Gene expression signatures

We analyzed the mean expression of multiple previously published gene signatures [7,15-19]. Briefly, these signatures include leukocyte-related signatures from Palmer et al. [17]: CD8 (n = 10 genes), B_Cell (n = 286), T_Cell (n = 178), and GRANS (n = 353). Stromal-related signatures were obtained from West et al. [16] (n = 402; DTF and SFT signatures combined) and Beck et al. [19] (n = 174). Genes enriched more than twofold in mammosphere-derived cells compared with differentiated cells were obtained from Dontu et al. [15] (n = 58). From Shipitsin et al. [18], we calculated the mean expression of the upregulated (n = 357) and downregulated (n = 353) genes from CD44⁺ versus CD24⁺ breast cancer cells from metastatic pleural effusions. From Creighton et al. [7], we calculated the mean expression of the CD44⁺/CD24^-/low/mammosphere signature reported (n = 119 upregulated genes; n = 279 downregulated genes). Finally, the proliferation and luminal gene cluster signatures were hand-picked (node correlation > 0.75) from the unsupervised intrinsic hierarchical clustering of the UNC337 using the intrinsic list of Parker et al. [9] and average linkage clustering using Cluster v2.12 (M. Eisen) [20] as shown in Figure S1 in Additional file 1. Gene lists from all genomic signatures are displayed in Supplemental Data in Additional file 2.

Breast cancer cell line microarray data set

We analyzed a data set that included Affymetrix U133A gene expression microarrays of 52 breast cancer cell lines [21]. Raw data were normalized using RMA, and genes were median-centered before analyses. Among the 52 breast cancer cell lines analyzed, DU4475, HCC1008 and HCC1599 cell lines were not included in Neve et al. [21].

Mouse breast tumor microarray data set

All mouse samples from the UNC were collected from fresh frozen invasive breast carcinomas as described previously [5] using Agilent mouse oligo microarrays. Data normalization and preprocessing were identical to that described for the UNC337 data set. We used only samples (n = 104) from our previous publication that were included in 1 of the 10 mouse classes [5].

Claudin-low and normal breast Euclidian centroid-based predictors

To robustly identify claudin-low samples, we built two predictors on the basis of either our human tumor data or the cell line data of Neve et al. [21]. To build a predictor, we first selected those genes that were significantly differentially expressed between claudin-low tumors defined by SigClust [22] (or cell lines) and all other subtypes using a two-class, unpaired SAM, with < 5% false discovery rate (FDR). Then we used these gene lists and calculated a claudin-low centroid and an "others" centroid from the training data. For every sample, we calculated the Euclidean distances to the two centroids and assigned the class of the nearest centroid. Using the same methodology, we also built a normal breast predictor by selecting those genes that were significantly differentially expressed between normal breast tissues and breast tumors using a two-class, unpaired SAM, with 0% FDR. Note that these gene lists are also included in Supplemental Data in Additional file 2.

Intrinsic subtype classification

For all human breast tumor studies, intrinsic subtype classification was performed using the PAM50 predictor [9]. Human claudin-low tumor samples were identified using either SigClust [22] or the nine-cell line claudin-low predictor. Samples identified by SigClust [22] or the nine-cell line claudin-low predictor were called claudin-low, regardless of the PAM50 call. For breast cancer cell lines, the claudin-low subtype classification was based on unsupervised hierarchical clustering using the intrinsic list of Parker et al. [9] and the node identified in Figure S4 in Additional file 1. The complete gene list of the nine-cell line claudin-low predictor can be found in Supplemental Data in Additional file 2.

