Identification of Tumor-Initiating Cells in a p53-Null Mouse Model of Breast Cancer

Using a syngeneic p53-null mouse mammary gland tumor model that closely mimics human breast cancer, we have identified, by limiting dilution transplantation and in vitro mammosphere assay, a Lin⁻CD29^HCD24^H subpopulation of tumor-initiating cells. Upon subsequent transplantation, this subpopulation generated heterogeneous tumors that displayed properties similar to the primary tumor. Analysis of biomarkers suggests the Lin⁻CD29^HCD24^H subpopulation may have arisen from a bipotent mammary progenitor. Differentially expressed genes in the Lin⁻CD29^HCD24^H mouse mammary gland tumor-initiating cell population include those involved in DNA damage response and repair, as well as genes involved in epigenetic regulation previously shown to be critical for stem cell self-renewal. These studies provide in vitro and in vivo data that support the cancer stem cell (CSC) hypothesis. Furthermore, this p53-null mouse mammary tumor model may allow us to identify new CSC markers and to test the functional importance of these markers. [Cancer Res 2008;68(12):4674–82]

Cancer stem cells (CSC), a limited subpopulation of tumor-initiating cells, are defined as cells that retain extensive self-renewal potential through a series of generations and have the ability to recreate the heterogeneity of the original tumor through asymmetric division. These cells have been posited to be responsible for resistance to conventional therapies, recurrence, and metastasis (1, 2). CSCs were first identified and characterized in acute myeloid leukemia (AML) using antibodies which recognized specific cell surface markers and fluorescence-activated cell sorting (FACS) to isolate a small subpopulation of cells capable of self-renewal and tumor formation after transplantation into the bone marrow of immunodeficient mice (3–5). By applying a similar strategy of transplanting FACS sorted single cells from solid tumors into immunodeficient mice, a small subpopulation of tumor-initiating cells has been identified from a variety of solid tumors, including breast (see review by Clarke and colleagues ref. 6).

Cells isolated from breast cancer pleural effusions with Lin⁻CD44⁺CD24^−/low phenotype displayed increased tumorigenicity using serial limiting dilution transplantation assays into immunodeficient mice (7). A 186 gene signature was identified by comparing human breast cancer Lin⁻CD44⁺CD24^−/low cells with normal breast epithelial and myoepithelial cells through gene expression profiling study (8). This signature, although derived from only a small number of patients, was able to predict the recurrence risk of breast, lung, and prostate cancers and medulloblastoma and also showed a strong correlation with overall and metastasis-free survival. The identification and characterization of tumor-initiating cells and the molecular pathways that account for their self-renewal and survival is critical to design therapies that preferentially target these cells and sensitize them to conventional chemotherapies and radiation therapies. From these studies, however, the relationship between the Lin⁻CD44⁺CD24^−/low cells in breast cancer to normal breast stem cells is unclear. For example, are these markers expressed on normal stem cells and are they conserved in the mammary gland in mice and other species?

There are several caveats concerning the transplantation studies using xenografts of human breast cancer cells in immunocompromised mice, which need to be considered. Potential differences in the stroma and microenvironment in mice and human, as well as defects in the immune system, may have profound effects on tumor initiation and progression in these models. Difficulties in obtaining a renewable source of cells from primary patient biopsies also present serious obstacles to performing detailed mechanistic studies and for the development of preclinical models. With the exception of the hematopoietic system, detailed functional characterization of normal stem cells is usually lacking, making it difficult to perform a direct comparison of the normal cells and CSCs.

In this regard, one notable exception is the mouse mammary gland. Stem cells have been identified in the mouse mammary gland using serial limiting dilution transplantation assays into the cleared mammary fat pad of syngeneic mice by two independent groups. A small proportion of cells isolated as β1 integrin (CD29)^hiCD24⁺Lin⁻ was able to reconstitute a complete and functional mammary gland (9). In parallel studies, α6 integrin (CD49f) and CD24 were also identified as mouse mammary gland stem cell markers facilitating the isolation of cells with mammary repopulating activity (10). Both of these studies indicated that mammary gland stem cells, as defined by these markers, are predominantly cycling and possess basal characteristics.

