Hypoxia-Induced Gene Expression in Chemoradioresistant Cervical Cancer Revealed by Dynamic Contrast-Enhanced MRI

Magnetic resonance imaging (MRI) is an indispensable tool in cancer management, and is widely used for staging, treatment planning, and response monitoring (1). Functional MRI, such as dynamic contrast-enhanced (DCE)-MRI, provides biologic information related to tumor physiology (2). There is a growing interest to use dynamic contrast enhanced MR imaging (DCE-MRI) to improve patient care by integrating anatomical and functional features (3). Locally advanced cervical cancer is one of the malignant diseases for which the inclusion of DCE-MRI could be helpful in clinical decision-making (1). Radiotherapy, often combined with cisplatin, is the primary treatment of choice and is challenging because of the risk of severe radiotoxicity in critical organs within the pelvis (4). Studies have suggested that DCE-MRI of cervical cancer depicts functional features associated with progressive disease (5–8). The ability of the technique to visualize tumor aggressiveness has, however, hardly been used, partly because the biologic meaning of the images is not completely understood (9).

DCE-MRI measures the temporal uptake pattern of a contrast agent in the tumor. When gadopentetate dimeglumine (Gd-DTPA) is used as contrast agent, the uptake mostly depends on the blood perfusion and the volume of the extracellular space (2). Biologic information can be extracted from the images by pharmacokinetic analysis of the uptake curves, where the Tofts and Brix models are most commonly applied (10). The Brix model has been recommended because the concentration of contrast agent in blood, that is, the arterial input function, is not needed in the calculations, making it particularly attractive for use in clinical routines (9). To fully exploit the potential of such models, a robust DCE-MRI biomarker should be identified and its molecular background clarified. Global gene expression profiles provide a detailed picture of the transcriptional program in tumors and thereby insight into the molecular phenotype. A few studies combining imaging and gene expression data have so far emerged, conveying results that encourage further work using this approach (11, 12).

In this study, we aimed to clarify the prognostic potential of the DCE-MRI Brix model parameters in patients with cervical cancer subjected to curative chemoradiotherapy and explore their molecular background by use of gene expression profiles. In a cohort of 78 patients, we carried out pharmacokinetic analyses of pretreatment DCE-MR images and generated parameter histograms of each tumor to account for the intratumor heterogeneity in physiologic conditions. For the most prognostic parameter, the percentiles showing the strongest association with survival were used to extract a DCE-MRI parameter that was combined with global gene expression data. We further searched for a gene signature that reflected this parameter and contained molecular information about the potential aggressive phenotype visualized in the images. Our study shows the feasibility of this approach to decipher the molecular background of functional images and to identify a gene signature that reflects an aggressive cancer phenotype.

A total of 187 patients with cervical carcinoma, prospectively recruited to our chemoradiotherapy protocol at the Norwegian Radium Hospital from 2001 to 2006, were included (Supplementary Table S1). Pretreatment DCE-MRI images suitable for analysis were available for 78 patients (DCE-MRI cohort). Three additional patients in our previous descriptive DCE-MRI study (7) were excluded because of unsatisfactorily fitting of the pharmacokinetic model. The remaining 109 patients served as an independent validation cohort based on their gene expression profiles. Possible pathologic lymph nodes in the pelvis at the time of diagnosis were evaluated by MRI, according to the response evaluation criteria in solid tumors version 1.1 (13). All patients were treated with external radiation of 50 Gy to the tumor, parametria, and adjacent pelvic wall, and with 45 Gy to the remaining part of the pelvic region. This was followed by brachytherapy of 21 Gy to point A. Adjuvant cisplatin (40 mg/m²) was given weekly during the course of external radiation. Follow up consisted of clinical examinations, and when symptoms of relapse were seen, MR imaging of pelvis and retroperitoneum as well as X-ray of thorax were carried out. Relapse (progressive disease) was classified as locoregional (regression within the irradiated field), distant, or both.

One to 4 tumor biopsies, approximately 5 × 5 × 5 mm in size, were taken before the start of therapy, immediately snap frozen, stored at −80°C, and used for gene expression analysis. A separate specimen was fixed in 4% buffered formalin, paraffin-embedded, and used for immunohistochemistry. The study was approved by the regional committee of medical research ethics in southern Norway, and written informed consent was achieved from all patients.

