Comprehensive analysis of mitochondria-associated genes in glioblastoma via single-cell and bulk RNA sequencing

Abstract

Glioblastoma (GBM) is a highly aggressive brain tumor with limited treatment options and poor survival. Mitochondrial dysfunction and metabolic reprogramming, particularly the Warburg effect, are increasingly recognized as critical drivers of GBM progression. Here, we integrated single-cell (scRNA-seq) and bulk RNA-sequencing (bulk RNA-seq) data to comprehensively examine mitochondria-associated genes in GBM. We identified differentially expressed mitochondrial genes with prognostic significance and constructed a 9-gene risk signature—ACOT7, THEM5, MTHFD2, ABCB7, PICK1, PDK3, ARMCX6, GSTK1, and SSBP1—using LASSO and Cox regression. This signature robustly stratified patients into high- and low-risk groups in both The Cancer Genome Atlas (TCGA) and Chinese Glioma Genome Atlas (CGGA) cohorts, remaining an independent prognostic factor in multivariate analyses. Time-dependent ROC AUCs were 0.729, 0.813, and 0.828 at 1, 2, and 3 years in TCGA, and 0.597, 0.650, and 0.546 in CGGA. A nomogram integrating the signature with clinical variables achieved AUCs of 0.649, 0.820, and 0.854 at 1, 3, and 5 years, with good 1/2/3-year calibration. Functional enrichment and clustering analyses revealed distinct metabolic phenotypes and survival differences between subtypes. Single-cell pseudotime analysis showed a transition from oxidative phosphorylation to glycolysis in malignant cells, aligning with the Warburg effect and implicating metabolic reprogramming in immune modulation. Our findings underscore the prognostic value of mitochondria-associated genes and suggest potential therapeutic targets for disrupting GBM metabolism. Overall, these results establish a mitochondria-centric prognostic model supported by single-cell context; findings are hypothesis-generating and warrant prospective and experimental validation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03795-3.

Introduction

Glioblastoma (GBM) is the most prevalent and aggressive primary malignant brain tumor in adults. It accounts for approximately 14–15% of all central nervous system tumors and nearly half (approximately 48–50%) of malignant brain tumors. The prognosis of GBM remains dismal, with a median survival of only about 15 months despite aggressive multimodal therapeutic interventions [1]. Beyond its lethality, GBM imposes a substantial burden on patients’ families and society. The tumor itself, along with the current standard treatments—including surgical resection, radiotherapy, and chemotherapy—inflicts severe damage to the brain, leading to cognitive, emotional, and behavioral impairments. Many patients experience loss of independence, inability to work, and exorbitant medical expenses. The current standard of care involves maximal safe surgical resection followed by radiotherapy combined with temozolomide chemotherapy [2]. Although adjunctive modalities (e.g., Tumor Treating Fields and immunotherapies) have emerged, their impact on long-term survival in unselected GBM populations remains limited; therefore, a biologically grounded, clinically translatable prognostic framework is urgently needed [3, 4].Metabolic reprogramming is a hallmark of malignant tumors, including GBM, and has been closely linked to therapeutic resistance [5]. GBM cells predominantly rely on glycolysis for energy production even in the presence of oxygen, a phenomenon known as the “Warburg effect.” Beyond aerobic glycolysis, GBM exhibits metabolic plasticity, dynamically toggling between oxidative phosphorylation (OXPHOS) and glycolysis to sustain growth and therapy resistance [6–8]. Importantly, perturbations in mitochondrial structure and function modulate apoptosis, redox homeostasis, and biosynthetic flux, thereby promoting invasive growth and therapeutic resistance in GBM [9, 10]. Given this centrality, mitochondria-targeted strategies are being explored as potential therapeutic avenues in GBM [5, 11]. However, many published GBM prognostic signatures remain bulk-only, are not mitochondria-focused, and often lack external validation and single-cell context, which limits biological interpretability and clinical utility [12–14]. These gaps motivate mitochondria-centric, multi-scale analyses that bridge cohort-level prognostic modeling with single-cell context. This approach aligns with the broader evolution of biomarker-driven oncology as outlined in recent perspectives by [15].