Genetically engineered mouse (GEM) models have been extremely important in helping to uncover the pathoetiology of human diseases. Numerous GEM models of breast cancer have been developed and characterized in detail (11). Many, but not all, of these generate diploid tumors, which have a uniform histology and thus may not provide models in which to investigate the biology of tumor-initiating subpopulations (12). Tumorigenic cancer cells have recently been identified from the mouse mammary tumor virus (MMTV)–Wnt-1 mice, in which Thy1⁺CD24⁺ cancer cells were more capable of regenerating tumors than non-Thy1⁺CD24⁺ cells (13). Tumor-initiating cells have also been reported in the homogeneous MMTV-Neu transgenic mouse model (14). However, both of these models involve the targeted overexpression of oncogenes with specific MMTV that may target the more differentiated cells types, although in the case of the MMTV-Wnt model, the secreted Wnt ligand may act on surrounding cells. As an alternative to these transgenic models, we have used a p53-null mammary gland transplant model, in which p53 is deleted in the germline of BALB/c mice, and the mammary epithelium is transplanted into the cleared fat pads of syngeneic wild-type recipient mice (15, 16). Unlike many other mouse breast cancer models, a subset of the tumors arising in this model retains the estrogen receptor (ER) and antiestrogens inhibit the formation of these tumors (17). Based upon extensive gene expression analyses of p53-null mammary epithelium as well as primary tumors, it has been suggested that tumors from this p53-null BALB/c breast cancer model mimic human tumors more closely than many other models (18–20). Finally, the p53 tumor suppressor is frequently (20–40%) mutated in human breast cancers and is a marker of poor prognosis as well as for chemotheraphy and radiation resistance (21, 22).

Mammary epithelial cells (MEC) from p53-null tumors were FACS sorted based upon the mouse mammary gland stem cell markers CD29 and CD24. The sorted populations were transplanted into cleared mammary fat pads of wild-type BALB/c mice. Using this approach, a small subpopulation (5–10%) of tumor cells with stem-like properties, including self-renewal and regeneration of heterogeneous tumors upon serial transplantation, was identified from 10 independent primary tumors. In vitro mammosphere assays provided further evidence that this subpopulation possesses stem-cell properties. Finally, RNA microarray analyses on isolated p53-null tumor cell subpopulations, as well as MECs from virgin BALB/c mice, have permitted the identification of genes and regulatory pathways of potential importance in understanding tumor initiation and progression, as well as therapeutic resistance.

Antibodies. R-phycoerythrin (R-PE)–conjugated rat anti-mouse CD44, FITC-conjugated and R-PE–conjugated rat anti-mouse CD24, biotin-conjugated mouse lineage panel (anti-CD3e, CD11b, CD45R/B220, Ly-6G/C, and TER-119), biotin-conjugated rat anti-mouse CD31, SAv-APC conjugate antibody, FITC-conjugated CD49f, and their corresponding isotype controls, FITC-conjugated rat IgG2a, R-PE–conjugated rat IgG2aκ, and R-PE–conjugated rat IgG2bκ, were all purchased from BD Biosciences. FITC anti-mouse/rat CD29 and FITC Armenian hamster IgG isotype control were from BioLegend. Anti-cytokeratins 5 and 14 (anti-K5 and anti-K14) rabbit polyclonal antibodies were from Covance. Anti-cytokeratin 8 (anti-K8) was from the Developmental Studies Hybridoma Bank, University of Iowa. Anti-ERα was obtained from Santa Cruz Biotechnologies. Biotin-conjugated CD140a was purchased from eBioscience. Biotinylated goat anti-rabbit or anti-rat secondary antibodies were obtained from Vector Laboratories, Inc. Texas Red–conjugated and Alexa 488–conjugated secondary antibodies for immunofluorescence were from Molecular Probes.

Preparation of single mammary tumor cells. All animal protocols were reviewed and approved by the Animal Protocol Review Committee at Baylor College of Medicine. Ten p53-null mammary tumors were generated as described (15), all without hormone treatment except tumor T3 with 2-wk hormone stimulation and tumors T6, T7, and T10 with pituitary isograft treatment. They were anesthetized with Avertin (Sigma) before tumors were removed. Tumors were minced using razor blades and digested in 10 mL digestion media/1 g tissue [digestion buffer containing DMEM/F12, 100 μg/mL gentimycin, antibiotic-antimycotic from Invitrogen, and collagenase type III (225 units/mL; Worthington)] at 37°C. Samples were pipetted every 30 min for 2 to 2.5 h while shaking on a rotary shaker at 125 rpm. Cells were filtered through 40-μm cell strainers and washed with washing buffer [F-12 medium (Invitrogen)/5% fetal bovine serum (FBS; JRH Biosciences)/50 μg/mL gentimycin] until the supernatant was clear. Then, cells were resuspended in HBSS⁺ [HBSS (Invitrogen) containing 2% FBS and 10 mmol/L HEPES buffer (Invitrogen)] before labeling with antibodies.

Preparation of single normal MECs. This protocol is as described by Welm and colleagues (23). Eight-wk-old to ten-wk-old virgin female BALB/c mice were used to isolate single MECs. Briefly, the #4 inguinal mammary glands (lymph nodes removed) were minced into small pieces. The digestion buffer mentioned above was used to digest the gland for 1.5 h at 37°C on a rotary shaker shaking at 125 rpm. Samples were pipetted every 30 min, and centrifuged at 600 × g for 10 min to pellet the organoids. Samples were washed four times in washing buffer at 425 × g for 2 s and washed once with PBS before trypsinization in 0.5 g/L trypsin/0.2 g/L EDTA in saline 10 min. HBSS⁺ was added to dilute out trypsin. Cells were filtered through 40-μm filters and centrifuged at 671 × g for 5 min before labeling with antibodies.

Flow cytometry. Cells were labeled with antibodies at a concentration of 10 million cells/mL under optimized conditions (1:200 for CD29-FITC, and 1:100 for CD24-PE) and were subjected to FACS analysis and sorting on a triple laser MoFlo (Cytomation). Dead cells were excluded by using propidium iodide (2 μg/mL; Sigma). Data analysis was performed on FlowJo version 6.4.7, Tree Star, Inc.

Transplantation into the cleared fat pad. Clearance of MECs and transplantation procedures were performed as previously described (24). After FACS, the designated number of cells were washed once with PBS and transplanted into the cleared fat pads of 21-d-old female BALB/c mice (Harlan).

Mammosphere assays. The protocol for mammosphere assays was as described by Dontu (25) with modification. Briefly, 10,000 sorted cells per well from distinct subpopulations were grown in six-well Ultra Low Attachment plates (Corning, NY) with 2 mL serum-free mammosphere medium (DMEM/F12 with 20 ng/mL basic fibroblast growth factor, 20 ng/mL epidermal growth factor, B27, 100 μg/mL gentimycin, antibiotic-antimycotic, all from Invitrogen). The cells were fed every 3 d and passaged using 0.05% trypsin/0.53 mmol/L EDTA-4Na. 2,000 dissociated cells per well were replated in 2 mL mammosphere medium for passages thereafter. Mammospheres were counted using a Leica Dissecting scope.

Colony forming assay. Serial passaged mammospheres grown in the serum-free mammosphere medium, and collagenase-dissociated tumor cells after 1 wk of growth on plastic with mammosphere medium supplemented with 5% FBS were dissociated with trypsin as described above. Viable cells from these two groups were sorted using Sytox Red (5 μmol/L, Invitrogen) into 96-well plates with FACSAria (BD Biosciences). Eight hundred cells per well were placed in mammosphere medium supplemented with 5% FBS. The plates were irradiated, 0, 2, 4, or 6 Gy, immediately after being sorted. The cells were allowed to grow undisturbed for 2 wk. The cells were then fixed with methanol/glacial acetic acid (2:1), stained with crystal violet, and then washed with water. The colonies were then counted, and data were graphed using SigmaPlot (Systat Software, Inc.).

Immunohistochemistry. See Rosen Lab Web site⁵

for detailed procedures and the dilution concentration for each antibody.

Microarray analysis. Total RNA was isolated from the sorted subpopulations based upon Lin, CD29, and CD24 expression using the PicoPure RNA isolation kit (Arcturus) and then followed by mRNA amplification using a T7 global amplification method (Two-Cycle Target Labeling kit, Affymetrix), DNA fragmentation, biotinylation, and hybridization onto Affymetrix 430 2.0 array chips. Microarray analysis was done with Affymetrix MG 430 2.0 chip with 45,037 probe sets. Statistical analysis was done with dChip⁶

and BRB Array Tools⁷ software packages. Expression was estimated with dChip (26, 27) perfect match model with quantile normalization. Differentially expressed genes were found with BRB Array Tools using paired (for tumors) or unpaired (for normal) t test and ANOVA analysis. Random variance model (RVM) was used in all cases because of the small sample size (28). The method of Benjaminin and Hochberg (29) was used for the estimation of false discovery rate (FDR). The complete array data can be accessed at National Center for Biotechnology Information Gene Expression Omnibus Database.⁸

p53-null mammary gland mouse tumors are heterogeneous. Ten tumors from p53-null mouse model were studied as summarized in Supplementary Table S1 and depicted in Fig. 1. Histologic analysis showed that the resulting tumors are heterogeneous. Immunohistochemical staining using myoepithelial (K5, K14) and luminal markers (K8, ERα) was performed to characterize the various epithelial cell types in the p53-null mammary tumors (Fig. 1A). Some tumors (T1–T5) express K5, K14, and K8. In contrast, other tumors displayed expression of K8, very low K14, but not K5 (T6–T10). Coexpression of K8 and K14 detected by immunofluorescence staining (Fig. 1B), as shown in T1, T3, and T6, suggested the expansion of a putative bipotent progenitor, which has been suggested previously to be a potential target cell in Wnt-1 tumors (30, 31). To identify a tumor-initiating subpopulation, we generated a tumor bank and transplanted these primary tumors into the cleared fat pads of syngeneic hosts permitting the generation of sufficient numbers of tumors for further studies. Based upon marker expression, the phenotypes of these tumors remained stable during transplantation (Supplementary Fig. S1).

Identification of tumorigenic subpopulations with stem cell properties. To determine if Lin⁻CD44⁺CD24⁻ were also the appropriate cell surface markers to identify the tumorigenic cells in the p53-null mouse model, subpopulations of cells obtained by FACS based upon expression of CD44 and CD24 (Supplementary Fig. S2) were transplanted into the cleared mammary fat pad of syngeneic mice. No significant differences were found among the four Lin⁻ subpopulations (CD44⁺CD24⁻, CD44⁺CD24⁺, CD44⁻CD24⁺, and CD44⁻CD24⁻) as shown in Supplementary Table S2. Histologic analysis also showed no difference between these tumors (data not shown). Therefore, Lin⁻CD44⁺CD24⁻, which represents a highly tumorigenic human breast cancer population, was not a suitable marker set to identify tumor-initiating cells in the p53-null mouse model. Thus, cell surface markers, which have been used previously to identify mouse mammary stem cells, were used to isolate subpopulations of p53 tumor cells.

Independent p53-null mammary tumors were dissociated with collagenase followed by FACS analysis based upon the cell surface markers, Lin, CD29, and CD24. Tumors displayed distinct FACS profiles (Fig. 2A), again demonstrating the heterogeneity of tumors generated from this mouse model. Limiting dilution transplantation experiments into the cleared mammary fat pad of syngeneic BALB/c mice showed that the Lin⁻CD29^HCD24^H (Lineage^NegativeCD29^HighCD24^High) subpopulation, representing overall ∼5% to 10% of the total cell population, displayed significantly increased tumorigenic potential compared with the other subpopulations (Table 1 and Supplementary Table S3). Thus, as few as 100 Lin⁻CD29^HCD24^H cells resulted in tumors (8 of 14 transplants), whereas no tumors were observed from 100 cells of the corresponding Lin⁻CD29^HCD24^L (Lineage^NegativeCD29^HighCD24^Low), Lin⁻CD29^LCD24^L (Lineage^NegativeCD29^LowCD24^Low), or Lin⁻CD29^LCD24^H (Lineage^NegativeCD29^LowCD24^High) subpopulations. FACS analysis and immunostaining showed that the resulting tumors generated by Lin⁻CD29^HCD24^H subpopulation displayed similar FACS profiles (Fig. 3A) and marker (K5, K14, K8, and ERα) expression (Fig. 3B) when compared with the original tumors from which they were derived, suggesting that the tumor-initiating subpopulation was able to generate the heterogeneous characteristics of the original tumor. The Lin⁻CD29^HCD24^L subpopulation, which was usually <5% of the total, also displayed increased tumorigenicity when compared with the Lin⁻CD29^LCD24^L and Lin⁻CD29^LCD24^H subpopulations which represented the bulk (>90%) of the tumor cells. The purity, as indicated by FACS analysis, of the Lin⁻CD29^HCD24^H subpopulation was estimated to be ∼70% to 88% (Fig. 2C). Therefore, the tumors observed when a large number (>2,500) of Lin⁻CD29^LCD24^L and Lin⁻CD29^LCD24^H cells were injected are most likely a result of the contamination with Lin⁻CD29^HCD24^H cells. This is supported by the observations that the Lin⁻CD29^LCD24^L-generated and Lin⁻CD29^LCD24^H-generated tumors had similar FACS profiles, as well as expression of specific biomarkers when compared with the parental tumors (data not shown). Due to the distinct FACS profile of Lin⁻CD29^HCD24^L, contamination from Lin⁻CD29^HCD24^H is not likely the reason for the observed tumorigenicity of the Lin⁻CD29^HCD24^L subpopulation. FACS analysis of tumors arising from Lin⁻CD29^HCD24^L indicated that they did not recapitulate the phenotype of the parental tumor (Supplementary Fig. S3). Notwithstanding, the Lin⁻CD29^HCD24^H subpopulation displayed approximately a 70-fold enrichment in tumorigenicity compared with the total Lin⁻ tumor cell population. The Lin⁻CD29^HCD24^H tumorigenic subpopulation from individual p53-null mammary tumors was enriched for K8⁺/K14⁺ cells compared with other CD29CD24 subpopulations as shown by immunofluorescence costaining (Supplementary Fig. S4). CD49f, another putative mouse mammary gland stem cell marker, together with CD24, also defined a small percentage of tumorigenic cells by FACS and serial transplantation, as shown in Supplementary Fig. S5 and Supplementary Table S4.

Tumorigenic cells generate increased numbers of secondary mammospheres. Self-renewal is a unique property of stem cells which distinguishes them from more differentiated cells. A mammosphere assay has been developed to study self-renewal potential by plating cells in a serum-free medium with growth factor supplementation on a nonadherent substrata followed by subsequent quantitation of sphere formation (25). Secondary mammospheres from the Lin⁻CD29^HCD24^H subgroup as determined from seven independent tumors were larger in size (up to 400 μm in diameter) and number (average 110 mammospheres per 2,000 Lin⁻CD29^HCD24^H cells) compared with all other subpopulations (Fig. 4A and B). Lin⁻CD29^HCD24^L cells can also form mammospheres, but with a reduced number, while barely any mammospheres were observed from Lin⁻CD29^LCD24^L. One potential limitation of the mammosphere assay is the possible occurrence of cell aggregation. To address this concern, single Lin⁻CD29^HCD24^H cells were plated into individual wells. Under these conditions, the mammosphere forming efficiency was similar as that observed for the bulk Lin⁻CD29^HCD24^H cells (Supplementary Fig. S6), suggesting that at the cell concentrations used, the mammosphere assay is measuring the potential for self-renewal and not cell aggregation. FACS analysis on passaged mammospheres derived from tumor T7 revealed that p53-null mammospheres are enriched in the Lin⁻CD29^HCD24^H tumor-initiating cells, whereas cells cultured on plastic with serum are not (Supplementary Fig. S7). To test whether the passaged mammospheres were resistant to radiation treatment, serial passaged mammospheres from a p53-null mammary tumor (T7) were treated with increasing doses of radiation up to 6 Gy, a lethal dose for the bulk of the p53-null mammary tumor cells when plated with serum on plastic, as shown in Fig. 4C. Our colony forming assay, results from passage 10 of tumor T7, suggested that mammospheres, which mainly consist of Lin⁻CD29^HCD24^H tumorigenic cells, displayed radiation resistance. The result has been repeated with an independent mammosphere passage.

Microarray analysis. To begin to identify signaling pathways that distinguish the tumorigenic and nontumorigenic subpopulations, microarray analysis was performed on the different subpopulations. For this analysis, subpopulations from three tumors (T1, T2, and T7) were collected by FACS based upon expression of CD29 and CD24, and RNA was isolated from each of the subpopulations. Genes significant at P < 0.01 using the univariate t test with RVM were identified in the comparisons of Lin⁻CD29^HCD24^H versus each of the rest of the three groups (Lin⁻CD29^HCD24^L, Lin⁻CD29^LCD24^H, and Lin⁻CD29^LCD24^L). This P value corresponded to FDR of ∼10%. There were 710 common probe sets (ps) in those three comparisons (representing 1.6% of the total 45,037 ps) differentially expressed in the Lin⁻CD29^HCD24^H subpopulation compared with the other three subpopulations, Lin⁻CD29^HCD24^L, Lin⁻CD29^LCD24^H, and Lin⁻CD29^LCD24^L (Fig. 5A). Among 710 ps identified, 527 ps were up-regulated (from 1.4-fold to 4-fold), and 183 ps were down-regulated (from 2-fold to 17-fold; see Supplementary Table S5). A heat map was generated from microarray data of various subpopulations based on the Lin⁻CD29^HCD24^H cells differentially expressed 710 ps (Fig. 5B), which were applied to ingenuity pathway analysis (IPA) for further analysis. Among 710 ps, 690 of them were mapped to known genes and 462 of them were found in the IPA knowledge database and labeled “focus” genes. Functional characterization of these differentially regulated genes assigns them to diverse biological processes, including cell cycle; cellular assembly and organization; DNA replication, recombination, and repair; gene expression; and cell-to-cell signaling and interaction. Those top five IPA-identified molecular and cellular functions, in which either up-regulated or down-regulated genes were involved, are depicted in Fig. 5B with the most significant genes among the networks highlighted below.

Lin⁻CD29^HCD24^H subpopulation showed increased expression of the polycomb genes Bmi1 and Ezh2, whose levels were up-regulated in a number of malignancies, including breast cancer. Bmi1 plays an important role in regulating the self-renewal capacity of hematopoietic, as well as human mammary gland stem cell (32, 33). One of the more significant observations is the increased expression of DNA damage response and repair genes among Lin⁻CD29^HCD24^H cells. Nek1, Brca1, Chek1, Hus1, Ung, Xrcc5, Sfpq, and Uhrf1 are among these genes. The altered expression of a subset of interesting genes was validated by quantitative PCR, as shown in Supplementary Fig. S8.

The ultimate goal of using mouse models is to investigate how the tumor-initiating cells identified in the tumors relate to normal stem cells and which pathways might be deregulated to provide more insight into the cell of origin, as well as to identify possible mechanisms involved in therapeutic resistance. The p53-null normal mammary epithelium has been shown to exhibit very similar global gene expression profiles compared with p53 wild-type BALB/c mammary epithelium through SAGE analysis (20). Thus, this allowed a comparison of the gene expression by microarray analysis of p53-null mammary tumor versus that of p53 wild-type epithelium.

Total RNAs from FACS-sorted subpopulations of wild-type BALB/c mice MECs, Lin⁻CD29^HCD24^P(MEC) [Lin⁻CD29^HighCD24^Positive, putative mouse mammary gland repopulating unit (MRU), ∼12%], Lin⁻CD29^PCD24^H(MEC) (Lin⁻CD29^PostiveCD24^High, 20%), Lin⁻CD24^L(MEC) (Lin⁻CD24^Low, ∼46%) based upon expression of Lin, CD29, and CD24 were isolated (Supplementary Fig. S9). The expression profiling of the MRU was compared with other subpopulations using a similar protocol to that used in the analysis of the p53-null tumor samples. Probe sets at P < 0.01 were chosen, and IPA was used to identify target genes differentially expressed within the MRU. Consistent with the findings of Stingl and colleagues (9, 10), MRU cells expressed basal/myoepithelial markers with high expression of Krt14, Krt5, and smooth muscle genes Actg2, Cnn1, Myh11, Myl9. On the other hand, expression of Krt8, Krt19, and Muc1 was high in the Lin⁻CD29^PCD24^H(MEC) subpopulation.

Evidence to support the CSC hypothesis has been derived primarily from studies of a variety of leukemias and solid human tumors in which FACS-sorted cell populations were transplanted into immunocompromised mice. In a recent report, however, based upon the study of three syngeneic mouse lymphoma tumor models (34), the authors concluded that the majority of the cells isolated from these relatively homogeneous mouse tumors were able to initiate tumor formation. These results contrasted with those of John Dick and colleagues (5), who had identified a rare leukemia initiating population of CSCs based primarily on xenografts of AML. Kelly and colleagues (34) questioned whether the overall microenvironment in xenograft models using immunocompromised mice might have accounted for the differences observed in these studies, thus casting doubt on the CSC hypothesis. However, an alternative interpretation is that these apparently disparate results most likely reflect the importance of selecting appropriate genetically engineered heterogeneous mouse models in which to perform these studies. The present study using the heterogeneous p53-null mouse mammary tumor model provides direct evidence using limited dilution transplantation experiments for the existence of a tumor-initiating subpopulation of CSCs in a syngeneic mouse model.

The anti-CD44 antibody successfully used for the isolation of tumor-initiating cells from breast cancer cells present in pleural effusions (7) did not facilitate the isolation of a tumor-initiating subpopulation in the p53-null mouse model. This may reflect differences in the epitopes recognized by the specific anti-CD44 antibody expressed in human breast cancer compared with mouse mammary tumor cells, possibly due to alternative splicing or posttranslational modifications (35). CD44 has been identified as an important regulator of AML stem cell homing (36). However, the functional importance of CD44 in mouse mammary tumor-initiating cells has yet to be determined. Whereas human breast cancer CSCs have been defined operationally as CD24^−/Low, the tumor-initiating population of p53-null tumors was predominantly CD24^H, although some activity was observed in the Lin⁻CD29^HCD24^L subpopulation. Unlike in human, where CD24 was only expressed on luminal epithelial cells (37), in mice, the metastasis-associated CD24, a glycosyl phosphatidylinositol–linked membrane protein, is expressed on major hematopoietic lineages, developing neural and epithelial cells, and has been suggested to be a marker for luminal MECs (38). The significance of this observation remains to be determined but may be a consequence of the possible origin of these tumors from a bipotent progenitor, as discussed below.

Breast cancer is not a disease that is driven through a simple mechanism, but through a complex set of changes (both genetic and epigenetic) in many pathways. Bmi1 has been reported to play an important role in maintaining stem cell self-renewal. Up-regulation of Bmi1 within the tumorigenic subpopulation may partially explain the increased mammosphere formation efficiency of such cells, but this remains to be shown directly (32, 39). Bmi1 can be recruited to histone H3 by Ezh2, which was also increased in this tumorigenetic Lin⁻CD29^HCD24^H subpopulation. Overexpression of Ezh2 maintained long-term repopulating potential by preventing exhaustion of hematopoietic stem cells (40). Kamminga and colleagues identified 46 genes, categorized in three groups, that intereacted with Ezh2 in epigenetic chromatin modification of hematopoietic stem cells. Among these, 15 genes are present in the up-regulated gene list in the current study, suggesting a similar mechanism may be active.

Bao and colleagues (41) showed that after radiation, DNA damage checkpoint proteins were more activated in tumor cells bearing CD133⁺ (a marker for both neural and brain CSCs) than in CD133⁻ tumor cells, suggesting that such tumor-initiating cells could be responsible for the recurrence of tumors after radiation therapy. Specific inhibitors of the checkpoint kinases, Chk1 and Chk2, were shown to sensitize CD133⁺ tumor cells to radiation treatment. At cell cycle checkpoint, DNA damage response and repair proteins are highly up-regulated in Lin⁻CD29^HCD24^H tumor-initiating cells. Interestingly, the proportion of Lin⁻CD29^HCD24^H tumor-initiating cells seems to increase upon successive mammosphere passaging, suggesting that the loss of p53 may promote symmetric division and expansion of this subpopulation. Taken together with the observation of radiation resistance of the mammospheres, a more efficient DNA damage repair mechanism may exist in Lin⁻CD29^HCD24^H cells compared with the other subpopulations. Ongoing studies are directed at testing this hypothesis.

There are controversial reports as to ER status of mammary gland stem cells. Stem cells are believed to be slowly proliferating cells, and it has been reported that majority of long-lived MECs that retained the tritiated thymidine incorporation are ER⁺ luminal cells (42). Side population (SP) cells, a small percentage of cells with mammary stem cell properties, proportionally, contained six times as many ER⁺ cells as non-SP cells (43), although, thus far, no in vivo transplantation experiments have supported the self-renewal property of the SP population. In contrast, two independent studies have reported that ER⁺ cells exhibit few stem cell properties and that, instead, the basal population, which is enriched in mouse mammary stem cells, did not express ERα (44, 45). In agreement with these results, the mouse mammary gland MRU Lin⁻CD29^HCD24^P(MEC) cells exhibit basal features, with increased expression of K5 and K14 and decreased expression of ERα. However, Lin⁻CD29^HCD24^H cells from p53-null mammary tumors contain some ERα-positive cells and cells with a mixed basal and luminal lineage (Supplementary Fig. S4) supporting the hypothesis that they may have been derived from bipotent progenitors present in normal epithelium. The loss of p53 followed by other genetic changes seems to result in the deregulation of stem cell self-renewal and possibly an expansion of this progenitor population, which would then result in both basal and luminal, as well as ER⁺ and ER⁻ cells in the tumorigenic Lin⁻CD29^HCD24^H subtypes. Therefore, the cell of origin for these tumors might be a bipotent progenitor, which possibly may be ER positive.

In summary, this study has identified a tumorigenic subpopulation within the p53-null mammary tumors as supported by both in vivo transplantation and in vitro mammosphere assay and has identified genes that are preferentially expressed in the putative mouse mammary tumor-initiating cell population. The correlation of mammosphere with their tumorigenic outgrowth potential validates the use of the assay as in vitro surrogate and suggests that it may be used for high throughput screens of small molecules and RNA interference to identify pathways which are essential for the self-renewal of these cells. This may allow us to identify new CSC markers, to test the functional importance of these markers in a syngeneic mouse model, and ultimately to improve the prognosis and treatment of breast cancer.

No potential conflicts of interest were disclosed.

Grant support: NIH grants CA16303 and U01-CA84243 (J.M. Rosen) and Komen Foundation grant PDF0504283 (M. Zhang).

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.

We thank the Texas Children's Hospital Flow Laboratory for flow cytometry assistance; the Baylor Microarray Core Facility for microarray analysis; Drs. Margaret Goodell, Tracy Vargo-Gogola, Mike Lewis, Yi Li, and Stefan Siwko for constructive criticisms on this manuscript; and Drs. Edmund Chang and Heidi Weiss for help with the array analysis and statistical analysis.