Association of TP53 and CDKN2A Mutation Profile with Tumor Mutation Burden in Head and Neck Cancer

Head and neck squamous cell carcinoma (HNSCC) affects over 800,000 people annually, including 65,000 in the United States (1, 2). The main epidemiologic factors causing HNSCC are the use of tobacco and alcohol, or infection with a transforming type of human papillomavirus (HPV; refs. 3, 4). HPV-negative HNSCC has a worse prognosis than HPV-associated (HPV+) HNSCC. While HNSCC diagnosed at an early stage is often cured with surgery or radiation, almost all patients with distant metastases require systemic therapy and nonetheless succumb to this cancer (5). The recent advent of immune checkpoint inhibition (ICI) has been transformative for some patients with advanced HPV-negative HNSCC, but even for patients expressing the favorable biomarker programmed death ligand 1 (PD-L1), the median time to progression is only 5 months and median survival only 13.4 months (6). Better predictive biomarkers to aid in treatment selection for ICI, as well as more effective therapeutic strategies, are required.

Incidence, presentation, and treatment of HPV-negative HNSCC are each strongly influenced by innate capacity for DNA damage response. This disease is typically associated with chronic use of tobacco, which provides a potent mutagenic stimulus. The significant majority of HPV-negative HNSCC tumors accumulate damaging mutations in the tumor suppressor gene TP53 (7, 8). In turn, abrogation of p53 function impairs initiation of cell-cycle checkpoints and apoptosis in the face of unrepaired DNA damage, contributing to acquisition of additional mutations and allowing tumors to gradually accumulate a significant tumor mutation burden (TMB). Loss of p53 activity also influences survival of tumors after DNA-damaging treatments including radiation or chemotherapy (9, 10), and influences prognosis (11). Across tumor types, elevated TMB (often defined as ≥10 mutations/MB or ≥15 mutations/MB, but varying among studies; ref. 12) predicts benefit from ICI. The predictive value of high TMB has been demonstrated in HNSCC specifically for the programmed cell death 1 inhibitor pembrolizumab (13), and for the PD-L1 inhibitor durvalumab (14). These relationships suggest TP53 status may be predictive of response to immunotherapies, but this remains to be further established.

Notably, TP53 mutations observed in HNSCC differ in their effect on p53 functionality, ranging from benign, to partial or complete loss of function (LOF), while some mutations (e.g., R248W, R273H, R175H) lead to gain of function (GOF; refs. 11, 15, 16). As p53 acts as a tetramer, some mutations can have dominant-negative (DNE) activity, if they result in modification of p53 that disables function but maintains protein capacity for oligomerization. Numerous groups have proposed systems for classification of TP53 mutations, based on the degree to which they disrupt protein sequence and function. For HNSCC, one study has suggested that the rules first defined by Poeta and colleagues to identify highly damaging mutations (11), when supplemented by splice site mutation information, are most prognostic of reduced survival following surgery for HNSCC, relative to numerous other algorithms for classifying these mutations (17). An evolutionary action score, computationally derived to classify mutations in highly conserved residues as high risk, was also prognostic in HNSCC (18). The relationship of specific TP53 mutation classes to prognosis for patients with HNSCC receiving DNA-damaging therapies or immunotherapies is of high interest, but not yet well understood (10).

In contrast to HPV-negative HNSCC, TP53 is very infrequently mutated in HPV+ HNSCC (19). However, p53 activity is greatly reduced in these tumors, because p53 is posttranslationally degraded by an HPV-encoded oncogene, the ubiquitin ligase E6. In parallel, the HPV oncoprotein E7 inactivates the pRB tumor suppressor protein, removing a second cell-cycle checkpoint; both E6 and E7 activities are required for cell transformation by HPV, emphasizing the importance of dual disruption of p53 and pRB. Although Rb is rarely mutated in HPV-negative HNSCC, a significant percentage of HNSCC tumors mutate or otherwise inactivate CDKN2A, encoding the tumor suppressor p16. p16 is an essential mediator of pRB tumor suppressor activity, so that loss of CDKN2A similarly compromises cell-cycle checkpoints important for repair of DNA damage (20). However, whether CDKN2A and TP53 mutations interact in a manner that influences TMB or other properties pertinent to therapeutic response to chemotherapy and ICIs, or overall prognosis of patients with HPV-negative HNSCC is unknown.

