Autophagy as a means of cell killing was first advanced by Clark's phenotypic description of ‘Type II autophagic cell death’ in 1990. However, this phenomenon later came into question, because the presence of autophagosomes in dying cells does not necessarily signify that autophagy is the cause of demise, but rather may reflect the efforts of the cell to prevent it. Resolution of this issue comes from a more careful definition of autophagy-dependent cell death (ADCD) as a regulated cell death that is shown experimentally to require different components of the autophagy machinery without involvement of alternative cell death pathways. Following these strict criteria, ADCD has been validated in both lower model organisms and mammalian cells, highlighting its importance for developmental and pathophysiological cell death. Recently, researchers have defined additional morphological criteria that characterize ADCD and begun to explore how the established, well-studied autophagy pathway is subverted from a survival to a death function. This Review explores validated models of ADCD and focuses on the current understanding of the mechanisms by which autophagy can kill a cell.
Autophagy is a regulated, catabolic process in which double-membrane vesicles, called autophagosomes, are formed de novo to engulf cytoplasmic content, which is then degraded upon fusion of the autophagosome with the lysosome. The molecular mechanism of autophagy, executed by the Atg genes, is well delineated (see Box 1), and has been extensively reviewed elsewhere (Bento et al., 2016; Mizushima et al., 2011; He and Klionsky, 2009). In the growing cell, autophagy is maintained at low basal levels, where it serves as a quality control pathway, eliminating long-lived proteins and damaged organelles. Autophagy is also induced in response to cellular stress, such as nutrient starvation, growth factor withdrawal and energy depletion. Through degradation and recycling of cellular components, autophagy supplies a continual source of metabolic building blocks to overcome the cellular deficiency. In both its basal and induced states, autophagy is necessary for cell survival and maintaining and/or restoring homeostasis.
Autophagosome initiation and elongation are controlled by the Ulk1 kinase complex, phosphatidylinositol-3-phosphate (PI(3)P) generation by the class III PI3K Vps34, the membrane-spanning protein Atg9 (green area in the figure) and ubiquitin-like conjugation pathways that ultimately produce lipidated LC3/Atg8 (LC3-II) at the surface of the phagophore membrane (blue area). The conjugation pathway involves covalent linkage of Atg12 to Atg5 by the E1 and E2-like enzymes Atg7 and Atg10, respectively, followed by conjugation of phosphatidylethanolamine (PE) to LC3; this requires Atg7, E2-like Atg3 and a complex between Atg16 and the Atg5–Atg12 conjugate, which functions as an E3 ligase, and is recruited to membranes by binding to WIPI2, a PI(3)P binding protein. The protease Atg4 primes LC3 for conjugation by cleavage of its C-terminus, and then facilitates its removal from autophagosomal membranes by deconjugating PE (pink area). Trafficking proteins, including Rabs and SNAREs, control the further maturation and movement of the autophagosome as it fuses with various endosomal components and eventually, the lysosome, to form the autolysosome, within which cargo is degraded and then released to the cytoplasm (pink area). Lysosomes are reformed from the autolysosomal membrane in a process known as autophagic lysosome reformation (ALR), setting the stage for a new cycle of autophagosome initiation (Chen and Yu, 2017).
Autophagy initiation is regulated by the central sensor of nutrients and energy, the mTOR kinase, from within the mTORC1 complex (Dunlop and Tee, 2014) (compare grey and green areas). mTOR phosphorylates several components of the Ulk1 complex, thereby inhibiting its activity, and suppressing autophagy (inhibitory phosphorylation; shown as a red ‘P’ in the Box figure). Upon nutrient starvation and/or growth factor withdrawal, mTOR is inactivated and Ulk1 is de-repressed. AMPK, which is activated following metabolic stress, modulates the autophagy pathway at several levels, including inhibiting mTOR activity directly, or by activating DAPK2 (activating phosphorylation; shown as a green ‘P’), which can inhibit mTOR by phosphorylating Raptor (Kim et al., 2013; Shiloh et al., 2018; Ber et al., 2015). AMPK also directly phosphorylates and activates Ulk1 (Alers et al., 2012) and activates the autophagy functions of PI3K by directly phosphorylating the Vps34-interacting protein Beclin-1 (Kim et al., 2013), or through DAPK2-mediated phosphorylation of Beclin-1, which blocks its interaction with the inhibitory Bcl-2 and Bcl-XL proteins (Shiloh et al., 2018). Phosphorylation of Beclin-1 is but one of multiple regulatory mechanisms modulating the Vps34–Beclin-1 complex, which include protein–protein interactions and post-translational modifications that affect its assembly, intracellular localization and kinase activity (see Backer, 2016).
The shut-down of the autophagic pathway is also regulated to restore autophagy to basal levels once the stress passes (yellow area). Upon its activation, Ulk1 phosphorylates and inactivates AMPK, thereby dampening the initiation signal (Alers et al., 2012). Furthermore, during starvation, the E3 ligase Cul3–KLHL20 is recruited to activated Ulk1, Vps34 and Beclin-1, which triggers their proteasome-mediated degradation (Liu et al., 2016a). As amino acid levels increase owing to autophagic activity, mTOR is reactivated, thereby not only shutting down the generation of autophagosomes, but also promoting ALR (Chen and Yu, 2017). Thus, the feedback mechanisms that terminate autophagy flux are built into the signaling pathway. Note that for simplification and due to space limitations many details are not included.
In certain circumstances, however, autophagy can lead to cell death. The recent recommendations of the Nomenclature Committee on Cell Death defines autophagy-dependent cell death (ADCD) as regulated cell death that depends on the autophagy machinery (i.e. pharmacological or genetic manipulations of autophagy genes block cell death), without involving alternative death pathways (Galluzzi et al., 2018). This is consistent with previous definitions of autophagic cell death (Shen and Codogno, 2011). The establishment of these criteria is critical, because autophagy is often observed in cell death scenarios, where it is activated in a failed effort to mitigate cell damage (Denton et al., 2012). In these latter cases, inhibition of autophagy promotes, rather than protects from cell death. However, proving causality is not sufficient, as in some cases autophagy activates other death pathways, resulting in a type of death that is autophagy dependent, but does not conform to the strict criteria of ADCD.
The acceptance of ADCD as a genuine death pathway is due to the development of specific tools for visualizing autophagy flux in a variety of contexts, and especially, to recent studies that have described the unique morphological characteristics of the process and identified molecular mediators of this pathway. This Review further extends the definition of ADCD, by surveying the recent developments in the field that have begun to explain mechanistically how a survival pathway becomes lethal, and how cell death is executed by over self-consumption.
Where does ADCD occur – validated incidents of ADCD
There are several examples in the recent literature of regulated cell death that fit the strict definition of ADCD. It should be noted, however, that ADCD is not as prevalent as a PubMed search for ‘autophagic cell death’ would suggest, as the vast majority of the studies claiming ADCD do not in fact meet the accepted criteria. In this section, we will survey examples of death that have been proven to require autophagy, in both lower organisms and mammals, during development and following pathophysiological insults.
Programmed cell death pathways, in particular apoptosis, are essential for removal and remodeling of tissues during development. Although mice with deletions of key autophagy genes (knockout, KO) develop normally and survive until the neonatal stage when they succumb to perinatal starvation (Mizushima and Levine, 2010), autophagy is needed for removal of unwanted cells as a back-up mechanism when apoptosis is blocked. For example, mice deficient for apoptosis [double knockout (DKO) of pro-apoptotic molecules Bak (also known as Bak1) and Bax] were viable, but when autophagy was also blocked (by Atg5 KO), mice showed embryonic lethality by day 13.5 with enhanced brain exencephaly (Arakawa et al., 2017). Specifically, cell death-dependent loss of interdigital webbing was even more delayed in the Bax/Bak DKO/ATG5 KO mice compared to DKO mice. In tissue from the DKO mice, but not that of the Bak single KO, in which apoptosis progressed normally, numerous autophagosomes were evident. This implies that in the absence of apoptosis, autophagy is activated to facilitate developmental cell death of the interdigital web (Arakawa et al., 2017).
Lower organisms provide additional examples of developmental ADCD. Under starvation conditions, the protist Dictyostelium discoideum aggregates into a multicellular organism, which then differentiates to form a spore-producing fruiting body in response to cAMP production. Autophagy provides nutrients and energy during this process and is thus necessary for cell survival. The cells forming the stalk of this organism then undergo developmental cell death, but not by apoptosis, as Dictyostelium does not possess any apoptosis genes, but by autophagy (Giusti et al., 2010). This requires a second signal that induces specific gene expression, involving either the stalk differentiation-inducing factor DIF-1 (in cell monolayers) or cyclic-diGMP (in the intact organism) (Chen and Schaap, 2016). In the presence of this second signal, mutation of atg1 blocks cell death (Luciani et al., 2011). Interestingly, DIF family morphogens have anti-tumorigenic properties in mammalian cells. For instance, when DIF-3 is introduced to human cancer cells, it induces mitochondrial depolarization and fission, autophagy and caspase-independent cell death (Dubois et al., 2016).
During Drosophila metamorphosis, developmental cell death of the obsolete larval midgut and salivary glands requires autophagy (see Box 2). Removal of the midgut has been defined as ADCD, because in contrast to the salivary gland, where apoptosis and autophagy function concurrently, only autophagy is necessary for midgut degradation. Notably, apoptotic genes are induced (Yin and Thummel, 2004), and caspase activity (mediated by the non-conventional effector caspase Decay) is detected; however, caspase inhibition or knockdown of decay does not affect removal of midgut structures (Denton et al., 2009). This may suggest a non-death role for caspases in this tissue. Despite the complexity of the Drosophila models of developmental death involving autophagy, they provide a unique in vivo system that is easy to manipulate genetically with well-defined phenotypic readouts, and have led to identification of molecular regulators of the process, as will be detailed further below.
During Drosophila metamorphosis, the steroid ecdysone triggers the removal of obsolete larval tissue by programmed cell death following waves of steroid-induced transcription (Yin and Thummel, 2005). Both apoptosis and autophagic genes are upregulated in response to ecdysone. Autophagic cell death has been implicated in the degradation of two larval tissues: the salivary glands and midgut. The first wave of ecdysone, during the late larval stage, triggers puparium formation and midgut cell death in the prepupa. Interestingly, while staining for DNA fragmentation is observed (TUNEL; i.e. a marker for apoptosis), none of the main initiator and effector caspases, except for Decay, is active in dying midgut, and expression of the caspase inhibitor p35 or knockdown of decay does not affect removal of midgut structures (Denton et al., 2009). In contrast, knockdown of various autophagy genes result in persistence of the tissue (Denton et al., 2009). However, the contribution of apoptosis and/or apoptosis genes to larval cell death cannot be totally dismissed, since disruption or mutation of both reaper (rpr) and head involution defective (hid) lead to persistent midgut structures (Yin and Thummel, 2004). Rpr and Hid function to de-repress fly caspases by inhibiting Drosophila inhibitor of apoptosis (Diap1). Diap1, which restrains apoptosis in the larva (Yin and Thummel, 2004), is degraded immediately after puparium formation, as a result of ecdysone-triggered induction of the PTP52F tyrosine phosphatase and subsequent dephosphorylation of TER94, a regulator of proteasome-mediated degradation. Induction of PTP52F and reductions in the levels of Diap1 are also necessary for autophagosome formation within the dying midgut (Santhanam et al., 2014).
Salivary gland removal occurs later, from 12–16 h post-puparium formation, in response to a second wave of ecdysone generated in the prepupa. Both apoptosis and autophagy are independently necessary for the developmental cell death of the salivary gland; the combination of caspase inhibition and Atg gene mutation results in greater persistence of the salivary gland than either condition alone (Berry and Baehrecke, 2007; Martin and Baehrecke, 2004). Furthermore, expression of Atg1 (the ortholog of mammalian Ulk1) induced salivary gland degradation in the absence of caspase activation, implying that autophagy by itself is sufficient to induce developmental cell death in the larval salivary gland (Berry and Baehrecke, 2007).
ADCD in pathophysiological conditions
ADCD has also been observed in pathophysiological conditions in mammals, in particular following insults that involve ischemia of the brain or heart (summarized in Table 1). Since systemic KOs of non-redundant Atg genes are lethal (Kuma et al., 2017), and even tissue-specific KOs lead to pathology (e.g. Koike et al., 2008), in order to show the necessity of autophagy for cell loss, researchers have resorted to non-specific drugs such as the phosphatidylinositol 3-kinase (PI3K) inhibitor 3-MA (Wen et al., 2008; Puyal et al., 2009; Feng et al., 2018), or less efficient lentiviral-based shRNA-mediated knockdowns (Matsui et al., 2007; Zheng et al., 2009; Ginet et al., 2014). Furthermore, the read-out is overall loss of viable tissue (i.e. infarct size) and not a direct quantification of the number of dead autophagic cells. Notably, apoptosis and necrosis can also be activated within the injured ischemic tissue, and in some cases, inhibition of autophagy also blocks apoptosis and/or caspase activation (Matsui et al., 2007; Koike et al., 2008; Zheng et al., 2009; Xie et al., 2016). At first glance, our strict definition of ADCD would not apply to ischemic injury. However, one needs to keep in mind that these are complex in vivo models involving a heterogeneous tissue that responds to heterogeneous stimuli, including the ischemic insult, the consequent inflammation and signals released by neighboring dying cells. In fact, one study showed that different death phenotypes occurred in different regions of the injured brain (Puyal et al., 2009). Thus, although these models are not perfect, owing to limitations in genetic manipulation and in vivo analysis, they are important in that they help to establish the relevance of autophagic cell death in higher organisms. Moreover, these studies support the premise that ADCD contributes to the pathology of ischemic injury, and as such, provides a unique target for therapeutic intervention.
Autophagy has also been shown to be necessary for cell death in certain cancer cell lines, especially those that are defective in apoptosis and resistant to apoptosis-driven chemotherapy. Thus, while autophagy is often necessary for survival of advanced tumors (Chen and White, 2011), in culture, certain drugs or drug combinations lead to ADCD, suggesting that activating autophagy may at times be beneficial for directly killing tumor cells or sensitizing them to additional chemotherapy (Fulda and Kögel, 2015). Notably, this is context dependent and only occurs in certain cancer cell lines (Chen et al., 2016; Nozaki et al., 2016). We refer the reader to a thorough review that has highlighted many studies showing ADCD by various agents in different cancer cell types, such as the BH3 mimetics obatoclax and gossypol, histone deacetylase inhibitors, as well as the natural plant products resveratrol and betulinic acid (Fulda and Kögel, 2015). Additionally, more recent studies that have clearly demonstrated autophagy gene-dependent cell death in the absence of apoptosis include the treatment of glioma cells with a combination of the tricyclic antidepressant imipramine (IM) and the anti-platelet drug ticlopidine (TIC; an inhibitor of ADP receptor P2Y12) (Shchors et al., 2015), the treatment of lung carcinoma cells with cabazitaxel (Huo et al., 2016) and the treatment of hepatocellular carcinoma with the natural flavonoid kaempferol (Han et al., 2017). Importantly, ADCD triggered by IM and TIC has also been associated with slower tumor progression and improved survival in in vivo mouse glioma models; knockdown of Atg7 in the glioma cells attenuates the effects of the drugs on survival and tumor growth (Shchors et al., 2015).
