Ethics statement

The study was approved by the regional Ethics Committee in Lund (2012/89), waiving the requirement for informed consent for the present study.

Patient material

The fresh frozen MBC tissues were obtained from The Southern Sweden Breast Cancer Group’s tissue bank at the Department of Oncology, Skåne University Hospital, from Uppsala University Hospital, and Örebro Hospital, all in Sweden. Patients were diagnosed between 1983 and 2009. All cases with good quality aCGH [1] and GEX [2] data from our previous studies were included in the present study (n=53). A physician reviewed all patient charts and compiled all clinico-pathological data. A pathologist graded all tumors to current pathological standard when paraffin blocks were available; all histological grades were represented. ER, progesterone receptor (PR) and HER2 were re-evaluated (see 7,18 for further details). The patients had received different combinations of adjuvant treatment, including hormonal, chemotherapy and radiation treatment (see Table 1 for further information).

The FBC cohort included 359 tumors, representing all intrinsic subgroups of FBC and is described in [19].

Bioinformatic analyses

The computational framework CONEXIC, that integrates aCGH and GEX data for detecting candidate drivers, was used. CONEXIC uses an integrative Bayesian scoring approach [13] and is inspired by Module Networks [20]. The basic assumptions for CONEXIC are:

1. 1 Drivers should modulate the gene modules’ expression.
2. 2 The driver’s expression controls the malignant phenotype rather than the copy number, since also other mechanisms can cause altered expression of the driver.

For a detailed description of the CONEXIC method, see Akavia et.al. [13]. Briefly, the learning algorithm of CONEXIC consists of three main steps:

1. 1 Selection of candidate drivers from commonly aberrant regions of the genome.
2. 2 A Single Modulator step that finds the initial association between the candidate drivers and gene modules.
3. 3 Improvement of the Single Modulator modules by an iterative network learning step.

The 53 MBCs and 359 FBCs with paired GEX and aCGH data were analyzed in the same manner. First, JISTIC (identification of significant targets of cancer) was used to identify commonly aberrant regions (i.e. more often than what would be expected by chance) of the genome in the tumors [21]. JISTIC is an improvement of the GISTIC (Genomic Identification of Significant Targets In Cancer) algorithm [22], and uses both the frequency and magnitude of the alterations in the genome to calculate a statistical score. GISTIC has problems in identifying sub-regions when large regions are amplified or deleted, while JISTIC includes an algorithm with the capability to find these regions [21].

The GEX datasets were normalized as described before [2,19]. Probes with a standard deviation <0.25 on a log₂ scale were removed and probe sets for genes were merged (inconsistent genes were removed). Genes without aCGH data were filtered out, resulting in a gene set of 9,185 unique genes for the MBCs and a gene set of 7,512 unique genes for the FBCs. Genes in common between the MBC and FBC datasets were then selected, resulting in a final set of 5,243 genes. CONEXIC was applied to the MBCs based on both the ‘MBC/FBC common’ (5,243 genes) and the ‘MBC unique’ (9,185 genes) gene sets (Figure 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Flowchart outlining the steps in the CONEXIC analysis.

https://doi.org/10.1371/journal.pone.0078299.g001

The candidate driver lists were used in the Single Modulator step to establish an initial association to gene modules; non-parametric bootstrapping was used [23] and repeated 100 times. The candidate drivers that were selected in at least 90% of the runs were then used as candidate gene lists for the final run of the Single Modulator step. The result file was then split into two modules per candidate driver, one module each for genes positively and negatively correlated with the candidate driver. Genes that were not assigned to any gene module were clustered using K-means; 20 stable K-means clusters were found in the MBC dataset and five in the FBC dataset. The final modules from Single Modulator (after the split) and the K-means modules were used as a starting point in the Module Network step, and then by using an iterative approach each gene was re-assigned into a module that best modeled its behavior. The Bayesian score is improved at each iteration and also here a non-parametric bootstrap was used and repeated 100 times. Candidate drivers that were selected in at least 40% of the runs were used as the candidate gene lists for the final run of the Module Network step. To rank the candidate drivers from the final list, CONEXIC scores were calculated. The candidate driver with the highest CONEXIC score is the most probable driver. The CONEXIC score was calculated with a Normal Gamma score; a score is calculated for each module and the scores for all modules regulated by a candidate driver were summed up to generate the CONEXIC score.

Each driver is associated with gene modules and a threshold expression level (split value) of the driver for each gene module divides the tumors into two groups, one with tumors that express the driver below the threshold and one with tumors that express the driver above the threshold.

An automated procedure, Literature Vector Analysis (LitVAn), was used to find over-represented terms associated with the genes in the modules in scientific papers from the NCBI Gene database [13].

CONEXIC was also applied separately to FBCs in the respective intrinsic subgroups (101 Luminal A, 65 Luminal B, 78 Basal-like and 42 HER2-enriched, respectively) and it was run in the same manner as described above. Unfortunately, the limited sample size precluded this possibility in the two previously identified subgroups of MBC (Luminal M1 and M2) [2].

Finally, a ranked-based module activity score was calculated for each MBC tumor for eight gene modules reflecting biologically relevant transcriptional programs found in FBC as previously described [24].

The GEX and aCGH data have been deposited in NCBI’s Gene Expression Omnibus (GEO) and are available through GEO Series accession numbers [GSE31259] and [GSE50512], respectively.

Statistical analyses

All statistical calculations and figures were generated in R [25]. The survival and survcomp packages were used for the survival with distant metastasis free survival (DMFS) as end-point. All P-values are two-sided.

JISTIC results

A JISTIC analysis was performed to find significantly aberrant regions among the MBC tumors, whereby 67 significant genomic aberrations were identified (51 gains and 16 losses). The regions contain 1,223 genes (769 gained and 454 lost). By comparison, when GISTIC was applied to the same tumors, only 39 regions (25 gains and 14 losses) were found [1]. Correspondingly, 208 significant regions of genomic aberrations were found (89 gained and 119 lost), harboring 924 genes (450 gained and 474 lost) when JISTIC was applied to the FBC tumors.

CONEXIC results in MBC and FBC

Thirty candidate drivers were found among the MBC tumors using the ‘MBC/FBC common’ gene set (Table 2 and Table S1) and 67 among the FBC tumors (Table 3 and Table S2). The 30 candidate drivers identified in the MBC tumors regulated 97 modules and the 67 candidate drivers identified in the FBC tumors regulated 162 modules (Figure 1). Male and female breast cancers displayed remarkably different landscapes of candidate drivers, as only two candidate drivers were found in common (TAF4 and CD164). When CONEXIC was applied separately to the intrinsic subgroups of FBC, only three more candidate drivers in common with MBC were found (ARHGAP30, COG3 and SPAG5). This is in line with our previous results on transcriptional and copy number profiling, where we showed that MBCs appeared to differ from the intrinsic subgroups of FBC, and furthermore that MBCs could be divided into novel, distinct subgroups [1,2].

When CONEXIC was applied to MBCs using the ‘MBC unique’ gene set, 45 candidate drivers were identified (Figure 1) and 17 of these were shared with the 30 candidate drivers from the ‘MBC/FBC common’ gene set (17/30 (57%)). Reassuringly, among the top 15 candidate drivers from the ‘MBC unique’ gene set, 11/15 (73%) were also found in the ‘MBC/FBC common’ gene set (data not shown).

CONEXIC analyses of the intrinsic subgroups of FBC identified 28 candidate drivers in Luminal A tumors, 17 in Luminal B tumors, 22 in Basal-like tumors and 12 in HER2-enriched tumors (data not shown). Many of these candidate drivers were also detected when all FBC tumors were analyzed together. As an example, GATA3 was found as the top candidate driver among both Luminal A and Luminal B tumors. The landscapes of candidate drivers differed greatly between the intrinsic subgroups of FBC, and only ARHGAP30 (Basal-like and HER2), TAF4 (Basal-like), SPAG5 (HER2 and Luminal A) and COG3 (Luminal A) were shared with MBC.

CONEXIC identifies known drivers of FBC

In order to highlight known oncogenes and tumor suppressor genes in the candidate gene lists, these were compared with the cancer Gene Census list (COSMIC, Catalogue of Somatic Mutations In Cancer) and the Amplified and Overexpressed Genes In Cancer (AOGIC) list. COSMIC to date contains 487 genes and AOGIC contains 77 genes. Eight known cancer genes from these compiled gene lists were found among the FBC candidate drivers: GATA3, CCNE1, PRCC, MALM2, CDK4, GRB7, NFIB and VTI1A. Among these, GATA3, CCNE1, CDK4 and GRB7 are known drivers of breast cancer [26,27]. In contrast, only two known cancer genes, ZNF217 and LHFP, were identified among the thirty MBC candidate driver genes. When CONEXIC was applied to MBCs using the extended ‘MBC unique’ gene set, an overall similar landscape of candidate drivers for MBC was detected. In addition, MAP2K4, which was not included in the ‘MBC/FBC common’ gene set, was detected as a candidate driver in MBC (data not shown). MAP2K4 has previously been reported to be over-expressed in pancreatic, breast and colorectal cancer, while LHFP expression has been reported in lipomas [26]. Interestingly, GATA3 was not found to be a candidate driver for MBC in the present study, despite the fact that the majority of the MBC tumors were ER positive. GATA3 was however identified as a candidate driver in the analyses of all FBCs, and within the Luminal A and Luminal B subtypes of FBC.

Biological processes and pathways regulated by the candidate drivers

Each potential driver is associated with gene modules (examples of two modules for MBC are shown in Figures 2-3) and the tumors are split into two groups for each candidate driver with a gene expression split value for that candidate driver. The two groups for each candidate driver display differential expression patterns of the genes in the modules, and these modules can help to explain the biological processes and pathways regulated by the candidate drivers. This information makes it possible to identify the most biologically and clinically interesting candidate drivers. LitVAn was used to perform text mining on the 97 MBC modules. A number of over-represented terms were thus found for MBC, including ‘cyclin’, ‘p53’, ‘tgf’, ‘invasion’ and ‘metastasis’ (Tables 2-3 and Figures 4-5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Heatmap of the genes in the module regulated by THY1 in MBC.

Red corresponds to up-regulation and green to down-regulation. The 66 MBC tumors are sorted according to their gene expression level of THY1.

https://doi.org/10.1371/journal.pone.0078299.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Heatmap of the genes in the module regulated by SPAG5 in MBC.

Red corresponds to up-regulation and green to down-regulation. The 66 MBC tumors are sorted according to their gene expression level of SPAG5.

https://doi.org/10.1371/journal.pone.0078299.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Graphic output of the LitVAn result for the MBC THY1 module and corresponding Kaplan-Meier survival analysis.

A) Significantly over-represented terms are represented as red circles and their association (graph edges) with multiple genes (yellow circles) from the module. The green dots represent the publication that significantly associates between the gene and the term, and the numbers in the green dots are the PubMed IDs for the respective publications. B) Distant metastasis free survival of the 66 MBC patients stratified by THY1 gene expression. The numbers below the plots indicate the number of patients at risk in each group at the given time points.

https://doi.org/10.1371/journal.pone.0078299.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Graphic output of the LitVAn result for the MBC SPAG5 module and corresponding Kaplan-Meier survival analysis.

A) Significantly over-represented terms are represented as red circles and their association (graph edges) with multiple genes (yellow circles) from the module. The green dots represent the publication that significantly associates between the gene and the term, and the numbers in the green dots are the PubMed IDs for the respective publications. B) Distant metastasis free survival of the 66 MBC patients stratified by SPAG5 gene expression. The numbers below the plots indicate the number of patients at risk in each group at the given time points.

https://doi.org/10.1371/journal.pone.0078299.g005

When gene set enrichments using gene ontology (GO) terms for the modules were studied to identify terms that were significantly enriched among MBCs, terms such as ‘cell differentiation’, ‘cell cycle’, ‘division’ and ‘signal transduction’ were found. Consequently, many of the candidate driver genes identified are known to be involved in processes associated with tumorigenesis, such as proliferation, motility, invasion, metastasis and cell differentiation.

THY1, a candidate prognostic invasion marker in MBC

We used the results from the LitVAn text mining and GO enrichment analysis for the selection of candidate driver genes of biological and clinical relevance for further validation. One of the most interesting modules was regulated at the first split by THY1 (Thy-1 cell surface antigen, or CD90) and contained 41 genes; the text mining revealed over-represented terms including ‘collagen’, ‘invasion’, ‘mmp’, ‘integrin’, ‘angiogenesis’, and ‘tgf’ (Figures 2 and 4). The THY1 module was enriched for GO terms including cell adhesion and collagen. Taken together, these findings suggest that THY1 plays a role in invasion and metastasis in MBC. The THY1 positive group in the present study was significantly associated with the previously described transcriptionally derived Luminal M1 group, and the THY1 negative group with the Luminal M2 group (Fisher’s exact test, P=0.0027). Moreover, the Luminal M1 group had a significantly higher module score for the invasion and metastasis module represented by PLAU, further supporting that THY1 is involved in invasion and metastasis in MBC [2]. THY1 maps to chromosome 11q22.3 and is a glycophosphatidyl-inositol (GPI)-anchored cell surface glycoprotein with a number of biological functions involving cell-cell and cell-matrix interactions and cell motility [28]. Furthermore, THY1 is included as an up-regulated gene in the EMT core signature defined by Taube et al., and the module that it regulates contains seven of these up-regulated EMT genes. The genes up-regulated in the EMT core signature were also significantly up-regulated in the THY1 positive MBCs, while the down-regulated genes were not differentially expressed across the MBC tumors (Figure S1) [29].

