A 26-Gene Hypoxia Signature Predicts Benefit from Hypoxia-Modifying Therapy in Laryngeal Cancer but Not Bladder Cancer

Tumor hypoxia confers resistance to radiotherapy and predicts for a poor treatment outcome (1, 2). A meta-analysis of clinical trials showed that there is level 1a evidence in favor of adding hypoxic modification to radiotherapy for head and neck squamous cell carcinoma (HNSCC; ref. 3). Despite this, hypoxia-modifying therapies have not been adopted as an international standard of care. A recent randomized phase III trial compared accelerated radiotherapy alone or with the hypoxia-modifying agents carbogen and nicotinamide (ARCON) in laryngeal carcinoma. The ARCON trial showed a significant improvement in 5-year regional control with CON (4). Evidence suggests that hypoxic tumors benefit most from hypoxia-modifying therapy (4–8). Similar results have been shown in bladder cancer. The bladder carbogen nicotinamide (BCON) phase III trial showed that addition of CON to radiotherapy significantly improved overall survival in bladder carcinoma (9, 10).

Several methods are used to measure tumor hypoxia. Oxygen electrode (2, 11), imaging approaches (7, 12) and hypoxia-specific markers such as pimonidazole (4, 6, 13) require prospective assessment, limiting the ability to validate findings. Measurement of endogenous markers of hypoxia such as carbonic anhydrase-9 (CA9; refs. 13–16) and hypoxia-inducible factor-1α (HIF-1α; refs. 14, 17, 18) has shown some success. However, approaches using multiplex markers such as gene signatures (8, 19–21) potentially better reflect the complex cellular response to hypoxia and account for the intratumor heterogeneity of hypoxia. Previous studies showed the potential of a 15-gene hypoxia signature to predict benefit from hypoxia-modifying therapy in HNSCC (8). A 121-gene hypoxia meta-signature was derived from 3 independent head and neck cancer datasets. The full and a reduced 26-gene signature (Table 1) had prognostic significance in head and neck, breast, and lung cancer (20). The 26-gene signature was taken forward for validation using a TaqMan low density array (TLDA) approach. Technical validation of the signature showed the approach was a reliable and reproducible measure of tumor hypoxia associated with low intratumor heterogeneity; the multiplex nature of the biomarker allows buffering of artifact effects which often contribute to variability of single measures (21).

The primary aim of this study was to investigate the ability of the 26-gene hypoxia signature to predict benefit from hypoxia modification using laryngeal and bladder cancer samples from the ARCON and BCON phase III clinical trials, respectively. As a secondary objective, the prognostic ability of the gene signature was also investigated in patients receiving nonexperimental treatment. Owing to its prognostic ability in multiple cancer types (20), similar gene expression in both cancer types was assumed and subsequently examined.

A retrospective study was carried out using formalin-fixed, paraffin-embedded (FFPE) pretreatment samples obtained from 2 prospective phase III clinical trials of radiotherapy alone or with CON. The target sample size was 150 patients. Analysis of 150 ARCON patients was required to detect a difference in HRs for hypoxic versus oxygenated tumors with P = 0.002 and 90% power. One hundred and fifty BCON patients were required to detect a difference in HR with P = 0.01 and 80% power.

First, tumor biopsies were obtained from 157 patients with T2–T4 laryngeal cancer who participated in the ARCON trial and randomized between April 2001 and February 2008. Samples from 6 centers in the Netherlands were obtained for 229 of the 345 patients enrolled in ARCON (Supplementary Fig. S1). Second, tumor resections were obtained from 185 patients with high-grade T1–T4a transitional cell bladder cancer who participated in the BCON trial and randomized between November 2000 and April 2006. Samples from 11 United Kingdom hospitals were obtained for 251 of the 333 patients enrolled in BCON (Supplementary Fig. S2). The study was approved by the Greater Manchester Research Ethics Committee (LREC 09/H1011/40 and LREC 09/H1013/24, respectively). REMARK guidelines for reporting tumor marker studies were followed.

