Changes in Gene Expression Associated With Response to Neoadjuvant Chemotherapy in Breast Cancer
http://www.100md.com
《临床肿瘤学》
the Division of Experimental Therapy, Division of Medical Oncology, Central Microarray Facility, Division of Diagnostic Oncology, Division of Surgical Oncology, Netherlands Cancer Institute, Amsterdam, the Netherlands
ABSTRACT
PURPOSE: At present, clinically useful markers predicting response of primary breast carcinomas to either doxorubicin-cyclophosphamide (AC) or doxorubicin-docetaxel (AD) are lacking. We investigated whether gene expression profiles of the primary tumor could be used to predict treatment response to either of those chemotherapy regimens.
PATIENTS AND METHODS: Within a single-institution, randomized, phase II trial, patients with locally advanced breast cancer received six courses of either AC (n = 24) or AD (n = 24) neoadjuvant chemotherapy. Gene expression profiles were generated from core-needle biopsies obtained before treatment and correlated with the response of the primary tumor to the chemotherapy administered. Additionally, pretreatment gene expression profiles were compared with those in tumors remaining after chemotherapy.
RESULTS: Ten (20%) of 48 patients showed a (near) pathologic complete remission of the primary tumor after treatment. No gene expression pattern correlating with response could be identified for all patients or for the AC or AD groups separately. The comparison of the pretreatment biopsy and the tumor excised after chemotherapy revealed differences in gene expression in tumors that showed a partial remission but not in tumors that did not respond to chemotherapy.
CONCLUSION: No gene expression profile predicting the response of primary breast carcinomas to AC- or AD-based neoadjuvant chemotherapy could be detected in this interim analysis. More subtle differences in gene expression are likely to be present but can only be reliably identified by studying a larger group of patients. Response of a breast tumor to neoadjuvant chemotherapy results in alterations in gene expression.
INTRODUCTION
Over the last three decades, neoadjuvant chemotherapy has been used for patients with locally advanced breast cancer.1 More recently, patients with large tumors also have become eligible for neoadjuvant chemotherapy to reduce the tumor size and increase the possibility of breast-conserving treatment.
Depending on the chemotherapy regimen used, the objective response rates, including both complete and partial remission (PR), amount to 60%, with up to 15% of the patients having a complete remission.2,3 In general, breast-conserving therapy can be performed in more than 50% of the patients treated with neoadjuvant chemotherapy.4
Patients achieving a pathologic complete remission (pCR) after neoadjuvant chemotherapy have a better survival compared with patients who do not achieve a pCR.5-7 Therefore, pCR can be used as a surrogate end point to assess the response to adjuvant chemotherapy.3,8
At present, a commonly used regimen for neoadjuvant chemotherapy in breast cancer is the combination of doxorubicin and cyclophosphamide (AC) because anthracycline-based regimens show high response rates. However, resistance to these therapies can exist or develop under primary therapy.9
Several trials have reported good response rates for taxane-containing regimens.10,11 Docetaxel is a potent taxane in monotherapy. It also shows high activity in combination with doxorubicin12 and activity in anthracycline-resistant disease.13,14 Recent studies describe significantly improved overall response rates in metastatic breast cancer patients treated with a combination of doxorubicin and docetaxel (AD) compared with patients treated with AC.15
Because the objective response rates after primary chemotherapy in locally advanced breast cancer are 40% to 60%, a part of the patients undergoing neoadjuvant therapy suffer from the side effects without benefiting from the treatment. Furthermore, in patients with progressive disease, valuable time for efficient treatment is lost. Given these facts, the identification of markers predicting response to primary chemotherapy would be of great advantage.
Several studies have shown that gene expression profiling is a successful tool that can not only be used for the classification of breast cancer,16 but can also help to distinguish prognostic subgroups17-19 and to predict response to neoadjuvant chemotherapy.20,21 In the study described here, we compared two different neoadjuvant chemotherapy regimens to identify gene expression patterns that can predict response to AD- versus AC-based chemotherapy. Additionally, gene expression profiles of primary breast tumors before treatment were compared with those in tumors remaining after chemotherapy.
PATIENTS AND METHODS
Selection of Patients
The majority of the patients were treated within a randomized, phase II, single-institution study performed at the Netherlands Cancer Institute. Patients with invasive breast cancer greater than 3 cm and/or at least one tumor-positive axillary lymph node were eligible for the study and randomized between AC- and AD-based neoadjuvant chemotherapy. In addition, patients who had undergone standard AC neoadjuvant chemotherapy identical to the treatment of the patients in the AC arm of the randomized study were included in the analysis.
Study Design
In total, 62 patients were included in this study; 57 participated in the randomized trial, and five were treated with AC outside the prospective trial. Fourteen patients had to be excluded from the analysis mainly because of the low percentage of tumor cells in the biopsy material. All patients gave informed consent, and the trial was approved by the Medical Ethical Committee of the Netherlands Cancer Institute. For all patients, 14-gauge biopsies of the breast tumor were taken under ultrasound guidance, snap-frozen in liquid nitrogen, and stored at –70°C. Patients were subsequently treated with six cycles of AC or six cycles of AD. In the AC arm, patients received six cycles of doxorubicin 60 mg/m2 and cyclophosphamide 600 mg/m2, whereas patients in the AD arm were treated with six cycles of doxorubicin 50 mg/m2 and docetaxel 75 mg/m2. Chemotherapy was administered every 3 weeks. After the last course of chemotherapy, patients underwent mastectomy or breast-conserving treatment according to the standard protocols used at our institute. From the 18 remaining tumors obtained at surgery, material was snap-frozen and used for microarray analysis. Response to chemotherapy was monitored by magnetic resonance imaging before treatment and after two and six courses of chemotherapy. Three patients did not complete chemotherapy because they showed progressive disease after four or five courses of chemotherapy.
Evaluation of Response
The response of the primary tumor to chemotherapy was evaluated clinically using imaging (mammography, ultrasound, and magnetic resonance imaging) and at pathologic evaluation. The principal end point was pCR, which was evaluated by pathologic examination of the surgical specimen. A pCR was defined as no residual tumor cells seen at microscopy. When a small number of scattered tumor cells was seen, the samples were classified as near pCR (npCR). Because the aim of this study was to identify a gene expression profile associated with a high sensitivity of the primary tumor to chemotherapy, tumors with a npCR were included in the group of complete remission for analytic purposes.
A partial response (PR) was defined as a reduction of the tumor mass of at least 50% and was evaluated by reviewing the combined clinical and imaging information. Patients with a residual tumor mass of more than 50% were classified as nonresponders.
Isolation and Amplification of RNA
RNA isolation and amplification were performed as previously described.22 One 5-μm tissue section (usually after 15 30-μm sections) of each biopsy and the first and the last section of each remaining tumor were hematoxylin and eosin stained to monitor the tumor cell percentage of the tissue. Only specimens with 50% of tumor cells were included in further analysis.
