Estrogen-Regulated Genes Predict Survival in Hormone Receptor–Positive Breast Cancers Daniel S

VOLUME 24 ⅐ NUMBER 11 ⅐ APRIL 10 2006

JOURNAL OF CLINICAL ONCOLOGY ORIGINAL REPORT

Estrogen-Regulated Genes Predict Survival in Hormone Receptor–Positive Breast Cancers Daniel S. Oh, Melissa A. Troester, Jerry Usary, Zhiyuan Hu, Xiaping He, Cheng Fan, Junyuan Wu, Lisa A. Carey, and Charles M. Perou

From the Departments of Genetics, ABSTRACT Medicine, and the Pathology & Labo- ratory Medicine, Lineberger Compre- hensive Cancer Center, University of Purpose North Carolina, Chapel Hill, NC. The prognosis of a patient with estrogen receptor (ER) and/or progesterone receptor (PR) –positive breast cancer can be highly variable. Therefore, we developed a gene expression–based outcome Submitted June 27, 2005; accepted ϩ ϩ December 14, 2005. predictor for ER and/or PR (ie, luminal) breast cancer patients using biologic differences among these tumors. Supported by funds from the National Cancer Institute (Bethesda, MD) Breast Materials and Methods SPORE program to the University of The ERϩ MCF-7 breast cancer cell line was treated with 17␤-estradiol to identify estrogen- North Carolina–Chapel Hill (Chapel Hill, regulated genes. These genes were used to develop an outcome predictor on a training set of 65 NC; Grant No. P50-CA58223-09A1), luminal epithelial primary breast carcinomas. The outcome predictor was then validated on three National Institutes of Health Grant No. independent published data sets. M01RR00046, and the National Insti- tute of Environmental Health Sciences Results Grant No. U19-ES11391-03; and by The estrogen-induced gene set identified in MCF-7 cells was used to hierarchically cluster a 65 University of North Carolina Lineberger tumor training set into two groups, which showed significant differences in survival (P ϭ .0004). Cancer Control Education Program Grant No. R25 CA57726 (M.A.T.). Supervised analyses identified 822 genes that optimally defined these two groups, with the poor-prognosis group IIE showing high expression of cell proliferation and antiapoptosis genes. Terms in blue are defined in the glossary, The good prognosis group IE showed high expression of estrogen- and GATA3-regulated genes. found at the end of this article and online ϩ ϩ at www.jco.org. Mean expression profiles (ie, centroids) created for each group were applied to ER and/or PR tumors from three published data sets. For all data sets, Kaplan-Meier survival analyses showed Authors’ disclosures of potential con- significant differences in relapse-free and overall survival between group IE and IIE tumors. flicts of interest and author contributions are found at the end of this Multivariate Cox analysis of the largest test data set showed that this predictor added significant article. prognostic information independent of standard clinical predictors and other gene expression–

Address reprint requests to Charles M. based predictors. Perou, PhD, Lineberger Comprehensive Conclusion Cancer Center, University of North This study provides new biologic information concerning differences within hormone receptor–positive Carolina at Chapel Hill, Campus Box breast cancers and a means of predicting long-term outcomes in tamoxifen-treated patients. 7295, Chapel Hill, NC 27599; e-mail: [email protected]. J Clin Oncol 24:1656-1664. © 2006 by American Society of Clinical Oncology © 2006 by American Society of Clinical Oncology

0732-183X/06/2411-1656/$20.00 sion profiling, breast tumors can be classified into INTRODUCTION ϩ DOI: 10.1200/JCO.2005.03.2755 five molecular subtypes (basal-like, HER2 /ER-, Breast cancers are traditionally stratified into hor- normal breast–like, and luminal A and B) with sig- mone receptor–positive and –negative groups to guide nificant differences in patient outcome.8,9 The two patient management. This is because almost all hor- luminal subtypes are composed almost entirely of mone antagonist (ie, tamoxifen) -responsive breast ERϩ and/or PRϩ tumors and are defined by the cancers are estrogen receptor (ER) and/or progester- high expression of a gene set (luminal epithelial one receptor (PR) -positive.1 However, a subgroup of ERϩ set) that includes ER and GATA3. Compared these patients experiences disease relapse irrespective with luminal A tumors, luminal B tumors have poor of standard therapy with tamoxifen.2-4 In many outcomes despite being clinically ERϩ.8,9 cases, the resistance mechanisms are unknown.5-7 A To date, several laboratories have developed method for identifying those ERϩ and/or PRϩ pa- gene expression–based predictors for ERϩ and/or tients who do not experience relapse in the presence PRϩ patients. All used supervised analyses based of tamoxifen versus those who do would be of sig- on patient outcomes/tumor response.10-13 Here, we nificant value. developed a gene expression–based predictor using Gene expression profiling is a powerful method an approach based solely on biologic characteris- for breast cancer prognostication. With gene expres- tics of the tumors. Because of the importance of

1656

estrogen signaling in breast epithelial cell biology, we hypothesized optimally differentiated group I versus group II tumors.20,21 The 65 lumi- that differential expression of estrogen-regulated genes would be nal tumors were then clustered using the 822 genes, and two groups useful in predicting outcome. (groups IE and IIE) were evident. By matching UniGene (NCBI, http://www.ncbi.nlm.nih.gov/entrez/ query.fcgi?dbϭunigene) identifiers, data for as many as possible of the 822 MATERIALS AND METHODS estrogen-SAM genes was obtained for ERϩ and/or PRϩ tumors (classified as provided in the primary publications) from three independent test data sets.8,9,12,22 The Ma et al,12 Sorlie et al,8,9 and Chang et al22 data sets con- Cell Culture and Collection of mRNA sisted of 60, 90, and 250 ERϩ and/or PRϩ tumors, respectively. Ma et al MCF-7 cells were maintained as described previously.14 Cells were plated tumors were uniformly treated with adjuvant tamoxifen alone. Sorlie et al in 150-mm dishes and grown until 50% confluence. Media was changed and tumors received adjuvant tamoxifen, with some also receiving neoadjuvant cells maintained for 48 hours in estrogen-free medium (phenol red-free chemotherapy. Chang et al tumors were heterogeneously treated; 24 of the RPMI-1640 with 10% charcoal-dextran-stripped fetal bovine serum) before 250 tumors we used for this data set were discussed in earlier publications.23 Ϫ treating for 2,4,8, or 24 hours with 10 6 M17␤-estradiol (Sigma-Aldrich To remove microarray platform/source systematic biases, we applied Inc, St Louis, MO). Cells were harvested, and mRNA isolated using a distance-weighted discrimination (DWD)24 to the training and test data sets. Micro-FastTrack kit (Invitrogen Corp, Carlsbad, CA). A reference mRNA From the DWD standardized luminal tumor training data set, centroids were sample was collected from cells maintained for 48 hours in estrogen-free created consisting of the average expression of the 822 estrogen-SAM genes for medium (ie, estrogen-starved cells). groups IE and IIE. We then classified each ERϩ and/or PRϩ tumor in the test Microarray Experiments data sets as group IE or IIE according to each sample’s nearest centroid as determined by Spearman correlation. Human whole-genome microarrays (Agilent Technologies Inc, Palo Alto, CA) were hybridized according to manufacturer’s protocol with Survival Analyses ␮ Cy3-CTP–labeled cRNA from estrogen-starved cells (2 g/sample) and Kaplan-Meier survival plots were compared using the Cox-Mantel ␤ ␮ Cy5-CTP–labeled cRNA from 17 -estradiol-treated cells (2 g/sample), log-rank test in Winstat (R. Fitch Software, Staufen, Germany) for Excel with dye-flip replicates for each time point. Microarrays were scanned 15 (Microsoft Corp, Redmond, WA). Two-way contingency table analysis was and image files analyzed as described previously. All primary microarray performed using Winstat for Excel. For the Chang et al data set, multivariate data are available via the University of North Carolina (Chapel Hill, NC; Cox proportional hazards analysis was performed using SAS (SAS Institute, UNC-CH) Microarray Database (https://genome.unc.edu/) and the Gene Cary, NC). Expression Omnibus (GEO; National Center for Biotechnology Informa- tion [NCBI], http://www.ncbi.nlm.nih.gov/geo/) with series number GSE2740 (GSM52882-GSM52909, GSM34423-GSM34568). RESULTS Analysis of Microarray Data to Identify GATA3 and Estrogen-Regulated Genes Data from microarray experiments were calculated as described.15 Genes Identification of Estrogen-Regulated Genes ϩ were excluded from data analysis if they did not have signal intensity of at least To identify estrogen-regulated genes, we used the ER hu- 30 in both channels for at least 70% of the experiments. To identify estrogen- man breast tumor-derived MCF-7 cell line as a model system. A regulated genes, we used one-class significance analysis of microarrays (SAM) one-class SAM supervised analysis16 with an FDR of 0.04% identi- to identify genes that changed in all estrogen-treated time points (as a single fied 383 induced and 574 repressed genes in microarray experiments 16 class) relative to the estrogen-starved cells. In our SAM analyses, we did not on MCF-7 cells treated with 17␤-estradiol for 2, 4, 8, or 24 hours. use the fold-change cutoff option to avoid the fold-change associated compli- 17 Many genes identified were previously known to be ER-regulated cations/pitfalls described by Larsson et al. Using a false discovery rate (FDR) 25-28 of 0.04%, SAM identified 383 estrogen-induced and 574 estrogen- including CCND1, PR, RERG, CTSD, and PDZK1. Using the repressed genes; for subsequent estrogen-SAM analyses, only the 383 in- program EASE (Expression Analysis Systematic Explorer; http:// duced genes were used. Average linkage hierarchical cluster analysis was david.niaid.nih.gov/david/ease.htm),29 the gene ontology (GO) cate- conducted and the results visualized in TreeView (http://taxonomy.zoolo- gories “sterol metabolism/biosynthesis,” “ribosome biogenesis/ 18 gy.gla.ac.uk/rod/treeview.html). assembly,” and “cytoskeleton structural constituent” were over- GATA3-induced genes were identified by microarray experiments on represented relative to chance in the set of 383 estrogen-induced genes. 293T cells transfected with GATA3 gene constructs, as detailed in Usary et al.19 One-class SAM analysis (0.58% FDR) identified 407 genes that were induced Estrogen and GATA3-Regulated Genes Are Present in in the GATA3 samples (as a single class) relative to empty vector controls. the Luminal/ER؉ Gene Cluster Analysis of Primary Breast Tumor Data Using the Estrogen- The luminal/ERϩ expression cluster is a gene set identified in Induced Gene Set many breast tumor profiling studies,23,30-32 includes GATA3 and ER, The primary breast tumor samples (collected with patient consent and is expressed in the luminal A and B tumor subtypes.8,9 To deter- and UNC-CH Human Investigations Review Committee approval) used in mine whether estrogen-regulated genes are present in the luminal/ the training data set are described in Hu et al (submitted for publication), ϩ except for 14 new tumor samples. A total of 118 fresh frozen breast tumor ER gene set, we first clustered 118 primary breast tumors using a and nine nontumor breast samples represented by 160 microarray experi- 1,300-gene “breast intrinsic” gene set developed by Hu et al (submit- ments were analyzed using the 1,300-gene “breast intrinsic” gene set devel- ted; Fig 1). Figure 1B shows that many of the estrogen-regulated genes oped by Hu et al (submitted), which identified 65 tumors as belonging to from our in vitro MCF-7 experiments were present in the tumor the luminal subtype. These luminal tumors included 61 ERϩ and/or PRϩ defined luminal/ERϩ gene cluster. To further define relationships tumors according to immunohistochemistry, three ER– and PR–, and one among genes in this cluster, we also ascertained the presence of genes not determined. regulated by GATA3, a transcription factor with an important role in The 383-gene MCF-7 estrogen-induced gene list was used to hierarchi- ϩ 19 cally cluster the 65 luminal tumors resulting in two groups, which we called ER breast cancer biology. Of the 407 genes induced by GATA3 in groups I and II. We used a two-class, unpaired SAM analysis (with 1% vitro, many were present in the luminal/ERϩgene cluster. Thus, genes FDR) to identify 822 genes (referred to as the estrogen-SAM list) that identified as being regulated by ER and/or GATA3 in vitro also cluster

www.jco.org 1657

Fig 1. Genes regulated by estrogen and/or GATA3 in vitro are present in the primary tumor luminal epithelial/estrogen receptor–positive (ERϩ) gene cluster. (A) Scaled-down representation of 118 tumors hierarchically clustered using the 1,300-gene intrinsic list developed by Hu et al (submitted for publication). (B) Luminal/ERϩ gene cluster. Blue, luminal epithelial subtype; pink, HER2ϩ/ER–; red, basal-like; green, nontumor breast–like.

1658 JOURNAL OF CLINICAL ONCOLOGY

Fig 2. Hierarchical cluster analysis of the 65 luminal tumors (identiﬁed in Fig 1) using the 822-gene estrogen–signiﬁcance analysis of microarrays–derived list. (A) Scaled-down representation of the complete cluster diagram. Group IE and IIE tumors are indicated by blue and orange, respectively. Gene clusters containing (B) XBP1, (C) ribosomal genes, (D) progesterone receptor, (E) FOXA1, (F) MAGE genes, (G) proliferation signature, and (H) apoptosis and interferon-response genes.

www.jco.org 1659

Downloaded from www.jco.org on April 8, 2006 . For personal use only. No other uses without permission. Copyright © 2006 by the American Society of Clinical Oncology. All rights reserved. Oh et al near these transcription factors in vivo and help to define an expres- grouping of samples when compared with that using the 383 estrogen- sion pattern seen in many studies.23,30,32,34 induced genes. Kaplan-Meier analysis showed that using the estrogen- Analysis of Luminal Tumors Using SAM list grouped the tumors into two groups (referred to as group IE ϭ Estrogen-Induced Genes and IIE) that predicted RFS (P .019; Fig 3A). We hypothesized that expression differences of estrogen- Group IE tumors showed high expression of XBP1, FOXA1, PR, induced genes may define clinically relevant subgroups within clini- and many ribosomal genes (Fig 2B-E). According to EASE, the GO cally defined ERϩ and/or PRϩ tumors. To test this hypothesis, we categories “transcriptional regulation,” “DNA binding,” and “extra- clustered the 65 tumors identified as luminal in Figure 1 (blue dendro- cellular” were over-represented relative to chance in group IE tumors. gram branch) using the 383 MCF-7 estrogen-induced genes. Two Group IIE tumors showed the high expression of a prominent prolif- 34,35 main groups resulted. Group I had higher expression of XBP1,PR, and eration signature including Ki-67, MYBL2, Survivin, STK6, and TFF, which are all known ER targets. Group II had higher expression CCNB2 (Fig 2G); these first four genes plus CCNB1 form the basis of a cluster of estrogen-induced genes that included CTPS, E2F6, and for the proliferation portion of the Paik et al recurrence score predic- 13 FANCA. Kaplan-Meier survival analysis showed that group I patients tor, which is a gene expression–based outcome predictor for ERϩ/ 10 had significantly better relapse-free survival (RFS) outcomes than node-negative, tamoxifen-treated patients. Recently, Dai et al group II (P ϭ .0004). performed a supervised analysis for genes that correlated with To further characterize the differences between group I and II outcomes in patients with high ER expression relative to age, and tumors, we performed a supervised analysis (two-class SAM with 1% identified this same proliferation signature as the main determi- FDR) using the major dendrogram branch division to define the two nant for predicting patient outcomes; however, they identified few supervising groups. This analysis identified 822 genes for which group genes associated with good outcomes. I and II tumors showed significant differential expression. This gene Group IIE tumors also showed high expression of a cluster of set, the estrogen-SAM list, was then used to hierarchically cluster the MAGE-A genes (Fig 2F), which have been associated with an increased 65 luminal tumors (Fig 2), which as expected, resulted in a very similar recurrence risk36 and poor tumor differentiation.37 Figure 2H shows

Fig 3. Kaplan-Meier survival curves of estrogen receptor– and/or progesterone receptor–positive tumors classiﬁed as groups IE or IIE using the 822-gene estrogen–signiﬁcance analysis of microarrays–derived list. Survival curves are shown for (A) the 65 Luminal epithelial tumor training data set, (B) the Ma et al,12 (C) Sorlie et al,8,9 and (D) Chang et al22 data sets. P values were calculated using the log-rank test.

1660 JOURNAL OF CLINICAL ONCOLOGY

Table 1. Multivariate Cox Proportional Hazards Analysis of Various Prognostic Factors in Relation to Relapse-Free Survival and Overall Survival for ERϩ and/or PRϩ Tumors in the Chang et al22 Data Set Relapse-Free Survival Overall Survival Variable Hazard Ratio 95% CI P Hazard Ratio 95% CI P to 4.92 < .0001 3.64 1.67 to 7.95 .001 1.71 2.90 ءGroup IIE v IE Age† 0.48 0.31 to 0.74 .001 0.53 0.30 to 0.93 .028 Size‡ 1.59 1.01 to 2.48 .044 1.45 0.80 to 2.64 .22 Tumor grade 2 or 3 v 1 1.80 0.99 to 3.3 .056 3.57 1.24 to 10.23 .02 Node status§ 2.11 1.08 to 4.11 .028 1.85 0.74 to 4.61 .19 Hormonal or chemotherapy v no adjuvant therapy 0.36 0.18 to 0.71 .003 0.47 0.19 to 1.19 .11

NOTE. Boldfacing indicates variables found to be significant (P Ͻ .05) in the Cox proportional hazards model. Abbreviations: ER, estrogen receptor; PR, progesterone receptor. .Tumors were classified as group IE or IIE using the estrogen–significance analysis of microarrays–derived listء †Age was a continuous variable formatted as decade-years. ‡Size was a binary variable: 0 ϭ diameter Ͻ 2 cm; 1 ϭ diameter Ͼ 2 cm. §Node status was a binary variable: 0 ϭ no positive nodes; 1 ϭ one or more positive nodes.

that group IIE tumors have high expression of genes with functions in Tumor Data Using the Estrogen-Induced Gene Set, under Materials the interferon pathway and apoptosis, such as FLIP/CFLAR, which is and Methods). Kaplan-Meier analysis (Fig 3B-D) showed that group an inhibitor of tumor necrosis factor receptor–mediated apoptosis.38 IE tumors had significantly better RFS in all test data sets. The group Several antiapoptosis genes including FLIP, AVEN, Survivin, and IE-IIE classification was also a significant predictor of overall survival BCL2A1 showed high expression in group IIE, suggesting an impaired (OS) for the test data sets in which OS data was available.9,22 Further- ability to undergo cell death. Recent reports have shown that high more, by decreasing the FDR in SAM, we were able to define groups IE expression of FLIP39 or BCL2A140,41 can directly contribute to and IIE using a reduced estrogen-SAM list of 113 genes without any chemoresistance, suggesting that functional inhibition of these pro- loss of predictive ability (Appendix A1, online only). teins may provide a therapeutic target for group IIE patients. Accord- Multivariate Analysis ing to EASE, the GO categories “cell cycle/mitosis,” “antiapoptosis,” Multivariate Cox proportional hazards analysis was performed and “MHC-I” were over-represented relative to chance in group IIE. on the Chang et al data set (Table 1). Using RFS and OS as the end Group IE-IIE Classification Predicts Outcome in ER؉ points, multivariate analysis showed that classifying tumors as -and/or PR؉ Tumors group IE or IIE provided significant prognostic power indepen To test the group IE-IIE classification as a clinically relevant dent of standard clinical factors (RFS P Ͻ .0001; OS P ϭ .001). The outcome predictor, we analyzed ERϩ and/or PRϩ tumors from three group IE-IIE designation had the strongest association of all vari- published breast tumor microarray data sets.9,12,22 We used a single- ables with RFS and OS. sample prediction algorithm to classify tumors in each test data set, In multivariate analyses that included Chang et al’s22 wound- which involved creating group IE and IIE centroids/average profiles response signature and van’t Veer et al’s23 70-gene signature along from the training data set (described in Analysis of Primary Breast with the clinical variables, the group IE-IIE classification continued to

Table 2. Multivariate Cox Proportional Hazards Analysis for ERϩ and/or PRϩ Tumors in the Chang et al22 Data Set Using Various Prognostic Factors Including the Group IE-IIE Classiﬁcation, the van’t Veer et al23 70-Gene Signature, and the Chang et al Wound-Response Signature Relapse-Free Survival Overall Survival Variable Hazard Ratio 95% CI P Hazard Ratio 95% CI P to 3.49 .014 2.31 1.03 to 5.19 .042 1.15 2.01 ءGroup IIE v IE 70-gene signature (poor v good) 2.76 1.50 to 5.06 .001 4.17 1.62 to 10.73 .003 Wound-response signature (activated v quiescent) 2.30 1.09 to 4.85 .028 2.80 0.82 to 9.55 .10 Age† 0.56 0.36 to 0.87 .010 0.64 0.36 to 1.14 .13 Size‡ 1.45 0.93 to 2.28 .10 1.34 0.74 to 2.45 .34 Tumor grade 2 or 3 v 1 0.93 0.47 to 1.82 .82 1.62 0.53 to 4.93 .40 Node status§ 1.72 0.89 to 3.33 .11 1.51 0.62 to 3.70 .36 Hormonal or chemotherapy v no adjuvant therapy 0.37 0.19 to 0.74 .005 0.46 0.18 to 1.14 .095

NOTE. Boldfacing indicates variables found to be significant (P Ͻ .05) in the Cox proportional hazards model. The 70-gene signature and the wound-response signature classifications were taken exactly as calculated in Chang et al,22 and their performances in multivariate analysis may be optimistic. .Tumors were classified as group IE or IIE using the estrogen–significance analysis of microarrays–derived listء †Age was a continuous variable formatted as decade-years. ‡Size was a binary variable: 0 ϭ diameter Ͻ 2 cm; 1 ϭ diameter Ͼ 2 cm. §Node status was a binary variable: 0 ϭ no positive nodes; 1 ϭ one or more positive nodes.

www.jco.org 1661

Downloaded from www.jco.org on April 8, 2006 . For personal use only. No other uses without permission. Copyright © 2006 by the American Society of Clinical Oncology. All rights reserved. Oh et al provide significant prognostic power independent of other variables and/or PRϩ patients. Good-outcome group IE tumors tended to be in the model (RFS P ϭ .014; OS P ϭ .042; Table 2). The performance more differentiated, and highly expressed a subset of estrogen- and of the 70-gene and wound-response signatures in this multivariate GATA3-regulated genes. Conversely, poor-outcome group IIE tu- analysis may be optimistically high because a subset of the patients in mors were more likely to be poorly differentiated. Association of the the Chang et al data set was used to train/optimize these two signa- group IE-IIE profile with grade is expected because grade includes a tures; therefore, the ability of the group IE-IIE classification to show measure of proliferation, which is an important determinant of out- independent prognostic power in a model containing these two pre- comes in ERϩ and/or PRϩ patients.8,10,13,19 However, because the dictors indicates its usefulness in predicting outcomes. group IE-IIE distinction was significant in a multivariate analysis with Group IE-IIE Associations With Clinical and grade included, this distinction adds prognostic information beyond Biologic Parameters what grade provides. To examine the hypothesis that group IE may be more differen- We used three published data sets as test sets and confirmed that ϩ tiated than group IIE tumors, we determined whether an association the Group IE-IIE classification was a significant predictor in ER ϩ existed between this classification and histologic grade. Two-way con- and/or PR patients. We note, however, that the relapse rates differed tingency table analysis showed significant association between grade between data sets, and that the Group IE tumors showed 7% to 40% and group IE-IIE class (Appendix Table A2, online only), with grade 1 relapse rates depending on the data set (Fig 3). This underscores the and 3 tumors more likely to be classified as group IE and IIE, respec- fact that relapse rates are dependent on the characteristics of the tively. Cramer’s V statistic, which measures the strength of association patient set used. For example, comparing relapse rates in the Chang between two variables in a contingency table, indicated a substantial et al22 data set to those observed in Paik et al13 may not be valid association (Cramer’s V Ͼ 0.36) between grade and group IE-IIE class because the Paik et al data set comprised tamoxifen-treated, node- for all data sets. For the Sorlie et al data set, p53 mutation data was negative patients, whereas the majority of the Chang et al patients available, and a two-way contingency table analysis showed a signifi- received no adjuvant therapy and many were node positive. However, cant association between p53 status and group IE-IIE class, with group the multivariate analysis we performed on the Chang et al data set IIE more likely to be p53 mutant (P ϭ .0019; Cramer’s V ϭ 0.44). indicated that our predictor had significant prognostic value independent of standard clinical factors and other gene expression– Comparison of Group IE-IIE Classification to Luminal A/B classification based predictors, and a hazard ratio of 2.90 for group IIE versus IE indicates that our predictor has potential clinical utility. By limit- We compared group IE-IIE classification to the luminal A/B classification.8,9 To identify Luminal A and B tumors in the three test ing the Chang et al data set to those patients who received adjuvant data sets, we used the single-sample predictor developed in Hu et al therapy and were stage I-II, we observed a relapse rate for group ϭ (submitted), which employs centroids for each of the five breast tu- IE patients of 12% (P .007) and significance for overall survival mor “intrinsic subtypes.” We then reclassified luminal A and B tumors outcomes (data not shown). This indicates that given a patient from each data set as group IE or IIE. Kaplan-Meier analyses showed population similar to Paik et al, our predictor’s “good group” can that the group IE-IIE classification did equally well or slightly better achieve outcomes similar to the Paik et al low-risk group. compared with the luminal A/B classification in separating luminal An important unanswered question is whether the group IE-IIE tumors into two groups with different survival outcomes (Appendix distinction predicts pure prognosis, responsiveness to endocrine ther- Table A3, online only). apy, or both. From analyses of patient subsets in the Chang et al data set, it is clear that the group IE-IIE distinction predicts outcome in ERϩ and/or PRϩ patient subsets either receiving or not receiving DISCUSSION adjuvant hormone therapy (data not shown). Paik et al observed similar results for their predictor.13 This is not surprising because half The search for markers that predict long-term outcomes in hormone (eight of 16) of the Paik et al genes were present in the estrogen-SAM receptor-positive tamoxifen-treated patients has been an intense area gene set. However, an advantage of our analysis is that it provides of study. Genomic analyses have contributed to this area, with the additional biologic information (eg, antiapoptosis genes) that the Paik development of several predictive gene sets and assays based on the et al and other predictors did not. Paik et al’s finding that their predic- selection of genes that directly correlate with patient/tumor tor also predicts benefit of chemotherapy42 may also apply to ours. outcomes.10-13,23 We took a different approach and selected genes Thus, the most pressing questions remaining regarding the group using no knowledge of outcomes, and instead selected genes on the IE-IIE classification are (1) whether group IE and IIE gain similar basis of regulation by estrogen and their natural patterns of expression benefits from chemotherapy, and (2) because group IIE tumors do in primary tumors. The 822-gene estrogen-SAM list identified many poorly in the presence of tamoxifen, might they do better if adminis- genes that may help explain the outcome differences seen in ERϩ tered alternative endocrine therapies.

monal therapy for early breast cancer on recurrence 4. Osborne CK: Tamoxifen in the treatment REFERENCES and 15-year survival: An overview of the randomised of breast cancer. N Engl J Med 339:1609-1618, trials. Lancet 365:1687-1717, 2005 1998 1. Osborne CK, Yochmowitz MG, Knight WA III, 3. Bardou VJ, Arpino G, Elledge RM, et al: Pro- 5. Ellis MJ, Coop A, Singh B, et al: Letrozole is et al: The value of estrogen and progesterone recep- gesterone receptor status signiﬁcantly improves more effective neoadjuvant endocrine therapy than tors in the treatment of breast cancer. Cancer 46: outcome prediction over estrogen receptor status tamoxifen for ErbB-1- and/or ErbB-2-positive, estro- 2884-2888, 1980 alone for adjuvant endocrine therapy in two large gen receptor-positive primary breast cancer: Evi- 2. Early Breast Cancer Trialists’ Collaborative breast cancer databases. J Clin Oncol 21:1973- dence from a phase III randomized trial. J Clin Oncol Group (EBCTCG): Effects of chemotherapy and hor- 1979, 2003 19:3808-3816, 2001

1662 JOURNAL OF CLINICAL ONCOLOGY

6. Osborne CK, Bardou V, Hopp TA, et al: Role of sion patterns. Proc Natl Acad Sci U S A 95:14863- receptor in breast cancer. Int J Cancer 84:122-128, the estrogen receptor coactivator AIB1 (SRC-3) and 14868, 1998 1999 HER-2/neu in tamoxifen resistance in breast cancer. 19. Usary J, Llaca V, Karaca G, et al: Mutation of 32. Sotiriou C, Neo SY, McShane LM, et al: Breast J Natl Cancer Inst 95:353-361, 2003 GATA3 in human breast tumors. Oncogene 23: cancer classification and prognosis based on gene 7. Shou J, Massarweh S, Osborne CK, et al: 7669-7678, 2004 expression profiles from a population-based study. Mechanisms of tamoxifen resistance: Increased es- 20. Bair E, Tibshirani R: Semi-supervised meth- Proc Natl Acad SciUSA100:10393-10398, 2003 trogen receptor-HER2/neu cross-talk in ER/HER2- ods to predict patient survival from gene expression 33. West M, Blanchette C, Dressman H, et al: positive breast cancer. J Natl Cancer Inst 96:926- data. PLoS Biol 2:E108, 2004 Predicting the clinical status of human breast cancer 935, 2004 21. Bullinger L, Dohner K, Bair E, et al: Use of by using gene expression profiles. Proc Natl Acad 8. Sorlie T, Perou CM, Tibshirani R, et al: Gene gene-expression profiling to identify prognostic sub- SciUSA98:11462-11467, 2001 expression patterns of breast carcinomas distin- classes in adult acute myeloid leukemia. N Engl 34. Perou CM, Jeffrey SS, van de Rijn M, et al: guish tumor subclasses with clinical implications. J Med 350:1605-1616, 2004 Distinctive gene expression patterns in human Proc Natl Acad SciUSA98:10869-10874, 2001 22. Chang HY, Nuyten DS, Sneddon JB, et al: mammary epithelial cells and breast cancers. Proc 9. Sorlie T, Tibshirani R, Parker J, et al: Repeated Robustness, scalability, and integration of a wound- Natl Acad SciUSA96:9212-9217, 1999 observation of breast tumor subtypes in indepen- response gene expression signature in predicting 35. Whitfield ML, Sherlock G, Saldanha AJ, et al: dent gene expression data sets. Proc Natl Acad Sci breast cancer survival. Proc Natl Acad SciUSA Identification of genes periodically expressed in the U S A 100:8418-8423, 2003 102:3738-3743, 2005 human cell cycle and their expression in tumors. 10. Dai H, van’t Veer L, Lamb J, et al: A cell 23. van’t Veer LJ, Dai H, van de Vijver MJ, et al: Mol Biol Cell 13:1977-2000, 2002 proliferation signature is a marker of extremely poor Gene expression profiling predicts clinical outcome 36. Otte M, Zafrakas M, Riethdorf L, et al: outcome in a subpopulation of breast cancer pa- of breast cancer. Nature 415:530-536, 2002 MAGE-A gene expression pattern in primary breast tients. Proc Natl Acad SciUSA65:4059-4066, 2005 24. Benito M, Parker J, Du Q, et al: Adjustment of cancer. Cancer Res 61:6682-6687, 2001 11. Jansen MP, Foekens JA, van Staveren IL, et systematic microarray data biases. Bioinformatics 37. Kavalar R, Sarcevic B, Spagnoli GC, et al: al: Molecular classification of tamoxifen-resistant 20:105-114, 2004 Expression of MAGE tumour-associated antigens is breast carcinomas by gene expression profiling. 