Two independent MIBC cancer cohorts (online supplementary table S1) were investigated and described previously: (1) the Comprehensive Cancer Center Erlangen- EMN cohort (CCC-EMN; n=187)18–21; and (2) The Cancer Genome Atlas cohort of bladder cancer (TCGA BLCA; n=407).5 18 In detail, the CCC-EMN cohort consisted of 187 consecutively collected MIBC cases of which 56 received adjuvant platinum-based chemotherapy. The TCGA BLCA cohort analyzed in the present study consisted of all 407 cases of the TCGA BLCA cohort with available RNAseq data. Clinical, pathological and molecular data of the TCGA cohort were updated accordingly.5 According to the clinical information of TCGA 69 patients of this cohort received platinum-based adjuvant chemotherapy without further details on the chemotherapy regimen.

The pooled adjuvant platinum-based chemotherapy cohort consisted of 56 patients from the CCC-EMN cohort (n=48: Cis-Platinum+Gemcitabine; n=8 Carbo-Platinum +Gemcitabine) and 69 patients from the TCGA BLCA cohort (no detailed information about the platinum-containing regime available). Detailed clinical-pathological data are depicted in online supplementary table S1.

The ImmuneTyper panel consists of three important genes related to interferon gamma induced T-cell driven antitumoral immunity: (1) CD3Z (CD247) codes for the zeta-chain of the T-cell receptor (TCR), (2) CD8A codes for the alpha subunit of the cytotoxic T-cell co-receptor where both genes represent stable markers for overall T-cell (CD3Z) and cytotoxic T-cell infiltration (CD8A) in MIBC as shown previously.18 22 23 Lastly (3) CXCL9 codes for a T-cell chemoattractant significantly induced by interferon gamma. CXCL9 not only correlates with interferon gamma induced T-cell driven antitumoral immune responses but also with an improved response to immune checkpoint inhibition in MIBC.5 11 18 22 23 The gene panel showed a comparable performance reflecting the general level of antitumoral inflammation compared with a T-cell-related inflammation signature consisting of 59 immune genes.18 To further validate the representativeness of the ImmuneTyper, correlational analyzes with stromal tumor infiltrating ICs (general immune infiltration marker; sTILs), specific IC populations (CD3⁺ T-cells, CD8⁺ cytotoxic T-cells, macrophages and NK-cells) and expression values of each specific gene were carried out in the CCC-EMN and the TCGA cohort.

Assessments of sTILs sTILs were analyzed on HE-stained whole mount sections in both cohorts by two independent observers as described previously according to a standard methodology of the International working group on TILs.18 24–27 sTIL assessment has been shown to reflect the overall immune infiltration status of several solid cancer entities.18 24–27

Spatially designed tissue microarrays (TMAs) covering the invasion front and the tumor center were prepared as described previously.18–20

Immunohistochemical stains were performed on 4 µm TMA tissue sections on a Ventana Benchmark Ultra autostainer (Ventana) for the following proteins: CD3 (F7.2.38, monoclonal mouse, ThermoFisher-Scientific, dilution 1:50), CD8 (C8/144B, mouse monoclonal, ThermoFisher-Scientific, dilution 1:50), CD68 (PG-M1, mouse monoclonal, ThermoFisher-Scientific, dilution 1:60), PD-1 (NAT105, mouse monoclonal; Ventana) and CD56 (MRQ-42, monoclonal mouse, CELL MARQUE, dilution 1:50) and PD-L1 (SP263 assay, Ventana).

PD-L1⁺ ICs and tumor cells (TCs) were scored by two pathologists (AH, ME) according to the distributor’s PD-L1 scoring algorithm.19 20 Positivity of TC was scored as a percentage of positive TC in relation to TC area. Number of ICs was scored as the percentage ratio of the area occupied by positive ICs per tumor area.19

Tumor, normal and stroma tissue as well as empty spaces were assessed as described previously.18 CD3⁺, CD8⁺, CD68⁺, PD-1⁺ and CD56⁺ IC were quantified (counts per mm²) and log2-transformed for further analysis.

