Integrative multi-omics analysis proposes a metabolic classification of gliomas: distinct metabolic states, immune infiltration, and prognosis

Abstract

Background

The tumor microenvironment (TME) of glioma harbors diverse cell types; however, cell metabolic heterogeneity remains to be explored. This study aims to characterize the metabolic features of different cell types in the TME by integrating multiple datasets, including genomics, bulk and single-cell transcriptomics, and metabolomics.

Methods

Unsupervised machine learning was used to construct an energy metabolic classifier based on the metabolic pathways identified from bulk RNA-seq of gliomas in the TCGA dataset. The classifier was externally validated using multiple datasets, including genomics, bulk RNA-seq, snRNA-seq, and the metabolomics data. Furthermore, metabolic heterogeneity associated with the classifier was further characterized at single-cell resolution.

Results

The energy metabolism-based classifier stratified patients into two prognostic clusters: patients in cluster 1 were characterized by high pathway activity of glycolysis, the pentose phosphate pathway (PPP), and fatty acid oxidation (FAO), whereas patients in cluster 2 exhibited higher activity in glutaminolysis. This metabolic classifier revealed both intratumoral and intertumoral metabolic heterogeneity, and the complexity was further validated by the metabolomics profiling and snRNA-seq data from the CPTAC dataset. Notably, OSMR, highly expressed in cluster 1, showed significant co-expression with key glycolytic enzyme genes. The OSM/OSMR/JAK1/STAT3 axis potently drives malignant progression of glioma cells, specially enhancing their invasive and migratory capabilities. Single-cell resolution analyses demonstrated that tumor metabolic heterogeneity is primarily driven by malignant cells rather than non-malignant components, while tumor microenvironment (TME) factors were also found to modulate malignant cell metabolism. Significantly, glycolytic activity in glioma cells increased during the phenotypic transition from PN (proneural) to MES (mesenchymal), with cluster 1 metabolic phenotypes predominating in the tumor core. Compared to cluster 2, cluster 1 patients exhibited higher mRNA expression of immunosuppressive checkpoint genes, which correlated with pronounced immunosuppression in the TME. Furthermore, various immune cells demonstrated distinct metabolic preferences at single-cell resolution.

Conclusions

This study developed an energy metabolic-based classifier for gliomas with prognostic and therapeutic potential. Metabolic reprogramming was linked with the PN-to-MES transition of glioma cells and immunosuppression in the tumor microenvironment. Multi-omics data, especially snRNA-seq, offered insights into metabolism heterogeneity at single-cell resolution, enabling personalized treatment strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-025-07602-z.

Introduction

Gliomas are the most common and aggressive primary intracranial malignancies, characterized by poor prognosis despite multimodal treatment involving surgery, chemotherapy and radiotherapy. Recent advances in molecular research and the rapid development of single-cell sequencing have uncovered remarkable heterogeneity within the tumor microenvironment [1, 2]. Building upon the transcriptional classification established by Verhaak et al. [3], Neftel et al. [2] further proposed distinct glioblastoma cell subtypes using single-cell sequencing. Different types of glioblastoma cells exhibit variations in their genetic, epigenetic, and microenvironmental factors, endowing them with distinct capabilities for proliferation and invasion. Energy metabolism reprogramming is now recognized as a hallmark of cancer [4], enabling tumor cells to adopt their metabolic pathways to fuel proliferation and progression [5]. In clinical specimens, Melin et al. has reported distinct metabolic phenotypes exist according to adult glioma subtypes [6]. Recent studies suggest that the plasticity and flexibility of tumor cell metabolism offer potential therapeutic opportunities [7–9].

Single-cell sequencing enables precise characterization of diverse cell types within the tumor microenvironment, including clonal diversity, metabolic profiles, and intercellular interactions [10]. Previous studies have reported that ovarian and hepatocellular tumor cells can be categorized into distinct metabolic clusters to predict the patient’s treatment response and overall prognosis [11, 12]. However, the mechanisms of glioma cell metabolism remain poorly understood, and a comprehensive metabolic classifier for gliomas is still lacking.

Immunotherapy has become a crucial component of cancer treatment, making it increasingly important to understand the metabolic heterogeneity of infiltrating immune cells and tumor cells in depth. An increasing numbers of studies have demonstrated that tumor metabolism not only plays a vital role in promoting the development and survival of tumor cells, but can also modulate anti-tumor immune responses by releasing metabolites that influence immune molecule expression, such as lactate, PGE2, and arginine [13, 14]. In glioblastoma, Veglia et al. revealed that PERK-driven glucose metabolism in monocyte-derived macrophages (MDM) promotes MDM immunosuppressive activity via histone lactylation, indicating that immune cell metabolism modulates their anti-tumor activity [15]. Additionally, tumor cells can inhibit anti-tumor immune responses by competing for and consuming vital nutrients, thereby impairing the metabolic fitness of tumor-infiltrating immune cells [16]. Understanding the metabolic heterogeneity and characteristics of immune cells in the tumor microenvironment facilitates the identification of therapeutic targets within metabolic pathways. This approach holds promise for enhancing cancer immunotherapies.

This study establishes an energy metabolism-based classifier for glioma using bulk RNA-seq, and validated it using multi-omics data, including genomics, bulk and single-cell transcriptomics, and metabolomics. Single-cell analysis characterized metabolic reprogramming during the PN-to-MES transition in glioma cells. Additionally, it elucidated the metabolic preferences of various immune cell types. Overall, by leveraging multi-omics data, this study uncovered glioma metabolic heterogeneity, providing novel insights and potential therapeutic targets.

Methods

Data accessibility

This study utilized several glioma datasets, including the TCGA dataset, the CGGA dataset, the CPTAC dataset [17], the IvyGAP dataset [18], and the GSE182109 dataset [19]. These datasets contain genomic data, bulk RNA-seq, single-cell or single-nucleus RNA-seq, and metabolome data. This study included both IDH-mutant and IDH-wild glioma patients, allowing for a comprehensive analysis across different molecular subtypes. The GSE182109 dataset and IvyGAP datasets provide transcriptome data from different anatomical regions within the same patients. In the CPTAC database, 16 patients were subjected to both bulk RNA-seq and single-nucleus RNA seq. The sequencing data are available for download from the following websites (https://portal.gdc.cancer.gov/, https://glioblastoma.alleninstitute.org, https://www.ncbi.nlm.nih.gov/geo/).

Construction of energy metabolic classifiers

Based on previous studies [20, 21] and MSigDB [22], we identified four critical metabolic pathways in gliomas: glycolysis, the pentose phosphate pathway, fatty acid oxidation, and glutaminolysis. We identified key genes associated with these metabolic pathways (Table S1). To assess the activity of these pathways in individual patient, we utilized the ssGSEA algorithm (GSVA package in R) [23].

