Tissue metabolites in diffuse glioma and their modulations by IDH1 mutation, histology, and treatment

The discovery of the oncometabolite 2-hydroxyglutarate in isocitrate dehydrogenase 1–mutated (IDH1-mutated) tumor entities affirmed the role of metabolism in cancer. However, large databases with tissue metabolites that are modulated by IDH1 mutation remain an area of development. Here, we present an unprecedented and valuable resource for tissue metabolites in diffuse glioma and their modulations by IDH1 mutation, histology, and tumor treatments in 101 tissue samples from 73 diffuse glioma patients (24 astrocytoma, 17 oligodendroglioma, 32 glioblastoma), investigated by NMR-based metabolomics and supported by RNA-Seq. We discovered comparison-specific metabolites and pathways modulated by IDH1 (IDH1 mutation status cohort) and tumor entity. The Longitudinal investigation cohort provides metabolic profiles of untreated and corresponding treated glioma samples at first progression. Most interestingly, univariate and multivariate cox regressions and Kaplan-Meier analyses revealed that tissue metabolites correlate with progression-free and overall survival. Thus, this study introduces potentially novel candidate prognostic and surrogate metabolite biomarkers for future prospective clinical studies, aiming at further refining patient stratification in diffuse glioma. Furthermore, our data will facilitate the generation of so-far–unanticipated hypotheses for experimental studies to advance our molecular understanding of glioma biology.


Introduction
Mutations in isocitrate dehydrogenase 1 (IDH1) or IDH2 are frequent in several cancer entities, including diffuse glioma (1)(2)(3). They lead to mutated enzymes with neomorphic activity, producing the oncometabolite 2-hydroxyglutarate (2-HG), which modulates cellular epigenetic programs, differentiation patterns, and metabolic profiles (4)(5)(6). Recent clinical trials in diffuse glioma have used the IDH1 mutation (IDH1-mut) as a therapeutic target (7,8). However, the therapeutic efficacy of these therapeutic approaches might critically depend on understanding metabolic alterations that occur in the context of IDH1 mutations. For example, 2-HG shapes the tumor immune microenvironment by suppressing T cell activity in glioma (9). Furthermore, the nucleotide synthesis utilization and DNA repair capacity is different in IDH1-mut and IDH1-WT glioma (10). These observations indicate that IDH1-mut glioma exhibit distinct therapeutic challenges and mechanisms for resistance to therapy.
Altered metabolism is a major hallmark of cancer, with the potential to promote tumor growth and acquired resistance to therapy (11). Glioma cells display abnormal energy metabolism, including alterations of glucose, amino acid, and fatty acid metabolism (12). Altered metabolic pathways including choline, taurine, hypotaurine, or glutamate/glutamine correlate with different WHO grades in diffuse glioma (13). Noninvasive detection of glutamate by magnetic resonance spectroscopy can serve as a metabolic imaging biomarker of response to temozolomide treatment in IDH1-mut glioma (14).
The discovery of the oncometabolite 2-hydroxyglutarate in isocitrate dehydrogenase 1-mutated (IDH1-mutated) tumor entities affirmed the role of metabolism in cancer. However, large databases with tissue metabolites that are modulated by IDH1 mutation remain an area of development. Here, we present an unprecedented and valuable resource for tissue metabolites in diffuse glioma and their modulations by IDH1 mutation, histology, and tumor treatments in 101 tissue samples from 73 diffuse glioma patients (24 astrocytoma, 17 oligodendroglioma, 32 glioblastoma), investigated by NMR-based metabolomics and supported by RNA-Seq. We discovered comparison-specific metabolites and pathways modulated by IDH1 (IDH1 mutation status cohort) and tumor entity. The Longitudinal investigation cohort provides metabolic profiles of untreated and corresponding treated glioma samples at first progression. Most interestingly, univariate and multivariate cox regressions and Kaplan-Meier analyses revealed that tissue metabolites correlate with progression-free and overall survival. Thus, this study introduces potentially novel candidate prognostic and surrogate metabolite biomarkers for future prospective clinical studies, aiming at further refining patient stratification in diffuse glioma. Furthermore, our data will facilitate the generation of so-far-unanticipated hypotheses for experimental studies to advance our molecular understanding of glioma biology.

