Oncogene (2012) 31, 3235–3243 & 2012 Macmillan Publishers Limited All rights reserved 0950-9232/12 www.nature.com/onc ORIGINAL ARTICLE RNAi screening in glioma stem-like cells identifies PFKFB4 as a key molecule important for cancer cell survival

V Goidts1,2, J Bageritz1, L Puccio1, S Nakata1, M Zapatka1, S Barbus1, G Toedt1, B Campos3, A Korshunov4,5, S Momma6, E Van Schaftingen7, G Reifenberger8, C Herold-Mende3, P Lichter1 and B Radlwimmer1

1Division of Molecular Genetics, German Cancer Research Center, Heidelberg, Germany; 2Department of Pediatric Oncology, Hematology and Immunology, Children’s Hospital, University of Heidelberg, Heidelberg, Germany; 3Division of Neurosurgical Research, Department of Neurosurgery, University of Heidelberg, Heidelberg, Germany; 4Department of Neuropathology, University of Heidelberg, Heidelberg, Germany; 5Clinical Cooperation Unit Neuropathology, German Cancer Research Center, Heidelberg, Germany; 6Edinger Institute, Frankfurt University Medical School, Frankfurt, Germany; 7Laboratory of Physiological Chemistry, de Duve Institute, Universite´ Catholique de Louvain, Brussels, Belgium and 8Department of Neuropathology, Heinrich-Heine-University, Du¨sseldorf, Germany

The concept of cancer stem-like cells (CSCs) has gained Introduction considerable attention in various solid tumors including glioblastoma, the most common primary brain tumor. Glioblastoma is the most common and most malignant This sub-population of tumor cells has been intensively primary brain tumor in adults. Despite multimodal investigated and their role in therapy resistance as well as aggressive treatment, comprising surgical resection, tumor recurrence has been demonstrated. In that respect, local radiation therapy and systemic chemotherapy, development of therapeutic strategies that target CSCs the average patient survival time is still in the range of 1 (and possibly also the tumor bulk) appears a promising year after diagnosis (Stupp et al., 2007). Several recent approach in patients suffering from primary brain tumors. studies suggested a hierarchical cellular organization in In the present study, we utilized RNA interference (RNAi) glioblastomas, which consist of heterogeneous cell to screen the complete human kinome and phosphatome populations that differ in their tumorigenic potential. (682 and 180 targets, respectively) in order to identify Indeed, it is believed that gliomas are initiated and and pathways relevant for the survival of brain maintained by a small sub-population of cells that is CSCs and thereby potential therapeutical targets for capable of extensive self-renewal and recapitulation of glioblastoma. We report of 46 putative candidates the original tumor in nonobese diabetic/severe com- including known survival-related kinases and phospha- bined immunodeficient mice (Singh et al., 2003; Lee tases. Interestingly, a number of genes identified are et al., 2006; Vescovi et al., 2006). These so-called brain involved in , especially , such as cancer stem-like cells (CSCs) have been implicated not PDK1 and PKM2 and, most prominently PFKFB4. only in tumor initiation, but also in therapy resistance In vitro studies confirmed an essential role of PFKFB4 and tumor recurrence. It was shown that brain CSCs are in the maintenance of brain CSCs. Furthermore, high enriched in recurrent malignant gliomas and are PFKFB4 expression was associated with shorter survival resistant to chemo- and radiotherapy (Bao et al., 2006; of primary glioblastoma patients. Our findings support the Eramo et al., 2006; Liu et al., 2006; Al-Hajj, 2007). importance of the glycolytic pathway in the maintenance Given their central role, it has appeared that of malignant glioma cells and brain CSCs and imply therapeutically targeting this particular population of tumor metabolism as a promising therapeutic target in cells is crucial for successful glioblastoma treatment glioblastoma. (Reya et al., 2001; Pardal et al., 2003; Park et al., 2009). Oncogene (2012) 31, 3235–3243; doi:10.1038/onc.2011.490; One possible direction is to uncover the mechanisms, by published online 7 November 2011 which brain CSCs escape apoptosis, and ultimately provide knowledge for the development of drug Keywords: loss-of-function screen; apoptosis; glioblas- therapeutics targeting the involved genes and pathways. toma; cancer stem-like cells; glycolysis Kinases and phosphatases control the reversible process of phosphorylation, which is involved in each intracellular pathway, including cell survival signalling. Therefore, identification of the kinases and phospha- tases whose inhibition induces the death of brain CSCs Correspondence: Dr V Goidts, Division of Molecular Genetics, may pave the way toward novel therapeutic targets. In German Cancer Research Center, Im Neuenheimer Feld, 280, 69120 this study, we performed a loss-of-function screen using Heidelberg, Germany. E-mail: [email protected] a lentiviral short-hairpin RNA (shRNA) library repre- Received 17 November 2010; revised 12 September 2011; accepted 15 senting the entire human kinome and phosphatome, and September 2011; published online 7 November 2011 identified 46 candidates that are essential for the survival Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3236 2.5 of brain CSCs in vitro. These candidates included key BMPR2 NTRK1 PAK2 PRKCA regulators of the glycolytic pathway, underlining the 2.0 central role of metabolism for the survival of brain CSCs. One of the glycolysis-associated candidate genes, 1.5 PFKFB4, was functionally characterized in brain CSCs 1.0 in vitro, and the importance of lactate and adenosine 0.5 triphosphate (ATP) production for the survival of brain