Probe Labeling and Microarray Hybridization
cRNA labeling and microarray hybridizations have been performed as described by Weigelt et al.22 The reference pool consisted of equal amounts of amplified cRNA of 100 invasive breast carcinomas. All experiments have been performed as dye-swaps on 18k human microarrays (Central Microarray Facility, NKI Amsterdam, Amsterdam, the Netherlands; http://microarray.nki.nl) containing 18,432 cDNAs selected from the Research Genetics Human Sequence Verified library. Fluorescent images were obtained by using the DNA Microarray Scanner G2565B (Agilent, Palo Alto, CA).
To validate the quality of the microarray data and the applied statistical routines, one in 10 hybridizations was a self-self hybridization. A self-self experiment comprises the hybridization of a RNA sample labeled with both dyes under the same conditions as all other arrays. These self-self experiments have been used to measure platform noise and as a negative control in the error model.
Microarray Data Analysis
Microarray images were analyzed with ImaGene Software (BioDiscovery Inc, El Segundo, CA), and data have been exported to R statistical package23 (NKI Microarray Facility) for normalization. Analyses were performed both on the whole data set and for the different chemotherapy regimens separately.
Normalization and Error Model
Background-corrected intensities were used to calculate log2 transformed ratios. The ratios were normalized using a lowess fit per subarray, as described earlier.24 After normalization, an error model tailored to our NKI platform was applied to identify significant regulation (http://www.cell.com/content/article/abstractuid=PIIS0092867400000155).25 The error model uses a dye-swapped pair of hybridizations to calculate the weighted average ratio per gene and a P value indicating the chance a gene is falsely classified as regulated. The error model was validated using self-self hybridizations that yielded no significantly regulated genes (data not shown). Genes were considered to be differentially expressed if they had a P < .01.
Data Selection and Filtering
To perform analysis on the different subsets, the number of genes was reduced by filtering out genes that did not significantly change in expression. The number of samples in which a gene should be significantly regulated was predefined and combined with a maximum number of missing values of four, to end up with a set of 5,000 to 7,000 genes because this number is optimal for most analysis steps.
Unsupervised Clustering
Two-dimensional unsupervised hierarchical clustering using Pearson correlation as distance function and Complete Linkage was performed with Genesis software (Institute for Genomics and Bioinformatics, Graz University of Technology, Graz, Austria; http://genome.tugraz.at/Software/).26
Supervised Classification
To identify genes differentially expressed in pretreatment biopsies of responders and nonresponders as well as in breast cancer specimens before and after treatment, we applied a supervised classification method as described previously.17,27,28 The pathologic and clinical data were used to match the patients to the different response groups. Patients with a pCR or npCR were termed responders; whereas nonresponders could have a clinical and/or pathologic PR, stable disease (SD), or progressive disease.
The signal to noise ratio (SNR) was calculated for every gene to identify the genes most relevant to the defined groups. SNR is defined as
where G1 and G2 are the groups containing the log2-transformed ratios for a gene; μ is the mean, and is the standard deviation.
To reduce the chances of including false-positive genes, we generated 2,000 random group definitions by shuffling the labels assigned to the samples and calculated the SNR values for all genes (Monte Carlo randomization). The SNR values of the permutations per gene are approximately normally distributed. When the SNR values of the defined groups are ordered and plotted together with their confidence levels resulting from the permutations, the graph quickly reveals the number of genes that can be selected with respect to the false-positive chance.
Once a set of genes has been identified for the two defined groups, a profile is determined by calculating the mean expression ratio for each gene. Classification is performed by calculating the Pearson correlation coefficient for all samples to both profiles and plotting them in two-dimensional space. Separation usually occurs at the x = y line, but other cutoffs can be chosen if minimizing false-positives or false-negatives has a higher priority.17
Although only individually significant SNRs are included, it is possible that less genes perform better at separating the two groups.17 To determine the optimal number of genes separating the groups, a cross-validation procedure is used as described earlier.17,29
Supplementary Information
Supplementary information on the methods and additional results are provided at the following Web site: http://microarrays.nki.nl/research/hannemann_ACAD_2005/.
RESULTS
Patients Characteristics
We included 62 patients in this study; 57 had participated in the randomized phase II trial, and five were treated with AC outside of this trial. Gene expression profiling was performed on tumor material obtained from 24 patients treated with neoadjuvant AC and from 24 patients treated with neoadjuvant AD. Gene expression of the pretreatment biopsy could only be determined for 31 patients. For two patients, only expression profiles of the remaining tumor after neoadjuvant chemotherapy were obtained; and for 15 patients, expression profiles of both the biopsy before treatment and the tumor after treatment were obtained. Patient characteristics are listed in Table 1.
Ten patients (21%) had a pCR or npCR of the primary tumor (five patients after AC and five after AD); 22 patients (46%) had a PR (11 patients after AC and 11 after AD); and 16 patients (33%) showed no response (eight patients after AC and eight after AD). For the 14 patients who could not be included in the analysis, the clinical and pathologic characteristics, as described in Table 1, were not statistically significantly different from those of the patients who could be analyzed.
Data Filtering
Data were filtered to obtain a reasonable number of significantly regulated genes (between 5,000 and 7,000). Filtering was individually performed on all datasets (supplemental Table 1; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
Prediction of Response to Neoadjuvant Chemotherapy
To identify a gene expression pattern in the prechemotherapy biopsy predicting response to neoadjuvant chemotherapy, we compared the gene expression profiles of patients with a pCR or npCR with the gene expression pattern in the biopsies of patients showing PR, SD, or progressive disease. Analyses were performed for the whole patient series as well as separately for the patients treated either in the AC arm or in the AD arm of the study.
Unsupervised two-dimensional hierarchical clustering groups tumors based on their overall similarities in gene expression. The biopsies of tumors that showed an npCR and those that did not show a npCR did not cluster in two distinct groups (data not shown).
We subsequently performed Monte Carlo randomizations to identify genes differentially expressed between the groups of interest. No significant SNRs were found.
In addition, we also applied Monte Carlo randomization to the AC and AD arms separately. Again, no significant SNRs were found (data not shown).
We also tested this approach on groups of patients who were subdivided in other ways (npCR v SD and SD v npCR and PR). Again, no gene expression signature significantly associated with response to neoadjuvant chemotherapy was identified.
