Metabolic expression profiling stratifies diffuse lower-grade glioma into three distinct tumour subtypes

Abstract

Background

Lower-grade gliomas (LGGs) show highly metabolic heterogeneity and adaptability. To develop effective therapeutic strategies targeting metabolic processes, it is necessary to identify metabolic differences and define metabolic subtypes. Here, we aimed to develop a classification system based on metabolic gene expression profile in LGGs.

Methods

The metabolic gene profile of 402 diffuse LGGs from the Cancer Genome Atlas (TCGA) was used for consensus clustering to determine robust clusters of patients, and the reproducibility of the classification system was evaluated in three Chinese Glioma Genome Atlas (CGGA) cohorts. Then, the metadata set for clinical characteristics, immune infiltration, metabolic signatures and somatic alterations was integrated to characterise the features of each subtype.

Results

We successfully identified and validated three highly distinct metabolic subtypes in LGGs. M2 subtype with upregulated carbohydrate, nucleotide and vitamin metabolism correlated with worse prognosis, whereas M1 subtype with upregulated lipid metabolism and immune infiltration showed better outcome. M3 subtype was associated with low metabolic activities and displayed good prognosis. Three metabolic subtypes correlated with diverse somatic alterations. Finally, we developed and validated a metabolic signature with better performance of prognosis prediction.

Conclusions

Our study provides a new classification based on metabolic gene profile and highlights the metabolic heterogeneity within LGGs.

Background

Diffuse lower-grade gliomas (LGGs) (World Health Organization (WHO) grades II and III) are infiltrative brain tumours that arise from glial or precursor cells and include astrocytomas, oligodendrogliomas and oligoastrocytomas.^1,2 Based on IDH mutation and 1p/19q status, LGGs are classified into three diagnostic and prognostic subtypes in 2016 WHO classification: IDH mutant and 1p/19q co-deleted (oligodendrogliomas), IDH mutant and 1p/19q non-co-deleted and IDH wild-type tumours.³ A subset of these gliomas will progress to glioblastoma (WHO grade IV) within months, whereas others remain stable for years. The survival ranges widely, from 1 to 15 years, and some LGGs have impressive therapeutic sensitivity.⁴ Current treatment varies with the histologic class, grade and the extent of resection, and includes neurosurgical resection, radio- and chemotherapy. Recurrence and tumour progression are inevitable because of their highly invasive nature and treatment resistance.⁵

Metabolic reprogramming has been a well-established hallmark of cancer.⁶ Tumour cells could modify their metabolic programmes to adapt to the energy and macromolecular requirements for uncontrolled cellular growth. The Warburg effect, tumour cells can shift from oxidative to fermentative metabolism, has been recognised for decades.⁷ More recently, it is increasingly clear that cancer cells and tumours have heterogeneous metabolic preferences and dependencies. With the given tumour, some cancer cells are predominantly glycolytic, whereas others display an oxidative phosphorylation metabolic phenotype.^8,9 Moreover, increasing studies find a metabolic symbiosis between glycolytic and oxidative cancer cells. Lactate and pyruvate generated by glycolysis can be transferred and used by neighbour cancer cells as substrates for tricarboxylic acid intermediates.¹⁰ In addition, fatty acids, ketones and glutamine can also be utilised by cancer cells through mitochondrial oxidative phosphorylation or tricarboxylic acid cycle.^11,12 Fully elucidating the metabolic reprogramming that occurs in human cancers will yield new insight into the mechanisms that drive tumour progression and provide potential therapeutic vulnerabilities that target metabolism.

The metabolic complexity and flexibility are commonly observed in gliomas. Glioma cells take up nutrients from the extracellular environment, such as glucose, acetate and glutamine, fatty acids and cholesterol, and use them for energy and biomass production by fuelling glycolysis, oxidative phosphorylation, tricarboxylic acid cycle as well as pentose phosphate pathway.¹³ Currently, it becomes clear that genetic alterations can contribute to metabolic heterogeneity in gliomas. EGFR mutation drives glioblastoma glycolysis through both AKT-dependent and -independent pathways involving mTORC2 and MYC, and also drives fatty acid synthesis through AKT-SREBP1-dependent mechanism.^14–16 Mutant IDH will result in the conversion of a-ketoglutarate (aKG) into the oncometabolite 2HG, which causes aberrant DNA and histone methylation.^17,18 However, most studies usually focus on a specific metabolic perturbation of gliomas and investigate it in isolation. It is urgent to uncover the metabolic profile and identify metabolic subtypes in gliomas, which will facilitate the development of treatments that target metabolism.

In this study, we classified diffuse lower-grade gliomas into three subtypes based on unsupervised clustering of metabolism gene expression profiles. The reproducibility and stability of this classification system was evaluated in three independent cohorts. Each of the three metabolic subtypes was associated with distinct clinical characteristics, metabolic signatures, immune infiltration and somatic alterations. Our findings highlight the metabolic heterogeneity of diffuse LGGs, and the obtained stratification may have clinical implications to drive the development of metabolism-targeted therapies.

Methods

Patients and datasets

This study collected 1176 diffuse lower-grade gliomas from two databases: TCGA and CGGA (Supplementary Table S1). For the TCGA cohort, the RNA-seq data, somatic mutation, copy number alterations (CNAs) and corresponding clinical information were downloaded from the TCGA database (http://cancergenome.nih.gov/).¹⁹ The molecular and immune features were also retrieved.²⁰ For the CGGA cohorts, which consisted of two RNA-seq datasets and a microarray dataset, the RNA-seq and microarray data accompanied with clinical information were obtained from the CGGA database (http://www.cgga.org.cn).^21,22 In addition, RNA-seq data and clinical information of 51 patients with initial and recurrent tumour samples (Supplementary Table S1) were gathered from glioma longitudinal analysis (GLASS) consortium (http://synapse.org/glass).²³ Two public single-cell RNA sequencing (scRNA-seq) datasets (GSE70630 and GSE89567) were collected.^24,25 This study was carried out in accordance with the Helsinki declaration and approved by the ethics committee of Tiantan hospital, and patient informed consents existed in these four public databases.

Identification and validation of metabolic subtypes

A previously published list of 2752 metabolism-related genes was used for subsequent clustering.^26,27 First, Cox regression analysis was performed to identify genes associated with overall survival (OS) in the TCGA cohort. Then, the candidate genes with high median absolute deviation (MAD) value (MAD ≥ 0.5) across all LGG patients were selected for consensus clustering (R package “consensusClusterplus”) to identify robust clusters.^28,29 The cumulative distribution function (CDF) and consensus heatmap were adopted to assess the optimal K.

To validate the metabolic subtypes in CGGA cohorts, we first trained a partition around medoids (PAM) classifier in the TCGA cohort. Then, the classifier was used to predict metabolic subtype for patients in the CGGA cohorts with R package “pamr”. Each sample was assigned to a metabolic subtype based on the Pearson correlation with the centroid.^22,30 The in-group-proportion (IGP) statistic with R package “clusterRepro” was performed to assess the reproducibility and similarity of the acquired metabolic subtypes between the training and validation cohorts.³¹

Estimation of immune infiltration

ESTIMATE was adopted to evaluate the immune cell infiltration (immune score) and stromal content (stromal score) for each sample.³² CIBERSORT algorithm was used to estimate the relative fraction of 22 immune cell types.³³ Single-sample gene-set enrichment analysis (ssGSEA) with R package “GSVA” was performed to quantify the enrichment levels of 23 immune signatures.^34,35

Computation of metabolism-relevant gene signatures

The 115 metabolism-relevant gene signatures were achieved from previously published studies.^27,36 By using R package “GSVA”, each sample received 115 scores corresponding to the metabolism-relevant signatures.

Identification of a metabolism-related signature

We first applied significance analysis of microarray (R package “samr”) to identify differentially expressed metabolism genes between subtypes. With the candidate genes, a Cox proportional hazards model (R package “glmnet”), which was suitable for high-dimensional regression analysis, was performed to identify an optimal gene set with prognostic significance.³⁷ Then, the regression coefficients were used to calculate the risk score for each sample.

Bioinformatic analysis

Gene ontology (GO) analysis was applied for functional annotation of differential genes between subtypes.³⁸ Gene set enrichment analysis (GSEA) was performed to determine gene sets with a statistical difference.³⁹ Principal components analysis (PCA) (R package “princomp”) was conducted to detect expression difference between subtypes.⁴⁰ Receiver-operating characteristic (ROC) curve analysis (R package “pROC”) was used for overall survival (OS) prediction. GISTIC2.0 analysis was conducted to assess CNA difference between subtypes, and locus with GISTIC value more than 1 or less than −1 was defined as amplification or deletion, respectively.⁴¹

Statistical analysis

All statistical analyses were conducted using R software, GraphPad Prism 6.0 (GraphPad Inc., San Diego, CA, USA), or SPSS 16.0 (IBM, Chicago, IL, USA). Kaplan–Meier analysis with log-rank test was performed to assess survival difference between subtypes. Chi-square test was carried out to determine the difference of clinical and molecular features between subtypes. For comparisons of three groups, one-way analysis and ANOVA test of variance were used. Univariate and multivariate Cox regression analyses were performed to determine factors with prognostic value. P < 0.05 was considered statistically significant.

