Origins of Metastatic Traits

Abstract

How cancer cells acquire the competence to colonize distant organs remains a central question in cancer biology. Tumors can release large numbers of cancer cells into the circulation, but only a small proportion of these cells survive on infiltrating distant organs and even fewer form clinically meaningful metastases. During the past decade, many predictive gene signatures and specific mediators of metastasis have been identified, yet how cancer cells acquire these traits has remained obscure. Recent experimental work and high-resolution sequencing of human tissues have started to reveal the molecular and tumor evolutionary principles that underlie the emergence of metastatic traits.

The molecular and tumor evolutionary basis of metastasis is a long-standing problem that particularly in the past few years has started to move toward a resolution. Important elements of the biologically complex metastatic cascade (Talmadge and Fidler, 2010; Valastyan and Weinberg, 2011), its dynamic and kinetic diversity across different cancer types and patient populations (Chiang and Massagué, 2008), and the strong stochastic nature at the cellular level (Kienast et al., 2010; Luzzi et al., 1998) are coming to light at a brisk tempo. Work during the past decade has demonstrated that the problem of metastasis can be tackled experimentally at the molecular level. Progress in experimental models and clinical research have led to the identification of genes that mediate various steps of the metastatic cascade in different tumor types and target organs (Ell and Kang, 2012; Guise, 2009; Nguyen et al., 2009a; Talmadge and Fidler, 2010; Valastyan and Weinberg, 2011). Increased genome sequencing output has facilitated the analysis of clonal relationships between primary tumors and secondary lesions in clinical samples (Campbell et al., 2010; Gerlinger et al., 2012; Yachida et al., 2010). Taken together, these developments have revealed a set of principles that illuminate fundamental questions on the origins of metastatic traits. We resort to specific examples in order to review here the nature and implications of these advances.

Early Ideas

The classical view of tumor progression, based on the clonal evolutionary theory of cancer (Nowell, 1976), postulated that the metastatic ability is conferred by random mutations in primary tumor cells that remain rare until clonally expanded and selected at secondary organ sites (Fidler and Kripke, 1977). This Darwinian view for metastatic progression was intuitively appealing and supported by evidence from cancer cell transplantation experiments in mice (Fidler and Kripke, 1977; Kripke et al., 1978). Clinical observations were also consistent with the view that metastasis is a rare achievement for cells in a primary tumor, although one that follows predictable patterns suggestive of specific mutations being responsible for its development (Hess et al., 2006). Identifying these driver mutations, however, remained an elusive goal.

Indications that the clonal selection model alone was not sufficient to explain the development of metastatic traits emerged with the advent of genome-wide transcriptomic techniques and their application to tumor samples. It became evident that the likelihood of metastasis in many types of cancer could be predicted from the overall gene expression profile of primary tumors (van ’t Veer et al., 2002; van de Vijver et al., 2002), suggesting that large segments of the cancer cell population contained traits that predisposed these tumors to metastasis. This was hard to reconcile with the idea of rare mutations as the main cause of metastatic progression. It was proposed that metastases may instead be an outcome of the same oncogenic forces that drive the emergence of primary tumors (Bernards and Weinberg, 2002). Experimental evidence for primary oncogenic mutations as drivers of metastasis had precedent (Staller et al., 2003). This view, however, had its problems too, as cancer so clearly was an evolutionary phenomenon (Greaves and Maley, 2012). It was hard to imagine how metastasis could not be the end result of strong selection imposed by different microenvironments. A complementary view invoked the “seed-and-soil” hypothesis that was first enunciated by Paget in the 19^th century and, in today’s language, stated that cancer cells may seed metastasis so long as they reach a compatible tissue microenvironment (Fidler, 2003). Compelling evidence has since accumulated that partially supports each of these early ideas.

Metastasis Steps

The formation of clinically detectable metastasis is the end result of a series of stochastic events that first allow cancer cells to disperse and survive in distant sites and later to grow as secondary tumors (Figure 1). This process comprises steps of cancer cell migration, local invasion, entry into the circulation, arrest at secondary sites, extravasation, and colonization. The use of “colonization” as a single term that belies the highly complex and demanding process that awaits infiltrated cancer cells in distant organs. Colonization can be parsed into steps of cancer cell survival upon entry into the tissue, formation of micrometastasis, adoption of latency states that can last up to decades, reactivation of growth in the latent micrometastases, aggressive overtaking of the host tissue, recirculation, and formation of tertiary lesions in the same or different organs (Figure 1).

Figure 1. Metastasis Steps and Bottlenecks.

The sequence of metastasis steps starts with the dissemination of cancer cells from the primary tumor and ends in the formation of clinically detectable macrometastases. Whereas dissemination is not severely rate limiting, only a small fraction of disseminated cells form micrometastases, and only a small minority of micrometastases become macrometastases. A cancer cell may reach distant organ capillaries within seconds of departing from a primary tumor and extravasate into the parenchyma within hours, but it may remain latent for decades. Progression of metastasis through each of these bottlenecks requires several specialized functions including cancer cell-autonomous functions and the cooption of various components of the target tissue stroma. The principal functions involved in each step are listed in each box. The thickness of the arrows reflects the relative rate of success of cancer cells at each transition. Residual tumor cells remaining after treatment (R_x) may rely on functions that had been previously selected during the micrometastatic state. Individual metastatic traits, e.g., the expression of a metastasis gene, mediate particular functions that slightly increase the probability that a cancer cell will perform a particular step of metastasis. All metastatic traits combined determine the probability that cancer cells will achieve overt metastasis. For an individual cancer cell this probability is very small, but for a large cancer cell population this probability can be absolute.