Unsupervised classification is a key machine learning approach. In this study, we used the K-means method to ascertain the optimal classification based on glioma metabolic profiles. Prior to clustering, samples were scaled to group them based on metabolic pathway patterns. We then reordered the samples based on metabolic clusters and the scaled ssGSEA values for heatmap visualization (using the Pheatmap package in R). Additionally, we applied the consensus clustering algorithm (ConsensusClusterPlus package in R) with 1000 iterations and 80% resampling to confirm the number of clusters (K). Principal component analysis (PCA) was also performed to highlight significant differences between clusters.

Calculation of immune signature scores

To evaluate cell proliferation, we utilized 31 validated genes associated with cell cycle progression (CCP) [24]. Immune infiltration levels were assessed by several immune signatures related to tumor-infiltrating lymphocytes (TILs), immune cytolytic activity (CYT), and interferon (IFN) responses [24]. ssGSEA scores were calculated to quantify the abundance of these gene sets. Tumor immune scores were estimated using the ESTIMATE algorithm [25]. The Microenvironment Cell Populations-Counter (MCP-counter) algorithm was used to quantify the populations of eight immune cell types and two stromal cell types [26]. The cluster score was computed by averaging the expression levels of all pathways within each cluster and subtracting the mean expression of four metabolic pathways. During this calculation, the values for each metabolic pathway were scaled across all samples. For scRNA-seq data from the CPTAC and GSE182109 datasets, tumor cells were classified into metabolic cluster 1 if their metabolic cluster 1 score exceeded that of metabolic cluster 2, and vice versa.

Preprocessing of scRNA-seq data

The preprocessing of the raw data for the CPTAC and GSE182109 datasets was described in detail in the publications [17, 19]. The Seurat package v5.0 [1] was used for downstream scRNA analysis. A series of quality filters were performed to remove the low-quality cells: too few total transcript counts (<300); possible debris with too few genes detected (<200) and too few UMIs (<1000); possible dead cells or sign of cellular stress and apoptosis with too high a proportion of mitochondrial gene expression over the total transcript count (>10%). Each sample was normalized and scaled by ‘SCTransform’ function in the Seurat package to correct for batch effects (parameters: vars.to.regress = c (“nCount_RNA”, “percent.mito”), variable.features.n = 3000). The Harmony algorithm was applied to integrate all samples for downstream analysis [27]. To reduce dimensions, principal component analysis (PCA) was initially conducted with 30 principal components (npcs = 30), followed by t-distributed stochastic neighbor embedding (t-SNE) using the first 20 dimensions (dims = 1:20). The Wilcoxon rank-sum test was used to identify the specific genes of each cluster through the FindAllMarkers function (logfc.threshold = 0.25).

Cell annotation

Using singleR [28], cellmarker [29], panglaodb [30] and marker genes reported in the literature [17, 31], all cell types were annotated through a combination of automated and manual methods.

Metabolic activity evaluation

The metabolic pathway activity of annotated cells was calculated as described in the previous article [32]. ssGSEA values were used to compare metabolic profiles across various cell types. Metabolic pathways with a P-value < 0.05 in GSEA were considered statistically significant. The OXPHOS gene set was extracted from the KEGG database. The gene sets of hypoxia and angiogenesis were retrieved from the MSigDB library.

Inference of CNV

To differentiate malignant cells from non-malignant cells, copy number variation (CNV) was calculated using the R package inferCNV (https://github.com/broadinstitute/inferCNV). Genes expressed in fewer than 10 cells and with a median expression below 0.1 were excluded from the analysis to focus on robustly expressed genes. Subsequently, genes were annotated based on their chromosomal positions, and CNV scores were calculated using moving averages across sets of 100 genes. Hierarchical clustering was applied to separate non-malignant cells from malignant cells based on distinct chromosomal deletions or amplifications.

Trajectory analysis

Trajectory analysis was facilitated by importing data from Seurat into the Monocle2 package in R [33]. Initially, genes with expression levels below 0.5 and cells expressing fewer than 200 genes were excluded from the dataset. The remaining data were reordered based on significantly differentially expressed genes identified with a q-value threshold of 1e-40. Subsequently, a DDRTree dimensional reduction algorithm was applied to interpret the data structure. Pseudotime analysis was then performed to map the temporal progression of cellular differentiation, pinpointing pathways that significantly changed along trajectories. Cells within the same branch were considered to reside at similar differentiation stages.

Glioma cell line and cell culture

The glioma cell lines of U87, LN229 and GL261 were purchased from ATCC and authenticated by STR assay. The patient-derived xenograft cell lines GBM1 and GBM2 were kindly provided by Professor XiuPing Zhou from the Institute of Nervous System Diseases, Xuzhou Medical University. The glioma cells were cultured in DMEM supplemented with 10% fetal bovine serum (FBS) and 1% antibiotics in a humidified incubator maintained at 37 °C with 5% CO₂. For transient knockdown experiments, siRNA was transfected into glioma cells using lipofectamine 8000 (Beyotime Biotechnology, China), followed by further experiments. The plasmids were obtained from Nanjing Corues Biotechnology Co, Ltd. The detail information of plasmids is listed in Table S2. The total RNA and protein were extracted 48 hours and 72 hours after transfection respectively.

Real-time quantitative PCR (RT-qPCR) analysis

Total RNA was extracted from glioma cell lines using TRIzol reagent, and reverse transcription was carried out using Roche reagent. cDNA was amplified by RT-PCR using Roche reagents. Genes expression levels were normalized to GAPDH expression. Primers used for target genes amplification are listed in Table S3.

Western blot

Total protein was extracted from glioma cells using pre-chilled RIPA buffer (Beyotime Biotechnology, China) containing a mixture of protease and phosphatase inhibitors, and protein concentration was quantified using a BCA protein assay kit (BL1784B, biosharp). Proteins were separated by gel electrophoresis and transferred onto a polyvinylidene fluoride (PVDF) membrane (Merck, Germany). The membrane was blocked in 5% non-fat milk for 1 hour and incubated with primary antibodies against OSMR (10982-1-AP, Proteintech), HK-2(22029-1-AP, Proteintech), LDHA (sc-137244, Santa Cruz), JAK1 (66466-1-lg, Proteintech), p-JAK1 (ab138005, abcam), STAT3 (ab68152, abcam), p-STAT3 (ab267373, abcam), β-actin (66009-1-Ig, Proteintech) overnight at 4 °C. Following three washes with 1×TBST, the PVDF membrane was incubated with a horseradish peroxidase (HRP)-conjugated secondary antibody (Sangon Biotech, China) at room temperature for 1 hour. The membrane was washed three more times with 1×TBST, followed by exposure and analysis.

Cell migration analysis

Cell migration was assessed using wound healing assay. For the wound healing assay, treated cells were evenly seeded into a 6-well plate and allowed to reach full confluence. A scratch was introduced in the cell monolayer, and the cells were then cultured in serum-free medium. Cell migration was monitored under a microscope, and images were captured and analyzed at 0, 24, and 48 hours.

Cell invasion analysis

Cell invasion ability was evaluated using Transwell chambers (8 μm pore size, Corning, USA) coated with Matrigel (BD Biosciences, USA). Briefly, 1 × 10⁴ cells in 200 μL serum-free medium were seeded into the upper chamber, while 500 μL complete medium containing 10% fetal bovine serum (FBS) was added to the lower chamber as a chemoattractant. After incubation for 48 h at 37 °C in a humidified atmosphere with 5% CO₂, non-invaded cells on the upper surface of the membrane were carefully removed with a cotton swab. Invaded cells on the lower surface of the membrane were fixed with 4% paraformaldehyde for 15 min, stained with 0.1% crystal violet for 20 min, and washed with PBS. The number of invaded cells was counted under an inverted microscope in five randomly selected fields at 200× magnification.

Tumor xenograft model

Female C57BL/6J mice (4 weeks old, body weight 19-21 g) were purchased from GemPharmatech Co, Ltd. (Jang Su, China) and housed in laminar airflow cabinets under specific pathogen-free (SPF) conditions. For orthotopic tumor implantation, 1 × 10⁵ GL261 cells stably transfected with sh-OSMR or sh-NC were stereotactically injected into the brain of each mouse. Intracranial tumor growth was evaluated by bioluminescence imaging. All animal studies were approved by the Institutional Animal Care and Use Committee of the First Affiliated Hospital of the University of Science and Technology of China.

Statistical analysis

The mean (standard deviation) was used for continuous variables with normal distribution, and the median and interquartile range (IQR) were used for other data. When comparing two groups of continuous variables, the t-test or the Mann–Whitney rank-sum test was used based on the distribution of data. The chi-square test was used for categorical variables. A P-value < 0.05 was considered statistically significant. The Kaplan-Meier survival curve was created using GraphPad Prism 9 software. A P-value < 0.05 from the log-rank test was considered indicative of a significant difference in survival times.

Results

Construction of energy metabolic classifier based bulk RNA-seq

A flowchart was created to systematically illustrate the process of establishing the energy metabolic classifier (Fig. S1). Based on previous studies of tumor metabolism, we identified four metabolic pathways: glycolysis, PPP, FAO, and glutaminolysis to describe the metabolic status of tumors (Fig. 1a). The ssGSEA algorithm was used to evaluate the activity of those metabolic pathways, and values were scaled using ‘scale’ function in RStudio. The heatmap revealed that in the TCGA dataset, patients in metabolic cluster 1 exhibited higher activity in the glycolysis, PPP, and FAO pathways, while patients in cluster 2 showed increased activity in the glutamine metabolism pathway (Fig. 1b). This same metabolic pattern was also observed in the CGGA dataset (Fig. 1c).

Fig. 1.

Gliomas exhibited significant metabolic heterogeneity between the metabolic clusters. (a) Schematics of the metabolic pathways of tumors. (b) Heatmap showing metabolic pathway activity between metabolic clusters in the TCGA dataset. (c) Heatmap showing metabolic pathway activity between metabolic clusters in the CGGA dataset. (d) Heatmap of consensus matrix showing the robust classification of consensus clustering (K = 2) in the TCGA dataset. (e) Heatmap of consensus matrix showing the robust classification of consensus clustering (K = 2) in the CGGA dataset. (f) Proportion of gliomas with different who grades between two metabolic clusters in the TCGA dataset. (g) Proportion of gliomas with different who grades between two metabolic clusters in the CGGA dataset. (h) Kaplan-Meier curves of the metabolic classifier in the TCGA dataset. (i) Kaplan-Meier curves of the metabolic classifier in the CGGA dataset. Log-rank test P-values are shown

The heatmaps of consensus matrix, consensus cumulative distribution function (CDF) plot, and delta area plot demonstrated the robustness of consensus clustering (K = 2) compared to other numbers of clusters in both the TCGA and CGGA datasets, based on the K-means unsupervised classification (Figs. 1d, e and S2a–h). Additionally, both the Proportion of Ambiguous Clustering (PAC) values and cluster consensus values indicated that K = 2 yielded the optimal clustering result (Tables S4 and S5). In terms of tumor grade, patients in metabolic cluster 1 had a higher proportion of high-grade gliomas in the TCGA and CGGA datasets (Fig. 1f, g). Univariate Cox analysis revealed that patients in cluster 1 had a significantly worse prognosis (Fig. 1h, the TCGA dataset: p < 0.001; Fig. 1i, the CGGA dataset: p < 0.001). Based on these results, two metabolic clusters were identified from the bulk transcriptome of gliomas: (1) dependent on glycolysis, PPP, and FAO; (2) dependent on glutaminolysis. Principal component analysis showed significant heterogeneity between the two clusters using this metabolic classifier, and the classification into two metabolic clusters was demonstrated to be robust (Fig. S2i, j). Moreover, there were significant differences in the mRNA expression of metabolism-related genes between the two metabolic clusters (Fig. S2k).

Clinical, genomic, and immune characteristics associated with metabolic classifier

To further analyze the clinical characteristics associated with the energy metabolic classifier, this study included several important clinical characteristics from the TCGA dataset (Fig. 2a). It is obvious that the age of patients in cluster 1 is significantly higher than that in cluster 2. In terms of molecular pathology, cluster 1 had a significantly higher proportion of IDH wild-type patients compared to cluster 2, and patients in cluster 2 have a higher proportion of MGMT methylation (Fig. 2a). Multivariate survival analysis revealed that the metabolic classifier was a significant risk factor in both the TCGA dataset (HR = 1.85[1.09, 3.13], p = 0.025) (Fig. S3a). Similar results were observed in the CGGA dataset (HR = 1.67[1.30, 2.13], p < 0.001) (Fig. S3b).

Fig. 2.

Clinical characteristics, genomic, oncogenic and immune differences between metabolic clusters in TCGA dataset. (a) Heatmap showing the distributions of age, sex, tumor grade, MGMT promoter methylation, and molecular subtypes between metabolic clusters. (b) Comparison of CNVs between two metabolic clusters. Significant amplifications or deletions of metabolism-related genes are indicated by arrows. (c) Quantitative analysis of ten classic oncogenic pathways between the metabolic clusters. (d) Comparison of the mRNA expression of immune checkpoint inhibitor genes between the metabolic clusters in the TCGA dataset. (e) Comparison of immune signatures between the metabolic clusters in the TCGA dataset. (f) Comparison of scores calculated with the estimate algorithm between the metabolic clusters. Boxplots display the median and interquartile range (IQR), with whiskers extending the maximum and minimum values within 1.5 times the IQR. The outliers are marked as individual points. Statistical significance was determined using the wilcoxon rank-sum test, with *p < 0.05, **p < 0.01, ***p < 0.001, and “ns” indicating non-significant results

Oncogene mutation events can stimulate cell-autonomous metabolic reprogramming. In this study, we found that several metabolic pathway-related genes were significantly amplified or deleted based on the CNV data (Fig. 2b), and the mRNA expression levels of these genes in the clusters were consistent with the amplification or deletion event in the genomic data (Fig. S4a, b). In addition, the overall copy number variation events in cluster 1 were significantly more frequent than those in cluster 2 (Fig. 2b).

