Targeting the Notch-regulated non-coding RNA TUG1 for glioma treatment

Targeting self-renewal is an important goal in cancer therapy and recent studies have focused on Notch signalling in the maintenance of stemness of glioma stem cells (GSCs). Understanding cancer-specific Notch regulation would improve specificity of targeting this pathway. In this study, we find that Notch1 activation in GSCs specifically induces expression of the lncRNA, TUG1. TUG1 coordinately promotes self-renewal by sponging miR-145 in the cytoplasm and recruiting polycomb to repress differentiation genes by locus-specific methylation of histone H3K27 via YY1-binding activity in the nucleus. Furthermore, intravenous treatment with antisense oligonucleotides targeting TUG1 coupled with a drug delivery system induces GSC differentiation and efficiently represses GSC growth in vivo. Our results highlight the importance of the Notch-lncRNA axis in regulating self-renewal of glioma cells and provide a strong rationale for targeting TUG1 as a specific and potent therapeutic approach to eliminate the GSC population.

G lioblastomas (GBMs) show heterogeneous histological features. These distinct phenotypes are thought to be due to their derivation from glioma stem cells (GSCs), which are a highly tumorigenic and self-renewing sub-population of tumour cells that have different functional characteristics 1,2 . As the self-renewing capability and multipotency may be closely associated with tumour progression and invasion, understanding the mechanisms underlying the regulation of stemness properties of tumour cells has been actively investigated 3-5 . In the neural stem cell, the Notch signalling pathway has a dominant role in inhibiting differentiation through the activities of its downstream effectors, such as Hairy and enhancer of split 1/5 (Hes1/5), which repress the implementation of neurogenic programs 6 . In the context of oncogenesis, Notch signalling has been shown to promote GSC self-renewal and to suppress GSC differentiation 7 . The mechanisms by which Notch regulates brain tumour stem cells appear to be similar to those governing regulation of neural stem cells during neural development 8 . However, the mechanism by which Notch signalling and its downstream effectors maintains the stemness properties of GSCs through the function of a certain set of genes, such as SOX2, MYC and Nestin, remains unresolved.
It has been increasingly demonstrated that large parts of the human genome are transcribed into non-coding RNAs (ncRNAs). A recent study showed that Notch-triggered oncogenic activity can be due to not only its ability to regulate proteins, but also long ncRNAs (lncRNAs) in leukaemia 9 . To date, studies have shown that lncRNAs affect the chromatin structure, mRNA stability and miRNA-mediated gene regulation by acting as competing endogenous RNA or natural miRNA sponges [10][11][12] .
Among these lncRNAs, Taurine Upregulated Gene 1 (TUG1) has been reported as cancer-related, and can bind to polycomb repressive complex 2 (PRC2) or PRC1 as well as repress gene expression 13,14 . TUG1 was originally identified as a transcript upregulated by taurine, whose function is associated with retinal development 15 . In the context of human malignancies, it is overexpressed in bladder cancer, gastric cancer and osteosarcoma [16][17][18] , whereas it is downregulated in non-small cell lung cancer 19 , suggesting context-dependent roles in different types of cancers. As the length of the TUG1 lncRNA is not short, B7.1 kb, it is plausible that TUG1 has multiple functions, which remain unknown.
Here, we show that TUG1, the expression of which is regulated by the Notch signalling pathway, was highly expressed in GSCs and maintained the stemness features of glioma cells. Furthermore, we developed new antisense oligonucleotides (ASO) targeting TUG1 coupled with a potent drug delivery system (DDS), which can be used intravenously to provide efficient and selective delivery to glioma cells at sufficient concentrations to acquire antitumour effects 20 . Our observations indicate that Notch-directed TUG1 is an effective epigenetic modulator that regulates the cancer stem cell population.

Results
Identification of TUG1 as a Notch-regulated lncRNA. The two GSC populations assessed (1228-GSC and 222-GSC) showed high levels of GSC markers, such as CD15 and SOX2, and GSC characteristics by functional analyses (in vivo limiting dilution assay) ( Supplementary Fig. 1), which agree with the previously described model 1,21 . To identify lncRNAs that are regulated by Notch signalling, we first performed RNA-sequencing (RNA-seq) analysis to investigate the lncRNA expression profile in these two GSC populations 5,22 . In Notch signalling, the Notch intracellular domain (NICD) translocates into the nucleus and binds the transcription factor, RBPJk 23 . Therefore, we selected lncRNAs that contain the binding motifs of RBPJk (-TTCCCAG/C-) around their transcriptional start site (TSS, À 2 kb to þ 1 kb, À indicates upstream of TSS and þ indicates downstream of TSS) using the JASPAR CORE database (http://jaspar.binf.ku.dk/) 24 . Commonly highly expressed lncRNAs with associated RBPJk motifs in both GSC populations were identified (Top 50, Fig. 1a).
To identify lncRNA potentially involved in Notch-induced GSC maintenance, we further performed lncRNA microarray analysis in GSCs with either treatment by small interfering RNA (siRNA) targeting Notch1 (si-Notch1), si-JAG1 (Notch ligand) 25 , or g-secretase inhibitors (N-S-phenyl-glycine-t-butyl ester (DAPT) and RO4929097, which is undergoing clinical trial in glioma) (Supplementary Data 1). Notably, the use of DAPT or RO4929097 may not be very specific to Notch signalling inhibition but rather inhibition of g-secretase. Therefore, we also examined lncRNA expression in si-JAG1 and si-Notch1 treated GSCs by RNA-seq analysis in addition to lncRNA microarray analysis and found that results of both analyses were quite concordant (R 2 ¼ 0.9014 and 0.9152 in 1228-and 222-GSCs, respectively) (Supplementary Fig. 2 and Supplementary Data 2). These analyses revealed that six lncRNAs were commonly downregulated in the two GSC populations following inhibition of Notch signalling (Fig. 1b, Supplementary  Fig. 3a-d and Supplementary Data 3). Among these, TUG1 (NR_002323.1) was the only lncRNA, which was also highly expressed as determined by RNA-seq analysis and possessed multiple RBPJk motifs around its promoter region (Fig. 1a,c). Indeed, Notch1 bound to RBPJk motifs within the upstream region of TUG1 TSS (Fig. 1d). We confirmed that level of TUG1 expression was efficiently reduced in GSCs following treatment with either si-Notch1, si-JAG1 or g-secretase inhibitors (DAPT and RO4929097) ( Fig. 1e-g and Supplementary Fig. 3a-d). Compellingly, downregulation of TUG1 expression by Notch signalling inhibition was rescued by ectopic NICD expression (Supplementary Fig. 3e-g). These data indicate that Notch signalling predominantly regulates TUG1 expression in glioma cells.
TUG1 maintains stemness features of GSCs. Inhibition of TUG1 by siRNA in GSCs reduced cell proliferation via induction of apoptosis as detected by FACS analysis with annexin V and 7-aminoactinomycin D staining (Fig. 2a,b). Notch signalling induces the promoter activity of Nestin, a well-characterized marker of stemness features 26 . We analysed the effect of TUG1 depletion in GSC lines that stably expressed enhanced GFP (EGFP) under the regulation of the Nestin promoter (GSC-pE-Nes; 5,27 ). In GSC-pE-Nes neurospheres, B80% of cells were positive for Nestin-EGFP. The majority of Nestin-EGFP-expressing cells were positive for Notch1, whereas Nestin-EGFP-negative cells were positive for JAG1 and lacked Notch1 expression ( Supplementary Fig. 4a). Although the viability of both EGFP-positive and EGFP-negative cells were comparable, Nestin-EGFP-positive cells were highly tumorigenic compared with Nestin-EGFP-negative cells ( Supplementary Fig. 4b,c). TUG1 depletion in GSC lines reduced Nestin promoter activity together with impaired neurosphere formation in GSC-pE-Nes (Fig. 2c,d). These morphological and functional changes were also observed after Notch1 and JAG1 inhibition ( Supplementary Fig. 4d,e). TUG1 depletion also effectively decreased the expression of stemness-associated genes (SOX2, MYC, Nestin and CD15) in two GSC lines (Fig. 2e,f). Taken together, it is evident that Notchregulated TUG1 affects the stemness features of GSCs.
TUG1 antagonizes miR-145 and regulates SOX2 and MYC. Some lncRNAs are known to act as miRNA sponges in the ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms13616 cytoplasm, where these lncRNAs bind to miRNAs and quench their activity 12 . As shown in previous reports, RNA-fluorescence in situ hybridization (FISH) analysis revealed that TUG1 localized to both the nucleus and the cytoplasm in GSCs ( Supplementary  Fig. 5a,b) 13 . We examined miRNA expression changes upon TUG1 inhibition in two GSC lines using a miRNA-based expression microarray approach. Twenty-one and 144 miRNAs were commonly upregulated and downregulated with a greater than twofold difference upon TUG1 inhibition, respectively (Supplementary Data 5). Among the upregulated miRNAs, miR-145 was most prominently increased by TUG1 inhibition (Fig. 3a-d). Intriguingly, the interaction between TUG1 and miR-
Using the TargetScan prediction algorithm (http://www. targetscan.org) 29 , we found predicted binding sites for miR-145 not only in TUG1 but also in SOX2 and MYC mRNAs, which encode for well-known stemness-associated transcription factors 30 . Ectopic expression of miR-145 impaired GSC growth and stemness properties ( Supplementary Fig. 5e-h and 6a). This phenotypic change was analogous to that observed after TUG1 inhibition in GSCs (Fig. 2a,b). To further analyse the possible role of TUG1 as a miRNA sponge, we used partial TUG1 transcripts (1-2,132 nucleotides, corresponding to TUG1 exon 1), which contained the predicted binding sites for miR-145 (Fig. 3e). Ectopic expression of these partial TUG1 transcripts efficiently  impaired the repressive effect of miR-145 on SOX2 and MYC expression, whereas mutated TUG1 or seed sequence-deleted TUG1 did not (Fig. 3e,f and Supplementary Fig. 6b). In addition, disruption of neurosphere formation and inactivation of the Nestin promoter by miR-145 were rescued by ectopic TUG1 expression (Fig. 3g,h). These data indicate that TUG1 protects the transcripts of stemness-related genes from miR-145-mediated degradation and aids in maintaining stemness properties.  TUG1-PRC2 complex suppresses neuronal differentiation genes.
Studies have demonstrated that TUG1 in the nucleus interacts with PRC components such as Pc2 and EZH2 via its exon 2 and represses target gene expression in trans 13,14 . To clarify the functional roles of nuclear-localized TUG1 in GSCs, we first transfected TUG1 RNA labelled with 5 0 -bromo-uridine (BrU) (2,133-2,910 nucleotides, corresponding to the TUG1 exon 2) into GSCs. Immunoprecipitation analysis using an anti-BrU antibody revealed that TUG1 bound to PRC2 components (EZH2, SUZ12) and the YY1 transcription factor (Fig. 4a). Intriguingly, interaction between PRC2 and YY1 was disrupted by use of siRNA against TUG1 or RNase treatment, suggesting that TUG1 functions as a scaffold molecule between PRC2 and YY1 (Fig. 4b,c). We further determined the region within TUG1 that interacts with YY1 and EZH2 using a series of TUG1 deletion mutants. EZH2 and YY1 bound TUG1 at the regions of 2,316-2,555 and 2,746-2,910 nucleotides, respectively ( Fig. 5a,b). Intriguingly, these two regions are well conserved  between human and mouse (86.7% and 82.4% homologies, respectively). Next, we performed a modified RNA pull-down assay coupled with promoter-microarray analysis using BrU-labelled TUG1-transfected cells. From this assessment, 3,183 and 2,333 genes were enriched with ectopic TUG1 in 1228-and 222-GSCs, respectively, and 630 genes were commonly identified as targets of TUG1 (Supplementary Data 6). Comparison between the TUG1 target genes (630 genes) and predicted YY1 target genes (1,714 genes) determined by Whitfield's criterion 31 revealed that 525 (83.3%) contained an YY1 motif (Fig. 4d). Gene expression microarray analysis revealed that the majority of those target genes (345/525 genes, 65.7%) were upregulated upon TUG1 inhibition (Fig. 4e,f, P ¼ 7.553 Â 10 -6 ). Those TUG1 target genes contained important regulators of neuronal differentiation, such as BDNF, NGF and NTF3. These target genes were silenced and modified via H3K27me3. Inhibition of TUG1 and YY1 derepressed their gene expression along with a reduction in the K27me3 level (Fig. 4g-i). This direct binding of endogenous TUG1 in the BDNF, NGF and NTF3 promoters was validated via chromatin isolation by RNA purification (ChIRP) analysis (Fig. 4j).
These findings were further supported by rescue experiments using ectopic expression of TUG1 and its deletion mutants (Fig. 5a). Notably, endogenous TUG1 was repressed and TUG1 target genes (BDNF, NGF and NTF3) were highly expressed in serum-differentiated GSC compared with GSCs (Fig. 5c,d). Transfection of TUG1 deletion fragments into serum-differentiated GSC revealed that exon 2 transcript (2,133-2,910 nucleotides) and the TUG1 fragments with EZH2and YY1-binding domains (2,316-2,910 nucleotides) could efficiently repress the expression of TUG1 target genes, whereas other deletion fragments could not (Fig. 5d). Taken together, these data suggest that TUG1 physically interacted with PRC2 and promoted locus-specific K27me3 via YY1 binding activity, which resulted in suppression of neuronal differentiation-associated genes.
Pivotal roles of the TUG1 exon 1 region in maintenance of GSC. To examine what are the relative and contextual contributions of miR-145 and PRC2 to the TUG1 effects on stemness features, we overexpressed each exon of TUG1 in GSCs treated with gamma-secretase inhibitor (RO4929097) (Fig. 5a, Fig. 6a-c). Interestingly, the TUG1 exon 1 transcript (1-2,132 nucleotides) rescued cell growth defects, apoptosis and the expression of stemness-associated genes, whereas neither the TUG1 exon 2 transcript (2,133-2,910 nucleotides) nor exon 3 transcript (2,911-7,115 nucleotides) rescued those features. In addition, TUG1 exon 1 transcript efficiently rescued Nestin promoter activity together with impaired neurosphere formation in GSC-pE-Nes (Fig. 6d,e). Consistently, the effects of Notch1 and JAG1 inhibition by siRNA on the stemness features of GSCs were rescued by ectopic TUG1 exon 1 expression, but were not rescued by either exon 2 or exon 3 ( Supplementary Fig. 7). Taken together, these data indicate the pivotal roles of the TUG1 exon 1 region in maintenance of stemness features in GSCs. self-renewal, as shown in previous studies 25 . In addition, B70% of Notch1-positive GBM cells showed high expression of TUG1 in the perivascular regions, whereas in areas distant from tumour vessels, the majority of GBM cells expressed neither Notch1 nor TUG1 (Fig. 7b-e). The positive correlation between TUG1 and Notch1 expression was further validated in an independent sample set using The Cancer Genome Atlas (TCGA) (Supplementary Fig. 8a).
The anti tumour effect and inhibitory activity against TUG1 expression mediated by TUG1-DDS were even stronger than those observed with a g-secretase inhibitor (RO4929097, oral administration) in our model (Fig. 7g,h, P ¼ 9.237 Â 10 À 3 ). RNA-FISH analysis showed that neither TUG1 nor Notch1positive cells were observed in residual tumours in mice with TUG1-DDS treatment, whereas in mice with CTRL-DDS   treatment, TUG1 and Notch1 expression was predominantly observed in the perivascular regions (Fig. 8a). Consistently, Nestin-EGFP and CD15-positive cells were diminished in TUG1-DDS-treated xenografts, suggesting that inhibition of TUG1 effectively eliminated the GSC population in vivo (Fig. 8b). Furthermore, in TUG1-DDS-treated xenografts, miR-145 was significantly increased and MYC and SOX2 were decreased (Fig. 8c, P ¼ 2.895 Â 10 À 3 ), as was found in GSCs in vitro following siRNA treatment against TUG1 (Fig. 3c,d).
Neuronal differentiation-associated genes (BDNF, NGF and NTF3) were reactivated in response to TUG1-DDS treatment (Fig. 8c,d). Taken together, these data indicate that both nuclear and cytoplasmic TUG1 cooperatively promote the stemness and tumorigenicity of GSCs in vivo.

Discussion
An increasing number of studies have focused on the impact of Notch signalling in GSCs, which may have plasticity and respond to signals from their microenvironment 8,32 . Because GSCs contribute to the growth, survival, invasion and recurrence of brain tumours, understanding the mechanisms that govern the specific regulation of a certain set of genes, which contribute to stemness properties, by the Notch signal pathway has been actively investigated 33 . In the current study, we identified TUG1, which was found via an unbiased approach to be commonly upregulated in two independent GSC populations and downregulated following inhibition of Notch signalling. TUG1 promotes self-renewal by sponging miR-145 in the cytoplasm and recruiting polycomb via YY1-binding activity to repress differentiation genes in the nucleus. We also found that MALAT1 and NEAT1 were highly expressed in GSCs. Although these are well known important functional lncRNAs associated with glioma behaviour 14,34,35 , they did not appear to be involved in the Notch-induced GSC maintenance system ( Supplementary Fig. 4f-h). Some lncRNAs contain miRNA-binding sites and may act as miRNA sponges, such as the abundant, cytoplasmic ncRNA 36 . TUG1 modulates miR-145 levels through its function as a miRNA sponge to trap miR-145, which is known to regulate the expression of core stemness-associated factors 37 . The miRNA sponge function of TUG1 is supported by three lines of evidence. First, we found that cytoplasmic TUG1 expression is extremely abundant compared with miR-145 in GSCs, whereas in TUG1 depleted GSCs, the level of miR-145 was found to be significantly increased. Second, downregulation of the stemness-associated factors by miR-145 was rescued by exogenous TUG1 expression, whereas TUG1 mutants that lacked specific miR-145 seed sequences could not. Third, both TUG1 and miR-145 were bound to wild-type AGO2 but not bound to a mutant form of AGO2 without the PAZ domain 28 . This interaction between TUG1 and miR-145 is consistent with a previous elegant study showing that abundant and cytoplasmic linc-RoR functions as an endogenous miR-145 sponge, which negatively regulates mature miRNA expression levels through a posttranscriptional mechanism and avoids miR-145 increases in self-renewing hESCs 30 .
Notably, TUG1 could prevent MYC mRNA from degradation by quenching miR-145. As MYC is an important downstream factor in Notch signalling 38 , as well as a transcriptional activator of TUG1 ( Supplementary Fig. 10), TUG1 and MYC may form a reinforcing activation loop leading to mutual stabilisation of their gene expression. In any case, the Jagged1-Notch1-TUG1 pathway can effectively enhance GSC self-renewal by antagonising miR-145 activity. Intriguingly, in addition to targeting stemness-associated factors, miR-145 is also known to act as a tumour suppressor by decreasing cell growth, apoptosis and angiogenesis in various cell types including those derived from GBM 16,39 . These tumour-suppressive functions of miR-145 may also explain why inhibition of TUG1 efficiently induced apoptosis after partial neuronal differentiation.
In the nucleus, TUG1 selectively regulated the epigenetic status of neuronal differentiation-associated genes, such as BDNF, NGF and NTF3. Notably, BDNF acts as a tumour suppressor in gliomagenesis 40 . Theoretically, RNA such as TUG1 may freely diffuse and transmigrate between chromosomes, but some DNA elements may be required for efficient loading onto certain genomic loci 41 . In this study, we illustrated that YY1 functional bridges facilitated the loading of regulatory TUG1 to its DNA targets. It is possible that YY1 functional bridges alone may not be sufficient to specify the target loci, as TUG1 was not simultaneously enriched at a large number of other YY1 sites.
However, we found that TUG1 fragments with EZH2-and YY1-binding domains could efficiently repress the expression of the PRC2-TUG1-YY1 target genes, whereas other deletion fragments, which lack either EZH2 or YY1 binding domains (dominant-negative constructs), could not. Therefore, TUG1 in the nucleus appeared to have interacted with PRC2 and promoted locus-specific methylation of histone H3K27, at least in part, via YY1-binding activity. Of interest, both the regions where PRC2 and YY1 interact with TUG1 are evolutionarily conserved between mouse and human, suggesting the importance of these regions 42 . Consistent with our findings, the concept of YY1 as docking protein has also been illustrated in an X chromosome inactivation model, in which YY1 tethers Xist RNA-PRC2 to the inactive X nucleation center 41 .
These regional distinct functions of TUG1 may raise the question of what are the relative and contextual contributions of each region via miR-145 and PRC2 on stemness features. Importantly, attenuation of the Notch-TUG1 axis by siRNA or g-secretase in regulating self-renewal of glioma cells was efficiently rescued by TUG1 exon 1 transcript where the miR-145-binding site locates indicating that there is a pivotal role for the exon 1 and miR-145 interaction in maintenance of stemness properties.
Many strategies to improve the delivery of drugs to GBMs are under investigation, as the blood-brain barrier and the blood-brain tumour barrier limit the enhanced permeability and retention effect, thereby resulting in poor efficacy due to inefficient drug delivery to GBM 43 . Brain tumour vessels show overexpression of receptors that mediate ligand-dependent drug delivery 44 . In particular, RGD peptides are promising ligand molecules for targeting a v b3 and a v b5 integrins, which are frequently overexpressed in GBM cells 45 . In order to inhibit TUG1 expression in a mouse xenograft model, we used cRGD ligand-conjugated polymeric micelles for delivery. These targetable polymeric micelles retained ASO accumulation within tumours, which is associated with transcytosis-mediated penetration and enhanced gene silencing activity. Although further investigations are required, cRGD-mediated drug delivery is a powerful strategy for targeting GBMs through facilitated ASO delivery beyond the blood-brain tumour barrier.
Emerging insights into gliomagenesis have revealed that GSCs have the potential to initiate and maintain the growth of gliomas, and may be crucial for their resistance to conventional therapies 1,46 . Further, the existence of GSCs and plastic epigenetic regulation may be linked to the morphological and lineage heterogeneity that is observed in GBM 2,5 . In the current study, we addressed the novel mechanism underlying the contribution of the Notch signalling pathway to the stemness features of GSCs in response to contact-dependent signals from adjacent cells expressing Jagged1. Further, we provide a new paradigm whereby targeting TUG1, especially coupled with a potent DDS, is an effective novel strategy for GBM treatment. One limitation of the current study is that both our GSC lines were classified as proneural type GBM by gene expression analysis (Supplementary Fig. 8b). Indeed, expression levels of TUG1 and Notch1 were prominently upregulated in proneural and classical GBM compared with two other subtypes (neural and mesenchymal) ( Supplementary  Fig. 8c). Therefore, it might be possible that targeting TUG1 is more promising for treatment of proneural or classical GBMs rather than other subtypes. However, given that inhibition of g-secretase-mediated Notch cleavage is a primary focus for the development of targeted therapeutics in GBM (including application in clinical trials such as RO4929097) 47 , our data may provide a strong rationale that targeting TUG1 is a more specific and potent therapeutic approach to eliminate the GSC population.
Methods GBM tissues, cell culture and drug treatment. GBM tissue samples were obtained from patients undergoing surgical treatment at Nagoya University hospital, Japan, after they provided written informed consent. Total RNA of normal brain was purchased from BioChain (Hayward, CA, USA). The procedures used for derivation of GSCs (1228-GSC and 222-GSC) were as described previously 5 . In brief, dissociated tumour cells were cultured in Neurobasal Medium (Life Technologies, Carlsbad, CA) containing N2 and B27 supplements (Life Technologies), along with human recombinant basic fibroblast growth factor and epidermal growth factor (20 ng/ml each; R&D Systems, Minneapolis, MN, USA). Serum-differentiated GSCs were established by culturing GSCs in DMEM (Life Technologies) containing 10% fetal bovine serum 5 . Serially transplanted GSCs were generated according to previously published methods 1 . GSCs were routinely tested for mycoplasma contamination and used in the experiments. For blocking of Notch signalling in GSCs, the g-secretase inhibitor DAPT (N-[N-(3,5-difluorophenacetyl)l-alanyl]-S-phenyl glycine t-butyl ester), Sigma-Aldrich, St Louis, MO, USA) or RO4929097 (Selleck Chemicals, Houston, TX, USA) were added, and the medium was changed every other day. For in vivo studies, RO4929097 was formulated as a suspension in 1% carboxymethyl cellulose with 0.2% Tween 80 for oral administration.
RNA extraction and preparation for RNA sequencing. Procedures utilized for next-generation sequencing were as described previously 48 . In brief, total RNA was extracted from GSCs using TRIzol (Life Technologies). The RNA-seq libraries were prepared using a paired-end RNA Sequencing Sample Prep Kit (Illumina, San Diego, CA, USA). Two microgram of total RNA was used as the starting material, and polyadenylated RNA was selected using Sera-Mag Magnetic Oligo (dT) Beads (Illumina). First-strand cDNA synthesis was performed using SuperScript II (Life Technologies), and second-strand cDNA synthesis was performed using DNA Pol I in the supplied GEX second-strand reaction buffer (Illumina). Paired-end adaptors were ligated to the cDNA fragments. PCR (eight cycles) was performed with Phusion High-Fidelity DNA Polymerase (Finnzymes Oy, Espoo, Finland). The PCR products were cleaned up with Agencourt AMPure XP magnetic beads (Beckman Coulter, Miami, FL, USA). cDNA was applied to the flow cell and paired-end 76 nucleotide-long reads were generated using Illumina GAIIx and Miseq instruments. We performed one experiment for the RNA-seq analysis in each cell line. Obtained data were aligned with TopHat 2.0.13 to hg19 with default parameters. Expression differences between two GSC lines were calculated with Cuffdiff 2.2.1.
lncRNA expression analysis via microarray technology. Total RNA was isolated from GSCs with TRIzol reagent (Life Technologies). RNA was amplified into cRNA and labelled according to the Agilent One-Colour Microarray-Based Gene Expression Analysis protocol (Agilent Technologies, Santa Clara, CA). Labelled samples were purified with the RNeasy Kit (Qiagen, Valencia, CA) and hybridized to SurePrint G3 Human GE 8 Â 60 K array slides (G4851B, Agilent Technologies) at 65°C with rotation at 10 rpm for 17 h. The arrays were scanned using an Agilent Microarray Scanner (G2565BA, Agilent Technologies). The scanned images were analysed using the Feature Extraction software, version 10.7.3.1 (Agilent Technologies) with background correction. Data analysis was performed with GeneSpring GX, version 12.6.0 (Silicon Genetics). Expression data were centred on a median with the use of the GeneSpring normalisation option with no substantial difference in results. The microarray analysis was performed in duplicate for each cell line. Statistical comparisons were made with the use of GeneSpring volcano plot filtering. Quantitative reverse transcription-PCR (qRT-PCR). Total RNA was isolated using TRIzol (Life Technologies), and 1 mg was used for reverse transcription with the SuperScript VILO cDNA Synthesis Kit (Life Technologies). Expression levels of target genes were determined using the delta Ct method, and normalized to the housekeeping gene GAPDH. Target genes were measured using at least three replicates by TaqMan PCR or SYBR Green quantitative PCR (qPCR). TaqMan PCR assays (Applied Biosystems) and oligonucleotide primers are shown in Supplementary Data 4.
Chromatin immunoprecipitation (ChIP) assay. ChIP assays were performed based on a modification of previously published methods 5 . In brief, GSCs were treated with 0.5% formaldehyde to cross-link histones to DNA followed by the addition of 125 mM glycine to stop the cross-linking reaction. After sonication of cell pellets, the lysate was incubated with 4 ul of anti-K27 trimethylated histone H3 (ref. 49) and 10 ml of either anti-Notch1 (ab27526, Abcam), anti-K27 trimethylated histone H3 (ab6002, Abcam), anti-YY1 (ab12132, Abcam) and anti-MYC (ab32, Abcam) antibodies. ChIP products were analysed by SYBR Green ChIP-qPCR using the primer sets shown in Supplementary Data 4.
Flow cytometry. Cells were analysed by fluorescence activated cell sorting using the PE Annexin V Apoptosis Detection Kit I (BD Bioscience, San Jose, CA, USA). In brief, single-cell suspensions were prepared from GSC spheres, and stained with PE Annexin V and 7-aminoactinomycin D. Data were collected on a FACS Calibur (BD Bioscience) and analysed using CELL Quest Pro (BD Bioscience).
Exogenous expression of TUG1. In vitro transcription of TUG1 RNA was performed with DNA templates that correspond to the different regions of the TUG1 RNA sequence (1-2,132 corresponding to exon 1 full-length, 2,133-2,910 corresponding to exon 2 full-length, 2,316-2,910, 2,556-2,910, 2,746-2,910, 2,133-2,745, 2,133-2,555 and 2,133-2,315 nucleotides) using the CUGA7 in vitro Transcription Kit (Nippon Gene, Tokyo, Japan). The DNA templates were amplified by PCR with the primer sets listed in Supplementary Data 4. To append the 5 0 -cap structure and 3 0 -polyA tail, the DNA template was ligated into the pTnT Vector (Promega, Madison, WI, USA). The seed sequence of miR-145 (5 0 -AAC TGGA-3 0 ) within TUG1 RNA exon 1 (1-2,132 nucleotides) was mutated or deleted using the QuikChange Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA, USA) with the primer sets listed in Supplementary Data 4. In addition, PCR-generated fragments containing cDNA for NICD (not including the si-Notch1 target region), TUG1 exon 1 (1-2,132 nucleotides), TUG1 exon 2 (2,133-2,910 nucleotides), TUG1 exon 3 (2,911-7,115 nucleotides) and luciferase were also ligated into a pcDNA3.1 vector (Life Technologies), which were subsequently transfected into the glioma cells in order to overexpress those TUG1 transcripts. Primer sequences for plasmid construction are summarized in Supplementary Data 4. Obtained plasmids were verified by conventional sequencing analysis. miRNA microarray. The SurePrint G3 Human miRNA 8x60K Microarray (G4872A, Agilent Technologies) was used according to the manufacturer's protocols. Total RNA was isolated from GSCs with TRIzol reagent (Life Technologies). In total, 100 ng of total RNA was labelled with pCp-Cy3 (Agilent Technologies) and 15 units of T4 RNA ligase (GE Healthcare, Little Chalfont, Buckinghamshire, UK) at 16°C for 2 h. Labelled samples were purified with MicroBio-Spin six columns (Bio-Rad, Richmond, CA, USA) and hybridized to microarrays at 55°C with rotation at 20 rpm for 20 h. The arrays were scanned using an Agilent Microarray Scanner (G2565BA, Agilent Technologies). The scanned images were analysed using the Feature Extraction software, version 10.7.3.1 (Agilent Technologies) with background correction. Data analysis was performed with GeneSpring GX, version 12.6.0 (Silicon Genetics). Expression data were normalized to the 75th percentile using the GeneSpring normalisation option with no substantial difference in results. The miRNA microarray analysis was performed in duplicate for each cell line. Statistical comparisons were made with the use of both the GeneSpring analysis-of-variance tool and volcano plot filtering.
