CDKN2A Alterations and Response to Immunotherapy in Solid Tumors

Immune checkpoint inhibition (ICI) using anti–PD-1/PD-L1 or anti–CTLA-4 antibodies has proven to be beneficial in the treatment of multiple cancers (1–5). Given the variability in response rates to immunotherapy agents, efforts have focused on identifying predictors of response to therapy, using cancer-specific as well as pan-cancer approaches. Expression of PD-L1 on cancer cells, high microsatellite instability, and tumor mutation burden (TMB) are the main predictors of response that have been identified, although the microbiome and a variety of immune, transcriptomic, metabolomic, and genomic biomarkers have been studied (4, 6–20). Identification of such biomarkers is both of clinical utility in choice and prioritization of treatment, and may provide insight into immune regulatory mechanisms and novel treatment approaches.

CDKN2A is a tumor suppressor gene on chromosome 9p21.3 that encodes two different proteins: p14ARF and p16INK4a. Translated in alternative reading frames, p14ARF and p16INK4a each act as regulators of the cell cycle. p14ARF interacts with MDM2 and degrades it, preventing the inactivation of p53 through ubiquitin-mediated proteolysis or transcriptional silencing (21). p16INK4a binds CDK4 and CDK6 to prevent phosphorylation of the RB protein. Loss-of-function (LOF) mutations or homozygous deletion of CDKN2A leads to loss of both proteins, releasing G₁–S and G₂–M cell-cycle checkpoints and resulting in uninhibited cell proliferation and tumor formation (22).

Along with TP53, CDKN2A is one of the most commonly altered genes in human cancers (22, 23). In The Cancer Genome Atlas (TCGA) studies, genomic alterations in CDKN2A were found in about 50% of head and neck cancers, 47% of pancreatic and esophageal cancers, 37% of melanomas and bladder cancers, 35% of gliomas, 30% of non–small cell lung cancers (NSCLC) and gastroesophageal junction and gastric adenocarcinomas, and 3% of renal cell carcinoma (RCC; refs. 24–31). Three different mechanisms can lead to inactivation of the gene: genomic deletion, point mutations, and CpG-island hypermethylation at the promoter region of CDKN2A. Genomic deletion, so-called “deep deletion” or homozygous loss, which typically extends to include other genes on 9p21.3, is the most common alteration affecting CDKN2A (32).

Multiple studies have examined the impact of CDKN2A alterations on the tumor immune microenvironment and on the response to ICI (33, 34). In melanoma, immune checkpoint blockade is less effective when CDKN2A is deleted, thought to be due, in part, to loss of senescence induced by CDKN2A deletion (35, 36). In a study of 909 NSCLC tumor specimens, CDKN2A/CDKN2B homozygous loss was significantly associated with negative PD-L1 expression by IHC (37). In pancreatic cancer and glioma, tumors harboring deletions or mutations in CDKN2A/CDKN2B had lower immune cytolytic activity, as defined by a cytolytic index that considers the expression signature of GZMA and PRF-1 (38, 39).

There is conflicting evidence regarding the association between CDKN2A alterations and ICI response. For example, in patients with advanced RCC infiltrated with CD8⁺ T cells, deletions in 9p21.3 (affecting CDKN2A) were associated with reduced clinical benefit and inferior survival when treated with ICI after tyrosine kinase inhibitor treatment (40). Similarly, in urothelial carcinoma, CDKN2B homozygous deletions (often concomitant with CDKN2A deletions) were associated with worse objective response rates and worse survival in the setting of ICI treatment (41). However, in melanoma, the evidence is less consistent, with some studies showing ICI resistance (36, 42, 43), and others demonstrating no effect of CDKN2A alteration on tumor response (44). In other tumor subtypes, associations between CDKN2A alterations and ICI response have not been examined in detail.

Herein, we assessed the association between LOF variants in CDKN2A and clinical outcomes in patients treated with ICIs in two large independent cohorts including six different cancer types.

The study included patients diagnosed with a solid cancer type for which ICIs were FDA approved: esophagogastric adenocarcinoma (EGC), head and neck squamous cell carcinoma (HNSCC), melanoma, NSCLC, RCC, and urothelial carcinoma. Two independent cohorts were analyzed: An internal cohort of 789 patients treated at the Dana-Farber Cancer Institute (DFCI; Boston, MA) and an external cohort of 1,250 patients treated at Memorial Sloan Kettering Cancer Center (MSKCC; New York, NY; Supplementary Fig. S1). All DFCI patients received an anti–PD-1/PD-L1 or anti–CTLA-4 agent in the metastatic setting and had next-generation targeted sequencing (NGS) of their tumor tissue performed (as described below). Patients were excluded if they were lost to follow-up, had no measurable disease, or had clinical deterioration within one week of the first ICI dose. Patients on combined treatment, for example, mutation-targeted or chemotherapy combined with non-ICI therapy, were excluded. Among the DFCI cohort, formalin-fixed paraffin-embedded (FFPE) specimens were collected between 2012 and 2020.

