Changes in Gene Expression Predicting Local Control in Cervical Cancer: Results from Radiation Therapy Oncology Group 0128

Imaging, Diagnosis, Prognosis

ChangesinGeneExpressionPredictingLocalControlinCervical Cancer: Results from RadiationTherapy Oncology Group 0128 Joanne B. Weidhaas,1Shu-Xia Li,2 Kathryn Winter,3 Janice Ryu,5 Anuja Jhingran,6 Bridgette Miller,7 Adam P. Dicker,4 and David Gaffney8

Abstract Purpose:Toevaluate the potential of gene expression signatures to predict response to treatment in locally advanced cervical cancer treated with definitive chemotherapy and radiation. Experimental Design: Tissue biopsies were collected from patients participating in Radiation Therapy Oncology Group (RTOG) 0128, a phase II trial evaluating the benefit of celecoxib in addition to cisplatin chemotherapy and radiation for locally advanced cervical cancer. Gene expression profiling was done and signatures of pretreatment, mid-treatment (before the first implant), and ‘‘changed’’gene expression patterns between pre- and mid-treatment samples were determined. The ability of the gene signatures to predict local control versus local failure was evaluated. Two-group t test was done to identify the initial gene set separating these end points. Supervised classification methods were used to enrich the gene sets. The results were further validated by leave-one-out and 2-fold cross-validation. Results: Twenty-two patients had suitable material from pretreatment samples for analysis, and 13 paired pre- and mid-treatment samples were obtained. The changed gene expression signatures between the pre- and mid-treatment biopsies predicted response to treatment, separating patients with local failures from those who achieved local control with a seven-gene signature. The in-sample prediction rate, leave-one-out prediction rate, and 2-fold prediction rate are 100% for this seven-gene signature.This signature was enriched for cell cycle genes. Conclusions: Changed gene expression signatures during therapy in cervical cancer can predict outcome as measured by local control. After further validation, such findings could be applied to direct additional therapy for cervical cancer patients treated with chemotherapy and radiation.

Cervical cancer is the most prevalent form of cancer in women just more than 11,000 cases of invasive cervical cancer in developing countries and the second most common cause of diagnosed in 2007 (American Cancer Society Facts and Figures, cancer death. In the Unites States, because of the sensitivity and 2007). Cervical cancer can be a lethal disease, however, with availability of Pap smears, the incidence is much lower, with the survival rate of patients diagnosed with locally advanced cervical cancer (stage II-IV) of only 50% at 5 years. Treatment for cervical cancer has improved through ran-

Authors’ Affiliations: 1Department of Therapeutic Radiology, Yale School of domized trials showing the benefit of sensitizing chemotherapy Medicine; 2Department of Biostatistics,Yale School of Public Health, New Haven, added to radiotherapy for locally advanced disease (1–3). Connecticut; 3Statistical Center, Radiation Therapy Oncology Group, Philadelphia, However, there remain a number of patients who still fail Pennsylvania; 4Department of Radiation Oncology, Jefferson Medical College of treatment, for whom salvage therapy has limited success 5 Thomas Jefferson University, Philadelphia, Pennsylvania; Department of Radiation because it is applied too late. Cancer treatment decisions are Oncology, University of California-Davis Cancer Center, Davis, California; 6Department of Radiation Oncology, The University of Texas M. D. Anderson based on the results of trials evaluating a group of patients with Cancer Center, Houston, Texas; 7Department of Obstetrics and Gynecology, a certain disease as a whole. Therefore, by following group Comprehensive Cancer Center of Wake Forest University, Winston-Salem, North guidelines, there is a subset of patients who will be overtreated 8 Carolina; and Department of Radiation Oncology, Huntsman Cancer Hospital, (who perhaps did not need the full treatment), as well as a University of Utah, Salt Lake City, Utah Received 8/29/08; revised 3/2/09; accepted 3/18/09; published OnlineFirst 6/9/09. subset of patients who will be undertreated (who will fail even Grant support: RTOG U10 CA21661, CCOP U10 CA37422, and Stat U10 with the standard treatment). A test identifying these two CA32115 grants from the National Cancer Institute. J.B.Weidhaas was supported groups could be used to individualize treatments, giving less by K08 grant CA124484. aggressive therapy to the first group and more aggressive Note: The contents of this article are the sole responsibility of the authors and do therapy to the second group. In other malignancies, such as not necessarily represent the official views of the National Cancer Institute. The costs of publication of this article were defrayed in part by the payment of page breast cancer, biological markers are already used to guide charges. This article must therefore be hereby marked advertisement in accordance treatment decisions. with 18 U.S.C. Section 1734 solely to indicate this fact. The biological signature of a cancer can be used to identify Requests for reprints: Joanne B. Weidhaas, Yale University, 333 Cedar Street, tumors that behave clinically different. Often these signatures New Haven, CT 06520. Phone: 203-737-4267; Fax: 203-785-6309; E-mail: [email protected]. are based on gene expression profiling (also referred to as F 2009 American Association for Cancer Research. cDNA microarray profiling), in which the cellular mRNA levels doi:10.1158/1078-0432.CCR-08-2257 are measured. Microarray profiling has been done for more

www.aacrjournals.org 4199 Clin Cancer Res 2009;15(12) June 15, 2009 Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Imaging, Diagnosis, Prognosis

these two time points to identify gene expression signatures Translational Relevance that would predict failure or success of current standard treatment consisting of chemotherapy and radiation therapy This article describes the results of gene expression for locally advanced cervical cancer. Although we were unable analysis correlated with outcome from a phase IIRTOGtrial. to identify a gene signature from patients before treatment that We find that gene expression changes are able to predict predicted outcome, we were able to identify a statistically response as measured by local-regional control in cervical significant seven-gene signature of gene expression changes that cancer to standard treatment with chemotherapy and separated patients by outcome, defined as local-regional radiation. This seven-gene signature, predicting outcome control or local-regional failure. Although others have reported from treatment, could be applied in the future to predict gene expression changes with treatment in cervical cancer (10), who with cervical cancer will fail standard treatment and this is the first study that we know of to show that these may require adjuvant treatment or surgery to achieve cure. changes can predict the response to treatment as measured by Whereas gene expression profiling is not novel, our finding local control. that changes in gene expression can be correlated with outcome has not been previously shown. Materials and Methods

Patient population. The tissue samples were obtained from patients than 15 years, and it generates enormous quantities of data that that participated in clinical trial RTOG C0128, a phase II study of the reflect at some level the innate biology of the tumors studied. cyclooxygenase-2inhibitor Celebrex (celecoxib) and chemoradiation in This approach has been successful in some cancer types in patients with locally advanced cervix cancer and who consented to have identifying groups of patients with more aggressive tumors who their samples used for research purposes. Briefly, 5-fluorouracil and are more likely to fail treatment, such as has been done in cisplatin chemotherapy was administered starting with the initiation of breast cancer (4). Gene expression profiling in cervical cancer radiation, according to the experimental arm of RTOG 9001 (1, 12). has been done and can differentiate the pathologic subtypes of Celecoxib was administered at 400 mg twice daily for 1 y. The radiation cancer versus benign conditions (5–7). There are additionally treatment consisted of 45 Gy given to the pelvis in 25 fractions delivered once daily followed by intracavitary radiation to a total dose cervical cancer gene signatures identified that predict resistance of f85 Gy to point A as is standard. to radiation alone (8, 9). A recent study evaluated patients by Biopsy collection and mRNA extraction. A biopsy of the cervix was microarray that were treated with modern therapy consisting of done with two to three passes of Tischler biopsy forceps. Tumor chemotherapy and radiation (10). Although the sample samples were obtained before treatment, referred to as the pretreatment numbers were small, they were able to identify gene expression sample, and at the first implant, referred to as the mid-treatment patterns that were different in patients that did or did not fail sample. Fresh tissue was placed immediately into RNAlater solution treatment. However, a significant gene signature separating and sent via overnight mail to the RTOG tissue repository. The tissue patients who failed and those who did not was not identified. was divided, with the large portion weighed and frozen in a small We performed gene expression profiling using tumor tissue aliquot of RNAlater and the small portion placed into a paraffin block from a national cervical cancer trial, Radiation Therapy for histologic analysis. Total RNA was extracted with TRIzol reagent from homogenized tumor tissue. Clean-up procedures were done on Oncology Group (RTOG) 0128. This study was a phase II trial the total RNA using RNeasy Midi kit (Qiagen) to ensure removal of any evaluating the benefit of adding Celecoxib to standard 5- remaining contaminants. Total RNA quality was assessed using a fluorouracil and cisplatin chemotherapy and radiotherapy for spectrometer, and ratio absorbance at 260 versus 280 nm was the treatment of locally advanced cervical cancer (stage IB2and determined. The quality of the amplified RNA was assessed using a IIB-IVA). In this study, biopsies were obtained both before 2100 Bioanalyzer (Agilent Technologies) to evaluate the purity of the treatment (pretreatment samples) and during treatment (mid- RNA. The ratios of the integrated 28S RNA peak to the 18S RNA peak treatment samples). Samples were obtained from 34 patients, were used as an indicator of RNA quality after amplification. and 22 pretreatment, 14 mid-treatment, and 13 paired samples cDNA microarray data collection. The hybridization assays and data had RNA of sufficient quality to perform gene expression pro- collection of expression values were done by the Microarray Core filing. As previously reported (11), f9,000 genes were studied Facility at Utah Health Sciences Center. This facility consists of an Amersham BioSciences GEN III Array Spotter and a Gen III Array for expression changes, and no differences in gene expression Scanner. To print microarrays, cDNA clones made from mRNA are PCR were measurable in pretreatment samples when comparing amplified and deposited onto a chemically modified glass surface to stage, age, or ethnic groups. There was a gene signature that form a high-density microscopic array of these genes. Hybridization of separated histologic subtypes, however. In addition, unsuper- labeled RNA samples from cervical cancer tissue versus Universal vised cluster analysis was able to separate gene expression Human Reference RNA to the microarray defines the genetic expression profiles between biopsies obtained pre- and mid-treatment in differences. Specialized microarray chips were designed to include the great majority of cases, showing 91 genes up-regulated and f9,000 particular genes of interest in reference to carcinoma, in 251 genes down-regulated. Gene expression patterns also duplicate. clustered patients into groups, although at the time of that Microarray data preprocess. Data extracted from the scanner were analysis, clinical data had not been collected, so correlation of imported into R. Appropriate R BioConductor (package limma) procedures were used for background correction, intra-array normali- these ‘‘gene signature groups’’ with outcome was not possible. zation, and inter-array normalization. Specifically, background correc- Because the clinical data from RTOG 0128 have now tion was done according to the method of Smyth (13). The corrected matured, in this report we were able to analyze the gene intensities are the expected values given the observed intensity values of expression profiles and their correlation with clinical outcome. the model of fitting a convolution of normal and exponential We studied gene expression patterns pretreatment, mid- distributions of foreground intensities using the background intensities treatment, and also the changes in gene expression between as a covariate. Within-array normalization was done by print-tip LOESS

Clin Cancer Res 2009;15(12) June 15, 2009 4200 www.aacrjournals.org Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Gene Expression Changes and Outcome in Cervical Cancer

(14, 15). Data between arrays were further normalized by the scale local-regional control (LRC) status, local control (LC) status, and normalization method (14–16). M values, short for log 2(R/G), were patient survival status. Because the LRC status was completely correlated averaged for duplicate spots. These M values were used for further data with patient survival status, we only report LRC in the results. analysis. Supporting vector machine methods implemented in R have been Statistical analysis of gene expression data. Unsupervised hierarchi- used to validate and refine the gene signature. Leave-one-out and 2-fold cal clustering analyses were done with Cluster software and visualized cross-validations were done. For the Mdiff efficacy signature validation, with TreeView (17) using complete linkage with appropriate distance the first 2-fold cross-validation subgroup includes all the possible test calculations indicated under corresponding figures. samples including at least one LC status 1 sample; the second subgroup Paired t tests were used to study the treatment-regulated gene includes all possible test samples with both samples having LC status 0 expression between pretreatment samples and mid-treatment samples. (local control). Because the sample is small, we can generate the whole Two-group t tests were used to discover various gene expression permutation set by relabeling response groups using the same number signatures as candidates for toxicity outcomes and efficacy outcomes. of responses. In this study, for the permutation set of 13 samples total Reported P values are the adjusted P values using the methods with 2with LC status 1 (local failure), we have a permutation set of 78. proposed by Benjamini and Hochberg (18) except in the case of the For each subset of genes, we try to generate a supporting vector machine initial finding of candidate genes for local regression efficacy markers. model for each of the 78 data sets. We define the permutation The cutoff of P values is generally 0.01 unless otherwise specified. separation rate as how many data sets can have a supporting vector In addition, gene signatures exclude genes that do not have any raw R machine that totally separates the two response groups. We use this as a values >100. benchmark to curb overparameterization. Heuristic methods used to Pretreatment gene expression, mid-treatment gene expression, and refine gene signatures will be elucidated with corresponding results.

