The first evidence in support of a pluripotent mammary epithelial stem cell can be found in the classic transplantation experiments of DeOme and colleagues 12. It was shown in the mouse that transplantation of small fragments from any portion of the mammary epithelial tree to a cleared mammary fat pad regenerates an entire branching mammary epithelial tree. Furthermore, a fragment can be taken from this regenerated gland and be serially transplanted to another cleared mammary fat pad 13. Formal proof for the existence of a pluripotent stem cell as opposed to multiple developmentally committed progenitors was not available, however, until a similar experiment was performed with retrovirally tagged cells. Neonatal mice were infected with mouse mammary tumour virus (MMTV), a retrovirus, and randomly chosen fragments of the gland were transplanted from multiparous adult animals to the cleared mammary fat pad of noninfected recipients. Since MMTV integrates at random into the nuclear genome, clones derived from individual infected mammary epithelial cells could be distinguished. It was seen that an entire mammary epithelial tree could be derived from the progeny of a single MMTV-infected cell 14.

Limiting dilution transplantation experiments also support the existence of pluripotent epithelial stem cells in the mouse mammary gland. In these experiments small numbers of dissociated mammary epithelial cells are transplanted so that only a minority of fat pads have epithelial outgrowths, with most of these outgrowths derived from a single mammary epithelial cell. Three types of outgrowths can be distinguished. The first generates a limited network of ducts without lobules, and another type produces a limited outgrowth of lobules without branching ducts. The final outgrowth type has both ductal and lobular development and completely fills the mammary fat pad 1516. Retroviral labelling confirms that each of these outgrowths is clonal 14.

It intuitively seems probable that the human mammary epithelium will be organised along similar lines to that of the mouse, and recent reports support this premise. Clonal patches of X-inactivation 41 and loss of heterozygosity 42 affect large, contiguous regions of breast epithelium, and this has been interpreted as reflecting origin from a common pluripotent stem cell – although the alternative possibility that these alterations arise during development 4344 cannot be formally excluded. In-vitro culture of dissociated human mammary epithelial cells yields a bipotent myoepithelial/luminal progenitor and progenitors restricted to basal/myo-epithelial and luminal lineages. These different progenitors can be prospectively purified according to expression of a range of cell surface antigens, including MUC-1, the common acute lymphoblastic leukaemia antigen, epithelial cell adhesion molecule and α₆-integrin 222324. In-vitro culture failed to identify a pluripotent progenitor capable of forming ductal cells in addition to luminal cells and basal/myoepithelial cells. Furthermore, recent reports indicate that the population of murine mammary epithelial cells initiating colony formation in vitro is distinct from the population that repopulates the cleared mammary fat pad 18. This strongly suggests that these in-vitro studies are characterising committed progenitors rather than pluripotent stem cells.

The gold standard for proving the existence of a subpopulation of tumourigenic stem cells is the demonstration that a minority of prospectively identified cells is able to transfer the tumour and recapitulate the phenotypic heterogeneity present within the parental tumour, whereas a much larger number of tumour cells lacking the relevant marker(s) cannot. This same subpopulation should again fulfil these criteria when re-isolated from a secondary tumour. This is feasible for haematopoietic malignancies since haematopoietic cells survive well in suspension and therefore tolerate the process of sorting and bone marrow transplantation well. Unfortunately for solid tumours in general, and for breast cancer in particular, the efficiency of tumour initiation in immunodeficient mice from primary dissociated tumour cells is very low. This most probably reflects anoikis in the absence of binding to a substratum in combination with the trauma of dissociation and cell sorting.

Two strategies have been successfully employed to circumvent these limitations and to identify tumourigenic cells within breast cancers. Al-Hajj and colleagues initiated tumours in immunodeficient mice with cells derived from eight human breast-cancer-associated malignant pleural effusions 7. These cells have already survived selection in vivo for the ability to survive in suspension and to tolerate low oxygen tensions, and this may account for their tolerance to cell sorting. Although one tumour was established directly from a primary breast cancer, cell sorting was only performed following passage in an immunodeficient mouse. It was found that a small proportion of tumourigenic cells could be isolated from dissociated tumour cells by flow cytometry. Following removal of nonepithelial cells, cells with the CD44⁺CD24^-/loLin^- phenotype were 10-fold to 50-fold enriched in their ability to initiate tumours compared with unsorted cells. For three of the malignant pleural effusions, further enrichment was possible according to expression of the ESA (epithelial cell adhesion molecule) antigen. As few as 200 CD44⁺CD24^-/loESA⁺Lin^- cells consistently initiated tumours, whereas 10,000 unsorted cells did so only rarely. CD44⁺CD24^-/loLin^- cells isolated from tumours that had been initiated by these cells were able to transfer the tumour to secondary and subsequent hosts, demonstrating the capacity for self-renewal.

The formation of mammospheres from normal mammary epithelial cells has already been described. Ponti and colleagues employed a similar approach to derive mammospheres from human breast cancers 52. Mammosphere cells were undifferentiated but, interestingly, a large majority had the same CD44⁺CD24^- phenotype reported by Al-Hajj and colleagues. When transferred to differentiating culture conditions, mammospheres produced cells of both luminal and basal/myoepithelial lineages. Mammospheres initiated tumours in the cleared mammary fat pad of immunodeficient mice at 1000-fold greater dilutions than established breast-cancer-derived cell lines. Following enzymatic dissociation, 10–20% of cells from primary tumour-derived mammospheres formed secondary mammospheres and several cultures were serially expanded for more than 40 passages, proving a capacity for long-term self-renewal. Mammospheres from more aggressive tumours formed a larger number of secondary mammospheres, suggesting a greater capacity for self-renewal.

The results of both Al-Hajj and colleagues and Ponti and colleagues suggest that breast cancer cells with the capability for long-term self-renewal are enriched within the CD44⁺CD24^- subset. We should not, however, infer from this that the CD44⁺CD24^- subset of tumour cells is a homogeneous population of cancer stem cells; instead, analogous to prospectively purified mammary epithelial stem cells, this population most probably represents a heterogeneous mix of cancer stem cells and early progenitor cells.

This is supported by the observation that, for certain of Al-Hajj and colleagues' malignant pleural effusions, an ESA⁺ subpopulation of CD44⁺CD24^- cells is more highly enriched in tumourigenic cells 7. Additionally, although virtually all of the cells in Ponti and colleagues' mammospheres were CD44⁺CD24^-, only 10–20% had the capacity for self-renewal.

Both of these earlier studies, by necessity, impose an intervening selection step between primary cancer cells and cell sorting – the ability to survive in a pleural effusion in the case of Al-Hajj and colleagues' experiments, and the ability to survive in nonadherent culture conditions for Ponti and colleagues' study. It therefore remains formally possible that only the subset of breast cancer stem cells that fulfil these criteria is identified in these studies.

A Hoescht dye-effluxing SP has been identified in AML patients that is highly enriched in the ability to transfer the disease to immunodeficient mice 53. This cell population is believed to contain a high proportion of leukaemia stem cells. A similar SP has been identified in primary neuroblastoma cells and cultured cell lines 54 and in certain established breast-cancer-derived cell lines 5455. For a variety of cultured cell lines it was shown that the SP was enriched in cells with the ability to initiate tumours in immunodeficient mice 55, although these studies are somewhat confounded by the higher concentration of (toxic) Hoescht dye present in non-SP cells. Future experimental strategies may include the characterisation of SP cells from primary breast cancers and/or breast-cancer-derived mammospheres.

For most mammalian malignancies, in particular epithelial malignancies, genetic alterations accumulate sequentially 56 until a set sufficient to induce malignant transformation coexists within a single cell 5758. Stem cell compartments are organised in a hierarchical manner with a small number of infrequently proliferating pluripotent stem cells, of rapidly dividing developmentally committed transit amplifying cells and of post-mitotic terminally differentiated cells 1. Under normal circumstances any mutation that occurs in a transit amplifying cell is lost from the tissue since these cells do not have the capacity for indefinite self-renewal. It is therefore likely that the majority of mutations responsible for malignancy will arise in tissue stem cells 30 or in their developmental precursors 44. If a transit amplifying cell sustains a mutation that confers indefinite self-renewal, however, and in particular if this cell already has additional mutations required for malignancy by virtue of descent from a tissue stem cell that contains these mutations, then malignancy can arise from a transit amplifying cell. In fact, since transit amplifying cells are both greater in number and proliferate more frequently than stem cells, this may be a common route to malignancy.

Most myeloid leukaemias are thought to arise directly from haematopoietic stem cells. There is evidence, however, that the M3 acute promyelocytic subtype arises from a transit amplifying cell that has acquired the capacity for self-renewal as a consequence of expression of the PML/RARα fusion protein 5. When transduced with a leukaemogenic MLL fusion gene in vitro both haematopoietic stem cells and committed myeloid progenitors promptly immortalise, and transplantation leads to AML with apparently identical phenotypes in vivo 59. This demonstrates not only that malignancy can arise from transit amplifying cells, but also that cancer stem cells do not necessarily correspond to tissue stem cells.

Similar studies with mammary epithelium would be most informative, but unfortunately promotors selectively expressed in stem cells and committed progenitors have yet to be identified; in lieu of these, most investigators have employed the MMTV promotor, which is expressed in a diverse array of cell types. Nevertheless, there is preliminary evidence that different oncogenes target different levels of the differentiation hierarchy. Tumours induced by expression of Wnt (and downstream members of this pathway) express markers of both luminal and basal/myoepithelial lineages 60. This is not seen with other oncogenes such as Neu and H-Ras, and suggests origin from a pluripotent precursor, presumably a mammary epithelial stem cell. In support of this hypothesis the number of CD24⁺CD29^hi stem cell candidates was increased in MMTV-wnt-1 transgenic mice but not in MMTV-neu mice 19. The observation that tumours initiated by carcinogen treatment of rats contain both luminal and basal/myoepithelial cells 61 is also consistent with a stem cell origin for these tumours.

There is an additional route via which mutations arising in transit amplifying cells can persist within the tissue and ultimately lead to malignancy. If dedifferentiation of transit amplifying cells to stem cells is a relatively frequent event, then mutations that occur in transit amplifying cells can be 'fixed'. In fact this mechanism may have particular relevance to mammary carcinogenesis since a population of persistent mammary epithelial cells has been identified that derive from differentiated cells produced during pregnancy, that survive the widespread apoptosis concurrent with post-lactational involution, that have the capacity for self-renewal on serial transplantation and that can differentiate to alveolar cells in subsequent pregnancies (see above) 3940. These parity-induced mammary epithelial cells are the cell of origin for a substantial proportion of tumours in transgenic mice expressing the unactivated Her2/neu (ErbB2) oncogene under the MMTV promotor 62, although this is not seen with the wnt1 oncogene.

Although we have focused on genetic alterations in breast epithelial elements, there is increasing evidence that stromal cells are not merely innocent bystanders in tumourigenesis 6364. The breast stroma accounts for 80% of the breast volume and comprises fibroblasts, adipocytes and endothelial cells. It has been demonstrated that tumour-associated stromal cells promote the development of epithelial malignancy in vitro and in vivo: premalignant mammary epithelial cells adopt a malignant growth pattern when cultured with tumour-derived stromal cells 65 or when transplanted to transplanted to irradiated mammary fat pads 66, but not when cultured in the presence of normal mammary stromal cells or transplanted to nonirradiated fat pads. Administration of the carcinogen NMU to mammary fat pad stromal cells is sufficient to induce tumour formation following repopulation with normal mammary epithelial cells 67.

The persistence of tumour promotion by breast cancer-associated stromal cells following passage in culture 65 suggests hereditable changes, and clonal genetic alterations 68697071 and epigenetic alterations 72 have been described. For an individual tumour the alterations observed in stromal cells differ from those in malignant epithelial cells, arguing against epithelial–mesenchymal transition. A genome-wide loss of the heterozygosity scan in stromal cells from a number of breast tumours revealed hotspots that were more frequently mutated 71. This indicates that selective pressures operate during the expansion of tumour-associated stromal cells and that alterations do not merely reflect genetic drift. The precise mechanisms mediating tumour promotion remain unclear; candidates include alterations in expression of growth factors and matrix remodelling enzymes, the recruitment of inflammatory cells and the elaboration of viral oncoproteins 63.

The efficacy of chemotherapy regimens, particularly in the neoadjuvant setting, is usually gauged by the proportion of tumour cells killed. This is reflected clinically by a reduction in size of the tumour. Since cancer stem cells represent a small proportion of the cells within breast tumours, a particular treatment modality could in principle kill a large proportion of differentiated cells while leaving tumour stem cells intact. In fact, CD34⁺CD38^- cells isolated from patients with AML are more resistant to daunorubicin than more differentiated progenitors 73, suggesting that cancer stem cells may have relative resistance to chemotherapy regimens. The ability to accurately quantify the proportion of cancer stem cells within a tissue sample will permit the identification of treatment regimens that efficiently eliminate cancer stem cells. It will also be instructive to determine whether the proportion of stem versus transit amplifying cells in a given tumour has implications for prognosis.

A related issue is bone marrow micrometastasis. Cytokeratin-expressing cells can be detected in the bone marrow of around one-third of patients with breast cancer and are an independent predictor of poor outcome 7475. Importantly these cells are also found in lymph-node-negative patients, and it has been proposed that this may account for the relatively high rate of systemic recurrence in this group of patients. Conversely, a substantial proportion of patients with bone marrow micrometastasis do not develop tumour recurrence. Cytokeratin-positive cells within the bone marrow may comprise both cancer stem cells and committed progenitors with limited proliferative potential. This may have prognostic significance with patients that have a larger proportion of breast cancer stem cells more likely to suffer recurrence. The ability to distinguish cancer stem cells from other tumour cells within the bone marrow may also have implications for the selection of adjuvant therapies.

With the emergence of the cancer stem cell paradigm has come the realisation that, since only a small proportion of tumour cells have the capacity to self-renew and propagate the tumour, it is this small proportion of tumour cells rather than transit amplifying cells or differentiated cells that must be targeted to achieve long-term cure.

The spectacular success of the tyrosine kinase inhibitor imatinib in chronic myeloid leukaemia provides a prototype for targeted therapeutics. While these therapeutics can induce dramatic tumour regression even in established malignancy, however, they do not in general lead to a cure but rather control the disease with recurrence on cessation. They are also vulnerable to the emergence of resistant clones with prolonged therapy 96. This most probably reflects genetic instability in established malignancy and a large number of cells that are targets for mutation. These observations raise fundamental questions about our approach to the control of malignancy. The emerging paradigm for multistep carcinogenesis in epithelial malignancy is a long period of clonal evolution in histologically normal tissue 58, for which a substantial proportion of genetic alterations may occur during development 44. This leads to histological atypia, dysplasia, pre-invasive disease and, finally, malignancy with the capacity for invasion and metastasis.

Nonsurgical treatment strategies for breast cancer are at present reserved for the treatment of established malignancy, but from first principles any stage in the evolution of malignancy can be targeted. In fact, therapies targeting pathways that govern self-renewal of preinvasive disease such as ductal carcinoma in situ, or even mutant stem cells in histologically normal tissue, may be more effective since the mutation rate is lower at this stage, since there is less likely to be coexistent mutations that confer resistance and since there is a smaller number of target cells for the emergence of resistance. Furthermore, metastasis may occur early in the history of breast cancer – frequently before the tumour becomes apparent clinically or on screening. Since it is probable that a large proportion of otherwise healthy women have cells containing these genetic alterations, this approach could only succeed if a large proportion of apparently healthy women are treated – necessitating a paradigm shift in our approach to breast cancer, and to epithelial malignancy in general.

We can envisage the development of specific inhibitors targeting the self-renewal of mammary epithelial stem cells. Since the mammary epithelium is unimportant cosmetically, such compounds could be administered to ablate the mammary epithelium in women carrying hereditary breast cancer susceptibility genes such as BRCA1/BRCA2. For women not carrying breast cancer susceptibility genes, however, such an approach may be unacceptable, and therefore in the longer term it is desirable to develop methods permitting repopulation with stem cells derived either from tissue stem cells expanded in vitro or differentiated from embryonic stem cells. Transplanted stem cells could be modified to confer a selective advantage over endogenous stem cells, as has been reported for haematopoietic stem cells 97. Transplanted cells could also be modified to reduce susceptibility to malignancy; for example, transgenic mice with extra copies of the p53 98 and INK4A 99 tumour suppressor genes have a substantially decreased incidence of malignancy in the absence of apparent adverse consequences. Although repopulation of mammary stem cells remains a rather distant prospect, intraluminal injection of donor spermatocyte stem cells into seminiferous tubules repopulates the testes of sterile mice leading to normal spermatogenesis 100, so there is no reason to suppose that it should not be technically feasible.