ARCON patients allocated to the accelerated radiotherapy arm received 68 Gy in 2 Gy fractions to the primary tumor over 36 to 38 days. Patients allocated to the ARCON arm received 64 Gy in 2 Gy fractions, carbogen (98% O₂ + 2% CO₂, 4 minutes before and during daily fractions), and oral nicotinamide (60 mg/kg, 1–1.5 hours before each fraction). BCON patients received 55 Gy in 20 fractions in 4 weeks or 64 Gy in 32 fractions in 6.5 weeks daily, 5 times per week. Carbogen was given 5 minutes before and during radiotherapy. Nicotinamide (40–60 mg/kg) was given 1.5 to 2 hours before each fraction.

One 4 μm hematoxylin and eosin-stained section from each FFPE block was analyzed. Clinical staging followed tumor-node-metastasis, American Joint Committee on Cancer/Unio Internationale Contra Cancrum (Italian) classifications, and grading was done according to the World Health Organization guidelines. Tumor necrosis (10, 22) and concurrent carcinoma in situ (pTis; ref. 23) were recorded (BCON) as recognized prognostic features. Samples with less than 10% tumor material were excluded from analysis.

RNA was extracted from FFPE samples (three 20 μm sections) using the RecoverAll Total Nucleic Acid Isolation Kit (Life Technologies), which included DNase I treatment. Reverse transcription of total RNA was conducted using the Omniscript Reverse Transcription Kit (Qiagen), oligo-dT primers (Qiagen), and random nonomers (Sigma-Aldrich). RNA yield from laryngeal biopsies was low due to small size, so preamplification was used to boost message. Preamplification of cDNA involved pooled TaqMan assays, Preamplification Master Mix (Life Technologies), and a PCR thermal cycler (Veriti 9902, Life Technologies).

Customized 384-well microfluidic TLDA cards were produced by Life Technologies. Reactions were carried out using TaqMan Gene Expression Master Mix (Life Technologies). Quantitative real-time PCR (qRT-PCR) was conducted using the ABI Prism 7900HT qPCR system as per manufacturers' protocol. Manual C_q values were determined retrospectively using ABI Prism Sequence Detection System (SDS) software (Life Technologies). Relative quantification of gene expression was calculated using the 2^−ΔCq method (24). Hypoxia scores (HS) were calculated as the normalized median expression of the 26 hypoxia genes. The geometric mean of 3 reference genes (GNB2L1, B2M, and RPL11) was used for normalization. The 3 best performing genes (best correlations and lowest median C_q values; Supplementary Figs. S3 and S4) were selected from 5 previously validated reference genes (21). A customized program was designed and used to automatically calculate TLDA-HS from SDS files (R package version 2.1.5; ref. 25). The group median was used to stratify patients as TLDA-HS low (≤median) or high (>median). This cutoff had the most discriminative prognostic power in multiple cancer types in previous analyses (20). Minus reverse transcriptase and no template controls were analyzed using intron-spanning TaqMan assays (CA9, GNB2L1, and RPL24; Life Technologies). All negative controls had C_q values more than 40 cycles. Scientific analysts were blinded to clinical outcome data.

Previous validation of the 26-gene hypoxia signature using fresh frozen HNSCC biopsies revealed one gene (CDC4A) to have nonlinear PCR reaction efficiency, which was removed from the signature (21). This gene was subsequently replaced by KRT17, the next in rank (20). The TLDA-HS (25-gene vs. 26-gene) for 33 biopsies (from 9 tumors) was significantly paired (Wilcoxon, P < 0.0001) and there was no significant difference in intratumor variation in TLDA-HS. Median coefficient in variance was 24.8% (range 7.1%–52.5%; 25 genes) versus 23.1% (range 10.1%–49.8%; 26 genes; paired Wilcoxon P = 0.07; Supplementary Fig. S5).

