The miR-199a/Brm/EGR1 axis is a determinant of anchorage-independent growth in epithelial tumor cell lines

In epithelial cells, miRNA-199a-5p/-3p and Brm, a catalytic subunit of the SWI/SNF complex were previously shown to form a double-negative feedback loop through EGR1, by which human cancer cell lines tend to fall into either of the steady states, types 1 [miR-199a(−)/Brm(+)/EGR1(−)] and 2 [miR-199a(+)/Brm (−)/EGR1(+)]. We show here, that type 2 cells, unlike type 1, failed to form colonies in soft agar, and that CD44, MET, CAV1 and CAV2 (miR-199a targets), all of which function as plasma membrane sensors and can co-localize in caveolae, are expressed specifically in type 1 cells. Single knockdown of any of them suppressed anchorage-independent growth of type 1 cells, indicating that the miR-199a/Brm/EGR1 axis is a determinant of anchorage-independent growth. Importantly, two coherent feedforward loops are integrated into this axis, supporting the robustness of type 1-specific gene expression and exemplifying how the miRNA-target gene relationship can be stably sustained in a variety of epithelial tumors.

C hromatin remodeling factors play vital roles in epigenetical regulation via genome-wide gene transcription 1 . On the other hand, microRNAs (miRNAs) are post-transcriptional regulatory molecules that are involved in diverse biological processes, including development, differentiation, and homeostasis 2 . Growing evidence indicates that the robustness of gene expression is often supported by coordinated transcriptional and miRNA-mediated regulatory networks 3,4 . In addition, improper use of these networks may lead to human diseases such as cancer. However, the interplay between chromatin remodeling factors and miRNA, as well as its biological outcome, is not fully understood in the context of gene regulatory networks common to a wide variety of cell lines.
The human SWI/SNF-A complex (also known as the BAF complex), a member of a family chromatin remodeling factors 5 composed of about 10 proteins, regulates gene transcription, either positively or negatively. The SWI/SNF complex contains a single molecule of either Brm or BRG1 as ATP-dependent catalytic subunits. Brm and BRG1 regulate target promoters that do not fully overlap and show clear differences in their biological activities [6][7][8][9] . This SWI/SNF complex interacts with various proteins, including transcriptional regulators, through many specific and varied associations with its several subunits. For example, the d4-family proteins DPF2 (REQ) and DPF3a/3b function as efficient adaptor proteins for RELB/p52 10 and RELA/p50 11 dimers to induce SWI/SNFdependent NFkB target genes.
In terms of human cancers, we and other groups have reported that Brm is frequently undetectable in various cancer cell lines 12 , and in primary tumors of the lung 13 , stomach 14 , and prostate 15 . We found in nuclear run-on transcription assays that a functional Brm gene was present and actively transcribed in all of the Brm-deficient cancer cell lines tested 12,16 , indicating that Brm expression is largely suppressed by post-transcriptional gene silencing. Brm was later shown to be efficiently targeted by both miR-199a-5p and miR-199a-3p 17 . In addition, Brm acts as a potent negative regulator of endogenous EGR1 gene expression. EGR1 activates the miR-199a (2) gene locus, which is mainly responsible for the biogenesis of mature miR-199a-5p and -3p in these cancer cell lines. Overall, these findings suggest that, in the cell lines examined, Brm and miR-199a form a robust doublenegative feedback loop that includes EGR1 17 . By examining a panel of human cell lines that were derived from a wide variety of cancer tissues, we found that they tend to fall into either of the steady states, miR-199(2)/Brm(1)/ EGR1(2) cells and miR-199a(1)/Brm(2)/EGR1(1) cells 17 , denoted hereafter as type 1 and type 2, respectively. These regulatory networks may explain why variable (either higher or lower) expression of miR-199a-5p/-3p 18 or EGR1 19 has been reported among many carcinomas when compared with the normal epithelial tissues from which they originated.
In the early stage of our current study, we noticed clear differences in the biological properties between type 1 and type 2 cells: all of the type 1 cell lines tested (8 lines), but no type 2 cell lines (4 lines), could grow in soft agar, providing us with an unprecedented opportunity to unravel the robust regulatory networks involved in anchorage-independent growth common to these cancer cell lines. Of course, the gene expression patterns of each cancer cell line would be expected to be largely cell line-specific and dependent on a wide variety of factors, including the originating tissue type, mutated genes, and pathological properties, such as the tumor stage. However, in our current study, we speculated that epithelial tumors would share regulatory networks that control their basic biological activities. In addition, we hypothesized that several genes would be specifically expressed in type 1 cancer cells, but not in type 2, and, further, that some of them would be crucial for their anchorage independency. Here, we have identified several genes specifically expressed in type 1 cells and show that single knockdown of some of these genes is sufficient to suppress the colony-forming activity of type 1 cells in soft agar. We further examined the underlying molecular mechanisms of the all-or-none regulation of these type 1-specific genes in the two cell types, leading to the identification of two coherent feedforward loops associated with the miR199a/Brm/EGR1 axis. We finally present evidence that these type-specific gene expression patterns can be recapitulated in tumor some lesions of non-small-cell lung carcinomas (NSCLCs).

Results
Type 1, but not type 2, cells grow efficiently in soft agar. For type 1 and type 2 cells, we chose 12 cell lines (8 for type 1 and 4 for type 2) originating from various human epithelial tumors (Fig. 1a). To ensure that we only examined cell lines originating from epithelial tumors, PA-1 (originating from tridermic teratocarcinoma), MDA-MB435 (recently discovered to originate from melanoma), and HEK-293FT (originating from human embryonic kidney) were removed from the panel of cancer cell lines previously used for categorization 17 . In addition, cell lines from one pancreatic (Panc-1) and three colon cancers (DLD-1, HT29, and HCT116) were added to the current analysis.
The results of a series of quantitative reverse transcription-polymerase chain reaction (RT-PCR) experiments ( Fig. 1a) confirmed EGR3, and EGR4 mRNA and mature miR-199a-5p, -3p, and miR-214 were determined by quantitative RT-PCR. The data represent the means 6 S.D. (n53). (b) Numbers of colonies formed in soft agar by each cell line. Numbers of colonies (more than 150 mm in diameter) in soft agar were counted 21-28 days after 1,000 cells were seeded in 60-mm plates. The data represent the means 6 S.D. (n 5 3) In a-b, P values were determined using Mann-Whitney test. (c) Four type 1 cell lines transduced with retroviral vectors expressing shBrm (#4 or #5) or shCre#4 (negative control) were seeded as in b. Colony numbers of shBrm-expressing cells were compared with those of shCre#4-expressing cells and the ratio was shown as a percentage. The data represent the means 6 S.D. (n 5 4). Asterisks indicate P value, compared with those transduced with shCre#4. *P,0.05, **P,0.01, ***P,0.001. www.nature.com/scientificreports SCIENTIFIC REPORTS | 5 : 8428 | DOI: 10.1038/srep08428 our previous observations that epithelial tumor cell lines can be classified into two types according to the expression levels of Brm, EGR1, and miR-199a; type 1 cells specifically express Brm mRNA, whereas expressions of EGR1 mRNA and miR-199a-5p and -3p, as well as miR-214, which is also generated from the miR-199a (2) gene locus, are restricted to type 2 cells. Notably, like EGR1, EGR2, EGR3, and EGR4, which are the other members of EGR family gene and which recognize the same DNA sequence 20 , were shown to be type 2specific ( Fig. 1a) by the parallel analysis. This might indicate EGR2, EGR3, and EGR4 are involved in the miR-199a/Brm axis in a similar manner to EGR1.
After several preliminary comparative analyses between type 1 and type 2 cells, we noticed clear differences in terms of anchorage-independent growth. type 1 cells formed 25-300 colonies (more than 150 mm in diameter) in soft agar when 1,000 cells were seeded per 60-mm plate and kept for 21-28 days (Fig. 1b). None of the four type 2 cells formed clear colonies in the same conditions. Notably, all of the type 2 cancer cell lines tested-SW13 21 , HuTu80 22 , H522 23 and C33A 24 -have shown clear tumor-forming activity in mouse xenograft models.
To test whether the anchorage-independent growth of type 1 cells requires Brm, we performed Brm knockdown experiments in several type 1 cells-A549, H1229, HeLaS3, and Panc-1-using a set of retroviral vectors containing shBrm (#4 and #5). We confirmed that all shBrm vectors significantly suppressed the levels of Brm mRNA and its product (see below) and that cells transduced with these vectors reduced colony-forming activity in soft agar when compared with that of negative control cells transduced with the shCre#4 vector ( Fig. 1c). These results reveal the pivotal role of Brm in anchorageindependent growth in type 1 cells. In several cases, strong suppression in colony formation in soft agar was observed by Brm knockdown, whereas the same culture grow normally when kept in monolayer culture (A549 cells expressing shBrm#5 and HeLaS3 and Panc-1 cells expressing either shBrm#4 or #5, Supplementary Fig. 1). In the case of H1299, however, we cannot exclude the possibility that reduction in the growth rate of monolayer culture partly contributed reduction in anchorage-independent growth. These findings provided us with an excellent opportunity to uncover the critical genes required for anchorage-independent growth of type 1 cells and suggested that these candidate genes would be expressed in a type 1-specific manner.
Several genes are preferentially expressed in type 1 cells. Whereas we know that suppression of the expression of a target protein by a certain miRNA is usually moderate and is not unconditionally retained in steady states, we selected the candidates of type 1specific genes from various targets of miR-199a-5p (10 genes tested), miR-199a-3p (11 genes tested), and miR-214 (6 genes tested). These target genes were identified in previous reports or were predicted by target prediction algorithms as well as by our own analysis (Supplementary Table 1). Of 32 candidate genes tested by quantitative RT-PCR ( Fig. 2a and Supplementary Figs. 2 and 3), CAV1 25 and KRT80 (both miR-199a-5p target genes) and CD44 26 , MET 27 , and CAV2 28 (all miR-199a-3p targets) were expressed in most of the type 1 cells but in none of the type 2 cells (Fig. 2a). We also performed FACS analysis for CD44 and MET using . P values were determined using Mann-Whitney test. (b) Protein expression levels of Brm, BRG1, and four type 1-specific genes products as well as five type-independent gene products were analyzed by western blotting using 10 gels. The fulllength blots were presented in the supplementary Figure 6. b-actin was used as the loading control for each gel. Analysis of KRT80 protein was not possible because of the lack of a specific antibody. Relative expression levels of each protein were quantified in Supplementary Figure 4. H1299 cells (type 1), and found that the entire population showed high levels of both CD44 and MET (data not shorn). We thus designated them as type 1-specific genes. Interestingly, two other epithelial-type keratin genes 29,30 , KRT7 31 (an miR-199a-3p target) and KRT19 (an miR-199a-5p target), were not expressed in any of the type 2 cells but were expressed in some of the type 1 cells (Supplementary Fig. 2 and 3).
The protein expression profiles of most of these type 1-specific genes, as well as four genes whose mRNA was expressed in both types, were examined by western blotting (Fig. 2b) and relative amounts of each protein were quantified ( Supplementary Fig. 4). The protein expression profiles were generally very similar to those of the corresponding mRNA (Fig. 2a). Mann-Whitney test further confirmed that these miR-199a target gene products (CD44, MET, CAV1 and CAV2) are specifically expressed in type 1 cells ( Supplementary Fig. 4). KRT80 protein was not analyzed because of the lack of a specific antibody. PTEN protein, a well-known miR-214 target, was clearly expressed in all of the cell lines except C33A cells, because C33A have homozygous nonsense mutation of this gene. It is noteworthy that among these type 1-specific proteins, CD44, MET, CAV1 and CAV2 function as plasma membrane sensors and signaling platforms and can be colocalized in caveolae in several physiological conditions 32 .
A single knockdown of CD44, MET, CAV1, or CAV2 is sufficient to suppress anchorage-independent growth of type 1 cell lines. To evaluate whether the type 1-specific genes identified above contribute to the anchorage-independent growth of this cell type, we developed pairs of shRNA constructs for CD44, MET, CAV1, and CAV2 capable of efficiently suppressing their target gene products ( Supplementary Fig. 5). Specific knockdown of KRT80 by short hairpin was not possible because there are too many conserved regions among the large keratin gene family paralogues. After A549, H1299, HeLaS3, and Panc-1 cells were transduced with the shRNAexpressing retroviral vectors, their colony-forming activity in soft agar was evaluated (Fig. 3). Single knockdown of CD44, MET, CAV1, or CAV2 efficiently suppressed colony formation compared with negative control cells expressing shRNA for Cre#4. As an exception, the colony number of A549 cells expressing shCAV2#2 was slightly increased, probably due to off-target effects. Knockdown of CD44, MET, CAV1, and CAV2 did not significantly affect cell growth. These results indicate that all of the four type 1-specific genes tested significantly contribute to the anchorage-independent growth of these type 1 cell lines.
All-or-none expression patterns of some type 1-specific genes in the cell line panel are supported by two coherent feedforward loops that associate with the axis. Given that miRNA usually suppresses its target protein in a modest manner, it might be somewhat unexpected that some miR-199a targets were regulated in an all-or-none manner between type 1 and type 2 cells (Fig. 2a,b). We speculated that this all-or-none phenomenon could be reflecting regulation by the molecular switch through miR-199a/Brm/EGR1 axis, where Brm and miR-199a expressions manifest a mutually exclusive pattern. Therefore, we first tested whether these type-1 specific genes, CD44, MET, CAV1, CAV2, and KRT80 genes are under the positive control of the Brm-type SWI/SNF complex, for which type 1 cells are competent.
To test whether type 2 cells can induce type 1-specific genes when Brm is exogenously introduced, we transfected SW13 cells (type 2) with Brm expression plasmid or empty plasmid (EV1). In these experiments, some parallel cultures were cotransfected with the expression plasmids for representative NFkB dimers-RELA/p50 (canonical pathway) and RELB/p52 (noncanonical pathway)-or empty plasmid (EV2) to determine whether the activation is enhanced by NFkB dimers. As shown in Fig. 4a, CD44, MET, CAV1, and KRT80 mRNA were induced by Brm, as judged by quantitative RT-PCR, although the Brm induction effect varied among the genes. CD44 and CAV1 mRNA levels were increased by cotransfection with NFkB dimers and Brm, whereas expression of MET gene was not NFkB dependent at all. CD44 expression was more strongly dependent upon the noncanonical dimer RelB/p52 than RelA/p50, consistent with a recent report 33 . CAV1 expression was further increased by cotransfection with RelA/p50 and Brm but was also significantly induced by RelA/p50 alone. Therefore, CAV1 would only require Brm for its full expression. On the other hands, the  expression levels of GSK3b and Sirt1 (miR-199a-5p targets), and PTEN (a miR-214 target), which did not show type 1-specific expression patterns, were not affected by high levels of either Brm or NFkB in SW13 cells (Fig. 4a). Next, A549, H1299, and HeLaS3 cells (type 1) were transduced with retroviral vector encoding Brm shRNA, and the steady-state expression of CD44, MET, CAV1, CAV2, and KRT80 was evaluated by quantitative RT-PCR (Fig. 4b). The levels of CD44, MET, and KRT80 mRNA were suppressed to various extents by Brm knockdown. Western blot analysis of parallel A549 cultures also indicated that CD44 and MET, but not CAV1 and CAV2, required Brm for expression (Fig. 4c) 7 . This strong Brm-dependency of CD44 expression is consistent with the previous report that assorted tissues from Brm null/BRG1-positive mice lack CD44 expression. Overall, these findings indicated that genes that are suppressed by miR-199a and simultaneously require the Brm-type SWI/SNF complex for efficient expression show distinct expression patterns: expression in type 1 cells but no expression in type 2 cells. But CAV1 and CAV2 expression failed to show clear Brm dependency in A549 and H1299 cells.
