The serine–threonine kinase AMPK is a heterotrimeric protein complex of catalytic α and regulatory β and γ subunits1,2,3. All AMPK subunits are required for AMPK stability and activity4. At lower cellular energy states, the γ subunits bind AMP/ADP and enhance AMPK activity to bring about energy homeostasis. AMP/ADP binding also enables the upstream metabolic kinases LKB1 and CAMKKβ to phosphorylate α subunits, fully activating AMPK5,6. AMPK is a metabolic hub1,2,3, yet its function in cancer cell metabolism remains undefined. AMPK inhibits biosynthetic kinases such as mammalian target of rapamycin (mTOR) and acetyl-CoA carboxylase (ACC)7,8,9. Therefore, AMPK is expected to play a suppressive role in cancer. Despite its tumour suppressive role in other cancers10, a potential oncogenic role of activated AMPK was alluded in astrocytic tumours of the brain11. Controversy surrounding its role in cancer stems in part due to the absence of genetic models and the use of non-specific pharmacological agents12. Contrary to early pharmacological studies13,14,15,16, some recent genetic studies showed that in some contexts, AMPK provides survival advantage critical for tumour growth17,18,19,20,21,22. In contrast, AMPKα1 knockout enhanced glycolysis and accelerated tumourigenesis in a lymphoma mouse model10, demonstrating species-specific and tissue-specific effects. Glycolysis is positively regulated by HIF1α. However, the role of AMPK in glycolysis and its relation with HIF1α is unclear. HIF1α that is degraded in the presence of O2 was found to be stabilized under normoxic condition in LKB1-deficient MEFS (where AMPK activity is reduced)23. In contrast, no such HIF1α stabilization was observed in AMPK-null MEFs21. While AMPK was found to inhibit the Warburg effect through HIF1α destabilization in mouse lymphoma10 and reduce glycolysis in mouse acute lymphoblastic leukaemia (ALL)24, AMPK promoted glycolysis in mitotically stressed cells, breast tumours, skin fibroblasts and astrocytes17,25,26,27,28. In this study, we present a mechanism by which AMPK regulates HIF1α transcription and glycolysis in GBM. We provide evidence that through phosphorylation of CREB1, a transcription factor highly expressed in GBM, AMPK, controls HIF1α and GABPA transcription to regulate GBM bioenergetics.


AMPK is highly expressed in GBM

While searching the Cancer Genome Atlas (TCGA) database for differentially expressed metabolic kinases in human cancer, we observed significantly higher expression of the AMPK α1, β1 and γ1 subunits (P ≤ 107) in GBM than in normal brain (Fig. 1a), and in GBM relative to lower-grade glioma (LGG) (Fig. 1b; P ≤ 10−14). Higher expression of AMPK-α1 (P = 0.0007), AMPK-β1 (P = 0.01), pAMPK (active phosphorylated AMPK, P = 0.003) and the AMPK substrate pACC (P = 0.01) also correlated with poor patient survival in LGG (Fig. 1c and Supplementary Fig. 1a), but not with GBM (Supplementary Fig. 1b), potentially due to higher basal expression of AMPK across all GBM relative to LGG. This is reminiscent of other genes (such as HIF1α and CREB1) that are highly expressed (sometimes uniformly) across GBM relative to LGG29,30,31, and are not prognostic but are important for GBM pathogenesis32,33,34. Biochemical analysis of human GBM and mouse high-grade glioma (HGG)12,35 (Fig. 1d–g and Supplementary Fig. 1c) revealed that active (phosphorylated) AMPK is high in tumours compared to normal brain tissue, and higher in tumour cells relative to infiltrating macrophages (Supplementary Fig. 1d,e). Compared to normal brain, the AMPK upstream kinases CAMKKβ and LKB1 were not upregulated (Fig. 1d). Of the AMPK subunits, the α and β but not γ1 were upregulated in GBM (Fig. 1d). The high pAMPK signature was retained in primary GBM stem cell lines (GSCs) derived from fresh tumour tissue of both proneural and mesenchymal subtypes (Fig. 1h).

Fig. 1: AMPK is highly expressed in GBM.
figure 1

a,b, Box plots (derived from TCGA Affimetrix data) showing high expression of AMPK isoforms in GBM compared to normal brain and low-grade glioma (LGG) (n = 10 normal adult human brain versus 548 GBM and 534 LGG). The edges on the box plots indicate the first and third quartiles (25th and 75th percentiles) of the data, with the line in the middle being the median. The whiskers on the boxplots extend another 1.5× of the interquartile range (between 25–75% range of data) from the edges of the boxes, respectively. *P < 0.001. c, Kaplan–Meier survival plots of LGG patients. d, Western blot (WB) showing levels of pAMPK and AMPK pathway genes in GBM and normal human brain (NHB). e, Immunohistochemical (IHC) analysis of normal brain and GBM using pAMPK antibody. Scale bars, 100 μm. f, Quantitation of pAMPK signal in human tissues (n = 15 GBM; 5 normal brain). *P = 0.0005. g, IHC showing pAMPK signal in mouse high grade glioma; N, normal tissue; T, tumour. Scale bars, 500μm. h, WB of pAMPK and pACC in primary human GSC lines and normal human astrocytes (NHAs). i, WB showing pAMPK and pACC in NHA and GSC lines in response to changes in glucose concentration. j, WB using PARP1 antibody showing cleaved PARP in control and AMPKβ1-shRNA-expressing GSCs. Actin was used as a loading control. k, Viability of GCSs and NHAs expressing non-target or AMPKβ1 shRNA (n = 3 independent experiments; *P < 0.005). Error bars are mean ± s.d. Statistical significance was assessed using Student’s two-tailed t-test, except in a,b, where Welch’s t-test was used. Source data are available in Supplementary Table 4. All WB represent data from 2–3 independent repeats. Unprocessed blots are available in Supplementary Fig. 9.

In normal cells, AMPK activity is stimulated by glucose starvation and reduced by glucose feeding1,2,3. Unlike in normal human glial cells, high AMPK activity in GSCs was chronically maintained and was insensitive to glucose levels (Fig. 1i). We questioned if high AMPK activity in these cells is due to oncogenesis-associated stress (OAS), which constitutes a variety of stress signals (such as ER stress, DNA damage response and oxidative stress) that are elicited as a coping response to oncogenic events36. GBM demonstrated high genotoxic and ER stress relative to normal brain cells and tissue (Supplementary Fig. 1f–h). Agents that mimicked OAS (Supplementary Fig. 1i,j) also activated AMPK (Supplementary Fig. 1k–m), suggesting a link between OAS and pAMPK in GBM. Consistent with this observation, expression of oncogenic EGFR or KRAS, or reducing PTEN levels, in detached human astrocytes increased AMPK phosphorylation (Supplementary Fig. 1n–p).

AMPK is required for viability of patient-derived primary GSCs

We investigated whether pAMPK is merely an indicator of OAS or is necessary for the viability of GSCs. AMPK silencing by AMPKβ1 shRNA induced apoptosis (Fig. 1j and Supplementary Video 1a,b). Depletion of AMPK by five independent genetic strategies (two AMPKβ1 shRNAs; three β1 siRNAs; dominant-negative AMPKα2; AMPKα1α2 shRNAs; and AMPKβ1 CRISPR) significantly reduced viability of primary GSC lines (by ~40–70%); normal human astrocytes (NHAs) remained relatively viable (Figs. 1k and 2a–h). Cre-mediated deletion of AMPKβ1 also reduced viability of oncogenic mouse neural stem cells (NPCs) (Fig. 2i). Strikingly, viability of the long-established GBM serum lines U87, A172 and T98G remained unaffected by AMPK shRNA (Fig. 2j). The effect of human AMPKβ1 shRNA was specific because expression of shRNA-resistant mouse AMPKβ1 cDNA in GSC9 and GSC10 lines rescued viability defects (Fig. 2k,l). Consistent with a previous report37, the AMPKβ1 subunit was specifically critical for survival, because depletion of the less-expressed β2 subunit did not reduce pAMPK levels or induce cell death, and its overexpression failed to rescue β1-silenced cells (Fig. 2m–o). Inhibition of upstream kinases that activate AMPK (LKB1 and CAMKKβ) also reduced the viability of GSC lines, consistent with the AMPK activation pathway being important for GBM (Fig. 2p).

