“Aerobic glycolytic imaging” of human gliomas using combined pH-, oxygen-, and perfusion-weighted magnetic resonance imaging

Highlights

• •Aerobic glycolytic imaging combines pH-, oxygen-, and perfusion-weighted MRI.
• •Aerobic glycolytic imaging depicts abnormal glucose metabolism of gliomas.
• •IDH wild-type gliomas show higher aerobic glycolytic index compared with mutants.
• •Aerobic glycolytic index in IDH wild-type glioma is correlated with glucose uptake.
• •Aerobic glycolytic index in IDH mutant glioma is correlated to lactate transporters.

Abstract

Purpose

To quantify abnormal metabolism of diffuse gliomas using “aerobic glycolytic imaging” and investigate its biological correlation.

Methods

All subjects underwent a pH-weighted amine chemical exchange saturation transfer spin-and-gradient-echo echoplanar imaging (CEST-SAGE-EPI) and dynamic susceptibility contrast perfusion MRI. Relative oxygen extraction fraction (rOEF) was estimated as the ratio of reversible transverse relaxation rate R₂′ to normalized relative cerebral blood volume. An aerobic glycolytic index (AGI) was derived by the ratio of pH-weighted image contrast (MTR_asym at 3.0 ppm) to rOEF. AGI was compared between different tumor types (N = 51, 30 IDH mutant and 21 IDH wild type). Metabolic MR parameters were correlated with ¹⁸F-FDG uptake (N = 8, IDH wild-type glioblastoma), expression of key glycolytic proteins using immunohistochemistry (N = 38 samples, 21 from IDH mutant and 17 from IDH wild type), and bioenergetics analysis on purified tumor cells (N = 7, IDH wild-type high grade).

Results

AGI was significantly lower in IDH mutant than wild-type gliomas (0.48 ± 0.48 vs. 0.70 ± 0.48; P = 0.03). AGI was strongly correlated with ¹⁸F-FDG uptake both in non-enhancing tumor (Spearman, ρ = 0.81; P = 0.01) and enhancing tumor (ρ = 0.81; P = 0.01). AGI was significantly correlated with glucose transporter 3 (ρ = 0.71; P = 0.004) and hexokinase 2 (ρ = 0.73; P = 0.003) in IDH wild-type glioma, and monocarboxylate transporter 1 (ρ = 0.59; P = 0.009) in IDH mutant glioma. Additionally, a significant correlation was found between AGI derived from bioenergetics analysis and that estimated from MRI (ρ = 0.79; P = 0.04).

Conclusion

AGI derived from molecular MRI was correlated with glucose uptake (¹⁸F-FDG and glucose transporter 3/hexokinase 2) and cellular AGI in IDH wild-type gliomas, whereas AGI in IDH mutant gliomas appeared associated with monocarboxylate transporter density.

1. Introduction

Metabolic reprogramming is a hallmark of cancer. Glycolysis is often augmented in malignant cells, even in the presence of abundant oxygen (Venneti and Thompson, 2017), in a process known as aerobic glycolysis or the ‘Warburg effect.’ This abnormal metabolism results in a substantial decrease in local extracellular pH due to increased production of lactic acid and other factors (Gatenby and Gillies, 2004). Increased extracellular acidity potentially contributes to tumor malignancy by inducing mutagenesis, invasion, angiogenesis, immunosuppression, chemoresistance, and promotion of glioma stem cells (Corbet and Feron, 2017, Morita et al., 1992). This phenomenon is in part mediated by hypoxic induction of hypoxia-inducible factor 1-alpha (HIF1α) (Brat et al., 2004), leading to overexpression of glycolysis-related proteins/enzymes, namely, glucose transporter, hexokinase, monocarboxylate transporter, and lactate dehydrogenase, thereby accelerating glycolysis (Fig. 1) (Venneti and Thompson, 2017). These prinicples in general apply to malignant gliomas (Venneti and Thompson, 2017). However, given inter-tumoral heterogeneity in glioma, these molecular processes vary. For example, in IDH mutant glioma (Koivunen et al., 2012), HIF1α response to hypoxia is blunted, resulting in lower expression of HIF1α related genes and shift of the glucose metabolism to oxidative phosphorylation (Chesnelong et al., 2014). Nevertheless, metabolic perterbations remain an important area of study in the biology of IDH mutant and wild-type gliomas.

Fig. 1.

Schematic overview of metabolic pathway in IDH wild-type and mutant gliomas. Similar to other malignant tumors, IDH wild-type glioma is characterized by a high level of aerobic glycolysis (Warburg effect). Hypoxia induces upregulation of HIF1α, leading to overexpression of glycolysis-related proteins/enzymes, namely, GLUT, HK2, MCT, and LDHA, thereby accelerating glycolysis. On the other hand, HIF1α response to hypoxia is blunted in IDH mutant glioma, shifting the metabolism to oxidative phosphorylation. In this study, the extent of aerobic glycolysis was estimated by pH- and oxygen-sensitive imaging combined with perfusion imaging. The proteins/enzymes investigated in our study are labeled in black. Notably, MTR_asym at 3.0 ppm increases not only in accordance with decrease in pH but also with increase in glutamine related to tumor activity. 2HG = 2-hydroxyglutarate, aKG = alpha-ketoglutarate, acetyl-CoA = acetyl coenzyme A, G6P = glucose-6-phosphate, GDH = glutamate dehydrogenase, GLS = glutaminase, GLUT = glucose transporter, HK2 = hexokinase 2, LAT = L-amino acid transporter, LDHA = lactic dehydrogenase A, MCT = monocarboxylate transporter, mTOR = mammalian target of rapamycin, PDH = pyruvate dehydrogenase, PDK1 = pyruvate dehydrogenase kinase 1, ROS = reactive oxygen species, TA = transaminase, TCA cycle = tricarboxylic acid cycle.

