In vivo Deuterium Magnetic Resonance Spectroscopy reveals Glucose fluxes through glycolysis and mitochondrial oxidation pathways in mouse Glioblastoma

Objectives: Glioblastoma multiforme (GBM), the most aggressive glial brain tumors, can metabolize glucose through glycolysis and mitochondrial oxidation pathways. While specific dependencies on those pathways are increasingly associated with treatment response, detecting 5 such GBM subtypes in vivo remains elusive. Here, we develop a dynamic glucose-enhanced deuterium spectroscopy (DGE 2H-MRS) approach for differentially assessing glucose turnover rates through glycolysis and mitochondrial oxidation in mouse GBM and explore their association with histologic features of the tumor and its microenvironment. Materials and Methods: GL261 and CT2A glioma allografts were induced in immunocompetent 10 mice and scanned in vivo at 9.4 Tesla, harnessing DGE 2H-MRS with volume selection and Marchenko-Pastur PCA (MP-PCA) denoising to achieve high temporal resolution. Each tumor was also classified by histopathologic analysis and assessed for cell proliferation (Ki67 immunostaining), while the respective cell lines underwent in situ extracellular flux analysis to assess mitochondrial function. 15 Results: MP-PCA denoising of in vivo DGE 2H-MRS data significantly improved the time-course detection (2-fold increased Signal-to-Noise Ratio) and fitting precision (-19±1% Cramér-Rao Lower Bounds) of 2H-labelled glucose, and glucose-derived glutamate-glutamine (Glx) and lactate pools in GL261 and CT2A orthotopic tumors. Kinetic modeling further indicated inter-tumor heterogeneity of glucose consumption rate for glycolysis and oxidation during a defined epoch of 20 active proliferation in both cohorts (19±1 days post-induction), with consistent volumes (38.3±3.4 mm3) and perfusion properties prior to marked necrosis. Histopathologic analysis of these tumors revealed clear differences in tumor heterogeneity between the two GBM models, aligned with metabolic differences of the respective cell lines monitored in situ. Importantly, glucose oxidation (i.e. Glx synthesis and elimination rates: 0.40±0.08 and 0.12±0.03 mM.min-1, respectively) strongly correlated with cell proliferation across the pooled cohorts (R=0.82, p=0.001; and R=0.80, p=0.002, respectively), regardless of tumor morphologic features or in situ metabolic characteristics of each GBM model. 5 Conclusions: Our fast DGE 2H-MRS enables the quantification of glucose consumption rates through glycolysis and mitochondrial oxidation in mouse GBM, which is relevant for assessing their modulation in vivo according to tumor microenvironment features such as cell proliferation. This novel application augurs well for non-invasive metabolic characterization of glioma or other cancers with mitochondrial oxidation dependencies.