The study was conducted in accordance with the ethical guidelines of the Declaration of Helsinki. The Institutional Review Board (IRB) of DFCI approved this work and informed written consent for cancer genetic analysis was obtained from each subject.

For the MSKCC cohort, we obtained deidentified publicly available data on 1,250 patients (Supplementary Fig. S1; ref. 13). Specimens were collected between 2013 and 2017.

We analyzed a publicly available dataset, IMvigor210 (33), which was a single-arm phase II study investigating atezolizumab (1,200 mg 3 times weekly) in patients with metastatic urothelial carcinoma (NCT02108652, NCT02951767). All 236 patients included had available tumor tissue taken in the 2 years prior to study entry. This tissue was used for PD-L1 analysis (SP142 antibody using the VENTANA platform, see below), immune cell infiltration assessment, and targeted NGS using the FoundationOne panel (Foundation Medicine) as described previously (33). We excluded patients without FoundationOne sequencing, RNA sequencing (RNA-seq), PD-L1 staining, or immune cell infiltration assessment by IHC. We also excluded patients with CDKN2A alterations of unknown significance (missense mutations, synonymous mutations, and single-copy losses).

For the TCGA Bladder Cancer (TCGA BLCA) cohort (26), we focused our survival analysis on patients with stage IV muscle-invasive bladder cancer (MIBC, n = 135 patients). For transcriptomic analysis, we included MIBC tumors across all stages (II–IV) for a total of 406 tumor samples with available bulk RNA-seq data.

Data about gender, cancer type, age at initiation of ICI therapy, type, and class of ICI-based regimen were collected for both the DFCI and MSKCC cohorts. Additional clinical variables available for the DFCI cohort included lines of therapy prior to ICIs and site of lesion subjected to targeted sequencing.

Details of the tissue collection, DNA extraction, and tumor targeted sequencing using the Oncopanel/PROFILE and MSK-IMPACT for DFCI and MSKCC cohorts, respectively, were previously described in detail (45–48). The Oncopanel gene panels include capture probes for 275–447 cancer-associated genes, as well as intronic portions of 60 genes for rearrangement detection (46). As described previously (13), MSK-IMPACT includes three versions and identifies somatic exonic mutations in a predefined subset of genes ranging between 341 and 468. The MSK-IMPACT analysis included both tumor-derived and matched germline normal DNA, so that somatic mutations were certain.

For the purposes of this study, we focused our mutational and copy-number variation (CNV) analyses on CDKN2A. TMB was defined as the number of exonic, nonsynonymous single nucleotide variants (SNV), and indel mutations per megabase of genome examined.

For the DFCI Oncopanel data, we excluded putative germline variants if they were previously found in either (i) a panel of historical normal samples sequenced via Oncopanel, (ii) the exome-sequencing project (ref. 49; http://evs.gs.washington.edu/EVS/; accessed May 2013), or (iii) in the Genome Aggregation database (gnomAD; ref. 50) at a population allele frequency ≥0.1%, unless the variant was also reported in COSMIC (45). SNVs and indels were called by the Oncopanel pipeline using MuTect v.1 0.27200 (https://software.broadinstitute.org/cancer/cga/mutect) and annotated using Oncotator (https://software.broadinstitute.org/cancer/cga/oncotator).

For the MSKCC cohort, germline variants were eliminated through the use of patient-matched blood DNA (47). SNVs were called using MuTect, VarDict (51), and Somatic indel detector (52), and reported for >5% allele frequency (novel variants) or >2% allele frequency (recurrent hotspots). Copy-number alterations were called using a custom pipeline and reported for fold-change >2 (47). Structural rearrangements were called using Delly (53).

In the IMvigor210 cohort, putative germline variants were filtered out using the gnomAD (50) at a frequency ≥0.1%, unless the variant was also reported in COSMIC (45).

For the DFCI, MSKCC, IMvigor210, and TCGA urotherlial carcinoma cohorts, LOF variants in CDKN2A were defined as either nonsense SNVs, out-of-frame and in-frame indels, splice-site variants affecting consensus nucleotides; or homozygous deletions. Missense SNVs were not included as we could not assess functional effects with confidence. For the DFCI cohort, CNVs were identified using a custom R-based tool (VisCap-Cancer; ref. 54) that compares read-depth at all genomic regions assayed among different samples. For all four cohorts, we focused on homozygous deletion CNVs in this analysis, and excluded heterozygous deletions, as the latter have uncertain functional effects, and can be confounded by background noise on both Oncopanel and MSK-IMPACT sequencing platforms (46, 47).

We obtained publicly available bulk RNA-seq data from the TCGA BLCA (55) and IMvigor210 cohorts (33). For each specimen, the transcripts-per-million (TPM) counts were calculated and the data was upper-quartile normalized using the R package edgeR. Tumor Immune Estimation Resource (TIMER; cistrome.shinyapps.io/timer) was used as previously described to investigate the molecular characterization of tumor–immune interactions as a function of CDKN2A genomic alteration status (56). Gene set enrichment analysis (GSEA) (http://software.broadinstitute.org/gsea/index.jsp) was performed on the TCGA dataset (n = 406 tumors) to test whether any biologically relevant gene sets (Hallmark, C6-oncogenic signatures, C7-immunologic signatures, C2-canonical pathways) were differentially expressed between CDKN2A-altered versus wild-type (WT) tumors.

We obtained data on PD-L1 and CD8 IHC staining from the IMvigor210 cohort of metastatic urothelial carcinoma (33). In this prior study, samples were scored for PD-L1 expression on tumor-infiltrating immune cells (IC), as IHC IC0, IC1, or IC2+, if <1%, ≥1% but <5%, or ≥5% of IC were PD-L1 positive, respectively. Tumor CD8 IHC was used to categorize tumors as “desert” when the prevalence of CD8⁺ cells was low (<10 CD8⁺ cells per 10 200x fields), “immune excluded” when CD8⁺ cells were present exclusively in tumor stroma, and “inflammatory” when CD8⁺ cells were in direct contact with malignant epithelial cells.

Statistical analyses were performed using R version 3.6.2. Statistical tests included χ² or Fisher exact tests for categorical variables and the Wilcoxon rank-sum test (two-group comparisons) or the Kruskal–Wallis exact test (three-group comparisons) for continuous variables.

Clinical outcomes included overall survival (OS) for the DFCI, MSKCC, and TCGA cohorts. For the DFCI and MSKCC cohorts, OS was calculated from the date of ICI initiation to the date of death. For the TCGA cohort, OS was calculated from the date of diagnosis to the date of death and progression-free survival (PFS) was calculated from the date of diagnosis to the date of progression or death. Alive patients were censored at the date of last follow-up. For the DFCI cohort, time-to-treatment failure (TTF) and overall response rate (ORR) were also available. TTF was calculated from the start date of ICI therapy to the start date of the next treatment or death. Patients alive and not started on next line were censored at the date of last follow-up. Response was investigator-assessed. ORR was defined as complete response (CR) or partial response (PR).

The distributions of OS and TTF were estimated with the Kaplan–Meier method along with 95% confidence intervals (CI), and their associations with genomic alterations were examined with the Wald χ² test from the Cox regression, adjusted for treatment type (single vs. combination) and number of lines of therapy prior to ICI for the DFCI cohort, and treatment type for the MSKCC cohort. For cohorts with significant associations, we further adjusted for TMB as a continuous variable.

The effects of CDKN2A alterations on ORR are presented as ORs estimated from logistic regression models, adjusted for prior lines of therapy, treatment type, and TMB in the DFCI cohort. All comparisons were conducted separately for each tumor histology. Two-sided P values are reported. Multiple comparison adjustment was not performed as these studies were exploratory in nature.

All relevant correlative data are available from the authors and/or are included with the manuscript. Clinical and genomic data for the DFCI and MSKCC cohorts are available in Supplementary Tables S1–S13. Targeted sequencing raw data were not deposited in a public repository since full sequencing data are not consented to be shared. The data that support the findings of this study are available from the corresponding author upon request.

All clinical and correlative data from the TCGA BLCA and IMvigor210 clinical trial are publicly available. All intermediate data from the RNA-seq analyses of the IMvigor 210 and TCGA cohorts are publicly available. Any other queries about the data used in this study should be directed to the corresponding authors of this study.

Algorithms used for data analysis are all publicly available from the indicated references in the Methods section. Any other queries about the custom code used in this study should be directed to the corresponding authors of this study.

Of 25,881 patients with available Oncopanel sequencing results at DFCI (Boston, MA) between September 2012 and October 2020, we identified 1,485 patients who received ICI treatment and did not harbor any SNV of unknown significance in CDKN2A, across the six included tumor types (Methods; Supplementary Fig. S1; Supplementary Table S1). From the MSKCC (New York, NY) cohort, we identified 1,250 patients who fit the same criteria (Supplementary Fig. S1; Supplementary Table S2).