The supervised classification approach used up to this point selects reporter genes on an individual basis based on the SNR. However, it is quite possible that response to neoadjuvant chemotherapy could only be predicted based on combinatorial effects between genes. For example, individual genes may not show any correlation with the outcome, but in the joint space spanned by a number of genes, responders and nonresponders could be separated. To investigate this possibility, we also evaluated the Liknon classifier30 on this data set in a cross-validation setting. This classifier simultaneously selects the genes and trains the classifier in high-dimensional spaces. Across a wide range of parameter settings of the classifier, it was impossible to separate the groups. The best performing classifier misclassified all nonresponders (ie, it simply predicted nonresponse, regardless of the gene expression profile). This strengthens the conclusion that no profile predictive of response to neoadjuvant chemotherapy could be found.
Changes in Gene Expression During Neoadjuvant Chemotherapy
Unsupervised clustering. Gene expression profiling was performed on 46 pretreatment biopsies and 17 posttreatment tumors. From 15 patients, tumor material obtained before and after treatment was available. Two-dimensional hierarchical clustering was used to order all samples according to their similarities in gene expression (Fig 1). Six tumors obtained after treatment clustered adjacent to the sample of the same patient obtained before treatment. Five of these six patients did not show any response to the chemotherapy but had SD.
In addition, hierarchical clustering was performed on 15 pairs of samples obtained from the same patients before and after chemotherapy. Clustering results were highly consistent with the ones obtained from the whole data set. Eight pairs clustered together, and six of them were already clustering next to each other by using the whole data set (supplemental Fig 1; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
Supervised classification. The results of the unsupervised clustering indicated that tumors responding to chemotherapy showed major changes in gene expression. Remarkably, no prominent changes in overall gene expression could be observed when tumors did not respond to chemotherapy. To identify genes that are differentially expressed between biopsy and postchemotherapy tumor tissue, we performed supervised classification. Figure 2 shows ranked SNR values and indicates the number of genes discriminating between untreated and treated tumor specimens. The optimal number of genes has subsequently been determined by cross validation. Using nine-fold cross validation on the whole data set containing 46 biopsies before treatment and 17 specimens after treatment, an optimal number of 30 genes was identified that distinguished before and after treatment samples with an average performance of 89% (Fig 3A).
By analyzing only the data of the patients treated in the AC arm (22 biopsies and nine tumors after treatment), a class predictor consisting of 71 genes with an average performance of 93% was found in a 10-fold cross-validation procedure (Fig 3B). In the AD arm (24 biopsies and eight specimen after treatment), a classifier of 17 genes with an average performance of 87% was identified in an eight-fold cross-validation procedure (Fig 3C).
Most of the genes separating the group of untreated biopsies from the tumors after chemotherapy are involved in the metabolism of the cell. A few genes play a role in proliferation or apoptosis (Table 2).
Supervised classification and cross validation were also performed on the 15 paired samples only. These results are provided at our Web site (supplemental Fig 2 and Table 2; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
DISCUSSION
We have performed gene expression profiling of RNA isolated from core-needle biopsies from primary breast carcinomas before treatment with neoadjuvant chemotherapy. Patients in this study received either AC or AD.
The primary goal of the study was to identify a classifier predicting the response of the primary tumor to either AC or AD chemotherapy. Furthermore, changes in gene expression patterns of the tumors induced by chemotherapy treatment were identified by comparing the profiles of the pretreatment biopsy with the profiles of the tumors after treatment.
No significant differences in the gene expression of untreated tumors from patients with npCR compared with all other patients were observed. Even when response to chemotherapy was defined in different ways (eg, any response v no response), no predictive gene expression profile could be identified. These findings indicate that there are no dominant gene expression signatures that account for the large differences in patient responses to AC and AD chemotherapy in breast cancer. However, this does not exclude the existence of a relatively subtle predictive profile, which may only become apparent when a larger set of patients can be evaluated. Of course, patient-associated factors, including drug metabolism, also have an important influence on chemotherapy response, which cannot be measured using gene expression profiling of the tumor.
Gene expression patterns predicting response of the primary breast tumor to neoadjuvant chemotherapy have been described by others. Chang et al21 studied the response to neoadjuvant docetaxel monotherapy in 24 patients and found a correlation between the expression of 92 genes and response. Ayers et al,20 using a similar study design, determined a profile of 74 genes predicting response to neoadjuvant paclitaxel/fluorouracil plus doxorubicin plus cyclophosphamide (T/FAC) in 24 patients, which was validated in a group of 18 patients.
Interestingly, both studies applied different definitions of patient response. Chang et al21 arbitrarily defined sensitive tumors as those with less than 25% residual disease because this cutoff divided the samples into two equally sized groups. Ayers et al20 only classified patients without any residual invasive cancer as responders. Additionally, these investigators isolated RNA from fine-needle aspirates from the primary tumor, rather than core-needle biopsies. There are also differences in tumor size of the patients included in the study of Chang et al (median tumor diameter, 8 cm) compared with our study (median tumor diameter, 4 cm). The classifiers described by Chang et al21 and Ayers et al20 predict patient's response to docetaxel monotherapy and T/FAC, respectively. In contrast, our patients were treated with either AC or AD. Given the fact that we could not determine a predictor of response in a similar number of patients compared with the other two studies, there might be no strong predictive profile for response to AC- or AD-containing chemotherapy.
In this pilot study, we are able to show that the administration of neoadjuvant chemotherapy causes major changes in gene expression in locally advanced breast cancer responding to treatment. However, no changes occur when the tumor does not respond.
These results are in contrast with a study by Buchholz et al.31 Biopsies from one patient obtained before treatment and 24 and 48 hours after initiation of treatment clustered more closely together than samples obtained from different patients. Also Perou et al16 showed that the gene expression profiles in two tumor samples from the same patient are more similar to each other than to samples from other individuals. Our data indicate that tumors that are sensitive to the chemotherapy administered to the patients develop significant changes in their gene expression profile during the courses of neoadjuvant chemotherapy, whereas resistant tumors do not.
These opposing results between our study and the one of Buchholz et al31 may be a result of the fact that the time points of the second and third biopsy in the latter study were 1 and 2 days after the initiation of treatment, respectively. A large part of the changes in gene expression may only occur later in time and would have been missed by using this approach.
Within our data set, the classifier separating treated from untreated samples consists of the first 71 genes in the AC group or the top 17 genes in the AD group. The classifier of the AD arm contains one gene involved in apoptosis; CD14 antigen is known to bind apoptotic cells and to be part of the mitogen-activated protein kinase signaling pathway. Most of the genes that were differentially expressed are involved in the metabolism of the cell (eg, Cox7B and LAMR1 in the AC arm and PolR2H and LSM7 in the AD arm). These results indicate that resistance or sensitivity of tumor cells to chemotherapy is not only dependent on apoptotic pathways and cell cycle regulation, but that other biologic processes also are required for this process. It may also be that the genes determined in these classifiers have additional, as yet unknown functions in regulating cell death, drug sensitivity, or cell communication, which contribute to sensitivity or resistance of tumors to neoadjuvant chemotherapy.
To implement our results in a clinical setting, a classifier should be established for early response of the tumor to chemotherapy by comparing the gene expression profile of the untreated biopsy with the gene expression profile of a biopsy obtained after one course of chemotherapy. If the response classifier is present in the tumor, the treatment could be continued, whereas the absence of the classifier would imply that patients should be treated with another chemotherapeutic regimen.
In summary, we showed that the differences in gene expression between responders and nonresponders to neoadjuvant chemotherapy must be rather subtle. No prominent gene expression profiles predicting patient's response to neoadjuvant chemotherapy emerge from the neoadjuvant studies carried out to date, including this interim report presented here. The correlation of gene expression profiles with response to neoadjuvant chemotherapy may currently represent the most feasible and powerful approach to identify response classifiers. However, relatively large studies will be required to achieve reliable and clinically useful predictive tests.
Authors' Disclosures of Potential Conflicts of Interest
The authors indicated no potential conflicts of interest.
Acknowledgment
We thank B. Weigelt for critical reading of the manuscript.
NOTES
Supported by the Dutch Cancer Society.
Presented in part at the European Breast Cancer Conference (Hamburg, Germany, March 2004) and at the 40th Annual Meeting of the American Society of Clinical Oncology, New Orleans, LA, June 5-8, 2004.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Schwartz GF, Hortobagyi GN: Proceedings of the consensus conference on neoadjuvant chemotherapy in carcinoma of the breast, April 26-28, 2003, Philadelphia, Pennsylvania. Cancer 100:2512-2532, 2004
Pierga JY, Mouret E, Laurence V, et al: Prognostic factors for survival after neoadjuvant chemotherapy in operable breast cancer: The role of clinical response. Eur J Cancer 39:1089-1096, 2003
Ferriere JP, Assier I, Cure H, et al: Primary chemotherapy in breast cancer: Correlation between tumor response and patient outcome. Am J Clin Oncol 21:117-120, 1998
Mauriac L, Durand M, Avril A, et al: Effects of primary chemotherapy in conservative treatment of breast cancer patients with operable tumors larger than 3 cm: Results of a randomized trial in a single centre. Ann Oncol 2:347-354, 1991
Fisher B, Bryant J, Wolmark N, et al: Effect of preoperative chemotherapy on the outcome of women with operable breast cancer. J Clin Oncol 16:2672-2685, 1998
Kuerer HM, Newman LA, Smith TL, et al: Clinical course of breast cancer patients with complete pathologic primary tumor and axillary lymph node response to doxorubicin-based neoadjuvant chemotherapy. J Clin Oncol 17:460-469, 1999
Wolmark N, Wang J, Mamounas E, et al: Preoperative chemotherapy in patients with operable breast cancer: Nine-year results from National Surgical Adjuvant Breast and Bowel Project B-18. J Natl Cancer Inst Monogr 30:96-102, 2001
Bonadonna G, Valagussa P, Brambilla C, et al: Primary chemotherapy in operable breast cancer: Eight-year experience at the Milan Cancer Institute. J Clin Oncol 16:93-100, 1998
Ravdin PM: Docetaxel (Taxotere) for the treatment of anthracycline-resistant breast cancer. Semin Oncol 24:S10, 1997 (suppl 10)
Moliterni A, Tarenzi E, Capri G, et al: Pilot study of primary chemotherapy with doxorubicin plus paclitaxel in women with locally advanced or operable breast cancer. Semin Oncol 24:S17, 1997 (suppl 17)
Holmes FA, Walters RS, Theriault RL, et al: Phase II trial of Taxol, an active drug in the treatment of metastatic breast cancer. J Natl Cancer Inst 83:1797-1805, 1991
Bissery MC, Vrignaud P, Lavelle F: Preclinical profile of docetaxel (Taxotere): Efficacy as a single agent and in combination. Semin Oncol 22:3-16, 1995
Sjostrom J, Blomqvist C, Mouridsen H, et al: Docetaxel compared with sequential methotrexate and 5-fluorouracil in patients with advanced breast cancer after anthracycline failure: A randomised phase III study with crossover on progression by the Scandinavian Breast Group. Eur J Cancer 35:1194-1201, 1999
Nabholtz JM, Senn HJ, Bezwoda WR, et al: Prospective randomized trial of docetaxel versus mitomycin plus vinblastine in patients with metastatic breast cancer progressing despite previous anthracycline-containing chemotherapy: 304 Study Group. J Clin Oncol 17:1413-1424, 1999
Nabholtz JM, Falkson C, Campos D, et al: Docetaxel and doxorubicin compared with doxorubicin and cyclophosphamide as first-line chemotherapy for metastatic breast cancer: Results of a randomized, multicenter, phase III trial. J Clin Oncol 21:968-975, 2003
Perou CM, Sorlie T, Eisen MB, et al: Molecular portraits of human breast tumours. Nature 406:747-752, 2000
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
Pusztai L, Ayers M, Stec J, et al: Gene expression profiles obtained from fine-needle aspirations of breast cancer reliably identify routine prognostic markers and reveal large-scale molecular differences between estrogen-negative and estrogen-positive tumors. Clin Cancer Res 9:2406-2415, 2003
van de Vijver MJ, He YD, van't Veer LJ, et al: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002
Ayers M, Symmans WF, Stec J, et al: Gene expression profiles predict complete pathologic response to neoadjuvant paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide chemotherapy in breast cancer. J Clin Oncol 22:2284-2293, 2004
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
Weigelt B, Glas AM, Wessels LF, et al: Gene expression profiles of primary breast tumors maintained in distant metastases. Proc Natl Acad Sci U S A 100:15901-15905, 2003
R Foundation for Statistical Computing: R: A Language and Environment for Statistical Computing. Vienna, Austria, 2003
Yang YH, Dudoit S, Luu P, et al: Normalization for cDNA microarray data: A robust composite method addressing single and multiple slide systematic variation. Nucleic Acids Res 30:e15, 2002
Hughes TR, Marton MJ, Jones AR, et al: Functional discovery via a compendium of expression profiles. Cell 102:109-126, 2000
Sturn A, Quackenbush J, Trajanoski Z: Genesis: Cluster analysis of microarray data. Bioinformatics 18:207-208, 2002
Khan J, Wei JS, Ringner M, et al: Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med 7:673-679, 2001
Simon R, Radmacher MD, Dobbin K, et al: Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J Natl Cancer Inst 95:14-18, 2003
Ambroise C, McLachlan GJ: Selection bias in gene extraction on the basis of microarray gene-expression data. Proc Natl Acad Sci U S A 99:6562-6566, 2002
Bhattacharyya C, Grate L, Rizki A, et al: Simultaneous classification and relevant feature identification in high-dimensional spaces: Application to molecular profiling data. Signal Processing 83:729-743, 2003
Buchholz TA, Stivers DN, Stec J, et al: Global gene expression changes during neoadjuvant chemotherapy for human breast cancer. Cancer J 8:461-468, 2002(Juliane Hannemann, Hendri)
ABSTRACT
PURPOSE: At present, clinically useful markers predicting response of primary breast carcinomas to either doxorubicin-cyclophosphamide (AC) or doxorubicin-docetaxel (AD) are lacking. We investigated whether gene expression profiles of the primary tumor could be used to predict treatment response to either of those chemotherapy regimens.