Introduction
Cancer is a highly heterogeneous disease, exhibiting multiple phenotypes that impose major challenges for clinical diagnosis and treatment, including assessment of treatment efficacy. Tumor heterogeneity can be characterized by many different aspects, such as cell proliferation, invasion, 5 and hypoxia. However, recent observations suggest that a crucial feature of tumor heterogeneity involves the underlying metabolism (Cantor and Sabatini, 2012). While aerobic glycolysis is a well-established hallmark of cancer metabolism (Warburg effect (Warburg, 1956)), oxidation of glucose through the mitochondrial tricarboxylic acid cycle (TCA) pathway is becoming increasingly associated with microenvironment adaptation and tumor progression (Faubert et al.,10 2020). Such metabolic heterogeneity is observed in glioblastoma multiforme (GBM), the most aggressive form of glial brain tumors (grade IV) (Wen and Kesari, 2008). Specifically, the tumor's proclivity to metabolize glucose through glycolysis and mitochondrial oxidation  is now being associated with pathway-specific dependencies for different GBM subtypes and their respective vulnerabilities to targeted treatments. Namely, a so-called mitochondrial GBM 15 subtype, with cellular bioenergetics relying exclusively on oxidative phosphorylation (OXPHOS), has demonstrated the highest sensitivity to OXPHOS inhibition and most favorable clinical outcome (Garofano et al., 2021). Moreover, different metabolic "rewiring" of glycolysis and oxidative metabolic pathways were recently reported in subpopulations of human GBM cells (U87MG) as reflected by their temozolomide resistance, the current gold-standard for adjuvant 20 chemotherapy in GBM (Immanuel et al., 2021). Accordingly, targeting mitochondrial metabolism for cancer therapy (Weinberg and Chandel, 2015) is currently being pursued with new treatment modalities in GBM (Molina et al., 2018;Shi et al., 2019).
Measuring metabolic fluxes in glycolysis and mitochondrial oxidation pathways simultaneously could thus harbinger metabolic phenotyping and early assessment of treatment efficacy. However, such characterization remains elusive due to a paucity of (invasive and noninvasive) methods with sufficient sensitivity and specificity towards properties associated with 5 such metabolic fluxes. For example, non-invasive molecular methods such as 18 FDG-PET can only detect tumor glucose uptake, but not its metabolic turnover. In vitro analysis of biopsy samples collected after in vivo administration of glucose tracers (e.g., 13 C-labelled) represents an invasive approach for detecting glycolytic and oxidative metabolism of glucose but not their respective in vivo fluxes, in clinical and preclinical GBM tumors Marin-Valencia et al., 10 2012). 1 H magnetic resonance spectroscopy ( 1 H-MRS) provides noninvasive information for tumor metabolic profiling, making it possible to distinguish relevant GBM subtypes such as IHDmut (Choi et al., 2012). However, it is unlikely that it could be used directly to assess metabolic turnover rates due to the crowded spectral areas. 13 C magnetic resonance spectroscopy ( 13 C-MRS) can overcome this limitation to some extent by detecting the non-toxic carbon isotope in labeled 15 glucose. Indeed, 13 C-MRS with infused 13 C-glucose tracers has been proposed for measuring glucose fluxes through glycolysis and mitochondrial oxidation in vivo in gliomas; however, these advanced methods have mostly been limited to very large voxel sizes, with significant non-tumor tissue contributions, as shown in human GBM (Wijnen et al., 2010) and patient-derived GBM xenografts (Lai et al., 2018). Furthermore, the temporal dynamic range of 13 C-MRS is inherently 20 very low due to the long carbon longitudinal relaxation constants (Lai et al., 2018).
Deuterium magnetic resonance spectroscopy ( 2 H-MRS) is a highly promising recently developed MRS modality based on intra-venous injection of non-toxic 2 H-labelled substrates, including 2 H-glucose. Compared to its 1 H-MRS counterpart, 2 H-MRS benefits from: (i) no detectable metabolic background signals, facilitating specificity and sensitivity (the natural abundance of deuterium is very low, and typically only trace amounts are observed from the ~110M proton signal in pure water); (ii) short metabolite longitudinal relaxation times, rapid sampling of the signals; and (iii) an internal reference for quantification that does not require presaturation -naturally abundant semi-heavy water, DHO (Lu et al., 2017). Recent studies have 5 demonstrated the potential of 2 H-MRS coupled with 2 H-glucose injection for tumor metabolic imaging, proving a non-invasive quantitative assessment of the Warburg effect in clinical and preclinical GBM (De Feyter et al., 2018) and early therapeutic response monitoring in mouse subcutaneous tumors (Kreis et al., 2020), also reported with 2 H-fumarate injection (Hesse et al., 2021). Importantly, Dynamic Glucose-Enhanced (DGE) 2 H-MRS has been used to measure 10 glucose consumption rates in the normal rat brain, linked to mitochondrial oxidation (Lu et al., 2017), and glycolysis rates in a mouse lymphoma model (Kreis et al., 2020). Therefore, we hypothesized the suitability of DGE 2 H-MRS for detecting both pathway fluxes simultaneously in GBM tumors.
Here, we selected two well-established syngeneic models of GBM -GL261 and CT2A - 15 recapitulating histologic, genetic, and immunogenic features of the disease (Martinez-Murillo and Martinez, 2007;Oh et al., 2014;Seligman and Shear, 1939;Seyfried et al., 1992;Zagzag et al., 2000), and harnessed a novel application of DGE 2 H-MRS with volume selection and spectral denoising based on Marchenko-Pastur PCA (MP-PCA) (Veraart et al., 2016) to: (i) demonstrate its ability to measure glucose consumption rates through glycolysis and mitochondrial oxidation 20 in mouse GBM in vivo; and (ii) explore potential modulations of glucose metabolism according to GBM microenvironment features, such as cell proliferation (Anderson and Simon, 2020;Charles et al., 2012).

Materials and Methods
As summarized in Fig. 1, this study included in vivo and post-mortem assessment of mouse GBM tumors, namely the GL261 and CT2A models, which were complemented by in situ analysis of these cell lines.  2.2 Mouse model of GBM 5 Tumors were induced in 12 mice (9 males and 3 females), as previously described (Simoes et al., 2008). Briefly, intracranial stereotactic injection of 1 x10 5 GL261 or CT2A cells was performed in the caudate nucleus; analgesia (Meloxicam 1.0 mg/Kg s.c.) was administered 30 min before the procedure. Mice were anesthetized with isoflurane (1.5-2.0% in air) and immobilized on a stereotactic holder (Kopf Instruments, Tujunga/CA, USA) where they were warmed on a 10 heating pad at 37 ºC, monitoring the body temperature with a rectal probe (WPI ATC-2000, Hitchin, UK). The head was shaved with a small trimmer, cleaned with Betadine, and the skull exposed through an anterior-posterior incision in the midline with a scalpel. A 1 mm hole was drilled in the skull using a micro-driller, 0.1 mm posterior to the bregma and 2.32 mm lateral to the midline. The tumor cells (1x10 5 in 4 μL PBS) were inoculated 2.35 mm below the cortical 15 surface using a 10 µL Hamilton syringe (Hamilton, Reno NV, USA) connected to an automatic push-pull microinjector (WPI Smartouch TM , Sarasota FL, USA), by advancing the 26G needle 3.85 mm from the surface of the skull (~1mm skull-to-brain surface distance), pulling it back 0.5 mm, and injecting at 2 μL/min rate. The syringe was gently removed 2 min after the injection had finished, the skin sutured with surgical thread (5/0 braided silk, Ethicon, San Lorenzo Puerto Rico) 20 and wiped with Betadine. The animals were kept at 25 ºC during recovery from anesthesia, and

Animals and cell line
given an opioid analgesic (Buprenorphine 0.05 mg/Kg s.c.) before returning to their cage.
Meloxicam analgesia was repeatedly administrated at 24-and 48-hours post-surgery.  , 1997-2018). For each animal, the tumor region was manually delineated on each slice, and the sum of the areas multiplied by the slice thickness to estimate the estimate the volume, which was averaged across the two 5 orientations acquired (coronal and axial). In addition, the pixel intensities from each slice were normalized to a reference region (ROI in cortex, from the contra-lateral hemisphere), and the pixel distributions for each tumor analyzed for skewness, kurtosis, and inter-quartile range (IQR).
The time-course changes of 2 H-labelled metabolite (Glc, Glx and Lac) concentrations were fitted using a modified version of the kinetic model reported by Kreis et al (Kreis et al., 2020), to 10 estimate the maximum rate of Glc consumption (total, V max ) for Glx synthesis (mitochondrial oxidation, V glx ) and Lac synthesis (glycolysis, V lac ), and the confidence intervals for all estimated parameters: (Eq 2) = + The coupled differential equations describing the concentration kinetics of each metabolite were: where: k g , apparent rate constant of glucose transfer between blood and tumor (min −1 ); k glx , apparent rate constant of Glx elimination (min −1 ); k lac , apparent rate constant of lactate elimination 20 (min −1 ); Glc concentration in plasma (mM); a 1 , the Glc concentration after the bolus injection (mM); k p , the effective rate constant of labeled glucose transfer to tissue (min −1 ); f, the fraction of deuterium enrichment; v, the extravascular-extracellular volume fraction; and k m , the constant for glucose uptake. As originally reported (Kreis et al., 2020), the fraction of deuterium enrichment (f) and constant for glucose uptake (k m ) were fixed: the former, using the 5 same estimation (f = 0.6, based on NMR of blood samples) since the injection protocol and dose/weight used were the same and also in mice; the latter, approximated to k m = 10 mM (Marin-Hernandez et al., 2011;Williams et al., 2012). All parameters were fitted without any restrictions to their range.

DCE T1-weighted 1 H-MRI
10 DCE T1-w MRI data were processed with DCE@urLab (Ortuno et al., 2013). ROIs were manually delineated for each slice and the time-course data was fitted with the Extended Tofts 2compartment model (Tofts, 1997), to derive the volume transfer constant between plasma and tumor extravascular-extracellular space (K trans ), the washout rate between extravascularextracellular space and plasma (k ep ), and the extravascular-extracellular volume fraction (v). The 15 measurements were averaged across 3 slices for each tumor, covering the whole lesion, and were consistent with the literature for mouse brain tumors, e.g. K trans (Boult et al., 2017).

Metabolic assessment in situ
GL261 and CT2A cells were seeded overnight on Seahorse XFp miniplates (1x10 4 cells/well). 20 After attachment/growth for 18 h, the RPMI compl medium in each well was changed to unbuffered medium (103575-100, Agilent, Santa Clara CA, USA) with the same glutamine and glucose concentration as in RPMI compl and incubated for 1 h in a CO 2 -free atmosphere. The cells were then studied in situ using a Seahorse XF HS Mini Analyzer (Agilent, Santa Clara CA, USA) using the

MitoStress Test Kit. The MitoStress Test Kit measures the extracellular acidification rate (ECAR)
and the oxygen consumption rate (OCR) in each well while specific inhibitors of the respiratory chain are sequentially injected. The data were processed with Seahorse Analytics (Agilent, Santa Clara CA, USA) and normalized to the cell number for each well. The latter were determined at 5 the end of the experiment from the cell lysates (RIPA buffer, prepared in-house) using the bicinchoninic acid protein assay (Pierce TM BCA Protein kit, ThermoFisher Scientific, Rockford IL, USA) and a microplate reader (BioTek ELMX800, Cole-Parmer, Winooski VT, USA), and assuming a total protein content 2500 µg/1 x10 7 cells, as before (Simoes et al., 2015). regions on each slide were manually defined with ROIs, followed by semi-automated counting of Ki67 +/cells to determine the total cell density and the labeling index (% Ki67 + cells). Finally, the total cell number for each tumor was estimated based on the total tumor volume (T2-w MRI data), the average cell count per surface area (histologic counting) and assuming a cell radius of 10 µm (as reported in mouse GL261 tumors (Roberts et al., 2020)). 5 Data were analyzed using the two-tailed Student's t-test. Differences at the 95% confidence level (p=0.05) were considered statistically significant. Correlation analyses were carried out with the Pearson R coefficient, unless indicated otherwise.

MRI assessment of GBM allograft tumors
Orthotopic GL261 and CT2A tumors were studied in vivo 19±1 days post-injection, during a well-described epoch of active cell proliferation before marked necrosis (Cha et al., 2003;5 Martinez-Murillo and Martinez, 2007;Zagzag et al., 2000). Volumetric (T2-w) and DCE T1-w 1 H-MRI of the entire cohort showed consistent tumor sizes (38.3±3.4 mm 3 ) and perfusion properties, respectively (Supplementary Table 1 and Fig. 3), in agreement with previous studies (Cha et al., 2003). Additional analysis revealed similar heterogeneity based on T2-w MRI contrast in both GBM models (skewness, kurtosis and inter-quartile range assessment of pixel distributions) 10 but significantly higher magnetic field homogeneity achieved in CT2A tumors (-

Quantification of glycolytic and oxidative consumption of glucose in GBM tumors in vivo
We then performed localized DGE 2 H-MRS ( Fig. 2-A) to monitor the deuterium-labelled glucose (Glc) dynamics; namely, conversion to its downstream products lactate (Lac) and glutamate-glutamine pool (Glx) (Fig. 2-B). The location of the voxel (shown in Fig. 2

-A for
representative GL261 and CT2A glioma-bearing mice) was chosen such that it encompassed as 20 much tumor volume as possible while avoiding peritumoral regions. The quality of the original spectra ( Fig. 2-C upper panel) reveals good spectral resolution and the ability to distinguish between the different 2 H-labelled metabolite pools. The signal-to-noise ratio for the natural abundance DHO peak (SNRi) in the basal original spectra was 20.0±0.9. To improve the detection limits and kinetic profiling, we harnessed a MP-PCA denoising strategy which clearly improved the spectral quality ( Fig. 2-C, lower panel). Specifically, MP-PCA indeed improved the SNRi to 43.9±3.7, demonstrating a 2-fold gain (p<0.001 in each tumor cohort) ( Fig. 2-D); and an average 19±1% improvement in spectral fitting precision ( Fig. 2-E) was noted. Specifically, MP-PCA 5 denoising consistently improved the time-course detection of Glx concentration changes without altering the kinetics (Fig. 2-F Adaptation of a previous kinetic model ( Kreis et al., 2020) enabled the estimation of maximum glucose consumption rates (V max ) in GL261 tumors in vivo, for synthesis of Lac -glycolysis (V lac ) -and Glx -mitochondrial oxidation (V glx ) ( Fig. 3-A). This was initially tested in parallel with two 10 variants of the model, including the extracellular volume fraction (v) either as a variable (v estimated) or fixed according to its estimation from the respective DCE T1-w MRI experiment (v fixed), using the Extended Tofts Model (Supplementary Figure 3). While most samples (e.g. G4) consistently rendered comparable V glx and V lac estimates taking v as either a model-free or modelfixed parameter, other samples (e.g. G6) did not ( Fig. 3-B). Specifically, v estimated and v fixed 15 models mostly converged to very close estimates of V lac and V glx (respectively fold-changes: 1.2±0.1 (p=0.140) and 1.1±0.1 (p=0.151), for GL261 (G1, G2, G3, G4, G5 and G7); and 1.2±0.1 (p=0.373) and 0.9±0.0 (p=0.374), for CT2A (C1,C2,C3,C4)). However, this was not verified in samples G6 and C5: while the v fixed model also converged to V lac and V glx estimates within the expected physiologic ranges, the v estimated model drifted to very large V lac and v estimates 20 without biological meaning, as demonstrated for the G6 sample ( Fig. 3-B). Therefore, the v fixed approached enabled the "stabilization" of the model for all the samples in each cohort. The precision of V glx and V lac estimations with the v fixed model was also demonstrated ( Fig. 3-C). for measuring the maximum rate of glucose consumption (V max ) for synthesis of lactate (V lac ) and glutamate-glutamine

Functional assessment of mitochondrial respiration and glycolysis in glioma cell lines in situ
To investigate whether our in vivo findings were related to basal metabolic properties of each tumor model, additional in situ analysis were performed for functional metabolic assessment of the respective cell lines. Specifically, we investigated glycolysis and mitochondrial oxidation metabolism at the cellular level during active proliferation in 2D cell culture (Fig. 5-A). The results clearly indicated simultaneous mitochondrial OXPHOS (Fig. 5-B) and glycolytic metabolism ( Fig. 5-C

Histopathology of GBM tumors
10 Given that our in situ findings could suggest differential metabolic adaptability of GL261 and CT2A cells to microenvironmental changes during GBM proliferation, potentially consistent with the heterogeneity of glucose metabolic fluxes detected in vivo (Fig. 4), histopathological analysis was performed. Tumors were screened for cytomorphological features, cell density, presence of hemorrhage, necrosis, and for cell proliferation index, blinded to in vivo MRI/MRS data (Fig. 6). 15 While no significant difference was seen in extent of necrosis between GL261 and CT2A tumors, only the latter displayed extensive morphologic heterogeneity ( Fig. 6- Fig. 4). Further assessment of cell 25 proliferation index based on ki67 immunostaining (Fig. 6-B) indicated a strong correlation of this parameter with the stromal-vascular fraction scores in GL261 tumors (Kendall's rank coefficient: τ=0.80, p=0.017), which was not observed in CT2A tumors ( Fig. 6-C). The latter displayed 2-fold higher cellularity than GL261 tumors ( Fig. 6-C

Modulation of glucose metabolism according to cell proliferation in GBM tumors
Finally, we explored potential associations between glucose metabolic heterogeneity and tumor microenvironment features in GBM allografts (Fig 7-A). Thus, glucose consumption rate correlated significantly with cell proliferation index across the pooled cohorts (n=12) and 10 regardless of the histopathologic phenotype (R=0.71, p=0.010 - Fig. 7-A). Importantly, this was associated specifically with glucose mitochondrial oxidation (i.e. Glx synthesis and elimination rates: R=0.82, p=0.001; and R=0.80, p=0.002, respectively - Fig. 7-B and 7-C) rather than glycolysis ( Fig. 7-D and 7-E), which was already detectable in the more heterogeneous GL261 model (R=0.84, p=0.018; and R=0.84, p=0.019, respectively -Supplementary Table 1). 15 Moreover, no association was observed between cell proliferation and other tumor parameters, such as volume, perfusion, heterogeneity based on T2w-MRI contrast, histopathologic phenotype, or cell density (Supplementary Table 1).

Fig. 7. Association between glucose metabolism and cell proliferation in pooled GL261 and CT2A tumor
cohorts. Cell proliferation index correlated significantly with glucose consumption rate (A), which was associated with its mitochondrial oxidation, i.e. rates of Glx synthesis (B) and elimination (C), rather than glycolytic turnover, 5 i.e. rates of Lac synthesis (D) or elimination (E). Plots: estimate±CI for glucose metabolism-derived metrics, and means±SE for cell proliferation index.

Discussion
The ability of DGE 2 H-MRS to quantify glucose metabolic rates through mitochondrial oxidation (normal rat brain (Lu et al., 2017)) or glycolysis (subcutaneous mouse tumors (Kreis et al., 2020)), has been rapidly gaining interest for in vivo metabolic flux assessment. Despite the 5 great promise for in vivo DGE 2 H-MRS, the typically low signal-to-noise ratio in these experiments, which is mainly incurred due to deuterium's low resonance frequency (compared with its 1 H counterpart), imposes significant boundaries on their ability to characterize the relevant metabolic rates and/or accurately define relatively weak signals, such as Glx. Thus, simultaneous characterization of both glycolytic and oxidative pathways remains limited. This work aimed to 10 harness the DGE 2 H-MRS methodology with volume selection and MP-PCA spectral denoising to overcome these major limitations, and investigate whether such simultaneous measurements could address the current need to detect specific metabolic pathway-dependencies in GBM tumors (Garofano et al., 2021). Furthermore, we investigated the potential relevance of DGE 2 H-MRS for characterizing the links between metabolic heterogeneity and histopathologic features, including 15 cell proliferation.
To test these features, we used two robust, immunocompetent mouse models mimicking clinical GBM -GL261 and CT2A (Griguer et al., 2005;Martinez-Murillo and Martinez, 2007;Oh et al., 2014;Seligman and Shear, 1939;Seyfried et al., 1992;Zagzag et al., 2000) -which evidenced pronounced histologic and metabolic heterogeneity and thereby were suitable for testing 20 our hypotheses. Our three-pronged strategy was designed to: (i) measure the in vivo glucose for cancer cell survival, adaptation, and proliferation in a rapidly shifting microenvironment (Fendt et al., 2020;Gillies et al., 2012;Lehuede et al., 2016), with a pivotal role for mitochondrial reprogramming between e.g. invasive and proliferative states (the latter supported by oxidative metabolism ), which are present in GBM (Rajapakse et al., 2020;Xie et al., 2014).
Altogether, our observations strengthen the relevance of DGE 2 H-MRS for in vivo detection of GBM dependencies on oxidative metabolism at any given progression stage. This could be helpful for early treatment assessment, by evaluating the response to: antiangiogenic therapies, which impact tumor perfusion (oxygenation) and therefore the ability to rely on oxidative metabolism 5 (Batchelor et al., 2014); OXPHOS-targeted treatments (Molina et al., 2018;Shi et al., 2019); or even the efficiency of chemosensitization to those or other therapies, e.g. with dichloroacetate (Michelakis et al., 2010;Shen et al., 2015).
Compared to the GL261 model, CT2A cells demonstrated more limited metabolic flexibility in situ. Namely, CT2A cells displayed markedly reduced respiration buffer capacity and no 10 apparent glycolytic response to acute inhibition of OXPHOS. This was consistent with a conserved, less aggressive stromal-vascular phenotype (score I) of CT2A tumors, and 2-fold higher cellular density than GL261's. Despite such metabolic and histopathologic differences between the two allograft models, pooling them strengthen our previous finding that non-necrotic GBM tumors with consistent sizes and perfusion properties had an increasing reliance on glucose 15 mitochondrial oxidation according to cell proliferation index. This is aligned with recent observations of increasing OXPHOS-dependance at more advanced stages of tumor progression (Faubert et al., 2020), and even polarization of tumor-associated macrophages towards an OXPHOS-dependent, pro-tumorigenic M2 phenotype (Van den Bossche et al., 2017). Thus, our results demonstrate the potential of DGE 2 H-MRS for non-invasive detection of clinically relevant 20 GBM phenotypes.
As in every study, we acknowledge the limitations of this work. Firstly, although the Glx peak region was assigned to the glutamate-glutamine pool, this region overlaps with other TCA-cycle intermediates/derivatives such as succinate (2.39 ppm), which has been detected e.g. in breast cancer cell lines with TCA-cycle truncations (Simoes et al., 2015). While no such truncations have been reported in the GL261 model, and are not expected according to the in situ Seahorse experiments performed, the CT2A model could potentially harbor them given the electron transport chain abnormalities reported (Kiebish et al., 2008). In any case, the Glx region assigned in DGE 2 H-MRS should still reflect de novo mitochondrial turnover of glucose. Secondly, the 5 normal variations in tumor shape within each cohort led to some discrepancies in the DGE 2H-MRS voxel vs. total tumor volumes. Specifically, tumors G4 and C4 -the smallest in their respective cohorts -presented more regular shapes that mostly fitted the total voxel volume. While this led to ~20% larger voxel volume than total volume, the same tumor/non-tumor proportion was kept during visual adjustment of the voxel in these tumors (as in all others); otherwise, lower rates 10 of glycolysis vs glucose oxidation would be expected in these samples due to stronger contaminations from non-tumor tissue, which was not the case (Supplementary Table 1, Vlac/Vglx: G4>G2~G1>G7~G5; C4~C1>C2). Moreover, the differences in voxel vs tumor volumes did not correlate significantly to glucose oxidation or glycolytic rates within each cohort or across pooled cohorts (Supplementary Table 4), further suggesting no association between 15 those variables, we performed the study at 9.4 Tesla, which represents a higher magnetic field strength compared to standard clinical scanners (1.5 to 3.0 Tesla). However, since deuterium spectroscopy and imaging (De Feyter et al., 2018) benefit significantly from increased field strengths, their translational application to the human brain has already been demonstrated at 7.0 Tesla (de Graaf et al., 2020) and could be extended to 9.4 Tesla human scanners.

Conclusion
This study demonstrated (i) the potential of DGE 2 H-MRS for in vivo assessment of glucose consumption rates through glycolytic and oxidative pathways simultaneously in mouse GBM, and potentially other aggressive tumors with OXPHOS dependencies (Mendez-Lucas et al., 2020); and 5 (ii) its relevance for metabolic characterization of GBM, as evidenced here by the strong association between the heterogeneity of glucose consumption rates and cell proliferation in two well-established allograft models prior to marked necrosis. The relevance of this novel application for non-invasive stratification and early assessment of treatment efficacy in GBM patients (Garofano et al., 2021) warrants its extension to additional tumor models and progression stages,  bottom, signal-to-noise ratio of the basal DHO peak (SNR DHOi ) before and after denoising. E Spectral fitting: top, time-domain fitting of MP-PCA spectra for absolute quantification (topcomponents; bottom -black, original; red, estimate; magenta, residuals), and kinetic fitting for Glc (red), Glx (green) and Lac (blue); bottom, Cramer-Rao lower bounds of the spectral fits in the 30 original data and after MP-PCA denoising (Glc, first 10 spectra after bolus injection; Glx and Lac, last 10 spectra). F Time-course profiles of metabolic concentrations in the original data and after MP-PCA denoising, which improved Glx detection. DHO, semi-deuterated water; Glc, 6,6'-2 H 2glucose; Glx, 4,4'-2 H-glutamate-glutamine; Lac, 3,3'-2 H-lactate; TCA, tricarboxylic acid. Plots: mean±SE, n=7 GL261 and n=5 CT2A. * p<0.05, ** p<0.01, *** p<0.001.  15 0.437, respectively). a 1 , Glc concentration after the bolus injection (mM); k p , effective rate constant of labeled glucose transfer to tissue (min −1 ); k g , apparent rate constant of glucose transfer between blood and tumor (min −1 ); V glx , maximum rate of Glc consumption for Glx synthesis (mM.min −1 ); k glx , apparent rate constant of Glx elimination (min −1 ); V lac , maximum rate of Glc consumption for Lac synthesis (mM.min −1 ); k lac , apparent rate constant of lactate elimination 20 (min −1 ); V max , maximum rate of total Glc consumption (mM.min −1 ).

Supplementary Figure 6. Effect of averaging DGE 2 H-MRS data for MP-PCA denoising.
Tumor G4 data displayed as an example. A Resampling the data from 8 to 128 averages (NA) did 20 not impact MP-PCA denoising performance with regards to SNR improvement, but impacted metabolic kinetic profiling and modulation (e.g. Lac at NA=8 vs. 64). B Simulated and real data (ground truth). Three approaches tested with simulated spectra: (first-from-left) adding different noise levels (NL) to the spectra; (second-from-left) changing the total number of spectra in the time-course (equivalent to number of repetitions, NR); or (third-from-left) reproducing the 25 averaging approach used for the real data (forth-from-left) by adapting NL according to NR to reproduce the SNR profile of the original data, in this case from NA= 4 to 256. Best performance highlighted (NA= 64, red arrow). -

-Marchenko-Pastur PCA denoising of 2 H-MRS spectra improved kinetic quantification -Metabolic kinetics revealed differential glucose pathway fluxes in non-necrotic tumors
-Modulation of glucose metabolism reflected tumor heterogeneity (proliferation) 15 Credit Author Statement: