Gene expression signatures of morphologically normal breast tissue identify basal-like tumors

Clinical data were collected for the samples from the Breast Cancer Functional Genomics Group clinical database. Cellular and fibrotic stroma were identified by visual inspection of hematoxylin and eosin stained tissue sections under a microscope. Cellular stroma was defined as tissue with more than 1,000 stroma cells uniformly distributed throughout the field of view (4× magnification), while fibrotic stroma was defined as tissue with less than 800 stroma cells in the field of view (4× magnification) and concentrated primarily around the ducts.

All tissue specimens and associated clinical data were collected at McGill University Health Center (Montreal, Canada) between 2000 and 2004 in accordance with the protocols approved by the research ethics committee. Patient consent was obtained on an individual basis for all patients participating in this study. Of 44 patients selected for the study, 34 patients had invasive ductal carcinoma and 10 were healthy donors undergoing reduction mammoplasty. Tissue samples were collected within 30 minutes after surgery, embedded in TissueTek OCT (Somagen, Edmonton, Alberta, Canada) and stored in liquid nitrogen until use. Frozen specimens were cryosectioned in 10-micron slices, stained using a hematoxylin and eosin staining protocol and dehydrated in ethanol and xylene as recommended by the LCM manufacturer (Arcturus, Mountain View, CA, USA). Following dehydration, the slides were air dried for 20 minutes and subjected to LCM. All normal tissues adjacent to tumor were microdissected from regions at least 2 mm away from tumor margins. Normal and adjacent stroma were sampled exclusively from the extralobular stromal compartment.

All tissues included in this study were re-examined by a clinical pathologist dedicated to the project. Tissue specimens were microdissected into epithelium and stroma using a PixCell IIe LCM system (Arcturus). All microdissections were performed within three hours following tissue staining. Total RNA was extracted from each population of microdissected cells using a GITC (guanidinium isothiocyanate) extraction protocol. Briefly, LCM caps were incubated for 5 minutes (room temperature) in 200 μl GITC extraction buffer (4 M GITC, 25 mM sodium citrate pH 7.0, 0.1 M β-mercaptoethanol, 0.5% N-lauroylsarcosine) supplemented with 1.6 μl β-mercaptoethanol. Subsequently, 20 μl of 2 M NaOAc, pH 4.0, 220 μl of water-saturated phenol and 60 μl of chloroform-isoamyl alcohol (23:1) were added to the extraction buffer. Following 15 minutes incubation on ice and centrifugation (12,000 rpm, 15 minutes) the aqueous phase was removed and RNA was precipitated with 2 μl glycogen (GenHunter, Nashville, Tennessee, USA) and 200 μl isopropanol. Samples were placed at -80°C for 30 minutes and centrifuged at 4°C (12,000 rpm) for 30 minutes to pellet RNA. Pellets were washed with 70% ethanol, air dried and subjected to DNAseI treatment (Roche, Basel, Switzerland). DNAseI treatment was performed in the presence of an RNase inhibitor (Invitrogen, Carlsbad, California, USA). Subsequently, samples were re-extracted as described above and re-suspended in 10 μl of diethylpyrocarbonate-treated water. RNA was quantified using a RiboGreen assay (Molecular Probes, Carlsbad, California, USA). Subsequently, 2 to 4 ng of total RNA was subjected to two rounds of T7 linear amplification using Ambion Amino Allyl MessageAmp kit (Ambion, Austin, Texas, USA) and labeled with Cy3 and Cy5 dyes according to the manufacturer's procedure. Prior to microarray hybridizations, amplified products were quantified using a spectrophotometer (Nanodrop, Wilmington, Delaware, USA) and subjected to BioAnalyzer to assay for quality (Agilent Technologies, Santa Clara, California, USA).

Whole Human Genome 44 K arrays (Agilent Technologies, product G4112A) were used for all experiments. RNA samples (500 ng) were subjected to fragmentation followed by 18 h hybridization, washing, and scanning (Agilent Technologies, model G2505B) according to the manufacturer's protocol (manual ID #G4140-90030). Samples were hybridized against Universal Human Reference RNA (Stratagene, ID #740000, La Jolla, California, USA). Duplicate hybridizations were performed for all samples using reverse-dye labeling.