Viewing these steps as orderly cell biological phenomena akin to developmental programs has been useful for mechanistic dissection of the metastatic process and has allowed the identification of a number of metastasis promoting and suppressing genes and functions (Valastyan and Weinberg, 2011). However, cancer is not a developmental program but an evolutionary process during which only a minority of malignant cells succeed. In this process, the mediators of metastasis function as factors that slightly increase, on a per cell basis, the probability of successful completion of one or more steps of the metastatic cascade.

Metastasis Bottlenecks

Biologically, metastasis is a highly inefficient process. Aggressive tumors are thought to release cancer cells into the circulation by the thousands each day, as can be inferred from the numbers of circulating tumor cells (CTCs) present in the blood of cancer patients (Baccelli et al., 2013; Nagrath et al., 2007; Stott et al., 2010) and experimental models systems (Yu et al., 2012). Much research has revealed in exquisite detail the molecular underpinnings of cell invasion, motility, and stromal interactions that lead cancer cells to enter the circulation and reach distant organs (Kessenbrock et al., 2010; Roussos et al., 2011). However, leaving a tumor is the easy part. The odds that an aggressive cancer cell in the circulation will form a metastatic colony in some organ are vanishingly small.

Most cancer cells that leave a tumor die, and much of this attrition happens as the circulating cancer cells infiltrate distant organs (Kienast et al., 2010; Luzzi et al., 1998). Even cell populations that are experimentally enriched for metastasis-initiating cells suffer extreme attrition in the organs that they invade (Chambers et al., 2002). For example, cancer stem cells isolated from patients with metastatic melanoma are capable of forming a tumor when individually implanted in the skin of a mouse (Quintana et al., 2008), but would probably not form a metastasis if individually inoculated in the general circulation. The same is true for colorectal cancer (CRC) cells when challenged to colonize the liver parenchyma (Calon et al., 2012).

The main bottlenecks for metastasis formation therefore seem to occur during the colonization of distant organs (Figure 1). The metastatic compatibility of certain tissues as envisioned by the seed-and-soil hypothesis is only relative. To infiltrating cancer cells, the best soil is still a deadly soil, just a bit less deadly than others. The stress of passing through endothelial barriers, a lack of survival signals and a supportive stroma in the host tissue, and an overexposure of solitary cancer cells to the perils of innate immunity challenge cancer cells that infiltrate distant organs (Chambers et al., 2002; Nguyen et al., 2009a; Vesely et al., 2011). Why cancer cells so easily die at distant sites is currently unknown, but this death en masse represents a major barrier to metastatic cancer progression. Clinically, it is also the most relevant barrier, because previous steps of cancer cell emigration from primary tumors and distribution to distant organs have already occurred for months at the time of cancer diagnosis. Identifying the natural mechanisms that eliminate disseminated cancer cells might facilitate the development of new therapeutic strategies to prevent or combat metastasis.

Metastatic Latency

Overt metastasis may eventually arise from residual populations of disseminated cancer cells that managed to survive in host tissues. Clinically, “metastatic latency” refers to the period elapsed between the diagnosis of a primary tumor and the emergence of detectable metastatic lesions. At the cell biological level, latency also refers to various states that disseminated cancer cells may adopt during this period of indolence. These states include growth arrest (“dormancy”) of solitary or micro-metastatic cancer cells and unproductive micrometastatic growth with cell proliferation counterbalanced by cell death or loss of stem cell fitness (“stemness”). These various states could coexist in the population of disseminated cancer cells residing in a patient. The biology of metastatic latency and reactivation is largely unknown, at least partially due to the lack of suitable experimental systems that would model this aspect of metastasis. However, recent work in mouse models of breast cancer provides insights into the kind of differentiation signals and stromal interactions that may be involved (Gao et al., 2012; Lu et al., 2011).

The clinical course of metastatic tumor progression can vary dramatically between tumor types and between patients. Some locally invasive cancers, such as glioblastomas, form distant metastases only rarely (Beauchesne, 2012; Lun et al., 2011), whereas other tumors of the brain, such as medulloblastoma, frequently metastasize (Wu et al., 2012). Lung and pancreatic cancers are frequently associated with metastasis at the time of diagnosis (Feld et al., 1984; Werner et al., 2013) whereas breast and prostate cancers are not (Lim et al., 2012; Popiolek et al., 2013). Some tumors metastasize so avidly that cancer in these patients is diagnosed as metastases of unknown primary source (Pavlidis and Pentheroudakis, 2012).

Assessing the true temporal patterns of primary tumor, latent metastasis, and overt metastasis only based on clinical observation is challenging. For example, pancreatic adenocarcinoma was thought to metastasize early because in many cases metastasis is already present at the time of disease diagnosis. However, mathematical modeling of exome sequencing data from matched primary and metastatic lesions suggests that this may be due to late diagnosis, not early metastasis, and the development of metastatic pancreatic cancer may in fact take decades (Yachida et al., 2010). Nevertheless, the fact that some tumor types, such as estrogen-positive breast cancer (Lim et al., 2012) and prostate cancer (Popiolek et al., 2013), are associated with long latency periods between primary tumor resection and the development of metastasis indicates that cancer cells in these tumors are disseminated long before they acquire the capabilities of metastatic colonization.

