improve the AUC compared with clinicopathological factors alone.Both markers improved the AUC, but not significantly (PD-1 CROI: p = 0.21; basal: p = 0.19; Fig. 5C).It should be noted that the ROC analyses are based on fewer patients than the odds ratios (Fig. 5A) because of the need of a complete overlap between all variables in the ROC analysis (152 vs 163 patients).

Discussion
Here, we present a large-scale study exploring the clinical impact of the immune contexture and cancer cell-intrinsic features in BC.Consistent with previous findings, we observed that immune infiltration in BC is highly stage dependent [23].We observed a significant increase in immune cell infiltration within the carcinoma cell region across stages.One explanation for the migration of immune cells into the tumor core concurrently with infiltration depth could be the much higher TMB found in MIBC than in NMIBC [24,25], resulting in a higher neoantigen burden, ultimately making MIBC more immunogenic.
In NMIBC, a high proportion of the different T-cell subsets (T helper cells, CTLs, and Tregs) were independently associated with a low risk of recurrence, consistent with previous reports [12], whereas only high CTL infiltration within the carcinoma cell region was independently associated with a low risk of progression.For MIBC, high infiltration with T helper cells in both regions and CTLs within the carcinoma cell region were independent prognostic markers of recurrence.Tregs favor a protumorigenic immune response and have been associated with a poor prognosis in several cancer types [26].This paradoxical prognostic effect of Tregs observed in NMIBC has been described by others [19,27,28], and may be explained by Treg-mediated downregulation of matrix metalloproteinase 2 [27].Taken together, our data might suggest that infiltration of T-cell subsets have a greater prognostic potential in NMIBC than in MIBC.Recent studies have focused on exhausted and dysfunctional T cells in cancer, which may explain why T cells are not as effective in later stages [29].Interactions with immunosuppressive cells, such as M2 macrophages and Tregs, are considered important factors mediating T-cell dysfunction [29].In our study, high stromal M2 infiltration was independently associated with a high risk of recurrence in MIBC, in line with previous findings [12,30,31].
Another known mechanism of T cell dysfunction is PD-1/ PD-L1 interactions [29].High expression of PD-L1 and its counterpart PD-1 is generally considered to be an indicator of a poor prognosis in both NMIBC and MIBC [6,32,33], but positive correlations with prognosis have also been reported [14,28].This discrepancy can partly be attributed to the large variability observed between antibodies (especially PD-L1) and nonstandardized cutoff values [34,35].In this study, high PD-1 and PD-L1 expression was associated with an increased risk of progression and recurrence in NMIBC.In MIBC, we observed an opposite trend, where high expression of PD-1 and PD-L1 correlated with a lower risk of recurrence.It seems counterintuitive that high PD-1 and PD-L1 expression is associated with a favorable prognosis in MIBC, and the biology underlying this needs to be explored in future studies.
Accumulating evidence suggests that the composition of the immune landscape in tumors may contribute to the cytotoxic effects of chemotherapy [36,37].In BC, the ratio of CD8 + to FOXP3 + has been linked to a neoadjuvant chemotherapy response in a small study of 41 patients [38].In our study, we observed that high stromal immune infiltration of both T helper cells, CTLs, and macrophages (M1 and M2) was associated with a better response to first-line chemotherapy.We also observed that basal tumors are less likely to respond to first-line chemotherapy as recently reported [20,39].Others have reported that luminal tumors and immune-infiltrative basal tumors respond better to neoadjuvant chemotherapy in MIBC [40].In our cohort, basal tumors were characterized by high infiltration of CTLs within the carcinoma cell part, and also a high level of immunosuppressive PD-L1-positive cells.Thus, T-cell dysfunction in our advanced cohort may explain the discrepancy.Additional studies are needed to fully establish the predictive value of the basal differentiation feature.
Multiple studies have investigated the immune landscape in BC, however using mainly single-marker immunohistochemistry and analog quantification, or not including spatial organization of immune cells.Our study is limited by the retrospective study design, use of TMAs, and potential batch-to-batch effect.Finally, it should be noted that our approach does not identify the cell types that expressed PD-1 and PD-L1.

Conclusions
Our results highlight several cancer cell-intrinsic and cellextrinsic features associated with clinical outcomes in BC, and importantly we show that our most promising markers (PD-1 and PD-L1) combined with clinical risk factors improved the prediction accuracy compared with clinical factors alone.If validated by additional studies, this could potentially aid future clinical management.