RNA-FISH analysis. RNA was visualized in paraffin-embedded sections using the QuantiGene ViewRNA ISH Tissue Assay Kit (Affymetrix, Frederick, MD, USA). In brief, tissue sections were rehydrated and incubated with proteinase K. Subsequently, they were incubated with ViewRNA probe sets designed against human TUG1, miR-145 and Notch1 (Affymetrix). Analysis of RNA-binding proteins by RNA pull-down assay. Nuclear extracts were obtained from GSCs (1 Â 10 8 ) and incubated with 50 pmol of BrU-labelled TUG1 RNA. RNA and nuclear extract conjugants were mixed with an anti-BrU antibody (MI-11-3, MBL International) for 2 h at 4°C. To collect the immunoprecipitated RNA-binding proteins, Dynabeads Protein G (Life technologies) were added and incubated for 1 h at 4°C. These TUG1 RNA-binding proteins were analysed by western blot analysis with anti-EZH2 antibody (#3147, Cell Signaling Technology), anti-YY1 antibody (ab12132, Abcam, Cambridge, UK), and anti-IgG antibody as a negative control (PM035, MBL International). HRP-linked anti-mouse IgG (#7076, Cell Signaling Technology) and HRP-linked anti-rabbit IgG (#7074, Cell Signaling Technology) antibodies were used as secondary antibodies.

Induction
Analysis of TUG1 RNA-binding loci by ChIP assay in vivo. Ten nM BrU-labelled TUG1 RNA (2,133-2,910 nucleotides) was transfected into GSCs using Lipofectamine 2000 (Life Technologies). After 72 h of transfection, GSCs were cross-linked with 0.5% formaldehyde, and chromatin fractions were extracted and homogenized. Chromatin and BrU-labelled TUG1 RNA mixtures were immunoprecipitated using an anti-BrU antibody and Dynabeads Protein G. Purified chromatin was eluted to yield DNA, which was then subjected to microarray analysis or ChIP-qPCR analysis.
Equivalent amounts of DNA, which was extracted after immunoprecipitation for BrU-labelled TUG1 RNA, and total input DNA were amplified in parallel using a random primer method with the GenomePlex Complete Whole Genome Amplification Kit (15 cycles), according to the manufacturer's instructions (Sigma-Aldrich). Amplified products were labelled with Cy5 for immunoprecipitated DNA samples and Cy3 for input DNA using the BioPrime Array CGH Genomic Labeling System (Life Technologies). A total of 4 mg of labelled DNA was then hybridized to the human promoter ChIP-chip microarray (G4874A, Agilent Technologies) for 48 h at 65°C. Data were extracted from scanned images using the Feature Extraction software, version 10.7.3.1 (Agilent Technologies). The text files were then imported into Genomic Workbench, standard edition 5.0.14 (Agilent Technologies), for analysis.
ChIRP assay. ChIRP assays were performed as described 50 . GSCs were cross-linked with 1% glutaldehyde for 15 min at room temperature and then quenched with 125 mM glycine for 5 min. The cross-linked chromatin was then isolated and hybridized with the probes, followed by streptavidin bead-based capturing and wash/elution steps. The antisense DNA probes targeting TUG1 and EGFP-ORF sequences were designed using ChIRP Probe Designer version 4.2 (https://www.biosearchtech.com/chirp-designer) and divided into odd and even pools. Probes used for ChIRP assays are listed in Supplementary Data 4. The ChIRP captured chromatin was then reverse cross-linked and analysed by qPCR using the primers listed in Supplementary Data 4.
Animal experiments. All experiments were performed under protocols approved by the Institutional Animal Care and Use Committee of Nagoya City University Graduate School of Medical Sciences. 222-GSC-pE-Nes were injected intracranially into 6-week-old female NOD/SCID mice (n ¼ 12, SLC, Shizuoka, Japan). Four weeks after the injection of GSCs, CTRL-DDS (n ¼ 4) or TUG1-DDS (n ¼ 4; 1 mg/kg per day) were intravenously injected twice a week for 4 weeks. RO4929097 (n ¼ 4; 20 mg/kg per day) was administered orally five times a week for 4 weeks. The accumulation of ASOs in tumour tissue was confirmed by an in vivo spectral imaging system (IVIS Lumina II, Xenogen, Alameda, CA, USA). Tumour volumes were calculated using the formula (width, height and length; W Â H Â L)/2.

Statistics.
For the statistical analysis of clinical data sets from TCGA, we used R Statistical Software (version 2.14.0; R Foundation for Statistical Computing, Vienna, Austria). The statistical significance of the differences between two groups was analysed by paired Student's t-test. Kruskal-Wallis analysis were used for to evaluate the extent of differences among more than three groups (StatView software version 5.0; Abacus Concepts, Berkeley, CA, USA). Microarray data analysis were performed by paired Student's t-test. All reported P values were two-sided, with Po0.01 considered statistically significant.