Genetic heterogeneity of Myc-induced mammary tumors reflecting diverse phenotypes including metastatic potential

Abstract

Human cancers result from a complex series of genetic alterations, resulting in heterogeneous disease states. Dissecting this heterogeneity is critical for understanding underlying mechanisms and providing opportunities for therapeutics matching the complexity. Mouse models of cancer have generally been used to reduce this complexity and focus on the role of single genes. Nevertheless, our analysis of tumors arising in the MMTV-Myc model of mammary carcinogenesis reveals substantial heterogeneity, seen in both histological and expression phenotypes. One contribution to this heterogeneity is the substantial frequency of activating Ras mutations. Additionally, we show that these Myc-induced mammary tumors exhibit even greater heterogeneity, revealed by distinct histological subtypes as well as distinct patterns of gene expression, than many other mouse models of tumorigenesis. Two of the major histological subtypes are characterized by differential patterns of cellular signaling pathways, including β-catenin and Stat3 activities. We also demonstrate that one of the MMTV-Myc mammary tumor subgroups exhibits metastatic capacity and that the signature derived from the subgroup can predict metastatic potential of human breast cancer. Together, these data reveal that a combination of histological and genomic analyses can uncover substantial heterogeneity in mammary tumor formation and therefore highlight aspects of tumor phenotype not evident in the population as a whole.

A hallmark of human cancer is genetic complexity, reflecting the acquisition of multiple mutations and gene rearrangements that give rise to the tumor phenotype. Indeed, recent large-scale DNA sequencing efforts have provided direct evidence for this complexity (1–5). An ability to model the complexity that gives rise to tumor heterogeneity would clearly enhance the understanding of the oncogenic process but also would enable the development and testing of combination therapeutics that might match this complexity. Mouse models of cancer have generally used the use of an activated oncogene or the disruption of a tumor suppressor gene to initiate the oncogenic process, but in most instances this single event is not sufficient to allow for tumor development. This can be seen in the often protracted latency of tumor development as well as the identification of specific additional genetic alterations that appear in these tumors. Interestingly, these genetic alterations and the initiating event are associated with characteristic histological patterns observed in the resulting tumors (6).

The study of cell signaling pathways that control cellular proliferation and cell fate has identified the Ras and Myc proteins as two critical components that are key for the control of normal cell growth. Moreover, various studies of oncogenic events that disrupt normal cell growth regulation have revealed the frequent alteration of Myc and Ras and further suggested a synergistic relationship between these two activities. In previous experiments, it has been shown that a serum-induced increase in Myc half-life levels was mediated by Ras activation, involving the Raf/ERK kinase pathway and leading to inhibition of Myc ubiquitin-mediated degradation (7, 8). This involves ERK mediated phosphorylation of Myc at Ser-62 to stabilize Myc. Conversely, phosphorylation of Thr-58, likely mediated by GSK-3 but dependent on the prior phosphorylation of Ser-62, is associated with degradation of Myc. Further analysis has shown that a Myc protein mutated at T58 (T58A) is insensitive to these Ras-mediated events and constitutively stable (8, 9). The interactions between Myc and Ras are also reflected in observations made in transgenic mice. For instance, interbreeding MMTV-Myc and MMTV-Ras mice results in a synergistic decrease in tumor latency (10). Moreover, in multiple Myc models, the resulting tumors harbored activating mutations in kras (11).

In light of these previous observations regarding Myc and Ras interactions, we have explored the role of Ras in Myc tumorigenesis. More broadly, we have investigated the extent of genetic heterogeneity seen in Myc-induced tumors, whether involving Ras or other events.

Results

Genetic Interaction of Ras and Myc in Mammary Tumorigenesis.

We started with a focus on the interactions involving Myc and Ras given past work highlighting the interplay of these two activities. We compared tumor development in transgenic mice expressing two versions of the Myc proto-oncogene. In the first, we placed the wild-type myc cDNA under the control of the MMTV promoter. The second construct was identical to the first except that the wild-type myc cDNA was replaced by the T58A myc cDNA. RNase protection expression analysis revealed that there were two wild-type and two T58A Myc transgenic lines expressing the transgene at high levels and tumor development was monitored in these lines.