A 1.5 T Signa Horizon LX tomography (GE Medical Systems) with a pelvic-phased array coil was used for MRI. Standard T1- and T2-weigthed imaging was carried out in addition to DCE-MRI. To record the DCE-MRI series, an axial T1-weighted fast spoiled gradient recalled sequence was used (7). Gd-DTPA (0.1 mmoL/kg body weight; Magnevist; Schering) was administered as a fast bolus injection. The sequence of DCE-MRI included 14 image series during a time period of 5 minutes, of which 1 series was recorded before the bolus injection and 13 after. The temporal resolution was 15 seconds for the first 11 image series and 1 minute for the remaining 2 series.

The tumor outline was assessed in axial T2-weighted images by an experienced radiologist, and transferred to the DCE series by coordinate mapping (Fig. 1A; ref. 7). The relative signal increase (RSI) was used as measure of the uptake of contrast agent and was calculated for each tumor voxel and time point, using the relation:

where S(t) is the signal intensity at time t, and S(0) is the signal intensity in the precontrast images. Using Levenberg–Marquardt least squares minimization (14), the Brix model (10) was fitted to the uptake curve of each tumor voxel, using the relation:

where A_Brix is the amplitude, k_ep the transfer rate of tracer from tissue to plasma, and k_el the clearance rate of the tracer from plasma. The parameters were allowed to vary freely in the fitting, except for the constraints A_Brix, k_ep, k_el ≥ 0. Parameter maps were created to visualize the distribution of parameter values within the tumors (Fig. 1B and data not shown). To investigate the correlation between the parameters and clinical outcome, a percentile screening method presented previously (7) was applied. In brief, for each tumor, a parameter histogram was generated and percentile values from the 1st to the 100th percentile were calculated (Fig. 1B and data not shown). On the basis of their nth percentile value, the patients were separated into 2 equally sized groups, and log-rank tests were employed to assess whether the risk of relapse differed significantly between the 2 groups. The relative risk (RR) and its 95% confidence interval (CI) were plotted as a function of percentile, and for A_Brix, the prognostic parameter was calculated as the mean value over the percentile interval with the most significant RR; that is, with the strongest association to outcome (Fig. 1C). For k_ep and k_el, the associations with survival were nonsignificant or significant only within a small percentile interval (Supplementary Fig. S1), and they were therefore not included in further analyses.

The HeLa, SiHa, and CaSki cervical cancer cell lines from American Type Culture Collection were used to generate cervical cancer–specific gene sets of hypoxia responsive genes. The cells were identity tested before the experiments by short tandem repeat (STR) profiling using Powerplex 16 (Promega), which amplifies 15 STR loci and amelogenin for gender identification: Penta E, D18S51, D21S11, TH01, D3S1358, F GA, TPOX, D8S1179, vWA, Amelogenin, Penta D, CSF1PO, D16S539, D7S820, D13S317, and D5S818.

The cells were incubated in Dulbecco's modified Eagle medium with GlutaMAX supplemented with 10% fetal calf serum and 100 U/mL penicillin/streptomycin (Gibco) under a 5% CO₂ atmosphere at 37°C. Cells were plated out in plastic dishes (3 × 10⁵ cells for 6 cm dishes, 9 × 10⁵ for 10 cm dishes) 24 hours before exposure to hypoxic (0.2% O₂, 5% CO₂) or normoxic (95% air, 5% Co₂) conditions for 24 hours at 37°C. The hypoxia treatment was carried out in an Invivo₂200 chamber (Ruskinn Technology Ltd.) with accurate O₂ and CO₂ controls. The treatment conditions of 24 hours with 0.2% O₂ were selected to mimic conditions with prolonged hypoxia and ensure a response by HIF1α and possibly by the unfolded protein response (UPR; ref. 15).

The cell-cycle distribution was analyzed on an LSR II flow cytometer (Becton Dickinson) after fixation in 70% ethanol and staining with 1.5 μg/mL Hoechst 33258. Western blots were carried out by lysing cells with 10 mmol/L Tris HCl lysis buffer (pH 7.5) containing 2% SDS and 100 μmol/L Na₃VO₄. The proteins were separated by 8% Tris-HEPES-SDS polyacrylamide gels (Pierce Biotechnology), blotted on a PVDF membrane and stained with the monoclonal mouse HIF1α antibody clone 54 (1:800, no. 610958; BD Transduction Laboratories) and donkey antimouse secondary antibody (715-001-003; Jackson ImmunoResearch Laboratories, Inc.).

Gene expression profiling of 155 patients (46 in the DCE-MRI cohort, 109 in the validation cohort) as well as normoxia- and hypoxia-treated cell lines was carried out using the Illumina bead arrays human WG-6 v3 (Illumina Inc.) with 48803 transcripts (16). In brief, total RNA was isolated from the frozen tumor specimens using Trizol reagent (Invitrogen) and from the cell lines using RNeasy MiniKit (Quiagen). cRNA was synthesized, labeled, and hybridized to the arrays. Signal extraction and quantile normalization were carried out by the software provided by the manufacturer (Illumina Inc.). Log-transformed data were used in the analyses. All clinical specimens had more than 50% (median 70%) tumor cells in hematoxilin and eosin stained sections derived from the central part of the biopsy. RNA from different biopsies of the same tumor was pooled. The Illumina data have been deposited to the GEO repository (GSE36562).

To explore biologic processes associated with A_Brix, a list of genes was generated from the expression profiles of 46 DCE-MRI patients, based on the Spearman rank correlation between gene expression and A_Brix. A cut off P-value of 0.05 was used to achieve an appropriate number of genes for further analysis. The gene ontology (GO) categories of the correlating genes were compared with those of all genes on the array using the master-target procedure with the Fisher's exact test in the eGOn software, where a multiple testing algorithm was used to control the false discovery rate (FDR; ref. 17). Furthermore, the gene set analysis tool Significance Analysis of Microarrays for Gene Sets (SAM-GS), which is based on the moderated t-statistic in SAM (18) and with the same multiple testing procedure as above to control FDR, was used for a supervised analysis on 16 gene sets comprising biologic processes that were significant in the GO analysis. Seven of the gene sets have been published previously (19–24), 5 sets were based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome pathway databases and 4 sets were generated from the gene expression data of the hypoxia-treated cervical cancer cell lines (Supplementary Methods S1). Since multiple gene symbols may exist for the same genes, the symbols in the various gene sets were altered to match those in our Illumina data set and only genes that could be found in our data set were included in the final gene sets (Supplementary Methods S2). For the gene set analysis, the patients were classified into a low A_Brix group or a high A_Brix group, based on the median value of A_Brix in the 78 DCE-MRI patients. The Illumina data file was reduced to one probe per gene by selecting the probe with the greatest average difference between the low and high A_Brix groups.

Immunohistochemistry (IHC) was carried out on 32 selected tumors with the monoclonal mouse HIF1α antibody clone 54 (1:25, no. 610958; BD Transduction Laboratories), binding the amino acid 610 to 727 on the HIF1α protein. The tissue sections (4 μm) were stained using the Dako EnVisionTM Flex+ System (K8012; Dako). For antigen retrieval, the PT-Link (Dako) and EnVisionTM Flex target retrieval solution at a high pH were used, and the sections were incubated with the HIF1α antibody for 30 minutes. A cervical tumor known to express HIF1α was used as a positive control, whereas as a negative control the antibody was substituted with mouse myeloma protein of the same concentration and subclass as the HIF1α antibody. Nuclear staining was scored on the basis of the percentage of positive tumor cells: 0, 0%; 1, 1% to 10%; 2, 11% to 25%; 3, 26% to 50%; 4, 51% to 75%; and 5, >75%. The cytoplasmic staining was generally weak or absent, and was therefore not quantified. The scoring was carried out by an experienced scientist at the Department of Pathology (R.H.) who was blinded to the DCE-MRI data.

Mann–Whitney U test was used for comparison of the protein expression level of HIF1α in the patient group with low A_Brix versus high A_Brix. In the survival analyses, the endpoint was progression-free survival, where the time from diagnosis to cancer-related death or to the first event of relapse was used, or locoregional control, which was defined as control within the irradiated pelvic region including lymph nodes. Thirteen patients died of causes not related to cancer and were censored. Cox proportional hazard analysis was used to evaluate the prognostic value of various parameters with respect to progression-free survival and locoregional control. Kaplan-Meier curves were compared using log-rank test. P-values <0.05 were considered significant.

The prognostic potential of the DCE-MRI A_Brix parameter was explored by evaluating the relative risk of all percentiles in the A_Brix histograms, which enabled us to judge different parts of the histograms separately and identify the one with the strongest relationship to outcome (7). The lower A_Brix values harbored the strongest prognostic information, and a broad range from the 6th to 59th percentile was associated with progression-free survival at a significance level below 0.08 (Fig. 1C), indicating considerable robustness in the A_Brix parameter (7). The interval from the 8th to 36th percentile had a significance level below 0.05, whereas the 20th to 30th percentiles were the most significant ones (RR < 0.44, P < 0.018; Fig. 1C). The mean of A_Brix values over the 20th to 30th percentiles was therefore used as the representative A_Brix imaging parameter. This A_Brix parameter varied considerably among the patients, ranging from 0.59 to 3.21, with a median of 1.50. Patients with an A_Brix below the median showed a poor outcome compared with the others, and the probability of survival differed with more than 30% at 5 years follow-up between the 2 groups (P = 0.004; Fig. 1D).

To examine whether the differences in A_Brix among the tumors were reflected in the transcriptional program of specific biologic processes, we combined the gene expression profiles of 46 tumors with the DCE-MRI data. We first used an unsupervised GO analysis to obtain a general picture of the biology in low A_Brix tumors, where the genes with the highest correlation between expression and A_Brix were included. Four major significant processes were identified; metabolism, cell cycle, cellular component organization and biosynthesis, and response to DNA damage stimulus (Supplementary Table S2). On the basis of these results, a supervised gene set analysis was carried out on all genes on the array to more specifically investigate phenotypes that could be reflected in A_Brix. The correlation between A_Brix and the categories metabolism, cell cycle, and cellular component organization and biosynthesis could indicate both altered proliferation and presence of hypoxia. Published gene sets associated with these phenotypes were therefore included (20, 22–24). To cover the process response to DNA damage stimulus, published gene sets representing intrinsic radioresistance were used (19, 21) as well as the KEGG DNA repair pathways for nonhomologous end-joining, homologous recombination, and base excision repair. The cell-cycle checkpoints pathway from Reactome was also included to represent both this process and the cell-cycle category. Moreover, the KEGG pathway for protein processing in endoplasmic reticulum (ER) was used to cover the metabolism category more thoroughly because the protein modification process was one of its significant subcategories. Finally, we included a published gene set associated with wound healing that has shown prognostic impact in several other tumor types (23).

The most significantly differentially expressed gene sets when comparing tumors with high (above median) and low (below median) A_Brix were those representing hypoxia (P = 0.025 and 0.032; Supplementary Table S3). The cell-cycle checkpoints pathway was also significant (P = 0.038), and a closer look at its genes revealed that several of those that were correlated with A_Brix are known to be activated under hypoxia, such as ATR and CHEK2 (25). Thus, A_Brix seemed to be associated with the transcriptional program regulated under hypoxia.

The transcriptional hypoxia response has been shown to differ among tumor types (26). A cervical cancer–specific hypoxia gene set could possibly be more strongly associated with A_Brix than the gene sets used above and thus contain genes that would be important for the creation of a robust gene signature that reflected A_Brix. We therefore generated hypoxia gene sets based on the expression profiles of cervical cancer cells grown under hypoxia. HIF1α protein was upregulated in all 3 cell lines by the hypoxia treatment, whereas only minor changes were observed in the cell-cycle distributions (Supplementary Fig. S2A and S2B). Four different gene sets were generated from the hypoxia induced expression changes; genes upregulated in all 3 cell lines, genes upregulated in one of the cell lines and confirmed as hypoxia regulated by the literature, and correspondingly for the downregulated genes (Supplementary Methods S1 and S2). In the second round of gene set analysis, the cervical cancer–specific hypoxia gene sets with upregulated genes were found to be the most significant ones (Table 1). These results supported the suggestion of a correlation between A_Brix and hypoxia, and showed that the cervical cancer–specific gene sets could be useful in the construction of an A_Brix gene signature.

To confirm the above finding of a correlation between A_Brix and hypoxia, protein expression of HIF1α was assessed by immunohistochemistry in the 32 remaining DCE-MRI patients who had not been evaluated in the gene expression study. HIF1α was selected as hypoxia marker because a hypoxia-induced HIF1α response was observed in the cell lines (Supplementary Fig. S2B). The patients with low A_Brix had a significantly higher HIF1α expression compared with those with high A_Brix (P = 0.004; Fig. 2A–C). In accordance with this result, several of the HIF1α targets were upregulated in tumors with low A_Brix in the DCE-MRI patients used for gene expression analysis. We thus concluded that low A_Brix was associated with tumor hypoxia and upregulation of hypoxia response genes.

