One hundred and forty-six patients represented by 161 breast tumor specimens (with 23 paired tumor samples) and 10 normal breast samples (195 total microarrays) were profiled. Most of these samples appeared in previous publications 181920, with 39 being new to this study, and all of which were collected using institutional review board-approved protocols. The clinical information for all samples is in the table in Additional file 1. Included within the 161 profiled tumors were 134 primary tumors, nine regional metastases and 18 distant metastases. Patients were heterogeneously treated in accordance with the standard of care dictated by their disease stage, estrogen receptor (ER) and HER2 status.

Total RNA isolation and microarray protocols are described in Hu et al 21. Each sample was assayed versus a common reference sample 22. The microarrays used were Agilent Human oligonucleotide microarrays that were scanned on an Axon GenePix 4000B, analyzed with GenePix Pro 4.1, and Lowess normalized. All microarray data have been deposited into the GEO under the accession number of GSE3521.

The background-subtracted, Lowess-normalized log₂ ratio of Cy5 over Cy3 intensity values were filtered to select genes that had a signal intensity of > 30 units in both the Cy5 and Cy3 channels. Only genes that met these criteria in at least 70% of the 195 microarrays were included for subsequent analysis. Next, each patient was classified according to the following metastasis scoring system (MetScore): MetScore = 1 were patients that had a primary tumor and were clinically node negative (N = 0) and distant metastasis negative (M = 0); MetScore = 2 were patients that had a regional metastasis (N = 1–3) and no distant metastasis (M = 0); MetScore = 3 were patients with confirmed distant disease at the time of diagnosis (M = 1 and any N) or that were represented by an actual distant metastasis sample. We next performed a multi-class significance analysis of microarrays (SAM) using a single sample from each patient, biasing the sample selection to use the actual regional or distant metastasis samples (146 arrays, see Additional file 1). We identified the gene set that was associated with the MetScore 1-2-3 distinction, which gave 1195 genes at a false discovery rate of 5%. This gene set was next used in a one-way average linkage hierarchical cluster using the program 'Cluster' 23, with the data being displayed relative to the median expression for each gene using 'Java Treeview' 24.

Relationships between the gene expression data and the MetScore classification was further examined using a 10-fold cross-validation (CV) analysis to identify a set of genes that might distinguish a MetScore group from the others. 10-fold CV using five different statistical predictors including PAM 25, a k-nearest neighbor classifier with either Euclidean distance or one-minus-Spearman-correlation as the distance function, and a class nearest centroid metric with either Euclidean distance or one-minus-Spearman-correlation as the distance function, were used as described in Chung et al 26. We performed 10-fold CV using the five different statistical predictors with the reported CV prediction accuracies being the average of the five predictors (Tables 1, 2, 3 and 4).

For the VEGF profile, an average expression value across all 13 genes (RRAGD, FABP5, UCHL1, GAL, PLOD, DDIT4, VEGF, ADM, ANGPTL4, NDRG1, NP, SLC16A3 and C14ORF58) was determined and the patients were placed into a three-group classification based their 13-gene average log₂ expression ratio from the University of North Carolina (UNC) training data set and using the cut off values (-0.63/0.08) that were identified using X-tile 27 and relapse-free survival as the endpoint. Analyses using the VEGF profile and the training set cutoffs were also applied to an independent test set of 295 patients assayed on Agilent microarrays (that is, NKI295 28), to a set of lung carcinoma samples from Bhattacharjee et al 29, and to the glioblastoma sample set from Nutt et al 30. To perform these across-data set analyses, for the NKI295 dataset we used the log ratio of red channel intensity versus green channel intensity and the data was median centered for every gene across the 295 arrays. The Netherlands Cancer Institute (NKI) dataset was then distant weight discrimination (DWD) normalized 31 with the UNC training dataset after collapsing by NCBI Entrez GeneID; after DWD normalization, the NKI data was also column standardized. For the Affymetrix datasets the probe level intensity .CEL files were processed by robust multi-chip average. The probe sets' log intensity was median centered for every gene across all the arrays. The Affymetrix datasets were also DWD normalized relative to the UNC training data after collapsing by NCBI Entrez GeneID, and were column standardized.

Univariate Kaplan-Meier analysis was performed with a log-rank test using WinSTAT for excel. Multivariate analysis of the NKI295 test set using Cox proportional hazards modeling was conducted in SAS version 9.1; a Cox hazard model was tested that included estrogen receptor status (coded as positive vs. negative), tumor size (coded as ≤ 2 cm vs. > 2 cm), lymph node status (codes as 0, 1–3, > 3 positive nodes or M = 1), age (continuous variable, formatted in decades), grade (coded as grade 1 vs. 2, and grade 1 vs. 3), and treatment (coded as yes if treatment with chemo and/or hormonal therapy, no if no adjuvant therapy was given), and the VEGF profile of low, intermediate or high as a single categorical variable. Another Cox model was also tested that included all the clinical variables, the VEGF profile, and other expression predictors 111318192835.

In situ hybridization (ISH) on tissue microarrays containing 250 different human breast tumors (not related to the 146 patients used for microarray analysis) was performed as previously described 36. Briefly, digoxigenin (DIG)-labeled sense and anti-sense RNA probes are generated by PCR amplification of approximately 450 bp products with the T7 promoter incorporated into the primers; the primer sequences used for amplification were VEGF (Forward-tctccctgatcggtgacagt, Reverse-tcgaaaaactgcactagagacaa), ANGPTL4 (Forward: gggaatcttctggaagacctg, Reverse-tacacacaacagcaccagca) and ADM (Forward-gtgtttgccaggcttaagga, Reverse-tcggtgtttccttcttccac). In vitro transcription was performed with a DIG RNA-labeling kit and T7 polymerase according to the manufacturer's protocol (Roche Diagnostics, Indianapolis, IN). Immunohistochemistry (IHC) was performed for HIF1α using Mouse Anti-Human HIF1α (BD Biosciences #610958) according to the protocol from Vleugel et al 37; the tumors were scored for perinecrotic and diffuse staining as described in Vleugel et al.

To identify gene expression patterns associated with breast cancer metastases, we performed 195 microarrays representing 134 primary tumors, nine regional metastases and 18 distant metastasis specimens (146 different patients and 10 normal breast tissues). Each patient was classified according to a MetScore, which is roughly analogous to stage except that tumor size was not considered (see Methods). As expected, this scoring system was highly predictive of patient outcomes (Figure 1A and 1B). Using the MetScore classifications, we performed CV analyses to determine if any MetScore group might be distinct relative to the others. Low accuracy rates (56% to 65%) for the prediction of MetScore 1 vs. MetScore 2 specimens were observed; however, when MetScore 1 vs. MetScore 3 (80% to 85%) or MetScore 2 vs. MetScore 3 samples (81% to 83%) were compared, high accuracy rates were obtained, which suggests that MetScore 3 was the most distinct group.

Next, we performed a multi-class significance analysis of microarray 38 analysis using a single sample from each of the 146 patients and the MetScore 1-2-3 grouping and obtained a 1195 genes at a 5% false discovery rate. This gene set was then used in a one-way average linkage hierarchical clustering analysis (Figure 2 and Additional file 2) where the samples were first ordered according to MetScore, and then according to their correlation to the average profile (that is, centroid) of the MetScore 3 class. This analysis demonstrates that some MetScore 1 and 2 samples actually have a MetScore 3 profile; a similar result has been shown before by Ramaswamy et al 9.