25. Altucci L, Addeo R, Cicatiello L, et al: 17beta- inversely correlated with tumour differentiation in J Clin Oncol 23:732-740, 2005 Estradiol induces cyclin D1 gene transcription, invasive ductal breast cancers: An immunohisto- 12. Ma XJ, Wang Z, Ryan PD, et al: A two-gene p36D1-p34cdk4 complex activation and p105Rb chemical study. Virchows Arch 439:127-131, 2001 expression ratio predicts clinical outcome in breast phosphorylation during mitogenic stimulation of 38. Srinivasula SM, Ahmad M, Ottilie S, et al: cancer patients treated with tamoxifen. Cancer Cell G(1)-arrested human breast cancer cells. Oncogene FLAME-1, a novel FADD-like anti-apoptotic molecule 5:607-616, 2004 12:2315-2324, 1996 that regulates Fas/TNFR1-induced apoptosis. J Biol 13. Paik S, Shak S, Tang G, et al: A multigene 26. Finlin BS, Gau CL, Murphy GA, et al: RERG is assay to predict recurrence of tamoxifen-treated, a novel ras-related, estrogen-regulated and growth- Chem 272:18542-18545, 1997 node-negative breast cancer. N Engl J Med 351: inhibitory gene in breast cancer. J Biol Chem 276: 39. Abedini MR, Qiu Q, Yan X, et al: Possible role 2817-2826, 2004 42259-42267, 2001 of FLICE-like inhibitory protein (FLIP) in chemoresis- 14. Troester MA, Hoadley KA, Sorlie T, et al: 27. Ghosh MG, Thompson DA, Weigel RJ: PDZK1 tant ovarian cancer cells in vitro. Oncogene 23:6997- Cell-type-specific responses to chemotherapeutics and GREB1 are estrogen-regulated genes ex- 7004, 2004 in breast cancer. Cancer Res 64:4218-4226, 2004 pressed in hormone-responsive breast cancer. Can- 40. Wang CY, Guttridge DC, Mayo MW, et al: 15. Hu Z, Troester M, Perou CM: High reproduc- cer Res 60:6367-6375, 2000 NF-kappaB induces expression of the Bcl-2 homo- ibility using sodium hydroxide-stripped long oligonu- 28. Jakowlew SB, Breathnach R, Jeltsch JM, et logue A1/Bfl-1 to preferentially suppress chemotherapy- cleotide DNA microarrays. Biotechniques 38:121- al: Sequence of the pS2 mRNA induced by estrogen induced apoptosis. Mol Cell Biol 19:5923-5929, 1999 124, 2005 in the human breast cancer cell line MCF-7. Nucleic 41. Cheng Q, Lee HH, Li Y, et al: Upregulation of 16. Tusher VG, Tibshirani R, Chu G: Significance Acids Res 12:2861-2878, 1984 Bcl-x and Bfl-1 as a potential mechanism of che- analysis of microarrays applied to the ionizing radia- 29. Hosack DA, Dennis G Jr, Sherman BT, et al: moresistance, which can be overcome by NF- tion response. Proc Natl Acad SciUSA98:5116- Identifying biological themes within lists of genes kappaB inhibition. Oncogene 19:4936-4940, 2000 5121, 2001 with EASE. Genome Biol 4:R70, 2003 42. Paik S, Shak S, Tang G, et al: Expression of 17. Larsson O, Wahlestedt C, Timmons JA: Con- 30. Gruvberger S, Ringner M, Chen Y, et al: the 21 genes in the Recurrence Score assay and siderations when using the significance analysis of Estrogen receptor status in breast cancer is associ- prediction of clinical benefit from tamoxifen in microarrays (SAM) algorithm. BMC Bioinformatics ated with remarkably distinct gene expression pat- NSABP study B-14 and chemotherapy in NSABP 6:129, 2005 terns. Cancer Res 61:5979-5984, 2001 study B-20. Presented at the 27th Annual San An- 18. Eisen MB, Spellman PT, Brown PO, et al: 31. Hoch RV, Thompson DA, Baker RJ, et al: tonio Breast Cancer Symposium, San Antonio, TX, Cluster analysis and display of genome-wide expres- GATA-3 is expressed in association with estrogen December 8-11, 2004 (abstr 24)

■■■

Acknowledgment We thank Phil Bernard, MD, Juan Palazzo, MD, and Funmi Olopade, MD, for providing tumor materials and advice, and Lynda Sawyer and Amy Drobish for collection of patient-associated data.

Appendix The Appendix is included in the full-text version of this article, available online at www.jco.org. It is not included in the PDF version (via Adobe® Reader®). Authors’ Disclosures of Potential Conﬂicts of Interest The authors indicated no potential conﬂicts of interest.

www.jco.org 1663

Author Contributions

Conception and design: Daniel S. Oh, Charles M. Perou Financial support: Charles M. Perou Provision of study materials or patients: Lisa A. Carey, Charles M. Perou Collection and assembly of data: Daniel S. Oh, Jerry Usary, Zhiyuan Hu, Xiaping He, Cheng Fan, Junyuan Wu, Charles M. Perou Data analysis and interpretation: Daniel S. Oh, Melissa A. Troester, Jerry Usary, Cheng Fan, Charles M. Perou Manuscript writing: Daniel S. Oh, Melissa A. Troester, Jerry Usary, Lisa A. Carey, Charles M. Perou Final approval of manuscript: Daniel S. Oh, Zhiyuan Hu, Lisa A. Carey, Charles M. Perou