PATIENTS AND METHODS: Within a single-institution, randomized, phase II trial, patients with locally advanced breast cancer received six courses of either AC (n = 24) or AD (n = 24) neoadjuvant chemotherapy. Gene expression profiles were generated from core-needle biopsies obtained before treatment and correlated with the response of the primary tumor to the chemotherapy administered. Additionally, pretreatment gene expression profiles were compared with those in tumors remaining after chemotherapy.
RESULTS: Ten (20%) of 48 patients showed a (near) pathologic complete remission of the primary tumor after treatment. No gene expression pattern correlating with response could be identified for all patients or for the AC or AD groups separately. The comparison of the pretreatment biopsy and the tumor excised after chemotherapy revealed differences in gene expression in tumors that showed a partial remission but not in tumors that did not respond to chemotherapy.
CONCLUSION: No gene expression profile predicting the response of primary breast carcinomas to AC- or AD-based neoadjuvant chemotherapy could be detected in this interim analysis. More subtle differences in gene expression are likely to be present but can only be reliably identified by studying a larger group of patients. Response of a breast tumor to neoadjuvant chemotherapy results in alterations in gene expression.
INTRODUCTION
Over the last three decades, neoadjuvant chemotherapy has been used for patients with locally advanced breast cancer.1 More recently, patients with large tumors also have become eligible for neoadjuvant chemotherapy to reduce the tumor size and increase the possibility of breast-conserving treatment.
Depending on the chemotherapy regimen used, the objective response rates, including both complete and partial remission (PR), amount to 60%, with up to 15% of the patients having a complete remission.2,3 In general, breast-conserving therapy can be performed in more than 50% of the patients treated with neoadjuvant chemotherapy.4
Patients achieving a pathologic complete remission (pCR) after neoadjuvant chemotherapy have a better survival compared with patients who do not achieve a pCR.5-7 Therefore, pCR can be used as a surrogate end point to assess the response to adjuvant chemotherapy.3,8
At present, a commonly used regimen for neoadjuvant chemotherapy in breast cancer is the combination of doxorubicin and cyclophosphamide (AC) because anthracycline-based regimens show high response rates. However, resistance to these therapies can exist or develop under primary therapy.9
Several trials have reported good response rates for taxane-containing regimens.10,11 Docetaxel is a potent taxane in monotherapy. It also shows high activity in combination with doxorubicin12 and activity in anthracycline-resistant disease.13,14 Recent studies describe significantly improved overall response rates in metastatic breast cancer patients treated with a combination of doxorubicin and docetaxel (AD) compared with patients treated with AC.15
Because the objective response rates after primary chemotherapy in locally advanced breast cancer are 40% to 60%, a part of the patients undergoing neoadjuvant therapy suffer from the side effects without benefiting from the treatment. Furthermore, in patients with progressive disease, valuable time for efficient treatment is lost. Given these facts, the identification of markers predicting response to primary chemotherapy would be of great advantage.
Several studies have shown that gene expression profiling is a successful tool that can not only be used for the classification of breast cancer,16 but can also help to distinguish prognostic subgroups17-19 and to predict response to neoadjuvant chemotherapy.20,21 In the study described here, we compared two different neoadjuvant chemotherapy regimens to identify gene expression patterns that can predict response to AD- versus AC-based chemotherapy. Additionally, gene expression profiles of primary breast tumors before treatment were compared with those in tumors remaining after chemotherapy.
PATIENTS AND METHODS
Selection of Patients
The majority of the patients were treated within a randomized, phase II, single-institution study performed at the Netherlands Cancer Institute. Patients with invasive breast cancer greater than 3 cm and/or at least one tumor-positive axillary lymph node were eligible for the study and randomized between AC- and AD-based neoadjuvant chemotherapy. In addition, patients who had undergone standard AC neoadjuvant chemotherapy identical to the treatment of the patients in the AC arm of the randomized study were included in the analysis.
Study Design
In total, 62 patients were included in this study; 57 participated in the randomized trial, and five were treated with AC outside the prospective trial. Fourteen patients had to be excluded from the analysis mainly because of the low percentage of tumor cells in the biopsy material. All patients gave informed consent, and the trial was approved by the Medical Ethical Committee of the Netherlands Cancer Institute. For all patients, 14-gauge biopsies of the breast tumor were taken under ultrasound guidance, snap-frozen in liquid nitrogen, and stored at –70°C. Patients were subsequently treated with six cycles of AC or six cycles of AD. In the AC arm, patients received six cycles of doxorubicin 60 mg/m2 and cyclophosphamide 600 mg/m2, whereas patients in the AD arm were treated with six cycles of doxorubicin 50 mg/m2 and docetaxel 75 mg/m2. Chemotherapy was administered every 3 weeks. After the last course of chemotherapy, patients underwent mastectomy or breast-conserving treatment according to the standard protocols used at our institute. From the 18 remaining tumors obtained at surgery, material was snap-frozen and used for microarray analysis. Response to chemotherapy was monitored by magnetic resonance imaging before treatment and after two and six courses of chemotherapy. Three patients did not complete chemotherapy because they showed progressive disease after four or five courses of chemotherapy.
Evaluation of Response
The response of the primary tumor to chemotherapy was evaluated clinically using imaging (mammography, ultrasound, and magnetic resonance imaging) and at pathologic evaluation. The principal end point was pCR, which was evaluated by pathologic examination of the surgical specimen. A pCR was defined as no residual tumor cells seen at microscopy. When a small number of scattered tumor cells was seen, the samples were classified as near pCR (npCR). Because the aim of this study was to identify a gene expression profile associated with a high sensitivity of the primary tumor to chemotherapy, tumors with a npCR were included in the group of complete remission for analytic purposes.
A partial response (PR) was defined as a reduction of the tumor mass of at least 50% and was evaluated by reviewing the combined clinical and imaging information. Patients with a residual tumor mass of more than 50% were classified as nonresponders.
Isolation and Amplification of RNA
RNA isolation and amplification were performed as previously described.22 One 5-μm tissue section (usually after 15 30-μm sections) of each biopsy and the first and the last section of each remaining tumor were hematoxylin and eosin stained to monitor the tumor cell percentage of the tissue. Only specimens with 50% of tumor cells were included in further analysis.
