Identifying strategies to target the metabolic flexibility of tumours

Plasticity of cancer metabolism can be a major obstacle to efficient targeting of tumour-specific metabolic vulnerabilities. Here, we identify the compensatory mechanisms following the inhibition of major pathways of central carbon metabolism in c-MYC-induced liver tumours. We find that, while inhibition of both glutaminase isoforms (Gls1 and Gls2) in tumours considerably delays tumourigenesis, glutamine catabolism continues, owing to the action of amidotransferases. Synergistic inhibition of both glutaminases and compensatory amidotransferases is required to block glutamine catabolism and proliferation of mouse and human tumour cells in vitro and in vivo. Gls1 deletion is also compensated for by glycolysis. Thus, co-inhibition of Gls1 and hexokinase 2 significantly affects Krebs cycle activity and tumour formation. Finally, the inhibition of biosynthesis of either serine (Psat1-KO) or fatty acid (Fasn-KO) is compensated for by uptake of circulating nutrients, and dietary restriction of both serine and glycine or fatty acids synergistically suppresses tumourigenesis. These results highlight the high flexibility of tumour metabolism and demonstrate that either pharmacological or dietary targeting of metabolic compensatory mechanisms can improve therapeutic outcomes. Metabolic compensation equips tumours with the plasticity to circumvent individual nutrient pathway perturbation. Méndez-Lucas et al. demonstrate the power of synergistic targeting of multiple metabolic pathways to stymie liver tumourigenesis.

M etabolism of tumours is different from metabolism in parental tissues. This difference is thought to favour maximal use of the resources with an adequate energy balance to allow cancer cells to survive and proliferate in a competitive tumour microenvironment 1,2 . This metabolic rewiring of tumours can induce metabolic dependencies that have the potential to be exploited therapeutically 3 . However, to date, only few existing strategies targeting tumour-specific metabolic vulnerabilities have proven useful 3 . One of the underlying challenges may be the metabolic plasticity of cancer cells. Furthermore, the selection of metabolic targets for therapeutic interventions has often been done in cell-culture systems, in which metabolism of initial tumour-derived cells can be substantially affected by culture conditions 4 . These systems also do not recapitulate tumour heterogeneity and complex intertumour and tumour-host interactions 5 .
One of the major regulators of cellular metabolism during proliferation and development is the proto-oncogene c-MYC (henceforth termed MYC), which is also one of the most frequently dysregulated lesions in human cancers 6,7 . The difficulty of targeting MYC itself 8 highlights the need to uncover and exploit therapeutic targets among its downstream effectors, including metabolic pathways. Using a genetically engineered mouse model of liver cancer 9 , we and others have demonstrated that MYC-induced liver tumours have increased catabolism of both glucose and glutamine in comparison with the normal liver [10][11][12] . Since both glucose and glutamine fuel pathways that are vital for cellular proliferation and survival 4,[13][14][15][16] , several enzymes that control these pathways have been considered to be attractive therapeutic targets 4,10,[17][18][19][20][21] . However, the inability to completely inhibit tumourigenesis in any of these approaches suggests that tumours engage compensatory mechanisms that allow them to survive and proliferate despite decreased activity of these enzymes. Although an understanding of the complexity of factors determining tumour metabolic vulnerabilities and flexibilities is emerging 1,[22][23][24][25] , the mechanisms of metabolic resistance of tumours in vivo still remain largely unknown. Here we identify multiple strategies that tumours engage in order to overcome metabolic perturbations, including expression of alternative enzyme isoforms, reliance on alternative metabolic pathways, and utilization of nutrients from the bloodstream. We demonstrate how blocking these compensatory mechanisms can lead to stronger inhibition of tumour growth in vivo, which provides a putative strategy for designing more effective therapeutic approaches.

Major pathways of glucose and glutamine catabolism are increased in MYC-induced liver tumours.
We have previously demonstrated that glucose catabolism through glycolysis and catabolism of both glucose and glutamine through the Krebs cycle are increased in MYC-induced liver tumours 9 in comparison with normal tissue counterparts 10 . We have now further evaluated these tumours in vivo, using either [U- 13 C]glucose or [U-13 C]glutamine in tumour-bearing and control mice. The [U- 13 C]glucose boluses resulted in elevated enrichment of lactate and citrate (Fig. 1a), whereas the [U-13 C]glutamine boluses resulted in enhanced enrichment of citrate, malate, fumarate and oxaloacetate-derived aspartate in tumours in comparison with normal livers (Fig. 1b). Furthermore, when the labelled nutrients were infused for 3 h, the pools of 13 C-glucose-derived pyruvate, lactate, alanine, citrate and glutamate, and the 13 C-carbon pools of malate, fumarate and aspartate derived from either [U- 13 C]glucose or [U- 13 C]glutamine, were increased in tumours (Fig. 1c). The extent of isotope enrichment, either from glucose or glutamine, demonstrated that they are major sources for the generation of Krebs cycle intermediates in both tumours and livers (30-40% of total isotope enrichment). The partial contribution of glucose and glutamine to the pools of most of Krebs cycle intermediates at steady state in tumours remained similar to that observed in normal livers, except for 13 C-glucose-derived citrate, the entry point of glucose carbons into the Krebs cycle (Extended Data Fig. 1a,b,d,e,g).
Together with fuelling the Krebs cycle, the increased level of glucose and glutamine catabolism is required to support the activity of other pathways essential for tumourigenesis (Fig. 1d). Indeed, we observed an accumulation of serine and glycine in MYC-induced liver tumours (Extended Data Fig. 1c,f), as well as a marked increase in their production from [U-13 C]glucose ( Fig. 1a and Extended Data Fig. 1b) and [α- 15 N]glutamine (Extended Data Fig. 1h). We also observed a strong elevation in the incorporation of both glucose and glutamine carbons into tumour fatty acids ( Fig. 1c and Extended Data Fig. 1i), consistent with the accumulation of neutral lipids (Extended Data Fig. 1j).
The increased metabolic-pathway activities in tumours were associated with an elevated expression of enzymes or a change in enzyme isoform patterns ( Fig. 1e and Extended Data Fig. 2a,b), consistent with previous observations by us and others 10,11 . These changes included (1) heightened expression of hexokinase 2, HK2, and depressed expression of glucokinase, GCK or hexokinase 4 (glycolysis), and (2) increased expression of GLS1 and decreased expression of the GLS2 isoform of glutaminase (glutaminolysis). The expression of all serine-biosynthesis enzymes was higher in tumours than in normal livers, most notably phosphoserine aminotransferase (PSAT1; Fig. 1e and Extended Data Fig. 2b). Finally, although the messenger-RNA-expression levels of enzymes involved in some normal liver functions, such as gluconeogenesis, were suppressed in tumours, the expression of some lipogenic enzymes, such as fatty acid synthase (FASN), remained high ( Fig. 1e and Extended Data Fig. 2b).
To evaluate whether the elevated activity of these major pathways of glucose and glutamine catabolism is required for MYC-induced liver tumourigenesis, we used the Albumin-CreER T2 mouse line to knock-out Hk2, Gls1, Psat1 and Fasn genes specifically in hepatocytes at the time of tumour initiation (Extended Data Fig. 2c), as verified by adeno-associated virus-mediated lineage tracing 26 (Extended Data Fig. 2d-g). In this case, tumours were generated by hydrodynamics-based transfection of a plasmid encoding a combination of MYC and MCL1 (refs. 12,27 ), encoded by another gene frequently overexpressed in human liver tumours (Extended Data Fig. 2h).

GLS2 partially compensates for GLS1 deletion in tumours.
