Cancer tissue functions as an ecosystem of a diverse set of cells that interact in a complex tumor microenvironment. Genomic tools applied to biopsies in bulk fail to account for this tumor heterogeneity, whereas single-cell imaging methods limit the number of cells which can be assessed or are very resource intensive. The current study presents methods based on flow cytometric analysis and cell sorting using known cell surface markers (CXCR4/CD184, CD24, THY1/CD90) to identify and interrogate distinct groups of cells in triple-negative breast cancer clinical biopsy specimens from patient-derived xenograft (PDX) models. The results demonstrate that flow cytometric analysis allows a relevant subgrouping of cancer tissue and that sorting of these subgroups provides insights into cancer cell populations with unique, reproducible, and functionally divergent gene expression profiles. The discovery of a drug resistance signature implies that uncovering the functional interaction between these populations will lead to deeper understanding of cancer progression and drug response.
Implications: PDX-derived human breast cancer tissue was investigated at the single-cell level, and cell subpopulations defined by surface markers were identified which suggest specific roles for distinct cellular compartments within a solid tumor. Mol Cancer Res; 15(4); 429–38. ©2016 AACR.
Cancer tissue is—like any tissue—by nature an ecosystem formed by a variety of phenotypically distinct cells. With the recognition of cancer as a genetic disease a few decades ago, the focus of drug discovery efforts almost sidelined these differences in the hope that the genetic differences between “healthy” and “diseased” would provide sufficient information to develop effective therapeutics and associated diagnostic tools to choose those drugs (1). Even today most molecular diagnostic tests treat tissue as a bulk entity from which a single genotype is determined in order to categorize the disease, derive a prognosis, and decide on the optimal therapy. As a plethora of targeted drugs were developed, a fast growing number of clinical trials led to increased progression-free survival while often reducing side effects compared with traditional chemotherapy or radiotherapy (2). While this lent preliminary confirmation to the general approach of aiming at one molecularly defined target, the hopes for a major improvement in overall survival were severely dampened as the clinical trials continued to show that increased survival time was almost invariably followed by relapse and emergence of drug resistance. Relapse, and in the large majority of cases concomitant metastasis, continues to be the main cause of mortality for many cancer types (2).
The role of the tumor microenvironment (TME) which has a strong influence on tumor fate (3) as well as the development of the cancer stem cell (CSC) model (4–6) offered arguments not to rely on a too simplistic tumorigenesis model and also the motivation to include phenotypic cellular differences in the investigation of cancer tissue. Deep phenotypic analysis of tumor-associated immune cells provides ample evidence that valuable information can be elucidated regarding the TME from cells embedded in and surrounding a cancer lesion (7). The initially rather static CSC concept was refined through the realization that even cells initially classified as nonstem cells can adopt a stem cell phenotype (8), and recent models indicate that the view of cells able to switch between various degrees of “stemness” may be more accurate (9).
The arguments for inclusion of phenotypic differences grew even stronger with a better recognition of a second source of intratumor heterogeneity (ITH)—in addition to TME cells colocalized with neoplastic cells—which develops as a consequence of cancer evolution. The basic concept of cancer evolution was defined almost 40 years ago (10), and several noteworthy experiments were carried out to investigate interactions between subclonal populations within a tumor (11). Partly due to technical limitations, ITH remained on the periphery for the next three decades. Observations of correlation between high ITH and poor prognosis in Esophageal cancer where evidence of heterogeneity was more easily accessible (12) revived this “dormant” branch of cancer research. Over the last decade, it has become clear that ITH is not just a peculiar side effect but likely reflects essential mechanisms enabling a tumor to react to dynamic selective pressures caused by the immune system, hypoxia, cancer drugs, and other environmental pressures (12, 13). Discovery of extensive genetic and transcriptional heterogeneity (14, 15) added necessary functional aspects to the model which has since been widely accepted. Finally, very compelling arguments for the clinical importance of an ITH-based model come from data tying heterogeneity to outcome across at least nine cancer types (16) as well as to major mechanisms of drug resistance either in the form of pre-existing resistant subclones or through de novo adaptive resistance driven by evolutionary mechanisms (17–19).
