The AMP-activated protein kinase (AMPK) is a sensor of cellular energy status that regulates cellular and whole-body energy balance. A recently reported crystal structure has illuminated the complex regulatory mechanisms by which AMP and ADP cause activation of AMPK, involving phosphorylation by the upstream kinase LKB1. Once activated by falling cellular energy status, AMPK activates catabolic pathways that generate ATP whilst inhibiting anabolic pathways and other cellular processes that consume ATP. A role of AMPK is implicated in many human diseases. Mutations in the γ2 subunit cause heart disease due to excessive glycogen storage in cardiac myocytes, leading to ventricular pre-excitation. AMPK-activating drugs reverse many of the metabolic defects associated with insulin resistance, and recent findings suggest that the insulin-sensitizing effects of the widely used antidiabetic drug metformin are mediated by AMPK. The upstream kinase LKB1 is a tumour suppressor, and AMPK may exert many of its antitumour effects. AMPK activation promotes the oxidative metabolism typical of quiescent cells, rather than the aerobic glycolysis observed in tumour cells and cells involved in inflammation, explaining in part why AMPK activators have both antitumour and anti-inflammatory effects. Salicylate (the major in vivo metabolite of aspirin) activates AMPK, and this could be responsible for at least some of the anticancer and anti-inflammatory effects of aspirin. In addition to metformin and salicylates, novel drugs that modulate AMPK are likely to enter clinical trials soon. Finally, AMPK may be involved in viral infection: downregulation of AMPK during hepatitis C virus infection appears to be essential for efficient viral replication.
The AMP-activated protein kinase (AMPK) was originally discovered as apparently distinct protein kinase activities that phosphorylated and inactivated two key enzymes of lipid biosynthesis, that is, acetyl-CoA carboxylase, involved in fatty acid synthesis , and 3-hydroxy-3-methylglutaryl-CoA reductase, involved in sterol synthesis . In 1987, we reported that these were functions of a single protein kinase that was activated allosterically by AMP, as well as by phosphorylation by a distinct upstream kinase ; at that time, we named it the AMP-activated protein kinase (AMPK) .
Since these early studies, it has become clear that AMPK has dozens of physiological targets and is a crucial regulator of energy balance, both at the cellular and at the whole-body levels. Given the crucial nature of energy balance in the function of cells and organisms, it is not surprising that the AMPK system has many implications for human health, with roles in diverse disorders such as heart disease, diabetes, cancer, inflammatory disorders and viral infection. Some drugs that are already widely used, including metformin and salicylates, now appear to act in part by activating AMPK, and several novel AMPK-activating drugs are under development. The purpose of this review is to discuss the implications of recent discoveries regarding the AMPK system in the development and treatment of human disease.
AMP-activated protein kinase: regulation by adenine nucleotides and calcium ions
AMP-activated protein kinase appears to exist throughout eukaryotes as heterotrimeric complexes composed of a catalytic α subunit and regulatory β and γ subunits  (Fig. 1, top). In mammals, each subunit is present as multiple isoforms (α1, α2; β1, β2; γ1, γ2, γ3) encoded by distinct genes. The site that is phosphorylated by upstream kinases to cause AMPK activation was identified within the rat α2 isoform as Thr172 . Phosphorylation of this site and of the critical site on the downstream target acetyl-CoA carboxylase [(ACC); i.e. Ser79 in rat ACC1 ], both usually monitored using phosphospecific antibodies, are now widely used as biomarkers for AMPK activation.
What is the physiological significance of activation of AMPK by AMP? The major source of AMP in the cell is the reaction catalysed by adenylate kinase (2ADP ↔ ATP + AMP), which appears to be maintained close to equilibrium in most eukaryotic cells. In unstressed cells, catabolism maintains the ATP : ADP ratio at around 10 : 1, and this drives the adenylate kinase reaction towards ADP, so that AMP is maintained at low levels. However, if the cells experience metabolic stress that causes the ATP : ADP ratio to fall, the adenylate kinase reaction will tend to be displaced towards AMP. Because the AMP concentration starts at such a low level, the changes in stressed cells are always much larger than the changes in ATP or ADP . Cellular AMP concentration is thus a sensitive indicator of energy stress.
The principal upstream kinase phosphorylating Thr172 was identified in 2003 to be a complex containing the protein kinase LKB1 [9-11]. This was an exciting discovery because LKB1 had previously been identified as the product of a tumour-suppressor gene mutated in patients with Peutz–Jeghers syndrome, an inherited susceptibility to cancer , although its downstream targets had not been identified. These findings introduced the first clear link between AMPK and cancer (see below for further discussion).
Binding of AMP to AMPK causes >10-fold allosteric activation and also promotes net phosphorylation of Thr172, both by stimulating phosphorylation by LKB1 and by inhibiting dephosphorylation by protein phosphatases . The latter effect of AMP, but not the other two, is mimicked by ADP , albeit only at 10-fold higher concentrations . All three effects are antagonized by ATP, but concentrations of AMP observed in stressed cells cause allosteric activation and promote phosphorylation even in the presence of physiological concentrations of ATP .