Molecular imaging is a promising method for non-invasively investigating the metabolic characteristics of different types of tumors and differentiating tumor progression from treatment effect (Wu et al., 2016). Recently, a pH- and oxygen-sensitive molecular imaging technique using amine chemical exchange saturation transfer spin-and-gradient-echo echoplanar imaging (CEST-SAGE-EPI) was introduced as a tool to explore tumor acidity and hypoxia simultaneously in human gliomas (Harris et al., 2018). Amine CEST exhibits sensitivity to tissue pH, along with other factors, through quantification of pH-dependent chemical exchange between amine protons on amino acids and bulk water (Sun et al., 2010). Further, the reversible transverse relaxation rate R₂′, which is sensitive to local deoxyhemoglobin concentration and hence tissue hypoxia (Fujita et al., 2003), can also be obtained using the multi-echo readout available using CEST-SAGE-EPI and can provide value into oxygen utilization (Harris et al., 2018). However, oxygen extraction is not completely echoed by measures of R₂′, as local oxygen extraction can be influenced by underlying blood volume. Even though ¹⁵O PET and ¹⁷O MRI can specifically detect oxygen consumption, they require radioactive tracers and stable isotopes, respectively (Atkinson and Thulborn, 2010, Lebrun-Grandie et al., 1983, Niesporek et al., 2018). We hypothesize a combination of the above-mentioned CEST-SAGE-EPI and perfusion imaging can be used to quantify the degree of aerobic glycolysis. Specifically, we theorize that an estimate of the relative oxygen extraction fraction (rOEF) can be quantified as the ratio of R₂′ to relative cerebral blood volume (rCBV) from perfusion imaging (Toth et al., 2013). Thus, an “aerobic glycolytic index,” or AGI, can then be derived as the ratio of tumor acidity, measured by amine CEST, to rOEF, assuming that aerobic glycolysis is reflected by production of lactate under inefficient use of oxygen.

The current study aims to demonstrate the association between AGI and glycolytic function in human gliomas. Using four approaches, we investigate (I) the difference in AGI between different tumor and tissue types; (II) the association between AGI and glucose utilization using ¹⁸F-fluorodeoxyglucose (FDG) PET; (III) the correlation between AGI and immunohistochemistry expression of key proteins or enzymes involved in aerobic glycolysis including HIF1α, glucose transporter 3, hexokinase 2, monocarboxylate transporter 1, and lactate dehydrogenase A, from image-guided biopsy specimens; and (IV) the relationship between AGI and measurements derived from extracellular flux bioenergetic analysis (Wu et al., 2007) of cultured cells obtained from resected tumor tissue.

2. Materials and methods

In order to characterize the association between AGI and glycolytic function in human gliomas, we performed four separate experiments: Study I compared AGI between IDH mutant and wild-type, and low-grade and high-grade gliomas, and between different tissue types; Study II investigated the correlations between AGI and glucose uptake using ¹⁸F-FDG PET; Study III examined the association between AGI and immunohistochemistry staining of critical proteins or enzymes related to aerobic glycolysis, namely, HIF1α, glucose transporter 3, hexokinase 2, monocarboxylate transporter 1, and lactate dehydrogenase A, from image-guided biopsy samples; and Study IV explored the correlation between AGI and measurements derived from extracellular flux bioenergetic analysis of cultured cells acquired from resected glioma tissue.

2.1. Patient population

This study was conducted from April 2015 to May 2018 with institutional review board approval (IRB# 14–001261; 10–000655). Written informed consent was acquired from all participants prior to study-related procedures. All de-identified patient information was stored on a secure research database. This study included 4 distinct patient populations. Table 1 shows the characteristics of these patients. Overlap in patients included: 8 and 3 patients in population I were also in population III and IV, respectively. Forty-four subjects were also included in a previous study that reported the metabolic features of diffuse gliomas by using CEST-EPI and CEST-SAGE-EPI without perfusion imaging (Yao et al., 2019). IDH mutation status was confirmed by genomic sequencing analysis using immunohistochemistry and/or polymerase chain reaction as previously described (Lai et al., 2011).

Table 1.

Patient demographics.

Study I	All patients	IDH mutant	IDH wild type
No. of patients*	51	30	21
Median age [range] (y)	52 [21–90]	40.5 [22–78]	59 [21–90]
Sex male/female	38/13	20/10	18/3
WHO grade (II/III/IV)	19/18/14	17/11/2	2/7/12
Scanner Prisma/Skyra	31/20	19/11	12/9
Study II	All patients	IDH mutant	IDH wild type
No. of patients**	8	–	8
Median age [range] (y)	60 [50–72]	–	60 [50–72]
Sex male/female	1/7	–	1/7
WHO grade (II/III/IV)	0/0/8	–	0/0/8
Scanner Prisma/Skyra	3/5	–	3/5
Study III	All patients	IDH mutant	IDH wild type
No. of patients*** (on/off treatment)	15 (2/13)	8 (1/7)	7 (1/6)
Median age [range] (y)	47 [29–62]	42.5 [29–49]	59 [44–62]
Sex male/female	10/5	4/4	6/1
WHO grade (II/III/IV)	5/5/5	4/3/1	1/2/4
Scanner Prisma/Skyra	9/6	5/3	4/3
Study IV	All patients	IDH mutant	IDH wild type
No. of patients (on/off treatment)	7 (3/4)	–	7 (3/4)
Median age [range] (y)	59 [51–68]	–	59 [51–68]
Sex male/female	6/1	–	6/1
WHO grade (II/III/IV)	0/1/6	–	0/1/6
Scanner Prisma/Skyra	3/4	–	3/4

* All patients in Study I were treatment-naïve or had been off any treatment for more than 2 years.

