|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Originally published as JCO Early Release 10.1200/JCO.2004.05.166 on May 10 2004 © 2004 American Society of Clinical Oncology. Gene Expression Profiles Predict Complete Pathologic Response to Neoadjuvant Paclitaxel and Fluorouracil, Doxorubicin, and Cyclophosphamide Chemotherapy in Breast CancerFrom Millennium Pharmaceuticals Inc, Cambridge, MA; and the Departments of Pathology, Biostatistics, Breast Medical Oncology, and Diagnostic Radiology, The University of Texas M.D. Anderson Cancer Center, Houston, TX Address reprint requests to Lajos Pusztai, MD, Department of Breast Medical Oncology, The University of Texas M.D. Anderson Cancer Center, Unit 424, 1515 Holcombe Blvd, Houston, TX 77030-4009; e-mail: lpusztai{at}mdanderson.org or Andrew.Damokosh{at}mpi.com
PURPOSE: The goal of this study was to examine the feasibility of developing a multigene predictor of pathologic complete response (pCR) to sequential weekly paclitaxel and fluorouracil + doxorubicin + cyclophosphamide (T/FAC) neoadjuvant chemotherapy regimen for breast cancer. PATIENTS AND METHODS: All patients underwent one-time pretreatment fine-needle aspiration to obtain RNA from the cancer for transcriptional profiling using cDNA arrays containing 30,721 human sequence clones. Analysis was performed after profiling, and 42 patients' clinical results were available, 24 of which were used for predictive marker discovery; 18 patients' results were used as an independent validation set.
RESULTS: Thirty-one percent of patients had pCR (six discovery and seven validation), defined as disappearance of all invasive cancer in the breast after completion of chemotherapy. We could identify no single marker that was sufficiently associated with pCR to be used as an individual predictor. A multigene model with 74 markers (P CONCLUSION: Our results suggest that transcriptional profiling has the potential to identify a gene expression pattern in breast cancer that may lead to clinically useful predictors of pCR to T/FAC neoadjuvant therapy.
Postoperative, adjuvant chemotherapy is widely used in the treatment of patients with stage I-III breast cancer because it improves overall survival.1,2 There are multiple combinations of cytotoxic drugs currently accepted as standard of care. The most effective combination regimens include anthracyclines, such as doxorubicin or epirubicine, which are topoisomerase II inhibitors. Paclitaxel, a microtubule stabilizer agent, has a different mechanism of action compared with anthracyclines and often produces tumor shrinkage in patients with advanced breast cancer with tumor cells that are resistant to anthracyclines.3 Therefore, it is hoped that the addition of paclitaxel to anthracycline-containing adjuvant chemotherapy will improve treatment efficacy for a subset of patients with breast cancer. Preliminary results of at least two large, randomized clinical trials support this hypothesis.4 A major clinical challenge is to identify the subset of patients, at the time of diagnosis, who benefit from these more prolonged, often more toxic, and more expensive regimens. Currently, there is no clinically useful molecular predictor of response to any cytotoxic drug used in the treatment of breast cancer.5 Clinical parameters such as tumor size, estrogen or HER-2 receptor status, histologic or nuclear grade, or the expression of single molecular markers (ie, Bcl-2, p53, MDR-1, and so on) show weak association with response and are not regimen-specific, which limits their utility in selecting chemotherapy treatment.6 Chemotherapy is applied empirically despite the observation that all regimens are not equally effective across the population of patients with breast cancer. However, recent technological advances have enabled researchers to scan the expression pattern of thousands of genes in individual tumors, and identify molecular signatures that are predictive of outcome.7-9 Administration of chemotherapy before surgery (neoadjuvant chemotherapy) provides a unique opportunity to identify molecular predictors of response to treatment in breast cancer. Pathologic complete response (pCR), defined as disappearance of all invasive cancer in the breast after completion of neoadjuvant chemotherapy, has been associated with improved long-term, disease-free, and overall survival.10-11 The purpose of this study was to assess whether gene expression patterns in breast cancer, at the time of diagnosis, could predict pCR to neoadjuvant sequential weekly paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide (T/FAC) chemotherapy. Our hypothesis is that a baseline, pretreatment gene expression pattern of breast cancer holds information about response to chemotherapy, and that this information can be extracted by transcriptional profiling, and formalized into a clinical outcome predictor by applying supervised machine learning methods.