Thr172 can also be phosphorylated by the Ca2+/calmodulin-dependent kinase kinases (CaMKKs), especially CaMKKβ [14-16], which represents an alternate upstream pathway by which AMPK can be activated by increases in intracellular Ca2+ in the absence of changes in AMP. There are conflicting findings concerning whether phosphorylation by CaMKKβ is promoted by AMP [8, 17], but as AMP binding inhibits Thr172 dephosphorylation, low concentrations of the AMP and Ca2+ can nevertheless act in an additive manner . Activation of AMPK by CaMKKβ is now known to occur in many physiological situations, including in depolarized hippocampal neurons , T cells activated by antigens  and cells treated with agonists that activate G protein-coupled receptors linked (via Gq/G11) to release of intracellular inositol trisphosphate and hence Ca2+. Examples of these agonists include thrombin acting via protease-activated receptor-1 in endothelial cells  and ghrelin acting via growth hormone secretagogue receptor-1 in hypothalamic neurons .
In summary, binding of AMP to AMPK promotes activation of the kinase by three complementary mechanisms: (i) allosteric activation, (ii) promotion of phosphorylation of Thr172 by LKB1 and (iii) inhibition of dephosphorylation of Thr172 by protein phosphatases, which in contrast to the other two mechanisms is mimicked by ADP. Phosphorylation of Thr172 can also occur in response to an increase in intracellular Ca2+. Next, I will discuss how structural studies of AMPK have illuminated this complex regulatory mechanism.
Structure of the AMPK complex
A crystal structure of a complete human α2β1γ1 complex was recently reported . The complex had been phosphorylated at Thr172 and crystallized in the presence both of the classical allosteric activator AMP and a novel pharmacological activator called ‘991’, and was therefore in a fully active conformation, although the nonspecific kinase inhibitor staurosporine was also included during crystallization. The layout of all of the major globular domains was clear, although some linking peptides were not fully resolved (Fig. 1).
The catalytic kinase domain (α-KD) is located at the N-terminal end of the α subunit and has the archetypal kinase domain structure, with small and large lobes joined by a hinge. As expected, the inhibitor staurosporine is located in the cleft between these lobes, where it occupies the site used by ATP during catalysis. Although a peptide substrate is not present in this structure, previous biochemical studies had suggested that the peptide sequence N-terminal to the phosphorylated residue on a substrate binds in a groove in the surface of the large lobe, placing the phosphoacceptor adjacent to the γ-phosphate of bound ATP . The α-KD is immediately followed by a small auto-inhibitory domain (α-AID), so-called because α-KD : α-AID constructs are much less active than those containing the α-KD alone ; reorientation of the position of the α-AID relative to the α-KD is likely to be involved in the activation mechanism . The C-terminal end of the α-AID marks the boundary between what can be termed the ‘catalytic module’ (top left in the lower part of Fig. 1) and the rather discrete ‘nucleotide-binding module’ (bottom right in Fig. 1), with the latter containing the C-terminal domains of the α and β subunits (α- and β-CTD) as well as the entire γ subunit. The critical phosphorylation site, Thr172, lies in the cleft between these two modules, and access of protein phosphatases to Thr172 would be restricted by the close proximity of the α-CTD in the active conformation shown in Fig. 1.
The β subunits contain two conserved globular domains, the carbohydrate-binding module (β-CBM) and the above-mentioned β-CTD. The β-CTD acts as the core of the heterotrimeric αβγ complex, forming a bridge between the α-CTD and the γ subunit. Although the connections between the β-CBM and the β-CTD are not fully resolved, the former lies on the opposite side of the kinase domain to the latter, with 991 and another AMPK activator A-769662 binding in the cleft between the β-CBM and the small lobe of the kinase domain. These two compounds are synthetic ligands derived from high-throughput screens designed to detect AMPK activators, and their binding causes both allosteric activation and inhibition of Thr172 dephosphorylation. The effects of their binding may be transmitted to the kinase domain via the ‘C-interacting helix’ on the β subunit, which immediately follows the β-CBM and which interacts with the C-helix on the small lobe of the kinase domain. On the opposite side of the β-CBM from the 991-binding site is the glycogen-binding site. Although this is known to be responsible for binding of AMPK to the surface of glycogen particles [25, 26], its exact physiological role remains uncertain.
The γ subunits contain an N-terminal region involved in interaction with the β-CTD, and four tandem repeats, termed CBS1 to CBS4, of a sequence motif known as a cystathione β-synthase (CBS) repeat, each represented in a different colour in Fig. 1. CBS repeats occur, usually as just two tandem repeats, in other human proteins where they often bind ligands containing adenosine . It is intriguing that mutations in conserved residues within CBS repeats disrupt ligand binding, and that several of these (including those in the AMPK-γ2 subunit, see below) cause inherited diseases in humans . The four CBS repeats in the AMPK-γ subunits assemble in a pseudosymmetrical manner to generate a flattened disc, with four clefts in the centre where ligands might bind. However, only three of these appear to be utilized, and these represent the sites where the regulatory ligands AMP, ADP and ATP bind in competition with each other. Of interest, the α-AID and the α-CTD are connected by an extended linker peptide, which in the active conformation shown in Fig. 1 wraps around one ‘face’ of the γ subunit, forming a structure called the ‘α hook’ that is in contact with AMP in site 3. Based on a previously proposed model , it was suggested that binding of AMP or ADP, but not ATP, in site 3 would allow interaction with the α hook. The consequent conformational change in the α-linker peptide when ATP binds was proposed to cause the nucleotide-binding and catalytic modules to move apart, allowing protein phosphatases better access to Thr-172 and thus explaining the ability of AMP and ADP binding to inhibit Thr172 dephosphorylation. There is, however, now some doubt about the details of this model, because the original assignment of sequence to electron density in the ‘α-hook’ region appears to have been incorrect [22, 28].