Despite the fact that some models of ADCD are more established than others and fit better to the definition of ADCD, the above literature survey shows that autophagy is a physiologically relevant mechanism for cell killing, and has inspired further investigation into the subcellular characteristics of the process and its underlying mechanisms. In particular, two validated models of ADCD, insulin withdrawal in adult hippocampal neural (HCN) stem cells (Ha et al., 2015; Yu et al., 2008) and resveratrol (RSV) treatment of the lung cancer cell line A549 (Dasari et al., 2017a), have been used to dissect the molecular regulation of ADCD. Both models fulfill the strict criteria of ADCD, as cell death depends on several autophagic genes and does not involve other cell death modalities. In the following sections, we will discuss the phenotypical and mechanistic aspects of ADCD, highlighting these studies and others.
How does ADCD occur – models of ADCD and their cellular hallmarks
Careful phenotypic analysis of cells undergoing ADCD has delineated the hallmarks of the process that distinguish it from other types of programmed cell death. Furthermore, these studies have led to possible explanations of how autophagy can actually kill the cell. These recent analyses have, for the most part, been performed in cell culture, which enables careful examination of the phenotype and its dynamics over time, an aspect that is more limited in vivo. Nevertheless, some of the phenotypic characteristics of ADCD have been observed in pathophysiological conditions in vivo as well. Of note, the scenarios described here should be differentiated from those in which autophagy is necessary to promote alternative death pathways, which, although dependent on autophagy gene function, do not fit the definition of ADCD (see Box 3).
In some scenarios, autophagy is not directly a cause of cell demise, but rather important for the activation of other cell death programs. Thus, this form of cell death can be mitigated by inhibition and/or genetic manipulation of autophagy and other cell death types, such as apoptosis or necroptosis. Although in these cases, cell death is autophagy dependent, it does not fit the definition of ADCD. For example, autophagy can be rerouted towards apoptosis or necroptosis when it is activated to high levels, but is defective in its last stages of fusion with the lysosome or autolysosomal degradation (see ‘stalled autophagy’ scheme in the figure) (Degtyarev et al., 2008; Gonzalez et al., 2012; Choi et al., 2016).
Autophagy can also activate alternative death pathways through the selective targeting of proteins that are limiting and necessary for cell function and viability via cargo-specific receptors, a scenario that requires full autophagy flux (see ‘selective autophagic degradation’). Examples of such degradation targets include K-Ras, which sensitizes cancer cells to tamoxifen owing to a reduction in MAPK signaling survival pathways (Kohli et al., 2013), the ROS scavenger catalase, which leads to ROS accumulation and necrosis (Yu et al., 2006), the intracellular iron sequestering protein ferritin, which gives rise to an intracellular iron overload and so triggers ferroptosis (Gao et al., 2016; Hou et al., 2016), and, conversely, the transferrin receptor, which leads to reduced iron uptake and apoptosis (Sirohi et al., 2013). Autophagy-mediated degradation of inhibitors of death pathways has also been shown to activate apoptosis or necroptosis. Such factors include the phosphatase FAP-1 (also known as PTPN13), which inhibits Fas signaling, degradation of which promotes apoptotic death in certain Fas-ligand treated human cells (Gump et al., 2014). The inhibitor of apoptosis protein (IAP) family is also targeted for selective degradation by autophagy: the Drosophila IAP dBRuce is degraded during Drosophila oogenesis, thus activating apoptotic death (Nezis et al., 2010). Similarly, autophagic degradation of IAPs in cancer cells has been shown to lead to the assembly of the RIPK1–RIPK3 necrosome complex, and consequently, to trigger necroptosis (He et al., 2014).
Finally, autophagy can lead to cell death independently of its normal catabolic role; here, activity of the full pathway down to the degradative step is not necessary (see ‘scaffolding function’). Instead, phagophore and/or autophagosome membranes can act as scaffolds for the assembly of death-inducing complexes, for instance, leading to caspase-8 activation and/or intracellular death-inducing signaling complex (DISC) formation (Young et al., 2012; Laussmann et al., 2011; Pyo et al., 2005; Jin et al., 2009), or the recruitment of RIPK1 and RIPK3 and intracellular necrosome formation (Goodall et al., 2016; Yu et al., 2006; Basit et al., 2013).
Death from over-eating – elimination of intracellular organelles and cytosol through excessive bulk autophagy
Our recent work thoroughly characterized the morphology of A549 cells in which ADCD was induced by RSV (Dasari et al., 2017a). At later time-points (48–72 h), as autophagy flux continued, we observed that the cytoplasm is overwhelmed by the presence of autophagic and empty vacuoles that were proven to be late-stage autolysosomes. Notably these cells are almost devoid of any intracellular organelles, including endoplasmic reticulum (ER), Golgi and mitochondria, reaching a final stage in which the area of autophagic vacuoles exceeded the cytoplasmic area. Membrane dysfunction is also apparent, specifically in the form of bulges in the nuclear membrane with enlarged perinuclear space and occasionally, nuclear shedding (Dasari et al., 2017a). These observations, conducted in a systematic manner in a classic ADCD model, support the long-held hypothesis that, during ADCD, lethality may result from extreme levels of autophagy flux that lead to overconsumption of cellular organelles and rerouting of cellular membrane sources to support autophagosome generation, to the point where cellular membrane homeostasis is disturbed (Fig. 1A).
Death from excessive mitophagy
Specific targeting of mitochondria by selective autophagy, known as mitophagy, may also lead to cell death when it is excessive, owing to the failure of mitochondria-depleted cells to generate energy (Fig. 1B). Mitophagy is a quality control process: damaged, depolarized mitochondria are removed, thereby limiting the generation of reactive oxygen species (ROS) and release of apoptogenic factors and blocking cell death (e.g. Wang et al., 2017; Qiao et al., 2016; Yu et al., 2018b; Di Rita et al., 2018; Li et al., 2018). However, mitophagy is only beneficial to the cells if it occurs in a limited and regulated manner, and its over-activation can be lethal, as outlined below. The most prominent hallmarks of death by mitophagy are the association of autophagy markers with the mitochondria, concomitant with the selective reduction of the mitochondrial compartment in the dying cells.
One of the first reports connecting mitochondrial damage to death by autophagy involved smARF, an alternative translation product of p19ARF (p14ARF in humans, also known as CDKN2A) that localizes to mitochondria. Overexpression of smARF led to loss of mitochondrial membrane potential and caspase-independent cell death that was blocked by knockdown of autophagy genes (Reef et al., 2006). Subsequently, it was shown that deregulated smARF expression induced PTEN-induced putative kinase 1 (PINK1)-Parkin-dependent mitophagy (Budina-Kolomets et al., 2013; Grenier et al., 2014). Ironically, the excessive autophagy induced by enhanced smARF protein levels, which results in the degradation of the p62 (also known as SQSTM1) cargo receptor, was shown to block further activation of p62-dependent mitophagy in response to stress, and in this manner may contribute to tissue loss and mortality during hypoxia and endotoxic shock (Zhang et al., 2015).
In other work, expression of the orphan nuclear receptor TR3 led to loss of mitochondrial membrane potential following TR3 translocation to the mitochondria, which resulted in the induction of mitophagy, and consequently, cell death (Wang et al., 2014). In HCN cells, insulin withdrawal leads to AMPK activation, which then phosphorylates the p62 cargo receptor on a unique site that promotes its translocation to the mitochondria and mitophagy (Ha et al., 2017). Likewise, treatment of neuronal cells with the neurotoxins 1-methyl-4-phenylpyridinium [MPP(+)] or 6-hydroxydopamine (6-OHDA) was shown to induce cell death and autophagy. Death was subsequent to the localization of extracellular signal-regulated kinases (ERK1/2) to the damaged mitochondria, which was sufficient to induce mitophagy (Dagda et al., 2008; Zhu et al., 2007). Significantly, KD of Atg genes partially blocked neuronal cell death that was induced by MPP(+) (Zhu et al., 2007). Another example of cell death caused by mitophagy involved the induction of ceramide stress in head and neck squamous cell carcinoma cell lines. In this system, the introduction of a C18 ceramide analog, or overexpression of ceramide synthase 1, led to mitophagy and caspase-independent cell death (Sentelle et al., 2012). Here, oxidative respiration and ATP production were impaired, which is consistent with the profound loss of mitochondria, providing a possible cause of cell lethality.
Hypoxia has been shown to induce mitophagy in cultured spinal cord neurons (Yu et al., 2018a). Knockdown of Bcl-2/E1B-19KD-interacting protein 3 (BNIP3), which recruited LC3 (also known as MAP1LC3) to the mitochondria, blocked mitophagy and restored declining ATP levels and cell viability following hypoxia. Furthermore, BNIP3 knockdown prevented loss of spinal cord neurons and improved long-term locomotor function following spinal cord injury in rats. Without showing mitophagy in the in vivo model, the authors extrapolated from the cell culture data and concluded that mitophagy also contributes to neuronal cell death in vivo (Yu et al., 2018a). Another recent study has provided evidence for mitophagy in infarcted tissue from adult rats following cerebral ischemia–reperfusion injury. Notably, infarct size was reduced by administration of either the non-specific autophagy inhibitor 3-MA, or a specific inhibitor of Drp1, a factor that is necessary for mitochondrial fission and the recruitment of the mitophagy regulators PINK1 and Parkin to the mitochondria (Feng et al., 2018). Of note, as it is much more difficult to demonstrate mitophagy in vivo, and, furthermore, to prove that it specifically causes cell death, there are not many unambiguous examples in the literature of lethal mitophagy in pathophysiological scenarios.
Death from autosis
Autosis is an interesting type of ADCD that is induced by the administration of high dosages of cell-permeable peptides that stimulate high levels of autophagy (Liu et al., 2013). Significantly, a small percentage of cells can be observed showing hallmarks of autosis in several models of nutrient starvation, i.e. starved cultured HeLa cells, neonatal rat hippocampal neurons following in vivo cerebral hypoxia-ischemia and the liver of anorexia-nervosa patients (Liu et al., 2013; Kheloufi et al., 2015). Autosis is morphologically distinct from the ADCD described above, although it shares some features (Fig. 1C). In its early stages, autosis is characterized by the accumulation of autophagosomes and autolysosomes and nuclear convolution with moderate chromatin condensation, as well as abnormal mitochondria structures and ER fragmentation and dilation (Liu et al., 2013). Eventually, the outer and inner nuclear membranes separate with ballooning of the perinuclear space, and a necrotic-like phenotype is observed, including swollen organelles and rupture of the plasma membrane at distant foci, yet cells remain strongly adherent to the plate. Inhibition of lysosome and/or autophagosome fusion does not block autosis or cell death (Liu et al., 2013); this is in contrast to other models of ADCD, for example, RSV treatment of A549 cells, in which KD of lysosomal enzymes rescue cell viability (Dasari et al., 2017a). Therefore, it is likely that autosis pathology does not involve full autophagy flux and additional factors elicit the lethal phenotype. A screen of bioactive compounds for possible inhibitors of autosis led to the identification of cardiac glycosides that act as inhibitors of the Na+/K+-ATPase, and, indeed, KD of one of the ATPase subunits similarly blocked autophagy and cell death induced by the peptide (Liu et al., 2013). Thus, in this model of ADCD, autophagosome accumulation, but not necessarily autophagy flux, is associated with a secondary necrosis-like phenotype, which may involve changes in ion transport and osmolarity that are mediated by the Na+/K+-ATPase pump, eventually leading to cell death.
Why does ADCD occur – molecular regulators of ADCD
Why does autophagy become lethal to the point where it over-consumes the cell and its components? This question cannot yet be fully answered, but several clues have emerged from the recent literature that suggest various possible mechanisms (summarized in Fig. 2). The first mechanism involves changes in the regulatory steps that affect autophagy induction, either through modulation of the Vps34–Beclin-1 complex (see below and Box 1; note Vps34 is also known as the class III PI3K or PIK3C3 in mammals), or by activation of AMP-activated protein kinase (AMPK), which has multiple targets within the autophagy initiation pathway (Ouyang et al., 2017; Law et al., 2017; Han et al., 2017; Ha et al., 2017; Shiloh et al., 2018) (mechanism 1 in Fig. 2). In other scenarios, autophagy is induced independently of its canonical regulators, presumably in a way that leads to hyper-activation of the down-stream executors (mechanism 2 in Fig. 2). Inherent differences in the molecular machinery that executes autophagosome generation have also been observed, and regulators that are specifically associated with lethal autophagy have been identified, although their precise functions are not yet understood (and therefore cannot yet be mapped to the scheme shown in Fig. 2). Some of these may serve as second signals that, in combination with the ‘normal’ induction of autophagy, lead to lethal autophagy (mechanism 3). Additional studies highlight the contribution of a misregulated termination of autophagosome biogenesis, which usually serves to limit the process (mechanism 4). For example, depletion of the E3 ligase KLHL20 prevented the degradation of Ulk1 (see Box 1), Vps34 and Beclin-1, leading to sustained autophagy; this contributed to starvation-induced cell death and exacerbated diabetes-associated muscle atrophy (Liu et al., 2016a). Thus, the amplitude and duration of autophagy, as determined by the protein levels and/or activity of its key regulators, can determine whether autophagy facilitates cell survival, or, in fact, leads to cell death, as discussed below in detail.
Hyper-activation of autophagy by the Vps34–Beclin-1 complex
Several studies are consistent with autophagic cell death being a hyper-activated form of autophagy that results from releasing the brakes on autophagy induction, specifically the function of the Vps34–Beclin-1 complex (Fig. 2, mechanism 1). BNIP3, which sequesters the Beclin-1 inhibitor Bcl-2 and thus activates Beclin-1, was shown to be a positive mediator of autophagic cell death (Daido et al., 2004; Azad et al., 2008; Kanzawa et al., 2005). Similarly, the Na+/K+-ATPase inhibitor ouabain-induced autophagic cell death in non-small cell lung carcinoma cells by down-regulating Bcl-2 (Trenti et al., 2014). Downregulation of Bcl-2 and Bcl-XL (encoded by BCL2L1) and upregulation of Beclin-1 was likewise observed in the ADCD model that involved insulin-withdrawal in HCN cells (Yu et al., 2008). In myeloma cells, autophagic cell death was suppressed by caspase-10, which cleaved BCLAF-1, a Beclin-1 activator that interferes with the interaction between Beclin-1 and Bcl-2 (Lamy et al., 2013). Interestingly, in a multiple myeloma cell line (IM-9) that overexpressed high levels of Bcl-2, betulinic acid-induced ADCD was mediated by activation of death-associated protein kinase 1 (DAPK1) by protein phosphatase 2 (PP2A) (Zhou et al., 2017). Consistent with previous results (Zalckvar et al., 2009), DAPK1 phosphorylated Beclin-1, which led to its dissociation from Bcl-2 and reciprocal association with Vps34, thus triggering autophagy flux (Zhou et al., 2017). Taken together, these studies imply that affecting the interaction between Beclin-1 and Bcl-2, or their relative expression levels, amplifies the activation of the autophagy pathway so that it becomes a lethal one.
Non-canonical activation of autophagy
In the case of RSV-induced ADCD in A549 cells, we observed that autophagy was independent of Ulk1 or Beclin-1, suggesting that a non-canonical pathway activated the downstream Atg proteins, possibly involving the Golgi as a membrane source for autophagosome generation (Dasari et al., 2017a). Similarly, ERK1/2-dependent mitophagy and cell death of SH-SY5Y neuroblastoma cells in response to MPP(+) or 6-OHDA also occurs independently of Beclin-1 or Vps34 activity, which contrasted with starvation-induced autophagy in these same cells (Dagda et al., 2008; Zhu et al., 2007). In both these systems, autophagy activation bypassed the critical regulatory step that is conferred by the Vps34–Beclin-1 complex (Fig. 2, mechanism 2), suggesting that autophagy proceeded in a manner that was less sensitive to a self-imposed brake. Indeed, under these conditions, autophagy was not only more likely to be over-activated at the induction stage, but was also uncoupled from the negative-feedback loops that involve Vps34 and Beclin-1, which normally serve as termination signals. However, to date, very little is known about the exact mechanism underlying non-canonical, Vps34- and Beclin-1-independent autophagy and how it might be regulated.
Death-specific autophagy factors
Autophagic cell death may require the activation of additional specialized mechanisms, which involve unique regulators that direct the pathway towards cell killing. Several studies in Drosophila have identified factors that are specifically necessary for autophagy in the dying salivary gland, but not for survival-associated starvation-induced autophagy in the larval fat body. These include the cell surface engulfment receptor Draper (drpr) (McPhee et al., 2010), its presumed activating ligand, macroglobulin complement-related (mcr) (Lin et al., 2017), and the Ral GTPase and all subunits of its effector, the exocyst complex (Tracy et al., 2016). Whether the role of these factors can be generalized to other systems of ADCD, specifically in mammals, is unknown. Interestingly, mammalian RalB and a subcomplex of the exocyst have been previously shown to activate Ulk1 and Vps34, as well as autophagosome nucleation in response to starvation, under conditions when autophagy functions as a survival pathway (Bodemann et al., 2011). Therefore, the function of Ral and exocyst in autophagic cell death may be specific to the fly.
ADCD that is induced by insulin deprivation in HCN cells involves the activation of glycogen synthase kinase-3β (GSK-3β) downstream of insulin receptor signaling (Ha et al., 2015), and also requires release of Ca2+ from ER stores, as a result of increases in expression of the Ryr3 ryanodine receptor (Chung et al., 2016). However, it is not known how either an increase in cellular Ca2+ or active GSK-3β links to lethal autophagy. Further investigations have implicated valosin-containing protein (VCP) as a factor that differentiates lethal autophagy from survival autophagy (Yeo et al., 2016). VCP is a multi-function hexameric AAA ATPase that is involved in the remodeling of its ubiquitylated client proteins to facilitate their extraction from membranes or protein complexes. Depending on different cofactors, it has been shown to be involved in a variety of processes, including protein degradation, activation of transcription factors and membrane fusion (Meyer et al., 2012). Normally, VCP is required for the proper maturation of autophagosomes into degradative autolysosomes (Ju et al., 2009); VCP inhibition during basal autophagy leads to accumulation of the lipidated membrane-associated form of LC3 (LC3-II; see Box 1), similar to inhibition of lysosomal function. However, knockdown or inhibition of VCP in insulin-deprived HCN cells led to a reduction in LC3-II, even in the presence of lysosomal inhibitors, and attenuated the formation of DFCP1 puncta (a biomarker for the accumulation of PI3P at the membrane), suggesting that VCP is necessary for an early stage of autophagosome initiation in this system (Yeo et al., 2016). Interestingly, the Drosophila VCP, TER94, has been previously linked to autophagy-dependent developmental cell death of the larval midgut (see Box 2) (Santhanam et al., 2014).
In addition to unique regulators of the initiation stages, there may be intrinsic differences in the pathways that execute autophagosome formation during pro-survival or maintenance autophagy and autophagic cell death. For example, autophagy associated with Drosophila developmental midgut cell death requires different components from survival-associated autophagy. Specifically, initiation of autophagy in the dying midgut occurs in the established manner (i.e. it requires the Atg1 complex, the Vps34 PI3K complex and Atg9), but does not use the canonical conjugation proteins (i.e. Atg4, Atg3, Atg7, Atg5, Atg12 and Atg16; see Box 1) to lipidate Atg8 (the homolog of LC3 in mammals) (Xu et al., 2015; Chang et al., 2013). Interestingly, the Uba1 E1 ubiquitylation enzyme was shown to be necessary for formation of Atg8 puncta (Xu et al., 2015). This contrasts with survival autophagy that is induced by starvation in the fat body or midgut, which required all Atg genes, including those involved in Atg8 conjugation, but not Uba1 (Xu et al., 2015; Chang et al., 2013).
Second signals – the sphingolipid and ceramide connection
We have applied an unbiased, signalome-wide shRNA viability screen to RSV-treated A549 cells in order to identify novel genes that are necessary for cell death in this system (Dasari et al., 2017a). One of the top hits was GBA1 (also known as GBA), which encodes the lysosomal enzyme glucocerebrosidase (GCase). GCase metabolizes glucosylceramide (GlcCer) to ceramide and glucose and is an important component of the lysosomal salvage pathway for ceramide production. Significantly, GBA1 has previously been linked to autophagy; loss-of-function mutations in the GBA1 gene are responsible for Gaucher disease and are also a risk factor for Parkinson disease, both of which are likely to involve defective autophagy (Kinghorn et al., 2017). Prolonged RSV treatment (48 h) leads to increases in the protein and enzymatic levels of GCase, which correlates with decreased cell viability (Dasari et al., 2017a). In parallel, the amounts of long-chain ceramides and of several ceramide metabolites also increased. The timing of GCase induction, and the fact that it is not observed following starvation, which induces survival autophagy, suggests that GCase is important for the sustained, excessive autophagy flux that becomes lethal. Furthermore, KD of GBA1 suppresses autophagosome formation and reverses the dramatic ultrastructural changes to cell morphology and organelles, thereby protecting cells from cell death. Likewise, the Drosophila GBA1 ortholog, Gba1a, is necessary for larval midgut regression, indicating a conserved function in mediating ADCD (Dasari et al., 2017b). This is one of the few factors identified in cell culture models so far that have been shown to be relevant for ADCD in vivo.
Interestingly, inhibition of GCase activity, used to model Gaucher disease, leads to changes in the biophysical properties of the plasma membrane, including decreased rates of clathrin-mediated endocytosis, as well as alterations in membrane fluidity and motility of lipids and membrane proteins (Batta et al., 2018). One can hypothesize that elevations in GCase activity would have opposite effects on membrane properties. Along these lines, KO of ceramide synthase 2, which leads to a shift from the production of very long acyl-chained sphingolipids (C22, C24) to long chain sphingolipids (C16, C18), resulted in tissue-dependent alterations of membrane properties that included changes in fluidity, membrane curvature and morphology (Silva et al., 2012). Because autophagy is highly dependent on membrane fusion and trafficking events from various membrane compartments, including plasma membrane, endosomes, ER and Golgi, for phagophore formation, membrane elongation and closure, and further maturation into the autolysosome (Bento et al., 2016), the composition and biophysical properties of cellular membranes should be critical factors for autophagosome generation and flux.
There may also be a more direct connection between high levels of ceramide, or changes in ceramide metabolism, to lethal autophagy, in that ceramide could serve as a second signal that leads to ADCD, in analogy to stalk cell death in Dictyostelium. Indeed, a recent study in myeloid cells showed that administration of an inhibitor of lysosomal acid ceramidase, which generated elevations in C16 ceramide, resulted in the accumulation of autophagosomes, and, ultimately, in apoptosis- and necrosis-independent cell death (Liu et al., 2016b). This is also consistent with the study discussed above, which showed that ceramide stress leads to ADCD, specifically through mitophagy (Sentelle et al., 2012). There, several long-chain ceramide species, such as C16, C18 and C24, as well as sphingosine, were shown to preferentially bind to lipidated LC3B (MAP1LC3B) and to also enhance its lipidation independently of the interaction. Notably, the levels of sphingosine and C16, C18 and C24 ceramide, among other long-chain species, increased upon RSV-mediated GCase activation (Dasari et al., 2017a). Furthermore, C18 ceramide, but not other ceramide species, localized to the mitochondria, where it functioned as a receptor for autophagosome-bound LC3-II to facilitate mitophagy, providing a precedent for the direct involvement of long-chain ceramides in autophagosome formation under cell death conditions (Sentelle et al., 2012). Further research into this aspect of ceramide signaling should be particularly enlightening.
Conclusions and future perspectives
Much progress has been made in the field of ADCD in recent years, establishing it as a genuine type of programmed cell death. The study of apoptosis, and later, programmed necrosis, has been driven forward by the discovery of molecular regulators of these pathways. Similarly, we expect that the switch in emphasis from descriptive to mechanistic studies on ADCD will advance its prominence as a genuine cell death program. More efforts need to be placed on understanding the nature of putative second signals that drive autophagy to become lethal, as well as the mechanism of non-canonical autophagy, and elucidating how a misregulation of autophagy initiation can lead to overactivation. The recognition that ADCD plays a significant role in the pathology of ischemia and suppression of certain tumor growth will certainly lead to further insights into the underlying mechanisms. Of note, with the exception of developmental cell death studies in Drosophila, our current preliminary understanding of the molecular regulation of ADCD has been derived from cell culture studies. For the most part, it is not yet known what relevance these factors have to in vivo pathophysiological and developmental scenarios. This will need to be established, even though such in vivo systems are more difficult to analyze and genetically manipulate than the currently used in vitro models.
To date, non-specific autophagy inhibitors, such as the commonly used anti-malaria drug chloroquine, are used in clinical and pre-clinical settings. The development of more specific autophagy inhibitors is warranted for the treatment of ischemia-related pathologies. As our understanding of how survival and lethal autophagy differ, inhibitors of lethal autophagy might be developed to specifically inhibit ADCD, without affecting maintenance autophagy. Conversely, the development of activators of ADCD in the context of cancer has the potential to greatly enhance the arsenal of anti-neoplastic chemotherapy. Thus, advances in ADCD are eagerly awaited on many fronts.
Our work in this area is supported by a grant from the European Research Council under the European Union's Seventh Framework Program (FP7/2007-2013/ERC, grant agreement 322709).
The authors declare no competing or financial interests.