Candidate tissue markers were validated by immunohistochemistry. Frozen tissue sections (10 μm thick) were defrosted at room temperature for 30 s, fixed in acetone (room temperature, 10 minutes) and air dried for 2 minutes. Subsequently, tissue sections were blocked with Peroxidase Blocking Reagent (DakoCytomation, Glostrup, Denmark). Primary antibodies were diluted at 1:50 and 1:15 for anti-c-kit (polyclonal rabbit anti-human CD117, DakoCytomation), and anti-CD31 (polyclonal mouse anti-human, DakoCytomation) and applied to the tissue sections for 45 and 15 minutes, respectively. Following a brief wash with TBS-T (tris-buffered saline tween-20), secondary antibodies were applied for 30 and 20 minutes, respectively. Labeled polymer-HRP anti-rabbit (EnVision+ System HRP(DAB), DakoCytomation) was used as a secondary antibody for c-kit staining and labeled polymer-HRP anti-mouse (EnVision+ System HRP(DAB), DakoCytomation) for CD31 staining. After a short wash with TBS-T, DAB Substrat-Chromogen Solution (EnVision+^® System HRP(DAB) DakoCytomation) was applied for up to 5 minutes for color development.

Microarray data were feature extracted using Feature Extraction Software (v. 7.11) from Agilent with the default parameters. Raw data were uploaded to the NCBI Gene Expression Omnibus database (GEO) and is accessible as data series GSE4823. Outlier features on arrays were flagged by the software. Arrays were required to have an average raw signal intensity of 1,000 in each channel, and a signal to noise ratio above 16 per channel. MvA plots were examined for signs of hybridization or labeling problems. Replicate arrays were required to have a concordance above 0.944. This level was established empirically using sets of known good replicate arrays in our database.

Data preprocessing and normalization were automated using the BIAS system [20]. Raw feature intensities were background corrected using the RMA background correction algorithm [21, 22]. Resulting expression estimates were converted to log2-ratios. Within array normalization was performed using spatial and intensity-dependent loess [23]. Median absolute deviation scale normalization was used to normalize between arrays [24].

Using class discovery under correlation distance and Euclidean distance metrics, 10,000 bootstrap iterations were performed to assess the significance of the observed clusters using the pvclust package for R[25]. Multidimensional scaling was applied to reduce the dimensionality of the data and permit visualization. Chi-square tests and logistic regression were applied to discrete and continuous variables, repsectively, to test for association with data partitions (clusters). The variables tested included estrogen receptor (ER) status, progesterone receptor (PR) status, lymph node (LN) status, HER2 receptor status, menopause status, age, grade, tumor size, and recurrence.

Both the linear models for microarray analysis (LIMMA) and significance analysis of microarrays (SAM) algorithms were used to identify differentially expressed gene sets from which to build class predictors [26‐29]. Genes from LIMMA were filtered for significance, (false discovery rate adjusted p value ≤ 0.01), fold change (≥2.0), intensity above background (A > 6.0), while genes identified by SAM were filtered by significance (q ≤ 0.3), fold change (≥2.0), and intensity (A > 6.0).

The prediction around medoids (PAM) algorithm was used to build predictors based on the filtered gene sets [30]. Cross validation was used to test the predictors. This procedure included independent selection of candidate gene sets for each cross validation step. Differentially expressed genes were mapped onto Gene Ontology (GO), and GO terms were tested for overrepresentation using the hypergeometric distribution [26].

We wanted to assess the relative contribution of different factors to the overall variability of gene expression observed in our data. Principal component analysis allows one to succinctly summarize data in a reduced number of dimensions (principal components) [31]. The principal components are ordered by the amount of variation (or signal) in the data that they explain. We performed principal component analysis on the patient matched adjacent stroma and epithelial data. Consecutive sequences of the first 10 principal components were tested for association with clinical characteristics using multivariate analysis of variance (MANOVA). Bonferroni multiple testing correction was applied to the resulting p values [31].

LIMMA was used to identify differentially expressed genes between tissues in individual patients and obtain expression estimates for the matched data ([28, 32]. Genes not exhibiting differential expression in at least 50% of samples were excluded from further analysis (B-statistic > 0). A paired t-test was used to identify genes whose patient-matched LIMMA expression estimates were significantly different from zero over the panel of patients (false discovery rate adjusted p value < 1e-5).

The expression of gene signatures from a number of publicly available datasets was examined in normal tissue.

The stroma-specific and epithelium-specific gene lists identified by Allinen and colleagues [33] contained 231 and 97 unique genes, respectively, of which 189 and 89 were located (mapped) successfully on the Agilent chip. The activated and inactivated core serum response (CSR) genes from Chang and colleagues [34] contained 228 and 233 genes, respectively, of which 209 and 211 were mapped to the Agilent array. The intrinsic breast cancer gene list of Sorlie and colleagues [35] contained 553 genes, of which 473 were mapped to the Agilent array. The desmoid type fibromatosis (DTF) and solitary fibrous tumor (SFT) specific gene lists from West and colleagues [36] contained 493 and 293 genes, respectively, of which 415 and 238 were mapped to the Agilent array. Genes that were likely to be expressed in normal breast tissue were selected from these gene sets by selecting genes with variance >1 in the normal tissue data; 7.3% of genes in the normal dataset have variance >1, and enrichment for high variance genes in the various gene sets was measured by a χ² goodness of fit test.

Genes from the Agilent whole genome arrays were mapped to the Agilent 24 K arrays used in the Netherlands cancer dataset [8]. The 24 K arrays used by Van de Vijver and colleagues [8] contained 24,498 features. Approximately 10,000 contigs on the 24 K array could not be mapped to GenBank identifiers. Of the remaining 14,339 identifiers, 12,112 were mapped to features on the 44 K Agilent array. Expression of the genes from the normal tissue signature was then examined in the 295 breast cancer samples from the Netherlands cancer dataset [8].

The GEO accession number of the array data series is GSE4823.

To determine the gene expression profiles of morphologically normal epithelium and stroma derived from reduction mammoplasties and breast cancer tissue, we integrated the use of LCM and T7-based RNA amplification with DNA microarrays. LCM provides an accurate means by which to isolate morphologically normal epithelium and stroma adjacent to breast cancer that is free from infiltrating tumor cells. This allows gene expression profiles to be generated from specific cell types rather than whole tissue [18]. LCM was used to isolate matched morphologically normal epithelial and stromal cells from 34 patients with invasive ductal carcinoma, and 10 patients who underwent reduction mammoplasty (Figure 1). Patient and tumor characteristics of the selected invasive ductal carcinoma patients are shown in Table 1 (and Additional file 4). In general, 2 to 5 ng of RNA were extracted from dissected normal epithelial ducts and stroma. We, as well as others, have established that T7 linear amplification preserves the ratios of mRNA abundance between mRNA species, provided all samples undergo the same number of amplification rounds [12, 37‐41].

Expression profiling was performed on cells isolated from morphologically normal epithelial and stromal tissue from 34 cases of invasive ductal carcinoma and 10 cases of reduction mammoplasty using Agilent whole genome arrays. A total of 66 samples were analyzed, of which 32 were isolated from stroma (26 from histologically normal ducts adjacent to tumor, and 6 from reduction mammoplasty), and 34 from epithelium (25 from histologically normal ducts adjacent to tumor, and 9 from reduction mammoplasty) (Table 1). Each of the LCM captured samples was interrogated in duplicate on a 44 K genomic feature microarray. Since several studies have suggested that normal stroma as well as morphologically normal terminal duct lobular units from cancer patients undergo loss of heterozygosity [42], we first performed a cluster analysis to determine whether the patient-matched stroma and morphologically normal epithelium were similar to those from reduction mammoplasty patients. After normalization, hierarchical clustering was applied to the 66 samples and the complete panel of genes (44 K genome features). Based on gene expression, the stroma and epithelium clustered according to tissue type (Figure 2a). Stroma surrounding histologically normal ducts from tumor specimens and stroma isolated from reduction mammoplasty clustered together. Similarly, morphologically normal epithelium from tumor specimens co-clustered with epithelium from reduction mammoplasties (Figure 2a). We observed similar tissue-specific clustering when using a multidimensional scaling class discovery approach (Figure 3; see Materials and methods). Only three adjacent stroma samples were found to behave as outliers, clustering with epithelial tissue at the whole genome level, an error rate comparable to other large scale microarray data sets (Figures 2a and 3).

To identify the genes responsible for the tissue-specific clustering observed in Figure 2a, class distinction was applied to identify all genes differentially expressed between tissues. Markers were defined based on patient matched stromal and epithelial samples (22 patients and 44 samples; see Materials and methods; Table 1). In total, 883 markers were identified that showed differential expression between matched epithelium and stroma in at least 50% of individual samples (LIMMA log odds >0), as well as differential expression between pooled epithelium and stroma samples (false discovery rate adjusted p value 1e-5; Additional file 8). Using these markers, hierarchical clustering was applied to the complete sample set (44 patients, 66 samples), and resolved the samples into epithelial and stromal clusters, including correct classification of the three outlier samples (Figure 2b). These genes define a normal tissue gene expression signature.

The complete list of GO terms overrepresented by genes in the normal tissue signature is located in Additional file 9 and summarized in Figure 4. Tissue specific genes in the normal signature include known fibroblast, endothelial, and epithelial genes, as well as potentially novel tissue markers. Epithelium-specific transcripts include genes associated with epithelial cell-cell junctions and the basal lamina, epithelial cell differentiation as well as epidermal growth factor receptor activity (Table 2). Stroma specific genes included extracellular matrix structural constituents, genes with collagen binding activity, and genes involved in angiogenesis and response to wounding (Table 2, Figure 4). Immunohistochemistry for selected proteins using commercially available antibodies demonstrated epithelial-specific expression of Kit, as well as elevated expression of von Willebrand factor and cd31 in stroma, and confirmed the microarray results (Figure 5).

Epithelial and stromal samples were analyzed separately to determine whether there were significant differences within tissue classes between reduction mammoplasty-derived tissue and adjacent morphologically normal tissues isolated from tumor sections. Samples in each tissue class were subjected to hierarchical clustering and subsequent bootstrapping to test for significance using all genes. Although epithelial and stromal samples each show two primary subclusters, these clusters were not statistically significant (Figure 6a,c, respectively). Importantly, adjacent and reduction samples were not associated with the subclusters in either tissue class (p = 0.732 and p = 0.075, respectively, χ² test for association). This analysis was repeated using a subset of genes (filtered by coefficient of variation ≥4), and showed similar results (data not shown).

Interestingly, morphologically normal adjacent stroma, without reduction mammoplasty samples, was found to consist of two significant subclusters whether using all genes (Figure 6d), or a filtered subset of genes (p = 1.64e-3, Fisher's exact test). These clusters were found to be associated with stromal 'cellularity' (Figure 7, defined in Materials and methods), which was assessed based on the hematoxylin and eosin staining of the normal tissues (p = 5.1e-3 and 2.1e-4, respectively, χ² test for association). A total of 669 genes were identified as differentially expressed between the adjacent stroma clusters using the LIMMA software with a false discovery rate less than 0.01, a fold change of at least 1.9, and a B statistic of at least 30. The majority of these genes were elevated in the pauci cellular fibrotic cluster when compared with the cellular stromal cluster. Furthermore, no association was found between clinical characteristics of the primary tumor, and statistically significant subclusters in either tissue class (p ≤ 0.01, data not shown) (Additional files 10 and 11). In contrast, no significant clusters were found within morphologically normal adjacent epithelium (Figure 6b, Additional file 3).

To determine whether gene expression patterns in normal breast epithelium or stroma derived from breast cancer patients can predict clinical or pathological features of the corresponding cancers, we applied a class prediction [30] approach and constructed tissue specific predictors for ER, PR, HER2, grade, tumor size, age, menopause status, recurrence, and lymph node status (Additional files 2 and 3). We used cross validation at every step of predictor construction [43], including the initial step of candidate gene selection. None of the predictors had low prediction error or low variance, with an average 50% mean prediction error by cross validation (Additional files 2 and 3). This analysis demonstrated that any gene expression differences detected in normal epithlium and stroma were neither associated with, nor predictive of, the clinical characteristics of the primary tumors.

Morphologically normal samples from different individuals are expected to show variations in gene expression due to a number of factors, including noise, differences in tissues, inter-individual variation, potential clinical differences, and the simple fact that different genes are expressed at different levels. Our goal was to identify the relative contribution of each of these sources of variation to our data (Additional file 12, panel A). Principal component analysis and multivariate analysis of variance revealed that the primary sources of variation in the data could be attributed to differences between tissues (Bonferroni corrected p = 7.9e-16, principal components 2 and 3), representing 3.98% of the variation between genes (Additional file 12, panel B), and differences between individuals (Bonferroni corrected p = 4.9e-6, principal components 3 through 8), representing 3.58% of the variation between genes. The majority of the variation in the data (84.58%) could be attributed to variations in expression between genes within a single sample. The strong correlation between arrays introduced by the common reference design of our experiment caused this variation to be common across all arrays (Additional file 13). Together, these effects accounted for 92.13% of the observed variation in the data. The remainder of the variation in gene expression was not associated with any known factors.

The identification of gene expression profiles for morphologically normal stroma and epithelium provide unique datasets that can be used to investigate breast cancer datasets for similarity to the normal tissue profile in order to gain a better understanding of breast cancer expression profiles. When our stroma and epithelium profile was compared to a dataset established by a serial analysis of gene expression (SAGE) approach from dispersed cells from one reduction mammoplasty sample [33], we observed a minimal overlap. Our normal stroma signature (562 unique genes) showed only a 25 gene overlap with that generated by SAGE for a mixture of fibroblast, endothelial and myofibroblast cells (mapped 189 unique genes), and a 2 gene overlap with the epithelium signature (mapped 89 unique genes), whilst our normal epithelium signature (321 unique genes) overlapped by 12 genes with the epithelium signature identified by SAGE. Although the overlaps are statistically significant (p = 1.33e-15 and p = 9.07e-12, respectively, hypergeometric test), the relatively low overlap between the signatures may be due to use of only a single patient in the SAGE data when compared to 44 patients in our dataset and our filtering criteria. However, the fact that no genes are in common between the epithelial gene set and that of the fibroblast data obtained by SAGE supports the purity of both cell populations in these studies.

To investigate the implication of the expression profiles generated from normal breast tissue in situ to those of tumor related genes in breast cancer, we analyzed the expression of genes in 295 breast carcinomas using a previously published dataset [8]. The normal tissue signature was mapped to 349 genes on the custom 24 K Agilent arrays used for the cancer study [8]. Hierarchical clustering of the 295 patient samples present in the cancer dataset using the genes in our normal tissue (stroma plus epithelium) signature, revealed two primary clusters of samples (Figure 8). Based on tissue specificity defined by our normal signature, the larger cluster showed enrichment for stroma specific genes (p = 0.0038, hypergeometric test) and showed an under-representation of epithelium specific genes (p = 0.001, hypergeometric test). However, this enrichment for stroma specific genes was not ubiquitously observed for all of the 257 tumor samples in the cluster, nor was it found to be associated with either the HER2, luminal A or luminal B tumor subtypes (Figures 8 and 9). In contrast, the smaller cluster showed enrichment for epithelium specific genes (p = 4.16e-13, hypergeometric test) and under-representation of stroma specific genes (p = 4.5e-42, hypergeometric test).

The smaller of the two clusters consisted of 38 samples, which were identified as ER negative, HER2 negative, and PR negative (Figure 9). This ER/HER2/PR negative cluster was found to express many normal and basal subtype specific genes as defined by Sorlie and colleagues [35], including keratin-5, keratin-17, and gamma-glutamyl hydrolase (GGH). Based on expression of these markers, we identified the samples in this cluster as consisting of basal-like and normal-like cancer subtypes as defined previously [35]. The remaining ER negative samples in the cancer dataset were HER2 positive and were located in the larger sample cluster. Notably, the cluster of basal-like and normal-like samples remained when the data was clustered using only our normal epithelium-specific gene set, whereas the cluster was not observed when normal stroma-specific genes were used in clustering (data not shown). This indicated that the basal subtype-specific patient cluster was enriched in genes expressed in normal epithelium when compared with other tumor subtypes.

Few datasets have been generated for stroma, and this is the first extensive dataset to be generated from normal stroma. To determine whether our normal stroma data set resembled other gene expression profiles for fibroblasts, a core set of genes shown to be differentially regulated when fibroblasts are stimulated with serum [44] was examined. We identified genes from the CSR profiles that were expressed in normal tissue (Additional files 6 (panel D) and 7) using a variance filtering criteria (see Materials and methods). Of the unstimulated fibroblast genes expressed in normal tissues, 84% were expressed in stroma, while 16% were expressed in epithelium, while the majority of genes activated in wounding were not expressed in either tissue (Additional file 6, panel C). These results indicate that both normal adjacent stroma and normal reduction stroma have expression profiles more similar to unstimulated fibroblasts.

To investigate the similarity of our normal stromal profile to that of fibroblastic tumors, normal stroma and epithelium expression profiles were compared to the gene signatures of DTF and SFTs [36]. Normal stroma samples expressed significantly more DTF-specific genes than expected by chance (p ≤ 2e-16, χ² goodness of fit test), while the number of SFT-specific genes was marginally significant (p = 0.038, χ² goodness of fit test) (Additional files 6 (panel A), and 7). Interestingly, normal stroma showed a statistically significant enrichment for expression of DTF-specific genes (p = 2.48e-5) (Additional file 6, panel B).