Downstream effects of AMPK activation
A full discussion of the known downstream targets for AMPK is beyond the scope of this article, and readers are referred to a previous review for further details . However, a summary of some of these targets, and the pathways they regulate, is shown in Fig. 2. Consistent with the findings that it is switched on by ATP depletion and has a key role in maintaining cellular energy balance, AMPK promotes catabolic pathways that generate ATP whilst inhibiting anabolic pathways involved in cell growth and other processes that consume ATP. It also causes cell cycle arrest in G1 phase , in part by phosphorylating MDM4 (MDMX), which is a component of the E3 ubiquitin ligase complex that regulates p53 turnover . Cell cycle arrest by AMPK is consistent with its role in maintenance of energy homoeostasis, because DNA replication (during S phase) and mitosis (M phase) are both energy-requiring processes. In general, AMPK achieves its effects both acutely via direct phosphorylation of metabolic enzymes or regulatory proteins involved in the process being regulated, and in the longer term via effects on gene expression. I will discuss here just a few examples of downstream effects of AMPK that will become relevant later in this review.
First, AMPK accelerates glucose uptake to drive catabolism of glucose during muscle contraction, which it does, at least in part, by phosphorylating the Rab-GTPase activator protein (Rab-GAP) TBC1D1 (Fig. 3). TBC1D1 is a member of the same family as TBC1D4 (also known as AS160), and phosphorylation by the insulin-activated kinase PKB/Akt plays a key role in insulin-stimulated glucose uptake . Whereas TBC1D4 appears to be the predominant form in adipocytes, TBC1D1 is the predominant form in most muscle types . Both TBC1D1 and TBC1D4 bind to intracellular vesicles containing the glucose transporter GLUT4, and their Rab-GAP domains promote the GTPase activity of members of the Rab family of small G proteins, thus maintaining them in their inactive, GDP-bound state. Phosphorylation of TBC1D1 at two sites, Ser237 and Thr596, promotes its binding to 14-3-3 proteins, which appears to repress its Rab-GAP activity. Rab proteins are thus converted to their active GTP-bound forms, and they then promote trafficking of the GLUT4-containing intracellular vesicles to the plasma membrane and thus increase glucose uptake [33-35]. Although AMPK can phosphorylate both Ser237 and Thr596 in cell-free assays, available evidence suggests that only Ser237 may be phosphorylated by AMPK in intact cells, with Thr596 being phosphorylated by other kinases including the insulin-stimulated kinase Akt .
Secondly, AMPK phosphorylates and inactivates both isoforms of acetyl-CoA carboxylase, ACC1 and ACC2, lowering the concentration of their reaction product malonyl-CoA. As malonyl-CoA is both an intermediate in fatty acid synthesis and an inhibitor of fatty acid uptake into mitochondria via the transport system involving carnitine : palmitoyl transferase-1, this has the dual effect of inhibiting fatty acid synthesis and enhancing fatty acid oxidation (Fig. 4) [36, 37].
Thirdly, AMPK enhances mitochondrial biogenesis by effects on the transcriptional co-activator PGC-1α. This may involve direct phosphorylation of PGC-1α , or activation of the lysine deacetylase SIRT1, which deacetylates and activates PGC-1α .
Fourthly, as well as its direct effects on acetyl-CoA carboxylase, AMPK inhibits fatty acid synthesis in the longer term by phosphorylating SREBP-1c, inhibiting its ability to promote transcription of lipogenic enzymes . AMPK also inhibits cholesterol and phospholipid/triacylglycerol synthesis, the former by phosphorylating and inactivating 3-hydroxy-3-methylglutaryl-CoA reductase (HMGR)  and the latter by inactivating glycerol phosphate acyl transferase; it remains unclear whether this is via direct phosphorylation .
Finally, AMPK inhibits two other major biosynthetic pathways required for cell growth, that is, protein and rRNA synthesis. It inhibits the former mainly by inhibiting the mechanistic target-of-rapamycin complex-1 (mTORC1), by phosphorylating both the upstream regulator TSC2  and the mTORC1 subunit Raptor  (Fig. 5). AMPK also inhibits rRNA synthesis by phosphorylating TIF-1A, a transcription factor for RNA polymerase-1 .
Role of AMPK in hypertrophic cardiomyopathy
The only disease-causing mutations currently identified in AMPK genes are in PRKAG2 (encoding the γ2 isoform of the nucleotide-binding subunit), which cause a form of the heart disease Wolff–Parkinson–White syndrome . This syndrome is characterized by ventricular pre-excitation (a premature excitation of the ventricles, readily detected by electrocardiogram), but when caused by PRKAG2 mutations it is usually, although not always , accompanied by cardiac hypertrophy. More than 10 different mutations in PRKAG2 have been identified, all of which cause amino acid substitutions within the CBS repeats of γ2 . Most are inherited in a dominant manner and are thus frequently observed in the affected families, although some (e.g. R384T in CBS2  and R531Q in CBS4 ) cause a more severe form of disease associated with death during infancy, and are therefore only found as de novo mutations that are not present in the parents. Many of the substitutions (R302Q, H383R, R384T, R531G and R531Q) affect basic side chains that occur at similar positions in CBS repeats 1, 2 and 4, and structural studies show that these positively charged side chains bind the negatively charged phosphate groups of the bound adenine nucleotides . These substitutions not only reduce allosteric activation, but also reduce the enhanced net Thr172 phosphorylation caused by AMP binding [27, 47, 48, 50]. Other substitutions (e.g. T400N and N488I), although not appearing to affect residues directly involved in nucleotide binding, still appear to reduce AMP binding. Reduced AMP activation is clearly a loss-of-function effect, which is difficult to reconcile with dominant inheritance of the mutations. However, because the mutations affect binding of the inhibitory nucleotide ATP, as well as the activating nucleotide AMP , the most likely explanation is that they reduce binding of ATP and thus increase basal Thr172 phosphorylation and AMPK activity (i.e. a gain-of-function effect). This has been directly demonstrated with the R531G and R531Q substitutions [48, 51], whilst the effects of mutations equivalent to T400N and N488I made in the γ subunit orthologue in budding yeast also suggest a gain-of-function effect . In addition, increased basal AMPK activity was observed in transgenic mice over-expressing the N488I mutation in the heart .
One feature of the cardiac myocytes of the patients bearing these mutations, based on post-mortem analysis, is the presence of large vacuole-like structures containing glycogen, which appear to disrupt the normal myofibrillar structure . The hearts of children bearing the R531Q and R384T mutations displayed gross hypertrophy and also contained up to 10 times the normal glycogen content [47, 48]. Transgenic mice over-expressing γ2 from a cardiac-specific promoter with R302Q , N488I  or R531G  mutations, but not wild-type γ2, also displayed gross cardiac hypertrophy, ventricular pre-excitation and increased cardiac glycogen content. Thus, the γ2 mutations appear to cause a glycogen storage disorder, but it is unclear why this should lead to ventricular pre-excitation. The atria and ventricles of normal adult hearts are separated by a fibrous ring containing collagen (the annulus fibrosis), which ensures that electrical excitation passes from the atria to the ventricles only through the atrio-ventricular node (where a short delay is imposed), and cannot then return to the atria. Without a fully intact annulus fibrosis, ventricular pre-excitation and potentially fatal arrhythmias can occur. Studies of transgenic mice over-expressing the N488I mutation showed that the fibrous rings were disrupted by glycogen-filled myocytes , which may account for the ventricular pre-excitation exhibited by these mice.
The increased glycogen content can therefore potentially explain the cardiac arrhythmias; but why do the mutations cause high glycogen levels? The most likely explanation is that the increased basal Thr172 phosphorylation and AMPK activity enhances GLUT4 translocation and glucose uptake. Evidence in favour of this idea comes from studies of transgenic mice over-expressing the N488I mutation , which have increased cardiac glucose uptake, glycogen synthesis and glucose-6-phosphate content (the latter is a key allosteric activator of glycogen synthase), as well as increased ACC phosphorylation and fatty acid oxidation. It is interesting that when the N488I over-expressing mice were crossed with mice carrying a point mutation in the GYS1 gene that rendered its product, glycogen synthase, resistant to glucose-6-phosphate activation, the changes in both the glycogen storage content and the ventricular pre-excitation were almost completely reversed . These results suggest that the increased basal AMPK activity causes enhanced glucose uptake into cardiac myocytes but that, in the absence of a genuine increased demand for glucose, glucose-6-phosphate concentrations rise thus driving the flux of glucose carbon into glycogen synthesis rather than glycolysis and oxidation. The cells are, in effect, behaving as if they were short of energy even though they are not. It is important to note, however, that although the glycogen storage and ventricular pre-excitation phenotypes of the N488I mice were reversed by crossing with the GYS1 mutants, their cardiac hypertrophy was not, showing that the latter has a distinct origin. The authors of the study suggested that this was due to enhanced insulin sensitivity of the cardiac myocytes, leading to hyperactivation of the mTORC1 pathway. Consistent with this, the hypertrophy could be largely reversed by treatment with the mTORC1 inhibitor rapamycin .
The γ2 isoform is expressed most abundantly in the heart, although it is also expressed elsewhere . Similar mutations have not been reported in the γ1 isoform, but they have been found in γ3, which is expressed almost exclusively in skeletal muscle . An R200Q substitution in γ3 in domestic pigs (equivalent to the R302Q substitution in human γ2) is associated with a high glycogen content in skeletal muscle, which adversely affects meat quality . Intriguingly, in a screen of candidate genes in 1500 lean and obese human subjects, a heterozygous mutation causing an R225W substitution in γ3 (R225 aligns with R302 in human γ2 and R200 in pig γ3) was found in two unrelated individuals. In these individuals, skeletal muscle glycogen content and basal AMPK activity were 2-fold higher and intramuscular triglyceride was 30% lower than in matched control subjects. In both the pigs and the human subjects, there was no obvious pathological problem associated with the γ3 mutation, and it is interesting to speculate that some endurance athletes might carry γ3 mutations because a high muscle glycogen content would be expected to confer an advantage in endurance events. In this regard, it is noteworthy that the AMPK activator 5-aminoimidazole-4-carboxamide riboside (AICAR), which increases muscle glycogen content  and improves endurance on treadmills when administered to sedentary mice over several weeks , has been banned for use in sport by the World Anti-Doping Agency.
Role of AMPK in type 2 diabetes
Type 2 diabetes occurs when increased release of insulin from the β cells of the pancreas can no longer compensate for insulin resistance, and the latter is strongly associated with increased triacylglycerol content, particularly in liver and skeletal muscle. One possible explanation for this association is that insulin resistance occurs in obese individuals when their capacity to store triacylglycerols in adipose tissue is exceeded, leading to accumulation of these lipids in liver and muscle. This is supported by the findings of studies in humans with lipodystrophy; these individuals lack adipose tissue and store triacylglycerols instead in liver and muscle, which then become severely insulin resistant . Despite the association between insulin resistance and triacylglycerols, in fact diacylglycerols, levels of which are increased at the same time, may be more important. Diacylglycerols activate novel isoforms of protein kinase C (PKCθ in muscle and PKCε in liver), which then appear to downregulate the insulin signalling pathway . This represents one mechanism underlying the insulin resistance associated with obesity, although other factors such as increased inflammation  or endoplasmic reticulum stress  may also play a part. However, AMPK activation may also be able to reverse both of these alternative pathways, as it has anti-inflammatory effects (discussed below) and inhibition of protein synthesis by AMPK has the potential to relieve endoplasmic reticulum stress.
By promoting the oxidation of fatty acids and inhibiting synthesis of fatty acids and triacylglycerols, treatments that activate AMPK would be expected to reduce lipid stores in liver and skeletal muscle and hence improve insulin sensitivity. By promoting glucose uptake by skeletal muscle [37, 65] and inhibiting gluconeogenesis in the liver [66, 67], they might also be expected to ameliorate more directly the hyperglycaemia associated with type 2 diabetes. Indeed, many of these effects are observed in vivo with pharmacological agents that activate AMPK. The first agent to be tested was AICAR, an adenosine analogue that is taken up into cells via adenosine transporters and converted to the phosphorylated nucleotide form, ZMP, which mimics all of the activating effects of AMP on AMPK . AICAR treatment was found to reverse many metabolic abnormalities in animal models of obesity and insulin resistance, such as fa/fa rats [68, 69], ob/ob mice  and rats fed a high-fat diet . At about the same time, it was reported that the biguanide drug metformin, currently the drug of choice for treatment of type 2 diabetes that is prescribed to >100 million patients worldwide, also activates AMPK both in intact cells and in vivo . Metformin activates AMPK indirectly by inhibiting the mitochondrial respiratory chain , but A-769662, a direct AMPK activator that binds at the same site as 991 (Fig. 1), also has similar in vivo effects in animals; A-769662 increases fatty acid oxidation in rats and decreases plasma glucose, body weight gain, plasma and liver triacylglycerols and hepatic expression of gluconeogenic and lipogenic enzymes in ob/ob mice . Similarly, the natural product berberine, which is used in traditional Chinese medicine, activates AMPK by inhibiting the respiratory chain in a similar manner to metformin  and also reduces body weight and improves glucose tolerance in db/db mice whilst reducing body weight and plasma triacylglycerols and improving insulin sensitivity in rats fed a high-fat diet .
Because (with the exception of A-769662) these agents activate AMPK indirectly, it has been important to establish whether their favourable metabolic effects are indeed mediated by AMPK. Arguing against this, the ability of AICAR or metformin to reduce glucose production by hepatocytes in vitro, or of metformin to improve glucose tolerance after short-term (30 min) treatment in vivo, was unaffected by knockout of both isoforms of the AMPK catalytic subunit (α1 and α2) in the liver . The AMPK-independent effects of AICAR may be due to the actions of its intracellular metabolite ZMP to directly inhibit the gluconeogenic enzyme fructose-1,6-bisphosphatase  and/or to lower cyclic AMP (which promotes transcription of gluconeogenic enzymes) due to inhibition of adenylate cyclase , whereas the AMPK-independent effects of metformin may in part be a consequence of ATP depletion, which was found to be more severe in cells lacking AMPK . Despite these negative findings, there is good evidence that the longer term insulin-sensitizing effects of metformin are mediated by AMPK. This evidence comes from a recent study of ‘double knockin’ (DKI) mice in which both endogenous genes encoding acetyl-CoA carboxylase (ACC1 and ACC2) were replaced by genes encoding single amino acid substitutions that eliminate the AMPK sites . As expected, the ACC1/ACC2 activities, and the cellular content of the product malonyl-CoA, were increased in the livers of these mice, and fatty acid oxidation was reduced. This was in turn associated with increased levels of hepatic di- and triacylglycerols and increased PKCε activity, whilst the mice had increased fasting blood glucose and insulin levels and were glucose-intolerant and insulin-resistant, and insulin was less effective at suppressing hepatic glucose production. Even more interesting was the effect of feeding a high-fat diet, which abolished the metabolic differences between the wild-type and DKI mice. However, whilst long-term (6 weeks) treatment with metformin improved metabolic parameters such as fasting blood glucose, insulin suppression of hepatic glucose production and gluconeogenic enzyme expression in the fat-fed wild-type animals, none of these effects was observed in the DKI mice. These results suggest that the insulin-sensitizing effects of metformin can be almost entirely attributed to its effects on fatty acid metabolism, mediated by AMPK.
Role of AMPK in cancer
The first link between AMPK and cancer originated from the finding that LKB1 is the upstream kinase required for AMPK activation in response to energy stress and biguanide drugs [9-11]. Mutations in the gene encoding LKB1 were already known to cause Peutz–Jeghers syndrome in humans. Patients with this syndrome are heterozygous for mutations that cause a loss of kinase activity, and they frequently develop benign intestinal tumours (polyps) that in mouse models appear to be due to haploinsufficiency . These patients also have an increased risk of developing malignant cancers at other sites, which in mice appears to be due to loss of heterozygosity in somatic cells . LKB1 is therefore a classical tumour suppressor, raising the question as to whether its tumour-suppressing effects are mediated by AMPK. LKB1 is also required for phosphorylation and activation of a family of 12 ‘AMPK-related kinases’ that have kinase domains closely related to those of AMPK . Although one or more of these AMPK-related kinases might also contribute to the tumour-suppressor effects of LKB1, AMPK remains the best candidate because of its ability to inhibit mTORC1 and almost all biosynthetic pathways required for cell growth, and to cause cell cycle arrest. In addition, tumour cells and other rapidly proliferating cells tend to have greatly increased rates of glucose uptake and glycolytic metabolism (termed the Warburg effect or aerobic glycolysis); this metabolic switch is required in part to provide precursors for biosynthesis . In the longer term, AMPK activation tends to promote the more energy-efficient oxidative metabolism utilized by quiescent cells, thus opposing the switch to aerobic glycolysis observed in many tumour cells.
Consistent with the idea that AMPK is a tumour suppressor are data showing that the AMPK activators metformin, phenformin and A-769662 all delay tumour onset in tumour-prone mice , and that a whole-body knockout of AMPK-α1 (the only catalytic subunit expressed in lymphocytes) accelerates the development of lymphomas in transgenic mice over-expressing the Myc oncogene product in B cells . In the latter case, this was accompanied by a shift towards aerobic glycolysis, which (as argued above) might be opposed if AMPK was present. This shift was dependent on hypoxia-inducible factor-1α (HIF-1α), a transcription factor that upregulates expression of glucose and lactate transporters and glycolytic enzymes . Translation of HIF-1α mRNA is upregulated by mTORC1 , a signalling pathway that is switched off by AMPK (Fig. 5).
The original finding that the tumour-suppressor LKB1 was part of the same pathway as AMPK led to epidemiological studies which showed that diabetic patients treated with metformin had a lower incidence of cancer than those on other medications ; this result has since been reproduced in many subsequent studies . There are several possible explanations for this association, and it is not yet certain that the effect is mediated by AMPK, nor that it is an effect of metformin on the tumour cells themselves. There is some evidence from a mouse xenograft model that it may be mediated in part by metformin acting on the liver, lowering plasma levels of glucose and insulin, with the latter considered to trigger increased tumorigenesis in insulin-resistant states . There is also evidence from mouse models that metformin and the related biguanide phenformin prolong survival by triggering a higher rate of apoptosis in tumour cells that have lost LKB1, because the lack of a functional AMPK pathway in such cells makes them more sensitive to ATP depletion caused by the biguanides [88, 89]. If this mechanism operates in humans, then biguanides might be most effective in treating cancers in which the LKB1–AMPK pathway has been lost, which is quite a common occurrence. It might also be reasonable for phenformin rather than metformin to be used for cancer treatment because, being more hydrophobic and less dependent on the organic cation transporter-1 OCT1 for cellular uptake , it can probably enter tumour cells more readily.
If the LKB1–AMPK axis is indeed a tumour-suppressing pathway that restrains tumour growth, it would be expected that tumour cells would be under selection pressure to downregulate the pathway. Consistent with this, biallelic loss-of-function mutations in the LKB1 gene are common in human cancers, occurring in up to 30% of non-small cell lung cancers [90, 91], 20% of cervical cancers  and 10% of melanomas . Loss of LKB1 activity leads to failure of AMPK activation during energy stress . Mutations in the subunits of AMPK itself seem to be much less common in human cancer, possibly because (unlike LKB1) each subunit is encoded by more than one gene. However, downregulation of expression of AMPK-α2 is relatively common in hepatocellular carcinoma and is associated with poor prognosis . Finally, the growth-promoting protein kinase B/Akt pathway is hyperactivated in many human cancers, by activating mutations in phosphatidylinositide 3-kinase (PI3K) or receptors upstream of it, or loss-of-function mutations in PTEN, which is the tumour suppressor that degrades phosphatidylinositol-3,4,5-trisphosphate (the PI3K product and second messenger for protein kinase B; also known as Akt) . It was shown in 2006  that PKB/Akt phosphorylates Ser485 on rat AMPK-α1, and that this inhibited subsequent phosphorylation of the activating site, Thr172, by LKB1. We have recently confirmed that Akt phosphorylates the equivalent site (Ser487) on human AMPK-α1 (although not the equivalent site on AMPK-α2, Ser491, which is modified instead by autophosphorylation) . These sites lie within a loop of about 50 residues that I have termed the ‘ST loop’ due to its serine-/threonine-rich nature. In the structure shown in Fig. 1, this loop was not resolved, but it would loop out of the rear side of the α-CTD (its approximate position is shown with a dashed line in Fig. 1). Because the complex may not be phosphorylated on the ST loop after expression in bacteria, it is likely to be mobile within the crystals; however we have provided evidence that, after phosphorylation on Ser487 and possibly other sites, the ST loop interacts with the C-helix on the kinase domain, thus blocking access of upstream kinases to Thr172 . We also showed that Ser487 was phosphorylated in three different human tumour cell lines in which Akt was hyperactivated due to loss of PTEN. In these cells, AMPK was moderately resistant to Thr172 phosphorylation and activation, but this could be rescued either by addition of an Akt inhibitor or by re-expressing PTEN, both of which triggered net dephosphorylation of Ser487 . These results show that a previously unrecognized effect of Akt hyperactivation in tumour cells is that it downregulates the AMPK pathway, thus reducing its restraining influence on cell growth and division. Because Ser487 phosphorylation inhibits but does not totally block Thr172 phosphorylation, the results also suggest that this downregulation might be overcome using AMPK-activating drugs, indicating a potential therapeutic role for these agents in tumours in which the Akt pathway is hyperactivated.
Role of AMPK in inflammatory disease
There is an increasing evidence that AMPK has anti-inflammatory effects, and that this may be mediated by its metabolic actions . Unactivated cells of the immune system, including dendritic cells, neutrophils and T cells, utilize mainly oxidative metabolism (including fatty acid oxidation) to generate ATP. However, once activated, they usually switch to aerobic glycolysis, analogous to the metabolic changes that occur in tumour cells. In dendritic cells, this switch is associated with reduced AMPK activation, is inhibited by pharmacological activation of AMPK and is promoted by AMPK downregulation . Similar effects are observed in macrophages: classically activated (M1) macrophages with a pro-inflammatory role utilize aerobic glycolysis, whereas alternatively activated (M2) macrophages, which are more involved in the resolution of inflammation, tend to utilize oxidative metabolism instead. Studies using macrophages in which AMPK has been downregulated suggest that AMPK normally attenuates production of inflammatory cytokines [100, 101]. The idea that the anti-inflammatory actions of AMPK are due to its metabolic actions was strengthened by the findings of studies in a mouse model with knockout of AMPK-β1, the predominant β subunit isoform expressed in macrophages. AMPK-β1-deficient macrophages displayed reduced ACC phosphorylation and mitochondrial content, as well as reduced rates of fatty acid oxidation that promoted the accumulation of pro-inflammatory diacylglycerols. This triggered M1 skewing in vivo in macrophages derived from both bone marrow and adipose tissue. Activation of AMPK with A-769662 increased fatty acid oxidation in wild type but not in β1-deficient macrophages, and the effects of A-769662 to suppress inflammation were impaired if fatty acid oxidation was blocked . These findings suggest that a major component of the anti-inflammatory action of AMPK is mediated via its effects on fatty acid oxidation. The role of AMPK in macrophages has also been studied using mice with global or myeloid-specific knockout of AMPK-α1, the catalytic subunit isoform expressed in macrophages. Regeneration of muscle in response to muscle injury was defective in these mice, and this was associated with reduced skewing towards M2 macrophages .
Interestingly, we have shown that AMPK is directly activated by salicylate, the natural product from which aspirin (acetyl salicylate) was derived . Salicylate and A-769662 appear to bind to the same site on AMPK; both are selective activators of β1-containing complexes and increase whole-body fatty acid oxidation in wild type but not in β1 knockout mice, showing that increased fat oxidation is mediated by AMPK activation in vivo. Although AMPK is not directly activated by aspirin itself , aspirin is rapidly broken down to salicylate. Indeed, following oral administration of aspirin, the plasma half-life and peak concentration of salicylate are orders of magnitude greater than those of aspirin itself. Although blockade of prostanoid biosynthesis, via inhibition of cyclo-oxygenases, has been considered to be the main mechanism of action of aspirin, there has always been a paradox in that, whilst aspirin and salicylate have similar anti-inflammatory potencies, salicylate is a very poor inhibitor of cyclo-oxygenase [105, 106]. Furthermore, although significant AMPK activation requires at least 1 mM salicylate, such concentrations are reached in humans taking high doses of aspirin or another salicylate derivative, salsalate, for rheumatoid arthritis. This raises the intriguing possibility that at least some of the anti-inflammatory effects of salicylate-based drugs may be mediated by AMPK. Of note, the regular use of aspirin is also associated with reduced incidence of cancer , although whether this effect is mediated by AMPK remains uncertain.
Role of AMPK in viral infection
Hepatitis C virus is a positive-sense, single-stranded RNA virus with a lipid envelope, and it is estimated that an infected hepatocyte produces up to 50 viral particles per day. Because this requires synthesis of viral protein and lipids as well as RNA, it would be expected that it would cause increased ATP turnover and hence AMPK activation, which might then restrain further viral replication. Paradoxically, however, AMPK activation was found to be downregulated in a cell culture model of hepatitis C virus infection . This appears to be because the viral protein NS5A binds to and activates PI3K, thus switching on the PKB/Akt pathway . PKB/Akt would promote protein and lipid synthesis and cell survival, but would also downregulate AMPK via phosphorylation of AMPK-α1 at Ser487 within the ST loop. Indeed, experiments involving transfection of an S487A mutant suggested that phosphorylation of Ser487 is required for efficient viral replication . Although investigation of the role of AMPK in viral infection is a relatively new area of research, there are indications that other viruses also interact with the AMPK system .
The potential application of novel AMPK-modulating drugs
An important consideration in the development of novel AMPK-activating drugs is that they would have to be significantly more effective than existing agents such as metformin or salicylates, which are both inexpensive and have a good safety record. One potential shortcoming with the use of metformin is that its effects appear to be largely confined to the liver. Because it is a cation that is not readily cell permeable, the uptake of metformin into cells requires membrane transporters of the organic cation transporter family, such as OCT1 . Further, because hepatocytes are exposed to higher concentrations of orally delivered metformin via the portal vein than most other cells , and also express high levels of OCT1, the effects of metformin in vivo may be largely restricted to the liver . In particular, metformin may not cause much activation of AMPK in skeletal muscle because of low peripheral drug concentrations and low OCT1 expression. The doses of metformin that can be administered to humans are limited by gastrointestinal side effects, which are likely to be AMPK independent and may be caused by inhibition of the respiratory chain and consequent lactate production in the liver or gut [112, 113]. The more hydrophobic biguanide phenformin is more cell permeable than metformin and therefore much less dependent on OCT1 for cellular uptake . Although formerly used for treatment of type 2 diabetes, it was withdrawn in most countries in the 1970s due to cases of lactic acidosis in patients taking the drug (with hindsight, this was an unsurprising side effect as both biguanides inhibit the respiratory chain and thereby activate AMPK ). However, the frequency of developing lactic acidosis with phenformin, whilst life-threatening and around 20-fold higher than with metformin, was still rare (≈60 cases per 100 000 patient-years) . This small risk might be more acceptable if the drug was being used to treat cancer, rather than diabetes .
The indirect AMPK activator AICAR, which is converted to the AMP mimetic ZMP within the cell , has the generic drug name acadesine and has been tested in humans . However, it is not orally available and has to be injected and, as discussed above, has clear ‘off-target’ AMPK-independent effects in mice, so its potential as an AMPK-activating drug in humans seems limited. New direct AMPK activators that are readily cell permeable, exemplified by 991 and A-769662 , may have additional benefits compared to metformin by activating AMPK and promoting glucose uptake in organs other than the liver, especially skeletal muscle. These compounds are being developed by pharmaceutical companies, and the results of studies of their potential toxicity are not yet available within the public domain. However, one potential problem with their use is that increased glucose uptake into skeletal and cardiac muscle, in the absence of a demand for increased catabolism, would be expected to increase their glycogen content, as observed with the γ2 mutations that increase basal AMPK activity. Nevertheless, whilst it is clear that increased glycogen content is harmful if it occurs during foetal development of the heart, it is not yet certain that it would be a problem in adults. Whether AMPK-activating drugs would also cause cardiac hypertrophy, which seems to be a secondary consequence of the activating γ2 mutations that is independent of their effects on glycogen accumulation , also remains unclear at present. However, adverse cardiac side effects would be a potential concern with any novel AMPK activators that would have to be carefully monitored, and the development of AMPK-activating drugs targeting isoform combinations not expressed in the heart, such as γ3 , might be advantageous.
Are there any indications for the development of inhibitors of AMPK? Such agents might, of course, be useful for the treatment of heart disease associated with γ2 mutations; studies using transgenic mice expressing the γ2 N488I mutation in the heart from a promoter switched off by tetracycline administration revealed that the clinical signs of the disease were at least partially reversible in mice . A larger market for AMPK inhibitors might be expected for treatment of those cancers in which the LKB1–AMPK pathway, rather than being downregulated as discussed above, has remained fully functional. There is evidence from mouse models that tumours with a functional LKB1–AMPK pathway are protected against cell death induced by addition of metformin or phenformin in vivo, compared with tumours lacking the pathway [88, 89]. In these cases, the biguanides may be acting as a type of cytotoxic drug via depletion of cellular ATP, which is less harmful when AMPK is available to mount a response. It seems possible that the existence of a functional LKB1–AMPK pathway within tumour cells may also protect them against the effects of other cytotoxic drugs used for cancer treatment, in which case such drugs might be effectively combined with an AMPK inhibitor. At present the only widely available AMPK inhibitor is compound C  (also known as dorsomorphin ), but that agent is in fact a relatively nonselective kinase inhibitor , and thus much more specific inhibitors are needed.
Conclusions and perspectives
It is clear from the above-mentioned discussion that AMPK plays an important role in several diseases that are highly prevalent within most human populations, including type 2 diabetes, cancer and inflammatory diseases. It also appears to be involved in one disorder (hepatitis C) that, whilst less common amongst the whole population, is nevertheless prevalent in certain subgroups. The fact that widely used drugs that have been in human use for 50 years (metformin) or even longer (salicylates) may work, at least in part, by activating AMPK suggests that AMPK activators may have potential for the development of novel drugs. There are many recent patents for such activators, and it seems likely that some of these drugs may enter clinical trials soon. However, a concern has emerged from the studies of mutations in AMPK-γ2 that cause excessive glycogen storage and ventricular pre-excitation in cardiac muscle: the adverse consequences of these mutations suggest that there could be some situations (e.g. in cardiac myocytes) in which the activation of AMPK in the absence of real energy deficit may be undesirable.
Conflict of interest statement
No conflicts of interest to declare.
Studies in the author's laboratory are supported by a Senior Investigator Award (097726) from the Wellcome Trust and by a Programme Grant (C37030/A15101) from Cancer Research UK.