|Home | About | Journals | Submit | Contact Us | Français|
The mechanistic underpinnings of metastatic dormancy and reactivation are poorly understood. A gain-of-function cDNA screen reveals that Coco, a secreted antagonist of TGF-β ligands, induces dormant breast cancer cells to undergo reactivation in the lung. Mechanistic studies indicate that Coco exerts this effect by blocking lung-derived BMP ligands. Whereas Coco enhances the manifestation of traits associated with cancer stem cells, BMP signaling suppresses it. Coco induces a discrete gene expression signature, which is strongly associated with metastatic relapse to the lung but not to the bone or brain in patients. Experiments in mouse models suggest that these latter organs contain niches devoid of bioactive BMP. These findings reveal that metastasis-initiating cells need to overcome organ-specific anti-metastatic signals in order to undergo reactivation.
Metastatic relapse of breast cancer and other carcinomas frequently occurs several years after initial surgery. Increasing evidence suggests that tumor cells that have disseminated from early lesions, including ductal carcinomas in situ, undergo an extended period of dormancy in the stroma of target organs (Nguyen et al., 2009; Valastyan and Weinberg, 2011). It is currently unclear if tumor cells become dormant as a consequence of intrinsic defects or in response to inhibitory signals that they encounter in foreign microenvironments.
Many malignancies, including breast cancer, are fuelled by a limited, although not necessarily small, number of cancer stem cells, which undergo self-renewal as well as generate rapidly proliferating progenitors and aberrantly differentiated post-mitotic cells (Clevers, 2011; Gupta et al., 2009). The metastatic capacity of human pancreatic and colorectal cancers is restricted to a subpopulation that includes cancer stem cells (Hermann et al., 2007; Pang et al., 2010). Furthermore, the Epithelial to Mesenchymal Transition (EMT) that facilitates tumor dissemination produces cells endowed with the capacity to self-renew, suggesting that these two cellular processes are intimately linked (Mani et al., 2008). Finally, the Id1/3 transcription factors and the miR200 and miR335 microRNAs promote the colonization phase of breast cancer metastasis at least in part by modulating cancer stem cell function (Gupta et al., 2007; Korpal et al., 2011; Shimono et al., 2009; Tavazoie et al., 2008). These results suggest that the cancer stem cells possess the migratory and self-renewal capabilities necessary to colonize distant organs, whereas the remaining tumor cells lack metastatic capacity.
The ability of metastasis-initiating cells to enter into, and eventually exit from, proliferative quiescence suggests an additional commonality with adult tissue stem cells. However, the relationship between cancer stem cell behavior and dormancy at metastastic sites is poorly understood. In this paper, we provide evidence that Coco, a secreted antagonist of TGF-β ligands, induces dormant metastasis-initiating cells to undergo reactivation in the lung. Mechanistic studies suggest that Coco exerts this function by blocking paracrine BMP signalling and thereby enhancing the self-renewal capability of metastasis-initiating cells.
We designed a gain-of-function cDNA screen that uses the mouse as a filter to isolate genes that mediate metastasis (Figure 1A) and applied it to a previously described series of mammary carcinoma cell lines, which appear to be arrested at defined steps of metastasis (Aslakson and Miller, 1992). Upon orthotopic injection, the 67NR cells give rise to non-invasive tumors, the 168FARN cells colonize locoregional lymphnodes but do not gain access to the vasculature, and the 4TO7 cells are able to disseminate but do not produce macroscopic metastases. In contrast, the 4T1 cells produce macroscopic metastases in the lung (Figure 1B). Upon transduction with cDNA libraries derived from 4T1 cells, the 67NR or 168FARN cells did not acquire the capability to give rise to lung metastases in 8 weeks, suggesting that the introduction of a single gene did not enable these cells to penetrate into the bloodstream and acquire the additional capabilities required for metastatic colonization. In contrast, the 4TO7 cells infected with the 4T1 libraries produced a total of 8 lung nodules in multiple mice (Figure 1B). After proviral rescue and re-introduction in 4TO7 cells, 3 of the 8 cDNAs isolated from individual lesions promoted lung metastasis without affecting primary tumor growth (Figures S1A; not shown). In contrast, 4TO7 cells infected with empty vector did not produce macroscopic lesions upon injection in 30 mice. This screening strategy can thus be used to identify mediators of the homing and outgrowth step of metastasis.
We focused on cDNA1 because it encoded an N-terminally truncated but potentially active version of Coco, a secreted inhibitor of TGF-β ligands (Bell et al., 2003) (Figure S1B). Rossant and colleagues had isolated the same transcript and termed it Dante (Pearce et al., 1999). Studies on frog development had shown that Coco binds directly to BMP and Nodal proteins, blocking their ability to bind to their cognate receptors, and interferes with Wnt signaling (Bell et al., 2003). Experiments in Xenopus embryos and 4TO7 cells indicated that cDNA1 possesses all the biological activities of full-length Coco (Figure S1C-F). Conversely, expression of full-length Coco induced the 4TO7 cells to colonize the lung after orthotopic injection, confirming that Coco promotes spontaneous metastasis similar to cDNA1 (Figure S1G).
The expression of Coco correlated with metastatic capability in the 4T1 progression series (Figure 1C, top). Of note, a large fraction of Coco remained associated with the cell layer rather than diffusing in the medium (Figure 1C, top). Treatment of live cells with modest amounts of highly purified trypsin led to the disappearance of Coco from the cell layer (Figure 1C, bottom), indicating that this inhibitor associates with the pericellular matrix and might therefore accumulate at high concentrations near the cell surface.
Interestingly, neither expression nor silencing of Coco modified the ability of tumor cells to proliferate in vitro (Figures 4D and S1H; not shown) or to form primary tumors when 1 × 105 cells were injected in the mammary fat pad (Figure 1E and 1F). Furthermore, Coco did not disrupt epithelial adhesion in normal murine mammary epithelial cells (Figure S1I) or induce completion of an EMT program and enhance invasion in 4TO7 cells (Figures S1J and S1K). These results suggest that Coco does not promote primary tumor growth or invasion.
To examine if Coco promotes the colonization step of metastasis, we performed tail-vein injection experiments. Expression of Coco enabled the 4TO7 cells to metastasize efficiently to the lung under these conditions (Figure 1G). Conversely, silencing of Coco using two distinct shRNAs suppressed the ability of the highly metastatic 4T1 cells to colonize the lung (Figure 1H). Similar inhibitory effects were observed upon depletion of Coco in 66cl4 cells, which were derived from the same spontaneous tumor as the 4T1 cells but are less aggressive (Aslakson and Miller, 1992) (Figure S1L; >99% inhibition), and in ErbB2-transformed mammary tumor cells isolated from MMTV-Neu(YD) mice (Guo et al., 2006) (Figure S1M; >90% inhibition). We concluded that Coco mediates lung colonization.
Confocal imaging of lung sections revealed that the 4TO7 cells extravasate in the stroma of the lung within 1 day after injection in the tail vein (Figures 2A and 2B). Coco did not enhance the ability of 4TO7 cells to infiltrate the lung. However, whereas control 4TO7 cells remained solitary and seemingly quiescent in the stroma of this organ, a small fraction of 4TO7-Coco cells started to proliferate from around day 14 and gave rise to metastatic outgrowths (Figures 2A and 2B). Most of the lesions were relatively small and had not yet undergone neoangiogenesis at day 21 but became large and vascularized by day 35 (Figure 2A). While the number of tumor cells within outgrowing micrometastases was comparable to that of solitary tumor cells at day 21, the tumor cells within macro-metastases far outnumbered the solitary tumor cells at day 35 (>9 fold). Anti-Ki67 staining suggested that a large majority of solitary 4TO7 and 4TO7-Coco cells were not actively cycling (>97%). In contrast, a large fraction of the 4TO7-Coco cells within metastatic lesions were actively proliferating (Figures S2A and S2B). The solitary 4TO7 and 4TO7-Coco cells were not apoptotic at any time point examined (Figure S2C), in agreement with the hypothesis that they had become dormant.
To confirm that the solitary tumor cells detected in the lung were quiescent, we performed 5-ethynyl-2’-deoxyuridine (EdU) incorporation experiments (Figure 2C). Confocal imaging after fluorescent conversion of EdU and anti-GFP staining indicated that > 95% of solitary 4TO7 and 4TO7-Coco cells did not enter into or traverse the S-phase over each of 3 distinct EdU incorporation periods. In contrast, a large fraction of 4TO7-Coco cells within outgrowing micrometastic lesions and overt metastases transited through the S phase over the same times (Figures 2D and 2E). Therefore, the 4TO7 cells undergo a protracted period of solitary tumor dormancy in the lung, but expression of Coco enables a fraction of these cells to exit from dormancy and produce metastatic outgrowths.
To examine if silencing of Coco induces dormancy, we examined the lungs of mice injected with control or Coco-silenced 4T1 cells. As anticipated, silencing of Coco did not inhibit extravasation or decrease the number of solitary tumor cells present on lung sections at day 14 and 28 (Figure S2D). Furthermore, whereas a fraction of control 4T1 cells initiated proliferation to give rise to metastatic outgrowths, virtually all Coco-silenced cells remained quiescent (Figures 2F and S2D), suggesting that depletion of Coco suppresses lung colonization by preventing the reactivation of dormant cells.
To define the molecular underpinnings of the pro-metastatic activity of Coco, we examined its ability to regulate BMP, Nodal, and Wnt signaling in mammary tumor cells. We excluded Nodal because the 4TO7 and 4T1 cells do not express nodal receptors or Cripto and do not respond to Nodal in reporter assays (Figures S2E and S2F; not shown). However, they express two type I and type II BMP receptors as well as Frizzled7 and LRP5 and 6 (Figures 2G and S2G). As anticipated, BMP4 induced robust C-terminal phosphorylation of BMP-responsive Smad proteins (1, 5, 8) in 4TO7 cells, and expression of Coco reversed this process (Figure 2G). In contrast, Coco did not inhibit but partially enhanced Wnt signaling (Figure S2H-K). Coco may exert this latter effect by alleviating the intracellular inhibitory crosstalk that BMP receptors exert on β-catenin (He et al., 2004; Kobielak et al., 2007).
Immunostaining of lung sections indicated that virtually all the solitary, dormant 4TO7 cells displayed strong nuclear accumulation of P-Smad over 35 days of observation, suggesting that BMP signaling was robustly activated in these cells (Figures 2H and 2I). Semiquantitative RT-PCR indicated that the 4TO7 and 4T1 cells express low levels of BMP proteins, whereas the cellular elements of the normal lung express significant levels of several BMP proteins (Figure S2L), in general agreement with prior data showing that both epithelial and mesenchymal cells in the lung produce BMP (Danesh et al., 2009). Metastatic seeding by 4TO7 or 4TO7-Coco cells did not modify the levels of BMP mRNAs in recipient lungs (Figure S2M). We thus inferred that the 4TO7 cells underwent protracted proliferative quiescence in response to BMP proteins that had been produced by both epithelial and mesenchymal cells and deposited in the stroma between alveoli.
Intriguingly, a small fraction (~2%) of the dormant tumor cells present in the lungs of mice injected with GFP-tagged 4TO7-Coco cells displayed no nuclear accumulation of P-Smad (Figures 2H and 2I). We speculate that these cells accumulated more Coco in their pericellular matrix or were exposed to lower amounts of BMP as compared to other cells. In agreement with the hypothesis that these rare P-Smad− cells were fated to produce metastatic outgrowths, both the incipient lesions and the macrometastases displayed no nuclear accumulation of P-Smad (Figures 2H and 2I). The solitary 4TO7 and 4TO7-Coco cells as well as the outgrowing 4TO7-Coco cells did not exhibit accumulation of β-catenin in the nucleus, suggesting that β-catenin signaling does not contribute to reactivation (Figures S2N and S2O).
In a reciprocal set of experiments, we evaluated if silencing of Coco induced reactivation of Smad signaling in solitary 4T1 cells that would have been destined to give rise to metastatic lesions. A small but sizeable fraction of solitary 4T1 cells (~ 5%) and all of those within outgrowing metastases did not display nuclear accumulation of P-Smad suggesting that the metastases arise from P-Smad-negative cells (Figure 2J). This subpopulation of P-Smad-negative cells was not detected in the lungs of mice injected with Coco-silenced 4T1 cells. In fact, > 99.3% of dormant Coco-silenced 4T1 cells displayed strong nuclear accumulation of P-Smad (Figure 2J). Together, these findings suggest that Coco promotes exit from dormancy by alleviating the ability of stromal BMP to enforce a dormant state.
Since the metastasis-initiating cells share functional properties with cancer stem cells (Nguyen et al., 2009; Valastyan and Weinberg, 2011), we considered the possibility that Coco promotes the manifestation of cancer stem cell traits. Phenotypic analysis indicated that the 4TO7 and 4T1 cells express high levels of the mammary epithelial stem cell markers CD24, CD29, and CD49f but do not express markers found in bi-potential progenitors or cells differentiated along the luminal or myoepithelial lineage (Figure S3A; Table S1). In addition, both types of cells are endowed with a program of gene expression similar to that of normal mammary epithelial stem cells (Table S2). Tumor-initiating cells isolated from mouse models of ErbB2-mediated mammary tumorigenesis exhibit a similar phenotype (Lo et al., 2011).
The cancer stem cells are defined by an inherent capability to undergo self-renewal, to give rise to an aberrantly differentiated progeny, and to seed tumors in vivo (Clevers, 2011; Gupta et al., 2009). To examine if Coco affects cancer stem cell function, we first examined its ability to influence self-renewal in vitro by using the mammosphere assay (Dontu et al., 2003). The 4TO7-Coco cells form about twice as many tumor spheres as compared to 4TO7 cells at each of three subsequent passages (Figure 3A and S3B). Administration of exogenous Coco produced a similar and dose-dependent effect in naïve 4TO7 cells (Figure 3B). In agreement with the observation that 4TO7 cells produce small amounts of BMP2 and 5 (Figure S2L), treatment of naïve 4TO7 cells with 10 ng/ml of exogenous BMP4 caused a profound inhibition of tumor sphere formation, which was reversed by concurrent administration of Coco (Figures 3C and 3D). Neither Coco nor BMP affected the ability of 4TO7 cells to survive or progress through the cell cycle under the conditions of the assay (Figure S3C; not shown). We infer that Coco expands their ability of 4TO7 cells to give rise to tumor spheres by blocking the small amounts of BMP that they produce.
We next evaluated if BMP can cause partial or aberrant differentiation of mammary tumor cells. Notably, treatment of 4TO7 cells with BMP4 induced expression of the transcription factor GATA3, a master regulator of luminal differentiation, whereas concurrent administration of recombinant Coco reversed this process (Figure S3F). In addition, whereas the 4TO7 cells that became dormant upon extravasation in the lung stroma displayed strong reactivity for GATA3 irrespective of whether they expressed Coco, the metastatic outgrowths formed by 4TO7-Coco cells were uniformly negative for GATA3 (Figure S3G). These results suggest that the high levels of BMP present in the lung stroma induce the 4TO7 cells to express GATA3, and they further imply that expression of Coco reverses this process in tumor cells fated to give rise to metastatic outgrowths. However, BMP4 did not induce expression of markers associated with differentiated luminal cells under standard culture conditions or in 3D Matrigel (Table S1; not shown). Thus, BMP upregulates GATA3, seemingly poising breast cancer cells toward luminal differentiation, but is insufficient to initiate partial or aberrant differentiation en face of oncogenic signaling.
Prior studies have suggested that transgenic ErbB2 mammary tumors follow a cancer stem cell model. Staining with the lipophilic dye PKH-26, which is diluted after each cell division, has indicated that the cells that retain the dye during the mammosphere assay possess the highest self-renewal capacity in vitro and the highest tumor initiation capacity in vivo (Cicalese et al., 2009). To examine if Coco affects self-renewal in vitro, we stained Coco-silenced and control primary tumor cells from MMTV-Neu(YD) mice with PKH-26 and subjected them to serial tumor sphere assay (Figure 3E). As previously reported, replating of the PKHHIGH, PKHLOW, and PKHNEG subsets led to tumor sphere formation in all cases, albeit with decreasing efficiency. Notably, knock down of Coco inhibited tumor sphere formation at each passage (Figure 3F). Treatment of naïve ErbB2-transformed cells with BMP4 exerted the same effect and Coco reversed it (Figure S3D). Silencing of Coco did not reduce the ability of ErbB2-transformed cells to survive or proliferate (Figure S3E) nor induced them to express differentiated genes in 3D Matrigel (not shown). These results suggest that Coco selectively sustains the ability of ErbB2-transformed cells to undergo self-renewal in vitro.
Prompted by the observation that Coco increases clonogenic outgrowth under standard culture conditions as well as in soft agar (Figures S3H and S3I), we examined if it also enhanced tumor initiation in vivo. We found that Coco significantly increases the ability of 4TO7 cells to seed tumors in vivo (Figure 3G). As anticipated, this effect was evident only when limiting numbers of tumor cells were injected in the mammary fat pad. Conversely, silencing of Coco inhibited the tumor-initiation capacity of 4T1 cells (Figure 3H). Semiquantitative RT-PCR experiments indicated that while the normal cell types present in the mammary fat pad express BMP7, the primary tumors generated by 4TO7 or 4T1 cells injected at this site did not express any of the BMP genes (Figure S2L). These results suggest that Coco promotes tumor initiation in the mammary gland as well as reactivation in the lung because it opposes the ability of stromal BMP to block clonogenic outgrowth.
The transcriptional program regulated by the embryonic stem cell transcription factors Sox2, Nanog, and Oct4 is often reactivated in aggressive and metastatic breast cancers (Ben-Porath et al., 2008; Wong et al., 2008). Furthermore, the transcriptional coactivator Taz, which is inhibited by the Hippo tumor suppressor pathway, has been recently implicated in breast cancer stem cell maintenance (Cordenonsi et al., 2011). Q-PCR experiments indicated that the 4TO7 cells express high levels of Sox2 and lower levels of Taz and Nanog, whereas the ErbB2-transformed cells express high levels of Oct4 and lower levels of the remaining three transcriptional regulators (Figure S3J and S3K). Intriguingly, Coco increased the levels of expression of Nanog, Sox2, and Taz in 4TO7 cells (Figure 3I). Conversely, BMP4 completely suppressed their expression (Figure 3J). Consistently, silencing of Coco significantly reduced the level of expression of Nanog, Oct4, and Taz and completely ablated expression of Sox2 in ErbB2-transformed cells (Figure S3L). Since Nanog, Sox2, and Oct4 are part of a self-sustaining circuit that powers stem cell maintenance (Young, 2011) and Taz appears to be specifically required in breast tumor progenitor cells (Cordenonsi et al., 2011), Coco may contribute to the manifestation of breast cancer stem cell traits by influencing the expression of these transcription factors.
To examine the mechanism through which Coco enhances the capability of mammary tumor cells to outgrow in the mammosphere assay and during lung colonization, we transduced 4TO7 cells with Smad6, which inhibits canonical BMP signalling (Hata et al., 1998), with the activated β-catenin mutant S4A, or with both Smad6 and β-catenin-S4A (Figure 4A, left). These mutant proteins exerted the anticipated signalling effects (Figures S4A-D). Whereas Smad6 increased tumor sphere formation by approximately twofold, similar to Coco, activated β-catenin did not induce this effect, either alone or with Smad6 (Figures 4A and 4B). In addition, Smad6 attenuated the induction of GATA3 in response to BMP (Figure S4E). These results suggest that Coco promotes self-renewal in vitro predominantly by inhibiting BMP signaling.
To examine if inhibition of BMP signaling promotes metastatic reactivation, we injected 4TO7 cells expressing Smad6 or activated β-catenin intravenously. Whereas Smad6 promoted lung colonization as efficiently as Coco, activated β-catenin induced this process to a modest extent (Figure 4C). Expression of a dominant negative form of BMPR-IB induced both 4TO7 cells and Coco-silenced 4T1 cells to colonize the lung. However, this construct inhibited BMP signaling and therefore promoted lung metastasis less efficiently as compared to Smad6 (Figures S4A, S4B, S4F and S4G). These results indicate that inhibition of BMP signaling rescues 4TO7 cells and Coco-silenced 4T1 cells from tumor dormancy.
As an alternative approach, we tested if BMP signaling decreased tumor initiation. 4T1 cells co-expressing an activated form of BMPR-IB (Q203D) together with BMPR-II (jointly termed CA-BMPR) were significantly less tumorigenic upon injection in the mammary fat pad as compared to controls (Figures 4D and S4H). These results indicate that BMP signaling suppresses the tumor initiating capacity of 4T1 cells.
To further study the connection between tumor initiating capacity and metastatic outgrowth, we examined the ability of 4T1 cells expressing CA-BMPR to metastasize to the lung upon tail vein injection. Whereas control 4T1 cells were highly metastatic in this assay, those expressing CA-BMPR were unable to colonize the lung, confirming that BMP signaling opposes metastatic colonization (Figure 4E). CA-BMPR exerted a similar effect in 4TO7-Coco cells (Figure S4I). These findings suggest that Coco induces metastasis-initiating cells to exit from dormancy by alleviating the capacity of lung-derived BMP proteins to activate canonical Smad signalling.
To explore the role of Coco in human breast cancer metastasis, we first examined a panel of 12 human breast cancer cell lines (Figure 5A). Immunoblotting indicated that the non-tumorigenic or non-invasive cells as well as the ER+ cells capable of colonizing the bone upon intracardiac injection exhibit low or undetectable levels of Coco, whereas the MDA-MB231 cells, which can colonize efficiently the lung and belong to the basal B gene expression cluster associated with a stem cell phenotype, express Coco (Figure 5A). Although, as revealed by additional experiments, the MDA-MB231 cells express the BMP inhibitors Noggin, DAN and Chordin-like 1 at levels comparable to those of Coco (Figure S5A-C), silencing of Coco was sufficient to suppress their ability to colonize the lung (Figures 5B and S5B). Monitoring for 6 additional weeks revealed that the Coco-silenced cells eventually gave rise to micrometastases (Figure S5D, left). These lesions did not arise as a result of re-expression of Coco but presumably through the acquisition of genetic or epigenetic modifications able to bypass its requirement (Figures S5D, right and S5E). Depletion of Coco also suppressed the lung colonization ability of CN34.2a cells, which were derived from the pleural effusion of a patient affected by metastatic breast cancer but are not as aggressive as the MDA-MB231 cells (Padua et al., 2008), suggesting that Coco is a general mediator of lung colonization.
Sorting according to the surface expression of CD44 and CD24 has been successfully used to identify cancer stem cells within primary breast tumors or mammary cell lines, including the MDA-MB231 cells (Al-Hajj et al., 2003; Cordenonsi et al., 2011). As anticipated, the large majority of control MDA-MB231 cells were CD44HIGH/CD24LOW/−, consistent with a cancer stem cell phenotype. Silencing of Coco led to the appearance of a large subpopulation of cells that did not express CD44, suggesting that expression of Coco is necessary to maintain the expression of this marker in MDA-MB231 (Figure 5C). In consonance with this observation, depletion of Coco profoundly inhibited the ability of MDA-MB231 cells to form tumor spheres in vitro without affecting their survival or proliferation (Figures 5D, S5G and S5H). Finally, silencing of Coco inhibited the capacity of MDA-MB231 cells to initiate tumorigenesis upon orthotopic injection in NOD-SCID-IL2γR−/− mice (Figure 5E). These results provide evidence that Coco promotes maintenance of various stem cell traits by MDA-MB231 cells.
To study the expression of Coco in human breast cancer, we stained 3 distinct TMAs comprising 15 normal or dysplastic glands and 126 breast cancers with affinity-purified antibodies reacting selectively with Coco (Table S3 and Experimental Procedures). Whereas normal or displastic mammary epithelial cells did not express detectable levels of Coco, the tumor cells and scattered stromal cells in a large fraction of breast tumors produced variable amounts of Coco (Figures 5F, 5G and S5I). Statistical analysis did not reveal any correlation between the intensity of staining for Coco and pathological grade, clinical stage, ER, or HER2 status in these samples (not shown). Interestingly, Kaplan Meyer analysis of the MSKCC TMA dataset, which is annotated for metastatic relapse, indicated that the patients with primary tumors exhibiting high levels of Coco had a significantly shorter overall survival (46.2 ± 7.7 months) as compared to the remaining patients (104.9 ± 23.2 months)(Figure 5H). We were not able to uncover other correlations, possibly because of the limited number of cases present in this TMA and the paucity of paired samples. These observations suggest that Coco confers a selective advantage during both tumor initiation and progression.
DNA microarray analysis was used to examine if Coco’s activity correlates with metastatic relapse in human breast cancer. Many signaling pathways, including the BMP pathway, are restrained by negative feedback loops, which ensure expression of target genes only after a defined threshold of signaling has been reached. We thus reasoned that a Coco-dependent signature of gene expression may have been more reflective of Coco’s activity than the level of expression of Coco mRNA. Furthermore, the Affymetrix HG-U133A and Agilent platforms do not contain probes for Coco.
DNA microarray analysis of control and Coco-silenced MDA-MB231 cells led to the definition of a signature comprising 56 genes (Figure 6A). By using a leave-one-out cross validation method, we identified the 14 genes most relevant in predicting overall metastatic relapse in the NKI295 dataset (Figure 6B, left), which comprises early stage, lymphnode negative cases (van de Vijver et al., 2002) (Table S4), and validated their predictive power on the EMC286 dataset (Figure S6A), which comprises early stage tumors from patients, who did not undergo adjuvant therapy after surgery (Wang et al., 2005) (Table S4). Finally, we tested the ability of the 14-gene signature to predict overall metastatic relapse in a combined dataset comprising the MSK82, EMC192, and EMC286 cohorts. This large dataset comprises patient populations with distinct clinical, pathologic and treatment characteristics, in proportions similar to those occurring in the breast cancer population at large (Bos et al., 2009) (Table S4). Interestingly, the 14-gene signature was strongly associated with overall metastatic relapse in this cohort (P=2.9E-4; n=560) (Figure 6B, right).
Further distillation of the 14-gene signature led to the identification of 2 of its component genes, KIAA1199 and NDRG1, whose combined overexpression predicted overall relapse in the EMC286 dataset with efficiency similar to that of the 14-gene signature (Figure S2B). Although both genes encode for proteins of unknown cellular function, prior studies have suggested that KIAA1199 sustains Wnt signaling in colorectal cancer (Birkenkamp-Demtroder et al., 2011) and NDRG1 is activated by HIF-1α, through inactivation of Myc-mediated repression, and by AP-1 (Ellen et al., 2008). Interestingly, expression of the 2 genes strongly correlated with poor prognosis in the largest cohort comprising the MSK82, EMC192, EMC286, and NKI295 datasets (P=2.9E-7; n=855) (Figure S6B).
To examine if expression of Coco correlates specifically with metastasis to the lung, we first compared the levels of Coco in lung and bone metastatic variants of MDA-MB231 cells (Zhang et al., 2009). Notably, Coco was upregulated in the lung metastatic variants LM2-4180 and LM2-4175 but not in the bone metastatic variants Bo-1833 and Bo-2287 (Figure 6C). None of 9 additional secreted BMP inhibitors had a similar pattern of upregulation in lung metastatic variants (Figure S5A). To corroborate the hypothesis that expression of Coco underlies organ-specific metastasis to the lung, we evaluated if the 14-gene and 2-gene signatures are able to selectively predict lung metastasis. Preliminary analyses indicated that both signatures were able to predict relapse to the lung but not to the bone or brain in the MSK82, EMC192, and EMC286 datasets (Figure S6C). To avoid the potentially confounding effect of patients with multi-site metastases, we repeated the analysis in the large combined dataset (MSK82 + EMC192 + EMC286) after exclusion of such patients. The results indicated that both signatures were strongly associated with lung but not bone or brain metastasis (Figures 6D and 6E) and predicted overall survival as well as lung metastasis independently of transcriptomic subtype (Table S5), tumor size, lymphnode positivity, ER status, HER2 status, pathological grade or expression of the NKI 70-gene poor survival signature (Tables S6 and S7). Finally, although the original Coco signature comprising 56 genes displayed target organ-specificity similar to that of the previously described Lung Metastasis Signature (Figure S6D) and was able to predict overall survival with similar efficiency (Table S8), the two signatures only shared 3 genes (Table S9), consistent with the involvement of biologically distinct mechanisms (Minn et al., 2005).
To examine if NDRG1 and KIAA1199 participated in lung colonization, we silenced each of the two genes, singly or in combination, in MDA-MB231 cells (Figure S6E-G). Although single silencing of NDRG1 or KIAA1199 did not inhibit lung colonization (Figures 6F, 6G, S6E, and S6F), simultaneous downregulation of both genes led to a significant inhibition of lung colonization (>80% inhibition at 10 weeks; Figure 6H and S6G), suggesting that NDRG1 and KIAA1199 are Coco-regulated genes involved in lung colonization.
To assess if Coco participates in metastasis to the bone or brain, we performed intracardiac injection experiments. Bioluminescent imaging showed that mice injected with control 4T1 cells develop bone, brain and adrenal gland metastases (Figures 7A and 7B). The bone lesions occurred in >80% of the mice and were osteolytic (Figures 7A and S7A). Notably, silencing of Coco did not reduce the rate of metastasis to bone, brain or adrenal gland or inhibit the growth of individual lesions (Figures 7A and 7B), suggesting that Coco is not required for colonization of these organs.
We hypothesized that Coco may mediate organ-specific metastasis to the lung because the stroma of this organ contains particularly high levels of bioactive BMP proteins. Since the action of BMP proteins is regulated by multiple secreted inhibitors, some of which can neutralize each other or also function as direct activators, as well as by complex additional mechanisms (Walsh et al., 2010), it is impossible to estimate the amount of bioavailable BMP present in a given tissue by using direct methods. Therefore, to gauge the amount of active BMP present in the bone marrow stroma, we examined P-Smad signaling in tumor cells that had infiltrated this microenvironment. Analysis of bone sections from control mice indicated strong nuclear accumulation of BMP-responsive P-Smad proteins in chondrocytes in the growth plate but not in most hematopoietic cells (Figure S7B-D). Intriguingly, a large fraction of the solitary 4T1 cells present in the bone marrow 7 days after intracardiac injection did not display nuclear accumulation of P-Smad (Figure 7C). Furthermore, virtually all the constituent cells in all micrometastases and osteolytic lesions detected at 5 weeks were similarly P-Smad-negative (Figure 7D), suggesting that these outgrowths had originated from P-Smad-negative solitary tumor cells. Similar results were obtained from examining the micrometastatic lesions that arose in the bone of 2 out of 5 mice injected with 4TO7 cells 7.5 weeks earlier (Figure 7D). Finally, also all of the solitary tumor cells detected in the brain parenchyma of mice injected 7 days earlier with 4T1 cells did not display nuclear accumulation of BMP-responsive P-Smad proteins (Figure 7C). In contrast, we had found that all the 4T07 and more than 94% of the 4T1 cells that had undergone dormancy in the lung were P-Smad-positive (Figure 2H and 2J). These results suggest that the levels of bioactive BMP proteins capable of engaging their cognate receptors on tumor cells are low in at least a subset of stromal microenvironments in the bone and brain. Together, these results suggest that Coco selectively mediates colonization of the lung because it enables tumor cells to overcome the inhibitory action of BMP proteins that they encounter upon infiltrating this organ.
We have found that breast cancer cells that have successfully extravasated in the lung and survived initial attrition remain dormant for an extended period because stroma-derived BMP proteins limit their ability to proliferate. Production of Coco enables a fraction of these cells to overcome inhibitory BMP signaling and to outgrow into macroscopic metastases. Thus, although most disseminated tumor cells may possess intrinsic defects that preclude them from surviving or undergoing active proliferation in the lung, those that are fated to give rise to clinical metastases, the metastasis-initiating cells, face strong anti-metastatic signals originating from the parenchyma of this organ (Figure 7E).
It is plausible that only a subpopulation of tumor cells – the so called cancer stem cells – possess the extensive self-renewal capability necessary for successful colonization of target organs (Clevers, 2011; Valastyan and Weinberg, 2011). Our results suggest that BMP enforces tumor dormancy by repressing two key cancer stem cell traits: self-renewal in vitro and tumor initiation in vivo. They further suggest that Coco induces exit from dormancy by reversing the ability of BMP to inhibit cancer stem cell function. This model is consistent with previous studies indicating that the BMP pathway inhibits self-renewal and promotes differentiation in pluripotent stem cells and various adult stem cells (Varga and Wrana, 2005). In addition, it is in general agreement with the observation that BMP signaling contributes to inhibit mesenchymal and stem cell states in human mammary epithelial cells (Scheel et al., 2011).
Of interest, BMP profoundly inhibited the expression of the transcription factors Nanog, Sox-2, and Oct-4, which comprise the core regulatory circuit that sustains embryonic stem cells (Young, 2011), and of the Hippo transducer Taz, which confers stem cell-related traits on breast cancer cells (Cordenonsi et al., 2011). In contrast, Coco enhanced the expression of these transcriptional regulators by reversing the effect of BMP. We presume that the interrelationships between these transcriptional regulators are complex, and that each component contributes to sustain a cancer stem cell program of gene expression, underlying the functions necessary for metastatic reactivation. In fact, the transcriptional modules that include as core components Nanog, Sox-2 and Oct-4 as well as Taz/Yap are overexpressed in aggressive and metastatic solid tumors (Ben-Porath et al., 2008; Wong et al., 2008).
The observation that Coco promotes colonization of the lung, but not of the bone or brain, reveals the existence of organ-specific barriers to metastatic reactivation. A large fraction of breast cancer cells that infiltrated the bone marrow stroma and virtually all of those that lodged in the brain parenchyma to give rise to metastatic outgrowths did not display nuclear accumulation of BMP-responsive P-Smads, suggesting that they had not been exposed to high levels of bioactive BMP. Thus, although they may face additional barriers, many tumor cells that infiltrate the bone or the brain do not need to neutralize locally produced BMP in order to outgrow. Since pre-treatment with BMP blocks the ability of MDA-MB231 cells to colonize the bone after intracardiac injection (Buijs et al., 2011), we speculate that breast cancer cells that infiltrate the bone are sensitive to the inhibitory action of BMP. They simply do not need to avert it through production of Coco because they lodge within sanctuaries devoid of the cytokine.
Several lines of evidence indicate that Coco is a particularly potent mediator of metastatic reactivation and that it exerts its effect by inhibiting lung-derived BMP (Supplementary Discussion). In particular, Coco was the only BMP inhibitor whose expression correlated with lung metastatic capacity. Furthermore, although the MDA-MB231 cells expressed Chordin-like 1, DAN, and Noggin at levels similar to those of Coco, knock down of Coco was sufficient to block the lung metastatic capacity of these cells. Although all secreted BMP inhibitors function as ligand traps, some have additional functions. Furthermore, individual members differ in oligomerization state, affinity for individual BMP proteins, and ability to bind to the extracellular matrix (Walsh et al., 2010). Therefore, Coco might be a particularly potent mediator of lung colonization because it has a very high affinity for BMP proteins or because it binds to the pericellular matrix and therefore reaches a very high concentration near the cell surface. However, we cannot exclude the possibility that Coco induces metastatic reactivation in the lung also through BMP-independent mechanisms, which remain to be defined.
The lag phase that separates the entry of tumor cells producing Coco into the lung stroma and their outgrowth suggests that additional adaptive mechanisms mediate lung colonization. Signals initiated by β1 integrins enable the initial outgrowth of tumor cells that have established productive interactions with the interstitial matrix of this organ (Shibue and Weinberg, 2009). In addition, the matrix proteins Tenascin C and Periostin organize niches that nurse outgrowing micrometastases by modulating Notch and Wnt signalling, respectively (Malanchi et al., 2012; Oskarsson et al., 2011). Although these signaling mechanisms have not been specifically linked to the reactivation of solitary tumor cells, it is possible that they cooperate with Coco to drive this process. Alternatively, they may act following initial outgrowth to foster the expansion of micrometastatic lesions, similar to the role of VCAM-1 in osteolytic bone lesions (Lu et al., 2011). Our observation that activated β-catenin does not mediate exit from solitary tumor cell dormancy is consistent with the hypothesis that Periostin falls in this latter class of prometastatic entities.
Attesting to the clinical relevance of our findings, patients with breast carcinomas expressing high levels of Coco exhibited reduced overall survival. Furthermore, expression of a 14-gene Coco signature predicted relapse to the lung but not to the bone or brain, in agreement with the notion that Coco specifically promotes colonization of the lung. Provocatively, further distillation of the signature led to the identification of 2 genes that maintained an intact ability to predict relapse to the lung and, in fact, participated in this process in a mouse model. Based on these observations, we suggest that the 14-gene and 2-gene signatures may be used to identify patients with a significant risk to develop lung metastases. Furthermore, we propose that monoclonal antibodies or other biological agents blocking Coco may exhibit anti-metastatic activity.
Coco has emerged from a gain-of-function retroviral cDNA screen in a mouse model of lung metastatic dormancy. The identification of Coco demonstrates that such screens could be used to define the genes that mediate the exit of solitary tumor cells from dormancy in the lung. Furthermore, similar screens using genome-wide shRNA libraries might enable the identification of genes that promote dormancy. Finally, both cDNA and shRNA screens could be applied to other mouse models to identify the mechanisms that regulate tumor dormancy at other metastatic sites or to investigate other steps of the metastatic cascade. By affording the advantage of rapid biological validation of single entities that are able to enforce dormancy or mediate exit from it, such gain-of-function genetic screens should lead to the identification of additional potential therapeutic targets for the treatment of metastatic disease.
Size-fractionated retroviral cDNA libraries were constructed essentially as described previously (Koh et al., 2004). 67NR, 168FARN, or 4TO7 cells were infected independently with the libraries at an MOI of 3:1 and injected in the mammary fat pad of BALB/c mice. Clonogenic tumor cells isolated from macroscopic metastases were expanded in selective medium (Aslakson and Miller, 1992). Proviral DNA was rescued and sequenced as described previously (Koh et al., 2002).
Tumorigenesis and metastasis assays were performed as previously described (Pylayeva et al., 2009). Animal studies were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee of MSKCC.
Mammosphere assays were performed essentially as previously described (Dontu et al., 2003). ErbB2 cells were stained with PKH-26 dye and subjected to sequential tumor sphere assays as described earlier (Cicalese et al, Cell 2009).
TMAs containing primary breast tumors and lung metastases were generated by the MSKCC Department of Pathology in compliance with protocols approved by the MSKCC Institutional Review Board and after the subjects gave their informed consent. Coco immunoreactivity was evaluated and scored by a clinical pathologist (E.B.). Overall survival was examined as a function of Coco abundance in breast tumor samples.
Comparisons between two groups were performed using an unpaired two-sided t test (p < 0.05 was considered significant). Results are reported as mean ± SD or ± SEM unless otherwise noted. All in vitro experiments were performed at least three times.
We thank G. Daley, J. Massagué, and F. Miller for reagents, X. Zhang and X. Jing for help with X-ray imaging, M. Akram for Coco staining of human TMAs, members of the Giancotti laboratory for discussions, and the staff of the Genomics and the Molecular Cytology Facilities for their help. This work was supported by the Geoffrey Beene Cancer Center at MSKCC and NIH grant P01 CA094060.
Gene expression data of MDA231-sh-control, sh-Coco #2, and sh-Coco #4 cells are deposited at Gene Expression Omnibus (GSE28049).
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.