Fig. 2: AMPK is essential for viability of primary GBM lines in vitro.
figure 2

a, WB using pAMPK, pACC and AMPKβ1/β2 common antibody in AMPKβ1-shRNA-treated GSC10 cells. b, Cell viability using AMPKβ1 shRNA no. 2 (n = 3; *P ≤ 0.007, #P ≤ 0.001). c, GSC viability using AMPKβ1 siRNA (n = 3; *P ≤ 0.001). Top: WB of β1/β2. Note that due to sequence homology, siRNAs knocked down both β1 and β2. d, WB of pAMPK and pACC in GSC lines expressing dominant-negative (DN) AMPKα2. Cont., control. e, GSC viability using DN AMPK (average of two independent experiments). f, GSC viability using AMPKα1/α2 shRNA (n = 3; *P = 0.0004, +P = 0.0002, **P = 0.0003). Top: WB of AMPKα1/α2. g, WB showing CRISPR knockout of AMPKβ1 in GSC10 cells. h, Viability of GSCs expressing AMPKβ1 CRISPR (n = 3). Note that CRISPR transfection efficiency in GSCs is low (~ 30–40%). *P = 0.004, +P = 0.002, **P = 0.01, ++P ≤ 0.006. i, Viability of oncogenic mouse neural stem cells (NPCs) from compound floxed mice (Ink4/Arf−/−; Pten lox/lox with AMPKβ1lox/lox or AMPKβ1+/+) treated with Adeno-Cre. (n = 3; *P = 0.0006). j, Viability of established GBM serum cell lines expressing AMPKβ1 shRNA (average of two independent experiments). NS, not significant. k, WB of GSC10 cells expressing human AMPKβ1 shRNA with or without mouse AMPKβ1. Note that AMPK α1/2 subunits are unstable in the absence of β subunits. l, Viability of GSCs expressing human AMPKβ1 shRNA with or without mouse AMPKβ1 (average of two independent experiments). m, WB in GSCs expressing AMPKβ2 shRNA. n, GSC viability in the presence of AMPKβ2 shRNA (average of two independent experiments). o, Viability of GSCs expressing human AMPKβ1 shRNA with or without mouse AMPKβ2 (n = 3; *P ≤ 0.003, +P ≤ 0.004, NS = non-significant). p, Viability of GSCs expressing LKB1 shRNA or treated with CAMKKβ inhibitor (STO) or ATM inhibitor (KU). (n = 3; *P ≤ 0.0008; +P = 0.0004; **P = 0.0006. Top: WB showing efficiency of LKB1 shRNA. q,r, Viability of GSCs or GBM serum lines at indicated conditions (n = 3; NS = non-significant, *P ≤ 0.0006; +P ≤ 0.003). Error bars are mean ± s.d. Statistical significance, two-tailed t-test; n values are independent experiments. Source data are available in Supplementary Table 4. WB represent data from two (d,g,k,m,p) or three (a,c) independent repeats. Unprocessed blots are available in Supplementary Fig. 9.

AMPK activation during metabolic and oncogenic stress provides protection36. Yet, unexpectedly, a tumour-mimetic metabolic stress-like condition (1 mM glucose or 1% O2) did not enhance death of AMPK-silenced GSCs beyond what occurred at the physiologically relevant 5 mM glucose level (at which our primary GSC lines are grown) (Fig. 2q). In fact, unlike long-established GBM serum lines, primary GSCs remained viable under stress (Fig. 2q,r). Thus, stress-tolerant primary lines are reliable surrogates of the tumour in maintaining an AMPK-dependent stress adaptation mechanism for survival in vitro.

AMPK is required for intracranial growth of GSC-derived tumours

To test if AMPK is necessary for tumour growth in vivo, we transplanted four primary GSC lines expressing non-target control (NT), AMPKβ1 shRNA or AMPKα1/α2 shRNA into the cerebral cortex of immunocompromised mice. AMPK depletion increased survival and reduced tumour growth in all GSC lines (Fig. 3a,b and Supplementary Fig. 2a–e), while no effect of AMPK silencing was observed in the GBM serum line U87 (Supplementary Fig. 2f). Control mice (non-target shRNA) lived for 17–29 days, while mice transplanted with GSCs expressing AMPKβ1 shRNA or AMPKα1/α2 shRNA either lived longer or remained tumour free (Fig. 3b,c and Supplementary Fig. 2e). Consistent with our in vitro results, expression of a shRNA-resistant cumate-inducible mouse AMPKβ1 completely rescued tumour growth and survival (Fig. 3d–f and Supplementary Fig. 2g,h). AMPK-silenced tumours showed significant residual AMPKβ1 expression and AMPK activity at both the tumour and cellular levels (Supplementary Fig. 2i–k), indicating inefficient silencing. The AMPK-silenced tumours were smaller and showed reduced proliferation and increased apoptosis compared to controls (Fig. 3c,g–l). We took one step further and tested whether repression of AMPK in established tumours enhances survival. Tumours were established using cells expressing both human AMPKβ1 shRNA and cumate-induced mouse AMPKβ1 (which is resistant to human shRNA). Withdrawal of cumate significantly increased survival of tumour-bearing mice (Fig. 3m).

Fig. 3: AMPK is essential for optimal GSC growth in vivo.
figure 3

a, Luciferase imaging of mice to monitor tumour growth of a GSC line (326) expressing AMPKβ1 or non-target (NT) shRNA at the indicated days. n = 8 mice per group (4 are shown). b, Kaplan–Meier survival data of four GSC lines expressing NT or AMPKβ1 shRNA. Note that the difference in survival between line 83 NTshRNA and 83 β1shRNA is not apparent due to the extended x axis, but is still significant. n = 326 (8 NT and 8 β1shRNA); AC17 (7 NT and 8 β1shRNA); 1123 (4 NT and 4 β1shRNA); 83 (8 NT and 6 β1shRNA)) c, H&E staining of tumours harvested at indicated days. df, In vivo specificity of AMPKβ1 shRNA was tested by using a cumate-inducible lentiviral expression system. The system works through the CymR repressor that binds the cumate operator sequences with high affinity. The repression is alleviated through the addition of Cumate, a non-toxic small molecule that binds to CymR. d, GSC326 transduced with cumate-inducible lentivirus and luciferase lentivirus were transplanted intracranially. Following confirmation of tumour growth by luciferase imaging, intraperitoneal delivery of water-soluble cumate (150 mg kg–1) rapidly turned on GFP. e, GSC326 expressing cumate-inducible mouse AMPKβ1 were transduced with human AMPKβ1 shRNA or NT shRNA and transplanted intracranially. 50% mice with AMPKβ1 shRNA received cumate or vehicle once every day. The images shown were captured on day 17 post transplantation. f, Kaplan-Meier survival data of three groups (n = 5 mice per group). P = 0.0001. g, h, IHC of Ki67 in NT and AMPKβ1 shRNA expressing tumours. Scale bar 100μm. i, Quantification of Ki67 positive cells (n = 3 mice per line per genotype; *P ≤ 0.001, +P ≤ 0.004). j,k, IHC of cleaved caspase-3 in NT- and AMPKβ1-shRNA-expressing tumours. Nuclei were stained with DAPI. Scale bar, 100 μm. l, Quantification of cleaved caspase-3-positive cells (n = 3 mice per line per genotype; *P ≤ 0.02, +P ≤ 0.002. m, Tumours were established using a GSC326 line expressing cumate-inducible mouse AMPKβ1 and human AMPKβ1 shRNA. Once tumours formed, cumate induction was continued in one group and stopped in another group. Kaplan–Meier survival data was plotted on two groups of mice (n = 5–6 mice per condition). Error bars are mean ± s.d. Statistical significance, two-tailed t-test. Source data are available in Supplementary Table 4.