Mammary developmental analyses

Public data sets from Raouf et al. [23] and Lim et al. [24] were downloaded from GEO and assigned NCBI Entrez gene identifiers as available in GEO. Samples were scaled to mean zero and variance of 1. Features were then collapsed to the mean of each gene identifiers. In Lim et al. [24], three epithelial cell-enriched subpopulations were profiled on DNA microarrays: mammary stem cells (MaSC), luminal progenitors (pL) and mature luminal cells (mL). We created a differentiation predictor for each sample as a measure of any sample's position along a MaSC → pL → mL axis as defined by gene expression. Distance-weighted discrimination (DWD) was used to determine the direction of greatest variation from MaSC to pL and pL to mL. To map a sample onto this axis of differentiation, the pL centroid was set as the origin, and the MaSC and mL centroids were transformed to length 1 (sum of squares equals 1) to map a sample onto this axis of differentiation. Before mapping a sample onto this axis, it is assumed that the test data set covers the range of differentiation, which allows median centering of genes to correct for platform bias. Test samples are then adjusted using the parameters for placing pL at the origin and are transformed to length 1. Each sample is then projected onto the MaSC → pL axis and the pL → mL axis by calculating the inner product of the sample and the MaSC or mL vectors identified by DWD. The difference of the two projected positions of each sample along the MaSC → pL → mL axis is referred to as the differentiation score. In the UNC microarray database web site [10], we have provided the detailed R commands and files regarding the differentiation predictor.

Mammospheres from normal breast tissues

Fourteen normal breast tissues were dissociated mechanically and enzymatically as described in Stingl et al. [25]. The samples were procured and used according to approved IRB protocols for research in human subjects. Mammospheres were cultured according to Dontu et al. [15], and single cells were plated onto ultralow attachment plates (Corning) at a density of 20,000 viable cells/ml. RNA was purified using RNeasy Mini kit (Qiagen) after 14 to 20 days in primary culture (first passage), and microarrays were performed as described above.

Immunohistochemistry

Formalin-fixed, paraffin-embedded tissue sections (~5 μm thick) were processed using standard immunohistochemistry methods as previously described [26]. The sections were incubated for 60 min at room temperature with primary antibody to claudin 3 (dilution 1:100; Invitrogen, Catalog No. 18-7340) or E-cadherin (Clone No. ECH-6, pre-diluted; Cell Marque). The slides were incubated for 45 min with diluted biotinylated secondary antibody (1:250 dilution) for 30 min with Vectastain Elite ABC reagent (Vector Laboratories). Sections were incubated in peroxidase substrate solution for visualization. Slides were counterstained with hematoxylin and examined by light microscopy. Tumor immunoreactivity was scored in a blinded fashion by two investigators (JIH and XH) into two categories: negative/weak positive and moderate/strong positive.

Immunofluorescence

Formalin-fixed, paraffin-embedded sections (~5 μm thick) were processed using standard immunostaining methods as previously described [5]. The primary antibodies and their dilution were vimentin (mouse anti-vimentin IgG1-κ, dilution 1:100; Invitrogen, Catalog No. 18-0052), keratin 5 (rabbit anti-human, dilution 1:500; Abcam, Catalog No. ab24647), and keratin 19 (Abcam, Catalog No. ab7754, mouse anti-human IgG2a; dilution 1:200). Secondary antibodies for immunofluorescence were conjugated with Alexa Fluor-568 (Red, keratin 5 and 19) or Alexa Fluor-488 (Green, vimentin) fluorophores (1:200; Molecular Probes, Invitrogen). Dual positivity was scored in a blinded fashion by XH into two categories: negative meaning no dual positive cells and positive meaning the presence of dual positive cells.

Cell lines

SUM159PT cells (Asterand) were maintained in Ham's F-12 medium with 5% fetal bovine serum (FBS), insulin (5 μg/ml), and hydrocortisone (1 μg/ml). MCF-7 was cultured in RPMI with 10% FBS [27], and SUM149PT was maintained in HuMEC media with supplements (Gibco) with and without 5% FBS [28]. All cell lines were grown at 37°C and 5% carbon dioxide.

Fluorescence-activated cell sorting (FACS) and microarray analysis

Nonconfluent cell cultures were trypsinized and filtered to produce single cell suspension, counted, washed with Hanks' balanced salt solution (Stem Cell Technologies) containing 2% FBS and stained with antibodies specific for human cell surface markers: EPCAM-fluorescein isothiocyanate (Stem Cell Technologies) and CD49f-phycoerythrin (BD Pharmingen). A total of 500,000 cells were incubated with antibodies for 30 min at 4°C. Cells were washed from unbound antibodies and immediately analyzed using Beckman-Coulter (Dako) CyAn ADP or sorted using BD FACScan. RNA was purified from sorted cells using the RNeasy Mini kit, and microarrays were performed as described above.

Identification of the claudin-low profile in a panel of breast cancer cell lines

To investigate if potential in vitro counterparts for these tumors exist, we analyzed a data set of 52 breast cancer cell lines [21] by hierarchical cluster analysis using our most recent human breast tumor intrinsic classification list [9]. The three major subgroups (luminal, basal B and basal A) identified previously by Neve et al. [21] were evident, with nine basal B cell lines clustering together with a node correlation of 0.59 (Figure S4 in Additional file 1). These cell lines showed low expression of the ER, HER2 and claudin 3, claudin 4 and claudin 7 (Figure S4 in Additional file 1). We identified those genes whose expression distinguished each human tumor subtype using significance analysis of microarrays (SAM) in our UNC337 tumor database, including a list defining the normal breast-like group (Figure ​(Figure2a).2a). These nine cell lines (MDA-MB-435, MDA-MB-436, Hs578T, BT549, MDA-MB-231, MDA-MB-157, SUM1315MO2, SUM159PT and HBL100) showed a gene expression pattern similar to the claudin-low tumor subtype with the lowest expression of genes involved in epithelial cell-cell adhesion (i.e., E-cadherin and claudins 3, 4 and 7), luminal differentiation (i.e., CD24, EpCAM) and high values for the CD44/CD24 and CD49f/EpCAM mRNA ratios (Figure S4, Table S3 in Additional file 1). To complement these clustering analyses, we developed a claudin-low centroid-based predictor using the UNC337 tumor data set and the SigClust-defined claudin-low group versus all others, and objectively classified the 52 cell lines as claudin-low or not; as expected, the human tumor-based claudin-low predictor identified these nine cell lines as claudin-low (Figure S5 in Additional file 1).

Open in a separate window

Figure 2

Identification of the claudin-low subtype in a panel of breast cancer cell lines. Gene clusters that characterize each primary human tumor subtype are shown in the human and cell line gene expression data sets. In both data sets, array trees have been derived by unsupervised hierarchical clustering using the intrinsic list from Parker et al. [9] as shown in Figure S1A and S4A in Additional file 1. (a) The top 50 upregulated genes associated with each molecular subtype, including the top 50 downregulated genes in claudin-low tumors, are shown in the UNC337 database. Top genes were selected after performing a two-class SAM (FDR = 0%) between each molecular subtype versus others. Luminal A and B subtypes were combined into the luminal subtype. In the tree, the yellow node denotes the claudin-low tumors defined by SigClust [22]. (b) Gene clusters characteristic of each tumor molecular subtype are shown in 52 breast cancer cell lines from Neve et al. [21]. Missing genes have been omitted. In the tree, the yellow node denotes the most highly correlated cell lines that best resemble the claudin-low subtype. 1 (yellow), claudin-low gene cluster of upregulated and downregulated genes; 2 (red), basal-like gene cluster; 3 (pink), HER2-enriched gene cluster; 4 (green), normal breast-like gene cluster; 5 (blue), luminal gene cluster.

The gene cluster that identifies the in vivo defined normal breast-like group was not differentially expressed by any of the 52 cell lines, suggesting overall that none of these cell lines show a normal breast-like phenotype (Figure ​(Figure2b).2b). The normal breast-like group of primary specimens is usually composed of actual normal breast samples and a small number of primary tumors [9,38], the latter of which we believe show the high expression of true normal tissue genes due to significant normal breast contamination [9]. In cell culture, however, there is no such contamination by normal epithelial or nonepithelial cells, and thus this lack of contamination may explain why normal breast-like tumor cell lines are not identified as such by the clustering. To be more objective, we applied a normal breast centroid-based predictor (normal breast versus all tumors using the UNC337 database) to the cell line data, and no cell line was classified as normal breast (Figure S5 in Additional file 1). Contamination with other tissues or cells is not uncommon in other tumor subtypes [1]. For example, within the top highly expressed genes in claudin-low tumors versus all others, there are several wound response-related genes (i.e., CD28, CD52) that are not expressed by the breast cancer cell lines (Figure ​(Figure2b),2b), potentially due to significant immune or stromal cell content in the tumor. Overall, these data suggest that the nine previously described breast cancer cell lines are most similar to claudin-low tumors in in vivo specimens.