Patients and Samples This study was conducted at the Nellie B. Connally Breast Center of The University of Texas M.D. Anderson Cancer Center (MDACC). Exploratory analysis was performed after transcriptional profiling, and clinical results have become available for 42 prospectively enrolled patients who received neoadjuvant chemotherapy. The first 24 patients were used for predictive marker discovery ("training cases"), and the next 18 patients were available as an independent validation set ("validation cases"). Three additional patients were available as validation cases but were excluded from the results because they had received trastuzumab, which is known to alter sensitivity to paclitaxel and anthracycline therapy. All patients underwent a single-pass, pretreatment fine-needle aspiration (FNA) of the primary breast tumor before starting chemotherapy. The aspiration was performed using a 23- or 25-gauge needle. Cells for cDNA array analysis were collected into vials containing RNAlater solution (Ambion, Austin, TX) and stored at –80°C. Every patient underwent surgery with axillary node sampling after completion of chemotherapy. Approximately 80% of the patients had T2 or T3 tumors, 57% were estrogen receptor–positive, and 17% had HER-2 amplification. The majority of the patients (88%) received weekly paclitaxel in 12 courses (x 12) followed by four courses of FAC. Two patients (10%) in each group had 3-weekly paclitaxel followed by FAC. One patient (2%) in the validation group received only weekly paclitaxel x 12. Clinical characteristics of the participants are presented in Table 1. At the completion of neoadjuvant chemotherapy, all patients had surgical resection of the tumor bed, with negative margins. Grossly visible residual cancer was measured, and representative sections were submitted for histopathologic study. Metallic markers had been placed under radiologic guidance in the shrinking tumor bed for any patient whose tumor was noticeably responding (< 1 cm residual cancer by imaging) during the course of treatment. Those metallic markers confirmed the site of the tumor bed in the specimen radiographs. When there was not grossly visible residual cancer, the slices of the specimen were radiographed, and all areas of radiologically and/or architecturally abnormal tissue were entirely submitted for histopathologic study. pCR was defined as no histopathologic evidence of any residual invasive cancer cells in the breast, whereas residual disease was defined as any residual cancer cells after histopathologic study. This study was approved by the institutional review board of MDACC, and all patients signed an informed consent for voluntary participation.
cDNA Array Hybridization RNA was extracted from a FNA sample using the RNAeasy Kit (Qiagen, Valencia, CA). The amount and quality of RNA was assessed with DU-640 UV Spectrophotometer (Beckman Coulter, Fullerton, CA), 1% Seakem LE agarose gel (Biowhittaker Molecular Applications, Rockland, ME), and Agilent 2100 Bioanalyzer RNA 6000 LabChip kit (Agilent Technologies, Palo Alto, CA). First strand cDNA synthesis was performed with Superscript II (Invitrogen, Carlsbad, CA) in the presence of 33P-dCTP (100 mCi/mL; Amersham, Little Chalfont, UK) from 1 to 2 µg total RNA. The generated cDNA probes were hybridized (without further amplification) to high-density nylon cDNA arrays as previously described.12 The arrays contain 30,721 human sequence clones. The cellular composition of the FNA samples, the RNA yield and hybridization success has been reported previously.13,14 In brief, FNA samples, on average, contain 80% neoplastic cells; the rest of the cells are infiltrating leukocytes. These samples contain no or little amounts of stromal cells (fibroblast, adipocyes) or normal breast epithelium. In our experience, 75% of single-pass FNA samples were suitable for cDNA array profiling.
