Zeb1 potentiates genome‐wide gene transcription with Lef1 to promote glioblastoma cell invasion

Abstract

Glioblastoma is the most common and aggressive brain tumor, with a subpopulation of stem‐like cells thought to mediate its recurring behavior and therapeutic resistance. The epithelial–mesenchymal transition (EMT) inducing factor Zeb1 was linked to tumor initiation, invasion, and resistance to therapy in glioblastoma, but how Zeb1 functions at molecular level and what genes it regulates remain poorly understood. Contrary to the common view that EMT factors act as transcriptional repressors, here we show that genome‐wide binding of Zeb1 associates with both activation and repression of gene expression in glioblastoma stem‐like cells. Transcriptional repression requires direct DNA binding of Zeb1, while indirect recruitment to regulatory regions by the Wnt pathway effector Lef1 results in gene activation, independently of Wnt signaling. Amongst glioblastoma genes activated by Zeb1 are predicted mediators of tumor cell migration and invasion, including the guanine nucleotide exchange factor Prex1, whose elevated expression is predictive of shorter glioblastoma patient survival. Prex1 promotes invasiveness of glioblastoma cells in vivo highlighting the importance of Zeb1/Lef1 gene regulatory mechanisms in gliomagenesis.

Introduction

Glioblastoma multiforme (GBM, World Health Organization grade IV astrocytomas) remains one of the most common and deadliest forms of brain tumors in adults (Ostrom et al, 2015). Treatment of GBM is extremely ineffective in part due to its highly invasive nature, resulting from single malignant cell invasion of normal brain parenchyma hindering complete tumor resection. Moreover, it is thought that GBM harbors cancer stem‐like cells (CSCs) with the capacity to seed an entire multi‐lineage tumor, meaning that even the smallest number of such cells left after surgery, chemotherapy and radiotherapy have the potential to cause tumor recurrence. GBM cell cultures containing CSCs with tumor‐initiating capability can be established and expanded from resected tumor tissue using defined serum‐free conditions (Singh et al, 2003, 2004; Lee et al, 2006). These cellular models show genetic profiles close to those of human GBM, and as opposed to conventional serum‐grown cell lines are strongly invasive when injected into the mouse brain (Lee et al, 2006; Xie et al, 2014; Paw et al, 2015). Thus, identifying the molecular players that regulate this important GBM trait may benefit from a better characterization of gene regulatory mechanisms that operate in GBM CSC cultures.

Zeb1 (zinc finger E‐box binding homeobox 1) is a transcription factor known for its ability to induce an epithelial–mesenchymal transition (EMT), a complex developmental program associated with loss of cell polarity, extensive extracellular matrix remodeling, and acquisition of a migratory behavior (Kalluri & Weinberg, 2009; Micalizzi et al, 2010). The importance of Zeb1 during embryonic development was revealed by the generation of Zeb1 null mice, which develop to term but die after birth displaying multiple developmental abnormalities (Takagi et al, 1998). Most of these can be attributed to its ability to activate an EMT‐like program in progenitors of various lineages. In the developing cerebellum, Zeb1 retains neural progenitors in the germinal zone by inhibiting their polarization, a required step during differentiation of granule neuron progenitors (Singh et al, 2016).

Most research performed on Zeb1 has focused on its role in tumor development, where EMT induction triggers cellular mobility and subsequent dissemination of cancer cells (Sánchez‐Tilló et al, 2011b). Accordingly, Zeb1 expression has been found at the migration front of a variety of epithelial tumors. In addition to promoting metastasis, recent studies have shown that Zeb1 is required for the tumor‐initiating capacity of both pancreatic and colorectal cancer cells, through the repression of stemness‐inhibiting microRNAs of the mir‐200 family (Bracken et al, 2008; Burk et al, 2008; Brabletz & Brabletz, 2010). Although a proper EMT is not at play in gliomas, recent studies described Zeb1 expression in this type of tumors, where it is correlated with tumor grade and infiltrative behavior (Siebzehnrubl et al, 2013; Kahlert et al, 2015). Furthermore, Zeb1 has been implicated in the processes of tumor invasion, chemoresistance, and tumorigenesis in GBM, by mechanisms that remain poorly characterized (Siebzehnrubl et al, 2013). Recently, Zeb1 was shown to be part, together with Sox2 and Olig2, of a transcriptional network that drives gliomagenesis (Singh et al, 2017).

Several studies have shown Zeb1 to inhibit the epithelial phenotype by repressing polarity and cell adhesion genes, via direct binding to so‐called E‐box motifs with the consensus sequence CAGGTG/A. Transcriptional repression by Zeb1 involves the recruitment of co‐repressor proteins like CtBP1/2, or chromatin modifying/remodeling factors such as HDAC1/2, BRG1, and LSD1 (Postigo & Dean, 1997, 1999; Postigo et al, 1999; Grooteclaes & Frisch, 2000; Shi et al, 2003; Ponticos et al, 2004; Wang et al, 2007; Sánchez‐Tilló et al, 2010). In addition to its well‐characterized role as repressor, few observations also implicate Zeb1 in transcriptional activation. When promoting EMT downstream of TGF‐beta, Zeb1 can bind activated SMAD proteins and CBP/p300 to enhance the transcription of the mesenchymal genes smooth muscle actin and myosin (Postigo, 2003; Postigo et al, 2003). Moreover, expression of Hippo pathway target genes in breast cancer cells is promoted by a functional interaction between Zeb1 and the Hippo cofactor YAP (Lehmann et al, 2016). In spite of recent findings, further progress on how Zeb1 coordinates such a complex genetic program has been hampered by the paucity of information on the identity of its direct target genes in the same cellular context.

Here we show that Zeb1 is highly expressed in GBM CSCs, and characterize its transcriptional program by combining binding site analysis with expression profiling upon Zeb1 loss of function. Surprisingly, we found that Zeb1 binding to DNA is associated genome‐wide with both activation and repression of gene expression, and that this dual role results from distinct modes of recruitment to regulatory regions. Target gene repression requires binding to E‐box sequences, while indirect recruitment together with Lef/Tcf factors potentiates gene expression, in the absence of active Wnt signaling. Notably, genes activated by Zeb1 include candidate mediators of tumor cell migration and invasion. These include the guanine nucleotide exchange factor Prex1, which we show to be important for invasion of GBM cells in vivo. Finally, we observe that Prex1 and Zeb1 expression levels are highly correlated in GBM patient tumor samples, with Prex1 levels being highest and indicative of a poor patient prognosis in the Classical GBM subtype. Overall, our work provides important mechanistic insights on how Zeb1 coordinately regulates an EMT‐like program that contributes to invasiveness in GBM.

Results

Characterization of Zeb1 target genes in GBM CSCs

In face of recent studies linking Zeb1 with invasiveness in various cancer types including GBM, we started by assessing the expression of Zeb1 protein in three GBM CSC models (NCH421k, NCH441, and NCH644) isolated and expanded in serum‐free conditions from resected GBM tumors. These have been previously characterized on their gene expression profile and tumor‐initiating capacity upon xenotransplantation (Campos et al, 2010; Fassl et al, 2012; Podergajs et al, 2015). The human fetal neural stem cell line Cb192 was also included for comparative purposes (Sun et al, 2008). Analysis by Western blot revealed strong Zeb1 expression in all cellular models, denoted by a band of the expected molecular weight, and which was reduced upon expression of a sequence‐specific shRNA against Zeb1 (Fig 1A). Immunostaining of NCH421k cells shows Zeb1 expression occurring in virtually all cells, similar to Cb192 cells (Fig 1B).

Figure 1. Characterization of Zeb1 transcriptional program in GBM CSCs.

