Combined deletion of Glut1 and Glut3 impairs lung adenocarcinoma growth

Glucose utilization increases in tumors, a metabolic process that is observed clinically by 18F-fluorodeoxyglucose positron emission tomography (18F-FDG-PET). However, is increased glucose uptake important for tumor cells, and which transporters are implicated in vivo? In a genetically-engineered mouse model of lung adenocarcinoma, we show that the deletion of only one highly expressed glucose transporter, Glut1 or Glut3, in cancer cells does not impair tumor growth, whereas their combined loss diminishes tumor development. 18F-FDG-PET analyses of tumors demonstrate that Glut1 and Glut3 loss decreases glucose uptake, which is mainly dependent on Glut1. Using 13C-glucose tracing with correlated nanoscale secondary ion mass spectrometry (NanoSIMS) and electron microscopy, we also report the presence of lamellar body-like organelles in tumor cells accumulating glucose-derived biomass, depending partially on Glut1. Our results demonstrate the requirement for two glucose transporters in lung adenocarcinoma, the dual blockade of which could reach therapeutic responses not achieved by individual targeting.


Introduction
Glucose transporters are the first and rate-limiting step for cellular glucose utilization, a process often exacerbated in tumor cells that enables their growth and proliferation (Ancey et al., 2018;Lunt and Vander Heiden, 2011). Although inhibiting glucose metabolism in lung tumors could become an efficient treatment strategy (Hensley et al., 2016;Patra et al., 2013;Xie et al., 2014), whether and which glucose transporter(s) should be targeted remains unclear because of their possible functional redundancy. Additionally, in solid cancers, tumor cell growth can be fueled by nutrients other than glucose, or by using transporter-independent metabolic processes including autophagy and macropinocytosis (Commisso et al., 2013;Karsli-Uzunbas et al., 2014;Romero et al., 2017;Son et al., 2013).
In this study, we exploited the Kras LSL-G12D/WT ; Trp53 Flox/Flox (KP) mouse model of lung adenocarcinoma to explore the importance of glucose transporters, expressed by tumor cells, in disease development. By glucose tracing and ultrastructural analyses, we identified the presence of lamellar body-like organelles in tumor cells, which are the primary site of glucose-derived biomass accumulation, occurring partly in a Glut1-dependent manner. We show that the deletion of Glut1 or Glut3 is not sufficient to decrease tumor progression, which is only affected significantly upon combined Glut1 and Glut3 loss.

Results and discussion
To interrogate the importance of particular glucose transporters in the most frequent subtype of lung cancer, lung adenocarcinoma (LUAD), we used The Cancer Genome Atlas specific for LUAD (TCGA-LUAD) to compare gene expression within the facilitated glucose transporter (GLUT, gene name SLC2A) family. SLC2A1 (GLUT1) followed by SLC2A3 (GLUT3) were the most expressed members, (Figure 1a) and high SLC2A1 expression was correlated with poor overall survival (Figure 1figure supplement 1a). From a human tissue microarray comprising 18 cases of stage IB-IIB LUAD and 18 cases of stage IA-IIB lung squamous cell carcinoma (LUSC), GLUT1 protein was detected in all tumors. Specifically, it showed intermediate and strong GLUT1 expression in 44% (8/18) and 56% (10/18) of the LUAD lesions, respectively ( Figure 1b). In LUSCs, GLUT1 exhibited strong staining in all samples (Figure 1b), confirming previous results obtained from squamous cell carcinomas of lung and other tissues (Goodwin et al., 2017;Hsieh et al., 2019). To obtain more information about SLC2A1 expression in LUAD, we applied a non-negative matrix factorization of TCGA-LUAD samples, which generated four distinct clusters (NMF1-4, In normal lung, Glut1 was only weakly expressed or undetectable in the alveolar compartment that comprises alveolar type-1 and 2 (AT1, AT2) cells, endothelial cells and alveolar macrophages (Figure 1-figure supplement 2b), AT2 cells being considered as the principal tumor cell-of-origin in this model (Desai et al., 2014;Sutherland et al., 2014;Xu et al., 2012). In contrast, Glut1 was strongly expressed in the bronchiolar epithelial compartment that contains Club cells, another cell type permissive to tumor initiation (Sutherland et al., 2014;Figure 1-figure supplement 2b). To determine the contribution of Glut1 for tumor development, we crossed KP mice to Slc2a1 Flox/Flox (G1) animals (Young et al., 2011), and initiated tumors in the resulting KPG1 mice and control KP mice by intratracheal lentiviral-Cre instillation (Lenti.PGK-Cre, Figure 1-figure supplement 2c-d). During tumor progression, we detected slightly but not significantly reduced changes in tumor growth rates monitored by X-rays micro-computed tomography (mCT) (Figure 1d). At sacrifice, although no significant difference was observed in the weight or in the number of lesions (Figure 1e   To detect if there are measurable changes in glucose utilization secondary to Glut1 deletion in vivo, we decided to monitor glucose-derived biomass accumulation in tumors with ultrastructural resolution. Specifically, we performed 13 C-glucose injections followed by nanoscale secondary ion mass spectrometry (NanoSIMS) imaging (Hoppe et al., 2013) correlated with electron microscopy (EM) (Figure 2a). Intense 13 C enrichments inside tumors revealed a compartmentalized intracellular accumulation of glucose-derived biomass, which can be attributed to a specific organelle of tumor cells visualized by the superposition of images obtained from scanning EM (SEM) and NanoSIMS ( Figure 2b). From SEM and transmission EM (TEM), these organelles resemble lamellar bodies (LBs) based on their size, secretory behavior and morphology, which exhibit more or less packed, occasionally visible lamellae (Figure 2b-c). LBs are lipid-rich secretory organelles specifically produced in Figure 1 continued Source data 1. Source files for tumor growth, grades and survival of KP and KPG1 mice. Figure supplement 1. Non-negative matrix factorization generates four LUAD subtypes, with NMF4 being the closest to KP tumors. Figure supplement 2. Glut1 is detectable in tumors and the bronchiolar epithelium but not in the alveolar compartment. 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2µ µ µ µ µ µ µ µ µ µ µ µ µ µ µ µ µ µ the lung by AT2 cells, responsible for surfactant production and release into the alveolar space. Within tumor cells, these LB-like organelles (LBOs) were typically from 200 nm to 2 mm in diameter, irregularly shaped and darkly stained. Often these contained more heavily stained regions and discernible tightly packed lamellae at higher magnifications ( Figure 2c). Quantitative NanoSIMS images of these organelles revealed a lower 13 C-enrichment signal intensity in KPG1 compared to KP tumor cells (Figure 2d), demonstrating the consequence of Glut1 loss in vivo.
Despite the differences measured by NanoSIMS, Glut1 deficiency only minimally affected tumor progression (see Figure 1d-h). We therefore hypothesized the existence of an alternative mechanism that sustains lung tumor growth. First, we isolated KP and KPG1 tumors, prepared single cell suspensions and tested their glycolytic capacity. In this ex vivo experimental setting, KPG1 tumors failed to stimulate glycolysis in response to glucose (Figure 3a and Figure 3-figure supplement 1a). As exception, one KPG1 tumor displayed the strongest glycolytic response of all tested tumors, which coincided with elevated expression of another high-affinity glucose transporter, Glut3 ( Figure 3a). To understand better the mechanisms of Glut1-deficient tumor progression, we also performed RNA sequencing analyses from the non-immune (CD45 -) fraction of KP and KPG1 tumors. To increase the relevance of this comparison, we analysed three cohorts of mice, where tumors were initiated with Lenti.PGK-Cre, Ad5.SPC-Cre or Ad5.CC10-Cre, the latter two providing SPC + -or CC10 + -cell restricted Cre expression for tumor initiation in distinct cell types (Sutherland et al., 2014) (see Figure 1-figure supplement 2c). Gene Set Enrichment Analysis (GSEA) highlighted, from all the Hallmark and KEGG up-and downregulated pathways, a unique significantly shared pathway between the three cohorts, which was upregulated in KPG1: 'Hallmark_bile acid metabolism' (Figure 3-figure supplement 2a). Upon closer examination of the genes included in this pathway, we identified several target genes of peroxisome proliferator-activated receptor alpha (PPARa) (Figure 3b), a crucial transcription factor in fatty acid catabolism (Reddy and Hashimoto, 2001). Furthermore, when interrogating the expression of known PPARa targets (Rakhshandehroo et al., 2010), GSEA highlighted their enrichment in KPG1 compared to KP tumors ( Figure 3c). Thus, KPG1 tumors have a stronger expression of known PPARa-target genes than KP lesions. Together, our metabolic and molecular analyses suggested two possible and distinct routes for Glut1-deficient tumor growth: the uptake of the same nutrient through another glucose transporter, Glut3, or via a PPARa-dependent metabolic shift ( Figure 3d).
To test the hypothesis of an involvement of PPARa in lung tumor progression, we cloned a mouse PPARa-dominant negative construct, PPARaD13 (Michalik et al., 2005), into a bi-promoter Lenti. Cre vector allowing doxycycline-inducible expression (Lenti.TRE-PPARaD13_PGK-Cre) in KP and KPG1 tumors (mice were crossed to CCSP-rtTA transgenic mice [Meylan et al., 2009; Figure  To test our second hypothesis, we initially monitored Glut3 expression in tumors. By real-time PCR, Slc2a3 was expressed to varying levels but not significantly differently between KP and KPG1 tumors (   Although this indicates that Glut3 deletion alone does not affect tumors in a detectable manner, we next wanted to evaluate its importance in a Glut1-deficient context. To do this, we generated KPG1G3 mice by interbreeding KP mice with Slc2a1 Flox/Flox and Slc2a3 Flox/Flox mice (see Figure 1-figure supplement 2c-d). Deletion of the two glucose transporters, Glut1 and Glut3, from the tumor epithelial cell compartment resulted in a decreased number of tumors ( Figure 4b).
Furthermore, contrasting with our data obtained from each separately-targeted Gluts, combined Glut1 and Glut3 deletion led to a significant reduction of tumor growth rates monitored by mCT scans, to a reduced tumor size with an absence of big (>50 mg) lesions at sacrifice, to a decrease of high grade tumors, and to an extended mouse survival ( Thus, in absence of Glut3, lesions might certainly incorporate 18 F-FDG via Glut1. The 18 F-FDG absorption of Glut1-and Glut3-deficient KP tumors was significantly reduced as compared to control KP lesions (Figure 4h-i). Because glucose uptake was not significantly lower in KPG3 compared to KP tumors, and in KPG1G3 compared to KPG1 tumors, a possible explanation is that most tumors analyzed had not yet gained Glut3 expression and become dependent on it, which agrees with an elevated Glut3 expression specifically in big lesions (see Figure 4-figure supplement 1c). Finally, to investigate if Glut1-and Glut3-deficient tumors might incorporate glucose via another glucose transporter, we performed a gene expression analysis of glucose transporter family members on KP and KPG1G3 lesions. This analysis showed that the expression of no other glucose transporter was changed upon dual Glut1 and Glut3 deletion in KP tumors (Figure 4-figure supplement 4a). Altogether, these data demonstrate that deleting Glut1 and Glut3 in KP tumors reduces glucose uptake, and that Glut1 highly contributes to glucose uptake in KP tumors. These results may also lead to the hypothesis that at least a part of the remaining 18 F-FDG-PET signal detected in KPG1G3 tumors might come from glucose uptake by the cells of the tumor immune microenvironment.
Recently, pharmacological inhibition of a secondary active, Na + -dependent glucose transporter, Sglt2, showed its role in premalignant KP lung lesions (Scafoglio et al., 2018), while a glycolytic switch was identified in advanced compared to early tumors (Kerr et al., 2016). These studies, together with our gene targeting approaches, definitively demonstrate the importance of glucose utilization for LUAD development in vivo. In the KL model, Glut1 deletion impairs LUSC development (Hsieh et al., 2019), and we find that Lkb1-deficient lung tumors lack detectable Glut3, precluding any compensation occurring through this transporter. In contrast, in pure LUAD our results    T2_B   T1_B   T3_B   T9_B   T8_A  T7_B   T4_B  T5_B  T6_B   T4_A  T5_A  T2_C   T1_C   T7_B   T11_A  T2_C   T8_B   T10_A   T9_A   T6_C   T5_C   T4_C   T3_C  T3_B   T4_B  T5_B  T6_A   T1_B   T7_A   KP  KPG1  indicate that Trp53-deficient tumor cells depend on glucose uptake based on a redundant system, whereby the expression of either Glut1 or Glut3 is sufficient to sustain tumor growth (Figure 4j). For clinical perspectives, we anticipate that dual Glut1 and Glut3 blockade will be necessary to achieve strong anti-tumor efficacy.

Materials and methods
Key resources (f) Percent of KP (n = 150) and KPG1G3 (n = 170) lesions classified by tumor grades, either detailed from alveolar hyperplasia (AH) to grade 5 or discriminated between alveolar hyperplasia and adenomas, and adenocarcinomas. Alveolar hyperplasia and adenomas include the AH and the tumor grades 1, 2, and 3. Adenocarcinomas contain the tumor grades 4 and 5. *: p < 0.05; **: p < 0.01. Fisher test was applied when comparing AH, grade 1, grade 2, grade 3, grade 4, and grade 5. Chisquare for trend was applied when comparing alveolar hyperplasia and adenomas, and adenocarcinomas.

TCGA data processing
The Cancer Genome Atlas (TCGA) Lung Adenocarcinoma (LUAD) dataset was retrieved from http:// cancergenome.nih.gov. Samples with clinical information and mRNA (RNASEQV2) expression data were selected (511 samples including 448 samples with survival data), and the data were downloaded using the RTCGAtoolbox (v2.12.1). RSEM gene quantifications as provided by TCGA were taken, counts were converted to log2 normalized counts expression and batch effect was removed using voom and removeBatchEffect functions from the limma package (v3.38.3).

Non-negative matrix factorization
Non-negative matrix factorization (NMF) was done with the NMF package (v0.21.0) and standard strategies applied to the matrix in 511 normalized TCGA-LUAD samples. Negative values present in the dataset were zeroed to obtain a non-negative matrix, and the NMF was performed for ranks 2 to 6 using the Brunet algorithm. Two hundred runs were performed per rank with random seeding. For each rank, the cophenetic correlation coefficient and silhouette score were computed from the 200 solutions and the factorization achieving the minimum residual error was kept as the solution for further analyses. The cophenetic correlation coefficients and silhouette scores were directly obtained using the plot function of the NMF package. As a control, the NMF was also performed on a randomized dataset built by taking within each patient a random permutation of the gene expression values, using the randomize function of the NMF package. After having selected rank four as the best one, the NMF was rerun only for rank four and for a superior number of runs (500 runs). The final NMF-based matrix is composed of 935 metagenes and 511 samples split into four subtypes (shown as a heatmap). Correspondence to the three subtypes found by the Cancer Genome Atlas Research Network (The Cancer Genome Atlas Research Network, 2014) (proximal proliferative, proximal inflammatory and terminal respiratory unit) and presence of TP53 mutation are shown on top bars. To identify which human NMF subtype corresponds the most to KP mouse bulk Lenti tumors, we combined mouse and human normalized expression data, removed batch effect using ComBat (package sva v3.30.1), fitted a logistic regression to the human data set using glmnet package (v2.0-18) and 10-fold cross-validation with default parameters, and applied the fitted model as prediction method to the mouse data. Human and mouse orthologous genes were obtained using biomaRt (v2.38.0) and only genes in common between mouse and human were used.

Human survival analysis
Univariate analyses of overall survival (OS) were performed using Cox proportional hazard regression models available in the survival package (v2.44-1.1). Survival curves were computed with the Kaplan-Meier method. p-values were computed with Wald test.

Study approval
All mouse experiments were performed with the permission of the Veterinary Authority of Canton de Vaud, Switzerland (license numbers: VD2391 and VD2663).