Liver-specific knockout of Gls1 (Gls1 KO ; Fig. 2a) resulted in increased latency of MYC-induced liver tumours (Fig. 2b), higher levels of glutamine (Fig. 2c) and lower incorporation of glutamine-derived carbons into Krebs cycle intermediates and amino acids (Fig. 2d). However, the total levels of Krebs cycle intermediates were unaffected upon Gls1 knockout (Extended Data Fig. 3a). These data demonstrated that GLS1 plays a significant role in glutaminolysis during MYC-induced tumourigenesis, consistent with previous results 18 . However, in the absence of GLS1, tumours were still able to develop and catabolize a substantial amount of glutamine. We next investigated which other enzymes could be responsible for this capability.
GLS2, whose expression was decreased in MYC-induced tumours in comparison with normal liver, appeared to be expressed in Gls1 KO tumours ( Fig. 2e and Extended Data Fig. 3b). Gls2 knockdown in Gls1 wild-type tumours, using short hairpin RNA (shRNA) interference (Extended Data Fig. 3c), had no effect on either tumourigenesis or glutamine catabolism (Extended Data Fig. 3d,e), demonstrating that GLS1 is the major glutaminase isoform supporting glutaminolysis in these tumours. In the absence of GLS1, however, inhibiting Gls2 expression (Gls1 KO /shGls2; Fig. 2e) resulted in a more remarkable delay of tumourigenesis than was observed in Gls1 KO alone (Fig. 2f). This delay was associated with increased levels of the apoptosis markers cleaved caspase-3 and poly-ADP ribose polymerase (PARP), and with reduced levels of a proliferation marker, proliferating cell nuclear antigen (PCNA), when compared with control (CT/shLuc) tumours (Fig. 2g). These data suggested that the growth of Gls1 KO /shGls2 tumours was inhibited not only at the initiation stage but also during tumour progression by affecting the balance between apoptosis and proliferation. Gls1 KO /shGls2 tumours showed a trend towards even higher accumulation of glutamine in comparison with that observed in Gls1 KO tumours (Extended Data Fig. 3f) and a further suppression of glutamine catabolism into both the Krebs cycle and amino acids ( Fig. 2h and Extended Data Fig. 3g,h). Nevertheless, the levels of Krebs cycle intermediates were still not supressed by the absence of both glutaminases (Fig. 2i). Interestingly, however, the total levels of non-essential amino acids (NEAAs), whose synthesis depends on glutamine, including glutamate, alanine and aspartate, showed  13 C enrichment of the indicated metabolites extracted from either normal livers or MYC-driven liver tumours after either a bolus of [U-13 C]glucose (a, normal livers n = 5, MYC tumours n = 6 mice) or [U-13 C]-glutamine (b, normal livers n = 4, MYC tumours n = 5 mice) measured by gas chromatography-mass spectrometry (GC-MS). Tumours were initiated in LAP-tTA/TRE-MYC mice by weaning them to regular chow. Mice maintained on a doxycycline-containing diet were used as controls. Data are presented as mean ± s.d. Statistical analysis was performed using a two-tailed Student's t-test. c, Quantification of the [U- 13 C] glucose and [U-13 C]glutamine carbon incorporation into some of the key intermediates of the central carbon metabolism in MYC-driven liver tumours and control normal livers after 3 h of infusion (n = 5 mice per group). The x-axis labels correspond to the indicated isotopologues of each molecule. The content of [ 13 C]palmitic acid from a triglyceride pool is compared for tumours and corresponding adjacent livers (n = 3 mice per group). Glutamine labelling is estimated from quantification of its spontaneous product pyroglutamate 66 , and glutamate from the non-pyroglutamate fraction. The decrease in tumoural [ 13 C]glutamate and [ 13 C]glutamine pools suggests their fast catabolism and increased anaplerosis. Labelling in all carbon positions of pyruvate, lactate and alanine from both [U-13 C]glucose and [U-13 C]glutamine in tumours suggests malic-enzyme-driven pyruvate cycling. The main serine isotopologues observed during [U-13 C]glucose infusion are +1 and +2, and not +3, which can be produced owing to carbon exchange during glycine synthesis and one-carbon metabolism 67 . Data are presented as mean ± s.d. Statistical analysis was performed using a two-tailed Student's t-test. Corresponding isotopologues in tumours and normal livers infused with the same label are compared. d, Diagram depicting the relative contribution of glucose (red) and glutamine (green) to the different pathways studied. e, Western blot comparing the expression levels of key enzymes involved in the metabolic pathways depicted in d in MYC-driven tumours and normal liver (n = 3 mice per group). The antibody against GLS1 specifically detects the kidney-type (KGA) isoform. See also Extended Data Figs. 1 and 2. a tendency to be lower in both Gls1 KO and Gls1 KO /shGls2 tumours than they were in control counterparts (Extended Data Fig. 4a,b).
The levels of NEAAs in tumours lacking glutaminases are maintained by exogenous uptake. To explore a possible mechanism for the delayed formation of Gls1 KO   aspartate and alanine were lower in both HCC MYC Gls1 KO /shLuc and Gls1 KO /shGls2 cells (Extended Data Fig. 4d). The extent of these changes correlated with the extent of proliferation delay observed in these cell lines. Interestingly, the levels of citrate were not affected (Extended Data Fig. 4d), apparently being supported by glucose catabolism. These results suggested that decreasing glutamine catabolism into the Krebs cycle may affect its biosynthetic capacity to support tumour cell proliferation. Consistently, the presence of all NEAAs in the medium improved slowed proliferation of HCC MYC Gls1 KO /shGls2 cells (Extended Data Fig. 4c).
To test whether a more substantial decrease in tumour Krebs cycle intermediates and amino acids levels observed upon the decrease in glutaminase activity in vitro in comparison with in vivo was due to the uptake of the amino acids by tumours from the bloodstream, we administered a bolus of [ 15 N]alanine, the second most abundant amino acid in serum after glutamine, to mice bearing either CT/ shLuc or Gls1 KO /shGls2 tumours. Interestingly, both types of tumour were able not only to assimilate alanine but also to convert it into glutamate, and consequently the remaining NEAAs by utilization of the alanine amino group, with an initial step catalysed by alanine aminotransferases (Extended Data Fig. 4f).
Together, these data suggested that limiting a biosynthetic capacity of the Krebs cycle may be one of the factors that determines the effect of glutaminase inhibition on tumour progression. Furthermore, the uptake and catabolism of circulating amino acids can support amino acid pools in tumours and may constitute one of the elements of tumour resistance to glutaminase inhibition.
Amidotransferase action compensates for the inhibition of glutaminase activity in tumours. Although Gls1 KO /shGls2 tumours exhibited significantly decreased flux of glutamine into the Krebs cycle, they were still able to incorporate approximately half of the glutamine-derived carbons into Krebs cycle intermediates, compared with control tumours, and at a level comparable to the adjacent livers ( Fig. 2h and Extended Data Fig. 3h). The presence of a substantial fraction of the +4 isotopologue of malate derived from [U-13 C]glutamine in Gls1 KO /shGls2 tumours reflected the direct glutamine catabolism (Fig. 2j). We therefore hypothesized that enzymes other than glutaminases could be responsible for the production of glutamate from glutamine and subsequent glutamine flux into the Krebs cycle, and could be potentially targeted together with glutaminase inhibition as a potential therapeutic intervention. Indeed, several other enzymes, including those involved in asparagine, nucleotide, glucosamine, NADH and aminoacyl-tRNA biosynthesis, can utilize glutamine as an amide nitrogen donor and generate glutamate 28 (Fig. 3a,b). Interestingly, the expression of most of these enzymes was upregulated in liver tumours driven by MYC (Fig. 3c).