**All patients in Study II had recurrent IDH wild-type glioblastomas at time of MRI

*** Patients prior to any treatment or who had been off any treatment for more than 2 years were considered to be off treatment. Patients who were either on active treatment or recently off treatment at the time of MRI scan were considered on treatment.

Patient population I included a total of 51 patients with pathology confirmed diffuse glioma (median age, 52 years; range 21–90 years). Inclusion criteria included: (i) adult patients with newly diagnosed glioma or recurrent glioma without any medical/surgical treatment for last 2 years; (ii) patients with CEST-SAGE-EPI and perfusion imaging.

Study II utilized baseline, pretreatment imaging data acquired as part of a prospective study to evaluate the metabolic treatment effect of osimertinib (epidermal growth factor receptor tyrosine kinase inhibitor, ClinicalTrials.gov # NCT03732352) in patients with recurrent IDH wild-type glioblastoma. Patient population II included 8 patients (median age, 60 years; range 50–72 years). All subjects were scanned with CEST-SAGE-EPI, perfusion imaging, and ¹⁸F-FDG PET at recurrence, prior to any treatment, for the current study.

Patient population III included 15 patients with histologically confirmed diffuse gliomas (median age, 47 years; range 29–62 years). A total of 38 tissue samples (21 from IDH mutant and 17 from IDH wild type) were acquired.

Patient population IV included 7 patients (median age, 59 years; range 51–68 years) with high-grade (WHO grade III and IV) IDH wild-type gliomas. The contrast-enhancing portions of tumor were biopsied and cultured.

2.2. MR acquisition

All patients were scanned with CEST-SAGE-EPI, dynamic susceptibility contrast (DSC) perfusion MRI, and anatomical imaging on 3-T scanners (Prisma or Skyra, Siemens Healthcare, Erlangen, Germany). Anatomical imaging was performed according to the standardized brain tumor imaging protocol (Ellingson et al., 2015). CEST imaging was performed prior to contrast administration. The CEST-SAGE-EPI sequence consisted of a saturation pulse train of three 100-ms Gaussian pulses with the peak amplitude B₁ of 6 μT and a SAGE-EPI readout consisting of 2 gradient echoes with echo times of 14.0 and 34.1 ms, one asymmetric spin-echo with a TE of 58.0 ms, and one spin-echo with echo time of 92.4 ms. The other acquisition parameters were: repetition time greater than 10000 ms; field of view = 217 × 240 mm; matrix size = 116 × 128; number of slices = 25; slice thickness = 4.0 mm with no interslice gap; partial Fourier encoding = 6/8, generalized autocalibrating partially parallel acquisition (GRAPPA) = 3; and bandwidth = 1628 Hz/pixel. A total of 29 z-spectral points was acquired at offset frequency from − 3.5 ppm to − 2.5 ppm; from − 0.3 ppm to + 0.3 ppm; and from + 2.5 ppm to + 3.5 ppm, all with respect to the water proton resonance frequency. An additional reference (S0) scan was obtained with 4 averages using the identical parameters without saturation pulses. The other details of the sequence are described elsewhere (Harris et al., 2018). The total acquisition time of CEST-SAGE-EPI was 7 min 30 s, benchmarked on a 3-T Prisma MR scanner (Software Version VE11C).

For DSC-perfusion MRI, a total dose of 0.1 mmol/kg of gadolinium diethylenetriaminepentacetate (Magnevist; Bayer HealthCare Pharmaceuticals, Wayne, New Jersey) was administered with 0.025 mmol/kg for the preload dosage to mitigate T1-based leakage contamination and the remaining 0.075 mmol/kg for dynamic bolus administration. The preload dose and the start of baseline DSC-MRI was separated for 2 min. The acquisition parameters were: echo time/repetition time = 17/2250–2000 ms; flip angle = 35–90°; field of view = 240 × 240 mm; matrix size = 128 × 128; slice thickness = 4–5 mm with no interslice gap; number of slices = 25–30; number of baseline acquisitions before contrast agent injection = 10–15; number of time points = 120; and acquisition time ≈ and approximately 4 min.

2.3. Postprocessing of MRI data

All CEST-SAGE-EPI data were corrected for motion using rigid transformation (mcflirt; Functional Magnetic Resonance Imaging of the Brain Software Library, Oxford,UK; http://www.fmrib.ox.ac.uk/fsl/) and corrected for B₀ inhomogeneities using a z-spectra based k-means clustering and Lorentzian fitting algorithm (Yao et al., 2018). Then, the integral of a width of 0.4 ppm was calculated around both the − 3.0 and + 3.0 ppm spectral points (−3.2 to − 2.8 ppm and + 2.8 to + 3.2 ppm, respectively). They were coupled with the corresponding S₀ image to quantify the asymmetry in the magnetization transfer ratio (MTR_asym) at 3.0 ppm, a measure related to pH (Harris et al., 2016), as defined by the following equation: MTR_asym(3.0 ppm) = S(−3.0 ppm)/S₀ − S(+3.0 ppm)/S₀, where S(ω) is the signal of bulk water available after the saturation pulse with offset frequency ω and S₀ is the signal available without application of saturation pulse. The first and second gradient echoes were averaged for the MTR_asym at 3.0 ppm to augment the available signal-to-noise. Calculation of R₂′ (=R₂* − R₂) was performed by solving a set of Bloch equations as previously described (Harris et al., 2018).

The dynamic time-series images of DSC MRI data were motion-corrected using mcflirt. Then, normalized relative CBV (nrCBV) was calculated using a leakage correction algorithm that accounts for bidirectional contrast agent exchange (Leu et al., 2016) and normalizing the CBV to the normal-appearing white matter tissue contralateral to the glioma. Creation of MTR_asym, R₂′, and nrCBV maps were performed with MatLab (release 2017b, MathWorks) using in-house programs. All resulting maps were registered to three-dimensional post-contrast T1-weighted images for subsequent analyses (tkregister2; Freesurfer; Massachusetts General Hospital, Harvard Medical School; https://surfer.nmr.mgh.harvard.edu).

rOEF map was calculated by dividing R₂′ by nrCBV as described previously (Toth et al., 2013). R2′ is the transverse relaxation due to local magnetic field inhomogeneities in a blood vessel network with a venous fraction of ≈0.75 × CBV (Yablonskiy et al., 2013). Assuming fully oxygenated arterial blood, the fraction of deoxygenated blood (1 $-$ Y) corresponds to the OEF:

$$\mathit{OEF}=1-Y=\frac{R2\text{'}}{0.75CBV\bullet \gamma \bullet \frac{4\pi }{3}\bullet {\mathrm{\Delta }\chi }_0\bullet Hct\bullet B_0}$$

In this equation, γ = 2.675 × 10⁸ s⁻¹ T⁻¹ is the gyromagnetic ratio, Δχ₀ = 0.264·× 10^− 6 is the susceptibility difference between fully deoxygenated and fully oxygenated blood (Spees et al., 2001), Hct = 0.35 is the small vessel hematocrit (Silvennoinen et al., 2003), and B₀ is the magnetic field strength. Then, relative measurement of OEF can be expressed as

$$\mathit{rOEF}=\frac{R2\text{'}}{\mathit{nrCBV}}$$

AGI was calculated by dividing MTR_asym at 3.0 ppm by rOEF, under the assumption that aerobic glycolysis is reflected by production of lactate under inefficient use of oxygen. In equation, AGI was calculated as:

$$\mathit{AGI}=\frac{{\mathit{MTR}}_{\mathit{asym}}at3.0ppm}{\mathit{rOEF}}=\frac{{\mathit{MTR}}_{\mathit{asym}}at3.0ppm\times nrCBV}{R2\text{'}}$$

2.4. ¹⁸F-FDG PET imaging (Study II)

All subjects in Study II underwent ¹⁸F-FDG PET imaging on the same day as the CEST imaging. ¹⁸F-FDG was synthesized using standard methods (Hamacher et al., 1986, Yu, 2006) to an average specific radioactivity of 200 GBq/mmol. After a minimum fasting period of 6 h, PET data were acquired for 15 min on PET/CT scanners (Biograph64-mCT or Biograph TruePoint, Siemens Healthcare, Erlangen, Germany) in 3D mode, 90 min after intravenous administration of ¹⁸F-FDG (5 mCi). The voxel size of FDG-PET imaging was 1.6 mm × 1.6 mm × 1.5 mm or 1.34 mm × 1.3 m × 1.5 mm. Also, CT imaging was performed and reconstructed using a standard filtered back projection algorithm to correct for photon attenuation of PET data.

PET images were then registered to the post-contrast T1-weighted images using 12-degree of freedom affine registration transformation with a normalized mutual information cost function (flirt; Functional Magnetic Resonance Imaging of the Brain Software Library). Uptake levels of ¹⁸F-FDG were normalized to the normal-appearing white matter tissue contralateral to the glioma to mitigate intersubject variability, resulting in normalized FDG (nFDG) maps (Ellingson et al., 2020).

2.5. Biopsy acquistion and immunohistochemistry (Study III)

Prospective biopsy targets were identified prior to surgery. Specifically, two to four MRI targets as spheres of 5 mm diameter were selected for each patient. These targets were selected based on the MTR_asym at 3.0 ppm and R₂′ maps in regions with high/low acidity and high/low hypoxia. Biopsy targets were uploaded to intraoperative navigation software (Brainlab, Munich, Germany). Standard of care tumor resection was carried out with acquisition tissue corresponding to biopsy targets using intraoperative neuronavigation guidance.

Immunohistochemistry analysis was performed on 5 μm formalin-fixed, paraffin-embedded tissue sections with a customizable immunohistochemistry detection system. This all-in-one detection system contained Antigen Retrieval Stock Solution (100x), Antibody Diluent, Peroxidase Block, Protein Block, Horseradish Peroxidase Polymer Conjugated Secondary Antibody, DAB Chromogen, DAB Substrate Buffer, and Hematoxylin and Bluing Solution. Heat-induced antigen retrieval was accomplished with Antigen Retrieval Buffer (pH 6.0) in a Decloaking Chamber at 95℃ for 30 min (Biocare Medical, Concord, CA, USA). Then, tissue sections were treated with Peroxidase Block and with Protein Block to reduce nonspecific background staining. Primary antibodies for HIF1 α (Sigma/Aldrich, HPA001275, 1:200), GLTU3 (Sigma-Aldrich, HPA006539, 1:200), hexokinase 2 (Sigma-Aldrich, HPA028587, 1:100), monocarboxylate transporter 1 (Sigma-Aldrich, SAB2702323, 1:200), lactate dehydrogenase A (Abcam, Ab47010, 1:500), and Ki67 (Sigma/Aldrich, 275R-15, 1:100) were applied for 60 min followed by detection with Horseradish Peroxidase Polymer Conjugated Secondary Antibody (anti-Rabbit or anti-Mouse). Subsequent immunodetection was completed using DAB Chromogen and counterstained with Hematoxylin and Bluing Solution. Tissue slides were scanned on a digital slide scanner (Aperio CS2; Aperio Technologies). The percentage of positive cells was calculated as the ratio of the number of positive cells to total cells in a tissue section, using semi-automated positive cell detection algorithm implemented in QuPath ver. 0.2.0-m8 (https://qupath.github.io/) (Bankhead et al., 2017).

2.6. Analysis of in vitro cellular metabolism (Study IV)

MRI-guided tissue biopsy of the contrast-enhancing region of a high-grade glioma using intraoperative image guidance was performed prior to surgical resection. Tumor cells from each sample (one cluster per each subject) were isolated as previously described (Nathanson et al., 2014). Briefly, ≈0.5 g of tumor tissue was digested via mechanical and enzymatic digestion (collagenase/DNAse), strained through a 70-μm filter, then further processed through myelin depletion by Myelin Removal Beads II (Miltenyi Biotec GmbH, Bergisch Gladbach, Germany), followed by CD45 + leukocyte depletion using CD45 MicroBeads human (Miltenyi Biotec GmbH) and red blood cell lysis. Dead cells were removed via the Dead Cell Removal Kit (Miltenyi Biotec GmbH), yielding ≈5.0 × 106 viable tumor cells/g tissue. After isolation, tumor cells were maintained as serum-free sphere cultures (gliomasphere) (Laks et al., 2016). Non-adherent gliomaspheres were dissociated, counted, and acutely attached to 25 μg/mL of Cell-Tak (Corning, New York, NY, USA) treated Seahorse assay plate via centrifugation (600 g for 5 min). Then, an Agilent Seahorse XF Analyzer (Agilent Technologies, Santa Clara, CA) was used to measure rates of oxygen consumption (OCR) and extracellular acidification (ECAR)(Divakaruni et al., 2014). Basal OCR was calculated by subtracting the rate insensitive to the electron transport chain inhibitors rotenone and antimycin A, and normalized to number of cells. Adenosine triphosphate production via oxidative phosphorylation and glycolysis were also estimated from OCR and ECAR, respectively (Divakaruni et al., 2014). This was done by correcting ECAR for respiratory acidification and applying established stoichiometry. Additional experimental details and considerations for adenosine triphosphate estimation are described elsewhere (Divakaruni et al., 2014). The fraction of adenosine triphosphate derived from glycolysis to the total adenosine triphosphate was calculated and termed as cellular AGI.

2.7. Region-of-interest analysis

For imaging data of subjects in Study I and II, three mutually exclusive regions-of-interest (ROIs) were defined using a semi-automated thresholding method (Ellingson et al., 2014). They included (a) contrast-enhancing tumor and (b) central necrosis defined by T1-weighted digital subtraction maps; and (c) T2 hyperintense regions on T2-weighted FLAIR images (non-enhancing tumor), excluding areas of contrast enhancement and necrosis. Contrast-enhancing tumor + non-enhancing tumor, contrast-enhancing tumor, and non-enhancing tumor ROIs were created from these ROIs. Normal-appearing white matter ROI contralateral to the glioma was additionally created for patients in Study II, by manually delineating frontal and paietal white matter. Further, we segmented the brain of patients in Study II using SPM12 software (Wellcome Department of Imaging Neuroscience, University College London, London, UK; http://www.fil.ion.ucl.ac.uk/spm), added the gray and white matter masks to created brain masks and thresholed them at 0.9, added contrast-enhancing tumor + non-enhancing tumor masks to the brain masks, and finally binarized them to create the whole brain ROI. For analyzing data of subjects in Study III and IV, MRI targets for biopsy and contrast-enhancing tumor ROIs, respectively, were used as ROIs. These ROIs were used to retrieve median AGI and nFDG.

2.8. Statistical analysis

For patients in Study I, median AGI between IDH mutant and wild-type, and between low and high grade (i.e. grade II vs III and IV) was compared by using Mann–Whitney U test. AGI between contrast-enhancing tumor, non-enhancing tumor, and normal-appearing white matter was also compared only among gliomas with contrast-enhancing tumor using Kruskal-Wallis test with post-hoc Tukey's Honest Significant Difference test. Spearman’s rank order correlation analysis was performed to investigate the correlation between AGI and nFDG of contrast-enhancing tumor + non-enhancing tumor, contrast-enhancing tumor, non-enhancing tumor, or the whole brain ROIs for subjects in Study II, AGI and quantitative immunohistochemistry measurements for subjects in Study III, and AGI and cellular AGI measured by bioenergetics analysis for subjects in Study IV. Confidence intervals of Spearman’s ρ correlation coefficients were determined by percentile bootstrapping. The absolute values of Spearman’s ρ correlation coefficients were classified as follows: 0–0.30, very weak; 0.30–0.50, weak; 0.50–0.70, moderate; 0.70–0.90, strong; and 0.90–1.00, very strong. A 2-sided P value<0.05 was considered statistically significant. Correction for multiple correlation analyses was not performed, to prevent overly conservative false negative detection rates considering the small sample sizes. All statistical analyses were performed by using MatLab or GraphPad Prism (Version 8.4.2; GraphPad Software, La Jolla, CA).

3. Results

3.1. Comparison of AGI between different tumor and tissue types

Figure 2 shows imaging findings of three representative cases that showed high, medium, and low AGI, and high and low ¹⁸F-FDG uptake. Fig. 3 shows comparison of AGI between different tumor and tissue types. AGI was significantly lower in IDH mutant than IDH wild-type gliomas (mean ± standard deviation, 0.48 ± 0.48 vs. 0.70 ± 0.48; P = 0.03), and significantly lower in low-grade than high-grade gliomas (mean ± standard deviation, 0.34 ± 0.27 vs. 0.70 ± 0.54; P = 0.01). When compared across different tissue types, AGI was the highest in contrast-enhancing tumor, followed by non-enhancing tumor and normal-appearing white matter (mean ± standard deviation, 1.25 ± 0.72 vs. 0.41 ± 0.36 vs. 0.08 ± 0.05; all post-hoc P-values < 0.05).

Fig. 2.

Patients with diffuse gliomas with high (A, from Study II), medium (B, from Study II), and low AGI (C, from Study I), and high (A) and low (B) ¹⁸F-FDG uptake. IDH wild-type glioblastomas (A and B) show higher AGI than IDH mutant grade II glioma (C). Contrast-enhancing region of the tumor is showing higher AGI than non-contrast-enhancing region of the tumor (A and B). MTR_asym at 3.0 ppm is grossly high in the entire tumors for both glioblastomas, while AGI seems to be correlated more with ¹⁸F-FDG PET than MTR_asym at 3.0 ppm. While ¹⁸F-FDG PET shows high physiological uptake in the cortex, quantitative maps, including AGI, derived from MRI do not show high values in the cortex. NAWM = normal-appearing white matter.

Fig. 3.

Comparison of AGI between different tumor and tissue types. Box plots are overlaid on violin plots. (A) Comparison between IDH mutant and wild type. Among IDH mutant gliomas, those with AGI greater than 1 are mostly high grade (4 out of 5). (B) Comparison between low-grade and high-grade gliomas. (C) Comparison across contrast-enhancing tumor, non-enhancing tumor, and normal-appearing white matter. * P < 0.05, ** P < 0.01 *** P < 0.001.

3.2. Correlation between MR metrics and ¹⁸F-FDG PET in IDH Wild-type glioblastoma

Figure 4 shows the scatterplots for comparison between AGI and nFDG. In comparison between AGI and nFDG, strong statistically significant correlations were observed in contrast-enhancing tumor + non-enhancing tumor region (ρ = 0.81; 95% CI 0.21–1.00; P = 0.01), non-enhancing tumor region (ρ = 0.81; 95%CI 0.32–1.00; P = 0.01), and contrast-enhancing tumor region (ρ = 0.81; 95% CI 0.10–1.00; P = 0.01). No significant correlation was found in the whole brain (ρ =  − 0.62; 95% CI − 1.00–0.20; P = 0.12).

Fig. 4.

Scatterplots of AGI compared with normalized ¹⁸F-FDG in the contrast-enhancing plus non-enhancing (A), contrast-enhancing (B), and non-enhancing (C) portions of IDH wild-type glioblastoma and (D) the whole brain. Lines show linear regression fit to the data. Strong correlations were found in all tumor areas, while no significant correlation was found in the whole brain.

3.3. Correlation between MR metrics and immunohistochemistry measurements in WHO grade II–IV diffuse gliomas

Figure 5 shows examples of MRI-based biopsy targets along with corresponding MR image and AGI map, and glucose transporter 3, hexokinase 2, or monocarboxylate transporter 1 expression within these areas. Fig. 6 shows the correlation matrix for comparison between AGI and immunohistochemistry measurements and the scatterplots of AGI compared with glucose transporter 3, hexokinase 2, and monocarboxylate transporter 1. The scatterplots of all comparisons are shown in Supplementary Material (Fig. S1). Strong significant correlations were found between AGI and glucose transporter 3 (ρ = 0.71; 95% CI 0.30–0.90 ; P = 0.006) or hexokinase 2 (ρ = 0.73; 95% CI 0.41–0.90; P = 0.004) in IDH wild-type glioma. In contrast, moderate significant correlation was found between AGI and monocarboxylate transporter 1 (ρ = 0.59; 95% CI 0.19–0.74 ; P = 0.01) in IDH mutant glioma. No significant correlation was found between AGI and HIF1 α, lactate dehydrogenase A, or Ki-67.

Fig. 5.

MR image and AGI map and corresponding H&E and immunohistochemistry staining for MRI-guided biopsy targets (circles). (A) Treatment-naïve IDH wild-type glioblastoma for which an area with low AGI was biopsied. Expressions of GLUT3 and HK2 are low in the slides from a 5-mm radius sample taken from the MRI-guided biopsy target. (B) Treatment naïve IDH wild-type glioblastoma for which an area with high AGI was biopsied. Expressions of GLUT3 and HK2 are high. (C) Post-surgical, chemoradiotherapy-naïve IDH mutant WHO grade III glioma for which an area with low AGI was biopsied. Expression of MCT1 is low. (D) Treatment-naïve IDH mutant WHO grade III glioma for which an area with high AGI was biopsied. Expression of MCT1 is high. For GLUT3, HK2, and MCT1 stains, cells positive for expression are brown. GLUT3 = glucose transporter 3, HK2 = hexokinase 2, MCT1 = monocarboxylate transporter 1.

Fig. 6.

(A) Correlation matrix of AGI compared with immunohistochemistry measurements shown with Spearman’s correlation coefficients. * P < 0.05, ** P < 0.01. Scatterplots of AGI compared with immunohistochemistry measurement of GLUT3 (B), HK2 (C), and MCT1 (D). Lines show linear regression fit to the data. GLUT3 = glucose transporter 3; HK2 = hexokinase 2, LDHA = lactate dehydrogenase A, MCT1 = monocarboxylate transporter 1.

3.4. Correlation between MR metrics and cellular metabolism in High-Grade glioma

Figure 7 shows imaging findings of two representative cases that showed high and low AGI, respectively, along with measurements by extracellular flux bioenergetic analysis. Fig. 8 shows the scatterplot of AGI derived from MRI compared with cellular AGI. A strong significant correlation was found between AGI and cellular AGI (ρ = 0.79; 95% CI 0.06–1.00; P = 0.04).

Fig. 7.

MR images/maps of two patients with IDH wild-type glioblastoma that showed higher (A) and lower (B) AGI derived from MRI and extracellular flux bioenergetic analysis. An ROI of contrast-enhancing portion is overlaid on each image/map. (C), (D) Bioenergetic analysis shows higher ECAR and lower basal OCR in (A), leading to higher cellular AGI that explains higher AGI derived from MRI, whereas ECAR and OCR are lower and higher, respectively, for (B), leading to lower cellular AGI that explains lower AGI derived from MRI.

Fig. 8.

Scatterplot of AGI compared with cellular AGI derived from bioenergetic extracellular flux analysis. Lines show linear regression fit to the data. A strong correlation was found between AGI and cellular AGI.

4. Discussion

The current study suggests aerobic glycolytic imaging derived from pH-, O₂–, and perfusion-weighted MRI can provide non-invasive information about the metabolic status of human diffuse gliomas. When comparing between IDH mutant and wild-type and between low- and high-grade gliomas, AGI was higher in IDH wild-type and high-grade gliomas than IDH mutant and low-grade gliomas, respectively. These observations are in line with the fact that IDH wild-type and high-grade gliomas have typical Warburg phenotypes, whereas IDH mutant and low-grade gliomas are more dependent on oxydative phosphorylation (Khurshed et al., 2017, Poteet et al., 2013). Further, contrast-enhancing tumor and non-enhancing tumor showed higher AGI than normal-appearing white matter, with contrast-enhancing tumor showing higher AGI than non-enhancing tumor. This may reflect the generally higher malignant status of contrast-enhancing tumor with a higher glycolytic activity than non-enhancing tumor.

Through careful metabolic investigation in Study II, we observed a strong correlation between the AGI and normalized ¹⁸F-FDG PET uptake in both enhancing and non-enhancing tumor in glioblastoma, while no significant correlation was found in the whole brain. While ¹⁸F-FDG PET is useful for assessing response of solid tumors in other organs, its utility is limited in the brain partly due to high physiological uptake of ¹⁸F-FDG in the normal brain. Because AGI exploits aspects of abnormal metabolism, it does not show high value in normal brain structures and therefore may be useful for both disease detection and treatment monitoring. Further, anatomical resolution is higher for AGI than ¹⁸F-FDG PET. AGI may enable us to differentiate radiation necrosis from true progression, with potentially better sensitivity than ¹⁸F-FDG PET due to small background AGI value (Hustinx et al., 2005). AGI may also have an indication for assessing treatment response likewise ¹⁸F-FDG PET (Verger and Langen, 2017), with its ability to capture abnormal tumor metabolism. Notably, gluco-CEST and gluco-chemical exchange sensitive Spin-Lock (CESL) that have been applied in glioma also do not exhibit high background signal (Paech et al., 2017, Schuenke et al., 2017, Xu et al., 2015). However, it is expected that these approaches primarily reflect perfusion and blood–brain barrier, similarly to conventional contrast agents.

In regard to association with immunohistochemistry expression measurements, AGI showed strong significant correlation with glucose transporter 3 and hexokinase 2 in IDH wild-type glioma, while these correlations were not significant for IDH mutant glioma. Interestingly, AGI showed a moderate correlation with monocarboxylate transporter 1 in IDH mutant glioma, while this correlation was not significant for IDH wild-type glioma. These results suggest that glucose transporter 3 and hexokinase 2, which are engaged in irreversible uptake of glucose by tumor cell (Venneti and Thompson, 2017), are likely the rate-limiting factors of aerobic glycolysis in IDH wild-type glioma, in which a larger fraction of glucose is shunted to glycolysis instead of oxidative phosphorylation. On the other hand, monocarboxylate transporter, which exports lactate (the end-product of glycolysis) from the cells into the extracellular space (Yu et al., 2016), is downregulated in IDH mutant glioma (Viswanath et al., 2016) and may modulate the rate of aerobic glycolysis in IDH mutant gliomas. This is consistent with recent studies implicating lower monocarboxylate transporter expression as a major contributor to lower lactate production in immortalized IDH mutant astrocytes (Viswanath et al., 2016).

Consistent with the hypothesis that AGI derived from pH-, O₂–, and perfusion-weighted MRI reflects the degree of aerobic glycolysis specific to the underlying tumor tissue, we found strong and significant correlation between AGI and the degree of aerobic glycolysis quantified using extracellular flux bioenergetic analysis in purified tumor obtained after resection. Even though the degree of aerobic glycolysis derived from both MRI and extracellular flux analysis aim to estimate the glycolytic status, caution should be paid to the fact that their theoretical underpinnings are different. Specifically, cellular measures directly reflect adenosine triphosphate production per cell per minute, while MRI measurements are averaged over a large volume of tumor tissue. How the distribution of cells with differing metabolic features within the tumor influence MRI measures remains to be fully investigated.

There are other metabolic imaging approaches, such as ¹⁵O PET and ¹⁷O MRI (Fan et al., 2020, Zhu and Chen, 2017), which can potentially capture Warburg effect of gliomas. For example, ¹⁷O MRI has detected decreased oxygen consumption in gliomas, presumably related to Warburg effect (Paech et al., 2020). The results of the current study warrant investigation of correlation of AGI with ¹⁵O PET or ¹⁷O MRI.

This study has a number of limitations. First, as mentioned above, the sample size for each experiment was small and MR acquisitions were performed on two different scanners. However, despite these sources of variability, results appeared to support the general hypothesis that aerobic glycolytic imaging derived from pH-, O₂–, and perfusion-weighted MRI reflects the general glycolytic status of the underlying tumor tissue. Second, AGI is a semi-quantitative approach and should not be considered as an absolute measurement of aerobic glycolysis. MTR_aym at 3.0 ppm and rOEF used for calculating AGI are both semi-quantiative measurements. Further, even though MTR_aym at 3.0 ppm is sensitive to tissue pH, it is not necessarily in a linear relationship with lactate concentration. Our approach may be oversimplified, and a more sophisticated approach could be more appropriate and needs to be explored in the future. Third, since AGI requires separate sets of image acquisitions (i.e. CEST-SAGE-EPI and DSC perfusion imaging), there may be some contamination from minor mis-registration, which could be problematic in evaluating small tumors or tumors residing in areas harboring large susceptibility-induced artifacts (Hagiwara et al., 2020). Fourth, even though a strong correlation was found between AGI and metrics related to glucose metabolism, AGI can also be affected by mechanisms unrelated to glucose metabolism. For example, MTR_asym at 3.0 ppm increases with increase in glutamine, which is consumed at a higher rate in malignant tumors (Harris et al., 2015). Correlation of AGI with abnormal glucose metabolism should be further investigated in future studies in diverse conditions. Lastly, the current study did not take into consideration the metabolic demand of immune cells that might be present in vivo within tumor tissue. Activated effector T cells also exhibit aerobic glycolysis, though the precise function in T cells is incompletely understood (Salmond, 2018). While results from the current study suggest AGI might be specific to abnormal metabolism from tumor cells, the influence of immune cell metabolism on AGI needs to be fully investigated, particularly if it is to be used in the context of immunotherapy.

In conclusion, aerobic glycolytic imaging showed a strong correlation with ¹⁸F-FDG uptake and the degree of aerobic glycolysis of purified tumor cells using bioenergetic analysis. Aerobic glycolytic imaging correlated with key metabolic enzymes as measured with immunohistochemistry, although there were distinct differences in these associations for IDH mutant compared to wild-type gliomas, perhaps due to particular rate limiting enzymes specific to the glioma genotype. Future studies are needed to investigate the utility of aerobic glycolytic imaging for both diagnosis and treatment monitoring.

Funding

This work was supported by American Cancer Society (ACS) Research Scholar Grant (RSG-15–003-01-CCE) (BME); University of California Research Coordinating Committee (BME); UCLA Jonsson Comprehensive Cancer Center Seed Grant (BME); UCLA SPORE in Brain Cancer (NIH/NCI 1P50CA211015-01A1) (BME, LML, PLN, AL, WBP, TFC); NIH/NCI 1R21CA223757-01 (BME); NIH-NIGMS Training Grant GM008042 (NSC); and American Brain Tumor Association Grant Jack & Fay Netchin Medical Student Summer Fellowship supported by BrainUp MSSF2100033 (NSC).

CRediT authorship contribution statement

Akifumi Hagiwara: Data curation, Formal analysis, Visualization, Writing – original draft. Jingwen Yao: Data curation, Formal analysis, Writing – review & editing. Catalina Raymond: Data curation, Writing – review & editing. Nicholas S. Cho: Investigation, Writing – review & editing. Richard Everson: Investigation, Writing – review & editing. Kunal Patel: Data curation, Writing – review & editing. Danielle H. Morrow: Data curation, Formal analysis, Writing – review & editing. Brandon R. Desousa: Investigation, Writing – review & editing. Sergey Mareninov: Formal analysis, Writing – review & editing. Saewon Chun: Data curation, Writing – review & editing. David A. Nathanson: Investigation, Writing – review & editing. William H. Yong: Investigation, Writing – review & editing. Gafita Andrei: Investigation, Writing – review & editing. Ajit S. Divakaruni: Investigation, Writing – review & editing. Noriko Salamon: Investigation, Writing – review & editing. Whitney B. Pope: Investigation, Writing – review & editing, Investigation, Writing – review & editing. Phioanh L. Nghiemphu: Investigation, Writing – review & editing. Linda M. Liau: Investigation, Writing – review & editing. Timothy F. Cloughesy: Investigation, Writing – review & editing. Benjamin M. Ellingson: Conceptualization, Visualization, Writing – review & editing.

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

ASD has previously served as a paid consultant to Agilent Technologies. BME is on advisory board of Hoffman La-Roche, Siemens, Nativis, Medicenna, MedQIA, Bristol Meyers Squibb, Imaging Endpoints, and Agios. BME is a paid consultant of Nativis, MedQIA, Siemens, Hoffman La-Roche, Imaging Endpoints, Medicenna, and Agios. BME has grant funding by Hoffman La-Roche, Siemens, Agios, and Janssen. BME holds a patent on this technology (US Patent #15/577664; International PCT/US2016/034886).

Appendix A. Supplementary data

The following are the Supplementary data to this article: