TCGA KIRC cohort. For TCGA KIRC cohort, data were abstracted from 533 patients with both RNA-seq data and clinical information in TCGA kidney clear cell renal cell carcinoma (KIRC) cohort. Clinical data were downloaded from cBioPortal12 and the TCGA Pan-cancer Clinical Data Resource, a database that provides processed high-quality clinical data.13 RNA-seq data were downloaded from the UCSC Xena. Mutational and arm somatic copy number alteration (SCNA) data were obtained from cBioPortal.

Zhongshan ccRCC cohort. The Zhongshan ccRCC cohort included patients with ccRCC who underwent nephrectomy at Zhongshan Hospital between January 2005 to June 2007. Data were censored until June 2017. Inclusion criteria were patients with pathologically proven ccRCC treated with nephrectomy and with available Formalin Fixed Paraffin Embedded (FFPE) specimens. Exclusion criteria were neoadjuvant or adjuvant systemic therapy, synchronous or metachronous bilateral RCC or histories of other malignant tumors. A total of 249 patients were included for further analyses.

Zhongshan fresh tumor sample cohort. Fresh tumor samples in the Zhongshan fresh tumor sample cohort were collected from 20 ccRCC patients between October 2017 and November 2018. The same inclusion and exclusion criteria of Zhongshan ccRCC cohort were applied, except that fresh tumor sample cohort did not need FFPE specimen.

Zhongshan metastatic ccRCC cohort. We retrospectively enrolled a cohort of 87 pathologically proved metastatic ccRCC patients treated with first-line sunitinib or sorafenib between March 2005 and June 2014 as the Zhongshan metastatic ccRCC cohort. The last follow-up time was January 2015. For inclusion criteria, patients had to have developed metastases, received sunitinib or sorafenib as first-line treatment in the metastatic setting, undergone their first CT scan assessment and processed available FFPE specimen. Exclusion criteria were the same as previously described. Response to treatment was assessed with RECIST 1.1.14

Immune cell compositions and pathway analysis. We evaluated the absolute and relative cell fractions of major types of tumor infiltrating immune cells with CIBERORT, a computational approach for inferring leukocyte representation in bulk tumor transcriptomes.15 Single sample Gene Set Enrichment analysis (ssGSEA) was chosen for immune deconvolution analyses of Immune Score16 and immune suppression score.17 Comparison of gene expression profiles were carried out with Gene Ontology (GO) analysis and Gene Set Enrichment Analysis (GSEA).18

Clustering and classifier construction. Unsupervised hierarchical clustering of normalized immune cell fractions with K-median identified the intrinsic pattern of immune cell infiltration, TMEcluster-A and TMEcluster-B. The optimal number of clusters was determined by Nbclust testing. To develop a robust immune cell composition classifier for assessing TME subtype, we applied prediction analysis for microarray (PAM),19 a centroid-based classification algorithm. PAM was widely used because of its reproducibility in subtype classification compared with other centroid-based prediction method.20 For example, the establishment of the well-recognized simplified version of ccA/ccB gene signature, ClearCode 34, was based on PAM.21 The Zhongshan ccRCC cohort, fresh tumor sample cohort and metastatic ccRCC cohort were classified into TMEcluster-A and TMEcluster-B with normalized immune cell densities obtained from immunohistochemistry using PAM model in this study.

Immunohistochemistry. We performed immunohistochemistry on Zhongshan ccRCC cohort, fresh tumor sample cohort and metastatic ccRCC cohort to evaluate infiltrations of macrophages (CD68, clone KP1, Dako), CD4⁺ T cells (CD4, ab213215, Abcam), Tregs (FOXP3, ab22510, Abcam), CD8⁺ T cells (CD8, clone C8/144B, Dako), B cells (CD19, ab31947, Abcam) and mast cells (tryptase, ab134932, Abcam) for further subtype classification with PAM model. SETD2 was stained with primary anti-SETD2 (HPA04245, Sigma-Aldrich Corp) antibody. The densities of each type of immune cells were evaluated in two representative areas at ×200 magnification. The mean value was adopted, changed into density as cells/mm2 and normalized. The PAM model was performed with normalized densities of each immune cell type in all three Zhongshan validation cohorts to assign the TME clusters. Two independent shots of SETD2 expression with the strongest staining at ×200 magnification were recorded and evaluated via the semi-quantitative immunoreactivity score (IRS) algorithm.

Flow cytometry. Cell suspensions were stained with fluorochrome-labeled antibodies specific for human CD3 (344 818 Biolegend), CD8 (301 006 Biolegend), PD1 (560 795 BD), CTLA4 (369610, Biolegend), TIM3 (345032, Biolegend) and TIGIT (372710, Biolegend). Cell suspensions were collected from tumors in Zhongshan fresh tumor sample cohort. We analyzed the stained cells on a flow cytometer using Flowjo software.22

Differences in continuous variables between two groups were analyzed by Student’s t-test or t’-test according to Levene’s test. Pearson’s χ² test, Cochran-Mantel-Haenszel χ2 test or Fisher’s exact test were used for categorical variables. Survival analyses were carried out with Kaplan-Meier method and Cox proportional hazards regression model. All statistical tests were two-sided and statistical significance was set at p<0.05. For TCGA KIRC cohort and Zhongshan ccRCC cohort, overall survival (OS) was defined as the length of time from the data of diagnosis to date of death or last follow-up time. Recurrence free survival (RFS) was defined as time from surgery to loss of follow-up or deaths from other causes without recurrence. Disease specific survival (DSS) was censored at deaths from ccRCC. Metastatic ccRCC patients were excluded from RFS analysis in both KIRC and Zhongshan ccRCC cohort. Overall survival was calculated from the time of therapy initiation to the time of death or last follow-up. Progression free survival (PFS) was defined as the time from therapy initiation to the time of disease progression or the last follow-up time. Four patients were excluded from PFS analysis for lack of progression information. Prognostic capabilities of different risk models were compared using time dependent receiver operating characteristic (ROC) analysis and Harrell’s concordance index (C-index).The R package ‘limma’ was used for analysis of differentially expressed genes. Statistical analyses were carried out with SPSS Statistics 21.0 and R software 3.51.

Analyses of major immune cell composition in KIRC cohort revealed that CD4⁺ T cells, CD8⁺ T cells, macrophages and Tregs increased while mast cells decreased with disease progression (Figure 1A, online supplementary figure 1A). The TME cell network demonstrated that macrophages, CD4⁺ T cells, CD8⁺ T cells, Tregs and B cells were positively correlated with each other. Moreover, macrophages, CD4⁺ T cells, CD8⁺ T cells, Tregs and neutrophils associated with shortened OS while mast cells associated with prolonged OS (Figure 1B, online supplementary table 1). Unsupervised hierarchical clustering of normalized immune cell fractions was performed and two major TME clusters with different clinical characteristics were identified (Figure 1C, online supplementary table 2). TMEcluster-A exhibited higher infiltration of mast cells. TMEcluster-B was characterized by increases in the infiltration of B cells, CD8⁺ T cells, Tregs, Macrophages and CD4⁺ T cells. TMEcluster-B correlated with higher stages and grades. The two TME subtypes showed significant differences in OS (p<0.001; HR 2.162, 95% CI: 1.546 to 3.023), RFS (p<0.001; HR 2.018, 95% CI: 1.373 to 2.967) and DSS (p<0.001; HR 2.907, 95% CI: 1.834 to 4.609) (Figure 1D, online supplementary figure 1B,C). After adjustment of stage and grade as binary variables, TME clusters remained an independent risk factor for OS (HR 1.482, 95% CI: 1.048 to 2.095, p=0.026).

Immune cell infiltrations in the KIRC cohort were taken as the training set in the PAM model. Macrophages, CD4⁺ T cells, Tregs, CD8⁺ T cells and B cells were extracted as phenotype signatures of TMEcluster-B while mast cells were extracted as phenotype signatures of TMEcluster-A (online supplementary figure 2A). We performed immunohistochemistry in Zhongshan RCC cohort to quantify densities of the phenotype signature immune cells (online supplementary figure 2B, online supplementary table 3). Applying the classifier to normalized immune cell density in Zhongshan ccRCC cohort, 100 samples were assigned as TMEcluster-A and 149 as TMEcluster-B. There were increased B cells, CD8⁺ T cells, Tregs, macrophages, CD4⁺ T cells and reduced mast cells in TMEcluster-B (Figure 1E, online supplementary figure 1D). TMEcluster-B associated with higher tumor stage (p=0.036) but not tumor grades (p=0.202) (online supplementary figure 1E), and experienced recurrence more frequently (p=0.012; HR 1.870, 95% CI: 1.141 to 3.066) (online supplementary figure 1F). TMEcluster-B cases had higher risk of overall mortality compared with TMEcluster-A (p<0.001; HR 2.629, 95% CI: 1.630 to 4.241) (figure 1F). In addition, overall deaths (HR 2.304, 95% CI: 1.401 to 3.790, p<0.001) remained significant between clusters after adjustment of stage, grade, Eastern Cooperative Oncology Group Performance Status (ECOG PS) and necrosis (online supplementary table 4).

We noticed that there were plenty of immune-related genes among the top 50 upregulated genes in TMEcluster-B in KIRC cohort, such as genes encoding components of CD8⁺ T receptor (CD8A, CD8B, CD27, CD3D), related to cytotoxic activities mediated by CD8⁺ T (IFNG, GZMK), suggesting a cytokine-rich microenvironment (CXCR3, IL20RB, CXCL13, CCL19) and associated with immune suppression (FOXP3, TGFB1, PDCD1, CTLA4, LAG3, TIGIT) (figure 2A). GO analysis of the most upregulated 200 genes in TMEcluster-B revealed enrichment of both pathways of immune activation and immune suppression among the 10 most significantly enriched biological pathways (figure 2B). GSEA analyses revealed that a large part of upregulated pathways were associated with immune parameters in classic pathway database REACTOME, BIOCARTA and KEGG (figure 2C). In particular, pathways representing immune evasion were upregulated in each of the database, including CTLA4 inhibitory signaling, immunoregulatory interactions between a lymphoid and a non-lymphoid cell, IL10 pathway and TGF-β signaling pathway (figure 2D). Using ESTIMATE,16 we found that TMEcluster-B had higher Immune Score (p<0.001), a marker of total immune infiltration (figure 2E). Immune deconvolution analysis via ssGSEA showed TMEcluster-B had higher immune suppression (p<0.001) score17 (figure 2F). T cell exhaustion markers unanimously increased in TMEcluster-B (online supplementary figure 3A). On the contrary, CD8 T cell/Treg ratio (p<0.001)3 and GZMB/CD8A ratio (p<0.001),23 ratio markers of anti-tumor immunity, significantly decreased in TMEcluster-B (online supplementary figure 3B,C). Our findings suggested a highly immune infiltrated but immunosuppressed tumor microenvironment in TMEcluster-B. We performed flow cytometry in Zhongshan fresh ccRCC cohort and observed higher expression of exhaustion markers on CD8⁺ T cells in TMEcluster-B including PD-1 (p=0.017), CTLA-4 (p=0.116) and TIM-3 (p=0.005) as expected (Figure 2G, online supplementary figure 2C). There were higher percentage of PD1⁺CD8⁺T cells (p=0.007), CTLA4⁺CD8⁺T cells (p=0.021) and TIM3⁺CD8⁺ T cells (p=0.004) in TMEcluster-B (online supplementary figure 2D).

Next, we explored differences in the frequency of mutations in common ccRCC cancer driver genes between the two clusters. Somatic SETD2 mutations were more frequent in TMEcluster-B with borderline statistical significance (p=0.054) (figure 3A). Low SETD2 protein expression previously demonstrated poor prognostic effects among TKI treated ccRCC patients.24 We analyzed SETD2 expression in Zhongshan metastatic ccRCC cohort and found that tumors in TMEcluster-B had lower SETD2 expression (p=0.027) (Figure 3B, online supplementary figure 2E). TMEcluster-B had lower overall tumor mutation burden compared with TMEcluster-A (p=0.008) (figure 3A), in accordance with a previous study showing that most immune signatures were upregulated in low-TMB subtype in ccRCC.25 We then analyzed the associations between TME clusters and prevalent arm level SCNAs (present in over 10% patients) with poor prognosis in a previous study.26 The heatmap showed that amplifications of 8q, 12, 20 and deletions of 9 p, 14q were enriched in TMEcluster-B. Total number of poor prognosis SCNAs (amplifications of 8q, 12, 20 and deletions of 4 p, 9 p, 14q, 19, 22q) was higher in TMEcluster-B as well (p<0.001) (figure 3A). TME clusters and ccA/ccB27 signature were positively correlated (p<0.001). Cell cycle proliferation score28 and 16-gene recurrence score,29 two well established molecular risk models, were both higher in TMEcluster-B (p<0.001 and p=0.055, respectively) (figure 3A).

Hakimi et al reported that TKI non-responders had higher macrophage infiltration.11 Macrophage infiltrations determined by CIBERSORT, immunohistochemistry and metagene value30 all elevated in TMEcluster-B (figure 1C and E, online supplementary figure 3D). Markers of macrophage chemotaxis and activation also increased in TMEcluster-B (figure 3C). They identified four distinct molecular subgroups in COMPARZ phase III patients significantly differed in response and survival, among which cluster 4 showed the worst TKI treatment response.11 Cluster 4 tumors had more frequent PD-L1 positivity and TMEcluster-B had higher PD-L1 mRNA expression (p<0.001) (online supplementary figure 3E). The hallmark inflammation signatures enriched in cluster 4 were upregulated in TMEcluster-B as well (online supplementary figure 3F). Another molecular subgroup with poor response to sunitinib, ccrcc4 identified by Beuselinck et al, was characterized by a strong inflammatory, Th1-orineted but suppressive immune microenvironment.31 TMEcluster-B displayed the same TME characterization as shown in heatmap (figure 3C).

Our previous findings suggested a highly potential correlation between TME clusters and TKI treatment response. We further explored the correlation in Zhongshan metastatic ccRCC validation cohort and found that TMEcluster-B had significant worse treatment response compared with TMEcluster-A (p=0.009) (figure 4A). The objective response rate was 50.0% for TMEcluster-A and 16.9% for TMEcluster-B. Clinical benefit rate was 87.5% for TMEcluster-A and 78.0% for TMEcluster-B. Kaplan-Meier analysis and univariate analysis revealed that TMEcluster-B cases experienced earlier disease progression (p=0.015; HR 2.7762, 95% CI: 1.530 to 4.986) and overall death (p<0.001; HR 2.223, 95% CI: 1.150 to 4.229) (figure 4B,C). Utilizing IMDC risk scores alone predicted 2-year PFS and OS with an area under the curve (AUC) of 0.74 and 0.76, respectively. IMDC risk scores incorporating TME clusters reached an AUC of 0.80 and 0.82 for PFS and OS (figure 4D,E). Integration of IMDC variables and TME clusters improved the C-index for PFS from 0.61 to 0.66 and OS from 0.67 to 0.71 (figure 4F,G).