GSC death occurs independent of AMPK-regulated mTOR and autophagy pathways

Cancer cell survival can be positively or negatively regulated by autophagy, depending on the context. AMPK positively regulates autophagy and, consequently, AMPK-silenced GSCs showed diminished basal autophagy and autophagy flux, as revealed by lower levels of LC3-II and P62 basally and in the presence of the autophagy inhibitor bafilomycin (Supplementary Fig. 3a,b). AMPK-regulated mitophagy has a broad role in mitochondrial homeostasis and could be crucial for survival in some contexts. However, reduced autophagy is unlikely to be the cause of cell death in AMPK-silenced primary GSCs because consistent with previous studies38,39, autophagy inhibition alone was insufficient to kill GSCs (Supplementary Fig. 3c). AMPK maintains cellular redox state40,41 by phosphorylating and inhibiting ACC1, which uses NADPH. In AMPK-silenced cells, hyperactivated ACC1 can deplete NADPH, causing oxidative stress and death that can be rescued by the reducing agent N-acetylcysteine (NAC)41. AMPK silencing raised levels of superoxide anion (Supplementary Fig. 3d), but neither the ACC inhibitor TOFA nor NAC rescued cell death, suggesting that superoxide is not the cause of death in primary GSC lines (Supplementary Fig. 3e). AMPK balances energy expenditure by inhibiting mTORC1, which can drain energy if uninhibited8,9. Indeed, AMPK depletion caused a significant drop in ATP (Supplementary Fig. 3f). However, AMPK silencing did not increase mTORC1 activity, and, consistent with previous work41, the mTORC1 inhibitor rapamycin failed to protect AMPK-silenced GSCs (Supplementary Fig. 3g). AMPK silencing did not significantly affect the expression of stem cell markers in GSCs (Supplementary Fig. 3h). Considering these results, we postulated that AMPK-depleted GSCs probably died not due to energy drainage but due to a deficit in energy production.

AMPK regulates GSC bioenergetics

To gain molecular insight, we performed RNA sequencing (RNA-seq). Analysis of differential gene expression and deregulated pathways showed that bioenergetics of cellular metabolism (glycolysis and mitochondrial function) was the most significantly downregulated pathway in AMPK-depleted GSCs (Fig. 4a). Transcriptional networks controlled by HIF1α and GABPA/NRF2 (two key transcription factors regulating glycolysis and mitochondria function, respectively42,43,44,45,46,47) were downregulated (Fig. 4a and Supplementary Fig. 4a–c), while other pathways (such as the SREBP-mediated lipid pathway, the Notch pathway, aminoglycan biosynthesis or MAP kinase scaffold activity) were upregulated in AMPK-silenced GSCs (Supplementary Fig. 4c). RNA and protein analysis of GSCs and tumours (Fig. 4b–d and Supplementary Fig. 4d,e) confirmed that HIF1α, GABPA and their targets were downregulated in AMPK-depleted cells. Phosphorylation of pyruvate dehydrogenase (PDH) by the HIF1α transcriptional target pyruvate dehydrogenase kinase (PDK) was also reduced (Supplementary Fig. 4e). Despite downregulation of the HIF1α target Glut1, glucose import was not affected (Supplementary Fig. 4f), probably due to compensation by the other glucose transporter Glut3 (Supplementary Fig. 4g).

Fig. 4: AMPK regulates GBM bioenergetics.
figure 4

a, Edge flux heat map showing pathways downregulated in GSCs expressing AMPKβ1 shRNA. Data was processed using NetWalker. bd, Relative expression of selective genes using qRT-PCR in GSCs (b,c) or tumours (d) expressing non-target (NT) or AMPKβ1 shRNA. Data was normalized to β-actin (n = 3; *P ≤ 0.0003, +P ≤ 0.006, **P ≤ 0.003, ++P ≤ 0.005, +*P ≤ 0.01, ***P ≤ 0.0002, #P = 0.001, ##P ≤ 0.009, +++P ≤ 0.008). ej, Extracellular acidification rate (ECAR; a measure of glycolysis) and oxygen consumption rate (OCR; a function of mitochondria) of GSCs (e,f), flank tumours (g,h) and established GBM serum lines (i,j) expressing NT or AMPKβ1 shRNA (n = 3 in e,f,i,j and n = 12 tumours per genotype in g,h; *P = 0.003, +P = 0.0008, #P = 0.0004 in e; *P = 0.008, +P = 0.0008, #P = 0.001, $P = 0.001 in f; **P ≤ 0.0001 in h; NS = non-significant in i,j. k, HPLC/mass spectrometric quantification of cellular energy levels in a GSC10 line expressing NT or AMPKβ1 shRNA (n = 3; *P = 0.03, **P = 0.01, +P = 0.001). l,m, Lactate and citrate released in media by GSCs expressing NT or AMPKβ1 shRNA (n = 3; *P = 0.01, +P = 0.03, #P ≤ 0.005). n, Kinetic flux analysis of U13C glucose by HPLC/mass spectrometry in GSCs expressing NT or AMPKβ1 shRNA (average of two independent experiments). o, Quantification of mitochondrial (Mito) mass in GSCs expressing NT or AMPKβ1 shRNA by PCR using β-actin and ND4 primers (n = 3; *P ≤ 0.0007). p, Quantification of mitochondrial complex activity in GSCs expressing NT or AMPKβ1 shRNA (n = 3; *P = 0.003, +P = 0.01). q, Electron micrographs of GSC326 expressing NT or AMPKβ1 shRNA. Scale bar, 2 μm. r, Quantification of mitochondrial number in two GSC lines expressing NT or AMPKβ1 shRNA (n = 10 cells per condition; *P ≤ 0.0002). Error bars ar mean ± s.d. Statistical significance, two-tailed t-test; n values represent independent experiments unless stated otherwise. Source data are available in Supplementary Table 4.

Functional analysis using two independent AMPKβ1 shRNAs confirmed that both glycolysis and mitochondrial respiration were reduced in AMPK-silenced GSCs (Fig. 4e,f) and tumours (Fig. 4g,h), but not in long-established GBM serum lines (Fig. 4i,j). Accordingly, energy levels (Fig. 4k) and glucose-induced lactate and citrate production (Fig. 4l,m) were also reduced. Kinetic flux analysis using 13C glucose showed reduced labelling of several metabolites of glycolysis and the tricarboxylic acid (TCA) cycle in AMPK-depleted cells (Fig. 4n and Supplementary Fig. 5a–h). Mitochondrial DNA analysis, activity assay and electron microscopy showed reduced mitochondrial mass/number and mitochondrial complex activity in AMPK-silenced GSCs (Fig. 4o–r). Although mechanisms are unclear, in some contexts PGC1α and SIRT1 have been shown to regulate mitochondrial biogenesis and respiration downstream of AMPK48. Total and acetylated PGC1α protein (a reaction regulated by the deacetylase SIRT1) remained unchanged in AMPK-silenced GSCs (Supplementary Fig. 4h,i). PGC1α overexpression did not rescue the mitochondrial defects of AMPK-silenced GSCs (Supplementary Fig. 4j–l). The SIRT1 activator SIRT1720 partially increased mitochondrial mass and oxygen consumption rate (OCR) in both WT AMPK-silenced GSCs (Supplementary Fig. 4m,n) perhaps through AMPK-independent mechanisms. Together, our findings indicate the presence of an AMPK-regulated transcriptional program that is important for GBM bioenergetics.

The HIF1α transcriptional program is downregulated in AMPK-silenced GSCs

Further molecular analyses showed that both steady-state and hypoxia-induced HIF1α protein levels were reduced in AMPK-silenced cells (Fig. 5a). HIF1α transcript level, in the presence of actinomycin D, was only slightly decreased, suggesting that HIF1α transcription and not RNA stability accounts for the decrease in mRNA (Fig. 5b). HIF1α protein levels also seemed more unstable in AMPK-silenced cells (Fig. 5c). Single-cell small-molecule RNA-FISH (sm-FISH)49 revealed that HIF1α active transcription was indeed significantly diminished in AMPK-silenced cells (Fig. 5d,e). Accordingly, HIF1α promoter activity and HIF1α DNA binding to its target genes32,33 were also reduced (Fig. 5f,g). Microvessel formation, potentially regulated by HIF1α-VEGF or other vasculogenic pathways, was also diminished in AMPK-silenced tumours (Fig. 5h,i). Consistent with previous reports32, HIF1α silencing reduced GSC viability (Fig. 5j).

Fig. 5: AMPK regulates HIF1α transcription in GSCs.
figure 5

a, WB of HIF1α in GSCs expressing NT or AMPKβ1 shRNA. PCNA was used as loading control. b, Quantification of HIF1α transcript in actinomycin-D-treated GSCs expressing NT or AMPKβ1 shRNA. Data was normalized to HPRT RNA (n = 3; *P ≤ 0.003). c, WB of HIF1α in cycloheximide (chx)-treated GSCs expressing NT or AMPKβ1 shRNA. d, Small-molecule FISH (sm-FISH) coupled with high content imaging of nascent HIF1α transcripts in GSCs expressing NT or AMPKβ1 shRNA. e, Quantification of HIF1α RNA in sm-FISH (n = 3; *P = 0.001). f, Quantification of HIF1α promoter activity in control (NT) and AMPK-silenced GSCs expressing hypoxia-responsive element (HRE)-firefly and renilla luciferase reporters (n = 3; *P = 0.007, +P = 0.01, ++P = 0.02). g, Chromatin immunoprecipitation (ChIP) using ChIP-grade HIF1α antibody and quantification of HIF1α binding to target gene promoters in control (NT) and AMPK-silenced GSCs (n = 3; *P = 0.005). h,i, Microvascular density in AMPK-silenced tumours using isolectin B4 IHC (h) and quantification (i) in GSC326 expressing NT or AMPKβ1 shRNA (n = 3 tumours per genotype). Scale bar, 100 μM. *P = 0.002, +P = 0.01. j, Viability of GSCs expressing NT or HIF1α shRNAs (n = 3; *P ≤ 0.005, +P = 0.02, **P = 0.007). k,l, WB of hydroxylated HIF1α, HIF1α and HIF2α in NT and AMPKβ1-silenced GSCs (k) and cells treated with 1 mM succinate (l). m, Viability of NT or AMPKβ1 shRNA GSCs overexpressing doxycycline (Dox)-inducible constitutively active HIF1α (n = 3; *P ≤ 0.0004, +P = 0.01). Top: WB of HIF1α in control (−Dox) and Dox-treated cells. Error bars are mean ± s.d. Statistical significance, two-tailed t-test; n, biological replicates. Source data are available in Supplementary Table 4. All WB represent data from two (c,k,i) or three (a,m) independent repeats. Unprocessed blots in Supplementary Fig. 9.

The TCA cycle, which produces succinate, was downregulated in AMPK-silenced cells. Since succinate inhibits prolyl hydroxylases50 that in turn hydroxylates and destabilizes HIF1α protein51, we tested if HIF1α is differentially hydroxylated in AMPK-silenced cells. More hydroxylated HIF1α was observed in AMPK-silenced GSCs, and the addition of succinate reduced HIF1α hydroxylation and increased HIF1α stability (Fig. 5k,l). HIF2α plays diverse roles in glioma32,52. We found that HIF2α is overexpressed in AMPK-knockdown GSCs, potentially to compensate (although insufficiently) for the loss of HIF1α (Fig. 5k). Finally, inducible overexpression of non-degradable HIF1α partially rescued viability of AMPK-silenced GSCs (Fig. 5m). Together, these results indicate that AMPK plays an important role in regulating HIF1α in GSCs.

The GABPA transcriptional program is downregulated in AMPK-silenced GSCs

Consistent with reduced levels of GABPA RNA and protein (Fig. 4c,d and Supplementary Fig. 4d,e), GABPA binding to its target genes45,46 was also diminished in AMPK-silenced cells (Fig. 6a). Transcript levels and target DNA binding of NRF1 (another transcription factor important for mitochondrial function)45 were upregulated (Fig. 6b,c), probably to compensate (although insufficiently) for the reduction in GABPA. Transcription of genes coordinated by both NRF1 and GABPA (such as COX4, mTERF and POLRMT) either did not change or increased slightly (Fig. 6d). GABPA transcribes TFAM, the transcription factor that critically regulates both transcription and replication of the mitochondrial genome45,46,47. Accordingly, silencing of GABPA reduced cell viability and respiration (Fig. 6e–g), and silencing TFAM reduced GSC viability (Fig. 6h,i). Consistent with the role of GABPA in regulating TFAM and mitochondrial function, overexpression of GABPA or TFAM partially rescued viability defects of AMPK-silenced GSCs, indicating their function downstream of AMPK (Fig. 6j,k). Since mitochondrial TCA cycle substrates were diminished in AMPK-silenced cells, we also tested if their replenishment rescue cell viability. Adding succinate but not pyruvate (as it is already present in the basal medium) partially rescued viability of AMPK-silenced GSCs. These results indicate that GABPA transcription is an important mechanism through which AMPK regulates GSC mitochondrial bioenergetics.

Fig. 6: AMPK regulates GABPA transcription in GSCs.
figure 6

a, ChIP using GABPA antibody and quantification of GABPA binding to target gene promoters in control (NT) and AMPK-silenced GSCs (n = 3; *P = 0.003, +P = 0.001). b, qRT-PCR analysis of NRF1 in GSCs expressing NT or AMPKβ1 shRNA (n = 3; *P = 0.006, **P = 0.005). Data was normalized to β-actin. c, ChIP using NRF1 antibody or nonspecific IgG and quantification of NRF1 binding to target gene promoters in cells expressing NT and AMPKβ1 shRNA (n = 3; *P = 0.003). d, qRT-PCR analysis of COX4, mTERF and POLRMT in NT and AMPKβ1-silenced GSC lines (average of two independent experiments). e,f, qRT-PCR of GABPA (e) and cell viability (f) in GSCs expressing NT or GABPA shRNA (average of two independent experiments). g, OCR and ECAR in GSCs expressing GABPA shRNA (n = 3; *P = 0.009, +P = 0.001; ns, not significant). h,i, WB of TFAM (h) and cell viability (i) in GSCs expressing NT or TFAM siRNA (n = 3; *P = 0.01, +P = 0.002, **P = 0.003). j,k, Viability of NT or AMPKβ1-expressing GSCs overexpressing GABPA (j) or TFAM (k). Top: WB of GABPA (j) and TFAM (k) with or without Dox treatment (n = 3; *P ≤ 0.001, +P = 0.009, **P = 0.005, #P = 0.03, ++P = 0.004; ***P = 0.0008). I, Viability of NT or AMPKβ1-shRNA-expressing GSCs treated with methylpyruvate (1 mM) or sodium succinate (1 mM) (n = 3; *P = 0.001, +P = 0.006). Error bars are mean ± s.d. Statistical significance, two-tailed t-test; n values represent independent experiments. Source data are available in Supplementary Table 4. Western blots represent data from two (h,j) or three (k) independent repeats. Unprocessed blots are available in Supplementary Fig. 9.

The AMPK–CREB1 axis regulates HIF1α and GABPA transcription in GSCs

We hypothesized that another transcription factor is a potential molecular link between AMPK and HIF1α/GABPA. Analysis of functional genomics data from ENCODE and Roadmap Epigenomics identified potential CREB1 binding sites in the HIF1α and GABPA promoter/enhancer regions (Supplementary Fig. 6a,b). Indeed, CREB1 chromatin immunoprecipitation (ChIP) showed significant CREB1 enrichment in the GABPA promoter and a potential enhancer distal to the HIF1α 3’ region. Importantly, this enrichment was significantly diminished following AMPK silencing (Fig. 7a). Accordingly, CREB1 knockdown reduced HIF1α promoter activity, HIF1α and GABPA transcriptional target expression, cell viability and GSC bioenergetics (Fig. 7b–f). TCGA data analysis showed that CREB1, EP300 (which augments CREB1 activity), HIF1α, GABPA and TFAM, as well as phosphorylated CREB131, are highly expressed in GBM relative to normal brain (Supplementary Fig. 7a–d). Although only CREB1 is prognostic in LGG (P = 0.002) (Supplementary Fig. 7e), all the genes mentioned are important for the biology of human cancer, including GBM31,32,33,34,53. These genes are therefore likely to be highly expressed uniformly across the large majority of GBMs compared to their variable expression in LGG, as reported for HIF1α and phosphorylated CREB129,30,31.

Fig. 7: AMPK regulates HIF1α and GABPA through CREB1.
figure 7

a, ChIP using CREB1 antibody and quantitative PCR showing CREB1 binding to promoter/enhancer regions of HIF1α and GABPA (n = 3; *P = 0.009, +P = 0.005). b, Quantification of HIF1α promoter activity in control (NT) and CREB1-silenced GSCs expressing HRE-firefly and renilla luciferase reporter plasmids (n = 3; *P = 0.01, +P = 0.008). c, qRT-PCR analysis of GSCs expressing NT or CREB1 shRNA (n = 3; *P = 0.002, +P = 0.001, #P = 0.004, $P = 0.007, **P = 0.01, ++P = 0.0005). d, Viability of GSCs expressing NT or CREB1 shRNA (n = 3); *P = 0.003, +P = 0.01). Top: WB of CREB1. e,f, ECAR and OCR in GSCs expressing CREB1 shRNA (n = 3; *P = 0.01, +P = 0.006). g,h, WB of pCREB1S133 and CREB1 (loading control) in GSCs expressing NT shRNA or two independent AMPKβ1 shRNAs. i, WB of pACC, pCREB1 and CREB1 in normal human astrocytes (NHAs) treated with AMPK activator A769662 or glucose starvation (30 min). j, IHC of pCREBS133 in tumours derived from GSCs expressing NT or AMPKβ1 shRNA. Scale bar, 100 μm. k, qRT-PCR analysis of GSC326 cells expressing control virus, CREB1S133A or CREB1S133E lentivirus (n = 3; *P = 0.01, +P ≤ 0.0003, $P = 0.002, #P = 0.02, **P = 0.03). I, WB of HIF1α and GABPA in GSCs expressing CREB1 shRNA, CREB1S133A and CREB1S133E mutants. Total CREB1 and actin are also shown. m, WB of HIF1α and GABPA in tumours derived from GSCs expressing CREB1S133A. Error bars are mean ± s.d. Statistical significance, two-tailed t-test; n values represent independent experiments. Source data are available in Supplementary Table 4. All western blots represent data from two (i,l,m) or three (g,h) independent repeats. Unprocessed blots are shown in Supplementary Fig. 9.

Exercise-induced stress activates AMPK to enhance muscle bioenergetics, and active AMPK phosphorylates CREB1 at S13354,55,56,57,58, a modification that activates CREB159. We tested if this pathway is hijacked during oncogenesis-associated stress in GBM. AMPK silencing by two independent shRNAs reduced CREB1S133 phosphorylation in human astrocytes (Fig. 7g,h), while pharmacological or physiological AMPK activation enhanced it (Fig. 7i). Accordingly, pCREB1S133 staining was also diminished in AMPKβ1-silenced tumours (Fig. 7j). To establish a more direct link between CREB phosphorylation, HIF1α and GABPA, we either silenced CREB1 or overexpressed CREB1S133A or CREB1S133E in NHAs, and examined levels of HIF1α and GABPA and their downstream glycolysis and mitochondrial targets. Several glycolytic and mitochondrial targets of HIF1α and GABPA were downregulated in CREB1S133A-expressing cells and upregulated in CREB1S133E-expressing cells (Fig. 7k). Consistent with these results, CREB1 knockdown and CREB1S133A repressed HIF1α and GABPA protein, while CREB1S133E increased HIF1α and GABPA (Fig. 7l). Accordingly, CREB1S133A-expressing tumours clearly showed reduced levels of HIF1α and GABPA protein (Fig. 7m).

Acute AMPK activation caused nuclear enrichment of pCREB1S133 (Supplementary Fig. 7f,g) in vitro, and in the tumour, nuclear pCREB1 co-localized with pAMPK (Supplementary Fig. 7h). Lastly, overexpression of AMPK α, β and γ subunits in normal astrocytes increased CREB1 phosphorylation (Supplementary Fig. 7i). However, AMPKβ1 overexpression did not significantly alter the proliferation of normal astrocytes (Supplementary Fig. 7j). Together, this evidence shows a direct link between active AMPK, CREB phosphorylation and HIF1α- and GABPA-regulated gene expression in GSCs. CREB1 is also phosphorylated at S133 by protein kinase A (PKA)59. However, AMPKβ1 shRNA reduced CREB1 phosphorylation but not that of other PKA substrates, indicating its specificity for CREB1. Importantly, the PKA activator forskolin phosphorylated PKA substrates in both NT shRNA and AMPKβ1 shRNA cells, but only CREB1 in NT shRNA cells (Supplementary Fig. 7k). The failure of forskolin to phosphorylate CREB1 in AMPKβ1 shRNA cells showed that AMPK plays a specific and dominant role in CREB1 phosphorylation in GBM.

To test if CREB1S133 phosphorylation plays a critical role in GBM pathogenesis, we expressed inactive inducible phospho-mutant CREB1S133A or phosphomimetic CREB1S133E and examined GSC viability. CREB1S133A reduced ATP levels and diminished viability (Fig. 8a,b), while CREB1S133E rescued viability defects of AMPKβ1-silenced cells (Fig. 8c). When transplanted in the brain, inducible expression of CREB1S133A significantly delayed tumourigenesis and improved survival in mice (Fig. 8d,e), indicating the significance of CREB1S133 phosphorylation in GBM.

Fig. 8: The AMPK–CREB1 transcriptional axis regulates GBM bioenergetics and tumour growth.
figure 8

a, ATP levels in GSCs expressing CREB1S133A mutant (n = 3; *P = 0.001, +P = 0.008). Top: WB of pCREB1 and CREB1 in GSCs expressing Dox-inducible CREB1S133A in the presence or absence of Dox. b, Viability of GSCs expressing CREB1S133A or control virus (n = 3; *P = 0.002, +P = 0.005). c, Viability of GSCs expressing NT or AMPKβ1 shRNA and overexpressing CREB1S133E (n = 3; *P ≤ 0.002). Right: WB of pCREB1 and CREB1 in GSCs expressing Dox-inducible CREB1S133E in the presence or absence of Dox. d, Growth of tumours expressing control virus or Dox-inducible CREBS133A in NSG mice fed with Dox-chow after tumour establishment. e, Kaplan–Meier survival data of Dox diet-fed mice with tumours expressing control virus or Dox-inducible CREBS133A (n = 6 mice per condition). f, Schematic showing the regulation of GBM bioenergetics by AMPK, which phosphorylates CREB at S133. This allows enhanced binding of CREB1 to its targets HIF1α and GABPA, which in turn augment the transcriptional program of glycolysis and mitochondrial biogenesis to regulate GBM viability and growth. Error bars are mean ± s.d. Statistical significance in these experiments (except e) was assessed using a two-tailed t-test; n values represent independent experiments. Source data are available in Supplementary Table 4. All western blots represent data from two (a,c) independent repeats. Unprocessed blots are available in Supplementary Fig. 9.

Systemic deletion of AMPK is tolerated in adult mice

We found that various types of cancer-associated stress chronically activate AMPK and that GSCs hijack a stress response pathway conserved in normal cells, using it to survive. This suggests that AMPK is an appropriate target in GBM, and so the effect of deleting AMPK in GBM mouse models and the effects of whole-body AMPK inhibition to check systemic tolerance warrant investigation. We did not use compound C, the only reagent sometimes used as an AMPK inhibitor because of its nonspecificity. Remarkably, adult mice with systemic AMPK knockout did not show overt metabolic phenotype and lived a normal life span (Supplementary Fig. 8) underscoring the potential use of AMPK inhibitors in the treatment of GBM and perhaps other human cancers where AMPK is activated.


In this study, we show that oncogenesis-associated stress chronically activates the cellular energy sensor AMPK in GBM. Our results are consistent with the high expression of active AMPK in glioma and its requirement in proliferation of oncogenic mouse astroglia11,12,60. Why the established GBM serum lines remained viable and formed tumours independent of AMPK is unclear. Decades of culture could have altered the genetic, epigenetic and metabolic landscape of these lines61, allowing them to adapt and evolve AMPK-independent growth and survival pathways. AMPK was also found to enhance glioma cell viability by inducing lipid import in vitro62. While this remains a possibility in vivo, our primary GSC lines are grown in serum-free media containing only two essential fatty acids, precluding their dependence on serum-derived non-essential fatty acids in vitro. In contrast to our results, the metabolic stressor AICAR, which activates AMPK, inhibited growth of the established glioma serum line U87MG16. While we could not directly compare their results with ours, we and others have shown that AICAR suppresses cancer including glioma through multiple AMPK-independent mechanisms12. Previous pharmacological approaches are not fully reconcilable with later genetic studies, and there is more recent appreciation that AMPK can play a context-dependent function in cancer3. Consistent with tissue- and species-specific effects of AMPK, AMPKα1 suppressed lymphomagenesis in a Myc mouse model10, while active AMPK was required for tumour growth in other models17,18,19,20,21,22,41,63. Interestingly, in a functional metabolic screen, AMPKβ1 was identified as a critical prostate cancer cell survival gene37, which together with our studies underscores a particularly important function of the AMPKβ1-containing complex in cancer cell survival and tumour growth.

