|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Journal of Clinical Oncology, Vol 24, No 19 (July 1), 2006: pp. 3039-3047 © 2006 American Society of Clinical Oncology. DOI: 10.1200/JCO.2006.05.6564 Novel Prognostic Immunohistochemical Biomarker Panel for Estrogen Receptor–Positive Breast Cancer
From Applied Genomics Inc, Burlingame, CA; Applied Genomics Inc; Comprehensive Cancer Institute of Huntsville, Huntsville, AL; Cleveland Clinic Foundation, Cleveland, OH; and Genetic Pathology Evaluation Centre, University of British Columbia, Vancouver, British Columbia, Canada Address reprint requests to Douglas T. Ross, MD, PhD, Applied Genomics Inc, 863 Mitten Rd #103, Burlingame, CA; e-mail: dross{at}applied-genomics.com
PURPOSE: Patients with breast cancer experience progression and respond to treatment in diverse ways, but prognostic and predictive tools for the oncologist are limited. We have used gene expression data to guide the production of hundreds of novel antibody reagents to discover novel diagnostic tools for stratifying carcinoma patients. PATIENTS AND METHODS: One hundred forty novel and 23 commercial antisera, selected on their ability to differentially stain tumor samples, were used to stain paraffin blocks from a retrospective breast cancer cohort. Cox proportional hazards and regression tree analysis identified minimal panels of reagents able to predict risk of recurrence. We tested the prognostic association of these prospectively defined algorithms in two independent cohorts. RESULTS: In both validation cohorts, the Kaplan-Meier estimates of recurrence confirmed that both the Cox model using five reagents (p53, NDRG1, CEACAM5, SLC7A5, and HTF9C) and the regression tree model using six reagents (p53, PR, Ki67, NAT1, SLC7A5, and HTF9C) distinguished estrogen receptor (ER) –positive patients with poor outcomes. The Cox model was superior and distinguished patients with poor outcomes from patients with good or moderate outcomes with a hazard ratio of 2.21 (P = .0008) in validation cohort 1 and 1.88 (P = .004) in cohort 2. In multivariable analysis, the calculated risk of recurrence was independent of stage, grade, and lymph node status. A model proposed for ER-negative patients failed validation in the independent cohorts. CONCLUSION: A panel of five antibodies can significantly improve on traditional prognosticators in predicting outcome for ER-positive breast cancer patients.
Management of breast cancer has been challenging because it is a biologically, histologically, and clinically heterogeneous disease. The most widely used algorithms for determining prognosis rely on tumor size, histologic grade, and extent of metastatic disease.1-3 Treatment guidelines combine this information with estrogen receptor (ER) and erbB2 expression status to stratify patients among surgery, chemotherapeutic, and radiation therapy options.4-6 However, individual risk assessment for early-stage patients is inadequate, and better identification of patients with a high risk of recurrence would help to make more informed decisions about treatment options.7,8 Microarray-based gene expression studies have brought to light some of the biologic diversity of breast cancer and have been used to identify gene expression signatures associated with outcome. A number of different statistical approaches have been used to identify minimal gene sets for prognostication or to predict response to therapy.9-14 Because of the wealth of gene expression studies, clinical development of gene expression–based assays has made some impressive initial progress, but it suffers the inherent limitation that the technology as a clinical tool will likely require specialized laboratories for quality assurance. The most clinically advanced test measures the expression of 21 genes in paraffin-embedded tumor material and uses a Cox proportional hazards model to predict risk of recurrence in early-stage breast carcinoma patients.15 The clinical introduction of these tests has generated much controversy concerning the degree of validation required for clinical use, but there is consensus that, if validated in independent cohorts across different institutions and patient populations, tests using multiple markers will help with better management of patients. Once prognostic gene signatures are identified, it is reasonable to conjecture that accurate measurement of a few of the genes would suffice to classify patients in clinical assays. We explored whether the insights from gene expression studies could be translated into robust immunohistochemistry (IHC) -based assays by generating hundreds of novel antibody reagents targeted to predicted protein products of genes selected on the basis of their gene expression patterns. We screened these reagents for utility in IHC and selected a panel of diversity composed of antisera that can be used with paraffin-embedded tissue to explore classification of carcinoma at the protein level. In this report, we used this panel of IHC antisera to identify a subset of only five or six reagents whose staining results can be combined using Cox or regression tree models, respectively, for prognostication of breast cancer patients. We validated the prognostic association of these tests by prespecifying the scoring system and Cox and tree model coefficients and prognostic category cutoff values, then by testing them on two independent breast cancer cohorts.
