Identifying NDUFB2 as a prognostic biomarker for glioblastoma through an exploratory analysis of an anesthesia-related gene signature

Abstract

Background

Glioblastoma (GBM) is the most aggressive primary brain tumor with poor prognosis despite multimodal therapy. Understanding the molecular and genetic characteristics of GBM and the influence of anesthesia drugs on tumor behavior is crucial for developing new therapeutic strategies.

Methods

We analyzed RNA-seq data from The Cancer Genome Atlas (TCGA)-GBM cohort and a Gene Expression Omnibus (GEO) dataset (GSE179004) of GBM samples treated with propofol and sevoflurane. Differential expression analysis identified anesthesia-related genes (ARGs), and their prognostic relevance was assessed using Cox regression. Consensus clustering stratified GBM patients into subgroups with distinct survival outcomes, immune cell infiltration, and pathway activities. An ARGs-based prognostic model was developed using Lasso-Cox regression and validated across cohorts. The hub gene NDUFB2 was identified and validated using single-cell sequencing, drug sensitivity assessment, and spatial transcriptome analysis. NDUFB2 expression levels were experimentally verified in our GBM samples.

Results

ARGs were significantly differentially expressed between GBM and normal tissues, with NDUFB2 identified as a hub gene associated with poor prognosis. Consensus clustering divided GBM patients into two subgroups with significant survival differences. An 11-ARGs-based prognostic model was established, demonstrating strong correlation with overall survival. NDUFB2 was predominantly expressed in malignant cells and associated with decreased survival and drug sensitivity.

Conclusions

Our study, based on an initial exploratory analysis of anesthesia-related genes (ARGs), highlights their potential prognostic significance in GBM. We propose NDUFB2 as a robust biomarker for prognosis and therapeutic response, supported by extensive validation. These findings offer insights into the molecular classification of GBM and suggest a possible impact of anesthesia drugs on tumor progression, warranting further validation in larger cohorts.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-15312-4.

Introduction

Glioblastoma (GBM) is the most common and malignant primary brain tumor in the central nervous system [1]. The treatment of GBM has always been a major challenge in the field of neuro-oncology because of its highly aggressive nature, poor efficacy and poor prognosis. Despite a combination of surgical resection, radiotherapy, and adjuvant chemotherapy, the median survival of GBM patients is still less than 15 months [2, 3]. In recent years, research for GBM has focused on understanding its molecular and genetic characteristics, as well as developing therapeutic strategies against specific molecular targets [4, 5]. However, these therapeutic approaches still face many challenges due to tumor heterogeneity and the complex tumor microenvironment.

Anesthesia management plays a crucial role in GBM surgical treatment. While anesthesia drugs ensure patient safety and comfort during surgery, recent studies suggest they may also affect tumor biology [6, 7]. Clinical and experimental studies indicate that some anesthetic drugs could indirectly influence tumor growth and metastasis. They may do this by altering the tumor microenvironment, immune responses, and the balance of cell proliferation and apoptosis [8, 9]. Until now, the impact of propofol and sevoflurane anesthesia-related genes (ARGs) on GBM prognosis has not been explored.

Although research in this area is still in its preliminary stages, the exploration of the correlation between anesthetic drugs and GBM offers the possibility of developing new therapeutic strategies. In this study, we embarked on the exploration of ARGs and their prognostic relevance in GBM, aiming to devise a novel ARGs-based model. Moreover, through the analysis of clinicopathological specimens from GBM patients at the First Affiliated Hospital of Guangxi Medical University (GXMU), we have identified NDUFB2 as a gene intimately associated with GBM prognosis for the first time.

Materials and methods

Data acquisition and differential expression analysis of ARGs

We sourced RNA-seq expression profiles and corresponding clinical information for GBM patients from the TCGA-GBM cohort. The TCGA-GBM dataset, part of the larger TCGA initiative, provides a detailed repository of genomic information with the goal of delineating the genetic alterations characteristic of glioblastoma multiforme, which is recognized as the most prevalent and lethal type of primary brain tumor in adults [10]. RNA-sequencing (RNA-seq) expression profiles (in Fragments Per Kilobase of transcript per Million mapped reads, FPKM) and corresponding clinical follow-up data for glioblastoma multiforme (GBM) were downloaded from The Cancer Genome Atlas (TCGA) database using the GDC Data Portal (https://portal.gdc.cancer.gov/) [11]. To establish a well-defined cohort for prognostic analysis, a rigorous selection process was implemented. First, we included only samples designated as primary solid tumors (TCGA sample code: ‘01’). Next, we excluded patients who met the following criteria: (1) missing information on overall survival (OS) status or survival time; (2) an OS time of zero days. This filtering process was designed to ensure that the cohort was suitable for survival analysis and to minimize bias from non-cancer-related events. After applying these criteria, a final cohort of 150 GBM patients was retained for all subsequent analyses. In addition, there were five normal brain samples used as controls.

RNA-seq data from GBM samples treated with propofol (n = 3) and sevoflurane (n = 3) were obtained from the Gene Expression Omnibus (GEO) database under the accession number GSE179004. This dataset was used exclusively to identify differentially expressed Anesthesia-Related Genes (ARGs). We established a false discovery rate (FDR) of less than 0.05 and an absolute log2 fold change (log2FC) greater than 1 as our criteria for significance. Heatmaps were generated to highlight the genes with the most significant expression changes, and volcano plots were created to provide a comprehensive view of the differential expression profile. Additionally, we conducted Gene Ontology (GO) enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis on the identified ARGs to further understand their biological implications [12–14].