Data Analysis To identify individual informative genes with respect to pCR versus lesser response, we explored several methods, including the nonparametric Wilcoxon rank sum test, two-sample independent t tests, and gene ranking by signal-to-noise ratio (SNR), all with bootstrap adjustment for multiple comparisons. Hierarchical clustering was performed using rank order correlation for the similarity metric, and complete linkage was performed for clustering to visualize the discriminating power of these genes (Spotfire Decisionsite 7.0; Spotfire, Somerville, MA). For multigene model discovery, we further filtered the data by selecting the markers with the highest SNR scores in an effort to reduce the computational resource required for the model discovery process without excluding potentially informative genes. SNR was selected because of its long-standing history and popularity in the microarray literature and its interpretation as a discrimination score.17 This resulted in a data set of 500 markers. There was a greater than 70% agreement with a data set comprised of markers whose t test scores had raw P < .05, which provided additional evidence that we had included the most informative markers. Multigene models of sizes up to and including 100 genes were built using a combination of two feature selection methods (SNR and support vector machine [SVM] feature selection [SVM-FS]), two class prediction algorithms (k nearest neighbors [k-NN, with k = 5 and Euclidian distance metric]), and an SVM with quadratic polynomial kernel.18 SVM-FS is a feature selection method developed as an extension to the SVM classifier.19 SVM is a maximum margin classifier that maximizes a hyperplane between two classes while simultaneously minimizing the number of misclassified training samples. SVM-FS incorporates weights to the kernel function of the SVM classifier, providing a method for ranking genes by checking the term corresponding to a given gene to assess its contribution to maximizing the margin between the two classes. To assess model performance, we used a combination of cross-validation and permutation testing. We used four-fold cross-validation repeated 20 times (for improved estimation accuracy) to estimate classification error and subsequently to rank the multigene models. During this procedure, the training set was randomly split into four equal subsets of approximately six cases each. Model estimation was conducted on three fourths of the data, and tested on the remaining fourth of the cases that were held out. This was repeated until each group of a fourth of the cases played the role of test data once. This process was then repeated 20 times, with the average representing the classification error for the given model. It should be emphasized that this method of cross validation to estimate classification error can only be used as a criterion for discriminating among the candidate models. It is not an unbiased estimate of classification error that one expects to observe when tested on an independent validation set. For this purpose, we set aside a validation data set of 18 patients. To assess if the classification error rates differ significantly from what chance alone could produce for a given classifier or feature selection combination, a permutation test was applied to the model discovery process. In a permutation test, outcome labels are randomly assigned to each patient, and the entire model discovery process is repeated, as outlined beginning with the 19,813 clones, using the permuted data set. One hundred such permuted data sets were produced from which we generated a random distribution of classification error rates. The error rates from the observed data were then compared with the distribution of random error rates to assess the probability of observing such an error rate by chance alone (P value). Finally, the top prediction model was defined as the most parsimonious multigene predictor with the lowest P value. The reason for selecting the most parsimonious model is to minimize the potential effects of "overfitting," (ie, the inclusion of genes with no biologically relevant information). Increasing the number of genes in a model often improves predictive accuracy in the training set, but including too many genes may lead to overfitting. The result of this is a model that has been partly (proportional to the number of irrelevant genes) trained on biologically irrelevant data, the characteristics of which are unlikely to be duplicated in the validation data. This leads to higher classification error rate in the independent data than might have occurred if the model excluded the irrelevant genes. The top model(s) was tested on an independent data set of 18 patients, and prediction results were presented using standard metrics (sensitivity, specificity, and so on) with 95% binomial CIs. The model discovery process and validation schema are depicted graphically in Figure 1.
Development of Multigene Predictor pCR in the breast was achieved by 13 (31%) of the 42 patients—a response rate that is consistent with our previous experience in a larger randomized study using similar preoperative therapy.20 Six cases of pCR were included in the training set, and seven cases, in the validation set (Table 1). Of 19,183 genes or any single gene examined, we could not identify any single clinical variable that was statistically significantly associated with pCR after bootstrap (or Bonferroni) adjustment for multiple comparisons in the 24 training cases. Hierarchical clustering of all 19,813 genes using the training data showed no separation of the complete responders from the incomplete responders (data not shown). This suggests that the information about response to therapy may be encoded in a relatively small set of genes compared with the entire transcriptome. Hierarchical clustering provides a visual display of the similarities between gene expression profiles, but it is not an appropriate method to predict outcome in new cases. We therefore examined multigene predictors of pathologic response using statistical learning algorithms. These learning algorithms are trained on cases with known outcome to formulate classification rules that connect gene expression profiles to patient outcome. We trained our predictors on the first 24 cases, six of which were pCR. Several commonly used algorithms were tested in combination with the informative genes to estimate the general performance of predictors and to identify the best predictor for validation on independent cases. The two prediction algorithms performed similarly in repeated four-fold cross-validation in the training set. Genes identified by the SVM-FS algorithm as the most informative produced the best cross-validation rates when used in combination with a k-NN (k = 5) classification algorithm. Figure 2 shows the change in classification error rate relative to the included number of markers ordered by SVM-FS for the k-NN classifier. The figure also shows the P values for each model from the permutation test. The results show that as the number of genes increases in the model the misclassification error rate declines. When the number of genes in the model increases above 70, the probability that chance alone could produce similar low error rates also drops below 10%. Based on these results, we selected a 74-gene k-NN model as our most promising predictor. The 74 informative genes and gene ontology is presented in Table 2. 21 The list includes 51 known genes (68%) and 23 unknown genes (32%).
Validation on Independent Cases We tested our 74-gene model on a validation data set (n = 18) that was excluded from the model discovery process to estimate expected true predictive accuracy. We expected that a robust learning-based method would effectively recognize expression patterns that are similar to those that it was trained on, and therefore, would show high positive predictive value (ie, when the model predicts pCR, the outcome will indeed be pCR). Conversely, since the model was only trained on 24 cases, it is reasonable to assume that there may be many possible expression profiles that are associated with pCR that were not included in our training data; thus, the model may have low sensitivity (ie, miss many cases of pCR). In our 74-gene model, overall prediction accuracy was 78% (14 of 18; 95% CI, 52% to 94%; Table 3). All three of the patients predicted to have pCR had a complete pathological response, a positive predictive value of 100% (95% CI, 29% to 100%). Negative predictive value was 73% (95% CI, 45% to 92%), sensitivity was 43% (95% CI, 10% to 82%), and specificity was 100% (95% CI, 72% to 100%). We also calculated prediction accuracy for two other models that performed comparably in repeated four-fold cross validation in the training data: a 29-gene SVM model using SNR feature selection, and a 49-gene SVM model using SVM-FS. Somewhat worse prediction accuracy was observed for both of these models in the validation set than for what seen with the SVM-FS–kNN model. For all three models, the average prediction accuracy was 69% (range, 61% to 78%) and is the most reliable estimate of expected prediction accuracy for unobserved cases.
The goal of this study was to examine the feasibility of developing a multigene predictor of pCR to a complex chemotherapy regimen and to estimate the predictive accuracy of such a test. This marker discovery trial was conducted prospectively and used cDNA array technology to identify potential molecular markers of pCR to sequential weekly paclitaxel/FAC neoadjuvant chemotherapy. We have chosen pCR as our clinical end point because retrospective analyses of several large clinical studies have shown that complete eradication of the cancer from the breast after neoadjuvant therapy is associated with improved long-term, disease-free, and overall survival.10,11 pCR is currently considered to be the best, though imperfect, early surrogate for cure after neoadjuvant therapy. The association between favorable long-term outcome and clinical complete response or clinical partial response is less robust. We propose that a test that could predict greater than average probability of pCR to one particular regimen has clinical importance because several different adjuvant or neoadjuvant chemotherapy regimens are currently used, and not all of these regimens are likely equally effective for all patients. The chemotherapy regimen that is most likely to induce pathologic CR as neoadjuvant therapy is also likely to be the regimen that will improve cure rates the most when given as adjuvant treatment. This concept is examined by a large randomized study, NSABP-B27, and a definitive answer regarding the value of pCR as early surrogate of long-term benefit will become available shortly. Using the first 24 cases for marker discovery, we identified differentially expressed genes between the two response groups (pCR v residual disease). Our analysis revealed that no single marker from the 19,813 clones was significantly associated with pathologic response. We applied machine-learning algorithms to discover a multigene model for response prediction. By assessing the expression of several markers simultaneously, we could predict pathologic response in the primary tumor remarkably accurately in a small independent cohort of patients. A 74-gene model selected by SVM-FS and using a k-NN (k = 5) classifier resulted in a prediction accuracy of pathologic response (pCR v no-pCR) of 78% in a validataion data set of 18 cases. In particular, the model showed a high positive predictive value of 100% (three of three patients predicted to have pCR indeed experienced pCR). The probability of any unselected patient with breast cancer to experience pCR after any of the best available neoadjuvant chemotherapy regimens has been shown to be no greater than 25% to 30%.20,22,23 A test that could predict a significantly greater than average chance of pCR to the T/FAC regimen would be of clinical value and could assist in selecting the most effective therapy for an individual. We believe that a test with a positive predictive value of 60%, indicating that test-positive individuals have twice as high probability for pCR than unselected patients, could be clinically useful for positive selection of a chemotherapy regimen. Our result suggests that this degree of accuracy may be a realistic expectation for multigene predictors. The development of multigene predictors for clinical use is a process, somewhat similar to the development of new therapeutic agents. The feasibility of the method needs to be tested; the accuracy of the test has to be estimated; the predictor will have to be optimized; and finally, the clinical utility needs to be proven in sequential series of clinical studies analogous to phase I, II, and III clinical trials. In the context of this process, we consider our results as encouraging positive findings from a "phase I" marker discovery study. Our current predictor needs to be optimized, particularly to improve sensitivity and to better define the true predictive accuracy with narrower CIs. Based on the characteristics of learning-based predictors, we expect that a second-generation predictor trained on a larger training set of cases will have improved accuracy, and a larger training set will help to better estimate predictive accuracy.16 It is provocative to speculate on the biologic function of the 74 genes that form our current predictor. However, many of these genes need not have a proximal role in the biologic function that they predict. They may be robust but distant downstream transcriptional effects of biologic events that influence drug sensitivity. Furthermore, informative gene lists can change substantially as the training set size from which they are generated increases. The rank order of genes is particularly susceptible to change from one list to another. Therefore, from the vantage point of gaining mechanistic insight into the biology of chemotherapy sensitivity or resistance, these results