In vivo genetic dissection of tumor growth and the Warburg effect

A well-characterized metabolic landmark for aggressive cancers is the reprogramming from oxidative phosphorylation to aerobic glycolysis, referred to as the Warburg effect. Models mimicking this process are often incomplete due to genetic complexities of tumors and cell lines containing unmapped collaborating mutations. In order to establish a system where individual components of oncogenic signals and metabolic pathways can be readily elucidated, we induced a glycolytic tumor in the Drosophila wing imaginal disc by activating the oncogene PDGF/VEGF-receptor (Pvr). This causes activation of multiple oncogenic pathways including Ras, PI3K/Akt, Raf/ERK, Src and JNK. Together this network of genes stabilizes Hifα (Sima) that in turn, transcriptionally up-regulates many genes encoding glycolytic enzymes. Collectively, this network of genes also causes inhibition of pyruvate dehydrogenase (PDH) activity resulting in diminished ox-phos levels. The high ROS produced during this process functions as a feedback signal to consolidate this metabolic reprogramming. DOI: http://dx.doi.org/10.7554/eLife.18126.001


Introduction
During cancer formation, developing tumor cells acquire multiple biological capabilities that ultimately lead to malignancy. These events include, sustained proliferation, resistance to cell death, induction of angiogenesis, cellular metastasis and a reprogrammed energy metabolism (Hanahan and Weinberg, 2011;Pavlova and Thompson, 2016). Warburg discovered that many cancer cells reprogram their glucose metabolism by transitioning from oxidative phosphorylation to glycolysis even in the presence of oxygen (Warburg, 1956a(Warburg, , 1956bVander Heiden et al., 2009). Such a metabolic state was termed 'aerobic glycolysis' and the ability of cancer cells to acquire this new metabolic state has since been referred to as the 'Warburg effect'. The phenomenon was understudied for decades, as it became clear, that contrary to Warburg's assertions, cancers were largely attributable to oncogenes and tumor suppressors, rather than to exclusive changes in metabolic status (Huebner and Todaro, 1969;Stehelin et al., 1976;Martin, 2001;Knudson, 1971;Shih and Weinberg, 1982). However, more recent studies have explored the link between metabolic processes and oncogenesis, and have noted that altered metabolism is an important element that contributes to the etiology of cancer (Pavlova and Thompson, 2016). Drugs targeting key regulators of aerobic glycolysis are being developed to be included in the cancer therapy regimen (Weinberg and Chandel, 2015;Galluzzi et al., 2013). Several oncogenic pathways, including PI3K/ TOR, JNK, Ras/ERK, regulate the catalytic activity or expression of key metabolic enzymes (Chambers and LoGrasso, 2011;Jones and Thompson, 2009;DeBerardinis et al., 2008).
Perhaps not any longer universally supported by modern evidence in cancer-metabolism, Warburg had also proposed that cancer cells undergo a glycolytic shift for the purpose of generating the bioenergetic makeup of the rapidly dividing cell (Warburg, 1956a). Pyruvate is the key metabolite that is used to control the last step of glycolysis in a tumor, and in the presence of lactate dehydrogenase (LDH), pyruvate is converted to lactate. In contrast, oxidative phosphorylation requires the mitochondrial enzyme complex pyruvate dehydrogenase (PDH) that converts pyruvate to acetyl-CoA, essential for the initiation of the tricarboxylic acid (TCA) cycle (Linn et al., 1969;Leiter et al., 1978). PDH is rendered inactive when it is phosphorylated by pyruvate dehydrogenase kinase (PDHK) and is activated when dephosphorylated by the phosphatase, PDHP (Harris et al., 2002;Bowker-Kinley et al., 1998). Many cancers maintain high ox-phos as well as glycolysis, maximizing the anapleuretic functions of the cell that provide the building blocks for lipid, protein and nucleotide synthesis.
The mammalian transcription factor, Hypoxia-inducible factor-1a (Hif-1a; called Sima or Hifa in Drosophila), regulates a number of target genes that promote various aspects of cancer, including metabolism, angiogenesis, cell survival, drug resistance, and invasive motility (Wykoff et al., 2000;Carmeliet et al., 1998;Ryan et al., 1998;Pennacchietti et al., 2003;Ema et al., 1997;Semenza, 2003;Dang and Semenza, 1999). Hif-1a participates in this process as hypoxia favors glycolysis over oxidative phosphorylation for ATP generation (Zhong et al., 2000;Keith and Simon, 2007;Bertout et al., 2008;Dang and Semenza, 1999;Kim et al., 2006). Hypoxia has been the proposed mechanism for oncogenes to effect a change in metabolic state (Finley et al., 2013;Ying et al., 2012;Vander Heiden et al., 2009;Levine and Puzio-Kuter, 2010;Fukuda et al., 2002). Mammalian studies often involve immortalized cell-lines with a variable and often unknown genetic background. Furthermore, while initiation of glycolysis has been studied (Lunt and Vander Heiden, 2011), the mechanism for the maintenance of the altered metabolic state under normoxic conditions is not as clear. Using Drosophila as a model system, we provide here, a complete genetic dissection of one mechanism that leads to and sustains a metabolic reprogramming in which Hifa, but not hypoxia, plays an important role.