In-house cohort for experimental validation: A total of 22 GBM tissue samples and corresponding adjacent normal brain tissues (when available) were collected from patients who underwent surgery at The First Affiliated Hospital of Guangxi Medical University (GXMU) between January 2018 and December 2023. This study was approved by the Institutional Ethics Committee, and informed consent was obtained from all patients (Institutional Review Board approval number, 2023-K102-01). These samples were used for RT-qPCR, Western Blot, and Immunohistochemistry (IHC) to validate the expression of the identified hub gene.

Prognostic relevance of ARGs

To investigate the prognostic impact of ARGs on GBM patient survival, we conducted an analysis within the TCGA-GBM cohort. Prognostically significant ARGs (univariate Cox regression analysis, with a significance threshold of p < 0.05.) were identified using univariate Cox regression analysis via the “survival” package in R. A network diagram was constructed to depict the interrelationships between the expression levels of these ARGs with prognostic potential. Gene copy number data for the TCGA-GBM cohort were retrieved from the Genomic Data Commons via the University of California Santa Cruz (UCSC) XENA platform. The “RCircos” R package allowed us to assess the prevalence of copy number variations (CNV frequency%) and to explore the genomic alterations present in ARGs linked to prognosis.

Cluster analysis of ARGs

Within the TCGA-GBM cohort, the prognostic value of ARGs was initially screened using univariate Cox regression analysis. Genes with a p-value < 0.05 were considered prognostically significant. To stratify GBM patients, consensus clustering was performed on the expression profiles of these prognostic ARGs using the “ConsensusClusterPlus” R package [15]. The optimal number of clusters (k) was determined based on the consensus matrix and cumulative distribution function (CDF) plots. Survival differences between these groups were evaluated through Kaplan-Meier analysis. To confirm the stability of our clustering, we employed Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) plots. Furthermore, we examined the variance in immune cell infiltration among the GBM subtypes. The R packages “GSVA” and “GSEABase” were instrumental in conducting Gene Set Variation Analysis (GSVA) and Gene Set Enrichment Analysis (GSEA), respectively, to discern KEGG pathway variances between the clusters.

Development of an ARGs-based prognostic model

The “survival” R package facilitated univariate Cox regression analysis to pinpoint ARGs with prognostic relevance. These genes were then integrated into a Lasso-Cox model, and cross-validation was performed using the “glmnet” package in R to formulate an ARG-based prognostic index. The prognostic index was calculated using the formula ∑(βi×Expi), where ‘i’ denotes the number of ARGs. Lasso regression, which is particularly suited to high-dimensional genomic data, was used to refine the model by reducing certain coefficients, thereby enhancing its predictive accuracy.

The TCGA-GBM cohort was randomly divided into a training set (70% of patients) and a testing set (30%). Patients in the training, testing, and entire TCGA cohorts were then stratified into high- and low-risk groups based on the median risk score of the respective cohort. Kaplan-Meier survival analysis was then applied to compare the overall survival between these groups. The model’s predictive significance was assessed through multivariate Cox regression, and a heatmap was generated to illustrate the relationship between risk scores and the signature ARGs.

Construction and assessment of a predictive nomogram

To enhance clinical applicability, a nomogram was constructed integrating the ARG-based risk score and key clinical variables, including patient age and gender. Calibration plots were created to juxtapose the nomogram’s predicted survival probabilities against the observed rates. To ascertain the clinical applicability of the nomogram, decision curve analysis (DCA) was utilized, a method frequently employed in medical decision-making assessment [16]. The DCA plots, with threshold probability on the x-axis and net benefit on the y-axis, allowed us to visually appraise the performance of various clinical strategies in terms of minimizing the risk of overtreatment and misdiagnosis.

Identification and validation of the hub gene of ARGs

We scrutinized differentially expressed ARGs using the Search Tool for the Retrieval of Interacting Genes (STRING) database to establish a protein-protein interaction (PPI) network. The Cytoscape software was then utilized to examine the PPI network and pinpoint the central ARG.

The ARGs within the PPI network were ranked using the Maximal Clique Centrality (MCC) algorithm provided by the cytoHubba plugin, selecting high-scoring genes as potential key players [17]. A Venn diagram was employed to identify common genes between the candidate sets, with the gene possessing the highest MCC score deemed the central ARG. We further explored the relationship between the central ARG and overall survival in pan-cancer analysis from the TCGA database, as well as its prognostic significance in GBM. The Tumor Immune Estimation Resource (TIMER) database was accessed for tumor-infiltrating immune cell analysis, with immune cell fraction data retrieved from the TIMER website [18]. Correlations between genes and Tracking Tumor Immunophenotype (TIP) scores, and autocorrelations between TIP scores were calculated using spearman correlation analysis and visualized using the linkET package [19].

Single-cell sequencing and drug sensitivity assessment

Single-cell sequencing, a powerful tool for analyzing rare cell types or phenotypic shifts within cellular populations, is increasingly applied in cancer research. To discern cell subtypes and investigate the distribution of CRGs-based signature genes across various single-cell subtypes, we analyzed a GBM single-cell RNA-seq dataset (GSE102130). Samples were randomly chosen and subjected to rigorous quality control using the “Seurat” R package. After normalization and downscaling genes with large variance coefficients through GBM, we performed clustering analysis and subtype annotation with “SingleR” to determine CRGs-based signature gene expression levels [20].

Drug response prediction is crucial for personalizing treatment strategies. Utilizing the Genomics of Drug Sensitivity in Cancer (GDSC) [21], The Cancer Therapeutics Response Portal (CTRP, https://portals.broadinstitute.org/ctrp) [22] and PRISM (https://depmap.org/portal/prism/) databases [23], we predicted the response of GBM samples in high- and low-risk groups to anticancer drugs using the “oncoPredict” and “ggplot2” R packages. We also assessed the expression of the central ARGs across different immunotherapy datasets.