The frequency of LOF genomic alterations (GA) in CDKN2A was similar in the DFCI and MSKCC cohorts (Fig. 1; Supplementary Fig. S1; Supplementary Tables S1, S3, and S4), but differed by histology. HNSCC and melanoma had the highest rate of CDKN2A GAs (HNSCC: 30/77, 39%, melanoma: 124/375, 33.1%) in the DFCI cohort. Homozygous deletions were the most common CDKN2A GA for all cancer types in both cohorts except for HNSCC in the DFCI cohort (Fig. 1; Supplementary Tables S1–S3).

DFCI cohort patients were excluded if they had: heterozygous loss of CDKN2A, received ICIs with concurrent chemo- or other therapy, received ICIs in the adjuvant setting, or deteriorated within a week of ICI initiation (Methods; Supplementary Fig. S1). Seven hundred and eighty-nine patients met the eligibility criteria and constituted the DFCI clinical cohort (Supplementary Fig. S1; Supplementary Tables S5–S7). For the MSKCC cohort, these exclusion criteria could not be applied because: (i) heterozygous loss data was not available and (ii) clinical data was limited. Thus, the entire cohort of 1,250 MSKCC patients was used for comparative analyses (Supplementary Tables S1 and S2).

Median follow-up after initiation of ICI therapy was 26.1 months (95% CI, 22.3–31.1; Table 1) and 18 months (95% CI, 17.0–19.0) for the DFCI and MSKCC cohorts, respectively (Table 1; Supplementary Tables S2–S5). The most common tumors in the DFCI and MSKCC cohorts are shown in Supplementary Fig. S2.

Median age at first dose of ICI was 65 (range: 25–93) years for the DFCI cohort. Thirty-three percent of patients from MSKCC were aged 61–70 years (N = 412). Nearly all of the patients received anti–PD-1– or anti–PD-L1–based therapy (DFCI: N = 773, 98%; MSKCC: N = 1,181, 95%; Table 1), most commonly as single agents (DFCI: N = 690, 87.4%; MSKCC: N = 963, 77%; Table 1; Supplementary Tables S2 and S5). Among 789 patients treated with ICI at DFCI, 71 (9%) had their tumor sequencing performed on biopsies obtained after initiation of ICI.

On multivariable analysis of the DFCI cohort (see Methods), patients with urothelial carcinoma with LOF CDKN2A GAs treated with ICI had significantly shorter overall survival (OS) compared with those without such alterations (median OS = 8.8 months vs. 25.2 months, respectively; HR, 1.8; 95% CI, 1.1–3.1; P = 0.019; Fig. 2A). Similarly, patients with melanoma whose tumors had LOF CDKN2A GAs also had worse OS compared with those with wild-type (WT) CDKN2A (median OS = 27.2 months vs. not reached, respectively; HR, 1.8; 95% CI, 1.1–3.1; P = 0.026; Fig. 2B). Similarly, in both the urothelial carcinoma and melanoma cohorts, CDKN2A alterations were associated with shorter time-to-treatment failure (TTF): for urothelial carcinoma, median TTF 4.2 months vs. 8.4 months, respectively (HR, 2.0; 95% CI, 1.3–3.1; P = 0.002; Fig. 2C); for melanoma, median TTF 10.0 months vs. 20.1 months, respectively (HR, 1.8; 95% CI, 1.2–2.9; P = 0.003; Fig. 2D). In contrast, there was no association between CDKN2A mutation status and overall survival or TTF for EGC, HNSCC, NSCLC, and RCC in the DFCI cohort (Supplementary Figs. S3, S4A, and S4B). However, all of these DFCI cohorts were much smaller than the DFCI urothelial carcinoma and melanoma cohorts with the exception of NSCLC, and therefore had lower power of detection. We also examined the possibility that the type of CDKN2A alteration would have an effect on ICI response. Although this comparison was underpowered by the small number of cancers with one type of alteration versus the other for several cancer types, every cancer type showed similar OS and TTF for patients with homozygous deletions in comparison with those with LOF point mutations (all P values > 0.05).