Animal models
Kras LSL-G12D/WT (K) (RRID:IMSR_JAX:008179) and Trp53 Flox/Flox mice (P) (RRID:IMSR_JAX:008462) in a C57BL6/J background were purchased from The Jackson Laboratory, and were crossed to obtain Kras LSL-G12D/WT ; Trp53 Flox/Flox (KP) mice. Stk11 Flox/Flox mice in a mixed background (FVB; 129S6) were obtained from R. DePinho (The University of Texas MD Anderson Cancer Center) through the National Cancer Institute mouse repository, were backcrossed seven times to C57BL6/J and were bred with Kras LSL-G12D/WT mice to obtain Kras LSL-G12D/WT ; Stk11 Flox/Flox (KL) mice. The generation of the Slc2a1 Flox/Flox (G1) mice and Slc2a1 genotyping were described previously (Young et al., 2011). Slc2a3 Flox/Flox (G3) mice were produced at the Duke University Transgenic Facility upon blastocyst microinjection of targeted ES cells obtained from KOMP Repository (project CSD48048). The resulting mice were then crossed to Flp transgenic mice to remove the neomycin resistance cassette and generate mice with a Slc2a3 Flox allele, in which exon six is flanked by LoxP sites. Slc2a3 genotyping was performed by PCR (94˚C 5 min, followed by 40 cycles of 94˚C 30 s, 56˚C 30 s and 72˚C 30 s, followed by 72˚C 7 min) using oligos 5'-TAGTGCCCAGGAATGTGAGGTCAG-3' and 5'-GCCCCAC-CAGATTTACCAAAGG-3'. While heterozygous mice were identified this way, in cases with only one detected band, DNA sequencing was further performed to discriminate without ambiguity between wild-type and floxed using 5'-TAGTGCCCAGGAATGTGAGGTCAG-3'. KP mice were bred to G1 or G3 to obtain KPG1, KPG3 or KPG1G3 mice. KL mice were bred to G1 to obtain KLG1 mice. Control mice used throughout the study carry two wild-type (WT) copies of Slc2a1 (or Slc2a3) or are heterozygotes (Slc2a1 WT/Flox ), as we did not notice any difference of tumor development comparing Slc2a1 WT/WT to Slc2a1 WT/Flox conditions. CCSP-rtTA mice were obtained from The Jackson Laboratory and were crossed to KP and KPG1 mice to enable doxycycline-mediated PPARaD13 transgene expression. For transgene induction, mice were fed on a diet of doxycycline-containing pellets (625 mg kg À1 ; 3242 diet + DOX 0.625 g kg À1 ; Kliba Nafag) for 15 days, or on a regular diet as control.

Tumor volume measurements
Longitudinal tumor volume monitoring was performed by X-rays micro-computed tomography (mCT; Quantum FX; PerkinElmer). Mice were anaesthetized using isoflurane (Piramal, 56.761.002) before the scanning procedure. Animals were then maintained under anesthesia during lung imaging set at 50 mm voxel size, with retrospective gating. Individual tumor volumes were measured and calculated either with Analyze 12.0 software (PerkinElmer) or with OsiriX MD software (Pixmeo; RRID:SCR_ 013618).

Mouse survival analysis
Overall mouse survival was evaluated considering actual mouse death or when lung tumor burden compelled to sacrifice. Control KP mice were the same for the three Kaplan-Meier analyses performed.

Assessment of tumor parameters
Tumor grading was assessed from haematoxylin and eosin (H and E) stained lung sections by an experienced veterinary pathologist (CG), based on a previous classification system (Jackson et al., 2005). Briefly, lesions were categorized as alveolar hyperplasia (AH), or as tumor on a 1-5 severity grading scale from grade one adenomas to grade five adenocarcinomas. Glut1 staining (defined as weak, intermediate or strong) was monitored in all grades. Tumor numbers were counted from lung sections. Tumor areas were calculated from lung sections using QuPath (Bankhead et al., 2017;RRID:SCR_018257). To evaluate Glut3 staining, tumors were categorized into small or big lesions based on their diameter on histology sections (small <0.82 mm< big); this diameter was chosen because it was the median diameter of all tumors from a test KP cohort. To determine tumor weights, the biggest visible tumors per mouse at autopsy were macro-dissected and immediately weighed. To determine Glut1-3 correlation, KP tumor-bearing lung serial sections stained for Glut1 or Glut3 were analyzed.

Preparation of lung tumors for electron microscopy (TEM and SEM) and NanoSIMS
Tumor-bearing KP and KPG1 mice at 21 weeks post-tumor initiation were injected intraperitoneally with a solution containing 13 C-glucose (or 12 C-glucose as control), four times every 45 min, beginning 3 hr prior to sacrifice, for a total of 7.5 g kg À1 . Mice were then perfused via the heart with a buffered mix of 1.0% glutaraldehyde and 2.0% paraformaldehyde in 0.1 M phosphate buffer (pH 7.4). After 2 hr, the lungs were removed and placed in the same fixative overnight. They were then sectioned with a vibratome (VT1200, Leica Microsystems) at a thickness of 100 mm, and the sections of interest, showing distinct tumors, washed thoroughly with cacodylate buffer (0.1 M, pH 7.4). These were then postfixed for 40 min in 1.0% osmium tetroxide with 1.5% potassium ferrocyanide in cacodylate buffer, and then 40 min in 1.0% osmium tetroxide alone (in cacodylate buffer). They were finally stained for 30 min in 1% uranyl acetate in H 2 O before being dehydrated through increasing concentrations of EtOH and then embedded in Durcupan ACM (EMS, USA) resin. The resin-embedded sections were hardened for 24 hr in an oven at 65˚C. Regions of interest were sectioned either at 50 nm thickness, and collected on single slot grids, for transmission electron microscopy (TEM), or at 0.5 mm thickness, and collected on silicon wafers, for scanning electron microscopy (SEM) and NanoSIMS analysis. Sections collected on grids were further contrasted with lead citrate and uranyl acetate solutions, and images taken using a transmission electron microscope (Spirit TEM, operating at 80 kV, FEI Company) with digital camera (Eagle CCD camera; FEI Company). SEM imaging of sections on silicon wafers for correlation with NanoSIMS images was performed using a Zeiss Gemini-SEM 500. Imaging was done at 3 kV, using an energy selective backscatter detector (EsB) with a grid bias of 1500 V. NanoSIMS images of cells selected from SEM images were acquired using a Cs + primary beam. Selected areas (usually 50 Â 50 mm for entire cell images or 25 Â 25 mm for specific areas within a cell) were first implanted in order to clean up the surface and reach stable emission conditions. The areas were then imaged by scanning the Cs + -beam focused on the surface of the sample; beam diameter was about 200 mm and current 1.8 pA for the 50 Â 50 mm images and about 100 mm with a current of 0.5 pA for the 25 Â 25 mm images. Number of pixels were always 256 Â 256 and dwell time 5 ms/pixel. For each image, eight layers were acquired and processed using Loo-katNanosims software (Polerecky et al., 2012). The layers were aligned and stacked. Using a SEM image aligned with the NanoSIMS images, areas of interests were drawn for every LBO.

Human microarray and immunohistochemistry
A tissue microarray (TMA) comprising 36 randomly selected human non-small cell lung cancers (18 stage IB-IIB LUAD, 18 stage IA-IIB LUSC based on the 7 th Edition of the UICC TNM classification) was generated by SB using the Translational Research Unit Platform of the Institute of Pathology, University of Bern, as described previously (Zlobec et al., 2013). Tumors were resected between 1990 and 2007. Representative H&E slides from all cases were scanned using a digital slide scanner (Pannoramic P250, Budapest, Hungary). Using a digital TMA annotation tool, each scan was marked using eight cores per tumor of 0.6 mm diameter each. Annotated images were aligned with the tumor block and cored out automatically (TMA Grand Master, 3DHistech, Hungary), producing a next-generation Tissue Microarray (ngTMA). The ngTMA block was cut at 4 mm and a double IHC using anti-GLUT1 (AEC chromogen, red staining) and anti-HIF1a (DAB chromogen, brown staining) was performed. Data on HIF1a were not considered in this study. Anti-GLUT1 (07-1401, Millipore, 1:5000; RRID:AB_1587074) was incubated for 30 min. Antigen retrieval was done in citrate buffer for 30 min at 100˚C. GLUT1 membrane staining on tumor cells was classified by an experienced pathologist (CN) into score 0 (no or weak staining in tumor cells in all eight cores), score 1 (intermediate staining in tumor cells in at least 1 of the eight cores) or score 2 (strong staining in tumor cells in at least 1 of the eight cores). No score 0 was present. Visual scoring was done using Scorenado, an automated core-wise digital image display and documentation tool based on QuPath elements (Lytle et al., 2019).