Fig. 1 -Fig. 2 -
Fig. 1 -Overview of the included cohort, clinical endpoints and study design.(A) Flowchart showing clinical outcome of the 785 patients with bladder cancer included in this study.Recurrence was defined as any local, regional or distant recurrence determined by pathology or cross-sectional imaging.Progression was defined as pathologically verified MIBC, or the development of metastasis determined by cross-sectional imaging.Response to first-line chemotherapy was evaluated according to RECIST1.1, and was defined as complete (CR) or partial response (PR), whereas no response was defined as stable disease (SD) or progressive disease (PD).(B) Eight tissue microarray (TMA) sections per patient were stained, scanned and aligned in the presented order.(C) A digital image analysis protocol was used to identify the carcinoma and stromal region of interest (ROI) based on pan-cytokeratin (A1/A3) staining.The border region (20 lm wide) was defined to exclude immune cells that were in contact with both regions.(D) Cells were classified based on the co-localization of selected markers as shown.To evaluate MHC-I expression we classified low, moderate and high staining intensity, and calculated a Histo-score = H score ; [1 Â (% area low intensity) + 2 Â (% area moderate intensity) + 3 Â (% area high intensity)]).TURBT = trans urethral resection of bladder tumor; CX = cystectomy; mIF = multiplex immunofluorescence; IHC = immunohistochemical.

Fig. 3 -
Fig. 3 -Immune subpopulations and immune evasion markers association with stage and differentiation features.Heatmap showing the spatial organization of immune cell subsets and immune evasion markers (PD-1, PD-L1 and MHC-I) stratified by tumor stage, annotated by the immune cell proportion (top panel).Columns are sorted by the degree of immune infiltration.Cell fractions were normalized using z-scores z= (xÀl)/r.(B-C) Heatmap showing the median cell fraction score (normalized using z-scores across tumor stage or differentiation features) stratified by tumor stage and differentiation features (muscle invasive bladder cancer only), respectively.Asterisks indicate significant association between a cell type and T-stage (Ta vs. all other stages ect.) or differentiation subtypes (basal vs. all other subtypes ect.).P values were calculated using the Wilcoxon rank sum test and corrected for multiple comparisons by the Bonferroni method.* p<0.05 ; ** p<0.01 ; *** p<0.001.

Fig. 4 -B
Fig. 4 -Immune subsets and cell-intrinsic features predict clinical outcomes in NMIBC.(A) Forest plot based on univariate Cox regression analyses for recurrence (black) and progression (blue) in NMIBC.The different immune cell subsets and immune evasion markers are stratified by optimized cut-offs (Supplementary Table3-4).Dots represent hazard ratios (HR) and lines the corresponding 95% confidence intervals (95% CI).Multivariable Cox regression analyses were performed separately for each variable together with the European Association of Urology (EAU) risk score and the variable BCG treatment either before or after the analyzed tumor.P values were calculated by the Wald test.The multivariable p values (p.adjust) were corrected by the Bonferroni correction (p.adjust*).Luminal = CK5/6-GATA3+, Double positive = CK5/6+GATA3+.(B-C) Time dependent receiver operating characteristic (ROC) curves censored at 24 months for different Cox regression models predicting recurrence (n=145) and progression (n=153), respectively.The two strongest predictors in the multivariable analysis (Fig.4A) are included.P values in B and C are calculated by the Inverse Probability of Censoring Weighting and test if the Area Under the Curve (AUC) changes significantly when an additional variable is included in the model.*p < 0.05, **p <0.01, NS = not significant (>0.05); yr inc.= years increment; M/F = males/females; C-index = concordance index.

Fig. 5 -
Fig. 5 -Immune subsets and cell-intrinsic features predict clinical outcomes in MIBC.(A) Forest plot based on univariate Cox regression analyses for recurrence (black) and logistic regression for response to first-line chemotherapy (blue) in muscle invasive bladder cancer (MIBC).The different immune cell subsets and immune evasion markers are stratified by optimized cut-offs (Supplementary Table5-6).Dots represent hazard ratios (HR) for recurrence and odds ratios (OR) for response.The corresponding lines are 95 % confidence intervals (95% CI).Multivariable Cox regression and logistic regression analyses were performed separately for each variable with the clinicopathological factors significant in the univariate analysis and the date of the FFPE material (90's/00's/10's).P values are calculated by the Wald test for Cox regression and the Students test for logistic regression.The multivariable p values (p.adjust) were corrected by the Bonferroni correction (p.adjust*).Basal=CK5/6+GATA-, Not Basal=CK5/6+GATA3+ or CK5/6-GATA3+-.(B) Time dependent receiver operating characteristic (ROC) curves censored at 24 months for different Cox regression models predicting recurrence (n=307).The two strongest predictors in the multivariable analysis (Fig.5A) are included.P values are calculated by the Inverse Probability of Censoring Weighting and test if the Area Under the Curve (AUC) changes significan when an additional variable is included in the model (C) ROC curves for different logistic regression models predicting response to first-line chemotherapy (n=152).The strongest predictor in the multivariable analysis (Fig.5A) is included together with the differentiation marker variable (Basal/Not Basal).P values are calculated by the DeLong test and test if the AUC changes significantly when an additional variable is included in the model.*p < 0.05, **p < 0.01, NS=not significant (>0.05); yr inc.= years increment; M/F = males/females; aBC = advanced bladder cancer; C-index = concordance index; PS = performance status.

5
Fig. 5 -Immune subsets and cell-intrinsic features predict clinical outcomes in MIBC.(A) Forest plot based on univariate Cox regression analyses for recurrence (black) and logistic regression for response to first-line chemotherapy (blue) in muscle invasive bladder cancer (MIBC).The different immune cell subsets and immune evasion markers are stratified by optimized cut-offs (Supplementary Table5-6).Dots represent hazard ratios (HR) for recurrence and odds ratios (OR) for response.The corresponding lines are 95 % confidence intervals (95% CI).Multivariable Cox regression and logistic regression analyses were performed separately for each variable with the clinicopathological factors significant in the univariate analysis and the date of the FFPE material (90's/00's/10's).P values are calculated by the Wald test for Cox regression and the Students test for logistic regression.The multivariable p values (p.adjust) were corrected by the Bonferroni correction (p.adjust*).Basal=CK5/6+GATA-, Not Basal=CK5/6+GATA3+ or CK5/6-GATA3+-.(B) Time dependent receiver operating characteristic (ROC) curves censored at 24 months for different Cox regression models predicting recurrence (n=307).The two strongest predictors in the multivariable analysis (Fig.5A) are included.P values are calculated by the Inverse Probability of Censoring Weighting and test if the Area Under the Curve (AUC) changes significan when an additional variable is included in the model (C) ROC curves for different logistic regression models predicting response to first-line chemotherapy (n=152).The strongest predictor in the multivariable analysis (Fig.5A) is included together with the differentiation marker variable (Basal/Not Basal).P values are calculated by the DeLong test and test if the AUC changes significantly when an additional variable is included in the model.*p < 0.05, **p < 0.01, NS=not significant (>0.05); yr inc.= years increment; M/F = males/females; aBC = advanced bladder cancer; C-index = concordance index; PS = performance status.
Fig. 5 -Immune subsets and cell-intrinsic features predict clinical outcomes in MIBC.(A) Forest plot based on univariate Cox regression analyses for recurrence (black) and logistic regression for response to first-line chemotherapy (blue) in muscle invasive bladder cancer (MIBC).The different immune cell subsets and immune evasion markers are stratified by optimized cut-offs (Supplementary Table5-6).Dots represent hazard ratios (HR) for recurrence and odds ratios (OR) for response.The corresponding lines are 95 % confidence intervals (95% CI).Multivariable Cox regression and logistic regression analyses were performed separately for each variable with the clinicopathological factors significant in the univariate analysis and the date of the FFPE material (90's/00's/10's).P values are calculated by the Wald test for Cox regression and the Students test for logistic regression.The multivariable p values (p.adjust) were corrected by the Bonferroni correction (p.adjust*).Basal=CK5/6+GATA-, Not Basal=CK5/6+GATA3+ or CK5/6-GATA3+-.(B) Time dependent receiver operating characteristic (ROC) curves censored at 24 months for different Cox regression models predicting recurrence (n=307).The two strongest predictors in the multivariable analysis (Fig.5A) are included.P values are calculated by the Inverse Probability of Censoring Weighting and test if the Area Under the Curve (AUC) changes significan when an additional variable is included in the model (C) ROC curves for different logistic regression models predicting response to first-line chemotherapy (n=152).The strongest predictor in the multivariable analysis (Fig.5A) is included together with the differentiation marker variable (Basal/Not Basal).P values are calculated by the DeLong test and test if the AUC changes significantly when an additional variable is included in the model.*p < 0.05, **p < 0.01, NS=not significant (>0.05); yr inc.= years increment; M/F = males/females; aBC = advanced bladder cancer; C-index = concordance index; PS = performance status.