Their expression also correlates with the expression of Gls1 in human HCC (Extended Data Fig. 5a), indicating that GLS1 inhibitors may not be efficient as a mono-therapy in human patients and could better be combined with amidotransferase inhibitors.
Investigations of these pathways can contribute to the synthesis of glutamate from glutamine using amido-[ 15 N]glutamine in either tumours or tumour-derived cells revealed that glutamine-derived 15 N was incorporated into nucleotides (both purines, AXP and GXP; and pyrimidines, UXP, TXP and CXP), NADH, and glucosamine, among other metabolites ( Fig. 3d and Extended Data Fig. 5b,c). Interestingly, we also observed the incorporation of the amido-[ 15 N]glutamine-derived nitrogen into amino acids ( Fig. 3d and Extended Data Fig. 5b,c), suggesting the presence of ammonia recycling, probably through the glutamate dehydrogenase (GDH) reverse reaction 29 . This labelling pattern was dramatically reduced in Gls1 KO /shGls2 cells, consistent with the lack of ammonia generation, but relatively preserved in Gls1 KO /shGls2 tumours probably due to the uptake of the labelled amino acids that were detected in serum (Extended Data Fig. 5d).
To demonstrate that the combined output of transamidase reactions could be feeding the Krebs cycle with glutamine-derived carbons in the absence of glutaminases, we treated tumour-bearing mice with 6-diazo-5-oxo-l-norleucine (DON), a glutamine analog, and pan-amidotransferase inhibitor. Consistent with its function, DON ablated the incorporation of the amido-[ 15 N] glutamine-derived nitrogen into nucleotides in MYC-driven liver tumours (Extended Data Fig. 5e). In Gls1 KO /shGls2 tumours, a single dose of DON administered prior to the [U-13 C]glutamine bolus substantially suppressed the incorporation of 13 C into Krebs cycle intermediates and amino acids (Fig. 3e), increased the [ 13 C]glutamine level while decreasing the levels of [ 13 C]glutamate and total levels of alanine, malate and fumarate (Extended Data Fig. 6a-c). Interestingly, DON minimally affected glutamine-derived anaplerosis in control tumours (Fig. 3e), suggesting that glutaminases are relatively insensitive to DON, and that only the combination of the administration of DON with glutaminase inhibition can completely suppress glutamine catabolism. Consistent with the in vivo results, DON treatment almost completely ablated glutamine utilization in HCC MYC Gls1 KO   . P value was calculated by Mantel-Cox test. c, Glutamine levels in CT and Gls1 KO tumours, determined by heteronuclear single-quantum coherence spectroscopy nuclear magnetic resonance (HSQC-NMR) (n = 5 mice per group). d, 13 C enrichment from [U-13 C]glutamine in the indicated metabolites of CT and Gls1 KO tumours after [U-13 C]glutamine boluses were administered (n = 5 mice per group). GC-MS. e-j, Gls2 knockdown in Gls1 KO tumours increases the repressive effect on tumour glutaminolysis and tumour burden. Gls1 KO shGls2 tumours are compared with CT/shLuc and Gls1 KO /shLuc tumours expressing shRNA against luciferase. e, Western blot demonstrating shRNA-mediated reduction of GLS2 protein levels in tumours (n = 3 mice per group). f, Kaplan-Meier survival curve (CT/shLuc n = 10 mice, Gls1 KO /shLuc n = 6 mice, Gls1 KO /shGls2 n = 7 mice). P value was calculated by Mantel-Cox test. g, Western blot showing the protein levels of cleaved PARP, cleaved Caspase-3 and PCNA in the indicated tissues (normal livers n = 2, CT/shLuc n = 6, Gls1KO/shGls2 n = 6 mice). β-Actin was used as a loading control. h, 13 C enrichment of the indicated metabolites after a [U-13 C]glutamine bolus was administered (n = 6, 5, 6, mice; group labels as above). i, Total levels of Krebs cycle intermediates (CT/ shLuc n = 6, Gls1KO/shLuc n = 5, Gls1KO/shGls2 n = 6 mice). GC-MS. j, 13 C enrichment of malate isotopologues in the experiment shown in Fig. 2h (n = 6 mice per group). GC-MS. All data are presented as mean ± s.d. Statistical analysis was performed using a two-tailed Student's t-test. See also Extended Data Figs. 2, 3 and 4.
( Fig. 3h and Extended Data Fig. 6f). This effect was rescued by either the addition of a mix of NEAAs or a combination of four amino acids absent from DMEM, that is, alanine, aspartate, proline and asparagine (AAAP; Extended Data Fig. 6g,h). The combination of DON and CB-839 also showed a synergistic effect in the human cancer cell line HepG2 (Fig. 6i). These data demonstrate that decreasing Krebs cycle anaplerosis and amino acid biosynthesis plays a substantial role in the suppression of tumour cell proliferation downstream of combined inhibition of glutaminases and amidotransferases.    Altogether, our data demonstrate that high levels of glutamine catabolism are required for MYC-induced tumourigenesis in vivo, and that deletion of Gls1 can be partially compensated by the presence of Gls2. Furthermore, a substantial proportion of glutamine-derived glutamate can be produced by enzymes that utilize glutamine as the amide nitrogen donor, thereby fuelling the Krebs cycle.
Synergistic inhibition of Gls1 and amidotransferases synergizes in models of human colon cancer. Next, we evaluated whether the synergistic interaction between glutaminases and amidotransferases in supporting the activity of the Krebs cycle and liver tumourigenesis also exists in other tumour types. To this end we used in vitro and in vivo models of colorectal cancer, a cancer type in which GLS1 plays a role in supporting tumour growth 30 and MYC has been identified as one of the regulators of the colon cancer metabolic reprograming 31 . The combination of CB-839 and DON had a substantially stronger inhibitory effect on the proliferation of the human colon cell line, HCT116 (Fig. 4a), and proliferation and colony formation of primary mouse Apc min/+ tumour organoids (Fig. 4b,c), than CB-839 or DON alone. Importantly, this effect was also observed in primary human colon cancer organoids (Fig. 4d,e), a model demonstrating a strong predictive power for therapy efficiency in humans 32 . Finally, a combination of an orally bioavailable selective Gls1 inhibitor (compound 27) 33 and DON was the most efficient in suppressing the growth of HCT116 xenografts (Fig. 4f). Accordingly, this combination also had the strongest effect on the levels of Krebs cycle intermediates and their labelling from [ 13 C]glutamine (Fig. 4g). Interestingly, citrate levels were maintained, which reinforced the role of glucose as the main contributor to the citrate pool. Importantly, these results also demonstrated that DON alone does not suppress the entry of glutamine carbons into the Krebs cycle. DON treatment resulted in elevated levels of aspartate, which may be a result of asparagine synthetase inhibition and may increase glutamine availability to glutaminases. Together these data confirm that inhibiting both glutaminases and other enzymes utilizing glutamine as an amide nitrogen donor is required for the maximal inhibition of glutamine catabolism in tumours.

Targeting glycolysis and glutaminolysis together efficiently inhibits tumourigenesis.
Our results demonstrated that the pools of Krebs cycle metabolites were preserved in Gls1 KO /shGls2 tumours (Fig. 2h), and glucose contribution to the Krebs cycle was higher in Gls1 KO /shGls2 cells than in CT cells, and even more so when inhibition of glutaminase expression was combined with DON treatment (Fig. 3f and Extended Data Fig. 6d,e). We next tested whether concomitant inhibition of glucose and glutamine catabolism would supress Krebs cycle activity and affect MYC-induced tumourigenesis. Elevated glycolysis in MYC-induced liver tumours was associated with increased expression of HK2 (ref. 10 ) (Fig. 1e), whose deletion was previously shown to have a profound effect on mammary gland and lung tumourigenesis 17 . In contrast, Hk2 deletion at the time of tumour initiation (Hk2 KO ; Fig. 5a) did not affect the latency of MYC liver tumours (Fig. 5b), probably owing to the observed compensatory expression of glucokinase (Fig. 5a). Nevertheless, Hk2 KO tumours showed decreased catabolism of glucose into lactate ( Fig. 5c and Extended Data Fig. 7a) associated with lower levels of glucose 6-phosphate and fructose 6-phosphate (Extended Data Fig. 7b). The incorporation of glucose into Krebs cycle intermediates was also decreased in Hk2 KO tumours (Fig. 5d). However, only the level of citrate, the entry point of glucose into the Krebs cycle, was decreased, and not the levels of other Krebs cycle intermediates (Extended Data Fig. 7c). Altogether, these data suggested that glutaminolysis could be compensating for a decreased glycolytic flux and vice versa. We next deleted both Gls1 and Hk2 (Gls1 KO Hk2 KO ) at the initiation stage of MYC-induced liver tumourigenesis. This combination strikingly impaired tumour formation (Fig. 5e), with 37.5% of mice failing to develop tumours even after 1 yr. In contrast to either Gls1 KO or Hk2 KO tumours, Gls1 KO Hk2 KO tumours had lower levels of Krebs cycle intermediates ( Fig. 5f and Extended Data Fig. 7d). They had a higher level of glutamine consistent with the inhibition of glutamine catabolism (Extended Data Fig. 7e), but maintained the levels of lactate (Extended Data Fig. 7f) and amino acids (Extended Data Fig. 7g). Importantly, the AMP/ATP and GMP/GTP ratios were depressed in Gls1 KO Hk2 KO tumours in comparison with CT tumours (Fig. 5g), suggesting that Gls1 KO /Hk2 KO tumours may fail to maintain a proper energy balance. Interestingly, Gls1 KO /Hk2 KO tumours also had lower levels of pentose phosphate pathway intermediates (Extended Data Fig. 7h). In general, the tumours that developed showed a wide range of metabolite concentrations, consistent with a heterogeneous pattern of GLS2, GCK and HK1 expression (Extended Data Fig. 7i,j), suggesting different mechanisms of metabolic adaptation. Consistent with the previous findings 17 , some tumours exhibited a residual expression of HK2 suggesting that they originated from cells with incomplete deletion of the gene. This observation reinforced the idea that the deletion of both HK2 and GLS1 is highly detrimental for tumourigenesis. Consistently, treatment with either UK5099, an inhibitor of the mitochondrial pyruvate transporter, or 2-deoxy-d-glucose (2DG) affected the proliferation of Gls1 KO /shGls2 cells more than of CT cells (Fig. 5h,i). Overall, these results demonstrate that, when the catabolism of either glucose or glutamine into the Krebs cycle is individually blocked, the other metabolite can compensate to support Krebs cycle activity and downstream pathways. Concomitantly, reducing both anaplerotic pathways simultaneously could supress tumourigenesis more efficiently than inhibiting each pathway individually.
Dietary interventions synergize with the inhibition of tumour biosynthetic pathways. One of the pathways that is fuelled by both glucose and glutamine in MYC-induced liver tumours is fatty acid biosynthesis (Fig. 1c,d). We tested whether this major pathway is required for the formation of MYC-induced tumours by deleting the Fasn gene at the onset of tumourigenesis (Fasn KO ). Fasn knockout delayed the onset of palpable tumours but did not lengthen the total time to a clinical endpoint (Fig. 6a,b). A total block of fatty acid labelling and unaffected labelling of citrate from [U-13 C]glucose confirmed the inhibition of de novo lipogenesis in tumours at the FASN step (Fig. 6c,d; Extended Data Fig. 8a). Interestingly, the defect in lipogenesis in Fasn KO tumours was not associated with the lowering of the total content of fatty acids but rather with and altered fatty acid composition (lower c16:0/c18:0 and c16:0/c18:2 ratios; Extended Data Fig. 8b,c). This outcome suggested that in the absence of Fasn and endogenous lipid production tumour development was sustained by different exogenous lipids and the tumour lipidome was determined by the lipid composition of the diet. Consistently, MYC-induced tumours express a wide range of lipid transporters (Extended Data Fig. 8d) and both CT and Fasn KO tumours imported fluorescently-labelled lipids from serum more avidly than the adjacent liver and most other tissues (Fig. 6e,f).
To test whether modulating lipid availability would synergize with Fasn deletion, we placed mice bearing Fasn KO and CT tumours onto a diet with a lower fat content (LFD, 5.1% in comparison with 9% in our normal chow diet). The LFD remarkably increased the latency of Fasn KO tumours (Fig. 6g), demonstrating that the uptake of circulating lipids was supporting the growth of Fasn KO tumours at a rate comparable to CT tumours. Paradoxically, Fasn KO tumours of mice on the LFD showed a tendency to have enlarged lipid pools measured at the endpoint (Extended Data Fig. 8e), but again with the change in fatty acid composition, reflected by decreased 16:0/18:0 and 16:0/18:2 ratios (Extended Data Fig. 8f). Consistent with the in vivo results, cells isolated from Fasn KO tumours showed a lower proliferation rate compared than that of CT cells in complete medium, and stopped proliferating in lipid-depleted medium, both of which were rescued by fatty acid supplementation (Extended Data Fig. 8g). Notably, a previously published Fasn inhibitor, Fasnall 34 , was only able to markedly decrease the growth of HCC MYC allografts when combined with the LFD (Fig. 6h). Together, these results demonstrate that lipid requirements of tumours in vivo are sustained by both de novo lipogenesis and lipid uptake, and that strategies that do not target both pathways could be less effective. Another major pathway fuelled by both glucose and glutamine, demonstrated to support the proliferation of cancer cells, is serine biosynthesis 35 . Serine and its immediate metabolic product glycine are essential precursors for protein, nucleic acid, folate and lipid synthesis [35][36][37][38] . PSAT1 catalyses one of the key steps in serine biosynthesis when glucose-derived carbon is combined with glutamine-derived nitrogen (Fig. 7a). MYC-induced liver tumours had heightened serine biosynthetic capacity and substantially elevated PSAT1 levels (Figs. 1a,c,e and 7b and Extended Data Figs. 1b and 2b). Thus, we tested whether directly inhibiting serine biosynthesis would affect MYC-induced tumourigenesis. Surprisingly, Psat1 deletion at the time of tumour initiation (Psat1 KO ) had no effect on a tumour onset or progression (Fig. 7b,c). [U-13 C]Glucose infusions and [amino- 15 N]glutamine boluses demonstrated the ablation of de novo serine and glycine synthesis in Psat1 KO tumours ( Fig. 7d and Extended Data Fig. 9a,b). Interestingly, only serine, and not glycine, levels were reduced (Fig. 7d), suggesting that serine pools were more dependent on serine synthesis, whereas the glycine level may be sustained by direct import from the bloodstream. To test whether a reduction in serine/glycine (SG) availability in vivo would synergize with the absence of PSAT1, we placed mice with either Psat1 KO or CT tumours on a SG-deficient diet (-SG; Fig. 7e) that has been shown to affect tumour progression in some mouse models of cancer 39 . Remarkably, the -SG diet had no significant effect on the latency of CT tumours (Fig. 7f), in which the glycine levels, but not serine levels, were substantially reduced (Fig. 7g). Only the combination of Psat1 deletion with the -SG diet achieved a reduction in the SG levels, and a striking delay in tumour progression (Fig. 7f). Consistent with the in vivo results, wild-type Psat1 tumour cells were able to proliferate in the absence of both serine and glycine, while Psat1 KO tumour cells required serine but not glycine for proliferation (Extended Data Fig. 9c). In the absence of serine, Psat1 KO cells were able to proliferate only when formate, a carbon donor for one-carbon metabolism normally present in serum, was supplied (Extended Data Fig. 9c). These data demonstrate the need for targeting both the synthesis and the dietary supply of serine and glycine, and also suggest a limitation in the interconversion of these amino acids. Indeed, one of the factors preventing the production of serine from glycine can be the depletion of methyl groups from one-carbon metabolism, which would limit nucleotide biosynthesis. Surprisingly, a very profound delay in the formation of Psat1 KO / -SG tumours was associated with only a few detectable metabolic changes. Psat1 KO /-SG tumours had a lower level of α-ketoglutarate, but not any other Krebs cycle intermediates (Extended Data Fig.  9d). While Diehl et al. 40 reported a decrease in purine nucleotide biosynthesis in cancer cell lines deprived of serine, we did not observe any effect of the -SG diet on the levels of purine nucleotides, possibly owing to 50% of serine still being present in serum of animals on the -SG diet (Extended Data Fig. 9e). The appearance of [ 15 N]serine and [ 15 N]glycine in the blood of tumour-bearing animals injected with [amino-15 N]glutamine demonstrated that in the absence of dietary SG, blood SG levels can be sustained by synthesis and export from other organs (Extended Data Fig. 9e). One of the differences that was observed, however, upon Psat1 KO and -SG diet was in the level of alanine, which was significantly increased in Psat1 KO /-SG tumours (Extended Data Fig. 9a, d). Serine depletion has been shown to affect ROS balance 39 , and alter the metabolism of glycerophospholipids and sphingolipids 38 , leading to intracellular accumulation of cytotoxic deoxysphingolipids 41 derived from alanine instead of serine. Therefore, an elevated alanine-to-serine ratio could be one of the causes of the increased latency of Psat1 KO / -SG tumours.
