Transcriptional subtypes of glottic cancer characterized by differential activation of canonical oncogenic programming

Abstract

Background:

There is a paucity of data concerning molecular heterogeneity among glottic squamous cell carcinoma, and the clinical implications thereof.

Methods:

Data corresponding to glottic squamous cell carcinoma were derived from The Cancer Genome Atlas. The Onco-GPS computational methodology was levied to derive four patterns of transcriptional activity and three functional subtypes of glottic cancer.

Results:

Thirty glottic cancer samples stratified to 3 distinct oncogenic states (S0-S2) based on a Onco-GPS model containing 4 transcriptional components (F0-F3). Membership in S2 and association with transcriptional component F0 conveyed an invasive phenotype, with transcriptional activity strongly reflecting EMT programming (including TGF-B and NF-KB signaling). S2 membership also correlated with inferior disease-specific survival (HR 9.027; 95% CI 1.021–79.767), and higher incidences of extracapsular spread and perineural invasion.

Conclusions:

We present a functional taxonomy of glottic cancer, with subtypes demonstrating differential up-regulation of canonical oncogenic networks and survival implications.

Introduction

Glottic squamous cell carcinoma (GSCC) comprises approximately 60% of new laryngeal cancer diagnoses, and is associated with the highest 5-year relative survival rate (77%) across stages among the three laryngeal subsites. The more favorable prognosis correlated with GSCC is attributable in part to the high proportion of localized disease at the time of diagnosis (83%) and the lower overall propensity for regional and distant metastatic disease.¹ However, as is common to all cancer types, a subset of GSCC progresses or recurs despite receipt of stage-appropriate therapy. Both clinicodemographic and molecular features putatively correlated with treatment failure or inferior survival among patients with laryngeal cancer have been described in the extant literature.^1–11 This empirical data, however, has not been, or has not been able to be, leveraged to affect clinical outcomes--a recent Surveillance, Epidemiology, and End Results (SEER) analysis, for example, stipulated a 25% relative increase in case-fatality rate between 1986 and 2016, based on a larynx cancer dataset wherein the majority of tumors were of glottic origin.¹² Though a few published efforts to systematically sequence the laryngeal cancer genome have led to the identification of discrete genetic perturbations^13–15, the complex interplay between the wide spectrum of tumor molecular features and both patient-centric and disease-centric clinical characteristics necessitates a deeper, coherent approach to understand the functional consequences of these genetic alterations to guide improvements in GSCC treatment and prognostication. Moreover, investigations into prognostic molecular changes among laryngeal cancers that group all subsites together neglect the potentially disparate clinical implications and heterogenous downstream effects of these genetic perturbations; in fact, studies have identified significant molecular differences among tumors originating from the different laryngeal subsites.¹³

The Cancer Genome Atlas (TCGA) profiled a series of laryngeal squamous cell carcinomas, furnishing the full spectrum of multi-omic data with correlated clinical information, making available for public analysis an extensive array of somatic and epigenetic alterations among laryngeal tumors.¹⁶ Using this dataset and the OncoGenic Positioning System (Onco-GPS) data-analytic and cancer classification methodology, we aimed to determine if an explicitly computationally derived transcriptional taxonomy and associated subtypes of glottic squamous cell carcinoma might convey prognostic information and elucidate the specific transcriptional features underlying divergent molecular and clinical phenotypes.¹⁷ Our approach, which is driven primarily by a search for transcriptomic subtypes of GSCC, is distinct from a number of extant publications that simply report discrete genetic perturbations that are independently associated with survival implications, a methodologic tactic that ignores the inextricable context-dependencies of individual genetic alterations. We aim for the data and analysis presented in the paper herein to serve as a platform for future, prospective investigations into the molecular complexities of GSCC.

Materials and Methods

Data Acquisition and General Description of Data

Patient data from the TCGA head and neck cancer analysis (clinicodemographic characteristics, level 3 DNA mutation packager calls, and mRNA-seq gene expression data) were derived from a public repository furnishing the dataset (Broad Institute’s Firehose Database).^16,18 Although all TCGA multi-omic characteristics were considered, our analysis was primarily contingent on transcriptomic data, accounting for the presence of missing RPPA, methylation, and mutation data among glottic tumors. Clinical annotations do not contain information regarding laryngeal tumor subsite; as such, pathology reports derived from laryngeal cancer patients were reviewed (BAP and HJ) to identify patients with primary glottic tumors. T-classification, N-classification, and M-classification data were based on pathologic staging. Additional clinical and demographic factors considered included age at diagnosis, race, sex, and empirical disease-centric prognosticators (rates of lympho-vascular invasion, perineural invasion, extracapsular extension); and, tobacco pack-years. Differences in clinical and demographic features among the defined oncogenic states were assessed using the t-test for continuous variables and the chi-squared test for categorical variables.

OncoGenic Positioning System (Onco-GPS) Detailed Methodology

Onco-GPS is a computational method designed to use Non-Negative Matrix Factorization (NMF) to decompose gene expression profiles into summary components that are used in order to define functional tumor subtypes aptly referred to as “oncogenic states.” A full description of the methodology can be found in Kim et al.¹⁷ Only expression data corresponding to genes represented in a selected set of a priori oncogenic pathway (C6) as characterized in the Molecular Signatures Database (MSigDB) were included in the decomposition process to emphasize the most relevant oncogene-driven transcriptional environment and mitigate the effects of other transcriptional activity.¹⁹ The list of oncogenic pathways is otherwise unfiltered and collectively represent a wide array of biological processes that facilitate identification of transcriptomic relationships among GSCC samples in an unbiased fashion. The resulting Onco-GPS model or “map” provides a robust framework to examine the complex, wide-ranging functional differences and similarities, originating from differences in molecular features including gene and pathway expression among tumor samples.

Specifically, mRNA-seq gene expression data was used as the base criteria for the Onco-GPS map. Non-negative matrix factorization (NNMF) was applied to the collection of TCGA GSCC tumor sample gene expression data to identify discrete patterns of transcriptional activity, each of which constituted a transcriptional component. Tumor samples are then assigned scores corresponding to their transcriptional similarity to each of the components; tumor samples with similar component correlations stratify into oncogenic states, each of which represents a functionally distinct molecular subtype of our GSCC tumor cohort. The number of components and states comprising our Onco-GPS map was explicitly chosen to produce a model with an appropriate degree of granularity to decisively illustrate the molecular heterogeneity of our cohort, and was corroborated by a robust cophenetic correlation coefficient (CCC). Each transcriptional component and oncogenic state were subsequently biologically annotated, using primarily transcriptomic and clinicodemographic information, to understand the practical implications of the cancer classification model. (Supplementary File A) Transcriptomic annotations were prioritized in annotating the Onco-GPS map given missing protein expression and mutational data. Single-sample gene set enrichment analysis (ssGSEA) was performed to obtain relative oncogenic and cellular pathway expression scores, which were used to annotate each oncogenic state and transcriptional component, to provide a more coherent interpretation of discrete gene expression data.^19,20 Drug response gene sets were derived from DSigDB, a manually and computationally curated collection of drug response data, and also used to annotate the Onco-GPS map.²¹ Multi-omic correlations, according to the Information Coefficient, with components and oncogenic states are reported, and the threshold for statistical significance was considered as p<0.01.

Survival Analysis

Cox proportional hazards regression analysis was run to assess predictors, including oncogenic state, of disease-specific survival (DSS). DSS was prioritized over overall survival as the survival outcome, given the absence of patient comorbidity data that would otherwise be more critical to consider in an overall survival analysis. Multivariable regression models for each survival outcome were fit to include parameters found to be significant by univariable regression analysis, or otherwise deemed to be clinically pertinent. Regression analysis assessing the predictive value of selected, individual molecular perturbations most highly correlated with transcriptional components and oncogenic states was also performed, in order to assess the relative value of Onco-GPS component and oncogenic state assignments.

Results

Baseline Characteristics Corresponding to GSCC Samples

Thirty GSCC samples were identified in the TCGA laryngeal squamous cell carcinoma dataset. In this cohort, mean age at diagnosis was 62.69 (SE=2.08), proportion of patients of male sex was 90.0% (n=27); proportions of advanced stage disease were 13.33% (stage III, n=4) and 76.67% (stage IV, n=23); proportion of Caucasian patients was 70% (n=21); and, mean tobacco pack-years was 51.05 (SE=4.93). Majority of tumors were moderately differentiated (n=18; 62.07%) versus well differentiated (n=4; 13.79%) or poorly differentiated (n=7; 24.14%). Incidence of perineural invasion and lympho-vascular invasion were 6.67% (n=2; 7 with missing data) and 26.09% (n=6; 7 with missing data). (Table 1)

Table 1.

Clinicodemographic paramaters among GSCC patients in State 2 versus other.

Parameter	Overall	State 2	State 0 and 1	P-value
Age (years), mean (SE)	62.7 ± 10.6	63.6 ± 10.3	61.8 ± 11.2	0.67
Sex				0.386
Male	27 (90%)	11 (84.6%)	16 (94.1%)
Female	3 (10%)	2 (15.4%)	1 (5.88%)
Race				0.238
Caucasian	21 (70%)	8 (61.5%)	13 (76.5%)
Other	8 (26.7%)	5 (38.5%)	3 (17.6%)
Tobacco pack-years, mean (SE)	51.1 (4.9)	40.1 (5.9)	58.3 (6.6)	0.069
T-classification				0.693
T1	1 (3.33%)	1 (7.69%)	0 (0%)
T2	3 (10%)	1 (7.69%)	2 (11.8%)
T3	5 (16.7%)	2 (15.4%)	3 (17.6%)
T4	21 (70%)	9 (69.2%)	12 (70.6%)
N-classification				0.405
N0	17 (56.7%)	7 (53.8%)	10 (58.8%)
N1	6 (20%)	2 (15.4%)	4 (23.5%)
N2	5 (16.7%)	2 (15.4%)	3 (17.6%)
N3	2 (6.67%)	2 (15.4%)	0 (0%)
M-classification
M0	30 (100%)	13 (100%)	17 (100%)	--
Overall AJCC Stage				0.595
I	1 (3.3%)	1 (7.7%)	0
II	2 (6.7%)	1 (7.7%)	1 (5.9%)
III	4 (13.3%)	1 (7.7%)	3 (17.6%)
IV	23 (76.7%)	10 (76.9%)	13 (76.5%)
Surgical Margin Status				0.665
Negative	25 (83.3%)	10 (76.9%)	15 (88.2%)
Close or Positive	3 (10.0%)	2 (15.4%)	1 (5.9%)
Unknown	2 (6.7%)	1 (7.69%)	1 (5.9%)
Extracapsular Extension				0.024
No	23 (85.2%)	7 (53.8%)	16 (94.1%)
Yes	4 (11.1%)	4 (30.7%)	0
Unknown	3 (10.0%)	2 (15.4%)	1 (5.9%)
Perineural Invasion (PNI)				0.035
No	21 (70.0%)	6 (46.2%)	15 (88.2%)
Yes	2 (6.67%)	2 (15.4%)	0
Unknown	7 (23.3%)	5 (38.5%)	2 (11.8%)
Lymphovascular Invasion (LVI)				0.157
No	17	5 (38.5%)	12 (70.6%)
Yes	6	3 (23.1%)	3 (17.6%)
Unknown	7 (23.3%)	5 (38.5%)	2 (11.8%)
Tumor Grade				0.48
Well differentiated	4 (13.3%)	1 (8.3%)	3 (17.6%)
Moderate differentiated	18 (60.0%)	9 (75%)	9 (52.9%)
Poor differentiated	7 (23.3%)	2 (16.7%)	5 (29.4%)

Survival Analysis

Among GSCC samples, membership in State 2 was associated with a significant DSS disadvantage (HR 9.027; 95% CI 1.021–79.767, p=0.048). More advanced pathologic N-classification (HR 3.01; 95% CI 1.22–7.38, p=0.016), PNI (HR 9.38; 95% CI 1.42–62.19, p=0.020), and extracapsular spread (HR 9.09; 95% CI 1.50–55.11; p=0.016) among the spectrum of clinicodemographic parameters assessed, were predictive of DSS. A multivariable regression model (Table 2) including membership in State 2 and significant univariable predictors was fit, revealing more advanced pN classification (HR 3.51; 95% CI 1.02–12.07, p=0.047) as the only covariate that retained statistical significance. Membership in State 2 remained on the margin of statistical significance (HR 18.79; 95% CI 0.77– 456.53, p=0.072). A multivariable regression model was fit including all laryngeal tumors from the TCGA analysis (including supraglottic and subglottic tumors) to further explore the incremental survival implications of glottic tumor membership in State 2. Glottic tumor membership in State 2 (HR 3.63; 95% CI 1.11–11.89) conveyed worse disease-specific survival compared not just to glottic tumors, but against all laryngeal tumors, of which supraglottic and subglottic tumors empirically convey a worse prognosis, accounting for sex, more advanced nodal disease, and ECS. (Table 3)

Table 2.

Cox proportional hazards regression analysis assessing predictors of disease-specific survival.

Univariable				Multivariable
Parameter	Hazard Ratio	95% CI	P-value	Hazard Ratio	95% CI	P-value
Age	1.04	0.95–1.14	0.371
Sex	2.06	0.24–17.70	0.51
Race	0.28	0.05–1.44	0.128
Tobacco pack-years	0.99	0.94–1.04	0.623
pT	1.42	0.41–4.98	0.58
pN	3.01	1.22–7.38	0.016	3.51	1.02–12.07	0.047
Surgical margin status
Negative	--	--	--
Close or Positive	2.82	0.31–25.43	0.356
Unknown	2.56	0.29–22.95	0.401
ECS
No	--	--	--	--	--	--
Yes	9.09	1.50–55.11	0.016	0.98	0.04–24.49	0.991
Unknown	3.11	0.28–34.47	0.354	13.96	0.49–397.76	0.123
PNI
No	--	--	--	--	--	--
Yes	9.38	1.42–62.19	0.02	1.14	0.05–24.26	0.932
Unknown	1.21	0.13–11.68	0.867	0.22	0.01–6.08	0.374
LVI
No	--	--	--
Yes	1.79	0.30–10.75	0.524
Unknown	0.95	0.10–9.10	0.961
State 2	9.03	1.02–79.77	0.048	18.79	0.77–456.53	0.072

Table 3.

Cox proportional hazards regression analysis assessing predictors of disease-specific survival among all laryngeal tumors.

Univariable				Multivariable
Parameter	Hazard Ratio	95% CI	P-value	Hazard Ratio	95% CI	P-value
Age	1.05	0.99–1.11	0.078
Female sex	3.31	1.16–9.47	0.026	2.58	0.76–8.72	0.127
Race
Non-White	--	--	--
White	0.28	0.05–1.44	0.128
Tobacco pack-years	0.99	0.98–1.02	0.838
pT	0.802	0.46–1.41	0.446
pN	1.95	1.13–3.35	0.016	1.95	1.05–3.65	0.036
Surgical margin status
Negative	--	--	--
Close or Positive	2.63	0.57–12.08	0.213
Unknown	2.37	0.81–6.97	0.115
ECS
No	--	--	--
Yes	10.62	2.77–40.79	0.001	6.34	1.58–25.34	0.009
Unknown	5.57	1.38–22.48	0.016	6.52	1.41–30.27	0.017
PNI
No	--	--	--
Yes	2.85	0.71–11.42	0.141
Unknown	4.02	1.23–13.10	0.021
LVI
No	-	--	--
Yes	2.27	0.64–8.06	0.205
Unknown	3.02	0.88–10.37	0.079
State 2	3.19	1.12–9.10	0.03	3.63	1.11–11.89	0.033

To establish the relative value of oncogenic state assignment versus individual pathway or gene expression, gene sets among the most highly represented in S2 (including gene sets reflecting a pattern of gene expression associated with EMT) were separately assessed via Cox regression analysis. Level of VIM expression (HR 1.01; p=0.034) was marginally predictive independently of DSS; relative activation of HALLMARK_EPITHELIAL_MESENCHYMAL_TRANSITION (p=0.216) and WP_FOCAL_ADHESION_PI3KAKTMTOR)PATHWAY (p=0.182) were not independently predictive of DSS.

Onco-GPS Map and Constituency

We generated an Onco-GPS map of GSCC tumors stipulating three oncogenic states (State 0, State 1, State 2) and four transcriptional components (Component 0, Component 1, Component 2, Component 3). (Figure 1) Oncogenic state constituency was relatively even (State 0, n=8 samples; State 1, n=9; State 2, n=13), and the corresponding cophenetic correlation coefficient exceeded 0.9. Significant differences in proportions of both extracapsular spread (30.7% versus 0.0%) and perineural invasion (15.4% versus 0.0%) between State 2 and all other glottic samples were observed. All other clinicodemographic features were homogenous throughout glottic cancer samples. (Table 1)

Figure 1.

Onco-GPS map contingent on 4 transcriptional components (F0-F3) stipulating 3 oncogenic states (S0-S2). Samples aggregate on the map and into states based on their transcriptional similarity to each of the 4 components. The number of glottic squamous cell carcinoma samples in each state are represented in the figure. Representative clinical, transcriptomic, and pathway activation annotations corresponding to all states, and the components of interested (F0 and F3) are also represented in this figure.

Summary of Component Annotations

Component F0 was affiliated most strongly with samples in State 2 (Figure 1) and was characterized by relative parallel, activation of several oncogenic mechanisms. Relative pathway activation and individual gene expression patterns correlated with epithelial-to-mesenchymal transition (EMT) and a mesenchymal phenotype were overwhelmingly overrepresented (Figure 2) in the transcriptional patterns defining F0 (Figure 2), including TGF-β (information coefficient [IC]=0.84; p<0.0001) and NF-kB signaling (IC=0.76; p<0.0001; vimentin (VIM) expression (IC=0.80; p<0.0001), and relative under-expression of E-cadherin (CDH1) (IC=−.0.37; p<0.0001). Moreover, multiple gene sets correlated with downstream KRAS pathway activation (IC=0.90; p<0.0001), including PI3K/AKT/mTOR signaling (IC=0.90; p<0.0001), were robustly associated with F0. Component F3, the other component framing the transcriptional environment of samples in State 2, (Figure 1) was characterized by relative activation of oncogenic networks empirically associated with laryngeal carcinogenesis including the EZH2 pathway²² (IC=0.70; p<0.0001), and a spectrum of dysregulated miRNA signaling (miR3173, IC=0.83; miR494_5P, IC=0.0.79; miR410_5P, IC=0.79; miR4750_3P, IC=0.79). Component F1 was defined by a relative pattern of gene expression correlated strongly with a number of ultraviolet radiation (UV) response gene sets (IC=0.87; p<0.0001), oxidative phosphorylation (IC=0.79; p<0.0001), and activation of EIF4E (IC=0.78; p=0.004) and β-catenin signaling (IC=0.74; p=0.004). Component F2 was defined by a pattern of gene expression imputing lipid (IC=0.82; p<0.001), glutathione (IC=0.82; p<0.001), and α-amino acid metabolism (IC=0.83; p<0.001), and in NFE2L2 (NFR2) signaling (IC=0.75; p=0.004), a transcription factor hypothesized to confer a cytoprotective effect in tumor cells against oxidative injury.

Figure 2.

Heat map demonstrating correlation between representative gene expression sets (derived from MSigDB) and the EMT phenotype characterizing component F0. Information coefficients and P-values describing the strength of the correlations are reported. VIM=vimentin.

Summary of State Annotations

Transcriptional annotation of S2 correlated evidence of relative up-regulation of gene sets affiliated with KRAS signaling (IC=0.75; p=0.001); protein expression and mutational data (i.e. KRAS), however, did not consistently demonstrate perturbations correlated with commonly dysregulated components in these pathways, suggesting that parallel transcriptional events might underlie the upregulation of this oncogenic network. (Figure 1) Identification of common genes in pro-tumorigenic pathways most highly correlated with State 2 (e.g. KRAS and Rb) elucidated a series of genes empirically associated with invasiveness and metastasis, including inhibin A²³ (IC=0.70; p<0.001), ETS1²⁴ (IC=0.68; p<0.001), and DCBLD2²⁵ (IC=0.62; p<0.001). Additional gene expression sets corresponding to cellular motility (IC=0.66; p<0.001) and an invasive phenotype (i.e. TGF-β signaling [IC=0.64; p<0.001]) were also fairly well correlated with State 2. Moreover, genes targeted by empirically pro-tumorigenic mi-RNAs including miR-1179²⁶ (IC=0.71; p<0.001), hypothesized to levy an effect on tumor invasiveness via modulation of MEK/ERK and PI3K/AKT signaling; miR-337²⁷ (IC=0.70; p<0.001), and miR-605 (IC=0.67; p<0.001) were up-regulated among S2 samples. Hallmark transcriptomic correlations with State 1 strongly imputed an inflammatory phenotype, including non-canonical NF-kB activation (IC=0.78; p<0.001), UV radiation signature gene sets (IC=0.74; p<0.001), and drug response gene expression signatures corresponding to metformin hydrocholoride response (IC=0.75; p=0.001) and JAK inhibition (IC=0.64; p=0.001). Additional S1 correlated gene sets corresponded to mitochondrial respiration (IC=0.74; p<0.001). State 0 was characterized by negative relative expression of genes correlated with a mesenchymal phenotype and invasiveness, including VIM (IC=−0.57; p<0.001), MMP9 (IC=−0.56; p<0.001) and MMP13 (IC=−0.51; p<0.001), CNN3 (IC=−0.67; p<0.001), CAP1 (IC=−0.55; p<0.001), and relative, elevated expression of the p63 protein (IC=0.57; p=0.004). No robust correlations were observed between mutations or protein expression and oncogenic state.

Discussion

There are few investigations in the extant literature exploring the molecular heterogeneity of glottic cancers to better understand heterogeneity in clinical outcomes. We present preliminary data contingent on the TCGA laryngeal cancer dataset demonstrating three distinct transcriptional subtypes of GSCC; tumors in State 2 demonstrated coordinately up-regulated oncogenic programming corresponding to EMT and a more invasive phenotype, conveying clinically with significantly higher rates of both PNI and ECS and worse disease specific survival (HR 9.027; 95% CI 1.021–79.698) Importantly, these were findings from a cohort made up of patients and tumors with otherwise homogenous clinicodemographic characteristics (e.g. age, sex, TNM staging criteria at time of diagnosis, and tumor grade). The paradigm of this study, which was to target a clinically homogenous study population to elucidate molecular features correlated with worse DSS, more closely mirrors a clinical reality wherein a subset of GSCC, despite glottic tumors’ lower overall propensity for regional and distant disease, fails stage-appropriate therapy. By constructing an oncologic map (Onco-GPS) before considering the survival implications of molecular perturbations, we prioritized the identification of clinically relevant transcriptional trends and pathway activation. Importantly, glottic tumors in State 2 demonstrated significantly worse DSS in our multivariable model including both supraglottic and subglottic tumors, subsites which empirically demonstrate worse outcomes. The collective molecular and clinical characteristics of State 2 describe a more aggressive phenotype, and this analysis should remain a consideration for providers managing glottic tumors, wherein a set of coordinated, biological features might convey a higher probability of recalcitrant disease.

Samples in State 2 most closely resembled F0, the transcriptional milieu of which represented EMT (Figure 2), with multiple associated oncogenic signaling networks (TGF-β and PI3K/AKT/mTOR) coordinately upregulated, as well. Though EMT programming is an active area of oncology research, the survival implications of EMT in laryngeal cancer are well-documented in the extant literature; tumor cells acquire migratory and metastatic capacity, and via a spectrum of parallel effects (e.g. upregulation of metalloproteinases, immune tolerance, neovascularization, and acquisition of stemness), demonstrate increased invasiveness, and more aggressive and metastatic potential.²⁸ In fact, studies investigating immunohistochemical markers in glottic and supraglottic squamous cell carcinoma have correlated both partial-EMT and EMT with worse overall and disease-free survival.²⁹ Moreover, EMT programming has been associated with higher risk of PNI, a disease parameter correlated with State 2 membership and with poorer prognosis among patients with head and neck cancer.³⁰ However, in this study, we demonstrate the presence of a specific transcriptional subtype of largely locally advanced glottic tumors (T3/T4) wherein the dominant transcriptional phenotype corresponded to EMT signaling. Though a comprehensive discussion of EMT is well beyond the scope of this analysis, therapies including modified synthetic miRNAs to interfere with EMT transcription factors; therapies promoting re-differentiation or trans-differentiation of tumor cells (i.e. mesenchymal-to-epithelial transition [MET]); and, inhibitors or ligand-neutralizing antibodies blocking upstream signaling pathways (e.g. TGF-β, NF-kB), have been explored as modalities for tumors demonstrating an EMT spectrum phenotype. Anti-EMT treatment paradigms, however, are complicated by empirical resistance to chemotherapeutic drugs, and the risk of MET promoting distant colonization by circulating tumor cells participating in early, undetected metastasis.³¹ Importantly, relative over-expression of each of the representative EMT pathways was not independently, significantly correlated with DSS, despite their overwhelming representation among State 2 samples, indicating that EMT, as with all discrete molecular perturbations, must be considered within the context of the broader transcriptional environment, and that disease trajectory of glottic tumors is likely contingent on synergistic activation of multiple parallel oncogenic programs. Moreover, the MSigDB-defined oncogenic pathways most highly correlated with S2 stipulated KRAS signaling (KRAS.LUNG_UP.V1_UP, IC=0.75; KRAS.600.LUNG.BREAST_UP.V1_UP, IC=0.69, KRAS.LUNG.BREAST_UP.V1, IC=0.68); importantly, no correlation was seen between S2 membership and activating KRAS mutations, suggesting orthogonal activation of the transcriptional program underlying the KRAS pathway. Downstream effectors of KRAS programming are known to effect tumor progression via promotion of autophagy, glycolysis and central carbon metabolism, and protection against oxidative stress.³² Collectively, these observations corroborate the value of the present methodology in investigating the molecular complexities of glottic tumorigenesis, as opposed to considering individual transcriptional or mutational perturbations in isolation.

State 0 and State 1 were also characterized by distinctive oncogenic programming. State 1 was correlated with a transcriptional signature indicative of an underlying inflammatory phenotype (i.e. non-canonical NF-kB signaling and UV radiation response signature gene sets), and drug response gene sets corresponding to both metformin and JAK inhibition.^33–35 Metformin has been investigated as an adjuvant therapy in the management of head and neck cancers, and JAK inhibitors are classically used in the management of autoimmune disorders acting on a pathway mediated by potentially tumorigenic, pro-inflammatory cytokines.³⁴ Studies exploring metformin’s potential efficacy in preventing tumor progression have stipulated inhibition of NF-kB signaling and mitigation of mitochondrial respiration (resulting in intracellular energy depletion, and diversion towards less efficient metabolic pathways), both cellular processes associated with S1 membership.^36,37 As such, although no significant S1-specific survival implications were identified in this analysis with a relatively small sample size, S1 is shown to demonstrate potentially actionable patterns of transcriptional activity. Though a large proportion of pathways represented in State 0 annotations simply characterized basal cellular processes, a series of genes classically associated with a mesenchymal phenotype were significantly and negatively correlated with S0, imputing a spectrum among TCGA GSCC tumors between epithelial (S0) and mesenchymal (S2) phenotypes.

This preliminary in silico analysis provides potentially valuable insights regarding the molecular heterogeneity of GSCC, and, more importantly, identifies a significant survival disadvantage associated with an oncogenic subtype that was discerned in clinically agnostic fashion. Though we have highlighted several genes and pathways highly correlated with each oncogenic state and transcriptional component constituting the oncogenic map, the precise combination of actionable oncogenic processes driving the survival differences between oncogenic states may not have been fully, molecularly interrogated, given that our biological annotations of the map are biased towards a priori cellular and oncogenic pathways characterized in MSigDB, and on genes that have been empirically associated with tumorigenesis. This study is also characterized by a relatively limited sample size; there is, however, a stark paucity of transcriptomic data corresponding to glottic tumors specifically that would otherwise be available to serve as a validation dataset corroborating the analysis herein. The lead author is currently constructing a laryngeal cancer database (with precise subsite information indicated), with tumor and normal-adjacent tissue to be submitted for molecular profiling and prospectively collected clinical data. We aim to use the contemporary analysis as a platform for future investigations into the molecular intricacies of glottic carcinogenesis.

We posit that the current analysis represents a framework for future in vitro analyses to better understand the synergistic oncogenic pathways driving the survival discrepancies associated with the oncogenic states described herein. Future prospective studies including less advanced glottic tumors (T1/T2) with comprehensive clinical data (including more granular details regarding treatment history) are necessary to improve prognostication and to further elucidate actionable, synergistic oncogenic programming responsible for the molecular and clinical heterogeneity seen in GSCC.

Conclusion

We present an explicitly, computationally derived molecular subtype of glottic squamous cell carcinoma demonstrating significantly inferior disease-specific survival, with preliminary biological annotations representing relative activation of canonical pro-tumorigenic programming. This state demonstrated a pattern of transcriptional activity imputing greater tumor invasiveness/aggressiveness, a transcriptional feature that conveyed clinically via higher proportions of both PNI and ECS. Future investigations are necessary to validate the map presented herein, and further elucidate pathways, protein expression, and epigenetic modifications most responsible for differences in GSCC outcomes.