In this study, we have leveraged a large dataset of HNSCC genomic data from Caris Life Sciences (CLS), and a comparator dataset compiled from public sources, to explore these relationships. This work reveals a striking profile of interaction between mutations of TP53 and CDKN2A, with each other and with TMB, in overall HNSCC and in specific HNSCC subsites.

A total of 1,010 HNSCC tumors were analyzed by CLS (Phoenix, Arizona) as part of routine comprehensive molecular profiling, of which 875 were selected for further analysis, based on their occurrence in specific tumor subsites of interest. This study was conducted in accordance with guidelines of the Declaration of Helsinki, Belmont Report, and U.S. Common Rule. In keeping with 45 CFR 46.101(b; ref. 4), this study was performed using retrospective, deidentified clinical data. Therefore, this study was considered Institutional Review Board–exempt and no patient consent was necessary from the subjects.

Prior to molecular testing by CLS, tumor enrichment was achieved by harvesting targeted tissue from formalin-fixed paraffin-embedded (FFPE) tumor samples using manual microdissection techniques. For a small subset of the tumors arising in the oropharynx, expression of the CDKN2A-encoded protein p16 was tested by IHC (using monoclonal antibody clone E6H4) with a standard cutoff of 2+, >70% being considered p16+; for positive specimens, a confirmatory HPV High Risk ISH was carried out using mRNA probes (HPV 16, 18, and 33). Next-generation sequencing was performed on genomic DNA isolated from FFPE tumor samples using the NextSeq platform (Illumina, Inc.). Matched normal tissue was not sequenced. A custom-designed SureSelect XT assay was used to enrich 592 whole-gene targets (Agilent Technologies); a total of 1.4 MB was assessed. All variants were detected with >99% confidence based on allele frequency and amplicon coverage, with an average sequencing depth of coverage of >500 and an analytic sensitivity of 5%. Splice junctions are covered in the sequencing panel with mutations seen at ±30 nucleotides from the boundaries of BRCA1/2 genes and ±10 nucleotides of the other genes. Splicing variants were annotated only for mutations detected in ±2 nucleotides from the exon boundaries. The copy-number alteration of each exon was determined by calculating the average depth of the sample along with the sequencing depth of each exon and comparing this calculated result to a precalibrated value. TMB was measured by counting all nonsynonymous missense mutations found per tumor that had not been previously described as germline alterations. The threshold to define TMB-high was greater than or equal to 10 or 15 mutations/MB. Values for TMB are aligned with those generated by other sequencing platforms based on the use of common guidelines recently published in (21).

Comparison data sets for studies with information on TP53 and CDKN2A mutation status (including copy-number alterations, sex, age, and tumor subsite) were collected from the publicly available datasets (PAD) at cBioPortal (http://www.cbioportal.org/index.do) and American Association for Cancer Research (AACR) Project GENIE (https://genie.cbioportal.org) as of July 2021. Data selection criteria are shown in Supplementary Fig. S1. The resulting PAD dataset was used in some cases for comparison with the CLS dataset, and in other cases, was merged with the CLS dataset to increase statistical power.

TP53 variant functional assessment was derived from integration of multiple sources, in addition to mutations explicitly characterized in detail in the scientific literature. Functional annotation for damaging TP53 mutations is based on American College of Medical Genetics and Genomics (ACMG) designation as “pathogenic” or “likely pathogenic” (22), or from merged results of three datasets [the consensus between ACMG and two other commonly used resources: InterVar (http://wintervar.wglab.org/; ref. 23) and ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/; ref. 24)]. To characterize TP53 mutations based on impact on p53 protein structure and function (LOF, DNE, GOF), some TP53 mutations were classified as DNE, resulting in the inhibition of the wild-type (WT) p53 by mutant protein in transactivation or cell growth assays, as listed in the International Agency for Research on Cancer (IARC) database (http://p53.iarc.fr), based on two studies (25, 26). DNE TP53 mutations with loss-of-function effect (DNE LOF) annotation is based on designations in the IARC database (http://p53.iarc.fr), where functional classification for loss of growth suppression and DNE activities is based on Z-scores from Giacomelli and colleagues (16). A Poeta disruptive designation reflects the classification system detailed in Poeta and colleagues (11), which is heavily biased towards mutations associated with truncations, and disruption of DNA binding capacity. A GOF designation is based on mutations that have been specifically described in the literature. Known GOF TP53 mutations are listed in Supplementary Table S1, along with citations describing function. In a few cases, tumors had more than one mutation in TP53; these specimens were classified based on the phenotype of the most damaging mutation observed (typically a truncating mutation, leading to a designation as Poeta disruptive).

