Prediction of Docetaxel Response in Human Breast Cancer by Gene Expression Profiling
http://www.100md.com
《临床肿瘤学》
the Taisho Laboratory of Functional Genomics, Nara Institute of Science and Technology, Nara
Department of Surgical Oncology, Osaka University Medical School, Osaka
NTT Communication Science Laboratories, Kyoto, Japan
ABSTRACT
PATIENTS AND METHODS: A total of 44 breast tumor tissues were sampled by biopsy before treatment with docetaxel, and the response to therapy was clinically evaluated by the degree of reduction in tumor size. Gene expression profiling of the biopsy samples was performed with 2,453 genes using a high-throughput reverse transcriptase polymerase chain reaction technique. Using genes differentially expressed between responders and nonresponders, a diagnostic system based on the weighted-voting algorithm was constructed.
RESULTS: This system predicted the clinical response of 26 previously unanalyzed samples with over 80% accuracy, a level promising for clinical applications. Diagnostic profiles in nonresponders were characterized by elevated expression of genes controlling the cellular redox environment (ie, redox genes, such as thioredoxin, glutathione-S-transferase, and peroxiredoxin). Overexpression of these genes protected cultured mammary tumor cells from docetaxel-induced cell death, suggesting that enhancement of the redox system plays a major role in docetaxel resistance.
CONCLUSION: These results suggest that the clinical response to docetaxel can be predicted by gene expression patterns in biopsy samples. The results also suggest that one of the molecular mechanisms of the resistance is activation of a group of redox genes.
INTRODUCTION
Gene expression profiling, mainly using DNA microarrays, is expected to provide breakthroughs in this area. This approach has already been used to identify genes that could serve as prognostic markers for breast cancer,6,7 and recently, Chang et al8 have made a preliminary report on the application of DNA microarray analysis in the identification of predictive factors of response to docetaxel in breast cancer patients. In this report, we performed gene expression profiling of breast cancer samples to develop a method for prediction of a patients' response to docetaxel.9,10 Here, we measured gene expression using a high-throughput reverse transcriptase polymerase chain reaction (PCR) technique (adapter-tagged competitive PCR [ATAC-PCR]) instead of DNA microarrays. Tumor samples obtained before chemotherapy and available for gene expression analysis are usually limited in volume. Because ATAC-PCR requires a smaller sample volume than DNA microarray analysis, we believe it is a more appropriate method for this application. In this study, we show that a diagnostic system based on the supervised learning theory predicts the response of new patients at over 80% accuracy, providing promise for its future clinical application. In addition, we have found that genes controlling the cellular redox environment are associated with docetaxel resistance.
PATIENTS AND METHODS
Evaluation of Clinical Response
The chemotherapeutic response of primary breast tumors and locally recurrent tumors was clinically evaluated according to the WHO criteria11 as follows: (1) complete response (CR), disappearance of all known disease; (2) partial response (PR), 50% or more decrease in the entire tumor burden; (3) no change, less than 50% decrease or less than 25% increase in the entire tumor burden; and (4) progressive disease, ≥ 25% increase in the entire tumor burden or appearance of new lesions. Entire tumor burden was measured by magnetic resonance imaging before and after neoadjuvant docetaxel treatment in primary breast cancer patients. In cases of locally recurrent breast cancer, whole tumor burden was measured by computed tomography scan and/or caliper measurement before and every 4 to 8 weeks after docetaxel treatment. In this study, patients with CR and PR were defined as responders, and patients with no change and progressive disease were defined as nonresponders. Among a total of 70 patients, the clinical response rate (CR+PR) was 48.6% (34 of 70 patients), and the complete pathologic response rate was 4.5% (two of 44 patients). Pathologic evaluation of a response was performed only in locally advanced breast cancer patients who underwent surgery, and complete pathologic response was defined as a complete absence of tumor cells.
ATAC-PCR Assay and Data Processing
Total RNA was purified from clinical materials using Trizol reagent (Invitrogen, Carlsbad, CA). Selection of the 2,453 genes was based on an expressed-sequence tag-sequencing survey of the genes expressed in breast cancer as previously described.7 RNA purified from 44 separate specimens (22 responders and 22 nonresponders), obtained from February 1999 to December 2000, were subjected to gene-expression profiling by ATAC-PCR.12 ATAC-PCR is an advanced version of quantitative competitive PCR that is characterized by the addition of unique adapters for different cDNAs. The expression levels of these 2,453 genes in breast cancer tissues were measured by ATAC-PCR. A single ATAC-PCR reaction included five cDNA samples and two different amounts of a control cDNA with different adapter tags and measured the relative expression of the samples against the control. Grouping of samples was completely random. The control was a mixture of 78 other primary breast cancer tissues. Relative expression levels were calculated against this control. Details of the experimental procedure are described elsewhere.7 Like DNA microarrays with two fluorescent dyes, ATAC-PCR measures relative expression levels against the fluorescence intensity of a control RNA sample. After excluding 233 genes for which 20% of the data was missing, the raw data describing gene expression levels were divided by the median of each sample. The amount of cDNA template material is not proportional to that of total RNA because each total RNA preparation adjusted to the same amount may contain different fractions of mRNA. Because the median of each sample is expected to reflect the overall mRNA level, normalization by this value corrects for variation in mRNA level from sample to sample. Values less than 0.05 were converted to 0.05 as a minimum value, and the entire data matrix was converted to a logarithmic scale. Expression-profiling data and supplemental information are available at our Web site (http://genome.mc.pref.osaka.jp).
Cancer samples from 26 patients recruited from January 2001 to June 2002 were used for validation of the diagnostic system. Because of the small number of genes in the validation set, their medians may not represent accurately the overall mRNA level. Instead, we used genes whose expression patterns are similar to the median values in the data set of gene expression obtained from the 44 specimens. The similarity measure was the Euclidean distance. Normalization using the expression level of such genes should have the same effect as using the median of a large data set. The following six genes were selected because their expressions patterns best matched the median patterns obtained with the data set described in this section: ribosomal protein L30, ATP synthase alpha subunit, proteasome subunit alpha type 2, RNA polymerase II transcription factor SIII p18 subunit, and nucleophosmin. The relative expression levels of these six genes were measured, in addition to those of the 85 diagnostic genes. The raw data were divided by the average of the six genes. The subsequent logarithmic conversion and setting of the cutoff value were performed as for the large-scale analysis.
Construction of Diagnostic Systems
To establish a diagnostic system with stable PCR amplification, 1,125 genes for which fewer than nine samples exhibited the lower cutoff value were used for the selection of diagnostic genes. The weighted-voting (WV) algorithm was performed as described previously.13 Ranking of the genes for this algorithm was based on signal-to-noise ratio (SNR). In the WV algorithm, each gene belonging to a predictor set is assigned a vote value, and prediction is based on the relative vote value in responders and nonresponders. The vote value of gene a, va, and the weight of gene a, wa, are described by the following two equations:
Here, xa is the expression level of gene a in a test sample. ra, na, sra, and sna, are the mean expression level of gene a in responders in the learning set, the mean expression level in nonresponders, the standard deviation in responders, and the standard deviation in nonresponders, respectively. The SNR used for ranking the genes is the absolute value of wa; diagnostic genes are selected in descending order of SNR.
The total vote for responders, Vr, is the sum of positive votes of all diagnostic genes. The total vote for nonresponders, Vn, is the sum of negative votes. In the case of missing values, the vote of the gene is assigned as 0. Prediction strength (PS) is defined by the following equation in which positive and negative PS indicates prediction of responders and nonresponders, respectively:
The K-nearest neighbor (K-NN) algorithm was also applied as described previously.14 Ranking of genes for K-NN was based on permutation P value (PPV) calculated from 50,000 random permutations. At first, difference between averages of nonresponders and responders is calculated with original data. With each random permutation of labels (responders and nonresponders), difference between averages of nonresponders and responders is calculated. From the number of trials in which the difference exceeded the original, PPV is calculated. K-NN is a simple algorithm that stores all learning samples and classifies a test sample based on the number of learning samples of each class among K samples most similar to the test sample. In our case, K = 1, and the Euclidean distance was used as the similarity measure.
Support vector machine (SVM) and recursive feature elimination (RFE) were performed as previously described.15 SVM is a classification algorithm that defines a discrimination surface in the utilized-feature (gene) space that attempts to maximally separate classes of training data. A test sample's position relative to the discrimination surface determines its class. RFE is an iterative procedure that eliminates features (genes) with the smallest ranking criterion determined by a cost function and was used here to rank genes for SVM.
Except for SVM, the performance of the algorithms was evaluated by leave-one-out cross validation. SVM performance was evaluated using 100 random trials of leave-two-out cross validation; leave-one-out was not successful because of selection bias caused by the small number of samples. The leave-one-out cross validation was considered complete; in each turn of validation, leaving out one sample, a defined number of genes were selected by one of the ranking methods. Subsequently, a prediction algorithm was constructed, and prediction was judged using the left-out sample. Forty- four iterations of this validation process were used to calculate the accuracy of the algorithm. The data used for gene selection were the normalized data of the 1,125 genes described earlier. Validation of the diagnostic system was performed using the WV algorithm constructed with the data of the top 85 genes ranked by SNR using the normalized data of 1,125 genes in 44 patients.
Other Statistical Analyses
Hierarchical cluster analysis for the presentation of diagnostic genes was performed by the Ward method using GeneMaths (Applied Maths, Sint-Martens-Latem, Belgium). Results of test samples were evaluated by applying a Bayesian inference scheme on a Bernoulli trial. Ninety-five percent CIs were inferred by computing the inverse of the appropriate cumulative beta distribution. Details are described at the following Web site: http://www.causascientia.org/math_stat/ProportionCI.html.
Evaluation by false discovery rate (FDR)16 was performed as follows. At first, we set a desired limit q on the FDR. Then, P values were arranged in ascending order, P(1) ≤ P(2) ≤...P(i) ≤...≤ P(v), where v is the total number of genes. We let r be the largest i such that P(i) ≤ i/v x q and rejected all hypotheses corresponding to P(1), ..., P(r).
The chance of association to a functional group was calculated using the hypergeometric distribution, described as follows.
Here, N, M, n, and m are the number of genes in the total population, the number of genes within the functional group in the total population, the number of selected genes, and the number of genes within the functional group among the selected genes, respectively. In our case, N = 1,125, M = 20, n = 85, and m = 5. Annotations of genes are supplied with gene expression data described earlier. Identification of redox genes was based on annotation by the RefSeq database (National Center for Biotechnology Information, Bethesda, MD).
Cell Culture and Transfection
MCF-7 breast cancer cells were grown in DMEM (Sigma, St Louis, MO) supplemented with 10% fetal bovine serum (Dainippon Pharm, Osaka, Japan) and 50 U/mL antimycotic-antimytotic (Invitrogen) at 37°C in a 5% CO2 humidified incubator and were split twice weekly. The full-length cDNAs were cloned into the gateway-adapted expression vector, pcDNA-DEST47, which is a gateway vector for cloning and expression of green fluorescent protein fusion proteins in mammalian cells. The control plasmid contained a five amino-acid peptide of random sequence as the insert. Transfection of MCF-7 cells was performed at 70% confluency using LipofectAMINE Plus Reagent (Invitrogen), following the manufacturer's protocols.
Terminal Deoxynucleotidyl Transferase-Mediated Deoxyuridine Triphosphate-Biotin Nick-End Labeling Assay
Transfected cells were grown on 24-well culture dishes and treated with different concentrations of docetaxel. After 24 hours, cells were fixed with 4% paraformaldehyde in phosphate-buffered saline and stained with TMR-red, following the manufacturer's protocols (In Situ Cell Death Detection Kit; Roche Molecular Biochemicals, Mannheim, Germany). The experiments were carried out in duplicate.
RESULTS
We compared three popular class-prediction algorithms, WV,13 K-NN,14 and SVM.15 Because a small number of genes, fewer than 100, were desired for diagnostic application, we selected genes using popular ranking methods. We used SNR to rank genes for WV and RFE to rank genes for SVM because these ranking methods are closely associated with each algorithm. PPV is a more general ranking method but was used only for K-NN in this study and not for the other algorithms. Gene selection was performed from a pool of 1,125 genes exhibiting stable PCR amplification (see Patients and Methods). The performance of each algorithm was evaluated with up to 100 genes by leave-one-out cross validation or a comparable method. The accuracy curves of these prediction methods with a variable number of up to 100 genes are shown in Figure 1. Our criterion for selection of the diagnostic system was the greatest accuracy with the least number of diagnostic genes. The results indicated that the WV algorithm generally performed best, and, with 85 or 95 genes, its accuracy was highest (72.7%; Fig 1). Therefore, we chose to use the WV algorithm with 85 diagnostic genes.
The 85 diagnostic genes were selected using data from all 44 patients in the descending order determined by SNR. Classification of the 44 patients by the final gene set is summarized in Figure 2. Sixty-one genes were elevated in nonresponders, and 24 genes were elevated in responders (Table 2). Patients were sorted vertically in order of the strength of their respective predictions. Setting the boundary at 0 clearly separated responders from nonresponders, with only four misclassifications (Fig 2). It should be noted that complete responders and patients with progressive disease were not located near the boundary, suggesting that their expression patterns clearly allowed them to be classified as responders or nonresponders.
The WV algorithm constructed with the 85 genes was tested with 26 new patients. We measured the expression levels of the 85 genes by ATAC-PCR and predicted the degree of clinical response to docetaxel. The prediction accuracy was 80.7% (95% CI, 63.5% to 92.5%; Table 3). The sensitivity, specificity, positive predictive value, and negative predictive value were 91.7% (95% CI, 68.1% to 99.5%), 71.4% (95% CI, 46.7% to 89.5%), 73.3% (95% CI, 49.5% to 90.3%), and 90.9% (95% CI, 65.9% to 99.4%), respectively.
In a small validation set, random chance could produce a significant number of correct predictions. However, the probability of observing 21 or more correct guesses is 0.12%, as calculated by the binomial distribution.
Relationship Between Response to Docetaxel and Clinicopathologic Parameters
The relationship between the response to docetaxel and tumor size, histologic grade, and estrogen receptor (ER) status was studied in locally advanced breast cancer patients (n = 44 of 70). Response rates of tumors measuring 5 cm or less and more than 5 cm were 44.5% (10 of 22 patients) and 54.5% (12 of 22 patients), respectively. Response rates of histologic grades 1 and 2 and grade 3 tumors were 46.7% (seven of 15 patients) and 51.7% (15 of 29 patients), respectively. Response rates of patients with ER-positive and ER-negative tumors were 47.4% (nine of 19 patients) and 52.0% (13 of 25 patients), respectively. None of these clinicopathologic parameters was significantly associated with tumor response.
Molecular Features Characterizing the Docetaxel Response
To identify molecular features that underlie docetaxel sensitivity, we examined the functions of the genes that were differentially expressed between responders and nonresponders. Considering the multiplicity of gene selection, we evaluated differentially expressed genes by FDR.16 The top 71 genes ranked in the ascending order of their permutation P values were selected when the FDR was set at 0.5. All of these genes were included among the 85 predictor genes and are marked in Table 2. It should be noted that 85 genes are ranked by the descending order of SNR, and the ranking is similar but not the same as that by PPV. Expression of tubulin was elevated in nonresponders, which may contribute to resistance to docetaxel, an antimicrotubule agent.18,19 However, the most prominent characteristic in nonresponders is elevated expression of genes controlling the cellular redox environment, that is, those of the glutathione and thioredoxin systems. These genes included glutathione S-transferase pi 1, glutathione peroxidase 1, thioredoxin, and peroxiredoxin 1 (thioredoxine peroxidase 2); in addition, glutathione peroxidase 4 was among the 85 predictor genes. Although more than half of the genes on the list are false-positives, the probability that five redox genes would appear together in the 85 selected genes is low (P = .013). We concluded, then, that these redox genes may be involved in docetaxel resistance, and we confirmed their function further by transfection experiments.
Protection of MCF-7 From Docetaxel-Induced Cell Death by Redox Genes
A docetaxel dose-response curve was constructed on a human mammary cell line, MCF-7, that is normally docetaxel sensitive20 but that was transfected with various redox-associated genes (Fig 3). Three genes (glutathione-S-tranferase pi, thioredoxin, and peroxiredoxin 1) were cloned as fusion proteins with green fluorescent protein under the control of a constitutive promoter. Cell death was assayed by the terminal deoxynucleotidyl transferase-mediated deoxyuridine triphosphate-biotin nick-end labeling assay, which detects chromosomal DNA fragmentation. All genes protected MCF-7 from docetaxel-induced cell death; for example, at a 10-nmol/L dose of docetaxel, only 60% of control cells survived, whereas the three cell lines overexpressing redox genes exhibited greater than 90% cell survival.
DISCUSSION
Some may argue that none of the genes selected for diagnosis in the previous study8 overlap with those in the present study except for calreticulin, glutathione peroxidase 4, and ATP synthase. However, this overlapping is unlikely to be obtained by chance; its probability is 0.02 assuming that the 2,453 genes of our study were included among the genes on the Affymetrix GeneChip (Santa Clara, CA). In addition, several factors make such a direct comparison difficult. The Affymetrix GeneChip has a wider coverage of genes, but recent analysis revealed that the quantitative data were obtained only with abundant genes, even in budding yeast, in which the mRNA population is much less complex than that of human tissues.22 Therefore, genes with low expression were excluded from the analysis.8 ATAC-PCR can be used to measure expression of rare genes, but its gene coverage is not as large as that of the GeneChip. Consequently, the gene populations that the two studies examined are not likely to be the same. In addition, the criteria for discriminating responders and nonresponders are different. Therefore, it is not surprising that only a few diagnostic genes overlap in the two studies.
Both the clinical and pathologic response rates to docetaxel observed in the present study (48.6%, 34 of 70 patients; and 2.5%, two of 44 patients, respectively) are lower than those reported for other neoadjuvant chemotherapies consisting of both an anthracycline and a taxane, such as doxorubicin and cyclophosphamide, followed by docetaxel (90.7% and 26.1%, respectively), suggesting the superiority of sequential chemotherapy with a taxane and an anthracycline over taxane alone.23 In addition, the fact that the approved dose of docetaxel in Japan (60 mg/m2) is lower than that (100 mg/m2) commonly used in the Western countries seems to explain, at least in part, the lower response rates in the present study.
Our results, both the gene expression patterns observed in patients and the in vitro transfection findings, strongly suggest that redox genes play a major role in docetaxel resistance. The glutathione and thioredoxin systems represent two major mechanisms for maintaining the intracellular redox environment through the reduction and oxidation of thiol groups.24-26 These genes have been found to be involved in resistance to other anticancer agents; inhibition of glutathione S-transferase pi activity by antisense cDNA increased sensitivity to anticancer agents in a colorectal cancer cell line,27 and elevated expression of glutathione S-transferase pi was correlated with resistance to cisplatin in lung cancer and head and neck cancer.28,29 Furthermore, increased thioredoxin levels have been reported to reduce the sensitivity of anticancer drugs in various cases.30-32 Our findings with docetaxel are consistent with these observations and are likely to apply generally to malignancies other than breast cancer. Thus far, the glutathione and thioredoxin systems have been studied individually, but both systems seem to work cooperatively in docetaxel resistance.
Because expression of redox genes is low in responders to docetaxel, administration of chemicals inhibiting redox enzymes to nonresponders could improve their response profile. This is a new concept in anticancer drug design, and the combination of these agents with the diagnostic system of this study may substantially advance current docetaxel-based chemotherapy regimens. Several inhibitors of thioredoxin have already been identified, and one of them, PX-12,33 has entered phase I clinical trials as an anticancer agent. These inhibitors are promising candidates for use in combination therapy. This speculation indicates that gene expression profiling of a well-designed clinical trial could be an important resource for pharmaceutical developments.
Authors' Disclosures of Potential Conflicts of Interest
Acknowledgment
We thank Y. Kurokawa, MD, S. Oba, PhD, and S. Ishii, PhD, for their helpful suggestions in the analysis of the data. We also thank J. Yodoi, MD, PhD, for his critique of the manuscript, H. Kita-Matsuo, PhD, for technical advice, and S. Maki-Kinjo, K. Miyaoka-Ikenaga, and Y. Ishida for their valuable technical assistance.
NOTES
Supported by a Grant-in-Aid for the Development of Innovative Technology from the Ministry of Education, Culture, Sports, Science and Technology of Japan, and by a grant from Taisho Pharmaceuticals Co, Ltd.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
1. Ringel I, Horwitz SB: Studies with RP56976 (Taxotere): A semisynthetic analogue of taxol. J Natl Cancer Inst 83:288-291, 1991
2. Goble S, Bear HD: Emerging role of taxanes in adjuvant and neoadjuvant therapy for breast cancer: The potential and the questions. Surg Clin North Am 83:943-971, 2003
3. Egawa C, Miyoshi Y, Takamura Y, et al: Decreased expression of BRCA2 mRNA predicts favorable response to docetaxel in breast cancer. Int J Cancer 95:255-259, 2001
4. Hasegawa S, Miyoshi Y, Egawa C, et al: Mutational analysis of the class I beta-tubulin gene in human breast cancer. Int J Cancer 101:46-51, 2002
5. Burkhart CA, Kavallaris M, Horwitz SB: The role of beta-tubulin isotypes in resistance to antimitotic drugs. Biochim Biophys Acta 1471:O1-O9, 2001
6. van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536, 2002
7. Iwao K, Matoba R, Ueno N, et al: Molecular classification of primary breast tumors possessing distinct prognostic properties. Hum Mol Genet 11:199-206, 2002
8. Chang JC, Wooten EC, Tsimelzon A, et al: Gene expression profiling for the prediction of therapeutic response to docetaxel in patients with breast cancer. Lancet 362:362-369, 2003
9. ten Bokkel Huinink WW, Prove AM, Piccart M, et al: A phase II trial with docetaxel (Taxotere) in second line treatment with chemotherapy for advanced breast cancer: A study of the EORTC Early Clinical Trials Group. Ann Oncol 5:527-532, 1994
10. Heys SD, Hutcheon AW, Sarkar TK, et al: Neoadjuvant docetaxel in breast cancer: 3-year survival results from the Aberdeen trial. Clin Breast Cancer 3:S69-S74, 2002 (suppl 2)
11. Miller AB, Hoogstraten B, Staquet M, et al: Reporting results of cancer treatment. Cancer 47:207-214, 1981
12. Kato K: Adaptor-tagged competitive PCR: A novel method for measuring relative gene expression. Nucleic Acids Res 25:4694-4696, 1997
13. Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring. Science 286:531-537, 1999
14. Pomeroy SL, Tamayo P, Gaasenbeek M, et al: Prediction of central nervous system embryonal tumour outcome based on gene expression. Nature 415:436-442, 2002
15. Furey TS, Cristianini N, Duffy N, et al: Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics 16:906-914, 2000
16. Benjamini Y, Hochberg Y: Controlling the false discovery rate: A practical and powerful approach to multiple testing. J R Stat Soc B 85:289-300, 1995
17. Muro S, Takemasa I, Oba S, et al: Identification of expressed genes linked to malignancy of human colorectal carcinoma by parametric clustering of quantitative expression data. Genome Biol 4:R21, 2003
18. Schiff PB, Horwitz SB: Taxol stabilizes microtubules in mouse fibroblast cells. Proc Natl Acad Sci U S A 77:1561-1565, 1980
19. Stein CA: Mechanisms of action of taxanes in prostate cancer. Semin Oncol 26:3-7, 1999
20. Dan S, Tsunoda T, Kitahara O, et al: An integrated database of chemosensitivity to 55 anticancer drugs and gene expression profiles of 39 human cancer cell lines. Cancer Res 62:1139-1147, 2002
21. Gordon GJ, Jensen RV, Hsiao LL, et al: Using gene expression ratios to predict outcome among patients with mesothelioma. J Natl Cancer Inst 95:598-605, 2003
22. Holland MJ: Transcript abundance in yeast varies over six orders of magnitude. J Biol Chem 277:14363-14366, 2002
23. Bear HD, Anderson S, Brown A, et al: The effect on tumor response of adding sequential preoperative docetaxel to preoperative doxorubicin and cyclophosphamide: Preliminary results from national surgical adjuvant breast and bowel project protocol B-27. J Clin Oncol 21:4165-4174, 2003
24. Filomeni G, Rotilio G, Ciriolo MR: Cell signaling and the glutathione redox system. Biochem Pharmacol 64:1057-1064, 2002
25. Arner ES, Holmgren A: Physiological functions of thioredoxin and thioredoxin reductase. Eur J Biochem 267:6102-6109, 2000
26. Nakamura H, Nakamura K, Yodoi J: Redox regulation of cellular activation. Annu Rev Immunol 15:351-369, 1997
27. Ban N, Takahashi Y, Takayama T, et al: Transfection of glutathione S-transferase (GST)-pi antisense complementary DNA increases the sensitivity of a colon cancer cell line to Adriamycin, cisplatin, melphalan, and etoposide. Cancer Res 56:3577-3582, 1996
28. Bai F, Nakanishi Y, Kawasaki M, et al: Immunohistochemical expression of glutathione S-transferase-Pi can predict chemotherapy response in patients with nonsmall cell lung carcinoma. Cancer 78:416-421, 1996
29. Shiga H, Heath EI, Rasmussen AA, et al: Prognostic value of p53, glutathione S-transferase pi, and thymidylate synthase for neoadjuvant cisplatin-based chemotherapy in head and neck cancer. Clin Cancer Res 5:4097-4104, 1999
30. Yokomizo A, Ono M, Nanri H, et al: Cellular levels of thioredoxin associated with drug sensitivity to cisplatin, mitomycin C, doxorubicin, and etoposide. Cancer Res 55:4293-4296, 1995
31. Sasada T, Iwata S, Sato N, et al: Redox control of resistance to cis-diamminedichloroplatinum (II) (CDDP): Protective effect of human thioredoxin against CDDP-induced cytotoxicity. J Clin Invest 97:2268-2276, 1996
32. Kahlos K, Soini Y, Saily M, et al: Up-regulation of thioredoxin and thioredoxin reductase in human malignant pleural mesothelioma. Int J Cancer 95:198-204, 2001
33. Oblong JE, Chantler EL, Gallegos A, et al: Reversible inhibition of human thioredoxin reductase activity by cytotoxic alkyl 2-imidazolyl disulfide analogues. Cancer Chemother Pharmacol 34:434-438, 1994(Kyoko Iwao-Koizumi, Ryo M)
Department of Surgical Oncology, Osaka University Medical School, Osaka
NTT Communication Science Laboratories, Kyoto, Japan
ABSTRACT
PATIENTS AND METHODS: A total of 44 breast tumor tissues were sampled by biopsy before treatment with docetaxel, and the response to therapy was clinically evaluated by the degree of reduction in tumor size. Gene expression profiling of the biopsy samples was performed with 2,453 genes using a high-throughput reverse transcriptase polymerase chain reaction technique. Using genes differentially expressed between responders and nonresponders, a diagnostic system based on the weighted-voting algorithm was constructed.
RESULTS: This system predicted the clinical response of 26 previously unanalyzed samples with over 80% accuracy, a level promising for clinical applications. Diagnostic profiles in nonresponders were characterized by elevated expression of genes controlling the cellular redox environment (ie, redox genes, such as thioredoxin, glutathione-S-transferase, and peroxiredoxin). Overexpression of these genes protected cultured mammary tumor cells from docetaxel-induced cell death, suggesting that enhancement of the redox system plays a major role in docetaxel resistance.
CONCLUSION: These results suggest that the clinical response to docetaxel can be predicted by gene expression patterns in biopsy samples. The results also suggest that one of the molecular mechanisms of the resistance is activation of a group of redox genes.
INTRODUCTION
Gene expression profiling, mainly using DNA microarrays, is expected to provide breakthroughs in this area. This approach has already been used to identify genes that could serve as prognostic markers for breast cancer,6,7 and recently, Chang et al8 have made a preliminary report on the application of DNA microarray analysis in the identification of predictive factors of response to docetaxel in breast cancer patients. In this report, we performed gene expression profiling of breast cancer samples to develop a method for prediction of a patients' response to docetaxel.9,10 Here, we measured gene expression using a high-throughput reverse transcriptase polymerase chain reaction (PCR) technique (adapter-tagged competitive PCR [ATAC-PCR]) instead of DNA microarrays. Tumor samples obtained before chemotherapy and available for gene expression analysis are usually limited in volume. Because ATAC-PCR requires a smaller sample volume than DNA microarray analysis, we believe it is a more appropriate method for this application. In this study, we show that a diagnostic system based on the supervised learning theory predicts the response of new patients at over 80% accuracy, providing promise for its future clinical application. In addition, we have found that genes controlling the cellular redox environment are associated with docetaxel resistance.
PATIENTS AND METHODS
Evaluation of Clinical Response
The chemotherapeutic response of primary breast tumors and locally recurrent tumors was clinically evaluated according to the WHO criteria11 as follows: (1) complete response (CR), disappearance of all known disease; (2) partial response (PR), 50% or more decrease in the entire tumor burden; (3) no change, less than 50% decrease or less than 25% increase in the entire tumor burden; and (4) progressive disease, ≥ 25% increase in the entire tumor burden or appearance of new lesions. Entire tumor burden was measured by magnetic resonance imaging before and after neoadjuvant docetaxel treatment in primary breast cancer patients. In cases of locally recurrent breast cancer, whole tumor burden was measured by computed tomography scan and/or caliper measurement before and every 4 to 8 weeks after docetaxel treatment. In this study, patients with CR and PR were defined as responders, and patients with no change and progressive disease were defined as nonresponders. Among a total of 70 patients, the clinical response rate (CR+PR) was 48.6% (34 of 70 patients), and the complete pathologic response rate was 4.5% (two of 44 patients). Pathologic evaluation of a response was performed only in locally advanced breast cancer patients who underwent surgery, and complete pathologic response was defined as a complete absence of tumor cells.
ATAC-PCR Assay and Data Processing
Total RNA was purified from clinical materials using Trizol reagent (Invitrogen, Carlsbad, CA). Selection of the 2,453 genes was based on an expressed-sequence tag-sequencing survey of the genes expressed in breast cancer as previously described.7 RNA purified from 44 separate specimens (22 responders and 22 nonresponders), obtained from February 1999 to December 2000, were subjected to gene-expression profiling by ATAC-PCR.12 ATAC-PCR is an advanced version of quantitative competitive PCR that is characterized by the addition of unique adapters for different cDNAs. The expression levels of these 2,453 genes in breast cancer tissues were measured by ATAC-PCR. A single ATAC-PCR reaction included five cDNA samples and two different amounts of a control cDNA with different adapter tags and measured the relative expression of the samples against the control. Grouping of samples was completely random. The control was a mixture of 78 other primary breast cancer tissues. Relative expression levels were calculated against this control. Details of the experimental procedure are described elsewhere.7 Like DNA microarrays with two fluorescent dyes, ATAC-PCR measures relative expression levels against the fluorescence intensity of a control RNA sample. After excluding 233 genes for which 20% of the data was missing, the raw data describing gene expression levels were divided by the median of each sample. The amount of cDNA template material is not proportional to that of total RNA because each total RNA preparation adjusted to the same amount may contain different fractions of mRNA. Because the median of each sample is expected to reflect the overall mRNA level, normalization by this value corrects for variation in mRNA level from sample to sample. Values less than 0.05 were converted to 0.05 as a minimum value, and the entire data matrix was converted to a logarithmic scale. Expression-profiling data and supplemental information are available at our Web site (http://genome.mc.pref.osaka.jp).
Cancer samples from 26 patients recruited from January 2001 to June 2002 were used for validation of the diagnostic system. Because of the small number of genes in the validation set, their medians may not represent accurately the overall mRNA level. Instead, we used genes whose expression patterns are similar to the median values in the data set of gene expression obtained from the 44 specimens. The similarity measure was the Euclidean distance. Normalization using the expression level of such genes should have the same effect as using the median of a large data set. The following six genes were selected because their expressions patterns best matched the median patterns obtained with the data set described in this section: ribosomal protein L30, ATP synthase alpha subunit, proteasome subunit alpha type 2, RNA polymerase II transcription factor SIII p18 subunit, and nucleophosmin. The relative expression levels of these six genes were measured, in addition to those of the 85 diagnostic genes. The raw data were divided by the average of the six genes. The subsequent logarithmic conversion and setting of the cutoff value were performed as for the large-scale analysis.
Construction of Diagnostic Systems
To establish a diagnostic system with stable PCR amplification, 1,125 genes for which fewer than nine samples exhibited the lower cutoff value were used for the selection of diagnostic genes. The weighted-voting (WV) algorithm was performed as described previously.13 Ranking of the genes for this algorithm was based on signal-to-noise ratio (SNR). In the WV algorithm, each gene belonging to a predictor set is assigned a vote value, and prediction is based on the relative vote value in responders and nonresponders. The vote value of gene a, va, and the weight of gene a, wa, are described by the following two equations:
Here, xa is the expression level of gene a in a test sample. ra, na, sra, and sna, are the mean expression level of gene a in responders in the learning set, the mean expression level in nonresponders, the standard deviation in responders, and the standard deviation in nonresponders, respectively. The SNR used for ranking the genes is the absolute value of wa; diagnostic genes are selected in descending order of SNR.
The total vote for responders, Vr, is the sum of positive votes of all diagnostic genes. The total vote for nonresponders, Vn, is the sum of negative votes. In the case of missing values, the vote of the gene is assigned as 0. Prediction strength (PS) is defined by the following equation in which positive and negative PS indicates prediction of responders and nonresponders, respectively:
The K-nearest neighbor (K-NN) algorithm was also applied as described previously.14 Ranking of genes for K-NN was based on permutation P value (PPV) calculated from 50,000 random permutations. At first, difference between averages of nonresponders and responders is calculated with original data. With each random permutation of labels (responders and nonresponders), difference between averages of nonresponders and responders is calculated. From the number of trials in which the difference exceeded the original, PPV is calculated. K-NN is a simple algorithm that stores all learning samples and classifies a test sample based on the number of learning samples of each class among K samples most similar to the test sample. In our case, K = 1, and the Euclidean distance was used as the similarity measure.
Support vector machine (SVM) and recursive feature elimination (RFE) were performed as previously described.15 SVM is a classification algorithm that defines a discrimination surface in the utilized-feature (gene) space that attempts to maximally separate classes of training data. A test sample's position relative to the discrimination surface determines its class. RFE is an iterative procedure that eliminates features (genes) with the smallest ranking criterion determined by a cost function and was used here to rank genes for SVM.
Except for SVM, the performance of the algorithms was evaluated by leave-one-out cross validation. SVM performance was evaluated using 100 random trials of leave-two-out cross validation; leave-one-out was not successful because of selection bias caused by the small number of samples. The leave-one-out cross validation was considered complete; in each turn of validation, leaving out one sample, a defined number of genes were selected by one of the ranking methods. Subsequently, a prediction algorithm was constructed, and prediction was judged using the left-out sample. Forty- four iterations of this validation process were used to calculate the accuracy of the algorithm. The data used for gene selection were the normalized data of the 1,125 genes described earlier. Validation of the diagnostic system was performed using the WV algorithm constructed with the data of the top 85 genes ranked by SNR using the normalized data of 1,125 genes in 44 patients.
Other Statistical Analyses
Hierarchical cluster analysis for the presentation of diagnostic genes was performed by the Ward method using GeneMaths (Applied Maths, Sint-Martens-Latem, Belgium). Results of test samples were evaluated by applying a Bayesian inference scheme on a Bernoulli trial. Ninety-five percent CIs were inferred by computing the inverse of the appropriate cumulative beta distribution. Details are described at the following Web site: http://www.causascientia.org/math_stat/ProportionCI.html.
Evaluation by false discovery rate (FDR)16 was performed as follows. At first, we set a desired limit q on the FDR. Then, P values were arranged in ascending order, P(1) ≤ P(2) ≤...P(i) ≤...≤ P(v), where v is the total number of genes. We let r be the largest i such that P(i) ≤ i/v x q and rejected all hypotheses corresponding to P(1), ..., P(r).
The chance of association to a functional group was calculated using the hypergeometric distribution, described as follows.
Here, N, M, n, and m are the number of genes in the total population, the number of genes within the functional group in the total population, the number of selected genes, and the number of genes within the functional group among the selected genes, respectively. In our case, N = 1,125, M = 20, n = 85, and m = 5. Annotations of genes are supplied with gene expression data described earlier. Identification of redox genes was based on annotation by the RefSeq database (National Center for Biotechnology Information, Bethesda, MD).
Cell Culture and Transfection
MCF-7 breast cancer cells were grown in DMEM (Sigma, St Louis, MO) supplemented with 10% fetal bovine serum (Dainippon Pharm, Osaka, Japan) and 50 U/mL antimycotic-antimytotic (Invitrogen) at 37°C in a 5% CO2 humidified incubator and were split twice weekly. The full-length cDNAs were cloned into the gateway-adapted expression vector, pcDNA-DEST47, which is a gateway vector for cloning and expression of green fluorescent protein fusion proteins in mammalian cells. The control plasmid contained a five amino-acid peptide of random sequence as the insert. Transfection of MCF-7 cells was performed at 70% confluency using LipofectAMINE Plus Reagent (Invitrogen), following the manufacturer's protocols.
Terminal Deoxynucleotidyl Transferase-Mediated Deoxyuridine Triphosphate-Biotin Nick-End Labeling Assay
Transfected cells were grown on 24-well culture dishes and treated with different concentrations of docetaxel. After 24 hours, cells were fixed with 4% paraformaldehyde in phosphate-buffered saline and stained with TMR-red, following the manufacturer's protocols (In Situ Cell Death Detection Kit; Roche Molecular Biochemicals, Mannheim, Germany). The experiments were carried out in duplicate.
RESULTS
We compared three popular class-prediction algorithms, WV,13 K-NN,14 and SVM.15 Because a small number of genes, fewer than 100, were desired for diagnostic application, we selected genes using popular ranking methods. We used SNR to rank genes for WV and RFE to rank genes for SVM because these ranking methods are closely associated with each algorithm. PPV is a more general ranking method but was used only for K-NN in this study and not for the other algorithms. Gene selection was performed from a pool of 1,125 genes exhibiting stable PCR amplification (see Patients and Methods). The performance of each algorithm was evaluated with up to 100 genes by leave-one-out cross validation or a comparable method. The accuracy curves of these prediction methods with a variable number of up to 100 genes are shown in Figure 1. Our criterion for selection of the diagnostic system was the greatest accuracy with the least number of diagnostic genes. The results indicated that the WV algorithm generally performed best, and, with 85 or 95 genes, its accuracy was highest (72.7%; Fig 1). Therefore, we chose to use the WV algorithm with 85 diagnostic genes.
The 85 diagnostic genes were selected using data from all 44 patients in the descending order determined by SNR. Classification of the 44 patients by the final gene set is summarized in Figure 2. Sixty-one genes were elevated in nonresponders, and 24 genes were elevated in responders (Table 2). Patients were sorted vertically in order of the strength of their respective predictions. Setting the boundary at 0 clearly separated responders from nonresponders, with only four misclassifications (Fig 2). It should be noted that complete responders and patients with progressive disease were not located near the boundary, suggesting that their expression patterns clearly allowed them to be classified as responders or nonresponders.
The WV algorithm constructed with the 85 genes was tested with 26 new patients. We measured the expression levels of the 85 genes by ATAC-PCR and predicted the degree of clinical response to docetaxel. The prediction accuracy was 80.7% (95% CI, 63.5% to 92.5%; Table 3). The sensitivity, specificity, positive predictive value, and negative predictive value were 91.7% (95% CI, 68.1% to 99.5%), 71.4% (95% CI, 46.7% to 89.5%), 73.3% (95% CI, 49.5% to 90.3%), and 90.9% (95% CI, 65.9% to 99.4%), respectively.
In a small validation set, random chance could produce a significant number of correct predictions. However, the probability of observing 21 or more correct guesses is 0.12%, as calculated by the binomial distribution.
Relationship Between Response to Docetaxel and Clinicopathologic Parameters
The relationship between the response to docetaxel and tumor size, histologic grade, and estrogen receptor (ER) status was studied in locally advanced breast cancer patients (n = 44 of 70). Response rates of tumors measuring 5 cm or less and more than 5 cm were 44.5% (10 of 22 patients) and 54.5% (12 of 22 patients), respectively. Response rates of histologic grades 1 and 2 and grade 3 tumors were 46.7% (seven of 15 patients) and 51.7% (15 of 29 patients), respectively. Response rates of patients with ER-positive and ER-negative tumors were 47.4% (nine of 19 patients) and 52.0% (13 of 25 patients), respectively. None of these clinicopathologic parameters was significantly associated with tumor response.
Molecular Features Characterizing the Docetaxel Response
To identify molecular features that underlie docetaxel sensitivity, we examined the functions of the genes that were differentially expressed between responders and nonresponders. Considering the multiplicity of gene selection, we evaluated differentially expressed genes by FDR.16 The top 71 genes ranked in the ascending order of their permutation P values were selected when the FDR was set at 0.5. All of these genes were included among the 85 predictor genes and are marked in Table 2. It should be noted that 85 genes are ranked by the descending order of SNR, and the ranking is similar but not the same as that by PPV. Expression of tubulin was elevated in nonresponders, which may contribute to resistance to docetaxel, an antimicrotubule agent.18,19 However, the most prominent characteristic in nonresponders is elevated expression of genes controlling the cellular redox environment, that is, those of the glutathione and thioredoxin systems. These genes included glutathione S-transferase pi 1, glutathione peroxidase 1, thioredoxin, and peroxiredoxin 1 (thioredoxine peroxidase 2); in addition, glutathione peroxidase 4 was among the 85 predictor genes. Although more than half of the genes on the list are false-positives, the probability that five redox genes would appear together in the 85 selected genes is low (P = .013). We concluded, then, that these redox genes may be involved in docetaxel resistance, and we confirmed their function further by transfection experiments.
Protection of MCF-7 From Docetaxel-Induced Cell Death by Redox Genes
A docetaxel dose-response curve was constructed on a human mammary cell line, MCF-7, that is normally docetaxel sensitive20 but that was transfected with various redox-associated genes (Fig 3). Three genes (glutathione-S-tranferase pi, thioredoxin, and peroxiredoxin 1) were cloned as fusion proteins with green fluorescent protein under the control of a constitutive promoter. Cell death was assayed by the terminal deoxynucleotidyl transferase-mediated deoxyuridine triphosphate-biotin nick-end labeling assay, which detects chromosomal DNA fragmentation. All genes protected MCF-7 from docetaxel-induced cell death; for example, at a 10-nmol/L dose of docetaxel, only 60% of control cells survived, whereas the three cell lines overexpressing redox genes exhibited greater than 90% cell survival.
DISCUSSION
Some may argue that none of the genes selected for diagnosis in the previous study8 overlap with those in the present study except for calreticulin, glutathione peroxidase 4, and ATP synthase. However, this overlapping is unlikely to be obtained by chance; its probability is 0.02 assuming that the 2,453 genes of our study were included among the genes on the Affymetrix GeneChip (Santa Clara, CA). In addition, several factors make such a direct comparison difficult. The Affymetrix GeneChip has a wider coverage of genes, but recent analysis revealed that the quantitative data were obtained only with abundant genes, even in budding yeast, in which the mRNA population is much less complex than that of human tissues.22 Therefore, genes with low expression were excluded from the analysis.8 ATAC-PCR can be used to measure expression of rare genes, but its gene coverage is not as large as that of the GeneChip. Consequently, the gene populations that the two studies examined are not likely to be the same. In addition, the criteria for discriminating responders and nonresponders are different. Therefore, it is not surprising that only a few diagnostic genes overlap in the two studies.
Both the clinical and pathologic response rates to docetaxel observed in the present study (48.6%, 34 of 70 patients; and 2.5%, two of 44 patients, respectively) are lower than those reported for other neoadjuvant chemotherapies consisting of both an anthracycline and a taxane, such as doxorubicin and cyclophosphamide, followed by docetaxel (90.7% and 26.1%, respectively), suggesting the superiority of sequential chemotherapy with a taxane and an anthracycline over taxane alone.23 In addition, the fact that the approved dose of docetaxel in Japan (60 mg/m2) is lower than that (100 mg/m2) commonly used in the Western countries seems to explain, at least in part, the lower response rates in the present study.
Our results, both the gene expression patterns observed in patients and the in vitro transfection findings, strongly suggest that redox genes play a major role in docetaxel resistance. The glutathione and thioredoxin systems represent two major mechanisms for maintaining the intracellular redox environment through the reduction and oxidation of thiol groups.24-26 These genes have been found to be involved in resistance to other anticancer agents; inhibition of glutathione S-transferase pi activity by antisense cDNA increased sensitivity to anticancer agents in a colorectal cancer cell line,27 and elevated expression of glutathione S-transferase pi was correlated with resistance to cisplatin in lung cancer and head and neck cancer.28,29 Furthermore, increased thioredoxin levels have been reported to reduce the sensitivity of anticancer drugs in various cases.30-32 Our findings with docetaxel are consistent with these observations and are likely to apply generally to malignancies other than breast cancer. Thus far, the glutathione and thioredoxin systems have been studied individually, but both systems seem to work cooperatively in docetaxel resistance.
Because expression of redox genes is low in responders to docetaxel, administration of chemicals inhibiting redox enzymes to nonresponders could improve their response profile. This is a new concept in anticancer drug design, and the combination of these agents with the diagnostic system of this study may substantially advance current docetaxel-based chemotherapy regimens. Several inhibitors of thioredoxin have already been identified, and one of them, PX-12,33 has entered phase I clinical trials as an anticancer agent. These inhibitors are promising candidates for use in combination therapy. This speculation indicates that gene expression profiling of a well-designed clinical trial could be an important resource for pharmaceutical developments.
Authors' Disclosures of Potential Conflicts of Interest
Acknowledgment
We thank Y. Kurokawa, MD, S. Oba, PhD, and S. Ishii, PhD, for their helpful suggestions in the analysis of the data. We also thank J. Yodoi, MD, PhD, for his critique of the manuscript, H. Kita-Matsuo, PhD, for technical advice, and S. Maki-Kinjo, K. Miyaoka-Ikenaga, and Y. Ishida for their valuable technical assistance.
NOTES
Supported by a Grant-in-Aid for the Development of Innovative Technology from the Ministry of Education, Culture, Sports, Science and Technology of Japan, and by a grant from Taisho Pharmaceuticals Co, Ltd.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
1. Ringel I, Horwitz SB: Studies with RP56976 (Taxotere): A semisynthetic analogue of taxol. J Natl Cancer Inst 83:288-291, 1991
2. Goble S, Bear HD: Emerging role of taxanes in adjuvant and neoadjuvant therapy for breast cancer: The potential and the questions. Surg Clin North Am 83:943-971, 2003
3. Egawa C, Miyoshi Y, Takamura Y, et al: Decreased expression of BRCA2 mRNA predicts favorable response to docetaxel in breast cancer. Int J Cancer 95:255-259, 2001
4. Hasegawa S, Miyoshi Y, Egawa C, et al: Mutational analysis of the class I beta-tubulin gene in human breast cancer. Int J Cancer 101:46-51, 2002
5. Burkhart CA, Kavallaris M, Horwitz SB: The role of beta-tubulin isotypes in resistance to antimitotic drugs. Biochim Biophys Acta 1471:O1-O9, 2001
6. van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536, 2002
7. Iwao K, Matoba R, Ueno N, et al: Molecular classification of primary breast tumors possessing distinct prognostic properties. Hum Mol Genet 11:199-206, 2002
8. Chang JC, Wooten EC, Tsimelzon A, et al: Gene expression profiling for the prediction of therapeutic response to docetaxel in patients with breast cancer. Lancet 362:362-369, 2003
9. ten Bokkel Huinink WW, Prove AM, Piccart M, et al: A phase II trial with docetaxel (Taxotere) in second line treatment with chemotherapy for advanced breast cancer: A study of the EORTC Early Clinical Trials Group. Ann Oncol 5:527-532, 1994
10. Heys SD, Hutcheon AW, Sarkar TK, et al: Neoadjuvant docetaxel in breast cancer: 3-year survival results from the Aberdeen trial. Clin Breast Cancer 3:S69-S74, 2002 (suppl 2)
11. Miller AB, Hoogstraten B, Staquet M, et al: Reporting results of cancer treatment. Cancer 47:207-214, 1981
12. Kato K: Adaptor-tagged competitive PCR: A novel method for measuring relative gene expression. Nucleic Acids Res 25:4694-4696, 1997
13. Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring. Science 286:531-537, 1999
14. Pomeroy SL, Tamayo P, Gaasenbeek M, et al: Prediction of central nervous system embryonal tumour outcome based on gene expression. Nature 415:436-442, 2002
15. Furey TS, Cristianini N, Duffy N, et al: Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics 16:906-914, 2000
16. Benjamini Y, Hochberg Y: Controlling the false discovery rate: A practical and powerful approach to multiple testing. J R Stat Soc B 85:289-300, 1995
17. Muro S, Takemasa I, Oba S, et al: Identification of expressed genes linked to malignancy of human colorectal carcinoma by parametric clustering of quantitative expression data. Genome Biol 4:R21, 2003
18. Schiff PB, Horwitz SB: Taxol stabilizes microtubules in mouse fibroblast cells. Proc Natl Acad Sci U S A 77:1561-1565, 1980
19. Stein CA: Mechanisms of action of taxanes in prostate cancer. Semin Oncol 26:3-7, 1999
20. Dan S, Tsunoda T, Kitahara O, et al: An integrated database of chemosensitivity to 55 anticancer drugs and gene expression profiles of 39 human cancer cell lines. Cancer Res 62:1139-1147, 2002
21. Gordon GJ, Jensen RV, Hsiao LL, et al: Using gene expression ratios to predict outcome among patients with mesothelioma. J Natl Cancer Inst 95:598-605, 2003
22. Holland MJ: Transcript abundance in yeast varies over six orders of magnitude. J Biol Chem 277:14363-14366, 2002
23. Bear HD, Anderson S, Brown A, et al: The effect on tumor response of adding sequential preoperative docetaxel to preoperative doxorubicin and cyclophosphamide: Preliminary results from national surgical adjuvant breast and bowel project protocol B-27. J Clin Oncol 21:4165-4174, 2003
24. Filomeni G, Rotilio G, Ciriolo MR: Cell signaling and the glutathione redox system. Biochem Pharmacol 64:1057-1064, 2002
25. Arner ES, Holmgren A: Physiological functions of thioredoxin and thioredoxin reductase. Eur J Biochem 267:6102-6109, 2000
26. Nakamura H, Nakamura K, Yodoi J: Redox regulation of cellular activation. Annu Rev Immunol 15:351-369, 1997
27. Ban N, Takahashi Y, Takayama T, et al: Transfection of glutathione S-transferase (GST)-pi antisense complementary DNA increases the sensitivity of a colon cancer cell line to Adriamycin, cisplatin, melphalan, and etoposide. Cancer Res 56:3577-3582, 1996
28. Bai F, Nakanishi Y, Kawasaki M, et al: Immunohistochemical expression of glutathione S-transferase-Pi can predict chemotherapy response in patients with nonsmall cell lung carcinoma. Cancer 78:416-421, 1996
29. Shiga H, Heath EI, Rasmussen AA, et al: Prognostic value of p53, glutathione S-transferase pi, and thymidylate synthase for neoadjuvant cisplatin-based chemotherapy in head and neck cancer. Clin Cancer Res 5:4097-4104, 1999
30. Yokomizo A, Ono M, Nanri H, et al: Cellular levels of thioredoxin associated with drug sensitivity to cisplatin, mitomycin C, doxorubicin, and etoposide. Cancer Res 55:4293-4296, 1995
31. Sasada T, Iwata S, Sato N, et al: Redox control of resistance to cis-diamminedichloroplatinum (II) (CDDP): Protective effect of human thioredoxin against CDDP-induced cytotoxicity. J Clin Invest 97:2268-2276, 1996
32. Kahlos K, Soini Y, Saily M, et al: Up-regulation of thioredoxin and thioredoxin reductase in human malignant pleural mesothelioma. Int J Cancer 95:198-204, 2001
33. Oblong JE, Chantler EL, Gallegos A, et al: Reversible inhibition of human thioredoxin reductase activity by cytotoxic alkyl 2-imidazolyl disulfide analogues. Cancer Chemother Pharmacol 34:434-438, 1994(Kyoko Iwao-Koizumi, Ryo M)