Expression Profiles of Osteosarcoma That Can Predict Response to Chemotherapy

Research Article

Tsz-Kwong Man,1 Murali Chintagumpala,1 Jaya Visvanathan,1 Jianhe Shen,1 Laszlo Perlaky,1 John Hicks,2 Mark Johnson,3 Nelson Davino,3 Jeffrey Murray,4 Lee Helman,5 William Meyer,6 Timothy Triche,7 Kwong-Kwok Wong,1 and Ching C. Lau1

1Departments of Pediatrics, Texas Children’s Cancer Center; Departments of 2Pathology and 3Orthopedic Surgery, Texas Children’s Hospital/Baylor College of Medicine, Houston, Texas; 4Cook Children’s Medical Center, Fort Worth, Texas; 5Pediatric Oncology Branch, National Cancer Institute, Bethesda, Maryland; 6University of Oklahoma Health Sciences Center, Oklahoma City, Oklahoma; and 7Children’s Hospital Los Angeles, Los Angeles, California

Abstract the resected tumor specimen is assessed for the degree of necrosis, Osteosarcoma is the most common malignant bone tumor in which is a reliable and the only significant prognostic factor in children. After initial diagnosis is made with a biopsy, patients with nonmetastatic disease and is used to guide the choice treatment consists of preoperative chemotherapy followed by of postoperative chemotherapy. Patients whose tumors display z definitive surgery and postoperative chemotherapy. The 90% necrosis (good or favorable response) have an excellent degree of tumor necrosis in response to preoperative prognosis and continue to receive chemotherapy similar to the chemotherapy is a reliable prognostic factor and is used to preoperative regimen. Patients whose tumors display <90% guide the choice of postoperative chemotherapy. Patients necrosis (poor or unfavorable response) have a much higher risk with tumors, which reveal z90% necrosis (good responders), of relapse and poor outcome even after complete resection of the have a much better prognosis than those with <90% necrosis primary tumor (2). To improve the outcome of the poor (poor responders). Despite previous attempts to improve the responders, attempts are usually made to use postoperative outcome of poor responders by modifying the postoperative chemotherapy regimens that are different from the preoperative chemotherapy, their prognosis remains poor. Therefore, there regimen by the addition or replacement of a chemotherapeutic is a need to predict at the time of diagnosis patients’ response agent. Such attempts in the past have been unsuccessful (1, 3) to preoperative chemotherapy. This will provide the basis for partly because the degree of necrosis is known only after 8 to 10 developing potentially effective therapy that can be given at weeks of preoperative therapy. It is possible that resistant tumor the outset for those who are likely to have a poor response. cells have additional time to either metastasize to the lungs or Here, we report the analysis of 34 pediatric osteosarcoma evolve further during the period when ineffective preoperative samples by expression profiling. Using parametric two-sample chemotherapy is given. Therefore, there is a need to identify at the t test, we identified 45 genes that discriminate between good time of initial diagnosis the patients who are likely to have a poor and poor responders (P < 0.005) in 20 definitive surgery response to standard preoperative therapy and therefore a poor out- samples. A support vector machine classifier was built using come eventually. Therapies tailored to improve the outcome for these predictor genes and was tested for its ability to classify those patients identified at the time of diagnosis to have a poor out- initial biopsy samples. Five of six initial biopsy samples that come can then be instituted at the outset when the chance for had corresponding definitive surgery samples in the training success is potentially higher. Although several other prognostic set were classified correctly (83%; confidence interval, 36%, factors have been proposed for predicting the long-term outcome 100%). When this classifier was used to predict eight of osteosarcoma patients, most are still controversial or have not independent initial biopsy samples, there was 100% accuracy been tested in large prospective studies (4–11). (confidence interval, 63%, 100%). Many of the predictor genes Recently, application of microarray technology to classify and are implicated in bone development, drug resistance, and diagnose various types of tumors has yielded promising results tumorigenesis. (Cancer Res 2005; 65(18): 8142-50) (12, 13). However, the use of this technology to predict response to chemotherapy in pediatric solid tumors is still in its infancy. In this Introduction study, we developed a multigene predictive model to classify good and poor responders of osteosarcoma in response to preoperative Osteosarcoma is the most common malignant bone tumor in chemotherapy using gene expression profiling. We used a slightly f children and accounts for 60% of malignant bone tumors different approach from previously published works (14). We first diagnosed in the first two decades of life (1). After the diagnosis is identified a molecular signature of chemoresistance by comparing made by an initial biopsy, standard treatment involves the use of the expression profiles of the definitive surgery samples of the good multiagent chemotherapy, definitive surgery of the primary tumor, responders with those of the poor responders, which, in principle, and postoperative chemotherapy. At the time of definitive surgery, have been enriched for resistant cells. We then tested the hypothesis that this predictor signature of chemoresistance could recognize the resistant cells present in the initial biopsy of some of the same Note: Supplementary data for this article are available at Cancer Research Online cases used in the first analysis (definitive surgery samples), although (http://cancerres.aacrjournals.org/). Requests for reprints: Ching C. Lau, Texas Children’s Hospital, 6621 Fannin Street, these resistant cells might have constituted only a small fraction of MC 3-3320, Houston, TX 77030-2399. Phone: 832-824-4543; Fax: 832-825-4038; E-mail: the primary tumor. Finally, we tested the ability of this signature to [email protected]. I2005 American Association for Cancer Research. predict chemoresistance in an independent set of initial biopsies doi:10.1158/0008-5472.CAN-05-0985 and found that there was 100% accuracy in its prediction.