Hotspot mutations and mutation-enriched stretches along the primary protein sequence were identified using previously described methods (27). Briefly, to determine if a frequently mutated site on the p53 protein constitutes a mutational hotspot, we have used a binominal distribution model with a P-value cutoff 0.005. Similarly, to calculate whether non-hotspot mutations are enriched in certain linear stretches along the TP53 primary sequence, we used sliding window of 5 aa and a binominal distribution model to identify larger regions of the primary structure that were, in sum, more commonly mutated than expected. Two-dimensional representation of TP53 hotspot mutations was prepared using MutationMapper tool available at https://www.cbioportal.org/visualize. For display of the distribution of mutations on the folded protein structure, figures were prepared using the program PyMOL (https://pymol.org/2/), based on a p53 structure deposited in the PDB, 1TUP (28).

Relationships between mutations and patient characteristics were assessed using Fisher exact tests (including determination of significant difference between the mutational spectra of dichotomized gender or subsite groups), and/or a Wilcoxon test (where indicated in legends). To allow for multiple mutations within a patient, the predictors used in these models were binary indicators of the presence versus absence of particular mutations of interest. To account for multiple comparisons of various types, we have lowered the threshold for statistical significance tenfold, to 0.005. Co-occurrence or mutual exclusion of mutations was calculated using Fisher exact test. TMB comparisons among groups in the CLS cohort were compared using non-parametric Kruskal–Wallis testing. Potential differences in TP53-CDKN2A co-occurrence across different sites were evaluated using a Breslow-Day homogeneity test. Potential differences in TMB status (as defined by cut-off values) across different TP53-CDKN2A mutation groups were evaluated using a χ² test for homogeneity. A multiple regression model linking TMB as continuous outcome variable to the mutation status of TP53 and CDKN2A, and tumor subsite, was built using the lm function in R.

Web resources used in this paper include: The cBioPortal for Cancer Genomics, https://www.cbioportal.org; AACR Project GENIE, https://genie.cbioportal.org; InterVar, a bioinformatics software tool for clinical interpretation of genetic variants by ACMG/AMP 2015 guidelines, http://wintervar.wglab.org/; ClinVar, a public archive reporting the relationships among human variations and phenotypes, https://www.ncbi.nlm.nih.gov/clinvar/. Academic researchers can gain access to CLS data in this study by contacting the corresponding author and filling out a study review committee form to CLS. Researchers and their institutions will be required to sign a data transfer agreement.

Two cohorts were analyzed for this study (Supplementary Table S2). The first was a group of 875 patients with HNSCC genomically profiled by CLS. The second was a group of 794 patients with HPV-negative HNSCC for whom data were available in cBioportal from The Cancer Genome Atlas PanCancer and Broad Institute data sets, and from AACR-GENIE public resources (Supplementary Table S2; Supplementary Fig. S1), based on exome sequencing or panel testing. Reflecting the known sex bias of HNSCC (29), the majority of specimens were from males (75.5% for CLS, 70.5% from the PAD). Tumors sequenced by CLS were typically derived from patients with late-stage or metastatic/recurrent cancers (Supplementary Table S2). In contrast, the tumors from the PAD were more diverse in stage, including some earlier-stage tumors.

Tumor subsite has been associated with differences in mutational profile for HNSCC (8). To perform analyses based on tumor subsite, we omitted tumors arising in subsites with very few cases observed and considered separately those tumors where site of origin was unclear (designated “advanced” or “occult primary” in the CLS dataset). We also omitted nasopharyngeal tumors from further analysis because, although many of these tumors arise based on prior infection with Epstein-Barr virus (EBV; ref. 30), analyzed specimens were not tested for EBV. We therefore analyzed in detail the combined set of tumors in five subsites: the hypopharynx, larynx, oral cavity, oropharynx, and sinonasal cavity (733 specimens from CLS and 794 specimens from the PAD), as well as in an additional 142 CLS specimens designated as advanced or occult primary.

