To identify genes associated with survival from antiestrogens, both serial analysis of geneexpression and gene expression microarrays were used to explore the transcriptomes of antiestrogen-responsive (MCF7/LCC1) and -resistant variants(MCF7/LCC9) of the MCF-7 human breast cancer cell line. Structure of the gene microarray expression data was visualized at the top level using a novel algorithm that derives the first three principal components,fitted to the antiestrogen-resistant and -responsive gene expression data, from Fisher’s information matrix. The differential regulation of several candidate genes was confirmed. Functional studies of the basal expression and endocrine regulation of transcriptional activation of implicated transcription factors were studied using promoter-reporter assays.
The putative tumor suppressor interferon regulatory factor-1 is down-regulated in resistant cells, whereas its nucleolar phosphoprotein inhibitor nucleophosmin is up-regulated. Resistant cells also up-regulate the transcriptional activation of cyclic AMP response element (CRE) binding and nuclear factor κB (NFκB) while down-regulating epidermal growth factor receptor protein expression. Inhibition of NFκB activity by ICI 182,780 is lost in resistant cells, but CRE activity is not regulated by ICI 182,780 in either responsive or resistant cells. Parthenolide, a potent and specific inhibitor of NFκB, inhibits the anchorage-dependent proliferation of antiestrogen-resistant but not antiestrogen-responsive cells. This observation implies a greater reliance on their increased NFκB signaling for proliferation in cells that have survived prolonged exposure to ICI 182,780.
These data from serial analysis of gene expression and gene microarray studies implicate changes in a novel signaling pathway, involving interferon regulatory factor-1, nucleophosmin, NFκB, and CRE binding in cell survival after antiestrogen exposure. Cells can up-regulate some estrogen-responsive genes while concurrently losing the ability of antiestrogens to regulate their expression. Signaling pathways that are not regulated by estrogens also can be up-regulated. Thus, some breast cancer cells may survive antiestrogen treatment by bypassing specific growth inhibitory signals induced by antagonist-occupied estrogen receptors.
ERs 7 are nuclear transcription factors, their activities being affected by the nature of the ligand bound and the pattern of genes/proteins expressed within cells (cellular context; Ref. 1 ). Antiestrogens compete with endogenous estrogens for activation of ER, and induce both cell cycle arrest and apoptosis in responsive cells (2) . Neither the genes regulated by antiestrogens that signal to apoptosis nor those genes that confer an acquired antiestrogen resistance have been identified. Nonetheless, antiestrogenic drugs are effective in both premenopausal and postmenopausal breast cancer patients, and in the metastatic and adjuvant settings (3) . The most widely used antiestrogen in current clinical practice is the triphenylethylene TAM. Clinical experience with this drug likely now exceeds 10 million patient years. When patients with metastatic disease are selected for treatment based on the ER and PgR content of their tumors, responses are seen in up to 75% of tumors expressing both receptors (2) . TAM also reduces the incidence of ER-positive breast cancers in high risk women (4) .
Other antiestrogens have emerged recently, most notably the benzothiophene Raloxifene and the steroidal ICI 182,780 (Faslodex). Both drugs appear to have significant clinical activity and may have better toxicological profiles when compared with TAM (2) . Faslodex has significant activity in TAM-resistant patients (5) , consistent with data obtained previously with TAM-resistant human breast cancer cells selected in vitro (6) .
Despite the utility of antiestrogens, most tumors that initially respond to these drugs will recur and require alternative systemic therapies (2) . Unfortunately, the precise mechanisms that confer resistance remain unknown. Change to an antiestrogen-stimulated phenotype has been described in some animal models (6 , 7) . This phenotype may occur in up to 20% of breast cancer patients but a loss of responsiveness to antiestrogens may be the more common phenotype (2) . The expression of mutant ER proteins and splice variants has been reported but the functional role of these in endocrine resistance remains unclear (2) . Most tumors acquiring antiestrogen resistance do so while retaining expression of ER (8) . Thus, whereas lack of ER expression is a major form of de novo antiestrogen resistance, other mechanisms must be active in most instances of acquired resistance (2) . The persistent expression of ER in tumors with acquired resistance suggests that some cells expressing this phenotype may either require ER expression and/or reflect the altered expression of otherwise estrogen-regulated genes.
Because ER-mediated transcription is directly affected by antiestrogens, we initially hypothesized that antiestrogen resistance might include perturbations in the patterns of expression and/or regulation of a subset of all of the ER-regulated genes (1) . To address this hypothesis, we first generated a novel series of human breast cancer variants from the MCF-7 human breast cancer cell line. These cells have different growth requirements for estrogen and exhibit differential sensitivities to TAM and ICI 182,780 (9, 10, 11) . In this study, we focus on MCF7/LCC1 cells (estrogen-independent, TAM-responsive, and ICI 182,780 responsive) and MCF7/LCC9 cells (estrogen-independent, ICI 182,780 resistant, and TAM cross-resistant; Ref. 11 ). Because the cells exhibit comparable cell cycle profiles 8 and are both MCF-7 variants, we can exclude the altered expression of genes related solely to differences in both genetic background and cell cycle distribution. A direct comparison of these respective transcriptomes should identify genes associated with survival from long-term antiestrogen exposure.
Several techniques are now available to explore the transcriptomes of tumors and experimental models. However, the most effective approach remains a matter of debate (12) . Studies in breast cancer have been limited, most simply attempting to identify the genes expressed in breast cancers. For example, a recent study by Perou et al. (13) explored data from excisional breast biopsies from 42 individuals. Gene clusters, identified by exploration of the data structure, include those associated with ER, HER-2, and IFN-induced genes. A similar cluster of IFN-regulated genes was identified in the breast cancer cell lines included in the NIH drug screening program (14) . Studies comparing the gene expression profiles of specific breast cancer phenotypes include an examination of histologically different samples from a single breast cancer lesion (15) and a preliminary analysis of a TAM-stimulated xenograft model (16) . None of these reports directly addressed either the function or potential role of the specific genes identified. We have used two different but complementary approaches, SAGE and gene expression microarrays. These approaches would not be expected to provide identical data because not all of the genes identified by SAGE are on the microarrays, some genes identified on the cDNA arrays may be confounded by cross-hybridization to homologous RNAs, and the ability to detect significant differences between the SAGE databases is affected by the relative abundance of the tags and the size of the databases. We approached both technologies as means to sample the transcriptomes of MCF7/LCC1 and MCF7/LCC9 cells, and to generate data that would allow us to begin testing our hypothesis implicating estrogen-regulated genes in antiestrogen resistance. We now show that cells can survive prolonged antiestrogen treatment by altering the expression, patterns of regulation, and functional activation of specific estrogen-regulated genes.
MATERIALS AND METHODS
MCF7/LCC1 cells were derived from the estrogen-dependent MCF-7 human breast cancer cell line after selection for growth in ovariectomized nude mice (9 , 17) . MCF-7/LCC9 cells were obtained by an in vitro stepwise selection of the estrogen-independent but antiestrogen-responsive MCF7/LCC1 cells against the steroidal antiestrogen ICI 182,780 (Faslodex). MCF7/LCC9 cells are ICI 182,780 resistant and TAM cross-resistant, express ER and PgR, and exhibit an estrogen-independent but responsive phenotype (11) . MCF7/LCC1 and MCF7/LCC9 cells were routinely passaged in Improved Minimal Essential Medium without phenol red (Biofluids, Bethesda, MD) supplemented with 5% CCS-IMEM. Serum was stripped of endogenous estrogens as described previously and is estimated to contain ≤10 fM estrogen (18) . Vehicle for all of the hormone/antihormone treatments was ethanol (final concentration <0.1% v/v). All of the cell cultures were maintained at 37°C in a humidified 5% CO2:95% air atmosphere and shown to be free of contamination with Mycoplasma species as determined by solution hybridization to Mycoplasma-specific, radiolabeled, RNase riboprobes (Gen-Probe Inc., San Diego, CA).
SAGE was performed as described previously (19) . Polyadenylic acid mRNA was harvested from cells using biotin labeled-oligodeoxythymidylic acid magnetic beads (Promega PolyATract System 1000 kit; Promega, Madison, WI) and treated with DNase I enzyme to remove any contaminating DNA. mRNA (5 μg) was converted to double-stranded cDNA using the Life Technologies, Inc. cDNA Synthesis kit (Life Technologies, Inc., Rockville, MD). Biotinylated cDNA was completely cleaved with Nia III and the 3′-end digested fragments extracted with magnetic streptavidin beads. The cDNA was evenly divided and ligated, one half to linker A and the other half to linker B (19) . Cleavage of the cDNA by BsmF1 produced 11–13 bp oligo DNA tags with linkers, which were blunt-ended with T4 polymerase. Linkers A and B were ligated together to form ditags, which were then amplified by PCR using primers to linkers A and B. Ditags (22–26 bp) were gel purified and ligated into concatenated polytags. The polytags were purified and cloned into the Sph1-digested pZeor1 vector, which was transferred to competent TOP10F′ cells by electroporation. Positive clones were selected overnight at 37°C for growth on low-salt Luria-Bertani bacterial plates supplemented with Luria-Bertani-Zeocin (50 μg/ml) and isopropyl β-d-thiogalactopyranoside (1 mm). Colonies were screened for plasmids containing appropriate inserts by size fractionating PCR products, obtained using M13 forward and reverse primers, in agarose gels. PCR products containing concatamers of >600 bp were purified and sequenced.
Characteristics of the SAGE databases are shown in Table 1 ⇓ . We compared the MCF7/LCC1 and MCF7/LCC9 databases, using the SAGE version 1.00 software (kindly provided by Dr. K. W. Kinzler, Johns Hopkins University, Baltimore, MD), to identify putatively differentially expressed genes. Only a representative sample of these can be presented. The genes presented in Table 2 ⇓ were primarily selected based on: (a) fold difference ≳2-fold; (b) that the Tags compared should represent ≤2 genes; and (c) that a Tag found in either the MCF7/LCC1 and/or MCF7/LCC9 SAGE libraries must represent ≳0.10% of the database. Evidence that a gene was already known to be expressed in breast cancers also was considered. None of these criteria were considered an absolute requirement for gene selection. Whereas 2-fold was selected as the cutoff, biologically critical events can be controlled by genes that exhibit a fold regulation as small as 50% (20) . As described recently by Man et al. (21) , χ2 analyses were used to compare the proportions of specific tags in each database.
RNA Isolation, Generation of Probes, and Hybridization of Gene Microarrays.
Each probe was generated from an independent cell culture, each culture being grown on a different day but using identical cell culture conditions. Six MCF7/LCC1 and five MCF7/LCC9 cell cultures were used. RNA was isolated from proliferating, subconfluent monolayers of each cell line using the TRIzol reagent (Life Technologies, Inc., Grand Island, NY). RNA quality was determined by standard spectroscopic and gel electrophoresis analyses.
Probes for the Clontech Atlas gene microarrays (Clontech, Palo Alto, CA) were made as described by the manufacturer. Briefly, 1 μg of Dnase-treated mRNA was primed with the Clontech cDNA Synthesis Primer mix. The product was reverse transcribed into radiolabeled cDNA with [γ-32P]dATP (Amersham Life Science Inc., Arlington Heights, IL), and the reaction incubated at 50°C for 25 min and terminated by adding 0.1 m EDTA (pH 8.0). Radiolabeled cDNA was purified and eluted through a NucleoSpin Extraction Column (centrifuged at 14,000 rpm). The cDNA probe was denatured with 1 m NaOH and 10 mm EDTA, and incubated at 68°C for 20 min. c0t-1 DNA and 1 m NaH2PO4 (pH 7.0) were added to the denatured probe, and incubated at 68°C for an additional 10 min.
Each Atlas Array (Clontech) was prehybridized with 5 ml of ExpressHyb buffer (Clontech) and 0.5 mg of denatured DNA from sheared salmon testes at 68°C for 30 min with continuous agitation. The cDNA probe, prepared as described above, was then added and allowed to hybridize overnight. The array was washed four times with 2× SSC containing 1% (w/v) SDS for 30 min at 68°C and once with 0.1× SSC containing 0.5% (w/v) SDS for 30 min at 68°C. One final wash was performed with 2× SSC for 5 min at room temperature. The Atlas Array was sealed in plastic and signals detected by phosphorimage analysis using a Molecular Dynamics Storm phosphorimager (Molecular Dynamics, Sunnyvale, CA). Each filter was used only once.
Measuring NPM and EGF-R Protein Levels.
Established methods were used for performing and quantifying Western analyses of NPM (22 , 23) . Briefly, 10 μg of protein was loaded onto an SDS-PAGE gel and fractionated under reducing conditions [5% (v/v) β-mercaptoethanol]. To account for within-gel differences, samples were loaded in a random sequence onto each gel. Proteins were blotted onto nitrocellulose membrane and the blots probed with an anti-NPM monoclonal antibody (kindly provided by Dr. Pui-Kwong Chan, Baylor College of Medicine, Houston, TX; Ref. 24 ). After transfer to the membranes, equal protein loading was confirmed by staining the nitrocellulose with Ponceau S as is widely reported (22 , 23 , 25) . Any material remaining in the gels were stained by Coomassie Blue. This approach provides an adequate and appropriate estimate for equivalence of protein loading (22 , 23 , 25) . Immunoreactivity was visualized using a horseradish peroxidase-linked goat antimouse IgG and the enhanced chemiluminescence detection system (Amersham Life Science Inc.). Chemiluminescence was densitometrically measured using a Quantity One Scanning and Analysis System (pdi, Huntingdon, NY).
EGF-R is expressed at low levels in MCF-7 cells and cannot readily be detected/quantified by Western blot. Consequently, we measured immunofluorescently labeled EGF-R protein by FACS. For each cell line, EGF-R immunofluorescence was performed by rinsing 5 × 106 cells once in PBS and pelleting cells by centrifugation at 1000 rpm for 5 min at room temperature. Cell pellets were resuspended in 100 μl of an anti-EGF-R mouse monoclonal antibody that recognizes the extracellular domain of the receptor (EGF-R antibody-1; NeoMarkers, Lab Vision Corp., Fremont, CA; 200 μg/ml diluted 1:50 in PBS), and incubated at room temperature for 1 h. Cell pellets were then resuspended in 1:50 dilution of R-phycoerythrin-conjugated goat antimouse IgG-2a (CALTAG Laboratories, Burlingame, CA) and incubated in the dark for 30 min. After rinsing in PBS, cells were again pelleted, fixed by resuspending in 1% paraformaldehyde, and fluorescence measured by FACS. Control cells were treated either with secondary antibody alone or with no antibody. FACS was performed on a FACStarPlus flow cytometer (Becton-Dickinson, Mountain View, CA) at 488 nm.
RNase Protection Analysis of IFN Regulatory Factor-1 mRNA Expression.
Total RNA was isolated using the TRIzol reagent (Life Technologies, Inc.) according to the manufacturer’s instructions. The IRF-1 riboprobe was made by in vitro transcription of a 360-bp fragment of the IRF-1 cDNA. The 36B4 loading control riboprobe was similarly obtained from a 220-bp fragment of the 36B4 cDNA (17) . Riboprobes were labeled by the addition of [32P]UTP (Amersham Life Sciences Inc.) in the transcription buffer. To achieve bands for the two genes with similar intensities, the 36B4 riboprobe was made with a specific activity of ∼20% that of the IRF-1 riboprobe. The RNase protection assays were performed as described previously (26) . Briefly, total RNA (30 μg), the IRF-1 riboprobe, and the 36B4 riboprobe were hybridized overnight at 50°C. After digestion with RNase A, the protected fragments were size fractionated on 6% acrylamide Tris-borate EDTA-urea minigels (Novex, San Diego, CA). The gels were dried and the respective signals quantified by phosphorimager analysis (Molecular Dynamics).
Estimation of the Transcriptional Activation of CREs and NFκB.
Two commercially available promoter-reporter assays were used to measure NFκB and CRE transcriptional activities. Experiments were performed as described by the manufacturer (Stratagene, La Jolla, CA). Briefly, firefly luciferase reporter constructs, under the control of the appropriate enhancer elements and trans-activator constructs, were provided in the PathDetect in vivo signal transduction pathway cis-reporting system (Stratagene). Cells were grown to 90% confluence in 5% CCS-IMEM medium and seeded at 8 × 104 cells into each well of 24-well tissue culture dishes. After incubation for 12–24 h, cells were transiently transfected with the appropriate plasmids using the Qiagen Superfect transfection reagent as described by the manufacturer (Qiagen, Valencia, CA). The ratio of plasmid to Superfect reagent was 250 ng:1 μl, with a transfection time of 2.5 h.
