The gene expression landscape of breast cancer is shaped by tumor protein p53 status and epithelial-mesenchymal transition

Gene expression modules were calculated from a dataset compiled from 10 independent studies, in total representing 1,608 breast cancer samples hybridized to Affymetrix HG-U133A arrays (U133A set; Additional file 1). The data were MAS5 normalized, mean centered across assays and samples were classified into molecular subtypes based on gene expression centroids from Hu et al. [6] as described [17]. Cross-hybridizing probes, defined as probes referring to more than one unique Entrez Gene ID or marked as cross-hybridizing by Affymetrix (x_at probes), were removed, and features were subsequently merged by calculating the mean expression of probes relating to the same Entrez Gene ID resulting in 12,208 gene-representative transcripts. Distant metastasis-free survival (DMFS) was not available for GSE3494 and GSE1456 and for these datasets relapse-free survival was used as a substitute for DMFS in survival analysis (Additional file 1). Clinical co-variates for the U133A set are described in Additional file 1. For validation of network modules a second gene expression breast cancer dataset representing 676 breast cancer samples was compiled from 12 independent studies performed on the Affymetrix HG-U133Plus2 platform (MAS5 normalized; Additional file 1). In addition, the NKI breast cancer dataset of 295 samples, representing an independent array technology, was used (Additional file 1). Additional datasets representing colon, ovarian, lung and bladder cancer, melanoma, diffuse large B-cell lymphoma and acute myeloid lymphoma are described in Additional file 2. For U133Plus2, data probes overlapping with the U133A platform were selected and expression data were merged based on Entrez Gene ID. Probe mapping between array platforms was done based on Entrez Gene IDs.

Prior to calculating correlations the data were filtered to remove non-varying genes. A standard deviation above 1 as cut-off criteria left the 3,824 (approximately 30%) most varying genes for further analyses. All pair-wise gene correlations were calculated for the 3,824 genes using a leave-one-out strategy: Pearson correlations between all possible gene pairs were calculated while excluding one dataset at a time; thus rendering a total of 10 correlation calculations. Only positive correlations above a set cut-off level across all these 10 calculations were used for further analyses; thereby, confounding factors inherent to single datasets were eliminated (Figure S1 in Additional file 3). Calculation of correlations between the 3,824 genes and a matrix with permuted class labels, repeated 1,000 times, gave a maximum random correlation of 0.14. Thus, a correlation above r = 0.14 could be considered significant (P = 0). Expression networks were created by connecting genes (nodes) by edges representing a minimum correlation across the 10 leave-one-out calculations above a set cut-off level, and then removing genes with less than 5 neighbors. To generate a gene expression landscape, we included genes from a network based on a correlation cut-off of 0.3, and visualized the network in Cytoscape using the pair-wise correlations as weights in a spring-embedded layout [20]. Next, each gene was placed in x-y space according to the r = 0.3 network layout and given a z-value based on the highest correlation cut-off at which it is in a network (using r = 0.3, 0.4, 0.5, 0.6 and 0.7 as cut-offs). Finally, the transcriptional landscape was visualized in R using the Krig, and Tps packages. Analysis using Spearman's rank correlation metric gave similar networks (data not shown). Modules within the created networks were mined for biological relevance using BINGO [21] and further text mining based analyses were performed using LitVan [22].

Module co-expression was evaluated by calculating the average pair-wise Pearson correlation between all genes for a module in a specific dataset. Co-expression values in validation data sets were compared to co-expression of 1,000 random gene sets of the same size (data not shown). In addition, the network average clustering co-efficient (NACC) was used as defined [23], that is, the fraction of the actual number of network connections within a defined gene module at a certain correlation cut-off level in relation to the maximum possible number of connections that could be obtained within that module. Module expression across samples was analyzed using a rank-based module activity score as previously described [24]. Relationships between module activity scores and sample annotations were analyzed using t-tests or ANOVA. For all survival analyses, patients were dichotomized on module expression above or below the module average and survival analyses were performed using the survival package in R. To control for dataset bias in survival analysis in the U133A set, robustness of results was evaluated in a leave-one-out analysis excluding one dataset at the time (data not shown). Correlations between network modules and individual genes were assayed using Spearman's rank correlations. All calculations presented were performed using PERL and R. All statistical tests were two-sided unless otherwise stated.

KPL-4 breast cancer cells were seeded and grown on an array-based siRNA screening platform, and each siRNA was assayed for effects on Ki-67 immunohistochemistry staining intensity as previously described [25]. KPL-4 was a kind gift from Dr Junichi Kurebayashi, Department of Breast and Thyroid Surgery, Kawasaki Medical School, Japan [26]. The data from the siRNA screen are available in Additional file 4. Log-transformed raw intensities were used as Ki-67 staining intensities in all analyses. Group-wise effects on Ki-67 staining intensity for genes in the two proliferation modules were analyzed per module by comparing the mean module Ki-67 intensity to a random intensity distribution based on 10,000 sampled gene groups of the same size as the assayed module. Mean module intensities were for visualization purposes centered to the mean of the respective random intensity distribution. As a comparison, the same calculation was performed for the five siRNA controls present on the array platform.

Many gene expression based studies of cancer have been hampered by small sample sets, but combining data from independent studies can potentially increase the power of such investigations [17, 27‐29]. We hypothesized that with a large number of samples, correlation in expression between genes becomes a powerful measure to identify core cancer-specific transcriptional programs. Therefore, we utilized a breast cancer gene expression dataset representing 1,608 samples combined from multiple sources (U133A set; Additional file 1) [17, 30]. For this large sample size, even a very low correlation between genes was significant (Pearson's r > 0.14, P = 0, 1,000 permutations). However, when constructing gene expression networks by connecting correlated genes we observed that, even though the connections are statistically significant, extraction of distinct modules leading to biological interpretation of the network is difficult (Figure 1a). To address this issue, we generated a gene expression landscape by visualizing the network as a heat map to identify regions with higher correlations (Figure 1b). A common concern with high throughput data is batch effects [31]. Importantly, we found that the influence of data source on results decreased with increasing correlations and became negligible at r > 0.6 (Figure S1 in Additional file 3). A network derived from correlations larger than 0.6 contained 187 genes with 1,272 connections distributed in eight visually distinct modules (Figure 1c, Additional file 5). We validated the co-expression of the modules in two independent breast cancer datasets representing 676 and 295 samples, respectively (Figure 1d, Additional file 1). Surprisingly, when testing in excess of 5,000 functionally annotated sets [4] none reached the level of co-expression observed for our modules (Figure S2 in Additional file 3), supporting the value of identifying cancer-specific transcriptional programs.

Based on published associations to breast cancer-specific tumor biology, a steroid response module (SR), a basal breast cancer module (basal), and a module containing genes (for example, FOS and EGR1) [32] related to early response to growth factor or serum stimulation (early response) were identified (Figure 1c, Additional file 5). Furthermore, one module (lipid) was representative of adipocytes, containing markers of terminal differentiation along that lineage (for example, ADIPOQ, PLIN) [33]. Additional mapping of module genes to known pathways and ontology terms suggested the remaining four modules to be associated with the cell cycle (mitotic checkpoint and mitotic progression), immune response (IR) and extracellular matrix-related processes (stroma) (Figure 1c, Additional files 5 and 6). Hence, gene expression landscape analysis is an intuitive approach for identifying biologically relevant transcriptional programs.

In order to relate module gene expression to clinical parameters and breast cancer subgroups, a rank-based module activity score [24] was calculated for each of the eight modules in each breast cancer sample (Figure S3 in Additional file 3). The SR module contained known ER-status-related genes, such as GATA3, CA12, XBP1 and FOXA1 [34‐37], and by correlation to module activity scores the expression of ER-alpha (ESR1) and the progesterone receptor (PGR) were strongly associated to this module (Spearman's rho = 0.65 and 0.50, respectively). The activity scores for the SR module showed a distinct bimodal distribution with a high activity in ER-positive as compared to ER-negative cases (P = 5*10^-75, t-test) (Figure 1e). Intuitively, one would expect the SR module to be specific for ER-positive tumors; however, some ER-negative cases also showed high SR activity (Figure 1e). A comparatively high expression of the androgen receptor (AR) within this subgroup of ER-negative samples (P = 4*10^-44, t-test; Figure S4 in Additional file 3) suggested these cases to be of the apocrine breast cancer subtype [38]. Thus, high activity of the SR module can act as a functional indicator for a general steroid response.

The basal module, containing known basal cell keratins KRT5, KRT14 and KRT17 [39], showed a subtype-specific bimodal activity score distribution with high module activity in the basal-like and normal-like subtypes (P = 6*10^-114, ANOVA) (Figure 1f). The IR module showed the highest activity in the basal-like and HER2-enriched subtypes (Figure S5 in Additional file 3), and within those subtypes high IR module activity was significantly associated with more favorable prognosis (P = 0.005 and P = 0.003, respectively, log-rank tests) as previously reported [17, 27].

Our gene expression landscape showed two distinct modules (mitotic checkpoint and mitotic progression) that both contained genes related to central mitotic processes (Figure 1c). These two cell cycle modules were difficult to differentiate with respect to function. Genes in both modules were in general annotated to similar gene ontology terms and, in particular, the majority of genes in both modules were annotated to the term M-phase (Additional file 6). However, when focusing on the differences between the two modules, we observed that in the mitotic checkpoint module there were four genes annotated to spindle checkpoint (MAD2L1, TTK, BIRC5, CENPE) and six genes annotated to regulation of cell cycle (CKS2, MAD2L1, TTK, BIRC5, CENPE, DLGAP5), whereas no genes were annotated to these terms in the other module. However, in the mitotic progression module, six genes were annotated to the microtubule cellular compartment (KIF4A, KIF15, KIF18A, KIF18B, KIF20, NUSAP1, and PRC1, of which five were annotated to microtubule-based movement), and six genes were annotated to DNA binding (E2F8, HJURP, EXO1, ERCC6L, KIF15, KIF4A, NUSAP1), whereas no genes in the mitotic checkpoint module were annotated to these categories. These differences indicated that one module is more related to regulation of the M-phase and the mitotic checkpoint, while the other module seemed more related to carrying out the M-phase. Literature mining [22] corroborated these differences (Figure S6 in Additional file 3).

To experimentally investigate the functional differences of the two mitotic modules suggested by our computational analyses, we utilized a high-throughput RNAi-based cell spot microarray screening method [25]. KPL-4 breast cancer cells were reverse transfected with a library of siRNAs targeting 5,760 genes and Ki-67 intensity was assayed as a marker for cellular proliferation [25]. By combining Ki-67 intensity data for genes in the mitotic checkpoint and progression modules separately, we could investigate module level effects on proliferation. As expected, knockdown of the mitotic progression module genes resulted in significantly lowered Ki-67 staining (Figure 2a, left) as compared to a group of unspecific control siRNAs (Figure 2a, right), suggesting that the mitotic progression genes are pivotal for progression through the cell cycle. However, knockdown of the mitotic checkpoint module genes did not result in lowered Ki-67 intensity (Figure 2a, middle), suggesting that knockdown of the mitotic checkpoint genes does not hinder mitotic progression. These results support our annotation of the modules to separate cell cycle processes and to denote them mitotic progression and mitotic checkpoint, respectively.

Elevated expression of mitotic checkpoint genes has been associated with chromosomal instability in breast cancer cells [40, 41], and the mitotic checkpoint module genes showed a considerable overlap with a signature for chromosome instability in tumors [42]. Moreover, high expression of TTK (MPS1) in our mitotic checkpoint module has been reported to promote aneuploidy in breast cancer [43]. Since the mitotic checkpoint and progression genes have been shown to be co-expressed in normal tissue [44], we suspected that they were separated in breast cancer because a subgroup of tumors challenged by chromosomal instability contained cells with a halted progression through the cell cycle [45]. To identify such tumors we investigated correlation between the mitotic checkpoint and progression modules within subgroups of breast cancer [23]. While the modules remained distinct in ER-positive samples as well as the luminal A and B subtypes, they were more interconnected in ER-negative samples and the basal-like subtype (Figure 2b, Figure S7a in Additional file 3).

Cells with a stressed mitotic checkpoint accumulate genomic aberrations [40, 41], but are subject to the p53-dependent G₁ post-mitotic checkpoint, which acts as an additional barrier against proliferation of aberrant cells [46]. Furthermore, proliferation of aneuploid daughter cells is strongly linked to p53 status [47]. Therefore, we investigated whether the separation of proliferation genes into two distinct modules in luminal tumors was related to p53 functional status. To this aim, we calculated the network average clustering co-efficient (NACC) between the mitotic checkpoint and progression modules in luminal samples with known TP53 status [48]. Indeed, we observed that while the mitotic checkpoint and progression modules were separated in TP53-wildtype samples they were connected in TP53-mutated samples (P < 0.05) (Figure 2c). Importantly, this finding was validated in an independent dataset (Figure 2d) [49]. In addition, low levels of TP53 gene expression correlated to increased interconnectivity between the two mitotic modules in luminal samples (Figure 2e), and TP53-mutated luminal samples showed elevated activity of the mitotic checkpoint and progression modules (P = 5*10^-4 and P = 9*10^-4, t-tests). Furthermore, the vast majority of basal-like tumors has dysfunctional p53 and displays high chromosomal instability, and the two mitotic modules were not separated in these tumors (Figure S7a in Additional file 3). Together, these analyses suggest a subgroup of genomically unstable luminal tumors with proliferation hindered by functional p53.

To investigate whether elevated activity of the mitotic checkpoint and progression transcriptional programs translated into disease aggressiveness, we performed survival analyses within the luminal A and B subtypes separately. The mitotic progression module only showed marginal prognostic capability within these subtypes (Figure 2f). However, high activity of the mitotic checkpoint module correlated significantly to unfavorable prognosis in both luminal subgroups (luminal A P = 3*10^-5, luminal B P = 0.01, log-rank tests) (Figure 2g). Thus, genes in the mitotic checkpoint module relate to a more aggressive disease phenotype within the otherwise low proliferating luminal A tumors, but also within the more highly proliferating luminal B tumors. Correspondingly, the mitotic checkpoint module was predictive for distant metastasis free survival (DMFS) within both grade 1 and grade 2 tumors (P = 0.007 and P = 1*10^-4, respectively, log-rank tests), whereas the mitotic progression module only was predictive for grade 2 tumors (P = 0.001, log-rank test) (Figure S7b in Additional file 3).

Cell lines have previously been shown to emulate molecular breast cancer subtypes, especially with regard to basal-like and luminal disease [50]. We calculated activity scores for the eight modules in gene expression data representing 51 breast cancer cell lines [50]. Hierarchical clustering of the module activity scores clearly separated the cell lines into luminal and basal groups (Figure 3a). The luminal cell lines showed exclusively high activity of the SR module, whereas the basal A and B cell lines showed comparatively higher activity of the proliferation-related modules (Figure 3a). Furthermore, low activity of the basal module together with high activity of the stroma module gave a cluster highly enriched for the basal B classified cell lines and cell lines recently defined as claudin-low (Figure 3a), suggesting that high activity of the stroma module relates to a more mesenchymal cell phenotype [51]. The stroma module was enriched for genes related to matrix remodeling processes (for example, VCAN, FBN1, DCN, MMP2; Additional file 5) and literature mining suggested an association to TGF-beta signaling (Figure S8 in Additional file 3), a pathway known to be involved in epithelial-mesenchymal transition (EMT) [52]. In order to further investigate a relationship between the stroma module and EMT, we used microarray data derived from induced expression of known EMT-inducing factors SNAI1, TWIST, GSC or TGF-beta1 in an immortalized human mammary epithelial cell system [53]. All of the four EMT-inducing factors clearly up-regulated genes from the stroma module, while, interestingly, genes in the basal module showed reduced expression (Figure 3b). In the clinical breast cancer data we observed a similar expression pattern of the stroma module as for a previously reported EMT-signature [53], that is, higher expression in luminal A as compared to basal-like tumors (Figure S3 in Additional file 3). However, for luminal A tumors high stroma module activity correlated to more favorable prognosis (P = 0.04, log-rank test) (Figure 3c), whereas the opposite trend was observed for basal-like tumors (P = 0.07, log-rank test) (Figure 3d). Furthermore, stroma module activity was higher in node positive as compared to node negative patients of the basal-like subtype (basal-like P = 0.007, luminal A P = 0.7, t-tests) (Figure 3e). In contrast, for luminal A samples high stroma module activity reflected small tumor size (P = 8*10^-4, basal-like P = 1, ANOVA) (Figure 3f), indicative of less aggressive disease. While a majority of the genes in the stroma module were regulated by EMT-inducing factors, many of the stroma genes are also well known fibroblast markers. Therefore, we investigated the expression of the stroma module genes in data representing primary breast fibroblasts [54]. Indeed, several of the stroma genes were also highly expressed in primary breast fibroblasts (Figure 3g). In conclusion, due to the heterogeneity of breast cancer, a transcriptional program may reflect different processes and have opposite biological effects in different breast cancer subtypes. Thus, the interpretation of a gene expression signature is highly dependent on subtypes, and both intra- and intertumoral heterogeneity should be considered.

Since several of the identified gene expression modules represented processes of broader influence on tumor progression, we assayed module co-expression in seven different cancer forms, including four carcinomas (colon, non-small cell lung carcinoma (NSCLC), ovarian and bladder), stage IV malignant melanoma, diffuse large B-cell lymphoma (DLBCL) and acute myeloid leukemia (AML) (Additional file 2). As expected, the proliferation-related modules were co-expressed across all assayed cancer forms (Figure 4a) and, in line with this, activity scores for the two mitotic modules showed a significant correlation to increasing tumor grade in ovarian carcinoma (P = 2*10^-14 and 5*10^-15, respectively. ANOVA) (Figure 4b). However, some modules were co-expressed only in certain cancer forms. For instance, the SR module was found only in breast and bladder cancer. Interestingly, it has been reported that a subgroup of bladder cancer have high AR expression [55], suggesting a gene expression scenario similar to AR-positive apocrine breast cancer. The breast cancer-derived stroma module was co-regulated in several of the assayed tumor datasets, including colon carcinoma (Figure 4a). As EMT is known to be involved in the canonical colorectal adenoma-carcinoma sequence [52], we tested whether activity of this module related to colon carcinoma patient outcome. Indeed, patients with high activity of the stroma module showed poorer disease-specific survival than patients with low stroma activity (P = 0.003, log-rank test; Figure 4c). Moreover, stroma module activity was independent of tumor stage or grade in this dataset (P = 7*10^-4, HR 3.0, 95% CI 1.6 to 5.6, Cox regression). A previous report has shown that a gene expression signature relating to tumor infiltrating lymphocytes is prognostic in advanced melanoma [56]. Correspondingly, the high activity of our IR module, mainly containing genes related to activated cytotoxic T-lymphocytes (Additional files 5 and 6), correlated to more favorable prognosis in patients with stage IV melanoma (Figure 4d). These analyses show that not only are certain gene expression modules conserved across several cancer forms, but also suggest that the biology reflected by these transcriptional programs is generally descriptive for tumor biology and clinical outcome.

Despite overlap with known markers for squamous cell morphology (for example, KRT5, KRT14, KRT17) [39, 57], the breast cancer basal module did not show strong co-expression in any of the other cancer forms. To investigate this, we created a gene expression network originating from the genes in the breast cancer-specific basal module using data representing 91 NSCLC [58]. We observed that while a core set of genes from the breast cancer module retained their high correlations, a large proportion of the gene-gene correlations were not present in the NSCLC data (Figure 4e). Using this core basal module (Figure 4e), we calculated expression sums for these genes in the NSCLC data and compared to tumor morphological type. Squamous cell lung carcinomas showed a strikingly higher expression of genes in the core basal module as compared to both adenocarcinomas and large cell carcinomas (P = 5*10^-24, ANOVA) (Figure 4f). Moreover, the core basal module showed higher co-expression in colon, ovarian and bladder cancer, as well as in DLBCL, suggesting this gene expression motif is highly conserved in cancers encompassing subtypes with basal or squamous morphology (Figure S9 in Additional file 3). These results show that a transcriptional program that is common to several cancer types contain a core set of genes that are correlated to additional genes in a cancer-specific manner. This may reflect that conserved cancer processes are regulated by distinct spectra of aberrations in different cancer forms.