Building a Cell-line claudin-low centroid-based predictor

Because primary human claudin-low tumors are highly enriched for immune system genes (and lymphocytes) when supervised analyses are performed, the expression of nonepithelial cell genes would likely increase the false positivity of the human claudin-low predictor when applied to other tumor data sets, because, for instance, tumors would be called claudin-low or not based upon the high expression of immune cell genes. Therefore, we developed a claudin-low centroid-based predictor using the cell line database of Neve et al. [21]. First, we evaluated its accuracy by applying the predictor onto our human tumor database (UNC337). The nine-cell line claudin-low predictor identified 37 samples (~11%) as claudin-low, including the 28 claudin-low samples previously identified by hierarchical clustering. The remaining nine samples identified by the predictor were previously called basal-like (n = 7), normal breast-like (n = 1) and HER2-enriched (n = 1) by the PAM50 predictor [9]. Overall, the nine-cell line claudin-low predictor showed 87.5% sensitivity, 97.0% specificity, and 75.7% and 98.7% positive and negative predictive values, respectively, if the SigClust-defined claudin-low group is considered the gold standard.

We then used the nine-cell line claudin-low predictor to identify this subtype in a mouse tumor database [5], which represents 13 different mouse models grouped in 10 different classes on the basis of their gene expression profiles (Groups I to X). Interestingly, all claudin-low samples identified by the centroid predictor (n = 9) were included in Group II, which we previously highlighted for its mesenchymal features (i.e., spindle-shaped cells). These nine murine claudin-low tumor samples were derived from six different mouse models (Brca1Co/Co;TgMMTV-Cre;p53+/-, DMBA-induced, p53-/- transplant, p53+/- IR, TgMMTV-Neu and TgWAP-T121) and overall showed similar enrichment for EMT markers and human mesenchymal and stem cell-like signatures, including decreased expression of proliferative- and luminal-associated genes (Figure S6 in Additional file 1). No murine normal mammary tissue sample was classified as claudin-low. These analyses suggest that a cell line centroid-based approach, by focusing on genes expressed in epithelial cells only, might give more accurate classification of tumors, which could be evaluated in future studies focused on tumor subtyping.

Clinical-pathological characteristics of claudin-low breast tumors

To determine for the first time the clinical-pathological characteristics of human claudin-low breast tumors, we evaluated our breast cancer patient database (UNC337) and two independent gene expression data sets (NKI295 and MDACC133) [11-13] using the nine-cell line claudin-low predictor and the previously published PAM50 subtype predictor (Figure ​(Figure3a)3a) because these two objective centroid predictors have been demonstrated to be the most robust to classify breast tumors into discrete subtypes. Across all three databases, claudin-low tumors showed a prevalence of 7 to 14%, and were mostly ER-/PR-/HER2- (also known as triple-negative tumors, 61 to 71%). Conversely, the majority of triple-negative tumors were either basal-like (39 to 54%) or claudin-low (25 to 39%), followed by HER2-enriched (7 to 14%), luminal B (4 to 7%), luminal A (4 to 5%) and normal breast-like (1%). In terms of prognosis, Kaplan-Meier survival analysis revealed that claudin-low tumors have a worse prognosis compared to luminal A tumors in both the UNC337 (hazard ratio [HR] of 2.83 and 5.66 for relapse-free survival [RFS] and overall survival [OS], respectively; P < 0.05) and NKI295 data sets (HR of 4.71 and 17.98 for RFS and OS, respectively; P < 0.001). Conversely, similar survival curves were observed between claudin-low tumors and the other poor prognosis subtypes such as luminal B, HER2-enriched and basal-like tumors (Figure ​(Figure3b).3b). Normal breast-like samples were omitted from survival analyses because they are likely significantly contaminated by normal breast tissue, and thus their true tumor biology is masked.