Estrogen (5 nm) and ICI 182,780 treatments (10 nm) were administered for 48 h after transfection in CCS-IMEM. Transfected cells were harvested and firefly luciferase activity measured using the Stratagene assay system. Activity is expressed in relative light units from a 20-μl sample as detected by luminometry. Each measurement is from duplicate samples, independent experiments being repeated on different days. Normalization of transfection efficiency was made to the Renilla luciferase reporter construct, under the control of the cytomegalovirus promoter (Promega). The Renilla luciferase assay was performed using the Promega Dual-luciferase reporter assay system.
Assessment of Growth Response to Parthenolide.
MCF7/LCC1 and MCF7/LCC9 cells were plated in 96-well tissue culture plates and incubated for 24 h in 0.2 ml of 5% CCS-IMEM. Medium was removed and replaced with fresh 5% CCS-IMEM containing either vehicle (0.1% DMSO) or parthenolide (300 nm and 600 nm). Cells were refed every third day with the appropriate cell culture medium. Cell growth was determined on day 6, using a crystal violet assay where dye uptake is directly related to cell number (27) . Cells were incubated for 30 min with crystal violet stain [0.5% (w/v) crystal violet in 25% (v/v) methanol] at 25°C. Unincorporated stain was removed with deionized water and the cells allowed to dry at room temperature. Incorporated dye was extracted into 0.1 ml of 0.1 m sodium citrate in 50% (v/v) ethanol for 10–15 min at room temperature. Absorbance was read at 570 nm using a Molecular Devices Vmax kinetic microplate reader.
Statistical Analyses and Analysis of Gene Expression Microarray Data.
t tests were used to compare control and experimental groups as appropriate for the RNase protection, Western blot, promoter-reporter, and cell proliferation assays. All of the tests were two-tailed, with statistical significance established at P ≤ 0.05, unless stated otherwise.
For the gene array studies, background signal was estimated locally and subtracted from the signal obtained from its target cDNA, producing the background-corrected data. These corrections were done using the algorithms in Pathways 4.0 (Research Genetics Inc., Huntsville, AL). Background-corrected data were normalized to account for differences in probe-specific activity, hybridization, and other variables among replicates (28) . Normalization was accomplished using the mean value of all of the background-corrected signals on each array.
Different approaches have been used to analyze data from gene array studies. Some methods are simply based on fold-regulation (29) , others are more statistically based (16 , 30) , and/or apply an informatics-based exploration of data structure (31 , 32) . The optimal approach remains a subject of considerable debate (30) . As with most gene microarray studies, our data set is high in dimensionality (597 dimensions) but the number of replicates is limited by the resource-intensive nature of the technology. The relatively few replicates limits the applicability of normal mixture models and other analyses that can operate in high dimensional data space (33 , 34) and often generates noisy data sets.
Previously, we have reported a hierarchical visualization algorithm that can reveal all of the major aspects of the multimodal data points, which concurrently exist in a high dimensional gene expression space (35 , 36) . Using this algorithm, our data can be projected from 597 dimensions to two or three dimensions (multidimensional scaling). This is accomplished by respectively deriving the first three principal components fitted to the antiestrogen responsive (MCF7/LCC1) and resistant (MCF7/LCC9) gene expression data (Fig. 1) ⇓ . Thus, we evaluate the data structure subsets visually and assess whether these contain differentially expressed genes that may contribute to the respective phenotypes.
Because we can visualize data structure, our next priority was to identify a simple, supervised approach for reducing the dimensionality of the data without affecting its structure. Thus, we applied geometric and simple descriptive statistical approaches to the normalized data before and after a logarithmic transformation of these data. As noted previously, the distribution of the expression data for each gene is unknown (30) , and it is unclear whether these violate the normal distribution required for parametric analyses. Indeed, it seems likely that the distribution assumption required will be normal for some genes and not for others. Whereas most investigators analyze data transformed by a logarithmic function, those genes with values that appear normally distributed before transformation may no longer have this distribution once transformed.
To be inclusive, we used simple statistics (t tests) to explore the data. The inflated type-1 error from multiple comparisons should overestimate (false positive) significant differences. We considered this preferable to a high incidence of false-negative estimates, which would lead to the exclusion of potentially informative genes. The inclusion of uninformative genes (false negatives) is less problematic at this stage of the exploration. We used Student’s t test, a t test for unequal variance (assumes normal distribution) and the nonparametric (distribution-free) Wilcoxon signed rank test. Logarithm transformed and nontransformed data were explored. This approach is similar to using a F test as described recently by Hedenfalk et al. (37) .
t test results were evaluated and candidate genes selected with which to reconstruct a lower dimensional data set that should retain most of the information apparent in the top level visualization. However, the t test results were only one of several criteria used to guide gene selection, and only a subset of those genes that appear to be differentially regulated are presented. These genes were selected by comparing the results of t tests on logarithm transformed and untransformed data, fold-regulation (∼2-fold or greater was selected because this difference is likely to be confirmed in independent analyses), the distribution of the background-corrected and normalized data for each gene (some genes appeared strongly differentially regulated but did not generate statistically significant differences because of heterogeneity in the data), and the probable relevance to breast cancer of each gene.
Where the gene subsets (reduced dimensional data) provide a reasonable description of the entire expression data, the replicate profiles of the resistant and responsive cells should exist in separable data space (35 , 36) . Furthermore, if the profiles are adequately defined by a small, rational gene subset, some of its members likely represent differentially expressed and functionally relevant genes. We acknowledge that our approach is limited, and is probably only applicable to simple comparisons within related cell culture models.
Genes Implicated by SAGE.