Probe Labeling and Microarray Hybridization
cRNA labeling and microarray hybridizations have been performed as described by Weigelt et al.22 The reference pool consisted of equal amounts of amplified cRNA of 100 invasive breast carcinomas. All experiments have been performed as dye-swaps on 18k human microarrays (Central Microarray Facility, NKI Amsterdam, Amsterdam, the Netherlands; http://microarray.nki.nl) containing 18,432 cDNAs selected from the Research Genetics Human Sequence Verified library. Fluorescent images were obtained by using the DNA Microarray Scanner G2565B (Agilent, Palo Alto, CA).
To validate the quality of the microarray data and the applied statistical routines, one in 10 hybridizations was a self-self hybridization. A self-self experiment comprises the hybridization of a RNA sample labeled with both dyes under the same conditions as all other arrays. These self-self experiments have been used to measure platform noise and as a negative control in the error model.
Microarray Data Analysis
Microarray images were analyzed with ImaGene Software (BioDiscovery Inc, El Segundo, CA), and data have been exported to R statistical package23 (NKI Microarray Facility) for normalization. Analyses were performed both on the whole data set and for the different chemotherapy regimens separately.
Normalization and Error Model
Background-corrected intensities were used to calculate log2 transformed ratios. The ratios were normalized using a lowess fit per subarray, as described earlier.24 After normalization, an error model tailored to our NKI platform was applied to identify significant regulation (http://www.cell.com/content/article/abstractuid=PIIS0092867400000155).25 The error model uses a dye-swapped pair of hybridizations to calculate the weighted average ratio per gene and a P value indicating the chance a gene is falsely classified as regulated. The error model was validated using self-self hybridizations that yielded no significantly regulated genes (data not shown). Genes were considered to be differentially expressed if they had a P < .01.
Data Selection and Filtering
To perform analysis on the different subsets, the number of genes was reduced by filtering out genes that did not significantly change in expression. The number of samples in which a gene should be significantly regulated was predefined and combined with a maximum number of missing values of four, to end up with a set of 5,000 to 7,000 genes because this number is optimal for most analysis steps.
Unsupervised Clustering
Two-dimensional unsupervised hierarchical clustering using Pearson correlation as distance function and Complete Linkage was performed with Genesis software (Institute for Genomics and Bioinformatics, Graz University of Technology, Graz, Austria; http://genome.tugraz.at/Software/).26
Supervised Classification
To identify genes differentially expressed in pretreatment biopsies of responders and nonresponders as well as in breast cancer specimens before and after treatment, we applied a supervised classification method as described previously.17,27,28 The pathologic and clinical data were used to match the patients to the different response groups. Patients with a pCR or npCR were termed responders; whereas nonresponders could have a clinical and/or pathologic PR, stable disease (SD), or progressive disease.
The signal to noise ratio (SNR) was calculated for every gene to identify the genes most relevant to the defined groups. SNR is defined as
where G1 and G2 are the groups containing the log2-transformed ratios for a gene; μ is the mean, and is the standard deviation.
To reduce the chances of including false-positive genes, we generated 2,000 random group definitions by shuffling the labels assigned to the samples and calculated the SNR values for all genes (Monte Carlo randomization). The SNR values of the permutations per gene are approximately normally distributed. When the SNR values of the defined groups are ordered and plotted together with their confidence levels resulting from the permutations, the graph quickly reveals the number of genes that can be selected with respect to the false-positive chance.
Once a set of genes has been identified for the two defined groups, a profile is determined by calculating the mean expression ratio for each gene. Classification is performed by calculating the Pearson correlation coefficient for all samples to both profiles and plotting them in two-dimensional space. Separation usually occurs at the x = y line, but other cutoffs can be chosen if minimizing false-positives or false-negatives has a higher priority.17
Although only individually significant SNRs are included, it is possible that less genes perform better at separating the two groups.17 To determine the optimal number of genes separating the groups, a cross-validation procedure is used as described earlier.17,29
Supplementary Information
Supplementary information on the methods and additional results are provided at the following Web site: http://microarrays.nki.nl/research/hannemann_ACAD_2005/.
RESULTS
Patients Characteristics
We included 62 patients in this study; 57 had participated in the randomized phase II trial, and five were treated with AC outside of this trial. Gene expression profiling was performed on tumor material obtained from 24 patients treated with neoadjuvant AC and from 24 patients treated with neoadjuvant AD. Gene expression of the pretreatment biopsy could only be determined for 31 patients. For two patients, only expression profiles of the remaining tumor after neoadjuvant chemotherapy were obtained; and for 15 patients, expression profiles of both the biopsy before treatment and the tumor after treatment were obtained. Patient characteristics are listed in Table 1.
Ten patients (21%) had a pCR or npCR of the primary tumor (five patients after AC and five after AD); 22 patients (46%) had a PR (11 patients after AC and 11 after AD); and 16 patients (33%) showed no response (eight patients after AC and eight after AD). For the 14 patients who could not be included in the analysis, the clinical and pathologic characteristics, as described in Table 1, were not statistically significantly different from those of the patients who could be analyzed.
Data Filtering
Data were filtered to obtain a reasonable number of significantly regulated genes (between 5,000 and 7,000). Filtering was individually performed on all datasets (supplemental Table 1; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
Prediction of Response to Neoadjuvant Chemotherapy
To identify a gene expression pattern in the prechemotherapy biopsy predicting response to neoadjuvant chemotherapy, we compared the gene expression profiles of patients with a pCR or npCR with the gene expression pattern in the biopsies of patients showing PR, SD, or progressive disease. Analyses were performed for the whole patient series as well as separately for the patients treated either in the AC arm or in the AD arm of the study.
Unsupervised two-dimensional hierarchical clustering groups tumors based on their overall similarities in gene expression. The biopsies of tumors that showed an npCR and those that did not show a npCR did not cluster in two distinct groups (data not shown).
We subsequently performed Monte Carlo randomizations to identify genes differentially expressed between the groups of interest. No significant SNRs were found.
In addition, we also applied Monte Carlo randomization to the AC and AD arms separately. Again, no significant SNRs were found (data not shown).
We also tested this approach on groups of patients who were subdivided in other ways (npCR v SD and SD v npCR and PR). Again, no gene expression signature significantly associated with response to neoadjuvant chemotherapy was identified.
The supervised classification approach used up to this point selects reporter genes on an individual basis based on the SNR. However, it is quite possible that response to neoadjuvant chemotherapy could only be predicted based on combinatorial effects between genes. For example, individual genes may not show any correlation with the outcome, but in the joint space spanned by a number of genes, responders and nonresponders could be separated. To investigate this possibility, we also evaluated the Liknon classifier30 on this data set in a cross-validation setting. This classifier simultaneously selects the genes and trains the classifier in high-dimensional spaces. Across a wide range of parameter settings of the classifier, it was impossible to separate the groups. The best performing classifier misclassified all nonresponders (ie, it simply predicted nonresponse, regardless of the gene expression profile). This strengthens the conclusion that no profile predictive of response to neoadjuvant chemotherapy could be found.