Open in a separate window

Figure 3

Clinical and pathological characteristics and prognosis of all intrinsic subtypes, including claudin-low tumors, across three independent breast cancer data sets. (a) Percentages of the different clinical-pathological characteristics in the UNC337 data set and two publicly available data sets (NKI295 and MDACC133). ER/PR/HER2 scores of the UNC337 database were based on clinically validated methods. (b) Survival data of the different molecular subtypes are shown for the UNC337 database and NKI295. Normal breast-like samples have been removed from this analysis. The UNC337 set represents a heterogeneously treated group of patients treated in accord with the biomarker status, whereas NKI295 is predominantly a local therapy only cohort.

We also evaluated the potential association between claudin-low tumors and treatment response by using the MDACC breast cancer patient data set (133 pretreated samples) of tumors treated with neoadjuvant anthracycline/taxane-based chemotherapy [13]. Notably, claudin-low tumors showed a lower pathological complete response (pCR) rate after anthracycline/taxane-based chemotherapy compared to basal-like tumors (38.9% vs. 73.3% pCR rates; P = 0.08, Fisher's exact test), but their pCR rate was higher than luminal A or B tumors; interestingly, the apparent pCR rate of the basal-like tumors increased from 59% (reported in Parker et al. [9]) to ~73% when the claudin-low subtype was included, because among 18 claudin-low tumors identified in this set, 12 (67%) of 18 of them were previously identified as basal-like and the response rate of this subgroup of patients was 41.7%. These findings suggest that claudin-low tumors show some chemotherapy sensitivity but overall have a poor prognosis and may not be managed effectively with existing chemotherapy regimens.

Twenty-one claudin-low samples were examined histologically by a pathologist (CL; Table S4 in Additional file 1) to further clinically characterize claudin-low tumors. Among the samples evaluated, 9 (43%) of 21 had noteworthy histological differentiation patterns, including medullary-like features (5/21, 24%) such as pushing margins and brisk tumor lymphocytic infiltration, two samples (2 of 21, 10%) showed metaplastic differentiation, one sample showed mixed ductal/lobular features, and one sample was a pure micropapillary carcinoma. The remaining 12 samples (12 of 21, 57%) were invasive ductal carcinomas not otherwise specified. Overall, lymphoid infiltration was evident in a total of seven samples (total of 7 of 19, 37%), which might explain the high mRNA expression of immune response genes present in these tumors. Among the other claudin-low samples that could not be examined histologically (n = 16), 50% (8/16) had a previous diagnosis of metaplastic tumors. It is interesting to note that two claudin-low cell lines, Hs578T and MDA-MB-157, were derived from metaplastic [39] and medullary [40] carcinomas.

Since the pathological examination of claudin-low tumors identified special histological features, we applied the nine-cell line claudin-low predictor to a publicly available database of histologically diverse subtypes of breast cancer (n = 113, NKI113) [14], which includes 10 medullary carcinomas, 20 metaplastic carcinomas and 22 invasive lobular carcinomas (ILC). Indeed, 8 (57%) of 14 and 2 (14%) of 14 claudin-low samples were identified as metaplastic and medullary carcinomas, respectively (Figure ​(Figure4f4f and Table S5 in Additional file 1). Conversely, only 2 of a total of 22 ILCs were identified as claudin-low despite the lack of E-cadherin expression in lobular tumors [41].

Open in a separate window

Figure 4

Epithelial differentiation score analysis of normal mammary tissue, human breast tumors, human cell lines and mouse mammary tumors. (a) Differentiation axis based on Lim et al. [24] data. (b) Mammospheres (MMS; n = 14) derived from normal breast tissue. Yellow crosses identify claudin-low MMS (n = 6, 43%) as defined by the nine-cell line claudin-low predictor. (c) Tumors and the normal breast-like group from the UNC337 database. (d) Breast cancer cell lines. Except for the nine claudin-low cell lines, we used the subtype calls (luminal [L] and basal [B]) as reported in Neve et al. [21]. (e) Mouse tumors from Herschkowitz et al. [5]. (f) Histological special types of breast cancer obtained from the NKI113 database [14]. Colored dots or boxes denote the subtype cells. IDC with OGC, invasive ductal carcinoma with osteoclastic giant cells; ILC, invasive lobular carcinoma; BL (red), basal-like; CL (yellow), claudin-low defined by the nine-cell line claudin-low predictor; H2 (pink), HER2-enriched; LA (dark blue), luminal A; LB (light blue), luminal B; NBL (green), normal breast-like. *P < 0.0001.