Mdiff values were evaluated in relation to toxicity and efficacy end Functional analysis. To corroborate our results with biological points. Mdiff values are calculated as mid-treatment M values minus plausibility, genes that met the P value cutoff criteria indexed by pretreatment M values for paired samples and thus are measures of the Unigen cluster id along the P values were uploaded into the Ingenuity gene expression changes. The toxicity outcomes were dichotomized by Pathway Analysis (IPA) software. Most of the reported gene names (a) whether patients experienced any grade 4 toxicity; (b) whether were the IPA annotation mapped through Unigen Cluster id. Genes patient’s median toxicity grade is z2;(c) whether patients experienced were associated with biological functions and/or diseases in the IPA any grade 3 or higher gastrointestinal toxicity; and (d) whether patients knowledge base. Fisher’s exact test was used to calculate a P value experienced any genitourinary toxicity. The cutoff points for toxicity determining the probability that each biological function and/or dichotomization are rather arbitrary with the goal to have a relatively disease assigned to that data set is due to chance alone. The cutoff for balanced sample size in two groups. Efficacy binary end points include P values is 0.05.

Table 1. Pretreatment and mid-treatment differentially expressed genes

Cluster Gene symbol* Pretreatment Mid-treatment Expression average M value average M value fold change 1 ENPP2 -0.68 -1.72 -2.06 1 VCAM1 -0.95 -2.43 -2.79 1 MGC5370 -0.25 -1.19 -1.92 1 MMP2 -0.48 -2.36 -3.68 1 GUCY1A3 -0.37 -1.39 -2.03 1 CXCL12 -0.39 -2.41 -4.05 1 CXCL12 -0.33 -1.55 -2.34 1 NOPE -0.12 -1.23 -2.15 1 RRM2B -0.06 -1.26 -2.30 1 XPC 0.01 -0.31 -1.25 1 C1ORF54 -0.44 -0.81 -1.29 1 CSF2RB -1.25 -2.65 -2.63 1 CDKN1A -1.61 -3.07 -2.76 1 CDKN1A -1.49 -2.92 -2.70 2 UNG 0.18 1.06 1.85 2 AURKA 0.61 1.39 1.71 2 DKFZP762E1312 1.27 2.16 1.85 2 FAM72A (includes EG:729533)1.08 1.99 1.88 2 CDCA5 1.21 2.60 2.62 2 CDCA5 1.14 2.34 2.30 2 FOXM1 1.21 2.50 2.44 2 MAD2L1 0.96 1.91 1.92 2 CHEK1 0.99 1.69 1.62 2 RPS19BP1 0.92 2.54 3.07 2 CIT 0.22 0.63 1.33 2 RFC3 0.56 1.15 1.51 2 HMGB2 0.66 1.67 2.02 2 KIF23 0.60 1.39 1.73 2 KIAA0101 0.76 2.01 2.37 2 CEP55 0.44 1.62 2.26 2 EXOSC9 0.53 1.79 2.39

*The gene symbol column is in the same order as that of the Fig. 1 for easier reference.

www.aacrjournals.org 4201 Clin Cancer Res 2009;15(12) June 15, 2009 Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Imaging, Diagnosis, Prognosis