(relative to scrambled shRNA) 0.0 CSCs was assessed. Fold change of PI-positive cells Kinases and phosphatases Figure 1 Identification of survival kinases and phosphatases. NCH421k cells were seeded in 96-well plates and transduced with Results shRNAs directed against all known and putative kinases and phosphatases in the . Cells were incubated for 6 days to allow target knockdown and cell death was measured Kinome and phosphatome RNA interference (RNAi) by PI staining and analyzed by FACS. The graph depicts the screen to induce brain CSC death median fold change of cell death for all kinases and phosphatases Lentiviral short-hairpin libraries were used recently to normalized to control shRNA. The genes whose loss of function conduct loss-of-function screens in mammalian cell lines increased cell death by 41.5-fold were considered as positive hits. The red line represents the median of all shRNAs plus 2.5-fold and turned out to be particularly useful in slowly median absolute deviation (MAD). dividing cells (Wurdak et al., 2010). We performed a large-scale silencing screen in vitro to identify kinases and phosphatases that are essential for the survival of Table 1 Putative survival kinases identified in the RNA interference brain CSCs. We transduced NCH421k, whose stem-like (RNAi) screen cell properties have been previously characterized (Campos et al., 2010), with a lentiviral vector-based Symbol Accession number Cell death fold change shRNA library that targets each of the 682 and 180 PAK2 NM_002577 2.08 known and putative kinases and phosphatases, respec- PRKCA NM_002737 2.05 tively (average of 4.5 shRNAs per target, The RNAi NTRK1 NM_002529 2.05 Consortium, Broad Institute from MIT and Harvard, BMPR2 NM_001204 2.02 MGC75495 XM_292160 1.85 MA). Viral stocks were produced in a 96-well plate TTBK2 NM_173500 1.85 format and used at a sufficient titer to transduce brain PRKCI NM_002740 1.85 CSCs. PFKFB4 NM_004567 1.82 In order to assess the efficiency of the viral transduc- ETNK2 NM_018208 1.79 tion, cells transduced with lentiviral particles produced EPHA5 NM_004439 1.77 RIPK1 NM_003804 1.73 from a control plasmid encoding green fluorescent SBK1 XM_370948 1.72 protein (GFP) were included in each plate. Viral PKM2 NM_182471 1.72 transduction was efficient as indicated by 70–90% ITK NM_005546 1.72 GFP-expressing cells. As positive control for the cell MAPKAPK5 NM_003668 1.69 PRKX NM_005044 1.68 death phenotype, brain CSCs were transduced with a NME3 NM_002513 1.68 specific shRNA targeting BCL2. After 6 days, cell death PIP5K1A NM_003557 1.68 was measured by propidium iodide staining and FLT1 NM_002019 1.65 fluorescence-activated cell sorting (FACS). In the BRD2 NM_005104 1.65 screen, we considered only those genes as positive hits CABC1 NM_020247 1.64 FN3KRP NM_024619 1.63 whose median over all shRNAs targeting this LCK NM_005356 1.62 showed an increased level of cell death of X1.5-fold MAP4K4 NM_145687 1.61 (median of all shRNAs plus 2.5-fold median absolute RIOK2 NM_018343 1.59 deviation) relative to that of the control (scrambled) ITPKC NM_025194 1.57 SPHK1 NM_182965 1.57 shRNA. Overall, 39 kinases and 7 phosphatases fulfilled MAP3K1 XM_042066 1.56 this criterion (Figure 1 and Tables 1 and 2) and were PDK1 NM_002610 1.55 defined as survival genes. Moreover, each of these genes SPHK2 NM_020126 1.55 was targeted by at least two shRNAs that showed a fold SGK NM_005627 1.54 increase of cell death rate X1.5-fold. STK25 NM_006374 1.54 CHEK2 NM_007194 1.53 ADRBK1 NM_001619 1.52 Validation of survival genes BMPR1B NM_001203 1.51 LIMK2 NM_016733 1.51 Next, we generated a focused library comprising the top STK32C NM_173575 1.51 10 candidates from the first screen, each represented by TP53RK NM_033550 1.50 2 to 3 independent shRNAs that gave the highest rate of TAOK1 NM_020791 1.50 cell death. To ensure that the differences observed were not caused by ‘off-target’ shRNA effects, we verified the level of silencing of each shRNA by quantitative real- independent shRNAs that decreased their target mRNA time reverse transcription–PCR (qRT–PCR). Of these transcript levels by 2- to 10-fold, suggesting that the top 10 genes, 9 were represented by at least two observed phenotype was because of the silencing of the

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3237 Table 2 Putative survival phosphatases identified in the RNA S2) showed an increased fold change of apoptosis in all interference (RNAi) screen cell lines. Symbol Accession number Apoptosis fold change

CDC25C NM_001790 1.86 Gene expression of PFKFB4 is associated with PPM1B NM_002706 1.82 glioblastoma prognosis PPP1R3C NM_005398 1.62 To assess the clinical relevance of the validated candi- PSTPIP1 NM_003978 1.60 dates, we compared their mRNA expression among a DUSP13 NM_016364 1.53 series of astrocytic gliomas of different grades (WHO PSTPIP2 NM_024430 1.50 PTPRM NM_002845 1.50 (World Health Organization) grades II, III and IV, including primary and secondary glioblas- tomas) using previously published microarray data

8 (GSE15698510) (Toedt et al., 2011). Remarkably, the BLV4 mRNA expression of PFKFB4, whose gene product plays 7 NCH421k NCH441 a key role in energy metabolism, significantly differen- 6 NCH644 tiated IDH1 wild-type primary glioblastomas from the

5 secondary glioblastomas as well as diffuse and anaplastic astrocytomas, which mostly carried IDH1 gene mutations 4 (Figure 3a). Moreover, primary glioblastoma patients 3 whose tumors demonstrated PFKFB4 mRNA expression 2 levels above average had a significantly (Po0.0001) Fold change of apoptosis

1 shorter overall survival time than glioblastoma patients whose tumors displayed lower than average PFKFB4 0 expression (Figure 3b). We confirmed the results using the