Cancer Res 2005; 65: (18). September 15, 2005 8142 www.aacrjournals.org

Materials and Methods initial biopsy or a definitive surgery specimen. The initial biopsy samples were obtained at the time of diagnosis before the initiation of preoperative Patients and tumor samples. The clinical information of the chemotherapy. The definitive surgery samples were collected after the osteosarcoma samples used in this study is summarized in Table 1. All completion of preoperative chemotherapy. The good responders were samples were collected through institutional review board–approved defined as patients whose tumors had z90% necrosis in response to protocols at four centers (Texas Children’s Hospital/Baylor College of preoperative chemotherapy as determined by histologic examination at the Medicine, Cook Children’s Medical Center, Pediatric Branch of the National time of definitive surgery and poor responders had <90% necrosis. In this Cancer Institute, and University of Oklahoma Health Sciences Center) after study, the percentage necrosis in the poor responders ranged from 5% to informed consents were signed. All but three patients (10, 20, and 26) were 86%. Five of the 28 patients were diagnosed with metastatic disease at enrolled in the same treatment protocol, the schema of which is shown in presentation. Immediately after collection, tumor specimens were snap Fig. 1. All patients, including the three not enrolled in the protocol, received frozen in liquid nitrogen and stored at À80jC until RNA extraction. All the same preoperative chemotherapy consisting of cisplatin, doxorubicin, samples used for RNA extraction were immediately adjacent to the frozen and high-dose methotrexate, except patients 3, 4, 8, and 24, who received sections used for diagnostic purpose and were representative of the only cisplatin and doxorubicin. All pathologic diagnoses were centrally corresponding tumors. All initial biopsy specimens were confirmed to reviewed by a single pathologist (J.H.). A total of 34 samples (14 initial contain >80% tumor cells. biopsies and 20 definitive surgery specimens) were included in this study, RNA extraction. Total RNA was extracted from tissues and cultured cells which were obtained from 28 individual patients, 18 males and 10 females. using TRIzol reagent (Invitrogen, Carlsbad, CA) following the manufac- The age of the patients ranged from 9 to 22 years. Six patients contributed turer’s protocol. Osteosarcoma samples were homogenized in TRIzol two samples each, both initial biopsies and definitive surgery specimens, reagent using 1 mL per 50 to 100 mg tissue. Normal human osteoblast whereas the remaining 22 patients contributed one sample each, either an primary culture cells obtained from Clonetics (San Diego, CA) were used as

Table 1. Clinical information of the osteosarcoma samples used in this study

Patient Tumor ID IB/DS Age (y) Gender Metastatic Histologic Recurrence Follow-up (mo) Status status response

1 197 IB 12 F N PR Yes 60 NED* 221 DS 2 207 IB 13 M N PR No 59 NED 236 DS 3 278 IB 13 M N GR No 51 NED 308 DS 4 289 IB 13 F N PR Yes 18 DOD 311 DS 5 345 IB 11 M N GR Yes 28 DOD 394 DS 6 410 IB 16 M N GR Yes 38 AWD 452 DS 7 204 IB 18 F N PR No 60 NED 8 241 DS 15 F N PR Yes 58 NED 9 252 DS 14 F M PR Yes 38 DOD 10 274 IB 13 M N PR Yes 16 DOD 11 299 IB 11 M M GR Yes 26 DOD 12 300 DS 15 F N GR Yes 44 DOD 13 342 DS 12 M N PR No 48 NED 14 386 DS 13 M N GR No 45 NED 15 392 DS 13 M N PR Yes 14 DOD 16 464 IB 14 F N PR Yes 14 DOD 17 479 IB 15 F N PR Yes 20 DOD 18 481 IB 19 M M PR Yes 8 DOD 19 483 DS 15 M M PR No 36 NED 20 545 IB 9 M N GR No 32 NED 21 591 DS 14 M M PR Yes 25 DOD 22 654 IB 22 M N GR No 19 NED 23 680 DS 15 M N PR No 30 NED 24 691 DS 15 F N PR Yes 6 DOD 25 759 DS 11 M N GR Yes 23 NED 26 760 DS 17 M N PR Yes 23 NED* 27 761 DS 21 M N PR No 25 NED 28 771 DS 15 F N GR Yes 48 NED*

Abbreviations: IB, initial biopsy; DS, definitive surgery. For metastatic status: M, metastatic; N, nonmetastatic. For histologic response: GR, good responder; PR, poor responder. For status: NED, no evidence of disease; AWD, alive with disease; DOD, died of disease. *Patient underwent resection of lung metastases and salvage chemotherapy.

www.aacrjournals.org 8143 Cancer Res 2005; 65: (18). September 15, 2005

Figure 1. Schema of the protocol for this study. Dexrazoxane is given as an adjuvant with doxorubicin for cardioprotection. In this protocol, all patients with localized disease and good response to preoperative chemotherapy will receive the same chemotherapeutic agents in postoperative chemotherapy. Poor-risk patients, including the poor responders and those that have unresectable primary tumors or metastatic disease, will receive the high-dose cytoxan and melphalan with peripheral stem cell rescue as postoperative therapy. CDDP, cisplatin; DOX, doxorubicin; MTX, methotrexate.