Results analysis revealed that the major molecular and cellular functions of these genes are cell cycle (12of 29),DNA Pretreatment or mid-treatment gene expression signatures do replication, recombination and repair (16 of 29), and cell death not predict treatment outcome. Biopsies were obtained both (12of 29).Many of these genes are also known to be associated before (pretreatment samples) and during (mid-treatment with cancer (16 of 29; Table 2). samples) treatment from 34 patients, and 22 pretreatment, 14 A pattern of gene expression changes correlates with treatment mid-treatment, and 13 paired samples had RNA of sufficient outcome. In the 13 patients with paired pretreatment and post- quality to perform gene expression profiling. We began by treatment samples, the LC, LRC, and survival status classifica- evaluating whether pretreatment or mid-treatment gene expres- tion overlapped. We therefore used only the LC classification in sion patterns alone could predict binary treatment outcomes, all of the following analysis. Two patients are in the LC status 1 namely the LRC, LC, and/or patient survival status. Whereas we class [indicating they failed locally (median time to failure, 18.8 found that there were patterns of gene expression that varied months)] and 11 patients are in the LC status 0 class [indicating across these outcomes, we found no genes expressed differen- that they achieved cure (median follow-up, 26.6 months)]. We tially using the adjusted P value cutoff we applied to these studied the Mdiff values, generated as the M value difference studies. Therefore, there is not a significant gene signature that between mid-treatment and pretreatment gene expression, in can predict outcome from either the pretreatment or mid- relation to the LC classification. treatment samples alone in this study. Three hundred thirty-eight genes were identified that have Changes in gene expression with treatment. Using the 13 different Mdiff values between the two LC status classes with a paired pre- and mid-treatment samples with a multiple false discovery rate–adjusted P value cutoff of 0.01. Of these, comparison adjusted P value cutoff of 0.01, we discovered 29 287 could be mapped to IPA id. The top molecular and cellular genes whose expression changed 2- to 5-fold pre- to mid- functions associated with these genes are cellular growth and treatment. The unsupervised hierarchical clustering grouped proliferation (83), cell death (63), gene expression (63), these genes into two distinct groups. The first group of genes, cellular development (58), and cell cycle (38; data not shown). labeled as cluster 1 in Table 1, started with lower expressions The top diseases and disorders that these genes are associated than normal tissue controls (Universal Human Reference RNA) with include cancer (78), connective tissue disorders (48), and pretreatment; interestingly, the expression of these genes went hematologic disease (33; data not shown). further down by the mid-treatment biopsy (B versus A, Fig. 1). We were able to define a smaller subset of 35 genes that met The mean fold of down-regulation was 2.43 with a range of the adjusted P value cutoff of 0.001. Thirty-two of them were 1.25 to 4.05. In contrast, the second group of genes, labeled as mapped to IPA ids. The top molecular and cellular functions cluster 2in Table 1, started with higher expression compared associated with these genes are cell growth and proliferation with normal tissue controls, and on treatment, their expression (11), cell death (9), cell morphology (8), cellular movement went further up (B versus A, Fig. 1). The mean fold up- (7), and cell-to-cell signaling and interaction (6). The top regulation was 2.05 with a range of 1.33 to 3.07. The functional diseases and disorders these groups of genes are associated with

Fig. 1. Treatment-regulated gene expression. A set of genes decrease expression and a set of genes increase expression when comparing gene expression levels between pretreatment (A) and posttreatment (B).

Clin Cancer Res 2009;15(12) June 15, 2009 4202 www.aacrjournals.org Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Gene Expression Changes and Outcome in Cervical Cancer

Table 2. Functions of treatment-regulated genes

Function Molecules Molecular and cellular functions Cell cycle MAD2L1 KIF23 DKFZP762E1312 CDCA5 AURKA CXCL12 FOXM1 XPC CIT CHEK1 CDKN1A RRM2B DNA replication, recombination, KIF23 HMGB2 DKFZP762E1312 CDCA5 RFC3 CXCL12 MMP2 CIT CDKN1A MAD2L1 AURKA and repair GUCY1A3 FOXM1 XPC CHEK1 UNG Cell death MAD2L1 AURKA VCAM1 FOXM1 CXCL12 MMP2 CIT XPC CHEK1 CDKN1A RRM2B UNG Disease and disorder Cancer HMGB2 KIF23 DKFZP762E1312 ENPP2 VCAM1 CXCL12 MMP2 CDKN1A RRM2B MAD2L1 AURKA KIAA0101 FOXM1 XPC CHEK1 UNG Gastrointestinal disease MAD2L1 HMGB2 ENPP2 VCAM1 AURKA CXCL12 CHEK1 CDKN1A RRM2B UNG

are cancer (11) and reproductive disease (9; Table 3). The two as is cervical cancer, our findings support protocols to attempt LC classes of samples can be clearly separated into hierarchical this approach because it seems to allow the identification of clustering using 338 genes (image not shown) or 35 genes significant gene signatures even with fewer patient numbers. (Fig. 2). Knowing the future outcome in a patient midway through We were able to further obtain a validated gene signature treatment would allow the application of additional therapy to consisting of seven genes heuristically. Specifically, 31 data sets those who are destined to fail, such as surgical intervention, were set up by including the top 5 up to 35 genes ranked by when possible, or additional adjuvant chemotherapy. These ascending P values, respectively. For example, the first data set decisions will of course need to be made only with additional includes the top 5 genes, the second the top 6, etc. The self- validation of the findings of this work. prediction rate for all data sets is 100%. For each gene set, the In cervical cancer, because of the ease of obtaining tissue from prediction rate of leave-one-out cross-validation prediction rates patients and the need to better classify patients as poor risk, gene for the two subgroups with 2-fold cross-validation and the expression profiling has been previously done. Several groups permutation separation rate are depicted in Fig. 3A. The self- have obtained microarray findings using techniques similar to prediction rate, 2-fold cross-validation rates, increases to and those used in this study to predict resistance to radiation alone plateaus at 100%. The permutation separation rate is low at in the treatment of cervical cancer (8, 9). Both of the studies seven genes and increases when we include more genes. have more vigorous cross-validation than this study due to their Therefore, we chose the top seven genes as the final gene bigger sample size and more balanced sample number in signature of seven. The hierarchical clustering could still cleanly different response classes. However, none of them have used the separate the two efficacy classes with these seven genes (Fig. 3B). Mdiff measures because both use biopsies from a single time point, pretreatment. In this study, we obtained a later tissue Discussion biopsy, and we were able to show that altered gene expression is a powerful predictor of response to treatment, even in this small In this study, we have analyzed the association of gene study. We hypothesize that because there are several levels of expression signatures in cervical cancer with outcome, mea- translational control, perhaps changing gene expression levels sured as local control, in patients from a national trial treated might be more accurate surrogates for gene expression response with chemotherapy and radiation, the standard of care. than stable mRNA levels, allowing significant findings even with Although we were unable to find a gene signature from either small patient numbers. the pre- or mid-treatment biopsy as a group that could predict We were able to find similarities between our Mdiff signatures outcome, we were able to identify a seven-gene signature whose and findings from both studies predicting radioresistance. The expression changes significantly predict local failure. This is the caveat is that all the cDNA arrays are custom made and may not first study of gene expression changes that we are aware of that necessarily represent the same set of transcripts. For example, predicts response to cancer treatment. Although other tumors 16 genes are described in the study by Wong et al. (9), but only may not be as amenable to a biopsy midway through treatment 7 of them are represented on our cDNA arrays based on

Table 3. Gene functions of Mdiff gene signatures, 35-gene set

Function Molecules Molecular and cellular functions Cellular growth and proliferation FBLN5 CD58 DKK1 HGF CTBP1 IGBP1 EFS PTPRE FER1L3 FIS1 IGFBP3 Cell death DKK1 HGF MYCT1 RCAN2 CTBP1 PTPRE GSN FIS1 IGFBP3 Cell morphology DKK1 HGF MYCT1 IGBP1 PTPRE GSN FIS1 IGFBP3 Cellular movement FBLN5 HGF IGBP1 EFS TRIP6 GSN IGFBP3 Cell-to-cell signaling and interaction FBLN5 CD58 HGF PTPRE GSN IGFBP3 Disease and disorder Cancer DKK1 FBLN5 HGF RCAN2 CTBP1 IGBP1 TRIP6 PTPRE GSN FIS1 IGFBP3 Reproductive System Disease DKK1 HGF RCAN2 PTPRE GSN FIS1 IGFBP3

www.aacrjournals.org 4203 Clin Cancer Res 2009;15(12) June 15, 2009 Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Imaging, Diagnosis, Prognosis

Fig. 2. Hierarchical clustering of Mdiff signature gene set 35. Hierarchical clustering is done using complete linkage with Pearson correlation. LC1indicates local failure; LC0 indicates local control.

National Center for Biotechnology Information accession (fibulin 5), SPARC (SPARC/osteonectin), CREM (cyclic AMP number. Of these seven genes, DDB1 (damage-specific DNA responsive element modulator), and DDX1 [DEAD/H(Asp-Glu- binding protein 1) was also identified in our study with Ala-Asp/His) box polypeptide 1] as predictors of local control significant Mdiff values between the two response groups. The versus failure. Whereas the similarities in our findings are cDNA microarrays used in the Kitahara et al. (8) study have the reassuring, we believe that in the era of standard therapy most extensive gene presentation; also, the study was based on consisting of chemotherapy and radiation, our findings are more patient samples. Our long Mdiff gene signatures do currently most applicable to predict response to treatment. overlap with this study as well; both studies found SRF (serum A recent report from M. D. Anderson Cancer Center also response factor, c-fos serum response element-binding), FBLN5 described a study where biopsies were taken at two time points,