We analyzed the data using multiple approaches to characterize TP53 mutations (Supplementary Table S3). First, we evaluated non-synonymous mutations, considering any that altered amino acid sequence. Second, we applied ACMG criteria (31), commonly used by CLS and others to designate TP53 mutations as pathogenic. Third, we considered mutations that were independently identified by ACMG criteria as well as reported in two other commonly used resources, Clinvar (24) and Intervar (ref. 23; designated x3, based on three sources of input). Fourth and fifth, we considered TP53 mutations designated as DNE or DNE plus LOF (DNE-LOF) based on criteria of the IARC (32). Sixth, we considered TP53 mutations viewed as damaging based on the Poeta algorithm, which selects for highly disruptive mutations (including those involving stop codons, or missense mutations severely compromising the integrity of the DNA binding domain; ref. 11), supplemented with mutations disrupting splice sites; this approach was previously demonstrated to have superior prognostic capacity (17). Finally, we additionally considered mutations, which had been described in various publications as GOF, a group typically representing relatively conservative missense modifications (ref. 15; Supplementary Table S1).

In the overall cohort (combined CLS and PAD for five subsites, advanced primary, and occult), 68% of tumors bore a non-synonymous mutation in TP53 (Fig. 1A), most of which were considered pathogenic by ACMG criteria. Significantly, fewer mutations were called using the x3, DNE, DNE-LOF, and Poeta approaches, with frequencies ranging from 36% to 50%. Relatively few mutations (19%) were identified as GOF. A higher frequency of TP53 mutations occurs in males versus females for a number of classes of solid tumors (32, 33). However, comparison of mutation profile in males versus females specifically in oral and laryngeal tumors, where sample size allowed testing for statistical significance, indicated no difference by sex (Supplementary Fig. S2).

We then performed detailed comparison of the frequency of TP53 mutations in the five anatomic subsites in the independent CLS and PAD cohorts (Fig. 1B). Notably, while the frequency of non-synonymous and ACMG mutations was comparable in both cohorts, the fractions of highly damaging mutations (x3, DNE, DNE-LOF, and Poeta) in TP53 were lower in the PAD datasets than in the CLS datasets. In addition, CLS tumors diagnosed as advanced or with unknown primary (often reflecting larger, advanced tumors) had a similar frequency of TP53 mutations as did tumors from the five merged sites (Fig. 1C).

One potential explanation for the incomplete agreement of datasets is that TP53 mutational frequency reflected differing proportions of tumors originating from specific HNSCC subsites in the CLS versus PAD cohorts (for instance, with CLS having a higher proportion of tumors from the oropharynx). Although HPV-negative HNSCC is often considered as a single disease, a growing number of publications have identified differences between tumors arising in distinct anatomic subsites. For instance, in oral cavity tumors, the composition of the microbiome is associated with disease pathogenesis (34), whereas laryngeal tumors have a greater mutational burden (35). In addition, the tissues that collectively form the structures of the head and neck arise from distinct developmental primordia (36), which may affect gene expression profile and dependence on specific pathways.

For both the CLS and PAD cohorts, the highest frequency of TP53 mutations, for all classes of mutation, was observed in the hypopharynx, larynx, and oral cavity, and the lowest in tumors of the sinonasal cavity (Fig. 2A and B). Overall, in the PAD, differences in frequency of TP53 mutation were less marked between tumor subsites, and for each subsite, there were fewer damaging mutations than observed for the same subsites in the CLS. For the PAD, data on deep deletions (loss of both copies) for TP53 was available; inclusion of this data did not affect estimate of gene LOF, as there were few cases of TP53 loss (Supplementary Fig. S3A). For oropharyngeal tumors, the CLS dataset suggest a lower frequency of TP53 mutation than observed in the PAD, although the relatively small total number of oropharyngeal tumors in the PAD limits statistically significant conclusions. This difference most likely reflects lack of testing for HPV in the majority of the CLS specimens, given the common occurrence of HPV+ disease in this specific tumor subsite, and the documented low frequency of TP53 and CDKN2A mutations in HPV+ tumors.

Using the PAD, where information was available, we then compared the frequency of mutations called by the ACMG criteria in the oral and laryngeal subsites, based on tumor stage (Supplementary Fig. S3B). This indicated no significant differences in the frequency of ACMG-called mutations based on tumor stage, in either the oral or laryngeal subsite, in accord with the established idea that TP53 mutation occurs early in tumor formation (37). This interpretation is supported by comparison of CLS and PAD data for individual tumor sites (Fig. 2A and B) versus CLS data for advanced and occult primary tumors (Fig. 2C), reflecting large and extremely advanced tumors, as the latter do not have a higher frequency of TP53 mutation than observed in the subsites.

We separately analyzed mutations in CDKN2A. These occurred at a lower frequency than mutations in TP53 (Fig. 1B and C), but almost all mutations were predicted to significantly impair or eliminate CDKN2A protein function, with 64% representing frameshifts or nonsense mutations. Overall, 35.7% of tumors bore a non-synonymous mutation in CDKN2A (32.2% in the CLS, and 39.54% in the PAD). Of TP53 WT tumors, the great majority were also WT for CDKN2A, a higher frequency than would be expected based on random assortment (P < 0.0001). Among TP53-mutated tumors of the five merged subsites, between 37% and 48% also bore CDKN2A mutations, across the various TP53 mutation categories; this frequency was also observed in HNSCC diagnosed as advanced or with occult primary (Fig. 1B and C; Supplementary Table S4). Based on analysis of the PAD, the frequency of CDKN2A mutations did not vary dependent on tumor stage (Supplementary Fig. S3B), implying mutation occurred early in tumorigenesis.

We also considered the relationship of TP53 and CDKN2A mutations in the context of individual disease subsites (Fig. 2A–C; Supplementary Fig. S3A; Supplementary Table S4). Here, although some consistent features of mutation profile were identified, significant differences emerged discriminating subsite profiles and datasets.

Employing the ACMG criteria for both the CLS (Fig. 2A) and PAD (Fig. 2B) cohorts across five tumor subsites, the highest proportion of TP53 mutations were called as damaging and the lowest as GOF. For both datasets, there was a very low frequency of tumors with a CDKN2A mutation in the absence of an accompanying TP53 mutation. Among tumors with a TP53 mutation, the likelihood of an accompanying CDKN2A mutation was typically lower in the PAD than in CLS cohort, even though pre-screening (Supplementary Fig. S1) had eliminated AACR-GENIE datasets with exceptionally low reported CDKN2A mutation frequency. The greatest likelihood of co-occurring CDKN2A and TP53 mutations was typically in laryngeal and oral cavity tumors. In the CLS dataset (Fig. 2A), the likelihood of co-mutation was most notable in laryngeal tumors bearing moderately to severely damaging classes of TP53 mutations. For the PAD, inclusion of available deep deletion information (Supplementary Fig. S3A) slightly increased the number of tumors with loss of CDKN2A, particularly in the oral and laryngeal subsites. Otherwise, the relative likelihood of co-occurring mutations was not affected by mutation calling method. Both the CLS and PAD cohorts had lower frequency of co-occurrence in oropharyngeal tumors, although the differed regard to observed frequency (PAD, 3.6% of TP53-mutated tumors, versus CLS, 17.9% of TP53-mutated tumors), with the higher value in the CLS dataset reflecting the likely inclusion of some HPV+ tumors in the dataset. Based on rigorous statistical testing of homogeneity, levels of mutational co-occurrence of CDKN2A and TP53 (pathogenicity annotated according to ACMG criteria) in PAD dataset (Supplementary Fig. S3C) did not significantly differ between tumor subsites. We also performed similar analysis for the CLS dataset of advanced and occult-primary tumors (Fig. 2C). The patterns of co-occurrence of TP53 and CDKN2A mutations in each of these groups and based on each calling method were similar to those found in the laryngeal and oral cavity cancers.

For the CLS dataset, TMB data were available. Using these data (Figs. 3 and 4), we evaluated whether TP53 and CDKN2A mutation status correlated with TMB. In the merged CLS cohort (five subsites), tumors bearing mutations in TP53 had significantly higher average TMB than did tumors lacking TP53 mutations; however, in the case of GOF TP53 mutations, the increase in TMB was not significant at a threshold of P < 0.005 (Fig. 3A). In this analysis, the presence or absence of a CDKN2A mutation did not correlate with average TMB (Fig. 3A). We assessed if a co-occurring CDKN2A mutation increased the proportion of tumors with a TMB above cutpoints of 10 (Fig. 3B) or 15 (Fig. 3C) mutations/Mb (14). Tumors with ACMG, x3, and IARC-specified mutations in TP53 were more likely to exceed those cutpoints if they bore a simultaneous mutation in CDKN2A, although this did not retain significance upon multiple comparison correction. In contrast, no evidence of a contribution of a CDKN2A mutation was observed in tumors bearing Poeta+splicing mutations. We note this observation of differential effect of CDKN2A mutation on correlation of distinct classes of TP53 mutation with elevated TMB would be in line with interpretation of a driver effect of the damaging TP53 mutation on the higher TMB; otherwise, we would have anticipated comparable results for all TP53 class combinations with CDKN2A. Interestingly, in tumors with no TP53 mutation, the presence of a CDKN2A mutation alone was associated with a significantly elevated TMB average, and increased frequency of tumors with a TMB in excess of 10 (Fig. 3A and B); however, almost no tumors lacking a TP53 mutation exceeded a TMB of 15, regardless of CDKN2A mutational status (Fig. 3C).

The ACMG classification system is standardly used for clinical reports indicating a pathogenic TP53 mutation. Therefore, using this TP53 mutation classification as basis for further analysis, we then examined the correlation of TP53 mutation, CDKN2A mutation, and TMB separately in each of the tumor subsites (Fig. 4). This revealed striking differences. For tumors of the oropharynx, advanced tumors, and those arising from an occult primary, co-occurrence of CDKN2A and ACMG-defined TP53 mutations correlated with a higher average TMB than TP53 or CDKN2A mutation alone, or no mutation (Fig. 4A). Focusing on TMB cutpoints, a significant elevation in the frequency of tumors exceeding TMB 15 was observed in advanced or occult primary tumors bearing mutations in CDKN2A, or CDKN2A plus ACMG-designated TP53 mutation; however, TP53 mutation alone did not correlate with higher TMB (Fig. 4B). No statistically significant correlation was observed between TP53 mutation and a greater frequency of tumors with high TMB was seen in hypopharyngeal and laryngeal tumors.

Previous studies have identified TP53 mutational hotspots, including in HNSCC (38, 39). The large size of the combined cohort analyzed in this study, coupled with the questions of the relation between HNSCC mutation class, co-occurrence of a CDKN2A mutation, and TMB, motivated analysis of specific TP53 mutational loci (Figs. 5 and 6; Supplementary Fig. S4). For this analysis, we only included oropharyngeal tumors that were confirmed to be HPV negative based on direct assessment (21 specimens). Cumulatively, 1,015 missense mutations, 37 in-frame deletions, and 158 truncating mutations were observed in 1,465 samples. Over the complete cohort, missense mutations were clustered within the sequences encoding the DNA binding domain (Fig. 5A; Supplementary Fig. S4). From this analysis, we identified 55 residues (Supplementary Table S5) mutated at higher than expected frequency (≥7 alterations, given the length of the TP53 coding sequence versus the total number of mutations). On the basis of analysis of tumor subsites (Fig. 5B), there was a marked difference in TP53 hotspot profile at distinct tumor subsites, likely arising at least in part from the distinct mutagenic environments characterizing oral cavity, larynx, and other regions.

In sum, a comparison of the complete set of TP53 mutations found with or without accompanying CDKN2A mutations (Fig. 5C and D; Supplementary Table S5) indicated that the overall pattern of mutation did not significantly differ between the two groups. However, a focus on the profile of the most common hotspots for TP53 missense mutations designated as pathogenic by ACMG criteria suggested some notable differences based on presence or absence of CDKN2A mutation, including mutations at hotspot areas near codon 277, more common in the absence of CDKN2A co-mutation, and near codons 192 and 157, more common when CDKN2A mutations co-occur (Fig. 6A). Interestingly, in tumors without CDKN2A mutation, there was higher frequency of TP53 truncating mutations within the first 150 amino acids of coding sequence (24.4% vs. 15.5% of total mutations; Fig. 6B). These would be associated with total loss of p53 function. In other analysis, focusing on the most commonly mutated residues (occurring at a frequency of >1% of the total mutations) observed in tumors with TP53 alone mutated, versus TP53 plus CDKN2A, we plotted those which occurred at a frequency >1.5-fold higher in one or the other genotype on the solved structure of the p53 DNA binding domain (Fig. 6C: 3D model). Intriguingly, this revealed a tendency for the most common TP53 hotspots co-occurring with CDKN2A mutations biased towards occurrence at sequences of p53 that directly contact DNA, whereas TP53 hotspots occurring in the absence of CDKN2A mutation were more likely to localize away from the DNA, in regions associated with control of p53 protein stability (Supplementary Table S5).

Although previous studies have investigated profiles of TP53 mutation in HNSCC and other tumor types, this is the first to analyze the interaction pattern of TP53 mutations with CDKN2A mutational profile, and with TMB. The data support several conclusions. First, considering HPV-negative HNSCC overall, the quantitative likelihood of a CDKN2A mutation accompanying a TP53 mutation is not affected by the degree of impairment of TP53 as predicted by functional class. Second, the frequency of co-occurrence of a CDKN2A mutation with a TP53 mutation varies by HNSCC tumor subsite, being highest in laryngeal and oral tumors. Third, the frequency of TP53 and CDKN2A mutations, and the likelihood of their co-occurrence, is not elevated in advanced versus early tumors, confirming occurrence of these mutational patterns as early tumor events. Fourth, across a combined set of HPV-negative HNSCC, all classes of TP53 mutation except GOF are associated with higher TMB versus TP53 WT tumors. In addition, CDKN2A mutation is associated with higher TMB, either occurring as a single mutation or in combination with a TP53 mutation. Fifth, detailed analysis by tumor subsite based on ACMG mutations indicates that the presence either of a CDKN2A mutation or a TP53 mutation is associated with a TMB >10 in laryngeal or oral cavity tumors. Sixth, there is a nonequivalent pattern of TP53 mutations in tumors with versus without a co-occurring CDKN2A mutation, and alignment of observed missense mutations with the TP53 protein structure suggests that the missense mutations may associate with distinct domains, dependent on the presence or absence of CDKN2A mutation. Finally, available sets of public data considered for inclusion in this study report different, and in some cases significantly lower levels of CDKN2A mutations than does at least one optimized clinical platform (CLS).

Masica and colleagues first reported that the combined set of tumors with TP53 mutations defined by the Poeta algorithm, or with TP53 splicing mutations, had worse outcome after margin-negative surgery and risk-based (predominantly radiation) postoperative therapy (17). Because the recognition that these most function-damaging TP53 mutations were associated with therapeutic response and prognosis, it has been of considerable interest to determine whether these mutations have a driver or passenger function, and might be relevant to other therapies. Given the reported relationship between high TMB and response to ICI, and the central role of p53 in mediating response to DNA damage, defining the relationship of TP53 mutation and TMB was a priority. Overall, our data are compatible with the idea that highly damaging TP53 mutations are not primary drivers of the highest TMB levels, as similar TMB ranges are observed across all classes of disruptive TP53 mutation (Fig. 3). Rather, higher TMB appears to be associated with large, later-stage tumors (such as those annotated as occult primary or advanced; ref. 40), and the presence of both CDKN2A and TP53 mutations (Fig. 4). The association of CDKN2A mutation with higher TMB—as an independent factor, as well as in combination with TP53—was surprising, as there is no specific precedent for a driver function (Figs. 3 and 4). Interestingly, CDKN2A mutation status was also correlated with a difference in location of TP53 missense mutations, localizing more to TP53 DNA-binding sequences, rather than those associated with protein stability (Fig. 6).

Although further work is required to determine if the relationship between CDKN2A mutation, TP53 mutation spectrum, and TMB reflects a functional role of CDKN2A-encoded proteins in governing mutation status, the observed correlation has interesting implications if a functional role is confirmed. Notably, CDKN2A encodes two distinct coding proteins: p16(INK4a) and p14(ARF). p16 is heavily studied in HNSCC based on its role in inhibiting the Rb pathway and its overexpression in the face of Rb inactivation in HPV+ disease; loss of CDKN2A is targetable with CDK4/6 inhibitors such as palbociclib (41, 42). Some recent studies have suggested that mutation of CDKN2A is associated with resistance to ICIs in solid tumors (43, 44), which is surprising given the correlation we identified with higher TMB; however, no mechanism connecting CDKN2A to ICI resistance has yet been identified. CDKN2A also encodes p14(ARF), which compared with p16 has attracted little study in HNSCC. p14(ARF) regulates p53 function (45), contributing to p53 stabilization by inhibiting MDM2; CDKN2A mutations leading to loss of p14(ARF) function would also have the effect of p53 destabilization. In one early study, an active role of exogenously expressed p14(ARF) in suppressing the growth of HNSCC tumors retaining TP53 was identified (46). More investigation of the interaction of p14(ARF) and p53 is merited. In future studies, it would also be of interest to incorporate information about copy-number loss of CDKN2A and TP53 (here available only for the PAD), and about tumors with hypermethylation and silencing of CDKN2A, given this is a frequent mechanism of inactivation of this tumor suppressor (47).

On the basis of these data, it would also be of interest to evaluate whether a combination signature of TP53 and CDKN2A mutation may be predictive of therapeutic response or prognostic of outcome for HNSCC, given results in this study and identification of such a role in other cancer types (48). The relation of dual CDKN2A/TP53 mutations to higher TMB may also be of interest in the context of gaining insight into which tumors are likely to respond to immunotherapies. However, response to immune checkpoint inhibitors depends on additional factors beyond TMB, including among others an “immune hot” tumor microenvironment, and levels of expression of PD-L1 on individual tumor cells; moreover, this study identifies more HNSCC tumors as having elevated TMB than the typical response rate to ICIs in this disease (49, 50). More work on this relationship is merited.

Finally, this work emphasizes the importance of considering HPV-negative HNSCC tumors based on subsite and clinical designation, rather than as a merged group. Recent studies have detected significant differences between tumor mutational profiles in oral, laryngeal, and other tumor subsites (35, 51–53), associated with differences in the prognostic potential for distinct mutations. Although the relatively large cohort analyzed here for the first time allows us to identify a number of intriguing trends for HNSCC tumors, the need to consider individual tumor subsites still restricts statistical power for some analyses. Notably, this work suggests the need for caution in basing assessments of mutational frequency on PAD data, at least for the genes studied here. Given evidence of consistent mutation frequency of TP53 and CDKN2A across tumor stages (Supplementary Fig. S2), we view the most likely factor contributing to the higher frequency of CDKN2A mutation frequency in the oral and laryngeal tumors in the CLS dataset as increases in detection efficiency with more recent technological advances. Conversely, a potential limitation of the current study is the incomplete characterization of HPV status in the CLS dataset; although analyses omitting oropharynx and sinonasal tumors were performed to reduce this concern, availability of large cohorts of tumors from these subsites with confirmed HPV-negative status will be a valuable addition. As the number of sequenced HNSCC tumors continues to increase, and technologies continue to improve, we anticipate the next few years will yield many further insights into the unique biology of distinct HNSCC subgroups.

No disclosures were reported.

A.Y. Deneka: Conceptualization, data curation, investigation, visualization, methodology, writing–original draft, writing–review and editing. Y. Baca: Conceptualization, data curation, software, investigation, visualization, writing–original draft. I.G. Serebriiskii: Conceptualization, data curation, software, formal analysis, investigation, visualization, writing–original draft. E. Nicolas: Data curation, visualization, methodology. M.I. Parker: Data curation, software, investigation. T.T. Nguyen: Investigation, writing–original draft, writing–review and editing. J. Xiu: Conceptualization, resources, formal analysis, supervision. W.M. Korn: Conceptualization, resources, formal analysis, writing–review and editing. M.J. Demeure: Validation, writing–original draft, writing–review and editing. T. Wise-Draper: Formal analysis, validation, writing–review and editing. A. Sukari: Formal analysis, validation, writing–review and editing. B. Burtness: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. E.A. Golemis: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.

We are grateful for support from the Fox Chase Cancer Center facilities for Genomics, and for Biostatistics and Bioinformatics. The authors were supported by NCI Core Grant P30 CA006927 (to Fox Chase Cancer Center), by DOD CA201045/W81XWH2110487 and by the William Wikoff Smith Charitable Trust (to E.A. Golemis), and by NIH P50 DE030707 (to B. Burtness).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.