Together, these data demonstrate that both SG synthesis and uptake can support the progression of MYC-induced liver tumours, and that a 50% reduction in the blood serine levels is sufficient to synergize with the inhibition of serine biosynthesis and profoundly impairs tumourigenesis.

Discussion
A number of compounds that inhibit metabolic enzymes, including GLS1 and FASN, are currently in different phases of clinical trials as potential cancer therapeutics. However, tumour metabolic plasticity could be a major obstacle to their success. Our in vivo results illustrate how cancer cells can withstand interventions that target metabolism, and explore different options for combinatorial targeting of specific metabolic adaptations for greater impact on tumour development.
Our data build on the previous work that has shown promising but heterogeneous outcomes with the inhibition of individual metabolic enzymes in different mouse models 4,[17][18][19][20][21]42 , suggesting that both the tissue of origin and the genetic drivers can influence not only the metabolic reprogramming during oncogenic transformation 10 but also the mechanisms of metabolic adaptation. Deleting those isoforms that drive glucose and glutamine catabolism in a single tumour system allowed us to reveal several mechanisms of metabolic compensation that should be considered when designing metabolism-based therapeutic strategies. Inhibiting the expression of either Hk2 or Gls1 at the onset of MYC-induced liver tumourigenesis resulted in the formation of tumours with re-expression of, respectively, Gck or Gls2-the isoforms expressed in the normal liver and repressed in tumours. We cannot exclude that the expression of liver isoforms in tumours is the result of the selection of transformed cells that maintain their expression during transformation. Nevertheless, these observations demonstrate that the expression of the enzyme isoforms in a parental tissue can determine the degree of tumour sensitivity to a metabolism-focused therapy. The underlying mechanisms that drive the expression of isoforms are yet to be elucidated.
Furthermore, our results emphasize the role of glucose and glutamine as anaplerotic nutrients that support Krebs cycle activity. The  KO mice on the LFD n = 10). The P value was calculated by Mantel-Cox test. Gene deletion was achieved during tumour induction by co-injection of pT3-CMV-Cre with c-MYC and MCL1-expressing plasmids in Fasn fl/fl and wild-type mice. Animals were moved onto the indicated diet 1 week later. h, The effect of combining the LFD and Fasn inhibitor, Fasnall, on the progression of MYC tumour-cell-derived allografts (Veh/9% fat diet n = 9 mice; Fasnall/9% fat diet n = 9 mice; Veh/4% fat diet n = 11 mice; Fasnall/4% fat diet n = 11 mice). All data are presented as mean ± s.d. Statistical analysis was performed using a two-tailed Student's t-test. In c, *P < 0.05, **P < 0.01 and ***P < 0.001, with respect to its corresponding adjacent liver; # with respect to the normal liver; α with respect to Fasn KO tumours. See also Extended Data Fig. 2 and 8. flux of carbons from either substrate can compensate and support Krebs cycle intermediate pools when the other is limited. Indeed, only reducing both glucose and glutamine catabolism considerably lowered these pools and affected tumour formation. These results further substantiate previous studies on the interactions between glucose and glutamine catabolism that support the survival of cancer cells 22-25 . Our results demonstrate that not only GLS1 but also synergistic action of glutaminases and amidotransferases can fuel glutamine-derived carbons to the Krebs cycle. This is clinically relevant because the expression of GLS1 and different amidotransferases is strongly correlated in human HCCs, suggesting that such patients may benefit from only a combination of compounds that inhibit these enzymes.  as an inhibitor of both amidotransferases and glutaminases, on the basis of previous results in vitro that employed relatively high doses 43,44 . In contrast, our results suggest that at therapeutically relevant doses, DON does not inhibit glutaminases. This finding opens the door to exploring the combination of GLS1 inhibitors with inhibitors of other glutamine-catabolising enzymes, such as DON. Importantly, DON prodrugs have been developed to enhance the delivery of an active compound to tumour tissues and circumvent gastrointestinal toxicity 45,46 .
Our results also shed light on the key roles played by dietary intake and host tissues in supplying nutrients that allow tumour cells to shift to an auxotrophic phenotype, emphasizing the potential importance of dietary intervention as an adjuvant cancer therapy. Strategies that individually target serine synthesis, or reduce its availability, have been the focus of recent studies. The -SG diet affects tumour growth in some models, including Eμ-Myc lymphoma and Apc min -induced intestinal tumours, while leaving KRas-induced pancreatic and intestinal tumours intact 39 . The -SG diet also enhanced the antineoplastic activity of biguadines on colon adenocarcinoma allografts, but had only a minor effect as a single therapy 47 . Our results suggest that the ability of tumours to synthesize serine makes them resistant to the -SG diet, and inhibiting serine biosynthesis is necessary to broaden the spectrum of cancer types that respond positively to this dietary regime. Importantly, inhibitors of serine biosynthesis pathway have been developed 36,48 , and the sensitivity towards these inhibitors may also be determined by the concentration of serine and glycine in a specific tissue 49 . Together with inhibiting SG biosynthesis in tumours, these compounds should also be able to suppress the synthesis of SG by the host tissues. In this context, a partial reduction of dietary serine and glycine, which is more feasible to implement in a clinic than total depletion, might suffice to deplete serine/glycine serum levels and affect tumour progression.
Similar to the situation with serine biosynthesis, different tumours have varying requirements for de novo fatty acid synthesis 21,42 . The FASN inhibitor TVB2640, alone or in combination, is currently being tested in different clinical trials (ClinicalTrials. gov identifiers: NCT02223247, NCT03032484, NCT02980029 and NCT03179904), and appears to shows anti-tumour activity 50 . Our results suggest that decreasing lipid availability or inhibiting lipid uptake should be considered as an adjuvant for therapies that target de novo lipogenesis. The inhibition of lipid transporters is one option 51 . However, the multitude of uptake mechanisms could present a major obstacle. Therefore, even though low-fat diets on their own seem to have minor effects on human cancer progression 52 , they represent an attractive alternative.