In the MSKCC urothelial carcinoma cohort (N = 206), there also was a significantly shorter OS among tumors with LOF CDKN2A alterations versus WT CDKN2A after adjusting for TMB and treatment type (see Methods, median OS = 11 months vs. OS = 19 months, respectively; HR, 1.7; 95% CI, 1.1–2.6; P = 0.017; Fig. 2E). We also noted a significant association between LOF CDKN2A GAs and worse OS in both the MSKCC RCC and EGC cohorts [RCC: median OS of 13 months vs. 50 months, respectively (HR, 5; 95% CI, 1.8–14.1; P = 0.002); EGC: median OS of 8 months vs. 17 months, respectively (HR, 2.4; 95% CI, 1.3–4.6; P = 0.006; Supplementary Figs. S3C and S4A–S4E]. We could not detect any significant association between LOF CDKN2A alterations and OS (after adjusting for covariates; see Methods) among MSKCC patients with HNSCC, melanoma, or NSCLC (Supplementary Figs. S4C and S5). Notably, for the MSKCC cohort, we were unable to adjust for other clinical variables such as lines of therapy prior to ICIs as this data was not available. Similarly, TTF and response data were not available in the MSKCC cohort.

In the urothelial carcinoma and melanoma DFCI cohorts, tumors harboring CDKN2A alterations were less likely to respond compared with WT tumors, but the associations were only statistically significant for urothelial carcinoma after adjusting for TMB, type of therapy, and prior lines of systemic treatment (see Methods, urothelial carcinoma: OR, 4.0; 95% CI, 1.2–18.5; P = 0.04; melanoma: OR, 2.0; 95% CI, 0.9–4.1; P = 0.06). CDKN2A alterations were not associated with ORR in any of the other five cancer types (Supplementary Fig. S4D).

In the DFCI urothelial carcinoma cohort, PD-L1 IHC was performed for 22 of 109 (20%) patients treated with ICI. Using a cutoff of ≥1% for positivity, 10 were PD-L1 positive and 12 were negative. No association between tumor cell PD-L1 status and OS, TTF, or ORR to ICI therapy was seen in this very small subset. Moreover, prior reports have shown that PD-L1 expression on tumor cells was not reported to be associated with response to ICI in urothelial carcinoma (33).

We were also interested in examining the potential association between heterozygous loss of CDKN2A and ICI response in the DFCI urothelial carcinoma cohort. Twenty-six patients with urothelial carcinoma with heterozygous loss of CDKN2A who were treated with ICI were compared with the original CDKN2A-altered and WT groups. Overall survival was slightly longer for patients with heterozygous loss of CDKN2A in comparison with CDKN2A WT patients, but neither comparison was statistically significant (P_adj = 0.16 and 0.82 for OS and TTF, respectively; Supplementary Fig. S6; Supplementary Table S8). Moreover, patients with urothelial carcinoma with the earlier definition of LOF alteration in CDKN2A had significantly worse OS and TTF compared with tumors with heterozygous deletions in CDKN2A (OS P = 0.0003; TTF P = 0.009).

To assess whether the association between CDKN2A alterations and clinical outcomes in the urothelial carcinoma cohort treated with ICIs was a general prognostic feature, we analyzed a DFCI cohort composed of 56 patients with urothelial carcinoma treated with platinum-based therapy (38 CDKN2A WT and 18 CDKN2A altered; Supplementary Table S9). Although the number of subjects was relatively small, there was no association with OS in CDKN2A-altered versus WT tumors among patients with urothelial carcinoma treated with platinum-based therapy (CDKN2A-altered median OS = 14.9 months; CDKN2A WT median OS = 13.4 months; HR, 0.78; 95% CI, 0.66–2.5; P = 0.5; Fig. 3A). To examine this further, we analyzed the TCGA data for 131 patients with stage IV urothelial carcinoma (87 CDKN2A WT and 44 CDKN2A LOF). In this patient population as well, there was no significant association between CDKN2A status and OS (Fig. 3B; Supplementary Table S10) or PFS (Fig. 3C; Supplementary Table S11). Note that the TCGA tumors were collected over 5 years ago, prior to the use of ICI therapy for urothelial carcinoma.

We examined the possibility that CDKN2A GAs were associated with differences in the tumor microenvironment in urothelial carcinoma. Bulk RNA-seq data was obtained for two independent urothelial carcinoma cohorts (IMvigor210 and TCGA; see Methods) at baseline and used for this purpose. First, GSEA (http://software.broadinstitute.org/gsea/index.jsp) was performed, comparing the CDKN2A LOF urothelial carcinoma subset (used as reference) with WT urothelial carcinoma samples in the IMvigor210 cohort (CDKN2A altered n = 43; CDKN2A WT n = 193; Supplementary Table S12). Antigen processing, IFNγ, and T-cell receptor signaling pathways, and other immune/inflammatory pathways were reduced in expression in CDKN2A-altered tumors (FDR-corrected P < 0.1 for T-cell receptor, antigen processing, and IFNγ response; Fig. 4A; Supplementary Fig. S7; Supplementary Table S13). Similar analysis performed using the TCGA BLCA RNA-seq data (CDKN2A altered n = 142; CDKN2A WT n = 264), was not statistically significant (Fig. 4A).