Clin Cancer Res 2009;15(12) June 15, 2009 4204 www.aacrjournals.org Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Gene Expression Changes and Outcome in Cervical Cancer with gene expression measured at both points in locally that grouped in cell cycle regulation. In fact, in our list of advanced cervical cancer treated with chemotherapy and differentially expressed genes, 7 of the 35 gene signature are cell radiation (10). In this study, there was a pretreatment biopsy cycle associated, and the odds of 7 cell cycle genes appearing in and a biopsy obtained 3 days after the initiation of a random list of 35 is P =4.3Â 10e-7. These findings suggest chemotherapy and radiation. The authors also confirmed gene that control of cell cycle after exposure to chemotherapy and expression changes, yet with unsupervised clustering, they were radiation is an important factor in ultimate tumor cell death unable to identify gene signatures that predicted recurrence for and local control in cervical cancer. The genes remaining in the their patients. In our study, the second biopsy occurred seven-gene signature also were interesting. One example is approximately halfway through treatment, and it is possible CD58, which has linkages with T-cell immunity and, possibly, that the gene expression changes we are detecting reflect cell p53 function (19). High CD58 has been associated with death, and this could account for our ability to use these gene progression of cervical intraepithelial neoplasia (20) to invasive expression changes as biomarkers of local control. cervical cancer. In addition, the WFS1 gene was found in our The cluster analysis of the raw gene list whose expression seven-gene signature and has been implicated in cancer and is changes predicted local control versus failure identified several part of the endoplasmic reticulum stress response (21). Finally,

Fig. 3. Identification and validation of seven-gene set. A,Visualization of cross-validation and further gene signature refinement. L00CV: leave one samples out cross validation. 2-fold CV (1) and 2-fold (2): see Results for details. Permutations Separation Rate: percentage of permutation samples out of all possible permutation sets (relabeling classes) that can have a total separation model. B, Hierarchical clustering of Mdiff signature gene set 7. LC1indicates local failure; LC0 indicates local control.

www.aacrjournals.org 4205 Clin Cancer Res 2009;15(12) June 15, 2009 Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Imaging, Diagnosis, Prognosis

NKG7 was identified in the signature and has been shown to be will need to be validated in a large independent study. induced by bovine papillomavirus type 1, suggesting that this However, our findings do support the future study of gene gene might be regulated by papilloma virus (22). Because expression changes in cervical cancer, as well as other cancers, human papillomavirus infection has been shown be associated as potential new ways to predict local control. with a better response to radiation (23), one could hypothesize that one mechanism may be through regulation of NKG7. Disclosure of Potential Conflicts of Interest Some of these genes could be potential future targets to enhance response to therapy. No potential conflicts of interest were disclosed. This study has some major limitations that should be noted, including our inability to study survival status separately from Acknowledgments local control or to perform multivariate analysis due to the small sample size. The gene signature identified in this study We thankTony Magliocco for assistance in interpreting the gene findings.