The data in Table 1 ⇓ show the number of different genes identified. Most genes were commonly expressed, and were not differentially expressed between the MCF7/LCC1 and MCF7/LCC9 cells. A selection of the genes identified by SAGE, and predicted to be differentially expressed in MCF7/LCC1 and MCF7/LCC9 SAGE databases, is shown in Table 2 ⇓ . Presentation of all of the genes expressed and/or differentially expressed is beyond the scope of a single, focused study. 9 The criteria applied for gene selection are described in “Materials and Methods.” NPM was included because we already know it to be both estrogen regulated (23) and indirectly associated with TAM treatment in patients (38) . Confirmation of the differential expression of NPM (see Table 2 ⇓ and Fig. 2B ⇓ ) and altered CRE binding activity (the function of XBP-1; see Table 2 ⇓ and Fig. 3B ⇓ ) indicate that these represent reasonable criteria for gene selection. Currently, the XBP-1 and NPM are the only genes from the SAGE database comparisons for which we have attempted to confirm differential expression/activation.
Comparing the SAGE databases identifies several genes that are up-regulated in MCF7/LCC9 cells compared with MCF7/LCC1 cells. These genes include XBP-1, NPM, cathepsin D, HSP-27, and n-ras. Increased CRE activity is indicated by the up-regulation of XBP-1, which regulates gene transcription through these response elements (39) . XBP-1 is involved in regulating the expression of several tissue-specific genes including tissue inhibitor of metalloproteinases, osteopontin, and osteocalcin (40) . Significantly, both Perou et al. (13) and West et al. (41) recently identified XBP-1 as being associated with ER gene expression clusters in human breast tumor biopsies. NPM is induced by estrogen in MCF-7 cells and is up-regulated in estrogen-independent cells (23) . NPM also provokes an autoimmune response in breast cancer patients, the magnitude of which is associated with TAM therapy (38) .
The altered expression of cathepsin D is consistent with our data published previously, showing increased secretion of this protein in several of our hormone-independent MCF-7 variants (42) . Cathepsin D expression in breast tumors also is associated, at least in some studies, with a poor prognosis (43) . HSP-27 expression has been implicated in refining the diagnosis of suspicious fine-needle aspirates of breast tissues (44) . Vitamin B12 binding proteins are expressed in breast tumors (45) , and vitamin B12 deficiency is a likely risk factor for breast cancer (46) . Altered expression of the n-ras-related gene is consistent with the elevated ras signaling reported in some breast cancer cell lines and tumors (47) .
SAGE also identified genes expressed at higher levels in the parental, antiestrogen-responsive cells (MCF7/LCC1) when compared with MCF7/LCC9 cells. These include ferritin, death-associated protein-6, and the eukaryotic elongation factor-γ. Ferritin is expressed in breast cancers, and breast tumor-derived ferritin may be a more useful tumor marker than measuring levels of ferritin in serum (48) .
Structure of the Gene Microarray Data.
It has been suggested that the cost required to perform gene microarray studies can be reduced by combining RNA populations from several replicates and performing a single hybridization on an Atlas array (16) . However, we found heterogeneity among replicate experiments, which often remained after normalization. Logarithmic transformation of these data reduced this heterogeneity but not to the point where a single replicate could be used to obtain an adequate description of the data. Consequently, multiple replicates are required to provide a more reliable estimate of the putative gene expression profiles. These observations on filter microarrays are consistent with recent reports for glass slide-based and oligonucleotide array-based gene expression microarrays (49 , 50) .
Fig. 1A ⇓ is a visual representation of the multidimensional data (597 dimensions) in three dimensions. This visualization allows for an inspection of the data structure, and the likely comparability of the replicates among each other and between the two experimental groups (antiestrogen-responsive MCF7/LCC1 and antiestrogen-resistant MCF7/LCC9). For this top level visualization, the replicate gene expression profiles for MCF7/LCC1 and MCF7/LCC9 exist within linearly separable regions of the gene expression data space after elimination of one outlier array from each experimental group. The top three principal components capture 81.2% of the cumulative variance in the data (597 dimensions). Thus, the data structure is consistent with differences in the gene expression profiles as predicted by the known differential antiestrogen responsiveness of the two variants.
Genes Implicated by Gene Microarray Studies.
The data in Table 3 ⇓ show the fold-differences in expression of selected genes identified in the Clontech Atlas gene microarray studies selected using the criteria described in “Materials and Methods.” The selection was not intended to describe fully the data set, only to assist in an initial exploration of the data. This small but rational subset of genes could be additionally evaluated in focused studies to confirm the differential expression patterns and establish potential functional relevance. Furthermore, if members of this subset were truly differentially expressed, we could begin to understand how cells perceive antiestrogens and adapt to this selective pressure.
To determine whether these genes are broadly representative of the differences between the gene expression profiles of MCF7/LCC1 and MCF7/LCC9 cells, we generated a three-dimensional projection from the seven-dimensional gene expression data space (Fig. 1B) ⇓ . This was necessary because we used several criteria to construct the subset, including some genes where fold-regulation or distribution of the data were given more weight than formal statistical significance. Consequently, we could not assume that we had maintained the linear separability of the data, at the top level, as seen in all 597 dimensions.
We might not expect this small subset of expression data (<2% of the information) to prove as effective in representing the respective phenotypes as the full data set (597 genes). Nonetheless, as for the 597-dimension visualization, after elimination of outlier data the seven-dimensional MCF7/LCC1 and MCF7/LCC9 profiles remain in linearly separable, three-dimensional data space. The top three principal components capture 98.9% of the cumulative variance in the data (seven-dimensions). This observation suggests that these data contain information that contributes to the differences in the molecular profiles of these two variants, that these genes may contribute to the respective biological phenotypes, and that additional studies of their potential functional relevance are warranted.