Image-based methods can reveal the presence of phenotypically different cells in tissue. Metrics to assess ITH based on imaging have also been developed (20), but while they allow basic quantification of the degree of ITH, their restrictions in sampling (one slice at a time) and number of measured parameters as well as the lack of downstream analysis options severely limit the depth of information extractable from a biopsy. Separation of cells according to specific properties prior to applying downstream analysis methods, such as next generation sequencing, gene expression, epigenetics, and other “omics,” on the other hand provides much deeper insight. Blood-borne cancers for example provide evidence that subclones display functional differences between each other when tested in mouse tumor models (21). Once identified, such subpopulations can be subjected to downstream analysis to elucidate cellular function (22). Some large-scale screening campaigns of cancer cell lines identified a range of potential surface markers that can be used in this fashion (23). More systematic tools and methods are urgently needed to make this approach accessible to research and clinical laboratories investigating solid tumor biopsies, and a more refined understanding of heterogeneity is necessary to positively affect patient care (24).
To assess solid tumor heterogeneity in human breast cancer, we established a workflow to extract tissue biopsies, dissociate the cells, and subject them to flow cytometric analysis and sorting (Fig. 1). We generated surface marker profiles of biopsies obtained from patient-derived xenograft (PDX) mouse breast cancer models using a range of reported CSC markers (4, 25–27). CD24 and CD44 as the earliest reported breast CSC markers (4, 25) and ERBB2 (HER2) as prominent breast cancer target reported to show different expression levels on the surface of breast cancer cells (28) were obvious choices to include in our panels. ITGA6 (CD49f), CD90, PROM1 (CD133), CD184, and EPCAM (CD326) have been found to be heterogeneously expressed in breast cancer cell lines (26). ITGB1 (CD29) was reported to have a role in breast cancer metastasis (29), and ALCAM (CD166) was also found to be associated with breast CSCs (30). Our results provide a profiling tool potentially complementing existing breast cancer molecular subtyping systems. Despite showing some value in predicting therapy responses, the systems in use are still limited in their accuracy, and improvements are needed, especially in triple-negative breast cancers (28, 31). To investigate whether the variation in surface markers reflects an underlying phenotypic divergence, we sorted cells from triple-negative breast cancer PDX mice and discovered evidence for profound differences in the expression profile of two distinct subpopulations.
Our results suggest that flow cytometry provides a useful tool for characterization of solid tumor biopsies and can serve as a basis to identify multiple tumor cell subpopulations of interest.
Materials and Methods
The PDX mouse tumor model system provides a renewable source of human tumor tissue displaying similar tissue structures as clinical material (32, 33). Mice were sourced from The Jackson Laboratory. The properties of the mouse cohorts used in the study are listed in Table 1.
PDX tumor extraction
Female NOD/SCID/ILIIrg−/− (NSG) mice age 6 to 7 weeks were purchased from The Jackson Laboratory with tumor material already implanted. All animal studies were performed under an Institutional Animal Care and Use Committee–approved protocol. Tumors were assessed weekly for growth and measured with a digital caliper. Tumor volume was estimated by the formula 0.5 × (length × width2). Once tumor size reached >500 mm3 (several weeks to several months), they were harvested for analysis. Mice were humanely euthanized, and the tumor was immediately removed via blunt dissection and placed in cold PBS for dissociation and analysis.
Preparation of single-cell suspensions from tumor tissue
Excised tissue was weighed and then minced using a scalpel or scissors to a size not exceeding 1 to 2 mm3. Tissue was then enzymatically dissociated into single cells as previously described (34–36) in a 37°C water bath for 30 minutes with frequent agitation. The enzymatic reaction was stopped by rinsing in PBS (Cellgro) and 1% BSA (Sigma-Aldrich). Dissociated tissue was filtered through a 70 μm sieve and treated with ACK Buffer (Life Technologies) to remove contaminating red blood cells. Cells were resuspended in PBS, without magnesium and calcium.
Cell staining for flow cytometry and analysis
Before staining, single-cell suspensions were blocked for 30 minutes, on ice in FcR block (Miltenyi Biotec). For surface marker characterization, cells were diluted in PBS and distributed into wells of a 96-well plate containing viability dye (LIVE/DEAD Fixable Near-IR or Aqua, Invitrogen) and Hoechst 33342 (Invitrogen; at 0.2 μg/mL) for viable nucleated cell identification. Monoclonal antibodies used are listed in Table 2.
Cells were incubated with antibody solutions for 30 minutes at 4°C, in the dark, and then rinsed twice in PBS before acquisition on BD LSRII flow cytometer. For sorting, cells were stained using a four-color panel combining amine viability reagent, anti-mouse cocktail, and markers identified from the primary characterization plate. After staining, cells were rinsed, and resuspended in BD FACS Pre-Sort Buffer, filtered again through a 70 μm sieve, and populations of interest were sorted on the BD FACSAria II (BD Biosciences) at 20 p.s.i. using a 100-μm nozzle. Cell doublets or clumps were removed with electronic doublet discrimination gating. Post-sorting analysis of sorted subpopulations regularly showed purity >95%. Analysis of results was performed using FACSDiva software (version 6.1.3 LSRll and version 8.0 FACSAria II). For gating strategy, see Supplementary Fig. S1.
RNA isolation and RNASeq
A total of 80,000 or more cells chosen for a specific surface marker status were collected and processed for RNA extraction. Quality assessments were determined using Agilent BioAnalyzer and NanoDrop readings. RNA specimens underwent analysis to determine their suitability for target labeling. Analyses included determination of concentration and volume, A260/A280 ratio, and 28S/18S rRNA ratio (where applicable), and an RNA integrity summary score (Agilent RIN or Caliper RQS). RNASeq was performed using the Illumina RNASeq TruSeq Stranded mRNA SQ905 Kit. For comparison with other breast cancer gene expression profiles, a collection of 1,033 breast cancer samples was used as reference dataset (38). Pathway analysis based on the RNASeq data for the two cell populations sorted with CD184 was performed using PathVisio3 (39, 40). Significant biological functions were determined using the elim method of the Kolmogorov–Smirnov test with the topGO R package: Adrian Alexa and Jorg Rahnenfuhrer (41).
Flow cytometry analysis of PDX breast cancer biopsies
We analyzed human breast cancer cells obtained from 54 different PDX breast tumor biopsies (1 per mouse) from 9 different mouse cohorts by flow cytometry using a panel of 10 antibodies to human surface markers. The 9 cohorts were derived from 8 individual patients; BRC3 and -12 were different passages of the same original tumor model (Table 1). The antibody panel comprised surface markers previously identified on cell populations enriched in tumor-initiating cells or cells with enhanced metastatic capability or signs of epithelial–mesenchymal transition (EMT; refs. 26, 27, 29, 30, 42–44). We developed a gating strategy tailored for the detection of human markers on cells from tumor tissue excised from host mice. The percentage of cells positive for a given surface marker within all human cells of the excised tumor tissue was determined after successive exclusion of debris, dead cells, and cell aggregates as well as mouse cells from the events detected by the flow cytometry (Fig. 2 and Supplementary Fig. S1). Cell viability varied between cohorts and low viability generally correlated with signs of necrosis within a biopsy (data not shown).
These surface marker measurements provide an immunophenotypic profile for each PDX cohort which was reproducible and distinct among the different PDX tumor models (Fig. 2). The interpretation that this profile may be a specific and stable property of a given tumor is further supported by the high similarity between the profiles of BRC3 and -12—two passages of the same PDX tumor model (Figs. 2 and 3). In addition, the tumor sizes and times that different mice from each cohort were analyzed varied in a range that suggests that the stability of the profile also extends through much of the tumor development.
We built a model from this dataset using the 9 most significant surface markers and performed a distance tree categorical analysis that demonstrated that for 53 of the 54 individual tumors analyzed, each is grouped with the correct cohort (Fig. 3). Tumor T177 from BRC5 was an exception because the expression pattern clustered with the closely related BRC10 cohort (marked by arrow in Fig. 3). A possible explanation is that CD29 was not measured for T177 and was missing for the clustering. Alternatively, adding other surface markers to the survey might allow further separation of individual cohorts.
A surprising finding was that clustering based on the surface markers chosen did not correlate well with their hormone receptor status. Although the BRC 5 and 10 (only non-TNB) clustered together as might be expected, the fact that several triple-negative (TN) cohorts (BRC 7, 9, 13) clustered more tightly with those two cohorts than with other TN cohorts (especially BRC8) is a direct indication of the intertumor heterogeneity of TN breast cancer. We did not observe any correlation between the immunophenotype and growth or tissue appearance (data not shown).
Next, we examined the dataset for pairwise correlations between markers. We performed permutation testing (10,000 permutations) for each pair of markers and adjusted the P values to correct for multiple testing in order to apply a meaningful statistical cutoff. As an example, we detected a negative correlation between the expression of CD24 and CD90 (Fig. 4; R2 = 0.77; P < 10−13). We investigated the 66 potential two-way correlations between the 10 human markers, fraction of murine cells in the biopsy, and % cell viability (Fig. 5). There was very low correlation of viability with any of the markers (P > 0.05), which is consistent with the notion that none of the surface markers is uniquely affected by our processing methods. On the other hand, 22 correlations (11 positive, 11 negative) proved to be statistically significant (two-tailed, alpha = 0.05), suggesting that the expression levels of several marker pairs are coordinated. This coordination would be consistent with both markers being regulated by connected cellular processes (e.g., proliferation, migration, etc.) because independent processes presumably lead to random co-occurrence.
Analysis of these correlations can thus generate specific hypotheses, and the surface marker surveys can contribute functional information in addition to the value as a categorization tool.
Transcriptional regulation in sorted cell populations
Profiling of PDX samples as described can be used as a tool in itself, but the data also revealed that each cancer analyzed here shows strong indications of ITH, which can be the starting point of deeper molecular investigations. Defined as the simultaneous presence of cells with and without a marker in question, ITH was detected in each cohort examined (Fig. 2). Interestingly, the markers which show such divisions varied among the cohorts. To investigate whether this immunophenotypic difference can reflect a more general subclonal organization of the tumor tissue, we chose one of the cohorts which reproducibly contained distinct populations either positive or negative for CD 184 (aka CXCR4) and positive for CD24. Both markers have been used to identify cells with CSC properties (25, 45) and our observation that distinct positive and negative populations for CD184 could be detected consistently in cohort BRC12. We subjected cells dissociated from tissue extracted from three different BRC12 mice to FACS sorting, creating four different cell pools from each mouse for RNA isolation: (1) unsorted bulk tumor cells; (2) bulk cells depleted of mouse cells; (3) CD24+/CD184lo cells; and (4) CD24+/CD184hi (Supplementary Fig. S2). RNASeq analysis on the whole transcriptome was performed, and gene expression differences were analyzed. A principal component analysis (PCA) of the entire dataset indicated a clear separation between populations based on species identity (PC2) and surface marker expression (PC1), which was very consistent among the triplicate samples (Fig. 6).
We compared our transcriptome data with a published dataset comprising 1,033 breast cancer samples of various intrinsic molecular subtypes (38). Our samples clustered with basal breast cancer, showing that the PDX subpopulations as well as the parent population retained the basal phenotype (Fig. 6B). However, subpopulations defined by CD184 surface expression showed major differences in their transcriptional profiles. We detected 1,197 genes which differed by more than 2-fold and below a false discovery rate (FDR) of 0.05, 246 of which were below FDR of 0.01. More intriguing even is the fact that most of these genes fell into a few distinct categories. Many of the genes highly expressed in the CD184hi cells are associated with cell proliferation, whereas genes overexpressed in the CD184lo pool are associated with regulatory processes within the cellular environment such as hypoxia, EMT, extracellular matrix (ECM) regulation, and angiogenesis (Fig. 6C). A group of genes known to be related to drug resistance in melanoma (46) showed significant upregulation on average in this pool as well (Fig. 7B), which is consistent with the notion that drug resistance is frequently associated with less proliferative cells and may point further to an important role of this subpopulation. Although these two distinct expression profiles reside in the same tumor tissue, the properties of the CD184lo subpopulation (see Fig. 2) would likely be undetectable in a bulk analysis.
A comprehensive analysis of the pathways upregulated in the respective populations expanded upon this finding; the pathways identified for CD184hi cells relate to cell proliferation. In contrast, the pathways active in the CD184lo population include downregulation of proliferation and are mostly related to extracellular communication, for example, ECM organization, cell–cell junction organization, and signal transduction (Fig. 7A).
The utility of surface markers for identifying and isolating specific cell types of interest within tumor samples has been previously demonstrated (4, 8). These studies show the benefit of this approach, offering more detailed characterization and analysis than with conventional bulk methods (7, 23, 47). Our study shows that a systematic flow cytometry survey with an 11-color panel resulted in a novel characterization tool capable of phenotypically profiling a range of breast cancer samples beyond the current standard (e.g., hormone receptor status). This presents an opportunity especially for triple-negative breast cancers where there is a need for further subgrouping (31). Surface marker profiles may offer useful functionally relevant categorization with comparatively low resource commitment. We demonstrated that our method allows correct clustering of replicate PDX mice with similar and dissimilar tumors through analysis of their surface marker phenotype with high accuracy (Fig. 3). This finding indicates that the surface marker phenotype may be a reflection of tumor intrinsic properties which can be detected reproducibly by our methods. This notion is further supported by the tight clustering of BRC3 and BRC12, which represent two passages from the same model. Interestingly, the two cohorts which are not triple-negative breast cancer (TNB) models (BRC5 and BRC10) cluster very tightly as well, suggesting that there is overlap between hormone receptor/IHC-based categorization and surface marker patterns. Because all but two models investigated are TNB, our data do not support any strong conclusions on other breast cancer types. Surprising in this regard however is the finding that one of the TNB models (BRC8) has an immunophenotype that is unique from all the others, and the relationship between the other TNBs and the hormone receptor cohorts is closer than between BRC8 and any other cohort. This observation indicates either that surface immunophenotyping provides independent phenotypic information beyond the hormone receptor status or that PDX breast cancer models may not maintain their hormone receptor status detected in the patient material. Because 5% to 40% of breast cancer patients have discordant ER status when comparing metastases with primary lesions or among multiple primary foci (48), a switch in hormone receptor status during tumor graft development cannot be excluded.
A particular strength of the classification by immunophenotype is the ability to capture sample heterogeneity which is not reflected as exhaustively in any of the current classifications. IHC reveals a small amount of this information but is limited with regard to number of parameters and suitable objective metrics. Some classification systems (e.g., based on gene expression patterns) treat a biopsy as a clonal population, assigning a single-cell property to the entire tissue while more sophisticated sequencing methods may detect the presence of subclonal mutations, but only provide cell- or population-specific information with resource commitments prohibitive for medium to high sample throughput. In contrast, immunophenotyping specifically calls out the presence of subpopulations and their respective fraction within the sample. There are multiple recent studies demonstrating the functional relevance of these subpopulations, such as cell–cell communication within a cancer lesion, which further suggests that their assessment may have clinical importance (49–52). Use of a novel classification system reflecting on quantified heterogeneity seems very compelling. Supporting this argument is our finding of multiple statistically significant two-way correlations between markers which are consistent with coordinated cellular states rather than independent regulation of single entities. These correlations may spark new hypotheses and guide the search for relevant cellular pathways controlling specific states within the context of tumor development. Interestingly, one of the strongest negative correlations detected is CD24 or CD44 (Supplementary Table S2). Given that both markers have been found to be ambiguous with regard to their ability to identify CSCs (25), it is conceivable that in some cases (as in the original report of CD44+CD24− being a breast cancer CSC signature; ref. 4), the success of finding CSCs through a combination may be based on higher sensitivity with two antibodies rather than a bona fide role as CSC marker for both proteins.
Prompted by this apparent dynamic regulation of subpopulations, we performed a whole transcriptome analysis of subclones identified by the presence or absence of the cell surface chemokine receptor CD184 (CXCR4). This marker plays a role in tumor communication with the microenvironment (44), in metastasis (53), and intratumor communication (52). CXCR4/CD184 is also being pursued as cancer drug target (54). We identified distinct cell subpopulations, CD184hi and CD184lo, in our tumor model BRC12 (Supplementary Fig. S2) and performed gene expression analysis on the subpopulations. Although a comparison of the samples with a database of breast cancer–intrinsic subtypes grouped all within the basal breast cancer category, comparison between the CD184hi and CD184lo populations revealed stark contrasts in very specific cell signaling pathways. Proliferation-related genes were highly upregulated in cells expressing CD184, whereas the CD184lo profile showed increased activity in ECM organization, angiogenesis, and promotion of differentiation of adjacent cell types. This contrast is consistent with the hypothesis that tumor tissue employs organizational elements present in healthy tissues to pursue various aspects of the hallmarks of cancer (55). Our observation that similar proportions of CD184hi and CD184lo populations were detected in two different passages of the same tumor model (BRC3 and BRC12; Fig. 2) supports the argument that these two subpopulations reflect a stable equilibrium. Reports by other investigators on more efficient engraftment of co-inoculation with CD184hi and CD184lo cells compared with either population alone and evidence for communication between those two populations (43, 52) support this view. Despite these demonstrations of the value of CD184, our results show variability in the proportion of CD184hi cells. Thus, the nature of heterogeneity regarding CD184 is likely more complex and dependent on other factors. Further studies are needed to clarify the nature of this association.