reference for microarray experiments and were lysed in 1 mL TRIzol reagent random hexamers and SuperScript reverse transcriptase II (Invitrogen). per 5 Â 106 to 10 Â 106 cells. RNA concentrations were determined by Quantitative real-time PCR was then carried out using the ABI Prism 7000 spectrophotometry, and RNA integrity was assessed by either agarose gel Sequence Detection System (Applied Biosystems, Foster City, CA) following electrophoresis or Bioanalyzer (Agilent Technologies, Palo Alto, CA). the manufacturer’s protocol and using gene-specific primers. All log ratio T7 RNA amplification. RNA amplification of total RNA was done based values were corrected for rRNA and referenced to normal human osteoblast on the modified Eberwine protocol (15). For first-strand cDNA synthesis, cells using DDCT method (Applied Biosystems). The Pearson and Spearman 1 Ag total RNA from each sample was mixed with T7-oligo(dT) primer (5V- correlation coefficients were calculated using SPSS statistical software GCATTAGCGGCCGCGAAATTAATACGACTCACTATAGGGAGATTTTTTT- (SPSS, Inc., Chicago, IL). TTTTTTTTTTTVN-3V), where V and N represent mixtures of G/C/A and G/ Data analysis. Raw quantification output of all array experiments were C/A/T, respectively. The RNA/primer mixture was denatured followed by subjected to data analysis using BRB Array Tools 3.1.0 developed by first-strand synthesis using Moloney murine leukemia virus (Invitrogen) at R. Simon and A. Pang Lam (http://linus.nci.nih.gov/BRB-ArrayTools.html). 37jC for 75 minutes. Second-strand synthesis was done using Escherichia Fluorescence intensities of the arrays were preprocessed and filtered by coli DNA ligase, E. coli RNase H, and E. coli DNA polymerase I (Invitrogen). removing spots with low signal intensity and low sample variance (P > The mixture was incubated at 16jC for 2 hours and then at 70jC for 0.01) as well as those that were missing in >50% of the experiments. (For 15 minutes. cDNA was purified using ChromaSpin TE-200 spin columns (BD details of sample variance calculations, refer to http://linus.nci.nih.gov/ Biosciences, Palo Alto, CA), dried, and resuspended in 8 AL DEPC water. BRB-ArrayTools.html.) Intensities were then normalized by intensity- In vitro transcription of double-stranded cDNA was done using the dependent local weighted regression (LOWESS) method. This method has Ampliscribe kit (Epicentre Technologies Corp., Madison, WI) according to been shown to perform better than other methods for normalizing cDNA the manufacturer’s instructions. Amplified RNA was purified using RNeasy microarrays (16). After normalization, intensity ratios were log trans- Mini kit (Qiagen, Valencia, CA) and quantified by fluorescence spectroscopy formed before any further analysis. using a RiboGreen RNA quantitation assay (Molecular Probes, Eugene, OR). Slightly >3,000 informative spots (3,018) remained after preprocessing Microarray experiments. Labeling of T7-amplified RNA samples was and filtering and were used for supervised classification. The classification carried out using an amino-allyl labeling kit (Ambion, Inc., Austin, TX). In algorithms tested include compound covariate predictors (CCP; ref. 17), brief, 0.5 Ag tumor or reference (normal human osteoblasts) amplified RNA k-nearest neighbor, nearest centroid, support vector machine (SVM), and was mixed with spiked control RNA (see below; Perkin-Elmer, Boston, MA) diagonal linear discriminant analysis (LDA; refs. 18, 19). These algorithms and random decamers. The mixture was denatured and reverse transcribed explore different aspects of the data to perform classification. For example, with amino-allyl dUTP. Cy3 or Cy5 monoreactive dye (Amersham Pharmacia CCP and LDA use linear combinations of different weighted expression Biotech, Arlington Heights, IL) was then added for labeling. Labeled tumor ratios for classification, whereas k-nearest neighbor and nearest centroid and reference cDNAs were combined and purified by NucAway spin columns are nonlinear and nonparametric methods (17, 18). SVM finds the optimal (BD Biosciences), dried, and redissolved in hybridization solution (240 AL hyperplane that is able to separate the data by projecting them into a high-

H2O, 250 ALof20Â SSC, 10 AL of 10% SDS, 500 AL formamide per 1 mL dimensional space (19). The feature extraction (predictor gene selection) solution) and combined with human Cot-1 DNA (10 mg/mL) and yeast tRNA was done using two-sample t test and cutoff of P V 0.005. We divided our (4 mg/mL). Probes were denatured and hybridized to cDNA microarrays that 34 samples into two sets: a training set, which contained 20 definitive had been prehybridized in buffer containing 5Â SSC, 0.1% SDS, and 1% bovine surgery samples from different patients, and a validation set, which serum albumin at 42jC for 45 minutes. After 16 to 24 hours of hybridization in contained 14 initial biopsy samples. Six patients contributed paired initial a humidified chamber, slides were washed in 0.2% SDS plus 1Â SSC at 42jC biopsy and definitive surgery samples (see Table 1). The leave-one-out cross- for 4 minutes, 0.2% SDS plus 0.1Â SSC at room temperature for 4 minutes, and validation (LOOCV) was done to test the robustness of our classifiers using 0.06Â SSC at room temperature for 4 minutes. Slides were then scanned using the training set. To prevent overly optimistic estimation of prediction error, a ScanArray 4.0 scanner (Packard Bioscience, Meriden, CT) using both Cy3 honest assessment of the classification error was carried out by repeating and Cy5 channels and quantified using the ScanArray Express software. the feature selection for each step of the cross-validation (17, 18, 20). After The cDNA microarrays used in this study were fabricated with the the training and LOOCV analysis, we picked the best classifier, the SVM Easy-to-Spot Human UniGene 1 PCR products (Incyte Genomics, Palo classifier, for validation. A two-step validation procedure was carried out to Alto, CA) that consist of a total of 9,216 unique elements. The PCR test the ability of the SVM classifier to predict chemoresistance at the time products were printed in duplicate along with three Arabidopsis cDNA for of diagnosis. This was first done using six initial biopsy samples that have detecting spike-in controls using an OmniGrid Accent Arrayer (GeneMa- corresponding definitive surgery samples in the original training set (paired chines, San Carlos, CA) equipped with 24 SMP3 pins (TeleChem samples). The final validation was done by using the SVM classifier to International, Sunnyvale, CA). All PCR products were resequenced before predict the response of eight independent initial biopsy samples that printing to validate their annotation. neither have been seen by the SVM classifier nor have their corresponding Quantitative real-time PCR. Total RNA samples were pretreated with definitive surgery samples been used in the training set (independent DNase for real-time PCR analysis. Synthesis of cDNA was carried out using samples). Gene ontology analysis of the predictor genes was done using