RNA was isolated and purified as described previously.18 We selected 21 genes according to the MDDACC subtyping approach, where these genes are known to represent specific differentiation markers identifying (1) luminal tumors expressing KRT20, KRT7, GATA3, FGFR3, CYP2J2, ERBB2, ERBB3 and CDH1 as well as (2) luminal ECM-like/‘p53-like’ tumors expressing MYH11, ACTG2, CNN1, MFAP4, VIM and (3) basal tumors expressing KRT5, KRT6A-C, SNAI2, CDH3, CD44. Thus, using these 21 genes we were able to classify our tumor samples into the two major MIBC subtypes (basal and luminal).6 7 18 Gene counts were normalized using two reference genes (SDHA, HPRT1) and log2-transformed for further analysis with the nSolver 4.0 software. Selected genes are depicted in online supplementary table S2.

RNA was extracted based on a magnetic bead technology approach (RNXtract, STRATIFYER, Germany) as described previously.18 20 CD3Z (CD247), CD8A and CXCL9 expression was assessed in triplicates on a Versant qPCR system (Siemens, Erlangen, Germany). Data were normalized, using three housekeeping genes (RPL37A, B2M, CALM2) according to MIQE guidelines.28 Final values were generated using ΔCT from the total number of cycles to ensure that normalized gene expression obtained is proportional to the corresponding mRNA expression levels (40- ΔCT).

Relative fraction of IC populations based on the complete TCGA RNAseq data set were calculated using the CIBERSORT algorithm (https://cibersort.stanford.edu/) with 1000 permutations.29

Survival curves were estimated by Kaplan-Meier regression. Median potential follow-up was calculated by reverse Kaplan-Meier analysis. Patients at risk at various time points are shown below each survival plot. Statistical comparisons of subgroups included nonparametric Wilcoxon rank-sum test for continuous variables, Pearson's chi-square test for nominal variables and log rank test for time-to-event outcomes. Multivariable survival analyzes were conducted using Cox proportional hazards regression modeling, to assess the magnitude of impact (the HR) of the ImmuneTyper-signature, while adjusting for clinical-pathological variables (pT-Stage, pN-Stage, lymphovascular invasion (L), age, gender, presence of distant metastasis and tumor grading (WHO2016 and WHO1973). Overall survival (OS) was defined as survival time until death by any reason (disease related, non-disease related; OS). DSS was defined as survival time until DSS. Disease-free survival was defined as survival time until local or distant disease recurrence (DFS). Variables for final multivariable models were selected by stepwise back and forward selection which is depicted in detail in online supplementary table S3. All p values were two sided and a p<0.05 was considered statistically significant. Cluster analysis was performed by unsupervised hierarchical clustering based on average link algorithm (WPGMA, weighted pair-group method using arithmetic averages) using Euclidean distance as metric scale. All statistical analyzes were performed by GraphPad Prism V.7.2 (GraphPad Software, La Jolla, California, USA) and JMP SAS V.13.2 (SAS).

We tested whether expression of a minimum of three genes, CD3Z, CD8A and CXCL9, involved in cytotoxic activity of T-cells correlated with distinct amounts of IC populations within the TIME. Results showed strong correlations of CXCL9 (Spearman r: 0.51–0.66; figure 1A), CD3Z (Spearman r: 0.43–0.56; figure 1B) and CD8A gene expression (Spearman r: 0.44–0.64; figure 1C) with amounts of CD3⁺ and CD8⁺ T-lymphocytes, macrophages (CD68⁺) and NK-cells (CD56⁺; figure 1D) as well as with sTILs (Spearman r: 0.63–0.74; figure 1E) a biomarker for overall IC infiltration in MIBC.18 These findings indicate that levels of anti-tumor activity within the TIME can be accurately assessed by gene and cell surface protein marker expression of lymphocytes and macrophages. Intergene correlations of CD3Z (CD247), CD8A and CXCL9 are depicted in online supplementary table S4.

Unsupervised hierarchical clustering of CD3Z, CD8A and CXCL9 revealed three clusters (figure 2A): cluster A (‘inflamed: high’; median sTILs 40%), cluster B (inflamed: Low’; median sTILs 15%) and cluster C (‘uninflamed’; median sTILs 5%). Total amounts of CD3⁺ and CD8⁺ T-cells as well as of NK-cells and macrophages were highest in cluster A and second highest in cluster B (eg, median log2 value of CD8⁺-T-Cells in Cluster A 9.21 vs 8.29 in Cluster B vs 6.52 in Cluster C; p<0.0001; figure 2B). Tumors within cluster A (‘inflamed: high’) showed a predominance of basal subtype (69.0%) while tumors with lower inflammation levels were predominantly luminal (eg, 59.6% of cluster C/‘uninflamed’; figure 2A). In addition, expression of PD-L1 on IC and TC, as well as of PD-1 on IC was significantly higher in ‘inflamed: high’ and ‘inflamed: low tumors’ compared with ‘uninflamed’ tumors (figure 2C).