Primary endpoints for analyses were those which reached significance in the clinical trials: regional control (RC; ARCON) and overall survival (OS; BCON). For ARCON patients, RC was defined as freedom from first regional recurrence. Secondary endpoints were local control (LC), disease-free survival (DFS), and OS. LC was defined as freedom from first recurrence at the site of primary tumor, DFS was defined as the time to local, regional, or distant recurrence, and OS was defined as time-to-death from any cause. For BCON patients, OS was defined as death from any cause. The secondary endpoint of relapse-free survival (RFS) was also reported. RFS was taken as time to tumor recurrence in bladder (muscle-invasive lesions only) or locoregional failure. Time was set to zero for those with persistent muscle-invasive tumor after radiotherapy or if no cystoscopic assessment was ever conducted. All endpoints were taken from randomization to event; patients were censored at the last time seen or at 5 years whichever was earlier. Median follow-up times for the cohorts studied were 59 (radiotherapy only) and 60 (CON) months and 61 (accelerated radiotherapy) and 61 (ARCON) months for BCON and ARCON patients, respectively.

Statistical analyses were conducted using SPSS version 19 for Mac. Survival estimates were conducted using the Kaplan–Meier method and differences compared using the log-rank test. HRs and 95% confidence intervals (CI) were obtained using Cox proportional hazard model. For ARCON patients, the number of events for these outcome parameters was low; hence, multivariate analyses were not possible. For BCON patients, in multivariate analyses, prognostic features were entered as covariates. The χ² test was used to compare proportions across the levels of categorical factors and Yates correction was used for 2 × 2 tables; the Mann–Whitney U test was used to compare median values for continuous variables between 2 groups. Spearman rank correlation coefficients were used to assess statistical relationships. P values were two-sided and significance set to P ≤ 0.05.

Tumor sections from 229 patients were received from Nijmegen (the Netherlands). The median RNA yield was 52.2 ng/μL (range 1.9–339.0 ng/μL). As a power calculation indicated 150 samples needed to be analyzed, 162 patient samples with the highest RNA quantity were selected. The samples were preamplified and TLDA-HS generated for 157 patients (5 patient samples failed to generate a TLDA-HS; 96.9% success rate). Hypoxia gene expression for 8 matched nonpreamplified and preamplified samples correlated well (Spearman rank ρ = 0.88, P < 0.0001; Supplementary Fig. S6).

In the subset of 157 patients, 77 (49.0%) received accelerated radiotherapy and 80 (51.0%) ARCON (Table 2). There were similar clinical characteristics for patients in the subset and the main trial (Supplementary Table S1). In the accelerated radiotherapy arm, radiotherapy was given as planned to most patients (76; 98.7%). In the ARCON arm, 59 patients (73.8%) received all doses of radiotherapy and CON. All analyses were conducted on an “intention-to-treat” basis. There were no statistically significant differences in clinicopathologic features between treatment arms. Supplementary Figure S7 shows the distribution of TLDA-HS according to randomization arm. The median TLDA-HS for the 157 patients was 0.034 (range 1 × 10⁻⁵–0.19) with no statistically significant differences in clinicopathologic features between TLDA-HS high and low groups (Supplementary Table S2). TLDA-HS did not correlate with percent tumor in the samples analyzed (ρ = −0.105, P = 0.217; Supplementary Fig. S7).

Patients with TLDA-HS high (“more hypoxic” tumors) receiving accelerated radiotherapy had a significantly lower 5-year RC of 81% (7/7 events) compared with 100% (0 events) for those who received ARCON (log-rank P = 0.009). Patients with TLDA-HS low did not benefit from the addition of hypoxia-modifying therapy to radiotherapy. Respective 5-year RC rates for accelerated radiotherapy and ARCON were 91% (3/7 events) and 90% (4/7 events; log-rank P = 0.90; Fig. 1A and B). Patients with less hypoxic tumors (RC = 90%) did less well with ARCON than those with more hypoxic tumors (RC = 100%; P = 0.056). A test for heterogeneity in treatment effect by TLDA-HS strata was not possible due to the low event rate; however, survival analysis revealed a distinct difference in treatment efficacy according to each stratum so an interaction may be assumed. TLDA-HS did not predict benefit from hypoxia modification for 5-year local control, DFS, or overall survival.

For prognostication in accelerated radiotherapy patients, there was no significant difference in 5-year RC between TLDA-HS high (80.5%, 7/10 events) and TLDA-HS low (91.0%, 3/10 events; n = 77; log-rank P = 0.29; Fig. 2A). Of the other clinicopathologic features, only N stage was prognostic; increasing N stage reduced 5-year RC (log-rank P < 0.001). TLDA-HS high was not an adverse prognostic factor for 5-year LC or DFS. There was a trend towards poorer OS in patients with TLDA-HS high (40.7%, 23/34 events) versus TLDA-HS low (67.5%, 11/34 events; log-rank P = 0.06; Fig. 2B). There were no significant differences in clinicopathologic features between TLDA-HS high and low groups.

RNA was extracted from tumor sections from 251 patient blocks available. The median RNA yield was 90.3 ng/μL (range 1.1–502.5 ng/μL). Samples from 41 patients were excluded for: low RNA yield (<33 ng/μL; 39 samples) or reclassification as T4b (2 samples). In total, 185 TLDA-HS were generated (25 samples failed; 88.1% success rate). Higher RNA yields meant preamplification before TLDA was not necessary.

In the subset of 185 patients, 97 (52.4%) patients received radiotherapy and 88 (47.6%) patients radiotherapy+CON (Table 3). There were similar clinical characteristics for patients in the subset and the main trial (Supplementary Table S1). In the radiotherapy arm, radiotherapy was given as planned to all patients (97; 100%). In the radiotherapy+CON arm, 60 patients (68.2%) received all doses of radiotherapy and CON. There were no differences in clinicopathologic features between treatment arms, except a higher proportion of concurrent pTis cases in the radiotherapy+CON arm (χ²P = 0.002). Supplementary Figure S7 shows the distribution of TLDA-HS according to randomization arm. The median TLDA-HS for the 185 patients was 0.021 (range 3 × 10⁻⁴–0.15). There were no statistically significant differences in clinicopathologic features in TLDA-HS high and low groups (Supplementary Table S3). TLDA-HS did not correlate with percentage of tumor in the samples analyzed (ρ = 0.003, P = 0.97; Supplementary Fig. S7).

Prediction of treatment benefit from CON showed no significant improvement in 5-year OS in patients with TLDA-HS high (36.3%, 31/53 events radiotherapy vs. 43.4%, 22/53 events radiotherapy+CON, log-rank P = 0.40). Five-year OS was 44.3% (24/46 events; radiotherapy) and 53.4% (22/46 events; radiotherapy+CON; log-rank P = 0.51) for TLDA-HS low (Fig. 1C and D). TLDA-HS did not predict benefit from hypoxia modification for 5-year RFS.

For prognostication in radiotherapy patients, there was no significant difference in 5-year OS between TLDA-HS high (35.4%, 29/55 events) and TLDA-HS low (44.0%, 26/55 events; P = 0.33; Fig. 2C). Adverse prognostic features of OS were necrosis (log-rank P = 0.008) and concurrent pTis (log-rank P = 0.03), which were also independent predictors of poor OS (necrosis P = 0.002 and pTis P = 0.004). Patients with TLDA-HS high showed a trend towards a poorer RFS (39.0%, 24/41 events, TLDA-HS high vs. 52.1%, 17/41 events, TLDA-HS low, log-rank P = 0.06; Fig. 2D). Necrosis was the only adverse prognostic feature using this endpoint (log-rank P = 0.002, multivariate P = 0.001). There were no significant differences in clinicopathologic features between TLDA-HS high and low groups.

Exploration of a new cutoff for TLDA-HS using all BCON patients (n = 185) showed the lowest quartile (5-year OS, 59.5%) had a better prognosis compared with the upper 3 quartiles (39.0%; log-rank P = 0.02; Supplementary Fig. S8). TLDA-HS was an independent predictor of 5-year OS (P = 0.007 vs. pTis P = 0.009 and T stage P = 0.01) and RFS (P = 0.01 vs. pTis P = 0.009 and T stage P = 0.02). Independent verification of this new cutoff using another cohort is necessary to draw conclusions.

Comparison of laryngeal and bladder cancer data showed similar expression in the 2 tissues types (Spearman's rank ρ = 0.73, P < 0.0001), but greater differential expression in laryngeal samples (Fig. 3). The median fold increase of the 26 genes (between highest and lowest 5 samples) was 57.2 (range 5.7–8.3 × 10³; laryngeal) and 43.3 (range 6.2–1.3 × 10³; bladder). For TLDA-HS, fold increase was 6.5-fold higher in laryngeal than in bladder cancer (429.6 in laryngeal vs. 66.5 in bladder; Mann–Whitney U, P < 0.0001).

Clinical trials of hypoxia-modifying treatments have not allocated treatment according to hypoxic status despite large intertumor variability in hypoxia (13, 26, 27). In this study, classification of patients as “more” or “less” hypoxic based upon expression of a 26-gene hypoxia signature showed that hypoxic laryngeal tumors benefited from hypoxia-modifying CON treatment. Patients had a 19% improvement in 5-year RC rate versus those who received radiotherapy alone. Although nodal bulk at diagnosis was an important predictor of poor prognosis and a multivariate analysis could not be conducted, there was no difference in N stage distribution for patients with TLDA-HS high versus low scores (P = 0.90) suggesting that the apparent benefit of hypoxia modification was not due to patients in the hypoxia group having smaller bulk nodal disease at diagnosis.

In a recent study of 323 patients with HNSCC, classification of patients as more hypoxic based upon expression of a 15-gene hypoxia signature independently predicted benefit from the hypoxic radiosensitizer nimorazole (8). Four genes (P4HA1, SLC2A1, KCTD11, and ALDOA) overlap with the 26-gene HNSCC hypoxia signature.

Because of the prominent role the larynx plays in swallowing, respiration, communication, and protection of the lower airway, quality-of-life considerations are paramount. The larynx preservation rate of 87% for patients receiving ARCON is comparable with patients receiving radiotherapy and concurrent cisplatin (28). In addition, accelerated radiotherapy and ARCON produce equal levels of toxicity (4). ARCON patients with more hypoxic tumors as measured by TLDA-HS even had a 5-year regional control of 100%. A previous translational study using the hypoxia marker pimonidazole in 79 ARCON patients also showed this trend (4) indicating the need for a companion biomarker for treatment selection. Lack of translation of treatment benefit from regional control to OS or DFS in ARCON patients reflects effective salvage therapy by laryngectomy and neck dissection and patient comorbidities with consequent increased mortality. The lack of prediction for LC is consistent with the findings from the ARCON trial, which suggested that the lack of benefit from CON for LC might be due to the extra 4 Gy given in the radiotherapy alone arm (4).

Despite similar hypoxia gene expression profiles in laryngeal and bladder cancer, the 26-gene signature did not predict benefit from CON in patients with bladder cancer. Differential gene expression is lower in bladder cancer and this reduced intertumor variation may explain the loss in predictive ability. It is known that a significant proportion of the transcriptional response to hypoxia is cell-type dependent (29, 30). The rest comprises a conserved hypoxia response independent of hypoxic output (20, 29–31). Chi and colleagues (2006) examined core hypoxia signatures in primary cells (30), Lendahl and colleagues (2009) identified a 30-gene core hypoxia signature in cancer cell lines (31) and Buffa and colleagues (2010) derived a 51-gene common hypoxia signature by creating coexpression networks in 3 head and neck and 5 breast cancer datasets (20). Comparison of these core hypoxia targets with our 26-gene HNSCC signature shows an overlap of only 3 genes (HIG2, P4HA1, and VEGFA) with the Lendahl signature and 15 genes (CA9, MPRS17, SLC16A1, C20orf20, ENO1, ALDOA, TPI1, LDHA, SLC2A1, PGK1, P4HA1, VEGFA, PGAM1, ANLN, and CDKN3) overlap with the Buffa signature. The highly conserved core hypoxia genes are primarily metabolic genes (e.g., PGK1, LDHA, ENO1, and ALDOA). Future work using a conserved core signature or an independently derived bladder-specific signature is under investigation. It is likely that the different cell-specific response to hypoxia is at least in part due to variations in transactivation by HIF, the master transcriptional regulator of the hypoxia response, and epigenetic factors. Supplementary Figure S9 shows how the TLDA can minimize intratumor variation in hypoxia of bladder tumors as seen in HNSCC (27). Hence, modification of the signature for use in this patient population has great potential. Another possible explanation for the lack of predictive value of the signature in bladder cancer is that the patterns of failure are different from patients with laryngeal cancer (mainly locoregional vs. <10% locoregional for bladder). Although hypoxia modification improves locoregional control and survival, the head and neck cancer meta-analysis showed the benefit for OS was less than that for locoregional control (3).