should be regarded as hypothesis-generating only. However, it was encouraging to see that several of the genes in our marker set of 74 have been shown previously to be involved with chemotherapy response. DD96 (epithelial protein upregulated in carcinoma, or membrane-associated protein 17) is associated with poor response to neoadjuvant T/FAC chemotherapy. This gene has been reported to be expressed in breast tumors, and it forms a protein cluster with the multidrug resistance protein (MRP2) and PDZK1, suggesting that DD96 may play a role in multidrug resistance to cytotoxic therapy.24,25 We also identified markers involved with cellular differentiation (KRT13) and migration (NFAT5) to be associated with poor response.26 Increased expression of genes involved in lipid metabolism, ACLY (ATP citrate lyase), FAAH (fatty acid amide hydrolase), and APOE (apolipoprotein E), were also associated with poor response, possibly providing the tumor cells with an energy resource for growth. We identified several known genes positively associated with pCR. One of these was FCGR3A (Fc fragment of IgG, low-affinity IIIa, receptor) a proinflammatory molecule often expressed on the surface of natural killers (NK) cells.27 Recently it has been shown that paclitaxel can induce expression of a wide variety of inflammatory and anti-inflammatory genes similar to those induced by lipopolysaccharide.28 This pro-inflammatory effect of paclitaxel could possibly augment anitumor activity in vivo. KPNA2 (karyopherin alpha 2) is another gene associated with pCR that functions as a nuclear transporter of proteins and has been reported to be involved with the microtubule assembly process, suggesting that it may also have a role in determining the sensitivity of tumor cells to microtubule stabilizing agents.29 Other markers associated with pCR are involved in the negative regulation of cell proliferation including NME1 (nonmetastatic cells 1, protein), NME2 (nonmetastatic cells 2, protein), and CREM (cAMP responsive element modulator). Additional experiments will be needed to determine the exact role of each marker and pathways that contribute to the complex phenotype associated with drug sensitivity and resistance. The potential clinical implications of our results are important. Currently, the only chance to improve survival for breast cancer patients is timely, aggressive, and optimal therapy at the time of diagnosis. Any improvement in selecting patients who have a better than average chance to benefit from a given chemotherapy regimen is an important improvement over the current unselected empirical use of various adjuvant chemotherapy regimens. Our results suggest that gene expression profiling may be developed into a diagnostic tool to identify patients who experience pCR after T/FAC chemotherapy. Our goal is to develop a second-generation more sensitive predictor by using a larger number of cases for training. It is also imperative to build a portfolio of predictors for various individual regimens each of which could be tested on the gene expression profile of newly diagnosed breast cancer to select the optimal neoadjuvant regimen that has the greatest probability to induce pCR with the least toxicity and cost.
The following authors or their immediate family members have indicated a financial interest. No conflict exists for drugs or devices used in a study if they are not being evaluated as part of the investigation. Owns stock (not including shares held through a public mutual fund): M. Ayers, J. Stec, A. Damokosh, E. Clark, J. Ross, J. Metivier, Millenium Pharmaceuticals. Served as an officer or member of the Board of a company: M. Ayers, J. Stec, A. Damokosh, E. Clark, J. Ross, J. Metivier, Millenium Pharmaceuticals.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
1. Carlson RW, Anderson BO, Bensinger W, et al: NCCN practice guidelines for breast cancer. Oncology 14:33-49, 2000 2. National Institutes of Health Consensus Development Conference, November 2, 2000. http://www.consensus.nih.gov 3. Rivera E, Holmes FA, Frye D, et al: Phase II study of paclitaxel in patients with metastatic breast carcinoma refractory to standard chemotherapy. Cancer 89:2195-2201, 2000[CrossRef][Medline]
4. Henderson IC, Berry D, Demetri G, et al: Improved outcomes from adding sequential paclitaxel but not from escalating doxorubicin dose in an adjuvant chemotherapy regimen for patients with node-positive primary breast cancer. J Clin Oncol 21:976-983, 2003