The LitVAn text mining also revealed ‘tgf’ as an over-represented term in MBC and TGFβ signaling is known to play an important role in EMT [30], further supporting that the THY1 module is involved in EMT. Fredlund et al. identified an EMT-induced stroma module correlating with aggressive disease in the Basal-like subtype of FBC [31]. The THY1 module identified among the MBCs in the present study contained 50% of the genes in this stroma module and the THY1 positive MBC tumors displayed significantly higher activity for the stroma module (Figure S2). Furthermore, the THY1 negative tumors displayed low activity for the steroid response module, which is also in agreement with our earlier results, where the Luminal M1 tumors displayed low activity for the ER module [2].

EMT has been shown to be closely associated with the claudin-low subtype of breast cancer in earlier studies [29,32]. A claudin-low signature was therefore used to cluster the MBC tumors in the present study, but no correlation with the claudin-low subtype was found (data not shown) [33]. Next, the 66 MBC tumors included in our original transcriptional profiling study were split into two groups using the gene expression split value for THY1, resulting in two equally sized groups containing 33 patients each [2]. The 13 MBC tumors not included in the CONEXIC analysis in the present study displayed the same pattern as the 53 tumors included in the CONEXIC analysis (Figure 2). Importantly, THY1 was found to be significantly associated with survival among MBC patients, with high expression of THY1 predicting a significantly shorter time to DMFS (P=0.0087; Figure 4B).

In order to validate the association between THY1 expression and the genes in its module, an independent dataset with 37 MBC tumors was used [4]. The MBC tumors in the validation cohort were sorted according to increasing expression of THY1, whereupon a comparison of the heatmaps for the two datasets displayed highly similar patterns (Figure S3), adding support for THY1 as a potential candidate driver in MBC. Moreover, the genes in the THY1 module were significantly correlated to THY1 compared to all other genes in both datasets (Wilcoxon test, P=2.2e-16 and P=6.0e-11, respectively).

SPAG5, a potential driver of cell proliferation in MBC

Another biologically and clinically interesting module was regulated at the first split by SPAG5 (sperm-associated antigen 5) and contained 48 genes (Figure 3). LitVAn found ‘cyclin’, ‘mitot’, ‘mitosis’, ‘spindl’, ‘checkpoint’, ‘g2’, ‘nocodazol’, ‘g1’, ‘p53’ and ‘cdk2’ to be overrepresented terms (Figure 5A). The SPAG5 module had enrichment of GO terms including cell cycle, cell division, M phase, mitosis, spindle and regulation of cell cycle. These results indicate that the SPAG5 module is involved in the mitotic checkpoint and progression of MBC, and that SPAG5 is a potential driver of proliferation in MBC. SPAG5 maps to chromosome 17q11.2 and encodes a protein, also called astrin, associated with the mitotic spindle machinery [34]. Fredlund et al. identified two modules involved in proliferation in breast cancer, the mitotic checkpoint module and the mitotic progression module [31], and the SPAG5 module identified herein among MBCs contained 70% of the mitotic checkpoint genes and 100% of the mitotic progression genes. The highly expressing SPAG5 tumors displayed a significantly higher activity for the mitotic checkpoint and mitotic progression modules, while there were no differences for the other modules (Figure S4), further supporting that this module is involved in the proliferation of MBC. SPAG5 has been shown to be prognostic in ER positive, but not in ER negative breast cancer in women, while it has been shown to be predictive for chemotherapy response in the latter group [35]. Analysis of SPAG5 in 1,379 FBCs in GOBO (Gene expression-based Outcome for Breast cancer Online) [36] confirmed that it was prognostic for distant disease free survival in ER positive, but not ER negative FBC (data not shown).

Next, the 66 MBC tumors included in our original transcriptional profiling study were split into two groups using the gene expression split value for SPAG5, resulting in one group containing 30 tumors with low expression of SPAG5 and one group containing 36 tumors with high expression of SPAG5. The 13 MBC tumors not included in the CONEXIC analysis in the present study showed the same pattern as the 53 tumors included in the CONEXIC analysis. The SPAG5 groups were however not associated with the transcriptional Luminal M1 and M2 groups (Fisher’s exact test, P=0.18) [2]. Tumors in the SPAG5 high group had significantly increased levels of Ki67 compared to tumors in the SPAG5 low group (Wilcoxon test, P<0.05). However, when comparing survival between the groups, no significant difference could be seen (Figure 5B), which is in line with our earlier study where no prognostic value of Ki67 was seen in MBC [37].