Histological Heterogeneity of Myc-Initiated Mammary Tumors.

Examination of the histology of Myc induced tumors revealed substantial heterogeneity for both wild type and T58A transgenics. Indeed, nearly 40% of the MMTV-Myc tumors (97 mice and 146 tumors) were microacinar with the remainder exhibiting various other morphologies (Fig. 1A). Microacinar tumors are characterized by small acini lined with cuboidal cells with basalar myoepithelial cells (Fig. 1B), with both cell types being neoplastic. Previous reports indicate that a progenitor of both luminal and myoepithelial cells could be targeted by oncogenes under the control of the MMTV promoter (12). In sharp contrast less than 10% of the MMTV-Myc T58A transgenic tumors were microacinar; rather, 40% of the T58A tumors exhibited a papillary histology (92 mice, 113 tumors). Papillary tumors are characterized by fibrovascular stalks lined by columnar cells that extend into gland-like spaces (Fig. 1C). Interestingly, both microacinar and papillary tumors contained prototypical MMTV-Myc cytoplasmic amphophilia with large nucleoli and course nuclear chromatin.

Fig. 1.

Varied initiating oncogenic events causes differences in tumor histology. Tumor histology from the MMTV-Myc (WT) and MMTV-Myc T58A (T58A) lines was examined. The various dominant histological patterns were scored for the various tumor lines and are shown (A). These patterns included microacinar, papillary, squamous, EMT, solid, adenocarcinoma and tumors with a mixed population of various types of tumors (Mixed). The percentage of tumors for a given class and transgenic line is shown. Examples of the histological types scored in (A) are shown and include; microacinar tumors (B), papillary (C), squamous (K14 IHC) (D), adenocarcinoma (E), mixed (F), and EMT (G).

We also noted tumors with a squamous morphology, with cells forming distinctive keratin pearls with cytokeratin 14-positive cells (Fig. 1D) and adenocarcinomas characterized by irregular glands lined by tall columnar epithelium and filled with bright pink acellular fluid (Fig. 1E) and solid tumors. In addition, there were tumors with a mixture of the various histological patterns (Fig. 1F). A number of epithelial-mesenchymal-transition (EMT) type tumors were also observed, characterized by bundles and sheets of fusiform spindle cells with large pleomorphic nuclei with open chromatin and pale polar cytoplasm (Fig. 1G). To confirm the EMT nature, immunohistochemistry for cytokeratin 18, vimentin, and smooth muscle actin (SMA) was completed, indicating that these tumors fit the dual-staining criteria for EMT type tumors (13). Together these data demonstrate that tumors with quite diverse histological phenotypes can develop from an otherwise similar initiating event.

Genetic Heterogeneity of Myc-Induced Tumors Revealed by Gene Expression Analysis.

Given the histological heterogeneity of Myc-initiated tumors, we explored the complexity through genome-scale gene expression analysis. As a starting point, we performed unsupervised hierarchical clustering to identify patterns of gene expression evident in the tumor samples. As shown in Fig. 2, at least five distinct clusters were evident from this analysis that separated the tumors coincident with the genetic initiating event as well as the histological subtype. In this initial analysis, the Neu tumors exhibited very little variation in expression profiles, forming a tight cluster distinct from the other tumors. In contrast, the wild-type Myc tumors exhibited considerable heterogeneity, segregating into at least four distinct clusters of gene expression patterns, one of which was shared by the Myc T58A samples.

Fig. 2.

Unsupervised hierarchical clustering differentiates tumors based on histological pattern and initiating oncogenic events. Unsupervised hierarchical clustering reveals the diversity arising in the tumors from wild type MMTV-Myc, MMTV-Myc T58A, and MMTV-Neu tumors. The various histological subtypes and genotypes are shown in the legend key at the bottom and are represented in the heat map style legend for each data point at the top of the figure. Samples with an activating mutation in kras are denoted with an x. Various clusters that were examined further are shown on the right axis and are labeled A–G. The clusters that were define a histological subtype are labeled at the bottom in colored boxes.

