Keywords

Introduction

Cancer evolution is now better understood at both a pan-cancer and tissue-specific level, with the work of international consortia such as The Cancer Genome Atlas (TCGA) [1] and Pan Cancer Analysis of Whole Genomes (PCAWG) [2, 3] providing us with increasingly rich molecular datasets, including for squamous cell cancers of the head and neck (HNSCC). Darwin’s theory of evolution was first directly applied to cancer by Peter Nowell, who hypothesized that from a single cell of origin, new cell lineages could evolve through genetic instability, with selection at a population level influenced by factors such as interaction with the immune system, metabolic adaptation to the microenvironment, and anti-cancer treatment [4, 5]. At the time Nowell made his observations, the most granular means of examining the genome was through karyotyping, direct visualisation of metaphase chromosomes within cancer cells allowing semi-quantitative assessment of chromosomal number and structure. Though Sanger sequencing enabled DNA characterisation at a greater degree of scale than had been previously possible, it has been the advent of massively parallel next generation sequencing in the last two and a half decades that made feasible the first sequencing of the human genome [6, 7], a project that remains ongoing [8].

These advances in sequencing technology, combined with complex computational analyses, now allow detailed examination of cancers at a genomic and transcriptomic level, through comparison to the germline DNA which the cancer genomes are derived from. The wealth of available cancer genomic data has allowed new insights into cancer biology, and revealed some of processes that underpin cancer evolution, such as the influence of mutational signatures [9,10,11], copy number and structural variation [12], and interaction with the immune system [13, 14]. Studies involving high-depth whole genome sequencing, and those with sequencing data from matched multi-regional samples can allow the timing of genomic aberrations to be inferred, revealing changes that occur early and late in tumour development [15]. At present, studies like these are extremely resource-intensive and challenging, and although the associated costs have been reducing as technology develops, much work remains to be done to demonstrate their clinical utility [16, 17].

The genomic landscape of HNSCC is now well-described, at least at a whole exome sequencing level [18]. However, there are limited whole genome data available for HNSCC as they are under-represented in the existing pan-cancer datasets [3, 18, 19]. There remains an ongoing paucity of data for metastatic disease. This article will discuss how new analytical approaches allow us to unpick the evolution of HNSCC before and during treatment, providing opportunities for novel therapeutic strategies.

Evolution of Treatment Naïve HNSCC

Genomic Landscape of HNSCC

There are now enough genomic data from a few sizeable clinical cohorts to confidently outline the genomic landscape of HNSCC, a key starting point for understanding the cancer-specific evolutionary processes. An important milestone for this was the initial publication from the TCGA in 2015, which included 279 cancers with whole exome sequencing of tumours and matched germline DNA, a much larger cohort than it had been previously possible to assess [20, 21]. The TCGA cohort has a strong bias towards HPV-negative disease, with 36/279 (12.9%) classified as HPV-positive, as defined by a significant number of mapped reads to E6 and E7 in the RNA data. Other biases included the dominance of oral cavity cancers (n = 172/279, 62%) and the heavy smoking history in the cohort, with a mean pack years of 51 [18]. The TCGA cohort has now been increased to 523 patients, largely confirming the previous observations and with a similar split between HPV-negative versus HPV-positive cancers (HPV + n = 72, 13.8%) [22, 23]. Nonetheless, this study has delivered key insights into HNSCC genomics and highlights differences in the biology of HPV-positive and HPV-negative cancer. As such it is an important starting point to understand HNSCC evolution.

Mutations in TP53 were confirmed to be near ubiquitous among patients with HPV-negative disease (86%), an observation hinted at in smaller cohorts [21, 24, 25], while only a single case of HPV-positive disease had a non-synonymous mutation in TP53 identified. Other genes noted to be more frequently mutated in HPV-negative disease were CDKN2A, the gene encoding the p16 protein, a key mediator of the cell cycle and of the G1/S checkpoint apparatus, and FAT1, a gene important in Wnt signalling, but which has also been implicated in resistance to cell cycle inhibitors through activation of CDK6 [26]. These findings confirm a phenotypic convergence in HNSCC towards dysregulation of cell cycle control, in HPV-negative disease through the functional loss of genes vital for cell cycle control but achieved in HPV-positive disease through expression of viral oncoproteins. Mutated NOTCH1 (17.1%) and CASP8 (10.1%) were prevalent in HPV-negative cancers but rarer in HPV-positive disease, CASP8 in particular being prevalent in oral cavity disease [18]. Compared to HPV-negative cancers, HPV-positive cancers demonstrate a higher prevalence of mutations in PIK3CA, the gene encoding the catalytic subunit of phosphoinositide 3-kinase (PI3K), an important intracellular signal transduction protein [19]. PIK3CA is remarkable for being the only frequently observed mutated oncogene in HNSCC, a cancer that is otherwise dominated in mutation terms by tumour suppressor genes. Increased prevalence of mutations in PTEN, the phosphatase and tensin homolog, upstream of PI3K signalling, highlights the importance of the PTEN/PI3K/mTOR pathway in HPV-positive HNSCC [19]. Other genes that may be more frequently mutated in HPV-positive disease as compared to HPV-negative disease include ZNF750, CASZ1, EP300 and FGFR3 [19], though much of the functional biology related to these specific alterations within the context of HNSCC is poorly understood at this time.

Combining the TCGA cohort with the whole genome data available from an HPV-positive cohort published by Gillison et al. demonstrates copy number changes that are commonly seen in both HPV-negative and HPV-positive cancers, such as loss of 3p, gain of 3q (seen across all squamous cell carcinomas [1, 27]), gain of chromosome 8 and loss of 9p [18, 19]. In some cases, such as with 9p, the loss is often observed to be relatively focal, specifically including the 9p21 locus that harbours CDKN2A, commonly observed to be mutated and providing evidence for phenotypic convergence towards loss of functional p16 resulting in dysregulated cell cycle. Copy number changes observed to be significantly different between HPV-positive and HPV-negative cancers included the 11q region containing CCND1, the gene encoding cyclin D1, which was gained in 14% of HPV-negative cases and lost in 17% of HPV-positive cases. A separate study of 108 HPV-negative cancers using a combination of whole exome and whole genome sequencing identified gain of 11q13.3 as being mutually exclusive with truncating mutations of FAT1, with concurrent proteomic analysis suggesting this signified convergent evolution towards dysregulated actin dynamics [28]. Losses in 11q, 13q, 14q and 16 were more commonly observed in HPV-positive cancers in the Gillison et al. analysis [19]. Interestingly, within the HPV-positive cohort, deletions in the region of RB1 were identified in approximately a third of HPV-positive cases (34%), a counterintuitive finding in light of expression of the E7 oncoprotein presumed to abrogate the inhibitory control exerted on the cell cycle by Rb [29], and perhaps an indication that further fitness advantage can be gained through additional attrition on the function of Rb even in the presence of the E7 oncoprotein.

Mutational and Copy Number Signatures

The availability of large pan-cancer sequencing datasets has made possible the development of a number of analytical approaches that shed light on the processes that drive mutagenesis and genomic variation in cancer genomes [30]. These approaches examine the specific mutations that occur across the genome within their genomic context, that is, the sequence immediately preceding and following the observed mutation [9, 31, 32]. Observing associations between these signatures and long understood clinical associations with various cancers (for example UV light [33, 34], smoking), along with experimental models, has allowed inference of causation and estimates of the relative importance for specific mutagenic processes within specific cancers [9, 11, 31, 35,36,37,38]. It is possible to extract mutational signatures from whole exome sequencing, but the power to discriminate between subtle differences is much greater with whole genome sequencing [39].

The mutational signatures of HNSCC have mostly been analyzed in the pan-cancer setting, but the available data do suggest some clear patterns. Some features are common to many cancers, such as the prominence of single base substitution (SBS) 1, one of the mutational signatures associated with ageing [35], and SBS 8, 16, 17 and 18, the cause of which remain unknown. An unsurprising finding is that SBS 4, related to tobacco, is prevalent in HNSCC [10, 11]. A single base substitution signature has not been confidently identified for alcohol, although certain patterns of mutations have been observed in cohorts of esophageal cancers [40, 41], and experimental models have shown that acetaldehyde exposure, an oxidation product of alcohol can lead to a double base substitution pattern [42]. Interestingly, one study examining a possible role for a new E. coli-related mutational signature in colorectal cancer also identified the same signature in one case of HNSCC [43].

In addition to these signatures, a further key evolutionary process identified for HNSCC, specifically HPV-positive disease, is the apolipoprotein B mRNA-editing enzyme, catalytic polypeptide (APOBEC), found in a high proportion of HPV-positive HNSCC [11, 19]. This family of cytidine deaminases are hypothesized to have evolved as a mechanism of cellular defence against DNA viruses, such as HPV, by causing mutagenesis in single stranded DNA during viral replication or transcription [44]. The mutational signatures, SBS2 and SBS13 [9, 10], have been ascribed to the APOBEC 3A and APOBEC3B enzymes deaminating cytosines, preferentially those immediately preceded by a thymine, though there may also be a role for DNA secondary structure for bases not following a thymine [45] (Fig. 6.1).

Fig. 6.1
2 bar graphs of percentage of single base substitutions for different trinucleotide contexts. In the one for S B S 2, the bars are plotted only under the C greater than T trinucleotide set with the highest percentage for T C A. For S B S 13, the majority of the bars are plotted under C greater than A and C greater than G sets. The highest percentage is for T C T under C greater than G set.

The two APOBEC-related single base substitution signatures, SBS 2 and SBS 13. Frequency of base changes is shown within a specific trinucleotide context. Data accessed from https://cancer.sanger.ac.uk/signatures/

Full size image

It is hypothesized that HPV infection drives APOBEC activity, potentiating mutagenesis indirectly as well as activating cellular growth and proliferation, this leading to increased genomic diversity and thus adaptability [46]. This is supported by modelling that suggests APOBEC activity could explain the excess of PIK3CA E542K and E545K mutations observed in HPV-positive HNSCC, both mutations being consistent with APOBEC activity [47, 48]. Mutations in PIK3CA are one of the most commonly identified in all cancers, usually clustered in one of two hotspots in the exon 9 helical domain and the exon 20 kinase domain [49]. In HNSCC compared to other cancers with a high frequency of PIK3CA mutations [50] there is a preponderance of the canonical helical domain mutations, E542K and E545K. This is also seen in the predominantly HPV-positive driven cervical carcinomas, favouring a role for associated HPV-related evolutionary processes. A large study of 1001 cell lines and 577 xenografts exposed to a variety of mutagenic stimuli observed marked variation in the APOBEC mutation rate over time, with multiple events clustered in time, for reasons that are unclear [51], but consistent with the phenomenon of kataegis, a process by which a large number of similar mutations occur in a focused area of the genome [52, 53].

More recently, in addition to unpicking mutational signatures by investigating genomic context, a number of groups have developed analogous techniques for copy number changes [54]. This involves identifying various features of copy number changes, often referred to as ‘genomic scars’ such as numbers of breakpoints, segment size and copy number aberration distribution across the genome and resolving them into cohesive patterns. This was first attempted in ovarian cancer and revealed considerable complexity in terms of the relationship between features of each signature, with seven different signatures identified in total [12]. The methodology was expanded upon in sarcoma, and later a pan-cancer cohort of approximately 10,000 to develop a total of 21 copy number signatures [55]. How copy number signatures may relate to the evolution of HNSCC remains an area that needs to be explored.

Intra-tumoural Heterogeneity and Timing Evolutionary Events

Though the gold standard for tracking cancer evolution is through analysis of longitudinally collected clinical samples, this is logistically challenging, often limited by the impossibility of obtaining samples from before the time of cancer diagnosis. Efforts to elucidate the critical early evolutionary events in head and neck cancer have been focused on pre-malignant disease, such as leukoplakia in the setting of oral cavity disease [56,57,58,59,60,61]. This approach is predicated on the model popularised by Fearon and Vogelstein in colorectal cancer [62] of incremental genomic changes which eventually promote outgrowth of a clonal population. This model is complicated by the observation that many pre-malignant lesions never progress to invasive malignancy for reasons that are poorly understood. The ‘field effect’ conceptual framework for HNSCC [63, 64], where cancers arise within a wider population of abnormal but not malignant cells, has been recently updated by findings of considerable genomic diversity within normal tissues [65, 66].

Intra-tumour heterogeneity can be inferred from sequencing a single tumour sample through deconvolution of the clonal architecture. At its simplest this can be achieved through analysis of the distribution of observed variant allele frequencies, with the largest peak consistent with clonal variants—those that are present in all cells of the cancer (Fig. 6.2). Peaks at lower frequency describe the presence of subclones. The spread of allele frequencies can be used as a crude metric for heterogeneity within a tumour, and has been shown in an analysis of the HNSCC TCGA to associate with poor clinical outcome [67, 68]. The accuracy of subclonal deconvolution can be improved through the integration of copy number, ploidy and tumour content, all of which can influence the measured allele fraction [31, 69,70,71]. In one study of whole exome sequencing of HPV-positive oropharyngeal cancers, increased heterogeneity based on a single-sample analysis was associated with a poorer relapse free survival [72].

Fig. 6.2
A histogram plots for density versus fraction of tumor cells. It has 4 vertical lines for 4 peaks to indicate clusters 7, 6, 5, and 4 with 538, 331, 174, and 8971 mutations, respectively. It shows a fluctuating trend and has a high peak at cluster 4.

An example of subclonal deconvolution using variant allele fraction. The clusters of mutations identify different populations of cancer cells, with the mutations in the largest peak being present in all of the cancer cells sampled

Full size image

The best data on which to attempt clonal deconvolution is high-depth whole genome sequencing, where adequate coverage to call tumour variants in addition to germline single nucleotide polymorphisms (SNPs) allows finer discrimination of subclonal mutations and changes in copy number number [73]. In the PCAWG analysis, subclonal deconvolution was possible in 34 cases of head and neck cancer, suggesting that only 25% of these had no subclones identifiable. Single sample whole genome or, to a lesser extent, whole exome sequencing can also be used to infer the order in which genomic aberrations were acquired [73]. Mutations and copy number changes that are identified as subclonal must be preceded by clonal changes, and mutations that can be ascribed to a particular copy number can be inferred to have occurred before or after a particular gain (Fig. 6.3) [73]. In the PCAWG cohort of head and neck cancer (n = 57) events that occurred early in evolution using this approach included loss of 9p, along with mutations in NOTCH1 and TP53, consistent with previous studies of pre-malignancy [58,59,60].

Fig. 6.3
A diagram depicts inferring time of genomic events. The samples in the left and middle has mutations acquired before gain. The Single Sample W G S in the right has mutations acquired both before and after number gain.

Example of inferring timing of genomic events. Data from a single sample are observed (far right). Subsequently it can be inferred that the yellow mutation was acquired before the copy number gain, whereas the green and red mutations were acquired after

Full size image

Sampling and sequencing multiple areas of a cancer can allow more detailed ordering of the observed genomic events [74, 75], categorising them into shared or private, and capture additional populations that would be missed with a single sample [76, 77]. This can also shed light on heterogeneity within a single cancer and provide potential clues to drivers of convergent evolution, as exemplified in the TRACERx study of lung cancer [76, 78, 79]. There are few studies that have directly assessed the sub-clonal heterogeneity of HNSCC with multiple sampling of the same tumour, and these are all from oral cavity cancers, where resection is the primary treatment and thus multi-regional samples easier to obtain. One small study of 5 patients with oral cavity SCC using whole exome sequencing found that the vast majority of mutations were conserved in the ~3 areas they sampled, though did not systematically examine copy number [80]. Another study of oral cavity SCC examined 44 biopsies from 13 cancers using shallow-depth whole genome sequencing to compare regions for copy number, finding relatively low levels of variation [81]. This seems at odds with the inferred sub clonal architecture seen in PCAWG, though all the datasets in question are probably too small to draw any firm conclusions. A further area of uncertainty is whether significant differences exist between primary HNSCC and lymph node metastases, with only a few studies directly examining this. One study using whole exome sequencing of matched primary and lymph node metastases in 13 patients found most mutations were shared (86%) [82], with another study involving single cell RNA sequencing of 5 cases of matched primary and lymph node suggesting tumours exhibiting signalling consistent with partial epithelial to mesenchymal transition were more associated with lymph node metastases. Larger studies along these lines will be needed to clarify intra-tumour heterogeneity and the molecular relationship between primary HNSCC and lymph node metastases.

Evolution and Considerations for Therapy in HNSCC

Developing Biomarkers and Therapies Informed by Evolution

There are currently no licenced therapies for HNSCC that incorporate genomic or evolutionary elements, but there are avenues along these lines that have shown some promise for the future. As discussed above, the PCAWG evolutionary timing analysis for the 57 patients in the head and neck cancer cohort highlighted with high confidence the loss of 9p as an early, clonal event in HNSCC [2]. Loss of 9p has been identified at a high prevalence within premalignant oral cancer lesions and associated with a higher risk of progression to malignancy [57, 58, 83], further highlighting its potential functional importance. A major pan-cancer study that included the HNSCC TCGA cohort found an association between aneuploidy and reduced host immune response, with this further associating to poorer outcomes on immune checkpoint inhibitors in a melanoma cohort [84]. Seeking to explore the immune consequences of 9p loss in HNSCC, William et al. examined a cohort of 188 cases of HPV-negative oral pre-malignant disease for copy number changes including 9p21.3 in addition to CD3+ , CD8+ and CD68+ cells assessed with multicolour immunofluorescence to characterise the immune infiltrate [85]. In this cohort, loss of 9p21.3 was not associated with increased infiltrate, although chromosomal gains such as trisomy and tetrasomy were. Recapitulating their analysis in the TCGA dataset, inferring immune infiltrate from RNA expression, the investigators observed that 9p loss in HNSCC was associated with reduced T cell infiltrate, so-called ‘immune-cold’ tumours, an effect that appeared to be driven by cases where there was 9p loss at the chromosomal arm level, rather than the more focused 9p21 loss typically associated with deletion of CDKN2A and the IFNA1 genes. Interestingly, these associations appeared to hold only for more advanced cases of HNSCC. The authors hypothesized that loss of 9p may be an important switch leading to a change in the microenvironment from immune-hot to immune-cold. Examining 9p as a potential biomarker, the investigators looked at 9p loss in a mixed clinical cohort treated with immune checkpoint inhibitors, finding that loss of 9p was associated with poorer prognosis, an effect not observed in an unrelated observational chemotherapy cohort [85].

A further example of evolutionary considerations applied to therapy is provided by the development of HRAS mutation-targeted therapies in HNSCC. In HNSCC, mutations in HRAS occur in 4–8% of cases, clustered around the activating hotspots in codons 13, 13 and 61 [18, 19, 86]. Tipifarnib is a farnesyl transferase inhibitor, a family of drugs that were developed to indirectly target the oncogenic activity of Ras through preventing its farnesylation, a key step in its localisation to the cell membrane and a pre-requisite for activation of its signalling. Though results in KRAS and NRAS-mutated cancers were disappointing, preclinical data suggested efficacy for tipifarnib in HRAS-mutant HNSCC cell line and xenograft models [87] leading to the KO-TIP-001 trial, an open label, phase II study of tipifarnib in HRAS-mutated HNSCC [88]. After an ad hoc analysis of the first 16 patients recruited to the trial, the protocol was amended to limit eligibility to patients with a variant allele fraction of >20% in their cancers, with 11/20 evaluable patients experiencing at least a partial response for an overall response rate of 55%. Selection of patients with specific mutations based on high variant allele fraction increases the chance that the mutation in question is clonally dominant, in theory improving the rationale for targeting the change, as subclonal populations with different molecular characteristics can exhibit varying responses to treatment [89]. This consideration of how clonally dominant specific mutations are may well be important when considering targeted therapies.

Moving beyond single molecular alterations, it is also possible that more abstracted evolutionary processes could be used to inform treatment in HNSCC. Tumour mutational burden (TMB) is effectively a composite output of the sum of mutational signatures acting upon a cancer genome, and potentially of neoantigen burden, and has been put forward as a candidate biomarker for response to immune checkpoint inhibitors [90]. This is supported by a large meta-analysis of over 1000 patients treated with immunotherapy which identified tumour mutational burden as the strongest predictor of response [91]. Two retrospective cohorts of HNSCC treated with immune checkpoint inhibitors have also demonstrated improved outcomes with high tumour mutational burden [92, 93]. Though these data seem to confirm a signal for tumour mutational burden and improved outcome on immune checkpoint inhibition in HNSCC, it is uncertain how this can be usefully integrated into the clinic, not least as there is debate around how best to define TMB [90]. Moreover, with ~20% of HNSCC qualifying as ‘high’ [94] and most patients with HNSCC eligible for access to checkpoint inhibitor therapy in the first or second line based on CPS anyway, it is not clear how TMB would be best integrated into the standard of care pathway.

Considering copy number signatures as a potential biomarker for treating HNSCC, Essers et al. examined a cohort of 173 patients with HPV-negative oropharyngeal, laryngeal and hypopharyngeal cancers who were treated with definitive chemoradiotherapy and performed low coverage whole genome sequencing to assess copy number signatures as defined by Macintyre et al. [12, 95]. Subsequent analysis and validation in the TCGA cohort identified a number of clinically relevant associations with copy number signatures. High signature 1 and 6 were associated with better and worse outcomes respectively, with 5 and 7 associated with increased frequency of distant metastases. Further work is needed to understand better the underlying biology that is driving these associations, but this study provides support for the concept of investigating evolutionary processes pertaining to copy number variation within the clinical paradigm.

Recurrent and Metastatic HNSCC—Evolution on Treatment

A knowledge of the genomic landscape of recurrent and metastatic HNSCC is an important starting point for understanding the relationship between molecular characteristics and treatment outcomes. Unfortunately, the genomic landscape of recurrent/metastatic HNSCC remains poorly defined at present, with available datasets limited by small size and heterogenous cohorts. The largest available cohort of recurrent/metastatic head and neck cancer with genomic characterisation comes from Memorial Sloan Kettering, with sequencing data from 151 patients using a 410 gene panel for mutations combined with low-depth whole genome sequencing for copy number [96]. Of these, 53/151 were HNSCC, the rest being accounted for by other head and neck malignancies, limiting the scope for characterising HNSCC. Nonetheless, some interesting comparisons could be made with primary HNSCC, including increased frequency of TERT-promoter mutation in HPV-negative HNSCC, most notably in tongue SCCs where it was seen in 91% (10/11) cases. HPV-positive tumours were observed to have fewer subclonal populations than HPV-, though data from targeted sequencing such as used in this study are not optimal for assessing this. Of note, recurrent/metastatic HPV-positive tumours were found to have a higher prevalence of features more commonly associated with HPV-negative disease, such as whole genome doubling and concurrent loss of 3p with TP53 mutation, suggesting these might be associated with a poorer prognosis ‘HPV-negative-like’ phenotype. The other available cohort to consider with regards the genomic landscape of recurrent/metastatic HNSCC is the Hartwig Foundation whole genome sequencing project for metastatic cancer, which included 42 cases of head and neck cancer in its pan-cancer analysis [97]. The spectrum of mutations and copy number variations in this set does not appear to depart significantly from the landscape of treatment-naïve HNSCC, but is too limited by size to make a meaningful comparison. It is likely that more will be learned as these cohorts of recurrent/metastatic HNSCC grow and genomic assays become more accessible.

The question of which genomic aberrations, if any, select for treatment resistance is best examined by using longitudinal matched tumour samples with matched clinical annotation. Due to the logistical challenges presented by generating these datasets, the available data are mostly limited to small cohorts, though these nonetheless present a valuable resource for hypothesis generation. Hedberg et al. examined a cohort of 10 HNSCC patients with whole exome sequencing from matched primary and metachronous recurrence samples, one of these cancers being HPV-positive [82] and five of them treated with radiotherapy. All the recurrences were in the upper aerodigestive tract and 9/10 were within 12 months of the initial treatment. Comparison of genomic profiles for primary versus metastasis was possible in 8 patients. Provocatively, only 60% of mutations were shared between the primary and metachronous occurrences, suggesting significant biological differences. In this particular study, many of the recurrences were from a distinctly separate anatomical site in the upper aerodigestive tract, inviting the hypothesis they had arisen from a distinct, but related population of either premalignant or treatment resistant cells. Similarly, a slightly larger cohort of 19 mainly HPV-negative patients subjected to the lower resolution technique of targeted sequencing with a 257 gene panel also found a significant proportion of the cohort had a markedly different mutational profile on relapse (31.9%, 6/19) [98]. Data on HPV-positive disease are even more limited, but a cohort of 7 matched primary/recurrence samples also showed substantial variation in the mutational spectrum between primary and relapse [99]. Focusing on patients treated with definitive chemoradiotherapy, de Roest et al. conducted low-coverage whole genome sequencing with targeted sequencing of 12 genes in 10 HPV-negative paired primary and relapse samples [81]. Here again, significant differences were seen in the genomic profiles of the primary versus the relapsed disease samples, even to the point where an algorithm trained on multi-regional data from primary cancers designated many of the relapses as ‘genetically unrelated’, even in some cancers which had relapsed within a few months of treatment [81].

The largest cohort of primary/relapse patients comprises 38 patients with HPV-negative disease, who had relapsed more than 6 months but less than 3 years after radiotherapy [100]. Using whole exome sequencing on DNA derived from FFPE samples, the investigators complemented their genomic analysis with paired RNA sequencing, additionally categorising the cancers sampled before and after relapse using the transcriptomics profiles described by Keck et al. [101]. Additionally, they performed single cell RNAseq on three patient-derived in vitro models. In addition to finding heterogeneity of transcriptional subtypes at a single cell level in the in vitro models, the investigators found variation in the longitudinal transcriptional subtypes observed between primary and relapsed, with a relative reduction in the frequency of the inflamed-mesenchymal subtype [100]. Of note, those patients in whom different transcriptional subtypes were observed tended to have longer time to relapse. Paired genomic analysis was conducted in 28 pairs and 79% were inferred to be genetically related, with an overall overlap of 70.9% of the top 20 mutated genes seen in paired samples [100]. Taken all together, these studies suggest a complex evolutionary relationship between primary and relapsed HNSCC, with some molecular elements that remain consistent with others changing or being selected through the treatment process. More insights will follow if analyses such as the above can be scaled in size and ideally integrated into clinical trials for high-quality clinical annotation.

Circulating Tumour DNA

Unpicking questions related to evolution in cancer on treatment ideally requires longitudinal tissue sampling, a process that involves discomfort and a degree of clinical risk for patients undergoing these procedures. However, in patients with cancer, a proportion of their cell-free DNA is derived from cancer, and as such can be used as a ‘liquid biopsy’ for the purposes of molecular characterisation [102,103,104]. This potentially offers a major opportunity to overcome some of the challenges of longitudinal tissue biopsies, with the blood tests required for plasma DNA collection being substantially less invasive than tissue biopsies. In addition to ease of sampling, circulating tumour DNA (ctDNA) may confer additional advantages over tissue biopsies for certain analyses, particularly with regards assessment of heterogeneity, where ctDNA may provide representation of a number of different metastatic deposits, rather than a single sample from a tissue biopsy [105, 106]. Novel analytical techniques have been able to move beyond the calling of mutations and copy number in ctDNA and integrate other characteristics such as fragmentation pattern and inferred nucleosomal occupancy to allow further exploration of tumour biology [107, 108]. That said, there are a number of technical challenges inherent in ctDNA analysis, the fraction of cell-free DNA that is derived from a patient’s cancer is often extremely low, especially in localised disease, meaning sampling effects can dominate assessments [109] and variant calling, particularly of copy number, can be challenging.

For HNSCC, ctDNA has predominantly been investigated as a tool to predict relapse, within the context of minimally residual disease. One particularly attractive area for this is in HPV-positive and EBV-positive disease, where the presence of viral DNA in the plasma provides a much easier target to differentiate for detection than mutated human DNA [110,111,112,113,114]. However, there are also data supporting this approach in non-virally driven HNSCC [115, 116]. There are few data at the present time for using ctDNA analysis to track longitudinal evolution of HNSCC on treatment, and this is an opportunity that needs to be exploited in the future.

Conclusions

Cancer cannot be understood without evolution. Evolutionary concepts are critical to understanding both how cancer develops and how it responds to treatment. Thanks to technological advances in sequencing and the efforts of international academic consortia we now have an approximate outline of the genomic landscape of treatment-naïve HNSCC, a starting point to unpick how those cancers develop, and why they respond so differently to treatment. The next steps are to improve our knowledge of recurrent/metastatic disease and begin to associate molecular characteristics with clinical phenotypes. More longitudinal studies of cancer evolution on treatment will be important in achieving these goals, with circulating tumour DNA a useful tool in taking this forward. To ultimately improve patient outcomes we need biologically-directed clinical trials embedded in a paradigm of forwards and reverse translation.