Establishment of a TGFβ-Induced Post-Transcriptional EMT Gene Signature

A major challenge in the clinical management of human cancers is to accurately stratify patients according to risk and likelihood of a favorable response. Stratification is confounded by significant phenotypic heterogeneity in some tumor types, often without obvious criteria for subdivision. Despite intensive transcriptional array analyses, the identity and validation of cancer specific ‘signature genes’ remains elusive, partially because the transcriptome does not mirror the proteome. The simplification associated with transcriptomic profiling does not take into consideration changes in the relative expression among transcripts that arise due to post-transcriptional regulatory events. We have previously shown that TGFβ post-transcriptionally regulates epithelial-mesenchymal transition (EMT) by causing increased expression of two transcripts, Dab2 and ILEI, by modulating hnRNP E1 phosphorylation. Using a genome-wide combinatorial approach involving expression profiling and RIP-Chip analysis, we have identified a cohort of translationally regulated mRNAs that are induced during TGFβ-mediated EMT. Coordinated translational regulation by hnRNP E1 constitutes a post-transcriptional regulon inhibiting the expression of related EMT-facilitating genes, thus enabling the cell to rapidly and coordinately regulate multiple EMT-facilitating genes.


Introduction
Traditional gain-of-function and loss-of-function approaches have yielded an enormous amount of information in regards to gene function in mammalian development and disease. However, changes in mRNA levels are not always correlative with changes in protein abundance, underlying the importance of post-transcriptional regulation during control of gene expression and activity [1]. Indeed, during germ cell development, it has been demonstrated that the 39-untranslated regions , when fused to a reporter, are sufficient to confer temporo-spatial specificity for 80% of genes tested [2]. Thus, it is clear that the UTRs of mRNA transcripts can significantly impact gene expression. The 'human genome project' reported the mean lengths of 59-untranslated regions (59-UTRs) and 39-UTRs of human mRNAs as 300nt and 770nt, respectively, compared to the mean coding length of 1340 nt [3,4], generating renewed interest in the 39-UTRs of mRNAs to map post-transcriptional regulatory activities.
The epithelial-mesenchymal transition (EMT), in which cells undergo a developmental switch from a polarized, epithelial phenotype to a highly motile fibroblastic or mesenchymal phenotype, has emerged not only as a fundamental process during normal embryonic development and in adult tissue homeostasis, but is also aberrantly activated during metastatic progression [5,6,7]. EMT is associated with changes in cell-cell adhesion, remodeling of extracellular matrix, and enhanced migratory activity, all properties that enable tumor cells to metastasize [5,6,7]. Numerous cytokines and autocrine growth factors, including TGFb, have been implicated in EMT [8,9]. Our previous studies [10,11] and those of others [12,13] have shown that regulation of gene expression at the post-transcriptional level plays an indispensable role during TGFb-induced EMT and metastasis. We identified a transcript-selective translational regulatory pathway in which a ribonucleoprotein (mRNP) complex, consisting of heterogeneous nuclear ribonucleoprotein E1 (hnRNP E1) and eukaryotic elongation factor 1A1 (eEF1A1), binds to a 39-UTR regulatory BAT (TGFb activated translation) element and silences translation of Dab2 and ILEI mRNAs, two transcripts which are involved in mediating EMT [10,11]. TGFb activates a kinase cascade terminating in the phosphorylation of hnRNP E1, by isoform-specific stimulation of protein kinase Bb/Akt2, inducing the release of the mRNP complex from the 39-UTR element, resulting in the reversal of translational silencing and increased expression of Dab2 and ILEI transcripts.
We have previously shown that shRNA-mediated silencing of Dab2 and ILEI in normal murine mammary gland (NMuMG) cells is sufficient to inhibit TGFb-mediated EMT as analyzed morphologically and by loss of upregulation of N-cadherin and vimentin, mesenchymal cell markers, whereas their overexpression does not induce constitutive EMT, independent of TGFb signaling [10,14]. Thus Dab2 and ILEI are required, but not sufficient, for TGFb-induced EMT. Hence, we, and others based on our studies [15], hypothesized that there are other mRNAs that are being silenced by hnRNP E1 in a similar fashion, and which cumulatively contribute to TGFb-induced EMT. To address this hypothesis, we adopted a combinatorial approach involving polysome profiling and RIP-Chip analyses using hnRNP E1 and filtered the array data based on the regulatory mechanism of Dab2 and ILEI. This led to the identification and validation of a cohort of target mRNAs that follow the same pattern of regulation as Dab2 and ILEI. Similar to Dab2 and ILEI, the identified target mRNAs harbor a structural BAT element in the 39-UTR as revealed by in silico analysis. This cohort of mRNAs may represent a new TGFb responsive and hnRNP E1-mediated regulon, operative at a post-transcriptional level in order to mediate TGFb-induced EMT in a temporal and expedited fashion.

