Relapse-Free Survival in Breast Cancer Patients Is Associated with a Gene Expression Signature Characteristic for Inflammatory Breast Cancer

Breast cancer is a complex genetic disease characterized by the accumulation of multiple molecular alterations. Recent advances in human genome research and high-throughput molecular technologies have made it possible to tackle the molecular complexity of breast cancer. This resulted in reclassifying breast tumors (1–4) and refining the prognostic and predictive capacity (5, 6). Some concepts of cancer biology, like the influence of tumor-associated stroma (7–9) and the presence of cancer stem/progenitor cells (10), are associated with relapse-free survival (RFS) or distant metastasis-free survival, thereby linking them to aggressive breast tumor behavior. Altogether, transcriptome analysis has provided leads on genes and biological processes that may drive breast tumor aggressiveness and could potentially lead to the development of new therapeutic targets.

Inflammatory breast carcinoma (IBC) is a clinical diagnosis designated as the T4d category in the tumor-node metastasis classification of the American Joint Committee on Cancer (11). It is a distinct clinical subtype of locally advanced breast cancer with a particularly aggressive behavior, characterized by explosive local growth and early lymphatic and hematogenous dissemination. Accordingly, the prognosis for patients with IBC is poor. When comparing patients with IBC to patients with stage IIIA breast cancer, the overall survival interval for patients with IBC is significantly shorter (10-year overall survival rates of 20% in IBC and 50% in stage IIIA breast cancer; refs. 12–14). Previously, we showed that IBC is characterized by a distinct gene expression signature compared with non-IBC (nIBC; refs. 15, 16). This IBC-specific gene expression pattern may hold the fingerprints of the mechanisms responsible for the aggressive behavior and elevated metastatic potential of IBC tumor cells. Evidence in that direction is provided by the identification of nuclear factor-κB hyperactivation (15, 16) and augmented insulin-like growth factor signaling (17) in IBC through genome-wide gene expression profiling. Both concepts have been linked to (breast) cancer metastasis. It has been shown that nuclear factor-κB activation mediates cellular transformation, proliferation, invasion, angiogenesis, and metastasis of cancer (18). The insulin-like growth factor pathway mediates cell motility through Rho GTPases via the phosphatidylinositol 3-kinase system, leading to an increased invasive and metastatic potential in breast cancer (19).

In the present study, we hypothesize that a small set of genes, predictive for the IBC phenotype, reflects a set of phenotypic traits associated with aggressive tumor cell behavior. We reason that this set of phenotypic traits not only determines aggressive tumor cell behavior in IBC but also in nIBC, thereby lending credit to the study of IBC as a model for aggressive tumor cell behavior in breast cancer in general. To test our hypothesis, we applied this set of genes onto several nIBC breast cancer data sets and compared the classification with RFS data, as well as with the classification of these breast tumors according to several known molecular classifiers and clinicopathologic variables. The aspects of aggressive tumor behavior, represented by the IBC signature, were identified by gene set enrichment for tumor-related concepts, as defined in the Oncomine database.

Patients and samples. Tumor samples were obtained from patients with breast adenocarcinoma treated in the Breast Clinic of General Hospital Sint-Augustinus. Each patient gave written informed consent. This study was approved by the local institutional review board. All samples were stored in liquid nitrogen within 15 min after excision (median delay of 9 min). Breast tumor samples included 17 pretreatment samples of patients with IBC, diagnosed by strictly respecting the criteria of the American Joint Committee on Cancer (11). The presence of tumor emboli was, as an isolated pathologic finding, not sufficient for the diagnosis of IBC. Clinicopathologic characteristics for IBC and nIBC are provided in Table 1. The study population used in the present study is a subpopulation of a former study by Van Laere et al., analyzing 19 IBC samples and 40 nIBC samples (17). The replicate samples used for analyzing reproducibility were left out. In addition, two IBC samples were excluded due to unreliable expression profiles. RNA from these samples was hybridized onto Affymetrix HGU133 plus 2.0 chips, and normalization, summarization, and gene filtering were done as described before (17).

