Background

Despite advances in multi-modal treatment of head and neck squamous cell carcinoma (HNSCC), mortality rates for advance disease remain high [1]. Thus there is an urgent need to identify novel chemicals with high activity in this disease. As with other tumor types, however, the time- and resource-intensive, multi-step clinical trial process remains a tremendous barrier to rapid drug development. Moreover, only specific molecular subtypes of tumors may respond to any given target agent [2], thereby decreasing the number of patients eligible for a particular study.

Targeted therapy has become an important method in personalizing treatment for cancer patients based on the genetic mutations present in their tumor(s). Such therapies enable the use of drugs to specifically target molecules within the tumor that are responsible for the malignancy. A search of the literature, as well as clinical trials that are currently underway in HNSCC, revealed a variety of agents being investigated that target various cellular molecules (e.g. epidermal growth factor receptor [EGFR], members of the phosphatidylinositide 3-kinase [PI3K] pathway, mammalian target of rapamycin [mTOR], cyclin-dependent kinases, vascular endothelial growth factor receptor [VEGFR], retinoblastoma protein [pRB], toll-like receptors and Aurora kinases) (clinicaltrials.gov). However, despite the multiple trials, only EGFR tyrosine kinase inhibitors and EGFR monoclonal antibodies (e.g. cetuximab) have been approved for clinical use and demonstrate only modest activity in a subset of patients [3]. New strategies are needed not only to identify active molecules, but also to define the target population that is most likely to benefit from therapy.

Cell lines are imperfect models of cancer: they tend to be generated from more aggressive, often metastatic tumors, can demonstrate genetic and epigenetic changes relative to the parent tumors, and lack interactions with the surrounding stroma and immune system [4–7]. However, they remain an invaluable discovery tool as they provide an unlimited source of self-replicating material, are easily manipulated and can be screened in a cheap and high-throughput way with large panels of drugs. Moreover, relationships between drug sensitivity and tumor genotypes observed in patient samples are also reflected in cell lines [8].

The advent of next generation sequencing has allowed complete, affordable and rapid genomic characterization of both patient samples and of cell lines. In parallel, the development of high-throughput robotic drug screening platforms has facilitated the rapid testing of a large number of drugs. Together these techniques provide the ability to correlate mutation status, copy number variation and expression levels with drug response. Two recent, large-scale studies, involving hundreds of cell lines of different tissue types [8, 9] have confirmed well known genetic markers of drug response (e.g. response to BRAF inhibitors in BRAF mutant cell lines) and identified novel associations such as the marked sensitivity of Ewing’s sarcoma cells harboring the EWS-FLI1 gene translocation to poly(ADP-ribose) polymerase (PARP) inhibitors [8]. However, given the large volume of data generated, only a limited analysis of the HNSCC cell lines involved in either study was presented. We endeavoured to reanalyze the data presented in these studies to provide a mutational landscape of HNSCC cell lines and to identify markers of drug sensitivity and resistance in HNSCC.

Methods

Defining the mutational and copy number landscape of HNSCC cell lines

The study by the Broad-Novartis group (Barretina et al.) included 31 HNSCC cell lines (of 947 total), seven of which were screened with 24 anticancer agents [9]. The cell lines were characterized by sequencing of ~1500 genes, as well as with array-based copy number variation (CNV) analysis and using mRNA abundance microarrays. A second study, by Garnett and coworkers, evaluated 639 cell lines (22 HNSCC lines) treated with 131 agents and characterized by targeted sequencing of 60 cancer genes, as well as array-based assessment of CNVs and mRNA abundance [8]. Note that eleven identically named HNSCC cell lines were common to both studies yielding a total of 42 uniquely named cell lines when both studies were combined. We integrated the CNV and mutational analysis of the most commonly altered genes from the two studies into Figures 1 and 2 and correlated them with the changes reported from patient samples by Stransky et al. [10]. CNV levels from Garnett et al., were simply reported as 0 (deletion), between 0 and 8 (copy-number neutral), and greater than 8 (amplification). Barretina et al. reported CNVs as continuous variables, relative to control genes with 0 considered “non-amplified”. We considered values greater than 2 (reflecting at least 2 extra gene copies) as amplifications and less than -2 (representing homozygous deletion) as this appeared to agree with the TCGA data from http://cbioportal.org and correspond best with the amplifications and deletions noted in the study by Garnett et al. (Additional file 1: Table S1).

Figure 1
figure 1

Genetic landscape of head and neck cancer cell lines based on data from Barretina et al., Nature 2012.

Full size image
Figure 2
figure 2

Genetic landscape of head and neck cancer cell lines based on data from Garnett et al., Nature 2012.

Full size image

Identification of biomarkers of chemotherapeutic sensitivity and resistance in HNSCC cell lines

Due to the small number of HNSCC cell lines that were treated with drugs in Barretina et al. (7 lines), we restricted drug sensitivity analysis to the data from Garnett et al. [8]. All statistical analysis was performed with the R statistical environment, version 2.15.2 (R Foundation for Statistical Computing) with the fdrtool package [11] version 1.2.10, to control the rate of false discovery due to multiple testing. We compared the half-maximal inhibitory concentration (IC50 in μM) for each drug between HNSCC cell lines and non-HNSCC cell lines via one-way analysis of variance (ANOVA) as computed through t-tests. Specifically, for each drug i, cell lines were partitioned into two groups j = {HNSCC, non-HNSCC} as per their cell-line type. Letting k denote replicate number, the linear model for each t-test was the standard y ijk  = μ j  + ε ijk , where y ijk represents the observed log2(IC50), μ j represents the mean response of group j, and each ε ijk represents a realization of ε~N 0 , σ i 2 . To control the false discovery rate, the “local false discovery rate” (LFDR) was estimated via computed p-values using Strimmer’s fdrtool [11, 12]. The LFDR has been championed by Efron and others for genomic studies as it is directly interpretable as posterior probability, and not a “corrected p-value” [13, 14]. A LFDR <0.05 was considered significant and a LFDR <0.1 was considered to be approaching statistical significance. We then looked for associations of copy number changes and mutations with response to drug treatment by two-way ANOVA including factor interaction, again using the LFDR to control false discovery rates. Specifically, the linear models used for the ANOVAs was y ijk  = μ + α i  + β j  + γ ij  + ε ijk where group i = {copy-number unchanged, copy-number changed}, group j = {wild-type, mutant}, and k again denotes replicate number. As per standard ANOVA, α i and β j represent the mean additive responses of their respective groups, γ ij represents any non-additive interaction effect, ε ijk represents a realization of ε ~ N(0, σ 2), and μ represents the grand-mean effect. The standard constraints ∑ i α i  = 0, ∑ j β j  = 0, and ∑ ij γ ij  = 0 were used to ensure that all parameters of each model were identifiable.

Results

The genetic landscape of HNSCC cell lines is similar to HPV-negative tumors

The mutational landscape of the 42 HNSCC cell lines, all of which were HPV-negative [7], demonstrated similarities with primary tumor samples from HPV-negative patients; including frequent mutations in tumor suppressor genes TP53 (74% of cell-lines [9]; 62% of tumors [10]) and CDKN2A, and less frequent ones in PTEN, SMAD4, NOTCH1 and NOTCH2 (Figures 1 and 2). Other similarities were rare activating mutations in oncogenes PIK3CA and HRAS, deletions of CDKN2A and amplifications of CCND1, epidermal growth factor receptor (EGFR), MYC and PIK3CA. A complete listing of mutations identified in HNSCC cell lines in Barretina et al. is provided in Additional file 2: Table S2. However, there were multiple, recurrent mutations in genes rarely or not identified in the patient samples (Additional file 2: Table S2 and Additional file 3: Table S3). In fact, there were 22 genes more frequently mutated than TP53, which was the most commonly mutated gene found in tumor samples (Additional file 3: Table S3). Most of these mutations were identical in all cell lines, such as two 5′ UTR mutations observed in neural cell adhesion molecule 1 (NCAM1) (insertion of adenine at position 112832307 (dbSNP ID: rs117108942) and deletion of cytosine at position 112832340) in virtually every HNSCC cell line in Barretina et al. Of note, 11 cell lines with the same name were characterized in both studies. However, two of these lines had significant discrepancies in terms of mutations between the studies (BHY, SCC9) bringing the true identities of the lines into question (Additional file 1: Table S1). Personal correspondence with the authors of Barretina et al. and the methods section of Garnett et al., have confirmed that the identification of their cell lines were confirmed with genoty**. The genoty** results are not provided in the supplementary data to allow direct comparison.

Chemicals with high and low activity in HNSCC cell lines

Four chemicals, Docetaxel (anti-mitotic chemotherapy), Bosutinib (combined SRC/ABL inhibitor), Afatinib (an EGFR and HER2 inhibitor), and Gefitinib (an EGFR inhibitor) were found to have significantly increased activity in HNSCC cell lines compared with the remainder of the cell line pool (Table 1, Figure 3). Two drugs, methotrexate and PD-173074 (an inhibitor of the fibroblast growth factor receptor [FGFR] and VEGFR) were found to have significantly lower activity in HNSCC lines (Table 1). A complete listing of the associations between drug response and HNSCC cell line type can be found in Additional file 4: Table S4.

Table 1 Drugs demonstrating significantly increased or decreased activity in HNSCC cell lines compared with non-HNSCC lines
Full size table
Figure 3
figure 3

Drug activity of HNSCC vs. non- HNSCC (“Other”) cell lines from Garnett et al ., Nature 2012. Points represent individual observations, while boxes show estimates of the respective median, interquartile range, and extrema.

Full size image

Of note, there were five identically named cell lines in both studies that were tested with identical drugs (Additional file 5: Table S5), four of which had similar mutational profiles suggesting that they were indeed identical lines. SCC-9 was excluded due to discrepancies in the mutational profiles reported by Barretina et al. and Garnett et al. (Additional file 1: Table S1). We sorted the comparable IC50s into three groups, representing cell lines that were exquisitely sensitive (IC50 < 3 μM) to the drug, responders (IC50 3.1-7.9 μM) and resistant cell lines (IC50 > 8 μM). We found that the majority were comparable (Additional file 5: Table S5).

Activating PIK3CA mutations are correlated with response to the PI3K inhibitor, AZD6482

The complete listing of drug sensitivity and gene status can be found in Additional file 6: Table S6, with the significant findings summarized in Table 2. Only mutations and copy number changes in EGFR, TP53, CDKN2A, PIK3CA and SMAD4 were present with sufficient frequency (>10%) in the cell lines to allow analysis. It should be noted that not all cell lines were treated with every drug and some genetic changes occurred in a very small number of cell lines, which resulted in exclusion of analysis of certain drugs with a particular genetic change. Of the 131 drugs tested, three were PI3K inhibitors including AZD6482, GDC0941, and the combined mTOR and PI3K inhibitor NVP-BEZ235. We calculated a robust increase in sensitivity to AZD6482, explainable by the interaction of PIK3CA mutation status and HNSCC cell-line type (LFDR <0.023, Figure 1). In addition, an increase in AZD6482 sensitivity was shared by all PIK3CA mutants (p <0.037, Figure 4A) regardless of cell line type. No association was observed for the other PI3K inhibitors, GDC0941 and NVP-BEZ235, and PIK3CA mutation status (LFDR ≈ 1). There were too few PIK3CA amplified cell lines to examine the effect of amplification alone on drug response, however when these were pooled with the PIK3CA mutant lines, no drugs were found to be preferentially active when compared to PIK3CA wild-type cell lines (Additional file 6: Table S6).

Table 2 Significant associations of mutations and amplifications with drug response in HNSCC cell lines
Full size table
Figure 4
figure 4

Drugs with differential activity by mutational status. (A) PI3K inhibitor AZD6482 demonstrates increased activity in PIK3CA mutant versus wild-type cell lines. (B) When analysis was restricted to HNSCC cell lines, AZD6482 and FAK inhibitor PF-562271 demonstrated increased activity in PIK3CA mutant lines. (C) AZD6482 and JNK Inhibitor VIII had increased activity in EGFR amplified cell lines relative to wild-type lines. Points represent individual observations, while boxes show estimates of the respective median, interquartile range, and extrema.

Full size image

When examining responses to inhibitors of upstream and downstream members of the PIK3CA pathway, the FAK inhibitor PF-562271 (upstream) demonstrated a trend towards selective inhibition of PIK3CA HNSCC mutant cell lines (LFDR = 0.079, Figure 4B), while no effect was observed for downstream inhibitors including three AKT inhibitors (AKT inhibitor VIII, MK-2206, A-443654) and four mTOR inhibitors (Rapamycin, Temsirolimus, JW-7-52-1, AZD8055).

There was a trend towards increased sensitivity to AZD6482 and JNK Inhibitor VIII in cell lines with EGFR amplifications (LFDR = 0.056, Figure 4C). The strongest association observed was increased activity of the retinoid receptor antagonist ATRA in TP53 mutant lines (LFDR = 0.007, Table 2 and Additional file 6: Table S6).

Discussion

Following decades of active research, only one class of targeted molecular agents, epidermal growth factor receptor (EGFR) inhibitors, have been approved for use in head and neck cancers [15]. Despite a modest survival benefit when administered concurrently with radiation, response rates to EGFR inhibitors are low when given alone (13%) and of limited duration (2–3 months). More effective drugs are needed in order to improve outcomes and reduce treatment-induced morbidities for HNSCC patients.

By pairing next-generation sequencing of cell lines with high-throughput drug screening techniques, the impressive studies by Garnett et al. and Barretina et al. [8, 9], confirmed, in multiple tissue types, known associations of genetic alterations with drug sensitivity and uncovered a multitude of new ones. Their sequencing findings were in agreement with preliminary data from The Cancer Genome Atlas (TCGA) HNSCC study, where a multitude of potentially druggable targets including amplifications (e.g. FGFR1, CCND1, MYC, EGFR), deletions (e.g. PTEN), activating mutations (e.g. PIK3CA) and fusions (e.g. FGFR3/TACC3) were observed. Despite a relatively limited number of cell lines, our HNSCC-specific reanalysis of these studies shows that the spectrum of mutations observed in HNSCC cell lines is similar to that of primary HNSCC patient samples [10, 36], and are thus less likely to recapitulate the treatment-sensitive HPV-positive tumors encountered in clinical practice. We suggest that the development of further HPV-positive cell lines and their incorporation into large-scale HNSCC cell line drug screening studies has the potential to identify novel effective agents and the mechanisms of drug sensitivity and resistance in HNSCC. Hopefully, this will lead to significant improvements in survival that has eluded us to date.

The disagreements noted in the cell line sensitivities and mutations between the two studies have significant implications for future work of this type. They are many factors that can explain these discrepancies, however the most likely is that the identically named lines are in fact different, despite the fact that genoty** was completed in both studies. Other possible sources of disagreements in the data are differences in screening techniques, different drug concentrations, and different statistical models to calculate IC50 values from the dose–response curves. Ideally, the genoty** data can be compared to determine the discrepancies and provide the definitive genotype for cell lines. We also suggest that a standard methodology of cell line drug screening needs to be developed to allow external validation of future findings.

Conclusions

High throughput drug screening of molecularly characterized HNSCC cell lines has the potential to rapidly identify promising agents to improve therapies for patients suffering with head and neck cancer. An expanded HNSCC specific study including HPV-positive cell lines has the potential to identify effective agents, as well as mechanisms of resistance and sensitivity to molecular agents.

Ethics approval

No ethics approval was required as this present work represents an investigation and analysis of publicly accessible data.