Genes expressed at a higher level in the MCF7/LCC1 cells include EGF-R, EGR-1, IRF-1, and both TNFα and its R1 receptor (TNF-R1). A well-established inverse relationship exists between the expression of EGF-R and ER in breast tumors (51) . EGF-R can induce expression of EGR-1 (52) , and expression of both genes is lower in MCF-7/LCC9 cells. EGR-1 is a transcription factor with proapoptotic activity (53) that can block NFκB function (54) and repress TGF-β receptor expression (29) . EGR-1 expression is down-regulated in 7,12-dimethylbenz(a)anthracene-induced mammary adenocarcinomas in rats (55) . IRF-1 is an IFN-regulated transcription factor that functions as a tumor suppressor gene (56 , 57) and is induced by TNFα (58) . A TNFα-mediated pathway for signaling to apoptosis occurs in MCF-7 human breast cancer cells (59 , 60) , and measuring serum TNF concentrations may be a useful prognostic marker in breast cancer patients (61) . Furthermore, HER-2/neu can block resistance to TNFα-induced apoptosis in breast cancer cells, using a mechanism that involves activation of NFκB (62) . We have previously implicated overexpression of superoxide dismutase in resistance to TNFα in MCF-7 cells (63) . Superoxide dismutase appears to be up-regulated in MCF7/LCC9 cells (Table 3) ⇓ and in TAM-stimulated MCF-7 xenografts (64) . NFκB (p65/RelA) appears expressed at higher levels in MCF7/LCC9 cells. NFκB is overexpressed in ER-negative breast cancer cells (65) and has an important role in the development of the normal mammary gland (66) .
NPM, EGF-R, and IRF-1 Are Differentially Expressed in MCF7/LCC1 and MCF7/LCC9 Cells.
The data in Table 2 ⇓ and Table 3 ⇓ predict differential expression of NPM, EGF-R, and IRF-1 between MCF7/LCC1 and MCF7/LCC9 cells. To confirm these observations, we measured the levels of the EGF-R (immunofluorescence) and NPM proteins (Western blot) and IRF-1 mRNA (RNase protection). The data in Fig. 2A ⇓ show that MFC7/LCC9 cells express lower amounts of EGF-R than MCF-7/LCC1 cells. NPM protein expression is significantly increased in MCF7/LCC9 cells compared with MCF7/LCC1 cells (Fig. 2B ⇓ ; P < 0.02), consistent with the predicted data from the SAGE analyses (Table 2) ⇓ and our previous studies (23 , 38) . The higher levels of IRF-1 mRNA, seen in the antiestrogen-responsive MCF7/LCC1 cells in Table 3 ⇓ , are confirmed by RNase protection analysis (Fig. 2C ⇓ ; P = 0.005). Both the gene microarray and RNase protection analyses show an ∼2-fold higher level of IRF-1 expression in MCF7/LCC1 cells, when compared with the antiestrogen-resistant MCF7/LCC9 cells.
Transcriptional Regulatory Activities of NFκB and CRE Are Increased in MCF7/LCC9 Cells.
The increased expression of NFκB (gene expression microarray) and XBP-1 (SAGE) imply increased transcriptional activation of promoters containing NFκB and CRE response elements, respectively. We confirmed these observations directly, using commercially available promoter-reporter assays to measure transcriptional activities. The data in Fig. 3 ⇓ show that the basal activity of both promoters is increased in MCF7/LCC9 cells; ∼10-fold for NFκB and 4-fold for CRE (P < 0.02). The increase in transcriptional activation of the NFκB constructs is greater than that predicted by the gene array data, but mRNA, protein, and protein/DNA binding activities can be poor predictors of the functional activation of some transcription factors (67) . This prediction is not problematic for XBP-1, where the 4-fold increase in mRNA expression identified by SAGE (Table 2) ⇓ compares well with the 4-fold increase in basal transcriptional activation (Fig. 3B) ⇓ .
We next assessed whether ICI 182,780, the antiestrogen used to generate the MCF7/LCC9 cells, could regulate the transcriptional activities of NFκB and CRE. Whereas ICI 182,780 inhibits NFκB activity in the MCF7/LCC1 cells (TAM- and ICI 182,780-responsive), this regulation is lost in the TAM and ICI 182,780 cross-resistant MCF7/LCC9 cells (Fig. 4A) ⇓ . In contrast, ICI 182,780 treatment does not alter the transcriptional regulatory activities of the CRE promoter in any of these variants (Fig. 4B) ⇓ .
MCF7/LCC9 Cells Are Specifically Responsive to an Inhibitor of NFκB Activity.
The increased activation of NFκB and loss of its estrogenic regulation in MCF7/LCC9 cells suggests that these cells might now be partly dependent on NFκB signaling for survival/growth. Consequently, we compared the growth response of MCF7/LCC1 and MCF7/LCC9 cells to parthenolide, a potent and specific inhibitor of NFκB that can inhibit the inhibitor of NFκB kinase repressor of NFκB (68 , 69) and also binds NFκB in a highly stereospecific manner to block DNA binding (70) . Parthenolide produces a dose-dependent inhibition of MCF7/LCC9 cells, with an apparent IC50 of ∼600 nm (Fig. 5) ⇓ . In contrast, parthenolide does not significantly affect growth of MCF7/LCC1 cells at these concentrations. MCF7/LCC9 cells are significantly more dependent on the transcriptional regulatory activities of NFκB than their ICI 182,780-responsive parental cells (P < 0.01 for MCF7/LCC9 versus MCF7/LCC1 at both 300 nm and 600 nm parthenolide).