5. Bast RC, Ravdin P, Hayes DF, et al: 2000 Update of recommendations for the use of tumor markers in breast and colorectal cancer: Clinical practice guidelines of the American Society of Clinical Oncology. J Clin Oncol 19:1865-1878, 2001 6. Hortobagyi GN, Hayes D, Pusztai L: Integrating newer science into breast cancer prognosis and treatment: Molecular predictors and profiles. American Society of Clinical Oncology 2002 Annual Meeting Summaries 191-202, 2002 7. van’t Veer LJ, Dai H, van de Vijver, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536, 2002[CrossRef][Medline] 8. 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[CrossRef][Medline] 9. Shipp MA, Ross KN, Tamayo P, et al: Diffuse large B-cell lymphoma outcome prediction by gene expression profiling and supervised machine learning. Nat Med 8:68-74, 2002[CrossRef][Medline] 10. 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[Abstract]
11. 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
12. Chiang LW, Grenier JM, Ettwiller L, et al: An orchestrated gene expression component of neuronal programmed cell death revealed by cDNA array analysis. Proc Natl Acad Sci U S A 98:2814-2819, 2001
13. Pusztai L, Ayers M, Stec J, et al: Gene expression profiles obtained from single passage fine needle aspirations (FNA) of breast cancer reliably identify prognostic/predictive markers such as estrogen (ER) and HER-2 receptor status and reveal large scale molecular differences between ER-negative and ER-positive tumors. Clin Cancer Res 9:2406-2415, 2003 14. Symmans WF, Pusztai L, Ayers M, et al: Total RNA yield and microarray gene expression profiles from fine needle aspiratons and core needle biopsy samples of breast cancer. Cancer 97:2960-2971, 2003[CrossRef][Medline]
15. Hwang D, Schmitt WA, Stephanopoulos G, et al: Determination of minimum sample size and discriminatory expression patterns in microarray data. Bioinformatics 18:1184-1193, 2002 16. Mukherjee S, Tamayo P, Rogers S, et al: Estimating dataset size requirements for classifying DNA microarray data. J Comput Biol 10:119-142, 2003[CrossRef][Medline]
17. Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class predication by gene expression monitoring. Science 286:531-537, 1999 18. Vapnik VN: Statistical Learning Theory. New York, NY, Wiley, 1998 19. Mukherjee S, Tamayo P, Slonim D, et al: (2000) Support vector machine classification of micro array data. MIT Report AI Laboratory and Center for Biological Computational Learning. AI Memo NO 1667 CBCL, Paper No. 82 20. Green MC, Buzdar AU, Smith T, et al: Weekly paclitaxel followed by FAC in the neoadjuvant setting provides improved pathological complete remission rates compared to standard paclitaxel followed by FAC therapy: Preliminary results of an ongoing prospective randomized trial. Proc Am Soc Clin Onc 20:33a, 2001 (abstr 129) 21. NCBI LocusLink Web site, Gene symbol and ontology, http://www.ncbi.nlm.nih.gov/LocusLink 22. NSABP: The effect on primary tumor response of adding sequential Taxotere to Adriamycin and cyclophosphamide: Preliminary results from NSABP Protocol B-27. Breast Cancer Res Treat 69:210, 2001 (abstr 5)
23. Smith IC, Heys SD, Hutcheon AW, et al: Neoadjuvant chemotherapy in breast cancer: Significantly enhanced response with docetaxel. J Clin Oncol 20:1456-1466, 2002 24. Kocher O, Comella N, Gilchrist A, et al: PDZK1, a novel PDZ domain-containing protein up-regulated in carcinomas and mapped to chromosome 1q21, interacts with cMOAT (MRP2), the multidrug resistance-associated protein. Lab Invest 79:1161-1170, 1999[Medline] 25. Kocher O, Cheresh P, Lee SW, et al: Identification and partial characterization of a novel membrane-associated protein (MAP17) up-regulated in human carcinomas and modulating cell replication and tumor growth. Am J Pathol 149:493-500, 1996[Abstract] 26. Jauliac S, Lopez-Rodriguez C, Shaw LM, et al: The role of NFAT transcription factors in integrin-mediated carcinoma invasion. Nat Cell Biol 4:540-544, 2002[CrossRef][Medline]
27. Anderson P, Caligiuri M, O'Brien C, et al: Fc-gamma receptor type III (CD16) is included in the zeta NK receptor complex expressed by human natural killer cells. Proc Natl Acad Sci U S A 87:2274-2278, 1990 28. Zaks-Zilberman M, Zaks TZ, Vogel SN: Induction of proinflammatory and chemokine genes by lipopolysaccharide and paclitaxel (Taxol) in murine and human breast cancer cell lines. Cytokine 15:156-165, 2001[CrossRef][Medline]
29. Oliver J, Gruss, Rafael E, et al: Ran induces spindle assembly by reversing the inhibitory effect of importin Submitted May 23, 2003; accepted February 2, 2004. Related Editorial
This article has been cited by other articles:
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
TOP
ABSTRACT
INTRODUCTION
.09) was built using data from the discovery samples and tested on the validation samples. Overall, a 78% (14 of 18) predictive accuracy was observed, with a 100% (three of three) positive predictive value for pCR, a 73% (11 of 15) negative predictive value, a sensitivity of 43% (three of seven), and a specificity of 100% (11 of 11). The expected response rate to T/FAC neoadjuvant therapy in unselected patients is 28%. 
on TPX2 activity cell. 104:83-93, 2001