BCL2 PAK2 NEK5 TTBK2 PRKCI National Cancer Institute’s REMBRANDT (Repository PFKFB4 PPM1B BMPR2 PRKCA CDC25C of Molecular Brain Neoplasia Data) database (http:// Figure 2 Candidate genes identified induce apoptosis upon rembrandt.nci.nih.gov). Kaplan–Meier survival analysis silencing. Verification of candidate gene knockdown-induced apoptosis as determined by caspase assays in three different brain revealed that increased PFKFB4 expression 43-fold CSC lines (NCH421k, NCH441 and NCH644) and in normal compared with normal brain was associated with NSCs (BLV4). Apoptosis was measured 4, 5, 6 and 7 days after significantly reduced overall survival of glioblastoma transduction. The time points that showed the highest caspase patients (P ¼ 0.0201; Supplementary Figure S3A). To test activity are shown. The fold change of apoptosis was normalized to whether the prognostic value of PFKFB4 is independent negative control scrambled shRNA and is depicted as mean of independent biological triplicates for one shRNA (shRNA-1) per of IDH1 mutation status in grade IV gliomas, we targeted gene. BCL2 shRNA was used as positive control. Error calculated Cox hazard models for the overall survival bars represent the s.d. of three independent biological replicates. time considering death of disease only and censoring all Gene knockdown induced a significantly increased rate of other events. A model with PFKFB4 alone with n ¼ 39 apoptosis in brain CSCs when compared with normal NSCs showed a significant effect of PFKFB4 expression on (Po0.05), with the exception of PAK2 shRNA (P ¼ 0.59). Statistical significance was calculated using Student’s t-test for survival (P ¼ 0.022). IDH1 alone showed no significant ranked transformed data. association with survival (n ¼ 68, P ¼ 0.95). In a model (n ¼ 39) stratified for IDH1, PFKFB4 was associated with survival (P ¼ 0.017). Proportional hazard assumption was intended targets (Supplementary Figure S1A). Only fulfilled for all models. silencing of NTRK1 was not confirmed, as its basal level of expression in NCH421k was too low to be measured by qRT–PCR. Thus, its detection in our screen was Overexpression of PFKFB4 isozyme is cancer specific most probably because of an ‘off-target’ effect. PFKFB4 is one of the four genes that encode 6-phos- We performed a validation screen on two additional phofructo-2-kinase/fructose-2,6-biphosphatase (PFK2/ brain CSC cultures (NCH441 and NCH644), as well FBPase-2), and originally was found to be expressed as one normal neural stem cell (NSC) line (BLV4), in the testes (Sakata et al., 1991). It is a bifunctional monitoring the activation of caspases 3 and 7, as they having both kinase and phosphatase functions, are key mediators of apoptosis. Measurements were and plays a key role in energy metabolism by controlling done 4, 5, 6 and 7 days after transduction in three glucose utilization. independent experiments (Figure 2 and Supplementary To determine the cancer specificity of PFKFB4 Figures S1B and S2). The results confirmed that expression, we compared its mRNA expression levels knockdown of each candidate gene resulted in a in the different brain CSC lines with that in a pool of higher fold change of apoptosis when compared with normal brain samples. As shown in Figure 4a, qRT– scrambled shRNA. Moreover, as shown in Figure 2, PCR analysis demonstrated that brain CSCs have a the silencing increased significantly the fold change of PFKFB4 mRNA expression that is B10-fold higher apoptosis of brain CSCs compared with normal NSCs when compared with that of adult normal brain tissue. (Student’s t-test for ranked transformed data; Po0.05). This result was confirmed at the protein level (Figure 4b). However, shRNA-1 targeting PAK2 (Figure 2) and We next asked whether the overexpression of shRNA-2 specific for PRKCA (Supplementary Figure PFKFB4 was related to the expression of PFKFB3,

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3238 a b PFKFB4 ** 1.2 PFKFB3 PFKFB3 PFKFB4 1

15 0.8

0.6 NCH644 Normal brain 0.4 NCH421k NCH441 PFKFB4 0.2 α

10 mRNA expression level -Tubulin 0 (relative to housekeeping genes)

Normal NCH441 NCH644 NCH421k brain 5 Figure 4 Differential expression of PFKFB3 and PFKFB4 in brain CSCs. (a) Analysis of PFKFB3 and PFKFB4 mRNA expression as * determined by qRT–PCR. Total mRNA was extracted from three different brain CSC lines (NCH421k, NCH441 and NCH644). As reference, a pool of normal brain mRNA was included. Results mRNA expression normalized to normal brain 0 were normalized to mRNA levels of four housekeeping genes. (b) PFKFB4 protein expression in three brain CSC cultures. As AII AAIII sGBIV pGBIV reference, a pool of normal brain samples was used. n = 8 n = 13 n = 10 n = 38

p<0.0001 decreased in brain CSCs when compared with a pool of 80 Low expression (n=24) normal brain mRNAs (Figure 4a). It should be noted High expression (n=11) that, based on microarray data (Toedt et al., 2011), no 40 expression was observed for the PFKFB1 isoform and that PFKFB2 was equally expressed in the glioma and Percent survival 0 normal brain samples (Supplementary Figure S3B). 0 200 400 600 800 1000 Days Silencing of PFKFB4 suppresses viability of brain CSCs Figure 3 Expression of PFKFB4 and PFKFB3 in different glioma and inhibits lactate and ATP production entities. (a) PFKFB4 and PFKFB3 mRNA expression levels in the To determine the effect of PFKFB4 knockdown on glioma expression data set GSE15698510. Box plot shows expression levels of PFKFB4 and PFKFB3 in diffuse astrocytomas: overall brain CSC population in vitro, transduction of World Health Organization (WHO) grade II (AII, n ¼ 8), NCH421k with PFKFB4 shRNA was performed in six- anaplastic astrocytomas WHO grade III (AAIII, n ¼ 13), secondary well plates. The knockdown level was monitored by glioblastoma WHO grade IV (sGBIV, n ¼ 10) and primary western blot analysis (Figure 5a) and cell death was glioblastoma WHO grade IV (pGBIV, n ¼ 38) relative to the assessed by propidium iodide (PI) staining and FACS. normal brain control tissue (n ¼ 4). Note that the median PFKFB3 and PFKFB4 mRNA expression levels are significantly higher in shRNA-2 induced 495% cell death, whereas shRNA-1 primary glioblastoma compared with the normal brain tissue showed B50% PI-positive cells (Figures 5b and c). samples and all other tumor groups (*Po0.02 and **Po0.002). Malignant glioma cells are known to be resilient to (b) Kaplan–Meier analysis showing the association of PFKFB4 anaerobic conditions created by a hypoxic microenvir- expression with overall survival of patients with IDH1 wild-type primary glioblastoma. Low- and high-expression groups were onment, thanks to their potential to increase glycolysis separated according to the mean expression value of all samples and thereby production of (Floridi et al., studied. The P-value was calculated applying the Mantel–Haenszel 1989). This metabolic adaptation is mediated by changes test (Po0.0001). Using a bootstrapping approach with 1000 in gene expression patterns triggered by the activation of draws, the P-value for differences in survival based on the hypoxia-inducible factor 1a (HIF-1a)(Luet al., 2002). expression of PFKFB4 was confirmed by reselecting the mean expression value. Previous studies have highlighted the immediate target- ing of PFKFB4 by HIF-1a (Minchenko et al., 2004). Analysis of HIF-1a protein and mRNA expression on another isozyme of PFK2 that is expressed in normal microarrays in brain CSCs and normal NSCs (Ernst brain and overexpressed in numerous tumors (Manzano et al., 2009) revealed a higher expression level in brain et al., 1998; Kessler and Eschrich, 2001; Atsumi et al., CSCs (Supplementary Figure S4A). This observation 2002; Gomez-Brouchet et al., 2007). As depicted in was made regardless of the normoxic culture conditions, Figure 3a, PFKFB3 showed a 1.3-fold higher mean suggesting that glycolysis is increased despite the mRNA expression level in primary glioblastomas presence of oxygen, a phenomenon described as the relative to normal brain tissue and was not over- Warburg effect (Lu et al., 2002). To verify that silencing expressed in the other glioma groups investigated. of PFKFB4 impairs metabolism and thereby induces Moreover, the level of PFKFB3 expression was not apoptosis of brain CSCs, we measured the level of significantly related to overall survival of primary lactate and of ATP at 3 and 4 days after PFKFB4 glioblastoma patients (P ¼ 0.518, data not shown). In knockdown with two independent shRNAs. Lactate contrast to PFKFB4, PFKFB3 mRNA expression was secretion and ATP level was significantly reduced in

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3239 3 days 4 days 5 days NCH421k, confirming that PFKFB4 silencing induces metabolic changes (Figures 6a and b). As shown previously, inhibition of glycolysis induces the activa- tion/phosphorylation of the adenosine monophosphate shRNA-2 Ctl shRNA-1 shRNA-2 shRNA-1 Ctl shRNA-1 shRNA-2 Ctl Ctl (AMP)-activated protein kinase (AMPK), which inhibits PFKFB4 the mTOR (mammalian target of rapamycin) pathway α-Tubulin and therefore sensitizes tumor cells to death receptor- induced apoptosis (Pradelli et al., 2010). This activation is mainly due to a loss of balance between AMP and Scrambled shRNA ATP content because of a drop of the ATP level. As shown by western blot analysis in Figure 6c, silencing of PFKFB4 with shRNA-1 induces the phosphorylation of AMP kinase. This phenomenon could not be demon- Counts strated when using shRNA-2, most likely because of the high apoptosis rate induced by shRNA-2. To determine PI whether the mTOR pathway has a relevant role in brain CSC survival, we treated NCH421k cells with rapamycin PFKFB4 shRNA-1 at different concentrations. Cell viability was assessed by MTS assay 3 days after treatment. The 20–40% decrease of cell viability upon rapamycin treatment suggests a potential role of mTOR in the survival of brain CSCs Counts (Supplementary Figure S4B).

PI PFKFB4 shRNA-2 Discussion

Several reports have provided evidence that glioblasto- mas are cellularly heterogeneous tumors containing the Counts so-called brain CSCs that show stem-like features in vitro and give rise to tumor development when xenotransplanted in immunocompromised mice. PI Furthermore, brain CSCs have been implicated in the resistance of glioblastomas to chemo- and radiotherapy, PI-positive cells PI-negative cells which emphasizes the importance to target these cells to 60000 achieve a more effective therapy (Bao et al., 2006; Eramo et al., 2006; Liu et al., 2006; Al-Hajj, 2007). In 40000 this respect, an appealing treatment approach would be to specifically target proteins encoded by the genes that are essential for the survival of brain CSCs. Events 20000 In this study, we performed a kinome- and phospha- tome-wide loss-of-function screen in order to identify genes whose silencing induces apoptosis in brain CSCs. 0 Our results indicate that B5% of the kinome and Scrambled PFKFB4 PFKFB4 phosphatome is required for the survival of brain CSCs. shRNA shRNA-1 shRNA-2 However, false-positive and false-negative results due to Figure 5 PFKFB4 knockdown induces an increase of PI-positive technical pitfalls inherent in large-scale screening cannot brain CSCs. (a) Western blot analysis of PFKFB4 knockdown in be excluded. As expected, known survival kinases and NCH421k cells. Levels of silencing were assessed 3, 4 and 5 days phosphatases, such as SPHK1, FLT1 and SGK1 after lentiviral-mediated transduction with scrambled shRNA (Kasahara et al., 2000; Brunet et al., 2001; Gomez- (Ctl), PFKFB4-shRNA1 and -shRNA2. a-Tubulin antibody was used as loading control. (b) Induction of cell death 6 days after Brouchet et al., 2007), were identified, as well as lentiviral-mediated transduction with scrambled shRNA, PFKFB4- DUSP13, CHEK2, ADRBK1, MAP3K1, PPM1B and shRNA1 and -shRNA2. Left panels show representative phase PRKCA, reported as survival genes in previous screens contrast microphotographs of transduced NCH421k. Scale bar, (MacKeigan et al., 2005; Giroux et al., 2006). This 500 mm. Histogram plots in the right panels depict the PI fluorescence intensity, acquired with a FACSCanto flow cytometer. reflects the robustness of our screening settings. Transduction efficiency, as determined by GFP-positive cells, was Interestingly, each gene knockdown that was vali- 495% (data not shown). (c) Quantitative analysis of PI dated in the secondary screen demonstrated a similar fluorescence intensity for 60 000 cells, 6 days after lentiviral phenotype in the three independent brain CSC lines transduction. evaluated, but at variable levels, which might be explained by the heterogeneity existing between patient samples. Nonetheless, the phenotype obtained after

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3240 [%] ** * [%] ** * 3 days 3 days 4 days 4 days 100 100

80 80

60 60

40 40

20 20

Relative ATP concentration 0 Relative lactate concentration 0 Scrambled BCL2 PFKFB4 PFKFB4 Scrambled BCL2 PFKFB4 PFKFB4 shRNA shRNA shRNA-1 shRNA-2 shRNA shRNA shRNA-1 shRNA-2 shRNA-1 Ctl Ctl shRNA-2 Phospho-AMPK AMPK α-Tubulin

Figure 6 PFKFB4 knockdown induces cellular metabolism changes. (a, b) Lactate and ATP contents were measured using a lactate assay and an ATP assay, respectively, 3 and 4 days after transduction with scrambled shRNA, shRNA specific for BCL2 and shRNA-1 and -2 targeting PFKFB4. Results are depicted as mean of biological triplicates. Data were normalized to 20 000 living cells and are shown as percentage to the lactate content after transduction with scrambled shRNA. shRNA targeting BCL2 was used as an additional negative control, ensuring that the effect observed was not due to apoptosis initiation. Error bars represent the minimum and the maximum values of the biological triplicates (*P ¼ 0.06, **P ¼ 0.01). (c) NCH421k cells were transduced with scrambled shRNA (Ctl) and shRNA-1 and -2 targeting PFKFB4. Total AMPK and phosphorylation of AMPK on Thr172 were assessed by western blot analysis 5 days after transduction. a-Tubulin antibody was used as loading control.

silencing, which was observed in each of the brain CSC sphorylating phosphoenolpyruvate to pyruvate, which lines tested, and was specific to the tumor cells, as leads to a net production of ATP (Christofk et al., 2008). indicated by the comparison with normal NSCs, is of PDK1 is another important gene whose gene product great interest for further characterization and ultimately has an effect on pyruvate metabolism and is responsible for therapeutics development. phosphorylation and concomitant inactivation of pyruvate Among the putative kinases and phosphatases that we dehydrogenase (PDH). Its expression has been shown to identified as relevant for brain CSC survival, several are be associated with poor prognosis in head and neck known to be activated by HIF proteins and thereby squamous cell carcinoma (Wigfield et al., 2008). Suppres- regulated by the hypoxic microenvironment of the sion of PDH by PDK1 inhibits the conversion of pyruvate tumor. Indeed, FLT1 (Fms-related tyrosine kinase 1), to acetyl-CoA, thereby attenuating mitochondrial function the vascular endothelial growth factor receptor 1, is and increasing the production of lactate. Recently, regulated by HIF proteins and plays a key role in dichloroacetate (DCA), an inhibitor of PDK1, has been vascular development (Okuyama et al., 2006). Increased shown to induce apoptosis in vitro, thus confirming the FLT1 expression has been shown in various types of phenotype we observed after knockdown of PDK1 in brain carcinomas, glioblastoma, multiple myeloma and ne- CSCs (Wigfield et al., 2008; Michelakis et al., 2010). phroblastoma (Fischer et al., 2008). In addition, our The third kinase that has elicited a strong cell death approach highlighted three other kinases, namely phenotype after silencing in brain CSCs was PFKFB4, PFKFB4, the pyruvate dehydrogenase kinase 1 which encodes the bifunctional enzyme PFK2/FBPase-2 (PDK1) and the M2 splice isoform of (6-phosphofructo-2 kinase/fructose-2,6-biphosphatase (PKM2), which are direct targets of HIF proteins (Kress (PFKFB1-4)). This enzyme modulates the intracellular et al., 1998; Minchenko et al., 2004; Wigfield et al., concentration of the allosteric glycolytic activator, 2008). This response to low level of oxygen is parti- fructose-2,6-biphosphate (F2,6BP), which is a key cularly important in solid tumors, such as malignant regulator of glycolysis (Yalcin et al., 2009). It has been gliomas, that are characterized by aberrant neovascu- originally identified in the testes and shown to be highly larization and poor oxygen diffusion. Most importantly, expressed in both colon and breast cancers (Minchenko recent studies have shown that hypoxia also influences et al., 2004; Gomez et al., 2005). The homologous the survival and the proliferation of brain CSCs within PFKFB3 encodes an inducible form of PFK2/FBPase-2 the tumor and promotes reprogramming toward a stem- that is ubiquitously expressed in many normal tissues, like phenotype (Keith and Simon, 2007; Heddleston such as brain and liver, and overexpressed in different et al., 2009; Soeda et al., 2009). Another important cancers (Manzano et al., 1998; Kessler and Eschrich, common feature of these three kinases is their involve- 2001; Atsumi et al., 2002). Both PFKFB3 and PFKFB4 ment in the glycolytic pathway. Indeed, PKM2 has been are induced by hypoxia in various tumors. Interestingly, reported to be overexpressed in tumor tissue and to be primary glioblastomas showed a significant higher advantageous for tumor cell growth. It is required for expression of PFKFB3 and PFKFB4 when compared aerobic glycolysis and catalyzes the last step by depho- with secondary glioblastomas as well as with the lower-

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3241 grade astocytomas. Furthermore, high expression levels from intraoperative ventricular tissue after informed consent of PFKFB4 correlated with poor survival in primary was given for the scientific use of anonymized samples and glioblastoma patients, whereas PFKFB3 expression was after approval by the ethics committee of the Neurosurgery not related to the clinical outcome of these patients. In Department/Charite´, University Hospital Berlin, Germany addition, mRNA expression of both PFKFBs showed (registration number 125/2001). The normal genotype was an opposite pattern in brain CSCs, in which PFKFB4 verified at passage 10 (data not shown). All cell lines were cultivated at 37 1C in a humidified was overexpressed whereas PFKFB3 was downregu- incubator with 5% CO2 as floating aggregates, the so-called lated when compared with normal brain tissue. These neurospheres on uncoated tissue culture dishes. Brain CSC findings suggest that PFKFB4 is the main PFK2/ medium consisted of Dulbecco’s modified Eagle’s medium/ FBPase-2 isozyme that regulates glycolytic flux in F-12 medium containing 20% BIT serum-free supplement, malignant glioma cells. This can be partly explained basic fibroblast growth factor and epidermal growth factor by the common loss of 10p in glioblastoma at a concentration of 20 ng/ml each (all from Provitro, Berlin, that contains the locus of PFKFB3. However, we Germany). Normal NSCs were maintained in Dulbecco’s cannot exclude that PFKFB3 plays a role in brain modified Eagle’s medium/F-12 medium supplemented with CSCs survival as suggested by its mRNA expression serum-free B27 (Life Technologies, Darmstadt, Germany), 1 M level in glioblastoma, but it is also required for normal HEPES (Gibco, Darmstadt, Germany), basic fibroblast growth factor and epidermal growth factor (20 ng/ml each). brain tissue homeostasis, as shown by its expression in normal brain. Our results thus suggest PFKFB4 High-throughput lentivirus production and transduction as a key molecule for the survival of brain CSCs. The Mission RNAi library was purchased from Sigma-Aldrich Moreover, its inhibition might affect not only brain (St Louis, MO, USA). High-quality DNA preparations were CSCs, but also the tumor bulk, as implied by the expres- obtained using a large-scale plasmid purification kit (Qiagen, sion level in glioblastoma. To test this hypothesis, brain Hilden, Germany and Roche, Mannheim, Germany). The CSCs should be treated with small-molecule inhibitors, lentivirus particles were produced in a 96-well format such as 3PO, that bind the substrate binding site of according to the TRC (The RNAi Consortium) protocol. PFKFBs (Clem et al., 2008). Such studies were initiated Briefly, 293T packaging cells were co-transfected with the but are hampered by the limited specificity of 3PO, as all pLKO.1 vector encoding the shRNA, as well as the necessary four PFKFBs share a high homology in their substrate- helper plasmids for virus production (psPAX2 and pMGD2), binding domain. Therefore, further efforts are necessary using Trans-IT (Mirus, Madison, WI, USA). The titer was measured using lentiviral particles that contained the pLKO.1 to develop inhibitors specifically targeting PFKB4. vector expressing GFP and ranged from 3 Â 106 to 5 Â 106 IU/ Silencing of PDK1 and PKM2 resulted in increased ml. Before transduction, cultured spheres were dissociated apoptosis of brain CSCs. Knockdown of PFKFB4 with trypsin or accutase treatment. For cell death analysis, showed a downregulation of lactate and ATP production NCH421k cells were seeded in 96-well plates at 20 000 cells per in brain CSCs and ultimately induced apoptosis possibly well in a final volume of 100 ml and transduced at a multiplicity by the phosphorylation/activation of AMPK because of of infection (MOI) of 5 in the presence of 8 mg/ml polybrene. the higher ratio of AMP/ATP. The identification of At 24 h after transduction, medium was replaced with fresh glycolytic kinases in our screen reflects the importance of brain CSC medium. Cells were analyzed 6 days after viral the Warburg effect by liberating the cells from oxygen transduction using PI staining and a flow-cytometer equipped dependence for ATP production. This adaptation to the with a high-throughput sampler. Our experimental settings ensured a maximal efficiency, but resulted in some transduc- tumor microenvironment has been shown to confer a tion- and readout-associated cytotoxicity and an average of survival advantage to cancer cells (Hsu and Sabatini, cell survival of 70.15% for all 3838 shRNAs (s.d.±10.25%) 2008). Our findings highlight the importance of the lactate and of 26.24% in wells containing scrambled shRNAs and ATP production in the maintenance of brain CSCs in (n ¼ 768; s.d.±11.38%). the tumor microenvironment and support a promise for To minimize potential artifacts due to position effects, the therapies that target proteins involved in the regulation of rows 1 and 12 were not used. In addition, a plate that contained tumor metabolism. a scrambled shRNA in each well was included and normal- ization was performed by matching the cell death rate to the corresponding well from that control plate. The production of single shRNAs was performed in 6-cm petri dishes. After 72 h, Materials and methods produced lentiviruses were concentrated by ultracentrifugation of the HEK293T supernatant. Transductions were performed Glioma patients at the MOI of 5 with 8 mg/ml polybrene for brain CSCs, and at See Supplementary Materials and methods. the MOI of 10 without polybrene for normal NSCs.

RNA extraction and qRT–PCR Cell culture See Supplementary Materials and methods. The investigated three brain CSC lines, NCH421k, NCH441 and NCH644, were established from primary glioblas- toma patients undergoing surgical resection according to the Gene expression analysis research proposals approved by the institutional review board See Supplementary Materials and methods. at the Medical Faculty, University of Heidelberg. The brain CSC lines were characterized genotypically and phenotypi- FACS and cell death analysis cally in a previous study (Campos et al., 2010). An adult Flow cytometry analysis was performed for NCH421k cells 6 days human neural stem cell line (BLV4) was originally cultivated after shRNA transduction in 96-well cell culture plates. Cells were

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3242 transferred onto a 96-well filter plate (Millipore, Bellerica, MA, Lactate and ATP assays USA), centrifuged, trypsinized and resuspended in 100 ml staining Brain CSCs were seeded at 2 Â 104 cells per 100 ml in 96-well solution (phosphate-buffered saline, 1.5 mg/ml PI). Samples were plate and immediately transduced with lentiviruses at an MOI analyzed using a flow cytometer equipped with a high-throughput of 5. Lactate levels were measured 3 and 4 days later using a sampler (FACSArray, BD Biosciences, San Jose, CA, USA). lactate colorimetric assay read at 570 nm according to the Validation of the cell death phenotype upon silencing of manufacturer’s instructions (MBL International, Woburn, PFKFB4 with shRNA-1 (50-CCTGTGGCATATGGTTGTA MA, USA). ATP levels were measured 3 and 4 days after AA-30) and shRNA-2 (50-GACGTGGTCAAGACCTACAA silencing using the ATP luminometric determination kit A-30) was performed in a six-well plate. Samples were run on a according to the manufacturer’s instructions (Invitrogen, FACSCanto flow cytometer and analyzed with FACS Diva Carlsbad, CA, USA). Concentration was calculated based on software (Becton Dickinson, San Jose, CA, USA). a standard curve and normalized to the total number of living cells. The cell number was estimated by the MTS-based Apoptosis and cell viability assays CellTiter 96 AQueous Assay (Promega) and calculated based To study apoptosis we used a caspase assay. Cells were seeded on a standard curve. at 20 000 cells per 100 ml per well in 96-well clear plates and transduced with lentiviral shRNA. Caspase-3/7 activity was Statistical analysis measured 4, 5, 6 and 7 days after transduction, using a R 2.12 (http://www.r-project.org) was used for statistical homogeneous luminescent method (Caspase-Glo 3/7 Assay, analyses. The significance was calculated using Student’s t-test Promega, Madison, WI, USA). Transduced cells were assuming unequal variances. The fold changes in apoptotic transferred to white/solid bottom assay plates and 100 mlof rate were compared after rank transformation (Zimmerman Caspase-Glo 3/7 reagent was added. Plates were incubated at and Zumbo, 1993). room temperature for 1 h, and luminescence intensity deter- For the comparison of the Kaplan–Meier curves, the mined using the luminometer MITHRAS plate reader Mantel–Haenszel test was applied. Using a bootstrapping (PerkinElmer, Shelton, CT, USA). approach with 1000 draws, the P-value for differences in To assess brain CSC viability, the MTS (3-(4,5-dimethyl- survival based on the expression of PFKFB4 was confirmed by thiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H- reselecting the mean expression value and calculating the tetrazolium)-based CellTiter 96 AQueous Assay (Promega) Mantel–Haenszel P-value. Multivariate analyses were per- was used following the manufacturer’s recommendations. formed calculating Cox hazard models. All samples were assayed in three biological replicas. The 3PO (3-(3-pyridinyl)-1-(4-pyridinyl)-2-propen-1-one) was kindly provided by Dr J Chesney (Louisville, KY, USA). Rapamycin InSolution was provided in dimethyl sulfoxide Conflict of interest (Merck, Darmstadt, Germany). The authors declare no conflict of interest. Western blot analysis Cell lysates were separated in 10% sodium dodecyl sulfate– polyacrylamide gel electrophoresis and transferred to nitrocel- Acknowledgements lulose membranes. The monoclonal mouse antibody against a-tubulin (Sigma-Aldrich) was used in a dilution of 1:2000, the We thank Dr S Pfister and D Haag for their helpful comments. polyclonal rabbit antibody against PFKFB4 (Sigma-Aldrich) We thank Mrs V Lang and G Iren for their technical help. We in a dilution of 1:1000 and the monoclonal rabbit and mouse thank Dr J Chesney for kindly providing 3PO. This work was antibodies against phospho-AMPK (Thr172) and total AMPK, supported by the German Federal Ministry of Education and respectively (Cell Signaling Technology, Danvers, MA, USA), Research (BMBF) within the National Genome Research in a dilution of 1:1000. Horseradish peroxidase-conjugated Network (NGFNplus; 01GS0883, 01GS0884), a stipend from secondary antibodies (Abcam, Cambridge, UK) were used in a the Alexander von Humboldt foundation (to SN) and by the 1:5000 dilution before the chemiluminescent detection of bound grant ‘Young Investigator Fellowship’ of the Medical Faculty antibodies (ECL plus kit; GE Healthcare, Little Chalfont, UK). of Heidelberg (to VG).

References

Al-Hajj M. (2007). Cancer stem cells and oncology therapeutics. Curr Campos B, Wan F, Farhadi M, Ernst A, Zeppernick F, Tagscherer KE Opin Oncol 19: 61–64. et al. (2010). Differentiation therapy exerts antitumor effects on Atsumi T, Chesney J, Metz C, Leng L, Donnelly S, Makita Z et al. stem-like glioma cells. Clin Cancer Res 16: 2715–2728. (2002). High expression of inducible 6-phosphofructo-2-kinase/ Christofk HR, Vander Heiden MG, Harris MH, Ramanathan A, fructose-2,6-bisphosphatase (iPFK-2; PFKFB3) in human cancers. Gerszten RE, Wei R et al. (2008). The M2 splice isoform of Cancer Res 62: 5881–5887. pyruvate kinase is important for cancer metabolism and tumour Bao S, Wu Q, McLendon RE, Hao Y, Shi Q, Hjelmeland AB growth. Nature 452: 230–233. et al. (2006). Glioma stem cells promote radioresistance by Clem B, Telang S, Clem A, Yalcin A, Meier J, Simmons A et al. (2008). preferential activation of the DNA damage response. Nature 444: Small-molecule inhibition of 6-phosphofructo-2-kinase activity 756–760. suppresses glycolytic flux and tumor growth. Mol Cancer Ther 7: Brunet A, Park J, Tran H, Hu LS, Hemmings BA, Greenberg ME. 110–120. (2001). Protein kinase SGK mediates survival signals by phosphor- Eramo A, Ricci-Vitiani L, Zeuner A, Pallini R, Lotti F, Sette G et al. ylating the forkhead transcription factor FKHRL1 (FOXO3a). Mol (2006). Chemotherapy resistance of glioblastoma stem cells. Cell Cell Biol 21: 952–965. Death Differ 13: 1238–1241.

Oncogene Brain cancer stem-like cells require PFKFB4 for survival V Goidts et al 3243 Ernst A, Hofmann S, Ahmadi R, Becker N, Korshunov A, Engel F Michelakis ED, Sutendra G, Dromparis P, Webster L, Haromy A, et al. (2009). Genomic and expression profiling of glioblastoma stem Niven E et al. (2010). Metabolic modulation of glioblastoma with cell-like spheroid cultures identifies novel tumor-relevant genes dichloroacetate. Sci Transl Med 2: 31ra34. associated with survival. Clin Cancer Res 15: 6541–6550. Minchenko O, Opentanova I, Minchenko D, Ogura T, Esumi H. Fischer C, Mazzone M, Jonckx B, Carmeliet P. (2008). FLT1 and its (2004). Hypoxia induces transcription of 6-phosphofructo-2-kinase/ ligands VEGFB and PlGF: drug targets for anti-angiogenic fructose-2,6-biphosphatase-4 gene via hypoxia-inducible factor- therapy? Nat Rev Cancer 8: 942–956. 1alpha activation. FEBS Lett 576: 14–20. Floridi A, Paggi MG, Fanciulli M. (1989). Modulation of glycolysis in Okuyama H, Krishnamachary B, Zhou YF, Nagasawa H, Bosch- neuroepithelial tumors. J Neurosurg Sci 33: 55–64. Marce M, Semenza GL. (2006). Expression of vascular endothelial Giroux V, Iovanna J, Dagorn JC. (2006). Probing the human kinome growth factor receptor 1 in bone marrow-derived mesenchymal for kinases involved in pancreatic cancer cell survival and cells is dependent on hypoxia-inducible factor 1. J Biol Chem 281: gemcitabine resistance. FASEB J 20: 1982–1991. 15554–15563. Gomez-Brouchet A, Pchejetski D, Brizuela L, Garcia V, Altie MF, Pardal R, Clarke MF, Morrison SJ. (2003). Applying the principles of Maddelein ML et al. (2007). Critical role for sphingosine kinase-1 in stem-cell biology to cancer. Nat Rev Cancer 3: 895–902. regulating survival of neuroblastoma cells exposed to amyloid-beta Park CY, Tseng D, Weissman IL. (2009). Cancer stem cell-directed peptide. Mol Pharmacol 72: 341–349. therapies: recent data from the laboratory and clinic. Mol Ther 17: Gomez M, Manzano A, Navarro-Sabate A, Duran J, Obach M, 219–230. Perales JC et al. (2005). Specific expression of pfkfb4 gene in Pradelli LA, Beneteau M, Chauvin C, Jacquin MA, Marchetti S, spermatogonia germ cells and analysis of its 50-flanking region. Munoz-Pinedo C et al. (2010). Glycolysis inhibition sensitizes FEBS Lett 579: 357–362. tumor cells to death receptors-induced apoptosis by AMP kinase Heddleston JM, Li Z, McLendon RE, Hjelmeland AB, Rich JN. activation leading to Mcl-1 block in translation. Oncogene 29: (2009). The hypoxic microenvironment maintains glioblastoma stem 1641–1652. cells and promotes reprogramming towards a cancer stem cell Reya T, Morrison SJ, Clarke MF, Weissman IL. (2001). Stem cells, phenotype. Cell Cycle 8: 3274–3284. cancer, and cancer stem cells. Nature 414: 105–111. Hsu PP, Sabatini DM. (2008). Cancer cell metabolism: Warburg and Sakata J, Abe Y, Uyeda K. (1991). Molecular cloning of the DNA and beyond. Cell 134: 703–707. expression and characterization of rat testes fructose-6-phosphate,2- Kasahara Y, Tuder RM, Taraseviciene-Stewart L, Le Cras TD, kinase:fructose-2,6-bisphosphatase. J Biol Chem 266: 15764–15770. Abman S, Hirth PK et al. (2000). Inhibition of VEGF recep- Singh SK, Clarke ID, Terasaki M, Bonn VE, Hawkins C, Squire J tors causes lung cell apoptosis and emphysema. J Clin Invest 106: et al. (2003). Identification of a cancer stem cell in human brain 1311–1319. tumors. Cancer Res 63: 5821–5828. Keith B, Simon MC. (2007). Hypoxia-inducible factors, stem cells, and Soeda A, Park M, Lee D, Mintz A, Androutsellis-Theotokis A, cancer. Cell 129: 465–472. McKay RD et al. (2009). Hypoxia promotes expansion of the Kessler R, Eschrich K. (2001). Splice isoforms of ubiquitous 6- CD133-positive glioma stem cells through activation of HIF-1alpha. phosphofructo-2-kinase/fructose-2,6-bisphosphatase in human Oncogene 28: 3949–3959. brain. Brain Res Mol Brain Res 87: 190–195. Stupp R, Hegi ME, Gilbert MR, Chakravarti A. (2007). Chemor- Kress S, Stein A, Maurer P, Weber B, Reichert J, Buchmann A et al. adiotherapy in malignant glioma: standard of care and future (1998). Expression of hypoxia-inducible genes in tumor cells. J directions. J Clin Oncol 25: 4127–4136. Cancer Res Clin Oncol 124: 315–320. Toedt G, Barbus S, Wolter M, Felsberg J, Tews B, Blond F et al. Lee J, Kotliarova S, Kotliarov Y, Li A, Su Q, Donin NM et al. (2006). (2011). Molecular signatures classify astrocytic gliomas by IDH1 Tumor stem cells derived from glioblastomas cultured in bFGF and mutation status. Int J Cancer 128: 1095–1103. EGF more closely mirror the phenotype and genotype of primary Vescovi AL, Galli R, Reynolds BA. (2006). Brain tumour stem cells. tumors than do serum-cultured cell lines. Cancer Cell 9: 391–403. Nat Rev Cancer 6: 425–436. Liu G, Yuan X, Zeng Z, Tunici P, Ng H, Abdulkadir IR et al. (2006). Wigfield SM, Winter SC, Giatromanolaki A, Taylor J, Koukourakis Analysis of gene expression and chemoresistance of CD133+ ML, Harris AL. (2008). PDK-1 regulates lactate production in cancer stem cells in glioblastoma. Mol Cancer 5: 67. hypoxia and is associated with poor prognosis in head and neck Lu H, Forbes RA, Verma A. (2002). Hypoxia-inducible factor 1 squamous cancer. Br J Cancer 98: 1975–1984. activation by aerobic glycolysis implicates the Warburg effect in Wurdak H, Zhu S, Romero A, Lorger M, Watson J, Chiang CY et al. carcinogenesis. J Biol Chem 277: 23111–23115. (2010). An RNAi screen identifies TRRAP as a regulator of MacKeigan JP, Murphy LO, Blenis J. (2005). Sensitized RNAi screen brain tumor-initiating cell differentiation. Cell Stem Cell 6: of human kinases and phosphatases identifies new regulators of 37–47. apoptosis and chemoresistance. Nat Cell Biol 7: 591–600. Yalcin A, Telang S, Clem B, Chesney J. (2009). Regulation of glucose Manzano A, Rosa JL, Ventura F, Perez JX, Nadal M, Estivill X et al. metabolism by 6-phosphofructo-2-kinase/fructose-2,6-bisphospha- (1998). Molecular cloning, expression, and chromosomal localiza- tases in cancer. Exp Mol Pathol 86: 174–179. tion of a ubiquitously expressed human 6-phosphofructo-2-kinase/ Zimmerman DW, Zumbo BD. (1993). Rank transformations and fructose-2, 6-bisphosphatase gene (PFKFB3). Cytogenet Cell Genet power of the Student t test and Welch t test for nonnormal 83: 214–217. populations with unequal variances. Can J Exp Psychol 47: 523–539.

Supplementary Information accompanies the paper on the Oncogene website (http://www.nature.com/onc)

Oncogene