Results
Study cohorts and comparisons based on clinical characteristics. We investigated distinct comparisons of tissue metabolites derived from NMR analyses, complemented with gene expression data from the RNA-Seq in selected comparisons, to investigate modulations by the IDH1 mutation and histology in the IDH1 mutation status cohort ( Figure 1A), and we examined tumor treatments in tissue pairs from newly diagnosed and corresponding progressive glioma with or without treatment between resections in the Longitudinal investigation cohort ( Figure 1B).
Furthermore, we performed ROC analysis of the top 4 significant metabolites in the histology-dependent comparisons (Supplemental Figures 2 and 3) and detected carnitine, o-phosphoethanolamine, ethanolamine, and histidine for comparison IV (Supplemental Figure 2A); citrate, glutathione, histamine, and  I  II  III  IV  V  VI  VII  VIII  IX  X Univariate metabolomics comparisons (fold change threshold > 1.2, A P < 0.0001, B P < 0.01, C P < 0.05, D P < 0.001). The arrows denote if a metabolite was elevated () or lowered (). According the consort diagram ( Figure  formate for comparison V (Supplemental Figure 2B); sarcosine, glutamate, taurine, and AMP for comparison VI (Supplemental Figure 3A); and scyllo-inositol, 4-aminobutyrate, alanine, and betaine for comparison VI (Supplemental Figure 3B). The PLS-DA score plots of all comparisons I-VII are illustrated in Supplemental Figure 5.  Table 2).
Comparison VIII ( Figure 1B) with untreated tissue pairs did not show any separation or clustering in the corresponding scores. We did not detect any patient-specific alterations -i.e., longitudinal alterations in tumor tissues from the same patient. In addition, we investigated general tissue metabolite differences in comparison VIII (Figure 7, A and B) between both time points -i.e., a pooled comparison of all newly diagnosed tissues and all tissues at first progression without any treatment between both resections. We identified in the PLS-DA 11 differentiating metabolites with VIP scores > 1.0, including myo-inositol, scyllo-inositol, N-acetylaspartate, glutamate, and O-phosphoethanolamine ( Figure 7, A and B).  I  II  III  IV  V  VI  VII  VIII  IX  X Univariate metabolomics comparisons (fold change threshold > 1.2, A P < 0.05, B P < 0.0001, C P < 0.001, D P < 0.01). The arrows denote if a metabolite was elevated () or lowered (). According the consort diagram (  All 3 patient-specific sample pairs showed a decrease in lactate after treatment. Furthermore, we observed 15 general tissue metabolite differences with VIP scores > 1.0 in the PLS-DA, including scyllo-inositol, myo-inositol, cystathionine, lactate, and formate ( Figure 8D).
Tissue metabolites correlate with clinical outcome parameters. Finally, we examined the relationship of tissue metabolites at initial diagnosis with progression-free survival (PFS) and overall survival (OS). First, we determined the relevant threshold for each metabolite (Supplemental Figure 4). For correlations with PFS, we identified 13 metabolites: PFS increased, with increasing concentrations of 2-HG, creatine phosphate, myo-inositol, and scyllo-inositol, while PFS decreased, with increasing concentrations of alanine, homocysteine, isoleucine, leucine, mannitol, o-acetylcholine, serine, threonine, and sn-glycero-phosphocholine (Supplemental Figure 4, A and B).
For correlations with OS, we discovered 15 metabolites: OS increased with increasing concentrations of 2-HG, acetate, formate, myo-inositol, and sn-glycero-phosphocholine. Furthermore, OS decreased with increasing concentrations of alanine, glutamate, glycine, homocysteine, isoleucine, leucine, mannitol, phenylalanine, threonine, and tyrosine (Supplemental Figure 4, C and D). The Kaplan-Meier curves indicated that prognostic values of distinct tissue metabolites (2-HG and myo-inositol for PFS, and 2-HG and acetate for OS groups 2 and 3) were clinically more favorable than the prognostic values of IDH1 mutation or histology (Figure 9).

Discussion
IDH1 mutations occur in a variety of cancers, and further insight into their effects on cellular processes -including metabolism in particular cancer entities -is an important area of development. These insights will improve an understanding of context-dependent alterations, which is crucial for successful clinical translation. We performed a comprehensive tissue metabolomics and pathway analysis approach (supported by RNA-Seq) in a large cohort of diffuse glioma patients and a high number of samples. We identified distinct metabolites and metabolic pathways that were modulated by IDH1 mutation status, histology, and tumor therapies. The correlation of metabolite with gene expression data allowed us to have a complementary look at the functional tumor biology level by an independent method (e.g., increased 2-HG levels matched with L2HGDH upregulation in IDH1-mut; ref. 15). The alterations of sn-glycero-3-phosphocholine in IDH1-mut (increased level) and IDH1-WT (decreased level) correlated well with alterations of LYPLA1 expression (Supplemental Figure 1). Low expression of LYPLA1 led to the accumulation of sn-glycero-3-phosphocholine (16), supporting our metabolite findings. Of note, many metabolites are substrates for more than 1 enzyme; thus, single comparisons of enzymatic activity and metabolite levels do not necessarily correlate -e.g., in the case of glutamate where we identified in the IDH1-WT glioblastoma group a total of 3 upregulated (OPLAH, EARS2, GOT1) and 2 downregulated (GLUD1, GCLC) enzymes ( Figure 4C).
