Metastasis-Initiating Cells and Ecosystems

Clinically evident metastasis remains largely incurable, but progress is being made in understanding its basic biological principles and laying the foundation for improving long-term outcomes in patients with advanced cancer (1, 2). In spite of the aggressiveness of metastatic cancer, only a small proportion of cells released from a tumor end up forming a distant lesion. This is partly due to the extensive attrition that cancer cells suffer as they confront the stresses of cancer cell dissemination, dormancy, outgrowth, and resistance to therapy (2). Cancer cells must accumulate numerous phenotypic traits to survive these stresses and become proficient at initiating metastasis.

It is increasingly clear that metastasis is initiated by a subset of cancer cells with a stem cell–like phenotype that makes them more competent than the bulk cancer cell population at reinitiating tumor growth in distant sites (3–5). Cell “stemness” in mature tissues represents a phenotypically fluid condition rather than a developmentally imprinted and rigidly defined state. Stem cells in different tissues have distinct properties and developmental trajectories, as do stem cells that support tissue homeostasis versus those that are mobilized in response to injury to regenerate the disrupted tissue. Adult stem cells that maintain homeostasis may suffer oncogenic mutations and become tumor-initiating cancer stem cells (CSC). However, the cells that initiate metastasis are not necessarily CSCs that acquired a few extra metastatic traits. Recent progress indicates that metastatic colonization of distant organs is initiated by stem-like cancer cells that are distinct not only in terms of expressing certain specialized prometastatic traits but also in terms of their overall phenotype, which is more akin to that of regenerative stem cells. Hence, we refer to these cells as “metastasis-initiating cells” (MIC), a term that denotes what these cells can do and why they are of interest. Cells with MIC features are enriched in residual disease that survives cancer therapy, underscoring their importance not only in the original formation of distant metastases but also during relapse of increasingly lethal metastases after multiple treatment cycles in the clinic.

Understanding the origins, traits, and vulnerabilities of MICs and their host ecosystems is critical for developing strategies to improve the prevention and treatment of metastasis. Here we leverage the conceptual framework and open questions from previous reviews on the metastatic process (1, 2) and the cells and microenvironments that make it possible (3, 5) to offer an update on recent progress toward the resolution of these questions. We present a summary of the metastatic process and the traits that cancer cells must possess to become MICs. We then address the question of how MICs acquire these traits. We consider the hypothesis that MICs emerge by the progressive accumulation of an array of prometastatic traits through evolution of highly heterogeneous cancer cell populations under the selective pressures of the multiple steps of the metastatic process. We also consider the alternative hypothesis that cancer cells acquire many of their critical traits at once by adopting a regenerative progenitor phenotype with intrinsic tumor regrowth capacity. Finally, we offer a perspective on the implications of the growing understanding for improving the treatment of patients with metastatic cancer.

Metastasis starts with a phase of cancer cell dissemination from the primary tumor to regional and distal sites (Fig. 1A). In certain cancers, such as breast cancer, dissemination may start during early stages of tumor progression while the tumor is still small and appears to be confined by basement membrane within a ductal structure—i.e., ductal carcinoma in situ (6, 7). Dissemination can continue until the source tumor is removed, as demonstrated by the presence of circulating tumor cells (CTC) in patients undergoing surgical resection of the tumor (8). CTCs culminate their journey by exiting from blood capillaries (a process called extravasation) and infiltrating the new host parenchyma.

The dissemination phase is followed by a phase of metastatic dormancy during which disseminated cancer cells do not yet grow to form a clinically manifest lesion because the host stroma presents physical, metabolic, and immune barriers that prevent this growth (9). Disseminated cancer cells survive under these conditions by entering a state of proliferative quiescence. During dormancy, metastasis remains a cryptic process that may eventually progress toward overt tumor growth or disappear by exhaustion of the disseminated cancer cell population under these barriers or by the effect of therapy.

The final phase of the metastatic process involves overt outgrowth by colonization of the host organ (Fig. 1A). Clinically detectable metastases may not become evident until months or decades after initial diagnosis and removal of a primary tumor. In some cancers overt metastasis occurs in multiple sites without marked preference for one or another organ. In other cancers, metastasis shows strong preference or “tropism” for certain organs. Notably, 2% to 5% of metastatic cancers are diagnosed as carcinomas of unknown primary, in which the primary tumor is not detected either because of its small size or because it may have spontaneously regressed after seeding aggressive metastases (10).

The three phases of the metastatic process—dissemination, dormancy, and outbreak—may coexist in time (Fig. 1B). Cancer cell dissemination may persist as long as the source tumor remains in place, long after the first disseminated cells accumulated as metastatic seeds in distant organs. Precocious metastatic lesions may break out from dormancy while the primary tumor is still in place, as occurs in patients whose cancer is first diagnosed as widely metastatic stage IV disease (Fig. 1B). When distant relapse occurs in a particular site long after the surgical elimination of the primary tumor, other organs may still harbor dormant metastatic cells that can independently evolve and eventually initiate metastatic outgrowth with their own set of organ-specific metastatic traits. An example is provided by late brain recurrence in patients with HER2-positive breast cancer who were treated with anti-HER2 agents that suppressed visceral metastasis (11).

A combination of invasive cell migration, extracellular matrix (ECM) remodeling, and the presence of leaky vasculature at the tumor invasion front enables the dispersion of cancer cells through the circulation, lymphatics, perineural, and perivascular routes, or via direct infiltration of adjacent body cavities (ref. 12; Fig. 2). The hematogenous circulation is the principal route in most types of cancer, but lymphatic dissemination and forms of extravascular spread are prominent routes in various types of cancer as well.

Malignant cells remodel the basement membrane to invade into tissue parenchyma, induce neoangiogenesis, move directionally toward blood vessels, and undergo transendothelial migration to enter the portal or systemic venous circulations (refs. 13, 14; Fig. 2A). These steps are enabled by dynamic changes in the repertoire and function of cell–cell adhesion molecules (cadherins and CAM family cell adhesion molecules), ECM remodeling enzymes (matrix metalloproteinases and ADAMTS endopeptidases), and cell–matrix adhesion molecules (integrins and syndecans; refs. 15, 16). Carcinoma cells undergo epithelial-to-mesenchymal transition (EMT), a phenotypic conversion process that enables cell migration and entry into the circulation (see below; refs. 1, 17). Stromal cells, including macrophages, fibroblasts, and neutrophils, can further stimulate invasion and intravasation by direct interactions with carcinoma cells (18, 19). Cancer cell–derived fibrogenic signals stimulate the deposition of desmoplastic matrix by mesenchymal cells in the tumor stroma (20, 21), which in turn facilitates the invasion of cancer cells and their proliferation through mechanosignaling cues (15, 16). The physical stress of migrating through tight interstitial spaces leads to nuclear envelope rupture and chromatin shearing, thus causing further genomic and phenotypic diversification of cancer cells that survive such stress (22, 23).

Unlike tumor tissues, which can be accessed only via invasive biopsies, CTCs derived from primary or metastatic tumors can be readily and repeatedly sampled via simple blood draws from patients with cancer. This provides a unique window to study metastasis biology and for use as longitudinal clinical biomarkers of metastasis and therapy response (24, 25). For these reasons, CTCs have become a topic of intense study. Although the sensitivity and specificity of current CTC detection assays remain limiting for some applications, the prospect of delineating CTC characteristics that discriminate between high and low metastatic capacity is of biological and clinical significance.

Cancer cells circulate singly and in clusters (ref. 26; Fig. 2B). CTC clusters are observed in the blood of patients and tumor-bearing mice and have a superior ability to seed metastasis in experimental models (27). CTC clusters may also include tumor-derived stromal cells (28). CTC clustering provides intercellular spaces that concentrate paracrine growth-promoting signals (29) and enable cooperation between different cancer cell states within a cluster (30). CTC clusters are enriched in genome methylation patterns that denote a stem-like cancer cell state (31), and therefore have features of MICs. MICs upregulate expression of certain cell–cell and cell–matrix adhesion genes (32), raising the question: Are CTC clusters more metastatic because clustering augments the aggressiveness of cancer cells or because MICs are intrinsically prone to clustering?

Blood platelets coat CTCs, providing protection from shear stress and immune attack, particularly attack by natural killer (NK) cells. Platelets can also facilitate cancer cell extravasation by mediating adhesion to selectins on the endothelium and by enforcing EMT in the cancer cells. Platelets release factors including TGFβ and platelet-derived growth factor (PDGF) that suppress NK cells and promote EMT of CTCs (14, 33, 34). A small proportion of CTCs in patients with breast cancer are bound to neutrophils, an abundant white blood cell type; these CTCs are more proliferative and effective at forming metastases in mice than are neutrophil-free counterparts (35).

Within seconds to minutes of leaving a primary tumor, CTCs may already become trapped in the capillaries of distant organs. In experimental systems, CTCs can remain lodged in capillaries for days until they either extravasate or disappear (36). The half-life of cancer cells in the circulation is estimated to be just a few hours, based on the rate of decline in the CTC numbers following surgical removal of primary tumors in patients (8).

Differences in the tightness of capillary endothelial walls determine the difficulty of seeding metastasis in different organs, contributing to organ-specific patterns of metastatic spread (Fig. 2C). Liver and bone, the most common sites of metastasis from many cancers, contain sinusoid capillary beds with large intracellular fenestration gaps as well as gaps in the underlying basement membrane, which facilitate CTC extravasation. At the other end of the range, the blood–brain and blood–choroid plexus barriers restrict the movement of molecules and cells between the systemic circulation and the brain parenchyma and cerebrospinal fluid, respectively. Accordingly, metastasis to the brain and leptomeningeal spaces selects for MICs expressing factors that increase barrier permeability or disrupt the barrier altogether (37–39).

Extravasation can be facilitated by platelets coating CTCs (34) and cytokines secreted by perivascular macrophages in target organs or by the cancer cells themselves (ref. 13; Fig. 2C). As an alternative to squeezing through tight interstitial spaces between endothelial cells, cancer cells may induce endothelial cell death. Human and mouse cancer cells have been shown to induce necroptosis of endothelial cells in coculture and orthotopic lung metastasis models, whereas inhibition of RIPK1/3-dependent endothelial necroptosis reduced extravasation and metastasis in these models (40).

Following extravasation, metastatic cells remain situated near blood capillaries and migrate along the abluminal surface on the capillaries (36, 41–44). Integrins mediate interactions between cancer cells and capillaries (45). This is most apparent in experimental models of metastasis to the brain, where the elongated nature of capillaries facilitates imaging studies (41, 43). However, perivascular residence of extravasated cancer cells is also observed in other sites, including the liver, lungs, and bone marrow (43).

Cancer cells disseminate via the lymphatic circulation, but whether lymph nodes are an obligatory staging post for metastasis remains controversial and likely depends on tumor type (46, 47). Comparative genomics of regional lymph node and distant metastases from colorectal and breast cancers suggest that in the majority of cases, distinct clones within primary tumors give rise to nodal and distant metastases, with the latter undergoing more stringent clonal selection (48–50). In patients, highly aggressive triple-negative breast cancers often metastasize without apparent regional lymph node involvement (47). Consistent with a correlative, not causal, relationship between lymph node and distant metastasis, recent clinical trials have shown no survival benefit for regional lymph node dissection in patients with breast and ovarian cancers and melanoma (51–53). Even when injected directly into the lymph nodes of mice, cancer cells preferentially metastasize via the blood rather than through the lymphatics (54, 55). However, transient passage through the lymphatics may confer long-term survival benefits to MICs as they join the blood circulation (56).

Melanomas implanted in mice lacking dermal lymphatic vessels exhibited lower cytokine expression and leukocyte infiltration than those implanted in wild-type controls, suggesting that cancer cell interactions with immune cells in lymphatic structures are important for antitumor immunity (57). The distinctive metabolic milieu of lymph may also enable cancer cell adaptation to the stresses of metastasis. For example, exposure to oleic acid in lymph protectedmelanoma cells from ferroptosis and increased their ability to initiate metastasis during subsequent hematogenous dissemination (58).

The metastatic spread of some tumors occurs largely or at least partly through routes that obviate the need to enter and exit the circulation (Fig. 2D). In ovarian cancer, cell clusters carried by the peritoneal fluid attach on the abdominal peritoneum where the cells grow (59). Peritoneal metastasis is also frequent in colorectal and pancreatic cancers. Melanomas and other cancers can spread without entering and exiting the circulation by engaging in extravascular migratory metastasis, a process of migration over the abluminal surface of vessels that is common during embryogenesis (60). Perineural spread is a long recognized, if underappreciated, form of metastatic spread in head and neck cancers and other squamous cell carcinomas (61). Regional spread of lung adenocarcinoma can occur by spread through air spaces where cancer cells detach from a primary tumor, migrate as clusters through air spaces, and reattach to the alveolar walls through vessel co-option to resume tumor growth, which is associated with tumor recurrence and poor survival of stage I lung adenocarcinomas (62, 63).

Dormant metastasis and the ability to predict the likelihood of an eventual outbreak in patients diagnosed with early-stage cancer are major considerations in the clinic (64). For example, patients with estrogen receptor–positive early-stage breast cancer who undergo complete surgical removal of the primary tumor continue to experience metastatic relapse as late as 20 years from their original diagnosis (64). Dormant metastasis is an understudied problem, owing to a dearth of appropriate experimental models. Models using aggressive metastatic cells sidestep the dormancy phase altogether. Some models of dormant metastasis rely on comparisons between nonisogenic cancer cell lines of varying propensity to develop metastatic colonies after inoculation in mice. Spontaneously arising models of dormancy have been characterized (9). New experimental models have been developed by in vivo selection of early-stage tumor-derived MICs that are competent to establish dormant metastasis and spontaneously generate late outbreaks, thus mimicking the course of the disease in patients (65). Studies on these various models have begun to shed light on the biology of MIC entry and exit from dormancy (Fig. 3).

Clinical dormancy implies one of two complementary scenarios: either a cell-autonomous “temporary mitotic exit” as originally advanced by Geoffrey Hadfield in 1954 (66) or, alternatively, the existence of persistent attempts at colonization that are aborted by cell-extrinsic limitations on metastatic outgrowth. Although several studies have explored the niches, signals, and mechanisms that maintain both modes of dormancy (reviewed in refs. 9, 67, and 68), multiple lines of experimental and clinical data are converging on the view that dormancy reflects a dynamic equilibrium of MICs with antagonistic immune surveillance (Fig. 3A).

Cancer cells suffer extensive attrition after infiltrating distant organs, as shown in mouse models of metastasis and as inferred from the low counts of disseminated cancer cells relative to the number of CTCs that a primary tumor likely shed while it was present in a patient. Elimination of cancer cells after extravasation is driven by passive factors such as high oxidative stress in the lungs (69, 70), a dearth of growth factors and nutrients in the cerebrospinal fluid (38, 71), or, more generally, a lack of the appropriate microenvironment cues that had nursed the growth of MICs in the primary tumor. These passive mechanisms of disseminated cancer cell elimination are accompanied by the proactive role of tissue-resident immunity and other defenses, such as astrocytes in the brain (36), against intruding cancer cells. Tissue-resident macrophages, such as microglia in the brain and Kupffer cells in the liver, and NK cells attack infiltrated cancer cells (18, 72, 73). In experimental models, depletion of T cells and NK cells increases the survival and outgrowth of disseminated cancer cells (66, 74, 75). Tumor lymphocyte infiltration, indexed in the form of an “immunoscore,” is inversely correlated with metastatic burden in patients (76), and metastatic clones can show branched evolution as a result of different immunoediting pressures (77). Dramatically, immunosuppressed recipients of transplant organs from donors deemed cured of localized melanoma or glioblastoma developed donor-derived metastasis (78, 79). Underscoring the importance of immune surveillance of MICs, in patients with completely resected localized melanomas, postsurgical adjuvant therapy with immune-checkpoint inhibitors to boost antitumor immunity reduced later metastatic relapse and prolonged survival (80–83).

Disseminated MICs that enter proliferative quiescence may be particularly equipped to survive these threats. Quiescence allows cancer cells to evade the innate and adaptive arms of immunity mediated by NK and T cells, respectively (65, 84–87). As an added benefit, dormancy renders disseminated cancer cells resistant to antimitotic therapy. The ability to enter a slow-cycling state is a property of some adult stem cells. Thus, to the extent that MICs are primed to enter protective dormancy, dormancy will enrich for MICs in disseminated cancer cell populations.

Entry of MICs into quiescence is stimulated by growth-inhibitory signals in the host microenvironment (Fig. 3A). TGFβ has cytostatic effects on normal epithelial cells and carcinoma cells by inducing the expression of cyclin-dependent kinase inhibitors, whereas it triggers tumor-suppressive apoptosis in premalignant epithelial cells. Carcinoma cells bypass the proapoptotic effect of TGFβ through various mechanisms while retaining the ability to respond to TGFβ with prometastatic effects, including entry into immunoprotective quiescence (88, 89). TGFβ has been shown to stimulate dormancy in experimental models of prostate, breast, and head and neck carcinomas (9). The three isoforms, TGFβ1, TGFβ2, and TGFβ3, are known to be produced in tumor microenvironments by various stromal components and cancer cells. Latent TGFβ1 in particular is present in perivascular niches of the bone marrow and brain parenchyma of mice and is activated by thrombospondin (TSP) to drive extravasated breast cancer cells into proliferative quiescence (ref. 44; Fig. 3A). In this context, TGFβ likely functions through canonical binding to the paired receptor kinases TGFBR1 and TGFBR2 with betaglycan, also known as type III TGFβ receptor, serving as a coreceptor that is particularly important for TGFβ2 binding (88).

Entry of disseminated MICs into dormancy is not merely a passive process resulting from the effect of local growth-inhibitory signals from the tumor microenvironment. Entry into dormancy can also be initiated by MICs themselves such that MICs that actively avoid mitogenic stimulation are selected for because quiescence enables immune evasion (65). MICs derived from early-stage lung and breast tumors actively secrete DKK1, an inhibitor of WNT factors that are potent mitogens for adult stem cells (Fig. 3A). This autocrine DKK1 enforces MIC dormancy in these models, and depleting DKK1 releases MICs from dormancy and delivers them to NK-mediated clearance (65).

These insights suggest that dormant metastasis results from a dynamic equilibrium in which MICs enter proliferative quiescence and thus avoid immune-mediated killing or enter the cell cycle for tumor growth at the risk of being eliminated (Fig. 3A). Another form of dormancy is tumor mass dormancy in which MICs start proliferating but are unable to satisfy the nutrient or oxygen requirements for growth beyond a small tumor mass due to lack of angiogenesis or other causes (9, 90). However, large clinical trials have shown little success with antiangiogenic drugs in the adjuvant or neoadjuvant setting when targeting latent micrometatases, suggesting that (with the limitations of the drugs available) the angiogenic switch may not be a major limiting factor to metastatic outgrowth in patients, or that current therapies are inadequate to suppress angiogenesis (91).

Little is known about the signals that lead disseminated cancer cells to exit from dormancy and develop clinically detectable metastases. It is possible that disseminated cancer cells are “awakened” from years of quiescence by signals that jolt them into entering the cell cycle. However, the cited studies on depleting mice of immunity suggest that there is always a fraction of the disseminated cancer cell population attempting to proliferate, perhaps due to a growth-permitting balance between local mitogenic and antimitogenic signals (Fig. 3A). In line with this idea, depletion of DKK1 in models of dormant lung adenocarcinoma and breast cancer metastasis enables exit from dormancy, implying that WNT ligands in the perivascular niche reactivate the proliferation of dormant cells (65). Bone morphogenetic proteins can trigger differentiation of disseminated breast cancer cells in mouse models, which is prevented by BMP sequestering factors (92). Cell metabolism may also play a role. Cell autophagy in response to metabolic stress promotes the survival of dormant breast cancer cells in mouse models (93), and depleting autophagy related 3 (ATG3) or p62/sequestosome-1 promotes cancer cell exit from dormancy with expression of 6-phosphofructo-2-kinase, which supports glycolysis (94).

MICs have certain requirements for continued growth during exit from dormancy. MICs derived from breast, lung, colorectal, and renal cell carcinomas express L1CAM and use this cell adhesion molecule to bind to laminin in perivascular basement membranes, spreading on the abluminal surface of capillaries (Fig. 3B). Once engaged, L1CAM cooperates with β1 integrins, the integrin–actin bridging protein ILK, and p21-activating kinases (PAK) to activate transcription factors YAP and MRTF for the initiation of metastatic outgrowth (32, 43). ILK and β-parvin regulate actin filaments supporting the formation of filopodia-like protrusions, which are required for metastatic outgrowth (95). The L1CAM-expressing MICs grow forming sheaths around the vessels, engulfing the capillary network before the micrometastatic mass acquires a spheroidal shape (32, 43). The requirement for L1CAM has been shown in experimental models of dormant MIC outgrowth following depletion of NK cells as well as models of aggressive metastatic outgrowth without an intervening dormancy phase (43).

Recent studies have revealed a role for neutrophil extracellular traps (NET) in promoting YAP activation during cancer cell exit from dormancy (96, 97). NETs are protease-loaded chromatin webs that neutrophils release as an ultimate maneuver to entrap and clear pathogens (98). NETs are of interest for their roles in inflammation, autoimmune diseases, and cancer (99). Studies using intravital microscopy documented the presence of NETs around disseminated breast cancer cells in mouse models of metastasis and an inhibitory effect of DNAse I on metastatic growth, implying that cancer cell–triggered NETs promote metastasis (100). Indeed, NETs produced during inflammatory responses were shown to reactivate the growth of local dormant cancer cells through cleavage of basement membrane laminin by NET-associated elastase and matrix metalloproteinase 9 (MMP9; Fig. 3B). The cleaved laminin binds α3β1 integrin to activate YAP and proliferation of the cancer cells (97).

Colonization of distant organs is the most dramatic phase of the metastatic process clinically and the most complex biologically. The outgrowth of disseminated MICs into large metastatic colonies depends on the successful avoidance of organ-specific barriers and the productive interactions with specific components of the host tissue ecosystem. Many of these barriers and interactions vary with the site of metastasis. The lung, liver, or brain parenchyma, the lymph nodes, bone marrow, and leptomeninges present widely different tissue microenvironments. Not surprisingly, studies on metastatic colonization of these major target sites have uncovered different factors, metabolic changes, cell–cell interactions, and other determinants of metastatic colonization depending on the tumor type and the metastasis organ site. Detailed accounts of organ-specific metastatic colonization factors have recently been provided for metastasis to different organs (101, 102), and specifically to bone (103), central nervous system (39, 104), lung (105), and liver (106). Here we briefly summarize current concepts on general and organ-specific colonization processes.

Metastatic tropism is manifest in the tendency of tumors to preferentially relapse in certain organs. Some cancers (e.g., small cell lung carcinomas and metastatic melanomas) form metastasis in multiple sites without marked organ preference, whereas other cancers preferentially relapse in certain organs (e.g., the bones in prostate and hormone receptor–positive breast carcinomas, the liver in colorectal carcinomas and ocular melanomas, or the lungs in soft-tissue sarcomas).

Metastatic tropism results from the combination of two distinct phenomena. One is the ability of cancer cells to reach particular organs through the circulation. Colon cancer is thought to predominantly metastasize to the liver because the mesenteric blood draining the intestines flows via the hepatic portal vein into the liver. The majority of CTCs derived from an intestinal tumor become trapped in the hepatic sinusoids and initiate metastases in the liver with far higher frequency than in the lungs, brain, or bones. It has been proposed that cells derived from certain tumors use organ-specific endothelium adhesion molecules to concentrate in the capillaries of those organs and preferentially extravasate there (107). However, experimental (65, 108, 109) and clinical (110) evidence suggests that, in most types of cancers, disseminated cells reach all organs and leave dormant cells everywhere. This suggests that metastasis develops from latent surviving MICs in the least antagonistic site first, and in other sites later. Thus, the metastatic organ tropism of a particular type of cancer is a function of the organs that MICs can travel to and, most critically, the probability that MICs will grow in a particular organ site depending on their intrinsic or acquired ability to sort the barriers and utilize the supportive niches in that site.

As we noted earlier, differences in the permeability of capillary beds in different organs determine the ability of disseminated cancer cells to accumulate in these organs. Several mediators of cancer cell extravasation have been identified in experimental model systems, including cytokines (e.g., ANGPTL4 and HBEGF), secreted enzymes (e.g., cathepsin S), and monocytes that facilitate extravasation in the lung capillary walls (1, 2, 13) or through the blood–brain barrier (refs. 37–39, 111; Fig. 2C). Metastatic colon cancer cells that colonize the liver can gain lung metastatic ability by expressing the cytokine PTHrP, which mediates extravasation of these cells in the lungs after they reenter the circulation (112). In some instances, metastatic cells disrupt endothelial barriers to allow the influx of nutrients from the circulation. Breast and lung carcinoma cells that infiltrate the leptomeninges produce complement component 3 to activate the C3a receptor in the choroid plexus epithelium and disrupt the blood–CSF barrier, thereby allowing influx of growth factors and nutrients from blood into the nutritionally barren CSF (38).

The important role of immunity in suppressing the outgrowth of disseminated MICs can give rise to organ-specific differences depending on the ability of MICs from different types of cancers to evade tissue-resident immunity (18, 72, 73). The ability of disseminated cancer cells to regulate their own exposure to immune attack by modifying the expression of MHC class I molecules (113), NK-cell and effector T-cell ligands (114), cytosolic DNA-sensing pathways (115), and damage-associated RNAs as ligands for pattern-recognition receptors (116) is under study given their likely importance in metastatic colonization. Moreover, the defenses that invading cancer cells must neutralize to seed distant organs are not limited to classic immunity. Breast cancer and lung adenocarcinoma cells that infiltrate the brain must contend with the release of proapoptotic FAS ligand by reactive astrocytes. FAS ligand is released from its membrane-bound precursor in astrocytes by the protease plasmin, and cancer cells can prevent this by expressing serpin inhibitors of the plasminogen activator (36).

Metabolic adaptation is an important aspect of metastasis. Unlike glycolytic primary tumors, MICs in multiple cancer types undergo a bioenergetic shift into an oxidative phosphorylation state (69, 117, 118). Organ-specific metabolic adaptations enable carcinoma MICs to reinitiate growth in tissues with variable bioavailability of oxygen, including in the lungs, where high oxygen levels cause oxidative stress that cancer cells can withstand by overexpressing NRF2 (refs. 118–120; Fig. 3C). Metastatic melanoma cells also undergo metabolic adaptations to withstand high levels of oxidative stress in the blood and lungs (69). Adaptations to the bioavailability of amino acids (121, 122), nucleotides (123), and other bioenergetic substrates (124, 125) have also been noted in metastasis models and human tissue samples and are likely to vary depending on the host tissue. In turn, metabolic stresses may shape the genomic and epigenomic reprogramming of cancer cells to enable metastasis (126). Other cancer cell adaptations serve to source scarce nutrients. For example, metastatic cells secrete the iron-chelating factor lipocalin A to scavenge scarce iron in the CSF (71).

Lipids are important for tumor growth. CD36, a scavenger receptor that can transport fatty acids and other molecules, is upregulated in MICs from oral carcinomas and other cancers and required for metastasis in various organs (127). Brain metastasis appears to place higher demands on MICs with regard to procuring their own lipids. The brain has a markedly lower level of triacylglycerol stores than other tissues. Breast cancer cell lines that are metastatic to the brain upregulate sterol response element binding transcription factor 1 (SREBF1) to promote lipid synthesis and fatty acid metabolism or alternatively upregulate CD36 for metastasis growth (128).

MICs that survive the organ-specific stresses described above additionally require support from the host stroma. Stromal support is thought to be important for MICs immediately after their infiltration of new tissues and establishment as long-term seeds for metastases. This may depend on a combination of cellular and extracellular cues that fulfills the survival and self-renewal requirements of MICs by mimicking the native niches that these cells or their predecessors occupied in their tissue of origin. Various cellular and extracellular components of natural stem cell niches support the viability of disseminated MICs (3, 5), including the ECM interacting components periostin and tenascin C, which cancer cells and fibroblasts produce to promote WNT and Notch signaling in MICs (refs. 129, 130; Fig. 3B). Evidence that disseminated cancer cells intrude into natural stem cell niches in the bone marrow has been provided in models of prostate (131) and breast cancer metastasis (132).

In addition to biochemical signals from the niche, biophysical and mechanical features of stromal niches can also promote metastasis (133). An expanding primary tumor mass exerts mechanical force on its tissue microenvironment, in turn triggering differentiation of myofibroblasts to produce extracellular proteins such as collagens and tenascin C (134), which further increase stiffness of the ECM. In turn, increased ECM stiffness can promote integrin-mediated focal adhesion assembly, RHOA–ROCK-mediated cell contractility, and mechanosignaling via the TGFβ, WNT, and YAP pathways (reviewed in ref. 133). Stiffness-mediated mechanosignaling may thus support MIC activities, including motility, anoikis evasion, and EMT. Consistently, in patients with breast and pancreatic cancers, tumors with increased ECM stiffness were associated with increased metastasis, therapy resistance, and poor prognosis (135–137).

Support from the stroma is also essential during the expansion of macrometastatic colonies, maintenance or aggressive metastatic lesions, and resistance to therapy. Most parenchymal components in sites of metastasis have been implicated in the support of survival, outgrowth, and/or chemoresistance of MICs through unidirectional or reciprocal signaling interactions with the cancer cells. This includes osteoblasts and osteoclasts in bone metastasis (101, 103, 138–140); astrocytes (141, 142), microglia (143), and neurons (144) in brain metastasis (39, 104); monocytes (145, 146), neutrophils (97, 129, 147), fibroblasts (130, 148), and epithelial cells (149) in lung metastasis (105); and mesenchymal, epithelial, macrophage, and other cells in liver metastasis (106). Cytokines and enzymes released by a primary tumor in soluble or vesicular form can alter the parenchyma of distant organs and establish “premetastatic niches,” zones that favor the immediate growth of cancer cells reaching these sites (150). It is unclear if premetastatic niches established by a primary tumor could support MICs as dormant seeds for long periods after a primary tumor is surgically removed.

Most studies on stromal interactions with metastatic cells have utilized experimental models in which aggressive cancer cells proceed to form metastatic colonies without an intervening dormancy phase. Therefore, with few exceptions, these studies do not directly distinguish whether a supportive interaction benefits MICs prior to dormancy, during the incipient outgrowth of micrometastasis after dormancy, or during the continued expansion of macrometastases. In spite of this shortcoming, several organ-specific metastatic mediators identified to date are accompanied with evidence of clinical relevance, providing potential targets for therapy in the treatment of metastasis.

EMT, referred to in passing above, is a key process during tumor progression and metastasis (1). Epithelial cells undergoing an EMT lose polarity and downregulate key cell-adhesion molecules, resulting in the loss of contacts with neighboring cells and an increased ability to migrate and invade adjacent tissue. The ability to undergo an EMT is a property of epithelial stem and progenitor cells, both normal and neoplastic, and reflects an intrinsic phenotypic plasticity of these cell states. EMTs play important roles in development and epithelial regeneration. During gastrulation, epiblast cells undergo an EMT to migrate to appropriate locations as they differentiate into mesoderm and endoderm progenitors. EMTs are also important in later morphogenic events and during wound healing in the adult (17, 21, 151). Cells that undergo an EMT return to an epithelial state through a mesenchymal-to-epithelial transition (MET; refs. 17, 21, 151). The regained epithelial state after an EMT–MET sequence may be different from the starting one, as it occurs during gastrulation.

In normal and neoplastic cells, EMTs are driven by SNAIL, SLUG, ZEB, TWIST, and other transcription factors (referred to as EMT-TF) that repress epithelial genes and stimulate the expression of mesenchymal components, together with micro-RNAs that balance this regulatory network (21). Stromal signals that trigger EMTs include TGFβ, WNT, and NFκB activators (17, 21, 151). TGFβ induces the expression of EMT-TFs in cooperation with RAS–MAPK signaling; TGFβ-activated SMAD transcription factors and MAPK-activated RREB1 jointly drive expression of SNAIL and SLUG coupled to cell type–dependent differentiation or fibrogenesis gene sets, in both normal and neoplastic cells (152). WNT-activated TCF transcription factors and NFκB transcription factor also converge on EMT-TF gene enhancers to regulate EMTs (21).

Carcinoma cells at the tumor invasion front undergoing an EMT can drive collective migrating by pulling cohesive ensembles of cancer cells that remain connected by cell–cell contacts (153). After efficiently disseminating and infiltrating distant organs using the mesenchymal properties acquired with the EMT, disseminated MICs undergo an MET to reinitiate tumor growth (154, 155). In human and mouse carcinomas, loss of cell-surface E-cadherin during EMT activates prosurvival signaling that is necessary in the detached state, but E-cadherin reexpression is required for the outgrowth of metastatic colonies (34). What signals trigger MET in disseminated MICs, and whether MET occurs before or after MICs pass through dormancy, remain open questions.

Epithelial and mesenchymal states are now viewed not as discrete endpoints but as a continuum of highly fluid, interconverting phenotypic states. Each particular epithelial progenitor cell type undergoes its own form of EMT, involving a distinct combination of EMT-TFs and different changes in cell morphology, adhesion, and motility (17). The extent of mesenchymal conversion achieved by an epithelial progenitor during an EMT also depends on the developmental state of this cell and reflects the phenotypic space that this state is programmed to sample. Many epithelial progenitors undergo what is referred to as a “partial” or “hybrid” EMT. However, the current trend is to consider every departure of epithelial cells from the polarized phenotype an EMT regardless of how far toward a stereotypical (“full”) mesenchymal phenotype this departure may reach (17). Although EMT is a common mechanism for metastatic dissemination, the cells that are most competent at forming metastasis are not necessarily those that adopt the most mesenchymal-like phenotype in their EMT (152, 156, 157). Some forms of dissemination, such as collective migration, may enable cancer cells to retain epithelial identity and intercellular contacts during migration and may not require an EMT.

Cancer cells require a large number of traits to perform each step of the metastatic process and additionally survive therapy if present. Some of the traits required to become a MIC are imparted by the basic tumor-initiating phenotype resulting from genetic alterations that activate driver oncogenes and that disrupt tumor suppressor genes. In addition to uncontrolled survival and proliferation functions, these traits include the ability to migrate, invade, and remodel ECM, all of which are conducive to metastasis. The basic traits or “hallmarks” of cancer included with the neoplastic phenotype remain essential throughout the course of tumor progression (158). Moreover, growing evidence indicates that cancer is initiated by oncogenic mutations in stem cells whose normal function is to maintain tissue homeostasis. Turned into CSCs, these cells retain self-renewal capacity, or “stemness,” which is essential for long-term tumor growth (4).

But even with these powerful traits in place, a vast majority of cancer cells leaving a tumor fail to form distant metastasis, instead succumbing to the stresses of the metastatic process. The many obstacles that MICs must sort on their way to establishing metastasis imply a need for additional traits beyond those provided by the CSC phenotype alone. Indeed, work over the past two decades has identified a large number of “metastasis genes” whose expression (or repression) enhances the ability of cancer cells to enter the circulation, extravasate, avert stromal attack, resist metabolic stress, enter dormancy, evade immune surveillance, reinitiate tumor growth, secure a supportive niche, and co-opt organ-specific stromal components (2, 5, 101).

Adaptability is another key trait of the MIC phenotype. Disseminated cancer cells must adapt to a succession of rapidly changing physical, metabolic, and immune challenges. To survive these stresses, MICs adjust their phenotype accordingly. Phenotypic plasticity, the ability of cells to dynamically change their gene-expression programs, is an overarching hallmark of tumor progression and metastasis, and a source of adaptability for MICs (159, 160). Phenotypic plasticity enables cancer cells to undergo EMT, adapt to specific microenvironments, and resist therapy (161, 162). Moreover, plasticity accentuates the phenotypic heterogeneity of genomically diverse cancer cell populations (163).

In sum, MICs acquire different classes of traits from different sources (Fig. 4A): basic prometastatic traits acquired with the neoplastic phenotype, additional specialized traits acquired under selection for tumor regeneration in specific organs, and a plastic phenotype adept at deploying these traits under rapidly changing conditions as MICs attempt to regenerate the tumor in new locations.

How do MICs acquire their unique adaptability and other specialized traits? Heterogeneous cancer cell populations evolve under continual selective pressures during tumor progression and metastasis. The identification of large numbers of tumor-associated mutations by large-scale genomic sequencing of matched normal and primary tumor tissues provides a basis for the notion that metastasis-specific traits may emerge by selection from the genetic heterogeneity present in cancer cell populations. However, the search for mutations that specifically drive metastasis has met with few examples to date.

Studies seeking to identify metastasis-specific mutations pose several challenges (164, 165). Metastasis typically evolves in patients over several years, and this temporal dimension cannot be adequately reproduced in mouse models. On the other hand, patient-based studies are logistically cumbersome, complicated by interpatient heterogeneity, and subject to sampling bias if only a small region of a large tumor is biopsied for genomic analysis (166). Therapy is another important confounder, because patients who experience metastatic relapse after initial removal of primary tumors will have typically received systemic anticancer therapies, and such metastases reflect therapy resistance and may not represent the natural history of metastasis per se. Despite these hurdles, several recent studies involving multiregion sampling of matched primary and metastatic cancers biopsied from large numbers of patients have illuminated the genetics of metastatic evolution (162, 165, 167–172). Notably, across tumors, there appear to be few recurrent metastasis-specific somatic mutations, including single-nucleotide variants (SNV), insertions, and deletions. Where such mutations have been identified, their functional significance is frequently linked to expanding the activity of the oncogenic driving pathway or conferring resistance to therapy (166, 168, 173, 174). However, activation of certain general growth-promoting pathways may translate into organ-specific metastatic advantage. For example, mutations in PIK3CA have been proposed to favor brain metastasis of breast cancer by driving lipid synthesis (128).

The association of genetic mutations with metastatic progression reflects a selection of prometastatic traits or an association of bystander mutational events with clones that gain a prometastatic phenotype by other means. A recent study of 1,421 samples from 394 tumors across 22 tumor types identified increased clonality of somatic copy-number alterations (SCNA) in metastases versus primary tumors, potentially indicating an as yet inadequately understood role for large-scale genomic rearrangements in driving metastasis (175). In non–small cell lung cancer, elevated copy-number heterogeneity across multiple sequenced regions of early-stage primary tumors was associated with increased risk of relapse (176), suggesting that ongoing chromosomal instability is an important mechanism to generate evolutionary space for the emergence of prometastatic traits within tumors.

Computational analyses of SCNAs and SNVs enable reconstruction of phylogenetic trees to determine the sequence and selection of mutations during metastasis (177). Two models for metastatic evolution have long been considered (178). The “linear evolution” model envisions that primary tumors gradually accumulate mutations, with a late subclone acquiring metastatic capacity (Fig. 4B). The “parallel evolution” model, on the other hand, argues that cancer cells that disseminate early are more effective at forming metastasis because they evolve in the host microenvironment, unlike clones that remain in the primary tumor (179). In HER2⁺ mouse models of metastatic breast cancer, early disseminating cells were found to be more metastatically competent than late disseminators (180, 181). Using sensitive cell-free DNA technology, circulating DNA with tumor-associated mutations can be detected in patients with subclinical primary tumors in a variety of cancers, suggesting that early dissemination is possible. However, the rarity of metastasis-specific SNVs in patients argues against the parallel evolution model (165, 167). Other studies have identified both metastasis-specific recurrent alternations and alterations that are shared between primary and metastatic tumors, suggesting that both early and late metastatic dissemination can occur (162). Moreover, recent work suggests the presence of “punctuated evolution” in some tumors, with acquisition of mutations that rapidly clonally dominate the primary tumor and then seed metastasis in an evolutionary “Big Bang” (169, 182–185). Other tumors may evolve more gradually, with multiple clones of low metastatic fitness competing within the primary tumor.

Taken together, the evidence from clinical and murine studies suggests that metastasis imposes an evolutionary bottleneck that only select cells within primary cancer cell populations are able to overcome. Such metastases may be monoclonal or polyclonal (169, 182–187), suggesting potentially distinct modes of seeding or selection. It is worth noting that the prevalent linear interpretation of the clonality and timing of metastases may be confounded by the possibility of ongoing “self-seeding” of metastases, wherein cells from metastatic tumors not only may recirculate and seed new metastatic lesions, but may also infiltrate preexisting primary and metastatic tumors (188, 189).

Intriguingly, in clear cell renal cell carcinoma, dominant clones resulting from punctuated evolution seed aggressive multiorgan metastasis with little intermetastatic heterogeneity, whereas oligometastases with indolent behavior arise from tumors displaying gradual evolution (ref. 184; Fig. 4B). Unlike multiorgan metastasis, oligometastasis in a single organ may be amenable to curative surgery or other local treatments, and thus provides an opportunity to use tumor clonality as a predictive biomarker for such therapies (190). In patients, there is a strong correlation between tumor size (a proxy for tumor age) and the probability of future metastatic relapse, which is inconsistent with the possibility of exclusive early metastatic seeding. However, in a large cohort of patients with breast cancer treated with mastectomy and no systemic therapy, tumor size was associated with early local recurrence and distant metastasis, but not with late metastasis (191). This observation suggests the possibility that in some tumors both early and late dissemination may coexist, with late, more proliferative disseminating cells driving mostly rapid relapse. In contrast, early disseminating cells may be particularly prone to latency for extended periods of time, either because of a capacity to enter a deep quiescent state or because those that grow are rapidly cleared by the immune system.

Although genomic alterations generate selectable survival traits in cancer cells, mutations alone cannot account for the fundamentally distinct biology of MICs. Unlike primary tumors, which are thought to largely arise through oncogenic transformation of homeostatic tissue stem and progenitor cells in their native niche (192–196), MICs must rapidly change their phenotypes to detach from the primary tumor, survive mechanical, metabolic, and immunologic stresses during dissemination to distant organs, adapt to the foreign microenvironments, access proliferative cues to reinitiate tumor growth in distant organs, and resist therapy. Dynamic adaptation to changing circumstances during the metastatic process requires faster change than a purely genetic mechanism, which allows for selection of variants only once per cell cycle, might support.

Although the dynamic expression of specific genes has long been associated with metastatic phenotypes, the master transcription factors and epigenetic remodeling mechanisms required for the different steps of the metastatic process in various tumors are only beginning to emerge (197). Genes required for chromatin modification (e.g., ARID1A/B and KMT2C/D) and DNA methylation (DNMT) are frequently mutated in advanced tumors, underscoring the importance of epigenetic modulation in metastasis (171). A fundamental question is whether metastatic adaptability involves greater overall chromatin accessibility, and hence the potential to activate the expression of multiple genes, as implied by studies in genetically engineered mouse models of lung cancer (198–200), or whether the ability to transition to one or more prometastatic phenotypic states by activation of specific transcription factors is sufficient, as shown in models of pancreatic ductal adenocarcinoma (201) and renal cell carcinoma (202). Recent work illuminated the role of histone H3 variants (203) and specific histone marks, such as H3K36me2 (204), as critical requirements for epigenomic reprogramming during metastasis. The question remains whether epigenetic alterations favor metastasis by randomly activating selectable genes or by enabling cancer cells to enter progenitor states that have a superior fitness to initiate metastasis.

Several key traits of the MIC phenotype are normally present in regenerative progenitor cells. In wound-healing processes (Fig. 5A), epithelial regenerative progenitors show intense phenotypic plasticity, undergo EMT, display migratory activity, interact with and remodel the ECM, and secrete signals that co-opt mesenchymal, endothelial, and immune stromal components to support the restoration of epithelial barrier integrity (205, 206). Recent work raises the possibility that carcinoma MICs may capture these traits all at once by adopting a regenerative phenotype.

Phenotypic plasticity, a key feature of tumor progression and metastasis (159, 160), is also a feature of normal adult stem cells, including those of the gut, lung, and skin. These cells are not locked in a fixed preexisting cell fate but rather enter and exit “stem-like” states in response to demands to preserve the structural and functional integrity of the tissue (4, 160, 207–209). In the intestinal mucosa, cells expressing the WNT signaling pathway component LGR5 function as stem cells in the maintenance of epithelial homeostasis (210). However, when LGR5⁺ cells are lost by genetic ablation or injury, LGR5⁻ progenitor cells can become LGR5⁺ cells (206, 209). Moreover, unique stem cell states are mobilized by damage to revive the epithelium through transcription factor YAP (211).

Cells from the mouse small intestine can grow as organoids, three-dimensional structures that form in laminin-rich basement membrane matrix and mimic the cellular composition and architecture of the intestinal epithelium (212). Both LGR5⁺ and LGR5⁻ cells sorted from normal mouse intestinal epithelium can form organoids (209, 213). Moreover, during the initial stages of organoid formation, LGR5⁺ cells convert to an LGR5⁻ state that includes YAP activation, and later reexpress LGR5 and downregulate YAP to establish a differentiated epithelium (214). Thus, intestinal organoid formation appears to start as a process of epithelium recovery driven by LGR5⁻ progenitors followed by restoration of differentiated structures driven by reemerging LGR5⁺ progenitors.

Dynamic phenotypic changes are also observed in cancer metastasis. In colorectal cancer, intestinal LGR5⁺ stem cells harboring constitutively activating mutations in the WNT signaling pathway act as tumor-initiating CSCs (195). However, intravital microscopy of experimental colorectal cancer metastasis in mice revealed that MICs disseminating from the primary tumor and seeding metastasis in the liver are predominantly LGR5⁻ (214). To initiate metastatic tumor growth in the liver, some progeny of LGR5⁻ cells reacquire LGR5 expression (214), whereas other cancers may acquire LGR5-independent regenerative capacity (215, 216).

Recent work has shown that MICs in colorectal cancer and other epithelial cancers dynamically express L1CAM (32). Originally identified as a neuronal cell-adhesion molecule, L1CAM is expressed in carcinomas at the primary tumor-invasion front and is associated with poor prognosis in multiple tumor types, including those of the breast, lung, colon, ovary, and kidney (217). L1CAM is not expressed in intact normal epithelia or required for epithelial homeostasis, but its expression is induced by the breach of epithelial intercellular adherens junctions during colitis and tissue injury, and is required for the survival of detached epithelial progenitors and for tissue repair (32). L1CAM expression in dissociated colorectal cancer cells is required for the initiation of organoid formation, where most L1CAM⁺ cells are LGR5⁻. L1CAM⁺ cells are highly competent to initiate growth of transplanted colorectal cancer tumors and liver metastases. As MICs initiate tumor growth and regain epithelial structures, L1CAM expression is downregulated, but can be reactivated again when tumor architecture is disrupted by therapy. Repeated cycles of therapy select for cells with the capacity to upregulate L1CAM expression, which can more efficiently reinitiate tumor growth (32). Thus, L1CAM seems to mark certain regenerative progenitor states, and its expression in these cells mediates the regrowth of intestinal epithelium after injury and tumor regrowth during metastasis.

Several studies highlight a biphasic process of epithelial repair that is co-opted during metastasis. Upon wounding, an immediate emergency response is activated in which epithelial progenitors enter distinct migratory and proliferate states to seal the breach in the mucosa (refs. 218, 219; Fig. 5A). This “band-aid” phase is followed by a slower regenerative phase in which normal tissue stem/progenitor hierarchies are recreated. Borrowing from repair, dissociation of epithelial structures in invasive colorectal cancer induces malignant progenitors to adopt a highly plastic regenerative phenotype endowed with the capacity to undergo EMT, migrate, avert anoikis, and drive L1CAM-dependent regrowth (refs. 32, 43; Fig. 5B).

Further evidence for the regenerative nature of metastatic progression comes from single-cell RNA-sequencing studies on tumor tissue from patients with lung adenocarcinoma as well as lesions from a mouse model of metastatic latency and outbreak (220). Whereas primary tumors showed developmental lineage promiscuity and cell states that were intermediate between the various cell states of the regenerative lung epithelium, metastatic cells reverted to earlier states and recapitulated lung organogenesis and regeneration phenotypes with striking fidelity.

The adoption of a regenerative phenotype by MICs has several implications. Instead of having to select a predetermined rigid MIC fate or randomly acquiring a multitude of independent metastatic traits, colorectal cancer cells seem to sample among a range of phenotypically transient states, some of which, such as the L1CAM⁺ state, are adept at supporting survival of disseminated cells and the initiation of metastatic growth and are hence selected for during the metastatic process (32, 43).

The highly specialized nature of most organ-specific metastatic traits suggests that these are acquired through selection from heterogeneous MIC populations that complete their evolution after seeding distant organs, rather than being intrinsic properties of the stem cell states that give rise to MICs. However, not all organ-specific traits result from clonal evolution after dissemination. The microenvironment of a primary tumor can preselect clones with properties that are compatible with the environment of a particular distant organ, thus influencing the metastatic tropism of the disseminating cancer cells. For example, primary breast tumors with high TGFβ activity are associated with lung relapse, and this is linked to the ability of TGFβ to induce the expression of the cytokine ANGPTL4 in the cancer cells. ANGPTL4 causes endothelium disjunction to promote extravasation of circulating breast cancer cells that lodge in lung capillaries (221). Primary tumors may also preselect for MIC clones that thrive in distant organs whose stroma the primary tumor mimics. For example, breast carcinomas rich in mesenchymal cells that produce CXCL12 select for MIC clones that thrive in a CXCL12-rich microenvironment, and thus are predestined to colonize the CXCL12-rich stem cell niches of the bone marrow (222).

Metastasis is subject to important organ-specific barriers and adaptation, as mentioned in previous sections and recapitulated in Fig. 6A. However, metastasis is by definition a systemic disease (223). Despite the growing understanding of MIC-intrinsic and metastatic niche–dependent mechanisms, relatively little is known about the complex ways in which systemic physiologic factors influence MIC plasticity and selection, and vice versa. However, recent studies have provided evidence for the potential role of various systemic signals in the development of metastasis (Fig. 6B). Although the molecular basis for a majority of these effects remains to be determined, and their causal relationship with metastatic progression remains to be firmly established, this is an area of rising interest.

Both primary and metastatic tumors can release cytokines, chemokines, and hormones either directly into the systemic circulation or packaged into exosomes, which in turn can modulate distant niches to favor or resist metastatic colonization, including organ-specific metastatic dissemination (146, 150, 224, 225). Although many critical mediators and mechanisms of premetastatic niche formation remain to be uncovered, studies to date suggest that organ-specific stromal cell reprogramming induces changes in ECM and matrix-associated proteins, which in turn enable recruitment of specific myeloid populations that support future MIC growth (226). Indeed, bone marrow–derived myeloid cells can play critical roles in either limiting or promoting MIC survival and growth, and the mechanisms enabling their context-specific recruitment, differentiation, and activity at various steps in the metastatic process have been a focus of growing interest (18, 97, 129).

The critical role of systemic immunity in controlling metastasis is now firmly established. Physiologic and pathologic phenomena that modulate systemic immunity indirectly control metastasis, in addition to having direct effects on MICs (Fig. 6). Stress and depression have been associated with early metastasis and poor prognosis in patients with breast, lung, and hepatobiliary carcinomas (227). In cell line xenograft models of metastatic breast cancer, increased corticosteroid stress hormone levels generated during tumor progression resulted in increased glucocorticoid receptor activation, downstream ROR1 upregulation, metastatic growth, and chemoresistance (228). In addition to direct effects on MICs, glucocorticoids have widespread immunosuppressive effects, which in turn can promote MIC outgrowth (229, 230). In addition to steroids, increased levels of the stress-induced catecholamine norepinephrine induced changes in neutrophils that enabled MIC escape from dormancy in syngeneic lung cancer models (231).

Systemic inflammation and immunosuppression resulting from a range of physiologic and pathologic conditions, including aging, postpartum breast involution, asthma, and osteoarthritis, have been associated with increased metastasis risk and reduced survival from a range of epithelial cancers (232–236). In a recent study, inflammation induced by surgical wounding was shown to promote MIC outgrowth from dormancy in breast cancer mouse models (237), suggesting a potential explanation for the early peak in metastatic recurrence within one year of curative breast surgery in patients (238). Based on these observations, several groups have proposed the prophylactic use of nonsteroidal anti-inflammatory agents and beta-blockers in the perioperative setting, although clear clinical evidence for a significant antimetastatic effect is lacking (237, 239).

Tumors in turn may alter host metabolism, immunity, and physiology locally as well as systemically (Fig. 6). Circulating cytokines and metabolites derived from metastatic tumors can have a dramatic impact on tissues such as skeletal muscle and adipose. Indeed, many patients with metastatic cancer develop cachexia, a debilitating muscle-wasting syndrome that is associated with poor prognosis and accelerated death (240). A better understanding of the molecular basis for how metastatic disease promotes cachexia is important for improving the care of patients with metastatic cancer.

The nervous system is emerging as a potential systemic modulator of metastasis (241, 242). Innervation is essential for physiologic regeneration (243), but the role of nerves in metastasis had long been surmised to be only as passive conduits for tumor dissemination in cases of perineural invasion (61). Growing evidence, however, suggests that neurons can be active determinants of cancer progression and outcome (refs. 241, 242; Fig. 6). Direct innervation of tumors can contribute to tumor progression at both the primary and metastatic sites. In mouse cancer models, autonomic denervation of the prostate, stomach, and pancreas inhibited tumor progression and cell dissemination from these sites (244). In addition to effects mediated by direct tumor innervation, increasing understanding of neuronal regulation of immunity suggests a systemic means of neuronal metastatic regulation, although little mechanistic insight is currently available (245). For example, in a syngeneic metastatic mammary carcinoma model, denervation of sensory neurons suppressed NK cell activity, resulting in increased lung and cardiac metastases (246).

Epidemiologic and experimental studies have identified correlations between lifestyle factors such as diet and exercise with cancer incidence, therapy resistance, and survival (247–249). Proposed mechanisms of tumor modulation include direct effects on cancer cells (for example, by altering insulin signaling or lipid metabolism), indirect effects on tumor inflammation and immune surveillance, or, especially for diet, tertiary effects mediated by the gut microbiome (127, 250–255). Obesity, which is associated with both poor diet and low exercise, is an independent risk factor for the future development of distant metastasis and death in patients with surgically resected primary breast cancer (254). With some notable exceptions (e.g., refs 127, 255), the majority of animal studies of diet, exercise, and microbiome modulation to date have been performed in models of tumor initiation or subcutaneous tumor cell transplantation, and the functional and mechanistic significance of these studies to the critical steps of MIC dissemination, dormancy, and outbreak, as well as therapy resistance in patients with advanced cancer, remain unresolved.

It is important to discern whether lifestyle factors that may influence tumor progression represent modifiable, and potentially therapeutically targetable, risk factors in patients who have already been diagnosed with primary tumors. A meta-analysis of 49,095 early-stage breast and colon cancer survivors suggested that survivors who increased their physical activity from before diagnosis and surgery showed a statistically significant decreased total mortality risk in comparison with those who did not change their activity level (256). Prospective clinical studies randomizing newly diagnosed patients with cancer to specific dietary or exercise interventions will help clarify the actionability of these factors in preventing and treating metastasis (257).

The role of the microbiome as a player in influencing cancer biology is becoming increasingly evident (refs. 258, 259; Fig. 6). The mucosal microbiome can interact directly with epithelial cells in niches such as the luminal gastrointestinal tract, respiratory tract, and skin to influence tumor progression through local inflammation or immunosuppression (258, 259). Gut pathobiotic strains of fusobacterium have been shown to colonize colorectal tumors and drive metastasis to the liver (260). Mouse xenografts derived from such human liver metastases continue to be colonized by fusobacterium, and antibiotics that target fusobacterium can inhibit metastasis in mouse models (260). Intriguingly, many cancers that are not directly exposed to environmental bacteria, including breast and bone cancers, have recently been shown to also harbor intratumoral microbiota, although the functional significance of such colonization remains unclear (261). In addition to local effects, metabolites secreted by microbiota may have diverse systemic effects on immunity, hormones, and organ function, most of which are currently poorly understood (262, 263). Other systemic factors, such as diet, can rapidly alter microbiome composition (264, 265), and patient-specific microbiota can in turn influence responses to therapy (266, 267). Given this complexity, there is an increasingly evident need to characterize and control for the microbiome composition as an independent variable in animal and human studies of metastasis.

In some cancers, such as melanoma, age at cancer diagnosis is associated with increased tumor aggression and metastasis (268). Three broad mechanisms have been proposed for age-related local and systemic acceleration of metastasis (269, 270). First, age-dependent accumulation of senescent cells, and their associated secretomes, in the tumor microenvironment may promote tumor invasion, premetastatic niche formation, and metastatic outgrowth (271–273). Coinjection of senescent fibroblasts, but not normal fibroblasts, promoted metastatic progression in mouse models (274). Second, age-related deterioration of extracellular collagen matrices may facilitate tumor cell invasion and promote transendothelial migration (275, 276). Third, age-associated systemic low-grade chronic inflammation, termed “inflammaging” (277), can disrupt tissue architecture and promote tumor dissemination (275, 278), whereas immunosenescence may prevent effective immune editing of disseminated MICs emerging from dormancy (270, 279). Further delineation of the mechanisms by which aging intersects with metastatic progression is needed to define therapeutic strategies to prevent and treat cancer in aging populations.