Relationship between IBC phenotype and poor prognosis signatures. Each of the 57 IBC/nIBC breast tumor samples was classified according to the wound healing response (WHR) signature (7, 8), the fibroblastic neoplasm signatures (FNS; ref. 9), the invasiveness gene signature (IGS; ref. 10), and the 70-gene prognostic signature (5) using correlation-based classifiers. Briefly, a signature-specific centroid for breast tumors with an activated WHR, for IGS-positive breast tumors, and for breast tumors with favorable prognosis was calculated on the original data sets. For the FNS subtypes, two centroids were calculated for each subgroup. Samples were classified according to each signature by comparing the signature-specific gene expression profiles of each sample to the signature-specific centroids using Pearson correlation coefficients. Samples were classified according to the correlation values, with the mean correlation coefficient as threshold. For the FNS-subtype classification, the strongest correlation coefficient between a sample and the signature-specific centroids was used. Resulting correlation values were statistically compared between IBC and nIBC using the Mann-Whitney U-test. In addition, the genomic grade index (4) was calculated when possible and statistically compared with IBC and nIBC.

Identification of an IBC-specific gene expression pattern. To compose a gene expression signature predictive of the IBC phenotype, we used the nearest shrunken centroid algorithm implemented in the prediction analysis of microarray (PAM) package (20) on a training set composed of 40 samples (12 IBC and 28 nIBC specimens). The remaining samples (5 IBC and 12 nIBC samples) were left out to serve as an independent validation set. Using a 10-fold cross-validation on the training set, a δ value was selected so that the misclassification error rate for IBC and nIBC samples was equal. Hence, an IBC signature of 263 probe set IDs, corresponding to 205 unique genes, was identified. This IBC signature was applied to the independent validation set, and sensivity, specificity, and overall test error rate were assessed.

Covariate analysis. Because IBC is generally estrogen receptor (ER)–negative, ErbB2-positive, and of high histologic grade, we did a covariate analysis to ensure that the IBC signature was primarily associated with the IBC/nIBC distinction. Therefore, we did an unsupervised hierarchical complete linkage clustering analysis using the IBC signature genes. Using the silhouette algorithm, we assessed the significance of clustering by randomly permuting class labels for tumor type (IBC/nIBC), ER status (ER+/ER−), ErbB2 status (ErbB2+/ErbB2−), and genomic grade (grade 1/grade 3) 100 times. The difference in average silhouette width for true and random labels was assessed for significance using a one-sample t test. In addition, a conditional backward linear regression analysis was done with the clustering output as dependent variable and tumor type, ER status, ErbB2 status, and genomic grade as independent variables.

Application of IBC signature to nIBC breast cancer data sets. From the National Center for Biotechnology Information website, seven publicly gene expression data sets (GSE2990, ref. 4; GSE2034, ref. 6; GSE1456, ref. 21; GSE4922, ref. 22; GSE3744, ref. 23; GSE5327, ref. 24; and GSE5460) were retrieved comprising gene expression data for a total of 1,157 nIBC breast tumor samples. Data regarding tumor grade, ER status, ErbB2 status, lymph node status, tumor size, presence of lymphovascular invasion, and patient age at diagnosis were present for, respectively, 683, 901, 129, 838, 607, 129, and 525 nIBC breast tumor samples.

Each of the 1,157 nIBC breast tumor samples were classified according to the IBC signature using a centroid-based classifier. Therefore, the IBC centroid resulting from the prediction analysis of microarray analysis was used. Genes represented multiple times in the IBC signature were averaged. In addition, each of the samples was classified according to the WHR signature, the FNS, the IGS, and the 70-gene prognostic signature, as described before. Pearson χ² statistics were used to compare classification results of the IBC signature with classification results of the remaining signatures.

When possible, the association between the IBC signature classification and several known clinicopathologic variables was evaluated. Therefore, Pearson correlation values, denoting the strength of the IBC signature expression in each individual nIBC breast tumor, were statistically compared between different clinicopathologic subgroups using the Mann-Whitney U-test. Using Pearson correlation coefficients, the tumor size (mm) and the age (y) of a patient at diagnosis were correlated with the IBC signature correlation values. Four of the breast cancer data sets contained information on disease-free survival, comprising a total of 881 samples. For each individual data set, the association between the classification according to the IBC signature and the disease-free survival interval was investigated. In addition, univariate and multivariate Cox regression analyses were carried out for all 881 samples grouped together.

Oncomine analysis. The set of 206 unique genes from the IBC signature was split in a set of 57 genes predictive for IBC and a set of 149 genes predictive for nIBC. Each set was uploaded to the oncomine database and was separately analyzed for overrepresented molecular concepts, as defined by the Oncomine Concepts Map tool. Briefly, a molecular concept is defined as a group of genes with common traits, for example, genes commonly involved in one signaling pathway, genes with a common promoter sequence, or genes with a common microRNA (miRNA) target site. These groups of genes are analyzed for overrepresentation in the IBC signature using gene set enrichment analysis (25). Using gene set enrichment analysis, two gene lists are compared by analyzing which genes are commonly present/absent in both gene lists (concordant results) and which genes are present in one gene list and absent in the second list (discordant results). Based on these data, a Fisher exact test is done and a P value is calculated. The odds ratio presents the product of the concordant results divided by the product of the discordant results (25).

IBC phenotype is associated with poor prognosis signatures. Each sample in the IBC/nIBC data set was correlated with each of the signature-specific centroids, and for each sample, genomic grade index was calculated. These correlation values were compared between IBC and nIBC breast tumor samples using Mann-Whitney U tests. The corresponding box plots are displayed in Fig. 1. These data significantly associate IBC with an activated WHR (P = 0.047), with high genomic grades (P = 0.002), with the presence of the IGS profile (P = 0.0004) and with poor prognosis (P < 0.0001). For the fibrotic neoplasm signatures, no significant associations were found (data not shown).

Identification and validation of an IBC gene expression signature. An IBC signature was identified using the nearest shrunken centroid classification algorithm implemented in the PAMR package in Bioconductor. The original data set of 57 samples was randomly divided into a training set of 40 samples (12 IBC and 28 nIBC) and a test set of 17 samples (5 IBC and 12 nIBC). Using a 10-fold cross-validation, a δ value was chosen, resulting in an equally small classification error rate for both the IBC and nIBC classes (Supplementary Fig. S1). In this way, a set of 263 probes, corresponding to 205 unique genes, was identified (Supplementary Table S1). The cross-validated probabilities for each sample in the training set are plotted in Supplementary Fig. S2. Most samples, except one IBC sample and two nIBC samples, have very high probabilities of being classified into the correct class.

To estimate specificity and sensitivity and test error rate of the IBC signature, the IBC signature was applied onto the test set composed of 5 IBC and 12 nIBC samples. All IBC samples were correctly classified, resulting in a sensitivity of 100%. Out of 12 nIBC samples, one sample was misclassified, resulting in a specificity of 91%. The overall test error rate was 5.8%.

Covariate analysis. Using unsupervised hierarchical complete linkage clustering, no sample cluster entirely composed of IBC samples was identified, indicating that some nIBC samples showed an IBC-like gene expression profile (Supplementary Fig. S3A). The significance of this clustering was calculated by randomly permuting the tumor class label 100 times and comparing the mean average silhouette width for true and random class labels (P < 0.0001). We did the same analysis for ER status, ErbB2 status, and tumor grade because IBC is generally ER−, ErbB2+, and of high tumor grade. Although the clustering was associated with ER status (P = 0.005), for ErbB2 status and genomic grade, no association was found, suggesting that the IBC signature is primarily associated with the distinction between IBC and nIBC (Supplementary Fig. S3B-E). A logistic regression model with the clustering output as dependent variable and tumor type, ER status, ErbB2 status, and genomic grade as independent variables supported this interpretation. Tumor type (IBC/nIBC) was identified as the sole predictive variable for the clustering [Exp(B) = 44.63, confidence interval 6.52-305.75, P < 0.0001].

To apply the IBC centroid onto publicly available gene expression data sets, we used the IBC centroid identified using the prediction analysis of microarray analysis. The centroid expression is reported in Supplementary Table S2. Application of the IBC centroid classification algorithm onto the IBC/nIBC data set resulted in a correct classification of 77% of the samples.

Association between IBC signature and other known breast cancer signatures in nIBC. To investigate the potential prognostic relevance of the IBC signature in nIBC, we first compared the IBC signature classification of 1,157 nIBC samples with the classification of the same samples according to other signatures with proved prognostic relevance, including the cell-of-origin subtypes, WHR, IGS, FNS, 70-gene prognostic signature, and genomic grade index. Results are shown in Table 2.

Association between IBC signature and clinicopathologic variables. Within the group of 1,157 nIBC breast tumors, the correlation values with the IBC centroid for each individual breast tumor were compared between subgroups stratified by ER status, HER2 status, grade, lymphovascular invasion, and the presence of lymph node metastasis using the Mann-Whitney U test. In addition, a Pearson correlation was calculated between the correlation coefficients and the age of a patient at diagnosis (y) and tumor size (mm). Boxplots and scatterplots are displayed in Fig. 2.

As younger patients tend to be more likely ER-like or of the basal-like phenotype, we checked the cell-of-origin subtype classification as confounding variable for the association between the IBC signature classification and patient age at diagnosis. Within the group of the basal-like (n = 89), luminal A (n = 228), and luminal B (n = 135) tumors, inverse correlations were found between the correlation values with the IBC centroid and the age of a patient at diagnosis (respectively, R_s = −0.189, P = 0.076; R_s = −0.177, P = 0.077; and R_s = −0.122, P = 0.128). No association was found between the IBC signature and age at diagnosis within the normal-like group (n = 153, R_s = −0.004, P = 0.964), and a positive correlation was found within the ErbB2+ subtype (n = 107, R_s = 0.174, P = 0.073). No associations were found with the presence of lymph node metastasis (n = 838), lymphovascular invasion (n = 129), or tumor size (n = 607).

Prognostic relevance of the IBC signature. To assess the clinical significance of IBC signature, we did a Kaplan-Meyer analysis for four data sets containing RFS data. The resulting Kaplan-Meyer plots are shown in Fig. 3. For three data sets, the difference in RFS between nIBC tumors with IBC-like characteristics and nIBC tumors with nIBC-like characteristics reached significance (P < 0.05). In all data sets, breast tumors with IBC-like characteristics showed a shorter RFS interval. We also did analyses for patients only belonging to the luminal A (n = 324) and the ER+ (n = 567) subgroups. Resulting Kaplan-Meyer plots are also shown in Supplementary Fig. S4. Our IBC signature retains its prognostic capacity within these subgroups (luminal A, P = 0.025; ER+ breast tumors, P < 0.0001). In addition, Cox regression analyses were done for each of the clinicopathologic variables and the classifications according to the different gene expression signatures on all 881 samples together. RFS was used as outcome variable. Variables significantly associated with RFS in univariate analyses were subjected to stepwise backward multivariate analysis. Results are described in Table 3. The WHR and the 70-gene prognostic signature were independently associated with RFS.

Oncomine analysis of IBC signature. The results from the oncomine analysis confirmed our observation that the IBC signature is associated with poor prognosis in breast cancer, as well as in other types of cancer. The set of 57 and 149 genes predictive for, respectively, IBC and nIBC was significantly enriched with genes denoting poor (CTSC, CXCL9, CXCL10, S100A8, THOC4) and good (ALDH6A1, VAV3, ERBB4, LZTFL1, MYB) prognosis in multiple types of cancer. In addition, a set of genes expressed downstream of cRel, a nuclear factor-κB transcription factor family member, was also overrepresented in the list of 57 IBC genes, thereby corroborating our previous work (odds ratio 3.07, P = 0.004).

Interestingly, evidence was found for a differential regulation of gene expression between IBC and nIBC by miRNAs. Using gene set enrichment analysis, statistically significant associations were found between the 149 nIBC-predictive genes on the one hand and an Oncomine-defined gene list, consisting of 743 genes with target sites for MIR-20, MIR-106B, and MIR-18 in the 3′ untranslated region of their mRNA molecules (odds ratio 1.59, P = 0.0007). Thirteen genes were in common between both groups (TBX3, TBC1D9, MYB, FOXA1, ITPR1, MAP7, APBB2, CCND1, CELSR2, CPEB2, ENPP5, ESR1, and TRPS1). In addition, significant associations were also found between a second Oncomine-defined gene list of 52 genes with target sites for MIR-221 and MIR-222 in the 3′ untranslated region of their mRNA molecules and the set of 149 nIBC-predictive genes (odds ratio 3.33, P = 0.002). Five genes were in common between both groups (MYLIP, GARNL1, TRPS1, IRX5, and ESR1).

In the past, genome-wide gene expression profiling has been successfully applied to gain novel insights into breast carcinoma biology. Gene expression signatures specific for activated stroma (7–9) and cancer stem/progenitor cells (10) all showed their prognostic capacity, thereby linking these concepts of cancer biology to aggressive tumor cell behavior. One particular form of aggressive breast carcinoma is IBC, characterized by an explosive local growth and elevated invasive and metastatic potential. Accordingly, patient survival for patients with IBC is poor (12–14). Several hypotheses can be raised explaining the dramatic features associated with IBC, such as augmented IBC tumor cell motility and the contribution of breast cancer stem/progenitor cells. The IBC-specific gene expression profile should bare the fingerprints of the processes involved in this pernicious breast cancer phenotype. In the present study, we use IBC as a model of aggressive breast cancer behavior in general. We hypothesize that biological concepts underlying aggressive tumor cell behavior in IBC may also be relevant in nIBC.

Our data show that IBC is indeed a suitable model to study aggressive tumor cell behavior. First, classification of the nIBC breast tumors according to prognostically relevant signatures showed a remarkable overlap with the classification according to the IBC signature. The fact that these signatures show extensive overlap in predicting the clinical outcome for individual patients suggests that these signatures reflect a common set of phenotypic traits of aggressive tumor cells, with each trait defined by one gene expression signature (26). Second, nIBC breast tumors with IBC-like properties have a reduced RFS compared with their nIBC-like counterparts. Third, patients with IBC-like characteristics more frequently show clinicopathologic features associated with aggressive tumor cell behavior (12–14). IBC-like breast tumors are frequently of higher grade, ER-negative, and ErbB2-positive. In addition, nIBC patients with IBC-like breast tumors are younger at time of diagnosis. However, despite the fact that virtually all IBC patients have lymph node metastasis at diagnosis and frequently display lymphovascular invasion, no association between the IBC signature and these variables was found in the nIBC population. In conclusion, our data suggest that IBC and nIBC tumors show overlapping phenotypic traits with respect to aggressive tumor cell behavior, which lends credit to the study of IBC as a model for aggressive tumor cell behavior.