However, the precise role of mitochondria-associated genes in GBM pathogenesis and progression remains inadequately understood. This study aims to conduct a comprehensive analysis of mitochondria-associated genes in GBM by integrating single-cell RNA sequencing (scRNA-seq) and bulk RNA sequencing (bulk RNA-seq) data, thereby elucidating their biological functions and clinical relevance. Single-cell transcriptomic analysis enables the characterization of tumor heterogeneity at a high resolution, revealing molecular features that may be masked in bulk analyses. By leveraging this approach, we aim to elucidate the role of mitochondrial gene expression patterns in GBM molecular subtyping and prognostic evaluation while identifying key mitochondrial-associated targets with therapeutic potential. To address these gaps with methods of proven utility, we leveraged penalized Cox modeling (LASSO) for robust feature selection, time-dependent ROC for dynamic discrimination, and unsupervised consensus clustering to resolve metabolic subtypes [16–18]. We explicitly integrate bulk and single-cell levels to balance cohort-level robustness with cellular resolution, acknowledging their complementary strengths and limitations [19].

Against this backdrop, we integrate single-cell and bulk transcriptomic data to systematically investigate the expression profiles and functional roles of mitochondria-associated genes in GBM. We emphasize their significance in tumor molecular classification and prognosis assessment while exploring their potential as therapeutic targets. Finally, we situate mitochondria-centric programs within broader oncogenic processes relevant to GBM—such as neovascularization, autophagy, and established biomarkers—to contextualize prognosis and therapeutic opportunities [20–22]. Through this research, we aspire to provide novel insights and a scientific foundation for GBM molecular subtyping and personalized therapeutic strategies. A preprint of this manuscript has previously been posted on Preprints.org (https://www.preprints.org/manuscript/202504.1775/v1) [23].

Study overview We followed a hypothesis-to-mechanism sequence to keep the narrative coherent: (i) Discovery: starting from MitoCarta, we identified mitochondria-associated genes that are both differentially expressed and prognostic in bulk datasets, yielding the candidate pool for modeling (Sect. 3.1, 3.2). (ii) Prognostic modeling: we distilled these candidates into a nine-gene risk score, validated it across cohorts, and integrated it into a clinical nomogram to quantify patient-level risk (Sect. 3.2, 3.3). (iii) Biological interpretation: to explain why risk differs, we performed unsupervised subtyping on mitochondria-associated expression and contrasted pathways by GSVA/GSEA to uncover metabolic phenotypes (Sect. 3.4, 3.5). (iv) Cellular context and dynamics: to anchor bulk-derived signals in tissue biology, we mapped the programs to scRNA-seq cell types and quantified OXPHOS→glycolysis shifts along pseudotime in malignant cells (Sect. 3.6, 3.7).

Materials and methods

Data sources and processing

We obtained single-cell RNA sequencing (scRNA-seq) data (GSE139448) from the Gene Expression Omnibus (GEO) database, comprising three GBM samples with a total of 8,360 cells (download date: 2024-11-19).

Bulk RNA-seq data and clinical information for GBM were sourced from The Cancer Genome Atlas (TCGA). Data were downloaded with the R package TCGAbiolinks: 1 GDCquery: Specified data type “STAR - Counts”. 2 GDCdownload: Retrieved the queried data. 3.GDCprepare: Processed and converted queried data into a SummarizedExperiment (SE) object for downstream R analysis. Clinical information was also downloaded via GDCquery_clinic in TCGAbiolinks (retrieval date: November 19, 2024). Adjacent-normal samples were identified by TCGA barcode convention (digits 14–15 = “11”), with all others labeled as tumor (implemented in code as substr(barcode, 13, 15) == “-11”).

We additionally collected GBM expression and clinical data from the Chinese Glioma Genome Atlas (CGGA), a specialized resource for GBM genomic data in Chinese patients, including expression profiles, genomic variations, and DNA methylation. Specifically, we used the dataset CGGA.mRNAseq_325.RSEM-genes.20,200,506 to obtain expression profiles and clinical annotations (download date: 2024-11-19).

Identification of differentially prognostic mitochondrial genes

We first downloaded 1,136 human mitochondria-associated genes from MitoCarta3.0 (https://www.broadinstitute.org/mitocarta) and identified 1,105 of these genes in the TCGA-GBM dataset. Next, using the R package edgeR, we discovered 94 differentially expressed genes (DEGs) (adj. p < 0.01 and |log2FC| > 1) between 94 GBM samples and normal control tissue, including 28 upregulated and 66 downregulated genes in GBM. Finally, by applying the survival R package, we examined the association between overall survival (OS) and each of the 1,105 mitochondrial genes, ultimately identifying 10 that were both differentially expressed and prognostically significant.

Construction and validation of the mitochondria-associated gene signature

Consistent with prior LASSO-anchored pipelines in neuro-oncology and systems pharmacology, we used cross-validation to control overfitting while retaining biologically coherent features [24–26]. From the TCGA-GBM dataset, we used the glmnet R package to perform regression and further select genes for computing a mitochondrial risk score. Nine genes remained in the final model. After constructing this risk model, samples from the TCGA cohort were split into high-risk and low-risk groups. We used the survival and survminer R packages to compare survival between these two groups and employed the timeROC R package to conduct time-dependent receiver operating characteristic (ROC) curve analyses at 1, 2, and 3 years. We then validated the model in an external cohort derived from the CGGA database by applying the same risk scoring, subgroup classification, survival analyses, and ROC calculations. In the external validation cohort, risk groups were additionally defined using an optimal cutpoint based on maximally selected rank statistics (survminer::surv_cutpoint).

Independent prognostic analysis and nomogram construction

To determine whether the mitochondrial gene signature could serve as an independent prognostic factor for GBM, we conducted univariate and multivariate Cox regression analyses. A nomogram was subsequently developed using the rms and regplot R packages to visualize the relationships between clinical variables and the prognostic model. We employed 1-, 2-, and 3-year calibration curves to assess and predict the nomogram’s performance. Time-dependent ROC for the nomogram was reported at 1, 3, and 5 years (Fig. 1).

Fig. 1.

Nomogram construction for GBM survival prediction. A, B Forest plots of univariate and multivariate Cox regression analyses; C The nomogram based on the 9-gene risk score; D Calibration at 1/2/3 years; E KM survival curves; F time-dependent ROC (1/3/5-year; AUCs = 0.649/0.820/0.854)

Clustering analysis

Using the ConsensusClusterPlus R package, we classified the TCGA-GBM cohort based on mitochondria-associated gene expression to investigate whether these genes were associated with GBM. When k = 2, we observed strong intracluster correlations and moderate intercluster correlations. Further survival-related analyses showed a significant difference in prognosis between the two clusters. Lastly, we conducted GSEA and GSVA analyses to explore differences in metabolism-related pathways between the two subgroups.

Functional enrichment analysis

Using the ClusterProfiler R package, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses. We visualized the GO and KEGG results with the circlize R package. We further examined three Warburg effect–related gene sets obtained from MSigDB using the GSVA algorithm, identifying enrichment differences between the subgroups.

Single-cell data processing

We employed the MAESTRO v1.5.1 [PMID: 32767996 ] standardized workflow to process all collected datasets, including quality control, batch effect correction, cell clustering, differential expression analysis, and cell-type annotation. Two metrics were used to assess cell quality: the total number of counts (UMIs) per cell (library size) and the number of detected genes per cell. Cells with a library size < 1,000 or fewer than 500 detected genes were discarded. We identified the top 2,000 variable features, performed PCA for dimensionality reduction, and applied KNN and Louvain algorithms for clustering. To better capture cell differences and variability across datasets with different cell numbers, we set the number of principal components to 30 and the graph-based clustering resolution to 1, then used Uniform Manifold Approximation and Projection (UMAP) for further dimensionality reduction and cluster visualization. In total, 17 clusters were identified. We used the Wilcoxon test, with log fold-change (|logFC| ≥ 0.25) and false discovery rate (FDR < 1e-05), to detect differentially expressed (DE) genes in each cluster compared with all other cells. For cluster annotation, we applied a marker-based approach implemented in MAESTRO, which references DE genes against published marker gene resources. Ultimately, four types of immune/tumor-related categories were identified. We conducted pseudotime trajectory analysis using the default settings in Monocle 3, which infers cell differentiation paths by placing single cells along a trajectory that corresponds to biological processes (e.g., cell differentiation).

Statistical analysis

We used the timeROC R package to compute time-dependent ROC curves and AUCs. Risk-score ROC horizons were 1, 2, and 3 years (Fig. 2), whereas nomogram ROC horizons were 1, 3, and 5 years (Fig. 1); calibration was assessed at 1, 2, and 3 years. The rms package was used to construct and evaluate nomograms and calibration curves. GO and KEGG enrichment analyses were conducted with the clusterProfiler R package. Multiple testing in enrichment analyses was controlled using the Benjamini–Hochberg false discovery rate (FDR). Continuous variables between two groups were compared using the Wilcoxon rank-sum test, and continuous variables among three or more groups were compared via the Kruskal–Wallis test. Categorical variables were compared using Fisher’s exact test or the chi-square test as appropriate. Unless otherwise specified, p < 0.05 was considered statistically significant. All statistical analyses were performed in R version 4.2.3.

Fig. 2.

Construction and validation of the predictive risk model. A, C, E In the training cohort: distribution of the risk model with corresponding survival status and expression patterns of the 9 genes, KM survival curves, and time-dependent ROC (1/2/3-year; AUCs = 0.729/0.813/0.828); B, D, F In the external validation cohort: similar analyses depicting survival status, gene expression, KM survival curves, and time-dependent ROC (1/2/3-year; AUCs = 0.597/0.650/0.546)

Results

Identification of differentially prognostic mitochondria-associated genes

Differential expression analysis between TCGA-GBM samples and control tissues identified 94 mitochondria-associated genes exhibiting significant changes (FDR < 0.01, |log₂FC| > 1), with 28 genes upregulated and 66 downregulated in GBM (Fig. 3A, volcano plot). Subsequently, prognostic analysis revealed 87 genes significantly associated with overall survival (OS) (Fig. 3B). The intersection of these 87 prognostic genes with the differentially expressed mitochondria-associated genes yielded 10 genes, which were selected for further analysis (Fig. 3C). The nine genes retained in the final model were ACOT7, THEM5, MTHFD2, ABCB7, PICK1, PDK3, ARMCX6, GSTK1, and SSBP1 (coefficients and HR directions summarized in Table S1).

Fig. 3.

Identification of differentially prognostic mitochondria-associated genes. A Volcano plot showing the differential expression of mitochondria-associated genes between GBM and control samples; B Forest plot of the top 20 prognostically significant mitochondria-associated genes; C Venn diagram illustrating the selection of 10 differentially prognostic genes

Construction and validation of the mitochondria-associated Rissk model

After excluding TCGA-GBM samples with missing survival time or status, 155 cases were retained as the training set. Based on the 10 differentially prognostic mitochondria-associated genes, we further refined the gene set and constructed a prognostic model. LASSO Cox regression analysis was performed using the glmnet R package to streamline the model. Examination of the coefficient trajectories indicated that as the lambda parameter increased, more coefficients converged to zero. Ten-fold cross-validation was employed to determine the optimal lambda value (λ = 0.0339602; see Figure S1A and S1B), resulting in the following risk score formula:

Risk scores were calculated for each sample in both the training and external datasets, and the distribution of mtRNAscores demonstrated that samples with higher risk scores had markedly shorter OS compared to those with lower scores. Analysis of the nine signature genes revealed that high expression of MTHFD2, ABCB7, and ARMCX6 was associated with lower risk (protective factors), whereas high expression of SSBP1, GSTK1, PDK3, PICK1, THEM5, and ACOT7 correlated with higher risk (risk factors) (Fig. 2A and B). By stratifying samples into high- and low-risk groups using the median mtRNAscore, Kaplan–Meier (KM) survival analysis showed a significant difference in OS in the training cohort (log-rank p = 3.37e − 07, HR = 2.781 [95% CI: 1.877–4.12]; 78 high-risk vs. 77 low-risk samples) with time-dependent ROC AUCs of 0.729, 0.813, and 0.828 at 1, 2, and 3 years, respectively (Fig. 2E). Validation in the external CGGA cohort confirmed the model’s robustness (log-rank p = 0.0396, HR = 1.579 [95% CI: 1.022–2.439]; 107 high-risk vs. 30 low-risk samples) with time-dependent ROC AUCs of 0.597, 0.650, and 0.546 at 1, 2, and 3 years, respectively(Fig. 2F).

Development and validation of a nomogram based on the 9-gene model

To translate the 9-gene model into clinical practice, we integrated the risk score with clinical variables to develop a nomogram for predicting patient survival. Univariate and multivariate Cox regression analyses were performed using the 9-gene risk score and clinical factors (age, sex, race, and IDH mutation status). Univariate analysis showed that both IDH mutation status and the 9-gene model were significantly associated with OS, while multivariate analysis confirmed that race, IDH mutation status, and the 9-gene model are independent prognostic factors (Fig. 1A and B). Consequently, a nomogram was constructed to estimate GBM patients’ survival probabilities (Fig. 1C). Calibration curves, which compare predicted and actual survival probabilities at different time points, demonstrated good concordance (Fig. 1D). The nomogram showed time-dependent ROC AUCs of 0.649, 0.820, and 0.854 at 1, 3, and 5 years, respectively (Fig. 1F), with calibration assessed at 1, 2, and 3 years. KM analysis further validated the effective stratification of patients into high- and low-risk groups.

Clustering analysis

To explore the potential mechanisms by which mitochondria-associated genes influence GBM prognosis, we performed consensus clustering based on their expression profiles. GBM samples were divided into two clusters (Fig. 4A–E). Kaplan–Meier survival analysis demonstrated that patients in Cluster 2 had significantly poorer survival compared to those in Cluster 1 (Fig. 4F).

Fig. 4.

Consensus clustering-based classification of GBM samples using mitochondria-associated gene expression. A–E Cumulative distribution function (CDF) curves and corresponding area under the CDF curves for k = 2–10, with heatmaps for clustering into 2, 3, 4, and 5 subgroups; F KM survival curves comparing the two clusters

Subsequently, using the limma package, we identified differentially expressed genes (DEGs) between the two clusters and conducted enrichment analysis. KEGG analysis revealed significant differences in pathways, including p53 signaling, TNF signaling, and IL-17 signaling(Fig. 5B), while GO enrichment analysis is summarized in Fig. 5A. Furthermore, GSVA analysis showed that Cluster 1 was primarily enriched in metabolic pathways such as Terpenoid_backbone_biosynthesis and Glyoxylate_and_dicarboxylate_metabolism, whereas Cluster 2 was enriched in Phenylalanine_metabolism and Starch_and_sucrose_metabolism (Fig. 5C). GSEA of the DEGs between clusters identified significant enrichment of the “GALLUZZI_PREVENT_MITOCHONDRIAL_PERMEABILIZATION” pathway in Cluster 2 (Fig. 5D).

Fig. 5.

Functional enrichment analyses based on subgroup-specific differentially expressed genes (DEGs). A Gene Ontology (GO) enrichment analysis of DEGs between molecular subgroups. B Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis. C Gene Set Variation Analysis (GSVA) highlighting differences in hallmark pathways across subgroups. D Gene Set Enrichment Analysis (GSEA) showing enrichment of Warburg effect–related metabolic pathways

Comparison of Warburg effect gene expression between subtypes

To assess the influence of mitochondria-associated genes on the effectors of the Warburg effect, we extracted 74 genes from three Warburg effect–related pathways. Analysis showed that 35 of these genes were significantly differentially expressed between Cluster 1 and Cluster 2 (Fig. 6A), while the remaining 39 genes did not exhibit significant differences (Fig. 6B). Kaplan–Meier analysis further revealed that five genes were significantly associated with OS: low expression of BID and USP44 correlated with higher risk, whereas high expression of PMAIP1, VDAC1, and SLC25A4 was linked to increased risk (Fig. 6C).

Fig. 6.

Warburg effect–related gene expression and survival analysis across GBM clusters. A, B Expression heatmaps of the 74 Warburg effect–related genes across different GBM molecular clusters identified by consensus clustering. C Kaplan–Meier survival curves evaluating the association between Warburg effect–related gene expression and overall survival (OS) in GBM patients

Single-cell data analysis

Following quality control, 17,029 genes and 8,360 cells were retained (see Figure S2 for assessments of gene and mitochondrial content). The MAESTRO pipeline was employed to identify the top 2,000 variable features, perform PCA for dimensionality reduction, and cluster cells using KNN and the Louvain algorithm. To effectively capture cell-to-cell variability across datasets with differing cell counts, 30 principal components and a graph-based clustering resolution of 1 were used, followed by UMAP for visualization. The distribution of cells across the three samples is shown in Fig. 7A, and UMAP visualization revealed 17 distinct clusters (Fig. 7B). Subsequent cell-type annotation identified 461 immune cells, 87 stromal cells, and 7,812 malignant cells (Fig. 7C). These clusters were summarized into four immune/tumor-related categories for downstream analyses (Fig. 7D).

Fig. 7.

Single-cell data processing and cell-type annotation. A Distribution of cells across the three integrated GBM samples. B UMAP plot displaying 17 clusters identified through graph-based clustering. C Classification of cell types based on cluster-specific marker gene expression. D Annotated UMAP plot showing four major immune/tumor-related cell populations

Next, we examined the differential expression of mitochondria-associated genes and Warburg effect-related genes across cell types. Mitochondria-associated genes exhibited distinct expression patterns in malignant versus immune cells (Fig. 8A), with further differences among immune cell subtypes (Fig. 8B). Similarly, the expression of Warburg effect-related genes in malignant compared to immune cells is shown in Fig. 8C, with additional details across immune cell subgroups in Fig. 8D. Notably, while PMAIP1 was expressed at relatively low levels across cell subgroups, VDAC1 and SLC25A4 were markedly overexpressed in malignant cells compared to stromal, endothelial, and monocyte populations. These findings corroborate earlier observations (Fig. 5).

Fig. 8.

Differential Expression of Mitochondria-Associated and Warburg Effect-Related Genes Across Cell Types. A, B Expression patterns of mitochondria-associated genes; C, D Expression patterns of Warburg effect-related genes

Pseudotime analysis

Pseudotime analysis was conducted using Monocle 3, which employs reverse graph embedding to quantitatively estimate pseudotime and reconstruct trajectories of biological progression [PMID: 24658644, PMID: 28114287, PMID: 28825705 ]. Initially, the function was used to identify cell trajectories (Fig. 9A). Subsequently, manually selected starting points allowed us to refine the trajectory (Fig. 9B). To explore metabolic changes along the trajectory, AUCell was used to calculate metabolic enrichment scores based on a mitochondrial gene set (mtDNA-encoded genes) and a Warburg effect-related gene set. In malignant cells, the mitochondrial gene (OXPHOS-related) score showed an overall declining trend along the pseudotime trajectory, indicating high OXPHOS activity at the trajectory’s origin that diminished over time. Conversely, the Warburg effect score increased in later stages, suggesting a metabolic reprogramming from oxidative phosphorylation toward glycolysis (Fig. 9C and D).

Fig. 9.

Single-cell trajectory analysis and metabolic activity scoring. A Pseudotime trajectory analysis of immune cells using Monocle 3. B Refined trajectory revealing progressive cell state transitions. C AUCell score distribution of the mitochondria-associated gene set across the trajectory. D AUCell score distribution of the Warburg effect–related gene set across the trajectory

Further, the graph_test function was applied to identify genes that significantly varied along the pseudotime trajectory. Among the nine model genes, six were identified as differentially expressed (q-value < 0.0001) along the trajectory. Their dynamic expression patterns across different cell types are shown in Fig. 10A, with corresponding UMAP visualizations in Fig. 10B. These results indicate that these genes may play crucial roles in cell state transitions and metabolic reprogramming, reinforcing the concept of dynamic transcriptional changes at the single-cell level.

Fig. 10.

Expression dynamics of the six model genes along the immune cell trajectory. A Expression trends of the six genes across pseudotime, highlighting their dynamic regulation during immune cell state transitions. B Feature plots showing the spatial expression patterns of the six genes on the UMAP projection

Finally, using the find_gene_modules function, co-expression module analysis was performed on pseudotime DEGs, generating a file (Genes_Module.csv) detailing the composition of each module. As shown in Fig. 11A, Modules 13, 7, and 14 exhibited the lowest expression in malignant cells (dark blue), suggesting potential suppression—possibly involving tumor suppressor genes, cell differentiation, or immune-related pathways—while these modules were most highly expressed in mono/macro cells, implying roles in monocyte/macrophage functional regulation (e.g., inflammatory signaling). Modules 7 and 14 were most highly expressed in endothelial cells (dark red), suggesting involvement in endothelial function, angiogenesis, or tumor microenvironment regulation. In contrast, Modules 9 and 16 were suppressed in endothelial cells (dark blue). In mast cells, Modules 14, 9, and 16 were highly activated, indicating possible roles in allergic responses, inflammatory modulation, or angiogenesis.

Fig. 11.

Cell-type–specific distribution of gene modules and metabolic pathway activity. A Distribution of co-expression gene modules across different immune cell types. B AUCell-based activity scores of representative metabolic pathways across cell types

Subsequently, average expression values for each cell type were calculated to construct a matrix, and GSVA was employed to assess the distribution of metabolic pathway scores across cell types (Fig. 11B). The results revealed that both Oxidative phosphorylation (OXPHOS) and Glycolysis/Gluconeogenesis were significantly elevated in malignant cells, consistent with metabolic plasticity rather than a strict binary switch, and supporting the relevance of the Warburg effect in glioblastoma.

Discussion and conclusion

Glioblastoma (GBM) is the most common and aggressive malignant glioma of the central nervous system, accounting for approximately 50% of all primary malignant brain tumors [27]. Despite aggressive treatments—including maximal safe resection, radiotherapy, and temozolomide chemotherapy—the five-year survival rate remains below 7%, which underscores the urgent need for novel prognostic biomarkers and therapeutic strategies [28]. The biological complexity and metabolic heterogeneity of GBM contribute to its poor prognosis, making it a persistent challenge for both clinicians and researchers [29].

In this study, we aimed to address this challenge by integrating bulk RNA sequencing data from The Cancer Genome Atlas (TCGA) and the Chinese Glioma Genome Atlas (CGGA) with single-cell transcriptomics to develop and externally validate a nine-gene mitochondrial risk signature (ACOT7, THEM5, MTHFD2, ABCB7, PICK1, PDK3, ARMCX6, GSTK1, and SSBP1). Our analysis revealed a shift from oxidative phosphorylation (OXPHOS) to glycolysis along malignant trajectories, consistent with the classical Warburg effect [2]. This shift provides a cellular explanation for the greater biological aggressiveness observed in high-risk patients and highlights the importance of mitochondrial reprogramming in GBM. Such metabolic rewiring has also been linked to inflammatory responses via inflammasome and ER-stress pathways [30].

Methodologically, bulk and single-cell RNA sequencing offer complementary strengths and distinct biases. Bulk profiling provides cohort-level stability and power for survival modeling but averages intratumoral heterogeneity; single-cell profiling resolves rare subpopulations and dynamic states yet introduces technical noise, batch effects, and smaller sample sizes [31, 32]. Our design leveraged both: we discovered and externally validated a nine-gene mitochondrial risk score at the cohort level, then anchored it to single-cell programs and pseudotime trajectories to rationalize biology. Beyond our own cross-validation and external replication, the use of LASSO/Cox to derive prognostic signatures in GBM has substantial precedent, supporting our modeling choice [33–35]. In parallel, recent work catalogs technical and biological biases in bulk transcriptomic mining—reinforcing why we anchor cohort-level signals in single-cell context to mitigate averaging artifacts [36]. Integrative single-cell + bulk pipelines have also proven effective for nominating and contextualizing biomarkers (e.g., CENPA), which further validates our study design [37]. Such pipelines have been successfully deployed across cancer types—for instance, Rasteh et al. profiled mitotic checkpoint kinases pan-cancer, while Dong et al. focused on autoubiquitination-related transcription factors [38, 39].

These data suggest therapeutic levers. Forcing pyruvate flux back into mitochondria by inhibiting pyruvate dehydrogenase kinase (PDK) is biologically coherent; dichloroacetate (DCA) radiosensitizes high-grade gliomas and counteracts facets of Warburg metabolism in preclinical models [40, 41]. Interestingly, local anesthetics such as lidocaine have also been explored as metabolic and ion-channel modulators with anti-cancer effects [42]. At the same time, GBM progression reflects intersecting oncogenic axes beyond metabolism. Neovascularization—with contributions from vessel co-option, vasculogenic mimicry, and canonical VEGF pathways—can blunt the durability of anti-angiogenic therapy [43, 44]. Autophagy exerts context-dependent (“double-edged”) effects in GBM, intersecting with stem-like states, immunity, and metabolism [45, 46]. Our subtype-resolved enrichment differences (e.g., p53, TNF, IL-17 pathways) reinforce the idea that risk-high tumors sit at the nexus of metabolic, inflammatory, and cell-death programs.

At the membrane interface, ion-channel biology intersects with metabolic stress in ways directly relevant to GBM invasion and microenvironmental acid–base dynamics. Voltage-gated sodium channels (VGSCs)—including Nav1.7—facilitate motility/invasion programs, and emerging agents against Na_v1.7 provide mechanism-linked tools to test in metabolism-aware combinations [47–49]. More broadly, Nav channels are known to exert noncanonical roles in cancer progression beyond electrophysiology [50]. Recent structural/glycosylation insights into Na_v isoforms sharpen targetability and may inform isoform-specific strategies [51].

Positioned within the biomarker landscape, established determinants (e.g., MGMT methylation, IDH mutation, EGFR alterations; circulating VEGF-A isoforms for therapy monitoring) already inform prognosis and treatment sensitivity [51, 52]. Transcriptomic signatures such as the 5-gene model proposed by Li et al. further illustrate the value of mRNA-based risk stratification in GBM [53]. A mitochondria-associated transcriptomic signature adds an orthogonal, metabolism-grounded dimension that may refine risk-adapted combinations. Embedding metabolic signatures within companion-diagnostic (CDx) pathways is feasible: roadmaps have emphasized fast-tracked drug development via biomarker-informed CDx frameworks [54]. In parallel, forward-looking reviews have highlighted mitochondria-centered signaling as a frontier in oncology’s future directions [55]. Together, our model and single-cell anchoring argue for precision metabolism in GBM—Emerging deep learning approaches, including generative adversarial networks (GANs), may enhance expression modeling and simulation of rare transcriptomic states in GBM.

Limitations

While our study provides a robust integration of bulk and single-cell RNA sequencing data to elucidate mitochondrial gene signatures in glioblastoma (GBM), several limitations should be noted. First, the study did not incorporate methods like single-cell reference-based deconvolution or orthogonal validation with spatial transcriptomics or proteomics, which could have further validated our findings at the cellular and sub-cellular levels. These approaches, though outside the scope of the current work, remain promising avenues to enhance the robustness of our conclusions in future studies. Additionally, while TCGA and CGGA cohorts provide valuable insight, the external validation cohort was relatively small, particularly for low-risk groups, which could limit the generalizability of the findings. Future prospective studies should address these limitations by incorporating these complementary techniques and validating the findings in larger, independent cohorts to strengthen the predictive power of the model.

Author contributions

F.W. performed software development, data curation, formal analysis, methodology design, visualization, and wrote the original draft. Z.J. contributed to software development, formal analysis, data curation, and co-wrote the original draft. Q.W., M.L., B.L., A.H., C.Z., Y.G., and H.L. were responsible for data curation. W.L. and W.H. provided supervision, resources, and participated in manuscript review and editing. X.F. supervised the study, acquired funding, and contributed to manuscript review and editing. All authors reviewed and approved the final manuscript.

Funding

This work is funded by Science and Technology Department of Henan Province (Item number: 232102311134), Health Commission of Henan province (Item number: YXKC2022020), National Natural Science Foundation of China‌ (Item number: 82303029), and the Zhengzhou Science and Technology Innovation Project for Healthcare (Item number: 2024YLZDJH027).

Data availability

The single-cell RNA-seq dataset supporting the findings of this study is publicly available at the Gene Expression Omnibus (GEO) under accession number GSE139448. The bulk RNA-seq data and corresponding clinical information for glioblastoma (GBM) were obtained from The Cancer Genome Atlas (TCGA) database (specifically, the TCGA-GBM project) via the NCI Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov/). Additional GBM expression and clinical data were acquired from the Chinese Glioma Genome Atlas (CGGA) (http://www.cgga.org.cn/). The list of human mitochondria-associated genes was retrieved from MitoCarta3.0 (https://www.broadinstitute.org/mitocarta). All datasets and resources are publicly accessible, and details regarding data processing and analysis steps are provided within the article. Further information is available from the corresponding author upon reasonable request. Codes and processed data are identical to those deposited with the preprint (doi: 10.20944/preprints202504.1775.v1).

Declarations

Ethics approval and consent to participate

No new human participants were directly involved in this study. All data utilized were obtained from publicly accessible databases (TCGA, CGGA, and GEO), which have existing ethics approvals or consent protocols in place. Accordingly, additional institutional review board approval or individual participant consent was not required for this analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.