A cardinal feature of metastatic disease is its ability to resist therapy. We recently reviewed the clinical contexts and approaches for targeting metastasis (280). Here we focus on the properties of MICs that enable therapy resistance and potential approaches to overcome these.

Because tumors comprise numerous genetically distinct subclones, cells harboring therapy resistance mutations can be selected for during metastasis as the patient undergoes treatment for initially localized disease. Genetic analyses of matched primary tumors and metastases show that metastases from patients who had received adjuvant therapy for treatment of initially localized cancer had reduced clonal heterogeneity in comparison with metastases from previously untreated patients, suggesting that treatment imposes an evolutionary bottleneck for MIC emergence (171). This is particularly evident in the case of treatments that target specific oncogenic signaling pathways, where therapy with inhibitors of these pathways selects for the emergence of cancer cells with specific recurrent resistance mutations. Examples include EGFR^T790M mutations in patients treated with the first-generation tyrosine kinase inhibitors gefitinib and erlotinib (281), and ESR1 mutations in patients with estrogen receptor–positive breast cancer treated with antiestrogen therapy (173). Clinical and experimental studies characterizing such targeted therapy–specific resistance mutations have enabled the rapid development of improved drugs that anticipate and target resistance mutations, with improved therapeutic results (282).

In addition to these genetically acquired mechanisms of drug resistance, MIC phenotypic properties may confer intrinsic therapy resistance (153, 280, 283). In most cancers, the mainstay of the treatment of metastasis remains cytotoxic chemotherapy, which preferentially targets rapidly proliferating cells. Thus, MICs that are capable of entering quiescence may more readily evade chemotherapy than do proliferative cancer cells. When therapies result in partial destruction of glandular metastatic tumors, they may in turn disrupt epithelial adherens junctions and induce MICs to enter a regenerative progenitor state, as observed in colorectal cancer (32). Successive rounds of therapy may further select for residual cells that are epigenetically primed to reenter regenerative states (284), suggesting an explanation for the successively restricted duration of clinical response typically observed with sequential lines of therapy in patients with advanced cancer. In addition, the stem-like properties of MICs may include the expression of ATP-binding cassette drug efflux transporters that could confer multidrug resistance (285). Dying or senescent cancer cells induced by therapy that partially debulks metastatic tumors can release cytokines as secretomes, to which MICs can respond by activating growth factor signaling programs to regenerate tumors (286–288).

The immune-evasive properties of MICs are likely also of importance in resisting both tumor cell–directed therapy and systemic immunotherapy. In addition to MIC-intrinsic mechanisms of immune escape, desmoplasia in tumors is an important mechanism of therapy resistance. In pancreatic ductal adenocarcinoma, Hedgehog signaling induces fibrosis, which in turn poses a physical barrier for drug delivery to tumor cells that are surrounded by fibroblasts (20, 289, 290). In many cancer types, immune cells are recruited to tumors from the systemic circulation but are spatially segregated from tumor cells by fibroblast barriers, resulting in immunologically “cold” tumors that do not respond to PD-1 checkpoint immunotherapy (291, 292). Targeting these fibroblast barriers using Hedgehog inhibitors disrupts desmoplasia in mouse models of pancreatic ductal adenocarcinoma (290). TGFβ is strongly immunosuppressive and fibrogenic (88, 89); the use of TGFβ inhibitors in colorectal and breast cancers increases drug and immune cell infiltration into the tumor, reversing immunotherapy resistance in preclinical models and correlating with superior immunotherapy responses in patients (293, 294). However, the translation of these approaches to patients has been challenging to date due to complex effects of drugs that target developmental signaling pathways on the tumor microenvironment, as well as broader systemic toxicities (295–297). Improved understanding of the mechanisms of tumor fibrogenesis and immune cell communication may yield more specific targets for therapeutic intervention.

The plasticity of MICs arguably poses the greatest challenge to therapeutic success, because MICs that are capable of sampling among numerous transcriptional and epigenetic states may be surmised to rapidly select for gene-expression programs that enable therapeutic escape. Therefore, understanding the molecular mechanisms of MIC plasticity may yield improved therapeutics (160, 298). For example, epigenetic modulators could be used to inhibit or reverse MIC plasticity or to force MIC differentiation into terminal, nonproliferative states. Targeting genotype-agnostic conserved regenerative states may block the relapse of metastatic disease that is being debulked by antineoplastic chemotherapy. Characterizing the endpoints of plasticity, such as neuroendocrine lineage plasticity in prostate cancer, may enable the development of new drugs that target the emergent neuroendocrine state instead of or in addition to the preexisting androgen-driven epithelial state (161).

In summary, recent work is illuminating the fundamental principles of MICs and their ecosystems, and importantly shedding light on the mechanisms by which metastatic tumors emerge and develop as phenotypically distinct, therapy-resistant entities in comparison with primary tumors. In brief:

• Metastasis develops through three distinct phases, namely dissemination, dormancy, and outbreak, which can coexist until the removal of the primary tumor.

• Each phase of metastasis reflects a dynamic equilibrium of MICs with host immunity, with the eventual outbreak of MICs as macrometastatic colonies reflecting a failure of immune surveillance.

• Although preexisting genetically distinct tumor subclones with high metastatic propensity are selected for in primary tumors, phenotypic plasticity, enabled by epigenetic and metabolic adaptations, is an overarching hallmark of MICs.

• Phenotypic plasticity endows MICs with the ability to sample multiple transcriptional states and dynamically adapt to diverse stresses during each phase of metastasis.

• MICs can adopt epithelial regenerative progenitor states that possess stem-like, immune-evasive, tissue-regenerative, and therapy-resistance capacities.

• The dynamic equilibrium between MICs and their organ-specific microenvironmental niches can be modulated by systemic factors that affect either or both components.

• Identifying the vulnerabilities of conserved regenerative MIC states, and the molecular determinants of communication between MICs and their ecosystems, could yield improved therapeutics to prevent and treat metastasis, and thus improve outcomes for patients with advanced cancer.

With the advent of immunotherapy and continued advances in targeted therapy, the last two decades have witnessed considerable improvements in outcomes of patients with some types of metastatic cancers (268). These advances are driven by increasingly closer collaborative bidirectional interactions between preclinical and clinical scientists, and improved, more representative patient-derived and animal models of metastatic cancer. Significant therapeutic advances are resulting from the rapid recognition, molecular characterization, and clinical translation of biologically discrete phenotypes of metastatic disease. Examples include curative local therapy for oligometastatic disease, immune activation phenotypes in extreme outlier patients with exceptional response to therapy, and mechanism-based therapies for metastases in specific organs, such as the brain, where even small improvements in tumor shrinkage can yield significant improvements in patient survival (280).

Further improvements in patient outcomes will emerge from a deeper understanding of the mechanisms that underlie the plasticity and dormancy of MICs as well as the complex interactions between the MICs, their niches, and systemic factors. To do so, preclinical models used to study metastasis and test therapies for metastatic cancer must recapitulate the genomic complexity, plasticity, and niche biology of advanced human cancers, rather than relying exclusively on simpler systems that may fail to capture these critical features.

In 1889, Steven Paget first hypothesized that metastasis might result from a fertile “seed” landing on congenial niche “soil” (299). Although investigations in the subsequent 132 years have vindicated Paget's view, a third dimension of the metastasis ecosystem—the “forest” of systemic factors—is emerging. By understanding and targeting the interactions between the seeds, soil, and forest, we may be able to convert metastasis, which is today still typically a death sentence, into a manageable illness.

J. Massagué reports grants from NIH during the conduct of the study; and personal fees from Scholar Rock Inc and Biogen Inc. outside the submitted work; in addition, J. Massagué has a patent for US7829066 issued, a patent for EU2831593 issued, a patent for EU3047039 issued, a patent for US10413522 issued, a patent for PCT/US2016/062880 pending, and a patent for PCT/US2017/045145 pending; and owns Scholar Rock Inc. stock. K. Ganesh reports grants from NIH, the Damon Runyon Cancer Research Foundation, and Burroughs Wellcome Funds during the conduct of the study; in addition, K. Ganesh has a patent for PCT/US2017/045145 pending.

The authors' work in the subject area of this review is supported by NIH grants R35CA252978 (J. Massagué), K08CA23021 (K. Ganesh), and P30-CA008748 (MSKCC), and the Alan and Sandra Gerry Metastasis and Tumor Ecosystems Center at MSKCC (J. Massagué).