We next tested whether type 1-specific genes are under the negative control of EGR1. When HeLaS3 and A549 cells were stably transduced with EGR1-expressing retrovirus, endogenous miR-199a-3p levels were elevated as expected from the axis (Fig. 5a). In HeLaS3, levels of MET, CAV1 and CAV2 mRNA (Fig. 5a) and their gene products ( Fig. 5b and Supplementary Table 2) were reduced by exogenous EGR1 expression. In the case of A549 cells, slight reduction of CAV1 and CAV2 mRNA and reduction of MET, CAV1 and CAV2 proteins were observed ( Fig. 5ab and Supplementary Table2). These results are consistent with previous reports indicating that MET 34 and CAV1 35 genes are negatively regulated by EGR1: the MET and CAV1 promoters have one and three EGR/SP-1 binding sites, respectively. We also found EGR1 binding sites on the CAV1, CAV2 and MET promoter regions (from -1600 to 1500bp of TSS) by using ChIP-seq data obtained by ENCODE. These results suggest that CAV1/2 and possibly MET are specifically expressed in type 1 cells by evading transcriptional suppression by EGR1 proteins and also post-transcriptional suppression by miR-199a-5p/3p.
The miR-199a/Brm/EGR1 axis persists in an extended panel of cell lines originating from epithelial tumors. Because our panel of cancer cell lines used for the development of the cell line typing was limited to 14 cell lines, we intended to increase the number of cell lines by directly performing quantitative RT-PCR of Brm mRNA (using totally 4 PCR primer pairs), EGR1 mRNA (using totally 4 PCR primer pairs) and miR-199a-3p by adding 12 new cell lines using the same experimental protocol as used for Fig. 1a (Fig. 7). The 4 independent PCR primer pairs for Brm or EGR1 gave essentially the same expression profile, respectively (Fig. 7a). Out of the total 26 cell lines, 23 could be categorized as either type 1 (17 lines) or type 2 (6 lines), according to the criteria shown in Supplementary Table 3. The remaining 3 lines, which were originated from gastric carcinomas and mammary tumors, cannot be categorized into either type 1 or type 2 (designated type 3). These results indicate that the miR-199a/Brm/EGR1 axis is largely retained in variety of epithelial tumor cell lines.
Since microarray data of Brm and EGR1 mRNA for 17 among these cell lines categorized as type 1 and type 2 were available from Sanger database (Genomics of Drug Sensitivity in Cancer http:// www.cancerrxgene.org), and their expression profiles obtained from the database was compared with those of the qRT-PCR data shown in Fig. 7a (Supplementary Fig. 9). We found that the expression profiles of Brm and EGR1 are not correlates well between them. Since Brm mRNA levels of even Brm-deficient cell lines such as SW13, H522, C33A, A427, and H23-previously reported by our 6,12 and other groups 7,36 by RT-PCR or Northern blotting analysis-were significantly high according to Sanger database, there would be limitations Relative expression levels of miR-199a-3p and type 1-specific mRNAs were determined by quantitative RT-PCR in A549 and HeLaS3 cells which were transduced with retroviral vectors expressing EGR1. The expression levels of cells transduced with empty vector was taken as 1.0. The data represent the means 6 S.D. (n 5 3). Asterisks indicate P value, compared with those transduced with empty vector. *P,0.05, **P,0.01, ***P,0.001 (b) Analysis of type 1-specific gene products and EGR1 in the parallel cultures prepared in a by western blotting. b-actin was used as the loading control. The full-length blots were presented in the supplementary Figure 8 and relative expression levels of each protein including two additional sets of blots were quantified in Supplementary in microarray data to estimate mRNA levels of such transcriptional regulatory genes as Brm and EGR1 accurately.
Since we found expression profiles of CD44, MET, CAV1 and CAV2 mRNA by our qRT-PCR and those obtained from Sanger Database are correlated well, we showed both of them in Fig. 7b. The expression levels of CD44 and MET were high in type 1 cell lines, whereas they were mostly undetectable in the type 2 cell lines even in these extended panels, and specific CD44 and MET expression in type 1 cells were statistically supported in both data of qRT-PCR and the Database. CAV1 and CAV2 expression was not detected in most type 2 cells with a clear exception of A427. Because of this, type 1 specific expression of CAV1 and CAV2 was not supported statistically. Relatively low EGR1 expression in A427 among type1 cells (Fig. 7a, Supplementary Table 3) might partly explain this exception.
Expression patterns observed in type 1 or type 2 cell lines are recapitulated in some cancer lesions of NSCLCs. We finally examined whether the distinct expression patterns observed between the two cell types are reflected in human primary tumors. Since the cell lines originating from NSCLCs in the cell line panel used here can be categorized as both type 1 and type 2 (Supplementary Table 3), we pathologically analyzed surgically resected, formalin-fixed, paraffin-embedded tissues from human cancer lesions of NSCLCs. Among NSCLCs, we especially focused upon squamous cell carcinoma (SCC), because in this type of cancer, we can easily understand activity of proliferation or the status/ direction of differentiation two-dimensionally in the histological section. After preparing sequential thin sections of total 21 SCC cases, they were immunohistochemically stained with antibodies against Brm, CD44, MET, and CAV1 and also probed for miR-199a-5p by in situ hybridization and interrelationships among their expression patterns in the coincident area of the each section were analyzed by comparing lower and higher differentiation status.
In the area of lower differentiation status where cancer cells are crowded by the active proliferation and have increased nuclear/cytoplasmic ratio without keratinization, we clearly observed a Brm 1 , CD44 1 , MET 1 , and CAV1 1 phenotype in almost all cases. In some of these areas, expression of miR-199a was undetectable as shown in Fig. 8 (surrounded by solid line), which recapitulates the expression patterns of type 1 cells. However, in the other areas expressing these 4 proteins, we detected also miR-199a expression, indicating expression heterogeneity in cancer legions.
As for the areas of highly differentiation status, we observed them in many so-called cancer pearls in 4 cases of SCC, where cancer cells are sparse with large cytoplasm. Even in cancer pearls, significant population at the periphery retains clearly recognized nuclei indicating that cells are still alive, but in the central regions, cells are gradually losing their nuclei on their process of keratinization. We detected a Brm 2 , CD44 2 , MET 2 , and CAV1 2 , and miR-199a 1 phenotype in all of the cancer pearls where the cell retained nuclei, which recapitulated that of type 2 cells (Fig. 8 within the broken line). In the area between solid line and the cancer pearl in Fig. 8, where tumor cells assumed intermediate differentiation status, these 4 proteins and miR-199a were weakly expressed with various extents. At least in these regions, tumor cells might be undergoing changes from the type 1 cells into the type 2 cells through the process of cellular differentiation.

Discussion
Using 12 cell lines that were strictly derived from human epithelial tumors, we can confirm the findings of our previous report that these cells can be classified into type 1 [mir-199a(2)/Brm(1)/EGR1(2)] (8 lines) or type 2 [miR-199a(1)/Brm(2)/EGR1(1)] (4 lines) cells. In our current study, we were able to efficiently identify the type 1specific genes by setting the reported miR-199a and miR-214 target genes as the candidates. Some of the identified type 1-specific genes (CD44, MET, and KRT80) required Brm, whereas others (CAV1, CAV2 and probably MET) required the absence of EGR1 for their efficient expression, indicating that two coherent feedforward loops are formed (Fig. 6a). These two feedforward loops are integrated into the robust double-negative feedback loop forming a regulatory network that functions as an efficient switch that determines the expression levels of these type 1-specific genes in an all-or-none manner (Fig. 6b). Thus, the current situation would be a good example of a network formed by multiple miRNA-mediated feedback and feedforward loops 37,38 , which are commonly present in a wide variety of cell lines. Importantly, we observed regions whose expression patterns recapitulated those of type 1 or type 2 cells by pathological analysis of SCC legions of NSCLC tissues. It should be pointed out, however, that there are several lesions whose expression patterns do not belong to either of them. We speculate that in the process of cell line establishment from primary tumors, they would tend to fall into either of steady states, type 1 or type 2 cells. Among the type 1-specific genes shown here, CD44, MET, CAV1, and CAV2 alone significantly contributed to anchorage-independent growth of type 1 cells when tested in knockdown experiments using four type 1 cell lines (Fig. 3). Several previous reports indicated that CD44 39 , MET 40 , and CAV1 41,42 are potentially important for colony formation in some epithelial tumor cell lines. We observed that these four genes are all simultaneously suppressed to a marginal level in type 2 cells by the regulatory network shown here, ensuring the anchorage dependency of type 2 cells. Whereas CD44 43,44 , MET 45 , CAV1 46 , and CAV2 47 have their own multiple downstream signaling  Fig. 1 (Brm-1, EGR1-1). The relative expression levels are shown by taking the highest as 1.0. Two red vertical break lines indicate the boundary among each type cell lines. Detailed criteria for type1-3 cells were indicated in Supplementary Table 3  pathways, their interplay would also contribute to anchorage-independent growth or metastasis 48,49 . Importantly, CD44 and MET were reported to co-localize with CAV1 in caveolae 50 , and CAV2 is also localized in caveolae when it forms hetero-oligomers with CAV1 51 . In normal epithelial cells, caveolae function as plasma membrane sensors, responding to changes in extracellular matrix via integrin signaling and also as interacting domains with cytoskeltones 32 . By unregulated expression of these four proteins in type 1 cells, their intimate and coordinated interactions would generate strong downstream signaling preferable to grow in soft agar. Therefore, in normal epithelial cells, we expect miR-199a-5p and -3p would fine-tune caveolin function such as homeostasis for plasma membrane integrity, signaling platforms, cytoskeleton remodeling and cell migration. In addition to overlapping subcellular localizations of these gene products, it should also be pointed out that the genetic loci of MET, CAV1, and CAV2 all reside in the fragile chromosomal region, FRA7G at 7q31.2 52 . These regions seems to be neither amplified nor deleted based on the copy number data compiled in CCLE (https:// ccle.ucla.edu/); copy number data of eight cell lines (data of SW13, C33A, HeLaS3 and KB are not available) and the copy numbers of this region are distributed from 1.72 to 2.61.
Currently, we cannot fully explain why type 2 cells can form tumors when tested in xenograft models using immunodeficient mice but fail to form colonies in soft agar in vitro. Our preliminary analyses indicated that type 2 cells express TNF-a, IL-1a, or CCL5, whereas they do not express Brm-dependent genes, IL-6 and IL-8, whose products support cell-autonomous growth in soft agar 53 . When type 2 cells are introduced into mice, their production of TNF-a, IL-1a, or CCL5 is expected to induce the production of cytokines such as IL-6 and IL-8 in the associated fibroblast-like cells, which might in turn function as paracrine factors in tumor formation.
Stable Brm knockdown in type 1 cell lines by short hairpin retroviral vectors reduced the expression levels of several type 1-specific genes when examined within 2 weeks after the transduction, but cloning of these shBrm-expressing cells in culture is usually difficult 54 . These genes are induced by the transient expression of Brm in type 2 cells (Fig. 4a), but type 2 cells stably expressing exogenous Brm are also difficult to establish, as observed by our group 6 and others 55 . These observations indicate that transition from either type 2 to type 1 or from type 1 to type 2 is inevitably partial and transient, suggesting that both type 1 and type 2 cells are strongly tied to their own state after the loops are established and therefore stable switching to the opposite type would be difficult. This robustness of each type may reflect some aspects of ''oncogene addiction'' 56 or ''oncomiR addiction'' 57 .
The results in our present study reveal that in normal cells, the interplay between chromatin remodeling factors and miRNAs would fine-tune plasma membrane sensors by several motifs including the miR-199a/Brm/EGR1 axis and two feedforward motifs detected here. These motifs, once misapplied during the process of carcinogenesis, would finally fix the cancer cells to extreme steady states, which cannot be easily reversed. We believe our current findings will give us clues to elucidate how the homeostatic balance is abrogated at cancer initiation to establish type 1 or type 2 tumors and how to guide the development of distinct therapeutic strategies in each case.
Plasmid preparation for retroviral vectors expressing shRNA. Pairs of oligonucleotides encoding gene-specific short hairpin RNA (shRNA) were synthesized (Supplementary Table 4) and inserted between the BbsI/EcoRI sites of pmU6. The pmU6 derivatives shCre#4 59 [used as negative control (NC)] and shBrm#4 17 were previously described. These pmU6-based plasmids were digested with BamHI and EcoRI and inserted between these sites in pSSSP for the retroviral vectors.
DNA transfection and preparation of retrovirus. For the transfection of plasmid vectors into cell lines, Lipofectamine 2000 (Invitrogen Corp.) was used in accordance with the manufacturer's instructions. The preparation and transduction of vesicular stomatitis virus-G (VSV-G) pseudotyped retroviral vectors were performed as described previously 10 .
Quantitative RT-PCR. Total RNA was extracted using a mirVana microRNA Isolation Kit (Ambion). To detect coding gene mRNAs, cDNA was synthesized with a PrimeScript TM RT Reagent Kit with gDNA Eraser (Perfect Real Time) (TaKaRa Bio) in accordance with the manufacturer's instructions. Quantitative (real-time) RT-PCR was performed using a SYBRH Select Master Mix (Applied Biosystems). GAPDH was used as an internal control. The primer pairs used are listed in Supplementary Table 4. For the detection of miRNA, miRNA-specific looped RT-primers and TaqMan probes were used as described by the manufacturer's protocol (Applied Biosystems). RNUB6 RNA was used as an internal control. PCRs were performed in triplicate using a 7300 Real-Time PCR system (Applied Biosystems).
Immunohistochemical staining. Deparaffinization, endogenous peroxidase inactivation, and antigen retrieval of formalin-fixed, paraffin-embedded clinical tissues and immunostaining of Brm were performed as described previously 17,60 . For CD44 (#3570; Cell Signaling), MET (ab51067; Abcam), and CAV1 (sc-894; Santa Cruz) immunostaining, the sections were incubated overnight at room temperature with the corresponding antibody used for western blotting and washed in phosphatebuffered saline. N-HistofineH Simple Stain TM MAX PO (MULTI) (414154F; Nichirei Biosciences Inc.) were then applied to the slides for 30 min at room temperature, followed by three washes in PBS. The reaction products were visualized using a 50 mg/dL 3,39-diaminobenzidine tetrahydrochloride solution containing 0.003% H 2 O 2 . The immunostained sections were evaluated independently by two pathologists in conjunction with hematoxylin and eosin-stained sections from the same lesions.
In situ hybridization. In situ hybridization analysis to detect miR-199a-5p in formalin-fixed, paraffin-embedded sections was performed using LNA-modified oligonucleotide probes as described previously 17,60 . Use of the clinical tissue sections in this study was approved by the Fujita Health University ethical review board for human investigation.
Statistical analysis. Results are presented as means 6 S.D. Statistical significance for quantitative RT-PCR assays was determined using a two-tailed Student's t-test. Statistical significance for the differences of a parameter between type 1 and type 2 cell lines was determined using Mann-Whitney test. In both cases, P-values,0.05 were considered statistically significant.