Forward genetic screens in mice uncover mediators and suppressors of metastatic reactivation

Significance

Insights into the mechanisms that enable disseminated cancer cells to survive during dormancy and then outgrow into life-threatening lesions may lead to the identification of novel therapeutic targets for the prevention or treatment of metastatic disease. We have developed flexible and high-throughput functional genetic screens, which enable the identification of single genetic entities that mediate metastatic reactivation of breast cancer in mice. These screens promise to facilitate the identification of the core signaling pathways that govern metastatic dormancy and reactivation.

Abstract

We have developed a screening platform for the isolation of genetic entities involved in metastatic reactivation. Retroviral libraries of cDNAs from fully metastatic breast-cancer cells or pooled microRNAs were transduced into breast-cancer cells that become dormant upon infiltrating the lung. Upon inoculation in the tail vein of mice, the cells that had acquired the ability to undergo reactivation generated metastatic lesions. Integrated retroviral vectors were recovered from these lesions, sequenced, and subjected to a second round of validation. By using this strategy, we isolated canonical genes and microRNAs that mediate metastatic reactivation in the lung. To identify genes that oppose reactivation, we screened an expression library encoding shRNAs, and we identified target genes that encode potential enforcers of dormancy. Our screening strategy enables the identification and rapid biological validation of single genetic entities that are necessary to maintain dormancy or to induce reactivation. This technology should facilitate the elucidation of the molecular underpinnings of these processes.

The majority of cancer-related deaths are caused by metastatic relapse (1). The process through which cancer cells acquire metastatic capacity is complex. Unrestrained proliferation, resistance to proapoptotic insults, and invasion through tissue boundaries are not sufficient for metastasis. To colonize distant organs, cancer cells must also adapt to the local microenvironment of the target organ and finally outgrow (2, 3). Mathematical modeling of clinical data and experiments in mouse models suggest that cancer cells disseminating from prevalent cancers, such as those of the breast and prostate, undergo an extended period of dormancy at premetastatic sites (4). Insights into the mechanisms that enable disseminated cancer cells to survive during dormancy and then outgrow into life-threatening lesions may lead to the identification of novel therapeutic targets for the prevention or treatment of metastatic disease.

Advances in genomics and mouse modeling have fostered a renaissance of studies on metastasis (2, 3). Current approaches to the study of metastasis can be divided in two categories. In the first, genetic methods are used to modify the function of a candidate gene in intact mice or in cells that are subsequently transplanted in mice (5, 6). In the second, genomic methods, such as DNA microarray analysis or array comparative genomic hybridization (aCGH), are used to identify a restricted number of candidate genes, which are then tested in appropriate mouse models (7, 8). Although these approaches have been extremely successful, they are very laborious and do not necessarily yield biologically potent mediators of metastasis. Functional genetic screens can lead to the rapid identification of strong mediators of a selectable phenotype (9, 10). In agreement with this notion, recent studies have revealed that RNAi screens in vivo can identify mediators of skin carcinogenesis (11). However, formidable practical barriers have so far prevented successful application of gain-of-function screens to a process as complex as metastasis. First, metastasis is a highly selective process. Only a small percentage of cancer cells acquire the genetic alterations necessary to complete each of the sequential steps of metastasis, reducing the throughput of screens in which genetically modified tumor cells are injected at the primary site. Second, the completion of each of the steps of metastasis requires the acquisition of multiple capabilities, suggesting the possible involvement of several genes. We have recently identified a mouse xenograft model that effectively mimics metastatic dormancy and reactivation of breast cancer (12). Here, we describe a flexible and high-throughput functional genomic platform that enables the identification of single genes that enforce dormancy or mediate metastatic reactivation of breast cancer.

Results

Feasibility Studies.

To examine the feasibility of a functional genomics approach to the study of metastasis, we designed a cDNA screen that uses the mouse as a filter to isolate prometastatic genes (Fig. S1A). In this screen, cDNA libraries derived from highly metastatic breast-cancer cells are introduced into nonmetastatic cells, and the recipient cancer cells are injected in the mammary fat pad of mice. After a lag time, cancer cells that have acquired the capacity to metastasize to the lung will eventually form lesions in this organ. Integrated proviruses are rescued from the cells that constitute these lesions and sequenced. Finally, the provirus is reintroduced in low metastatic cells to confirm its prometastatic activity.

To model metastasis as faithfully as possible, we elected to use a progression series that comprises four mouse mammary carcinoma cells seemingly arrested at defined steps of the metastatic cascade (Fig. 1A). Whereas the 67NR cells give rise to noninvasive primary tumors, the 168FARN cells are able to colonize loco-regional lymph nodes but do not gain access to the vasculature, and the 4TO7 cells enter into the vasculature and seemingly home to the lung but do not produce visible metastatic lesions. In contrast, the 4T1 cells complete all these steps and produce macroscopic lung metastases (13).

Fig. 1.

Screening platform for the isolation of genes that mediate metastatic reactivation. (A) Schematic representation of the metastatic capabilities of the mouse mammary carcinoma cells constituting the progression series used here. (B) 4TO7-TGL cells transduced with empty vector or Coco were inoculated i.v. into syngeneic mice. Mice were killed after 21 d. Lung sections were subjected to immunofluorescent staining of GFP (tumor cells) and PECAM-1 (endothelial cells), followed by confocal analysis and 3D reconstruction. (C) Design of the tail-vein cDNA screen. After transduction with the 4T1 libraries, the 4TO7 cells are injected i.v. in syngeneic mice. Candidate mediators of metastasis are recovered from lung metastases.

We constructed three high-complexity cDNA libraries from size-fractionated mRNA isolated from 4T1 cells and cloned them in the modified retroviral shuttle vector pEYK3.1, which uses the MoLV-LTR promoter to drive expression of N-terminally Flag-tagged proteins (total ∼2 × 10⁶ independent clones) (Fig. S1B). FACS analysis indicated that a multiplicity of infection (MOI) of 3:1 leads to efficient transduction of more than 85% of 67NR, 168FARN, and 4TO7 cells (Fig. S1 C and D). The cDNA libraries from the 4T1 cells were thus transduced at an MOI of 3:1 into each of the nonmetastatic cell types (Fig. S2A). Interestingly, upon transduction with 4T1 libraries and injection into the mammary fat pad of syngeneic mice, the 67NR or 168FARN cells did not give rise to macroscopic lung metastases, suggesting that the introduction of a single gene did not enable these cells to penetrate into the bloodstream and acquire all of the additional capabilities required for metastatic colonization. In contrast, the 4TO7 cells infected with the 4T1 libraries produced a total of nine lung nodules in multiple mice within 60 d (Fig. S2A) (12). Thirty mice injected with 4TO7 cells infected with empty vector did not produce macroscopic lung metastases, indicating that insertional mutagenesis does not contribute to metastasis in this system.

Sequencing revealed that three of the cDNAs isolated were in the reverse orientation and that one of them was out of frame (Table S1). The remaining three cDNAs were able to promote lung metastasis upon reintroduction in 4TO7 cells (12). One of these cDNAs was recovered from two distinct lesions and found to encode for an N-terminally truncated form of the secreted TGF-β ligand inhibitor Coco (encoded by Dand5) (Fig. S2B). Subsequent studies indicated that Coco promotes lung colonization by inhibiting paracrine bone morphogenetic proteins (BMPs) signaling and validated the relevance of Coco for human breast-cancer metastasis (12). The other two cDNAs encoded full-length mitochondrial ribosomal protein L3 (Mrpl3) and a 5′ fragment of the long noncoding RNA Malat1 (metastasis-associated lung adenocarcinoma transcript 1) (Fig. S2 C and D). Consistent with their ability to mediate a cancer-specific function, MRPL3 is overexpressed in non-small cell lung cancers (14) and MALAT1 in non-small cell lung cancers and hepatocellular carcinomas (15, 16). Although we have not examined the mechanism by which these genes promote metastasis, MRPL3 promotes the translation of mRNAs encoding mitochondrial enzymes (17) and the biogenesis of miRNA-like small RNAs (miRNAs) (18), and MALAT1 is involved in alternative splicing (19) and transcriptional activation of growth-control genes (20). These results indicate that a gain-of-function cDNA screen in a mouse model of mammary-tumor progression enables the identification of single genes implicated in lung colonization.

Confocal imaging indicated that control 4TO7 cells infiltrate the lung as efficiently as those expressing Coco after tail-vein injection. However, whereas the 4TO7 cells enter into proliferative quiescence immediately upon penetrating into the lung stroma and remain in this state for the entire observation period, those expressing Coco undergo reactivation at around day 7 postinjection (Fig. 1B) (12). Additional experiments demonstrated that the number of 4TO7 cells remain constant for at least 2 mo in the lung stroma. These cells are viable, as indicated by their intact nuclear morphology and absence of reactivity with antibodies to cleaved Caspase 3 (Fig. S3 A–D). Moreover, they are not senescent because they do not exhibit senescence-associated heterochromatic foci or phosphorylated histone γH2AX (Fig. S3 E and F). These results demonstrated that the 4TO7 cells enter into bona fide solitary tumor dormancy upon infiltrating the lung.

Phenotypic and functional analyses demonstrated that the 4TO7 and 4T1 cells exhibit traits associated with cancer stem cells, such as the ability to form oncospheres in vitro and to initiate tumors in vivo (12). Moreover, immunofluorescent staining indicated that the 4TO7 cells express Vimentin, but not E-cadherin, both in vitro and during solitary tumor dormancy in the lung, indicating that they have passed through an epithelial–mesenchymal transition (EMT) and retain this state during dormancy (Fig. S4). The fully metastatic 4T1 cells exhibited similar EMT traits in vitro and during the initial phase of latency in the lung (Figs. S4A and S5). After reactivation, a fraction of outgrowing 4T1 cells exhibited reduced, but not absent, levels of Vimentin (Fig. S5 A and B). However, virtually none of the 4T1 cells, within either incipient or frankly outgrowing lesions, displayed detectable levels of E-cadherin (Fig. S5C). These results suggest that the 4T1 cells undergo reactivation without reverting to an epithelial phenotype.

A Gain-of-Function Platform for the Identification of Canonical Genes Involved in Metastatic Reactivation.

Building upon the identification of a faithful model of metastatic dormancy and reactivation, we designed an improved strategy to screen for genes specifically involved in metastatic reactivation (Fig. 1C). We generated 4TO7 cells expressing the triple-modality TGL reporter vector, which encodes for thymidine kinase, GFP, and luciferase transduced them with the libraries, and finally injected them directly into the tail vein of syngeneic mice. This approach bypasses the significant attrition that occurs during tumor initiation, primary tumor growth, and intravasation, thus increasing the throughput of the screen. In addition, bioluminescent imaging of luciferase activity enables a quantitative assessment of colonization, eliminating the need to dissect a large number of mice. Finally, expression of GFP facilitates the identification of small metastatic lesions under a fluorescent dissection microscope.

Preliminary experiments indicated that the 4TO7-TGL cells do not produce lung metastases when they are injected at 3 × 10⁵ in the tail vein of syngeneic mice (n = 20). After transduction with the three high-complexity cDNA libraries, the 4TO7-TGL cells were then injected at this dose. To improve the screening output, the recipient cells were inoculated in different numbers of mice depending on the level of complexity of the library with which they were transduced (Fig. 2A). Bioluminescent imaging and dissection indicated that several mice developed one or more lung metastases (Fig. S6). Identical inserts encoding full-length TM4SF1, a divergent tetraspanin, were recovered from 44 lesions, and identical inserts encoding ArhGEF2, a Rho-GEF, were recovered from 5 lesions (Table 1). In contrast to control 4TO7 cells, those stably transduced with TM4SF1 efficiently colonized the lung after tail-vein injection, suggesting that expression of TM4SF1 is sufficient to promote reactivation in the lung (Fig. 2B and Fig. S7A). In contrast, ArhGEF2 did not pass validation. Previous studies have indicated that TM4SF1 is associated with poor prognosis in squamous lung carcinoma (21) and mesothelioma (22). In addition, TM4SF1 combines with synthenin-2, which promotes melanoma and breast-cancer metastasis by activating protein kinase C and focal adhesion kinase/Src family kinase signaling (23–25).

Fig. 2.

Design of the gain-of-function cDNA library screen. (A) The flowchart describes the throughput and theoretical redundancy of the cDNA screen. (B) 4TO7-TGL cells stably transduced with empty vector or TM4SF1 were inoculated i.v. into syngeneic mice. Lung colonization was measured by bioluminescent imaging. The panels show representative images (Left), and the graph shows the normalized photon flux at the indicated times (Right). Note that the scale for normalized photon flux is logarithmic.

Table 1.

Results of the cDNA screen

Tm4sf1 encodes for the divergent tetraspanin TM4SF1 [also known as tumor-associated antigen L6 (TAL6)] and Arhgef2 for the Rho/Rac GEF 2 (also known as GEF-H1). Identical inserts encoding full-length TM4SF1 or Rho/Rac GEF 2 were recovered from 44 metastatic lesions. Two metastases, which presumably arose from the confluence of two distinct outgrowths, contained both inserts. Tm4sf1 passed validation and is indicated in red. Numbers in parentheses indicate the lesion numbers, which contain inserts encoding TM4SF1 or ArhGEF2. Mets, metastases; N/A, not available.

Gain-of-Function Screens Identify microRNAs That Promote Metastatic Reactivation.

Because of their ability to simultaneously regulate the level of expression of multiple genes, microRNAs may be particularly adept at coordinating gene-expression changes that are required to complete several steps of the metastatic cascade (26). Although prior studies have indicated that expression of miR-335 and miR-126 can suppress micrometastatic outgrowth by blocking reinitiation and neo-angiogenesis, respectively (27, 28), it is not currently known whether microRNAs can promote exit from dormancy and reactivation.

To identify microRNAs able to promote breast-cancer reactivation in the lung, we screened a library of mouse microRNAs cloned in the retroviral vector pMSCV (29). The original library was subdivided in five pools, each containing ∼70 microRNAs, and supplemented with an additional pool containing 27 microRNAs, which we had identified as up-regulated more than twofold in 4T1 cells compared with 4TO7 cells by using microRNA microarray analysis (Fig. 3A). Because pMSCV drives expression of GFP from an internal ribosome entry site (IRES), we could directly determine the titer of each pool and transduce the 4TO7 cells at an MOI able to drive expression of an average of one microRNA per cell. Cells transduced with each of the six pools of microRNAs were injected in the tail vein of five syngeneic mice to ensure complete screening of each collection of microRNAs (Fig. 3A). Five of the 30 mice developed multiple metastases in their lungs and 2 mice developed single lesions. Genomic sequencing of tumor cells isolated from these lesions revealed that they harbored nine integrated microRNA genes (Table 2). The microRNAs miR-340 and miR-346 and the combination of miR-138 and miR-223 were recovered from multiple lesions from different mice. To validate these hits, we constructed retroviral miR-Vec vectors expressing each of the nine microRNAs from the CMV promoter and infected 4TO7-TGL cells using single vectors or combinations of vectors. Q-PCR (quantitative PCR) was used to verify that the microRNAs were adequately expressed. Tail-vein injection followed by bioluminescent imaging indicated that expression of miR-138 or miR-346 efficiently induces lung colonization in the majority of mice (3/5), indicating that these microRNAs promote exit from dormancy in the lung (Fig. 3B and Fig. S7 B and C). In contrast, the other microRNAs did not pass validation. These results indicate that our screening platform permits the isolation of microRNAs that mediate metastatic reactivation.

Fig. 3.

Gain-of-function microRNA library screen. (A) The flowchart describes the throughput and theoretical redundancy of the microRNA screen. (B) 4TO7-TGL cells stably transduced with empty vector or the indicated microRNAs (miR-138, Left; miR-346, Right) were inoculated i.v. into syngeneic mice. Lung metastasis was measured by bioluminescent imaging. The panels show representative images (Top), and the graphs show the normalized photon flux at the indicated times (Bottom). Note that the end time point is 8 wk for miR-138 and relative controls and 6 wk for miR-346 and relative controls.

Table 2.

Results of the microRNA screen

The microRNAs miR-25 and miR-27a were recovered from a single lesion and did not pass validation upon reexpression, either alone or in combination. The microRNAs miR-138 and miR-223 were similarly recovered together from three lesions. miR-346 and miR-138 passed validation and are indicated in red. Numbers in parentheses indicate the lesion numbers, which contain inserts encoding corresponding miRNAs. BLU, bioluminescence units; N/A, not available.

shRNA Screens Identify Genes That Enforce Tumor Dormancy.

We reasoned that screening a library of shRNAs might enable the identification of genes that antagonize metastatic reactivation. Because they need to be inactivated for successful outgrowth, these genes can be viewed as enforcers of the dormant state. To test this hypothesis, we screened a modified version of the Cancer 1,000 mouse library (30, 31). The base library comprises ∼2,300 shRNAs targeting a list of ∼1,000 cancer-relevant genes (2–3 shRNAs per gene). To ensure potent knock down as single-copy integrants, the shRNAs were cloned in the context of an miR-30 backbone in an MSCV-based, GFP-expressing vector (32). Based on initial “reconstruction” studies, the Cancer 1,000 library was subdivided in six pools, each containing 288 shRNAs, and supplemented with a seventh pool comprising 260 additional shRNAs against cancer-related genes (1–7 shRNAs per gene) (33). After transduction with the seven pools of shRNAs from the modified Cancer 1,000 library, the 4TO7-TGL cells were injected into the tail vein of 49 recipient mice (Fig. 4A). A large fraction of mice (∼25%) developed metastatic lesions after 6 wk. Metastatic cells were isolated from 45 lung lesions, and the potentially prometastatic shRNAs that they harbored were sequenced and identified. Primary hits were prioritized based on their rate of recurrence. The target genes of shRNAs identified in 2 or more independent lesions encoded a variety of signaling proteins and transcription factors (Table 3). To validate these results, we conducted a secondary screen under similar conditions. Interestingly, shRNAs targeting three genes were recovered in both rounds of screening, strongly suggesting that these genes must be inactivated for exit from dormancy. Intriguingly, the genes targeted by these shRNAs constitute the Notch inhibitor Numb (34, 35), the histone methyl transferase Smyd5 (36), and the Smad E3 ubiquitin ligase Smurf2 (37). Because it is well established that Smurf2 antagonizes TGF-β signaling, which promotes metastasis (2, 37), we focused on Numb and Smyd5. Customary specificity controls in RNAi screens involve testing a suite of shRNAs targeting the gene or rescuing the effect produced by the original shRNA through expression of an shRNA-insensitive cDNA (9). Because we were interested in simplifying the overall phase of validation, we decided to examine whether ectopic expression of Numb or Smyd5 suppressed the ability of metastatic breast-cancer cells to colonize the lung. Vectors encoding these genes were introduced in mammary-tumor cells from MMTV-Neu mice because these cells provide a faithful model for HER2⁺ metastatic breast cancer (12, 38) (Fig. S7 D and E). Interestingly, bioluminescent imaging revealed that expression of Numb and Smyd5 almost completely suppresses the ability of ErbB2-transformed cells to metastasize to the lung, consistent with the hypothesis that Numb and Smyd5 mediate tumor dormancy (Fig. 4B). Although we cannot at present exclude that some of the other shRNAs identified silence genes that enforce dormancy, these results indicate that our screening strategy can lead to the rapid identification of single genes that enforce tumor dormancy.

Fig. 4.

Loss-of-function shRNA library screen. (A) The flowchart describes the throughput and theoretical redundancy of the shRNA screen. (B) ErbB2-TGL cells stably transduced with empty vector or the indicated genes (Numb, Left; Smyd5, Right) were inoculated i.v. into syngeneic mice. Lung metastasis was measured by bioluminescent imaging. The panels show representative images (Top), and the graphs show the normalized photon flux at the indicated times (Bottom).

Table 3.

Results of the shRNA screen

The primary screen yielded 45 hits. The table shows the targets of shRNAs recovered from more than two distinct lesions. Hits are rank-ordered based on their cumulative frequency of occurrence. The genes targeted by shRNAs also recovered in a secondary screen are indicated in red.

Bioinformatic Analysis of Existing Datasets Can Provide Useful Information on the Genes Identified.

Oncoprint analysis of the provisional breast cancer TCGA dataset indicated that TM4SF1, and the long noncoding RNA MALAT1 are amplified or overexpressed in 11% and 7% of the cases, respectively (Fig. S8A). Oncomine (www.oncomine.org/resource/login.html) analysis revealed that TM4SF1 is expressed at higher levels in basal-like compared with luminal-like tumors in the Farmer dataset. In contrast, SMYD5 is up-regulated in apocrine tumors compared with other tumors (Fig. S8B). NUMB was down-regulated in both basal-like and luminal-like tumors in the Farmer dataset and in all tumors in the Curtis dataset (Fig. S8C). Finally, MALAT1 was up-regulated in breast carcinoma compared with normal breast, and in lymph-node metastases compared with breast carcinoma in the Zhao dataset (Fig. S8D). No useful inference could be made regarding the other genes identified. Comprehensive mechanistic studies and clinical validation experiments will be necessary to establish the biological and clinical relevance of each of the genes identified.

Discussion

We have developed a screening strategy that uses the mouse as a filter to identify genetic entities that promote metastatic reactivation. Validation experiments have provided evidence that the canonical genes and microRNAs identified through this method are potent mediators of reactivation whereas the targets of shRNAs, and potentially those of microRNAs, prevent reactivation and are thereby implicated in enforcing dormancy. Thus, forward genetic screens in mice can identify genes that govern metastatic dormancy and reactivation.

The screening platform we have developed has several notable attributes. First, it enables a specific analysis of metastatic dormancy and reactivation. The recipient 4TO7 cells infiltrate the lung efficiently but immediately enter into solitary tumor-cell dormancy, enabling genetic analysis of this process and the subsequent reactivation step (12). In contrast, prior studies have not distinguished between solitary tumor-cell dormancy and micrometastatic dormancy, or they have focused on the latter (39–41). Moreover, the 4TO7 cells display features of cancer stem cells, such as the ability to form oncospheres in vitro and to initiate tumors in mice (12). Because of their migratory and invasive ability, as well as their ability to enter into quiescence and to undergo reactivation in response to specific stimuli, the 4TO7 cells provide a faithful model for migrating cancer stem cells (42). Our screening platform should therefore facilitate the identification of the core signaling pathways that drive the reactivation of these metastasis-initiating cells.

Second, our screens seem to be very stringent. The background is virtually negligible. The parental 4TO7 cells or those transduced with empty vector never give rise to macroscopic metastases, indicating that stochastic genetic or epigenetic events or insertional mutagenesis cannot convert these cells into fully metastatic cells at a detectable frequency. This conclusion is consistent with the hypothesis that a limited number of genes govern metastatic dormancy and reactivation. Furthermore, although we have injected the library-transduced 4TO7 cells directly into the bloodstream, thus bypassing the steps of metastasis leading to dissemination, the colonization phase of metastasis is characterized by significant attrition. Only a small fraction of the tumor cells, which are inoculated intravenously, survive in the bloodstream, extravasate into the stroma of the target organ, and avoid apoptosis (43–45). In fact, only tumor cells that adhere to the abluminal endothelial basement membrane receive signals that promote their survival, but not their proliferation, allowing entry into dormancy (46). Based on prior results (12), we have estimated that only ∼8% of injected 4TO7 cells survive initial attrition and successfully enter into dormancy in the lung. Moreover, only a small fraction of the tumor cells that have entered into dormancy eventually undergo reactivation, suggesting that formidable barriers oppose metastatic reactivation (43–45). We suspect that the power of our screens stems from both the low level of background and the selectivity of the step of metastasis examined.

Third, our screening strategy is scalable. Currently, genome-wide screens are limited by constraints in the generation of libraries encoding or targeting the entire exome, in the rate of transduction of recipient cells, and in proper regulation of gene expression. Despite the suboptimal coverage of our libraries, we have been able to identify several strong mediators of reactivation by performing a limited number of screens. Based on these results, we envision that repeated or larger-scale screening of existing libraries or screening of newly generated genome-wide libraries, comprising full-length cDNAs or guide RNAs for CRISPR-Cas9 genome editing (47, 48), will enable isolation of additional genes that govern metastatic dormancy and reactivation.

Finally, the platform we have described is flexible and can be adapted to the study of metastatic reactivation in other organs and by other cancer types or to the study of other steps of metastasis. Unlike DNA microarray-based approaches for the identification of candidate mediators of metastasis, our gain-of-function genetic screening strategy affords the advantage of rapid biological validation of single prometastatic entities. Because of its attributes, our screening platform should contribute to the identification of the core signaling pathways that regulate metastatic dormancy and reactivation and possibly facilitate the development of novel therapeutic strategies.

Materials and Methods

Retroviral cDNA, microRNA, and shRNA screens were performed in 6- to 12-wk-old BALB/c mice. Animals were imaged in an IVIS 100 chamber, and data were recorded using Living Image software (Xenogen). All mouse studies were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee of the Memorial Sloan Kettering Cancer Center (MSKCC). Results are reported as mean ± SD or SE. Comparisons between two groups were performed using an unpaired two-sided t test (P < 0.05 was considered significant).