We have begun to identify the molecular changes associated with cell survival after prolonged ICI 182,780 treatment in breast cancer cells. Whereas we have not attempted to confirm the altered expression of all implicated genes, some expression patterns are consistent with the activities we have confirmed. Here we discuss only those genes for which altered mRNA, protein, and/or transcriptional activation have been confirmed, and that are known to interact with each other in various cellular models, i.e., IRF-1, NPM, NFκB, and CRE.
IRF-1 can function as a tumor suppressor and can signal to apoptosis through both p53-dependent and p53-independent pathways (71) . These observations may partly reflect the ability of IRF-1 to induce a caspase cascade through activation of either caspase 1 (ICE; Ref. 72 ) and/or caspase 7 (73) . Caspase 1 is involved in regulating apoptosis in normal mammary epithelial cells (74) , and overexpression of caspase 1 is lethal in MCF-7 human breast cancer cells (75) . Preliminary data from our laboratory demonstrate that overexpression of IRF-1 inhibits anchorage-dependent colony formation and that the rate of cell proliferation in MCF-7 cells is inversely related to the level of IRF-1 expression (76) . These data suggest that the down-regulation of IRF-1 in MCF7/LCC9 cells may protect these cells from IRF-1-induced inhibition of proliferation and/or induction of apoptosis.
NPM can function as an oncogene, its overexpression fully transforming NIH 3T3 cells in a standard assay for oncogenic potential (77) . We have shown that levels of autoantibodies to NPM increase in breast cancer patients 6 months before their recurrence. Consistent with an estrogenic/antiestrogenic regulation of NPM, the levels of these autoantibodies are lower in breast cancer patients that have received TAM (38) . The increased NPM expression in MCF7/LCC9 cells compared with MCF7/LCC1 cells may reflect oncogenic potential of NPM, an activity potentially related to its ability to inhibit IRF-1 function (see below).
NFκB has been implicated in resistance to cytotoxic drugs and can function as a survival factor in various cell types (78) . Several aspects of normal mammary gland development appear dependent on NFκB activity (66) , perhaps partly reflecting its estrogenic regulation (65) . Elevated NFκB activity arises early during neoplastic transformation in the rat mammary gland (79) . Widely expressed in breast cancer cells and tumors, elevated NFκB activity is associated with estrogen-independence (65 , 66) . Currently, NFκB is the only protein known to induce BRCA2 expression (80) . ICI 182,780 cannot suppress the increased NFκB activity in MCF7/LCC9 cells, despite inhibiting this function in ICI 182,780-responsive cells (MCF7/LCC1). The functional relevance of this observation was tested directly using parthenolide, which both specifically binds NFκB and blocks degradation of the endogenous NFκB inhibitor IκB, resulting in the inhibition of NFκB transcriptional regulatory activities (68 , 70) . This activity of parthenolide has been used to evaluate the functional role of NFκB in several recent studies (68 , 69 , 81 , 82) . MCF7/LCC9 cells are significantly more sensitive to growth inhibition by parthenolide than their MCF7/LCC1 parental cells. This observation is consistent with a greater functional reliance on NFκB activation for cell growth/survival, and implies that one option for surviving antiestrogen exposure is the up-regulation of an estrogen-regulated survival factor(s) concurrent with the loss of its ER-mediated regulation. Furthermore, parthenolide is now in clinical trials, and our data suggest that it may prove useful in combination with Faslodex or other antiestrogens to either increase responsiveness and/or delay the appearance of resistant disease.
XBP-1 has been identified recently in clusters of genes associated with ER-positive breast tumors in two independent studies (13 , 41) , and its expression is increased in MCF7/LCC9 cells. XBP-1 is a transcription factor that binds and activates CRE (39) . The importance of CRE-regulated events is widely reported in many cell types (83 , 84) . These events include a likely role in signal transduction either at or downstream of ER and PgR (85) . The relevance of increased CRE activity in MCF7/LCC9 cells is additionally supported by recent evidence that CRE-decoy oligonucleotides inhibit the growth of MCF-7 cells (86) . We detected a 4-fold increase in CRE transcriptional activation in MCF7/LCC9 cells. Importantly, ICI 182,780 cannot regulate CRE activity in either MCF7/LCC1 (ICI 182,780-responsive) or MCF7/LCC9 (resistant) cells. These data imply an additional option available to breast cancer cells, a switch to signaling pathways that are normally independent of ER-mediated signaling.
IRF-1, NPM, NFκB, and CRE are known to affect cell proliferation, apoptosis, and/or carcinogenesis. Two critical protein-protein interactions directly link the IRF-1, NFκB, and NPM proteins. Direct binding occurs between IRF-1 and NPM (77) , and between IRF-1 and NFκB (87 , 88) . In both cases, the interactions with IRF-1 have important effects on gene transcription and cell signaling. NPM binding inhibits the transcription regulatory activities of IRF-1 (77) . A coordinated perturbation in the regulation of these two genes has occurred in the MCF7/LCC9 cells; NPM is up-regulated and IRF-1 is down-regulated. Thus, overexpression of NPM could additionally reduce the remaining lower levels of IRF-1, potentially blocking/eliminating its ability to initiate an apoptotic caspase cascade through caspase 1 and/or caspase 7. Such an effect would likely also eliminate the ability of IRF-1 to induce p21cip1/waf1 (89) and cooperate with wild-type p53 in signaling to apoptosis (56 , 57) . Changes in the amount of available IRF-1 will directly affect the number of IRF-1:NFκB heterodimers available to regulate an additional series of genes. Whereas NFκB will compete with NPM for IRF-1 binding, their relative affinities for IRF-1 are unknown, and the preferred IRF-1 heterodimer remains to be established. IRF-1:NFκB protein-protein interactions or other cooperative interactions are implicated in the induction of ATF-2/jun (90) , RANTES (91) , VCAM-1 (88) , interleukin 6 (92) , and MHC class 1 genes (87) . A functional IFN-β enhanceosome has been described that includes IRF-1, NFκB, and ATF2/jun (93) . The importance of both IRF-1 and NFκB in IFN-induced signaling may contribute to the ability of IFNs to increase responses to antiestrogens (94, 95, 96) .