GLOSSARY

Centroid: Mean expression profile of a group of samples for a Luminal epithelial tumors: Human breast tumors that are char- given set of genes. acterized by the expression of genes typically expressed in the cells that line the ducts of normal mammary glands; these genes include the estro- Cramer’s V statistic: This statistic provides a quantitative gen receptor (ER), GATA3, X-box binding protein 1, FOXA1, and cyto- measure of the strength of association between the two variables keratins 8 and 18. in a contingency table (which cannot be obtained from the P value). Cramer’s V values range from 0 to 1, with 0 indicating no SAM (significance analysis of microarrays): A statistical relationship and 1 indicating perfect association. Traditionally, technique using established software that determines the significance in values between 0.36 and 0.49 indicate a substantial relationship, changes of gene expression seen in microarray analysis (eg, cDNA and and values greater than 0.50 indicate a very strong relationship. oligonucleotide microarrays), which measures the expression of thou- The V statistic is a generalization of the more familiar ␾ statistic sands of genes and identifies changes in expression between different to non–2 ϫ 2 contingency tables, and for 2 ϫ 2 tables the V sta- biologic states. On the basis of changes in gene expression relative to the tistic is equal to the ␾ statistic. standard deviation of repeated measurements, SAM assigns a score to DWD (distance-weighted discrimination): A method each gene. When scores are greater than an adjustable threshold, permu- to identify and adjust systematic biases that are present within tations of repeated measurements are used by SAM to estimate the per- microarray datasets. These systematic biases are a result of many centage of such genes identified by chance, the false discovery rate different experimental features including different microarray (FDR). In addition, SAM correlates gene expression data to a wide range platforms and different production lots of microarrays. DWD is of clinical parameters, including treatment, diagnosis categories, and useful in merging/comparing two tumor microarray datasets survival time. completed on different microarray platforms. Supervised analysis: Analysis of gene expression profiling data in EASE (Expression Analysis Systematic Explorer): A which external information such as survival status is used as a guide to computer program used for rapid biologic interpretation of gene select genes from microarray data. lists that result from the analysis of microarray or other high- throughput genomic data. The main function of EASE is to per- form theme discovery for a given list of genes. Two-way contingency table: A tabular representation of cate- gorical data, it shows the frequencies for particular combinations of val- Gene ontology: Allows for annotating genes and their products ues of two discrete variables X and Y. Each cell in the table represents a with a limited set of attributes, with the three organizing principles being mutually exclusive combination of X-Y values. A ␹2-based P value can molecular function, biological process, and cellular component. The be calculated for a contingency table and is used to determine whether development of structured, controlled vocabularies (ontologies) that there is a statistically significant association between the two variables in describe gene products in terms of these organizing principles in a the contingency table. A 2 ϫ 2 contingency table refers to a table in species-independent manner is a constantly evolving process. which X and Y are binary variables.

1664 JOURNAL OF CLINICAL ONCOLOGY

JOURNAL OF CLINICAL ONCOLOGY EDITORIAL

Functional Analysis of the Breast Cancer Genome

Matthew J. Ellis, Division of Oncology, Department of Internal Medicine, Washington University in St Louis, St Louis, MO

Clinicians have long recognized that a diagnosis of breast lated genes. As an interesting side issue in the article, Oh et al cancer encompasses different tumor types with very different clin- compared this list of genes with a list induced by transfection of the ical outcomes. The first critical step toward a biomarker-based GATA3 transcription factor into a GATA3 null cell line and found subclassification schema to aid disease management occurred with considerable overlap. Thus GATA3 and ER must work in concert the widespread introduction of tumor estrogen-receptor (ER) to regulate this large gene repertoire. There was also a good deal of measurement three decades ago. It is worth revisiting the litera- overlap with genes found to be expressed in ERϩ breast cancers in ture of the time, because the introduction of ER testing was very earlier studies (referred to as “luminal” in keeping with Oh et al’s controversial. In particular, investigators worried that a benefit earlier work on breast cancer classification) as well as genes repre- from endocrine therapy in ER–disease could not be excluded.1 sented in the only clinically available multigene test that predicts It took the weight of a meta-analysis to resolve this issue,2 but tamoxifen resistance.7 even today we worry about excluding a patient from appropri- To ascertain the role these estrogen-induced genes might play ate endocrine therapy because of false-negative ER results. The in clinical outcomes, Oh et al first performed hierarchical cluster- single-sample prediction problem continues to be central to the ing of the estrogen-induced genes in a 64–luminal tumor training current debate regarding new and more complex biomarker set and identified two major groups, groups I and II. All of the gene approaches. Apparently robust group predictions typical of information available from these cases was subsequently compared these analyses does not necessarily require great measurement between groups I and II (a supervised analysis), with a 1% false- accuracy. Prospective clinical testing, on the other hand, needs discovery rate, to produce an 822-gene list. The tumors were then to be highly precise to avoid the potentially tragic consequences reclustered with the 822 differentially expressed genes to reproduce of tumor misclassification. two groups (called IE and IIE). Information on as many of the 822 Further subclassification for ERϩ disease to prospectively genes as possible was then sought from three publicly available, identify endocrine therapy–resistant ERϩ tumors is of longstand- clinically annotated breast cancer microarray databases. After data ing interest. The introduction of HER2 testing has already taken us manipulation to remove “platform and source systemic bias” an some way down this road. It is clear that the activation of HER2 by average expression index or “centroid” was used to classify clini- gene amplification drives endocrine therapy resistance,3 and the cally ERϩ and/or PgRϩ tumors into two groups with significant introduction of the HER2-targeting monoclonal antibody trastu- differences in survival. From a functional genomics standpoint, it zumab into the adjuvant setting improves outcomes for patients is interesting to note that the favorable-prognosis group IE gene list with ERϩ HER2ϩ tumors.4 HER2 gene amplification does not, was largely populated by transcription factors XBP1, FOXA1, and however, explain all or even most cases of endocrine therapy fail- PgR and ribosomal genes. Unfavorable Group IIE tumors were ure, and new insights and predictive models are needed. One rich in proliferation genes (Ki-67, MYBL2, Survivin, STK6, and longstanding idea is to consider the ER in the context of an entire CCNB2), as well as apoptosis regulators FLIP/CFLAR, AVEN, and pathway; as our insight into the role of estrogen in tumorigenesis BCL2A1. These gene signatures imply, not surprisingly, that endo- deepens, we will be able to move beyond reliance on progesterone crine therapy resistance is associated with a deregulated cell cycle receptor (PgR) measurement for ER activity assessment. The late and resistance to programmed cell death. William McGuire, who championed progesterone receptor mea- So what are we to make of these data? The first issue, of course, surement in breast cancer,5 would therefore have been very inter- is that Oh et al’s extensive gene list is far from a clinical test. ested in the paper by Oh et al6 in this issue of the Journal of Clinical Although the signature is robust, and survives the data averaging Oncology. The University of North Carolina–Chapel Hill (Chapel algorithms necessary for cross-platform assessment, clinical testing Hill, NC) team assessed the biomarker potential of hundreds of needs to be formulated into a reproducible technology that can be estrogen-regulated genes (including PgR) encoding proteins with a repeated thousands of times in a cost-effective manner. Here, we wide variety of cellular functions. are faced with a dilemma that is becoming more acute every day. To briefly summarize their findings, Oh et al identified an Although the systematic evaluation of expression from 822 genes extensive list of genes that were found to be regulated by estradiol at in a clinical environment sounds expensive, actually, by using multiple time points in the MCF7 breast cancer cell line. Although smaller customized microarrays, it could be achieved for a couple they found both estrogen-repressed genes and estrogen-induced hundred dollars with a turnaround of 1 or 2 days. However, the genes, their subsequent work in the paper focused on only upregu- starting material for RNA extraction (at least currently) must be a

