News and Views

Nature Biotechnology 27, 44 - 46 (2009)

Looking ahead in cancer stem cell research

John E Dick1

  1. John E. Dick is in the Division of Cell and Molecular Biology, University Health Network, and Department of Molecular Genetics, University of Toronto, Toronto, Ontario M5G 1L7, Canada. e-mail:

The history of the stem cell field offers pointers for future research on cancer stem cells.

The idea that many cancers are organized as hierarchies sustained by cancer stem cells (CSCs) at their apex has generated almost breathless excitement in many quarters of the cancer research community. Hardly a week goes by without a CSC spotting in yet another cancer or an assertion that CSCs are responsible for a particular treatment failure. In the rush to jump on the CSC bandwagon, many have lost sight of the historical development of this field and the fundamental stem cell principles that should underpin experimental strategies. A recent report in Nature from Quintana et al.1 provides a jarring caution to uncritical acceptance of the CSC hypothesis as universal for all tumors. These authors demonstrate that the original observation of the rarity of CSCs in melanoma (on the order of 1 in 105 cells) was an artifact of the xenotransplant tumor-initiation assay, caused by residual innate immunity of the recipient mouse. When a more profoundly immune-deficient recipient was used, 15–25% of the tumor cells exhibited CSC activity, a frequency that argues against any meaningful hierarchical organization in human melanoma and raises the question of whether CSCs exist for this tumor type.

To place the report by Quintana et al.1 into context, one must understand why the CSC hypothesis was first proposed many decades ago2. Since the beginning of the era inaugurated by the great experimental pathologists, microscopic examination has revealed that many tumors, be they liquid or solid, exhibit morphologic heterogeneity. Tumors are also heterogeneous functionally, as demonstrated most dramatically by studies in which murine or human tumors were re-transplanted into syngeneic or xenogeneic immune-compromised recipients. Remarkably, such experiments have also included human autotransplants. The collective conclusion from these studies was that tumor re-initiation is variable and often rare, requiring 103 to 107 cells.

These observations of cellular heterogeneity, together with the finding that malignant teratocarcinoma tumors contain highly tumorigenic cells as well as nontumorigenic cells that are differentiated progeny of the tumorigenic cells, led to the idea that tumors are caricatures of normal development with a hierarchical organization. In my mind, the strongest evidence of tumor heterogeneity to date has come from cellular radiolabeling with tritiated thymidine in human cancer patients, which enabled precise measurements not only of proliferation but also of cellular life span and hierarchical relationships. For leukemia patients, these studies established that the vast majority of leukemic blasts were post-mitotic and continuously replenished from a small fraction of cells (~5%) that cycled rapidly. Interestingly, they also revealed a rare cellular fraction that remained dormant for weeks to months before beginning to cycle.

Two models have been put forward to explain tumor heterogeneity. Although a detailed description of these models is beyond this commentary, they have been widely reviewed and their basic elements are shown in Figure 1(Ref. 2). The stochastic model argues that tumors are biologically homogeneous. Any functional heterogeneity is due to random or stochastic influences that alter the behavior of individual cells in the tumor. These influences can be intrinsic (e.g., levels of transcription factors, signaling pathways) or extrinsic (e.g., host factors, microenvironment, immune response). By contrast, the hierarchy model argues that tumors are caricatures of normal tissues such as skin, colon and blood, many of which are cellular hierarchies maintained by stem cells. In this model, cellular heterogeneity is due to the hierarchy that emanates from a class of cells at the apex, biologically distinct from the rest of hierarchy, that possesses self-renewal capacity and can be thought of as CSCs.

Figure 1: Models of tumor heterogeneity.

Figure 1 : Models of tumor heterogeneity.

Tumors are composed of phenotypically and functionally heterogeneous cells. There are two theories as to how this heterogeneity arises. According to the stochastic model, tumor cells are biologically equivalent but their behavior is influenced by intrinsic and extrinsic factors and is therefore both variable and unpredictable. Thus, tumor-initiating activity cannot be enriched by sorting cells based on intrinsic characteristics. In contrast, the hierarchy model postulates the existence of biologically distinct classes of cells with differing functional abilities and behavior. Only a subset of cells has the ability to initiate tumor growth; these cancer stem cells possess self-renewal and give rise to nontumorigenic progeny that make up the bulk of the tumor. This model predicts that tumor-initiating cells can be identified and purified from the bulk nontumorigenic population based on intrinsic characteristics. This figure and figure legend were originally published in Ref. 2.

Full size image (67 KB)

Although the stochastic model can accommodate the existence of 'functional CSCs'—cells that behave as CSCs—the essential difference from the hierarchy model is that every tumor cell is thought to have the potential to behave like a CSC given the right influences. If the developmental rigidity implicit in the hierarchy model is correct, it should be possible to fractionate a tumor into fractions devoid of CSCs and other fractions highly enriched for CSCs. If the stochastic model is correct, a tumor cannot be cleanly fractionated because all tumor cells have the same potential to initiate tumors. The challenge is devising a tumor-initiation assay that can correctly score the potential of a cell to be a CSC; if some cells revealed their intrinsic CSC potential only under special microenvironmental conditions not provided by the assay, they would be falsely scored as negative.

For human cancers, investigation of CSCs requires the use of xenotransplant systems. These have evolved greatly beyond the original nude mouse model and now include recipients with more profound defects in adaptive immunity (deficient in T cells and B cells) and, more recently, with defects in innate immunity (deficient in natural killer cells and macrophages). Quintana et al.1 first transplanted human high-grade melanoma cells (a highly aggressive tumor) into nonobese diabetic/severe combined immunodeficient (NOD/SCID) mice and found that this is not a reliable assay as only four of seven samples generated tumors within 8 weeks. Using a limiting dilution approach, they determined that the frequency of the melanoma cancer–initiating cell (MC-IC) was 1 in 106 cells. However, increasing the time of observation to 32 weeks increased the number of tumors scored, raising the MC-IC frequency to 1 in 105 cells. These findings seemed in keeping with recent reports of CSCs in other solid tumors3. However, as NOD/SCID mice retain innate immunity, the authors turned to a newer mouse model that is depleted in natural killer cells (NOD/SCID/Il2rg−/−). With these mice and the use of Matrigel to provide a more supportive microenvironment, the MC-IC frequency rose to between 1 in 9 and 1 in 4 cells. Clearly, innate immunity in the NOD/SCID recipients was restricting a large proportion of melanoma cells that had CSC potential.

A priori, neither the stochastic model nor the hierarchy model makes any predictions as to the frequency of CSCs: they could be frequent or rare4. However, if a tumor is functionally homogeneous (that is, every cell is tumorigenic or proliferative), the hierarchy model becomes meaningless. So, does the work of Quintana et al.1 suggest that melanoma tumors are functionally homogeneous and therefore do not follow the hierarchy model? Some may argue the fine point that even with a CSC frequency of 1 in 4, 75% of the tumor cells still lack CSC activity. And if the CSCs respond differently to cancer therapy, it would still be meaningful to purify them and investigate their unique biology. Quintana et al.1 attempted this, but no markers that could fractionate the tumor were found. However, 1 in 4 is a very high frequency, and considerable efforts would be needed to prove beyond doubt that a true hierarchy exists, including discovery of a cell-surface marker or metabolic pathway that enables fractionation followed by detailed quantitative analysis of transplanted mice.

As in other areas of cancer research, we should be guided by experimental murine cancer models, which enable syngeneic transplantation, thereby overcoming immune rejection. In a recent study of three murine breast cancer models—MMTV-Wnt-1, MMTV-neu and p53+/−—the CSC frequency was different for each model: 1 in 177, 1 in 112 and 1 in 1,090, respectively5. Importantly, this study demonstrated tumor fractionation into CSC and non-CSC fractions. Wide variation in the frequency of tumor-initiating cells has also been observed in murine leukemia models, with some lymphoid leukemia models (Eμ-myc, bcr/abl/CDKN2 and Mll-Enl) having frequencies higher than 1 in 10 cells6 and others (MOZ-TIF or Pten−/−) showing frequencies of 1 in 104 to 1 in 106 cells4. Like mouse tumors, human tumors will likely exhibit wide variation in the frequency of tumor-initiating cells, even in immune-deficient models, as has now been found for melanoma by Quintana et al.1. We should be open to the existence of human cancers that have high frequencies of tumor-initiating cells and that do not follow the hierarchy model.

How should the field move forward? Stem cell principles developed over many years, which established that the hematopoietic system is organized as a hierarchy, provide a useful guide2. The first principle is the importance of developing assays capable of reading out any tumor cell that possesses stem cell function. Thus, emphasis should be placed on developing xenotransplant models that are devoid of immunity. It is perhaps not surprising that innate immunity has such a strong effect in melanoma, as this type of tumor is known to be highly immunogenic. It is unlikely that a similar effect will be observed in brain tumors, given that the brain is an immune-privileged site. In addition, as the murine microenvironment may not always be sufficient to reveal CSC potential, efforts should be made to humanize recipient mice. In parallel, more CSC studies in murine models, like those described above, should be undertaken to guide xenotransplant experiments. Finally, two recent reports7,8 demonstrate how sophisticated genetic tools can be used to carry out lineage tracking of CSCs in murine models. Although these studies did not attempt cell fractionation, they provide clear evidence of a hierarchy in colon cancer.

The second principle is to develop tumor endpoints that will enable discrimination between CSCs (or progenitors) that retain extensive proliferative potential but have limited self-renewal capacity and CSCs that possess extensive capacity for self-renewal. CSCs are often thought of as homogeneous. However, clonal tracking studies have shown that many human leukemia stem cells can make a large graft of leukemia but then extinguish, whereas a minority of these stem cells exhibit long-term repopulation with the capacity to extensively self-renew2. Thus, although all human leukemia stem cells generate a 'tumor', they vary considerably in their self-renewal capacity. It is highly likely that solid-tumor CSCs are similarly heterogeneous. Accordingly, CSC assays must extend long enough, or through enough rounds of serial transplantation, to distinguish CSCs with different capacities for self-renewal. Such tests of heterogeneity must be carried out on single CSCs using single-cell transplants or lentiviral vector–mediated clonal tracking. Ultimately, it is the CSCs with extensive self-renewal capacity that will be the most crucial to understand and to link to disease relapse.


Finally, the most significant question is the relevance of the hierarchy and stochastic models to human patients. The most important prediction to come from the CSC field is that CSCs are inherently resistant to a variety of chemo- and radiotherapies. Recent clinical studies have begun to test the response to chemotherapy of biomarkers associated with CSCs (distinct cell surface phenotypes and in vitro tumorigenesis assays). Notably, Li et al.9 demonstrated that, after undergoing neoadjuvant chemotherapy, patients showed significant increases in cells with properties that xenotransplant studies had associated with CSC activity (CD44+CD24 phenotype and tumorigenic sphere formation). The non-CSCs were killed by the chemotherapy, resulting in tumor response, but the CSC-like fraction was relatively resistant and thus constituted a higher proportion of the remaining tumor. More clinical studies are needed to assess how responses to therapy and other clinical disease parameters correlate with CSC biomarkers, as are new imaging technologies for CSCs. Additionally, studies on large numbers of human tumors with optimized CSC assays could determine whether CSC properties (e.g., number, gene expression signatures, pathway biomarkers) correlate better than bulk tumor properties with clinical outcomes. Ultimately, the relevance of the CSC model will be determined by clinical data.

In 1937, Furth and Kahn10 established that a single cell from a murine cancer cell line rather than a transmissible agent could initiate a cancer graft when transplanted into mice. That a similar experimental approach can still shed new light on cancer biology, some 70 years later, underscores the importance of understanding our scientific history and the truth of the adage that those who do not know their history are condemned to repeat it.



  1. Quintana, E. et al. Nature 456, 593–598 (2008). | Article | PubMed | ChemPort |
  2. Dick, J.E. Blood 112, 4793–4807 (2008). | Article | PubMed | ChemPort |
  3. Dalerba, P., Cho, R.W. & Clarke, M.F. Annu. Rev. Med. 58, 267–284 (2007). | Article | PubMed | ISI | ChemPort |
  4. Kennedy, J.A. et al. Science 318, 1722; author reply 1722 (2007). | Article | PubMed | ChemPort |
  5. Vaillant, F. et al. Cancer Res. 68, 7711–7717 (2008). | Article | PubMed | ChemPort |
  6. Kelly, P.N. et al. Science 317, 337 (2007). | Article | PubMed | ISI | ChemPort |
  7. Barker, N. et al. Nature advance online publication, doi:doi:10.1038/nature07602 (17 December 2008). | Article |
  8. Zhu, L. et al. Nature advance online publication, doi:doi:10.1038/nature07589 (17 December 2008). | Article |
  9. Li, X. et al. J. Natl. Cancer Inst. 100, 672–679 (2008). | Article | PubMed | ChemPort |
  10. Furth, J. & Kahn, M. Amer. J. Cancer 31, 276–282 (1937).


These links to content published by NPG are automatically generated.