While the role of AMPK in cancer is complex3, several mechanisms are known by which it checks cellular growth and proliferation8,9,64,65. In contrast, very few mechanisms are known by which it promotes survival and growth in other contexts. Through comprehensive genetic analysis, we here show a mechanism by which AMPK supports tumour bioenergetics, growth and survival in human glioblastoma. We show that by phosphorylating CREB1, which occurs abundantly in GBM31, AMPK enhances HIF1α and GABPA transcription to support GBM bioenergetics (Fig. 8f). HIF1α can attenuate mitochondrial respiration through its transcriptional target PDK, which phosphorylates and inhibits PDH66. Therefore, HIF1α downregulation and reduced PDH phosphorylation was expected to augment mitochondrial respiration in AMPK-silenced cells. Perhaps downregulation of GABPA/TFAM, which are critical to mitochondrial transcription and replication, overrides the stimulatory effect of hypophosphorylated (active) PDH and diminishes mitochondrial function in AMPK-silenced GSCs.

The evidence that AMPK phosphorylates CREB at S133 (also a PKA site) in vitro and in vivo is strong54,55,56,57,58. AMPK also strongly phosphorylated a chimaeric peptide of CREB1 and ACC (a bonafide AMPK substrate) that lacked the AMPK phosphorylation site in ACC (S78), indicating specificity of AMPK for CREB1. In fact, the Vmax of AMPK for the CREB1 peptide is comparable to that for the ACC peptide (12.3 pmol min−1 versus 9.3 pmol min−1)58. High levels of nuclear CREB1 are expressed in GBM31. We speculate that the AMPKβ1 complex, which is enriched in the nucleus67, phosphorylates CREB1 and potentially other nuclear targets20. It remains to be seen what specific role AMPK or the CREB1 transcriptional program (that can be regulated independent of AMPK) plays during the step-wise evolution of human cancer.



The following reagents were used: oligomycin, FCCP, antimycin, rotenone, tunicamycin, thapsigargin, camptothecin B, hydroxyurea, puromycin, doxycycline, DAPI, hydrogen peroxide, formamide, STO-609, KU55933, NAC, TOFA, PKI, forskolin, methyl pyruvate, sodium succinate, cycloheximide and chloroquine, glucose, resveratrol and DMSO (all from Sigma), G418 (Invitrogen), compound 991 (a gift from D. Carling, Imperial College, London), A769662 and rapamycin (LC Laboratories) and SIRT1720 (Apen Biotechnology). The VectaStain ABC kit was from Vector Laboratories. D-Luciferin was from Perkin Elmer. Compound C was from EMD Millipore. Isolectin B4 (IB4; cat no. DL-1207) (Vector Laboratories) was a gift from E. Boscolo, Cincinnati Children’s Hospital, Ohio).

Cell culture

Human primary glioblastoma lines were established from freshly resected tumours either in our laboratory under a University of Cincinnati institutional review board (IRB) approved protocol, or obtained from our collaborators (University of Alabama). Informed consent was taken from all subjects. Cells were maintained as suspension cultures in UltraLow attachment plates in glioma stem cell (GSC) medium that contained glucose-free DMEM-F/12 supplemented with 5 mM glucose, B27, EGF (10 ng ml−1), bFGF (10 ng ml−1), GlutaMAX and heparin (5 mg ml−1). Adherent glioma serum lines (U87, T98G and A172) and HEK293T cells were purchased from ATCC and were not re-authenticated. The U87 line was recently identified as another GBM line. We completed our experiments long before this misidentification was published. These lines were maintained in glucose-free DMEM supplemented with 5 mM glucose and 10% FCS and were used in our earlier publication. Normal human astrocytes (NHAs) were purchased from Lonza Group Ltd. and immortalized with retroviral expression of large T-antigen (pBabe-puro TcDNA, no. 14088, Addgene). NHAs were maintained in glucose-free DMEM supplemented with 5 mM glucose and 10% FCS. Culture conditions of NHAs did not influence the effect of AMPK shRNA. We switched NHA culture medium temporarily to GSC medium and tested the effect of AMPK shRNA, and the results remained consistent to what we observed in FCS-containing medium. All lines were checked for mycoplasma every other week using mycoplasma-specific PCR. Primary GSC lines and NHA were analysed by STR analysis. For proliferation and viability analysis, both direct counting using Trypan blue method and a fluorescence-based method (CellTiter-Fluor; Promega) were used.

Western blotting

Western blot (WB) analysis was carried out following standard methods as described previously12. Each experiment was successfully carried out two to three times. Antiboides were validated by using positive and negative control tissues and cells. Antibody information is provided in Supplementary Table 1.

Immunohistochemistry (IHC)

Fresh frozen human tissues and paraffin sections were obtained from the tissue repository at the University of Cincinnati under a UC IRB-approved protocol. Mouse high-grade glioma tissue was obtained from a transgenic glioma mouse model35. IHC was done as described previously12. Mice were anaesthetized and perfused intracardially with 4% PBS. Tumours were then dissected out and processed for paraffin embedding and sectioning. Antibody validation was done using multiple positive and negative control tissues and cells. Antibody information is provided in Supplementary Table 1.


After 72 h of shRNA infection, cells were maintained at hypoxia in a controlled atmosphere chamber (Don Whitley Scientific) with a gas mixture containing 0.5% to 1% O2, 5% (vol/vol) CO2 and 94% (vol/vol) N2 at 37 °C for the indicated time depending on the experiment. Cells were harvested for downstream processing inside the chamber.

Glucose import

A glucose uptake assay was performed using 100 μM 2-(N-(7-nitrobenz-2-oxa-1,3-diazol-4-yl)amino)-2-deoxy-D-glucose (2-NBDG; Invitrogen), followed by FACS detection (BD Bioscience). After 72 h of shRNA infection, 105 cells ml−1 were incubated with 2-NBDG (100 μM) for indicated time points, and data from 10,000 single-cell events were collected. Original FACS plots are provided in the Supplementary Source File.

Quantification of superoxide anion

The intracellular levels of superoxide (O2) were measured with MitoSOX Red (Invitrogen). 1 × 105 cells were plated in a 12-well plate, treated with MitoSOX Red (5 μM) for 10 min at 37 °C, washed with PBS and analysed in FACScan flow cytometers. Original FACS plots are provided in the Supplementary Source File.

Lentivirus preparation and production of stable shRNA-expressing cell lines

Several independent shRNA sequences were screened for each of the used human target genes, and the sequences which exhibited maximal knockdown were used for the study. Most shRNA clones (in pLKO.1 plasmid) were from the Sigma Mission RNAi shRNA library. The AMPKβ2 shRNA clone was purchased from OriGene. pLKO.1-puro scrambled (NT) shRNA (Sigma) was used as a negative control. ShRNAs were prepared in 293T cells as before12. Overexpression clones in pInducer 20 were selected with G418 (700 μg m−1). Efficacy of knockdown/overexpression was assayed by WB or quantitative with reverse transcription PCR (qRT-PCR). The sequences that exhibited maximal knockdown/overexpression were used for the study. ShRNA sequences are provided in Supplementary Table 2.

CRISPR–Cas9 gene knockout

CRISPR–Cas9 allows for specific genome disruption and replacement in a flexible and simple system resulting in high specificity and low cell toxicity. Three guide RNAs (gRNAs) were designed to knockout target sequence of the AMPKβ1 locus. Donor and gRNAs were purchased from Blue Heron Biotech, WA, USA. The target sequences cloned into the pCas-Guide vector are provided in Supplementary Table 2. The efficiency of the CRISPR–Cas9 was assessed by immunoblot assay.

Mass spectrometry of metabolites

Profiling of glycolysis, the TCA cycle and nucleotide metabolites was carried out by solvent extraction followed by liquid chromatography mass spectrometry (LC-MS). Cells were treated with 5 mM D-glucose (U13C6, CLM-1396, Cambridge Isotope) for different time points up to 90 min. Metabolites were detected using LC-tandem-quadrupole mass spectrometry. The investigators were blinded to allocation during experiments and outcome assessment.

Bioenergetics experiments

ATP quantification in GSCs (1 × 105 cells) was done using the ApoSENSOR ATP Luminescence Assay Kit (BioVision). ECAR and OCR were analysed using the XFe96-Analyzer (Seahorse Biosciences) as before12. GSCs were attached to the wells with Cell Tak (Corning). Tumour tissues from flank xenografts were surgically harvested and one millimetre uniformly thick sections were incised using metal tissue matrices. From these brain sections, 2 mm diameter circular pieces were cut out using biopsy punches (Miltex, York, PA). Tissues of equal weight were analysed in the Seahorse analyzer as described67. Extracellular lactate and citrate were measured from culture supernatants using lactate and citrate assay kits (Biovision, Milpitas, CA).