Patients Comprehensive Cancer Institute of Huntsville (CCIH). Seven hundred twenty patients with primary invasive breast cancer seen at the CCIH between 1989 and 2002 were assembled into three tissue array blocks. Inclusion in these studies required that follow-up information be available for the patient and that sufficient excess material be available in one or more paraffin blocks such that follow-up clinical studies would be possible for patient care as clinically indicated. Clinical follow-up data were obtained by chart reviews for 466 patients and entered into an institutional review board–approved database cross referenced to the arrayed patient cases through an anonymous unique identifier. Cleveland Clinic Foundation (CCF). Two hundred ninety-nine consecutive primary invasive breast cancer patients seen at CCF between 1995 and 1996 were arrayed on a single tissue array. A duplicate block was constructed using independent cores from the same patient set. Each tissue array patient was assigned a unique identification that was then linked to a CCF institutional review board–approved database containing 5-year clinical follow-up information.16 British Columbia Cancer Agency (BCCA). Three hundred forty-four consecutive primary invasive breast cancer patients seen at Vancouver General Hospital between 1974 and 1995 (mean follow-up time, 11.7 years) were arrayed as a set of three tissue arrays. Sections were stained, and duplicate cores for each patient were reported as consensus scores. Each tissue array patient was assigned an anonymous unique identification that was then linked to an anonymous BCCA clinical research ethics board–approved database containing follow-up information.17
Methods
Prognostic Panel Scoring
Cohorts Three different cohorts were used in this study (Table 1). The discovery cohort was assembled from patients seen at the CCIH between 1989 and 2002. Validation data set 1 was assembled from 229 consecutive patients diagnosed at the CCF in 1995 and 1996 (follow-up at 5 years), and validation set 2 was assembled from 344 consecutive patients seen at Vancouver General Hospital between 1974 and 1995 (mean follow-up time, 11.7 years).16,17 The time to disease recurrence and the type of treatment stratified by ER and nodal status were quite similar between the CCIH and CCF cohorts (data not available for the BCCA study). The BCCA cohort documented more deaths compared with the CCIH cohort over the course of the study because of, in part, the greater follow-up time available for this study (data not shown).
Derivation of the Panel of Diversity Close to 700 gene targets for generation of polyclonal affinity-purified antipeptide antisera were chosen on the basis of interesting gene expression patterns in published data sets.12,18-26 We used an iterative process that subjectively assessed both the staining characteristics of the antisera and whether they identified two discreet populations in the CCIH cohort to select the 126 highest quality reagents, termed a panel of diversity, to stain and characterize clinical cohorts. Figure 1 depicts a k-means cluster of the breast cancer gene expression data from Sorlie et al25 in heat-chart form to which 199 antibodies, for which gene expression data were available, are mapped. The four measured categories are as follows: (1) total antibodies, meaning those antibodies that were produced or purchased and subsequently applied to preliminary titering breast tumor tissue; (2) the subset of these antibodies that exhibited apparently specific staining and was thus applied to the training CCIH cohort; (3) prognostic antibodies used in model building; and (4) antibodies in one of the prognostic models reported herein. The bar graph depicts that these panels of diverse antibodies, both those that stain breast tumor tissue and those with an association with tumor recurrence, target much of the biologic diversity of the carcinoma specimens as assessed by gene expression patterns.
Prognosticator Panel for Early-Stage Breast Cancer Patients We derived prognostic models for ER-positive, lymph node–negative patients using the 195 ER-positive, node-negative patients in the CCIH training cohort and the subset of 20 antibodies that had a significant association with outcome in this patient set (Fig 2). Included in this analysis with an equal opportunity to contribute to models were commercially available antibodies tested across the CCIH cohort, of which ER, PR, Ki67, and p53 showed a significant association with early recurrence (data not shown). A Cox proportional hazards model was selected that used four novel antibody reagents (HTF9C, CEACAM5, NDRG1, and SLC7A5) and one commercially available reagent (p53). Ten-fold cross validation suggested that the model was not overfit to the data (P = .019). We similarly selected a regression tree model comprised of three novel reagents (SLC7A5, NAT, and HTF9C) and three commercial reagents (PR, p53, and Ki67), the derivation of which was also upheld by 10-fold cross validation (P = .006).