Although targeting several metabolic pathways may seem to be more detrimental for normal cells and tissues than a mono-therapy, our results and those of others 24,25 demonstrate that a therapeutic window exists. One of the contributory factors is the enhanced likelihood that complementary metabolic activities co-exist and are elevated in tumour cells in comparison with a given normal cell arising out of their different metabolic demands. A deeper understanding of the differences in metabolic requirements and the response to metabolic interventions between different tumour cells and normal cells is required to design therapeutic approaches with minimal side effects.
In conclusion, our study illuminates how the understanding of metabolic adaptations in vivo is crucial for the design of effective therapeutic strategies for cancer, and that exploitation of combinatorial interventions against compensatory metabolic pathways may lead to more robust inhibition of tumour growth.

Mice.
Mice transgenic for both LAP-tTA and tet-o-MYC that overexpress MYC in the liver of males were generated as previously described 9 . To initiate MYC expression, male mice were weaned to regular chow. Male mice kept on a doxycycline-containing diet did not overexpress c-MYC and were used as controls (normal livers). Gls1 fl/fl mice 53 , Hk2 fl/fl mice 17 and Fasn fl/fl mice 42 have been previously described. Psat1 fl/fl mice were derived from Psat1 tm1a(KOMP)Wtsi mice, obtained from the KOMP repository (University of California, Davis, MGI:4363603). Gls1 fl/fl , Hk2 fl/fl , Psat1 fl/fl and Fasn fl/fl mice, originated in different backgrounds, were back-crossed for a minimum of ten generations with an FVBN/J strain to allow comparison among all the experimental groups. All lines were crossed with Alb-CreER T2 (ref. 54 ) and Rosa26eYFP lines 55 to generate Gene fl/fl / Alb-CreER T2 /Rosa26eYFP lines. Gene deletion was achieved by intraperitoneal (i.p.) administration of tamoxifen (Sigma; 10 mg per kg (body weight); dissolved in 1:10 ethanol:oil solution) to mice starting at 5-6 weeks of age. Mice were bred and maintained under specific-pathogen-free conditions at The Francis Crick Institute (Mill Hill laboratory and Midland Road Laboratory). Mice were palpated to detect liver tumours two times per week starting at 3 weeks post hydrodynamics-based transfection or doxycycline diet removal. All mice were closely monitored for any signs of distress, poor health or body-weight loss (weight loss of either 20% from the start of an experiment, or 15% in the last 72 h, was considered a humane end point). Tumours were identified as hard masses in an otherwise soft abdominal area or by an enlarged abdomen at later stages. A 20% increase in the normal abdomen diameter was considered a humane endpoint. All procedures and animal husbandry were carried out in accordance with the UK Home Office, under the Animals (Scientific Procedures) Act 1986, and the Local Ethics Committee under the Project license number P609116C5.

Constructs and reagents.
The constructs for mouse injection, including pT3-EF1α-c-MYC, pT3-EF1α-MCL1 and pCMV-SB (which encodes for sleeping beauty transposase), and pCMV-CRE (which encodes Cre recombinase), were previously described 56  Hydrodynamics-based transfection of DNA in the liver. All mice used in the experiments were males. Hydrodynamic-based transfection was performed as described 57 , with some variations. Briefly, 10 μg of pT3-EF1α-c-MYC, pT3-EF1α-c-MYC/shGls2 or pT3-EF1α-c-MYC/shLuc with 10 μg of pT3-EF1α-Mcl1 along with sleeping beauty transposase (SB) in a ratio of 25:1 were diluted in a volume of saline (0.9% NaCl) corresponding to 10% of body weight, and injected into the lateral tail vein of 7-to 9-week-old FVB/N mice in 7 to 9 s. Since the ectopic expression of Cre recombinase has been previously shown to exacerbate MYC-induced apoptosis and reduce the efficiency of MYC-induced tumourigenesis 58 , the ectopic expression of MYC by hydrodynamics-based transfection was combined with the expression of MCL1. In these cases, a gene deletion was achieved by co-injection of 40 μg of Cre-recombinase (pT3-CMV-Cre) with c-MYC and MCL1-expressing plasmids and 4 μg of SB-encoding plasmid. Titration of the amount of SB allowed us to ensure there was a reliable amount of integration events and induction of tumourigenesis. Mice used as controls (normal livers, Figs. 2-7) were injected with an empty plasmid (pT3-EF1α-MCS). The timescale for tumour burden is specific for different methods of c-MYC induction and gene knockout.
Xenograft experiments. We subcutaneously injected 5 × 10 6 human HCT116 into the flanks of 6-to 8-week-old male SCID mice. We injected 5 × 10 6 mouse HCC MYC cells into the flanks of FVB/N mice. Xenografts were allowed to grow 2-3 mm before mice were randomized into groups for treatment. Tumours were measured with a calliper every 3 d, and the volume was calculated using the formula for a hemiellipsoid (volume = (π / 6) × length × width 2 ). A tumour diameter of 1.5 cm was considered a humane end point. All mice were closely monitored for any signs of distress, poor health or body weight loss (weight loss of either 20% from the start of an experiment, or 15% in the last 72 h, was considered a humane end point). All tumours were collected at the end of the experiment and stored for further processing.
Diets and compound administration. For the dietary amino acid restriction, two synthetic diets were used: TestDiet Baker amino acid diet with no added serine and glycine 5W53, and the control diet Baker amino acid defined diet 5CC7. Mice were placed on diets 1 week before hydrodynamics.
For the low-fat diet experiment, a Low-Fat Control for Western Diet (5TJS) from Test Diet was used. We did not use a zero-fat diet to avoid its reported paradoxical effect on lipid accumulation in Fasn KO non-cancerous livers 59 . To improve acceptance of the diet, in this experiment 5TJS was initially mixed with 5CC7 pellets (both diets have 5.1% fat content). Mice were placed on diets 1 week after hydrodynamics.
For the treatment with the Fasn inhibitor Fasnall, 4% and 9% Fat diets from Teklad Global Diets were used (2914 C and 2919, respectively). Diet was changed 6 d after injection of tumour cells, when tumours reached 0.3 mm diameter, and 25 mg per kg (body weight) Fasnall was administered every 72 h i.p from one day after diet change. The vehicle was 50% DMSO, 25% saline, 25% PBS.
For short-term DON administration, 50 mg per kg (body weight) DON was administered 4 h before labelling. For the treatment of mice bearing HCT116 xenografts, 25 mg per kg (body weight) DON and 100 mg per kg (body weight) GLS1 inhibitor (compound 27) were used. DON was dissolved in saline and administered every 72 h i.p. Compound 27 (ref. 33 ) was dissolved at 10 mg ml -1 in 1% tween 80, the pH was adjusted to 3.5 and was then kept at room temperature under constant mixing for a maximum of 7 d. Compound 27 (and vehicle) was administered every 24 h via oral gavage.
Stable-isotope labelling in vivo. Two types of in vivo label administration were performed: either bolus injections (once for glucose, endpoint: 15 min; twice for glutamine, endpoint: 30 min; and one for alanine, endpoint: 30 min), or a long-term infusion (3 h). Bolus allowed us to evaluate the capacity of tissues to take up and catabolize these nutrients. Intravenous infusions allowed us to measure a tracer incorporation into various metabolites, including lipids, and to calculate the steady state partial contribution of glucose and glutamine to different pathways.