Journal of Clinical Oncology, Vol 24, No 11 (April 10), 2006: pp 1649-1650 1649 DOI: 10.1200/JCO.2005.05.3520 Information downloaded from www.jco.org and provided by UNIV OF NC/ACQ SRVCS on April 8, 2006 from 152.2.181.221. Copyright © 2006 by the American Society of Clinical Oncology. All rights reserved. Matthew J. Ellis snap-frozen tumor biopsy whose quality and tumor content have for this vision has only just begun, and the excitement in the breast been assessed by histopathology. On the other hand, the measure- cancer research community is palpable. ment of 822 genes by quantitative polymerase chain reaction from RNA extracted from a paraffin block is impractical at this stage, and © 2006 by American Society of Clinical Oncology a step to reduce the gene number closer to 100 or so without loss of predictive abilities would be required for the clinical implementa- REFERENCES 1. Chamness GC, Mercer WD, McGuire WL: Are histochemical methods for tion of Oh et al’s signature. We may have to face the fact that, like estrogen receptor valid? J Histochem Cytochem 28:792-798, 1980 our colleagues who treat lymphoma, breast cancer physicians will 2. Tamoxifen for early breast cancer: An overview of the randomised trials— have to insist on frozen tumor acquisition in the near future. We Early Breast Cancer Trialists’ Collaborative Group. Lancet 351:1451-1467, 1998 achieved fresh tissue acquisition for biochemical ER testing before 3. De Laurentiis M, Arpino G, Massarelli E, et al: A meta-analysis on the interaction between HER-2 expression and response to endocrine treatment in and I predict that, despite the barriers, we will be doing it again. An advanced breast cancer. Clin Cancer Res 11:4741-4748, 2005 obvious move in this direction is the prospective evaluation of the 4. Piccart-Gebhart MJ, Procter M, Leyland-Jones B, et al: Trastuzumab after Amsterdam 70-gene prognostic signature that requires frozen tis- adjuvant chemotherapy in HER2-positive breast cancer. N Engl J Med 353:1659- 8 1672, 2005 sue RNA analysis. 5. McGuire WL, Horwitz KB, Pearson OH, et al: Current status of estrogen and One of the most interesting aspects of the Oh et al article con- progesterone receptors in breast cancer. Cancer 39:2934-2947, 1977 cerns the glimpses of the molecular complexities of the biologic system 6. Oh DS, Troester MA, Usary J, et al: Estrogen-regulated genes predict that we are trying to understand. Certain genes, such as STK6, keep survival in hormone receptor-positive breast cancers. J Clin Oncol 24:10.1200/ JCO.2005.03.2755 surfacing in breast cancer profiling studies. This gene is present in the 7. Paik S, Shak S, Tang G, et al: A multigene assay to predict recurrence of National Surgical Adjuvant Breast and Bowel Project (NSABP) recur- tamoxifen-treated, node-negative breast cancer. N Engl J Med 351:2817-2826, 2004 rence score published by Paik et al.7 As a side note, gene nomenclature 8. van de Vijver MJ, He YD, van’t Veer LJ, et al: A gene-expression signature is extremely frustrating: STK6 is actually assigned to the mouse gene, as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002 9. Zhou H, Kuang J, Zhong L, et al: Tumour amplified kinase STK15/BTAK whereas the human homolog is Aurora-A/AURKA/STK15/ induces centrosome amplification, aneuploidy and transformation. Nat Genet BTAK. STK15 is deregulated in a wide spectrum of poor-prognosis 20:189-193, 1998 cancers9-11 and of interest, Aurora kinase inhibitors are in development.12 10. Forozan F, Mahlamaki EH, Monni O, et al: Comparative genomic hybrid- ization analysis of 38 breast cancer cell lines: A basis for interpreting comple- The power of the HER2-trastuzumab paradigm is that we mentary DNA microarray data. Cancer Res 60:4519-4525, 2000 have successfully paired a somatic mutation with a targeted ther- 11. Tanner MM, Grenman S, Koul A, et al: Frequent amplification of chromo- apy. It seems likely that the final formulation of a biologically based somal region 20q12-q13 in ovarian cancer. Clin Cancer Res 6:1833-1839, 2000 classification of breast cancer will have a strong bias toward genetic 12. Gadea BB, Ruderman JV: Aurora kinase inhibitor ZM447439 blocks chromosome-induced spindle assembly, the completion of chromosome con- abnormalities that create opportunities for targeted treatments. densation, and the establishment of the spindle integrity checkpoint in Xenopus The functional annotation of the breast cancer genome required egg extracts. Mol Biol Cell 16:1305-1318, 2005

■■■

Acknowledgment Supported by National Cancer Institute (Bethesda, MD) Grants No. U01 CA114722 and R01 CA095614.

Author’s Disclosures of Potential Conﬂicts of Interest The author indicated no potential conﬂicts of interest. Author Contributions

Manuscript writing: Matthew J. Ellis Final approval of manuscript: Matthew J. Ellis

1650 JOURNAL OF CLINICAL ONCOLOGY