This study strongly supports the growing body of evidence advocating pretreatment measurement of tumor hypoxia for selection of patients most likely to benefit from hypoxia modification in HNSCC. Measurement of the TLDA-HS is a simple procedure; most hospital laboratories are equipped and can conduct routine qPCR analyses. There is a high success rate in generating data—96.9% with preamplification. Multigene tests are already available in outpatient settings for breast (32–36) and colon (37) cancer. It was estimated that 12,360 people were diagnosed with and 3,650 people died of laryngeal cancer and 73,510 were diagnosed with and 14,880 people died of bladder cancer in the United States in 2012 (38). Use of this multi-gene test could potentially assist in providing treatment options for patients with hypoxic tumors.

C.M.L. West, A.L. Harris, C.J. Miller, and F.M. Buffa have ownership interest (including a patent - WO 011/076895). No potential conflicts of interest were disclosed by the other authors.

C.M.L. West, A.L. Harris, C.J. Miller, and F.M. Buffa are registered as inventors of a patent covering the use of the hypoxia gene signature. The patent reference number is WO 011/076895 A1 and is fully accessible at www.patentlens.net including the list of genes comprising the signature. The patent is registered in the name of Cancer Research Technology. Study sponsors had no involvement in study design, collection, analysis and interpretation of data, or writing or submission of the manuscript.

Conception and design: A. Eustace, N. Mani, G.N.J. Betts, H. Denley, P.J. Hoskin, A.L. Harris, J.H.A.M. Kaanders, C.M.L. West

Development of methodology: A. Eustace, N. Mani, G.N.J. Betts, H. Denley, C.J. Miller, F.M. Buffa, C.M.L. West

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Eustace, N. Mani, P.N. Span, J.J. Irlam, G.N.J. Betts, H. Denley, J.J. Homer, P.J. Hoskin, A.L. Harris, J.H.A.M. Kaanders

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Eustace, N. Mani, P.N. Span, J.J. Irlam, J. Taylor, H. Denley, A.M. Rojas, A.L. Harris

Writing, review, and/or revision of the manuscript: A. Eustace, N. Mani, P.N. Span, J. Taylor, G.N.J. Betts, H. Denley, C.J. Miller, J.J. Homer, A.M. Rojas, P.J. Hoskin, F.M. Buffa, A.L. Harris, J.H.A.M. Kaanders, C.M.L. West

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Eustace, N. Mani, P.N. Span, J.J. Irlam, J. Taylor, H. Denley, A.L. Harris, J.H.A.M. Kaanders

Study supervision: A. Eustace, F.M. Buffa, J.H.A.M. Kaanders, C.M.L. West

The authors thank ARCON and BCON trial investigators, the University of Manchester Clinical Immunology and Molecular Medicine Group for use of Good Clinical Practice facilities, and Ric Swindell for statistical support.

This work was supported by the Medical Research Council of the UK (grant no. G0801525; to C.M.L. West). C.M.L. West was also funded by Cancer Research UK (grant C1094/A11365) and the Experimental Cancer Medicine Centre. J. Taylor and C.J. Miller were funded by Cancer Research UK (grant C480/A12328). F.M. Buffa and A.L. Harris were funded by Cancer Research UK, EU Metoxia, and the NIHR Biomedical Research Centre, Oxford.

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.