Boluses of stable-isotope-labelled compounds dissolved in saline (0.9% NaCl) were administered through the tail vein (i.v.). [U- 13  Infusions were performed in animals under isofluorane anaesthesia, through a tail vein catheter, using an Aladdin AL-1000 pump (World Precision Instruments). The infusions protocol was based on the experiments described previously 60 . For [U-13 C]glucose, mice received a bolus of 0.4 mg per g (body weight), followed by a 0.012 mg per g per min infusion for 3 h at 0.15 ml h -1 . For glutamine-derived stable-isotope infusions, mice received a bolus of 0.187 mg per g (body weight), followed for a 0.005 mg per g per min infusion for 3 h at 0.15 ml h -1 . At the endpoint, blood was obtained by cardiac puncture under terminal anaesthesia, and tissues were then collected and freeze-clamped with a Wollenberger-like device pre-cooled in liquid nitrogen. The tissues were stored at −80 °C until analysis.
For GC-MS analysis of the polar metabolites, part of the polar fraction was washed twice with methanol, derivatized by methoximation (Sigma, 20 µl, 20 mg ml -1 in pyridine) and trimethylsilylation (20 µl of N,O-bis(trimethylsilyl) trifluoroacetamide reagent (BSTFA) containing 1% trimethylchlorosilane (TMCS), Supelco), and analysed on an Agilent 7890A-5975C GC-MS system 61,62 . Splitless injection (injection temperature 270 °C) onto a 30 m + 10 m × 0.25 mm DB-5MS + DG column (Agilent J&W) was used, using helium as the carrier gas, in electron ionization (EI) mode. The initial oven temperature was 70 °C (2 min), followed by temperature gradients to 295 °C at 12.5 °C per min and then to 320 °C at 25 °C per min (held for 3 min). Amino-terminal glutamine residues spontaneously cyclize to become pyroglutamic acid (5-oxoproline) in the ionization source. The levels of pyroglutamic acid were used as a reference for the levels of glutamine in GC-MS experiments. It should be considered that in these experiments, a fraction of glutamic acid can also contribute to pyroglutamic acid pool.
For GC-MS analysis of fatty acids, part of the apolar fraction was washed twice with methanol, derivatized by methoximation (Sigma, 20 µl, 20 mg ml -1 in pyridine) and analysed on an Agilent 7890A-5975C GC-MS system 61,62 . Splitless injection (injection temperature 270 °C) onto a 30 m + 10 m × 0.25 mm DB-5MS + DG column (Agilent J&W) was used, using helium as the carrier gas, in electron ionization (EI) mode. The initial oven temperature was 50 °C (1 min), followed by temperature gradients to 190 °C at 20 °C per min (held for 3 min), then to 242 °C at 4 °C per min, then to 292 °C at 10 °C/min and then to 320 °C at 20 °C per min. Triglycerides were purified by solid-phase extraction on aminopropyl silica columns (Biotage Insolute NH2), with tripentadecanoin as an internal standard.
Metabolite quantification and isotopologue distributions were corrected for the occurrence of natural isotopes in both the metabolite and the derivatization reagent. Data analysis and peak quantifications were performed using MassHunter Quantitative Analysis software (B.06.00 SP01, Agilent Technologies). The level of labelling of individual metabolites was corrected for natural abundance of isotopes in both the metabolite and the derivatization reagent 62 . Abundance was calculated by comparison to responses of known amounts of authentic standards.
NMR spectra were acquired at 25 °C with a Bruker Avance III HD instrument with a nominal 1 H frequency of either 700 or 800 MHz using 3-mm tubes in a 5-mm CPTCI cryoprobe. For 1 H 1D profiling spectra, the Bruker pulse program noesygpr1d was used with a 1-s presaturation pulse (50-Hz bandwidth) centred on the water resonance, 0.1-ms mixing time and 4s acquisition time at 25 °C. Typically 128 transients were acquired. To monitor the fate of 15 N atoms derived from 15 N-glutamine labelling, we acquired 2D 15 N-1 H heteronuclear multiple-bond correlation (HMBC) spectra at 1 H 700 MHz using the Bruker pulse program hmbcf3gpndqf, adapted to include solvent water resonance irradiation during the relaxation delay. Typically, the acquisition parameters employed were sweep widths 13 ppm ( 1 H) and 220 ppm ( 15 N), with offsets on the solvent water signal ( 1 H) and at 120 ppm ( 15 N). Acquisition times were 0.86 s ( 1 H) and 0.012 s ( 15 N), and pulse widths 7.4 μs ( 1 H) and 25 μs ( 15 N). There were 16 transients collected for each increment, yielding a total measurement time of 4 h 16 m. The spectra were processed and plotted using Bruker TopSpin 3.5. The raw data were apodized with 1 Hz line broadening in the 1 H dimension and unshifted sine bell in the 15 N dimension and zero-filled to a matrix size of 32 K × 512 points. The spectra are presented in a magnitude mode. 13 C incorporation was assessed using 2D 13 C-1 H heteronuclear single-quantum coherence (HSQC) spectroscopy with the pulse sequence hsqcetgpsisp2 using sweep widths 14 ppm ( 1 H) and 165 ppm ( 13 C) and offsets on the solvent water signal ( 1 H) and 70 ppm ( 13 C). Acquisition times were 0.16 s ( 1 H) and 0.0176 s ( 13 C), and pulse widths 7.4 μs ( 1 H) and 11 μs ( 13 C). Eight transients were collected for each increment. Non-uniform sensing (35%) of the data points in the indirect dimension was employed, yielding a total measurement time of 1 h 46 m. The spectra were reconstructed using the compressed sensing algorithm in Topspin, using Lorentz-to-Gaussian transformation (LB = −1 Hz; GB = 0.08) and cosine apodization in the 1 H and 13 C dimensions, respectively, and zero-filling to 4 K ´ 1 K points. NMR spectra were analysed with rNMR software 63 .
Metabolite analysis was performed by LC-MS using a Q-Exactive Plus (Orbitrap) mass spectrometer coupled to a Vanquish UHPLC system (both Thermo Fisher Scientific). The chromatographic separation was performed on a SeQuant Zic pHILIC (Merck Millipore) column (5-μm particle size, polymeric, 150 × 4.6 mm). The injection volume was 10 μl, the oven temperature was maintained at 25 °C and the autosampler tray temperature was maintained at 4 °C. Chromatographic separation was achieved using a gradient program at a constant flow rate of 300 μl per min over a total run time of 25 min. The elution gradient was programmed as decreasing percentage of B from 80% to 5% during 17 min, holding at 5% of B during 3 min and finally re-equilibrating the column at 80% of B during 4 min. Solvent A was 20 mM ammonium carbonate and solvent B was acetonitrile. Metabolites were identified and quantified by accurate mass and retention time and by comparison to the retention times, mass spectra, and responses of known amounts of authentic standards using TraceFinder 4.1 EFS software (Thermo Fisher Scientific). Label incorporation and abundance was estimated using TraceFinder 4.1 EFS software. The level of labelling of individual metabolites was estimated as the percentage of the metabolite pool containing one or more 13 C atoms after correction for natural abundance isotopes. Abundance was given relatively to the internal standard.