Results

Consensus clustering identifies three subtypes in diffuse lower-grade gliomas

In order to characterise metabolic heterogeneity within diffuse LGGs, previously reported 2752 metabolism-relevant genes were used for clustering analysis.²⁶ The workflow is shown in Fig. 1a, and clinical characteristics of patients of four cohorts were listed in Supplementary Table S1. We first performed Cox regression analysis to identify genes associated with OS, and a total of 1233 candidate genes were retained for following clustering analysis in the TCGA cohort. By applying consensus clustering on the gene expression profile of candidate genes, three resulting clusters, M1, M2 and M3, were defined (Fig. 1b and Supplementary Fig. S1). To validate the assignments of subtypes, we performed PCA and further confirmed the robust difference in expression portraits within these three metabolic subtypes (Fig. 1c). Of note, a significant prognostic difference was observed among subtypes (P < 0.001), with shorter overall survival for M2 than M1 and M3 (Fig. 1d).

Fig. 1. Identification of distinct metabolic subtypes in diffuse lower-grade glioma through metabolism gene profiling.

a Flow chart shows the computational method to classify tumour samples into metabolic subtypes. TCGA cohort was used as a discovery set, three CGGA cohorts were collected as validation sets. MAD, median absolute deviation. b Heatmaps of three metabolic subtypes defined in four cohorts. The metabolic subtypes of CGGA cohorts were predicted by a partition around medoid (PAM) classifier derived from the TCGA cohort. Patients were arranged based on the predicted metabolic subtypes. c Principal component analysis (PCA) of three metabolic subtypes using candidate genes. d Survival analyses show significant differences between the three metabolic subtypes in four LGG cohorts. P value was calculated by the log-rank test among subtypes.

To validate our findings in the TCGA cohort, we evaluated the reproducibility of metabolic subtypes in three CGGA cohorts. The centroid of each metabolic subtype was calculated, and each sample in CGGA cohorts was assigned to a subtype according to the Pearson correlation of centroid³⁰ (Fig. 1b). In-group-proportion (IGP) statistical analysis showed high consistency between subtypes of TCGA and CGGA cohorts (Supplementary Table S2). Furthermore, the obtained metabolic subtypes displayed a similar pattern of expression and prognostic features with the TCGA cohort (Fig. 1c, d).

Correlation of the metabolic subtypes with clinical characteristics in TCGA and CGGA cohorts

To assess the clinical relevance of the metabolic expression subtypes identified above, we next determine correlations with molecular and pathological features. The statistical analysis (Chi-square test) revealed that histologic grade III and IDH wild type were associated with M2 subtype, pathological-type oligodendroglioma and 1p/19q co-deletion were associated with M1 subtype (P < 0.001) (Fig. 2a and Supplementary Table S3). We also compared our metabolic classification with previously reported TCGA transcriptome and methylation molecular subclasses.^19,42 In the TCGA cohort, M2 subtype was significantly associated with transcriptome clusters CL/ME, and methylation clusters LGm1/LGm4/LGm5 (P < 0.001). M1 was linked to transcriptome cluster NE and LGm3, whereas M3 was associated with LGm2 (P < 0.001) (Fig. 2a). Similarly, the correlations of metabolic subtypes with clinicopathological features were also observed in three CGGA cohorts (Fig. 2b, Supplementary Fig. S2 and Supplementary Tables S4–6).

Fig. 2. Clinical characteristics of metabolic subtypes in TCGA and CGGA cohorts.

a Correlation of our classification with clinical characteristics and previous LGG subclasses in the TCGA cohort. b Correlation of metabolic subtypes with clinical characteristic and previous LGG subclasses in CGGA cohort2. ME mesenchymal, NE neural, CL classical, PN proneural. Chi-square test was used for statistical analysis.

Transcriptome analysis of the metabolic subtypes

To further characterise the three metabolic subtypes, gene expression differences were enquired with R package “samr”. Differentially expressed genes (DEGs) (FDR < 0.05) of each subtype were used for functional enrichment analysis. GO analyses revealed that the specific DEGs of M1 were mainly enriched in chemical synaptic transmission, neurotransmitter secretion, glutamate secretion and nervous system development (Supplementary Fig. S3A). Instead, the upregulated genes of M2 were annotated to cell division, mitotic nuclear division, immune response, T-cell co-stimulation, cell–cell adhesion and DNA replication (Supplementary Fig. S3B). For M3, it was significantly enriched in translation initiation, rRNA processing, regulation of transcription, inflammatory response and translation (Supplementary Fig. S3C). Besides, GSEA was also conducted to confirm the functional enrichments of each subtype, and similar results were observed (Supplementary Fig. S3D–F).

Immune infiltration of metabolic subtypes in diffuse LGGs

Due to the significant difference in immune-related enrichments among subtypes, we further resorted to immune-related tools published previously to decipher the immune infiltration of metabolic subtypes. We first computed stromal and immune scores based on the ESTIMATE method.³² Compared with M1 and M3, M2 subtype showed higher immune and stromal scores but lower purity (Fig. 3a). CIBERSORT analysis³³ revealed that M1 had an increased percentage of lymphocytes, whereas M2 and M3 displayed a higher level of M2 macrophages (Fig. 3b and Supplementary Table S7). Moreover, ssGSEA scores³⁵ were applied to quantify the enrichment levels of immune cells and functions. Most immune cells tended to display an increased level in M1 and M3 subtypes, such as Tfh, DCs, Th2, Th1, B cells, neutrophils and Treg. M2 was associated with higher levels of CD8 + T cells, cytolytic activity, as well as elevated inflammation-promoting and T-cell co-stimulation, M1 showed higher levels of Th17, CD8 + T cells and mast cells, whereas M3 had relatively higher levels of T-helper cells, macrophages, along with increased functions of APC and T-cell co-inhibition (Fig. 3C). To validate these findings, we next sought to dissect the immune infiltration of each subtype in three CGGA cohorts and obtained consistent results (Fig. 3a–c, Supplementary Fig. S4 and Supplementary Table S7).

Fig. 3. Immune infiltration of three metabolic subtypes in TCGA and CGGA cohorts.

a Violin plots of immune, stromal and tumour purity scores from ESTMATE for different metabolic subtypes in the TCGA cohort and CGGA cohort2 (ANOVA test). b Violin plots show the lymphocytes and macrophage (M2) proportion from CIBERSORT for the three metabolic subtypes in the TCGA cohort and CGGA cohort2 (ANOVA test). Lymphocytes contain B cells, CD4 T cells, CD8 T cells, follicular helper T cells, Tregs, gamma delta T cells, NK cells and plasma cells. The middle bar represents the median value. ***P < 0.001, ****P < 0.0001. c Heatmaps show differential enrichments of immune-related signatures in three metabolic subtypes. ANOVA test was used for statistical analysis, and the P values were labelled.

Recurrent gliomas with altered metabolic subtype exhibit corresponding changes in immune infiltration

For further validating the association between metabolic subtype and immune microenvironment, we collected RNA-seq data of 51 glioma patients with initial and recurrent samples from the GLASS cohort.²³ With the same method, each sample was assigned into a metabolic subtype (Supplementary Fig. S5A). Samples in M1 were mainly recurrent and belonged to transcriptome cluster PN, whereas M2 was associated with histologic grade IV, IDH wild-type and transcriptome clusters CL/ME. PCA further confirmed the difference in expression profile (Supplementary Fig. S5B). As expected, patients in the M2 subtype showed worse outcome (Supplementary Fig. S5C). Then, the immune infiltration of each metabolic subtype was inquired and assessed with the tools mentioned above. As well as increased cytolytic activity, the M2 subtype displayed higher immune and stromal scores but lower purity. In contrast, M1 showed a higher percentage of lymphocytes and purity, but lower immune and stromal scores (Supplementary Fig. S5D, E), which exhibited high consistency with our previous findings.

After recurrence, we found 21 patients with altered metabolic subtype (Supplementary Fig. S5F). Considering the distinct difference between M1 and M2 subtypes in immune infiltration, we focused on four patients who turned from the M2 subtype into the M1 subtype. Compared with initial tumours, recurrent tumours of these four patients had an increased level of purity and lymphocytes, but decreased immune, stromal scores and M2 macrophage percentage (Supplementary Fig. S5G). These results indicated that metabolic alterations might shape the immune microenvironment in glioma.

Metabolic characteristics of acquired subtypes in diffuse LGGs

To further explore the metabolic characteristics of each subtype, 115 metabolism-relevant gene signatures were achieved from previously published studies,^27,36 and GSVA was conducted to quantify the enrichment levels of metabolism processes. Differential analyses (Supplementary Table S8) revealed that higher levels of metabolism processes in M1 subtype were mainly lipid and amino acid, whereas M2 displayed higher levels of metabolism processes that were related to carbohydrate, nucleotide and vitamin. Of note, oxidative phosphorylation, citric acid cycle, gluconeogenesis and pyruvate metabolism were relatively enriched in M1. M3 subtype, by contrast, showed few enrichments of metabolism signatures, which implied low metabolic activities (Fig. 4). We further interrogated three CGGA cohorts and computed these metabolic scores, similar difference was observed (Supplementary Figs. S6–8 and Supplementary Table S8).

Fig. 4. Association between metabolism-relevant signatures and LGG subtypes.

Heatmaps of differential enrichments of metabolism-related signatures in the TCGA cohort. Lipid, carbohydrate, amino acid, nucleotide and vitamin metabolism signatures were presented. The statistical difference was compared through the ANOVA test, and the P value <0.05 was considered significant.

Metabolic subtypes are associated with diverse somatic variations

Since genetic aberrations are associated with metabolism reprogramming in gliomas,^14–17 we further identified somatic variations that potentially drive metabolic expression subtypes. Somatic mutation data from the TCGA cohort were enquired and the correlations of metabolic subtypes with potential somatic drivers were examined. The genes with high mutation frequency or in critical pathways are shown in Fig. 5a and the statistical analysis was performed with the Fisher test (Supplementary Table S9). M1 was enriched in 1p/19q co-deletion, IDH1, CIC and FUBP1 mutations. M2 was enriched in mutations in driver genes, such as PTEN, EGFR and NF1. M3 was enriched in IDH1, TP53 and ATRX mutations (Fig. 5a). Similarly, we explored the correlations of subtypes with copy number variations (CNAs) and GISTIC2.0 analysis found distinct CNAs among metabolic subtypes. M2 displayed more frequently deleted or amplified regions, such as PTEN, EGFR, MET and CDKN2A/B (Fig. 5a). In addition, M2 tumours had higher aneuploidy, copy number burden and homologous recombination deficiency scores, as well as increased tumour mutation burden (Fig. 5b).

Fig. 5. Somatic variations associated with metabolic expression subtypes in the TCGA cohort.

a Oncoprint of mutation status and copy number variations within metabolic subtypes, Fisher test was adopted for comparison analysis. b Violin plots describing the DNA damage measures within metabolic subtypes of the TCGA cohort (ANOVA test). The middle bar represents the median level of the corresponding features. ****P < 0.0001.

Identification of a metabolic signature with prognostic significance

Based on the metabolic expression classification we developed, we next built a prognostic signature with Cox proportional hazards model (R package “glmnet”).³⁷ First, SAM analysis identified 1118 differentially expressed metabolism genes between M2 and M1/M2 subtype in the TCGA cohort, wherein 1055 genes were available in three CGGA cohorts. Then, Cox proportional hazards modelling was performed for selecting a gene set with best prognostic value (Fig. 6a). Subsequently, a 14-gene signature was obtained and scores were calculated with regression coefficients (Fig. 6b, c). Kaplan–Meier analysis revealed that patients with high scores had significantly shorter overall survival (OS) (Fig. 6d). High scores were enriched in M2 subtype, grade III or IDH wild-type tumours (Fig. 6e). Besides, the developed signature had higher AUC compared with factor of age (Fig. 6f). Furthermore, multivariate Cox regression analysis also revealed that the metabolic signature was an independent prognostic factor for diffuse LGGs (Supplementary Table S10). We next validated this signature in three CGGA cohorts and obtained consistent results (Supplementary Fig. S9 and Supplementary Tables S11–13), which implied the superior performance of metabolic signature for prognosis prediction.

Fig. 6. Identification of a metabolic signature associated with overall survival by Cox proportional hazards model.

a Cross-validation for tuning parameter selection in the proportional hazards model. b Heatmap showing the expression levels of 14 signature genes. Patients were arranged by metabolic clusters. c Coefficient values of each signature gene. d Kaplan–Meier survival analysis of the metabolic signature in patients of LGGs. P value was calculated by the log-rank test. e Comparison of the metabolic scores in patients stratified by metabolic subtype, grade and WHO subtype (ANOVA test). Sub1 stands for IDH mutant and 1p/19q co-deleted tumours, Sub2 for IDH mutant and 1p/19q non-co-deleted tumours, Sub3 for IDH wild-type tumours. ****P < 0.0001. f ROC curve analyses show the association between OS and age, metabolic signature. AUC area under the curve.

We further determined whether the metabolic signature was derived from tumour or stromal cells. Two scRNA-seq datasets were obtained from the GEO database.^24,25 High expression of gene sets corresponding to markers of particular cell types classified cells as tumour cells (PTPRZ1), macrophages (CSF1R), oligodendrocytes (MOBP) and T cells (CD3D) (Supplementary Fig. S10A–D). Then, we calculated the metabolic score of each cell using “GSVA” package, and found that the distribution of metabolic scores showed nonspecific among tumour and stromal cells (Supplementary Fig. S10E, F). Dotplots further displayed the expression levels of 14 signature genes in different cells. Tumour cells expressed higher levels of genes COX15, SULF2, EXTL3 and PCCA. CLIC1, HEXB and MTAP were enriched in macrophages. SLC25A28 showed high level in oligodendrocytes (Supplementary Fig. S10G). These results indicated that our metabolic signature was not specific to a certain population, but derived from whole tumour and stromal cells.

Discussion

Given the high heterogeneity in glioma, previous studies stratified glioma patients through unsupervised clustering of tumours based on genomic and transcriptomics data, which led to the valuable discovery of many patient group-specific differences.^{4,19,22,42,43} Recently, metabolite profiling has become an informative approach to elucidate tumour heterogeneity. Peng et al. analysed a cohort of 9125 TCGA samples across 33 cancer types to characterise tumour subtypes based on the expression of seven metabolic pathways.⁴⁴ Daemen et al. conducted broad metabolite profiling and identified three subtypes that showed distinct metabolite profiles associated with glycolysis, lipogenesis and redox pathways.⁴⁵ Bidkhori et al. identified three HCC subtypes with distinct differences in metabolic and signalling pathways and clinical survival.⁴⁶ In this study, we established a new classification of diffuse lower-grade gliomas based on 2752 metabolic genes previously reported. Three metabolic subtypes were identified, and clinical characteristics, immune infiltration, metabolic signatures and somatic variations were explored. Our findings extended the molecular subtyping of diffuse LGGs and deepened our understanding of metabolic heterogeneity within this tumour.

Growing studies suggest that metabolic heterogeneity within the tumour environment influences local immune cell function and might contribute to immunotherapy failures. For example, lactate derived from cancer cells suppresses T-cell and NC-cell function and inhibits monocyte activation and differentiation of the dendritic cell.^47,48 Excessive consumption of nutrients (e.g., glucose and glutamine) by cancer cells will suppress T-cell activation, inhibiting glucose and glutamine metabolism impairing T-cell proliferation and function.^49,50 Here, we performed metabolism enrichment and immune infiltration analyses, and revealed distinct immune states among metabolic subtypes. Consistently, M1 and M3 subtypes, which showed a higher level of d-glutamine and d-glutamate metabolism, were associated with higher enrichment of Tfh, Th2, Th1 and Treg cells. In contrast, M2 subtype had more carbohydrate metabolism enrichments that displayed low T-cell activation (Figs. 3 and 4). Furthermore, we compared the metabolic classification with the immune subclasses we identified previously.²² The Sankey diagrams showed high concordance between these two classifications (Supplementary Fig. S11). These results demonstrated a close connection between metabolism reprogramming and immune phenotype.

When we associated metabolic subtype with WHO 2016 categorisation in the TCGA cohort, 97.8% (134/137) of IDH mutant and 1p/19q co-deleted (Sub1) tumours were enriched in M1 subtype. In all 89% (65/73) of IDH wild-type LGGs (Sub3) were enriched in M2 subtype, whereas only 64.6% (124/192) IDH mutant and 1p/19q non-co-deleted (Sub2) tumours were assigned into M3 subtype (Supplementary Fig. S12A). Three CGGA cohorts were also enquired and similar results were observed (Supplementary Fig. S12A). Supplementary Fig. S12B showed the survival curves of TCGA and CGGA cohorts based on WHO 2016 classification. We further applied our metabolic classification into Sub2 and Sub3 subtypes. In CGGA cohorts, our metabolic classification could distinguish the overall survival of Sub2 and Sub3 tumours, and M2 had worse outcome (Supplementary Fig. S12C, D). These results indicated that IDH mutant astrocytomas and IDH wild-type/GBM-like tumours were a mix of M1, M2 and M3 metabolic subtypes, except for oligodendrogliomas. These tumours could be further divided into groups with distinct prognosis based on our metabolic classification. In addition, we looked into 1p and 19q genes and differentially expressed genes (DEGs) between M1 and M2&M3, and only 146 genes were shared. Moreover, we detected the shared genes between centroid genes and 1p/19q genes, and only 107 genes were observed (Supplementary Fig. S13A, B). These results demonstrated that the metabolic classification was independent of 1p/19q genes. Our findings further broadened our understanding of molecular subtyping of diffuse lower-grade gliomas.

It becomes clear that metabolism subtyping is significantly correlated with clinical outcome and has potential prognostic value. Notably, the upregulated subtype of carbohydrate, nucleotide and vitamin metabolism is associated with poor prognosis, whereas upregulated TCA and lipid metabolic subtype implies better outcome.^44,51 Consistently, M1 subtype, which showed high levels of lipid metabolism and citric acid cycle, was associated with longer overall survival. M2, upregulated subtype of carbohydrate, nucleotide and vitamin metabolism, had poor outcome. Besides, we also determined the correlation between metabolic signatures and clinical outcome in diffuse LGG. Out of these metabolic signatures, pyrimidine biosynthesis and carbohydrate metabolisms, such as N-glycan biosynthesis, glycosaminoglycan degradation, galactose and cyclooxygenase arachidonic acid metabolism, were associated with poor prognosis, whereas improved survival was observed in patients with high scores of glutamine and lipid metabolism, including cholesterol biosynthesis, cardiolipin, propanoate, glyoxylate and dicarboxylate metabolism (Supplementary Fig. S14 and Supplementary Table S14). These results further demonstrated and highlighted the prognostic relevance of metabolic expression subtype.

As analysis of cancer metabolism and genetics merged over the past two decades, metabolic reprogramming can be largely viewed as a consequence of oncogenic driver events.^51,52 Frequent amplification of growth factor receptor tyrosine kinase-encoding genes and IDH mutations has emerged as critical player in altered glioma metabolism.¹³ Mutant EGFR drives GBM glycolysis through both AKT-dependent and -independent pathways involving mTORC2 and MYC, and ERK-dependent nuclear translocation of phosphorylated PKM2.^14,15,53 In IDH mutant glioma cells, the oncometabolite 2HG inhibits the function of BCAT1 and BCAT2 and makes cells more dependent on glutaminase-driven glutaminolysis, suggesting glutamine as a primary source of tumours with IDH mutation.^54,55 Similarly, M2 subtype with enriched EGFR mutation and amplification was associated with a high level of carbohydrate metabolism. M1 and M3 subtypes with an increased level of d-glutamine and d-glutamate metabolism were enriched in IDH1 mutation.

In summary, this study classified diffuse lower-grade gliomas from the metabolic perspective and proposed three subtypes with distinct prognosis, immune infiltration, somatic variation and metabolic phenotype. Our results demonstrated the metabolic heterogeneity of diffuse lower-grade glioma and provided valuable stratification for the design of metabolism-targeted therapies.

Author contributions

W.Z., Z.Z. and W.M.: conceptualisation and supervision; F.W., Y.L. and G.L.: methodology, data curation and writing—original draft preparation; Y.Z. and Y.F.: data collection, software and writing—reviewing and editing.

Ethics approval and consent to participate

This study was carried out in accordance with the Helsinki declaration and approved by the ethics committee of Tiantan hospital, and patient informed consents existed in these two public databases.

Consent to publish

Not applicable.

Data availability

All data in this study are available in TCGA, GLASS, GEO and CGGA datasets.

Competing interests

The authors declare no competing interests.

Funding information

This work was supported by National Natural Science Foundation of China (81672479, 82002994, 82002647 and 81702460) and Natural Science Foundation of Anhui Province (1908085QH335).

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-021-01418-6.