Traditional prognosticators, such as the Nottingham Prognostic Index (NPI), use clinical and pathologic parameters to identify risk groups. Figure 3 depicts outcomes in the CCIH cohort stratified by the NPI and by the derived Cox and regression tree models. Using the NPI as a prognosticator identified the expected distinct outcome curves using all ER-positive patients (Fig 3A). When the population was limited to the ER-positive, node-negative subgroup, NPI could only classify patients as good or moderate risk, with a relatively equivalent 5-year disease-free survival rate of approximately 90% (Fig 3B). The Cox model identified a group of patients as either poor or moderate outcomes with a 5-year disease-free survival rate of approximately 75% as opposed to patients classified as good with a 5-year disease-free survival rate of approximately 95% (P < .001; Fig 3C). The tree model identified a group of patients with poor outcomes with a 5-year disease-free survival rate of approximately 77%, whereas good and moderate outcomes were similar at 5 years with disease-free survival rates between 90% and 95% (P < .001; Fig 3D).
A Cox model was also derived using all ER-negative patients and the 21 antisera significant by univariate analysis to design a recurrence predictor. Robust separation of good, moderate, and poor curves was derived using three antisera (HTF9C, SLC7A11, and AKR1C1; P < .001). This algorithm missed significance after 10-fold cross validation (P < .065) and failed independent validation (data not shown).
Validation in Independent Cohorts
In both the BCCA and CCF cohorts, the fraction of ER-positive patients without metastases was too small to test the significance of these models. Instead, we tested the power of the models using all ER-positive patients. In the CCF data set, the Cox model identified poor patients with a 5-year disease-free survival rate of 50% compared with approximately 70% for patients classified as moderate and 87% for patients classified as good (P = .0008; Fig 3E). In the BCCA data set, the Cox model distinguished ER-positive patients classified as poor with overall survival rates of 55% compared with 75% for patients classified as moderate and 90% for patients classified as good (P = .0039; Fig 3G). Importantly, in both cohorts by multivariable analysis, the Cox model was independent of stage, grade, and lymph node status (Table 2). In the combined CCF and BCCA data sets, for the subset of patients who received either a poor or good prognostic call (82%), the sensitivity for a poor prognostic call for predicting disease progression was 36%, whereas its specificity was 88%. The positive predictive value of a poor prognosis call was 38% (95% CI, 32% to 44%), whereas the negative predictive value was 88% (95% CI, 84% to 92%).
The similarity in the prognostic powers of the Cox model to the NPI in the BCCA cohort (Fig A1, online only) indicates that clinical stage and this Cox model give complementary information when all stages of ER-positive breast cancer are assessed. This suggests that the Cox model will be most useful for treatment stratification in the setting of identifying early-stage patients (node negative), where models based solely on clinical stage and pathologic grade are less useful. The tree model was less predictive of outcome but nevertheless validated. In the CCF data set, the tree model distinguished poor patients with a 5-year disease-free survival rate of 75% compared with 85% for moderate and good patients (P = .052; Fig 3F). In the BCCA data set, the tree model distinguished poor and moderate patients, who both had an overall 5-year survival rate of 75% compared with 90% for good patients (P = .00024; Fig 3H). The tree model was independent of node status in the BCCA cohort, but it was not independent of node status, stage, or grade in the CCF cohort; therefore, it is likely less useful than the Cox model (Table 2). In the combined CCF and BCCA data sets, for the subset of patients who received either a poor or good prognostic call in these datasets (94%), the sensitivity for a poor prognostic call for predicting disease progression was 47%, whereas the specificity was 76%. The positive predictive value of a poor prognosis call was 28% (95% CI, 23% to 32%), whereas the negative predictive value was 88% (95% CI, 85% to 91%).
