Next-generation sequencing technologies have provided us with a precise description of the mutational burden of cancers, making it possible to identify targetable oncogene addictions. However, the emergence of resistant clones is an inevitable limitation of therapies targeting these addictions. Alternative approaches to cancer treatment are therefore required. We propose here a novel approach, based on the notion of conflicting signals and on a phenotypic description of cancer cells. “Phenotype” is an inherently complex notion that we describe in the conceptual framework of the epigenetic landscape, with a view to bridging the gap between theory and practice at the patient's bedside. By passing from theory to the description of several examples, we will illustrate how this approach can facilitate data analysis and the design of new strategies for cancer treatment. Cancer Res; 76(23); 6768–73. ©2016 AACR.
Over the last 10 years, next-generation sequencing technologies have unraveled the mutational landscape of most tumors. Several success stories relating to the use of treatments targeting specific mutations have been reported, concerning for instance IDH2 in leukemia (1), BRAF in melanoma (2), and HER2 and EGFR in breast and lung cancer (3, 4). However, the clinical efficacy of such approaches remains limited. Indeed, in most cases, the response to targeted therapy is followed by relapses due to the emergence of subclones of tumor cells resistant to the treatment or to the activation of alternative pathways (5). The postgenomic era of cancer research should, therefore, make use of new conceptual frameworks, such as ecology-inspired approaches (6), to improve patient treatment. In this perspective, we adopt a phenotype-centered view of cancer cells, combining inputs relating to gene mutation status, gene expression regulation, and interactions with the microenvironment. The description of cancer cells in this way may favor new proposals for cancer treatment based on the concept of “conflicting signals.” To this end, we first need to propose an operational definition of cell phenotype.
Definition of Cell Phenotype
The cell phenotype results from a set of complex biological traits integrating inputs from the genetic material of the cell and from environmental cues. It can be modeled as a position in an N-dimensional space, in which each dimension is an axis between two archetypes for a given task. The archetypes are theoretical cells occupying the extreme positions on a given dimension of the phenotype. For example, if we consider the energetic metabolism of the cell as a phenotypic dimension, we can define the following two archetypes: a cell using only oxidative phosphorylation to produce ATP and a cell producing ATP exclusively by glycolysis. Any real cell from the N-dimensional phenotypic space would project on the segment between the two extreme archetypes of this axis (Fig. 1A). The complexity of the cell phenotype is sustained by a multitude of phenotypic axes, but the axes to be considered can be reduced to a small number of relevant axes amenable to experimental investigation (7, 8).
In this perspective, we will highlight two particular axes that appear to be determinant for the acquisition of the hallmarks of cancer. The first is the balance between proliferation and quiescence. Its relevance was highlighted by a recent study describing a tRNA pool signature efficiently discriminating between proliferative and quiescent cells in a multitude of models, including primary cells and cells in which proliferation or senescence was induced experimentally (9). The second axis relates to energetic metabolism, which has recently returned to the fore some 80 years after the work of Otto Warburg. Cancer cells have higher rates of glucose capture and breakdown via glycolysis than normal cells, enabling them to generate the building blocks of anabolic processes and to buffer oxidative stress (10). As a result, cancer cells are dependent on glycolysis (aerobic or anaerobic) rather than oxidative metabolism, the preferred process for generating energy in nonproliferating cells. The balance between glycolytic and oxidative phosphorylation metabolism can, therefore, be seen as an axis of particular relevance for the description of cancer cell phenotype. However, care should be taken to avoid a dichotomist vision of cellular metabolism, because proliferating cells can use aerobic glycolysis and oxidative phosphorylation simultaneously (11).
The coordinates of a given cell in this N-dimensional space define its phenotype, encoded by gene network constraints and stabilized by environmental selective pressure. In addition to proliferation state and energetic metabolism, other axes could be taken into account, such as oxidoreductive balance, the balance between anabolic and catabolic processes, or the intensity of autophagy.
Cell Phenotypes as Basins of Attraction in the Epigenetic Landscape
The position of the cell phenotype in the N-dimensional phenotypic space described above should be seen as a still from a movie rather than as a permanent characteristic, because cell phenotype is not stable. This complexity may be better understood by considering the notion of the “epigenetic landscape”, initially proposed by Waddington (12) and further refined by Huang (for more details, see ref. 13). This concept is based on a topographical description of different “basins” (stabilized phenotypes) separated by “chains of mountains.” For the phenotypic description of cancer cells proposed above, we could define a main basin of attraction corresponding to rapidly proliferating cells predominantly using glycolysis, another corresponding to a quiescent phenotype predominantly using oxidative phosphorylation, and a third basin of attraction, including differentiated phenotypes (quiescent cells with predominantly glycolytic metabolism; Fig. 1B). Of note, one could further distinguish sub-basins among these large basins. For instance, quiescence is not a unique state but rather a collection of heterogeneous states (14).
The representation of cell phenotypes as a multi-attractor landscape summarizes some essential features. Even though many molecules and functions, like oxidative phosphorylation or autophagy, are active in all cellular states, albeit with different intensities, the cellular states (quiescence, proliferation) are disjoint and mutually exclusive. Mutated oncogenes or tumor suppressor genes strongly influence the genetic networks stabilizing these phenotypes, by controlling the traits essential for a cancer phenotype, as defined above. For instance, activation of the mTORC1 complex is one of the most frequent genetic hits observed in cancer (directly or indirectly; ref. 15), whose activation results in cell proliferation and activation of the glycolytic pathway.
Transformed cells are, thus, trapped in the proliferation basin of attraction. This does not mean that their phenotype is stable and, indeed, cancer cells remain subject to stochastic phenotypic oscillations that may result in their migration to another basin, or sub-basin of attraction. The stochastic phenotypic oscillations of cancer cells have been well described in a cell line model of human breast cancer, including at least three phenotypically different subpopulations. After cell sorting to separate the different subpopulations, reconstitution of the initial phenotypic heterogeneity was systematically observed, attesting to the occurrence of stochastic phenotypic transitions (16). Another example of phenotypic oscillations of cancer cells has been proposed in cell lines exposed to various cytotoxic agents, where a reversible multidrug resistance phenotype was observed in rare cells (17).
When phenotypic oscillations are strong enough to enable the cell to change basin of attraction, the cell crosses the crest between the two basins and enters an unstable, indeterminate state. This indetermination might be resolved by three moves: (i) a return to the initial basin; (ii) a change of basin of attraction; or (iii) a fall in the senescence or apoptosis abyss, as represented in Fig. 1B. To take into account the latter situation, we need to refine the classical landscape description of phenotypes, which only considers basins of attraction. We thus propose here to add to this landscape an unstable stressed state, located on a crest, which is not supposed to be occupied for long durations. Excessive or prolonged stress, which can be defined as an excessive variation of the environment of the cell (in terms of oxygen concentration, nutrient availability, etc.), will thus lead to cell death if the cell is unable to adapt sufficiently (18).
The notion of “prolonged stress” implies that the cell fate depends on the kinetics of a stress signal, that is, on the rate at which the crest line is crossed, in Fig. 1B. Indeed, most cellular stresses trigger a biphasic response, with an initial prosurvival output, and lethal consequences in cases of persistent activation. For example, short pulses of p53 activation activate the prosurvival effects of DNA repair, whereas prolonged activation of p53 drives cells into senescence (19, 20). This apparently paradoxical observation can be understood by taking into account, as illustrated in Fig. 2, that a single cause may trigger two opposite effects with different kinetics (further discussed in ref. 21).
Within this conceptual framework, we will now consider how stress and conflicting signals could be exploited to reverse the cancer phenotype.
In our simplified model of the epigenetic landscape with three basins of attraction (differentiation, quiescence, and proliferation) and a nonattracting, stressed state located on crest lines, cancer occurs when the cells remain trapped in the proliferative attraction basin. Cancer treatments aim to empty this basin, by killing proliferating cells with chemotherapy or by inducing cells to move from the proliferation basin to the differentiation basin, as has been achieved with all-trans retinoic acid or arsenic trioxide to treat promyelocytic leukemia (22). Cancer cells could also be eliminated through the use of conflicting signals promoting simultaneously quiescence and proliferation.
Such unphysiologic signals can be mediated by the combination of gene mutations, extracellular signals (growth factors, cytokines, nutrients), or pharmacologic intervention. The notion of “conflicting signals” is not new (see e.g., refs. 23, 24) but is rarely used, despite its utility as a conceptual tool for interpreting complex situations, as described below. The logical essence of conflicting signals is summarized in Fig. 3. Two signals inducing incompatible phenotypical changes are fine when delivered separately, but not when delivered simultaneously.
Distinguishing conflicting signals and synthetic lethality
We would like to further clarify the notion of conflicting signals and their relationships with the more popular and related concept of synthetic lethality.
In the literature, synthetic lethality refers to a situation in which the simultaneous inactivation of two pathways is lethal. This corresponds to a situation of redundancy, in which a function essential for cell survival may be mediated by the products of two genes. Logically, cancer cells in which one of these genes was invalidated during oncogenesis will be highly sensitive to agents inhibiting the other gene. This approach has been employed in the use of PARP inhibitors to treat breast cancers with mutations of the breast cancer 1 (BRCA1) gene (25). In cancers bearing BRCA1 mutations, only PARP can repair DNA damage. Such tumors are, thus, highly sensitive to PARP inhibition, which results in mitotic catastrophe.
The concept of nononcogenic addiction (26) is tightly linked to this situation. This concept describes how oncogenic transformation creates a high level of addiction to diverse pathways that are essential for cell survival because of oncogene activity. For example, c-myc overexpression is associated with an exquisite dependency on spliceosome activity (27) and SUMOylation pathways (28), which are good candidates for therapeutic interventions.
In what follows, we reserve the term synthetic lethality only to such cases of concomitant inactivation of genes leading to cell death and the term conflicting signals for the cases of concomitant activation of genes leading to cell death. This point being clarified, let us give more examples of how such conflicting signals may be used against tumor cells. Table 1 summarizes the experimentally demonstrated conflicting signals detailed in this perspective, with a highlight of the conflicting signals involving pathways involved in quiescence and pathways required for cell proliferation.
|Signal 1 .||Signal 2 .||Ref. .|
|Nuclear FoxO||Proliferation signals||31|
|Signal 1 .||Signal 2 .||Ref. .|
|Nuclear FoxO||Proliferation signals||31|
NOTE: For each experimental example, conflicting signals (signal 1 and 2) are cited in two separated columns. Interestingly, there are numerous examples of negative feedbacks between genes in column 1 versus 2, whereas the genes in the same columns are prone to form positive feedbacks.
AMPK and mTORC1
We have recently reported an example of conflicting signals due to the simultaneous activation of mTORC1 (promoting cell proliferation, anabolic metabolism, and glycolysis) and AMP-activated kinase (AMPK; promoting cell quiescence, catabolic metabolism, and oxidative phosphorylation). Acute myeloid leukemia cells display stronger mTORC1 activation than their normal counterparts. This renders these cells highly sensitive to AMPK-activating compounds, because the activation of both pathways converges on the eIF2A–ATF4 stress pathway, leading to cell death (29). This is a first example illustrating how conflicting signals might be useful to uncover unsuspected therapeutic window for cancer treatment.
Nuclear FoxO and PI3K/AKT/mTORC1
Another example of the inability of tumor cells to withstand strong quiescence-inducing signals is provided by the FoxO family of transcription factors, which play a key role in inducing and maintaining quiescence (30). In addition to its major effect in inducing quiescence, it has been suggested that FoxO may also, paradoxically, act as an apoptosis inducer (for a review, see ref. 31). However, FoxO is excluded from the nucleus in primary proliferating cells (following the activation of PI3K/AKT) and cannot, therefore, trigger the transcription of genes encoding apoptotic factors in these cells. FoxO-induced apoptosis has only been described in cell lines overexpressing constitutively nuclear mutants of FoxO (32, 33). Molecular analysis would focus on the detailed description of the transcriptional program for cell death driven by FoxO expression in such artificial models. The interpretation at the phenotype level is more parsimonious: The observed cell death results from a conflict between signals for quiescence and proliferation. Interestingly, the apoptosis of several cells lines in vitro may be induced by resveratrol, which has been shown to induce quiescence by activating FoxO (34) and AMPK (35), two genes important for the quiescent state (Table 1).
p53 and NF-κB
The notion of conflicting signals is particularly helpful for understanding paradoxical effects of NF-κB, depending of the context. “Typical” activators of NF-κB (like inflammatory cytokines) exert prosurvival effects, whereas “atypical” activators of NF-κB (UV light and some chemotherapeutic compounds) are proapoptotic. Things become clear when one notes that "atypical" NF-κB activators systematically activate p53 as well. What matters in the "context" is whether p53 is simultaneously activated or not. The simultaneous activation of p53 and NF-κB is unphysiologic. p53 and NF-κB inhibit each other's ability to stimulate gene expression (36). They have opposite effects on cellular metabolism and cell fate: p53 promotes oxidative metabolism and inhibits the cell cycle, whereas NF-κB promotes glycolysis and survival (37). Their conflict is fundamental, as inhibition or loss of NF-κB activity abrogates p53-induced apoptosis, indicating that NF-κB is essential in p53-mediated cell death (38).
The conflicting signals generated by NF-κB and TP53 may account for the oncogene-induced senescence observed in the case of the Ras oncogene. Indeed, Ras overexpression triggers a biphasic response. The initial part of the response, involving mTORC1 and NF-κB, favors the glycolytic pathway. However, after some time, prolonged Ras activation also leads to the activation of p53 (39), which enters into conflict with NF-κB, particularly as regards the regulation of energetic metabolism. These observations illustrate the way in which natural conflicting signals can act as an efficient safety system, preventing excessive proliferation due to oncogene activation.
IFNβ on proliferating cells
Another example of conflicting signals is provided by IFNβ in proliferating cells. The detection of viral nucleic acids by the cell triggers the induction of IFNβ and, as an immediate consequence, IFNα induction. IFNβ/α has major pleiotropic effects, including p53 activation and transcription arrest (40), which play a key role in preventing viral replication. IFNβ is also specifically toxic to proliferating cells (for a review, see ref. 41), but not to quiescent cells. An attractive explanation for this differential sensitivity is that the inhibition of the glycolytic pathway by IFNβ and p53 creates a conflicting signal only in cycling cells. This conflict is exacerbated if IFNα/β is added to cells in which the PI3K/mTORC1 pathway is strongly stimulated. This may explain the surprising, misleading description that IFNα-induced apoptosis in tumor cells "was mediated through the PI3K/mTORC1 signaling pathway" (42). Once again, the conflicting signal concept allows a simpler interpretation: mTORC1 activation of glycolytic metabolism primes the lethal effects of IFNα/β–p53, due to a conflict in the regulation of energetic metabolism. Similarly, cells displaying mTORC1 overactivation due to TSC2 deletion have been shown to be particularly sensitive to glucose deficiency, which triggers the death of these cells (43). This situation provides another example of the effects of conflicting signals, as mTORC1 activation transduces the signal for a nutrient-rich environment (44), conflicting with glucose deficiency and creating an unsolvable energetic problem leading to cell death.
In this perspective, we have revisited the notion of cell phenotype thanks to the notion of epigenetic landscape (with disjoint basins of attraction and an unstable state) and showed how the notion of conflicting signals, in this phenotypic frame, could complement targeted therapies, which are genotype centered.
Most modern cancer treatments are focused on the genotype of tumor cells, for instance, in the oncogenic addiction model, according to which cancer cells are dependent on signals from their oncogenic pathways to survive. However, the strong selection pressure exerted on cancer cells by oncogene-targeting agents results in the inevitable emergence of subclones with resistance mutations conferring greater fitness in the presence of the targeted therapeutic agent.
Thus, genotype-centered views have led to success stories, followed by disappointments. In this perspective, we have proposed a phenotype-centered view of cancer cells. Phenotypes are multidimensional and cannot be reduced to a couple of regulation pathways. However, some genes occupy key positions at the nodes of the gene networks underlying phenotypes and represent interesting ways to create conflicting signals.
Combinations of therapies based on the addiction concept, with therapies based on the conflicting signals concept, may improve cancer treatment and might even lead to cure. Indeed, the resistant subclones that emerge after targeted therapies will probably be highly sensitive to quiescence inducers that are catastrophic for the cells that proliferate. We propose the testable hypothesis that quiescence inducers should be effective agents for provoking lethal conflicting signals in proliferating cells, without hampering nonproliferating cells. Activators of FoXO or AMPK, such as resveratrol or GSK621, or improved versions of these drugs, have interesting properties of this type and could be combined with targeted therapies. We expect that the identification of additional quiescence inducers would be of great interest for allowing development of the strategy presented here.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
Conception and design: P. Sujobert, A. Trautmann
Writing, review, and/or revision of the manuscript: P. Sujobert, A. Trautmann
Study supervision: P. Sujobert, A. Trautmann
We would like to thank J.P. Abastado, N. Bercovici, G. Bismuth, C. Charvet, J. Delon, L. Laplane, and E. Solary for their helpful comments on this manuscript, Naïs Coq for the artwork of Fig. 1B, and Vincent Feuillet for that of Fig. 2.