Plasmids and clonings

Mouse β1 and β2 open reading frames (ORFs) were purchased from Origene and subcloned into lentiviral vectors CMVTV and FCIV, respectively. Dominant-negative AMPKα2 (a gift from R. Jones, McGuill University, Canada) was cloned into FCIV lentiviral vector67. PLenti-PGK-KRAS4A (G12V) cat. no. 35634 was from Addgene and pLV EGFRvIII Hygro was gifted by F. Furnari (Cancer Research, San Diego, California). For overexpression studies, genes were subcloned into the doxycycline-inducible lentiviral vector pInducer 20. Human AMPKβ1 was also subcloned into cumate-inducible SparQ all-in-one lentivirus (gift from System Biosciences, Paolo Alto, CA). Human HIF1α (Acc no. KR710294.1), GABPA (Acc no. BC035031.2) and TFAM (Acc no. BC126366.1) were PCR amplified to add attB sites using TFAM, GABPA and HIF1α primers. Primer sequences are provided in Supplementary Table 3.

The PCR fragments were gel purified and transferred into pDONR221 and pInducer20 using gateway recombination as above. Sequences of all clones were confirmed by Sanger sequencing at the CCHMC DNA Core.

To generate mCreb1 S133A, mutagenesis was performed on mCreb1 (Acc no. BC021649) in pCMV-SPORT6 with the Phusion Site Directed Mutagenesis Kit (Thermo Scientific, catalogue no. F-541). The open reading frame of wt mCreb1 and mCreb1 S133A was PCR amplified to add attB recombination sites using Herculase II Fusion DNA Polymerase (Agilent Technologies, catalogue no. 600675-51).

The PCR fragments were gel purified using the QIAquick Gel Extraction Kit (Qiagen, cat. no. 28704) and transferred by BP recombination reaction into pDONR221 (Invitrogen, cat. no. 12536-017) using Gateway BP Clonase II Enzyme Mix (Life Technologies, catalogue no. 11789-020) and subsequently by LR recombination reaction into pInducer20 using Gateway LR Clonase II Enzyme Mix (Life Technologies, catalogue no. 11791-020). To generate mCreb1 S133E, mutagenesis was performed directly on mCreb1 in pInducer20 using the Phusion Site Directed Mutagenesis Kit and primers 5’-AAGGAGGCCTGAGTACAGGAAAATTT-‘ and 5’-GAAAGGATTTCCCTTCGTTTTTGG-3’

For cloning into cumate-inducible SparQ vector, mouse AMPKβ1 cDNA was amplified from pcMV-Sport6 mouse AMPKβ1. Restriction sites were added by PCR using the following primers: forward 5′-GCTAGCATGGGCAACACGAGC-3′ and reverse 5′-GCGGCCGCTCATCATATCGGCTTGTAGAGGAGGGT-3′. The PCR product and the recipient plasmid SparQ All-in-one Cumate Switch Vectors (System Biosciences Cat no. QM812B-1) were digested with NheI and NotI. The ligation reaction was transformed into GC10 Competent Cells (Sigma), and plasmid DNA was purified with QIAGEN Hi Speed Plasmid Midi Kit (Qiagen).

Sirt3 Flag (plasmid no. 13814) and pcDNA4 myc PGC-1α (plasmid no. 10974) were purchased from Addgene.

HIF1α promoter activity assay

Cells were seeded onto 6-well plates, co-transfected with the HRE luciferase reporter and control pSV40-Renilla (gifts from G. Huang) for 24 h. Transfection medium was washed off and cells were further incubated for 24 h. Cell extracts were analysed using the Dual-Luciferase Assay System (Promega).

Quantification of mitochondrial mass

Cells were treated with shRNA expressing viral particles and after 72 h, genomic DNA was prepared using QIAamp DNA kit (Qiagen). mtDNA and genomic DNA were detected using ND4 and β-actin primers, respectively. PCR was used to compare fold change in total mitochondrial content.

RNA extraction and qRT-PCR

Two micrograms of total RNA (RNeasy kit, Qiagen) was used for first strand cDNA synthesis with oligo-dT primers and Superscript II Reverse Transcriptase (Invitrogen). qRT-PCR was performed using SYBR Green PCR Master Mix (Applied Biosystems) and Quantitect primers (Qiagen) in an ABI PRISM 7900 Sequence Detection System (Applied Biosystems). Relative mRNA expression was calculated using the comparative Ct method after normalization to a loading control. Samples were run in triplicates with a primer-limited probe for the reference gene (actin or HPRT). Primers are provided in Supplementary Table 3.

Orthotopic xenograft

All animal procedures were carried out in accordance with the Institutional Animal Care and Use Committee (IACUC)-approved protocol of Cincinnati Children’s Hospital Medical Center (CCHMC; Cincinnati, OH). The study is compliant with all relevant ethical regulations regarding animal research. Animals were monitored daily by animal care personnel. For orthotopic implantation, 1 × 104 primary human GBM cells were stereotactically injected into the left striatum of NOD-SCID IL2Rgnull mice. Both male and female mice were used. Randomization of mice for this study was not necessary. For in vivo bioluminescent imaging, luciferase-expressing cells were established by infection with plenti-CMV-luc viral particles (a gift from S. Wells, Cincinnati Children’s Hospital, Ohio) and tumour growth was monitored using IVIS 200 system. Five minutes before bioluminescence imaging, mice were anaesthetized and injected (intraperitoneally) with luciferin (150 mg kg−1) and imaged using IVIS (Xenogen). All mice were euthanized following observation of lethargy and/or neurologic symptoms. For flank xenografts, 2 × 106 cells were injected subcutaneously and imaged using IVIS. These tumours were harvested for ex vivo bioenergetics experiments. In mice that required cumate injections, animals were injected with cumate (150 mg kg−1) from the day of orthotopic cell implantation every alternate day until animals were euthanized. The investigators were not blinded to allocation during experiments and outcome assessment.

Culture of human tumour cells from orthotopic xenografts

Tumour cells were isolated from NT shRNA or AMPKβ1 shRNA tumours and passaged once. Purity of human cells were determined by PCR with human and mouse-specific DNAPolE and intracisternal A particle primers, respectively (not shown). Determination of lentiviral copy number was determined by PCR showed a lower AMPKβ1 shRNA lentiviral integration compared to NT shRNA (3.15 ± 0.62 copies of NT versus 1.33 ± 0.31 copies of β1 shRNA).

Generation of whole-body AMPK-knockout mice

AMPKα1−/−, α2lox/lox mice (BL/6; gifts from B. Viollet, INSERM, Institut Cochin, Paris, France) and Rosa26–CreER (BL/6; gift from A. Kumar (Cincinnati Chilren’s Hospital, Ohio); originally from NCI), were crossed. AMPKβ1lox/lox mice obtained from the Sanger Center and AMPKβ2−/− mice that were generated in our lab (and crossed for >10 generations with BL/6 mice) were also crossed with Rosa26–CreER mice. Mice were injected with tamoxifen (225 μg g−1 body weight, intraperitoneal) at six months of age every day for three consecutive days. After one week, tissues were harvested (n = 2 mice pergenotype). Recombination efficiency was determined by western blot using AMPKα and AMPKβ antibodies. Body weight was recorded once a month and death records were used to generate survival plot. Randomization of mice for this study was not done. Genotyping primers are provided in Supplementary Table 3.

Electron microscopy

Cell pellets were fixed in 3% glutaraldehyde/0.2 M sodium cacodylate buffer pH 7.4 for at least 1 h at 4 °C. After fixation, the samples were washed 3× with cacodylate buffer and post-fixed with 1% osmium tetroxide/0.2 M sodium cacodylate buffer pH 7.4. After fixation, cells were washed 2× with cacodylate buffer, followed by 1× with ddH2O. Samples were gently resuspended in 1.5% agarose (type IX ultra-low gelling) and centrifuged at 1,000g for 5 min and processed as described previously68. All images were taken using a 120 kV transmission electron microscope (Hitachi, H-7650, V01.07, Tokyo, Japan)

Immunofluorescence microscopy

Tissue processing for frozen sections was done as described previously67,68. Fluorescent images were taken on a Nikon AZ-100 multizoom microscope equipped with a Nikon DS-Ri1 camera. Confocal images were taken in Nikon C2 confocal microscope. pACC and IBA1 were both rabbit antibodies and so a tyramide-based signal amplification was used (tyramide amplification kit; Thermo Fisher T20932). Antibody information is provided in Supplementary Table 1.

Mitochondrial complex activity

Cells were sonicated in 0.5 ml ice-cold 5 mM KH2PO4 (pH 7.5), 1% digitonin, and then used for electron transport chain (ETC) enzyme assays. ETC enzymes were assayed at 30° C using a temperature-controlled spectrophotometer Shimadzu UV-1600. Activity of complex I (NADH:CoQ reductase) was measured in 5 mM KH2PO4 (pH 7.5), 5 mM MgCl2, 0.24 mM CoQ1, 0.5 mM KCN, 1 mg ml−1 BSA and 2.4 μg ml−1 antimycin A. The reaction was initiated with 0.02 mM NADH and reduction of absorbance at 340 nm was recorded with spectrophotometer before and after addition of rotenone (final concentration 2 μg ml−1). Activity of complex I + III (NADH:cytochrome c reductase) was measured in 5 mM KH2PO4 (pH 7.5), 5 mM MgCl2, 0.24 mM CoQ1, 0.5 mM KCN, 1 mg ml−1 BSA and 0.12 mM cytochrome c (oxidized form). Reaction was initiated with 0.02 mM NADH and increase of absorbance at 550 nm was recorded with spectrophotometer before and after addition of antimycin A (final concentration 2 μg ml−1). Activity of complex IV (cytochrome c oxidase) was measured in 50 mM KH2PO4 (pH 7.5), 2 μg ml−1 rotenone and 0.03 mM reduced cytochrome c at 550 nm. Activity of complex V was measured in 50 mM Tris (pH 8.0), 5 mg ml−1 BSA, 20 mM MgCl2, 50 mM KCl, 15 μM FCCP, 5 μM antimycin A, 10 mM phosphoenol pyruvate, 2.5 mM ATP, 2 U ml−1 of lactate dehydrogenase and pyruvate kinase, and 0.02 mM NADH. Reaction was initiated by adding cell lysate followed by reduction of absorbance at 340 nm before and after addition of 2 μM of oligomycin. Citrate synthase (CS) assay media contained 0.1 mM 5,5′-dithiobis (2-nitrobenzoic acid); 3-carboxy- 4-nitrophenyl disulfide (DTNB), 0.25% Triton X-100, 0.5 mM oxaloacetate, 0.31 mM acetyl CoA, 50 mM Tris-HCl, pH 8.0. CS activity was calculated by increasing absorbance at 412 nm using extinction coefficient for TNB 13.6 mM−1 × cm−1.


One microgram of total RNA from NT or AMPKβ1 shRNA expressing GSCs was prepared three days after lentivirus transduction and was used for mRNA library preparation. Completed libraries were sequenced on an Illumina HiSeq2000 in Rapid Mode, generating 20 million or more high-quality 50-base-long single end reads per sample. RNA-seq analysis was based on the TopHat/Cufflinks pipeline. Data was processed through NetWalker69, an application platform that allows interactive comparative analysis of most active networks and functional processes. The reference annotation used was based on the UCSC Known Genes table. This method allows accurate quantification of expression of all transcripts, known or novel. BAM files have been deposited to GEO (GSE82183). The investigators were blinded to allocation during experiments and outcome assessment.

Computational identification of putative CREB1 binding sites

Analysis of functional genomics data from ENCODE and Roadmap Epigenomics identified potential CREB1 binding sites in the HIF1α and GABPA promoter regions. Since no ChIP-seq datasets exist describing CREB1 binding in cell types relevant to this study, we devised a computational method for identifying likely CREB1 binding sites in relevant cell types proximal to HIF1α and GABPA. We first compiled datasets indicative of likely regulatory regions from ENCODE and Roadmap Epigenomics, including DNase-seq, ChIP-seq for specific transcription factors, and models combining specific histone marks into likely regulatory states70. We restricted our analysis to experiments performed in cell types relevant to this study: glial cell lines (U87 and NH-A), glioblastoma cell lines (D54 and M059J), neuronal (PFSK-1 and T98G) cell lines, neuronal stem cells, and cortex and ganglion eminence-derived neurospheres. We identified likely regulatory regions located within 100 kb of either gene by taking the union of the genomic coordinates covered by these datasets. Next, for each of these putative ‘relevant’ regulatory regions, we restricted our attention to regions containing ChIP-seq peaks for CREB1 in any cell type. The resulting regions are therefore first bound by CREB1 in at least one experiment, and second, likely regulatory regions in relevant cell types. Using this approach, we identified three regions putatively bound by CREB1: one in the promoter of GABPA and two in HIF1α regulatory regions. We designed primers to capture the centre of each CREB1 ChIP peak.

Chromatin immunoprecipitation assays (ChIP)

ChIP assays were performed using 500 μg of chromatin. Chromatin was sonicated to fragments of ~500 bp and immunoprecipitated using 1 μg of ChIP-grade antibodies: HIF1α (NB100-134, Novus Biological), GABPA (sc-22810×, Santa Cruz Biotechnology), CREB1 (9104, Cell Signaling Technology) and with irrelevant IgG antibodies of mouse and rabbit origin and recovered using protein A/G magnetic beads (from Magna ChIPT A/G, Millipore). The precipitated DNA was amplified by real-time qPCR, using primer sets designed to amplify regions of the target genes. ChIP primer sequences are provided in Supplementary Table 3.

Single-molecule RNA FISH

Control or AMPKβ1 shRNA expressing GSC10 cells were grown on no. 1.5 cover glass and fixed with 3.7% formaldehyde in 1× PBS for 10 min at room temperature. Fixed cells were washed twice with 1× PBS, permeabilized with 70% EtOH for 2 h at 4 °C, then treated with wash buffer (2× SSC, 10% formamide) for 5 min at room temperature. Fixed, permeabilized cells were then hybridized to Stellaris sm FISH probes against human HIF1α conjugated with Quasar 570 dye in hybridization buffer (100 mg ml−1 dextran sulfate and 10% formamide in 2× SSC) overnight at 37 °C. Hybridized cells were washed in wash buffer 30 min at 37 °C, treated with DAPI, and mounted with Vectashield. Cells were imaged on a Nikon A1 laser scanning confocal microscope equipped with GaAsP detectors using 561 nm excitation for Quasar 570 dye and 405 nm excitation for DAPI. A 100× NA 1.45 objective was used for imaging and cells were sampled at Nyquist resolution (0.12um pixel size). Z-stacks of ~5 μm were acquired to capture the entire cell. Total number of transcripts per cell was quantified using Bitplane Imaris. Cells were analysed using the ‘Spots’ algorithm with a spot size of the diffraction limit (0.280um).

Statistics and reproducibility

For all in vitro and ex vivo experiments, three to ten technical replicates were used. Each experiment was repeated successfully two to three times as indicated in the individual figure legends. For in vivo mouse orthotopic xenograft studies, 4–8 mice per genotype were used, and for body weight and survival analysis of wild-type and AMPK-KO mice, 12 mice per genotype were used. Sample size was chosen with consideration to ensure adequate statistical power to detect prespecified effects. GraphPad Prism software was used to generate and analyse survival plot, and R was used to generate box plots from TCGA data. P values were generated using a two-sided t-test to calculate statistical significance with P < 0.05 representing a statistically significant difference. No statistical method was used to predetermine sample size. Kaplan–Meier analysis with log-rank posthoc test was used for survival studies. There was no need to exclude mice from analysis except the few that died during surgical transplantation of tumour cells. The number of indicated mice represents the total number of mice used and processed for each experiment. Because investigators were aware of the cell genotypes which they themselves transplanted in mice, there was no option for them to remain blinded to allocation for the in vivo experiments. The investigators were blinded to allocation for IHC analyses. For orthotopic xenograft studies, mice were euthanized at the ethical endpoint when they failed to meet the predetermined CCHMC IACUC quality-of-life guidelines. No mice that completed in vivo studies were excluded from analyses. There are no limitations in reproducibility for experiments.

Reporting Summary

Further information on experimental design is available in the Nature Research Reporting Summary linked to this article.

Data availability

Statistical source data is available in Supplementary Table 4. RNA-seq BAM files have been deposited to GEO (GSE82183). All other supporting data of this study are available from the corresponding author on reasonable request.