Because of improved surveillance, an increasing number of new breast cancer patients have an early stage of disease at diagnoses. Although these patients have a relatively good prognosis when treated with surgery alone, approximately 10% to 30% will experience recurrence in the absence of adjuvant treatment.5,27-29 Current prognostic tools that rely on clinical and pathologic variables have only marginal utility in this early-stage tumor subset of patients. We sought to determine whether we could discover individual antibodies or sets of antibodies used in combination that could predict clinical outcome and, therefore, could be used in combination with clinical parameters to help with treatment decision making. Gene expression profiling of tumors has clearly demonstrated that there is biologic heterogeneity in breast cancer and that gene expression–based subtypes are associated with different clinical behavior. In this study, we explored whether gene expression data could be translated into prognostic models using a small set of IHC markers. We generated several hundred novel antisera as well as screened many commercially available IHC reagents. Through a stepwise process, we identified a minimal set of five antisera that could be combined using a Cox proportional hazards model and a set of six antisera that could be combined using a regression tree model that could be used to predict outcome in ER-expressing breast cancer patients. We validated these prognosticators in two independent ER-expressing breast cancer cohorts (all patients) from different institutions. Although these validation studies were underpowered in the clinically important node-negative subset of patients, multivariable analysis demonstrated that the Cox model prognostic index was independent of clinical stage, lymph node status, and pathologic grade. This suggests that, although it is yet to be directly demonstrated, this test may be useful in early-stage, node-negative patients. Although the antisera used in the Cox and tree models overlapped, the Cox model outperformed the tree model in both validation cohorts by hazard ratio and statistical significance and by its independence from clinical and pathologic staging. The five antisera that comprise the Cox model are associated with distinct pathways in cells but, in general, relate to cell division physiology. p53 has been implicated in checkpoints governing progression through the cell cycle and is commonly mutated and/or overexpressed by IHC measurement in tumors. Use of p53 as a prognostic marker has been controversial in large part because of inconsistency in reagents and methods used for analysis.3,8,30 Use of IHC assessment of a p53 protein expression status is best supported for ER-positive node-negative breast cancer.3,30 SLC7A5 is part of a two-protein complex with SLC3A2, the heavy chain of a neutral amino acid transporter implicated in nutrient transport at the blood-brain barrier.31,32 By gene expression pattern, it clusters with markers of proliferation in breast cancer and has been previously noted to be highly expressed in some cancers.33-35 CEACAM5 is a glycosylphosphatidylinositol-anchored protein member of the immunoglobulin supergene family that has been explored as a serologic marker for breast cancer detection.36 NDRG1 is upregulated in response to stress, including hypoxia, and was previously found to be upregulated in tumors.37 HTF9C has weak homology to Saccharomyces cerevisiae exonuclease, and its steady-state transcript level varies across the cell cycle.26 Although all biomarkers in the prognostic panel can be linked to cell growth or differentiation, these markers distinguish overlapping but distinct tumor sets and are only weakly correlated with traditional proliferation markers (eg, Ki67 and PCNA) in our data sets. This suggests that these reagents together may be detecting distinct manifestations of aggressive tumor phenotypes that may be present in breast tumors. This study lends support to the premise that measurement of the expression of only a few well-chosen genes or their cognate protein products will suffice to identify clinically significant patient groups. A consistent theme is beginning to emerge in that most published prognostic gene expression algorithms for ER-positive breast cancer measure multiple markers that relate, in part, to different aspects of growth or differentiation of tumor cells. This includes the 21-gene prognostic assay, the luminal A versus luminal B gene expression pattern, and the microarray-based prognostic gene expression signatures identified in independent studies.9-12,15,38,39 These results build on a rich literature that has documented the association between tumor proliferation, differentiation, and outcome. The diagnostic test reported herein includes reagents that detect a protein responsible for cell cycle checkpoint regulation, a hypoxic response marker, a nutrient transporter, a cell cycle–regulated gene, and a carcinoembryonic marker. The fact that these different approaches have identified biomarker signatures quite different in detail but independently validated suggests that the biology that distinguishes tumors with differing outcomes is robust and measurable. Although our study failed to identify a prognostic panel for ER-negative tumors, recent progress in IHC classification of so-called triple-negative breast tumors suggests that different ER-negative patients subtypes might also be clinically distinct.40 The current combination of antisera in the ER-positive Cox model presented herein contributes complementary information to current prognosticators in ER-positive patients and is a significant improvement on current IHC prognosticators. We are encouraged to explore its utility in early-stage patients, where it has the potential to compete favorably with gene expression–based approaches with regard to both cost and ease of use. Further study is warranted to develop this biomarker assay into a clinical tool to significantly improve decision making about the use of adjuvant therapy for breast cancer.