The fact that IBC is more frequently ER-negative, ErbB2-positive, and of high tumor grade compared with nIBC is a possible confounding factor in this study. However, covariate analysis and regression analysis for these variables have shown that the IBC signature is primarily associated with the distinction between IBC and nIBC. Although ER status was identified as a significant covariate, the IBC signature was able to stratify patients with ER+ breast tumors in two groups with significant differences in RFS. This might be explained by the high propensity of luminal B tumors to be IBC-like, whereas luminal A tumors are more often nIBC-like. However, subgroup analyses on the luminal A breast tumors showed similar differences between IBC-like and nIBC-like breast tumors with respect to RFS. Also, in the current data set of 881 patients, ER status was not predictive for RFS in univariate analysis, whereas our IBC signature did show prognostic value in both univariate and multivariate analysis. Hence, the link between the IBC-specific gene expression profile and aggressive tumor cell behavior is not only attributable to the simple distinction between breast tumors regarding these clinicopathologic variables. Nevertheless, due to our study design (e.g., imbalance of ER status between IBC and nIBC and unknown survival status for both IBC and nIBC), at present, we cannot exclude confounding effects by ER status and survival. In addition, the possibility remains that the differences in gene expression between IBC and nIBC are sample set–specific and, therefore, cannot be generalized.

An important question is which aspects of aggressive tumor cell behavior are represented by the IBC signature. A hint is provided by the oncomine analysis, which showed that the set of genes predictive for IBC is enriched in mediators of cell motility. A clear picture arose when performing an exhaustive analysis of 455 significantly overexpressed genes in IBC (data not shown). Several molecular concepts linked to Arp2/3-driven Y-branching of actin filaments, actin polymerization, and lamellipodia extension were identified. Also, the phosphatidylinositol 3-kinase pathway is suggested as an upstream converter of extracellular signals to the Arp2/3 complex. The conversion of extracellular signals to cell migration via phosphatidylinositol 3-kinase–Arp2/3 route has been described previously (27). Our data suggest that the IBC signature is a signature also representing aspects of tumor cell motility, a feature generally linked to aggressive tumor cell behavior.

This study also confirms some of our previous data on IBC (28, 29). The Oncomine analysis identified a member of the nuclear factor-κB transcription factor family as an important transcription factor for the IBC gene expression profile. Also, genes up-regulated by nuclear factor-κB, a literature-defined concept, are significantly overrepresented in the list of IBC-predictive genes. Also, IBC samples more often have an activated WHR, indicating the importance of tumor stroma in this subtype of breast cancer. The strong association between the IBC phenotype and the IGS is indicative of the possible important contribution of breast cancer stem cells for the pathogenesis of IBC. In addition, the evidence suggests a differential regulation of gene expression profiles in IBC and nIBC by miRNAs. These small RNA molecules regulate gene expression by binding to a target sequence of the 3′ untranscribed region of mRNA molecules, thereby inducing degradation of the targeted mRNA molecule, which results in the down-regulation of gene expression of the gene corresponding with the targeted mRNA molecule. We observed a significant overlap between the list of nIBC-predictive genes only and two groups of genes with common miRNA target sequences in the 3′ untranslated region. This indicates that those genes with the specific miRNA target sequence in the 3′ untranslated region are overexpressed in nIBC, indicating a more elaborate presence of the specific miRNAs in IBC. One group of genes was regulated by MIR-221 and MIR-222. An interesting observation regarding this fact is that MIR-221 and MIR-222 both target the p27 Kip1 tumor suppressor and cell cycle inhibitor, thereby promoting cell proliferation (30, 31). Interestingly, Gonzalez-Angulo et al. observed p27 Kip1 down-regulation in 84% of tumor samples from patients with IBC (32). The second group of genes was regulated by miRNAs from the MIR-106B family. This family of miRNAs seems to target the cyclin-dependent kinase inhibitor p21/CDKN1A, thereby inducing progression through the cell cycle (33). In conclusion, our data show that the adverse phenotype characterizing IBC might be, in part, explained by the regulation of gene expression by specific miRNAs, thereby opening an entire new avenue for IBC research.

No potential conflicts of interest were disclosed.