Considering the significant difference in overall survival between patients in the two clusters, this study explored the classical oncogenic pathways in the two clusters. Based on previous literature reports, we identified ten classic tumor core pathways and quantified pathway activity using the GSVA algorithm [34]. In cluster 1, the activities of the cell cycle, Hippo, TGF-β, NRF2, and NOTCH were significantly higher than those in cluster 2, while the activities of PI3K, RAS, and WNT pathways were higher in patients in cluster 2 (Fig. 2c). Similar results were observed in the CGGA dataset (Fig. S5a).

Metabolic reprogramming has been shown to significantly influence tumor immunity. In the TCGA dataset, multiple targetable immune checkpoint genes were highly expressed in cluster 1 (Fig. 2d). A similar tendency was observed in the CGGA dataset (Fig. S5b). Several immune signatures, such as TILs, CYT, and IFN response, were significantly elevated in cluster 1 (Fig. 2e), and a similar tendency was also found in the CGGA dataset (Fig. S5c). Furthermore, the ESTIMATE algorithm [25] revealed significantly higher immune (p < 0.001) and stromal scores (p < 0.001) in metabolic cluster 1compared to metabolic cluster 2 (Fig. 2f). This result was further validated in the CGGA dataset (Fig. S5d). To further explore the relationships between tumor metabolism and immunity, we assessed the abundance of all ten immune and stromal cell types using the MCP-counter algorithm. In both the TCGA and CGGA datasets, metabolic cluster 1 exhibited significantly higher abundances of CTLs, CD8+ T cells, monocytes, and myeloid cells compared to cluster 2 (Fig. S6a, b). Given the substantial influence of IDH status on glioma metabolism reprogramming, we conducted a subgroup analysis stratified by IDH status. The metabolic classifier consistently retained its robust performance in classifying both IDH-wildtype and IDH-mutant gliomas, effectively identifying subgroups with distinct clinical outcome and immune profiles (Figs. S7, S8 and Tables S6, S7).

Overall, this study demonstrated significant differences between the two metabolic clusters in clinical characteristics, genomic mutations, molecular pathology, and tumor immunity, highlighting intrinsic heterogeneity between the two patient groups.

Multi-omics data further validated the energy metabolic classifier

The robustness of the metabolic classifier was further validated using the CPTAC dataset. Patients were assigned to corresponding metabolic clusters based on the established classifier. The heatmap revealed consistent metabolic differences in the CPTAC dataset compared to the TCGA and CGGA datasets (Fig. 3a). Metabolic cluster 1 exhibited higher activities in glycolysis, PPP, and FAO pathways, whereas cluster 2 showed elevated activity in glutamine metabolism (Fig. 3a). Consistent with this, key evaluation metrics-indicating the consensus heatmap, CDF plot, and delta area plot-also demonstrated the robustness of the K = 2 clustering in the CPTAC dataset (Figs. 3b and S9a–d). PCA revealed significant heterogeneity between the two clusters (Fig. S9e). Although no significant difference in overall survival was observed between the two clusters, patients in metabolic cluster 1 tended to have a worse prognosis compared to those in cluster 2 (p = 0.259) (Fig. S9f). In addition to bulk RNA-seq, metabolite levels were also quantified in the CPTAC dataset. The levels of lactate, alanine, and palmitic acid in metabolic cluster 1 were significantly higher than those in metabolic cluster 2 (Fig. 3c). The levels of glycerol-3-phosphate and glutamate in metabolic cluster 2 were significantly higher than those in metabolic cluster 1 (Fig. 3c). The observed metabolite levels difference aligned with the metabolic pathway activity difference. Furthermore, the Hippo, TP53, PI3K, and RAS pathways exhibited consistent differences with those observed in the TCGA and CGGA datasets (Fig. 3d).

Fig. 3.

Further validation of multi-omics data from the CPTAC dataset. (a) Heatmap of metabolic pathway activity between metabolic clusters. (b) Heatmap of consensus matrix showing the robust classification of consensus clustering (K = 2) in the CPTAC dataset. (c) Comparison of relative metabolites involved in metabolic pathways between the metabolic clusters in the CPTAC dataset. (d) Quantitative analysis of ten classic oncogenic pathways between the metabolic clusters in the CPTAC dataset. (e) Comparison of the mRNA levels of immune checkpoint genes between the metabolic clusters in the CPTAC dataset. (f) Comparison of diverse immune signatures between the metabolic clusters in the CPTAC dataset. (g) Comparison of scores calculated with the estimate algorithm between the metabolic clusters in the CPTAC dataset

In terms of the immune profile of the tumor microenvironment, metabolic cluster 1 exhibited higher mRNA levels of immunosuppressive checkpoint genes (Fig. 3e). Meanwhile, several immune features, including IFN, TILs and CYT, were significantly elevated in metabolic cluster 1 compared to cluster 2 (Fig. 3f). Moreover, the ESTIMATE algorithm revealed significantly higher immune and stromal scores in metabolic cluster 2 compared to cluster 2 (Fig. 3g). The MCP algorithm indicated significantly higher abundance of monocytes, myeloid cells, and neutrophils in metabolic cluster 1 (Fig. S9g), consistent with the findings in the TCGA and CGGA datasets. Further validation using multi-omics data from the CPTAC dataset, confirmed the robustness of the energy metabolic classifier, which accurately reflects glioma metabolic heterogeneity and highlights the significant impact of metabolic reprogramming on tumor immunity.

The role of the OSM/OSMR in glioma metabolism

To further explore whether a specific transcriptomic program is associated with metabolic phenotypes in metabolic cluster 1, differential gene analysis was performed between the two metabolic clusters, and results were visualized using a volcano plot (Fig. 4a). To identify the top upregulated genes in metabolic cluster 1, the top 100 highly expressed genes from the TCGA, CGGA, and CPTAC datasets were intersected, identifying OSMR as a key gene (Fig. 4b). OSMR, the receptor for OSM (Oncostatin M), is primarily secreted by macrophages and T cells and belongs to the interleukin-6 cytokine family. It has been reported that OSM/OSMR play a key role in the proliferation and differentiation of various tumors. In gliomas, it has been reported that OSM acts through OSMR to enhance glioma cell migration. Additionally, there are report suggesting that M2-type macrophages secrete OSM, which acts on OSMR to facilitate the proneural-to-mesenchymal transition (PMT) in glioblastoma. We found that OSMR gene was highly expressed in mesenchymal subtype and high-grade gliomas in the TCGA cohort (Fig. 4c, d). Similar trends were also observed in the CGGA cohort (Fig. S10a, b). However, the role of OSMR in glioma metabolism remains poorly understood. Survival analysis revealed that high OSMR expression is significantly associated with worse overall survival in glioma patients in the TCGA dataset (LGGs: p = 0.0013, GBMs: p = 0.0097) (Fig. 4e, f). In CGGA datasets, high expression of OSMR showed significant worse overall survival in LGGs (p < 0.0001) (Fig. S10c). For GBM, although high OSMR expression is not significantly associated with overall survival, there is still a tendency in that direction (p = 0.137) (Fig. S10d). We further explored the correlation between OSMR and glycolysis-related genes. Correlation analysis revealed significant positive associations between OSMR and both mRNA and protein levels of hexokinase 2 (HK2) and lactate dehydrogenase A (LDHA) in the CPTAC dataset (mRNA level: Pearson’s r = 0.60, p < 0.001, Pearson’s r = 0.60, p < 0.001; protein level: Pearson’s r = 0.54, p < 0.001, Pearson’s r = 0.29, p = 0.002) (Fig. 4g–j). In the TCGA dataset, OSMR also showed strong positive correlation with HK2 (Pearson’s r = 0.58, p < 0.001) and LDHA (Pearson’s r = 0.71, p < 0.001) at mRNA level (Fig. S10e, f). According to the published literature [35, 36], OSM was added to the culture medium at a concentration of 10 ng/ml for the subsequent experiments. Additionally, siRNA was used to knocked down OSMR in glioma cell lines to explore the biological function of OSMR (Fig. S10g) and observed a significant decrease in the invasion and migration ability of the U87 cell line (Fig. 4k–n). Moreover, we validated the physiological function of OSMR in the PDX glioma cell lines GBM1 and GBM2, and obtained consistent results (Fig. S11a–h). In terms of metabolism, OSMR knockdown significantly reduced the protein levels of HK2 and LDHA through western blotting in U87 cell line (Fig. 4o). Similarly, this consistent trend was observed in PDX glioma cell lines (Fig. S12a, b). As key enzymes in the aerobic glycolytic pathway, HK2 and LDHA suggest that OSM/OSMR plays a critical role in aerobic glycolytic metabolic reprogramming in glioma.

Fig. 4.

The role of the OSMR in glioma metabolism. (a) Volcano plots of differentially expressed genes between metabolic subtypes in the TCGA dataset. (b) Venn diagram showing OSMR as candidate genes across the TCGA, CGGA, and CPTAC datasets. (c) The scaled mRNA expression of OSMR across transcriptome subtypes of gliomas in the TCGA dataset. (d) The scaled mRNA expression of OSMR across normal brain and all who grades of gliomas in the TCGA dataset. (e) Kaplan-Meier curve analysis of low grade gliomas (LGGs) in TCGA dataset stratified by OSMR expression level. (f) Kaplan-Meier curves showing worse overall survival with high OSMR expression in glioblastoma (GBM) patients in the TCGA dataset. (g, h) The correlation analysis of OSMR mRNA expression with HK2 and LDHA mRNA levels in the CPTAC dataset. (i, j) The correlations analysis of OSMR protein levels of OSMR with HK2 and LDHA protein levels in the CPTAC dataset. (k, l) The invasion ability of the U87 cell line with si-OSMR-2 was significantly decreased compared with that of the control group. (m, n) The migration ability of the U87 cell line with si-OSMR-2 was significantly decreased compared with that of the control group. (o) Protein levels of the aerobic glycolysis enzymes HK2 and LDHA following the knockdown of OSMR in the U87 cell line

The OSM/OSMR/JAK1/STAT3 axis drives malignant progression of glioma

Pathway enrichment analysis revealed significant activation of the JAK/STAT pathway in glioma patients with high OSMR expression (Fig. 5a). Given JAK1’s predominant expression in the nervous system and previous reports on STAT3 with OSMR, we evaluated its phosphorylation status and that of downstream STAT3 via Western blotting. OSM treatment significantly upregulated p-JAK1 (Tyr1034/1035) and p-STAT3 (Tyr705) levels, indicating pathway activation, whereas pre-treatment with a JAK-STAT3 pathway inhibitor (Ganoderic acid A (GA-A), 1.5 μM) abolished this effect, confirming JAK1-STAT3 pathway dependency (Fig. 5b). Consistently, significant activation of the JAK1-STAT3 pathway was also confirmed in PDX cell lines (Fig. S12c, d). To functionally validate the axis, we performed rescue experiments in U87 cells (representing metabolic cluster 1). GA-A significantly attenuated OSM-induced cell invasion (p < 0.001) (Fig. 5c and d) and migration (p = 0.001) in U87 cell line (Fig. 5e, f), phenocopying OSMR knockdown effects. A similar inhibitory trend was reproduced in PDX cell lines in both invasion and migration assays (Fig. S13a–h). Complementary gain-of-function studies were conducted in LN229 cells (metabolic cluster 2) stably overexpressing OSMR, which exhibited similar results with U87 cell line (Fig. 5g–j). In vivo validation utilized a syngeneic orthotopic model, where 1 × 10⁵ luciferase-labeled GL261 cells (sh-OSMR vs. sh-NC) were stereotactically injected into the brain of C57BL/6J mice (n = 5/group). Bioluminescence imaging revealed a significant reduction in tumor burden (p = 0.0039) (Fig. 5l, m) and prolonged median survival in sh-OSMR group (p = 0.0011) (Fig. 5n). Collectively, these results establish OSMR as a critical mediator of glioma progression through JAK1/STAT3-dependent metabolic reprogramming, highlighting its potential as a therapeutic target.

Fig. 5.

The OSM/OSMR/JAK1/STAT3 axis drives progression of glioma. (a) enrichment analysis of different expressed genes between the high vs low expression group of OSMR in TCGA dataset. (b) Protein levels of the pathway of JAK1/STAT3 following the knockdown of OSMR in the U87 cell line. (c, d) The invasion ability of the U87 cell line treated with GA-A was significantly lower compared to the control group. (e, f) The migration ability of the U87 cell line treated with GA-A was significantly lower compared to the control group. (g, h) The invasion ability of the LN229 cell line treated with GA-A was significantly lower compared to the control group. (i, j) The migration ability of the LN229 cell line treated with GA-A was significantly lower compared to the control group. (k) The protocol of the orthotopic tumor implantation. (l) Bioluminescence imaging of intracranial tumor growth. (m) Radiance of bioluminescence imaging of intracranial tumor. (n) Kaplan-Meier curves showing better overall survival with sh-OSMR group of xenograft mouse compared to the sh-NC group

Metabolic profiles of malignant cells at single-cell level

This study aimed to explore metabolism at single-cell resolution using single cell RNA sequencing to better understand metabolic heterogeneity in the glioma microenvironment. Single-nucleus sequencing and bulk RNA sequencing were performed on 16 patients from the CPTAC dataset. The snRNA-seq data from 16 patients were quality-checked and integrated using the Harmony algorithm. Subsequent dimensionality reduction and clustering analyses were performed. The t-SNE projection showed excellent integration performance (Fig. S14a). As described in the methods, cells were annotated both automatically and manually, including tumor cells, immune cells, and brain-resident cells. The t-SNE visualization displayed the annotated cell types (Fig. 6a). A dot plot further illustrated the expression of cell type-specific marker genes (Fig. 6b). The inferCNV algorithm was used to identify malignant or non-malignant cells, confirming accurate annotation of glioma cells (Fig. S14b). For these 16 patients, they were assigned to metabolic clusters based on bulk RNA-seq and malignant cells from snRNA-seq data using the metabolic classifier (Fig. 6c, d). Notably, classifications based on bulk RNA-seq and snRNA-seq consistently grouped patients into the same metabolic clusters. The metabolic score accurately reflected the energy metabolic activity of each patient (Fig. 6e). As shown in Fig. 6f, the metabolic classification of most tumor cells was consistent with the metabolic clusters in the bulk RNA-seq data. Based on the above results, tumor metabolic heterogeneity is primarily driven by malignant cells.

Fig. 6.

Performance of the metabolic classifier at the single-cell resolution of the CPTAC dataset. (a) tSNE plot of snRNA-seq data from the CPTAC dataset with the corresponding cell types marked. (b) Dot plot displaying the expression of marker genes across various cell types. The dot size represents the proportion of cells within each type, while the color indicates the average expression level. (c) Metabolic classification based on bulk RNA-seq data of 16 patients in the CPTAC dataset who underwent both bulk rna and snRNA sequencing. (d) Metabolic classification based on mixed malignant cells of snRNA-seq for 16 patients in the CPTAC dataset using the metabolic classifier. (e) The robustness of the metabolic cluster scores in defining the metabolic classifier. The cluster score is calculated as the average of two cluster-specific pathways minus the average of four metabolic pathways. (f) Comparison of metabolic patterns between mixed malignant cells and bulk tumors. (g) Correlation of glycolysis, FAO, and glutaminolysis with hypoxia, and angiogenesis of malignant cells. (h) Correlation of FAO, and glutaminolysis with hypoxia, and angiogenesis of glioma cells from the CCLE database

In addition, we sought to investigate the impact of the tumor microenvironment on tumor metabolism. Given that environmental conditions cannot be directly integrated into the metabolic analyses, hypoxia and angiogenesis signatures were used as proxies for oxygen and nutrients supply within tumor microenvironment. Consistent with published research, the glycolysis pathway exhibits a significant positive correlation with hypoxia and angiogenesis (hypoxia, Pearson’s r = 0.67, p < 0.001; angiogenesis, Pearson’s r = 0.27, p < 0.001) (Fig. 6g). The fatty acid oxidation pathway is also positively correlated with hypoxia and angiogenesis (hypoxia, Pearson’s r = 0.22, p < 0.001; angiogenesis, Pearson’s r = 0.20, p < 0.001) (Fig. 6g). Glycolysis, free fatty acid oxidation, and glutamine metabolism were all significantly positively correlated with oxidative phosphorylation (glycolysis, Pearson’s r = 0.43, p < 0.001; fatty acids, Pearson’s r = 0.36, p < 0.001, glutamine metabolism, Pearson’s r = 0.34, p < 0.001) (Fig. S15a). Glioma cell lines from the CCLE database were used to explore the metabolism of glioma cells without nonmalignant cells in vitro. Although the limited number of glioma cell lines may introduce some statistical variability, intriguingly, certain metabolic patterns in these cell lines contrast with those observed in tumor tissues. Fatty acid oxidation tends to show a negatively correlation with hypoxia, though this did not reach statistical significance (Pearson’s r = −0.22, p = 0.42) (Fig. 6h, top). Glutamine metabolism was significantly negatively correlated with angiogenesis (Pearson’s r = −0.54, p = 0.037) (Fig. 6h, bottom), and also had a negative correlation trend with hypoxia (Pearson’s r = −0.46, p = 0.082) (Fig. S15b). This result suggests that nonmalignant cells in the tumor microenvironment may have a certain impact on the metabolism of tumor cells. In summary, tumor metabolic heterogeneity is primarily driven by malignant cells, our results underscore the significant roles of cell-cell interactions and the tumor microenvironment in shaping tumor metabolism.

Metabolic characteristics of malignant and nonmalignant cells in tumor microenvironment

To further elucidate the metabolic characteristics of different cell types within the tumor microenvironment, we assess the metabolic activities of five annotated cell types based on prior studies. Tumor cells exhibited the highest metabolic activity (Fig. 7a), and demonstrated elevated across the majority of metabolic pathways (Fig. 7b). In contrast, tumor-infiltrating lymphocytes had the lowest metabolic activity (Fig. 7a).

Fig. 7.

Metabolic profiles of various cell types at the single-cell level. (a) Activity of metabolic pathways across five distinct cell types in the CPTAC dataset. (b) Performance of five cell types in metabolic pathways extracted in KEGG in the CPTAC dataset. (c) Cluster scores of two metabolic clusters in malignant cells in the CPTAC dataset. (d) Cluster scores of two metabolic clusters in non- malignant cells in the CPTAC dataset. (e) The numbers of two metabolic clusters at different locations for patients in the ivy gap dataset (CThbv: cellular tumor of hyperplastic blood vessels, CTmvp: cellular tumor of microvascular proliferation, CTpan: cellular tumor of pseudopalisading cell around necrosis, CTpnz: cellular tumor of perinecrotic zone, it: infiltrating area, LE: leading edge, CTLs: cytotoxic T lymphocytes). (f) Proportion of metabolic cluster 1 and cluster 2 cells at single-cell level at different locations in the GSE182109 dataset. (g) Pseudotime analysis revealed the plasticity and dynamic transition of glioma cells. (h) Trajectory analysis demonstrated a significant transition starting from OPC- and NPC-like tumor cells in the CPTAC dataset. (i) The glycolytic activity of glioma cells increased with the PN-MES transition

To further investigate the relationship between the metabolic classifier and cell properties, we compared the metabolic pathway activities of malignant and nonmalignant cells. Among malignant cells, tumor cells from patients in metabolic cluster 1 exhibited significantly higher cluster 1 scores (p < 0.001) (Fig. 7c, left), while those in metabolic cluster 2 showed higher cluster 2 scores (p < 0.001) (Fig. 7c, right). In contrast, nonmalignant cells displayed no statistical differences in cluster 1 scores or cluster 2 scores between metabolic clusters (p = 0.067 (left), p = 0.078 (right)) (Fig. 7d). The findings further demonstrate that the metabolic heterogeneity of tumor tissue is primarily driven by malignant cells, and our metabolic classifier effectively captures this heterogeneity at single-cell resolution. We further explored the relationship between tumor metabolic heterogeneity and spatial location using the IvyGAP and GSE182109 datasets. Samples were divided into metabolic cluster 1 and metabolic cluster 2 based on the energy-related metabolism classifier. In the IvyGAP dataset, tumor tissue in the marginal regions, such as infiltrating and leading edge areas, exhibited higher cluster 1 score. Conversely, cellular tumors associated with hyperplastic blood vessels and microvascular proliferation were more likely to exhibit metabolic cluster 2 status. In contrast, cellular tumors in regions with pseudopalisading cells around necrosis and the perinecrotic zone tended to display metabolic cluster 1 status (Fig. 7e). In the GSE182109 dataset, samples were collected from multiple regions, including the core area, the MRI-enhancing area, and the marginal area, for scRNA-seq analysis (Fig. 7f). Based on the cell annotations provided in the original study, we calculated metabolic cluster scores for malignant cells using our classifier. Although statistical testing was not feasible due to the limited number of samples from the marginal areas (n = 2), we observed that the proportion of metabolic cluster 1 malignant cells was significantly lower in the tumor edge area compared to the core and MRI-enhancing areas (Fig. 7f). These results suggest that malignant cells are more likely to exhibit the metabolic cluster 1 phenotype in regions closer to the tumor core. Collectively, these findings highlight the spatial heterogeneity of glioma metabolism and demonstrate that our metabolic cluster accurately delineates the spatial metabolic characteristics of malignant cells at the single-cell resolution. This underscores the utility of our classifier in capturing both inter- and intra-tumoral embolic diversity.

Based on the intrinsic transcriptional characteristics of tumor cells and previous literature, glioma cells are typically classified into four subtypes: OPC-like, NPC-like, AC-like, and MES-like cells [2]. Trajectory analysis revealed a prominent transition starting from OPC- and NPC-like malignant cells, which correspond to the proneural (PN) subtype of GBM as classified by the TCGA dataset, to MES-like cells. This finding aligns well with previous findings (Fig. 7g, h). Additionally, glycolytic activity in glioma cells was observed to increase progressively during the PN-to-MES transition (Fig. 7i). Furthermore, hypoxia and angiogenesis within the tumor microenvironment were also found to be associated with the malignant transformation of tumor cells (Fig. S16a, b), suggesting that tumor metabolism plays a critical role in driving the mesenchymal transformation of glioma cells. Based on these findings, we propose that metabolic reprogramming in tumor cells, coupled with in influence of the tumor microenvironment, significantly contributes to the malignant transformation of glioma cells.

Metabolic characteristics of immune cells at single-cell level

In this study, we further explored the metabolic profiles of diverse immune cell populations. Utilizing singleR and leveraging insights from previous literature, we performed further dimensionality reduction and clustering analysis on myeloid cells (Fig. 8a) and lymphocytes (Fig. 8b). The marker genes of various immune cell subtypes were found to be specifically and highly expressed in their corresponding cell types (Fig. 8c, d).

Fig. 8.

Metabolic profile of immune cells at the single-cell level in the CPTAC dataset. (a) tSNE projections of myeloid cells. (b) tSNE projections of lymphoid cells. (c) Dot plot showing marker genes of different myeloid cell types. (d) Dot plot showing marker genes of different lymphoid cell types. (e) Violin plot showing the metabolic activities of glycolysis, oxidative phosphorylation, and FAO in myeloid cells. (f) Violin plot showing metabolic the activities of glycolysis, oxidative phosphorylation, and FAO in lymphoid cells. (g) Performance of metabolic pathways extracted from KEGG between CD4+ T cells and CD8+ T cells

Based on the cell annotation, this study revealed distinct metabolic characteristics across various immune cells types. Among myeloid cells, monocytes and macrophages exhibited significantly higher glycolytic activity compared to microglia (monocytes vs microglia: p = 0.002, microglia vs macrophages: p = 0.03 (top)) (Fig. 8e). Neutrophils and DC displayed relatively high glycolytic activity but low oxidative phosphorylation activity (Fig. 8e). Among lymphocytes, T regulatory cells (Tregs) demonstrated higher glycolytic activity than CD4+ T cells (Fig. 8f, p = 0.034). While CD8+ T cells showed significantly higher oxidative phosphorylation activity compared to CD4+ T cells (p = 0.002) (Fig. 8f). Additionally, plasma cells exhibited markedly higher oxidative phosphorylation and fatty acid oxidation activities than other lymphocyte subtypes (Fig. 8f). To further investigate the metabolic profiles of T cells, we conducted metabolic enrichment analysis of CD4+ and CD8+ T cells. CD4+ T cells displayed significantly elevated activity in lysine degradation, sphingolipid metabolism, glyceride metabolism, and glycerophospholipid metabolism (Fig. 8g). In contrast, CD8+ T cells exhibited higher activity in oxidative phosphorylation and cytochrome P450 metabolism. Overall, these findings highlight that different immune cell types exhibit distinct metabolic preferences (Fig. 8g).

Discussion

The metabolic heterogeneity of glioma remains a complex and multifactorial phenomenon. Advances in sequencing technology provide an unprecedented opportunity to explore the complexity of glioma metabolism. Leveraging large-scale multi-omics data, this study established an energy metabolism-based classifier for gliomas, identifying two prognostic clusters with distinct pathway activities: Cluster 1, characterized by high glycolysis, PPP and FAO activity, and Cluster 2, with elevated glutaminolysis. Our integrated analysis revealed that this metabolic heterogeneity is primarily intrinsically driven by tumor cells but is also significantly modulated by the tumor microenvironment, offering insights for developing personalized therapeutic strategies.

The metabolic heterogeneity and preference of glioma cells are increasingly recognized for their role in tumor progression and as therapeutic targets [37]. Relevant studies in other cancers support this concept, such as the glycolytic subtype linked to mesenchymalization in pancreatic cancer (Evangelista et al. [38]) and oxidative phosphorylation-driven heterogeneity in ovarian cancer (Schaeffer et al. [21]). Compared with previous studies, our research provides several novel insights: 1. By integrating multi-omics data (genomics, transcriptomics, and metabolomics), we established an energy-related metabolic classifier for gliomas using the TCGA dataset and validated its robustness across multiple external datasets. 2. Clinical characteristics, classic oncogenic pathways, and immune profiles between metabolic clusters were deeply explored. 3. We investigated the underlying causes of metabolic heterogeneity in gliomas at single-cell resolution and explored the relationship between glioma cell progression and metabolic reprogramming. 4. We further characterized the metabolic status of diverse immune cell populations at single-cell resolution.

Genome instability and mutations have long been recognized as fundamental hallmarks of the cancer [39]. The Cancer Genome Atlas (TCGA) research network has constructed a comprehensive catalog of genomic alterations driving tumorigenesis [3, 40]. Previous studies have further demonstrated that mutations in metabolism-related genes constitute a core component of metabolic reprogramming across various cancer types [8, 41]. The hypoxia conditions prevalent in glioma core regions has been shown to promote genomic instability and epigenetic modifications, ultimately contributing to tumor progression, malignant transformation, and chemoresistance [42, 43]. In line with this, we observed a higher mutation burden in metabolic cluster 1, with genes involved in metabolic pathways such as glycolysis harboring significantly more genomic mutations in the TCGA dataset. These observations suggest that metabolic gene mutations may represent an intrinsic driver of metabolic heterogeneity in gliomas.

Further exploration identified OSMR as the top upregulated gene in cluster 1 through differential gene analysis. While previous studies have reported that the OSMR would promote the proliferation and the differential of the certain tumors. As for glioblastomas, macrophage-secreted OSM acts on OSMR in glioblastoma cells to promote proneural–mesenchymal transformation [36, 44]. This study found that a decreased level of OSMR in the presence of OSM significantly reduced the expression of HK2 and LDHA, key genes in the aerobic glycolytic pathway, thereby affecting glycolytic activity. Furthermore, this study demonstrated that OSM/OSMR axis promotes malignant progression of glioma cells via JAK1/STAT3 signaling pathway. This finding is consistent with the prior reports showing that OSM binding to OSMR activates the JAK/STAT pathway, which plays a critical role in regulating glycolysis [45, 46]. In orthotopic xenograft mouse models, knockdown of OSMR significantly prolonged survival compared to sh-NC group. The concomitant immunosuppressive profile observed in cluster 1 patients suggests that OSMR may contribute to reshaping the immune microenvironment through metabolic reprogramming, a hypothesis that warrants future investigation.

Metabolism reprogramming, characterized by the adaptation of energy metabolism to support rapid cell growth and proliferation, has emerged as a critical hallmark of cancer [4]. In 1964, Professor Warburg first reported that tumor cells use glycolysis for energy under aerobic conditions (the Warburg effect) [47]. Recent researches have indicated that the activation of oncogenic pathways can lead to upregulation of metabolic pathways [8, 48]. For instance, under nutrient-deprived conditions (e.g., limited glucose, or glutamine availability), tumor cells activate the c-Myc pathway to modulate key metabolic enzymes (PHGDH, PSAT1 and PSPH) in the serine synthesis pathway, thereby maintaining cellular survival [49]. In our study, several classical oncogenic pathways were significantly activated in patients in cluster 1, such as pathways related to glycolysis metabolism. For example, Hippo and c-Myc exhibit significantly higher activity in cluster 1.

Single-cell transcriptomics from the CPTAC and GSE182109 datasets confirmed that metabolic heterogeneity is primarily driven by malignant cells, and trajectory analysis revealed escalating glycolytic activity during the transition from OPC-like to MES-like phenotypes. Spatial analysis of the IvyGAP and GSE182109 datasets indicated that metabolic cluster 1 status becomes more prevalent in tumor cells nearer the hypoxia core. This in vivo metabolic pattern differed markedly from CCLE glioma cell line profiles, underscoring the role of the tumor microenvironment, modeled here via hypoxia and angiogenesis signatures, in shaping metabolic phenotypes.

The observed metabolic differences between in vivo and in vitro models likely stem from the complex glioma microenvironment, which is not fully replicated in cultured cell models. Therefore, it is necessary to establish a model system recapitulating tumor microenvironments to explore tumor metabolism. There is growing evidence that immune responses are associated with changes in tissue metabolism, including nutrient consumption, increased oxygen consumption, and the production of reactive nitrogen and oxygen intermediates [50–52]. The metabolic phenotype of cluster 1, characterized by high lactate levels, was closely associated with an immunosuppressive microenvironment, as evidenced by enriched immune signatures and checkpoint gene expression. Furthermore, the higher proportion of IDH wild-type and mesenchymal subtype in this cluster is consistent with a more immunosuppressive context. These findings collectively suggest that patients in metabolic cluster 1, who exhibit features of immune suppression, may represent a candidate population for future investigation into combined immunotherapy.

Similar to tumor cells, immune cells also display functional metabolic heterogeneity that influence their anti-tumor activity. For instance, upon activation, T cells shift from oxidative phosphorylation (OXPHOS) to support their effector functions [53, 54]. Activated neutrophils, M1-polarized macrophages, and iNOS-expressing dendritic cells (DCs) predominantly rely on glycolysis to meet energy demands [55]. Therefore, understanding immune cell metabolic profiles and their functional consequences may enable the development of targeted therapies to modulate immunosuppressive microenvironment and improve clinical outcomes.

Conclusions

This study developed an energy metabolism-based classifier for gliomas with both prognostic and therapeutic potential. Metabolic reprogramming showed significant associations with both OPC-to-MES transition in glioma cells and immunosuppression within the tumor microenvironment. Multi-omics data, especially single-cell transcriptomes, provide us with a perspective to understand metabolism of gliomas at the single cell resolution, so as to better customize individualized treatment strategies for patients.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

TME

Tumor microenvironment

TCGA

The cancer genome atlas

CGGA

Chinese glioma genome atlas

PPP

The pentose phosphate pathway

FAO

Fatty acid oxidation (FAO)

MES

Mesenchymal

MDM

Monocyte-derived macrophages

CPTAC

Clinical Proteomic Tumor Analysis Consortium

Ivy GAP

Ivy glioblastoma atlas project

MSigDB

Molecular Signatures Database

SsGSEA

Single sample Gene Set Enrichment Analysis

PCA

Principal Component Analysis

CCP

Cell cycle progression

TILs

Tumor-infiltrating lymphocytes

CYT

Immune cytolytic activity

IFN

Interferon

CCLE

Cancer Cell Line Encyclopedia

KEGG

Kyoto Encyclopedia of Genes and Genomes

CNV

Copy number variation

WHO

World health organization

IDH

Isocitrate dehydrogenase

MGMT promoter methylation

O6-Methylguanine-DNA methyltransferase promoter methylation

IQR

Interquartile range

DCs

Dendritic cells

Oligs

Oligodendrocytes

GLUT

Glucose transporters

G-6-P

Glucose 6-phosphate

MCTs

Monocarboxylate transporters

LDH

Lactate dehydrogenase

TCA

Tricarboxylic acid cycle

α-KG

α-Ketoglutaric acid

G-6-P

Glucose-6-phosphatase

R-5-P

Ribose 5-phosphate

NADPH

Nicotinamide adenine dinucleotide phosphate

Lac

Lactate

TAMs

Tumor associated microglia or macrophages

SMCs

Smooth muscle cells

AC-like

Astrocyte like

MES-like

Mesenchymal like

NPC-like

Neural progeneitor

OPC-like

Oligdendrocytes like

CThbv

Cellular tumor of hyperplastic blood vessels

CTmvp

Cellular tumor of microvascular proliferation

CTpan

Cellular tumor of pseudopalisading cell around necrosis

CTpnz

Cellular tumor of perinecrotic zone

IT

Infiltrating area

LE

Leading edge

CTLs

Cytotoxic T lymphocytes

CDF

Cumulative distribution function

Author contributions

Qiang Zhu and Chaoshi Niu conceived and designed the study. Qiang Zhu and Wanxiang Niu were responsible for data acquisition and analysis. Qiang Zhu, Chengkun Ye, and Wanxiang Niu undertook the interpretation of data. Qiang Zhu prepared the original draft of the manuscript, which was further reviewed and edited by Qiang Zhu, Maolin Mu, and Chaoshi Niu. Chaoshi Niu provided supervision throughout the study.

Funding

This study was supported in part by grants from the National Natural Science Foundation of China [82273281]. This study was supported in part by grants from the National Natural Science Foundation of China [82203693].

Data availability

Not applicable.

Declarations

Ethics approval and consent to participate

All animal experiments were conducted in accordance with institutional recommendations. The procedures were approved by the Animal Welfare Committees of the First Affiliated Hospital of the University of Science and Technology of China.

Consent for publication

All authors consent to the publication of this manuscript.

Competing interests

The authors declare that they have no competing interests.