Next, we used a 5-gene expression signature described by Oh and colleagues (ABCB1, APBA2, GPR18, PEG10, and SLAMF7) to define CD4⁺ cytotoxic T cells, a subset of T cells thought to be highly active in the immune response to infiltrated bladder tumors (57). Out of 57 inflamed bladder samples for which signature scores were computed and available in the IMvigor210 cohort, CDKN2A WT samples were found to have a higher composite signature score than CDKN2A GA samples (Wilcoxon rank-sum test P = 0.06; Supplementary Fig. S8; Supplementary Table S12).

Finally, we checked tumor-infiltrating IC status and PD-L1 IHC on both IC and tumor cell (TC) expression in the IMvigor210 cohort. Out of 236 analyzed tumors, 66 had an inflammatory immune phenotype. IC1 (1–5%) and IC2+ (≥5% cutoff) PD-L1 expression was observed in 92 of 236 (39%) and 96 of 236 (40.7%) patients, respectively. Patients with CDKN2A alterations were less likely to have an inflammatory immune phenotype and less likely to express PD-L1 on ICs compared with the CDKN2A WT group (P = 0.02 for both analyses; Fig. 4B and C; Supplementary Table S12). However, we found no association between CDKN2A alteration status and PD-L1 expression on TCs (not shown).

Because IHC data were not available on TCGA urothelial carcinoma samples, we used TIMER immune deconvolution to check for associations between CDKN2A and immune infiltration. We showed that CDKN2A expression in urothelial carcinoma samples correlated significantly with CD8⁺ T-cell infiltration (Spearman correlation coefficient ρ = 0.259; P = 4E-7; Fig. 4D). A higher copy number of CDKN2A was also associated with higher CD8⁺ infiltration (Kruskal–Wallis P = 0.0022; Fig. 4E).

In this work, we evaluated the effect of LOF GAs in CDKN2A on clinical outcomes of ICI treatment across different cancer types in two independent cohorts from DFCI and MSKCC. CDKN2A GAs were consistently associated with poorer outcomes in the DFCI and MSKCC urothelial carcinoma cohorts treated with ICI. This is in accordance with prior reports suggesting that loss of CDKN2B, a neighboring gene located 25 kb centromeric to CDKN2A, was associated with unfavorable prognosis (41). Notably, CDKN2A GAs were not associated with poor clinical outcomes in patients with TCGA with metastatic urothelial carcinoma, or in a cisplatin-treated metastatic urothelial carcinoma cohort from DFCI. This supports the idea that CDKN2A status is a predictive rather than a prognostic biomarker in this patient population.

Poor outcomes to ICI in patients with RCC with deletion of 9p21.3, a chromosomal region that encompasses the CDKN2A gene, have recently been described (40). In contrast, no prior report has assessed the impact of CDKN2A alterations on clinical outcomes in patients with EGC treated with ICIs. In our dataset, CDKN2A alterations were associated with worse OS in both the MSKCC RCC and EGC cohorts, but not the corresponding DFCI cohorts (Supplementary Figs. S3G, S3H, and S5E). For EGC, the discrepancy may be due to the breakdown of samples by site along the upper gastrointestinal tract between the MSKCC and DFCI cohorts. A very small number of patients with RCC had CDKN2A alterations (inactivating mutations or homozygous copy number loss) in both DFCI (4/56) and MSKCC (4/151) cohorts which limits the power to detect any association. In contrast, the analysis by Braun and colleagues (40) included patients with heterozygous copy-number losses in 9p21.3 where the association with worse clinical outcomes was only found in infiltrated tumors. For EGC, further study is required to explain the discordant results.

Prior clinical reports have led to conflicting results regarding the association of genomic aberrations in CDKN2A with clinical outcomes in ICI-treated patients with melanoma (34–36, 43). In our study, findings in patients with melanoma were also inconsistent in these two cohorts. DFCI melanoma patients harboring CDKN2A LOF GAs responded poorly to ICIs when compared with patients with no CDKN2A GAs. In the MSKCC cohort, however, there was no appreciable difference between the two groups. One possible explanation is that the MSKCC cohort is more heterogeneous, a mix of different disease stages at the time of treatment, including patients being treated with adjuvant ICI therapy. This could clearly confound this analysis.