1. ZEB1 protein levels in the human fetal neural stem cell line CB192 and in three GBM CSC lines (NCH441, NCH644, and NCH421k), as determined by Western blot, upon infection with a lentiviral vector expressing a control (shCtr) or Zeb1 shRNA (shZeb1).
2. CB192 and NCH421k cells were immunostained for Zeb1 (red) and nestin (green). Nuclear staining with DAPI (blue). Scale bars: 40 μm.
3. Genome‐wide mapping of Zeb1 binding sites by ChIP‐seq was combined with transcriptional profiling upon Zeb1 knock‐down with a sequence‐specific shRNA against Zeb1 in order to identify Zeb1 direct target genes. Diagram shows design of experiment.
4. Examples of Zeb1 ChIP‐seq enrichment profiles in the vicinity of previously characterized Zeb1 target genes.
5. Location of Zeb1 binding events respective to various genomic features.
6. Location of Zeb1 binding events in relation to closest annotated TSS.
7. Venn diagram depicting the number of genes up‐ and downregulated upon Zeb1 knock‐down, and unique genes associated with at least one Zeb1 binding event following a nearest gene annotation.
8. Number of Zeb1 binding events associated with up‐ (red bar) or downregulated (blue bar) genes upon Zeb1 knock‐down.
9. Heat map displaying the cumulative fraction of deregulated genes upon Zeb1 knock‐down that are directly regulated by Zeb1. Number of transcripts with expression fold change > 1.2 are plotted against Zeb1 ChIP‐seq peaks with increasing P‐value (bin = 94). As control, the average result obtained with 100 sets of random binding events of equal size is shown.
10. Enrichment of GO Biological Process terms associated with Zeb1 direct targets as determined using DAVID bioinformatics resource. Modified Fisher exact value for each category is shown.
11. qPCR validation of deregulation of selected genes upon Zeb1 knock‐down in NCH421k cells.

Data information: (H) Test data comprised of 1,000 randomized datasets of equal size are represented as box with median of test and first and third quartiles; whiskers, ± 1.5 × interquartile range (IQR). (K) Data are shown as mean ± SD of triplicate assays (significance determined by unpaired two‐tailed t‐test). In all cases: n.s. for non‐significant, *P < 0.05, **P < 0.005, ***P < 0.001.Source data are available online for this figure.

To investigate the molecular mechanisms underlying the role of Zeb1 in GBM CSCs, we started by characterizing its transcriptional program in NCH421k cells by combining genomic location analysis with transcriptional profiling upon shRNA induced Zeb1 loss of function (Fig 1C). Genome‐wide mapping of Zeb1 binding profile by ChIP‐seq identified 6,879 binding events (P < 10E⁻¹⁰), associated with 4,430 unique genes following a nearest gene annotation (Tables EV1 and EV2). Binding events were found at close vicinity to previously identified Zeb1 repressed genes associated with epithelial cell polarity, such as Pard6b, Crb3, or Cdh1 (Fig 1D) (Eger et al, 2005; Aigner et al, 2007; Singh et al, 2016). Most binding events occur within introns or intergenic regions and at great distances from TSSs, consistent with binding predominantly to distal enhancers (Fig 1E and F). Because binding of a transcription factor does not necessarily denote a regulatory function, we next characterized changes in transcriptome resulting from Zeb1 knock‐down. NCH421k cells were infected with a lentivirus expressing a sequence‐specific shRNA against Zeb1, and gene expression changes were characterized 72 h later using DNA arrays. This identified 298 deregulated genes (fold change > 1.2; P < 0.05), of which 67% (n = 200) were downregulated and 33% (n = 98) upregulated (Fig 1G, Table EV3).

A dual role for Zeb1 promoting gene activation and repression

To gain an insight into the global transcriptional response mediated by Zeb1, we integrated location analysis and expression profiling results. We found Zeb1 binding events to be significantly associated with both downregulated and upregulated genes (P = 3.2E⁻²³ and P = 5.5E⁻¹⁵, respectively), when compared to the association with 1,000 randomized datasets of equal size (Fig 1H). We also determined the fraction of up‐ and downregulated genes associated with at least one Zeb1 binding event (direct targets), considering increasing P‐value cutoffs for Zeb1 binding. The significance of such association was assessed by comparison with the association with 100 randomly generated ChIP‐seq datasets of equal size. Both up‐ and downregulated genes are significantly enriched with Zeb1 direct targets when compared to control datasets, again highlighting a dual activity of Zeb1 in mediating both activation and repression of gene expression (Fig 1I).

Next, we performed a gene ontology (GO) search on a high‐confidence list of direct Zeb1 targets, composed of genes bound and activated (60) or repressed (42) by Zeb1 (Fig 1G and Table EV4). Genes repressed by Zeb1 were associated with terms (GO biological process) including “neuron development and morphogenesis” and “epithelial cell differentiation” (e.g., Myo6, Shroom3, Pard6b, Nrxn3), while activated targets were enriched in genes associated with “glycoprotein metabolic process” (Csgalnact1, Mgat4a, B3galt2), “actin cytoskeleton organization”, and “locomotory behavior” (e.g., Cmcm5, Itgb1, Nrp2, Prex1) (Fig 1J and K). Examples of targets previously implicated in cancer cell migration/invasion in various contexts included S100B (Pang et al, 2012; Tan et al, 2018), Spry4 (So et al, 2016), and Coro1C (Lim et al, 2017; 1; Mataki et al, 2015) (Fig 1K). Thus, Zeb1 coordinates various cellular processes altogether reminiscent of an EMT‐like program, by simultaneously activating and repressing distinct sets of target genes.

Two modes of recruitment of Zeb1 to gene regulatory regions

To understand the molecular basis for the dual activity of Zeb1 in regulating gene transcription, we investigated how Zeb1 is recruited to its target sites by performing a de novo search for DNA enriched motifs within 50bps of peak summits. In addition to the expected E‐box sequence recognized by Zeb1 (CAGGTG), the hexamer sequence (ACAAAG) that matches the consensus binding site for high mobility group box (HMG‐box) transcription factors was also strongly enriched (Fig 2A). Strikingly, while the E‐box was prevalent in Zeb1 peaks associated with low P‐values (high peaks) (Fig 2A, top), the HMG‐box binding sequence (herein referred to as “HMG motif”) was mostly enriched in less significant peaks (small peaks) at the bottom half (Fig 2A, bottom). Both motifs are sharply enriched at peak summits and are thus well positioned to mediate DNA binding (Fig 2B). Hierarchical clustering based on the presence of each motif at close vicinity to peak summits (50 bps) segregates most of Zeb1 peaks in two larger groups containing either motif, whereas only a minority of peaks (3.8%) contained both (Fig 2C). Notably, whereas the E‐box motif is enriched at peaks associated with both activated and repressed genes, the HMG motif is found exclusively at peaks associated with Zeb1‐activated genes (Fig 2D). In conclusion, these results suggest that smaller peaks result from indirect recruitment of Zeb1 to the HMG motif, and that these Zeb1 binding events result in activation of gene expression.

Figure 2. Two modes of Zeb1 recruitment to target sites differentially associate with gene activation and repression.

1. Density plot of Zeb1 ChIP‐Seq reads at 4‐kb genomic regions centered at peak summits (signal intensity represents normalized tag count). The top five motifs identified by a de novo search at 100‐bp regions centered at summits from top half or bottom half of list of Zeb1 peaks are shown, with respective statistical parameters.
2. Enrichment profile of E‐box and HMG motifs at 4‐kb genomic regions centered at Zeb1 peak summits.
3. Hierarchical clustering of Zeb1 peaks based on the presence of each motif within 50 bp from peak summits. Only peaks with at least one motif are shown.
4. Smoothed curves representing enrichment profiles of each motif centered on peak summits associated with gene repression (left) or activation (right).

Multiple Lef/Tcf factors are recruited to HMG‐type peaks at Zeb1 target genes

To further investigate the gene activation function of Zeb1, we focused our subsequent studies on the regulation of the Neuropilin 2 receptor protein (Nrp2) and the guanine exchange factor Prex1, two genes directly activated by Zeb1 (Fig 1K) and which have been previously linked with migratory behavior of malignant cells (Prud'homme & Glinka, 2012; Ebi et al, 2013; Dillon et al, 2015; Lucato et al, 2015). Multiple Zeb1 binding events occur at various intronic regions of the Prex1 and Nrp2 genes, enriched for the H3k4me1 and H3K27ac histone marks characteristic of active enhancers as determined by chromatin landscape profiling of GBM CSCs in a previous study (Fig 3A). In subsequent experiments, we focused on one region in each gene, containing two ACAAAG sequences in the absence of any E‐box, thus enabling us to investigate the regulatory activity resulting from Zeb1 recruitment exclusively by HMG motifs (Fig 3A).

Figure 3. Zeb1 and Lef/Tcf factors bind to regulatory regions associated with Nrp2 and Prex1 genes.

1. ChIP‐seq enrichment profiles of Zeb1 in genomic region spanning the Nrp2 and Prex1 genes. Enrichment profiles of H3K4me1 and H3K27ac as determined by Rheinbay et al (2013) are also shown. Regions used in transcriptional assays are marked with an asterisk and shown below, with black triangles marking location of primers used in ChIP‐PCR, centered on HMG motifs.
2. Expression of various Lef/Tcf factors in indicated cell types assessed by Western blot analysis. Histone H3 is shown as loading control.
3. ChIP‐qPCR showing Lef1, Tcf3, and Tcf4 recruitment to Prex1 and Nrp2 regulatory regions in NCH421k cells.
4. Expression qPCR validation of deregulation of selected genes upon Zeb1 knock‐down in NCH441 and NCH644 cells.
5. ChIP‐qPCR showing Lef1 recruitment to Prex1 and Nrp2 regulatory regions in NCH441 and NCH644 cells.
6. EMSA shows binding of Lef1 to various oligonucleotide probes containing one HMG motif each (selected from Nrp2 and Prex1 regulatory regions) or mutated versions: NRP2_HMG2 (1), NRP2_HMG2_mut (2), NRP2_HMG1 (3), Prex1_HMG1 (4), Prex1_HMG1_mut (5), Prex1_HMG2 (6). Arrowhead marks the Lef1 specific band.
7. Coimmunoprecipitation of Zeb1‐V5 with Lef1‐Flag, using protein extracts from transfected 293T cells.

Data information: (C, E) Data are shown as mean ± SD of triplicate assays (significance determined by one‐way ANOVA with Fisher's LSD test). (D) Data are shown as mean ± SD of two biological replicates (significance determined by unpaired two‐tailed t‐test). In all cases: n.s. for non‐significant, *P < 0.05, **P < 0.005, ***P < 0.001, ****P < 0.0001.Source data are available online for this figure.

Amongst HMG‐box transcription factors, members of the Lef/Tcf family (Tcf3 and Tcf4) were found to be overexpressed in gliomas (Chen et al, 2011; Zhang et al, 2011; Liu et al, 2012; Rheinbay et al, 2013; Gao et al, 2014; Pećina‐Šlaus et al, 2014; Li et al, 2016), with all members of the family described as promoting migration and invasion of glioma cells (Chen et al, 2011; Zhang et al, 2011; Liu et al, 2012; Rheinbay et al, 2013; Gao et al, 2014; Pećina‐Šlaus et al, 2014; Li et al, 2016). Thus, we proceeded by analyzing their expression in NCH421k cells, and assessing binding to regulatory regions. Western blot analysis shows Lef1 (aka Tcf7 l3), Tcf3 (aka Tcf7 l1), and Tcf4 (aka Tcf7 l2) are strongly expressed in NCH421K cells, with Tcf1 (aka Tcf7) expression only detected upon high membrane exposure (Fig 3B and unpublished observation). ChIP‐PCR performed on chromatin extracted from NCH421k cells using antibodies against Lef1, Tcf4, and Tcf3 showed all three factors bind strongly to Prex1 and Nrp2 regulatory regions in addition to the previously characterized Wnt targets Nkd1 and Tnfrsf19 (Buttitta et al, 2003; Koch et al, 2005; Zhang et al, 2007), as compared to two negative control regions (ORF1 and ORF2) (Fig 3C). Binding of Lef1 to a series of five additional “HMG‐type Zeb1 peaks” associated with activated genes S100b, Spry4, or Col9a1 was also detected by ChIP‐PCR (Figs 1K and EV1A) further demonstrating co‐recruitment of Lef/Tcf factors and Zeb1 to regulatory regions.

Figure EV1. Binding of Lef/Tcf factors to regulatory regions in various GBM CSC models.

1. ChIP‐qPCR assessing binding of Lef1 to various HMG‐type Zeb1 peaks, or control Wnt targets in NCH421k cells.
2. ChIP‐qPCR showing Lef1, Tcf3, and Tcf4 recruitment to Prex1 and Nrp2 regulatory regions in NCH441 and NCH644 cells, as indicated.

Data information: Data are shown as mean ± SD of triplicate assays (significance determined by one‐way ANOVA with Fisher's LSD test). In all cases: n.s. for non‐significant, ***P < 0.001, ****P < 0.0001.

Similarly to NCH421K cells, shRNA‐mediated knock‐down of Zeb1 expression in NCH441 and NCH644 cells resulted in reduced Nrp2 and Prex1 expression (Fig 3D), while binding of Zeb1 to the associated regulatory regions was also observed (Fig 3E). Moreover, binding of the Lef/Tcf factors expressed in each cellular model was also detected by ChIP‐PCR (Fig EV1B), indicating that in all three GBM CSCs tested, various Lef/Tcf family members co‐occupy the Zeb1 target regions at Nrp2 and Prex1 genes.

Given the degeneracy of the consensus binding sequence for HMG factors, we next investigated the ability of a representative member of the family (Lef1) to bind the specific ACAAAG sequence predominantly enriched at Zeb1 peak summits. Confirming this possibility, electrophoretic mobility shift assay (EMSA) using in vitro transcribed and translated Lef1 protein showed binding to all oligonucleotide probes spanning one of each four HMG motifs found at regulatory regions of Nrp2 and Prex1 genes, but not to probes in which the motif was mutated (Fig 3F). Lastly, we were able to recover V5‐tagged Zeb1 upon immunoprecipitation of Flag‐tagged Lef1 from protein extracts produced from 293T cells co‐transfected with expression vectors for both factors, suggesting that Zeb1 and Lef1 may physically interact (Fig 3G). Together, these results strongly suggest Lef/Tcf factors mediate recruitment of Zeb1 to activated target genes.

Transcriptional synergy between Zeb1 and Lef1

To better investigate the activity of Zeb1 in promoting Prex1 and Nrp2 expression, we performed transcriptional assays in transfected P19 cells using reporter constructs bearing the luciferase gene and a minimal promoter sequence, under the regulation of the selected regulatory regions. Strikingly, while expression of Zeb1 alone did not activate the reporter gene, Lef1 resulted in relatively weak activation that was significantly potentiated by Zeb1 co‐expression (Fig 4A). Similar results were observed using the serum‐grown glioblastoma cell lines LN229 and U87, although with overall reduced fold activation (Fig 4B and C). Moreover, co‐expression of both factors also resulted in transcriptional synergy in Cb192 cells, where analysis of the genome‐wide binding profile of Zeb1 revealed an identical recruitment of Zeb1 mediated by HMG motifs, similar to the situation observed in the GBM CSCs used in this study (Figs 4D and EV2). Importantly, mutations of each pair of HMG motifs abolished transcriptional activity of both regulatory regions, confirming the requirement of an HMG binding factor (Fig 4C and D). Furthermore, synergy between Zeb1 and Lef1 was also observed in a multimer build of seven HMG motifs ((HMG‐Box)x7), supporting the view that functional interaction between both factors does not require direct recruitment of Zeb1 to E‐box sequences (Fig 4A). Although some differences concerning the activity of each factor alone were observed across cell lines (e.g., P19 versus Cb192 cells) presumably due to differential expression of endogenous factors, these transcriptional assays altogether support a role for both Zeb1 and Lef1 regulating Prex1 and Nrp2 via the tested regulatory regions.

Figure 4. Synergy between Zeb1 and Lef1 in transcriptional assays.

• A–D
  Transcriptional assays in P19 cells (A), LN229 cells (B), U87 cells (C), or Cb192 cells (D) co‐transfected as indicated in figure with Prex1::luc and Nrp2::luc (or mutated versions where HMG motifs were disrupted individually or in combination) or (HMG)x7::luc, and expression vectors for Lef1 and/or Zeb1. Data are shown as mean ± SD of four biological replicates (significance determined by one‐way ANOVA with Bonferroni correction). In all cases: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Figure EV2. Genome‐wide location analysis in human neural stem cell line Cb192 suggests indirect recruitment of Zeb1 by an HMG‐box transcription factor.

1. Density plot of Zeb1 ChIP‐Seq reads at 4‐kb genomic regions centered at peak summits (signal intensity represents normalized tag count). DNA motifs identified by a de novo search at 100‐bp regions centered at summits from a high confidence (P < 10E⁻¹⁰) or lower confidence (P < 10E⁻¹⁵) are shown, with respective statistical parameters.
2. Enrichment profile of E‐box and HMG motifs at 4‐kb genomic regions centered at Zeb1 peak summits.
3. Location of Zeb1 binding events in relation to closest annotated TSS.
4. Location of Zeb1 binding events respective to various genomic features.

Gene activation by Zeb1/Lef1 does not require canonical Wnt signaling

To investigate whether beta‐catenin is required for the Zeb1‐mediated gene activation of Nrp2 and Prex1 genes, we first assessed the status of Wnt pathway in normal culture conditions of GBM CSCs. Immunofluorescence staining followed by confocal microscopy showed that in all three GBM CSC models analyzed, beta‐catenin is found associated with plasma membrane and excluded from cell nuclei in control conditions, while it can be very effectively translocated into the nucleus in the presence of the Wnt agonist CHIR99021 (Fig 5A). This finding suggests that although Wnt pathway is not active when GBM CSCSs are grown in normal culture conditions, it can be activated with the addition of a Wnt agonist. Supporting this same conclusion, we were able to detect an enrichment of beta‐catenin on Wnt target genes (Axin2 and Tnfrsf19) only after GBM CSC lines were exposed to the Wnt agonist CHIR99021 (Fig 5B, bottom). In control culture conditions, binding of beta‐catenin was not detected at regulatory regions of Zeb1‐activated targets (although binding to Prex1 was detected in two of the three cell models tested but only upon Wnt activation) (Fig 5B), altogether demonstrating that activation of Zeb1/Lef1 target genes, which occurs in control culture conditions, does not require canonical Wnt signaling.

Figure 5. Activation of Nrp2 and Prex1 gene expression by Zeb1 does not require active Wnt signaling.

1. Confocal microscopy imaging of NCH421k, NCH441, and NCH644 cells showing cellular localization of beta‐catenin (green) in control conditions (DMSO, left) or upon incubation with CHIR99021 (right). Blue (DAPI). A high magnification of the small dashed square is shown below, for NCH421k cells. Scale bars 10 μm.
2. ChIP‐qPCR assessing beta‐catenin recruitment to Wnt‐ and Zeb1‐activated targets in NCH421k, NCH644, and NCH441 cells in control conditions (DMSO, top) or upon exposure to the Wnt agonist CHIR99021 (bottom).
3. A Lef1 mutant that does not interact with beta‐catenin (Lef1(AK)) was generated by introducing two point mutations in residues previously shown to be required for this interaction, and which are conserved across the protein family (marked with asterisks) (Graham et al, 2000; Grumolato et al, 2013). Transcriptional assay in P19 cells assessing the activity of Lef1 and Lef1(AK) in the absence, or presence, of a stabilized version of beta‐catenin (S33Y) on TOP‐flash::luc reporter.
4. Transcriptional assay in P19 cells assessing the activity of Zeb1, Lef1, or the defective beta‐catenin interacting mutant Lef1(AK) on Nrp2::luc reporter (left) or Prex1::luc reporter (right).

Data information: (B) Data are shown as mean ± SD of triplicate assays (significance determined by one‐way ANOVA with Fisher's LSD test). (C, D) Data are shown as mean ± SD of four biological replicates (significance determined by one‐way ANOVA with Bonferroni correction). In all cases: n.s. for non‐significant, *P < 0.05, ***P < 0.001, ****P < 0.0001.

Finally, we mutated two residues in Lef1 (D21A and E29K) that are conserved across the Lef/Tcf family and known to be required, based on structural and functional data, for the interaction of these transcription factors with beta‐catenin (Graham et al, 2000; Grumolato et al, 2013), and tested the effect of this mutation in transcriptional assays. As expected, the resulting protein Lef1(AK) was unable to interact with a stabilized version of beta‐catenin (S33Y) on the Wnt signaling reporter TOP‐flash (Fig 5C). Notably, this was able to transactivate the Nrp2 and Prex1 regulatory regions on its own (with the same efficiency of its wild‐type counterpart) and to function in synergy with Zeb1 to very similar levels, confirming that beta‐catenin is not required for the cooperative transactivation of target genes by Zeb1 and Lef1 (Fig 5D). Very similar results were obtained with Tcf1 and the equivalent mutant Tcf1(AK), further suggesting redundancy amongst Lef/Tcf family members in transcriptional synergy with Zeb1 (Fig EV3). In summary, these results demonstrate that activation of Zeb1 target genes in synergy with Lef/Tcf factors is independent of the Wnt pathway, which is not active in the absence of exogenously added Wnt agonists in our GBM CSC models.

Figure EV3. A Tcf1 mutant unable to interact with beta‐catenin functions in synergy with Zeb1 in transcriptional assays.

A Tcf1 mutant previously shown to be unable to interact with beta‐catenin (Grumolato et al, 2013) is equally able to transactivate the regulatory regions of Nrp2 (top) or Prex1 (bottom) in synergy with Zeb1 as its non‐mutated counterpart, in transfected P19 cells. Data are shown as mean ± SD of four biological replicates (significance determined by one‐way ANOVA with Bonferroni correction). In all cases: n.s. for non‐significant, **P < 0.01, ****P < 0.0001.

Activation of Wnt signaling does not promote the expression of Zeb1‐activated target genes

While previous results show that Wnt signaling is not required for Zeb1‐mediated gene activation, it is still possible that inducing the pathway will further potentiate the synergy between Zeb1 and Lef/Tcf factors. Contrary to the previous observation that Wnt signaling promotes the expression of EMT factors, incubation of NCH421k cells with Wnt3a unexpectedly resulted in decreased expression of Zeb1 transcript (Fig 6A), a result also observed using two Wnt agonists CHIR99021 and 6‐BIO (Fig 6B and C). A 50% reduction in Zeb1 transcript observed upon exposure to CHIR99021 resulted in a similar decrease in protein level (Fig 6D). Concomitant with reduced Zeb1 expression, exposure to Wnt3a, 6‐BIO, and CHIR99021 resulted also in a decrease in Nrp2 and Prex1 transcript levels (Fig 6A–C), with the exception of Nrp2 expression in NCH421K cells that remained unchanged in CHIR99021‐treated cells (Fig 6B). This discrepancy may result from differences in the activity of various Wnt agonists, resulting in compensation of Nrp2 expression by a Zeb1‐independent mechanism in CHIR99021‐treated cells. As expected, all Wnt agonists resulted in strong activation of Wnt target genes Axin2 and Nkd1 (Fig 6A–C). Exposure to 6‐BIO resulted also in reduced Zeb1 expression in NCH441 and NCH644 cells (Fig 6E and F). The observed decrease in Zeb1 expression level upon Wnt signaling activation resulted in a diminished recruitment of Zeb1 to HMG motif target sites (Itgb1, Nrp2, Prex1) below detectable levels by ChIP‐qPCR (Fig 6G), while binding of Zeb1 to E‐box containing regions in Pard6b and Zeb1 promoters decreased but was still detected (Appendix Fig S1). Thus, activation of Wnt signaling in GBM CSCs reduces Zeb1 expression and its indirect recruitment to regulatory regions, therefore not potentiating the expression of Zeb1/Lef1 target genes.

Figure 6. Wnt signaling activation does not promote the expression of Zeb1‐activated target genes.

• A–C
  Expression qPCR of Wnt‐ and Zeb1‐activated targets in NCH421k cells upon incubation with Wnt3a, 6‐BIO, or CHIR99021 as indicated in figure.
• D
  Zeb1 protein levels in control conditions or upon exposure to the Wnt agonist CHIR99021, assessed by quantitative Western blot (left).
• E, F
  Expression qPCR of Wnt‐ and Zeb1‐activated targets in NCH441 and NCH644 cells upon incubation with 6‐BIO.
• G
  ChIP‐qPCR showing Zeb1 recruitment to Zeb1‐activated targets in NCH421k cells in control conditions or upon exposure to the Wnt agonist CHIR99021.

Data information: (A, B, C, E, F) Data are shown as mean ± SD of two biological replicates (significance determined by unpaired two‐tailed t‐test). In (D), data are shown as mean ± SD of three biological replicates (significance determined by t‐test). (G) Data are shown as mean ± SD of triplicate assays (significance determined by one‐way ANOVA with Fisher LSD test). In all cases: n.s. for non‐significant, *P < 0.05, **P < 0.005, ***P < 0.001, ****P < 0.0001.Source data are available online for this figure.

Correlative expression of Zeb1 and Prex1 in GBM samples

We next searched for evidence that the gene activation function of Zeb1 characterized in GBM CSCs could play a regulatory role in the context of GBM tumors. With this aim, we performed correlative studies using publicly available datasets profiling the transcriptomes of tumors from large cohorts of GBM patients. We started by focusing on a previously characterized dataset consisting of 139 primary GBM (G‐CIMP negative) samples (Gravendeel et al, 2009). Notably, from the 60 activated Zeb1 targets identified in our study, 23 genes were indeed positively correlated with Zeb1, 15 of which have expected roles in cell adhesion/migration based on available literature (Table 1) and are therefore candidate mediators of tumor invasion downstream Zeb1. Of these 23 genes, 18 were also positively correlated with Zeb1 when a similar analysis was performed using another pool of 445 tumor samples from the TCGA data base (Table 1). Notably, Zeb1 expression and Prex1 expression were highly correlated in both analyses (Pearson correlation = 0.309 and 0.370 in Gravendeel and TCGA datasets) and a similar result is observed when tumor samples from the Classical subtype (where Prex1 expression is highest) are considered (Pearson correlation = 0.413 and 0.326) (Fig 7A–D).

Table 1.

Zeb1 targets correlated with Zeb1 in GBM samples and expected function in cell adhesion/migration

Gene	r (Pearson)	P‐value	Significance	TCGA	Predicted function in cell adhesion/migration
PITPNC1	0.4041	< 1.00E‐04	****	YES	YES
HMGB1	0.3949	< 1.00E‐04	****	YES	YES
DEPDC1B	0.3360	< 1.00E‐04	****	NO	YES
PTAR1	0.3348	< 1.00E‐04	****	YES	NO
SPRY4	0.3335	< 1.00E‐04	****	YES	YES
PHLDA1	0.3330	< 1.00E‐04	****	YES	YES
NUP93	0.3273	1.00E‐04	***	NO	YES
COL9A1	0.3201	1.00E‐04	***	YES	YES
COL20A1	0.3161	2.00E‐04	***	YES	YES
NCALD	0.3088	3.00E‐04	***	YES	NO
PREX1	0.3087	3.00E‐04	***	YES	YES
HAPLN1	0.2850	8.00E‐04	***	YES	YES
SOBP	0.2754	1.20E‐03	**	YES	NO
S100B	0.2592	2.30E‐03	**	NO	YES
BNIP1	0.2369	5.50E‐03	**	NO	NO
ACSBG1	0.2260	8.20E‐03	**	YES	NO
B3GALT2	0.2164	1.14E‐02	*	YES	NO
TSPAN3	0.2083	1.49E‐02	*	YES	YES
PID1	0.2006	1.92E‐02	*	YES	NO
CORO1C	0.1898	2.69E‐02	*	NO	YES
PACSIN2	0.1743	4.25E‐02	*	YES	YES
HEY2	0.1733	4.37E‐02	*	YES	YES
USP54	0.1723	4.49E‐02	*	YES	NO

*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Figure 7. Expression of Zeb1 and Prex1 are correlated in GBM tumor samples.

• A, B
  Scatter plot showing correlative expression levels between Zeb1 and Prex1 genes in all GBM or classical subtype in Gravendeel (A) or TCGA (B) datasets. R is Pearson correlation coefficient.
• C, D
  Box plot representing levels of Prex1 transcripts in GBM tumor samples from the Granvendeel (C) or TCGA (D) datasets categorized in different subtypes.
• E–G
  Representative immunohistochemistry from GBM patient tumors showing strong overlap of staining for Zeb1 (red) and Prex1 (brown). Rectangles 1 and 2 in (F) denote vascular proliferates and pseudopalisading tumor cells, respectively. A necrotic region is delimited by the white dashed line in (G). Nuclear counterstain with hematoxylin (blue). Scale bars 200 μm (E); 100 μm (F); 50 μm (G).
• H
  Immunohistochemistry at higher magnification from the core of a GBM patient tumor. Scale bar 20 μm.
• I–K
  Immunofluorescent co‐staining of Zeb1 (green) and Prex1 (red) showing correlated expression with few cells positive only for an individual marker. Nuclear staining with DAPI (blue). Scale bar 10 μm.
• L
  Quantitative Western blot analysis (left) and corresponding quantification (right) of Zeb1 and Prex1 protein levels in control and Zeb1‐silenced NCH421k cells.

Data information: (C, D) Data distribution is represented as box with median and first and third quartiles. Whiskers go down to the 5^th and 95^th percentile, with points outside drawn as individual dots. (L) Quantification is shown as mean ± SD of three biological replicates (significance determined by t‐test), with *P < 0.05, **P < 0.01.Source data are available online for this figure.

Considering the potential confounding effects of a correlation study performed at a cell population level, we compared the expression of Zeb1 and Prex1 by coimmunostaining in sectioned GBM tumor samples. In all tumor cases examined, Prex1 and Zeb1 showed a strong overlap in their expression (> 95%) in the main tumor mass and a higher expression in palisading tumor around or close to necrotic areas (Figs 7F–H and EV4). Only in vascular proliferates, many Zeb1‐positive cells do not express Prex1 (Fig 7F). In one case with an infiltrative zone in brain parenchymal tissue, Prex1 expression was also indicative of Zeb1 expression, whereas Zeb1‐expressing cells are not always Prex1‐positive (Fig 7E). However, in infiltrating zones and adjacent parenchyma, it is not possible to clearly distinguish tumor from non‐tumor cells. Co‐expression of both genes at the cellular level was also confirmed by multicolor immunofluorescence staining. Although we observed cells that were positive for only one gene, most cells in the tumor mass that were Prex1‐positive also showed nuclear immunoreactivity for Zeb1 and vice versa (Fig 7I–K).

Figure EV4. Zeb1 and Prex1 are strongly co‐expressed at core of GBM tumor samples.

Representative images and table, describing features of all tumor cases analyzed in this study. Numbers shown are percentage of total number of cells that are Prex1 only (Prex1+) or Prex1/Zeb1 double positive (Prex1+/Zeb1+). Zeb1 staining (red), Prex1 staining (brown), hematoxylin counterstaining (blue). All cases were diagnosed as grade IV astrocytoma. Scale bar 100 μm.

Finally, we observed that Zeb1 knock‐down in NCH421k cells leads to a significant (28%) reduction in Prex1 protein expression (Fig 7L). In conclusion, our results are in line with the high correlation shown by the transcriptomics data and support the conclusion that Prex1 is positively regulated by Zeb1 in GBM tumors.

Prex1 expression promotes migration and invasion of GBM cells in vitro and in vivo and correlates with shorter patient survival

Given the predicted role of Prex1 in cell motility, we next assessed the importance of Prex1 for glioma cell migration. Using a sequence‐specific shRNA targeting Prex1, we performed a scratch assay on three different GBM CSC models growing in adherent conditions (Fig 8A–C). Reduced Prex1 expression resulted in a strong reduction in the total area covered by cells after 24 h in all cellular models tested, as compared to non‐transduced cells or cells expressing a control shRNA, suggesting that attenuation of Zeb1 expression decreases migration of GBM cells.

Figure 8. Prex1 promotes migration and invasion of GBM cells in vitro and in vivo .

1. Representative images of scratch assay of glioblastoma cells infected with a control or Prex1 specific shRNA lentiviral vector. Scale bar 100 μm.
2. Quantification of remaining exposed area at 24 h for three different cellular models, as indicated in figure. Data are shown as mean ± SD of four independent experiments (significance determined by unpaired two‐tailed t‐test). n.s. for non‐significant, *P < 0.05, **P < 0.01.
3. Quantification of Prex1 transcript levels in parental glioblastoma cells, or cells expressing a control or Prex1‐specific shRNA. Data are shown as mean ± SD of triplicate assays.
4. Representative images of tumorspheres from control, Prex1‐, and Zeb1‐silenced NCH421k cells seeded onto mouse brain slice cultures. Scale bar 200 μm.
5. Quantification of phenotypes of tumorspheres shown in (D).
6. Representative tumors resulting from intracranial transplantation of control or Prex1‐silenced NCH421k cells, as indicated in figure. Tumor cells were visualized by staining for the human intermediate filament Nestin. Right panels display a magnified view of region delimited by white insets in left panels. Scale bar 200 μm.
7. Quantification of tumor invasiveness was calculated by determining the ratio of the area covering the core tumor mass over the area of the invasive front of tumor cells. Data are shown as mean ± SD of 5 controls and 6 tumors with Prex1 knock‐down, with significance determined by unpaired two‐tailed t‐test (see Materials and Methods for details on the analysis).
8. Kaplan–Meier curve comparing survival amongst GBM patients with high or low levels of Prex1 transcript (median cut off). Analysis performed with all GBM patients from Gravendeel dataset, or restricted to Classical subtype from Gravendeel or Rembrandt datasets are shown, as indicated in figure. P‐value determined according to log‐rank test.

To further assess a role for Prex1 in glioma cell invasion, we seeded tumorspheres of Prex1‐ or Zeb1‐silenced and control non‐silenced NCH421k cells onto organotypic slice cultures of adult mouse brains as an ex vivo tumor model (Fayzullin et al, 2016). In our experimental setup, tumor spheres exhibited either an invasive phenotype with substantial numbers of cells migrating out‐ or completely dispersing into the parenchyma, or a non‐invasive phenotype with tumorspheres retaining a round and compact appearance and none to few cells migrating out. We observed that 80% of non‐silenced NCH421k spheres displayed brain parenchyma infiltrative behavior, compared to ~40% for the Prex1‐silenced and 14% for the Zeb1‐silenced NCH421k spheres (Fig 8D and E).

Next, dissociated Prex1‐silenced or non‐silenced NCH421k cells were orthotopically implanted into immunocompromised mice. Unexpectedly, tumors initiated by Prex1‐silenced NCH421k cells were on average 12 times larger than controls (Fig 8F). This was possibly due to an increased number of tumor‐initiating cells, as suggested by increased expression levels of alpha‐6 integrin, which is co‐expressed with other markers of and enriches for tumor‐initiating cells in GBM (Lathia et al, 2010) (Appendix Fig S2). In line with this, Prex1‐silenced cells displayed an increased tumor sphere forming capacity compared to non‐silenced NCH421k cells (141.5 ± 14.1%, SD, n = 4, P = 0.001). Importantly, taking the difference in tumor size in orthotopic transplantations into account, brain tumors arising from Prex1‐silenced cells displayed a significant decrease in invasive behavior, measured as the ratio of the area covered by tumor cells at the invasive front over the area of the main tumor mass, compared to non‐silenced cells (Fig 8F and G). Together, these results provide evidence that Prex1 supports the invasive behavior of GBM cells.

We finally asked whether Prex1 expression in GBM tumor samples could be of prognostic value. Prex1 expression was indeed correlated with reduced patient survival (median survival of 7.3 months against 9.65 months of patients with low Prex1 expression) considering the Gravendeel dataset, with this difference increasing to 4.6 months within the GBM Classical subtype (median survival of 7.6 against 12.2 months for patients with high and low Prex1 expression, respectively) (Fig 8H). Although we did not find any statistically significant differences with the cohort of patients represented in the TCGA dataset, high Prex1 expression was also predictive of shorter survival when classical subtype GBM patients present in the Rembrandt dataset (Madhavan et al, 2009) were considered (median survival of 11.8 against 14.0 months). Altogether, functional assays and patient survival data implicate Prex1 as an important player in GBM, highlighting the importance of the Zeb1 activation mechanism uncovered in this study.

Discussion

Although previous studies have shown the importance of Zeb1 as a regulator of EMT in various types of cancer cells, the nature of its transcriptional program remains poorly understood. To address this gap in knowledge, we performed a genome‐wide characterization of Zeb1 target genes in GBM. Our results establish a new paradigm of Zeb1 function, whereby indirect recruitment via HMG motifs present at regulatory regions results in gene activation. Genome‐wide mapping of Zeb1 binding sites revealed a large‐scale use of this mechanism, a surprising finding considering that the ability to promote EMT has been ascribed mostly to its transcriptional repressor activity.

Several lines of evidence converge on Lef/Tcf factors as important players in this novel mechanism, including mediating Zeb1 recruitment to regulatory regions. They are strongly enriched at all tested HMG‐type Zeb1 “peaks”, promote gene activation in synergy with Zeb1 in a Wnt‐independent manner, and can be part of a complex with Zeb1 when co‐expressed. However, besides Lef/Tcf factors, it is possible that additional HMG‐box transcription factors expressed in GBM CSCs may also function in the same gene regulatory regions. These could include members of the Sox family, which are known to display complex interactions with components of the Wnt pathway and bind to regulatory regions of genes regulated by Lef/Tcf factors (Kormish et al, 2010; Zhang et al, 2013; Zhou et al, 2016).

How indirect recruitment of Zeb1 to regulatory regions results in gene activation, and in particular what transcriptional cofactors mediate this Zeb1 function, remains to be elucidated. Zeb1 was recently shown to physically interact with Tcf4 to activate LAMC2 and uPA genes in colorectal cancer cells, via a mechanism that involves the recruitment of p300 (Sánchez‐Tilló et al, 2015). In sharp contrast with our findings however, this activity requires direct binding of Zeb1 to E‐box sequences and is strictly dependent on active Wnt signaling. Various lines of evidence converge in the conclusion that gene activation by Zeb1 via this novel mechanism does not require nuclear beta‐catenin. Other cases of Wnt‐independent Lef1 functions have been reported, namely in association with ATF2 transcription factors (Grumolato et al, 2013). It is possible, however, that in other cell types the synergistic activity of Zeb1 and Lef1 may be further enhanced by canonical Wnt signaling, defining a cross‐talk with this important pathway.

Our study resulted in the identification of transcriptional targets with expected functions in malignant cell invasion in GBM, as exemplified by Prex1. In line with the transcriptional regulation data, we found Zeb1 and Prex1 to be co‐expressed in large areas of both core and periphery of all tumors analyzed. The fact that Zeb1 expression is not restricted to the infiltrating tumor rim as previously suggested (Siebzehnrubl et al, 2013; Zhang et al, 2016) is in agreement with the view that GBM growth occurs concomitantly with cell migration in various regions across the tumor, in particular around hypoxic and pseudopalisading necrotic areas. In addition, it is also supported by the finding that Zeb1 expression is promoted by hypoxic conditions, more likely to be found at the tumor core (Kahlert et al, 2015; Zhang et al, 2016).

Our results conclude for the absence of Wnt signaling in the GBM CSC models analyzed when growing in the absence of Wnt agonists. We do not expect this to be a general feature of GBM CSC models, since previous work suggested that Wnt pathway status varies across cellular models representing different GBM subtypes (Huang et al, 2016). In addition, we found that activation of canonical Wnt signaling reduced Zeb1 expression in all tested models, an unexpected result considering that Wnt activation promotes Zeb1 expression and EMT in other cancer contexts (Sánchez‐Tilló et al, 2011a). One possibility is that culture conditions in our experiments do not fully recapitulate those existent in a tumor niche, such as its hypoxic environment. It should be noted, however, that observations linking Wnt signaling with GBM tumor invasion are still scarce and warrant further investigation. Moreover, the suggestion that Wnt activates Zeb1 in GBM CSCs was found so far in a single study describing the knock‐down of beta‐catenin in NCH421k cells (Kahlert et al, 2012), which results in decreased Zeb1 expression (an unpublished observation that we reproduced). Given evidence showing absence of Wnt signaling in our GBM CSC models in control conditions, we attribute this result to the consequence of knocking down the activity of non‐nuclear beta‐catenin.

Prex1 gene encodes a guanidine‐exchange factor (GEF) for the Rho family of small GTP‐binding proteins (RACs), which was shown to promote metastatic growth by a Rac1‐dependent mechanism in various cancer types (Ebi et al, 2013; Dillon et al, 2015; Lucato et al, 2015). Both Rac1 and downstream effectors of Rho‐GTPases have been shown to promote invasiveness and survival of malignant cells in GBM (Fortin Ensign et al, 2013). In line with that, our observations provide evidence for a role for Prex1 in supporting GBM cell invasion in vivo. The results of brain tumor xenograft studies also revealed an unexpected increase in the size of tumors initiated by Prex1‐silenced GBM CSCs. This in vivo observation was correlated with an increase in alpha‐6 integrin expression and increased sphere formation ability by the Prex1‐silenced cells in vitro, suggesting that Prex1 attenuation increases the relative fraction of cells with tumor‐initiating ability, thereby leading to enhanced tumor formation. Moreover, our observation that Prex1 silencing resulted in decreased cancer cell dissemination in vivo despite this increase in tumor size suggests the previously unrecognized possibility that Prex1 is involved in mechanisms controlling the balance between more invasive cancer cells with lower tumor‐forming ability and tumor‐initiating cells with lower invasive potential (Lee et al, 2006; Xie et al, 2014; Paw et al, 2015). Also in agreement with a function for Rac/Prex1 pathway in promoting the invasive phenotype of GBM tumors, we found Prex1 expression to be of prognostic value in the classical subtype, where its expression is highest. This may not reflect a subtype‐specific regulation of Prex1 by Zeb1, since their correlative expression is not exclusive to classical GBM tumors. Instead, it may be explained by the genetic aberrations that characterize this tumor class (high‐level EGFR amplification and partial or complete loss of PTEN) (Madhavan et al, 2009) that are expected to lead to an over‐activation of the PI3K pathway (Bjorge et al, 1990; Wee & Wang, 2017), previously shown to activate Prex1 by phosphorylation (Sosa et al, 2010; Ebi et al, 2013; Dillon et al, 2015; Lucato et al, 2015).

Although Rho‐GTPases and their associated proteins are known mediators of cell motility in EMT programs, this is to our knowledge the first direct link between a classical EMT inducer and Prex1. It will be important to understand whether Zeb1 also promotes expression of Prex1 in other types of cancer, and to which extend Rho/Rac pathway components are under the direct control of classical EMT activators.

Interestingly, a genome‐wide survey using Genomic Regions Enrichment of Annotations Tool (GREAT) (McLean et al, 2010) revealed a strong association of Zeb1 binding sites with genes enriched for cellular component terms associated with cell shape in the context of motility (“actin cytoskeleton”, “lamellipodium”, and “filopodium”), together with various pathway terms associated with cadherins (“N‐cadherin/E‐cadherin signaling events”) and activity of small GTPases (“regulation of CDC42/Rac1/RhoA”) (Fig EV5). Although this analysis is exclusively based on Zeb1 binding and therefore lacks evidence for gene regulation, it strongly suggests that the Zeb1 transcriptional program in GBM is more broadly implicated with mesenchymal cell motility, controlling this process at different levels. Future studies should validate additional Zeb1 target genes in this context.

Figure EV5. Genome‐wide binding profile of Zeb1 is enriched at genes associated with terms associated with EMT and mesenchymal cell migration.

The Genomic Regions Enrichment of Annotations Tool (GREAT) was applied to the list of genomic regions bound by Zeb1 in NCH421k cells. Enrichment of Gene Ontology terms associated with Cellular components, Pathways, and Disease is shown. See the Extended Experimental Procedures section in the Appendix for details of the analysis.

Overall, our findings provide a model whereby Zeb1 can simultaneously activate and repress gene expression in the same cellular context, and in this way coordinate a complex genetic program that shares many similarities with an EMT. We do not expect this mechanism to be restricted to GBM CSCs, as have observed recruitment of Zeb1 via HMG motifs in the human fetal neural stem cell line Cb192. Future studies should address the importance of this new paradigm in other cellular contexts, both during development and cancer.

Materials and Methods

Cell culture and infection of GBM CSCs

NCH421k, NCH644, and NCH441 and Cb192 cells were cultured in DMEM‐F12 GlutaMAX medium (GIBCO) supplemented with 1× N‐2 Supplement (GIBCO), 0.05× B‐27 supplement (GIBCO), Penicillin‐Streptomycin (100 U/ml) (Gibco), EGF (10 ng/ml) (Peprotech), bFGF (10 ng/ml) (Peprotech), and Laminin (1 μg/ml) (Sigma‐Aldrich) in T‐flasks, plates, or well plates (Corning) pre‐coated with sterile‐filtered Poly‐L‐Lysine (Sigma‐Aldrich). P19 embryonic carcinoma cells, human embryonic kidney cells (293T), LN229, and U87MG cells were maintained in Dulbecco's modified Eagle's medium (DMEM)/High glucose (Gibco) supplemented with Fetal Bovine Serum Heat Inactivated (10%) (PAA Laboratories, GE Healthcare), Penicillin‐Streptomycin (100 U/ml) (Gibco), and l‐Glutamine (2 mM) (Gibco). Replication‐incompetent lentiviruses were produced by transient transfection of 293T cells with lentiviral vectors expressing control or sequence‐specific shRNAs (see extended experimental procedures for vectors used), co‐transfected with the viral packaging vector psPAX2, and the viral envelope vector pCMV‐VSVG. Medium was replaced 14 h post‐transfection and 48 h after lentiviral particles were concentrated by ultracentrifugation at 90,000 g for 4 h and resuspended in 0.1% BSA PBS. NCH421k cells were infected 24 h after plating.

Immunofluorescence of cells

NCH421k, NCH644, and NCH441 cells were grown on glass coverslips coated with poly‐L‐Lysine (Sigma‐Aldrich) and fixed with 4% formaldehyde for 10 min. Immunofluorescence on fixed cells was performed using standard procedures (see Appendix Table S7 for antibodies used). Cell nuclei were stained with DAPI (4′,6‐diamidino‐2‐phenylindole; Sigma‐Aldrich) before mounting in Aqua Poly/Mount (Polysciences). Confocal images were acquired on a Leica SP5 confocal, using a 63× 1.3NA Oil immersion objective, and processed with ImageJ.

Western blotting

Crude cell lysates and immunoprecipitated samples were diluted in 2× Laemmli buffer (Sigma‐Aldrich) and denatured for 5 min at 95°C. Samples were separated in 10% SDS–PAGE gels and transferred to nitrocellulose membranes (GE Healthcare) using standard procedures. Blots were probed with the primary, HRP‐conjugated or fluorescent secondary antibodies listed in Appendix Table S7. For quantitative Western blot analysis, the ratio of Zeb1/Prex1 over tubulin was normalized against control samples, using images obtained with an Odyssey Scanner and quantified with Fiji image software.

Gene expression and DNA microarrays

Gene expression analysis of cultured cells by quantitative real‐time PCR using PerfeCTa SYBR Green FastMix, ROX (Quanta Biosciences) or iTaq Universal SYBR Green Supermix (Bio‐Rad) and 7900HT Real Time PCR System or QuantStudio 7 Flex Real‐Time PCR System (Applied Biosystems) was carried out according to the manufacturers' instructions using cDNA produced with High Capacity RNA‐to‐cDNA Master Mix (Applied Biosystems) after Trizol RNA extraction (see Appendix Table S5 of extended experimental procedures for primer sequences). Relative expression is normalized to reference genes TBP and GAPDH expression levels, and to untreated samples. Microarray analysis was performed on biological triplicates of NCH421k cells 72 h upon infection. RNA concentration and purity were confirmed using Agilent 2100 Bioanalyzer with RNA Nano Kit (Agilent Technologies). 100 ng of RNA was processed with Ambion WT Expression Kit (Life Technologies) and hybridized to the Affymetrix Mouse Gene 1.0 ST Array, according to the manufacturers' protocol. CEL files were analyzed using Chispter software (v 3.1.0, Kallio et al, 2011) using RMA normalization and empirical Bayes two‐group test with Benjamini‐Hochberg post hoc for P‐value correction.

Chromatin immunoprecipitation and ChIP‐Seq

Cells were fixed sequentially with 2 mM di(N‐succinimidyl) glutarate and 1% formaldehyde in phosphate‐buffered saline and lysed, sonicated and immunoprecipitated as described (Castro et al, 2011), using rabbit anti‐Zeb1 (ab1424, Abcam). DNA sequences were quantified by real‐time PCR (see Appendix Table S6 of extended experimental procedures for primer sequences) using PerfeCTa SYBR Green FastMix, ROX (Quanta Biosciences) or iTaq Universal SYBR Green Supermix (Bio‐Rad), and 7900HT Real Time PCR System or QuantStudio 7 Flex Real‐Time PCR System (Applied Biosystems). Quantities of immunoprecipitated DNA were calculated by comparison with a standard curve generated by serial dilutions of input DNA. For sequencing, DNA libraries were prepared from 10 ng of immunoprecipitated DNA according to manufacturers' protocol. Single‐end sequencing was performed using HiSeq‐2000. Raw reads were mapped to the human genome (GRCh37/hg19) assembly with Bowtie version 0.12.7 (Langmead et al, 2009). Data were further processed with MACS version 1.4.1 (Zhang et al, 2008), to define location of binding events (see extended experimental procedures for further details on data processing).

In vitro binding and transcriptional assays

Electrophoretic mobility shift assays were performed as described (Castro et al, 2006), using probes (see Appendix Table S4 for sequences) labeled with [γ‐32P] ATP (Perkin Elmer) using T4 polynucleotide kinase (New England Biolabs). Zeb1 and Lef1 proteins produced by coupled in vitro transcription and translation in rabbit reticulocyte lysates (TNT, Promega) were incubated with probe in 20 μL binding reactions (15% Glycerol, 20 mM HEPES pH 7.9, 5 mM MgCl₂, 50 mM KCl, 0.01% Triton X‐100, 10 mM DTT, 5 mM PMSF, 0.2 μg/μl herring sperm DNA (Sigma‐Aldrich D7290) for 20 min at RT, and loaded onto 6% non‐denaturing polyacrylamide gels in TBE running buffer (89 mM Tris‐base, 89 mM Boric Acid, and 2 mM EDTA). Reporter gene assays were performed as previously described (Castro et al, 2011), with various cell lines transfected with Lipofectamine 2000. Oligonucleotides used to generate the luciferase constructs, and to insert mutations with QuickChange Site‐Directed Mutagenesis Kit (Stratagene), are described in extended experimental procedures.

Analysis of Gravendeel, Rembrandt, and TCGA GBM datasets microarray data

Glioblastoma expression data of the Gravendeel glioma dataset (Gravendeel et al, 2009) (GEO Accession number GSE16011), the TCGA GBM dataset (McLendon et al, 2008), and the Rembrandt dataset (Madhavan et al, 2009) were analyzed and extracted through the GlioVis portal (Bowman et al, 2017). The Kaplan–Meier method was used to perform survival analyses on groups classified by Prex1 expression levels. In all survival analyses, the outcome variable was time from start of treatment until death. Subjects still alive at the time or analysis and subjects lost to follow‐up were considered censored. P‐value determined by log‐rank test.

Human tissue samples and immunohistochemistry

Intraoperative specimens of brain tumor patients were obtained from the Department of Neurosurgery or the Edinger Institute of the University Hospital. All work involving human tissue was approved by the local ethical committee (ethical votes No. GS‐04/09 and GS‐249/11). Tissue samples were dehydrated and embedded in paraffin prior to cutting into 2‐ to 3‐μm sections on a microtome (Leica SM2000R, Wetzlar, Germany). Clinical cases were stained using antibodies against Zeb1 (HPA027524, Sigma) and Prex1 (HPA001927, Sigma) on a fully automated Leica Bond III (Leica Biosystems, Germany) using the Leica Bond Polymer Refine Detection Kits (DS9800+ DS9390) (see extended experimental procedures for more details). For fluorescence microscopy, the Tyramide Signal Amplification (TSA) method was applied, using an Opal‐4 color kit (NEL794001K, PerkinElmer Inc., Waltham, USA). Fluorescent images were captured on a confocal LSM Nikon TE2000‐E microscope. Images were processed using EC‐C1 3.60 software and ImageJ (NIH).

Cell migration assay

For migration assays, cells expressing shRNA against Prex1 or control were dissociated and resuspended in DMEM, 10% FCS and 35,000 cells were pipetted into each chamber of silicone inserts with 500 μm gap (Ibidi, Munich, Germany). Cells were left to attach overnight, and the inserts were removed creating a 500 μm cell‐free gap. Phase contrast images of the same gap fields were captured at 0 h and 24 h of incubation using an inverted light microscope (Olympus IX70) with camera (Soft Imaging Solutions). The remaining cell‐free area was quantified using ImageJ software and values of area relative to timepoint 0 h set at 100.

Invasion assay using organotypic brain slices

Organotypic brain slices from wild‐type C57/BL6 mice (6–8 weeks) were prepared as described previously (Kraft et al, 2017). Mice were killed by cervical dislocation, and the brains were excised, washed briefly in ice‐cold Krebs buffer, and embedded in low melting agarose in plastic embedding molds. Slices were cut at 300–350 μm thickness on a vibratome (Leica, Germany). Brain slices were transferred onto porous poly‐carbonate membranes and cultivated in a humidified atmosphere in serum‐containing medium for 2 h, with medium replaced every day thereafter. Tumorspheres grown from control NCH421k cells or with silenced Prex1 or Zeb1 expression were pipetted onto the slices and their invasive phenotype determined at day 8 as “infiltrative” (showing significant cancer cell dispersion into surrounding tissue), or “non‐infiltrative” (with no or very few cells leaving the spheres). Images were acquired with a Nikon SMZ25 stereomicroscope fitted with a P2‐SHR Plan Apo 2× objective using the NIS Elements Software (Nikon). In total, 61 tumorspheres (20 for Prex1 knock‐down, 25 controls, 16 Zeb1 knock‐down) seeded on 16 brain slices were analyzed. Results are given as the percentage of spheres displaying an infiltrative phenotype.

Brain tumor xenografts

GBM‐derived NCH421k cells were transduced with lentivirus expressing either non‐silencing‐shRNA or Prex1‐shRNA sequences as described above. After 4 days of selection in the presence of puromycin (1 μg/ml), cells were tested for successful Prex1 silencing by quantitative PCR. See extended experimental procedures section of Appendix for details on FACS analysis and self‐renewal assays. Prior to stereotactic implantations into the corpus striatum of the right hemisphere of 6‐ to 8‐week‐old CB‐17 NOD‐SCID male mice (Charles‐River Laboratory), transduced cells were mechanically dissociated to single‐cell suspensions and viable cells were quantified by cell counting using trypan blue. Approximately 3 × 10⁴ cells in 2.5 μl of PBS were intracranially injected into each mouse using the following stereotactic coordinates: anteroposterior +1.0, mediolateral +1.5, and dorsoventral −2.5. Mice were euthanized 6 weeks after implantation and brains were collected, rinsed in ice‐cold PBS, followed by fixation in 4% paraformaldehyde, cryopreservation, and embedding in Tissue‐Tek OCT compound (Sakura Finetek) as described (Verginelli et al, 2013). Frozen samples were cryostat‐sectioned (14 μm), mounted onto SuperFrost glass slides (Fisher), and stored at −20°C until use. Sections were subjected to immunohistochemistry with either antibodies against human Nestin or anti‐human nuclear antigen (see extended experimental procedures), followed by counterstaining with hematoxylin. For analysis of tumor invasion, images were color‐deconvoluted with the Fiji (ImageJ, NIH) software to obtain images with DAB‐stained human Nestin only. The contrast of the staining was increased by setting a threshold, and then, the area of the high density core tumor mass was defined and measured. For the calculation of the infiltrated brain area in relation to the tumor core size, two to three images with a complete hemisphere at different anterior–posterior positions from each tumor were processed. Next, a second area along the border of the invasive front of tumor cells was defined. The quotient from the area of the invasive front and tumor core was calculated as a value reflecting infiltrating depth taking into account the size of the tumor. Animal procedures were conducted in accordance with the guidelines of the Canadian Council for Animal Care and were approved by the Animal Care Committee of the Montreal Neurological Institute of McGill University.

Statistical analysis

For orthotopic transplantations, five was chosen as the minimum final number of animals per condition according to previous studies in the same field, while simultaneously keeping into consideration the need to limit numbers of animals used (3R′s). No randomization process was used, and two animals were excluded from control group since tumors did not develop, following previously established criteria. Results were confirmed on selected cases (n = 6) by a researcher blind to the identity of samples being quantified. Values are depicted as the average in fold size of the area of tumor invasion relative to the core tumor area. Errors are expressed as ± standard deviation and significance determined by unpaired two‐tailed t‐test. In all other assays, results are plotted as mean ± standard deviation from four (transcriptional and migration assays), three (quantitative Western blotting), two (expression data), or one biological replicate from a representative experiment (ChIP‐PCR). We followed common standards in the field and assumed data met a normal distribution (although sample sizes are too small for this assumption to be tested), and significance was determined by parametric tests (as indicated in figure legends) without assuming similar variance, using GraphPad Prism 6.

Data deposition

All genomic datasets generated in this study have been submitted to the Array Express database (ebi.ac.uk/arrayexpress) with the following identifiers: Zeb1 ChIP‐seq in NCH421k and Cb192 cells (E‐MTAB‐5541) and DNA arrays of Zeb1 LoF in NCH421k cells (E‐MTAB‐5543).

Author contributions

PR conception and design, collection and/or assembly of data, data analysis, and interpretation and manuscript writing. SM concept and design, collection, and/or assembly of data. VP concept and design, collection, and/or assembly of data. VT collection and/or assembly of data. AASFR data analysis and interpretation. RW data analysis and interpretation. RB collection and/or assembly of data. YT collection and/or assembly of data. CH‐M provision of cell lines, data analysis, and interpretation. SS conception and design, data analysis, and interpretation. SM conception and design, data analysis and interpretation, manuscript writing. DSC concept and design, data analysis and interpretation, manuscript writing, and final approval of manuscript.

Conflict of interest

The authors declare that they have no conflict of interest.

Supporting information

Appendix

Expanded View Figures PDF

Table EV1

Table EV2

Table EV3

Table EV4

Review Process File

Source Data for Figure 1

Source Data for Figure 3

Source Data for Figure 6

Source Data for Figure 7

The EMBO Journal (2018) 37: e97115