Patients within cluster A and cluster B showed significantly longer 5 years DSS (cluster A: 67.3% (95% CI 48.2% to 82.1%); cluster B: 47.9% (95% CI 33.3% to 62.9%) and DFS (cluster A: 64.5% (95% CI 45.2% to 79.9%); cluster B: 45.1% (95% CI 31.0% to 60.0%)) compared with uninflamed tumors with lowest 5 years DSS (29.1%, 95% CI 20.1% to 40.1%) and DFS (23.9%, 95% CI 15.7% to 34.7%) rates (figure 2D). Multivariable Cox-regression analyzes revealed clusters A and B as independent predictors for improved DSS (multivariable p value 0.00001) and DFS (multivariable p value 0.00002; figure 2E; online supplementary table S3).

We employed gene expression data of 407 MIBC out of the TCGA data set to validate our findings.5 Gene expression of the three genes, CXCL9 (Spearman r: 0.56–0.75; figure 3A), CD3Z (Spearman r: 0.53–0.70; figure 3B) and CD8A (Spearman r: 0.62–0.82; figure 3C), correlated with IC populations derived by the CIBERSORT-algorithm. Furthermore, expression of these three signature genes significantly associated with sTILs (Spearman r: 0.58–0.71; figure 3D).

Unsupervised hierarchical clustering of CD3Z, CD8A and CXCL9 revealed three clusters (figure 4A): cluster A (‘inflamed: high’; median sTILs 40%), cluster B (‘inflamed: low’; median sTILs 15%) and cluster C (‘uninflamed’; median sTILs 5%). Total amounts of CD3⁺ and CD8⁺ T-cells as well as of NK-cells and macrophages were distributed in the same manner as our CCC-EMN cohort with highest numbers in cluster A and second highest in cluster B (figure 4B). Distribution of tumor subtypes regarding the different inflammation clusters was comparable to our CCC-EMN cohort: cluster A/‘inflamed high’ showed a predominance of basal squamous tumors (61.6%) while lowly inflamed tumors were predominantly luminal (79.7% of cluster C/‘uninflamed’; figure 4A). Highly (cluster A) and lowly (cluster B) inflamed tumors exhibited significantly higher levels of the immune checkpoint genes PD-L1 and PD-1 (PDCD-1) compared with uninflamed tumors (cluster C; figure 4C).

Congruent with the above described observations, we found that patients with highly inflamed tumors (cluster A) had significantly prolonged OS and DFS (figure 4C). Different to the CCC-EMN cohort, lowly inflamed tumors (cluster B) showed no superior outcome measures than uninflamed tumors (figure 4C). Furthermore, high inflammation (cluster A) emerged as an independent prognostic factor for better OS (multivariable p value 0.0006) and DFS (multivariable p value 0.007) in a multivariable Cox-regression model compared with lowly and uninflamed tumors (figure 4D;online supplementary table S3).

To investigate whether patients treated with platinum-based adjuvant chemotherapy show a dependency on inflammation levels, we combined 69 patients from the TCGA and 56 patients from the CCC-EMN cohorts (online supplementary table S1). We also found three cluster groups (Cluster A–C; figure 5A) with comparable distributions of tumor subtypes as described above: cluster A: 53.5% basal tumors (luminal tumors 39.6%); cCluster B 69.5% luminal tumors (basal tumors 23.9%), cluster C 69.5% luminal tumors (basal tumors 19.4%).

Patients within cluster A (‘inflamed high’) showed significantly prolonged OS (multivariable p=0.0015; multivariable HR=0.27 (Cl. A vs Cl. B)) and prolonged DFS after adjuvant platinum-based chemotherapy in multivariable Cox proportional hazard analyzes (multivariable p=0.0092; multivariable HR=0.41 (Cl. A vs Cl. B);figure 5B,C; online supplementary table S3).