Claudin-low subtype resembles the MaSC profile

We hypothesized, as did Lim et al. [24], that a mammary differentiation program starting from a MaSC → luminal progenitor (pL) → mature luminal cells (mLs) exists, and therefore we built a differentiation model using data from Lim et al. [24]. For this predictor, higher scores represent greater differentiation status along this axis that culminates in ER+ mLs (Figure ​(Figure4a).4a). First, we applied this predictor onto similar subpopulations of purified human mammary epithelial cells that were separated by fluorescence-activated cell sorting (FACS) and profiled as part of the independent study of Raouf et al. [23] using different surface makers (Figure S7 in Additional file 1). The differentiation predictor showed 100% (8 of 8) accuracy with the bipotent progenitor subpopulation from Raouf et al. [23] [CD49f⁺(MUC1/CD133)^-(CD10/THY1)⁺] showing the lowest differentiation score (and thus most similar to the MaSC fraction of Lim et al. [24]), followed by the luminal-restricted progenitor [CD49f⁺(MUC1/CD133)⁺(CD10/THY1)^-] and then the mLs [CD49f^-(MUC1/CD133)⁺(CD10/THY1)^-]. In addition, we established and expression-profiled mammosphere cultures obtained from 14 different normal breast tissues because these cultures enrich for cells with stem or self-renewal capacity [15]. As expected, mammospheres showed a low differentiation score close to the MaSC profile (Figure ​(Figure4b4b).

Notably, and as shown by Lim et al. [24], the breast cancer subtypes segregate along the normal mammary epithelial differentiation hierarchy starting with undifferentiated claudin-low tumors, followed by basal-like, then HER2-enriched tumors, and finally both luminal tumor subtypes (Figure ​(Figure4c).4c). As expected, we observed the same pattern using the breast cancer cell line data [21] (Figure ​(Figure4d).4d). Conversely, the nine-cell line claudin-low predictor identified the MaSC and stromal (CD49f^low/EpCAM^-) subpopulations of Lim et al. [24] as claudin-low, and as expected, both subpopulations showed the highest and lowest expression of the up- and downregulated genes that define the nine-cell line claudin-low predictor (Figure S8 in Additional file 1). Moreover, we applied the differentiation predictor to the mouse data set of Herschkowitz et al. [5] (Figure ​(Figure4e)4e) and observed that the previously identified claudin-low murine samples scored the lowest, while the MMTV-Neu and MMTV-PyMT models, which are known luminal mammary adenocarcinoma models, scored the highest. In addition, among 113 histological special types of breast cancer [14], medullary, adenoid cystic and metaplastic tumors showed the lowest score in the differentiation axis (Figure ​(Figure4f),4f), which is consistent with our previous reports of commonalities between metaplastic carcinomas, claudin-low tumors and breast cancer TICs [6].

Finally, we evaluated the prognostic ability of the differentiation predictor in the UNC337 and NKI295 breast cancer patient data sets. Low differentiation scores were statistically significantly associated with poorer RFS and OS in univariate (Figure ​(Figure5a)5a) and multivariate analyses after adjusting for the main clinical-pathological parameters (i.e., size, grade, node and ER status), including tumor subtype (Figure ​(Figure5b).5b). These data suggest that tumors with an undifferentiated phenotype similar to the normal MaSC and/or early progenitors have a poorer prognosis compared to tumors with a more mature luminal phenotype, and this association is independent of the luminal B and HER2-enriched subtypes and the common clinical variables.

Open in a separate window

Figure 5

RFS and OS of breast cancer patients based on the differentiation tumor status. (a) Kaplan-Meier RFS and OS curves for UNC337 and NKI295 cohorts. Patients were rank-ordered and divided into two equal groups (low scores/differentiation in red and high scores/differentiation in black). (b) A combined multivariate analysis stratified by cohort was performed to test for significance of the differentiation status (as a continuous variable) conditioned on tumor intrinsic subtype, tumor size, histological grade, node status and ER. HR, hazard ratio; CI, confidence interval.