Methods Immunohistochemistry.For all Applied Genomics Inc (Sunnyvale, CA) antisera, tissue arrays were deparaffinized and dehydrated by submerging in xylene three times for 10 minutes each, followed by rinsing three times in 100% ethanol and two times in 95% ethanol and then treated by microwave boiling for 11 minutes in 10 µM buffered citrate (pH 6.0). Slides were allowed to cool to room temperature and were then rinsed in distilled water followed by phosphate-buffered saline. Slides were dipped in 0.03% hydrogen peroxide followed by rinsing with phosphate-buffered saline and then stained using dilutions of antibodies in DAKO Diluent (DAKO Cytomation Inc, Carpinteria, CA) for 1 hour at room temperature. Secondary antibody was applied for 1 hour, and staining was visualized using the DAKO Cytomation Envision staining kit according to the manufacturer's instructions. For each antibody, dilutions were first tested on a small titer tissue array that had breast cancer samples, with positive and negative samples for all antibodies in the panel, in addition to a set of tumor-derived cell lines suspended in paraffin. For the Comprehensive Cancer Institute of Huntsville (CCIH) cohort, commercial antibody slides (ER, PR, EGFR, erbB2, c-KIT, CK5/6, CK17, p53, and Ki67) were stained by a commercial service (US Labs Inc, Irvine, CA). Commercial stains for the British Columbia Cancer Agency (BCCA) and Cleveland Clinic Foundation (CCF) cohorts were performed as previously described.16,17 Antisera generation and screening. Genes targeted for antisera generation were selected on the basis of gene expression patterns in diverse tumors as well as selected cell biologic experiments.12,18-24,26,41 One to three peptides from the open reading frame were selected and synthesized by standard procedures and conjugated to keyhole limpet hemocyanin, mixed, and immunized into two out-bred rabbits. Antipeptide titers were monitored, and when suitable titer was achieved, raw antisera was harvested, pooled, and affinity purified on a column conjugated with a mix of immunizing peptides. Novel antisera (n = 737) were screened by an iterative process that entailed first determining a concentration that discriminated tissue and/or cell types on titer arrays composed of an assortment of tumor and normal tissues. Antibodies determined to have an interpretable staining pattern (n = 404) were placed on tumor screening arrays composed of 300 to 400 breast tumor specimens (no clinical follow-up available) to identify those that distinguished a significant subset of patients. Images of each stained core were obtained using an automated scanning microscope with hardware and software designed for tissue array image archiving (Bacus Laboratories, Lombard, IL). Data sets and images were managed in a custom Oracle (Oracle; Redwood Shores, CA)-based database designed to retrieve and assemble data sets and images based on clinical annotations and other data parameters. Model building for prognostication. Model building was performed on the CCIH tumor sample set. For a patient to be included in model building, data was required to be present for at least 80% of the antisera. One hundred twenty-six of the 404 screened antibodies were selected to be stained and scored on the CCIH array for which follow-up data were available. Using a log-rank test, the potential significance of each antibody as a predictor of recurrence was assessed in the following four populations: all patients; estrogen receptor (ER) –positive patients; ER-positive, node-negative patients; and ER-negative patients. Antibodies shown to have a significant (P < .10, two sided) association with recurrence were used in Cox proportional hazards models or regression trees. For the ER-positive, node-negative patients, the antibodies deemed to have prognostic significance in the ER-positive and/or ER-positive, node-negative patient sets were used. For the ER-negative patients, the antibodies deemed to have significance in the ER-negative patients were used, along with antibodies that showed a positive correlation with poor outcome in the total patient population.
The Cox proportional hazards regression model building used initially all of the antibodies nominated by the univariate tests. The models were then iteratively pruned by removal of antibodies that did not add significantly to the fit of the model, until removal of antibodies caused a significant reduction in the strength of the model. A prognostic score for the final model for the ER-positive, node-negative patients was defined as equal to (1.543 x SLC7A5) + (1.124 x p53) + (1.057 x NDRG1) + (0.716 x HTF9C) + (0.504 x CEACAM) – 0.863, in which each antibody was given a value of 0 or 1 based on the staining, as described earlier. Patients were classified into the following three prognostic categories: patients with a prognosis score of Tree models were created using the RPART routines, wherein regression models are derived that can be visualized as binary trees. The trees were pruned to a minimal complexity (least number of terminal nodes without losing too much prognostic ability) by a cross-validation procedure in which models were built on a series of patient groups picked from the total patient set using a series of increasingly pruned trees.42 The results over all the groups defined by the decision trees are summed, and the minimally complex, least error prone model was chosen. The model was further simplified by binning nodes with a range of response values to the classes of good, moderate, and poor. Ten-fold cross validation was performed as an unbiased test for the potential overfitting of the Cox and tree models. In this test, the patients were assigned to 10 overlapping groups, each comprising 90% of the patients; the set of antibodies with a significant univariate association with outcome was identified, and new models were derived. These ten models were used to predict outcome for the 10% of patients absent from each model building episode such that, ultimately, all patients were assigned to two prognostic groups (good and poor). The statistical significance of the association of the composite predicted groups with outcome was assessed by the Wald test, and if nonsignificant (P > .05, two sided), it was deemed likely that the derived models were overfit. Validation in independent cohorts. The CCF and BCCA tumor sets were used for model validation. Antibody scoring rules, Cox and regression tree model terms, and criteria for categorization as good, moderate, and poor were prospectively defined before staining and scoring. For a patient to be included for model validation, data were required to be present for all biomarkers in the model. For the CCF data set, disease-free survival was assessed; recurrence was defined as any return of carcinoma, including locoregional, contralateral, new primary, or distant metastasis, whereas death from other causes was censored. For the BCCA data set, overall survival was assessed (disease-free survival was unavailable), and death from other causes was censored. Follow-up time in both validation cohorts was limited to 5 years. The hypothesis that our prognostication of patients would be replicated in the validation cohorts was tested by converting the prognostic categories of good, moderate, and poor into values of 0, 1, and 2, respectively, and then assessing the values in a univariate Cox model, with a P value (two sided) in the Wald test of less than .05 being considered significant.
Calculation of sensitivity, specificity, positive predictive value, and negative predictive value. For calculation of the performance metrics of the ER-positive Cox proportional hazards model, the patients deemed moderate by the prognostic score were removed from the cohort on the rational that a designation of a moderate prognostic risk is not likely to be clinically useful. Within the patient group predicted to be high risk, patients who had an event (recurrence or death depending on cohort) were designated true positives (TP), whereas patients without an event were designated false positives (FP). Within the patient group predicted to be low risk, patients without an event were designated true negatives (TN), whereas patients who had an event were designated false negatives (FN). Sensitivity was defined as follows: TP/(TP + FN). Specificity was defined as follows: TN/(TN + FP). Positive predictive value (PPV) was defined as follows: TP/(TP + FP). Negative predictive value (NPV) was defined as follows: TN/(TN + FN). The 95% CI for PPV was calculated as PPV ± 1.96 x (PPV x (1 – PPV)/N)
Calculation of Nottingham Prognostic Index. The Nottingham Prognostic Index (NPI) was calculated as follows: (0.2 x tumor diameter in cm) + lymph node stage + tumor grade. The lymph node stage is 1 if there are no nodes affected, 2 if up to three glands are affected, or 3 if more than three glands are affected. Similarly, the tumor grade is scored as 1 for a grade 1 tumor, 2 for a grade 2 tumor, or 3 for a grade 3 tumor. The prognostic categories of the NPI are defined as good outcome less than 3.4
Although all authors completed the disclosure declaration, the following authors or their immediate family members 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. For a detailed description of the disclosure categories, or for more information about ASCO's conflict of interest policy, please refer to the Author Disclosure Declaration and the Disclosures of Potential Conflicts of Interest section in Information for Contributors.
Dollar Amount Codes (A) < $10,000 (B) $10,000-$99,900 (C)
We thank the staff of Huntsville hospital and Pathology Associates of Huntsville, AL, for their help in assembling the Comprehensive Cancer Center of Huntsville cohort.
Authors' disclosures of potential conflicts of interest and author contributions are found at the end of this article.
1. Haybittle JL, Blamey RW, Elston CW, et al: A prognostic index in primary breast cancer. Br J Cancer 45:361-366, 1982[Medline] 2. Galea MH, Blamey RW, Elston CE, et al: The Nottingham Prognostic Index in primary breast cancer. Breast Cancer Res Treat 22:207-219, 1992[CrossRef][Medline] 3. Fitzgibbons PL, Page DL, Weaver D, et al: Prognostic factors in breast cancer: College of American Pathologists Consensus Statement 1999. Arch Pathol Lab Med 124:966-978, 2000[Medline] 4. Eifel P, Axelson JA, Costa J, et al: National Institutes of Health Consensus Development Conference Statement: Adjuvant therapy for breast cancer, November 1-3, 2000. J Natl Cancer Inst 93:979-989, 2001 5. Goldhirsch A, Wood WC, Gelber RD, et al: Meeting highlights: Updated international expert consensus on the primary therapy of early breast cancer. J Clin Oncol 21:3357-3365, 2003 6. Mincey BA, Palmieri FM, Perez EA: Adjuvant therapy for breast cancer: Recommendations for management based on consensus review and recent clinical trials. Oncologist 7:246-250, 2002 7. Wood WC, Anderson M, Lyles RH, et al: Can we select which patients with small breast cancers should receive adjuvant chemotherapy? Ann Surg 235:859-862, 2002[CrossRef][Medline] 8. Hayes DF, Isaacs C, Stearns V: Prognostic factors in breast cancer: Current and new predictors of metastasis. J Mammary Gland Biol Neoplasia 6:375-392, 2001[CrossRef][Medline] 9. 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[CrossRef][Medline] 10. 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 11. Wang Y, Klijn JG, Zhang Y, et al: Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet 365:671-679, 2005[Medline] 12. Sorlie T, Perou CM, Tibshirani R, et al: Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 98:10869-10874, 2001 13. Chang HY, Nuyten DS, Sneddon JB, et al: Robustness, scalability, and integration of a wound-response gene expression signature in predicting breast cancer survival. Proc Natl Acad Sci U S A 102:3738-3743, 2005 14. Chang HY, Sneddon JB, Alizadeh AA, et al: Gene expression signature of fibroblast serum response predicts human cancer progression: Similarities between tumors and wounds. PLoS Biol 2:E7, 2004[CrossRef][Medline] 15. Paik S, Shak S, Tang G, et al: A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 351:2817-2826, 2004 16. Yoder BJ, Tso E, Skacel M, et al: The expression of fascin, an actin-bundling motility protein, correlates with hormone receptor-negative breast cancer and a more aggressive clinical course. Clin Cancer Res 11:186-192, 2005 17. Makretsov NA, Huntsman DG, Nielsen TO, et al: Hierarchical clustering analysis of tissue microarray immunostaining data identifies prognostically significant groups of breast carcinoma. Clin Cancer Res 10:6143-6151, 2004 18. Chen X, Cheung ST, So S, et al: Gene expression patterns in human liver cancers. Mol Biol Cell 13:1929-1939, 2002 19. Chen X, Leung SY, Yuen ST, et al: Variation in gene expression patterns in human gastric cancers. Mol Biol Cell 14:3208-3215, 2003 20. Garber ME, Troyanskaya OG, Schluens K, et al: Diversity of gene expression in adenocarcinoma of the lung. Proc Natl Acad Sci U S A 98:13784-13789, 2001 21. Perou CM, Sorlie T, Eisen MB, et al: Molecular portraits of human breast tumours. Nature 406:747-752, 2000[CrossRef][Medline] 22. Pollack JR, Sorlie T, Perou CM, et al: Microarray analysis reveals a major direct role of DNA copy number alteration in the transcriptional program of human breast tumors. Proc Natl Acad Sci U S A 99:12963-12968, 2002 23. Ross DT, Scherf U, Eisen MB, et al: Systematic variation in gene expression patterns in human cancer cell lines. Nat Genet 24:227-235, 2000[CrossRef][Medline] 24. Schaner ME, Davidson B, Skrede M, et al: Variation in gene expression patterns in effusions and primary tumors from serous ovarian cancer patients. Mol Cancer 4:26, 2005[Medline] 25. Sorlie T, Tibshirani R, Parker J, et al: Repeated observation of breast tumor subtypes in independent gene expression data sets. Proc Natl Acad Sci U S A 100:8418-8423, 2003 26. Whitfield ML, Sherlock G, Saldanha AJ, et al: Identification of genes periodically expressed in the human cell cycle and their expression in tumors. Mol Biol Cell 13:1977-2000, 2002 27. Elledge RM, McGuire WL: Prognostic factors and therapeutic decisions in axillary node-negative breast cancer. Annu Rev Med 44:201-210, 1993[CrossRef][Medline] 28. Fisher B, Dignam J, Bryant J, et al: Five versus more than five years of tamoxifen for lymph node-negative breast cancer: Updated findings from the National Surgical Adjuvant Breast and Bowel Project B-14 randomized trial. J Natl Cancer Inst 93:684-690, 2001 29. Fisher B, Jeong JH, Dignam J, et al: Findings from recent National Surgical Adjuvant Breast and Bowel Project adjuvant studies in stage I breast cancer. J Natl Cancer Inst Monogr 30:62-66, 2001[Medline] 30. Allred DC, Harvey JM, Berardo M, et al: Prognostic and predictive factors in breast cancer by immunohistochemical analysis. Mod Pathol 11:155-168, 1998[Medline] 31. Kido Y, Tamai I, Uchino H, et al: Molecular and functional identification of large neutral amino acid transporters LAT1 and LAT2 and their pharmacological relevance at the blood-brain barrier. J Pharm Pharmacol 53:497-503, 2001[CrossRef][Medline] 32. Boado RJ, Li JY, Nagaya M, et al: Selective expression of the large neutral amino acid transporter at the blood-brain barrier. Proc Natl Acad Sci U S A 96:12079-12084, 1999 33. Kanai Y, Segawa H, Miyamoto K, et al: Expression cloning and characterization of a transporter for large neutral amino acids activated by the heavy chain of 4F2 antigen (CD98). J Biol Chem 273:23629-23632, 1998 34. Kim do K, Ahn SG, Park JC, et al: Expression of L-type amino acid transporter 1 (LAT1) and 4F2 heavy chain (4F2hc) in oral squamous cell carcinoma and its precursor lesions. Anticancer Res 24:1671-1675, 2004[Medline] 35. Kobayashi H, Ishii Y, Takayama T: Expression of L-type amino acid transporter 1 (LAT1) in esophageal carcinoma. J Surg Oncol 90:233-238, 2005[Medline] 36. Chu TM, Nemoto T: Evaluation of carcinoembryonic antigen in human mammary carcinoma. J Natl Cancer Inst 51:1119-1122, 1973[Medline] 37. Cangul H, Salnikow K, Yee H, et al: Enhanced overexpression of an HIF-1/hypoxia-related protein in cancer cells. Environ Health Perspect 110:783-788, 2002 (suppl 5)[Medline] 38. Sotiriou C, Neo SY, McShane LM, et al: Breast cancer classification and prognosis based on gene expression profiles from a population-based study. Proc Natl Acad Sci U S A 100:10393-10398, 2003 39. Dai H, van't Veer L, Lamb J, et al: A cell proliferation signature is a marker of extremely poor outcome in a subpopulation of breast cancer patients. Cancer Res 65:4059-4066, 2005 40. Nielsen TO, Hsu FD, Jensen K, et al: Immunohistochemical and clinical characterization of the basal-like subtype of invasive breast carcinoma. Clin Cancer Res 10:5367-5374, 2004 41. Nielsen TO, West RB, Linn SC, et al: Molecular characterisation of soft tissue tumours: A gene expression study. Lancet 359:1301-1307, 2002[CrossRef][Medline] 42. Therneau TM, Atkinson EJ: An introduction to recursive partitioning using the RPART routine. Rochester, MN, Mayo Clinic, Technical Report, 1997 Submitted January 18, 2006; accepted April 11, 2006. This article has been cited by other articles:
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
TOP
ABSTRACT
INTRODUCTION
0 were classified as good, patients with a score of less than 0.7 were classified as moderate, and all other patients were classified as poor. 
.5. The 95% CI for NPV was calculated as NPV ± 1.96 x (NPV x (1 – NPV)/N)
$100,000 (N/R) Not Required