Comparisons I-III illustrate potentially novel metabolite profiles modulated by IDH1 mutation that might further explain different clinical features and prognoses of glioma subtypes. For example, we identified increased levels of sarcosine, o-acetylcholine, and glutamate in IDH1-WT glioma compared with IDH1-mut glioma (Tables  1 and 2). Recent studies indicate that these neurotransmitters have tumor-promoting features; thus, this might help to further explain the unfavorable prognosis of IDH1-WT compared with IDH1-mut diffuse glioma (17)(18)(19)(20).
Increased hypotaurine level was another metabolic feature of IDH1-WT diffuse glioma (Tables 1 and  2). Recent preclinical studies suggest that hypotaurine activates hypoxia-related signaling and mediates tumor progression in glioma (21). Of note, correlations of ADO gene expression and hypotaurine metabolite levels did not correlate in IDH1-WT glioblastoma ( Figure 4C). We detected higher hypotaurine levels but reduced ADO gene expression in IDH1-WT glioblastoma. Two further enzymes that could be connected to changing hypotaurine levels were either not part of our RNA data set (glutamate decarboxylase like 1 [GADL1]) or only slightly increased in the IDH1-mut but not in IDH1-WT glioblastoma group (cysteine sulfinic acid decarboxylase [CSAD]). Thus, we cannot fully explain the lack of correlation between ADO/ GADL1/CSAD gene expression and hypotaurine levels. Another finding was the detection of higher cystathionine levels in IDH1-mut astrocytoma ( Figure  3C). This metabolite was recently reported as a magnetic resonance spectroscopy (MRS) marker for the noninvasive detection of 1p/19q codeleted gliomas (22) -e.g., oligodendroglioma (which, per definition, harbor IDH1 mutations). The identification of cystathionine in IDH1-mut astrocytoma in  our study might indicate that this MRS parameter reflects the presence of IDH1 mutation. The corresponding CBS gene expression levels were not statistically significantly altered in our study, probably due to the limited sample size.
The comparison III of IDH1-mut versus IDH1-WT glioblastoma ( Figure 4C) revealed several metabolite differences. Increased glutamate tissue levels in IDH1-WT glioblastoma correlated with several gene expression changes ( Figure 4C), indicating different roles of glutamate in multiple metabolic pathways. GLUD leads to glutamate turnover and is involved in metabolic pathways (nitrogen, alanine, aspartate and glutamate metabolism, arginine biosynthesis) (23). Low GLUD1 levels in IDH1-WT glioblastoma, thus, correlate well with increased glutamate tissue levels. Glutamate has tumor-promoting features in glioma (20). Furthermore, citrate and corresponding ACLY gene expression levels were increased. ACLY is involved in oxaloacetate, acetyl-CoA and citrate turnover and promotes cancer cell proliferation and progression (24,25). These metabolic differences between IDH1-WT (increased glutamate, citrate; Figure 4C) and IDH1-mut glioblastoma (decreased glutamate, citrate; Figure 4C) reflect different clinical outcome of these entities. Another finding included decreased levels of butyrate, a neuroprotective gut metabolite (26), in both IDH1-mut and IDH1-WT glioblastoma (Tables 1 and 2), suggesting further investigations on the role of microbial metabolites. We then investigated histology-dependent comparisons IV-VII in the IDH1 mutation status cohort (Figures 5 and 6 and Supplemental Figure 2 and 3). As outlined, the diagnoses of all different glioma entities in our cohorts from Tübingen and Cologne were based on the WHO classification 2016 (1) and the following parameter for molecular diagnostics: IDH1, ATRX, 1p/19q codeletion, and MGMT. For example, oligodendrogliomas were only diagnosed based on the presence of IDH1 mutation and 1p/19q codeletion. Of note, the upcoming new WHO classification will introduce further entities and refine existing diagnoses (27,28). This is particularly true for astrocytic tumors, e.g., the IDH1-mut glioblastoma (as diagnosed in our cohort based on the current WHO classification) will be replaced by the entity IDH1 mutated astrocytoma grade 4. Since our metabolite comparisons mainly focus on the ongoing parameters in diffuse glioma such as IDH1 mutation and histomorphology, the data portrayed here will still be of a high relevance and importance even after upcoming classification changes by WHO.
The Longitudinal investigation cohort comparisons VIII-XI ( Figure 1B, Figures 7 and 8, and Supplemental Table 2) investigated metabolic alterations during tumor progression either with (treated) or without treatment (untreated) between 2 resections. To overcome sample size limitations in the treated category, we pooled all tumor therapies. We did not detect patient-specific significant alterations in untreated longitudinal comparisons ( Figure 7A). In comparisons IX-XI, the PCA (Figures 7 and  8) revealed several outliers from the main clusters, with glucose, lactate, glycine, scyllo/myo-inositol, hypotaurine, and glutamine being the major drivers of separation. For 1 patient with glioblastoma (T10/ T11) ( Figure 8A), we found a significant increase of hypotaurine and glutamine, both of which have been previously reported as tumor-promoting metabolites (21,29). Of note, PFS and OS of this patient were lower compared with other patients in this comparison. Finally, we matched pairs of oligodendroglioma (comparison XI; Figure 8C and Supplemental Table 2); in the PCA, we identified 1 cluster of 3 patients driven by lactate, which decreased over time (Figure 8C), indicating a reduced Warburg effect. Clinical outcome in 2 of these patients who had an additional increase in scyllo-and myo-inositol concentrations showed increased PFS and OS compared with other patients in this comparison. Even with rather small sample sizes in the comparisons of the Longitudinal investigation cohort, we provide candidate metabolites for the design of prospective studies. Particularly, we provide an evaluation on longitudinal levels of hypotaurine, glutamine, and lactate, and their correlations with clinical outcome might identify novel surrogate markers. However, further investigations in larger prospective cohorts will be necessary to dissect whether these longitudinal patient-specific metabolite alterations are consequences of their treatment rather than tumor progression per se. Our present study suggests candidate tissue metabolites for such a prospective evaluation.
Finally, we investigated the potential utility of tissue metabolites as prognostic markers by correlating metabolite levels at initial diagnosis with PFS and OS ( Figure 9 and Supplemental Figure 4). A recent study (30) identified 9 metabolites that correlated significantly with OS but did not investigate any correlations of metabolites with PFS. We detected 13 metabolites that predicted PFS (Supplemental Figure 4A) and 15 metabolites that correlated with OS (Supplemental Figure 4C). We confirmed previous correlations with OS for myo-inositol, alanine, sn-glycero-3-phosphocholine, glycine, and glutamate. However, we did not detect any associations with OS for choline, creatine, and phosphocholine. In our study, increased acetate levels correlated positively with OS. In addition to these metabolites, we discovered 9 potentially novel metabolites that should be further evaluated as prognostic markers for OS (Figure 9 and Supplemental Figure 4). Our data suggest that systematic prospective assessments of tissue metabolic profiling and their correlation with PFS and OS in prospective large data sets are warranted. They might help to further augment and refine existing integrated diagnoses and prognostic stratification in given glioma subtypes.
Taken together, this in-depth analysis of tissue metabolomics and altered metabolic pathways, supported by RNA-Seq, in diffuse glioma provides insights into metabolite alterations in distinct and clinically relevant subgroups of diffuse glioma. It can serve as a starting point for multiple further experimental investigations and prospective clinical studies to further understand exact underlying molecular mechanisms and modes of action.

Methods
Patients and samples. This study was approved by the institutional ethics committee. We retrospectively identified 73 patients from the University Hospital Tübingen and the University Hospital Cologne, with diffuse glioma WHO grades II-IV with archival high-quality tissue samples, molecular routine diagnostics for IDH1 status, 1p/19q, ATRX, and MGMT gene methylation according to the WHO classification 2016 (1). Furthermore, the diagnosis of oligodendroglioma always included the presence of IDH1 mutation and 1p/19q codeletion. Among these patients, 28 of 73 patients had more than 1 resection, leading to a total of 101 samples. Patients had been treated at the University Hospital Tübingen between 2014 and 2017 (n = 53) and at the University Hospital Cologne (n = 20) between 1999 and 2014. All demographic and clinical characteristics were collected from the electronic clinical database. We defined 2 cohorts (Figure 1): The IDH1 mutation status cohort and the Longitudinal investigation cohort and performed distinct comparisons of tissue metabolites (comparisons I-XI; Figure 1).

JCI Insight
Metabolite and RNA extraction from archival biological samples. Archival tumor samples were available after tissue acquisition according to a standardized routine biobanking procedures. In brief, after removal of tumor tissue from the resection cavity in the operating room, samples were immediately frozen in liquid nitrogen, the tumor content was assessed on a cryo-slide by a neuropathologist, and only tissue samples with sufficient tumor content (>70%-80%) were stored as biological samples in the biobank at -80°C. Even though samples were immediately frozen after tissue removal, it is known that neurosurgical tissue collection might also involve occlusion of blood vessel supply and, thus, also lead to metabolic changes in the context of surgical procedures. We consider this as a natural limitation of investigations of tissue metabolomics. Deep-frozen biopsy sample (50-100 mg) were cryogenically pulverized (cryoPREP, Covaris Inc.) and dried in a lyophilisator overnight (Alpha 2-4, Martin Christ). A target mass of 5 mg dried sample was used for metabolite extraction by focused ultrasonication, and 3 mg of dry weight was used for RNA isolation.
NMR-based metabolomics. A target mass of 5 mg lyophilized glioma powder was transferred into 2 mL adaptive focus acoustics (AFA) glass tubes (milliTUBE, Covaris Inc.) and mixed with 400 μL of ultrapure methanol and 800 μL of MTBE (solvent grade). The tissue was then homogenized in a degassed water bath at 7°C with a focused ultrasonicator (E220evolution, Covaris Inc.). Two consecutive sonication programs with vertical sample movement each lasting 5 minutes were used for a maximum yield of homogenization and metabolite extraction. Afterward, the metabolite suspension was separated into a polar and lipid phase by adding 600 μL of ultrapure water. Then, the glass tubes were centrifuged for 5 minutes at 5000g at 4°C. The polar phase was evaporated to dryness overnight with a vacuum concentrator (Concentrator 5310, Eppendorf). Dried pellets were resuspended with deuterated phosphate buffer (200 mM K 2 HPO 4 , 200 μM NaN 3 , pH 7.4) containing 1 mM of the internal standard TSP (trimethylsilylpropanoic acid). To obtain a maximum dissolution, the plastic tubes were thoroughly vortexed and then centrifuged 5 minutes at 14,000g at 4°C. The clear supernatants were transferred into 5 mm and 1.7 mm NMR tubes and a 96-well rack placed into the cooled (4°C) NMR autosampler. Spectra were recorded on a 600 MHz ultrashielded NMR spectrometer (Avance III HD, Bruker BioSpin) equipped with a 5 mm (patients with 1 sampling time point) and 1.7 mm (patients with 2 sampling time points) triple-resonance ( 1 H, 13 C, 15 N/ 31 P) room temperature probe. For optimum water suppression and shim adjustment, a quick simple proton zero go (ZG) experiment was performed followed by a 2-hour lasting Carr-Purcell Meiboom-Gill (CPMG) experiment in order to suppress residual background signals from macromolecules (time domain, 64,000 points; sweep width, 20 ppm; 1024 scans). The recorded free induction decays (FIDs) were fourier transformed and were spectra properly phase corrected and baseline corrected (Topspin 3.6.1, Bruker Bio-Spin). Metabolite annotation and quantification was performed with Chenomx NMR Suite 8.5 (Chenomx Inc.).
RNA-Seq. RNA-Seq was performed at the Next Generation Sequencing Competence Center Tübingen (NCCT) on an Illumina NovaSeq6000 instrument (Illumina) with a sequencing depth of approximately 25 million clusters per sample.
RNA was purified from lyophilized tissue using the QIAsymphony RNA kit (Qiagen). Tissue was thawed at room temperate, and 800 μL of Buffer RTL Plus was added to the tissue. Homogenization was performed twice for 2 minutes at 25 Hz using the TissueLyser II (Qiagen). Resulting homogenates were centrifuged for 3 minutes at 14,000g and 4°C at maximum speed in a microcentrifuge, and 400 μL of the resulting lysate was loaded on the QIAsymphony SP (Qiagen) using the RNA CT 400 protocol. At the end of the protocol, the RNA was eluted in 35 μL of RNase-free water.
For library preparation, mRNA fraction was enriched using polyA capture from 200 ng of total RNA using the NEBNext Poly(A) mRNA Magnetic Isolation Module (NEB) with the liquid handler Biomek i7 (Beckman). Next, mRNA libraries were prepared using the NEB Next Ultra II Directional RNA Library Prep Kit for Illumina (NEB) according to the manufacturer's instructions. Library molarity was determined by measuring the library size (approximately 400 bp) using the Fragment Analyzer with the High NGS Fragment (1-6000 bp assay; Agilent) and measuring the library con- Read quality of RNA-Seq data in FASTQ files was assessed using ngs-bits (v.2020_06) to identify sequencing cycles with low average quality, adaptor contamination, or repetitive sequences from PCR amplification. Reads were aligned using STAR v2.7.3a to the GRCh37, and alignment quality was analyzed using ngs-bits and visually inspected in the Integrative Genome Viewer (v2.7.2). Normalized read counts for all genes were obtained using Subread (v2.0.0).
Selection and presentation of significant metabolites by NMR spectroscopy and RNA-Seq in comparisons I-XI. Using spectral analysis of 1 H-NMR data, we annotated and quantified a total of 66 metabolites, and 58 showed significant changes as defined by P < 0.05 (we additionally set a fold change threshold of > 1.2 to avoid the display of variations close to background) in at least 1 of the comparisons I-XI (Figure 1, A and B, and Tables 1 and 2). We also displayed metabolites that we detected with a lower threshold for significance -i.e., with P < 0.1 (Tables 1 and 2) for a total of 81 metabolites. This extended set is important because the sample sizes per comparison might have been too low to detect higher significance levels in univariate testing. In the IDH1 mutation status cohort ( Figure 1A and Tables 1 and 2, comparisons I-III), we found a total of 39 significant metabolites and another 25 metabolites with P < 0.1. Additionally, within the IDH1 mutation status cohort, we investigated metabolic tissue profiles with identical IDH1 mutation status and different histology, and we identified a total of 44 significant and additional 22 metabolites with P < 0.1 ( Figure  1A and Tables 1 and 2, comparisons IV-VII).
Finally, in the Longitudinal investigation cohort, we detected a total of 5 significant and 14 more metabolites with P < 0.1 ( Figure 1B and Tables 1 and 2, comparisons VIII-XI).
We used the gene expression data from RNA-Seq for a cross-correlation of metabolite/gene expression changes in the IDH1 mutation status cohort where metabolites of IDH1-mut were compared against IDH1-WT (comparisons I-II).
Statistics. All statistical analyses were performed with the MetaboAnalyst 4.0 toolbox with a threshold of maximum 60% missing values and with JMP 14.2.0 (SAS Institute Inc.; https://www.jmp.com/ en_us/home.html). In order to account for dilution and different tissue weights/cell numbers, we normalized data using the probabilistic quotient normalization (PQN) approach (31). Furthermore, for making metabolites of high magnitude changes comparable, the pareto scaling method was applied (mean-centered and divided by the square root of the SD of each variable), which can result in negative values (32). The preprocessed data were analyzed by univariate classical 2-tailed Student's t test (P < 0.05 and FDR < 0.1 were considered significant), multivariate PCA including combined scores and loadings plot and multivariate PLS-DA including VIP scores where metabolites with VIP scores > 1 represent a significant contribution to the PLS-DA model (VIP scores are calculated as a weighted sum of the squared correlations between the PLS-DA components and the original variable).
ROC analyses were performed to identify robust biomarkers that can distinguish between the effects of IDH1 mutation and histology entity. Top AUC (area under the ROC curve) metabolites were screened for related gene expression changes in the RNA data set for comparisons I-III.
Pathway analysis was performed with the quantitative enrichment analysis module of the MetaboAnalyst 4.0 toolbox. Metabolite abundance changes from the comparisons I-III and IV-VII ( Figure 1A) were matched with the Kyoto Encyclopedia of Genes and Genomes (KEGG) database with 84 metabolite sets based on human pathways.
For correlations of metabolites with PFS and OS, we performed univariate regressions for each metabolite. All metabolites that showed significant correlations (likelihood-ratio test, P < 0.05) for PFS and OS were considered for further analysis. One-stage classification and regression tree (CART) analyses identified the optimal thresholds for each metabolite to dichotomize patients into good and poor survival groups. Kaplan-Meier analyses estimated survival for all subgroups of patients stratified by the identified metabolites and their thresholds. Survival times are presented as median survival with 95% CI.