Changes in Gene Expression During Neoadjuvant Chemotherapy
Unsupervised clustering. Gene expression profiling was performed on 46 pretreatment biopsies and 17 posttreatment tumors. From 15 patients, tumor material obtained before and after treatment was available. Two-dimensional hierarchical clustering was used to order all samples according to their similarities in gene expression (Fig 1). Six tumors obtained after treatment clustered adjacent to the sample of the same patient obtained before treatment. Five of these six patients did not show any response to the chemotherapy but had SD.
In addition, hierarchical clustering was performed on 15 pairs of samples obtained from the same patients before and after chemotherapy. Clustering results were highly consistent with the ones obtained from the whole data set. Eight pairs clustered together, and six of them were already clustering next to each other by using the whole data set (supplemental Fig 1; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
Supervised classification. The results of the unsupervised clustering indicated that tumors responding to chemotherapy showed major changes in gene expression. Remarkably, no prominent changes in overall gene expression could be observed when tumors did not respond to chemotherapy. To identify genes that are differentially expressed between biopsy and postchemotherapy tumor tissue, we performed supervised classification. Figure 2 shows ranked SNR values and indicates the number of genes discriminating between untreated and treated tumor specimens. The optimal number of genes has subsequently been determined by cross validation. Using nine-fold cross validation on the whole data set containing 46 biopsies before treatment and 17 specimens after treatment, an optimal number of 30 genes was identified that distinguished before and after treatment samples with an average performance of 89% (Fig 3A).
By analyzing only the data of the patients treated in the AC arm (22 biopsies and nine tumors after treatment), a class predictor consisting of 71 genes with an average performance of 93% was found in a 10-fold cross-validation procedure (Fig 3B). In the AD arm (24 biopsies and eight specimen after treatment), a classifier of 17 genes with an average performance of 87% was identified in an eight-fold cross-validation procedure (Fig 3C).
Most of the genes separating the group of untreated biopsies from the tumors after chemotherapy are involved in the metabolism of the cell. A few genes play a role in proliferation or apoptosis (Table 2).
Supervised classification and cross validation were also performed on the 15 paired samples only. These results are provided at our Web site (supplemental Fig 2 and Table 2; see http://microarrays.nki.nl/research/hannemann_ACAD_2005/).
DISCUSSION
We have performed gene expression profiling of RNA isolated from core-needle biopsies from primary breast carcinomas before treatment with neoadjuvant chemotherapy. Patients in this study received either AC or AD.
The primary goal of the study was to identify a classifier predicting the response of the primary tumor to either AC or AD chemotherapy. Furthermore, changes in gene expression patterns of the tumors induced by chemotherapy treatment were identified by comparing the profiles of the pretreatment biopsy with the profiles of the tumors after treatment.
No significant differences in the gene expression of untreated tumors from patients with npCR compared with all other patients were observed. Even when response to chemotherapy was defined in different ways (eg, any response v no response), no predictive gene expression profile could be identified. These findings indicate that there are no dominant gene expression signatures that account for the large differences in patient responses to AC and AD chemotherapy in breast cancer. However, this does not exclude the existence of a relatively subtle predictive profile, which may only become apparent when a larger set of patients can be evaluated. Of course, patient-associated factors, including drug metabolism, also have an important influence on chemotherapy response, which cannot be measured using gene expression profiling of the tumor.
Gene expression patterns predicting response of the primary breast tumor to neoadjuvant chemotherapy have been described by others. Chang et al21 studied the response to neoadjuvant docetaxel monotherapy in 24 patients and found a correlation between the expression of 92 genes and response. Ayers et al,20 using a similar study design, determined a profile of 74 genes predicting response to neoadjuvant paclitaxel/fluorouracil plus doxorubicin plus cyclophosphamide (T/FAC) in 24 patients, which was validated in a group of 18 patients.
Interestingly, both studies applied different definitions of patient response. Chang et al21 arbitrarily defined sensitive tumors as those with less than 25% residual disease because this cutoff divided the samples into two equally sized groups. Ayers et al20 only classified patients without any residual invasive cancer as responders. Additionally, these investigators isolated RNA from fine-needle aspirates from the primary tumor, rather than core-needle biopsies. There are also differences in tumor size of the patients included in the study of Chang et al (median tumor diameter, 8 cm) compared with our study (median tumor diameter, 4 cm). The classifiers described by Chang et al21 and Ayers et al20 predict patient's response to docetaxel monotherapy and T/FAC, respectively. In contrast, our patients were treated with either AC or AD. Given the fact that we could not determine a predictor of response in a similar number of patients compared with the other two studies, there might be no strong predictive profile for response to AC- or AD-containing chemotherapy.
In this pilot study, we are able to show that the administration of neoadjuvant chemotherapy causes major changes in gene expression in locally advanced breast cancer responding to treatment. However, no changes occur when the tumor does not respond.
These results are in contrast with a study by Buchholz et al.31 Biopsies from one patient obtained before treatment and 24 and 48 hours after initiation of treatment clustered more closely together than samples obtained from different patients. Also Perou et al16 showed that the gene expression profiles in two tumor samples from the same patient are more similar to each other than to samples from other individuals. Our data indicate that tumors that are sensitive to the chemotherapy administered to the patients develop significant changes in their gene expression profile during the courses of neoadjuvant chemotherapy, whereas resistant tumors do not.
These opposing results between our study and the one of Buchholz et al31 may be a result of the fact that the time points of the second and third biopsy in the latter study were 1 and 2 days after the initiation of treatment, respectively. A large part of the changes in gene expression may only occur later in time and would have been missed by using this approach.
Within our data set, the classifier separating treated from untreated samples consists of the first 71 genes in the AC group or the top 17 genes in the AD group. The classifier of the AD arm contains one gene involved in apoptosis; CD14 antigen is known to bind apoptotic cells and to be part of the mitogen-activated protein kinase signaling pathway. Most of the genes that were differentially expressed are involved in the metabolism of the cell (eg, Cox7B and LAMR1 in the AC arm and PolR2H and LSM7 in the AD arm). These results indicate that resistance or sensitivity of tumor cells to chemotherapy is not only dependent on apoptotic pathways and cell cycle regulation, but that other biologic processes also are required for this process. It may also be that the genes determined in these classifiers have additional, as yet unknown functions in regulating cell death, drug sensitivity, or cell communication, which contribute to sensitivity or resistance of tumors to neoadjuvant chemotherapy.
To implement our results in a clinical setting, a classifier should be established for early response of the tumor to chemotherapy by comparing the gene expression profile of the untreated biopsy with the gene expression profile of a biopsy obtained after one course of chemotherapy. If the response classifier is present in the tumor, the treatment could be continued, whereas the absence of the classifier would imply that patients should be treated with another chemotherapeutic regimen.