The papillary and microacinar tumors share clusters of genes, designated on the right of Fig. 2 as group G and a portion of group D. Moreover, the microacinar tumors have expression of several genes in common with the MMTV-Neu tumors in cluster A while the papillary cluster does not. The microacinar tumors also have an elevation of genes in cluster E that is not shared by any other type of tumor. Interestingly, EMT and squamous tumors have lost expression of the microacinar and papillary clusters (D–G) and instead have elevated expression of genes in clusters B and C. We also note that there is a significant enrichment of genes with E2F sites in their promoters relative to the other clusters (Table S1). In examining the clusters for ontology and for individual genes, we noted that Cluster C contained many prototypical EMT markers such as SMA and vimentin. While interesting, the annotation of these various clusters was largely uninformative as to the genetic differences between the tumors.

To explore the relationship in gene expression patterns that were influenced by the genetic initiating event, we made use of expression data derived from 13 mouse mammary tumor models (14). This expanded analysis further emphasized the heterogeneity of the Myc mammary tumors (Fig. S1). The Wap-Myc tumors exhibited a profile very similar to that of the microacinar MMTV-Myc tumors and the MMTV-int3 tumors clustered with the papillary MMTV-Myc tumors. Nevertheless, the similarities in the gene expression patterns would suggest a common phenotype for these tumors initiated by distinct events. Finally, a mixture of other tumors, including the DMBA-induced tumors, the p53-null tumors, and conditional BRCA1 deletions clustered with the MMTV-Myc tumors with EMT pathology. This analysis highlights the heterogeneity in the MMTV-Myc tumors.

Heterogeneity in Activation of Cell Signaling Pathways in Myc-Induced Tumors.

To further explore Myc tumor heterogeneity, we used previously described signatures of cell signaling pathways. These signatures represent the activation of a cell signaling pathway, in the form of a pattern of gene expression unique to that circumstance. Importantly, this ‘phenotype’ can be quantitated and assessed in various biological samples, not unlike measuring activation of a kinase. Using these methods, our approach was to examine the pathway signatures in the context of the clusters defined in Fig. 2. We have therefore maintained the sample classification and sample order used in the unsupervised hierarchical clustering to enable a direct comparison.

As shown in Fig. 3, this pathway signature analysis revealed that the various subtypes defined by histology, gene expression, and mutation were also associated with distinct patterns of signaling pathway activity. For instance, Neu tumors exhibited a pattern of pathway activation quite distinct from the Myc-induced tumors that included an activation of TGFβ, RhoA, Stat3, and E2F1. In contrast, the papillary tumors also exhibited an activation of Stat3 and E2F1 but no activation of TGFβ and RhoA. In addition to separating the tumors based on the initiating genetic mechanism, this analysis also separated the tumors based on histological properties. The squamous and EMT tumors cluster very closely. In addition, microacinar and papillary tumors share activation of many pathways including E2F1 and Myc and both lack activation of pathways such as TGFβ and Ras. However, Stat3 (high in papillary) and β-catenin (high in microacinar) differentiate the two major tumor classes. Indeed, a number of pathways divide these histological types with statistical significance (Fig. S2). Taken together, unsupervised clustering and the probability of pathway activation has subdivided tumors into classes that correlate with the histological subtypes. To validate the signatures that split the two major classes of tumors, we performed immunohistochemistry for Stat3 and β-catenin, which was consistent with the computational probabilities (Fig. S3).

Fig. 3.

Pathway probabilities cluster tumor samples. Probability of pathway activation for MMTV-Myc, MMTV-Myc T58A, and MMTV-Neu tumors was calculated for the genes and conditions listed on the right axis of the heat map. Probability (from 0 to 1) is represented by a color scale with the probability of being like control represented in blue and probability of being like the signaling pathway of interest being red. The probabilities were clustered for pathways and the samples were ordered in the same sequence shown for unsupervised clustering in Fig. 4. The histology, genotype, and kras mutation status are shown for each sample with the color codes being identified in the legend and kras mutations being illustrated with an × (A). The predicted probability of Ras pathway activation was graphed against Myc pathway activation with r² = 0.5651 with P < 0.0001 (B). The probability of Ras pathway activation was also examined in tumors with wild type and mutant kras sequence. Using a Mann-Whitney test, the significance was illustrated with P < 0.0001 (C).

This analysis also revealed that EMT tumors exhibit a uniformly high probability of Ras pathway activation which is also associated with a high frequency of Ras mutation (denoted with an ×). Strikingly, these tumors also exhibited a uniformly low level of Myc pathway activity. In accordance with the low probability of Myc activation in EMT samples, analysis of probe level data reveals that Myc expression has returned to a level normally observed in a non-transgenic mammary gland using our data and previously published data (GSE 2528). In contrast, other histologic subtypes such as microacinar and papillary tumors showed the reverse pattern of pathway activation—high Myc activity and low Ras activity. We observed a strong negative correlation between Myc and Ras pathway status (r² = 0.5651, P < 0.0001) (Fig. 3B), despite the requirement for kras mutations in Myc tumors. To test whether kras mutations were associated with an elevation in Ras pathway activity, we divided samples into tumors with and without kras mutations and examined the probability of Ras pathway activation (Fig. 3C). This revealed a significant association between mutation and predicted pathway activation status (P < 0.0001).

Heterogeneity in Tumor Phenotypes Revealed from Expression Signatures.

Analogous to the development of pathway signatures, other recent work developed expression signatures reflecting other aspects of cancer biology, including metastatic potential. Although previous work has concluded that the MMTV-Myc model is poorly metastatic (15–17), we have re-examined this property in the context of the defined subgroups. We applied a previously developed signature for metastasis (18) to the MMTV Myc tumors (Fig. 4A). Strikingly, the squamous/EMT subgroup was strongly predicted as exhibiting the metastatic signature in contrast to most of the other tumors. To directly test if the prediction of MMTV-Myc tumors with a high probability of metastasis indeed reflected metastatic potential, pulmonary histological sections were examined for the presence of metastasis from 20 tumor samples with the highest predicted metastatic probabilities and 20 samples with the lowest probabilities. As shown in Fig. 4B, there was no evidence of metastasis in the low predicted probability group. In contrast, in the samples with a high predicted probability of metastasis, we observed 25% of samples with metastatic growth (Fig. 4B). Representative histological sections are shown (Fig. 4 C and D) from samples marked with an * in Panel 4A. Examination of a typical pulmonary metastasis revealed a poorly differentiated carcinoma that is invading the pulmonary parenchyma and is too poorly differentiated to otherwise categorize. However, it does not show squamous differentiation or the EMT phenotype pattern. It is important to note that mice with metastatic lesions had multiple tumors and we were not able to note a mouse with pulmonary metastasis and only EMT-type tumors.

Fig. 4.

Metastatic signature predictions identify samples with pulmonary metastasis. The probability of a metastatic signature was calculated and is illustrated in a heat map with a low probability of metastasis shown in blue and a high probability shown in red (A). Histological sections from the lung were examined for 20 samples with the lowest probability and 20 samples with the highest probability of the metastatic signature. The histology of these pulmonary samples was closely examined and the percentage of samples with metastasis calculated (B). Representative histology from a sample lacking pulmonary metastasis (C) and with numerous small tumor nodules in the lung (D) are shown for the tumors with a corresponding * in panel A.

Comparison with Human Breast Cancer.

Computational analysis of microarray data for human breast cancer has revealed a number of defined categories including luminal, basal and Her2-positive subgroups. To determine whether the subgroups we identified in the MMTV-Myc tumors are represented within a subpopulation of human breast cancers we have compared our tumor data with human data. Initially we predicted probabilities for the human data and clustered these pathway probabilities (Fig. 5A). We then generated and validated signatures for the EMT/squamous, microacinar, and papillary subtypes from the mouse tumor data. The human dataset was then interrogated with these signatures and the probability of the EMT signature is shown (Fig. 5B). This revealed an enrichment of samples with a high probability of an EMT signature in triple negative human breast cancer (ER, PR, and Her2-negative). Indeed, we illustrated that there was a significant elevation in the EMT probability in the triple negative group (Fig. 5C) (P = 0.0046).

Fig. 5.

EMT mouse subgroup similarities in human breast cancer. Pathway probabilities for human breast cancer samples were calculated and clustered and are represented in a heat map (A). Signatures for the major mouse subgroups were generated, validated and then were applied to the human breast cancer dataset in the same order as the clustered samples (B), revealing that the EMT signature was enriched in samples that were negative for ER, PR and HER2. When we compared the EMT probability between triple negative human cancers and all other breast cancer samples, we found a significant elevation of EMT probability in the triple negative group (C). We then examined the probability of lung metastasis in the various mouse tumor subtypes (D) and found that it correlated with the EMT subgroup. In addition, we found a strong correlation between the probability of EMT and lung metastasis in human breast cancer (E).

Given the poor prognosis associated with triple negative breast cancer and the predisposition for distant metastasis, we investigated whether there was a correlation between the Squamous/EMT subgroup and metastasis in human breast cancer. First we examined the differences in metastatic probability between the squamous/EMT, papillary, and microacinar phenotypes in the mouse microarray data (Fig. 5D) using a signature generated for breast to lung metastasis (19). This revealed a significant difference between the elevated probability of metastasis in the EMT group and the reduced probability in the papillary and microacinar subgroups (P < 0.0001 for both). We then examined the correlation between the mouse tumor EMT subgroup probability and the metastatic probability in the human breast cancer dataset (Fig. 5E). This revealed that there was as strong correlation between the mouse EMT subgroup prediction and the prediction for metastasis (P < 0.0001, r² = 0.1970). Together these data illustrate that the squamous/EMT subgroup shares attributes of triple-negative breast cancer and is associated with an increased risk of metastasis.

Discussion

Recognizing that breast cancer is not a single disease state but rather a heterogeneous collection of distinct disease entities, the ability to dissect this complexity is critical to developing an understanding of the underlying mechanisms as well as effective therapeutic strategies. To address this challenge, we have made use of mouse models with a particular focus on the role of the Myc oncogene in the initiation of mammary carcinogenesis. Our studies, making use of both genetic tools as well as genomic technologies, provides further evidence for the complexity of breast tumor development even in a context where tumors develop as a result of a single initiating event.

Although the genetic initiation through Myc and the functional interaction with Ras clearly contributes to the complexity of the resulting tumors, it was also clear that the development of Myc-initiated tumors exhibited greater complexity than simply an activation of Ras. This was seen in the substantial variation in histology but more directly in the complexity of the gene expression patterns. We extended this analysis through the use of gene expression signatures developed to represent the activation and deregulation of various pathways. Using this approach to examine the Myc-Ras relationship has revealed that in the EMT type tumors that there was a low probability of Myc pathway activation and a high probability of Ras activity, corresponding with activating mutations in Kras in these samples. The loss of Myc pathway activation in these samples is likely due to the epithelial specific nature of the MMTV promoter and corresponding lack of promoter activity in the mesenchymal tissue. However, the presence of activating Ras mutations raises the possibility that activated Ras has driven the tumor into the EMT state, leading to loss of Myc expression. Indeed, previous work has illustrated that overexpression or activation of Ras can contribute to EMT (20, 21). In addition to the loss of Myc expression in EMT samples, we noted a strong inverse correlation between Myc and Ras pathway activation status. While surprising in light of the dependence of Myc on Ras activation, the correlation reveals that once Myc-induced tumors activate Kras they are predisposed to losing Myc activity, indicating that Ras may play a dominant role in tumor progression. In support of this theory, previous studies illustrated that tumors with both Myc and Ras expression regress upon deinduction of Ras, but not Myc (22). Additional work has shown that Myc tumors that regress upon deinduction of the Myc transgene only do so in the absence of Ras (23). Taken together, this suggests that there is a requirement for Ras activation in Myc tumors, but upon Ras activation, the nature of the tumor is sufficiently changed so that the role for Myc is altered.