Cell experiments. The HCT116 colon cancer cell line and HepG2 hepatocellular carcinoma cell line were obtained from the Francis Crick cell lines repository (Cell cervices/Science Technology platform). The cell lines were tested for mycoplasma contamination and confirmed negative. All human cell lines undergo routine quality control, which includes STR profiling and species identification for validation. For isolation of tumour cells, collected tissue was maintained in the ice-cold serum-free DMEM until processing (not more than 10 min). Tumours were minced in a 10-cm Petri dish using a scalpel, and washed twice with ice-cold HBSS containing EGTA. Then, intensive additional mincing was performed and 10 ml of digestion medium (HBSS containing 4 mM Ca 2+ , 5.5 mM glucose, 2 mM glutamine and 40 μg per ml of Liberase TM (Roche)) were added, and incubated for 15-20 min at 37 °C. The digested mixture was then passed through a 100-µm sterile nylon mesh cell strainer into a sterile 50-ml conical tube. A 100-µm mesh allowed small groups of cells to pass through, which lead to better survival results in some genotypes. Filtered cells were then centrifuged for 5 min at 200g. The supernatant was removed, and cells were resuspended in the washing medium (MEM-Eagle for suspension culture plus 2 mM glutamine; Biological Industries). Cells were then seeded in DMEM containing 25 mM glucose, 2 mM glutamine, 10% FBS, 1% penicillin-streptomycin.
To knock-out Psat1 in cells isolated from MYC-driven tumours induced in Psat1 fl/fl mice, the retroviral vector MSCV-CreERT2 puro (Addgene plasmid no. 22776) was used. After transduction, cells were selected with puromycin, and then treated with 4-hydroxytamoxifen (4-OHT) to induce Cre activity.
Gls1 KO /shLuc, Gls1 KO /shGls2 cells, Fasn KO cells and their CT counterparts were directly isolated from the respective tumours. The experiments for the analysis by GC-MS analysis were performed in medium containing 10 mM glucose and 2 mM glutamine in 6-well plates and 10-cm plates for NMR analysis. Cells were collected at sub-confluency with the medium being refreshed 3 h before the extraction.
Metabolite extraction from cells was performed as described for tissues except the first step when the media was aspirated and cells were rapidly washed with ice-cold PBS and 600 μl of ice-cold methanol containing standards were added. Cells were scraped and transferred to an Eppendorf and 1.2 ml of chloroform were added. Cell proliferation was monitored by using the IncuCyte system (Essen Bioscience). In each experiment, each condition was plated in triplicate. Growth curves were generated from data points acquired during 4 h interval imaging.
For immunofluorescence, tissues were fixed for 24 h in 4% paraformaldehyde, equilibrated in 30% sucrose, embedded in OCT compound (VWR international) and stored at −80 °C until 7-μm-thick cryosections were obtained. Sections were permeabilized by incubating with 0.2% Triton X-100 in PBS for 10 min, and treated with the indicated primary antibody for 16 h. Primary antibodies used were: GFP (4745-1051, Bio-Rad) and PanCK (Z0622, Dako). After four washes with 0.05% Tween in PBS, cells were incubated with anti-rabbit Alexa Fluor 488 or 555 secondary antibodies (Invitrogen) for 1 h. Nuclear marker DAPI was included in the mounting medium. Samples were examined using a confocal microscope (TCS SP5 II Leica) and LCS Lite software (Leica) was used to collect digital images.
Mouse intestinal organoid culture. Organoids were stablished from tumours isolated from APC min mice using a previously described protocol 64 , with an additional collagenase and dispase digestion step after the EDTA-chelation step 65 where Matrigel was replaced with Cultrex BME, Type 2 RGF PathClear (Amsbio 3533-010-02). Organoids were cultured in Intesticult Organoid Growth medium (06005, Stem Cell Technologies). The Rho kinase inhibitor Y-27632 (Sigma) was added to the culture when trypsinized.

Nature MetabolisM
Extended Data Fig. 1 | Glucose and glutamine metabolism in MYC liver tumours. a-g, Mice bearing MYC-driven liver tumours and control mice (n=5 per group) were infused with [U-13 C]glucose (a-c), or [U-13 C]glutamine (d-f) and the label incorporation into tissue metabolites was analysed by GC-MS: percent enrichment in either serum glucose (a) or glutamine (d) in mice administered either [U-13 C]glucose or [U-13 C]glutamine bolus, respectively (*glutamine enrichment is estimated from quantification of its spontaneous product pyroglutamate); (b and e) percent enrichment; (c and g) total content of metabolites. Note that lower glutamine enrichment in Krebs cycle intermediates in tumours in comparison with normal livers is proportional to the difference in serum enrichment between control and tumour-bearing mice; normalized values for Krebs cycle metabolites from (b and e) are shown in (g). Data are presented as mean ± S.D. Statistical analysis was performed using a two-tailed Student's t-test. Complete list of exact p-values is provided as a source data file. h, 15 N-HMBC 2D NMR signals of the indicated metabolites in the indicated mouse tissues after amino-15 N-glutamine bolus. Spectra of three representative mice per group are shown (n=5 mice per group). (i) Percent enrichment of tissue total fatty acids (both free and esterified) after either [U-13 C]glucose or [U-13 C]glutamine infusions (n=5 mice per group). GC-MS. Data are presented as mean ± S.D. Statistical analysis was performed using a two-tailed Student's t-test. Complete list of exact P values is provided as a source data file. j, Nile red fluorescence visualised by confocal microscopy showing neutral lipid accumulation (in red) in MYC-driven tumours. Nuclei are shown in blue. Representative images out of four normal livers and tumour-bearing livers are shown. Fig. 3 | Metabolic consequences of the deletion of either Gls1 or Gls2 in MYC liver tumours. a, Total concentration of 13 C-labelled Krebs cycle metabolites in CT and Gls1 KO tumours after a [U-13 C]glutamine bolus (n = 5 mice per group). GC-MS. b, Western blot of the samples presented in Fig. 2a showing the expression of GLS2 in Gls1 KO tumours (normal livers n = 3 mice, CT tumours n = 5 mice, Gsl1 KO tumours n = 5 mice). c-e, shRNA mediated Gls2 knock down in MYC-driven liver tumours with intact Gls1 expression does not affect tumour burden or glutamine catabolism. Liver tumours were induced by hydrodynamics-driven co-transfection of plasmids encoding MYC that included a miR30 based shRNA targeting for Gls2 or Renilla Luciferase (pT3-EF1α-c-MYC/shGls2 and pT3-EF1α-c-MYC/shLuc, respectively), and a plasmid encoding MCL1 (pT3-EF1α-MCL1) (c) Western blot demonstrating efficient Gls2 knock down (normal livers n = 3 mice, shLuc tumours n = 5 mice, shGls2 tumours n = 5 mice). d, Kaplan-Meier survival curve (shLuc n=13 mice; shGls2 n=11 mice). Pvalue was calculated by Mantel-Cox test. e, 13 C-enrichment in the indicated metabolites extracted from either shLuc or shGls2 tumours (Gls1 wild type) after a [U-13 C]glutamine bolus (n = 5 mice per group). f, Total level of glutamine and glutamate in CT/shLuc (n = 6), Gls1 KO /shLuc (n = 5), and Gls1 KO /shGls2 (n = 6) tumours. LC-MS. g, Isotopologue distribution of the 13 C-enrichment of glutamine in the serum of mice shown in Fig. 2h-j, Extended Data Fig. 3f,h and 4b (CT/shLuc n = 6 mice, Gls1 KO /shLuc n = 4 mice, Gls1 KO /Gls2 KO n = 5 mice). GC-MS. Glutamine enrichment was estimated from quantification of its spontaneous product pyroglutamate. h, 13 C-enrichment of the indicated metabolites after a [U-13 C]glutamine bolus, related to Fig. 2h,i, shows the enrichment of glycolytic intermediates from [U-13 C]glutamine through gluconeogenesis in tumours and the respective adjacent livers (CT/shLuc n = 6 mice, Gls1 KO /shLuc n = 5 mice, Gls1 KO /Gls2 KO n = 6 mice). All data are presented as mean ± S.D. Statistical analysis was performed using a two-tailed Student's t-test. Complete list of exact P values is provided as a source data file.