In summary, we showed that the differences in gene expression between responders and nonresponders to neoadjuvant chemotherapy must be rather subtle. No prominent gene expression profiles predicting patient's response to neoadjuvant chemotherapy emerge from the neoadjuvant studies carried out to date, including this interim report presented here. The correlation of gene expression profiles with response to neoadjuvant chemotherapy may currently represent the most feasible and powerful approach to identify response classifiers. However, relatively large studies will be required to achieve reliable and clinically useful predictive tests.
Authors' Disclosures of Potential Conflicts of Interest
The authors indicated no potential conflicts of interest.
Acknowledgment
We thank B. Weigelt for critical reading of the manuscript.
NOTES
Supported by the Dutch Cancer Society.
Presented in part at the European Breast Cancer Conference (Hamburg, Germany, March 2004) and at the 40th Annual Meeting of the American Society of Clinical Oncology, New Orleans, LA, June 5-8, 2004.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Schwartz GF, Hortobagyi GN: Proceedings of the consensus conference on neoadjuvant chemotherapy in carcinoma of the breast, April 26-28, 2003, Philadelphia, Pennsylvania. Cancer 100:2512-2532, 2004
Pierga JY, Mouret E, Laurence V, et al: Prognostic factors for survival after neoadjuvant chemotherapy in operable breast cancer: The role of clinical response. Eur J Cancer 39:1089-1096, 2003
Ferriere JP, Assier I, Cure H, et al: Primary chemotherapy in breast cancer: Correlation between tumor response and patient outcome. Am J Clin Oncol 21:117-120, 1998
Mauriac L, Durand M, Avril A, et al: Effects of primary chemotherapy in conservative treatment of breast cancer patients with operable tumors larger than 3 cm: Results of a randomized trial in a single centre. Ann Oncol 2:347-354, 1991
Fisher B, Bryant J, Wolmark N, et al: Effect of preoperative chemotherapy on the outcome of women with operable breast cancer. J Clin Oncol 16:2672-2685, 1998
Kuerer HM, Newman LA, Smith TL, et al: Clinical course of breast cancer patients with complete pathologic primary tumor and axillary lymph node response to doxorubicin-based neoadjuvant chemotherapy. J Clin Oncol 17:460-469, 1999
Wolmark N, Wang J, Mamounas E, et al: Preoperative chemotherapy in patients with operable breast cancer: Nine-year results from National Surgical Adjuvant Breast and Bowel Project B-18. J Natl Cancer Inst Monogr 30:96-102, 2001
Bonadonna G, Valagussa P, Brambilla C, et al: Primary chemotherapy in operable breast cancer: Eight-year experience at the Milan Cancer Institute. J Clin Oncol 16:93-100, 1998
Ravdin PM: Docetaxel (Taxotere) for the treatment of anthracycline-resistant breast cancer. Semin Oncol 24:S10, 1997 (suppl 10)
Moliterni A, Tarenzi E, Capri G, et al: Pilot study of primary chemotherapy with doxorubicin plus paclitaxel in women with locally advanced or operable breast cancer. Semin Oncol 24:S17, 1997 (suppl 17)
Holmes FA, Walters RS, Theriault RL, et al: Phase II trial of Taxol, an active drug in the treatment of metastatic breast cancer. J Natl Cancer Inst 83:1797-1805, 1991
Bissery MC, Vrignaud P, Lavelle F: Preclinical profile of docetaxel (Taxotere): Efficacy as a single agent and in combination. Semin Oncol 22:3-16, 1995
Sjostrom J, Blomqvist C, Mouridsen H, et al: Docetaxel compared with sequential methotrexate and 5-fluorouracil in patients with advanced breast cancer after anthracycline failure: A randomised phase III study with crossover on progression by the Scandinavian Breast Group. Eur J Cancer 35:1194-1201, 1999
Nabholtz JM, Senn HJ, Bezwoda WR, et al: Prospective randomized trial of docetaxel versus mitomycin plus vinblastine in patients with metastatic breast cancer progressing despite previous anthracycline-containing chemotherapy: 304 Study Group. J Clin Oncol 17:1413-1424, 1999
Nabholtz JM, Falkson C, Campos D, et al: Docetaxel and doxorubicin compared with doxorubicin and cyclophosphamide as first-line chemotherapy for metastatic breast cancer: Results of a randomized, multicenter, phase III trial. J Clin Oncol 21:968-975, 2003
Perou CM, Sorlie T, Eisen MB, et al: Molecular portraits of human breast tumours. Nature 406:747-752, 2000
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
Pusztai L, Ayers M, Stec J, et al: Gene expression profiles obtained from fine-needle aspirations of breast cancer reliably identify routine prognostic markers and reveal large-scale molecular differences between estrogen-negative and estrogen-positive tumors. Clin Cancer Res 9:2406-2415, 2003
van de Vijver MJ, He YD, van't Veer LJ, et al: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002
Ayers M, Symmans WF, Stec J, et al: Gene expression profiles predict complete pathologic response to neoadjuvant paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide chemotherapy in breast cancer. J Clin Oncol 22:2284-2293, 2004
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
Weigelt B, Glas AM, Wessels LF, et al: Gene expression profiles of primary breast tumors maintained in distant metastases. Proc Natl Acad Sci U S A 100:15901-15905, 2003
R Foundation for Statistical Computing: R: A Language and Environment for Statistical Computing. Vienna, Austria, 2003
Yang YH, Dudoit S, Luu P, et al: Normalization for cDNA microarray data: A robust composite method addressing single and multiple slide systematic variation. Nucleic Acids Res 30:e15, 2002
Hughes TR, Marton MJ, Jones AR, et al: Functional discovery via a compendium of expression profiles. Cell 102:109-126, 2000
Sturn A, Quackenbush J, Trajanoski Z: Genesis: Cluster analysis of microarray data. Bioinformatics 18:207-208, 2002
Khan J, Wei JS, Ringner M, et al: Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med 7:673-679, 2001
Simon R, Radmacher MD, Dobbin K, et al: Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J Natl Cancer Inst 95:14-18, 2003
Ambroise C, McLachlan GJ: Selection bias in gene extraction on the basis of microarray gene-expression data. Proc Natl Acad Sci U S A 99:6562-6566, 2002
Bhattacharyya C, Grate L, Rizki A, et al: Simultaneous classification and relevant feature identification in high-dimensional spaces: Application to molecular profiling data. Signal Processing 83:729-743, 2003
Buchholz TA, Stivers DN, Stec J, et al: Global gene expression changes during neoadjuvant chemotherapy for human breast cancer. Cancer J 8:461-468, 2002(Juliane Hannemann, Hendri)