Myristoylation confers noncanonical AMPK functions in autophagy selectivity and mitochondrial surveillance

AMP-activated protein kinase (AMPK) plays a central role in cellular energy sensing and bioenergetics. However, the role of AMPK in surveillance of mitochondrial damage and induction of mitophagy remains unclear. We demonstrate herein that AMPK is required for efficient mitophagy. Mitochondrial damage induces a physical association of AMPK with ATG16-ATG5-12 and an AMPK-dependent recruitment of the VPS34 and ATG16 complexes with the mitochondria. Targeting AMPK to the mitochondria is both sufficient to induce mitophagy and to promote cell survival. Recruitment of AMPK to the mitochondria requires N-myristoylation of AMPKβ by the type-I N-myristoyltransferase 1 (NMT1). Our data support a spatiotemporal model wherein recruitment of AMPK in association with components of the VPS34 and ATG16 complex to damaged mitochondria regulates selective mitophagy to maintain cancer cell viability. AMP-activated protein kinase (AMPK) plays a role in starvation-induced autophagy, but a role in mitochondrial damage-induced mitophagy is not known. Here, Liang et al. show that AMPK is recruited to damaged mitochondria in an N-myristoylation-dependent manner and in turn recruits the ATG16 autophagy complex.

M acroautophagy, commonly referred to as autophagy, is a degradation mechanism that mediates the sequestration and destruction of a variety of endogenous and exogenous cellular components as well as microbial pathogens contributing to the maintenance of cell bioenergetic homeostasis, cytoplasmic quality control and defence against disease [1][2][3] . In mammalian cells, autophagy can be triggered by extracellular signals and intracellular perturbations mediated through multiple interconnected pathways with the mTOR and AMPK cascades playing important roles in nutrient-depletion-induced autophagy [4][5][6][7][8] .
While nutrient withdrawal triggers autophagy-mediated nonselective bulk degradation of cytoplasmic contents, autophagy can also selectively mediate removal of specific organelles 3,9 . Under starvation conditions, however, mitochondria can escape degradation by autophagy despite donating membrane structures for the biogenesis of non-selective autophagosomes 10 . Indeed, cell starvation induces mitochondrial elongation and tubulation through phosphorylation-dependent Drp1 downregulation, which protects mitochondria from autophagy-mediated degradation 11,12 . In contrast, autophagy can mediate selective mitochondrial removal, a process known as mitophagy, during lineage differentiation or in response to mitochondrial damage [13][14][15][16] . Studies with depolarized mitochondria provide important insights into autophagy-mediated selective degradation of damaged mitochondria. On depolarizing mitochondrial damage, the mitochondrial resident protein kinase PINK1 is activated and recruits the protein ubiquitin E3 ligase Parkin to damaged mitochondria, where Parkin catalyses ubiquitination of a variety of mitochondrial proteins 13,17,18 . Since ubiquitinated proteins can be recognized by p62/SQSTM1, which also interacts with LC3 via its C-terminal LIR motif, p62 could potentially bridge damaged mitochondria and autophagosomes [19][20][21] . However, p62 is not essential for mitophagy 22 , suggesting that additional mechanisms may deliver damaged mitochondria to autophagosomes or, alternatively, the autophagy machinery can be recruited to damaged mitochondria to mediate mitochondrial removal.
Although Parkin overexpression in cultured cells can effectively eliminate the majority of mitochondria on depolarizing stress, endogenous Parkin levels are not sufficient to mediate mitophagy in many cell types including murine embryonic fibroblasts (MEFs) 23,24 , which, similar to Parkin knockout cells, do not exhibit mitochondrial aberrations that are indicative of defects in mitochondrial surveillance [25][26][27][28] . Thus, the physiological role of Parkin-mediated mitophagy in mitochondrial homeostasis is not fully understood 14,29,30 . Indeed, mitophagy can occur in a PINK-Parkin cascade-independent manner [31][32][33][34][35] , although the mechanisms of Parkin-independent mitochondrial surveillance remain to be elucidated.
Despite the striking difference in substrate selectivity between starvation and mitochondrial damage-induced autophagy, our data show that AMPK is activated under both conditions. AMPKa knockout cells exhibit suboptimal autophagy induction in response to multiple stress signals including depolarizing mitochondrial damage and exhibit defects in removal of damaged mitochondria. We demonstrate that the AMPK complex associates with damaged mitochondria, which is dependent on N-myristoylation of AMPKb subunits. Further, mitochondrial damage induces physical interaction between AMPK and ATG16 complexes. AMPK is also required for retention of the ATG16 complex in damaged mitochondria. Thus, we provide evidence suggesting that association of AMPK with damaged mitochondria mediates recruitment of the ATG16 complex to the damage sites, serving as a new mechanism underpinning surveillance of mitochondrial damage and removal of damaged mitochondria through autophagy.

Results
AMPK plays a role in autophagy and mitochondrial removal. AMPK mediates adaptive cell functions in response to perturbations in nutrient metabolism and bioenergetics to maintain adequate intracellular energy levels 36 . AMPK plays an important role in regulating autophagy 5,37 , a cellular process that provides energy during nutrient starvation. Given the variety of signals in addition to nutrient depletion that induce autophagy, we speculated that AMPK might be required for autophagy generally or alternatively only a subset of autophagy contexts. In MEFs, short-term serum or glucose starvation and brief treatment with either rapamycin or the mitochondrion-depolarizing agent CCCP (carbonyl cyanide 3-chlorophenylhydrazone) for 4 h increased LC3-PE (Fig. 1a), indicating that these conditions trigger a rapid onset of autophagy. Rapamycin but not glucose starvation or CCCP led to a marked reduction in pS6 levels ( Supplementary Fig. 1A); thus, glucose starvation and CCCPinduced autophagy do not require substantial mTOR inhibition as represented by decreases in pS6 levels. Phosphorylation of AMPK (pAMPK) and AMPK substrates including ACC (pACC) and ULK1 (pULK) 38,39 was readily detected in wild-type, but not MEF cells lacking both AMPKa1 and a2 (AMPK À / À ) ( Fig. 1a; Supplementary Fig. 1A,L), indicating that these events are dependent on AMPK. In wild-type MEFs, CCCP induced pAMPK and pACC within 30 min ( Supplementary Fig. 1B,C). LC3 lipidation was reduced (approximately two-to sixfold) and high-molecular-weight ubiquitinated proteins were substantially increased in AMPK À / À cells as compared with wild-type cells (Fig. 1a), suggesting altered autophagy activity [40][41][42] . LC3-PE levels remained lower in AMPK À / À cells in the presence of bafilomycin, which was used to clamp autophagosomes, suggesting that autophagy flux is not increased in these cells ( Supplementary Fig. 1D).
AMPK promotes cell survival under nutrient starvation 5 . We observed rapid cell death in AMPK À / À but not wild-type MEFs treated with CCCP (Fig. 1b,c; Supplementary Fig. 1E), suggesting that AMPK also contributes to cell viability following mitochondrial damage. Similar results were obtained with RNA interference (RNAi)-mediated AMPKa depletion from H23 human lung cancer cells ( Fig. 1d; Supplementary Fig. 1F,G). Further, CCCP-induced cell death was effectively prevented by reactive oxygen species scavenging and cell death-associated poly (ADP-ribose) polymerase1 (PARP1) cleavage was largely abolished by transfection of green fluorescent protein (GFP)-AMPKa2 in AMPK À / À cells (Fig. 1b,c). Taken together, AMPK is required for effective autophagy and cell survival in response to multiple signals including mitochondrial stress.
Transmission electron microscopy was performed to examine whether AMPK is required for autophagy-mediated removal of damaged mitochondria. As compared with wild-type MEFs, increased mitochondrion numbers and elongated mitochondria were present in AMPK À / À cells ( Fig. 1e; Supplementary  Fig. 1H). A brief exposure to CCCP led to mitochondrial swelling and altered or loss of cristae in the majority of mitochondria in both wild-type and AMPK À / À cells ( Fig. 1e; Supplementary  Fig. 1H-J). However, large-sized (41 mM) mature autophagosomes containing mitochondrial vestiges (hereby termed mitophagosomes), often adjacent to clusters of curved or circular mitochondria, were present in wild-type but not AMPK À / À cells. Instead, relatively small non-selective autophagosomes devoid of recognizable organelles and autophagy-related structures (o0.5 mM) likely representing phagophores formed in AMPK À / À cells ( Fig. 1e; Supplementary Fig. 1H-J).
To determine whether preautophagosomal vesicles formed in AMPK À / À cells might eventually develop into mature autophagosomes or the autophagy process might terminate     Supplementary Fig. 1H. (f) Western analyses of protein levels from WT and AMPKa null (AMPKa À / À ) MEFs briefly exposed to CCCP (30 mM for 2 h) and then allowed to recover for 3-24 h after the removal of CCCP at 0 h. (g) Western analyses of protein levels from WT and ATG5 À / À MEFs treated without ( À ) or with ( þ ) CCCP for 3 h. Densitometry values of mitochondrial marker HSP60 and TOM20 by ImageJ (normalized against actin) were shown below each band.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms8926 ARTICLE prematurely after initiation, we performed a pulse-chase study. Following a transient exposure to CCCP, mitophagosomes and non-selective autophagic structures in wild-type and AMPK À / À cells, respectively, became less abundant following CCCP removal (Supplementary Fig. 1I-K). However, condensed mitochondrial fragments but not mitophagosomes were present in AMPK À / À cells, suggesting that cells lacking AMPK were unable to effectively remove damaged mitochondria through autophagy. Indeed, western analyses revealed reduced basal and CCCPinduced LC3-PE levels in AMPK À / À MEFs compared with wild-type cells. Following a transient CCCP pulse, LC3-PE returned to basal levels more rapidly (between 2 and 4 h) in AMPK À / À cells than in wild-type cells (B8 h) ( Supplementary  Fig. 1L). Thus, our data suggest that AMPK, while not obligatory for autophagy initiation, controls autophagy activity and signal duration. Consistent with inefficient elimination of damaged mitochondria, basal levels of the mitochondrial protein TOM20 were increased greater than twofold in AMPK À / À MEFs (Fig. 1f). Exposure to CCCP led to rapid loss of TOM20 and the mitochondrial chaperone HSP60 followed by a gradual recovery after CCCP removal in wild-type cells. In contrast, TOM20 and HSP60 levels underwent little change in AMPK À / À cells (Fig. 1f). Of note, loss of mitochondrial proteins preceded transient AMPK activation at the 4-h recovery time point ( Supplementary Fig. 1L), likely due to a transient decline of ATP levels following extensive mitophagy and the restoration of ATP levels afterword due to either compensatory activation of glycolysis or de novo mitochondrial biogenesis. TOM20 levels were not reduced in ATG5 À / À cells on CCCP treatment (Fig. 1g), suggesting that loss of TOM20 is dependent on autophagy. Taken together, our data are consistent with a role for AMPK in autophagy-mediated mitochondrial removal.
LKB1 but not AMPK activity is dispensable for mitophagy. We next sought to determine whether LKB1, which plays a critical role in the energy-sensing cascade 43 , is required for mitophagy. Following a brief exposure to CCCP, pACC was strongly induced in a panel of cell lines except for H23, which lacks functional LKB1 (Fig. 2a). Consistent with elevated pACC levels, LC3-PE was substantially increased following exposure to CCCP, whereas the 'autophagic flux' indicators p62 and NBR1 (in expressing cells) were decreased ( Fig. 2a; Supplementary Fig. 2A). CCCP treatment did not alter p62 messenger RNA (mRNA) levels ( Supplementary Fig. 2C).
The rapid autophagy induction in response to mitochondrial stress was accompanied by loss of mitochondrial proteins HSP60, COX-IV and TOM20 in H226 and H23 lung cancer cells ( Fig. 2b; Supplementary Fig. 2A). However, COX-IV and HSP60 were not reduced in MDA-MB231 cells, whereas TOM20 levels were slightly increased in MCF7 cells ( Fig. 2b; Supplementary Fig. 2A), despite comparable expression of Parkin, a protein ubiquitin E3 ligase that mediates ubiquitination-dependent mitophagy. Thus, MDA-MB231 and MCF7 cells undergo non-selective autophagy rather than mitophagy or, alternatively, mitochondrial biogenesis may compensate for the loss of specific mitochondrial proteins 22,44 . The substantial decreases in mitochondrial proteins following CCCP treatment in H23 cells lacking LKB1 is consistent with the LKB1-mediated energy-sensing function being dispensable for selective removal of damaged mitochondria. Phosphorylation of 4EBP1 remained unchanged in MDA-MD231 and H226 cells expressing functional LKB1 (Fig. 2a), suggesting that CCCP had little inhibitory effect on mTORC1. Unexpectedly, CCCP induced a substantial reduction in p4EBP1 in MCF7 and H23 cells (Fig. 2a), likely reflecting the loss of the 4EBP1 protein, the mechanism for which is currently unclear.
These data argue that neither LKB1 nor mTORC1 inhibition is required for CCCP-induced mitophagy. Conversely, rapamycinmediated mTOR inhibition resulted in loss of TOM20 in H226 but not H23 lung cancer cells ( Supplementary Fig. 2B), consistent with mTOR inhibition alone being insufficient to initiate mitochondrial removal 33 . Further, short hairpin RNA (shRNA)-mediated depletion of LKB1 in SKOV3 cells had no effect on TOM20 downregulation, providing additional evidence supporting the notion that the energy-sensing function downstream of LKB1 is not required for mitophagy (Fig. 2c). In contrast, transfection of a kinase-dead mutant version of AMPKa, GFP-AMPKa2K45R, prevented CCCP-induced TOM20 downregulation in H23 cells (Fig. 2d), consistent with a requirement of AMPK kinase activity but not LKB1 in this process.
It is currently unclear why CCCP-induced AMPK activation was not sufficient to inhibit the mTORC1-4EBP1 cascade. Since CCCP treatment also caused striking increases in AKT phosphorylation ( Supplementary Fig. 2B), mTORC1 activation might be maintained by a strong AKT activation following CCCPinduced mitochondrial stress, consistent with previous studies 45 . Indeed, allosteric inhibition of AKT with MK2206 suppressed S6 phosphorylation in CCCP-treated cells ( Supplementary Fig. 2B).
Previous studies suggest a role for AMPK in regulating the ATG1 complex 46 and AMPK-dependent ULK1 phosphorylation regulates starvation-induced autophagy 38 and mitophagy in differentiated hepatocytes 39 . However, ULK1 phosphorylation is generally reduced during starvation-induced autophagy 7 with a sevenfold reduction in pULK1-S555 on nutrient starvation in cancer cells 47 . Although ULK1 phosphorylation at S555 (pULK1-S555), S777 (pULK1-S777) and S467 (pULK1-S467) was detected only in wild-type but not AMPK À / À cells ( Supplementary  Fig. 1A,L), pULK was not increased on CCCP treatment in any cell line tested and indeed pULK1-S555 was decreased following CCCP treatment in most cell lines ( Fig. 2a; Supplementary  Fig. 2B).

CCCP induces AMPK association with damaged mitochondria.
To elucidate the mechanisms by which AMPK contributes to selective removal of damaged mitochondria, we determined whether AMPK might physically associate with mitochondria. We observed vesicle-like distribution of ectopically expressed GFP-AMPKa2 or RFP-AMPKa2 in the cytoplasm of approximately one-third of transfected cells. While these vesicular structures partially co-localized with the autophagosome marker, GFP-LC3 puncta, the vast majority of the vesicles overlapped with mCherry-p62 ( Supplementary Fig. 3A,B), morphologically resembling both mitochondrial aggregates and autophagosomerelated p62 bodies 19,21 . Similar subcellular distribution has been reported for AMPKb subunits 48,49 . Indeed, transfection of AMPKb1-GFP exhibited co-localization with MitoTracker Red (MTR) in Saos2 cells and the co-localization was increased by a short exposure to CCCP (Supplementary Fig. 3C).
To determine whether endogenous AMPK is recruited to the mitochondria in non-transfected cells, immunofluorescence staining was performed in MEFs. AMPKa subcellular distribution patterns overlapped with MTR in wild-type MEFs but not in AMPK À / À MEFs ( Supplementary Fig. 3D,E). Further, mitochondrial extracts from several cell lineages were analysed for AMPK content. Total and T172-phosphorylated AMPK were readily detected in the mitochondrial fraction of all cell lines assessed including MEF and MCF7 cells (Fig. 3a,b). CCCP treatment led to an approximately threefold increase in AMPK in mitochondrial extracts (Fig. 3b), suggesting that AMPK is recruited to the mitochondria in response to mitochondrial damage. The mitochondrial protein porin (VDAC1) was B10-fold enriched in the mitochondrial fraction and the cytosolic proteins, p70S6K, LKB1 and PDK1, were undetectable in the mitochondrial fraction (Fig. 3a,b), consistent with high purity of the mitochondrial preparation. To further validate the quality of mitochondrial isolation, reverse-phase protein array (RPPA) was performed to quantitatively determine the levels of multiple cellular and mitochondrial proteins. A number of mitochondrial proteins were enriched 2-12-fold in the mitochondrial extracts, whereas proteins not previously shown to be mitochondrial were detected at only background levels (Supplementary Fig. 3F,G). In addition, some proteins exhibited differential localization dynamics in a cell-type-dependent manner including the presence of pAKT in mitochondria in MCF7 (Fig. 3b), consistent with previous reports 50, 51 .
AMPK recruits mitophagy mediators. We next attempted to identify potential AMPK target(s) at the mitochondria that could mediate mitophagy. AMPK was previously shown to interact with the ULK1 complex 39,46 . However, dissociation of AMPK from the ULK1 complex is crucial for rapid autophagy induction on nutrient depletion in cancer cells 47 . We were unable to detect a steady-state interaction between endogenous ULK1 and AMPK in several cell types by co-immunoprecipitation; under the same condition, AMPKa was strongly associated with b subunits ( Fig. 3c; Supplementary Fig. 3H). Given our data suggesting a role for AMPK in regulating the elongation steps leading to the formation of mature autophagosomes rather than preautophagosomal structures ( Fig. 1e; Supplementary Fig. 1H-J), we then assessed potential physical interaction of AMPK with the CCCP -+ -+ -+ -+  VPS34 and ATG16 complexes. While there was no detectable interaction between VPS34 and AMPK, both ATG16 and ATG5-12 conjugates were present in AMPKa immunoprecipitates of endogenous proteins ( Fig. 3c; Supplementary Fig. 3H).
Notably, a brief CCCP treatment led to a two to over threefold increase in the ability to co-precipitate AMPK and ATG16 complexes ( Fig. 3c; Supplementary Fig. 3H). The GFP-AMPKa2 fusion but not GFP alone also associated with the ATG16 complex ( Supplementary Fig. 3I) with the interaction increasing greater than twofold on CCCP treatment (Fig. 3d). Further, to Note that only probing with ATG16 antibody (rabbit polyclonal) gives strong heavy chain signals, but not with ATG12 or the b1/2 antibodies (rabbit monoclonal) (Supplementary Fig. 3H). (d) Western analyses of proteins co-purified with GFP-AMPKa2 (GFP-a2) by GFP-TRAP assays from H23 stable lines following treatment with ( þ ) or without ( À ) CCCP (30 mM, 4 h). Mock assays were performed as controls in cells transfected with an empty mCherry (mCh) vector. Note that GFP only control was shown in Supplementary Fig. 3I. LE, long exposure; SE, short exposure. (e,f) Western analyses of protein levels in whole cell lysate (WCL) or mitochondrial extracts (mito) from wild-type (wt) or AMPKa À / À (a À / À ) MEFs treated with or without CCCP (15 mM) for 3 h (e) and from wild-type and ATG5 À / À MEFs (f). Cytosol and nuclear protein FOXO3A and mitochondrial proteins HSP60 and TOM20 were blotted as fractionation controls.
determine whether AMPK kinase activity is required for association with the ATG5 complex, we performed stable transfection of wild-type and kinase-dead GFP-AMPKa2. Although wild-type and kinase-dead proteins are expressed at different levels with the latter being approximately twofold higher ( Supplementary Fig. 3J), they form a complex with proportional amounts of b subunits (Supplementary Fig. 3K). Both fusion proteins are capable of association with ATG5-12, which is increased in response to CCCP. However, although AMPKa2KD was present in a complex with ATG5-12, relatively more ATG5-12 (approximately twofold) associates with wild-type AMPKa2 ( Supplementary Fig. 3K), suggesting that AMPK activity accelerates the recruitment of the ATG5 complex following CCCP-induced mitochondrial damage. We next interrogated whether AMPK regulates recruitment of the autophagy machinery. The core components of the class III PI3K complex VPS34-VPS15 and the ATG16-ATG5-12 complex, which mediate multiple steps of autophagy, were markedly increased in the mitochondrial fraction on CCCP treatment in wild-type cells ( Fig. 3e; Supplementary Fig. 3L). These proteins were more abundantly detected in mitochondrial extracts from non-treated AMPK À / À cells compared with wild-type cells, likely due to the accumulation of damaged mitochondria and ubiquitinated proteins in these cells (Fig. 1a,e). However, in contrast to wild-type cells, CCCP treatment led to a slight decrease in the VPS34 complex and a loss of ATG16 complex from the mitochondria in AMPK À / À cells ( Fig. 3e; Supplementary  Fig. 3L). Further, while recruitment of core autophagy regulators to mitochondria requires AMPK, both AMPKa and b were present in mitochondria in ATG5 À / À MEF cells (Fig. 3f), suggesting that AMPK association with mitochondria is not dependent on association with ATG5 or autophagy, which is dependent on ATG5. Thus, our data are consistent with AMPK being required for retaining VPS34 and ATG16 complexes in damaged mitochondria on depolarizing stress.
Mitochondrial association requires N-myristoylation of AMPKb. N-myristoylation of AMPKb subunits is essential for AMPK association with intracellular membranes 49 ; we posited that recruitment of the AMPK complex to mitochondria might be mediated through the same mechanism. In the majority (B72%) of transfected MEF and H23 cells, AMPKb1-GFP associated with vesicle structures, whereas the N-myristoylation-defective b1 mutant, AMPKb1G2A-GFP, exhibited either homogenous distribution or distinct perinuclear aggregates varying in size partially co-localized with mCherry-p62 ( Supplementary Fig. 4A). AMPKb1-GFP but not AMPKb1G2A-GFP co-localized with MTR, which was enhanced on CCCP treatment during autophagosome clamping with bafilomycin A1 (Supplementary Fig. 4B,C).
Consistent with the recruitment of AMPKb1 but not AMPKb1G2A to the mitochondria, we detected robust AMPKb1 myristoylation in H23 cells in the mitochondrial fraction following labelling with amidomyristate using click chemistry (Fig. 4a). A brief exposure to CCCP led to a 1.6-fold increase in myristoylated AMPKb1 in the mitochondrial fraction with a modest increase in overall levels of mitochondrial myristoylated proteins (Fig. 4b). We then examined whether protein myristoylation is required for damage-induced mitochondrial removal using the myristate analogue, 2-hydroxymyristic acid (2HMA). 2HMA inhibited mitochondrial protein myristoylation more effectively than cytosolic proteins as determined by click chemistry, with stronger inhibitory effects being achieved in the cells treated with CCCP (Fig. 4c). Strikingly, inhibition of myristoylation effectively abrogated membrane recruitment of the AMPK complex and the loss of COX-IV and TOM20 in response to mitochondrial damage (Fig. 4d). Note that although the Miro-1 GTPase, previously shown to be a target of Parkinmediated proteolysis 52 , also requires myristoylation for membrane association, its levels were not affected by CCCP arguing against a role of Parkin-dependent proteolysis in this setting ( Fig. 4d; Supplementary Fig. 4D).
While myristoylation could potentially affect multiple mitochondrial proteins, the myristoylation-deficient form of AMPKb1 (AMPKb1G2A) was not present in mitochondria even in the absence of myristoylation inhibitor ( Supplementary Fig. 4E), suggesting that the association of AMPK with mitochondria is intrinsic to N-myristoylation of the b subunits rather than being mediated through the myristoylation of other proteins. Strikingly, autophagosome clamping with bafilomycin revealed multiple mitophagosomes following a brief exposure to CCCP in H23 cells but not in cells pre-treated with 2HMA with 2HMA-treated cells containing a larger number of damaged mitochondria ( Fig. 4e;  Supplementary Fig. 4F). These data suggest a crucial role for protein N-myristoylation in mitophagy.

Association of AMPK with mitochondria mediates mitophagy.
Since AMPKb N-myristoylation does not affect AMPK complex formation 49 , we postulated that the myristoylation-incompetent AMPKb1G2A-GFP, due to its inability to associate with membranes and specifically mitochondria, might function in a dominant-negative manner if membrane association is required for AMPK to mediate mitophagy. We thus examined the effects of AMPK mislocalization on autophagy selectivity. AMPKb1-GFP showed partial co-localization with mCherry-LC3 puncta in B31% of co-transfected cells; co-localization was enhanced over twofold on autophagosome clamping with bafilomycin A1 (Fig. 5a). AMPKb1G2A-GFP did not co-localize with LC3 puncta in either the presence or absence of bafilomycin A1 (Fig. 5a). Further, CCCP-induced TOM20 loss was abolished in H23 cells transfected with AMPKb1G2A-GFP but not the wildtype b1 subunit, whereas AMPKb1G2A-GFP did not affect p62 turnover, LC3 lipidation or rapamycin-induced non-selective autophagy ( Fig. 5b; Supplementary Fig. 5A). Thus, while myristoylation-mediated membrane association of the AMPK complex contributes to mitochondrial degradation, this mechanism is dispensable for non-selective autophagy.
We next determined whether forced mitochondrial localization of AMPK is sufficient to trigger mitochondrial removal. To constitutively target AMPK to mitochondria, we tagged AMPKb1-GFP N terminally with the mitochondrial localization signal of TOM20 with the resultant construct being designated Mito-b1-GFP (Fig. 5c). Mito-b1-GFP exhibited co-localization with MTR (Fig. 5d), suggesting that Mito-b1-GFP is exclusively mitochondrial in the absence of mitochondrial stress. Strikingly, mitochondrial alterations were observed in all cells transfected with Mito-b1-GFP as compared with 39% of non-transfected cells (Fig. 5d) or 36% of the cells transfected with TOM20-GFP ( Supplementary Fig. 5B,C), including loss of the peripheral portion of the mitochondria and disruption of mitochondrial network. To determine whether these aberrations might be due to mitophagy, we compared the effects of Mito-b1-GFP on mitochondria in wild-type versus ATG5 À / À MEFs. Similar phenotypes including loss of the bulk of the mitochondrial network were observed in wild-type MEFs but not ATG5 À / À cells ( Fig. 5e; Supplementary Fig. 5D), suggesting that the mitochondrial alterations caused by mitochondrial-targeted AMPK are dependent on ATG5, which is essential for autophagy and mitophagy. The mitochondrial changes induced by Mito-b1-GFP were recapitulated when TOM20-mCherry was used to mark mitochondria (Supplementary Fig. 5E). Thus, our data suggest that direct targeting of AMPK to mitochondria is sufficient to mediate autophagy-dependent mitochondrial removal.
The PINK-Parkin pathway mediates mitophagy with Parkin being recruited to mitochondria. We next determined whether AMPK-mediated mitochondrial removal requires Parkin. HCC38 human breast cancer cells, bearing a homozygous deletion of the PARK2 gene 53 , lack detectable expression of Parkin ( Supplementary Fig. 5F). In HCC38 cells, CCCP induced loss of mitochondrial proteins TOM20 and CHCHD3 to an extent comparable to that observed in H23 cells ( Supplementary  Fig. 5F-H). Transfection of Mito-b1-GFP led to loss of the peripheral portion of the mitochondrial network ( Supplementary  Fig. 5I), similar to the phenotype found in Parkin expressing cells (Myr-b1) co-precipitated with GFP-AMPKa2 following metabolic labelling of H23 cells with azidomyristate and in vitro conjugation with biotin alkyne. Input protein levels were determined by immunoblotting. Note that biotin-conjugated azidomyristoylated AMPKb1 migrated relatively slowly. (b) Western analyses of myristoylated AMPKb1 (Myr-b1) levels in the whole-cell lysate (WCL) and mitochondrial fraction (Mito) of H23 cells treated with or without CCCP following azidomyristic acid labelling and click chemistry. WCL without click conjugation was loaded as controls. Bafilomycin (Baf) was used to prevent autophagy/mitophagy-mediated protein degradation. LDHB was blotted for loading control. The bottom panel shows overall levels of protein myristoylation (streptavidin blot) in the mitochondria. ImageJ quantification was performed to assess the relative abundance of the proteins (fold changes). (c) Detection of overall levels of protein myristoylation in mitochondria (Mito) and cytosol (sup) following metabolic labelling and click chemistry in H23 cells treated with ( þ ) or without ( À ) 2-hydroxy myristic acid (2HMA) and CCCP. (d) Western analyses of protein levels from H23 cells treated without ( À ) or with ( þ ) CCCP (30 mM) for 4 h in the presence ( þ ) or absence ( À ) of 2HMA (0.5 mM; added 20 h before CCCP). WCL, whole-cell lysate; MF, membrane fraction extracted using the Cell Fractionation Kit, HT (Ab109718) from Abcam (Cambridge, MA, USA). Note that AMPK complex was not enriched in the membrane fraction containing the plasma membrane and other organelles in addition to the mitochondria. ** denotes absence of vimentin in membrane fraction and presence of a faster migration band of non-characterized origin in 2HMA-treated cells. (e) Representative transmission electron microscopy micrographs of H23 cells treated with ( þ ) or without ( À ) 2HMA (0.5 mM) for 24 h before brief exposure to CCCP (15 mM) in the presence of bafilomycin A (100 nM). Red arrows, autophagosomes; blue triangles, mitochondria; scale bar, 500 nM. (Fig. 5d,e; Supplementary Fig. 5E). Further, knockdown of AMPKa impaired CCCP-induced loss of mitochondrial proteins in HCC38 cells ( Supplementary Fig. 5J,K), consistent with a role of AMPK in mediating Parkin-independent mitophagy.
Mitophagy represents a surveillance mechanism against damaged mitochondria, which release reactive oxygen species and protein factors including cytochrome c that trigger the intrinsic apoptosis pathway. We posit that the inability to remove damaged mitochondria in a timely manner would cause apoptosis and potentiate the effects of apoptosis inducers. Indeed, incubation of H23 cells with 2HMA to prevent myristoylation of AMPK and subsequent recruitment to the mitochondria, inhibited cell growth effectively with prolonged exposure for 2 weeks eliminating the entire cell population (Fig. 6a), whereas myristic acid exhibited no cytotoxicity (Fig. 6b). Forced mitochondrial localization of AMPK using Mito-b1-GFP markedly improved cell viability (Fig. 6c), suggesting a role for mitochondrial AMPK in protecting cells from 2HMA-induced cell death. Further, 2HMA treatment for 48 h led to PARP1 cleavage and substantial loss of AMPK proteins (Fig. 6d;  Supplementary Fig. 6A). While 2HMA alone induced a modest level of apoptosis-associated PARP1 cleavage and etoposide alone induced PARP1 cleavage only when used at a high dose (5 mM), the combination of 2HMA with a sub-lethal dose of etoposide (2.5 mM) caused a marked increase in PARP1 cleavage in H23 cells, leading to substantial loss of intact PARP1 (Fig. 6d;  Supplementary Fig. 6A). These data suggest that protein N-myristoylation of AMPK promotes cell viability. To identify the enzyme(s) responsible for AMPK N-myristoylation, RNAi was performed to knockdown myristoyl-CoA protein Nmyristoyltransferase 1 (NMT1) and NMT2. Depletion of NMT1 but not NMT2 led to AMPK loss (Fig. 6e), recapitulating the effects of 2HMA. Thus, it is likely that NMT1, but not NMT2, mediates AMPKb N-myristoylation. Further, knockdown of NMT1 in H23 cells, while having little effect on autophagic flux as indicated by a comparable extent of CCCP-induced loss of p62 and induction of LC3-II, partially prevented the loss of TOM20 (Supplementary Fig. 6B). These data suggest a role of NMT1mediated myristoylation in selective removal of mitochondria. Interestingly, NMT1 but not NMT2 is overexpressed in B1/3 of human lung adenocarcinomas and high NMT1 mRNA levels correlate with statistically significantly poorer overall patient survival (Fig. 6f). The association of higher NMT1 mRNA levels   with poorer patient outcomes was also found in ovarian ( Supplementary Fig. 6C) and kidney cancers with progressively worsened patient survival linked to increasing NMT1 levels in renal clear cell carcinoma (Supplementary Fig. 6D). In all three tumour types where NMT1 levels predict poorer prognosis, high NMT1 mRNA expression strongly correlates with increased ACC phosphorylation (Fig. 6g-i), consistent with the AMPK-ACC cascade being downstream of NMT1, as AMPKb myristoylation also regulates stress-induced AMPK activation in LKB1 proficient cells 49 . Indeed, the correlation between NMT1 mRNA and pACC levels was maintained in lung cancer samples with wild-type LKB1 (P ¼ 0.0038) and LKB1 is rarely aberrant in either ovarian or renal cell cancer. Furthermore NMT1 knockdown in cancer cells expressing functional LKB1 led to a reduction in stressinduced ACC phosphorylation ( Supplementary Fig. 6E). Thus, although LKB1 status and function remain to be determined in respective cancer tissues, there appears to be a broad relationship between NMT1 and pACC.

Discussion
The energy-sensing function of the LKB1-AMPK cascade can mediate starvation-induced autophagy, which lacks substrate selectivity 54 , as well as mitochondrial biogenesis under energy stress [55][56][57] . We now provide several lines of evidence that AMPK plays a significant role in surveillance of mitochondrial stress by mediating selective removal of damaged mitochondria through autophagy. Our data, however, suggest that the role of AMPK in mitochondrial surveillance is independent of LKB1 and hence its energy-sensing function. Such functional specification can in principle mitigate signalling ambiguity by allowing selective removal of damaged mitochondria in the absence of energy stress and, conversely, preventing inadvertent mitochondrial depletion under nutrient starvation when mitochondrial function becomes critical for efficient bioenergetics.
In conjunction with the findings that AMPK regulates both selective and non-selective autophagy 39,46 , we propose the existence of multiple mechanisms whereby AMPK regulates autophagy selectivity. In rapidly proliferating cells, AMPK appears to be required for the vesicle elongation step to form mature autophagosomes but not the initial step of autophagy that generates phagophores. Pre-autophagosome vesicles were present in greater numbers in AMPK À / À than in wild-type cells (Fig. 1e), suggesting that AMPK may function to achieve autophagy selectivity, in part, by limiting the formation of nonselective phagophores. In addition, we showed that AMPK along with the class III PI3K VPS34 and the ATG16-ATG5-12 complexes, which mediate vesicle elongation and autophagosome maturation, were actively recruited to mitochondria on depolarizing damage. Notably, mitochondrial stress triggers a rapid association between AMPK and the ATG16 complex, which appears to mediate the retention of the ATG16 complex and potentially additional proteins at damaged mitochondria. Both of the VPS34 and ATG16 complexes have multiple phosphorylated components with VPS34 and BECN1 being direct substrates of AMPK 58 ; future studies are warranted to determine whether AMPK-dependent phosphorylation of these and additional targets plays a role in mediating mitophagy. Nonetheless, dynamic recruitment of AMPK, the VPS34 and ATG16 complexes to the mitochondria appears to facilitate the recruitment and lipidation of LC3, which is a substrate of the ATG16 complex 59 . In addition to being essential for LC3 lipidation, homotypic fusion of ATG16L1 precursors represents a critical step in autophagosome biogenesis before LC3 lipidation 59 . This notion is supported by the observation that targeting AMPK specifically to mitochondria was sufficient to induce loss of mitochondria in a manner that is dependent on ATG5, an essential component of the ATG16 complex (Fig. 5e). Thus, recruitment of AMPK to mitochondria may act to sense mitochondrial damage allowing assembly and activation of the autophagy machinery in temporal and spatial proximity to damage sites. This mechanism, although reminiscent of the translocation of Parkin to damaged mitochondria, differs from the role of the PINK-Parkin cascade, which appears to target damaged mitochondria to pre-formed autophagosomes through ubiquitination of mitochondrial proteins 14 . In addition, overexpression of Parkin can completely eliminate mitochondria in cells, which was not the case for AMPK. Indeed, our data suggest that AMPK can mediate mitophagy independent of Parkin, consistent with AMPK-mediated recruitment of the autophagy machinery to damaged mitochondria, representing a novel mitophagy pathway.
Of note, we found spatial association of AMPK with mitochondria and autophagosomes to be reliant on N-myristoylation of AMPKb, which mediates membrane association 49 . In contrast, non-selective autophagy was not dependent on AMPK N-myristoylation compatible with the different roles for AMPK in selective and non-selective autophagy, being at least in part dependent on localization to specific organelles. Interestingly, N-myristoylation of Sip1, the yeast version of the AMPKb subunit, is essential for the Snf1 complex to localize to vacuolar membranes on glucose withdrawal in Saccharomyces cereviciae 60 . Thus, this mechanism appears to be evolutionarily conserved. It is currently unclear how N-myristoylation mediates AMPK association with damaged mitochondria. Since protein lipidation is one of the mechanisms of sensing membrane curvature and stress-induced membrane curvature is important for the recruitment of autophagy machinery to form autophagosomes 61,62 , a plausible hypothesis is that membrane depolarization and the subsequent membrane curvature may allow the integration of the myristic acyl-chain into damage sites. Indeed, CCCP-induced mitochondrial fragmentation is characterized by the formation of circular mitochondria ( Supplementary Fig. 1J), which markedly increases the surface areas of the mitochondrial membrane and alters the curvature of the membrane both of which may increase the integration of N-myristoylated AMPK complex into the mitochondrial membrane.
In addition to AMPK, the AMPK kinase LKB1 is also regulated by protein lipidation with prenylation of LKB1 targeting its C-terminal cysteine residue within the CaaX motif. LKB1 prenylation appears to be crucial for the ability of LKB1 to activate AMPK 63 . Intriguingly, our data suggest that the role of AMPK in selective removal of damaged mitochondria is independent of the canonical energy-sensing function that requires functional LKB1. Thus, the different types of lipidation appear to play a critical role in determining AMPK function potentially by compartmentalizing AMPK to distinct membrane structures, as the membrane affinity of lipids varies depending on their acyl-chain lengths 64 .
Since surveillance of damaged mitochondria plays a role in sustaining cancer cell viability by pre-empting the intrinsic apoptosis pathway, the underlying mechanisms could represent intervention targets for cancer treatment. Before our study, drugs that inhibit protein N-myristoylation, which is increased in a variety of cancers, have been assessed in preclinical studies 65 . It is poorly understood, however, whether individual myristoylated proteins can be targeted specifically. Our data suggest that AMPK is specifically regulated by NMT1 but not NMT2 despite both enzymes being expressed in lung cancer cells. NMT1 mRNA is overexpressed in B34% of lung adenocarcinomas and correlates with worsened patient outcome with a 1.5-year reduction in median patient survival. Further, inhibition of protein N-myristoylation is effective in eliminating lung cancer cells at least in cell culture and sensitizes lung cancer cells to sub-lethal doses of etoposide, providing a proof of concept for targeting protein N-myristoylation for lung cancer treatment. Future drug-discovery studies are warranted to identify NMT1specific inhibitors, which may expand the therapeutic arsenal for lung cancer and other cancers, where high NMT1 mRNA levels are also associated with poorer patient outcomes.

Methods
Immunofluorescence, co-focal and electron microscopy. Immunofluorescent staining, phase contrast, fluorescence microscopy were performed as described previously 5,69 . For transmission electron microscopy, samples were fixed with 0.1 mol l À 1 cacodylate buffer (pH 7.3) containing 3% glutaraldehyde and 2% paraformaldehyde for 1 h, washed and treated with 0.1% Millipore-filtered cacodylate-buffered tannic acid, postfixed with 1% buffered osmium tetroxide for 30 min and stained en bloc with 1% Millipore-filtered uranyl acetate. The samples were then dehydrated in increasing concentrations of ethanol, infiltrated and embedded in Spurr's low-viscosity medium, followed by polymerization at 70°C for 2 days. Ultrathin sections were stained with uranyl acetate and lead citrate. A JEM 1010 transmission electron microscope (JEOL, USA) was used at an accelerating voltage of 80 kV and digital images were captured using the AMT Imaging System (Advanced Microscopy Techniques).
Reverse-phase protein microarray analysis. Tissue lysate microarray was processed as previously described 70 . Briefly, serially diluted lysates were spotted onto FAST slides (Schleicher & Schuell BioSciences, Keene, NH, USA) using a robotic GeneTAC arrayer (Genomic Solutions, Inc., Ann Arbor, MI, USA). After printing, the slides were blotted sequentially with Re-Blot (Chemicon, Billerica, MA, USA), I-Block and a biotin blocking system (Dako, Capinteria, CA, USA), probed with primary antibodies and incubated with biotin-conjugated secondary antibodies. The signals were then amplified using a Catalyzed Signal Amplification kit (DakoCytomation, Carpinteria, CA, USA) according to the manufacturer's instructions. The processed slides were scanned and quantified using the Microvigene software (VigeneTech Inc., North Billerica, MA, USA) and the quantitative values of 5 consistently expressed proteins (p38, JNK, ERK, mTor and GSK) as internal controls.
Bioorthogonal labelling and click chemistry. Cells were maintained in serumfree medium for 24 h before metabolic labelling with azidomyristate in the presence of 1% BSA (fatty acid free) using the Click-iT Metabolic Labelling kit (Life Technologies, Grand Island, NY, USA) according to the manufacturer's protocol. Following biotin conjugation, labelled proteins were detected by western blotting with streptavdin-HRP (Cell Signaling, Beverly, MA, USA).