References 1. MorrisM,EifelP,LuJ.Pelvicradiationwithconcur- 8. Kitahara O, Toyomasa K, Tsunoda T, Harima Y, 16. Smyth GK. SpeedTP,Normalization of cDNA micro- rent chemotherapy compared with pelvic and para- Nakamura Y. Classification of sensitivity or resis- array data. In: Carter D, editor. Methods: selecting aortic radiation for high-risk cervical cancer. N Engl J tance of cervical cancers to ionizing radiation candidate genes from DNA array screens: application Med 1999;340:1137 ^43. according to expression profiles of 62 genes se- to neurosceince.Vol. 31. 2003. p. 265 ^ 73. 2. Pearcey R, Brundage M, Drouin P, et al. Phase III trial lected by cDNA microarray analysis. Neoplasia 17. Eisen MB, Spellman PT, Brown PO, Botstein D. comparing radical radiotherapy with and without cis- 2002;4:295 ^ 303. Cluster analysis and display of genome-wide expres- platin chemotherapy in patients with advanced squa- 9. Wong Y, Sahota D, Chung T, et al. Gene expression sion patterns. Proc Natl Acad Sci U S A 1998;95: mous cell cancer of the cervix. J Clin Oncol 2002;20: pattern associated with radiotherapy sensitivity in cer- 14863^8. 966^72. vical cancer. Cancer J 2006;12:189 ^93. 18. Benjamini Y, Hochberg Y. Controlling the false 3. Rose PG, Bundy BN, Watkins EB, et al. Concurrent 10. Klopp A, Jhingran A, Ramdas L, et al. Gene expres- discovery rate: a practical and powerful approach to cisplatin-based chemoradiation improves progression sion changes in cervical squamous cell carcinoma af- multiple testing. J Royal Stat Soc, Ser B 1995;57: free and overall survival in advanced cervical cancer: ter initiation of chemoradiation and correlation with 289 ^ 300. results of a randomized Gynecologic Oncology Group clinical outcome. Int J Radiation Oncology Biol Phys 19. Sampaziotis F, Kokotas S, Gorgoulis VG. P53 study. N Engl J Med 1999;340:1144 ^ 53. 2008;71:226^36. possibly upregulates the expression of CD58 4. SorlieT, Perou CM,Tibshirani R, et al. Gene expres- 11. Zempolich K, Fuhrman C, Milash B, et al. Changes in (LFA-3) and CD59(MIRL). Med Hypotheses 2002; sion patterns of breast carcinomas distinguish tumor gene expression induced by chemoradiation in ad- 58:136^40. subclasses with clinical implications. Proc Natl Acad vanced cervical carcinoma: a microarray study of 20. Mota F, Rayment N, Chong S, Singer A, Chain B. Sci U S A 2001;98:10869 ^ 74. RTOG C-0128. Gynecol Oncol 2008;109:275 ^ 9. The antigen-presenting environment in normal and 5. Contag S, Gostout B, Clayton A, Comparison of 12. Gaffney D, Winter K, Fuhrman C, et al. Feasibility of human papillomavirus (HPV)-related premalignant gene expression in squamous cell carcinoma and ade- RNA collection for micro-array gene expression analy- cervical epithelium. Clin Exp Immunol 1999;116: nocarcinoma of the uterine cervix. Gynecol Oncol sis in the treatment of cervical carcinoma: a scientific 33^40. 2004;95:610^ 7. correlate of RTOG C-0128. Gynecol Oncol 2005;97: 21. Ueda K, Kawano J,Takeda K, et al. Endoplasmic re- 6. Santin A, Zhan F, Bignotti E. Gene expression pro- 607 ^ 11. ticulum stress induces Wfs1 gene expression in pan- files of primary HPV16- and HPV18-infected early 13. Smyth GK. Bioinformatics and computational biolo- creatic h-cells via transcriptional activation. Eur J stage cervical cancers and normal cervical epithelium: gy solutions using R and bioconductor. New York: Endocrinol 2005;153:167 ^ 76. identification of novel candidate molecular markers for Springer; 2005. p. 397 ^ 420. 22.Yuan ZQ, Nicolson L, Marchetti B, Gault EA, Campo cervical cancer diagnosis and therapy.Virology 2005; 14. Yang YH, Dudoit S, Luu P, Speed TP. Normalization MS, Nasir L. Transcriptional changes induced by 331:269^91. for cDNA microarray data. In: Microarrays: optical bovine papillomoavirus type 1 in equine fibroblasts. 7. SongJ, LeeJ, LeeJ, Jung H, Kim S, Lee K. Microarray technologies and informatics. 2001. p. 141 ^ 52. J Virol 2008;82:6481^91. analysis of normal cervix, carcinoma in situ,andinva- 15. Yang YH, Dudoit S, Luu P, et al. Normalization for 23. Harima Y, Sawada S, Nataga K, Sougawa M, sive cervical cancer: identification of candidate genes cDNA microarray data: a robust composite method OhnishiT. Human papillom virus (HPV) DNA associat- in pathogenesis of invasion in cervical cancer. Int J addressing single and multiple slide systematic varia- ed with prognosis of cervical cancer after radiothera- Gynecol Cancer 2008;18:1051 ^ 9. tion. Nucleic Acids Res 2002;20:e15. py. Int J Radiat Oncol Biol Phys 2002;52:1345 ^ 51.

Clin Cancer Res 2009;15(12) June 15, 2009 4206 www.aacrjournals.org Downloaded from clincancerres.aacrjournals.org on October 1, 2021. © 2009 American Association for Cancer Research. Changes in Gene Expression Predicting Local Control in Cervical Cancer: Results from Radiation Therapy Oncology Group 0128

Joanne B. Weidhaas, Shu-Xia Li, Kathryn Winter, et al.

Clin Cancer Res 2009;15:4199-4206.

Updated version Access the most recent version of this article at: http://clincancerres.aacrjournals.org/content/15/12/4199

Cited articles This article cites 20 articles, 5 of which you can access for free at: http://clincancerres.aacrjournals.org/content/15/12/4199.full#ref-list-1

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Subscriptions Department at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://clincancerres.aacrjournals.org/content/15/12/4199. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.