Multi-omic characterization of 14q deletion in renal clear cell carcinoma identifies EDNRB as a predictor of immunotherapy response

Abstract

Background

Although chromosome 14q deletion (14q−) is present in up to 40% of clear cell renal cell carcinoma (ccRCC) cases, its immunologic and molecular impact remains unclear. Because TRAF3 and NFΚBIA, key regulators of nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) signaling, are located on 14q, we hypothesize that 14q− in ccRCC promotes increased inflammation and unique determinants of immune checkpoint blockade (ICB) response.

Methods

We collected data from 825 patients with ccRCC from publicly available sources, our own independent retrospective study, and the Oncology Research Information Exchange Network. These data include bulk RNA-sequencing (RNA-seq), single-cell RNA-seq (scRNA-seq), and whole exome sequencing (WES). We also include reverse phase protein array from DepMap cell lines (n=277). Next, we performed multiplex immunofluorescence for spatial analysis of immune cells on 23 patient slides with paired WES and ICB response. Finally, progression-free survival, 14q− status (assessed in bulk RNA-seq or whole exome sequencing), and kidney inflammatory marker EDNRB expression were assessed in two independent ICB immunotherapy trials.

Results

14q− status is associated with increased NF-κB-target transcription and p65 phosphorylation. Tumor scRNA-seq revealed increased expression of CCL20, an NF-κB target and lymphotactic protein. Simultaneously, single-cell and bulk RNA-seq demonstrated higher effector CD8+T cell infiltration in 14q- tumors compared with controls. We investigated myeloid panel mIF images from 14q− tumors and found dendritic cells (DCs) significantly aggregate with tumor cells in ICB responders compared with non-responders. Cell–cell communication analysis on DCs and tumor cells from a 14q− responder revealed tumor-derived EDN1 is uniquely interacting with EDNRB on DCs. In two ICB trials, 14q− combined with elevated EDNRB expression correlated with significantly improved progression-free survival.

Conclusions

This is the first evidence that 14q− in ccRCC is associated with an antitumor inflammatory phenotype, possibly driven by aberrant NF-κB activation. Furthermore, 14q− with high EDNRB expression (EDNRB+) may serve as a predictive biomarker for ICB response in ccRCC. Finally, our results indicate patients with 14q−/EDNRB− may respond poorly to immunotherapy and new treatment options should be explored for this cohort.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Heterozygous loss of chromosome 14q deletion (14q−) occurs in 20–40% of renal clear cell carcinoma (ccRCC) and is associated with poor prognosis. The molecular consequences of 14q loss and its effect on immunotherapy outcomes are unknown.

WHAT THIS STUDY ADDS

• We demonstrate 14q−deleted ccRCC is associated with a distinct inflammatory profile defined by increased nuclear factor kappa-light-chain-enhancer of activated B cells signaling, increased CD8+T cell infiltration, and unique determinants of immune checkpoint blockade response (EDNRB expression). This study highlights the potential of genomic stratification of ccRCC, potentially via 14q deletion and EDNRB status.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• With additional study, 14q deletion and EDNRB status may serve as predictive biomarkers for immunotherapy success in ccRCC.

Background

14q deletion (heterozygous loss of chromosome 14q, or 14q−) has previously been identified as a negative prognostic indicator for clear cell renal cell carcinoma (ccRCC) prior to the introduction of checkpoint immunotherapy.^{1 2} Clinically, 14q− tumors demonstrate rapid tumor growth, high metastatic potential, and poor prognosis.^{1 2} The prevalence of 14q− in ccRCC is 20–40%.³ Despite the high number of patients harboring these alterations, few studies have explored molecular consequences or associated clinical characteristics of 14q− after the broad adoption of checkpoint immunotherapy in genitourinary oncology.

Importantly, two genes essential for inhibiting nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) are located on chromosome 14q: TRAF3 and NFΚBIA. NF-κB is a group of five transcription factors that mediate the expression of various cellular processes. Activation of NF-κB in tumor cells can cause tumor cell proliferation, survival, epithelial-mesenchymal transition, and resultant distant metastasis.^{4 5} On the other hand, NF-κB mediates the expression of various chemokines and their receptors; thus, tumors with high activation of NF-κB often have a high degree of immune cell infiltration.⁵ We hypothesize that co-deletion of key NF-κB inhibitors, TRAF3 and NFΚBIA, in 14q− tumors results in uncontrolled NF-κB activation leading to the poor prognosis previously observed in 14q−deleted ccRCC, as well as increased immune infiltration (figure 1a). We believe this unique biology results in unique determinants of immune checkpoint blockade (ICB) response (figure 1a).

Figure 1. (A) Our proposed model of 14q deleted biology. 14q deletion is associated with NF-κB activation, increased inflammation, aggressive tumor features, and finally better response to immunotherapy when EDNRB expression is high. Created in https://BioRender.com. (B) We compiled many different data sources for different purposes in this paper, including DepMap, scRNA-seq, OSU retrospective cohort, TCGA, CheckMate 009/010/025 trials, and the Hugaboom et al study. The transparent circles represent data groupings for analysis, and solid boxes of the same color indicate the theme of that analysis. (C) Summary of data resources and associated data types used throughout the analysis. 14q−, chromosome 14q deletion; ICB, immune checkpoint blockade; mIF, multiplex immunofluorescence; NF-κB, nuclear factor kappa-light-chain-enhancer of activated B cells; OSU, our institution; RNA-seq, RNA sequencing; RPPA, reverse-phase protein array; scRNA-seq, single-cell RNA sequencing; TCGA, The Cancer Genome Atlas; WES, whole exome sequencing.

Nearly all patients with ccRCC receive ICB therapy. ICB options include programmed cell death protein 1/programmed cell death ligand 1 (PD-1/PD-L1) checkpoint blockade or combination PD-L1/PD-1 and cytotoxic T cell associated antigen 4 (CTLA-4) blockade.⁶ The objective response rate to checkpoint immunotherapy in ccRCC is 42% for combined ipilimumab plus nivolumab in the front line,⁷ and 25–30% for single agent ICB in subsequent lines.⁸ None of the typical biomarkers for immunotherapy success (tumor mutational burden, microsatellite instability high status, CD8+T cell abundance, and PD-L1 immunohistochemistry) have proven useful for predicting response in patients with metastatic ccRCC.⁹ In the adjuvant, or post-surgery setting, PD-L1 positivity and sarcomatoid features weakly predict ICB benefit (HRs 0.72 and 0.59, respectively).¹⁰ Despite these findings, more work is needed to identify precision biomarkers with a stronger signal in the metastatic setting.

In this work, we collected data from 825 patients with ccRCC from publicly available sources and our own independent retrospective study of multi-fluorescent immunohistochemistry (figure 1b,c, online supplemental table 1). Combining these data using artificial intelligence and multi-omic analysis, we discovered 14q− is associated with increased NF-κB signaling, increased CD8+T cell infiltration, and unique determinants of ICB response (ie, high expression of endothelin receptor type B: EDNRB). Our analysis is the first to reveal these unique molecular and clinical features associated with 14q− deletion. This work introduces new opportunities to personalize the treatment of ccRCC, particularly for 14q−/EDNRB− patients who do not respond to current standard of care therapy according to our analysis.

Methods

Bulk tumor sequencing data acquisition and pre-processing

We performed analyses on the Ascend cluster using RStudio V.4.4.0 within the Ohio Supercomputer Center (https://www.osc.edu/). We downloaded RNA sequencing (RNA-seq) count data, single nucleotide variant (SNV) data, and copy number variant (CNV) data from The Cancer Genome Atlas (TCGA) using GenomicDataCommons¹¹ and GDC client and batch corrected RNA-seq count data with ComBat-seq from R package sva¹² using institution of origin as the batch. We also downloaded RNA .bam files for analysis with trust4¹³ with GDC client. We downloaded genomic variant calling data in .maf files. We downloaded copy number alteration data in gene-based raw copy number. Chromosomal arm deletions were defined as continuous copy number deletions across ≥95% of the genes in a single chromosomal arm. We obtained whole-exome sequencing (WES) copy number results from the Oncology Research Information Exchange Network for our multiplex immunofluorescence (mIF) imaged samples. We directly determined 14q status from these results. Next, we downloaded transcripts per million (TPM) data and associated clinical characteristics from the CheckMate 009/010/025 trials from Braun et al’s¹⁴ supplemental data section. Finally, we downloaded sequencing data from Hugaboom et al¹⁵ from dbGaP accession number phs003618.v1.p1 using the SRA toolkit. We processed Hugaboom et al’s data using the nf-core’s rnaseq^16–24 and sarek.^{16 17 19 20 23–28} We also used VarScan2²⁹ as a secondary copy number caller to CNVkit²⁷ (included with sarek).^{16 17 19 20 23–28} Visualizations were performed in R with ggplot2,³⁰ ggpubr,³¹ maftools,³² and GenVisR.³³ All data were aligned to Hg38.

Single-cell RNA sequencing data acquisition and pre-processing

Single-cell RNA sequencing (scRNA-seq) from six studies^{15 34–38} were downloaded from the Gene Expression Omnibus in the form of .txt files, .rdata files, or .h5 files. Each study was loaded into R V.4.4.0 within Ohio Supercomputer Center and preprocessed. The data were analyzed using Seurat.^39–43 Quality control measures were performed as follows: remove cells with <5 × the SD below median feature count, >5 × the SD above median feature count, <5 × median total count, and <10% mitochondrial gene expression. We performed quality control steps for each sample individually. To mitigate experimental batch effect, we used harmony⁴⁴ and clustered using clustree.⁴⁵ We developed individual machine learning classifiers for each cell type in Li et al³⁸ using scPred.⁴⁶ For cell types with under 200 cells available, no classifiers were built. For cell types with over 1,000 cells available, a random sample of 1,000 cells were selected. We determined 72 potential models available from caret⁴⁷ could be appropriate for each cell type. We ran scPred⁴⁶ trainModels function using each model for each cell type. We then selected models with a minimum accuracy and specificity of 0.8 for each cell type. For cell types with multiple appropriate models after filtering for accuracy, we selected the model with the highest possible sensitivity and speed. Finally, we trained a final reference object using the best model for each cell type (defined above) (online supplemental table 2). We then normalized, scaled, and predicted the cell type for each of the other studies using this reference object. After this initial cell typing step, we clustered each object and used nearest neighbors to classify cells which remained unassigned. The most frequent neighboring cell type was selected. In the case of ties, if there was a common parent, we used “parent_unk” (ie,“CD8+T_unk”). If there was not a common parent, the cell remained unassigned and was ultimately removed from future analysis. We removed ~1% of total cells. Finally, we used CopyKat⁴⁸ to determine chromosomal status. We created three controls for each study: kidney cells (non-cancerous), B cells, and myeloid cells. We performed CopyKat⁴⁸ prediction for the tumor cells using each control. For each malignant cell, we took the mean CopyKat⁴⁸ score for each gene in the 14q or chromosome 3p deletion (3p−) region. When greater than 60% of the malignant cells in each sample had a mean CopyKat⁴⁸ score below −0.50 across either 14q or 3p, we determined that the tumor contains a chromosomal arm deletion at that locus. Visualizations were performed with ggpubr³¹ and ggplot2.³⁰

DepMap analysis

Chromosomal arm deletions and protein array data were downloaded from the DepMap database. Any cell line with a deletion or amplification in chromosome 4q was excluded due to aberrant expression of NF-κB itself. 3p and 14q status was determined with the chromosomal arm deletion information. Reverse-phase protein array (RPPA) data for phospho-p65 were extracted for each cell line and visualized. Comparison was made with a two-sided Wilcox test. All visualization was performed using ggpubr³¹ and ggplot2.³⁰

Malignant cells scRNA-seq analysis

We filtered each scRNA-seq project^{15 34–38} for just the malignant cells and combined them together mitigating batch effect using harmony. Then, we used a Wilcox test to determine if the percentage of cycling tumor cells was broadly increased in 14q− tumors. We used FindAllMarkers from Seurat^39–43 to compare all tumor cells across genotypes and to compare all tumor cell subtypes. Gene level information for NF-kB complex members and gene targets was extracted for visualization. Only significant targets (adjusted p value<0.05) were used for dendrograms. Top markers for cycling tumor cells were extracted from Li et al³⁸ and their differential expression information extracted for visualization. All visualization was performed using ggpubr,³¹ ggplot2,³⁰ and Seurat.^39–43

Cycling cells bulk RNA-seq analysis

TCGA bulk RNA-seq were extracted and pre-processed as described. DESeq2⁴⁹ was performed across genetic subsets (WT, 14q−, and 3p−) to generate differential expression statistics. The same genes as in the scRNA-seq malignant cell analysis for cycling tumor cells were extracted and visualized. All visualization was performed using ggpubr³¹ and ggplot2.³⁰

Immune microenvironment scRNA-seq analysis

We first explored the broad cell type composition of 14q− tumors and found them to be enriched with CD8+T cells by two proportion z test. Then we determined this increase was mainly composed of effector CD8+T cells using a Wilcox test. Finally, we used FindAllMarkers from Seurat^39–43 to determine the expression of key inflammatory markers and checkpoint molecules in CD8+T cell subtypes. All visualization was performed using ggpubr,³¹ ggplot2,³⁰ and Seurat.^39–43

Immune microenvironment bulk RNA-seq analysis

Using the quantiseqR⁵⁰ R package, we determined the estimated abundance of CD8+T cells in 14q− tumors. We also analyzed the RNA-seq bams using Trust4¹³ and default settings. T-cell receptor clonality and CD3Raa statistics were compiled using the RIMA_pipeline.⁵¹ All visualization was performed using ggpubr³¹ and ggplot2.³⁰

Multiplex immunofluorescence assays

Tumor sections were stained with two 7-plex immunofluorescence (IF) panels. The IF assays were optimized and performed on a Bond RX Autostainer (Leica Biosystems) using the Perkin Elmer Opal tyramide signal system. The T-cell panel included CD3, CD8, PD1, FoxP3, GzmB, T-bet and DAPI whereas the myeloid panel included CD68, CD86, CD163, CD11b, CD11c, panCK and DAPI. The performance of antibodies was validated on non-neoplastic tonsil tissue to obtain a staining pattern in accordance with available published data. The anti-GzmB antibody (1:500; clone EPR8260, rabbit mAb; Abcam; catalog no. ab134933) was detected using Opal Polaris 480 fluorophore (1:100); the anti-PD-1 antibody (1:1,000; clone EPR4877(2), rabbit mAb; Abcam; catalog no. ab137132) was detected using the Opal Polaris 690 fluorophore (1:100); the anti-Tbet antibody (1:1,500; clone D6N8B, rabbit mAb; Cell Signaling Technology; catalog no. 13232) was detected using Opal Polaris 520 fluorophore (1:100); the anti-CD3 antibody (1:150; clone SP7, rabbit mAb; Abcam, catalog no. ab16669) was detected using Opal Polaris 620 fluorophore (1:100), the anti-FoxP3 antibody (1:500; clone 236A/E7; mouse mAb; Abcam; catalog no. ab20034) was detected using Opal Polaris 570 fluorophore (1:100), and the anti-CD8 antibody (1:200; clone CD8/144B; mouse mAb; Biocare Medical; catalog no. ACI 3160 A, C) was detected using Opal Polaris 780 (1:25). The anti-CD68 antibody (1:100; clone KP1, mouse mAb; Biocare Medical; catalog no. CM 033 A, B, C) was detected using Opal Polaris 480 fluorophore (1:100), the anti-CD163 antibody (1:100; clone 10D6, mouse mAb; Biocare Medical; catalog no. CM353C/A) was detected using the Opal Polaris 690 fluorophore (1:100), the anti-CD86 antibody (1:400; clone E2G8P, rabbit mAb; Cell Signaling Technology; catalog no. 91882) was detected using Opal Polaris 520 fluorophore (1:100), the anti-CD11b antibody (1:1,000; clone EP1345Y, rabbit mAb, Abcam, catalog no. ab52478) was detected using Opal Polaris 620 fluorophore (1:100), the anti-CD11c antibody (1:75; clone Leu-M5; mouse mAb; Biocare Medical; catalog no. ACI 3122 A, B) was detected using Opal Polaris 570 fluorophore (1:100), and the anti-panCK antibody (1:100; clone AE1/AE3; mouse mAb; Biocare Medical; catalog no. CM011C/B) was detected using Opal Polaris 780 (1:25). Whole-slide multispectral images were acquired at 20× magnification using the PhenoImager HT automated quantitative pathology imaging system (Akoya Biosciences).

Multiplex immunofluorescence computational preprocessing

Images were processed with the nextflow pipeline MCMICRO.⁵² Our implementation of MCMICRO begins at the segmentation step. S3segmenter⁵² and Cellpose⁵³ were used to segment individual cell and tissue areas using computer vision. Next, cell-level quantification of marker intensity was performed using MCquant.⁵² Following this, cells were clustered with Leiden clustering based on marker intensity and clusters were identified manually with scimap.⁵⁴ Cells which could not be confidently cell-typed were removed (up to 12% of cells in some cases). Spatial distance, spatial interaction, and spatial aggregate scores were calculated for each image. Next, spatial proximity scores between all cell types identified were calculated. All these measures were saved for visualization in R.

Multiplex immunofluorescence for CD8+T cell characteristics

Spatial interaction scores from scimap⁵⁴ were extracted and analyzed in R. A t-test was performed to assess the differences between groups. Visualization was performed using ggpubr³¹ and ggplot2.³⁰

Multiplex immunofluorescence for dendritic cell characteristics

Proximity density from scimap⁵⁴ was extracted and analyzed in R. Distribution of proximity density scores for all cell types was visualized, and significantly increased dendritic cell (DC) interactions with tumor cells were noted in responder tissue. Visualization was performed using ggpubr³¹ and ggplot2.³⁰

Cell–cell communication network analysis

Single-cell data were processed according to the methods described above. First, single-cell data from responders and non-responders were loaded and filtered for just CD8+T cells and DCs. Scriabin⁵⁵ was used to quantify the number of interactions and types of interactions between tumor cells and DCs. Wilcox tests were used for each interaction type to determine if there was a difference between the responder and non-responder interaction intensities. It was also recorded when one group was the only one to display a certain interaction. Costimulatory and coinhibitory interactions between CD8+T cells and DCs were visualized. Next, this process was repeated for DCs and tumor cells. The quantity of DCs was recorded and visualized. Expression of EDNRB and EDN1 were assessed with the Seurat^39–43 function FindAllMarkers. All graphical visualizations were created with ggpubr³¹ and ggplot2.³⁰

Immunotherapy survival analysis

14q deletion was detected in CheckMate 009/010/025¹⁴ trial RNA-seq using a machine learning model developed in TCGA using DNA-level 14q deletion annotations. First, predictor genes were selected in an unbiased manner using ElasticNet⁵⁶ machine learning with the R package glmnet.⁵⁷ Next, an xgboost⁵⁸ prediction model was built and Bayesian hyperparameter optimization was performed with the R packages xgboost⁵⁸ and ParBayesianOptimization.⁵⁹ Accuracy with the final model was 89% (95% CI 0.81% to 0.95). This final model was used to predict 14q status in CheckMate 009/010/025.¹⁴ Model statistics can be found in online supplemental tables 3 and 4. EDNRB expression was extracted directly and a cut-off was selected at the near quartile with surv_cutpoint from the R package survminer.⁶⁰ Data from Hugaboom et al¹⁵ included WES and bulk RNA-seq. This was processed as described above. EDNRB expression was extracted from TPM data. A cut-off was also selected using surv_cutpoint near the first quartile. Survival statistics were calculated with survfit and Surv from the R package survival.⁶¹ All results were visualized with the R package survminer.⁶⁰

Results

14q− co-occurs with 3p deletion in ccRCC across six data sources

To describe the genetic landscape of 14q− ccRCC, we used data from TCGA and six scRNAseq studies.^{15 34–38} We visualized the copy number landscape of chromosome 14 along with other chromosomes with known chromosomal alterations in ccRCC: 3, 5, 8, and 9. The heterozygous deletion of 14q co-occurs with 3p deletion (3p−) 72% (n=71/98) of the time in TCGA (figure 2a). In the scRNA-seq data sets, we found 3p− always co-occurred with 14q− (figure 2b). Due to this frequent co-occurrence of 14q− and 3p−, we included both double wild type (3pWT/14WT) and 3p− only samples as controls for subsequent analyses. There was no other ubiquitous chromosomal abnormality in 14q− patients. The complete chromosomal abnormality landscape in TCGA is included in online supplemental figure 1a). We also investigated tumor mutational burden (TMB) in TCGA. There was no difference in TMB across 3pWT/14qWT (WT), 14q−, and 3p− samples (online supplemental figure 1b). Most scRNA-seq samples are primary kidney samples and treatment naïve (figure 2b). Several oncogenes and tumor suppressor genes are located on chromosome 14q, visualized in figure 2c, including NF-κB-related genes NFKBIA and TRAF3. No other mutations in NF-κB-related proteins are prevalent among 14q− samples (online supplemental figure 1c-e).

Figure 2. (A) Copy number landscape across chromosomes 3, 5, 8, 9, and 14 in TCGA-KIRC. The y axis shows the percentage of all samples with copy number loss (<1 copy, shown in blue) and copy number gain (>4 copies, shown in red). There are sample cohorts with no continuous deletion of either chromosome 14 or 3 (wild type, top, n=262), 14q deletion (14q−, middle, n=98), only 3p deletion (3p−, bottom, n=177). (B) CopyKat scores and clinical features of 51 scRNA-seq patients across six studies. In the top panel, average CopyKat scores (an approximation of copy number for scRNA-seq experiments) are shown for chromosomal arms 14q and 3p: darker blue color indicates a higher degree of deletion and darker red color indicates a higher level of amplification. The bottom panel represents the clinical characteristics of treatment status, sample location, and scRNA-seq project for each patient. (C) Here we show a schematic of the locations of different oncogenes and tumor suppressor genes identified from OncoKB⁸² which are located on chromosome 14q. Green bars represent locations only and do not represent gene area. 3p−, chromosome 3p deletion; 14q−, chromosome 14q deletion; ccRCC, clear cell renal cell carcinoma; KIRC, kidney renal clear cell carcinoma; scRNA-seq, single-cell RNA sequencing; TCGA, The Cancer Genome Atlas.

NF-κB activation and transcription are increased in 14q− tumors

NFΚBIA and TRAF3 are key inhibitors of NF-κB signaling located on chromosome 14q (figure 2c).^{4 5} IκB (encoded by NFKBIA) and TRAF3 must be degraded for transcription of NF-κB target genes to occur (figure 3a).^{4 5} Logically, it follows that deletion of these proteins will result in increased target gene transcription and NF-κB pathway activation. To investigate this at the protein level, we leveraged RPPA from DepMap to determine if NF-κB proteins are differentially activated in 14q− cell lines. We used available WES data to determine 14q and 3p status for 271 cell lines. Analysis of RPPA outputs from these cell lines revealed increased phospho-p65 protein abundance in 14q− cell lines indicating increased NF-κB pathway activation (figure 3b). The sample size for kidney cancer cell lines alone was too small (WT: n=5, 14q−: n=11, 3p−: n=2), so cell lines from all cancer types were considered (WT: n=173, 14q−: n=44, 3p−: n=60) (online supplemental figure 2). In addition to protein activation, we also wanted to assess the expression of key NF-κB transcription factor and regulatory genes. Because NF-κB is expressed and differentially activated in immune cells, scRNA-seq analysis of ccRCC tumor cells alone was used to assess the impact of 14q− on expression of NF-κB pathway genes and transcription factor targets. Cell typing was performed using artificial intelligence models built in R with scPred.⁴⁶ Relevant statistics can be found in online supplemental table 2. As expected, expression of TRAF3 and NFΚBIA was decreased in 14q− tumor cells compared with WT or 3p− only (figure 3c). Expression of additional key NF-κB pathway genes (RELA, RELB, NFΚB2, IΚBΚB, and CHUK) was increased in 14q− tumor cells (figure 3c). To examine the downstream impact of 14q− on NF-κB transcription factor function, we assessed 77 NF-κB transcriptional targets in scRNA-seq of tumor cells. Expression of NF-κB targets is increased in 14q− tumor cells compared with controls (figure 3d). Next, we investigated which tumor cell type, defined in Li et al,³⁸ is most common in 14q− tumors. We found cycling tumor cells (a known feature of NF-κB activation) were most prevalent in 14q− tumors (figure 3e,f). We explored which transcriptomic features are associated with the cycling tumor cell subtype and found increased expression of key cancer proliferation genes UBE2C,⁶² TOP2A,⁶³ PI3,⁶⁴ KIF20A,⁶⁵ and CEP55⁶⁶ (figure 3d). To confirm this finding of increased proliferation in 14q− tumors from TCGA, we investigated the expression of these same genes and found significantly increased expression in 14q− tumors compared with controls (figure 3h).

Figure 3. (A) IκB (NFΚBIA) and TRAF3 must be degraded for canonical and non-canonical NF-κB target genes to be activated. Created in https://BioRender.com. (B) RPPA shows phospho-p65, an important indicator of NF-κB activation, was increased in 14q− cell lines in DepMap pan-cancer (WT: n=173, 14q−: n=44, 3p−: n=60). (C) Expression of NF-κB genes RELA, RELB, NFΚB2, IΚBΚB, CHUK, TRAF3, and NFΚBIA in WT (n=5), 14q− (n=15), and 3p− (n=23) ccRCC tumor cells from scRNA-seq. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. (D) Dendrograms of NF-κB target expression in WT (n=5), 14q− (n=15), and 3p− (n=23) ccRCC tumor cells. Gene count is shown on the y axis and log fold change is shown on the x axis. The purple dendrogram (genes’ expressions from 14q− tumor cells) shows a rightward shift indicating broadly increased expression of NF-κB target genes. (E) Composition of all ccRCC tumor cell subtypes by sample is shown in the top panel. Individual percentages of the cycling subtype are shown in the bottom panel of green dots. Size of the dot indicated the percentage of cycling tumor cells for that sample. (F) Composition of the cycling tumor cell subtype in WT (n=5), 14q− (n=15), and 3p− (n=23). (G) Expression of key genes related to cancer cell proliferation in tumor subtypes cells from scRNA-seq. Cycling tumor cells have the highest expression. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. (H) Expression of the same genes in G in bulk RNA-seq from the TCGA database across WT (n=262), 14q− (n=98), and 3p− (n=177) TCGA-KIRC tumors. 14q− tumors display more cycling phenotypes than controls. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. Significance values: *p≤0.05, **p≤0.01, ***p≤0.001, ****p≤0.0001. 3p−, chromosome 3p deletion; 14q−, chromosome 14q deletion; ccRCC, clear cell renal cell carcinoma; KIRC, kidney renal clear cell carcinoma; NF-κB, nuclear factor kappa-light-chain-enhancer of activated B cells; RPPA, reverse-phase protein array; scRNA-seq, single-cell RNA sequencing; TCGA, The Cancer Genome Atlas; WT, wild-type.

CD8+ T cells are enriched in the tumor immune microenvironment of 14q− tumors

NF-κB activation is known to cause increased immune infiltration via increased expression of chemotactic molecules.⁵ We investigated whether this was the case in 14q− tumors. scRNA-seq demonstrated the expression of key lymphocyte chemotaxis mediators regulated by NF-κB (ICAM1,⁶⁷ CXCL16,⁶⁸ and CCL20⁶⁹) were increased in 14q− tumors (figure 4a). We then investigated the composition of the immune microenvironment in WT, 14q−, and 3p− tumors (excluding ICB treated samples) using cell-typing enabled by artificial intelligence tool scPred⁴⁶ (figure 4b). We observed increased CD8+T cells in 14q− tumors (figure 4b). Next, we wanted to determine the type of CD8+T cell which was predominantly enriched. Effector CD8+T cells were primarily enriched in 14q− samples (figure 4c). Using the definition from Li et al, we determined effector CD8+T cells using many genes’ expressions including RHOB, NR4A1, TNF, INFG, PDCD1, LAG3, and CD244 (figure 4d).³⁸ Our scRNA-seq analysis here is limited by a sample size of just 34 patients (only two WT patients), so we investigated bulk RNA-seq from the TCGA database to confirm our findings. TCGA analysis showed increased CD8+T cell infiltration from quantiseqR⁵⁰ (figure 4e) and increased number of unique CDR3 amino acid sequences, thus indicating increased CD8+T cell infiltration in 14q− tumors (figure 4f), and increased T-cell clonality (figure 4g) detected by Trust4.¹³ In addition to increased T-cell abundance, we wanted to determine whether CD8+T cells were interacting with tumor cells more often in 14q− tumors. To do this, we used a retrospective cohort from our institution with paired mIF and genomics. Chromosomal landscapes of each group (WT, 14q−, and 3p−) can be found in online supplemental figure 3. We investigated a T-cell panel of markers (FoxP3, GzmB, CD3, CD8, PD-1, and T-bet) using artificial intelligence software in the MCMICRO pipeline⁵² and scimap⁵⁴ to identify cell types and their spatial patterns from images. Processed images, marker clustering, and cell type definitions can be found in online supplemental figure 4. Regulatory T cells showed high positivity for FoxP3 and CD3. All CD8+T cells showed high expression of CD8 and CD3. Activated CD8+T cells also showed high GzmB positivity, exhausted CD8+T cells showed high positivity for PD-1, and naïve CD8+T cells showed little positivity for both PD-1 and GzmB. Lymphocytes show high CD3 positivity, but no positivity for CD8 or FoxP3. Finally, the category “Other Immune” was created for cells showing high T-bet positivity, but no positivity for other markers. Cells with low positivity for all markers were categorized “Tumor and Stroma”. Representative output images are shown in figure 4h. mIF images of immunotherapy responders show that 14q deleted tumors have increased interactions between CD8+T cells and tumor/stromal cells compared with other genotypes (figure 4i).

Figure 4. (A) scRNA-seq expression of key chemotactic genes regulated by NF-κB: ICAM1, CXCL16, and CCL20 in tumor cells from WT (n=5), 14q− (n=15), and 3p− (n=23) ccRCC tumors. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. (B) Composition of all immune subtypes by sample is shown in the top panel. Individual percentages of CD8+T cells are shown in the bottom panel of dark blue dots. Size of the dot indicated the percentage of CD8+T cells for that sample. (C) Composition of effector CD8+T cells in ICB-naïve WT (n=2), 14q− (n=11), and 3p− (n=21) tumors. (D) Expression of key genes (RHOB, NR4A1, TNF, INFG, PDCD1, LAG3, and CD244) used to define effector CD8+T cells in scRNA-seq. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. (E) Estimated CD8+T cell abundance by quantiseqR in WT (n=262), 14q− (n=98), and 3p− (n=177) tumors from TCGA-KIRC. (F) Number of unique CD3Raa sequences (indicating unique CD8+T cells) in WT (n=262), 14q− (n=98), and 3p− (n=177) tumors from TCGA-KIRC. (G) TCR clonality (indicating proliferating CD8+T cells) in WT (n=262), 14q− (n=98), and 3p− (n=177) tumors from TCGA-KIRC. (H) Representative images of processed, T-cell panel mIF for WT (n=5), 14q− (n=5), and 3p− (n=10) tumors. (I) Spatial interaction scores of CD8+T cells with tumor/stomal cells in ICB responder WT (n=3), 14q− (n=3), and 3p− (n=5) tumors. Significance values: *p≤0.05, **p≤0.01, ***p≤0.001, ****p≤0.0001. 3p−, chromosome 3p deletion; 14q−, chromosome 14q deletion; ccRCC, clear cell renal cell carcinoma; ICB, immune checkpoint blockade; KIRC, kidney renal clear cell carcinoma; mIF, multiplex immunofluorescence; NF-κB, nuclear factor kappa-light-chain-enhancer of activated B cells; OSU, our institution; scRNA-seq, single-cell RNA sequencing; TCGA, The Cancer Genome Atlas; TCR, T-cell receptor; Treg, regulatory T cell; WT, wild-type.

Dendritic cell interactions with tumor cells are associated with response in 14q− patients

Despite more infiltration of CD8+T cells without ICB treatment, 14q− patients have been shown to have a worse overall prognosis.^{1 2} Furthermore, neither CD8+T cell abundance nor location has proven predictive of immunotherapy success in ccRCC.⁹ Therefore, we sought unique determinants of ICB response within 14q− patients. We repeated mIF and artificial intelligence processing of resultant images for a myeloid cell panel (CD68, CD86, CD11c, CD11b, CD163, and panCK). Processed images, marker clustering, and cell type definitions can be found in online supplemental figure 5. DCs were defined as cells with high positivity for CD11b and CD11c. M1 macrophages were highly positive for CD86 and CD68. M2 macrophages, on the other hand, were highly positive for CD163. Mixed polarized macrophages showed high positivity for CD86, CD68, and CD163. Tumor cells were strongly positive for panCK. The “Other” category was created for cells negative for all these markers. Representative images for each genotype, responders and non-responders, are shown in the left panel of figure 5a–c. 14q− responders displayed uniquely and significantly increased proximity density between DCs and tumor cells (figure 5e).

Figure 5. (A–C) Representative images of processed, myeloid cell panel mIF for WT, 14q−, and 3p−, (respectively) responders (WT: n=3, 14q−: n=5, 3p−: n=5) (left) and non-responders (WT: n=2, 14q−: n=3, 3p−: n=5) (right). (D,E) Proximity density scores between dendritic cells and tumor cells across WT, 14q−, and 3p−, (respectively) responders in green (WT: n=3, 14q−: n=5, 3p−: n=5) and non-responders in red (WT: n=2, 14q−: n=3, 3p−: n=5) show significantly increased proximity in 14q− responders. Significance values: *p≤0.05, **p≤0.01, ***p≤0.001, ****p≤0.0001. 3p−, chromosome 3p deletion; 14q−, chromosome 14q deletion; DC, dendritic cell; ICB, immune checkpoint blockade; mIF, multiplex immunofluorescence; OSU, our institution; WT, wild-type.

Dendritic cell features predict ICB response in 14q− patients

First, we investigated the types of interactions between DCs and CD8+T cells in scRNA-seq of five 14q− responders and three 14q− non-responders. We found increased interaction intensity across many costimulatory and coinhibitory interactions between DCs and CD8+T cells across the responder and non-responder cohorts. Significantly different interactions can be found in figure 6a. Examining interactions between DCs and tumor cells, however, one interaction stuck out to us: EDN1 on tumor cells interacting with EDNRB on DCs. We explored sample-level interaction intensity and proportion in figure 6b. We found 14q− responders showed increased proportion of EDN1-EDNRB interactions than 14q− non-responders (two proportion z-test p value<2.2E−16). Next, we saw 14q− responders showed significantly increased EDN1-EDNRB interaction intensity (figure 6c). These findings were not explained by overall DC infiltration which was qualitatively similar across responders and non-responders (figure 6d). We then explored whether general expression of EDN1 or EDNRB could be used as a surrogate measure of this interaction. We saw DCs from 14q− responders expressed significantly more EDNRB than DCs from 14q− non-responders (figure 6e). Meanwhile, endothelial cells and malignant cells from 14q− non-responders displayed higher expression of EDNRB than respective cell types in responder tumors (figure 6e). We also compared EDN1 expression across cell types. EDN1 was not significantly different in DCs or in endothelial cells across ICB response. EDN1 expression was increased in malignant cells from 14q− responders compared with tumor cells from non-responders (figure 6e). Based on these findings, we hypothesized that EDNRB or EDN1 could stratify 14q− patients’ immunotherapy response.

Figure 6. (A) Cell–cell communication scores for key co-stimulatory and co-inhibitory interactions between dendritic cells and CD8+T cells in a 14q− NR versus 14q R. B Sample-level EDN1-EDNRB interaction proportion (across all interactions between tumor cells and dendritic cells within that sample) represented by dot size and mean intensity represented by dot color (higher intensity shown in red and lower intensity in blue) as detected by Scriabin. (C) Interaction intensity of EDN1-EDNRB detected by Scriabin within NR and R 14q− tumors and DCs. (D) Abundances (shown in dots) and composition (shown as a stacked barplot) of dendritic cell types in 14q non-responders and responders are similar according to scRNA-seq. (E) Expression of EDN1 and EDNRB in dendritic cells, endothelial cells, and malignant cells across 14q− non-responders and responders. Higher log fold change is indicated by dot color (increased in red and decreased in blue) and lower p value is indicated by dot size. (F) Kaplan-Meier curves with accompanying table for 14q−/EDNRB+ (n=36), 14q−/EDNRB− (n=25), 14qWT/EDNRB+ (n=90), and 14qWT/EDNRB− (n=30) patients in the CheckMate 009/010/025 trial treated with nivolumab. (G) Kaplan-Meier curves with accompanying table for 14q−/EDNRB+ (n=26) and 14q−/EDNRB− (n=6) patients in the Hugaboom et al study who received nivolumab first line. Significance values: *p≤0.05, **p≤0.01, ***p≤0.001, ****p≤0.0001. 3p−, chromosome 3p deletion; 14q−, chromosome 14q deletion; ccRCC, clear cell renal cell carcinoma; DC, dendritic cell; NR, non-responders; R, responders; scRNA-seq, single-cell RNA sequencing; WT, wild-type.

To determine whether 14q status and EDNRB expression could determine ICB response, we investigated the CheckMate 009/010/025 immunotherapy trials.¹⁴ We determined 14q status via our own 14q detection machine learning algorithm. Relevant statistics can be found in online supplemental tables 3 and 4. In the CheckMate 009/010/025 trials for PD-1 inhibitor nivolumab, 14q− patients who also have low EDNRB expression experienced worsened PFS survival (HR: 3.27, 97% CI 1.75 to 6.08) over 14q intact and 14q deleted EDNRB higher patients (figure 6f).¹⁴ We compared 14q status and EDNRB expression with other immunotherapy predictive measures. We investigated the predictive performance of several other markers of interest in CheckMate 009/010/025: 14q deletion on its own, EDN1 expression, TMB, PD-L1 (CD274) expression, CD8A expression, and CD83 expression (online supplemental figure 6). None of these markers achieved the predictive power of 14q status and EDNRB expression (online supplemental figure 6).

We also investigated a second cohort of patients from a separate trial of nivolumab in ccRCC.¹⁵ This cohort included WES, so we directly determined 14q status. In this study, 14q− patients comprised >80% of the total ccRCC cohort; thus, we focused our analysis only on 14q− patients due to sample size constraints. We found that, again, 14q−/EDNRB− patients display worsened progression-free survival (HR: 2.61, 97% CI 1.00 to 6.81) compared with 14q−/EDNRB+ (figure 6g).

Discussion

14q deletion was previously identified as a negative prognostic indicator in ccRCC before the introduction of ICB therapy.^{1 2} Clinically, 14q−deleted (14q−) tumors demonstrate rapid tumor growth, high metastatic potential, and poor prognosis without ICB treatment.^{1 2} The prevalence of 14q deletions in ccRCC is significant (20–40%),³ but few studies have explored 14q deletion’s molecular consequences or associated treatment outcomes in the immunotherapy era. In this study, we are the first group to connect NF-κB activation with 14q− (figure 3), to show higher antitumor CD8+T cell infiltration in 14q− tumors (figure 4), and finally to predict response to immunotherapy using EDNRB expression among 14q− patients (figure 6). We also identified a cohort of 14q−/EDNRB− patients who do not respond well to standard-of-care, ICB-based treatments in ccRCC and may benefit from new treatment paradigms (figure 6). These results may lead to the personalization of kidney cancer treatment.

First, we explored the impact of 14q deletion on tumor cell biology. Because TRAF3 and NFKBIA, genes encoding NF-κB inhibitors, are located on chromosome 14q, we hypothesized that 14q− deletion would result in increased NF-κB activation, due to lack of inhibitory proteins.^{4 5} NFKBIA is a key inhibitor of the canonical NF-κB pathway (figure 3a).⁵ In ccRCC, overactivation of the canonical NF-κB pathway has been implicated in tumor progression.⁷⁰ On the other hand, TRAF3 inhibits the non-canonical pathway (figure 3a). Less is known about the impact of non-canonical NF-κB signaling in ccRCC, but loss-of-function TRAF3 alterations are known to induce non-canonical NF-κB activation in other cancers, resulting in increased tumor growth.⁴ We investigated the expression of key NF-κB genes in both the canonical and non-canonical pathways within 14q− tumor cells. We found broadly increased expression of both NF-κB complex members and gene targets in 14q− tumor cells compared with controls (figure 3c–d). These results indicate increased activity of both the canonical and non-canonical NF-κB pathways in 14q− tumor cells.

NF-κB has numerous effects on tumors. First, NF-κB activation can increase tumor cell growth.⁵ To assess whether tumor growth was impacted in 14q− tumors, we studied features of proliferating tumor cells in scRNA-seq and bulk RNA-seq from TCGA in 14q− tumors compared with WT and 3p− controls (figure 3e–h). 14q− tumors displayed more expression of proliferation markers than controls, as expected and previously reported in clinical trials^{1 2} (figure 3e–h). Second, NF-κB activation can have a positive impact on the tumor immune microenvironment (TIME) by increased expression of chemotactic genes.⁵ Since nearly all patients with ccRCC are treated with immunotherapy, we were curious whether NF-κB was impacting immune infiltration in 14q− tumors. In fact, we saw increased NF-κB-regulated, T cell, chemotactic gene expression (ICAM1,⁶⁷ CXCL16,⁶⁸ and CCL20⁶⁹) by 14q− tumor cells in scRNA-seq compared with controls (figure 4a). Together, these results suggest 14q− tumor cells display phenotypic differences associated with increased NF-κB activity compared with controls.

Because chemotactic gene expression was altered, we explored the TIME of 14q− ccRCC in more detail. 14q− tumors were enriched in CD8+T cells compared with controls (figure 4b,c and e–g). Using mIF, we also showed that CD8+T cells from 14q− tumors interact more frequently with tumor or stroma cells (fibroblasts, endothelial cells, or non-malignant kidney cells) than in controls (figure 4i). Interactions between CD8+T cells and stromal cells are often immunosuppressive, while interactions with tumor cells can be exploited with ICB.⁷¹ Despite these findings, previous studies in ccRCC have demonstrated that CD8+T cell characteristics, such as abundance, location, and dispersion do not predict immunotherapy response in ccRCC.⁹ Nevertheless, we found significant differences in both the TIME and molecular landscape of 14q− tumors compared with WT and 3p− tumors. Based on these findings, we hypothesized that 14q− tumors may have unique determinants of ICB response. Therefore, we investigated mIF images visualizing myeloid cells and tumor cells from 14q− responders and non-responders. We observed that DCs are more proximal to tumor cells in 14q− responders compared with non-responders (figure 5e). The impact of DCs on immunotherapy response in ccRCC is an open area of research.

DCs are antigen presenting cells that contribute to both innate and adaptive immune responses.⁷² In tumors, DCs can impact tumor immunity via cross presentation of antigen to T cells or co-stimulatory/co-inhibitory cell–cell interactions.⁷² DCs can be affected by ICB, because PD-L1 and PD-1 are both expressed by DCs.⁷² Furthermore, another ICB target, CTLA-4, is specifically designed to increase the potency of cross presentation.⁷² Interestingly, our results show a higher number of co-stimulatory and co-inhibitory interactions between DCs and CD8+T cells in 14q− responder tumors than in 14q− non-responder tumors (figure 6a); however, the bulk expression or protein abundance of PD-L1, PD-1, or CTLA-4 do not predict ICB success in the literature⁹ nor in our own analysis (online supplemental figure 6). There have also been attempts to combine ICB therapy with DC vaccines with promising results.⁷³ DC vaccines prime naïve T cells for new immune responses, whereas ICB therapy augments existing immune responses.⁷⁴ Combination approaches are synergistic and in clinical trials in a variety of both solid and hematologic malignancies.⁷⁵ While these approaches are currently investigational, future work on the role of DCs in kidney cancer could shed light on such clinical trials. We investigated which factors may be impacting DC behavior to better understand how DCs may be impacting ICB response in 14q− patients.

We found EDN1 and EDNRB from tumor cells and DCs, respectively, were more frequently and intensely interacting in 14q− responder samples (figure 6b,c). We hypothesized that EDNRB expression coupled with 14q− could predict good response to ICB therapy. To prove this, we explored immunotherapy response of 14q− patients in CheckMate 009/010/025¹⁴ and Hugaboom et al¹⁵ immunotherapy trial data. In both datasets, we saw 14q−/EDNRB+ status positively predicts immunotherapy response. Importantly, in both studies, 14q−/EDNRB− patients benefited from immunotherapy the least (figure 6e,f). EDNRB encodes endothelin receptor type B with a variety of functions: release of relaxin factors, endothelin-1 clearance, and vascular smooth muscle contraction and cell growth.⁷⁶ The specific roles of EDNRB in ccRCC are unclear, but previous studies have recognized EDNRB expression as a possible prognostic indicator in the absence of immunotherapy.⁷⁷ The impact of high EDNRB on DCs is debated in the literature: some authors claim high EDNRB expression is an indicator of DC maturation,⁷⁸ whereas others claim ETb (encoded by EDNRB) dysfunction is associated with inflammation in the kidney.⁷⁹ These results are the first to include EDNRB as a marker for immunotherapy response in ccRCC.

Our study has several notable limitations. We are limited by low sample size in our exploration of determinants of ICB response in scRNA-seq data in 14q− patients. Additionally, we use CopyKat⁴⁸ to predict 14q− status in lieu of paired WES in our single-cell cohort and machine learning to predict 14q− in the CheckMate 009/010/025 trials. Although every effort was made to confirm results in cohorts with DNA-level data available (TCGA, mIF, and Hugaboom et al¹⁵), more comprehensive genetic characterization will be needed in future studies. CopyKat was not tested for its ability to predict 14q deletion status specifically, nor do we know how the clonality of 14q deletion may impact its detection in single-cell methods. The applicability of CopyKat and our own machine learning model to clinical systems is also unclear and untested in this manuscript. In the future, more comprehensive and consistent genomic profiling across all samples should be employed. Furthermore, our observations in 14q− tumors are descriptive in nature and could benefit from experimental validation such as functional experiments confirming NF-κB dysregulation and immune-promoting cytokine release following 14q loss, and perturbation or rescue experiments validating the proposed END1-EDNRB axes. Finally, the source of EDNRB transcript in the bulk data shown here is up for debate. EDNRB is expressed not only in DCs, but also in endothelial cells and smooth muscle cells. More work is needed to establish the cellular source of EDNRB transcripts in bulk RNA-seq. We are also aware of RNA signature scores for ICB response⁸⁰; however, these were developed for prediction of response to atezolizumab+bevacizumab and sunitinib, not nivolumab monotherapy shown here in figure 6e,f.

Aside from these limitations, our work raises several important questions for future study: (1) NF-κB inhibitors are being explored in several solid tumor types.⁵ Our results suggest that perhaps 14q− tumors may be sensitive to NF-κB inhibition, particularly 14q−/EDNRB− patients who experience the least benefit from current treatment options. (2) The impact of EDNRB on DCs in the TIME is poorly understood. More work is needed to unravel how high EDNRB expression is affecting DC behavior, perhaps to complement ongoing research on DC therapies,⁸¹ which may benefit 14q−/EDNRB− patients as well. (3) Although our results show promising data suggesting 14q−/EDNRB+ patients are exceptional responders to immunotherapy, more work is needed to prospectively confirm this result and to determine the best method of detection for this genomic subset of ccRCC. In the future, we hope to advance personalized immuno-oncology approaches for ccRCC through investigation of these three avenues and addressing the limitations explored above.

The results of our current study comprise the first comprehensive characterization of the genetic landscape, molecular associations, immune microenvironment features, and unique determinants of ICB response of 14q− in ccRCC. To our knowledge, this work is the first to suggest a connection between 14q−deletion, increased NF-κB activation, and increased CD8+T cell infiltration. Finally, we believe this work is the first to identify EDNRB expression, along with 14q status, and could predict immunotherapy response. This work advances personalized treatment for ccRCC.

The views expressed in the submitted article are our own and not an official position of The Ohio State University, Georgetown University, or the Pelotonia Foundation.

Data availability statement

Data are available in a public, open access repository. Data may be obtained from a third party and are not publicly available.