The presence of a cell compartment with apparently slower proliferative capability, but high activity, in intercellular communication raises the interesting question whether diagnostics focusing on the most abundant cells may miss significant properties residing in a minority of the cells. The apparent coexistence of a slow and fast proliferating cell compartment also implies that the tumor may use diverse strategies for greater resilience (56). Particularly any role in therapy resistance and recurrence warrants deeper investigation, and there are indeed numerous observations linking slower proliferating cells with drug resistance (57). The expression of 100 genes upregulated by the AXL kinase (Fig. 7B), which has been implicated in drug resistance in lung cancer and melanoma (46), is significantly higher in CD184lo cells, adding a further facet to this picture.
Some specific gene expression differences further suggest a distinct role of the CD184lo compartment within the tumor tissue. Integrin beta 6 mRNA expression, which is 7.7-fold higher in this subpopulation, has been shown as an independent predictor of survival, and high expression is associated with distant metastasis and poor survival in breast cancer (58) as well as in colon cancer where it also appears to confer chemoresistance (59). A number of upregulated secreted factors with roles in cancer metastasis, angiogenesis, and hypoxia (e.g., TGFB1, LOXL4, WNT4) add to the interpretation that these cells communicate actively with their surrounding and may have a more significant role in the interaction with the TME than their CD184hi counterparts. Finally, overexpression of CD274 (PD-L1) (5.8-fold; FDR = 0.047) suggests CD184lo cells may also be more resilient against the host immune system.
Analysis of the bulk tissue (unsorted sample) did not reveal any of those properties, demonstrating that important traits in cancer biology are missed by the current standard.
Our results have to be interpreted within the context of the inherent limitations of a model system. Differences between the xenotransplantation host animal and the actual patient situation inevitably exist, most of all in the presence of an intact immune system interacting with the patient cancer tissue. The methods described here should allow to uncover powerful biological insights from patient biopsies, especially if the added heterogeneity of the immune cells can be included. On the other hand, the experimental control in the model system allows to establish some basic principles as shown in our study. Because similarities between cancer tissue from a PDX model and clinical samples are well documented (32), one can reasonably expect some of these principles to apply to the clinical situation. We believe our study forms an important building block on the way to cell subtype-specific studies which allow a comprehensive exploration of the cellular networks comprising clinical cancer tissues.
We have described novel single-cell–directed immunophenotyping methods for solid tumor cells. The deep immunophenotyping provides a new method of subtyping breast cancer cells. Most importantly, we have demonstrated that solid cancers can include cell subpopulations with substantial and biologically impactful differences compared with the whole population. These properties are undetectable with current standard methods based on bulk gene expression analysis. Our findings could have major implications for cancer biology and monitoring and predicting therapy response.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed by the other authors.
Conception and design: E. Snowden, W. Porter, F. Hahn, A. Middlebrook, R. Blaesius
Development of methodology: F. Hahn, M. Ferguson, A. Middlebrook, S. Ghanekar, R. Blaesius
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. Snowden, W. Porter, F. Hahn, M. Ferguson
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E. Snowden, W. Porter, F. Hahn, F. Tong, J.S. Parker, A. Middlebrook, R. Blaesius
Writing, review, and/or revision of the manuscript: E. Snowden, W. Porter, F. Tong, J.S. Parker, S. Ghanekar, W.S. Dillmore, R. Blaesius
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E. Snowden, W. Porter, F. Hahn, M. Ferguson, A. Middlebrook
Study supervision: W.S. Dillmore, R. Blaesius
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
Note: Supplementary data for this article are available at Molecular Cancer Research Online (http://mcr.aacrjournals.org/).
Some tumor data for this paper were retrieved from the Mouse Tumor Biology Database (MTB), Mouse Genome Informatics, The Jackson Laboratory, Bar Harbor, Maine. http://www.informatics.jax.org. (October, 1998 i.e., data retrieved 8/4/16).
- Received August 26, 2016.
- Revision received November 23, 2016.
- Accepted December 21, 2016.
- ©2016 American Association for Cancer Research.