The gene expression patterns from this SAM analysis were complex and there were few, if any, that directly correlated with a simple progression from MetScore 1 to 2 to 3. Included within this gene set were many clusters and/or gene sets that have been identified previously, including a luminal/ER+ pattern 113940 and a proliferation signature 4142, both of which are integral parts of a gene expression assay that predicts the likelihood of recurrence in ER+ and patients treated with tamoxifen 13. In addition, many other biologically important gene sets were identified, including an 'immediate early' gene cluster containing c-FOS and JUNB (Figure 2A) 43, a set of fibroblast genes containing PLAU, THSB2 and multiple collagen genes (Figure 2B), a set of immune cell genes (Figure 2D), and a gene set containing CXCL12 (Figure 2C); CXCL12 was the top-ranked gene from this SAM analysis and was recently identified as a chemokine whose high expression promotes tumor cell proliferation, migration and invasion 17. Analysis of these individual clusters by EASE 44, with both EASE score and Bonferroni < 0.05 used as the cut off, identified many significant gene ontology categories that included 'transcription regulation' and 'DNA/nucleic acid binding' for the FOS-JUN cluster, while the fibroblast cluster was over-represented for 'extracellular matrix', 'cell adhesion and communication', 'organogenesis', 'development', and 'regulation of protease activity'. The CXCL12 cluster was over-represented for 'cell adhesion', 'cell migration' and 'extracellular matrix'. Lastly, a small 13-gene cluster containing VEGF, Adrenomedulin (ADM) and Angiopoietin-like 4 (ANGPTL4) was identified as the 'VEGF-profile' (Figure 2E), which is discussed below in greater detail.

Our previous work identified five 'intrinsic' subtypes of breast cancer that are of prognostic and predictive value 184145. Subtype classification of the tumors using the centroid predictor from Hu et al 18 showed significant outcome predictions (Figure 1C and 1D). A Chi-squared test (p = 0.0006) showed that intrinsic subtype was significantly correlated with MetScore, with the Basal-like and HER2-enriched groups being the most frequent in MetScore 3 and with no Luminal A samples being in the MetScore 3 group. Correlations between tumor subtype and stage have been described 4647, and were recapitulated here.

A small cluster of genes containing VEGF was identified (Figure 2E) that showed high expression in MetScore 3 tumors. This gene cluster contained several secreted proteins that have been implicated in endothelial cell (VEGF and ANGPTL4), lymphatic cell (ADM) and smooth muscle cell (GAL) dynamics. As a step in evaluating this profile, we performed ISH to determine what cell type was producing VEGF, ANGPTL4 and ADM. In the vast majority of cases that showed strong ISH positivity (which totaled approximately 10% of the 250 tumors tested), it was the tumor cells themselves that produced the mRNA for these three genes, and typically all three were produced (Figure 3). In a few cases, both tumor and fibroblasts showed ISH positivity, but this was rare.

As a second step in the evaluation of the VEGF profile, we created an average expression ratio of the 13 genes for each patient and looked for correlations with outcome. By dividing the patients into low, intermediate and high-expression groups using relapse-free survival (RFS) and cutoffs determined by X-tile 27, we saw that the VEGF profile was prognostic of RFS (Figure 4A) and overall survival (data not shown) with the high expression portending a poor outcome. Rank order expression classifications (two or three groups) were also robust methods of predicting outcomes (Additional file 3). Applying the VEGF profile classification rules to an independent test set of 295 patients (that is, NKI295) 1028 also significantly predicted outcomes (Figure 4B), as did rank order classifications (Additional file 3). This classification rule was also of prognostic value on a set of lung carcinoma samples (Figure 4C and Additional file 3E and 3F), although there were too few 'low' samples to be included into the Kaplan-Meier plot analysis, and on a set of patients with glioblastoma (Figure 4D and Additional file 3G and 3H); we noted that two genes (ANGPTL4 and C14ORF58) were not found on the Affymetrix platform for lung and glioblastoma test data sets. However, the Pearson correlation is 0.992 (UNC training dataset) and 0.986 (NKI295 test dataset) respectively between the average of the 13 genes and that of the 11 genes (omitting ANGPTL4 and C14ORF58). We repeated the survival analysis for the UNC dataset and NKI test set again using the 11 genes and the results were very similar (data not shown).

A multivariate Cox proportional hazards analysis on the NKI295 test set using RFS was performed using clinical variables and the VEGF profile, and it was determined that the VEGF profile was a significant predictor of RFS (Table 1). In Fan et al 48, we evaluated the prognostic powers and concordance across multiple expression predictors including the intrinsic subtypes, the NKI 70 gene signature, a microarray-based version of the Genomic Health Inc. Recurrence Score, and the wound-response profile using this same NKI patient data set, and we have also identified other profiles of prognostic significance including an estrogen pathway 19 and p53 mutation profiles 35; therefore, we performed a Cox proportional hazards analysis (Table 2) with backwards variable selection (Table 3) to evaluate a model that contained all of the aforementioned gene expression predictors and clinical variables. The final model contained both clinical parameters and multiple gene expression predictors including the VEGF profile (Table 3). Similar results were obtained when using time to distant metastasis formation, or overall survival (data not shown), or when treating the VEGF profile as a continuous variable (Table 4).

A biological implication of the VEGF profile is that it is related to a tumor's response to hypoxic conditions, which historically has been referred to as the Warburg effect 4950. A central tenant of the Warburg effect is that a tumor's metabolism becomes more dependent upon glycolysis due to anaerobic conditions. To examine glycolysis using a genomic approach, we created a 'glycolysis-profile' using the six most highly correlated glycolytic enzyme probes (GPI, PKM2, PFKP, PGK1, GAPD, ENO1, Figure 5A); the VEGF profile and the six best glycolysis probes were highly correlated (p < 0.001, Table 5).

A fibroblast/mesenchymal signature was another profile that changed with MetScore (Figure 2B), and thus to examine the potential fibroblast cell content present within each MetScore group we determined each patient's average expression value of the genes contained with the cluster presented in Figure 2B. This gene set contains fibrillin, fibroblast activation protein alpha, six collagen protein subunits and versican, which are genes and/or proteins that are typically produced by fibroblast and/or mesenchymal cells 55. This analysis shows that the fibroblast profile is correlated with intrinsic subtype (Table 5, p = 0.012) and that the MetScore 3 samples had the lowest average expression compared with the MetScore 1 and 2 samples (ANOVA p-value = 0.005, data not shown). Pathological examination of H&E sections of the distant metastasis samples also supports this conclusion and shows scant admixed mesenchymal cells in the distant metastasis samples versus their primaries that show abundant admixed mesenchymal cells (Figure 6).

We examined whether the intrinsic subtypes, the MetScore classification, and the VEGF signature correlated with any of the following expression profiles that have been associated with metastatic potential: a) the NKI 70-gene predictor 1011, b) the 'wound-response' profile 32, c) a cell line-derived hypoxia profile 33, d) an 11-gene BMI/stem cell signature 34, e) a bone metastasis signature 14, f) a lung metastasis signature 15, g) a hypoxia metagene 56, and h) the expression profile of three individual genes (HIF1α, Snail 3, and Twist 16). These analyses identified a large amount of concordance across profiles (Table 5). For example, the breast tumor subtype was significantly correlated with the bone and lung profiles, Snail expression, and the 11-gene stem cell signature; in particular, the bone and lung profiles were associated with both ER-negative subtypes (Basal-like and HER2-enriched), and Snail expression and the 11-gene stem cell signature were the highest within the Basal-like subtype. Similar results were also observed when the VEGF profile was compared with the other profiles. Two 'hypoxia signatures' have been described and shown to be of prognostic value across a variety of tumor types including breast 3356; the large signature of Chi et al 38 showed a four-gene overlap with the VEGF profile (ADM, NDRG1, DDIT4 and ANGPLT4) while the 'hypoxia-metagene' of Winter et al 56 showed a three-gene overlap (VEGF, NDRG1 and ANGPLT4); as might be expected, all three of these profiles were correlated (Table 5, p < 0.0001).