To construct a DCE-MRI signature with the most important genes reflected by A_Brix, we selected the 4 significant hypoxia gene sets (Table 1), and extracted the 31 genes with a negative correlation between the gene expression and A_Brix; that is, the genes that were upregulated in tumors with low A_Brix. Most of these 31 genes are known to participate in biologic processes affected by hypoxia, such as energy metabolism, cell cycle, and proliferation, including HIF1α targets such as ALDOA, STC2, and HMOX1 (27–29), and genes involved in the UPR (DDIT3, STC2, ERO1L, and AK2; refs. 30–32; Table 2), suggesting that several parts of the hypoxia induced transcriptional program were reflected by A_Brix.

To ensure that the signature sufficiently represented the A_Brix parameter, we carried out unsupervised clustering of the 46 DCE-MRI patients based on the expressions of the 31 genes. Clustering showed 2 groups with different outcome, for which the one with high expression of hypoxia responsive genes had low A_Brix (P < 0.001) and poor outcome (P = 0.011) as compared with the other (Fig. 3A). We further calculated a hypoxia score for each tumor by averaging the median centered expression levels for the 31 genes, as described (26). In line with the above results, patients with a high hypoxia score had low A_Brix (P < 0.001) and poor outcome (P = 0.011) compared with the others (Fig. 3B). The DCE-MRI hypoxia gene signature therefore seemed to satisfactorily represent the A_Brix parameter, including its relationship to clinical outcome.

The prognostic impact of the gene signature was further validated in an independent cohort of 109 patients with cervical cancer. In this validation set, the patients who clustered together because of high expression of the genes in the signature had a significantly worse outcome than the remaining patients (P = 0.002; Fig. 4A). When assessing the hypoxia score in this cohort, the patients with a high score had a poor outcome compared with those with a low score (P = 0.006; Fig. 4B). The prognostic value of the signature was therefore confirmed. To assess the importance of the signature in comparison with existing clinical markers, the patients from the DCE-MRI cohort and the validation cohort were merged, and a multivariate Cox analysis was carried out on the resulting group of 155 patients. The hypoxia score emerged as a prognostic factor independent of lymph node status, FIGO stage, and tumor volume for both progression-free survival and locoregional control (Table 3).

By integrating DCE-MRI and global gene expression data of cervical cancers, we found that the images reflect a transcriptional program regulated under hypoxia. This program included a gene signature with prognostic impact in an independent validation cohort, thus pointing to hypoxia-regulated pathways that may promote cervical cancer aggressiveness. Tumor hypoxia is a known prognostic factor in many types of cancer, including cervical cancer (33), and would therefore be valuable to implement in clinical decision-making. Our work is the first to provide molecular insight into the DCE-MR images that could facilitate this strategy, encouraging the use of DCE-MRI as a tool to handle hypoxia-induced chemoradioresistance in cervical cancer.

Several evidences for a relationship between low A_Brix and tumor hypoxia were provided by our work. The hypoxia gene sets were significantly associated with A_Brix, and revealed upregulation of hypoxia response genes including HIF1α targets in tumors with low A_Brix. In line with this, tumors with low A_Brix had higher protein expression of HIF1α compared with those with high A_Brix. The results from the GO analysis were also in accordance with these conclusions, as the biologic processes metabolism, cell cycle, and DNA damage signaling, which were associated with A_Brix, are important targets for the hypoxia response (25, 28, 34). In particular, genes from the metabolism category encoding the glycolytic enzymes ALDOA, GAPDH, and PFK4B4 were found to be upregulated in tumors with low A_Brix, consistent with the increased glycolytic activity often seen in hypoxic tumors (28). Loncaster and colleagues (5) showed a relationship between A_Brix and oxygen tension in cervical tumors, as measured by Eppendorf pO₂ histography, supporting our conclusion.

Extraction of genes from the hypoxia gene sets that were upregulated in tumors with low A_Brix resulted in a DCE-MRI hypoxia gene signature that was associated with the clinical outcome, independent of clinical markers for progression-free survival. This suggests that the gene signature associated with A_Brix provides information of disease progression that is not covered by the conventional clinical parameters. In addition, the gene signature appeared as the most significant factor for locoregional control, indicating that hypoxia as measured by A_Brix could be a well-needed biomarker of local recurrence in cervical cancer. Furthermore, while previous studies have demonstrated the importance of hypoxia in resistance to radiotherapy (33), the strong prognostic impact of this hypoxia gene signature indicates that hypoxia also influences the effect of chemoradiotherapy in cervical cancer patients.

The use of global gene expression data enabled detailed insight into the transcriptional program reflected by A_Brix. The response to hypoxia occurs through different pathways, and may involve activation of HIFs, the UPR or ER stress response, and signaling through the mechanistic target of rapamycin (MTOR) kinase (32). Eight of the 31 genes in the DCE-MRI hypoxia gene signature are known HIF1α targets, namely ALDOA, ERO1L, GAPDH, PFKFB4, P4HA2, C4orf3, HMOX1, and STC2 (27, 29, 32, 35, 36), although SCARB1 is a target of HIF2 (the endothelial PAS domain protein 1, EPAS1; ref. 36). The protein encoded by one of the other genes, RHOC, is indirectly associated with HIF1α through its interaction with the von Hippel-Lindau tumor suppressor (VHL; ref. 37), which degrades HIF1α. In addition, AK2, ERO1L, DDIT3 (also known as CHOP), and STC2 have been found to be involved in the UPR (30–32, 38). It thus seems that processes associated with both HIF activation and UPR may be important in the hypoxic phenotype depicted by low A_Brix. The involvement of HIF activation was further supported by the high HIF1α protein expression in these tumors.

Current knowledge of how these genes influence cervical cancer progression is sparse. However, several of them or their encoded proteins, including STC2, PFKFB4, RHOC, S100A2, HMOX1, and ISG15, have been associated with poor prognosis in other cancer types (39–44). Moreover, upregulation of ALDOA, GAPDH, and PFKFB4 suggest a high glycolytic activity, as mentioned above, which has been associated with aggressiveness (28). Furthermore, SCARB1, PVR, and RPL36A have been shown to promote rapid proliferation (45–47), and AK4 may play a critical role in both cellular survival and proliferation during stress conditions such as hypoxia (48). Both DDIT3 and STC2 are targets of the activating transcription factor ATF4, which is induced by the eukaryotic translation initiation factor kinase EIF2AK3 (PERK) in response to hypoxia (31, 32, 49). STC2 has been shown to inhibit apoptosis and to induce proliferation and invasiveness in response to HIF1α or ER stress during hypoxia (27, 31, 50). DDIT3 on the other hand, is a proapoptotic protein, but was recently shown to protect tumor cells under hypoxia through the regulation of autophagy, proposing a role in the balance between autophagy and apoptosis (49). It may thus be speculated that the combined action of these 2 genes in tumors with low A_Brix promotes the adaption to hypoxia by inhibiting apoptosis and inducing autophagy. Taken together, these studies support a role of the signature genes in the development of an aggressive hypoxic phenotype.

Using DCE-MRI to assess hypoxia associated aggressiveness in cervical cancer is appealing because MRI is already in routine use for patient diagnosis at many hospitals (1). Furthermore, the use of noninvasive imaging to assess the molecular hypoxia phenotype could potentially reduce the need of invasive biopsy procedures and be carried out during the course of treatment for response evaluation. A_Brix seems to be a valuable DCE-MRI parameter for this purpose because of its strong correlation to outcome in our study. The robustness of the parameter, as suggested by the percentile analysis of A_Brix histograms, is supported by the cervical cancer study of Loncaster et al. (5), where A_Brix also had prognostic impact although the patient characteristics and MRI protocol differed from ours. In the present work, we have showed how noninvasive imaging by means of DCE-MRI and A_Brix assessment may visualize hypoxia and its molecular basis in chemoradioresistant tumors. Our results may have important clinical implications in that they suggest the use of DCE-MRI to identify patients with treatment-resistant tumors that may benefit from additional or alternative therapy targeting hypoxia.

No potential conflicts of interest were disclosed.

Conception and design: C. Halle, E. Malinen, H. Lyng

Development of methodology: E. Andersen, G. Hasvold, E. Malinen, H. Lyng

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Halle, M. Lando, E.-K. Aarnes, G. Hasvold, R.G. Syljuåsen, K. Sundfør, G.B. Kristensen, E. Malinen, H. Lyng

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Halle, E. Andersen, M. Lando, G. Hasvold, M. Holden, R.G. Syljuåsen, R. Holm, E. Malinen, H. Lyng

Writing, review, and/or revision of the manuscript: C. Halle, M. Lando, E.-K. Aarnes, G. Hasvold, M. Holden, R.G. Syljuåsen, K. Sundfør, G.B. Kristensen, R. Holm, E. Malinen, H. Lyng

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E. Andersen, G.B. Kristensen, H. Lyng

Study supervision: H. Lyng

This work was supported by the Norwegian Cancer Society and the South-Eastern Norway Regional Health Authority.

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.