Cancer Res 2005; 65: (18). September 15, 2005 8144 www.aacrjournals.org

FatiGO (http://fatigo.bioinfo.cnio.es; ref. 21). Gene symbols of the robustness of each classifier in the training set (Table 2). Instead of 45 predictor genes were used to create the gene ontology plot at level 4 using resubstitution estimation, which underestimates the classi- of biological process. Kaplan-Meier analysis of the prognostic significance of fication error, we used a honest estimation method (external each gene was computed using the initial biopsy samples of the LOOCV) for the classification error (see Materials and Methods). nonmetastatic cases (n = 12). Overall survival was compared between the The correct classification rates of LOOCV using these algorithms high expression group (n = 6) and the low expression group (n = 6) using the median expression of the gene as cutoff. The analysis was done using were 65% to 70% (Table 2). Among the six algorithms, SVM had one SPSS and the significance was calculated using log-rank test. of the best performances (70% correct classification). Three good responders and three poor responders were misclassified using SVM. Two of the good responders (300 and 394) that were Results and Discussion misclassified as poor responders by the SVM classifier developed Classifying good and poor responders based on definitive recurrent disease 11 and 9 months, respectively, after completion of surgery specimens. With the goal to identify molecular signatures therapy. This suggests that there may be some residual resistant that can predict chemoresistance of osteosarcoma, we first cells in the definitive surgery specimens of these two cases that attempted to establish the expression profiles of resistant versus were recognized by the algorithm based on the predictor gene set. sensitive osteosarcoma cells. We hypothesized that the definitive One of the poor responders (680) that was misclassified as a good surgery samples from the poor responders should be enriched for responders by the multigene predictor remains free of disease after resistant tumor cells. Using expression profiles from definitive 30 months of follow-up. The other misclassified poor responder surgery samples would therefore enhance the sensitivity and power (761) had 86% necrosis, which is very close to our cutoff for good to detect the difference between chemosensitive and resistant cell response. populations compared with using initial biopsy samples in which SVM is a powerful pattern recognition technique with the ability resistant cells may only be present as a small fraction. Therefore, to build robust predictive models when the number of training we predicted that using definitive surgery samples would increase samples is small but the dimensionality is high. SVM has many the chance of identifying a molecular signature of chemoresistance appealing properties for classification of microarray data, including and that this signature could be used to identify the good and poor measures to prevent overfitting and local minima that are responders in the initial biopsy samples. We further hypothesized associated with other classification algorithms. Recently, it has that the nonspecific perturbations in the expression profiles caused been used successfully to classify microarray data (22–24). by preoperative chemotherapy would affect both good and poor Therefore, although the LOOCV showed only moderate success in responder samples and therefore would not contribute to the classifying the training set samples, we decided to use the SVM determinants of the chemoresistance signature. To test these classifier to classify initial biopsy samples. Furthermore, LOOCV is hypotheses, we examined if we could classify good and poor of low bias but high variance and tends to select for more responders using definitive surgery specimens only. We divided the conservative models (18). Applying the SVM classifier to an definitive surgery samples from 20 patients into two groups: good independent test set would be a better test of its accuracy. (n = 7) and poor (n = 13) responders. We first identified a set of 45 Use of multigene classifier to predict response to preoper- predictor genes that could discriminate the two classes (good and ative chemotherapy in initial biopsy. To test the SVM classifier, poor responders) in the definitive surgery samples using two- we divided our 14 initial biopsy samples into two groups. The first sample t test with a significant cutoff (P = 0.005). Figure 2 shows group consisted of six samples, which had corresponding definitive the relative expressions of these 45 predictor genes in good and surgery samples included in the training set (paired samples). poor responders. Most of these genes (91%) were overexpressed in Using these six cases, we attempted to verify that our classifier built poor responder specimens. from definitive surgery samples could predict the chemoresistance Various supervised classification algorithms (CCP, k-nearest of the corresponding initial biopsy samples based on the neighbor, nearest centroid, SVM, and LDA) were then applied to hypothesis that the molecular signature of chemoresistance as the training set to test if they could classify good and poor recognized in definitive surgery samples was already present in the responders using P of 0.005. LOOCV method was used to test the initial biopsy at the time of diagnosis. The second group consisted

Figure 2. The 45 predictor genes in the chemoresistance signature were selected based on two-sample t test to distinguish between good and poor responders in 20 definitive surgery samples (P = 0.005). Forty-one genes were overexpressed in poor responders, whereas only 4 genes were overexpressed in good responders. Two genes do not have symbols yet in the public databases, and their accession numbers are used instead. Bottom, the color scale represents log2 expression ratios of the genes.

www.aacrjournals.org 8145 Cancer Res 2005; 65: (18). September 15, 2005

Table 2. LOOCV of 20 definitive surgery osteosarcoma samples

Tumor ID Histologic Concordance of classification with histologic response response CCP LDA 1-NN 3-NN NC SVM

300 GR No No No No No No 308 GR Yes Yes Yes Yes Yes Yes 386 GR No No No No No No 394 GR No No No No No No 452 GR Yes Yes Yes Yes Yes Yes 759 GR Yes Yes Yes Yes Yes Yes 771 GR Yes Yes Yes Yes Yes Yes 221 PR Yes Yes Yes Yes Yes Yes 236 PR Yes Yes Yes Yes Yes Yes 241 PR Yes Yes Yes Yes Yes Yes 252 PR Yes Yes Yes Yes Yes Yes 311 PR Yes Yes Yes Yes Yes Yes 342 PR Yes Yes Yes Yes Yes Yes 392 PR Yes Yes Yes Yes Yes Yes 483 PR Yes Yes Yes Yes Yes Yes 591 PR Yes Yes Yes Yes Yes Yes 680 PR No No No No No No 691 PR No No No No No No 760 PR No Yes Yes Yes No Yes 761 PR No No No No No No % Correctly classified 65 70 70 70 65 70

NOTE: LOOCV was carried out with feature selection at each validation to minimize the overoptimistic estimation of error rate. The six classification algorithms used are CCP, LDA, 1-nearest neighbor (1-NN), 3-nearest neighbor (3-NN), nearest centroid (NC), and SVM. Yes denotes the classification by the algorithm was correct; No denotes the classification was wrong.

of eight initial biopsy samples that did not have matched definitive resistant cells in the definitive surgery samples was already present surgery samples included in the training set, thus representing a in the initial biopsy samples at the time of diagnosis. Our result is totally independent set of samples that had not been used in consistent with the notion proposed by Ramaswamy et al. (25) that building the classifier. the metastatic signature of metastatic tumors is already present in The SVM classifier misclassified one sample (of six) in the first the primary tumor. The high accuracy of our multigene classifier to group of paired samples, with a correct classification rate of 83% identify good and poor responders from two separate groups of (confidence interval, 36%, 100%; Table 3; see Supplementary initial biopsy samples suggests that response to chemotherapy can Table S1 for details of prediction by all algorithms). The only potentially be predicted at the time of diagnosis. This can misclassified sample was from a patient (410) who was classified as significantly affect the design of future therapeutic studies of a good responder based on histologic response but was predicted to osteosarcoma, in which intensified therapy could be given at the be a poor responder by the multigene classifier. Interestingly, this time of diagnosis to those patients who are predicted to be poor patient initially presented with localized disease but eventually responders to standard therapy to improve their outcome. developed recurrent disease in the lungs 25 months after Predictor genes. Gene ontology analysis of the predictor genes completion of therapy, suggesting that there were resistant cells suggest that most of them can be grouped into cell growth and/or present in the initial biopsy that were recognized by the multigene maintenance (38.24% of total predictor genes that have gene classifier; presumably, these resistant cells metastasized to the lungs ontology annotation at level 4 of biological processes); nucleobase, before definitive surgery and subsequently gave rise to the recurrent nucleoside, nucleotide, and nucleic acid metabolism (32.25%), tumor. Ironically, the multigene predictor classified this patient’s macromolecule metabolism (29.41%); response to stress (14.71%); definitive surgery sample (452) as good responder (Table 2), regulation of metabolism (14.71%); and signal transduction implying that either the definitive surgery sample used in our (14.71%; Supplementary Fig. S1). In particular, ubiquitin-mediated analysis was not representative of the primary tumor in that it did proteolysis and cell cycle regulation are two major pathways that not contain the resistant cells or the resistant cells had already some of the predictor genes are grouped under. More detailed metastasized before definitive surgery and therefore were no longer analysis revealed that many of the predictor genes have interesting detectable in the primary tumor. properties that are related to bone development, cancer biology, In the second group of independent initial biopsy samples, the and drug resistance (see Table 4 for the detailed information of the classifier correctly predicted eight of eight of the samples (100% 45 predictor genes). For instance, TWIST1, which encodes a helix- correct; confidence interval, 63%, 100%). These eight samples loop-helix transcriptional factor, has been implicated in Saethre- included five poor responders and three good responders. These Chotzen syndrome, radial aplasia, Robinow-Sorauf syndrome, results further indicate that the gene expression signature of the and craniosynostosis (26–29). Mice with heterozygous Twist1

Cancer Res 2005; 65: (18). September 15, 2005 8146 www.aacrjournals.org

Table 3. Classification of initial biopsy samples (paired Six of the predictor genes (AMPD2, CDC20, NDUFS5, SDHB, and independent) using SVM classifier (see Table 2 CDC2L2, and DDAH1) are located on chromosome 1p, which is a legend) recurrent region of chromosomal gain in osteosarcoma as reported by our group and others (44–46). The occurrence frequency of Tumor ID Histologic Concordance with histologic response chromosome 1p in the predictor list (14%) is almost thrice higher response than the expected frequency of 1p in the 3,018 filtered genes (5%). Paired Independent The relationship between DNA copy number changes and gene expression levels as well as their correlation with clinical outcome 410 GR No is currently under investigation. 197 PR Yes Quantitative reverse transcription-PCR validation. To 207 PR Yes validate the microarray results, transcript levels of seven predictor 278 GR Yes genes were measured by quantitative reverse transcription-PCR 289 PR Yes using unamplified RNA from seven samples that had sufficient 345 GR Yes 204 PR Yes quantities of RNA. These seven samples were made up of both initial 274 PR Yes biopsy and definitive surgery samples from good and poor 299 GR Yes responders (Table 5). The seven predictor genes were selected 464 PR Yes randomly and exhibited various degrees of statistical significance. 479 PR Yes All seven genes tested showed positive correlation between micro- 481 PR Yes array and quantitative reverse transcription-PCR results. Two of the 545 GR Yes genes, PDCD5 and TWIST1, also showed significant difference using 654 GR Yes Pearson correlation (P < 0.05). Although validation results of in- house cDNA microarrays usually have less correlation with NOTE: Only one sample (tumor ID 410) in the paired group was quantitative reverse transcription-PCR when compared with other misclassified, and the patient subsequently developed recurrent commercial microarray platforms (47), our validation results are disease (see text for details). comparable with or better than other published studies (48, 49). Predictor genes and overall survival. Because histologic response to preoperative chemotherapy is only a surrogate marker of survival, one way to cross-validate the chemoresistance mutation showed defects in craniofacial and limb development signature is to test if the predictor genes in the resistance signature (30), which resemble those found in Saethre-Chotzen syndrome. are also associated with the long-term outcome of osteosarcoma Although the role of TWIST1 in bone development is still under patients. Ultimately, we would like to directly identify prognostic study, it has been reported that TWIST1 affects CBFA1/RUNX2 markers that can predict long-term outcome of osteosarcoma expression (31) and DNA-binding ability of RUNX2 (32), an patients. We did Kaplan-Meier analysis on all 45 predictor genes important regulator of osteoblast differentiation and proliferation individually using only the initial biopsy samples from the (33). Thus, dysregulation of TWIST1 expression may play a role in nonmetastatic cases in our prediction set. Despite the small the pathogenesis of osteosarcoma. In addition, several studies sample size (n = 12), three of the genes (HSPA4, PIGF, and LMO6) indicated that TWIST1 functions as a potential oncogene, were found to be statistically significant in the prediction of cooperating with MYC and MYCN in suppressing p53-dependent survival (log-rank test, P < 0.05) and one gene (TREM2) showed a apoptosis pathways (34, 35). Recent findings also showed that strong correlation (P = 0.0784; data not shown). Interestingly, the TWIST1 was involved in Taxol resistance and metastasis, further expression pattern of the first three genes confirms the prediction implicating the important role of this gene in chemoresistance that high expression of these genes in poor responders will and tumor invasion in osteosarcoma (36, 37). It is also interesting correlate with poorer prognosis. TREM2 is the only gene that shows to note that, in addition to having antiapoptotic properties, correlation between low expression level and poor prognosis, overexpression of TWIST1 is associated with characteristics of which also matches our predictor model (i.e., lower expression of osteoprogenitor cells that have decreased proliferation and less TREM2 in poor responders). A recent article reported that down- mature phenotype (38). All of these are predicted properties of regulation of TREM2 may affect differentiation of osteoclasts, cancer stem cells (39), suggesting that the chemoresistant cells that leading to impaired bone resorption (50). Additionally, HSPA4 has show overexpression of TWIST1 may arise from cancer stem cells. been found to be overexpressed in human hepatocellular TMPO belongs to a group of ubiquitously expressed nuclear carcinoma and showed antiapoptotic activities (51). We recognize proteins that play an important role in nuclear envelope organiza- the limitations of this preliminary analysis because of small sample tion and cell cycle control. It is up-regulated in medulloblastoma size, relatively short duration of follow-up in some cases, and the when compared with normal cerebellum (40). One form of the problem with multiple testing. Therefore, these results need to be TMPO transcripts, TMPO-b, is highly expressed in human neuro- validated in a much larger sample set with long-term follow-up blastoma cell lines compared with normal nerve tissue (41). UBE2D2 data. encodes a member of the E2 ubiquitin-conjugating enzyme family Ochi et al. have recently reported a similar study of predicting also known as UBCH5B or UBC4. Inhibition of UBE2D2 inhibits chemotherapy response in a few osteosarcoma patients (14). human papillomavirus E6–stimulated p53 degradation (42). MCM2 Although our predictor gene list is different from theirs, most of the is part of the minichromosome maintenance (MCM) complex that is predictor genes in both studies had higher expression in tumors of involved in initiation of eukaryotic genome replication. It regulates poor responders than those of good responders. The different age the helicase activity of the MCM complex. High expression of distribution of the patients (more adults in their study), the MCM2 is associated with poor survival in prostate cancer (43). relatively smaller sample size (n = 19 in their study compared with www.aacrjournals.org 8147 Cancer Res 2005; 65: (18). September 15, 2005

Table 4. Detailed information of 45 predictor genes in the chemoresistance prediction model

Description Genbank Gene Map accession no. symbol location

1 Ubiquitin-specific protease 32 AA022783 USP32 17q23.2 2 CHK1 checkpoint homologue XM_006254 CHEK1 11q24-q24 (Schizosaccharomyces pombe) 3 Ras homologue gene family, member Q AI829684 PIGF 2p21 4 Twist homologue 1 (acrocephalosyndactyly 3; X91662 TWIST1 7p21.2 Saethre-Chotzen syndrome; Drosophila) 5 Cell division cycle 2–like 2 (PITSLRE proteins) BE268611 CDC2L2 1p36.3 6 Sec1 family domain containing 1 AF319958 SCFD1 14q12 7 Zinc finger protein 184 (Kruppel-like) BG254958 ZNF184 6p21.3 8 Matrin 3 BG179511 MATR3 5q31.2 9 Thymopoietin U09088 TMPO 12q22 10 LIM domain-only 6 AI017508 LMO6 Xp11.23-p11.22 11 Ligase I, DNA, ATP dependent XM_009118 LIG1 19q13.2-q13.3 12 SWI/SNF–related, matrix-associated, NM_003069 SMARCA1 Xq25 actin-dependent regulator of chromatin, subfamily a, member 1 13 RNA-binding motif protein 25 BG260998 RNPC7 14q24.3 14 CDC-like kinase 3 BE908195 CLK3 15q24 15 Ubiquitin-conjugating enzyme E2D 2 AI553806 UBE2D2 5q31.2 (UBC4/5 homologue, yeast) 16 Hydroxyacyl-CoA dehydrogenase, type II AU154299 HADH2 Xp11.2 17 Adenosine monophosphate deaminase 2 (isoform L) BE778615 AMPD2 1p13.3 18 Interleukin-6 signal transducer NM_002184 IL6ST 5q11 (gp130, oncostatin M receptor) 19 Chromosome 6 open reading frame 68 BG289236 C6orf68 6q22.1 20 Dimethylarginine dimethylaminohydrolase 1 AL078459 DDAH1 1p22 21 Succinate dehydrogenase complex, BE747410 SDHB 1p36.1-p35 subunit B, iron sulfur (Ip) 22 Transmembrane protein 1 XM_009794 TMEM1 21q22.3 23 Programmed cell death 5 AA452724 PDCD5 19q12-q13.1 24 Minichromosome maintenance deficient 2, mitotin BE250461 MCM2 3q21 (Saccharomyces cerevisiae) 26 GDP dissociation inhibitor 1 BG395127 GDI1 Xq28 25 SMC4 structural maintenance BF239180 SMC4L1 3q26.1 of chromosomes 4–like 1 (yeast) 27 Chromosome 10 open reading frame 7 BE792735 C10orf7 10p13 28 CDC20 cell division cycle 20 homologue (S. cerevisiae) BG256659 CDC20 1p34.1 29 PTD008 protein BG178791 PTD008 19p13.13 30 Praja 2, RING-H2 motif containing BF529933 PJA2 5q21.3 31 3-Oxoacid CoA transferase 1 NM_000436 OXCT 5p13.1 32 SNRPN upstream reading frame AL514750 SNRPN 15q11.2 33 Centromere protein A, 17 kDa AL555786 CENPA 2p24-p21 34 Heat shock 70-kDa protein 4 BE742483 HSPA4 5q31.1-q31.2 35 SWI/SNF-related, matrix-associated, BE778256 SMARCE1 17q21.2 actin-dependent regulator of chromatin, subfamily e, member 1 36 NADH dehydrogenase (ubiquinone) Fe-S protein 5, BF791786 NDUFS5 1p34.2-p33 15 kDa (NADH-coenzyme Q reductase) 37 Ubiquitin-conjugating enzyme E2A (RAD6 homologue) AL545489 UBE2A Xq24-q25 38 Fetal Alzheimer antigen NM_004459 FALZ 17q24.3 39 Mature T-cell proliferation 1 AL536910 MTCP1 Xq28 40 RNA-binding motif protein, X-linked BG167309 RBMX Xq26 41 Tubulin tyrosine ligase BE738204 TTL 2q13 42 Transcribed locus, moderately AI086307 AI086307 13 similar to XP_509796.1 similar to CGI-145 protein (Pan troglodytes) 43 Ephrin B2 AI127370 EFNB2 13q33 44 Triggering receptor expressed on myeloid cells 2 AF213457 TREM2 6p21.1 45 Membrane-spanning 4 domains, subfamily A, L35848 MS4A3 11q12 member 3 (hematopoietic cell-specific)

Cancer Res 2005; 65: (18). September 15, 2005 8148 www.aacrjournals.org

Table 5. Validation of array expression data by quantita- lists in these two studies. However, we would like to point out tive real time-PCR that the high prediction accuracy of independent samples in our study (eight of eight) suggests that the strategy of using definitive Gene Pearson correlation Spearman rank correlation surgery samples for training is at least comparable with, if not better than, using initial biopsy samples in the study by Ochi et al. rP r P This further justifies the use of definitive surgery samples in our training set. TWIST1 0.777 0.04 0.679 0.094 In summary, we have developed a multigene classifier that could PDCD5 0.776 0.04 0.75 0.052 predict at the time of diagnosis the response of osteosarcoma to OXCT 0.743 0.056 0.571 0.18 preoperative chemotherapy. This chemoresistance signature can TMPO 0.64 0.121 0.643 0.119 potentially be used to stratify patients to different preoperative UBE2A 0.441 0.322 0.107 0.819 chemotherapy regimens in future clinical trials, which could EFNB2 0.346 0.568 0.1 0.873 ultimately affect long-term survival. We also identified a set of AMPD2 0.308 0.501 0.393 0.383 predictor genes that could be used to predict clinical outcome. The encouraging results of this study warrant validation of the NOTE: Each validation is composed of seven samples, which consists predictor gene set with a larger number of patients and eventually of a mixture of initial biopsy and definitive surgery samples. Four were in a prospective trial to evaluate its utility in risk stratification. good responders and three were poor responders. r, correlation coefficient; P, significance level. Acknowledgments Received 3/23/2005; revised 6/5/2005; accepted 7/6/2005. Grant support: NIH grant CA88126, Gillson Longenbaugh Foundation, and Cancer Fighters of Houston, Inc. (C.C. Lau). 28 in our study), use of different chemotherapeutic agents, their use The costs of publication of this article were defrayed in part by the payment of page of two rounds of RNA amplification, variations in the arrays, and charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact. application of different training and classification strategies and We thank Richard Simon, Lisa Wang, Alison Bertuch, Paul Meltzer, and Javed Khan algorithms make it difficult to compare directly the predictor gene for helpful discussions.

References prognostic factor in patients with osteosarcoma. Cancer samples using microarray expression data. Bioinfor- Res 2003;63:1759–63. matics 2000;16:906–14. 1. Link MP, Gebhardt MC, Meyers PA. In: Pizzo P, Poplack 12. Liu ET. Classification of cancers by expression 25. Ramaswamy S, Ross KN, Lander ES, Golub TR. A D, editors. Principles and practice of pediatric oncology. profiling. Curr Opin Genet Dev 2003;13:97–103. molecular signature of metastasis in primary solid 4th ed. Philadelphia: Lippincott-Williams & Wilkins; 13. Ramaswamy S, Golub TR. DNA microarrays in tumors. Nat Genet 2003;33:49–54. 2002. p. 1051–89. clinical oncology. J Clin Oncol 2002;20:1932–41. 26. HowardTD,PaznekasWA,GreenED,etal. 2. Provisor AJ, Ettinger LJ, Nachman JB, et al. Treatment 14. Ochi K, Daigo Y, Katagiri T, et al. Prediction of Mutations in TWIST, a basic helix-loop-helix transcrip- of nonmetastatic osteosarcoma of the extremity with response to neoadjuvant chemotherapy for osteo- tion factor, in Saethre-Chotzen syndrome. Nat Genet preoperative and postoperative chemotherapy: a report sarcoma by gene-expression profiles. Int J Oncol 2004; 1997;15:36–41. from the Children’s Cancer Group. J Clin Oncol 1997;15: 24:647–55. 27. Yousfi M, Lasmoles F, El G, V, Marie PJ. Twist 76–84. 15. Van Gelder RN, von Zastrow ME, Yool A, et al. haploinsufficiency in Saethre-Chotzen syndrome induces 3. Meyers PA, Heller G, Healey J, et al. Chemotherapy for Amplified RNA synthesized from limited quantities of calvarial osteoblast apoptosis due to increased TNFa nonmetastatic osteogenic sarcoma: the Memorial Sloan- heterogeneous cDNA. Proc Natl Acad Sci U S A 1990;87: expression and caspase-2 activation. Hum Mol Genet Kettering experience. J Clin Oncol 1992;10:5–15. 1663–7. 2002;11:359–69. 4. Baldini N, Scotlandi K, Barbanti-Brodano G, et al. 16. Park T, Yi SG, Kang SH, et al. Evaluation of 28. Gripp KW, Stolle CA, Celle L, et al. TWIST gene Expression of P-glycoprotein in high-grade osteosar- normalization methods for microarray data. BMC mutation in a patient with radial aplasia and craniosy- comas in relation to clinical outcome. N Engl J Med Bioinformatics 2003;4:33. nostosis: further evidence for heterogeneity of Baller- 1995;333:1380–5. 17. Radmacher MD, McShane LM, Simon R. A paradigm Gerold syndrome. Am J Med Genet 1999;82:170–6. 5. BacciG,LonghiA,FerrariS,etal.Prognostic for class prediction using gene expression profiles. 29. Kunz J, Hudler M, Fritz B. Identification of a significance of serum alkaline phosphatase in osteosar- J Comput Biol 2002;9:505–11. frameshift mutation in the gene TWIST in a family coma of the extremity treated with neoadjuvant chemo- 18. Dudoit S, Friedman N. Classification in microarray affected with Robinow-Sorauf syndrome. J Med Genet therapy: recent experience at Rizzoli Institute. Oncol Rep experiments. In: Speed T, editor. Statistical analysis of 1999;36:650–2. 2002;9:171–5. gene expression microarray data. Boca Raton: Chapman 30. Bourgeois P, Bolcato-Bellemin AL, Danse JM, et al. 6. Gorlick R, Huvos AG, Heller G, et al. Expression of & Hall/CRC; 2003. p. 93–158. The variable expressivity and incomplete penetrance of HER2/erbB-2 correlates with survival in osteosarcoma. 19. Vapnik V, Chapelle O. Bounds on error expectation the twist-null heterozygous mouse phenotype resemble J Clin Oncol 1999;17:2781–8. for support vector machines. Neural Comput 2000;12: those of human Saethre-Chotzen syndrome. Hum Mol 7. Bielack SS, Kempf-Bielack B, Delling G, et al. 2013–36. Genet 1998;7:945–57. Prognostic factors in high-grade osteosarcoma of the 20. Simon R, Radmacher MD, Dobbin K, McShane LM. 31. Yousfi M, Lasmoles F, Marie PJ. TWIST inactivation extremities or trunk: an analysis of 1,702 patients Pitfalls in the use of DNA microarray data for diagnostic reduces CBFA1/RUNX2 expression and DNA binding to treated on neoadjuvant cooperative osteosarcoma study and prognostic classification. J Natl Cancer Inst 2003;95: the osteocalcin promoter in osteoblasts. Biochem group protocols. J Clin Oncol 2002;20:776–90. 14–8. Biophys Res Commun 2002;297:641–4. 8. Davis AM, Bell RS, Goodwin PJ. Prognostic factors in 21. Al Shahrour F, Diaz-Uriarte R, Dopazo J. FatiGO: a 32. Bialek P, Kern B, Yang X, et al. A twist code osteosarcoma: a critical review. J Clin Oncol 1994;12: web tool for finding significant associations of Gene determines the onset of osteoblast differentiation. Dev 423–31. Ontology terms with groups of genes. Bioinformatics Cell 2004;6:423–35. 9. Feugeas O, Guriec N, Babin-Boilletot A, et al. Loss of 2004;20:578–80. 33. Lian JB, Stein GS. Runx2/Cbfa1: a multifunctional heterozygosity of the RB gene is a poor prognostic 22. Zhao Y, Pinilla C, Valmori D, Martin R, Simon R. regulator of bone formation. Curr Pharm Des 2003;9: factor in patients with osteosarcoma. J Clin Oncol 1996; Application of support vector machines for T-cell 2677–85. 14:467–72. epitopes prediction. Bioinformatics 2003;19:1978–84. 34. Maestro R, Dei Tos AP, Hamamori Y, et al. Twist is a 10. Franzius C, Bielack S, Flege S, et al. Prognostic 23. Peng S, Xu Q, Ling XB, et al. Molecular classification potential oncogene that inhibits apoptosis. Genes Dev significance of (18)F-FDG and (99m)Tc-methylene of cancer types from microarray data using the 1999;13:2207–17. diphosphonate uptake in primary osteosarcoma. J Nucl combination of genetic algorithms and support vector 35. Valsesia-Wittmann S, Magdeleine M, Dupasquier S, Med 2002;43:1012–7. machines. FEBS Lett 2003;555:358–62. et al. Oncogenic cooperation between H-Twist and 11. Ulaner GA, Huang HY, Otero J, et al. Absence of a 24. Furey TS, Cristianini N, Duffy N, et al. Support vector N-Myc overrides failsafe programs in cancer cells. telomere maintenance mechanism as a favorable machine classification and validation of cancer tissue Cancer Cell 2004;6:625–30.

www.aacrjournals.org 8149 Cancer Res 2005; 65: (18). September 15, 2005

36. Wang X, Ling MT, Guan XY, et al. Identification of a 42. Rolfe M, Beer-Romero P, Glass S, et al. Reconstitution 19 in osteosarcoma. Cancer Genet Cytogenet 2001; novel function of TWIST, a bHLH protein, in the of p53-ubiquitinylation reactions from purified compo- 130:14–21. development of acquired taxol resistance in human nents: the role of human ubiquitin-conjugating enzyme 47. Jarvinen AK, Hautaniemi S, Edgren H, et al. Are data cancer cells. Oncogene 2004;23:474–82. UBC4 and E6-associated protein (E6AP). Proc Natl Acad from different gene expression microarray platforms 37. Yang J, Mani SA, Donaher JL, et al. Twist, a master Sci U S A 1995;92:3264–8. comparable? Genomics 2004;83:1164–8. regulator of morphogenesis, plays an essential role in 43. Meng MV, Grossfeld GD, Williams GH, et al. 48. Chang JC, Wooten EC, Tsimelzon A, et al. Gene tumor metastasis. Cell 2004;117:927–39. Minichromosome maintenance protein 2 expression expression profiling for the prediction of therapeutic 38. Lee M-S, Lowe GN, Strong DD, et al. TWIST, a basic in prostate: characterization and association with out- response to docetaxel in patients with breast cancer. helix-loop-helix transcription factor, can regulate the come after therapy for cancer. Clin Cancer Res 2001;7: Lancet 2003;362:362–9. human osteogenic lineage. J Cell Biochem 1999;75:566–77. 2712–8. 49. Kihara C, Tsunoda T, Tanaka T, et al. Prediction of 39. Dean M, Fojo T, Bates S. Tumor stem cells and drug 44. Lau CC, Harris CP, Lu XY, et al. Frequent amplifica- sensitivity of esophageal tumors to adjuvant chemoresistance. Nat Rev Cancer 2005;5:275–84. tion and rearrangement of chromosomal bands 6p12- therapy by cDNA microarray analysis of gene-expression 40. Yokota N, Mainprize TG, Taylor MD, et al. p21 and 17p11.2 in osteosarcoma. Genes Chromosomes profiles. Cancer Res 2001;61:6474–9. Identification of differentially expressed and develop- Cancer 2004;39:11–21. 50. Paloneva J, Mandelin J, Kiialainen A, et al. DAP12/ mentally regulated genes in medulloblastoma using 45. Simons A, Janssen IM, Suijkerbuijk RF, et al. Isolation TREM2 deficiency results in impaired osteoclast differ- suppression subtraction hybridization. Oncogene 2004; of osteosarcoma-associated amplified DNA sequences entiation and osteoporotic features. J Exp Med 2003;198: 23:3444–53. using representational difference analysis. Genes Chro- 669–75. 41. Weber PJ, Eckhard CP, Gonser S, et al. On the role of mosomes Cancer 1997;20:196–200. 51. Gotoh K, Nonoguchi K, Higashitsuji H, et al. Apg-2 thymopoietins in cell proliferation. Immunochemical 46. Zielenska M, Bayani J, Pandita A, et al. Compa- has a chaperone-like activity similar to Hsp110 and is evidence for new members of the human thymopoietin rative genomic hybridization analysis identifies overexpressed in hepatocellular carcinomas. FEBS Lett family. Biol Chem 1999;380:653–60. gains of 1p35 approximately p36 and chromosome 2004;560:19–24.

Cancer Res 2005; 65: (18). September 15, 2005 8150 www.aacrjournals.org

Tsz-Kwong Man, Murali Chintagumpala, Jaya Visvanathan, et al.

Cancer Res 2005;65:8142-8150.

Updated version Access the most recent version of this article at: http://cancerres.aacrjournals.org/content/65/18/8142

Supplementary Access the most recent supplemental material at: Material http://cancerres.aacrjournals.org/content/suppl/2005/09/22/65.18.8142.DC1

Cited articles This article cites 48 articles, 15 of which you can access for free at: http://cancerres.aacrjournals.org/content/65/18/8142.full#ref-list-1

Citing articles This article has been cited by 8 HighWire-hosted articles. Access the articles at: http://cancerres.aacrjournals.org/content/65/18/8142.full#related-urls

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://cancerres.aacrjournals.org/content/65/18/8142. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.