In metastatic urothelial carcinoma, high expression of PD-L1 on tumor-infiltrating immune cells is associated with enhanced ICI efficacy (33, 58). In the IMvigor210 cohort, we were able to show that patients harboring CDKN2A GAs were more likely to have lower PD-L1 expression on tumor-infiltrating immune cells and a less inflammatory immune microenvironment. This is in contrast with RCC, where tumors with 9p21.3 loss were found to have a stronger immune infiltration (40). CDKN2A-altered urothelial carcinomas also showed depletion of the T-cell receptor, antigen processing, and IFNγ pathways, replicating the findings seen in melanoma (42) although we were unable to confirm these findings in the TCGA BLCA cohort.

The association between CDKN2A GAs and response to ICI treatment of urothelial carcinoma raises the question of potential mechanisms of this effect. CDK4, a critical cell-cycle kinase, is normally restrained by p16INK4a (the protein product of CDKN2A). In the absence of p16INK4a, active CDK4-cyclin D1 phosphorylates SPOP, stabilizing it and protecting it from degradation by the anaphase-promoting complex activator FZR1. Phospho-SPOP is an adaptor protein that enhances the activity of Cullin 3–based ubiquitin ligases to enhance PD-L1 ubiquitination and ultimately degradation. Hence, it has been reported that CDKN2A activity leads to reduced PD-L1 surface expression on cancer cells. However, we observed no correlation between CDKN2A loss and PD-L1 expression on tumor cells by IHC analysis in the IMvigor210 cohort. This suggests that this mechanism is not operative in urothelial carcinoma, and further study is required to elucidate how CDKN2A loss reduces ICI treatment effect independent of PD-L1 expression in urothelial carcinoma. We note that the majority of CDKN2A GAs are homozygous gene deletions in urothelial carcinoma. In the TCGA BLCA cohort, 43% of samples with homozygous deletion of CDKN2A have codeletion of part or all of a nearby interferon gene cluster on 9p21.3. We hypothesize that this codeletion event may contribute in part to the poor ICI response in urothelial carcinoma with CDKN2A GAs. Unfortunately, neither the Oncopanel nor the MSK-IMPACT analyses include analysis of this IFN gene cluster, so a direct association with response cannot be examined.

One limitation to our study is that it is a retrospective analysis of clinical outcomes in select patient cohorts treated at two tertiary care centers. However, we included several variables that are thought to influence response to ICI, such as the number of prior treatment regimens, single versus combination therapy, and TMB, and all of our analyses were multivariate, taking into account these other factors. Second, targeted panel sequencing (OncoPanel and MSK-IMPACT) was used to identify mutations and copy-number events in tumor samples. While IMPACT detects germline variants by sequencing peripheral blood or normal adjacent tissue, OncoPanel does not; this makes it possible that some of the GAs captured in the DFCI cohort are germline rather than somatic. However, all variants that were observed at a frequency >0.1% in the gnomAD were excluded, and germline mutations in CDKN2A are very rare. We also excluded samples with only single-copy losses, missense and synonymous mutations as their effect on protein function is variable. Also, promoter hypermethylation, a mechanism of CDKN2A expression silencing, was not assessed in either cohort. Hence, a subset of our “wild-type” patient cohorts could, in fact, have aberrant p16 expression due to promoter methylation. Finally, the modest level of data granularity (genomic and clinical) available to us for the MSKCC dataset prevented exclusion of tumors harboring heterozygous deletions in CDKN2A, and adjustment for important clinical variables that may have impacted the results such as systemic lines of therapy prior to ICI. These issues may explain, in part, the differences in results between the DFCI cohort and the MSKCC cohort.

In conclusion, CDKN2A GAs were found to be associated with reduced benefit from ICI therapy in urothelial carcinoma, but not consistently in any of the five other cancer types. Significant heterogeneity of response association was seen in these two cohorts, which is likely partially due to the different criteria applied to identify CDKN2A GAs, as well as differences in the size and composition of the patient populations at the two sites.