Experimental Design and Identification of a TGFbinduced Post-transcriptional EMT Gene Signature
To identify potential target mRNA transcripts that are translationally regulated by hnRNP E1 in a TGFb-dependent manner, we adopted a combinatorial approach involving expression profiling analyses and RNA immunoprecipitation analysis (RIP-Chip). As shown (Fig. 1A), we performed a screen using: 1) total mRNA and 2) RNA isolated from monosomal (nontranslating) versus polysomal (translating) fractions from TGFbtreated (24 h) and non-treated NMuMG cells and from the hnRNP E1 knockdown derivative (E1KD), that undergo constitutive EMT even in the absence of TGFb [10,11]. In addition, we screened for transcripts that selectively interact with hnRNP E1 in NMuMG cells under unstimulated conditions and subsequently lose their temporal association following TGFb stimulation (Fig. 1A). The samples were individually hybridized to Affymetrix GeneChipH Mouse Genome 430 2.0 arrays.
Following normalization, data was filtered to produce three datasets representing 1) TGFb translationally regulated genes, 2) genes translationally activated following hnRNP E1 knockdown and 3) hnRNP E1 interacting transcripts (Fig. 1B). Genes from the TGFb translationally regulated dataset were selected as transcripts that displayed an enhanced ratio of association with the actively translating polysomal pool compared to the non-translating monosomal pool, with no or minor changes in total mRNA expression in NMuMG cells following TGFb stimulation. mRNA transcripts that displayed enhanced association with polyribosomes irrespective of TGFb-treatment in E1KD cells were candidates for translationally active genes in an hnRNP E1 knockdown context. Whereas, transcripts which displayed a decreased association with hnRNP E1 in NMuMG cells following TGFb stimulation, as determined by RIP-Chip analysis, were selected as hnRNP E1 interacting candidates (Fig. 1B).
To perform a functional interpretation of our array analysis, all three datasets were queried against GO, Panther and KEGG databases using DAVID and Panther platforms ( Table 1). Analysis of the TGFb translationally regulated dataset revealed significant enrichment of categories associated with cell cycle, transcription and ubiquitin-mediated proteolysis. Genes actively translated in E1KD cells are involved in cell cycle, translation and the regulation of the actin cytoskeleton, whereas, transcripts that displayed differential interaction with hnRNP E1 mapped to terms associated with transcription, ubiquitin-mediated proteolysis, in addition to enrichment of several signaling pathways including MAPK, Wnt, integrin and Ras pathway. This analysis is consistent with our findings that TGFb-mediated translational regulation plays a major role during EMT [10,11], as evidenced by enrichment of EMT-associated processes and pathways. In addition, our data indicates that EMT-associated processes are coordinately regulated at both the transcriptional and posttranscriptional level. Enrichment of EMT-associated pathways within the E1KD and RIP datasets also suggest that hnRNP E1 is a key effector of TGFb-mediated translational regulation.

Identification of Candidate mRNA Transcripts Translationally Regulated by hnRNP E1 in a TGFbdependent Manner
In order to identify genes whose expression is translationally regulated by TGFb through hnRNP E1, the intersection of our three data sets was utilized ( Fig. 1B) revealing 36 genes, which we have termed BAT genes ( Table 2). The translational status of the 36 putative BAT genes as determined by isolation of nontranslating monosomal (M) fractions (40S, 60S and 80S) and actively translating polysomal (P) fractions from cells treated 6 TGFb for 24 h ( Fig. 2A) is represented by the signal intensities of monosomal and polysomal association, and is displayed as a heat plot (Fig. 2B). The data reveal that the expression of these transcripts (total mRNA) did not vary significantly 6 TGFb in either the parental NMuMG or E1KD cells (Fig. 2B). However, in the NMuMG cells, these transcripts preferentially translocated from the M to P fractions following TGFb stimulation, whereas in the E1KD cells, these transcripts were associated with the P fraction irrespective of TGFb-treatment (Fig. 2B). This methodology accurately identified ILEI as one of the target transcripts, as demonstrated by semi-quantitative RT-PCR analysis of the sample cDNA used for the microarray hybridization (Fig. 2C). The ILEI mRNA is polyribosome-associated following TGFbtreatment in parental NMuMG cells, whereas it is found polyribosome-associated in the E1KD cells in the absence or presence of TGFb. Total ILEI mRNA levels were not affected by TGFb stimulation in either cell type (Fig. 2C).
Interestingly, several of the identified mRNAs have been previously implicated as targeted transcripts of TGFb-mediated translational regulation including calpastatin [16] and epidermal growth factor receptor [17]. Additionally, this approach identified several candidates that have been shown to be involved in the EMT process including Eukaryotic initiation factor 5A2 [18], Moesin [19], Egfr [20] and Inhibin beta-A [21]. These candidates were subsequently used for further validation studies.

Validation of Selected Genes from the Affymetrix Array
We next addressed whether the translational regulation of polysome-bound transcripts correlated with respective RNA and protein levels. Initially, we performed a polysome profile expression analysis independent of the pooled microarray samples to further demonstrate the translocation of mRNA from the nontranslating M fractions to the actively translating P fractions in non-stimulated and TGFb-treated cells. In parental NMuMG cells, the target mRNAs are primarily associated with the 80S fraction in non-stimulated cells, and display a complete shift to the actively translating polysomes after 24 h of TGFb treatment (Fig. 3A). These results are in agreement with our previous Figure 1. The experimental design involves a combinatorial approach using microarray analyses of polysome profiles and RNA Immunoprecipitation. (A) Flow chart representing the experimental design. For expression profiling, cytosolic extracts from untreated and TGFbtreated (24 hr) NMuMG and E1KD cells were fractionated by sucrose gradient centrifugation and RNA was isolated from the non-translating findings that hnRNP E1-directed translational regulation targets the 80S stage of translation elongation [11]. In contrast, the hnRNP E1 knockdown E1KD cells displayed abundant target mRNA in the actively translating polysomal fractions irrespective of TGFb treatment (Fig. 3B). As a control, semi-quantitative RT-PCR with b-actin specific primers displayed continuous association of the mRNA with the polysomes irrespective of TGFbtreatment ( Fig. 3A and Fig. 3B), demonstrating that the translational control is transcript-specific and not due to global regulation of translation.
TGFb has been shown to exert both an antiproliferative effect, functioning as a tumor suppressor during early stages of tumorigenesis, whereas at later stages it functions as a tumor promoter aiding in metastatic progression through an autocrine TGFb-loop [8]. For this reason we investigated whether the regulation of these BAT genes is specific to the induction of EMT by performing a polysome profiling analysis with a second in vitro model cell line, the well-polarized mammary epithelial cells (EpH4). EpH4 cells undergo cell cycle arrest and apoptosis in response to TGFb-treatment, whereas the oncogenic Ha-Ras derivate (EpRas cells) will undergo EMT in response to TGFb [10,22,23,24]. The results from the polsome profile analysis reveal that the expression of these transcripts (total mRNA) did not vary significantly 6 TGFb in either the parental EpH4 or the derivative EpRas cells (Fig. S1A). However, in the EpRas cells, these transcripts preferentially translocated from the monosomal (M) to polysomal (P) fractions following TGFb stimulation, whereas in the EpH4 cells, these transcripts were associated with the monosomal fraction irrespective of TGFb-treatment (Fig. S1B), suggesting that TGFb-mediated translational activation of these target mRNAs is specific to the EMT response.
We next investigated the temporal relationship between total mRNA levels and protein expression levels in TGFb-treated NMuMG and E1KD cells. With the exception of moesin, total mRNA levels for these target genes, as measured by quantitative real time PCR (qPCR), displayed only minor changes following TGFb-treatment in both the NMuMG and E1KD cells compared to a ,5 fold increase in Fibronectin (Fn1), a mesenchymal marker and target of TGFb-mediated transcriptional regulation (Fig. 3C, D). These results concur with the microarray data that demonstrate that total mRNA levels for these transcripts were only slightly induced by TGFb (Fig. 2D). However, it cannot be completely excluded from these results that transcriptional regulation is involved, albeit at a low rate. In contrast, protein expression levels, as analyzed by immunoblot analysis (Fig. 3E), revealed that non-stimulated NMuMG cells, despite having abundant message, have low levels of protein for these target genes, and display a rapid, and time-dependent increase in protein expression levels following TGFb-treatment (Fig. 3E). Furthermore, the increased protein expression levels of these transcripts were shown to correlate with acquisition of a mesenchymal phenotype as demonstrated by increased expression of the mesenchymal marker N-cadherin and decreased expression of Zona occludens 1 (ZO-1). In contrast, in the E1KD cells, although there is not an apparent reduction in the expression of epithelial cell marker ZO-1, the expression of the mesenchymal marker N-cadherin, as well as the protein expression levels of the BAT genes, were constitutive irrespective of TGFb-treatment (Fig. 3E).

Target mRNAs are Regulated through Interaction with hnRNP E1 and a Structurally Conserved BAT Element
According to the RIP-Chip data, the selected target genes displayed a decrease in association with hnRNP E1 following TGFb-treatment. The average signal intensity of the association with hnRNP E1 between control and TGFb-treated samples are represented as a heat plot (Fig. 4A). In each case, less of these mRNAs were immunoprecipitated by a-hnRNP E1 in the presence of TGFb compared to the control, unstimulated NMuMG cells (Fig. 4A). To further investigate the temporal association of hnRNP E1 with the selected target genes, we performed a RIP analysis independent of the microarray samples. As shown (Fig. 4B), hnRNP E1 interacts with the target transcripts. Immunoprecipitation with a-hnRNP E1 or mouse IgG from cytosolic extracts prepared from NMuMG cells treated with TGFb for the times indicated, followed by RT-PCR analyses, revealed that while target mRNAs were steadily expressed, hnRNP E1 interaction occurred primarily in non-stimulated cells. These results are in agreement with our previous findings that TGFb activates a kinase cascade terminating in the phosphorylation of hnRNP E1, by isoform-specific stimulation of protein kinase Bb/ Akt2, inducing the release of the hnRNP E1 from the 39-UTR cis regulatory element, resulting in the reversal of translational silencing and increased expression of EMT-facilitating transcripts [10].
We have previously identified the structural BAT element in the 39-UTRs of Dab2 and ILEI which binds hnRNP E1 and mediates TGFb-induced translational regulation of these transcripts [10]. The Dab2 and ILEI BAT elements consist of a proximal stem and an asymmetric bulge followed by a distal stem and terminal loop (Fig. 4C). In order to determine whether the selected target genes also contain a respective BAT element, we utilized a consensus BAT element pattern, based on the secondary structure of Dab2 and ILEI BAT elements, to query the non-redundant 39UTR sequences of the selected target genes using RNAmotif, an RNA secondary structure algorithm [25]. Putative BAT elements were identified in the target mRNAs with significant folding similarity as identified by the stem-loop and asymmetric bulge ( Fig. 4D and Fig. 4E). Furthermore, we performed an additional in silico analysis of several of the target genes to determine if the predicted BAT elements are conserved in humans by querying the non-redundant 39-UTR sequences of human ILEI, Eifa52 and Msn using the RNAmotif algorithm (Fig. S2). The results demonstrate that these target mRNAs share considerable sequence homology (.80%) as well as a conserved structural fidelity between the predicted mouse and human BAT elements (Fig. S2).
We next examined the temporal association of hnRNP E1 with the predicted BAT elements using an RNA affinity pull down assay (Fig. 4F). The respective BAT element cRNAs were coupled to sepharose beads, and used to precipitate hnRNP E1 from cytosolic S100 extracts isolated from TGFb-treated and nontreated NMuMG cells. As a negative control, we used the Dab2/ U10A element (BAT-M), which contains a U to A substitution at position 10 which unfolds the stem loop structure resulting in monosomal pool and actively translating polysomal pool, designated as M and P, respectively (n = 2). Total unfractionated RNA was isolated from NMuMG and E1KD cells treated 6 TGFb (24 hr) (n = 2). For the RIP-Chip analysis, cytosolic extracts from untreated and TGFb-treated (24 hr) NMuMG cells were immunoprecipitated with a-hnRNP E1 antibody or an isotype control (n = 2). (B) Venn diagram summarizing the results of the genome wide analysis. Intersection of the three data sets yielded 36 putative BAT genes whose expression is translationally regulated by TGFb through hnRNP E1. doi:10.1371/journal.pone.0052624.g001 Table 1. Number of genes from dataset assigned to a given biological process or pathway is compared to the number of genes expected by chance to map to the term. diminished binding affinity to hnRNP E1 [10,11]. Immunoblot analysis confirmed that hnRNP E1 was precipitated by the predicted Egfr and Eif5a2 BAT elements from non-stimulated NMuMG, but TGFb treatment induced the loss of hnRNP E1 binding in a time-dependent manner. Additionally, pre-treatment of NMuMG cells with the PI3K inhibitor LY294002, blocked the ability of TGFb to modulate hnRNP E1 interactions (Fig. 4F), consistent with our previous observation that inhibition of the PI3K/Akt pathway blocked hnRNP E1 Ser43 phosphorylation [10].

Discussion
Despite intensive transcriptional array analysis of human tumors, the identity and validation of 'EMT signature genes' remains elusive [26,27,28], partially because the transcriptome does not mirror the proteome [29]. To understand how the interplay of RNA-binding proteins affects the regulation of individual transcripts, high-resolution maps of in vivo protein-RNA interactions are necessary [30]. An alternative approach is expression profiling on a genome wide scale, whereby non-translating and actively translating pools of mRNAs are isolated by sucrose density gradient fractionation and subsequently subjected to microarray analysis [31]. RNA-Binding Protein Immunoprecipitation-Microarray (RIP-Chip) profiling is an advanced highthroughput analysis of mRNAs that co-immunoprecipitate with particular mRNA-binding proteins [32]. An mRNA-binding protein of interest is immunoprecipitated, and the associated mRNA is isolated and subsequently subjected to microarray analysis. A combinatorial approach involving expression profiling and RIP-chip analysis on a genome-wide basis will yield definitive information on a particular regulatory pathway.
Herein, we have identified a cohort of translationally regulated mRNAs that are upregulated during TGFb-induced EMT by using a combinatorial approach involving polysome profiling and RIP-Chip analysis. Filtering the Affymetrix array data based on the translational state of transcripts in non-stimulated and TGFbtreated NMuMG and E1KD cells, and intersecting these genes sets with the RIP-chip analysis led to the identification of a set of target mRNAs that follow the same pattern of regulation as Dab2 and ILEI, two transcripts necessary for EMT which were previously shown to be translationally regulated by TGFb through hnRNP E1 [10,11]. While our confidence in the establishment of this TGFb-induced post-transcriptional EMT signature was strengthened by the identification of several transcripts which have been previously shown to be translationally regulated by TGFb, including calpastatin [16] and epidermal growth factor receptor [17], our approach was not without some limitations. For example, this approach correctly identified ILEI mRNA, a wellcharacterized target for BAT-mediated translational silencing [10,11], however, another previously identified target, Dab2 remained unidentified. This result may be due to a low signal-to-noise ratio of Dab2 expression levels, as the microarray based approach requires that the level of expression of target mRNA exceeds the cutoff limit of detection with a high-signal-to noise ratio [33].
Protein expression levels depend on the rate of transcription, as well as other defined control mechanisms, such as mRNA stability [34], nuclear export and mRNA localization [35], translational regulation [36], and protein degradation [37]. Post-transcriptional regulation is mainly controlled by the association of trans-acting RNA binding proteins with cis-regulatory regions in the UTRs of  mRNAs. The bioinformatic prediction of putative BAT elements in the identified BAT mRNAs reveals a conserved structure-based homology based upon the functional structural motif previously identified for Dab2 and ILEI. These structures within the 39-UTRs of the selected target mRNAs all share a stem-loop motif with an asymmetric bulge, albeit with considerable sequence diversity. Although the Egfr and Eif5a2 BAT elements were validated for their ability to bind hnRNP E1, a more comprehensive analysis is still required. This includes, but is not limited to fine mapping and cloning of the 39-UTR of the candidate genes into reporter vectors to demonstrate functional gain-of-silencing potential. Furthermore, while in silico analysis demonstrated that the predicted BAT elements in several of the target mRNAs displayed considerable structural fidelity between mouse and human homologs, a more rigorous analysis is needed to determine if these conserved secondary structures are simply the result of of the high degree of intraspecies sequence homology in the respective 39-UTRs, or the result of stabilizing selection.
The BAT element provides further insights into the importance of regulatory elements in the maintenance of homeostasis. Our results are suggestive of a stimulus-dependent upregulation of a post-transcriptional regulon coordinated by the concerted action of a trans-acting mRNP complex and a cis-regulatory element in the 39-UTR of target genes. Eukaryotic regulons are defined as higher-order genetic units (quasi genome) consisting of monocistronic mRNA subsets under the control of a regulatory RNA binding protein [30]. RNA binding proteins have been shown to specifically bind transcripts encoding functional and colocalized protein classes [38,39,40]. Post-transcriptional regulons may have evolved as a mechanism to rapidly and coordinately suppress multiple EMT genes.
During the invasive phase of metastasis, a carcinoma cell activates EMT programs by different regulatory pathways. Differentiation to a mesenchymal phenotype enables the cancer cell with the ability to survive through the different steps of metastatic progression, including localized invasion by primary tumor cells, intravasation, translocation, extravasation and finally micrometastatic colonization at the secondary site [9]. Now, we demonstrate a cohort of selective transcripts that are posttranscriptionally upregulated by TGFb and are correlative with an induction of the EMT phenotype. Akt2-mediated hnRNP E1 phosphorylation post-TGFb stimulation is the regulatory mechanism mediating the TGFb-induced translational activation of EMT-facilitating transcripts. We have now shown that hnRNP E1 is a central moiety in this process, and may represent an important molecular target for the development of modulators of this translational regulatory pathway. Furthermore, the continued delineation of the role of the identified target transcripts during EMT will prove to be extremely useful and will allow for their interrogation and manipulation in physiological and pathological situations.

Reagents
Mouse a-hnRNP E1 and a-ZO-1 were obtained from Novus Biologicals. a-ILEI, a-Inhibin beta A and a-EIF5A2 were obtained from Abcam. a-EGFR was obtained from Cell Signaling Technology. a-Moesin was purchased from BD Biosciences. a-Hsp90 and normal mouse IgG were purchased from Santa Cruz Biotechnology. Secondary antibodies, a-mouse and a-rabbit-IgG-HRP were obtained from GE Healthcare Bio-Sciences.

Cell Culture and Treatments
Murine mammary epithelial NMuMG cells, obtained from American Type Culture Collection (ATCC; Manassas, VA, USA), were maintained in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum, 10 mg/ml insulin, and antibiotics/antimycotics (100 units/ml penicillin G, 100 mg/ml streptomycin, and 0.25 mg/ml amphotericin B). E1KD (previously termed SH14) were generated in the laboratory and have been described [10]. TGFb2 was a generous gift from Genzyme Corporation (Cambridge, MA, USA) and was used at a final concentration of 5 ng/ml. Where indicated, cells were treated with 10 mM of LY294002 30 min before TGFb treatment.

RNA Immunoprecipitation
RNA immunoprecipitation was performed as described previously [32]. Briefly, the cytosolic extract was incubated with 10 mg of mouse a-hnRNP E1 antibody or mouse a-IgG at 4uC overnight, and precipitated with Protein A-Sepharose (Invitrogen). The beads were washed three times with IP Wash Solution (150 mM NaCl, 50 mM Tris pH 7.5, 0.5% NP40), and immunoprecipitated RNAs isolated by Trizol (Invitrogen) and treated with RNase-free DNase I (Applied Biosystems).

Polysome Profiling
Polysome analysis was performed as described previously [42]. Briefly, cell lysates were layered onto a 10%-50% sucrose gradient and centrifuged at 100,0006g at 4uC for 4 h. Gradient fractions were collected using a fraction collector with continuous monitoring of absorbance at 254 nm. RNA was extracted with Trizol (Invitrogen) and purified with RNeasy minikit (Qiagen).

Microarray Data Processing
Affymetrix microarray analysis was conducted on two independent samples for each experimental condition. Samples were processed at the MUSC Proteogenomics Facility (http:// proteogenomics.musc.edu) using Affymetrix Mouse Genome 430 2.0 GeneChipsH in accordance with the manufacturer protocols. The resulting raw data files were deposited in the NCBI Gene Expression Omnibus (Accession #GSE40466). Hybridization data (CEL files) were normalized by RMA algorithm using Affymetrix Expression Console software; detection calls were obtained by Affymetrix MAS5 algorithm. Gene representations not receiving 'present' detection scores in $25% of all samples were excluded from further analysis.

Data Analysis
The average raw signal intensity values from two independent samples for each experimental condition were determined and used for multiparametric comparisons. Filtering of the genes sets expression levels. Data are presented as means 6 s.e., n = 3 (*P,0.05). (D) Immunoblot analysis examining protein expression levels of the potential targets, a-Hsp90 (control) and a-N-cadherin and a-ZO-1 (EMT markers), in NMuMG and E1KD cells treated with TGFb for the indicated times. doi:10.1371/journal.pone.0052624.g003 met the following criteria: i) the ratio of the average raw signal intensity of monosomal (M) versus polysomal (P) mRNAs from control NMuMG cells was filtered as (M control /P control $2), whereas in the E1KD cells the parameter was set at (P control / M control $2); ii) the ratio of P vs. M associated mRNAs from TGFb-treated NMuMG and E1KD cells was filtered as (P TGFb / M TGFb $2); iii) fold change in total RNA from control vs. TGFbtreated cells was determined by (Total RNA TGFb /Total RNA control #2); and, iv) the ratio of fold induction of polysomal RNA in NMuMG cells post TGFb-stimulation compared to control was filtered at ([(P TGFb /M TGFb )/(P control /M control )] $5), whereas in the E1KD cells the parameter was set at ([(P TGFb /M TGFb )/(P control / M control ) ,2). Finally, for the RIP-Chip, the ratio of the average raw signal intensity for mRNAs immunoprecipitated by hnRNP E1 in control NMuMG cells vs. TGFb-treated was filtered at (IP:E1 Control/TGFb $1.2), whereas for the IgG immunoprecipitation the parameter was set at (IP:IgG Control/TGFb #1).

Functional Pathway Search Analysis
Functional analysis was performed using the Database for Annotation, Visualization and Integrated Discovery (DAVID), Molecular Signature Database (MSigDB) and Protein Analysis Through Evolutionary Relationships (Panther) platforms. Biological processes and pathway terms from Gene ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), Panther, Reactome and Biocarta databases were utilized.

Real-time Quantitative PCR (Taqman system)
Total RNA was isolated by Trizol extraction. Reverse transcription was performed using the Superscript first strand synthesis system (Invitrogen). Quantitative PCR was performed using an Applied Biosystems 7500 Fast Real-Time PCR System and default cycle conditions. Briefly, reactions were prepared using 50 ng cDNA, TaqmanH Fast Universal PCR Master Mix and mouse-specific primers for Egfr (cat number Mm00433023_m1), Moesin (Mm00447889_m1), Eif5a2 (Mm00812570_g1), Fam3c (Mm00506842_m1), Inhba (Mm00434339_m1), Fn1 (Mm01256744_m1), and Gapdh (Mm99999915_g1) according to manufacturer's protocol (Life technologies). All samples were run in triplicate and normalized to Gapdh. Data analysis was performed using the relative quantification (DDC T ) method [43].

RNA Pulldown
RiboMax kit (Promega) was used to generate milligram quantity of BAT cRNA. cRNA was bound to CNBr-activated Sepharose beads and incubated for 1 h at 4uC with 50 mg of cytosolic extract (S100 Fraction) from NMuMG cells treated 6 TGFb. Following incubation, beads were washed with 0.2 M NaCl and resolved by SDS-PAGE.

Bioinformatic Prediction of BAT Elements
Analysis of select 39UTR genes was completed using RNAmotif software [25] on a MacBook Pro using an Intel Core-i7, 4 gb of RAM, and Mac OS-X 10.6. Software was compiled using GNU's GCC compiler (gcc.gnu.org) in the OS-X terminal utility. Descriptor file was written using Xcode version 3.2 in the C development module. Input file was created in a generic text editor, sequences were obtained from UTRdb (utrdb.ba.itb.cnr.it). Analysis was run from the OS-X terminal utility, output was sent to a generic text file to be used for later interpretation. Figure S1 Validation of the putative EMT signature gene targets in EpH4 and EpRas cells (related to Figure 3). (A) RT-PCR analysis using gene specific primers for the potential targets and b-Actin (control) on total RNA extracted from EpH4 and EpRas cells 6 TGFb for 24 hr. (B) Monosomal fractions (M; 40S, 60S, and 80S fractions) and polysomal fractions (P) from EpH4 or EpRas cells treated 6 TGFb for 24 hr were isolated by sucrose gradient centrifugation and pooled. RT-PCR analysis using gene specific primers for the potential targets and b-Actin (control). (DOC) Figure S2 In silico analysis of BAT elements from human homologs of target mRNAs (related to Figure 4).