Metastasis Organ Tropism

Solid tumors display dramatic variation in the pattern of metastasis. Some mainly relapse in one particular organ (e.g., prostate cancers in bone, ocular melanomas in liver, sarcomas in lungs) whereas others relapse in multiple organs (e.g., triple-negative breast cancers, skin melanomas, lung cancers, renal carcinomas). Blood circulation patterns can direct cancer cells to a particular organ, as in the case of the mesenteric circulation directing CRC cells to the liver. However, most solid tumors release cells into the general circulation reaching many organs. The fenestrated endothelium of bone marrow and liver capillaries is more permissive than are the contiguous capillary walls in other organs, particularly in the brain. The capacity of circulating cancer cells to pass through endothelial walls may, therefore, influence the organ tropism of tumors. Furthermore, because the vast majority of cancer cells infiltrating a distant organ die, the capacity of circulating cancer cells to resist decimation in specific organ microenvironments is another determinant of organ-specific metastasis (Nguyen et al., 2009a) (Figure 2A).

Figure 2. Determinants of Metastatic Organ Tropism.

(A) Metastasis as a whole, and organ tropism in particular, are determined by traits that operate in two consecutive, but temporally separate stages: establishment of a disseminated tumor cell population and development of overt metastasis. Depending on the tumor type, the separation between these two stages as determined by their clinical manifestation may be weeks (e.g., in lung adenocarcinoma) or decades (ER-positive breast cancer). The organ tropism of cancer cell dissemination and micrometastasis is determined by circulation patterns from the primary tumor to different distant organs, the permissiveness of capillary walls to extravasation, and the presence of locations (“niches”) in the invaded parenchyma that provide a supporting home for the survival and stemness of disseminated tumor cells. Capillary walls can be permissive for extravasation (e.g., in the fenestrated capillaries of the bone marrow), difficult (e.g., in the capillaries of the lungs) or very difficult (e.g., the brain-blood barrier capillaries). The organ tropism of the overt colonization is determined by the existence of signals that reactivate latent metastatic cells, the ability of these cells to co-opt specialized components of the parenchyma (e.g., osteoclasts in the bone marrow, astrocytes in the brain), and the ability to evade therapy owing to drug access or intrinsic drug resistance properties.

(B) Each type of cancer has a typical pattern of metastatic relapse, initially involving mainly one organ (e.g., prostate cancer metastasis to bone, sarcoma metastasis to lung) or multiple organs (e.g., lung and breast carcinomas). Therapeutic treatments may suppress metastasis in all organs uniformly (as in “Cancer A”) or in some organs more than in others (as in “Cancer B”). In the latter case, clinical management of the cancer alters the organ metastatic pattern of the disease.

Mediators of extravasation have been identified (Bos et al., 2009; Clark et al., 2000; Gupta et al., 2007; Padua et al., 2008; Weis et al., 2004; Wolf et al., 2012). Expression of these genes in cancer cells increases their accumulation as disseminated seeds in susceptible organs, thereby augmenting the probability of eventual relapse in those organs. The same applies to mediators of cancer cell survival in distant organs (Chen et al., 2011; Valastyan et al., 2009; Zhang et al., 2009). The expression of these mediators in broad segments of the cancer cell population in primary tumors is reflected in metastasis signatures that predict relapse to various organs in clinical tumor cohorts (Bos et al., 2009; Cheung et al., 2013; Minn et al., 2005; Nguyen et al., 2009b; Padua et al., 2008; Tavazoie et al., 2008; Valastyan et al., 2009; Vanharanta et al., 2013; Zhang et al., 2009). These findings have provided important insights into the biology of metastatic dissemination and the proteins and microRNAs mediating it.

Although circulation patterns, extravasation barriers, and survival on arrival are three key determinants of the capacity of particular tumors to seed specific organs, a different set of conditions determine the capacity of micrometastatic seeds to develop into macrometastases (Figure 2A). Overt colonization critically depends on the capacity of disseminated cancer cells to benefit from specific stromal components in a particular organ. For example, cancer cells exploit osteoclasts in the bone marrow for osteolytic metastasis (Ell and Kang, 2012; Guise, 2009) and astrocytes in the brain parenchyma for cerebral metastasis (Kim et al., 2011; Xing et al., 2013). These traits may come into play only when disseminated cells emerge from the latent state, months to decades after having seeded that organ. This implies that in addition to the tumor reinitiation phenotypes inherited from the cell-of-origin or acquired during earlier steps of tumor progression, some of the overt colonization traits may be selected only after micrometastatic cells initiate aggressive outgrowth in the target organ parenchyma.

Both the determinants of metastatic seeding as well as those of overt colonization of different organs underlie the patterns of metastatic organ tropism of each type of cancer (Figure 2A), that is, the probability that a particular cancer will relapse in specific organ sites. It is possible that in the absence of an effective treatment, and if given enough time, metastasis would emerge in all organs in every case. Indeed, advances in disease management are prolonging the life of patients with metastases, but also changing the patterns of metastasis in certain types of cancer (Figure 2B). For example, a current rise in the incidence of brain metastasis of HER2+ breast cancer is attributable to the fact that therapeutic agents targeting the HER2 oncoprotein are effective in controlling extracranial metastases but less effective against brain metastases (Sledge, 2010). Thus, differences in sensitivity to therapy of disseminated cancer cells in different organs constitute one more determinant of metastasis patterns.

Clonal Evolution and Tumor Heterogeneity

Genetic Analysis of Human Cancer

Cancers arise through cycles of mutation and clonal selection (Stratton et al., 2009). A classical genetic model of human cancer progression is provided by the analysis of CRC development. These tumors arise through progression of small adenomas into full-blown carcinoma with associated mutations that drive tumor progression (Fearon and Vogelstein, 1990). Inactivation of the APC tumor suppressor is the gatekeeper event, followed by mutations in KRAS, TP53, SMAD4, PIK3CA, TGFBR2, and others (Jones et al., 2008). Based on these observations, it was envisioned that yet other genetic alterations would make carcinomas metastatic (Fearon and Vogelstein, 1990). When technology enabled exome sequencing of human tumor samples, a comparison of primary tumors and matched metastases found no mutations in CRC that were specifically and recurrently associated with metastasis (Jones et al., 2008).

Two important conclusions emerged from these studies. First, the intervening period between tumor initiation, i.e., the time when normal cells turn into premalignant precursors of cancer, and the emergence of invasive carcinoma far exceeds the additional time that it takes for the development of metastasis (Jones et al., 2008). Second, mathematical modeling of tumor evolution based on the genetic data suggested that the selective advantage provided by individual driver mutations is on average rather low, only ~0.4% (Bozic et al., 2010).

These lessons are complemented with insights from other cancers. Instead of recurrent metastasis-specific mutations, alterations are found in genes that are commonly mutated in primary tumors. For example, comparative genomic sequencing of several tumor specimens of the same patient has revealed metastasis-associated mutations in renal, pancreatic, and lobular breast cancers (Gerlinger et al., 2012; Shah et al., 2009; Yachida et al., 2010). Metastatic clones in pancreatic cancer may harbor amplifications of the oncogenes KRAS^G12V, MYC, and CCNE1 (Campbell et al., 2010). Mutant alleles of the tumor suppressors TP53, SETD2, and KDM5C were found in renal carcinoma metastases (Gerlinger et al., 2012). In a patient with ER-positive breast cancer from whom metastatic cells were isolated 9 years after primary tumor resection, the most compelling metastasis driver mutation was in ERBB2 (Shah et al., 2009), which harbors similar activating mutations in primary breast cancers as well (Bose et al., 2013; Cancer Genome Atlas Network, 2012). In basal breast cancer, a similar approach revealed differences in mutant allele frequencies between primary and meta-static tumors, but the affected genes were the same with few exceptions (Ding et al., 2010). Single cell analysis of a triple negative breast cancer resulted in compatible observations (Navin et al., 2011).

Although no metastasis-specific, recurrent, driver mutations have been demonstrated thus far, metastatic cell clones clearly do not represent the whole primary tumor population, but only parts of it (Campbell et al., 2010; Gerlinger et al., 2012; Yachida et al., 2010). In a pancreatic cancer, lung metastases and peritoneal metastasis were shown to descend from different clones in the primary tumor (Campbell et al., 2010; Yachida et al., 2010), demonstrating in tissue samples what has long been clear in experimental models of metastasis: different metastatic clones differ in their ability to colonize specific organs (Nguyen et al., 2009a). In prostate cancer, on the other hand, metastatic tumors in different target organs seem to arise from a single metastatic clone (Liu et al., 2009).

The longer the latency period between the emergence of a primary tumor and that of distant metastases, the more likely it is that metastatic clones will carry specific genetic alterations. The pressure to adapt to, alter, and eventually overtake the host stroma combined with the genetic drift of genomically unstable cancer cells all but ensure that some genetic alterations will be enriched for in metastatic clones. The number of metastatic tumor specimens that have been analyzed in comparative genomic studies remains low, and it is not clear whether any of the metastasis-associated driver mutations identified have higher mutation frequencies in metastatic lesions more generally (Figure 3). An emerging conclusion suggests, however, that the mechanism of action of the metastasis-associated gene products is not likely to be metastasis-specific.

Figure 3. Genetic Alterations in Metastatic Cancer.

A schematic shows the possible classes of metastasis-associated cancer driver mutations as a continuum ranging from primary tumor-specific to metastasis-specific genetic lesions. Systematic cancer genome resequencing efforts are beginning to shed light on the mutational complement of metastatic cancers by comparing the genetic alterations in primary tumors and corresponding metastatic lesions in a procedure termed comparative lesion sequencing (Jones et al., 2008). Ideally, several regions of individual tumors are sampled in order to account for clonal heterogeneity. Mutations that would have a higher selective advantage at the metastatic site would also have a higher mutation frequency in metastatic lesions. Conversely, some mutations that are advantageous at the primary site might make cancer cells less fit for metastatic progression, in which case the mutation frequency would be higher in primary tumors. At extremes, a type of mutation would be either fully metastasis- or primary tumor-specific, respectively. Relative difference in mutational frequency between primary and metastatic cancer (Δf) is defined as [f(T) − f(M)]/[f(T) + f(M)], where f(T) and f(M) refer to the mutational frequency of a gene in primary tumors and corresponding metastasis, respectively. α, Mutations with higher frequency in primary tumor; β, mutations with equal frequency in both primary and metastatic tumors; γ, mutations with higher frequency in metastatic tumors; δ, metastasis-specific mutations.

Lessons from Mouse Models

Recent work combining analysis of human tissues and mouse models of CRC revealed a strong impact of nongenetic events on the probability of liver metastasis (Calon et al., 2012). CRC cells that frequently become insensitive to the tumor suppressive action of TGF-β via genetic inactivation of the TGF-β signaling pathway can afford to overexpress this cytokine, which enhances metastasis-formation in the liver by inducing the secretion of the prosurvival cytokine interleukin-11 from stromal fibroblasts (Calon et al., 2012). Thus, a genetic event in cancer cells enables the selection of a trait—high expression of TGF-β—that strongly favors metastasis by profitably engaging the stroma. What drives the high expression of TGF-β in TGFBR2 mutant CRC cells may not be another mutation but rather a diverse set of epigenomic regulators of TGF-β production, any of which underlies this highly valuable trait.

Several studies have identified pathways that promote or suppress metastasis in genetically engineered mouse models of cancer. Tumor initiation by Pten loss in the prostate can trigger activation of the TGF-β tumor-suppressor pathway through SMAD4 expression (Ding et al., 2011). In this context, yet unidentified signals can induce the expression of COUP-TF2, which can inhibit SMAD4 and consequently allow the tumors to metastasize to lymph nodes and the lungs (Qin et al., 2013). Simultaneous inactivation of Pten, Tp53, and Smad4 in the mouse prostate gives rise to aggressive tumors with some bone meta-static activity, whereas tumors with Pten and Tp53 mutations alone are more indolent (Ding et al., 2012b). Smad4 inactivation thus seems to be a critical event in the development of advanced prostate cancer that eventually may also form metastasis. Similarly, Nkx2.1 loss in Kras^G12D/Tp53^−/−-driven lung adenocarcinomas allows tumor progression and metastasis (Winslow et al., 2011). Activation of Notch signaling upon Aes deletion in mouse intestinal tumors promotes local invasion, extravasation, and metastasis (Sonoshita et al., 2011). These mouse studies point at important pathways as mediators of tumor progression, including metastasis, but they do not provide evidence for metastasis-restricted genetic alterations. A possible exception is provided by the results of a genetic screen that used transposon-based insertional mutagenesis in medulloblastoma: certain insertions promoted leptomeningeal metastasis on both Ptch mutant and Tp53 mutant backgrounds (Wu et al., 2012). Thus, a minority clone in a primary tumor was prompted by a particular genetic alteration to form metastasis, providing evidence that a clonal mutation can provide metastatic advantage.

In some mouse models, cancer cell dissemination is observed before primary tumors become overtly invasive (Hüsemann et al., 2008; Rhim et al., 2012). Even nontransformed stem cells placed in the circulation can infiltrate the lungs and survive for extended periods of time (Podsypanina et al., 2008). These observations raise the possibility that early-disseminated cancer cells may evolve independently of, and in parallel with, cancer cells in the primary tumor (Klein, 2009). However, this model is not independently validated by genetic data from human tumor samples. It also remains unknown whether the progeny of the cells that disseminate early form clinically manifest metastases later on. It could be that cancer cells that disseminate later after having become more aggressive in the primary tumor stand a better chance of forming metastases.

Amplifying Oncogenic Signaling

If not through genetic alterations, how do clones with high metastatic propensities emerge under selective pressure? Accumulating evidence from integrative analysis of metastatic xenograft models and large clinical gene expression data sets suggests that the selected traits augment the robustness of signaling pathways that already are active in primary tumors, increasing the odds that cancer cells will thrive in distant organs. In this model, metastatic fitness is a function of the signaling amplitude of certain oncogenic pathways.

In breast cancer cells of the triple-negative (ER–, PR–, HER2–) subtype, several prometastatic genes have been identified that amplify the output of cell survival and stemness pathways that are already selected for in the primary tumor. For example, high expression of the vascular cell adhesion molecule-1 (VCAM-1) in breast cancer cells hypersensitizes the PI3K-Akt cell survival pathway to activation by limiting external signals. VCAM-1 can be engaged by α4β1 integrins on tumor-associated macrophages, leading to activation of ezrin, a PI3K and Akt adaptor protein (Chen et al., 2011) (Figure 4A). Similarly, hyper-activity of the tyrosine kinase Src in breast cancer cells sensitizes the PI3K-Akt pathway to activation by CXCL12 (via its receptor CXCR4) and IGF1 in the bone marrow stroma (Zhang et al., 2009). Also, expression of the extracellular matrix protein tenascin C (TNC) in breast cancer cells augments the output of the Notch and Wnt pathways in support of metastasis-initiating cell fitness in the lungs (Oskarsson et al., 2011). The main interaction partner of TNC in the extracellular matrix of adult stem cells, periostin (POSTN), binds and concentrates Wnt to the same end (Malanchi et al., 2012) (Figure 4A).

Figure 4. Amplification and Expansion of Oncogenic Pathways as Metastatic Traits.

(A) Metastatic traits acquired by a quantitative gain in pathway output. Amplification of the signaling capacity of cell survival and proliferation pathways provides survival and stemness advantages to disseminated cancer cells. In this case, the level of metastatic fitness is proportional to the robustness of the signaling pathway. In primary tumors with a relatively high abundance of growth and survival signals, the signaling capacity of oncogenic pathways such as PI3K, Notch, and Wnt, is sufficient to support tumor growth. However, for disseminated cancer cells reaching distant tissue microenvironment where pathway-activating signals are scarce, the signaling capacity of these pathways is not sufficient for survival. The traits selected under such pressure include the expression of components that amplify the signaling capacity of the pathway in response to limiting levels of pathway activators in the host microenvironment. In triple-negative breast cancer examples include VCAM-1 and SRC as amplifiers of the PI3K pathway in cancer cells reaching the lungs or the bone marrow, respectively, whereas Tenascin C and Periostin act as amplifiers of the WNT and NOTCH pathways in cancer cells reaching the lungs.

(B) Metastatic traits acquired by a qualitative expansion of pathway output. Tumor-initiating pathways may additionally provide prometastatic traits by gaining access to subsets of target genes that enhance the homing of disseminated cancer cells to sites of survival. In this case, the level of meta-static fitness is not linearly proportional to the signaling strength of the pathway but depends of the pathway activating an additional set of effectors. One example is provided by the expansion of target genes that the HIF pathway activates in renal cell carcinoma as a result of epigenetic modifications that open the gene promoters to access by activated HIF.

An interpretation of these observations is that metastatic cells need to optimize the output of their vital intracellular stemness and survival pathways when venturing away from the primary tumor and infiltrating alone distant sites. The activity level of these pathways provided by the early oncogenic events may have allowed the tumor to progress locally, but in metastatic sites with limiting levels of trophic cues signal amplification by factors such as VCAM-1, Src, TNC, and POSTN might be required for cell survival. This same purpose may be served, on a lesion-specific basis, by the noted amplifications of KRAS^G12V, MYC, and CCNE1 in pancreatic cancer metastasis (Campbell et al., 2010), or the expression of certain mediators of cancer invasion in melanoma (Scott et al., 2011). Conversely, some of the genes associated with lung metastasis in breast cancer patients and in experimental systems were found to also be associated with primary tumor growth (Minn et al., 2007; Minn et al., 2005). Thus, the signal optimization that is ultimately required for metastasis can increase the fitness of cancer cells and get selected for already in primary tumors.

Expanding the Output of Oncogenic Pathways

The involvement of oncogenic pathways in conferring metastatic traits is not limited to quantitative gains in the robustness with which these pathways transmit stemness and survival signals. The target gene spectrum of a transcriptional program can also undergo qualitative modulation through alterations in chromatin accessibility of its target loci (John et al., 2011). In the context of metastatic progression, such modulation could expand the output of an already activated oncogenic pathway by allowing it to express, in addition to the basic tumorigenic functions, metastatic phenotypes as well (Figure 4B).

One example has been identified in renal cell carcinoma, a tumor type that is initiated by loss of the von Hippel-Lindau tumor suppressor (VHL) and consequent activation of hypoxia-inducible transcription factors (HIFs) (Kaelin, 2008). During tumor progression, histone H3 lysine 27 (H3K27) demethylation and DNA demethylation can allow the VHL-HIF pathway to access new target genes, such as the chemokine receptor CXCR4 and cytohesin 1-interacting protein (CYTIP), which mediate metastatic invasion and outgrowth in the lungs and bones (Vanharanta et al., 2013). Alterations in the distribution of the repressive H3K27me3 chromatin mark are also associated with metastatic progression in breast cancer, where high expression of the noncoding RNA HOTAIR can redistribute H3K27me3 across the genome in order to repress metastasis suppressor genes (Gupta et al., 2010). Moreover, the genome-wide binding patterns of the estrogen receptor evolve during breast cancer progression, with possible functional consequences (Ross-Innes et al., 2012).

These examples illustrate how the genetic activation of an oncogenic pathway during tumor initiation does not automatically lead to the manifestation of this pathway’s full malignant potential. Rather, a pathway that initially drives primary tumor formation may evolve through quantitative and/or qualitative amplification of its output, ultimately providing the cell with a metastatic advantage.

Sources of Influence: Germline, Plasticity, Stroma, and Therapy

The intrinsic metastatic traits of cancer cells may be further enriched or diminished by various sources of influence. Germline variants play a role in cancer predisposition (Pharoah et al., 2004). It is therefore also likely that human germline variants affecting tumor progression exist, as has been demonstrated by genetic mapping analysis in mice (Park et al., 2005). There is evidence for a heritable component in human cancer prognosis (Hemminki et al., 2011), but identifying specific poor prognosis variants has proven difficult. Certain gene polymorphisms seem to correlate with poor clinical outcome, one example being the FGFR4^G388R variant (Spinola et al., 2005). However, the difficulty of conducting genome-wide association studies on cancer progression have, at least for now, prevented a thorough assessment of the germline component of metastatic propensity.

The ability to adopt phenotypic changes, including changes into a more stem-like phenotype, can help cancer cells through metastasis progression bottlenecks. Cytokine signals induce the reactivation of developmental epithelial-to-mesenchymal transition (EMT) programs in cancer cells (Yang and Weinberg, 2008). EMT has been proposed to provide cancer cells with several prometastatic traits, including stemness, motility, and resistance to therapy, all at once (Mani et al., 2008; Yang et al., 2004). Although the role of EMT in developmental processes is well established (Thiery et al., 2009), its significance for cancer progression is still being defined. For example, EMT-inducing transcriptional regulators can both promote (Yang et al., 2004) and inhibit metastatic colonization (Ocaña et al., 2012; Tsai et al., 2012). Also, metastatic lesions in patients show epithelial, not mesenchymal features. It has been suggested that EMT provides a transient benefit in cancer cell dissemination but must be reverted by mesenchymal-to-epithelial transition (MET) at the metastatic site (Korpal et al., 2011; Ocaña et al., 2012). Although several molecular players in EMT and MET can impart metastatic traits in experimental systems, more work is required for an understanding of the specific changes that these mediators trigger as well as their specific contributions to metastasis.

In all tumors, a variety of non-neoplastic stromal cells interact with the cancer cells. This can result in both pro- and antitumorigenic effects through often complex molecular mechanisms (Hanahan and Coussens, 2012). Several specific examples ranging from systemic to juxtacrine signaling loops that modulate primary tumor progression and metastasis have been described, most of which represent reactive interactions between stroma and cancer cells (Acharyya et al., 2012; Calon et al., 2012; Chen et al., 2011; DeNardo et al., 2009; Gocheva et al., 2010; Granot et al., 2011; Karnoub et al., 2007; Labelle et al., 2011; Lin et al., 2001; Lu et al., 2011; Magnon et al., 2013; McAllister et al., 2008; Nieman et al., 2011; Pencheva et al., 2012; Png et al., 2012; Qian et al., 2011; Schwitalla et al., 2013; Sethi et al., 2011). Such tumor-shaping cues are not exclusively produced by stromal cells, but other environmental conditions, such as hypoxia, can also induce metastasis-promoting changes in cancers (Chaturvedi et al., 2013; Eisinger-Mathason et al., 2013; Erler et al., 2006; Gilkes et al., 2013). Tumor stromal signals also play a role in attracting circulating metastatic cells in a tumor self-seeding process that favors the clonal expansion of these cells (Comen et al., 2011; Kim et al., 2009).

In addition to the reactive and often reversible interactions with the components of the microenvironment, the stroma contributes to tumor progression through more subtle ways by altering local selective pressures (Gillies et al., 2012). This can lead to the selection of cells with enhanced metastatic potential. For example, many of the stromal cells in tumors originate from the bone marrow, potentially making the tumor stroma more “bone marrow-like” through secretion of soluble factors such as CXCL12 and IGF1. Cancer clones evolving under the influence of such factors can then become enriched for qualities that allow them to survive in and take advantage of the bone marrow already before reaching the bone. This phenomenon of “metastasis seed preselection” provides one possible explanation for the emergence of metastatic traits already at the primary tumor site (Zhang et al., 2013) and it could be further amplified by tumor seeding with CTCs released from metastatic lesions (Kim et al., 2009).

Metastatic cancers are also exposed to, and evolve under, therapy-induced pressures (Diaz et al., 2012). Increasing evidence suggests that there are molecular links between metastatic cancer progression and drug resistance. For example in breast cancer, the transmembrane protein metadherin (MTDH) provides cancer cells with protection against chemotherapy as well as enhanced capability to metastasize to the lung (Hu et al., 2009). Moreover, chemotherapy can modify the tumor microenvironment and consequently enhance the expression of factors such as the chemokine CXCL1 that can foster therapy resistance and promote metastasis via stromal interactions (Acharyya et al., 2012). The traits conferring resistance to therapies may therefore also drive metastatic tumor progression.

Deterministic versus Stochastic Emergence of Metastasis Clones

Even the most predictive markers of metastatic tumor progression only indicate a higher probability of metastasis at the organismal level. At the cellular level, the degree of randomness is even greater, as the tissues of patients who ultimately develop one lethal metastatic lesion may carry thousands of disseminated micrometastatic cells that never took off. Therefore, despite the identification of molecular mediators that confer a metastatic phenotype and constitute the deterministic component of the metastatic process, metastases also display strong stochastic features (Talmadge and Fidler, 1982). Both syngeneic mouse models as well as human xenograft systems have demonstrated that at least in experimental systems stable metastatic cancer cell clones do exist (Talmadge and Fidler, 2010), and analysis of clinical cancer data sets have shown that tumors that are prone to metastasis share gene expression traits with such clones (Bos et al., 2009; Minn et al., 2005; Nguyen et al., 2009b; Pencheva et al., 2012; Vanharanta et al., 2013). Also, certain well-characterized prometastatic functions are specifically required for colonization of particular organ sites, such as the bone marrow (Ell and Kang, 2012). All this suggests that specific heritable molecular alterations, both genetic and epigenetic, determine the capacity of cancer cells to form metastasis.

However, a stochastic nature of metastatic cancer progression is evident at several levels. First, tumor evolution feeds from random heritable alterations that increase diversity upon which selection can act. Of the possible evolutionary paths available for a given cancer cell clone, only some can result in a metastatic phenotype. As determined by the combined effects of their tissue of origin and acquired alterations, many cancer clones may end up in fitness peaks that are not compatible with metastases. A phenotype that is beneficial during early phases of tumorigenesis may thus prevent the same clone from becoming metastatic later on (Figure 5A). Second, any cancer phenotype can be reached through multiple evolutionary routes, which is reflected by the diverse mutational complements of individual cancers. As a consequence, no two cancer clones are identical (Figure 5B). Third, as demonstrated in experimental transplantation models, even within clones endowed with high metastatic fitness, few cells actually form a metastasis. The fate of phenotypically identical metastatic cells will therefore never be identical (Figure 5C). The existing genetic evidence would be compatible with a stochastic model whereby highly aggressive cancer cell clones that emerge from primary oncogenic events all have the potential to form metastasis at a very low frequency (Jones et al., 2008).

Figure 5. Levels of Stochasticity in Metastatic Cancer Progression.

The stochastic nature of the evolution of metastatic cancer clones is manifest at several levels of cancer progression.

(A) For every cancer clone, depending on their cell of origin and previous somatic alterations, only a defined set of fitness peaks are achievable, and only some of these peaks can reach high metastatic fitness (red peaks). Starting from a normal cell, cancer cell clones move within fitness landscapes by acquiring random sets of driver alterations. This leads each clone to one of the possible fitness peaks (red arrows). The ratio of possible metastatic and nonmetastatic evolutionary paths for a given cancer clone may vary greatly depending on the cell of origin or initial mutational events. Some tumors may have none whereas some may have several ways to become metastatic. This is reflected in the fact that the incidence of metastatic progression varies between tissue and tumor types.

(B) The (epi)genetic changes that allow different cancer clones to climb up the same metastatic fitness peak are never identical. Thus, even tumors that arise from identical normal cells through identical tumor-initiating mutations, and that ultimately display clinically similar metastatic phenotypes, have acquired those metastatic traits through a different evolutionary path (dotted red lines).

(C) Of the cellular progeny of a cancer clone that, in principle, have the qualities that are required for metastatic colonization, only a minuscule fraction (α) will ever form meaningful metastasis. Thus, even for cells that have reached a metastatic fitness peak, the attrition rates remain extremely high. This means that the fate of phenotypically identical metastatic cells is in most cases not identical.

Closing the circle, some of the observed randomness in the metastatic process might in fact be the consequence of factors yet to be discovered. For example, the formation of metastasis may be possible only at certain defined microenvironmental niches, such as in the hematopoietic stem cell niche in the bone marrow (Ding et al., 2012a; Shiozawa et al., 2011). The need of cancer cells to fall in a suitable niche in order to thrive could provide a partial explanation to the stochastic nature of metastasis-formation (Ghajar et al., 2013). At a more general level, systemic inflammatory reactions triggered by primary tumors may modulate the stroma at distant organs to unevenly increase metastatic colonization of otherwise resistant tissues (Kaplan et al., 2005).

The development of metastatic cancer clones is thus the end result of an intimate interplay between selection and stochastic events. As we learn more, an increasing fraction of the randomness in the process may turn out to be nonrandom. However, only by appreciating that metastasis results from a combination of nonrandom determinants and random chance we may arrive at a comprehensive understanding of the origins of metastatic traits.

Conclusions and Perspective

Progression from an early neoplastic lesion to a metastatic cancer is a long evolutionary process that often takes years, if not decades. Rather than involving mutations that individually provide strong competence to perform multiple metastasis steps, metastasis involves multiple alterations, each providing a slight selective advantage in at least one of the many steps of metastasis. Multiple such alterations must accumulate for a metastatic clone to emerge. The metastatic cell is a rare end result of extensive genomic and epigenomic tinkering. Identifying the most demanding steps in metastasis and therapeutically targeting the metastatic traits that mediate these steps are aims of future research.

Even with the most aggressive metastatic clones, the success rate of metastatic colonization remains low. The vast majority of disseminated cancer cells in any target tissue die before they can form metastatic colonies. The molecular mechanisms of this widespread death are for the most part unknown. Identifying what kills disseminated cancer cells en mass may provide clues for therapeutic intervention against the survivors.

Metastatic cancer cell clones are genetically distinct and do not necessarily represent the majority of a primary tumor. Clonal heterogeneity within primary tumors is a source for the selection of metastatic cancer cells. Sampling of multiple locations within a tumors and its metastasis is needed in order to establish the clonal evolution of metastatic processes.

High-throughput genetic analysis of primary human cancers and their metastatic counterparts have identified metastasis-associated mutations but not metastasis-specific mutations. Many, though not all, likely are passenger events without causal role in metastasis progression. Hence, from a genetic perspective, metastasis is an extension of primary tumor progression, not a distinct step with characteristic mutational determinants. The lack of singular mutational events that would selectively confer metastatic potential may reflect the fact that the phenotypes of metastatic cells are complex and not achievable through a single alteration.

Cancers evolve under strong microenvironmental and therapy-induced pressures. Both stromal cells and therapeutic agents can limit cancer growth but they can also skew the clonal distribution of a tumor and thus provide opportunities for the expansion and metastatic dissemination of selected cancer sub-clones. Therapy-resistance and metastasis are thereby molecularly linked.

Mutational activation of oncogenic pathways does not automatically lead to the manifestation of their full tumorigenic and metastatic potential. In metastatic cancer cells, signaling pathways have been tuned through selective pressures to become as prometastatic as possible, resorting to many of the principles enunciated above. Unlike developmental processes with layers of regulators providing robustness to the system, the emergence of highly metastatic cancer clones often seems to be accompanied by the loss of such safety nets in order to achieve maximal output of oncogenic pathways. This may lead to vulnerabilities that could be useful for therapeutic intervention.