Hif-1a and c-Jun N-terminal kinase (JNK) are associated together in many tumor types (Comerford et al., 2004;Laderoute et al., 2004;An et al., 2013). It is well established that reactive oxygen species (ROS) such as superoxide and peroxide radicals can cause both activation of the JNK pathway (Lo et al., 1996;Owusu-Ansah and Banerjee, 2009) and stabilization of Hif-1a (Drö ge, 2002;Chandel et al., 2000). It is increasingly apparent that persistent activation of JNK signaling is involved in cancer development, progression and perhaps cellular transformation (Manning and Davis, 2003;Raitano et al., 1995;Smeal et al., 1991;Wagner and Nebreda, 2009). In addition to the above functions, it is likely that JNK could have an indirect role in attenuating oxidative phosphorylation by activating PDHK, thus blocking PDH function (Zhou et al., , 2008. Determining how a variety of oncogenic pathways interact together to cause the metabolic reprogramming from oxidative phosphorylation to glycolysis is the central focus of this investigation. We achieve this by activating a single oncogene and show that this leads to a cascade of events that ultimately cause a glycolytic activation and allow maintenance of this altered metabolic state. There are multiple ways to model the 'Warburg effect'. This study takes advantage of the powerful genetic techniques in Drosophila used to identify epistatic relationships to provide a comprehensive and mechanistic basis for the establishment and maintenance of this metabolic transition in a receptor tyrosine kinase (RTK) induced tumor.

LDH activation and transcription by a specific RTK
Aerobic glycolysis in tumors is characterized by the conversion of pyruvate to lactate by the enzyme, lactate dehydrogenase (LDH). Importantly, LDH has been demonstrated to be a marker for poor prognosis in multiple malignancies such as renal cell carcinoma (Armstrong et al., 2012). The Drosophila genome contains a single gene encoding an LDH enzyme (ImpL3), and biochemical studies demonstrate that it functions most like LDHA, the human form predominantly expressed in skeletal muscle that favors the conversion of pyruvate to lactate (Rechsteiner, 1970). An increase in LDHA enzymatic activity has been observed in diverse malignant cancers (Dang and Semenza, 1999).
We adapted a classic biochemical enzymatic assay (Abu-Shumays and Fristrom, 1997) to visualize the activity of LDH in vivo. Endogenous LDH activity is not observed in epithelial imaginal tissues such as the wing and eye discs (Figure 1a; Figure 1-figure supplement 1a) that developmentally mature in the larva to give rise to the corresponding appendages in the adult (Swammerdam, 1737), although evidence for such endogenous expression can be observed in the brain and salivary glands ( Figure 1-figure supplement 1c-d). Using the Gal4/UAS system (Brand and Perrimon, 1993) to individually activate oncogenes in the wing disc, we found that the activated PDGF/VEGF receptor (known as Pvr) causes robust LDH activity (Figure 1b) while many other oncogenes do not (see later).
To determine whether the increase in LDH activity is due to an increase in LDH transcription, a GFP-based enhancer trap (Quiñones-Coello et al., 2007) was used to visualize LDH expression. This reporter (called LDH-GFP) consists of EGFP inserted within 50 base pairs upstream of the LDH transcriptional start site within its native locus. LDH-GFP is a direct insertion of GFP into the endogenous LDH locus and is not affected when combined with UAS/Gal4 constructs. As expected, no GFP is detected in the developing wild-type wing or eye disc of LDH-GFP larvae that are otherwise wild type ( Figure 2c; Figure 1-figure supplement 1b), and similar to the results of the activity assay, endogenous LDH-GFP expression is readily apparent in the brain and in the salivary gland (Figure 1-figure supplement 1c-d). However, in a dpp blk1 Gal4, UAS-Pvr act genetic background, in which Pvr act is mis-expressed along the anterior/posterior boundary of the wing disc, a large tumorous overgrowth is observed with the mutant cells exhibiting robust LDH-GFP expression ( Figure 1d). Ras1 act , functioning downstream,also causessome LDH expression, but this effect is Figure 1. LDH induction by a single activated oncogene. Wing discs are from wandering third instar larvae. Scale bars, 50 mm. (a-b) LDH enzymatic activity is induced by Pvr act (full genotype: dpp blk-1 -Gal4, UAS-GFP, UAS-Pvr act ) stained for LDH activity (brown precipitate). The green channel is not shown in this panel for clarity. (a) Control, (dpp-Gal4, UAS-GFP) no detectable LDH enzymatic activity. (b) Expression of Pvr act (dpp-Gal4, UAS-GFP, UAS-Pvr act ) induces LDH activity. (c-j) LDH-GFP transgene induction (shown in green) monitors the expression of LDH. Nuclei are marked with To-Pro (blue). Full genotype in each panel includes the driver (dpp blk-1 -Gal4), a UAS-transgene as indicated and mCherry to mark expressing cells (red channel omitted for clarity). (c) Control. No LDH-GFP expression is detected in dpp-Gal4, UAS-mCherry wing disc. (d) Pvr act (dpp-Gal4, UAS-mCherry, UAS-Pvr act ) causes robust LDH-GFP induction. (e-j) Either no expression or very mild expression (arrows in e, h) of LDH is seen when Ras act (e), wingless (f), unpaired (g), yorkie (h), InR act (i), or Egfr act (j) transgenes are expressed under the control of dpp-Gal4 in the wing disc. DOI: 10.7554/eLife.18126.002 The following figure supplement is available for figure 1: modest compared to that seen with activation of Pvr ( Figure 1e). Together, the enzyme and the reporter assay establish that activation of PDGF/VEGF Receptor leads to transcriptional up-regulation of LDH and the subsequent formation of active LDH enzyme in the tumor tissue.
Expression of many oncogenes cause large tumorous overgrowths that superficially look like the Pvr induced tumors. However, LDH is not expressed in every tumor that shows increased cell proliferation and over-growth. For example, over-expression of the secreted ligand, Wingless (Neumann and Cohen, 1996), causes overgrowth and duplication of the tissue but does not cause LDH expression (Figure 1f). Similarly, overexpression of the JAK/STAT pathway ligand Unpaired (Upd), promotes tissue growth (Chao et al., 2004;Rodrigues et al., 2012), but does not induce LDH expression (Figure 1g). The transcription factor Yorkie is a potent growth-promoting signal in the wing disc (Huang et al., 2005), but at best, it induces very weak and variable expression of LDH-GFP in the wing pouch ( Figure 1h). These results indicate that the up-regulation of LDH by Pvr is not a general consequence of stepped up cell proliferation seen in all tumors.
Furthermore, given Pvr's function as an RTK, it was initially a surprise to find that neither constitutively signaling form of the insulin receptor (InR act , Figure 1i) nor the activated form of epidermal growth factor receptor (Egfr act ; Figure 1j) can induce LDH-GFP expression even though both cause significant overgrowth of the tissue. This observed specificity of Pvr, compared with other RTKs, in the expression of LDH, presents a unique opportunity for investigating the mechanism by which an oncogene can regulate the metabolism of a tumor tissue.
Hif-1a is a well-known inducer of LDHA expression in many cancer cell lines (Dang and Semenza, 1999). As with the mammalian counterpart, Drosophila Hifa (called Sima) is rapidly degraded under  normoxic conditions and therefore no Sima protein is detected in the wild-type wing disc ( Figure 2a), but Sima protein is stabilized in a Pvr act background ( Figure 2b). Overexpression of Sima (but not of a Sima mutant lacking its DNA binding domain) robustly induces both LDH activity and expression (Figure 2c-d; Figure 2-figure supplement 1a). It also causes growth defects, similar to the phenotypes of the Hph mutant (Centanin et al., 2005). These growth defects associated with gross overactivity of Sima are likely independent of its role in hypoxia related responses.
Importantly, LDH expression induced by Pvr act is strongly suppressed upon silencing sima using either one of two independent transgenic RNAi constructs (Figure 2e-f). Thus, Sima is essential for LDH induction by Pvr act . Knockdown of sima does suppress some of the overgrowth caused by Pvr act , but this effect is small (about a 20% reduction in volume. see Figure 2-figure supplement 1b) and the tumor-growth loss is clearly not as robust as the complete loss of LDH-GFP seen using the same knockdown construct. We conclude that tumor growth is a result of multiple interacting events, and blocking glycolysis alone will not be sufficient to fully rescue the growth of a tumor.
The Pvr induced tumor is characterized by unrestrained, disorganized growth, and in principle, one mechanism for Sima stabilization could involve the hypoxic environment that might exist within the disorganized mass of a tumor tissue. Examples of such a mechanism in which deeper tissues within solid tumors become hypoxic have been demonstrated in human cancers (Bertout et al., 2008;Keith and Simon, 2007). However, it is also true that in addition to hypoxia, Hif can be stabilized under normoxic conditions by multiple metabolites such as NO, ROS, succinate, and fumarate (Chandel et al., 2000;Isaacs et al., 2005;Selak et al., 2005;Mateo et al., 2003). To address whether the Pvr induced LDH-GFP expression is specifically due to a hypoxic core within a tumor, we utilized the AyGal4 system (Ito et al., 1997) to generate flip-out clones expressing Pvr act . This technique allows analysis of LDH expression in single or small groups of cells expressing Pvr act that are surrounded by normal tissue. Under these conditions, there is no over-growth or tumor formation, yet these single/small number cell clones express LDH in a strikingly cell-autonomous fashion ( Figure 2g). We conclude that a large tumor with a hypoxic core is not a requirement for Sima stabilization or LDH expression.

Both ERK and PI3K pathways are necessary for LDH expression
Ras is the major effector of RTK signaling (Moodie et al., 1993;Vivanco and Sawyers, 2002), and indeed we have found that constitutively active dRas1 is sufficient for a small increase in LDH expression and activity ( Figure 1e). Also, induction of LDH by Pvr act is suppressed by a dominant-negative mutant allele of Ras (dRas1 N17 ) (Lee et al., 1996) (Figure 3h). Interestingly, although Ras is downstream of both the EGFR and InR (Insulin receptor) pathways, on their own, neither Egfr act nor InR act causes LDH expression even as they individually cause overgrowth (Figure 1i-j). We found that Egfr act causes phosphorylation of ERK but not AKT ( Similarly, when we blocked PI3K signaling in a Pvr act background using several independent ways, including co-expression of an RNAi against PI3K (Figure 3k), a dominant-negative mutant form of PI3K (Figure 3l), inactivation of Akt by RNAi (Figure 3m), or silencing dTOR by RNAi (Figure 3n), each combination suppresses the induction of LDH-GFP. Finally, Pvr act dependent accumulation of Sima protein is significantly suppressed when either Dsor1 or Akt is silenced (Figure 3-figure supplement 1a-b). Thus, the PI3K/Akt/TOR axis is also definitively needed for LDH regulation.
Importantly, individual overexpression of either an activated form of human Raf (hRaf act ) or a constitutively active form of PI3K (PI3K act , active due to a mutation in the p110a catalytic subunit) (Brand and Perrimon, 1994;Leevers et al., 1996) does not induce LDH-GFP expression although each genotype causes significant over-growth (Figure 3o-p). In combination, however, co-expression of hRaf act and PI3K act either using drivers or in small clones, leads to extensive LDH induction even more robustly than what is seen for Pvr act (Figure 3q-r). Furthermore, dual PI3K/ERK activation is also sufficient to increase Sima expression (  . Downstream of Pvr act , ERK and PI3K pathways control Sima translation. All wing discs shown are from wandering third instar larvae. All transgenes are expressed using the dpp blk-1 -Gal4 driver. Scale bars, 50 mm. (a-f) Differential activation of ERK and PI3K pathways by RTKs. Phospho-specific antibodies against dp-ERK (a-c, green) or p-Akt (d-f, green) are used to detect activation of the corresponding pathways. (a-c) dp-ERK staining is seen in in Egfr act (a) and Pvr act (c) backgrounds, but not upon activation of InR (b). (d-f) In contrast to (a-c), p-Akt is detected upon activation of InR (e) and Pvr (f) but not Egfr (d). Thus only Pvr act causes both the ERK (assayed by ERK phosphorylation) and PI3K (assayed by Akt phosphorylation) pathways to be activated. (g-v) Both PI3K and ERK pathways must be activated for LDH-GFP induction. Expression of LDH-GFP (green) marks glycolytic tissue. In Figure 3 continued on next page the critical regulator of glycolytic genes upon co-activating PI3K and ERK pathways. We conclude that a combination of ERK and PI3K signaling is both necessary and sufficient to induce Hifa-dependent LDH induction. On its own, each signal is necessary but not sufficient to cause LDH expression.
InR/TOR pathway is associated with translational control in diverse systems (Fukuda et al., 2002;Hay and Sonenberg, 2004). In particular, both translation and degradation critically control Hifa level (Bacon et al., 1998;Gorr et al., 2004). Consistent with this notion, the induction of LDH by Pvr act /Sima remains unaffected when Pointed (Drosophila ETS2), the well-established direct nuclear target of the ERK pathway (O'Neill et al., 1994), is mutated ( Figure 3v). In contrast, the ribosomal S6 kinase, a known target of TOR signaling that promotes translation of a subset of mRNA transcripts, including Hif-1a (Ma and Blenis, 2009) is prominently involved in this process since a dominant-negative form of S6k is a strong suppressor of Pvr-mediated LDH induction and Sima accumulation (Figure 3s and Figure 3-figure supplement 1f). The ERK signal also modulates protein translation through the Mnk kinase (LK6 in Drosophila). Lk6 binds ERK and phosphorylates the initiation factor eIF4E to promote Cap dependent translation (Arquier et al., 2005;Parra-Palau et al., 2005). Strikingly, we find that a single-copy loss of lk6 strongly suppresses LDH induction and Sima accumulation downstream of Pvr ( Figure 3t and Figure 3-figure supplement 1g). Finally, published literature (Miron et al., 2001) has established 4E-BP (Thor) as a downstream effector of the PI3K/Akt pathway that forms a complex with elF4E and inhibits translation. mTOR phosphorylates 4E-BP releasing it from eIF4E (Furic et al., 2010) allowing efficient translation. We found that co-expression of 4E-BP with Pvr act suppresses LDH induction ( Figure 3u). Taken together, these data strongly support a mechanism in which ERK and PI3K pathways converge at the level of translational control of gene products including Sima. We propose that initially, the increased translation of sima transcript generates sufficient Sima protein (in excess of the rate of Sima degradation), leading to LDH induction; later we will show that this reprogramming is further augmented by a feedback loop involving ROS.

Pvr act induced metabolic changes Intensification of glycolytic pathway
In order to gain better understanding of the transcriptional basis of Pvr-induced glycolytic transition at a genome-wide level and the role of sima in the regulation of this process, we carried out an RNA-Seq experiment in which transcriptomes of wild-type wing imaginal discs were compared with those with the genotypes Pvr act (glycolytic activity), or InR act (no detectable glycolytic activity) or Pvr act + sima RNAi (glycolytic activity suppressed). Each genotype was analyzed in triplicate. The data were filtered for the GO term 'metabolism', and further sorted for genes that are up-regulated in Pvr act but not in InR act background.Amongst this subset, transcripts that are down-regulated in Figure 3 continued (r), the Ay-Gal4 system is used to generate clones with nuclear b-Galactosidase staining (red) marking the clones. LDH-GFP (green) expression (g; Pvr act control) is suppressed upon loss of ERK pathway members Ras (using a dominant negative, h), Downstream of Raf1, (Dsor1, using an RNAi construct, i), and ERK (Rolled, using an RNAi construct, j). (k-n) PI3K pathway components are also essential for LDH-GFP expression. LDH-GFP (green) expression (g; Pvr act control) is suppressed upon loss of PI3K pathway members Ras (using a dominant negative, h), PI3K (using an RNAi construct, k or a dominant negative version, l), Akt (using an RNAi construct, m), or TOR (RNAi; n). (o-r) ERK and PI3K pathways must be co-activated for LDH expression. (o-p) Single activation of the ERK pathway (using hRaf act ; o), or of the PI3K pathway (using PI3K act ; p) is insufficient for LDH-GFP expression. (q,r) Coactivation of both pathways by co-expression of hRaf act and PI3K act together, either using a dpp-Gal4 driver (q) or in small Ay-Gal4-derived clones (r) induces robust LDH-GFP expression. (s-v) Translational regulators are essential for LDH transcriptional control. A dominant-negative form of S6k (s) or overexpression of Thor/4EBP (u), both acting downstream of PI3K as well as loss of a single copy of Mnk/Lk6 (t), functioning downstream of ERK, abolish transcription of LDH-GFP. In contrast, loss of the major transcriptional factor in the ERK pathway, the ETS domain protein, Pointed (v) has no effect on LDH transcription. DOI: 10.7554/eLife.18126.006 The following figure supplement is available for figure 3: Pvr act + sima RNAi background were analyzed further. Our results confirm the transcriptional up-regulation of LDH. Interestingly, the transcription of six of the ten glycolytic enzymes is up-regulated in a Pvr act background of which four -Hex-A, Pfk, Ald and Impl3-are regulated in a Sima dependent manner (Figure 4a-b). The 73 sima-dependent metabolic genes identified in this assay are listed in Figure 4-figure supplement 1. GO and network analysis reveal enrichment in biological function for genes involved in regulating rate-limiting steps of glycolysis. Hex-A (hexokinase-A) phosphorylates glucose to generate glucose 6-p in the first step of glycolysis. Similarly, Pfk (phosphofructo kinase) phosphorylates fructose 1-p to generate fructose 1,6-bp, whereas Ald (Aldolase) converts fructose 1,6-bp to glyceraldehyde 3-p in the next step. Impl3/LDH (lactate dehydrogenase) converts pyruvate to lactate in the final step of glycolysis. Pvr and Sima-dependent regulation of these key enzymes indicates a more efficient usage of the entire glycolytic pathway.

Attenuation of oxidative phosphorylation
In addition to displaying increased aerobic glycolysis, a subset of cancer cells also attenuate mitochondrial respiration (Pelicano et al., 2006;Warburg, 1956aWarburg, , 1956b. The RNA-seq data identifies a set of transcripts encoding mitochondrial proteins that is down-regulated in Pvr act background. Attenuation of ETC complex protein activity has also been demonstrated to up-regulate transcripts for all glycolytic enzymes suggesting a cross talk between the two primary modes of metabolism (Owusu-Ansah et al., 2008).
Just as LDH drives glycolysis by converting pyruvate to lactate, pyruvate dehydrogenase (PDH) drives the TCA cycle and oxidative phosphorylation by converting pyruvate to acetyl-CoA. PDH is inactive when phosphorylated by p-PDHK. When stained with appropriate antibodies, Pvr act cells are found to express PDHK total , active PDHK (phospho-PDHK), as well as the inactive form of PDH (p-PDH) (Figure 5a-f). The high p-PDH level is suppressed by co-expressing PDHK-dsRNA ( Figure 5-  . Src42A and JNK inhibit mitochondrial function by regulating PDHK activity. All wing discs shown are from wandering third instar larvae. All transgenes are expressed using the dpp blk-1 -Gal4 driver. Scale bars, 50 mm. (a-f) Pvr act up-regulates PDHK expression as well as its activity. Total PDHK protein expression detected using an a-PDHK1 antibody (a, d, green) while phospho PDHK-specific antibody detects only the activated form (b, e, green). p-PDH (c, f, green) is detected only if PDH is phosphorylated and inactivated by p-PDHK. Compared to control, Pvr act tissue shows elevated expression of PDHK total (a, d), p-PDHK (b, e), and p-PDH (c, f). (g-j) Pvr act activates Src and JNK. Compared with controls, Pvr act tissue shows increased expression of activated Src ((g, i) phospho Src specific, green) and activated p-JNK ((h, j) phospho JNK specific, green). (k) Co-activation of ERK and PI3K pathways dramatically increases PDHK total (green) expression, but this PDHK is inactive (see Supplementary Figure 5c An antibody that recognizes both the active and the inactive forms of PDHK shows that coexpression of hRaf act and PI3K act results in dramatic up-regulation of PDHK total . Similar to the results obtained for Sima, the regulation of PDHK protein is also at a translational level (Figure 5k; Figure 5-figure supplement 1o,q). Interestingly, while the combined activation of PI3K act and hRaf act is sufficient for PDHK production, the protein thus produced is inactive and is not recognized by a phospho-specific, p-PDHK antibody ( Figure 5-figure supplement 1c). This is unlike in Pvr act cells where both PDHK total and PDHK active are seen (Figure 5d-e). Consistent with this observation, p-PDH (the inactive form of PDH) is also not seen when hRaf act and PI3K act are co-expressed (Figure 5-figure supplement 1b), and yet is readily detectable in a Pvr act background (Figure 5f). These results establish that additional proteins independent of ERK and PI3K pathways function downstream of Pvr and are necessary to activate PDHK and thus inhibit PDH and mitochondrial oxidative phosphorylation.
Src42A, a Drosophila c-src proto-oncogene homolog, functions in epidermal closure during both embryogenesis and metamorphosis by regulating JNK signaling (Tateno et al., 2000). It has also been found that Src42A controls tumor invasion and cell death by activating JNK (Ma et al., 2013). In mammalian cancer studies, Src mediates tumor cell metastasis and angiogenesis induced by VEGF (Eliceiri et al., 1999;Weis et al., 2004). Using phospho-specific antibodies, we found that both Src42A and JNK are extensively phosphorylated in Pvr act cells (Figure 5g-j). Knockdown of Src42A suppresses phospho-JNK accumulation induced by Pvr act (Figure 5o), whereas over-expression of a dominant negative form of JNK (Bsk DN ) does not significantly reduce phospho-Src levels ( Figure 5s) indicating that Src42A functions upstream of JNK.
Knockdown of Src42A, or over-expression of Bsk DN partially suppresses the phosphorylation level of PDHK induced by Pvr act (Figure 5l-m). As shown earlier, co-expression of hRaf act and PI3K act is unable to increase p-PDHK levels ( Figure 5-figure supplement 1c), but the simultaneous activation of ERK, PI3K and JNK pathways by using hRaf act , PI3K act and active JNKK (hemipterous, hep act ), clearly up-regulates PDHK activity ( Figure 5-figure supplement 1d), and the resulting effect on PDH is also dramatic. A dominant negative form of JNK (Bsk DN ) or loss of JNKK (hemipterous, hep R-NAi ) suppresses the formation of p-PDH upon Pvr activation (Figure 5n; Figure 5-figure supplement 1e). Thus JNK is required for activation of p-PDHK and as a result, sufficient to inactivate PDH by causing its phosphorylation.
In summary, the above results show that:

Feedback loops to strengthen the metabolic reprogramming
Although over-active PDH can promote reactive oxygen species (ROS) production during high mitochondrial flux (Kaplon et al., 2013), dysfunction of PDH within the mitochondrion is invariably associated with increased production of ROS (Glushakova et al., 2011;Ambrus et al., 2011). Thus, Pvr act cells exhibit high ROS levels detected by direct staining with the DHE dye (Figure 6a-b) or as monitored by gstD-GFP expression (Robinson et al., 2006) (Figure 6c-d) frequently used as a surrogate for ROS, marking cells under oxidative stress (Ohsawa et al., 2012;Sykiotis and Bohmann, 2008). This ROS induction is suppressed when PDHK RNAi is co-expressed ( Figure 6-figure supplement 1a).
Co-expression of hRaf act and PI3K act does not cause ROS production unlike that seen for Pvr act (Figure 6e; Figure 6-figure supplement 1b). Instead, Pvr act induced ROS is effectively suppressed in a Bsk DN  Two notable facts about ROS have a bearing on the data presented here. The first is that ROS can activate the JNK pathway through interaction with the upstream component ASK (JNKKK) (Tobiume et al., 2001;Owusu-Ansah and Banerjee, 2009). The second is that high ROS conditions result in potent stabilization of the Hif protein even under normoxic conditions (Chandel et al., 2000;Fandrey et al., 2006). This provides the opportunity for ROS generated downstream of both Hif and JNK, to feedback and reinforce these two pathways. To test this hypothesis, several antioxidant (free radical scavenging) proteins were expressed to reduce the level of ROS. Of these, the strongest ROS scavenging activity was seen upon Peroxidasin (Pxn) expression. In addition to a decrease in ROS (Figure 6g-h), we found that JNK is no longer activated when Pxn is co-expressed in Pvr act discs (Figure 6i-j). Similarly, LDH-GFP expression (Figure 6k-l) and the accumulation of Sima protein are strongly suppressed by co-expression of Pxn (Figure 6m-n). These results establish ROS as the central player in enforcing the metabolic reprogramming. Initially established through the activation of three oncogenic pathways by Pvr, the shift to glycolysis is then reinforced when ROS is generated as a subsequent step. This establishes one means to maintain a stable Warburg effect by oncogene activation.

Discussion
A model (Figure 7) can be constructed that is consistent with all our observations on the Pvr act induced interacting network that leads to aerobic glycolysis and potentially also allows this new metabolic state to be maintained through later stages of tumor growth. This is a genetic model derived from analysis that allows placement of gene function according to their hierarchy along a pathway. Many genetic backgrounds that achieve full or partial Warburg effect have been described in the literature (Elstrom et al., 2004;Kim and Dang, 2005;Wang et al., 2011;Faubert et al., 2013;Hitosugi et al., 2009, Yang et al., 2012Hong et al., 2016;Jin et al., 2016). Although the details may seem to vary, the three critical components involved in this process are the two pyruvate-metabolism enzymes, LDH and PDH and the free radical metabolite class designated as ROS. During the acquisition of the aerobic glycolytic activity in a tumor environment, LDH actively converts pyruvate to lactate, and high PDHK inactivates PDH resulting in low pyruvate to acetyl-CoA conversion, low TCA flux and electron transport activity. Our results suggest that these events can be sustained over long periods of tumor growth only when coupled with a feed back signal from accumulating ROS in a mechanism that reinforces and enhances the high-LDH/low-PDH activity state. This genetic study was achieved in an in vivo context of an animal that is only mutant for the one specific activated oncogene that we introduce and is otherwise normal. Additional pathways are activated without accumulating new mutations. Also, while hypoxic conditions might favor a glycolytic activation in general, hypoxia is not an essential component for a tumor to become glycolytic, as in this system, where Hif is stabilized by alternative means in a normoxic environment.
Pvr act triggers two parallel pathways in a Ras dependent manner. The first is the PI3K/Akt pathway that targets the S6K protein important for translational initiation. The second is the Raf/ERK Figure 6. ROS strengthens the metabolic reprogramming. All wing discs shown are from wandering third instar larvae. All transgenes are expressed using the dpp blk-1 -Gal4 driver. Nuclei are marked with DAPI (blue; c-l). Scale bars, 50 mm. (a-f) JNK mediates ROS induction. DHE staining (green; a, b) directly monitors ROS levels and gstD-GFP expression (green; c-f) is a measure of oxidative stress caused by the generated ROS. No detectable ROS stress is seen in control discs (a, c), while Pvr act causes robust ROS and gstD generation (b, d). Co-expression of hRaf act and PI3K act does not phenocopy Pvr act in causing oxidative stress (e). A dominant negative form of JNK (Bsk DN ) effectively suppresses ROS generation (f). (g-m) A ROS feedback signal plays a central role in Pvr act induced metabolic reprogramming. Sima accumulation is detected by using an a-Sima antibody (m-n). Pvr act induced ROS generation (g), p-JNK expression (i), LDH expression (k), and Sima accumulation (m), are all strongly suppressed by expressing a scavenger gene, Peroxidasin (Pxn h, j, l and n). pathway targeting the L6K (MNK in mammals) protein that is also independently required for translational initiation. On their own, any one of these pathways can cause tumor growth, presumably due to independent downstream transcriptional events. But both pathways must act simultaneously in order to establish translational control of both Hifa (Sima), and PDHK. Interestingly, while the PI3K and ERK pathways acting together causes significant amounts of PDHK translation, the protein thus produced is inactive. In contrast, the PDHK translated in a Pvr act background is enzymatically active and capable of inactivating PDH. Our data show that yet another parallel pathway initiated by Pvr act that involves Src and JNK helps generate the active form of PDHK. Therefore all three downstream effectors of activated Pvr (Src/JNK, PI3K/S6K and ERK/Lk6, eIF4E) are essential in generating an active form of PDHK protein to down-regulate oxidative phosphorylation, while only the last two of these pathways are needed for LDH production and glycolysis. It is easy to visualize how, in the context of the triggering oncogene, the tumor caused could either be non-glycolytic, or be glycolytic Figure 7. A model for induction and maintenance of a metabolic reprogramming. Pvr act induced tumor tissues show a robust metabolic reprogramming from oxidative phosphorylation to aerobic glycolysis. The model describes the relationship between many pathways that are activated in response to the Pvr act signal. ROS, generated as a consequence of the initial metabolic reprogramming feeds back to upstream components, to consolidate and maintain the Warburg effect. See text for details. DOI: 10.7554/eLife.18126.015 but not lacking in oxidative phosphorylation (Pavlova and Thompson, 2016). In the system described in this paper, it is Inevitable that direct pyruvate to acetyl-CoA conversion within the mitochondrion, and therefore the TCA flux will be attenuated since PDH function is inhibited by PDHK in the tumor cells. However, we do not have direct evidence for or against any bypass mechanisms such as those using glutamine as the primary driver of anabolic processes being operational in the context of the Drosophila Pvr act induced tumors as they are for many cancers. This issue requires further direct measurement of metabolic activity of the tumor tissue in the near future.
While this sequence of events can initiate a switch in the metabolic profile of the cell, an additional mechanism is required to sustain this transition over a long period of time. The central player in this sustenance is ROS. The level of these free-radical species rises as the mitochondrion becomes dysfunctional (Sun et al., 2009;Glushakova et al., 2011;Owusu-Ansah et al., 2008). Importantly, this excess ROS functions as a feedback signal that has at least two important consequences. First, ROS stabilizes Hif using the same mechanism that is used during hypoxia (Chandel et al., 2000;Fandrey et al., 2006), thus reinforcing the glycolytic pathways. The second is that ROS can activate the JNK pathway (Tobiume et al., 2001;Owusu-Ansah and Banerjee, 2009). Thus, after the initial reprogramming in metabolism, the transition is made stable over time, as the generated ROS reinforces the upstream members. All of the results shown here are consistent with this positive feedback model that enforces the Warburg effect and sustains it over time. For example, scavenging ROS blocks phosphorylation of components that are upstream of it (Figure 6i-j). Unlike Pvr, activation of PI3K and Raf together does not raise ROS even as it increases LDH (Figure 3q and Figure 6e); and activation of Ras, which will not lead to high ROS causes only very weak metabolic reprogramming compared with Pvr act (Figure 1d-e). Thus, the positive feedback function of ROS is a critical step in tumors undergoing metabolic reprogramming or cell lines that are not necessarily under hypoxic conditions. Although tissue hypoxia does stabilize Hif (Wenger, 2002), this would be true of large non-vascularized solid tumors but not necessarily in circulating cells or in cancer cell lines. The mechanism detailed in Figure 7 is Hif dependent but hypoxia independent, a condition often observed in human cancers (Lee et al., 2008;Ivan et al., 2001;Zhong et al., 2000;Park et al., 2003).
Consistent with translational control, transcript levels of PDHK or Sima are not altered in Pvr act cells, whereas the PDHK and Sima protein levels are very significantly up-regulated by Pvr act through the translational module controlled by PI3K/ERK pathways. Unlike in several published mammalian studies (Kaplon et al., 2013;Kim et al., 2006), we did not find evidence for transcriptional control of PDHK by Hif. Analysis of our RNA-Seq data reveals that PDHK transcription levels remain unaltered in Pvr act background (Figure 4-figure supplement 1). This was also confirmed by quantitative RT-PCR ( Figure 5-figure supplement 1p), The reason for this discrepancy is unclear; perhaps Hif does control basal transcriptional levels of PDHK, but the oncogene induced up-regulation of activity results from a post-transcriptional control of PDHK activity as we observe here.
It is interesting that the activation of a single oncogene, the PDGF/VEGF Receptor (Pvr) induces multiple phenotypes associated with tumor formation, including overgrowth, cell shape change, local migration and importantly for this study, a likely shift in metabolism from oxidative phosphorylation to aerobic glycolysis. These tumors are not metastatic presumably because this requires complete loss of epithelial polarity (Pagliarini and Xu, 2003;Wu et al., 2010). Although only a single oncogene is activated, the observed Warburg effect is caused by the activation of several interconnected pathways that are precisely dissected in this genetically tractable model system. The distinct advantage of this analysis is that the tumor is generated in an otherwise wild-type background such that the pathways activated reflect primary drivers of the different aspects of the tumor and the analysis is not complicated by background mutations. In some cancer cell-lines, a Warburg effect is often entirely attributed to PI3K activation (DeBerardinis et al., 2008). This interpretation is complicated by the presence of background mutations that make cell-lines immortal, before they are transformed. For example, PI3K activated tumors are often seen only in backgrounds in which either Raf, ERK or Ras is activated (McCubrey et al., 2007;Steelman et al., 2008;Yuan and Cantley, 2008). With the advent of modern gene manipulation technologies, it is now possible to create in vivo models in mouse (Xue et al., 2014;Platt et al., 2014), but determining the epistatic relationships between all genetic components continues to be a challenge. Given the complete conservation of all the relevant components between Drosophila and mammals, it is reasonable to propose that a similar set of epistatic relationships to the one proposed here, will hold in mammalian cancers.

Drosophila stocks
The following fly stocks were used: w 1118 , dpp-Gal4 blk-1 , UAS-dRas1 V12 (Ras act ), MKRS hsFLP.86E, (Ras DN ) . Single-plane fluorescence images were obtained using a Zeiss LSM 700 confocal microscope with Zen 2009 acquisition software. Images and average intensity of images were processed and measured using ImageJ. At least 20 discs were used for each genotype per experiment. Three technical replications were done for each experiment.

Flip-out clones
The Ay-Gal4 system (Ito et al., 1997), combined with a heat shock-driven FLP recombinase, was used to generate clones in which oncogenes and/or RNAi and other transgenes were expressed. Expressing cells were marked by including UAS-driven transgenes for rat CD2 or nuclear b-Galactosidase. A 30-45 min heat shock (37˚C) was applied to larvae grown for approximately three days AEL (25˚C), and larvae were dissected after approximately 48 hr of subsequent growth at 25˚C.

RNA-seq library construction and analysis method
Total RNA was isolated using PureLink RNA Mini Kit (Life Technologies, Carlsbad, CA) according to manufacturer's protocol. Total RNA fraction was processed using Illumina's ''TruSeq RNA Sample Preparation v2 Protocol'' (Illumina, San Diego, CA). The resulting purified cDNA library was sequenced on the Illumina Hiseq 2000 by following manufacturer's protocols. 50 bp single-end RNA-seq reads were obtained, and sequence files were generated in FASTQ format. The quality score of RNA-seq reads was obtained by using the FastQC. Reads were then aligned to the Drosophila melanogaster Reference Sequences UCSC dm6 from Illumina iGenome files using TopHat v2.0.10 (Kim et al., 2013).Transcript assembly and estimation of their abundances were calculated with Cufflinks 2.1.1 (Trapnell et al., 2012) by using the Drosophila melanogaster Reference annotation dataset UCSC dm6 from Illumina iGenome files (https://support.illumina.com/sequencing/ sequencing_software/igenome.html). Differential expression for genes across the different conditions was calculated with Cuffdiff 2.1.1 (Trapnell et al., 2012). Heat maps showing gene expression levels (in FPKM, fragments per kilobase of transcript per million mapped fragments) through the different samples were drawn with package heatmap.2 in R studio for a subset of selected genes. Data deposited in ArrayExpress: E-MTAB-3808.

Quantitative real-time PCR analysis
Total RNA was extracted from approximately forty 3rd instar larval wing discs using PureLink RNA mini kit (Ambion, Waltham, MA). The SuperScript III First-Stand synthesis SuperMix kit (Invitrogen) was used for first-stand cDNA synthesis. Relative quantitative PCR was performed by comparative C T method using Power SYBR Green PCR master mix kit (Applied Biosystems, Waltham, MA) and a StepOne Real-Time PCR detection thermal cycler (Applied Biosystems). The levels of RpL10 were used to normalize total cDNA input.

Wing disc volume measurements and analysis
To synchronize larvae, eggs were collected on grape juice plates, and only larvae hatched within a three-hour period were used. After allowing newly hatched larvae to grow for 96 hr at 29˚C, wing discs were dissected, fixed, washed, and stained with TO-PRO-3. Discs were then mounted in a drop of Vectashield and slowly moved to edge of drop (in increasingly less mounting medium) until they were stuck in position to the glass slide. Small, square coverslips were placed on either side of drop, and a larger, rectangular coverslip was placed spanning the smaller ones. All slides were fixed with nail polish. This mounting creates a gap (of coverslip thickness) between the top coverslip and slide, preventing the coverslip from pressing and deforming wing discs. When imaging, z-stacks were obtained with a distance between slices of 14 mm, and the entire disc was covered. To measure volumes, plugin 'A 3D editing' was used (ImageJ). Briefly, each disc was manually outlined in each z-section (using the channel for mCherry), and the plugin then calculated total volume. Brightness and contrast were enhanced identically for each sample to aid in outlining. At least 10 discs were used for each genotype per experiment.

ROS assay
For ROS detection in wing discs, after incubation with DHE (Invitrogen) in RT, discs were rinsed twice in Schneiders medium, fixed for 5 mins in 4% Formaldehyde, and rinsed once with 1XPBS before confocal imaging.