G. Sonpavde reports personal fees from BMS, Genentech, EMD Serono, Merck, Sanofi, Seattle Genetics, AstraZeneca, Exelixis, Janssen, Bicycle Therapeutics, Pfizer, Immunomedics/Gilead, Scholar Rock, and G1 Therapeutics, as well as grants from AstraZeneca, Sanofi, Gilead, QED, and BMS outside the submitted work. G. Sonpavde also reports unpaid steering committee of studies for BMS, Bavarian Nordic, Seattle Genetics, QED, and G1 Therapeutics; paid steering committee of studies for AstraZeneca, EMD Serono, and Debiopharm; data safety monitoring committee for Mereo; travel costs from BMS and AstraZeneca; writing/editor fees from UpToDate and Elsevier Practice Update Bladder Cancer Center of Excellence; and speaking fees from Physicians Education Resource (PER), Onclive, Research to Practice, and Medscape. R.I. Haddad reports personal fees from Merck, BMS, GSK, Pfizer, Genentech, Instil Bio, Bayer, Eisai, Achilles, Io Biotech, AstraZeneca, Roche, BioNtech, and Kura during the conduct of the study. R.I. Haddad also reports grants from BMS, Merck Pfizer, Genentech, and AstraZeneca outside the submitted work, as well as Data Safety Monitoring Board at Nanobiotix and chair of thyroid cancer NCCN panel. K.W. Mouw reports other support from Pfizer outside the submitted work. M. Giannakis reports grants from Bristol Myers Squibb, Merck, Servier, and Janssen outside the submitted work. F.S. Hodi reports grants and personal fees from Bristol Myers Squibb and Novartis, as well as personal fees from Merck, EMD Serono, Sanofi, Surface, Compass Therapeutics, Apricity, Pionyr, Torque, Rheos, Bicara, Psioxus, Eisai, Checkpoint Therapeutics, Idera, Takeda, Genentech, Bioentre, Gossamer, Iovance, Trillium, Immunocore, and Millennium outside the submitted work. In addition, F.S. Hodi has a patent for Methods for Treating MICA Disorders pending, licensed, and with royalties paid; a patent for Tumor Antigens and Uses Thereof issued; a patent for Angiopoietin-2 Biomarkers Predictive of Anti-immune Checkpoint Response pending; a patent for Compositions and Methods for Identification, Assessment, Prevention, and Treatment of Melanoma Using PD-L1 Isoforms pending; a patent for Therapeutic Peptides issued; a patent for Methods of Using Pembrolizumab and Trebananib pending; a patent for Vaccine Compositions and Methods for Restoring NKG2D Pathway Function Against Cancers pending, licensed, and with royalties paid; a patent for Antibodies that Bind to MHC Class I Polypeptide-related Sequence A pending, licensed, and with royalties paid; and a patent for Anti-galectin Antibody Biomarkers Predictive of Anti-immune Checkpoint and Anti-angiogenesis Responses pending. S.A. Shukla reports other support from Agenus, Agios, BreakBio, Bristol Myers Squibb, and Lumos Pharma outside the submitted work. D.A. Braun reports non-financial support from Bristol Myers Squibb, as well as personal fees from LM Education and Exchange Services, Exelixis, Octane Global, Defined Health, Adept Field Solutions, Slingshot Insights, Blueprint Partnerships, Charles River Associates, Trinity Group, and Insight Strategy outside the submitted work. T.K. Chouieri reports grants, personal fees, and non-financial support from Merck, BMS, Roche, EMD Serono, and AstraZeneca during the conduct of the study. T.K. Chouieri also reports non-financial support from NCCN panel; personal fees and non-financial support from GU SC of the NCI; personal fees from Foundation Medicine, Up-to-Date, and CME Programs outside the submitted work. In addition, T.K. Chouieri has a patent for Biomarkers of Activity/Toxicity from IO pending and is a Division Chief, GU Oncology at DFCI and is the lead of several trials in GU cancers (National/International lead). No disclosures were reported by the other authors.

E. Adib: Conceptualization, data curation, formal analysis, visualization, methodology, writing–original draft. A.H. Nassar: Conceptualization, data curation, formal analysis, investigation, visualization, writing–original draft. E.W. Akl: Conceptualization, data curation, formal analysis, visualization, writing–original draft. S. Abou Alaiwi: Data curation, visualization, writing–review and editing. P.V. Nuzzo: Data curation, writing–review and editing. T.H. Mouhieddine: Data curation, writing–review and editing. G. Sonpavde: Resources, validation, writing–review and editing. R.I. Haddad: Validation, writing–review and editing. K.W. Mouw: Validation, writing–review and editing. M. Giannakis: Validation, writing–review and editing. F.S. Hodi: Validation, writing–review and editing. S.A. Shukla: Formal analysis, writing–review and editing. A. Gusev: Formal analysis, validation, writing–review and editing. D.A. Braun: Formal analysis, supervision, writing–review and editing. T.K. Choueiri: Supervision, project administration, writing–review and editing. D.J. Kwiatkowski: Conceptualization, resources, supervision, investigation, methodology, writing–original draft.

D.A. Braun acknowledges support by the DF/HCC Kidney Cancer SPORE Career Enhancement Program (P50CA101942–15), DOD CDMRP (KC170216, KC190130), and the DOD Academy of Kidney Cancer Investigators (KC190128). S.A. Shukla acknowledges support by the NCI (R50 RCA211482A). A. Gusev acknowledges support by the NCI (R01 CA227237, R01 CA244569) and the Louis B. Mayer Foundation.