Spatial transcriptome analysis of hub gene

To ascertain the cellular makeup of each spot on tissue slides, we applied deconvolution techniques, coupled with rigorous quality control of single-cell transcriptomic data. Criteria included the count of expressed genes, UMIs, and mitochondrial RNA percentages per cell. The CIBERSORTx software’s get_enrichment_matrix and enrichment_analysis functions enabled us to create an enrichment score matrix. The Seurat package’s SpatialFeaturePlot function illustrated these scores, with intensity of color reflecting the abundance of cell types at each spot [24]. Spearman correlation analysis, visualized using the linkET package, was performed to determine the relationships between cell abundance across spots and between cell abundance and gene expression.

RNA isolation and reverse Transcription-Quantitative Polymerase Chain Reaction (RT-qPCR)

To corroborate the expression levels of the key gene identified for the prognostic signature in both GBM and normal brain tissues, we conducted RT-qPCR analyses. Total RNA was isolated from tissue samples utilizing the Total RNA Extraction Kit (Cat: R1200, Solarbio Life Sciences, Beijing, China) following the provided protocol. We then synthesized cDNA from 500 ng of total RNA using the PrimeScriptTM RT reagent kit (Perfect Real Time). The PCR amplifications were performed with specific primers for NDUFB2 (forward: 5’-TGAGCCCCGGTATAGACAGT-3’, reverse: 5’-ATACGGAAAGTGACCCAGCAC-3’) and GAPDH (forward: 5’-GAAAGCCTGCCGGTGACTAA-3’, reverse: 5’-GCCCAATACGACCAAATCAGAG-3’). The relative expression of NDUFB2 was quantified using the Formula 2^−(ΔΔCt).

Western blot

For protein expression analysis, Western blotting was employed. Proteins from GBM samples were harvested and quantified using a BCA protein assay kit (Solarbio). The lysates were prepared with loading buffer (Solarbio), denatured at 95 °C for 5 min, and subsequently separated on 15% SDS-PAGE gels (Solarbio). Proteins were then transferred to PVDF membranes (0.22-µm pore size, Millipore, Billerica, MA, USA) using a transfer system (Bio-Rad, Hercules, CA, USA). The membranes were blocked with 5% BSA (Solarbio) for 1 h at ambient temperature and incubated with primary antibodies targeting NDUFB2 (1:500, Proteintech, Cat# 17614-1-AP) and GAPDH (1:5000, Proteintech, Cat# 10494-1-AP) overnight at 4 °C. Following three 5-minute washes with TBST (Solarbio), the membranes were treated with fluorescent dye-labeled secondary antibodies (1:10,000; LI-COR Biosciences, Lincoln, NE, USA) for 3 h at 4 °C. After three additional TBST washes, the protein bands were visualized using an Odyssey infrared imaging system (LI-COR) and analyzed densitometrically.

Immunohistochemistry (IHC)

The Human Protein Atlas (HPA) project, initiated in 2003 and funded by the Knut & Alice Wallenberg Foundation in Sweden, is a comprehensive resource that maps the distribution and abundance of over 24,000 human proteins across various tissues, including healthy and tumor samples, cell lines, and blood cells. This database (accessible at https://www.proteinatlas.org/) [25] presents its findings through immunohistochemical staining images, accompanied by expert annotations. We consulted the HPA to confirm the protein-level expression of the aforementioned hub gene.

We obtained tissue samples from 22 glioblastoma multiforme (GBM) patients who underwent surgical procedures between January 2018 and December 2023 at the First Affiliated Hospital of Guangxi Medical University (GXMU). All patients provided informed consent, and the study was conducted under a protocol approved by the Ethics Committee of the same institution. The samples, embedded in paraffin, were treated with NDUFB2 antibody (1:200, Proteintech, Cat# 17614-1-AP) overnight following deparaffinization and antigen retrieval via microwave. Immunostaining utilized the avidin-biotin peroxidase method, with hematoxylin used for counterstaining. Two independent pathologists evaluated each sample. They randomly selected five fields of view for each sample, grading them based on staining intensity (0: no staining, 1: pale yellow, 2: light brown, 3: dark brown) and coverage (0: 0%, 1: 1–25%, 2: 26–50%, 3: 51–75%, 4: 76–100%). A final score was calculated by multiplying the intensity score (0–3) and the coverage score (0–4). A final score ≥ 8 was defined as high expression (positive staining), while a score < 8 was defined as low expression (negative staining) for statistical analysis.

Cell culture

The human glioblastoma cell line U251 was purchased from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China). The cells were cultured in DMEM (Gibco, USA) supplemented with 10% FBS (Hyclone, USA) and 1% penicillin/streptomycin (Invitrogen, USA) in a humidified atmosphere with 5% CO₂ at 37 °C. Mycoplasma contamination was checked using the MycoAlert Mycoplasma Detection Kit.

Small interfering RNA (siRNA) transfection

To knock down the expression of NDUFB2, U251 cells were transiently transfected with small interfering RNAs (siRNAs). Three distinct siRNA sequences targeting NDUFB2 (si-NDUFB2#1, si-NDUFB2#2, si-NDUFB2#3) and a non-targeting negative control siRNA (si-NC) were designed and synthesized by GenePharma, Suzhou, China. Transfection was performed using Lipofectamine 3000 (Invitrogen, USA) according to the manufacturer’s protocol. The siRNA sequences were as follows:

• si-NDUFB2#1:

  • Sense: 5'-GCUGACCAGAUCCCAGGUGTT-3'
  • Antisense: 5'-CACCUGGGAUCUGGUCAGCTT-3'
• si-NDUFB2#2:

  • Sense: 5'-GGGUCACUUUCCGUAUCCUTT-3'
  • Antisense: 5'-AGGAUACGGAAAGUGACCCTT-3'
• si-NDUFB2#3:

  • Sense: 5'-AGGUAUCCCUCCUGAUGAUTT-3'
  • Antisense: 5'-AUCAUCAGGAGGGAUACCUTT-3'

Transwell migration and invasion assays

The migratory and invasive abilities of U251 cells were evaluated using Transwell chambers (8 μm pore size; Corning, USA). For the invasion assay, the upper chamber membrane was pre-coated with Matrigel (Corning, USA) and allowed to solidify. For the migration assay, no Matrigel was used. Briefly, 5 × 10⁴ transfected U251 cells were resuspended in 200 µL of serum-free DMEM and seeded into the upper chamber. The lower chamber was filled with 500 µL of DMEM containing 10% FBS, which served as a chemoattractant. After 24 h of incubation at 37 °C, cells that had migrated or invaded to the lower surface of the membrane were fixed with 4% paraformaldehyde for 30 min. Non-migrated cells remaining on the upper surface were gently removed with a cotton swab. Subsequently, the membrane was stained with 0.1% crystal violet for 30 min at room temperature. After washing with PBS and air-drying, images of the stained cells were captured using a light microscope. For quantification, the number of cells in at least five randomly selected fields of view was counted.

Wound healing assay

The migratory capacity of U251 cells was further assessed by a wound healing assay. Cells were seeded into 6-well plates and cultured until they formed a confluent monolayer. A sterile 200 µL pipette tip was used to create a uniform linear scratch down the center of the monolayer. The plates were then gently washed twice with PBS to remove detached cells and cellular debris. Subsequently, cells were cultured in serum-free medium to inhibit cell proliferation. Images of the scratch wound were captured at 0, and 24 h using an inverted microscope.

Statistical analysis

Statistical evaluations were conducted utilizing the R programming environment (version 4.0.2), with a threshold of p < 0.05 set for statistical significance. To assess differences between two distinct experimental conditions, we employed the Student’s t-test, while the one-way ANOVA was utilized for comparisons involving multiple treatment groups. In order to maintain the robustness and reliability of our statistical outcomes, we implemented adjustments for multiple comparisons during differential expression analysis. Additionally, batch effects were acknowledged and rectified during the integration of data from TCGA and GEO databases.

Results

Delineating ARGs expression patterns in GBM

A total of 964 differentially expressed ARGs were identified between GBM and normal tissues (Supplementary Table 1; Fig. 1). The distribution of these genes was shown in a volcano plot (Fig. 2A). GO enrichment analysis revealed that ARGs were primarily enriched in biological processes such as “axon ensheathment” (BP), cellular components like the “main axon” (CC), and molecular functions including “calmodulin binding” (MF) (all p < 0.05) (Fig. 2B). KEGG analysis showed significant enrichment in pathways including “Huntington’s disease” and “Arginine and proline metabolism” (all p < 0.05) (Fig. 2C). The protein-protein interaction network and chromosomal locations of these ARGs are presented in Fig. 2D and E (p < 0.0001), respectively. Analysis of copy number variation (CNV) frequency showed that PLEKHA6 (gain: 11.9%, loss: 0.3%) and LRRC43 (gain: 4.0%, loss: 1.3%) had a higher frequency of gain, whereas PRKAR1B (loss: 4.5%, gain: 2.9%) and GNA12 (loss: 3.5%, gain: 2.4%) had a higher frequency of loss (Fig. 2F). Moreover, NDUFB2 exhibited a CNV gain in 2.4% of the cases and a CNV loss in 2.1%.

Fig. 1.

Flowchart of the study. The diagram illustrates the comprehensive workflow of this study. The process includes (1) data acquisition from public databases, (2) identification of differentially expressed anesthesia-related genes (ARGs) between glioblastoma (GBM) and normal tissues, (3) functional enrichment analyses (GO and KEGG), (4) consensus clustering to identify molecular subgroups, (5) construction and validation of an ARG-based prognostic model, (6) identification and multi-level analysis of the hub gene NDUFB2, and (7) experimental validation of NDUFB2 expression

Fig. 2.

Identification and functional analysis of differentially expressed ARGs in GBM.A Volcano plot showing differentially expressed ARGs between GBM (n = 150) and normal brain tissues (n = 5). Red dots represent upregulated genes, and blue dots represent downregulated genes. The vertical dashed lines indicate a |log2(Fold Change)| >1, and the horizontal dashed line indicates a p-value < 0.05. B Gene Ontology (GO) enrichment analysis of the 964 ARGs. The bar plot displays the top enriched terms in Biological Process (BP), Cellular Component (CC), and Molecular Function (MF) categories (all p < 0.05). C Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis. The bubble chart shows the most significantly enriched pathways (p < 0.05). D Protein-protein interaction (PPI) network of the ARGs constructed using the STRING database and visualized with Cytoscape. The network was found to be statistically significant (p < 0.0001). E Chromosomal locations of the ARGs. F Frequency of copy number variation (CNV) alterations for the top ARGs, showing the percentage of samples with CNV gain (red) and loss (blue)

Stratification of GBM subgroups via consensus clustering

Based on the expression of 964 ARGs, consensus clustering stratified the GBM cohort into two optimal subgroups, designated Cluster A and Cluster B (k = 2) (Fig. 3A). Principal Component Analysis (PCA), t-SNE, and UMAP plots visually confirmed a clear separation between the two clusters (Cluster A and Cluster B) (Fig. 3A). Kaplan-Meier analysis showed that patients in Cluster B had a significantly poorer overall survival compared to those in Cluster A (log-rank test, p = 0.046) (Fig. 3B). The two clusters also exhibited distinct immune cell infiltration patterns, with significant differences observed in the abundance of Activated CD4 T cell (p < 0.01), Gamma delta T cell (p < 0.01), Mast cell (p < 0.01), Natural killer cell (p < 0.05), Regulatory T cell p < 0.01), and T follicular helper cell (p < 0.05) (Fig. 3C). GSVA revealed differential activation of KEGG pathways between the clusters (Fig. 3D). Specifically, Gene Set Enrichment Analysis (GSEA) showed that pathways such as GOCC_IMMUNOGLOBULIN_COMPLEX and KEGG_CHEMOKINE_SIGNALING_PATHWAY were significantly enriched in Cluster B (p < 0.05) (Fig. 3E, Supplementary Table 2).

Fig. 3.

Identification of GBM molecular subgroups based on ARG expression. Data is from the TCGA-GBM cohort (n = 150). A Consensus clustering of ARGs identified two optimal clusters (k = 2). The separation of Cluster A (blue) and Cluster B (yellow) is visualized using Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) plots. B Kaplan-Meier survival curves for overall survival (OS) between the two clusters. Patients in Cluster B exhibited significantly poorer OS (log-rank test, p = 0.046). C Boxplots comparing the infiltration levels of various immune cell types between Cluster A and Cluster B. Asterisks indicate the level of statistical significance from a Wilcoxon rank-sum test (*p < 0.05; **p < 0.01). D Heatmap of Gene Set Variation Analysis (GSVA) showing differential activation of KEGG pathways between the two clusters. E Gene Set Enrichment Analysis (GSEA) plots showing significant enrichment of the GOCC_IMMUNOGLOBULIN_COMPLEX and KEGG_CHEMOKINE_SIGNALING_PATHWAY in Cluster B (p < 0.05)

Formulation and assessment of ARGs-based prognostic model

To establish a predictive model for patient outcomes, we employed both COX regression and Lasso-Cox techniques. The patient cohort was segmented into training and testing subsets to facilitate the development and subsequent validation of the prognostic model. Lasso regression analysis resulted in a model with 11 ARGs, detailed in supplementary Table 3 (Fig. 4A and B). The expression patterns of these 11 ARGs are presented in a heatmap (Fig. 4C). Kaplan-Meier survival plots affirmed the model’s robust correlation with overall survival (OS) across all cohorts (p < 0.05) (Fig. 4D-F), linking elevated risk scores with diminished 5-year survival rates. Multivariate Cox regression analysis, including age, gender, and risk score, showed that the ARG-based risk score was an independent prognostic factor for OS in GBM patients (HR = 1.003, 95% CI = 1.001–1.005, p = 0.0093) (Fig. 4G).

Fig. 4.

Construction and validation of an 11-ARG prognostic signature. The TCGA-GBM cohort was divided into a training set (n = 105) and a testing set (n = 45). A LASSO coefficient profiles of the ARGs. The vertical dashed line indicates the optimal lambda value selected. B Ten-fold cross-validation for tuning parameter selection in the LASSO model, resulting in an 11-gene signature. C Heatmap showing the expression of the 11 model ARGs in high- and low-risk groups, along with patient survival status. D-F Kaplan-Meier survival curves for OS comparing high- and low-risk patients in the (D) entire cohort (n = 150) (p < 0.001), (E) testing cohort (n = 105) (p < 0.001), and (F) training cohort (n = 45) (p = 0.015). G Forest plot of multivariate Cox regression analysis including the risk score and clinical variables (age, gender). The risk score was an independent prognostic factor (HR = 1.003, 95% CI = 1.001–1.005, p = 0.0093)

Establishment of a prognostic nomogram for GBM patients

We devised a nomogram that amalgamates our ARGs-derived model with pertinent clinical variables (Fig. 5A). Calibration plots were employed to validate the nomogram’s predictive accuracy (Fig. 5B). The cumulative hazard curve demonstrated increased risk levels for patients classified within the high-risk nomoRisk category (Fig. 5C). Decision curve analysis (DCA) was applied to evaluate the nomogram’s effectiveness in forecasting 1-year, 3-year, and 5-year survival probabilities for GBM patients (Fig. 5D-F), showcasing its potential clinical utility.

Fig. 5.

A nomogram for predicting GBM patient survival. A Prognostic nomogram integrating the ARG-based risk score and clinical factors to predict 1-, 3-, and 5-year overall survival. B Calibration plots for the nomogram at 1, 3, and 5 years. The x-axis represents nomogram-predicted survival, and the y-axis represents actual survival. The dashed line represents the ideal prediction. C Cumulative hazard curves for patients in the high-risk and low-risk groups defined by the nomogram’s total points (nomoRisk). D-F Decision curve analysis (DCA) for the nomogram’s predictive performance at (D) 1 year, (E) 3 years, and (F) 5 years, assessing its clinical net benefit

Identification of the hub gene in ARGs

In our quest to unravel the network of differentially expressed ARGs in GBM, we pinpointed central or “hub” genes through the MCC algorithm within the cytoHubba plugin (Fig. 6A). We curated a set of candidate genes, selecting the top 15 ARGs with the most significant differential expression based on their |logFC| values, and those with the highest MCC scores from the PPI network (as listed in supplementary Table 4). This gene set was then cross-referenced with genes from our prognostic model, leading to the identification of NDUFB2 as the pivotal hub gene, distinguished by its top-ranking MCC score. A pan-cancer risk assessment underscored NDUFB2’s role as a prognostic marker for GBM patient outcomes (Fig. 6B). Expression analysis indicated that NDUFB2 levels were markedly elevated in GBM tumors compared to non-tumor tissue (p = 0.017) (Fig. 6C) and were also significantly higher across a broader spectrum of gliomas (p = 0.014) (Fig. 6D). Elevated expression of NDUFB2 correlated with reduced overall survival (OS) and progression-free survival (PFS) rates (p < 0.001 for both) (Fig. 6E, F).

Fig. 6.

Identification and characterization of NDUFB2 as a hub gene. A PPI network highlighting the top 15 hub genes identified by the MCC algorithm in cytoHubba. NDUFB2 was identified as the key hub gene present in the prognostic model. B Pan-cancer analysis showing the hazard ratio (HR) of NDUFB2 across various cancers, indicating its risk-associated role in GBM. C Boxplot of NDUFB2 expression in GBM tumors vs. normal tissues from the TCGA and GTEx databases (p = 0.017, Wilcoxon test). D Boxplot of NDUFB2 expression in gliomas vs. normal tissues (p = 0.014, Wilcoxon test). E, F Kaplan-Meier curves showing that high NDUFB2 expression is associated with poorer (E) overall survival (OS) and (F) progression-free survival (PFS) in GBM patients (n = 150, log-rank test, p < 0.001 for both). G Correlation plot illustrating the relationship between NDUFB2 expression and Tumor Inflammation Signature (TIP) scores. H Bubble plot showing the correlation between NDUFB2 expression and the infiltration levels of various immune cells. The size of the bubble represents the p-value, and the color represents the correlation coefficient. I Scatter plot showing the specific correlation between NDUFB2 expression and macrophage infiltration

Figure 6G illustrated the relationship between TIP scoring and NDUFB2 expression, including the interconnection among various scores. We employed multiple computational strategies to evaluate the association between NDUFB2 expression and the presence of immune-infiltrating cells. Notably, the correlation between NDUFB2 and different immune cell types was found to be significant, with varying cell types exhibiting positive or negative correlations with expression levels (Fig. 6H). Furthermore, Fig. 6I provides a detailed analysis of the specific association between NDUFB2 expression and macrophage infiltration.

Analysis of ARGs in single-cell data and drug sensitivity assessment

Utilizing single-cell RNA sequencing data from GSE102130, we delved into the expression profile of NDUFB2 within the tumor microenvironment (TME) (Fig. 7A). Our findings revealed a predominant expression of NDUFB2 in malignant tumor cells, hinting at its potential significance in GBM pathogenesis (p < 0.001) (Fig. 7B-F). We also studied how NDUFB2 expression relates to drug sensitivity using databases like GDSC, CTRP, and PRISM (all p < 0.001) (Fig. 7G-J). The analysis indicated that higher levels of NDUFB2 were associated with increased resistance to the drugs evaluated.

Fig. 7.

Single-cell and drug sensitivity analysis of NDUFB2. A t-SNE plot of single cells from GBM tissue (dataset GSE102130), colored by identified cell type. B-F Violin and t-SNE plots showing NDUFB2 expression is predominantly localized in malignant tumor cells compared to other cell types in the tumor microenvironment (p < 0.001). G-J Scatter plots showing the correlation between NDUFB2 expression and drug sensitivity (IC50) from GDSC, CTRP, and PRISM databases. A positive correlation indicates resistance (all p < 0.001). K ROC curves showing the AUC values for NDUFB2 expression in predicting immunotherapy response in datasets GSE126044 and GSE67501. L ROC curve for NDUFB2 expression predicting immunotherapy response in the GBM dataset PRJNA482620 (AUC = 0.623, 95% CI = 0.415–0.803). M Boxplot of NDUFB2 expression in responders vs. non-responders to immunotherapy from the PRJNA482620 dataset (p = 0.231, Wilcoxon test)

When predicting response to immunotherapy using ROC-AUC values based on NDUFB2 expression, our analysis uncovered notable variability across different datasets. NDUFB2 demonstrated superior predictive accuracy in datasets GSE126044 and GSE67501 (Fig. 7K). Within the GBM dataset PRJNA482620, NDUFB2 expression yielded an AUC of 0.623 for predicting immunotherapy response (95% CI = 0.415–0.803) (Fig. 7L). While not statistically significant (p = 0.231), a trend was observed where patients with elevated NDUFB2 expression generally exhibited a less favorable response to immunotherapy (Fig. 7M).

Spatial transcriptome analysis of NDUFB2

Spatial transcriptome analysis was performed to map the expression of NDUFB2 within the GBM tissue architecture. After deconvolution, cell types were mapped to their spatial locations (Fig. 8A-K). NDUFB2 expression was predominantly observed in regions identified as malignant tumor cells (Fig. 8M). In contrast, macrophages were localized in distinct, often non-overlapping regions (Fig. 8L). A significant inverse spatial correlation was found between NDUFB2 expression levels and macrophage abundance across the tissue spots (Pearson’s r =−0.4, p < 0.001) (Fig. 8N).

Fig. 8.

Spatial transcriptome analysis of NDUFB2 in GBM. A-J Spatial distribution plots showing the abundance of ten different deconvoluted cell types across the GBM tissue section. Color intensity indicates cell abundance per spot. K Spatial plot showing the predominant cell type assigned to each spot after deconvolution. L Spatial feature plot showing the distribution and abundance of macrophages. M Spatial feature plot showing the expression level of NDUFB2, primarily localized in tumor cell regions. N Scatter plot demonstrating a significant inverse spatial correlation between NDUFB2 expression and macrophage abundance per spot (Pearson’s r = −0.4, p < 0.001)

RT-qPCR, WB and IHC

To ascertain the expression levels of NDUFB2 in GBM and non-tumorous tissues, RT-qPCR analyses were carried out. The findings highlighted a marked discrepancy in NDUFB2 expression between GBM and normal samples, with an upregulation of NDUFB2 observed in GBM specimens (Fig. 9A). These observations were further corroborated by Western blot (WB) analysis (Fig. 9B, C). To complement these findings, the protein expression of NDUFB2 was examined using data from the HPA database, which confirmed its elevated presence in GBM tissues (Fig. 9D, E). We further validated the protein expressions of NDUFB2 in 22 GBM patients from our cohort by IHC and observed that NDUFB2 was aberrantly expressed in 72.7% (16/22) of GBM tissues (Fig. 9F, G).

Fig. 9.

Experimental validation of NDUFB2 expression. A RT-qPCR analysis of NDUFB2 mRNA levels in GBM tissues (n = 4) and non-tumorous brain tissues (n = 4). B Representative Western blot images showing NDUFB2 protein expression in GBM (n = 4) and normal tissue (n = 4) samples. GAPDH was used as a loading control. (C) Quantification of Western blot band intensities from (B). D, E Representative immunohistochemistry (IHC) images from the Human Protein Atlas (HPA) database showing NDUFB2 protein staining in normal brain and GBM tissue. F, G Validation of NDUFB2 protein expression in an in-house cohort of 22 GBM patients (GXMU-cohort). Scale bars = 200 μm. H Western blot analysis confirmed the knockdown efficiency of NDUFB2 protein expression in U251 cells transfected with si-NC, si-NDUFB2#1, si-NDUFB2#2, and si-NDUFB2#3. I Representative images from the Transwell assays showing cells that invaded through the Matrigel-coated membrane (Invasion) and migrated through the uncoated membrane (Migration). Cells were stained with crystal violet. (Scale bar = 100 μm). J Quantitative analysis of the number of invaded and migrated cells per field of view. The results demonstrate a significant reduction in both invasion and migration in the si-NDUFB2#1 group compared to the si-NC group. K Representative images from the wound healing assay at 0 h and 24 h after scratching the confluent monolayer of U251 cells transfected with si-NC. L Representative images from the wound healing assay at 0 h and 24 h after scratching the confluent monolayer of U251 cells transfected with si-NDUFB2#1

Silencing of NDUFB2 expression inhibits the migration and invasion of U251 glioblastoma cells

To investigate the biological function of NDUFB2 in glioblastoma, we first assessed the effect of its knockdown on the migratory and invasive capabilities of U251 cells. We designed three distinct small interfering RNAs (si-NDUFB2#1, #2, and #3) to specifically target NDUFB2, along with a non-targeting negative control (si-NC). After transient transfection into U251 cells, the knockdown efficiency was confirmed by Western Blot (Fig. 9H). Based on the silencing efficiency, si-NDUFB2#1, which showed the most significant reduction in NDUFB2 expression, were selected for subsequent functional assays.

We first evaluated the effect of NDUFB2 knockdown on the invasive potential of U251 cells using a Matrigel-coated Transwell invasion assay. As shown by crystal violet staining, the number of cells that successfully invaded through the Matrigel barrier and the porous membrane was substantially reduced in the si-NDUFB2#1 transfected group compared to the si-NC group (Fig. 9I). Quantitative analysis further confirmed that silencing NDUFB2 expression led to a significant impairment of the invasive capacity of U251 cells (P < 0.01) (Fig. 9J). Next, to determine whether NDUFB2 also affects cell migration, we performed a Transwell migration assay without the Matrigel coating. Consistent with the invasion results, the number of cells migrating to the lower chamber was markedly decreased in the NDUFB2-depleted group relative to the control group (Fig. 9I). The quantitative data verified that NDUFB2 knockdown significantly inhibited the migratory ability of U251 cells (P < 0.01) (Fig. 9J).

To further corroborate the findings on cell migration, a wound healing assay was conducted. A uniform scratch was made in a confluent monolayer of transfected cells. After 24 h, cells in the si-NC group had migrated effectively to close the wound gap (Fig. 9K). In contrast, cells transfected with si-NDUFB2#1 displayed a significantly delayed wound closure rate, indicating impaired migratory function (Fig. 9L).

Discussion

Glioblastoma (GBM) is an exceedingly invasive tumor of the central nervous system, and its study has been a focal point in the field of neuro-oncology. Despite certain advancements in diagnostic and therapeutic modalities in recent years, the median survival time for patients with GBM remains dismal, typically not exceeding 15 months. Current research priorities include elucidating the molecular underpinnings of GBM, identifying novel therapeutic targets, and developing more efficacious treatment strategies [26]. Advances in molecular biology have significantly enhanced our understanding of the biological characteristics of GBM. Recent research has identified many genes and pathways linked to GBM development, including TERT promoter, IDH1/2 mutations, and EGFR amplification [27, 28]. These molecular markers not only facilitate the classification of GBM but also lay the groundwork for personalized treatment approaches. However, the pronounced heterogeneity and the complexity of the tumor microenvironment in GBM pose formidable challenges to treatment [29].

In recent years, there has been a growing body of research focusing on the impact of anesthetic agents on tumor cells and the tumor microenvironment. Some retrospective studies and animal experiments suggest that general anesthetics, particularly inhalational agents, may promote tumor cell proliferation and metastasis [30]. For instance, certain inhalational anesthetics have been implicated in diminishing the body’s antitumor immune response by inhibiting the function of natural killer cells and T cells [31, 32]. Anesthetic drugs play a pivotal role in the surgical management of GBM, yet emerging studies suggest that these agents may also influence the biological characteristics of the tumor and the long-term prognosis of patients. In this study, we constructed an anesthetic drug-based model, ARGs, utilizing public datasets, which demonstrated commendable performance in predicting the prognosis of patients with GBM. Given the potential for a more complex role of anesthetic drugs in the treatment of GBM, future studies are warranted to delve deeper into the influence of these drugs on tumor biology.

A key finding of our study is the stratification of GBM patients into two molecularly distinct clusters based on ARG expression, with Cluster B exhibiting significantly poorer overall survival. This finding transcends simple classification; it suggests that a patient’s intrinsic response pathways to anesthetic agents may be indicative of their underlying tumor biology. Critically, the poorer prognosis of Cluster B was mechanistically linked to an altered tumor microenvironment, characterized by significant differences in immune cell infiltration and the enrichment of pathways like the “KEGG_CHEMOKINE_SIGNALING_PATHWAY”. This aligns with previous studies showing that chemokine signaling is pivotal for immune evasion and tumor progression in GBM [33, 34]. Unlike previous GBM classifications based on genomics or transcriptomics, our ARG-based approach provides a novel perspective, implying that perioperative management could potentially interact with a pre-existing molecular landscape to influence patient outcomes.

NADH: ubiquinone oxidoreductase subunit B2 (NDUFB2) is a gene within the human genome that encodes a protein component of mitochondrial complex I (also known as NADH: ubiquinone oxidoreductase) [35]. Complex I is the first and largest enzyme complex of the mitochondrial electron transport chain (ETC), primarily catalyzing the transfer of electrons from NADH to coenzyme Q (CoQ) and concurrently pumping protons from the mitochondrial matrix into the intermembrane space, thereby generating a transmembrane proton gradient. This gradient serves as the driving force for ATP synthesis, hence complex I plays a crucial role in ATP production [36]. Our identification of NDUFB2 as a key prognostic hub gene is particularly noteworthy. NDUFB2, a subunit of mitochondrial Complex I, plays a critical role in cellular metabolism and ATP production [35]. While its role in GBM is underexplored, dysregulation of mitochondrial function is a known hallmark of cancer. Previous studies have implicated NDUFB2 in other malignancies; for instance, it has been linked to prognosis in head and neck squamous cell carcinoma, urothelial carcinoma and Alzheimer’s disease [37–39]. Our findings align with this emerging evidence, suggesting that NDUFB2-mediated metabolic reprogramming could be a crucial driver of GBM aggressiveness. This provides a strong rationale for future studies investigating NDUFB2 as a therapeutic target aimed at disrupting tumor cell metabolism. More importantly, our spatial transcriptomic analysis revealed a significant inverse correlation between NDUFB2 expression and macrophage infiltration. This novel finding suggests a potential mechanism whereby NDUFB2-driven metabolic changes in tumor cells may create a non-permissive or “cold” immune microenvironment, possibly by excluding or repolarizing macrophages, thus contributing to GBM’s aggressiveness.

Our study has several important limitations that must be acknowledged. A primary and significant limitation is that the initial identification of ARGs was based on a GEO dataset with a very small sample size (n = 3 per group). This inherently limits the statistical power and generalizability of the initial gene set and means the findings should be considered exploratory. Consequently, the term “anesthesia-related genes” must be interpreted with caution, representing genes whose expression was altered in a specific experimental context rather than a universally validated set. This underscores the urgent need for validation in larger, prospectively collected clinical cohorts where the type and duration of anesthesia are meticulously recorded. Second, our prognostic model was developed and validated using retrospective data from public databases. While statistically robust, this bioinformatic approach does not establish causality. Third, while we provided preliminary experimental validation for NDUFB2 expression, our study is correlational and does not establish causality. A crucial limitation is the absence of functional studies to elucidate the precise molecular mechanism by which NDUFB2 influences GBM progression. Experiments such as gene knockdown or overexpression using CRISPR/Cas9 in GBM cell lines and patient-derived organoids are required to test our hypothesis that NDUFB2 drives metabolic reprogramming and modulates the immune microenvironment. Such studies are essential to confirm whether NDUFB2 is a driver of the malignant phenotype and a viable therapeutic target.

Conclusion

In conclusion, this study presents an exploratory analysis linking anesthesia-related gene expression to GBM prognosis. Despite the limitations of the initial gene discovery phase, our comprehensive validation pipeline identifies NDUFB2 as a robust and independent adverse prognostic factor. Our findings suggest a novel potential mechanism linking NDUFB2-mediated metabolic reprogramming to the creation of an immunosuppressive tumor microenvironment. While further functional validation and confirmation in larger clinical cohorts are essential, this work provides a valuable hypothesis and a promising biomarker, opening new avenues for investigating the interplay between perioperative factors and GBM biology.

Authors’ contributions

Conceptualization, Y.B.X.; Data curation, J.W.W. and L.Y.Z.; Funding acquisition, Y.B.X.; Project administration, Y.B.X.; Resources, Q.L. and Y.B.X.; Software, J.W.W.; Supervision, J.W.W.; Validation, J.W.W. and L.Y.Z.; Visualization, Q.L.; Writing – original draft, J.W.W.; Writing – review & editing, Y.B.X.

Funding

This study was supported by Special Fund of Neurotoxicity of General Anesthetics and Its Prevention and Treatment Innovation Team of the First Affiliated Hospital of Guangxi Medical University (No. YYZS2022001), Guangxi Clinical Research Center for Anesthesiology (No. GK AD22035214), Guangxi Key Research and Development Program (No. AB24010066).

Data availability

The TCGA-GBM cohort dataset is available at The Cancer Genome Atlas (TCGA) website (https://xena.ucsc.edu/). The GSE179004, GSE102130, GSE126044 and GSE67501 cohort datasets for this study is available at Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/gds/?term=).

Declarations

Ethics approval and consent to participate

Informed consent was acquired from all patients and the study was approved by the ethics committee of The First Affiliated Hospital of Guangxi Medical University (Institutional Review Board approval number, 2023-K102-01) and conducted in accordance with Good Clinical Practice and the Declaration of Helsinki and its latest amendments.

Competing interests

The authors declare no competing interests.