CRE activation also may interact with the pathways regulated by IRF-1, NFκB, and NPM interactions. Delgado et al. (97) described a cyclic AMP-dependent pathway that inhibits IRF-1 transactivation. Thus, the increased CRE activity in MCF7/LCC9 cells may explain, in part, the lower IRF-1 mRNA levels seen both in the gene expression arrays and in the IRF-1 RNase protection studies.
The concurrent changes in NPM, IRF-1, NFκB, and CRE suggest a novel integrated signaling pathway that may involve the ability of NPM and CRE to inhibit IRF-1 initiation of a caspase cascade to apoptosis, the altered ability of cells to induce genes dependent on IRF-1:NFκB, and an increased activation of survival pathways that involve both NFκB and CRE. Studies to additionally establish the nature, function, and regulation of this putative pathway are currently in progress, including an overexpression of NFκB in sensitive cells and a dominant-negative approach in resistance cells. Because we looked only at cells that survived long-term antiestrogen exposure, the ability of the changes implicated in the present study to protect from an initial or short term exposure have yet to be determined. For example, cells may or may not survive an initial antiestrogenic exposure by the same mechanisms that allow for long-term survival. Irrespective of whether these other genes are functionally involved, their patterns of expression may be important in better predicting the 25% of ER+/PgR+, 55% of ER-/PgR+, and 66% of ER+/PgR− breast tumors that do not respond to antiestrogens (2) .
It is not possible, in a single focused study, to define all of the potentially differentially expressed genes nor to establish their functional relevance firmly. Because the number of cellular models studied is small, additional functional studies where expression of the candidate genes is induced or repressed are in progress. Nonetheless, our data imply that breast cancer cells have highly plastic transcriptomes, with access to several signal transduction pathways for regulating the choice to differentiate, proliferate, or die. For example, MCF7/LCC9 cells have taken several possible interactive/interdependent approaches to circumvent the growth inhibitory effects of antiestrogens. This plasticity in gene expression patterns is consistent with the marked heterogeneity apparent in the clinical disease (2 , 98) .
In summary, our data suggest that one molecular profile associated with surviving prolonged antiestrogen exposure may include loss of ER-mediated signaling to apoptosis through IRF-1. This lost signaling is achieved both by down-regulation of IRF-1 and a coordinated up-regulation of its inhibitor NPM, and possibly another protein partner NFκB. Up-regulation of CRE activities also is implicated in this molecular profile. Other patterns of gene expression may provide alternative routes to the resistant phenotype or in cells that acquire a TAM-stimulated phenotype (2) . The identification of these molecular profiles and signaling pathways may ultimately allow us to understand ER-regulated signaling, facilitate the development of novel treatment strategies, and allow clinicians to better identify antiestrogen-responsive and -resistant breast tumors.
We thank Dr. K.W. Kinzler and his colleagues at Johns Hopkins University, Baltimore, MD, for their assistance in establishing the SAGE protocols and for providing their SAGE data analysis software.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
↵1 Supported in part by Public Health Service awards 5R01-CA/AG58022, 5P50-CA58185 (to R. C.), 5R33-CA83231 (to Y. W.) from the National Cancer Institute; Department of Defense Awards DAMD17-99-9189 (to K. B. B.), DAMD17-99-1-9191, BC010619, and BC990358 (to R. C.) from the United States Army Medical Research and Materiel Command; and the American Cancer Society IRG-97-1520-01 (to T. C. S.). Technical services also were provided by the Flow Cytometry and Cell Sorting, and Macromolecular Shared Resources funded through Public Health Service Award 2P30-CA51008 (Vincent T. Lombardi Cancer Center Support Grant).
↵2 Present address: Celera Genomics, 45 West Gude Drive, Rockville, MD 20850.
↵3 Present address: Indiana University, Department of Medicine, Indianapolis, IN 46202.
↵4 Present address: Laboratory of Clinical Investigation, National Institute on Aging, NIH, 5600 Nathan Shock Drive, Baltimore, MD 21224.
↵5 Present address: United States Patent and Trade Mark Office, Crystal Plaza 3, Washington, DC 20231.
↵6 To whom requests for reprints should be addressed, at Room W405A Research Building, Vincent T. Lombardi Cancer Center, Georgetown University School of Medicine, 3970 Reservoir Road, NW, Washington, D.C. 20007. Phone: (202) 687-3755; Fax: (202) 687-7505; E-mail:
↵7 The abbreviations used are: ER, estrogen receptor; CRE, cyclic AMP response element; CCS-IMEM, improved minimal essential medium supplemented with 5% charcoal calf stripped serum; EGF-R, epidermal growth factor receptor; IRF-1, interferon regulatory factor-1; NPM, nucleophosmin; PgR, progesterone receptor; SAGE, serial analysis of gene expression; TAM, Tamoxifen; XBP-1, X-box binding protein-1; FACS, fluorescence-activated cell sorting; NFκB, nuclear factor κB; EGR-1, early growth response factor-1; TNFα, tumor necrosis factor α.
↵8 R. Clarke, unpublished observations.
- Received August 31, 2001.
- Accepted May 2, 2002.
- ©2002 American Association for Cancer Research.