AMPK—Sensing Energy while Talking to Other Signaling Pathways

Summary The AMP-activated protein kinase (AMPK) is a sensor of cellular energy and nutrient status, expressed almost universally in eukaryotes as heterotrimeric complexes comprising catalytic (α) and regulatory (β and γ) subunits. Along with the mechanistic target-of-rapamycin complex-1 (mTORC1), AMPK may have been one of the earliest signaling pathways to have arisen during eukaryotic evolution. Recent crystal structures have provided insights into the mechanisms by which AMPK is regulated by phosphorylation and allosteric activators. Another recent development has been the realization that activation of AMPK by the upstream kinase LKB1 may primarily occur not in the cytoplasm but at the surface of the lysosome, where AMPK and mTORC1 are regulated in a reciprocal manner by the availability of nutrients. It is also becoming clear that there is a substantial amount of crosstalk between the AMPK pathway and other signaling pathways that promote cell growth and proliferation, and this will be discussed.

Given the great potential of AMPK activators as therapeutic agents ( Hardie, 2013 ), many high-throughput screens have been conducted and numerous activators found; the structures of some of these are shown in Fig. 1 . They can be divided into three major classes according to their mechanism of action: 1) Indirect activators that inhibit cellular ATP production and thus increase cellular AMP:ATP and ADP:ATP ratios ( Fig. 1a ). Most of these compounds reduce ATP production by inhibiting the respiratory chain, although resveratrol, quercetin and oligomycin inhibit the mitochondrial ATP synthase (F 1 F 0 -ATPase) ( Gledhill et al., 2007 ), while 2-deoxyglucose inhibits glycolysis. This mechanism is most clearly demonstrated using cells expressing a mutation in the AMPK-γ subunit (e.g. R531G in γ2) that prevents binding of AMP at the critical activating site; this class of compound fails to activate the mutant kinase. This approach has been used to show that the anti-diabetic drug metformin, the glycolytic inhibitor 2-deoxyglucose, and the plant products, resveratrol, berberine and galegine (metformin was derived from the latter), all act via this mechanism ( Hawley et al., 2010 ). In the last two or three years a bewildering variety of natural plant products, many of which (like berberine) have been used in traditional Asian medicine, have been shown to activate AMPK. The author recently compiled a list of over 50 such compounds ( Hardie, 2014 ). While their mechanism of action has not been established in most cases, the author suspects that many of them are mitochondrial poisons produced by plants as a defensive mechanism. Following the aphorism of Paracelsus that “the dose makes the poison”, at high doses these compounds may act as mitochondrial poisons that deter grazing of the plant by animals or insects, or infection by plant pathogens, while at lower doses they may have useful therapeutic properties by causing a mild inhibition of ATP synthesis that activates AMPK. 2) Compounds that are pro-drugs converted into AMP analogs inside cells. The founder member of this class is 5-aminoimidazole-4-carboxamide ribonucleoside (AICAR), a nucleoside that is taken up into cells by adenosine transporters ( Gadalla et al., 2004 ) and then phosphorylated to the equivalent nucleotide ZMP ( Fig. 1b ). ZMP is an AMP analog that mimics all three effects of AMP on the AMPK system ( Corton et al., 1995 ) and binds to the same site(s) as AMP, so the effects of AICAR are also abolished by the R531G mutation ( Hawley et al., 2010 ). For many years, AICAR was the only member of this class, but recently a new AMP analog, 5-(5-hydroxyl-isoxazol-3-yl)-furan-2-phosphonic acid or C2, was reported to be 100-fold more potent than AMP, and over 1000-fold more potent than ZMP, as an AMPK activator in cell-free assays. Since it carries a negatively charged phosphonate group, C2 is not cell-permeable, but it can be administered to cells in the form of the pro-drug C13 ( Fig. 1b ) in which the phosphonate group is esterified; C13 is converted into C2 by intracellular esterases. Unlike ZMP, C2 has the advantage that it does not affect other AMP-sensitive metabolic enzymes, such as glycogen phosphorylase or fructose-1,6-bisphosphatase ( Gomez-Galeno et al., 2010 ). Like AMP and ZMP, C2 causes both allosteric activation and protection against Thr172 dephosphorylation, and its effects are abolished by the R531G mutation, but unlike AMP and ZMP it is a selective activator of complexes containing AMPK-α1 rather than -α2 ( Hunter et al., 2014 ). 3) Direct AMPK activators whose binding involves the β subunit carbohydrate-binding module. The recently recognized binding site for this class of AMPK activator is described in more detail in the section on AMPK structure below. The first activator in this class was the thienopyridone A769662 ( Cool et al., 2006 ), which directly activates AMPK, mimicking the effects of AMP to cause allosteric activation and protect against Thr172 dephosphorylation ( Goransson et al., 2007 ; Sanders et al., 2007 ). A769662 acts in an AMP-independent manner, since it activates AMPK in cells expressing the AMP-insensitive R531G mutant ( Hawley et al., 2010 ). Even before the exact binding site had been clarified, it was clear that it involved the β subunit, because activation was more effective with β1 than β2 complexes ( Scott et al., 2008 ) and was abolished by a mutation within β1 (S108A) that prevents autophosphorylation of that serine residue ( Sanders et al., 2007 ). Recently, two new activators, “991” and “MT-63-78”, have been described ( Fig. 1c ). Crystallography showed that 991 binds in the same site as A769662 ( Xiao et al., 2013 ), while MT-63-78 also shows selectivity for β1 complexes ( Zadra et al., 2014 ) and is likely therefore to bind the same site. The presence of a well-defined binding site for these synthetic activators raises the intriguing question as to whether there are natural ligands that bind this

I will now discuss how the recent crystal structure of a complete human α2β1γ1 complex ( Xiao et al., 2013 ), together with analysis of previous partial structures, has illuminated the regulatory mechanisms described above. The structure of the heterotrimer can be divided into two major sections, i.e. the “catalytic module” ( Fig. 2 , bottom left) and the “nucleotide-binding module” ( Fig. 2 , top right). The critical phosphorylation site, Thr172, lies in the cleft between these two modules, where access to upstream kinases and phosphatases may be particularly sensitive to conformational changes. The structures of the three subunits within the heterotrimer will now be discussed in turn. The α subunit The kinase domain at the N-terminus of the α subunit has the small N-lobe and larger C-lobe of a typical serine/threonine kinase domain, with the active site that binds MgATP (which in this structure contained the kinase inhibitor staurosporine) located in the cleft between them. Despite this bound inhibitor the kinase domain was in an active conformation, because Thr172 had been phosphorylated prior to crystallization, and because the construct was also crystallized in the presence of two activating ligands, i.e. AMP and 991. In addition, the four hydrophobic residues that form the “regulatory spine” (Leu68 and Leu79 from the N-lobe, and His137 and Phe158 from the C-lobe) were aligned as in other active kinases ( Taylor and Kornev, 2011 ). Within the α subunit sequence the kinase domain (KD) is immediately followed by the auto-inhibitory domain (AID), so-called because KD-AID constructs are around ten-fold less active than constructs containing the KD alone, even after phosphorylation on Thr172 ( McBride et al., 2009 ; Pang et al., 2007 ). Structures for a KD-AID construct from the fission yeast Schizosaccharomyces pombe ( Chen et al., 2009 ), and for two isolated rat and human AIDs ( Chen et al., 2013 ), show that AIDs form a compact bundle of three α-helices. In the structure of the S. pombe KD-AID ( Chen et al., 2009 ), where the kinase is in an inactive conformation with the “regulatory spine” out of alignment, helix α3 of the AID interacts with the “rear” of the kinase domain (i.e. the opposite side to the active site), forming interactions with residues in the C-helix of the N-lobe and the E-helix of the C-lobe that appear to maintain the kinase domain in an inactive conformation ( Fig. 3 , left). Although the AID was not fully resolved in the human structure ( Xiao et al., 2013 ), it was clear by comparison with the S. pombe KD-AID structure that the AID had undergone a major rotation, so that helix α3 within it now interacts with the γ subunit, rather than with the N- and C-lobes of the kinase domain ( Fig. 3 , right). The first part of the linker between the AID and the C-terminal domain of the α subunit (the “α-linker”) forms the “hinge” between the “catalytic module” and the “nucleotide-binding module”, and the binding of this linker to the AMP-bound γ subunit appears to cause this reorientation of the AID. This is discussed in more detail below, in the section describing the γ subunit. The globular C-terminal domain of the α subunit terminates with an α-helix that carries a well-defined nuclear export sequence. A similar sequence is found in both the α1 and α2 isoforms, although it has only been shown to be functional as a nuclear export sequence in the case of α2 ( Kazgan et al., 2010 ). Just prior to this α-helix there is a long serine/threonine-rich loop that has not been resolved in any of the crystal structures. We refer to this as the “ST loop”, and its function will be discussed later.

It has been known for many years that the γ subunit contains the regulatory binding sites for adenine nucleotides ( Cheung et al., 2000 ; Scott et al., 2004 ). Following N-terminal regions that vary greatly in length and sequence between different isoforms, and a short segment providing the β-strand that interacts with the C-terminal domain of the β-subunit, γ subunits from all species contain four tandem cystathione β-synthase (CBS) repeats. CBS repeats occur in a small number of other human proteins (including cystathione β-synthase itself) ( Bateman, 1997 ), usually as just two repeats that form a structure known as a Bateman domain, with a cleft between the repeats that often binds ligands containing adenosine ( Scott et al., 2004 ). The two Bateman domains in the AMPK-γ subunits assemble in a pseudo-symmetrical head-to-head manner, forming a flattened disk with one CBS repeat in each quadrant. This arrangement provides four symmetrical clefts near the center of the disk where the regulatory nucleotides could potentially bind; by convention ( Kemp et al., 2007 ) these are numbered 1-4 according to the number of the CBS repeat providing a conserved aspartate side chain that interacts with the ribose ring of the bound nucleotide. However, in mammalian AMPK only sites 1, 3 and 4 are utilized, with site 2 always unoccupied (interestingly, CBS2 has an arginine in place of the conserved aspartate). The phosphate groups of adenine nucleotides bound at sites 1, 3 and 4 lie close together in the center of the disk. Some amino acid side chains interact with nucleotides at more than one site, or even with nucleotides in different sites depending on which nucleotide is bound ( Chen et al., 2012 ; Xiao et al., 2007 ; Xiao et al., 2011 ). This arrangement suggests that there would be complex interactions between the three sites; agreeing with this, mutations of any of the three sites reduces the effects of AMP on both Thr172 phosphorylation and allosteric activation ( Oakhill et al., 2010 ). Sites 1 and 3 are able to bind AMP, ADP and ATP in competition, but there is some disagreement about site 4. It was originally suggested that site 4 only binds AMP in a “non-exchangeable” manner, with sites 1 and 3 being the sites where ADP or ATP exchanged with AMP when they were soaked into crystals made using AMP ( Xiao et al., 2007 ; Xiao et al., 2011 ). However, it was subsequently reported that if the core of the heterotrimeric complex was crystallized in the presence of ATP rather than AMP, ATP was found at sites 1 and 4, rather than 1 and 3 ( Chen et al., 2012 ). Binding of ATP rather than AMP was also observed to cause a conformational change in which the two Bateman domains (formed by CBS1/CBS2 and CBS3/CBS4) rotated towards each other by 5°, which had not been observed in the previous crystal soaking experiments. Since AMPK complexes purified from bacteria in the absence of added nucleotides predominantly have AMP rather than ATP bound at site 4, Chen et al (2012) argued that this site must have a higher affinity for AMP than ATP. As discussed in the next paragraph, site 3 appears to be the site where AMP binding causes activation, but analysis of the crystals obtained in the presence of ATP also suggested that binding of ATP at site 4 would preclude binding of any other nucleotide at site 3 ( Chen et al., 2012 ). Thus, to obtain binding of AMP at the critical site 3, AMP may have to be already bound at site 4, suggesting a co-operative interaction between the two sites. Intriguingly, the linker peptide between the AID and the C-terminal domain of the α subunit wraps around the face of the γ subunit that contains sites 2 and 3 ( Xiao et al., 2013 ; Xiao et al., 2011 ). One well-conserved sequence motif within this linker (α-regulatory subunit interacting motif-1, α-RIM1) contacts the unused site 2, while another (α-RIM2) contacts AMP bound in site 3 ( Xin et al., 2013 ). These authors proposed that the binding of α-RIM1 and α-RIM2 to the γ subunit when AMP is bound in site 3 would force the AID to dissociate from the kinase domain, thus relieving its inhibitory effects. Since this linker also forms the hinge between the “catalytic” and the “nucleotide-binding” modules, this change may also pull these modules together, reducing the accessibility of Thr172 to protein phosphatases and protecting it against dephosphorylation. Although this model is quite compelling, more structures of mammalian heterotrimers in inactive states will be required to confirm it. The role of nucleotide binding at site 1, and the reason why ADP mimics only one of the three effects of AMP, remain unclear. Also unclear is the mechanism by which AMP binding promotes phosphorylation of Thr172 by LKB1, although the next section describes some intriguing new twists in this story.

One can make several evolutionary arguments at to why nutrient sensing might have developed at the lysosome. Many heterotrophic protists, such as flagellates, ciliates and amoebas, feed primarily by phagocytosis, which delivers macromolecules directly to lysosomes or vacuoles without passage through the cytoplasm. Lysosomes or vacuoles in unicellular organisms can therefore be regarded as the equivalent of the gut in a multicellular organism. All eukaryotic cells also carry out autophagy, the process by which organelles and macromolecular complexes, such as protein aggregates or glycogen particles, are engulfed by autophagosomes and delivered to lysosomes for digestion. Indeed, AMPK plays a fundamental role in switching on autophagy during starvation by phosphorylating ULK1, the protein kinase that initiates the process ( Egan et al., 2011 ). Plant and fungal cells contain vacuoles, which are essentially very large lysosomes that can be used for long-term storage of nutrients and other purposes. Interestingly, in the budding yeast Saccharomyces cerevisiae there appear to be two pools of glycogen, a cytoplasmic pool that can be rapidly mobilized by phosphorylase (encoded by GPH1 ) during early stationary phase, and a vacuolar pool mobilized by α-glucosidase (encoded by SGA1 ) during late stationary phase ( Wilson et al., 2010 ). These two pools can be distinguished by mutations in GPH1 or SGA1 (which both cause increased glycogen accumulation, but at different stages during stationary phase), and because the vacuolar pool does not form in strains with mutations in genes required for autophagy. Intriguingly, deletion of any one of twelve genes required for the structure or assembly of the v-ATPase also leads to glycogen accumulation, which may occur because acidification of the vacuole fails to occur, and this is necessary for the activity of the vacuolar α-glucosidase encoded by SGA1 ( Wilson et al., 2002 ). Glycogen breakdown can also occur in lysosomes in mammals, particularly during the neonatal period. Thus, human cells also express a lysosomal α-glucosidase, whose genetic loss is fatal during the first year of life. Interestingly, following the birth of rats numerous glycogen-containing autophagosomes rapidly appear in the liver and muscle and then disappear within a few hours ( Schiaffino et al., 2008 ). In addition, mice with knockouts in the autophagy genes Atg5 and Atg7 are born at the expected frequency, but die within one day of birth ( Komatsu et al., 2005 ; Kuma et al., 2004 ). Mice with a knock-in mutation in RagA that inhibits its GTPase activity and locks it in its GTP-bound state, thus activating mTORC1 irrespective of nutrient supply, also die within a few hours of birth. This appears to occur because they are unable to respond to the profound hypoglycemia that follows the loss of nutrient supply from the placenta, by switching off mTORC1 and inducing autophagy ( Efeyan et al., 2013 ). It appears that lysosomal glycogen breakdown may be particularly important in newborn mammals, which must rapidly switch from the placenta to the gut for their source of nutrients. All of these observations suggest that the lysosome or vacuole can be an important site where sensing of glucose has to occur.

A-Raf, B-Raf and c-Raf are paralogous protein kinases involved in the Ras-Raf-MEK-ERK pathway, by which the effects of many growth factor receptors on cell growth and proliferation are transduced. The role of B-Raf in cancer was highlighted by findings that a V600E mutation in its activation loop, which causes constitutive activation, is present in around 50% of all cases of malignant melanomas, and in smaller proportions of other cancers ( Davies et al., 2002 ). It was shown many years ago that AMPK phosphorylates a site near the C-terminus of c-Raf, Ser621, in cell-free assays ( Sprenkle et al., 1997 ), and recently this has been re-investigated. Despite the fact that the sequence around Ser621 is a good fit for the consensus recognition motif for AMPK ( Gwinn et al., 2008 ; Scott et al., 2002 ), treatment of MEFs with the AMPK activator A769662 did not appear to cause increased phosphorylation of c-Raf at Ser621. However, it did cause increased phosphorylation of B-Raf at the equivalent site, Ser729, as well as attenuation of ERK activation, and both effects were lost in AMPK knockout MEFs ( Shen et al., 2013 ). When B-Raf was re-expressed in B-Raf-deficient MEFs, the ability of AMPK activators to inhibit ERK activation was regained using wild type B-Raf but not an S729A mutant; a keratinocyte cell line expressing the S729A mutant B-Raf also proliferated more rapidly than cells expressing the wild type. Phosphorylation of Ser729 on B-Raf does not directly inhibit its kinase activity, but instead promotes 14-3-3 binding and thus prevents association of B-Raf with KSR1, a scaffold protein with a kinase-like domain related to the Raf kinases, on which the latter normally assemble with their downstream kinases MEK and ERK to trigger downstream signaling ( Shen et al., 2013 ) ( Fig. 6 ). Inhibitors of oncogenic B-Raf V600E , such as vemurafenib and dabrafenib, are initially very effective in treating melanomas carrying the mutation, but unfortunately resistance to these drugs develops rapidly and disease returns. However, following their findings suggesting that AMPK down-regulated B-Raf signaling, the same group showed that treatment of melanoma cells with a combination of PLX4720 (a relative of vemurafenib) and phenformin (a more cell-permeable relative of metformin) enhanced apoptosis and reduced cell viability in a synergistic manner, delayed the emergence of PLX4720 resistance in a cell culture model, and induced tumor regression in xenograft and genetically engineered (B-Raf V600E /PTEN -/- ) mouse models of melanoma ( Yuan et al., 2013 ). These results suggest that the AMPK activators metformin or phenformin might be useful adjuncts for therapy of human melanoma patients with B-Raf inhibitors. Interestingly, the AMPK signaling pathway was also reported to be down-regulated in cells expressing the B-Raf V600E mutation; this was associated with phosphorylation by ERK and RSK of two sites on LKB1 that were proposed to negatively regulate phosphorylation of AMPK ( Zheng et al., 2009 ). Thus, there appear to be two opposing negative feedback loops operating between B-Raf and AMPK ( Fig. 6 ). Although the molecular details are less clear, there are other possible connections between the Ras-Raf-MEK-ERK and AMPK pathways. KSR2 is a paralog of KSR1; protein interaction screens identified several AMPK subunits (α1, α2, β1, γ1) as binding partners that had a preference for KSR2 over KSR1 ( Costanzo-Garvey et al., 2009 ). Although only KSR1 is detectable in MEFs, neither ksr1 -/- nor ksr2 -/- MEFs were able to oxidize exogenous fatty acids, but this was restored by over-expression of either KSR1 or KSR2; this required the region of these proteins (between domains CA2 and CA3) that interact with AMPK. Interestingly, KSR2 knockout mice become obese, glucose-intolerant and insulin-resistant as adults, which appeared to be due to reduced energy expenditure rather than increased food intake. The phosphorylation of AMPK in response to AICAR in explants of white adipose tissue from wild type mice was lost in the KSR2 -/- mice, and genes involved in oxidative metabolism were also down-regulated in white adipose tissue. Interestingly, this function of KSR2 appears to translate into humans ( Pearce et al., 2013 ), because in a screen of around 2000 obese humans, a significantly higher proportion had heterozygous mutations in the KSR2 gene compared with lean controls. Heterozygous carriers of human KSR2 mutations had lower basal metabolic rates and higher respiratory quotients, the latter indicating a relative deficit in fat oxidation. Finally, fatty acid oxidation was enhanced in myotubes that had been differentiated from C2C12 myoblasts transfected with DNA encoding wild type KSR2, but this was partially or completely abolished using most of the human KSR2 mutants, although the defect could be restored in every case by treating the cells with metformin ( Pearce et al., 2013 ). Thus, KSR2 appears to be required for the abi