While tumor heterogeneity is well established in the context of human breast cancer, it was somewhat surprising to note the degree of heterogeniety evident in the Myc mouse model. Indeed, when we compared the Myc tumors to a number of standard mouse models through an unsupervised hierarchical clustering analysis, the extent of variation in the Myc induced tumors relative to other models became apparent. Through this analysis we noted that the papillary Myc tumors clustered together with Int3-driven tumors, rather than standard Myc tumors, possibly indicating a shared use of the Notch pathway. The analysis of predicted probabilities for a large number of signaling pathways in addition to Myc and Ras has provided a more detailed picture of the various morphological types of Myc induced tumors. For instance, Microacinar tumors have a high probability of β-catenin activity while papillary tumors do not. The reverse is true for the activation of the Stat3 pathway in these two tumor types. These distinctions provide a unique characterization of the signaling within the heterogenous Myc induced tumors and provide mechanistic insights into the development of various histological patterns.

The power of expression signatures in revealing discrete phenotypes was perhaps best illustrated by the finding of a subset of the Myc-induced tumors that exhibited evidence of a metastatic signature and indeed, showed evidence of lung metastasis. While we show that the squamous/EMT subtype has an increased potential for metastasis, it should be noted that the pulmonary metastases were not squamous or EMT in nature. This could be due to selection against EMT in the lung, or possibly that the metastasis occurs before the transition to a mesenchymal phenotype. However, our approach has provided evidence for the heterogeneity of the Myc induced tumors in the MMTV-Myc including those with significance to the study of human cancer. Indeed when we compared the various mouse subgroups with human breast cancer, we found that squamous/EMT probabilities were elevated in triple negative tumors (ER, PR, and Her2) and that the EMT tumor signature from the mouse model was correlated with the probability of metastasis. When subgroup identification is linked with the pathway analyses, this approach provides an ability to identify underlying events associated with this phenotype that may be exploited for therapeutic regimens.

Materials and Methods

DNA.

The constructs used to create the transgenic mice were based upon the previously published p206 vector (24). The myc cDNA and the myc T58A cDNA were cloned into the vector using standard cloning methods. Preparation of the linearized fragment for microinjection was completed as previously described (25).

Animal Use.

Animal use and husbandry was in accordance with institutional and federal guidelines. Transgenic mice were generated through standard methods. Once the four founding lines were identified and expanded, female transgenic mice were maintained in a continuous breeding program. Mice were monitored twice weekly for tumor development by palpation and tumors were measured twice weekly. Kaplan-Meier curves for tumor latency were generated through GraphPad Prism (www.graphpad.com).

Histology.

Tumor samples were fixed overnight in formalin and then were processed for routine histology, and thin sections were stained. Immunohistochemistry was performed as previously described (13) with P-Stat3; Cell Signaling Technology D3A7 and β-catenin; BD Transduction Labs 610153.

RNA.

Preparation of RNA samples from flash frozen tumor samples used the Qiagen RNeasy Midi Kit (Qiagen) after roto-stator homogenization. RNase protections were conducted as previously described (26).

RT-PCR of Kras for mutation sequencing was performed using the Qiagen 1-Step RT-PCR kit. Following the reaction, the products were run on a 2% gel, product was excised and gel purified with the Qiagen Gel Purification Kit. Primers for the RT-PCR were:

• 5′ GGA GAG AGG CCT GCT GAA 3′

• 5′ TCT TCT TCC CAT CTT TGC TCA 3′.

Fifty nanograms of the purified RT-PCR product were sequenced using the nested primer with the following sequence;

• 5′TAG AAG GCA TCG TCA ACA C 3′.

RNA from Myc tumors and 20 MMTV-Neu (NDL2–5) tumor samples was submitted to the Duke Microarray Core facility. Microarray data have been submitted to GEO and can be found under GSE15904. Unsupervised clustering was performed with Cluster 3.0 and results were visualized with JavaTreeView. Pathway predictions were conducted as previously described (27). Additional signatures for HSF and Metastasis (IGS) were used as described (18). All statistical tests and Kaplan-Meier plots were performed using GraphPad Prism 4 software.