RNA isolation, reverse transcription and real-time PCR
Single cell suspensions from macro-dissected tumors were obtained using a gentleMACS Octo Dissociator (Miltenyi Biotec). RNA was prepared for sequencing from bulk tumors (bulk Lenti) or from CD45 -(non-immune) tumor fractions (PGK-Cre, SPC-Cre and CC10-Cre). In the latter case, magnetic cell sorting to remove immune cells was performed using anti-CD45 MicroBeads (130-052-301, Miltenyi Biotec). RNA for real-time PCR or sequencing was extracted using TRIzol (15596018, Thermo-Fisher Scientific). For real-time PCR, 1 mg RNA was used for reverse transcription using High-Capacity cDNA Reverse Transcription Kit (4368814, ThermoFisher Scientific). Real-time PCR was done using Taqman universal PCR master mix (4324018, ThermoFisher Scientific) and Taqman  Mouse mRNA sequencing, differential expression, and pathway analyses Multiplexed libraries for mRNA-seq were prepared with the TruSeq stranded mRNA kit (Illumina) starting from 500 ng of good-quality total RNAs (RNA quality scores > 7 on the Fragment Analyzer). Sequencing was subsequently performed on a NextSeq 500 instrument (Illumina) on a high-output flow cell, yielding single-end reads of 85 nucleotides. Adapter sequences and low-quality ends were removed with Cutadapt (v1.9.1), trimming for TrueSeq and polyA sequences. Reads were aligned to mouse genome build mm10 using HISAT2 aligner (v2.0.3beta). Genes with low expression were filtered out (average transcripts per kilobase million < 20). Counts were normalized for library size using TMM method from EdgeR (v3.24.3) and voom from limma. Differential expression was computed with limma between Glut1-KO and WT control, or PPARaD13 and control.
We investigated pathways as defined in the Hallmark and KEGG collections of MSigDB (v6.0; Subramanian et al., 2005). Gene Set Enrichment Analysis (GSEA) was performed on genes ranked based on t-statistics obtained when comparing Glut1-KO and WT control, or PPARaD13 and control. Statistical significance was calculated by permutation tests (number of random permutations = 10 5 ).
In vivo measurements of the tumors' glucose uptake were performed on a group of animals (n = 14, i.e. 3 KP, 4 KPG1, 3 KPG3 and 4 KPG1G3, fasted overnight) with positron emission tomography (PET) as previously described in detail (Lanz et al., 2014). Mice were anesthetized with a mixture of 4% isoflurane (vol/vol) in 100% O 2 for 2 min and then maintained with a mixture of 1.5% isoflurane in 100% O 2 (0.9 L/min) during tail catheter insertion and initial glycaemia measurement. PET data were acquired on an avalanche photodiode-based LabPET-4 small-animal scanner (Gamma Medica, Sherbrook, Canada). Mice were prone positioned on a heat-regulated scanner bed with respiratory cushion and rectal temperature measurement probe, with the PET scanner field of view adjusted to cover the chest area. A bolus injection of 18 F-fluorodeoxyglucose ( 18 F-FDG) (~20 MBq) was administered through the tail vein catheter within the first 30 s of the PET scan, followed by 150-200 ml of saline chase solution. During the 60 min scan, mice were maintained under 1.5-2% (vol/vol) isoflurane anesthesia in oxygen (0.9 L/min). Monitored temperature and breathing rate were maintained within a physiological range. Raw data were reconstructed with a maximum likelihood expectation maximization (MLEM) algorithm on a circular FOV of 60 mm of diameter and a standard voxel size of 0.5Á0.5Á1.18 mm, in three time frames of 20 min. The last time frame (labeling steady-state) was used to derive standardized uptake values (SUV) images and measure maximal SUV (SUV max ) values over tumor regions.

SUV g=ml
where R ROI represents the radioactivity concentration measured in the region of interest (ROI), A inj is the injected activity corrected for the radiotracer decay and w is the mouse weight. PMOD 2.95 software (PMOD Technologies, Zurich) was used for the determination of the standardized uptake value (SUV) and for the delineation of the tumors' ROIs, based on a visual slice-byslice matching with the corresponding mCT images. Tumors appearing on consecutive slices were grouped as single VOIs for the calculation of the SUVmax.

Statistical analyses and graphic design
Statistical analyses were performed, and graphics produced using R version 3.5.1, Prism version eight or Microsoft Excel version 16.

Additional files
Supplementary files . Transparent reporting form

Data availability
The RNA sequencing data have been deposited to the GEO database (https://www.ncbi.nlm.nih. gov/geo/) and assigned the identifier: GSE138757.
The following dataset was generated: