Development and validation of a novel glioblastoma prognostic model: identifying STOX1 and ZNF248 as ferroptosis-related biomarkers

Abstract

Background

Despite recent advances in glioblastoma (GBM) therapies, patient survival remains dismal. Existing prognostic markers lack sufficient accuracy, and the resistance of GBM to chemotherapy underscores the need for both improved predictive models and new therapeutic targets.

Methods

We analyzed transcriptomic data from GBM and non-tumor brain tissues in the TCGA and GTEx databases to identify differentially expressed genes (DEGs). A prognostic signature was then constructed via univariate Cox regression, LASSO selection, and multivariate Cox analysis. The resulting prognostic model was validated in an independent GEO GBM cohort using Kaplan–Meier and log-rank testing. We further compared mutation patterns and immune infiltration between high- and low-risk groups. Multi-omics integration highlighted candidate tumor suppressors, which were functionally assessed in GBM cell lines and an orthotopic mouse model for the effects on ferroptosis sensitivity.

Results

A 14-gene prognostic signature was developed and robustly stratified GBM patients into high- and low-risk groups with significantly different overall survival. High-risk tumors showed elevated PTEN mutation frequency, enhanced immunosuppressive microenvironments with increased PD-L1 and regulatory T cells, and distinct co-mutation patterns. STOX1 and ZNF248 were remarkably downregulated in GBM tissues. Their overexpression in vitro suppressed GBM cell proliferation and sensitized cells to ferroptosis. In vivo, nude mice bearing STOX1-overexpressing GBM cells showed prolonged survival under ferroptosis-inducing conditions.

Conclusion

We present a validated 14-gene prognostic model that accurately predicts GBM patient outcomes and reveals STOX1 and ZNF248 as novel tumor suppressors related to ferroptosis. Targeting STOX1 and ZNF248 may overcome the resistance of GBM to ferroptosis and improve therapeutic efficacy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-025-04126-w.

Introduction

Glioblastoma (GBM) is the most common and aggressive primary malignant brain tumor in adults, accounting for approximately 15% of all central nervous system (CNS) tumors [1]. Despite the adoption of Stupp protocol, which combines maximal surgical resection with radiotherapy and temozolomide-based chemotherapy, over the past two decades [2], the prognosis for GBM patients remains dismal. The median overall survival following standard-of-care treatment is approximately 16.9 months [3, 4], with fewer than 5% of patients surviving beyond five years [5].

This poor prognosis is largely attributable to several hallmark features of GBM, including high intratumoral heterogeneity, extensive invasiveness, immune dysregulation, robust angiogenesis, and pronounced resistance to conventional therapies [6]. One defining characteristic of GBM is its dense and disorganized microvasculature, which plays a central role in supporting tumor growth, promoting therapeutic resistance, and enabling immune evasion [7]. The tumor microenvironment is full with pro-inflammatory signals, immunosuppressive cells, and abundant vascular endothelial cells that support neovascularization [8]. As a result, anti-angiogenic therapies, most notably bevacizumab, a monoclonal antibody targeting vascular endothelial growth factor A (VEGF-A), have been explored as treatment options. Although these agents have shown efficacy in improving progression-free survival (PFS), they have failed to significantly prolong overall survival (OS), underscoring the urgent need for novel therapeutic strategies [9, 10].

Over one decade ago, advances in high-throughput sequencing technologies have greatly expanded our understanding of GBM’s molecular landscape. These advances have led to the identification of key oncogenic drivers such as PTEN, TP53, EGFR, NF1, RB1, and IDH1, and enabled the classification of GBM into molecular subtypes with distinct prognostic implication [11]. More recently, the emergence of single-cell RNA sequencing has provided unprecedented insights into the cellular heterogeneity of GBM, offering a more refined understanding of tumor subtypes and dynamic molecular changes within the tumor microenvironment [12, 13]. However, despite these advances, the clinical utility of molecular markers in GBM remains limited. Robust and reliable biomarkers for predicting patient outcomes and therapeutic response are still lacking, highlighting a critical gap in current GBM management. Therefore, the development of reliable prognostic models and the identification of novel therapeutic targets based on these models are essential steps toward improving survival outcomes and quality of life for GBM patients.

To address this gap, our study aims to construct and validate a prognostic model for GBM. By comparing gene expression between GBM and normal brain tissues, we identified differentially expressed genes associated with patient survival. Using a comprehensive bioinformatics pipeline, we refined these candidates into a multi-gene prognostic signature capable of stratifying patients into risk subgroups with distinct survival outcomes. In addition, we conducted functional analyses of two key genes within the signature, STOX1 and ZNF248, to investigate their roles in modulating ferroptosis sensitivity. These insights may help uncover ferroptosis-related mechanisms and provide a potential strategy for overcoming GBM’s resistance to conventional therapies.

Materials and methods

Data acquisition and differential expression analysis

Transcriptomic data for GBM patients (N = 168) were obtained from the Cancer Genome Atlas Program (TCGA). Transcriptomic data for normal brain tissues (N = 1157) were obtained from the Genotype-Tissue Expression (GTEx) databases. TCGA data were downloaded using the GDC Data Transfer Tool (v1.6.1), whereas GTEx data were acquired directly from the GTEx Portal. To reduce noise, low-expression genes (average count < 1) were excluded from the raw count data. TCGA counts data were normalized using the variance stabilizing transformation (VST) provided by DESeq2 (v1.36.0) to generate a log-transformed count matrix, while GTEx transcripts per million (TPM) data were uniformly converted via log₂(TPM + 1). Due to the different sequencing platforms and pipelines of the TCGA and GTEx datasets, significant batch effects were anticipated. These batch effects were mitigated using the ComBat method within the SVAseq package (v3.44.0) by designating the sample source (TCGA or GTEx) as the batch variable. The integrated expression matrix, formatted as log₂(TPM + 1), comprised gene expression profiles from GBM tumor tissues, adjacent non-tumor tissues, and normal brain tissues, and served as the foundation for subsequent differential expression analyses. Additionally, GBM transcriptomic data and corresponding prognostic information were manually curated from the GEO dataset GSE108474 to further support the analysis.

Differentially expressed genes (DEGs) were statistically identified using the Wald test, and P-values were adjusted using the Benjamini-Hochberg method. Genes were classified as differentially expressed if they exhibited an adjusted P-value < 0.05 and |log₂ (FC)| > 1. Results were visualized through volcano plots generated with ggplot2 (v3.4.2) to illustrate the relationship between |log₂ (FC)| and adjusted P-values. The heatmaps were created with the pheatmap package v1.0.12 to depict sample expression patterns after Z-score normalization and hierarchical clustering. The complete list of DEGs, including gene names, log₂ (FC), raw P-values, and adjusted P-values, was exported for subsequent functional enrichment analyses and validation steps.

Gene set enrichment analysis (GSEA)

GSEA was performed using the Broad Institute’s GSEA software (version 4.3.2). The normalized expression matrix was reformatted into the required GCT file format, and the sample groupings (GBM vs. normal brain tissue) were specified in a corresponding CLS file. The analysis utilized the Hallmark gene sets from MSigDB (h.all.v2023.1.Hs.symbols.gmt), which comprise 50 predefined biological pathways. Genes were ranked by their Signal-to-Noise Ratio (SNR), defined as the difference in mean expression between GBM and normal tissues divided by the sum of their standard deviations. The enrichment score (ES) for each gene set was calculated using a random walk algorithm that traverses the ranked list. To assess statistical significance, 1000 permutation tests were performed, yielding nominal p-values that were then adjusted via the False Discovery Rate (FDR) method (with FDR < 0.05 considered significant). Normalized Enrichment Scores (NES) were determined for the comparisons across gene sets. Enrichment plots were generated to illustrate the distribution of each gene set along the ranked list. The leading-edge subset of genes that contributed most to the enrichment signal was extracted for further validation.

Cox regression and LASSO regression analyses

Transcriptomic data along with clinical information for GBM patients were obtained from the TCGA-GBM dataset. Clinical data included survival time (in days) and survival status. Samples lacking survival data or with non-positive survival times were excluded from further analysis. The DEGs identified in earlier analyses and provided in log₂(TPM + 1) format served as the starting point for regression analyses.

Univariate Cox proportional hazards regression was performed for each DEG using the survival package (v3.5–7) in R (v4.3.1) with gene expression treated as a continuous variable. The output included the hazard ratio (HR), 95% confidence interval (CI), and p-value with a significance threshold of p < 0.05. Genes that showed a significant association with overall survival were then incorporated into a multivariate Cox regression model that adjusted for clinical covariates such as age, sex, and tumor grade (e.g., IDH mutation status). Kaplan–Meier survival curves were generated using the survminer package (v0.4.9) by splitting patients at the median gene expression level, and differences between groups were assessed using the log-rank test.

To further narrow down the candidate prognostic genes and construct a robust risk prediction model, Least Absolute Shrinkage and Selection Operator (LASSO) regression was conducted using the glmnet package (v4.1–8) in R. The input for this analysis included the expression matrix of the DEGs that were significant in univariate Cox analysis, along with their corresponding survival data. The optimal penalty parameter (λ) was determined via 10-fold cross-validation by identifying the λ that minimized the partial likelihood deviance. Genes with nonzero coefficients at this optimal λ (λ_min) were selected to form the prognostic signature. For each patient, a risk score was calculated by summing the products of the selected gene expression values and their corresponding regression coefficients, as follows:

Patients were then stratified into high-risk and low-risk groups based on the median risk score. Kaplan–Meier survival analysis with the log-rank test was used to compare survival outcomes between the two groups. Finally, the predictive performance of the model was evaluated through time-dependent Receiver Operating Characteristic (ROC) curves using the timeROC package (v0.4), which provided Area Under the Curve (AUC) values at 1, 2, and 3 years.

In addition, survival analyses of STOX1 and ZNF248 were conducted in glioma patients with all grade or Grade IV GBM using the Chinese Glioma Genome Atlas (CGGA) database. Datasets from cohort 1 and cohort 2 are available in the China National Center for Bioinformation (Accession Numbers: PRJCA001746 and PRJCA001747), while microarray data from cohort 3 can be downloaded directly from the CGGA website.

Gene mutation analysis

Somatic mutation data for the TCGA-GBM cohort were obtained from the Genomic Data Commons (https://portal.gdc.cancer.gov/) by applying the following filters: project “TCGA-GBM”, data category “Simple Nucleotide Variation”, and data type “Masked Somatic Mutation” in MAF format. These files were downloaded using the GDC Data Transfer Tool (v1.6.1). Patients were classified into high-risk and low-risk groups based on median risk scores derived from the previous LASSO regression analysis. The raw MAF files were then imported into R via the maftools package (v2.16.0), where only functional exonic mutations (including missense, nonsense, frameshift, and splice site variants) were retained after filtering out non-coding and silent mutations.

Mutation rates for each gene were calculated as the proportion of samples with at least one mutation. Genes showing significantly different mutation frequencies between the risk groups were identified using Fisher’s Exact Test with p-values adjusted by the Benjamini-Hochberg method (FDR < 0.05). Data visualization was achieved through oncoplots (via maftools) to display common mutated genes, mutation frequency bar charts created with ggplot2 (v3.4.2), and heatmaps produced by the ComplexHeatmap package (v2.14.0) to illustrate the patterns of co-occurrence and mutual exclusivity, with significance evaluated using the hypergeometric test (P < 0.05). Additionally, the impact of representative EGFR, PTEN, and TP53 mutations on GBM patient survival was evaluated, and the regulatory effect of TP53 knockdown on STOX1 and ZNF248 expression was analyzed using the KnockTF 2.0 tool.

Immune infiltration analysis

TCGA-GBM primary tumor transcriptome data, normalized as log₂(TPM + 1), were used for immune infiltration analysis. Patients were divided into high-risk and low-risk groups based on the median risk score obtained from the previous LASSO regression analysis. The relative abundances of 28 immune cell types were estimated using the CIBERSORT algorithm (CIBERSORT v1.0.0) in R with the LM22 signature matrix and 1000 permutations. Only samples with P-values < 0.05 were included in further analysis. The group differences were assessed with the Wilcoxon rank-sum test, applying Benjamini–Hochberg adjustment to control FDR < 0.05. Single-sample gene set enrichment analysis (ssGSEA) was performed using the GSVA package (v1.48.0) in R, employing ImmuneSigDB gene sets that cover 28 immune cell markers and various immune-related functions. Differences in enrichment scores between the risk groups were evaluated using the Wilcoxon rank-sum test (FDR < 0.05). To display the results, heatmaps (using pheatmap v1.0.12), boxplots (using ggplot2 v3.4.2), and Pearson correlation scatter plots (using corrplot v0.92) were generated, illustrating the distribution of immune cell populations and their associations with risk scores. In addition, the relationship of STOX1 and ZNF248 with immunotherapy response was further assessed using the Tumor Immune Dysfunction and Exclusion (TIDE) platform.

Spatial transcriptomics data analysis

Spatial transcriptomics data for GBM tissues were obtained from Dryad and the 10x Genomics website. This dataset, generated on the 10x Genomics Visium platform, included FASTQ files containing spatial barcode and Unique Molecular Identifiers (UMI) data, a spatial coordinate file (tissue_positions_list.csv), and a high-resolution H&E-stained image (tissue_hires_image.png). Data integrity was verified via MD5 checksum, and alignment was performed against the GRCh38 reference genome (Ensembl v104). Raw FASTQ files were processed using 10x Genomics Space Ranger (v2.0.1) with the STAR aligner (v2.7.10a). The command spaceranger count was used to produce gene expression count matrices for individual spots, while spaceranger aggr combined multiple samples with their associated H&E images into a single normalized matrix. Spots with fewer than 500 UMIs or fewer than 200 detected genes were excluded from the analysis. The resulting gene expression matrix was normalized with SCTransform in Seurat (v4.3.0) to account for differences in sequencing depth and to correct for batch effects. Spatial clustering was performed using Seurat’s FindClusters function with the Louvain algorithm (resolution = 0.8), and clusters were annotated based on known GBM markers. Differential expression analysis between clusters was conducted using the Wilcoxon rank-sum test, with significant genes defined by an FDR < 0.05 and |log₂ (FC)| > 1. SpatialFeaturePlot in Seurat was then employed to overlay the expression patterns of key genes onto the H&E images.

Single-Cell sequencing analysis

Single-cell transcriptomic data were obtained from GEO (accession GSE84465) as well as the public 10x Genomics “Human Glioblastoma” dataset. Raw FASTQ files were processed via Cell Ranger (v7.1.0) with the “--include-introns” option, aligning reads to GRCh38 (Ensembl v104) to generate gene expression matrices. Cells with fewer than 500 UMIs, fewer than 200 genes, or with mitochondrial gene content exceeding 10% were excluded. The filtered data were normalized using SCTransform in Seurat (v4.3.0), followed by Principal Component Analysis (PCA) (using the top 20 principal components) and Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction and visualization. Clustering was performed using the Louvain algorithm (resolution = 0.6), with clusters subsequently annotated based on established marker genes (e.g., EGFR for tumor cells). Differential expression analysis between clusters was conducted using the Wilcoxon rank-sum test (with FDR < 0.05 and |log₂ (FC)| > 0.5 as significance thresholds), and the differences in cell abundances were evaluated using Fisher’s Exact Test (FDR < 0.05). Functional enrichment analysis was carried out with clusterProfiler (v4.8.1) using both Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) databases. Data visualizations included UMAP plots, dot plots, and enrichment graphs.

Cell culture

GBM cell lines (U251MG, LN229, T98G, and U87MG) were obtained from the Cell Resource Center of the Shanghai Institute of Biochemistry and Cell Biology (SIBCB), human astrocyte cell line (SVG P12) and human immature oligodendrocyte cell line (MO3.13) were purchased from Cyagen Oricell. SVG P12, MO3.13, LN229, and U87MG cells were maintained in high-glucose DMEM (Gibco, catalog no. 11965-092) with 10% fetal bovine serum (FBS) (Gibco, catalog no. 10099-141). U251MG and T98G cells were cultured in DMEM/F12 (Gibco, catalog no. 11320-033) and RPMI-1640 (Gibco, catalog no. 11875-093), respectively, each supplemented with 10% FBS. All media contained 1% penicillin-streptomycin (Gibco, catalog no. 15140-122). Cells were incubated in a humidified incubator at 37 °C with 5% CO₂, and the medium was changed every 2–3 days. When cultures reached 80–90% confluence, cells were washed with PBS (Gibco, catalog no. 10010-023) and detached with 0.25% trypsin-EDTA (Gibco, catalog no. 25200-056) at 37 °C for 2–5 min. The trypsinization was terminated by adding medium containing 10% FBS, and cells were collected by centrifugation at 300 xg for 5 min. Cells were subsequently reseeded at a split ratio of 1:5 and monitored using an inverted microscope (Olympus IX71). All cell culture procedures were performed under aseptic conditions within a Biosafety Level II cabinet (Thermo Fisher Scientific, model 1300 Series A2). Mycoplasma testing was routinely performed using a PCR-based detection kit (Takara, catalog no. 6601). For cryopreservation, cells were suspended in a freezing medium composed of 10% dimetilsolfossido (DMSO) (Sigma-Aldrich, catalog no. D2650) and 90% FBS, initially stored overnight at − 80 °C, then transferred to liquid nitrogen for long-term storage.

To determine the effects of STOX1 or ZNF248 on ferroptosis resistance, transfected cells (see below) were treated with 2 µM ferrostatin-1 (Fer-1) (Sigma‑Aldrich, catalog no. SML0583) in combination with varying concentrations of erastin (MedChemExpress, catalog no. HY‑15760) as indicated. Erastin was prepared as a 10 mM stock in DMSO and diluted to the final concentrations in medium containing 10% FBS. Fer-1 was also prepared as a 10 mM stock in DMSO and diluted to 2 µM for both cell lines. Control groups received an equivalent volume of 0.1% DMSO.

Plasmid transfection

Twenty-four hours before transfection, U251MG and T98G cells were seeded in 6‑well plates at a density of 1 × 10⁵ cells per well and cultured to 50%–70% confluence. Transfection was performed using Lipofectamine™ 3000 reagent (Invitrogen, catalog no. L3000015). For each well, 2 µg of plasmid DNA, either pcDNA3.1-STOX1 (based on the NM_001130159.3 sequence from the National Center for Biotechnology Information (NCBI) database), pcDNA3.1-ZNF248 (based on NM_001267597.2 from NCBI), or the pcDNA3.1 empty vector control (all obtained from Aikesheng Biotechnology), was diluted in 100 µL Opti‑MEM (Gibco, catalog no. 31985‑070); simultaneously, 5 µL Lipofectamine 3000 was also added to the mixture. After a 15‑minute incubation at room temperature to allow complex formation, the transfection mixture was added to the cells. Six hours post-transfection, the medium was replaced with fresh medium containing 10% FBS, and the cells were further cultured. Forty-eight hours after transfection, cells were harvested for downstream analyses.

Quantitative Real-Time PCR (qRT‑PCR)

Total RNA was isolated from cultured cells using the RNeasy Mini Kit (Qiagen, catalog no. 74104) following the manufacturer’s protocol. RNA purity and concentration were measured with a NanoDrop spectrophotometer (Thermo Fisher Scientific), ensuring A260/A280 ratios between 1.8 and 2.0. RNA integrity was verified by agarose gel electrophoresis, which showed distinct 28 S and 18 S rRNA bands. From 1 µg of total RNA, first-strand cDNA was synthesized using the iScript™ Reverse Transcription Supermix (Bio‑Rad, catalog no. 1708840) under the following conditions: pre-incubation at 25 °C for 5 min, extension at 42 °C for 30 min, and enzyme inactivation at 85 °C for 5 min. Primers were designed with Primer‑Premier software. The selection of endogenous controls was based on pre-experimental validation. GAPDH was used as the internal control for cell line comparisons, while ACTB was used as the internal control for overexpression and erastin treatment experiments. The primer sequences were as follows:

ACTB

Forward: 5’‑CACCATTGGCAATGAGCGGTTC‑3’.

Reverse: 5’‑AGGTCTTTGCGGATGTCCACGT‑3’.

CYP2E1

Forward: 5’‑ATGTCTGCCCTCGGAGTCA‑3’.

Reverse: 5’‑CGATGATGGGAAGCGGGAAA‑3’.

FTH1

Forward: TGAAGCTGCAGAACCAACGAGG.

Reverse: GCACACTCCATTGCATTCAGCC.

GAPDH

Forward: 5’‑GTCTCCTCTGACTTCAACAGCG‑3’.

Reverse: 5’‑ACCACCCTGTTGCTGTAGCCAA‑3’.

GPX4

Forward: ACAAGAACGGCTGCGTGGTGAA.

Reverse: GCCACACACTTGTGGAGCTAGA.

SLC7A11

Forward: TCCTGCTTTGGCTCCATGAACG.

Reverse: AGAGGAGTGTGCTTGCGGACAT.

STOX1

Forward: 5’‑TTGGGTGAAGTTCTTTGCTGT‑3’.

Reverse: 5’‑TTTCATCAAACGCTCCAAAAGTG‑3’.

ZNF248

Forward: 5’‑TTGGAAATGGAGCCGTATGGA‑3’.

Reverse: 5’‑CCCAGGTTTTGTCACTCACTTTA‑3’.

qRT‑PCR was performed on a CFX96 Real‑Time PCR System (Bio‑Rad) using the SsoAdvanced™ SYBR Green Supermix (Bio‑Rad, catalog no. 1725270) in a 10 µL reaction volume, which included 1 µL of cDNA, 0.5 µL of each primer (10 µM), and 5 µL of 2× SYBR Green Supermix. The amplification protocol consisted of an initial denaturation at 95 °C for 3 min, then 40 cycles of 95 °C for 10 s, 60 °C for 30 s, and 72 °C for 30 s. A melting curve analysis from 65 °C to 95 °C was conducted to confirm product specificity and ensure the absence of primer-dimers. Relative expression levels of STOX1 and ZNF248 were quantified using the 2^−ΔΔCt method with GAPDH as the endogenous control by comparing overexpression samples to those transfected with an empty vector.

Cell proliferation assay

Forty-eight hours post-transfection, the cells were trypsinized, counted, and seeded into 96‑well plates at 2 × 10³ cells per well in six replicates per condition. Cell proliferation was monitored at 24, 48, 72, 96, and 120 h using the Cell Counting Kit-8 (CCK-8) (Dojindo, catalog no. CK04). At each time point, 10 µL of CCK‑8 reagent was added to each well. Following a 2-hour incubation at 37 °C, absorbance was measured at 450 nm using a microplate reader (BioTek Synergy H1). Background readings from wells containing only medium and CCK‑8 were subtracted, and proliferation was normalized to the empty vector control group at 24 h.

Malondialdehyde (MDA) assay

Forty-eight hours after transfection, cells were trypsinized, counted, and seeded into 6‑well plates at a density of 1 × 10⁵ cells per well in six replicates per condition. Following a 24‑hour incubation at 37 °C under 5% CO₂ with different doses of erastin, cells were harvested for MDA analysis. MDA levels were quantified using a lipid peroxidation MDA assay kit (Beyotime, catalog no. S0131S) according to the manufacturer’s instructions. Briefly, cells were washed with PBS and lysed in 300 µL of lysis buffer containing 1% Phenylmethanesulfonyl fluoride (PMSF). The lysate was then sonicated (three cycles of 10 s at 20% power, with 30-second intervals) and centrifuged at 12,000 xg for 10 min at 4 °C. A 100 µL aliquot of the resulting supernatant was mixed with 200 µL of thiobarbituric acid (TBA) working solution and incubated at 95 °C for 40 min. After cooling to room temperature, the samples were transferred to a 96‑well plate, and absorbance was measured at 532 nm using a microplate reader (BioTek Synergy H1).

Reactive oxygen species (ROS) assay

Forty-eight hours post-transfection, the cells were trypsinized, counted, and seeded into 96‑well plates at 8 × 10³ cells per well in five replicates per condition. Following a 24-hour incubation at 37 °C under 5% CO₂ with or without erastin, cells were treated with ROS-detecting DCFH-DA probe (Beyotime, catalog no. S0033S), and washed three times with serum-free medium. The fluorescent signals were visualized under an inverted fluorescent microscope (NIKON, ECLIPSE Ti2) and analyzed by ImageJ.

BODIPY 581/591 C11 lipid peroxidation assay

The procedure was identical to the ROS assay, except that cells were treated with BODIPY 581/591 C11 (Beyotime, S0043S) for lipid peroxidation detection. Fluorescence was recorded with the same fluorescent microscope.

Orthotopic implantation and survival analysis

All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of Bayone Laboratory Animal Center (License No. BO-IRB-20251010-XJY-ZXZ-A0027) and conducted in compliance with institutional ethical guidelines. Six-week-old male BALB/c nude mice were housed under standard conditions with a 12-hour light/dark cycle and provided autoclaved food and water ad libitum. For orthotopic implantation, T98G glioblastoma cells overexpressing STOX1 were suspended in a 1:1 mixture of culture medium and Matrix Gel (Beyotime, C0371). A total of 5 × 10⁵ cells in 2 µL volume were stereotactically injected into the left hemisphere of the brain (coordinates: 2 mm lateral, 1 mm anterior to bregma; depth: 4 mm) at a constant rate of 0.1 µL/min. Mice were randomly assigned to two groups (N = 12). Ten days post-implantation, the treatment group received daily intraperitoneal injections of dihydroartemisinin (DHA; 50 mg/kg) for ten days, while the control group received an equivalent volume of vehicle (10% DMSO + 20% PEG300 + 5% Tween-80 + 65% PBS) on the same schedule. Following transplantation, animals were monitored daily for general health and survival until 40 days following implantation. Animals exhibiting > 20% body weight loss or severe neurological deficits, defined as a modified Neurological Severity Score (mNSS) > 10, were euthanized according to humane endpoint criteria.

Statistical analysis

All experimental data were analyzed using R software (v4.3.1) and are presented as the mean standard deviation mean (SEM) from at least three independent experiments. A significance threshold was set at P < 0.05. For the cell proliferation assays, absorbance values at each time point were first tested for normality using the Shapiro‑Wilk test. When data were normally distributed, comparisons were performed using a two‑tailed Student’s t‑test; otherwise, the Wilcoxon rank‑sum test was applied. In addition, two‑way ANOVA was conducted to evaluate the interaction between treatment groups and time points, with Bonferroni correction applied for multiple comparisons. For the MDA assays in U251MG and T98G cells under various treatment regimens, normality was again assessed via the Shapiro‑Wilk test; subsequently, normally distributed data were analyzed using one‑way ANOVA or two‑way ANOVA followed by Tukey’s post‑hoc test, while non-normal data were analyzed using the Kruskal‑Wallis test with Dunnett’s post‑hoc analysis. All P‑values were further adjusted using the Benjamini‑Hochberg method with significance defined at an FDR < 0.05.

Results

Construction of the GBM prognostic model

To develop a prognostic model for GBM, we conducted a comparative transcriptomic analysis between the normal brain tissues from the Genotype-Tissue Expression (GTEx) database, and the GBM tissues from The Cancer Genome Atlas Program (TCGA) database. This analysis identified 1923 genes that were significantly upregulated or downregulated (Fig. 1A).

Fig. 1.

Construction of the gene signature derived from differentially expressed genes (DEGs) between normal brain tissues and GBM samples. (A) Volcano plot illustrating 1923 DEGs identified by comparing normal brain tissues with GBM samples. (B, C) The top 10 enriched pathways in normal brain tissues (B) and GBM samples (C). (D) DEGs with prognostic significance were selected through univariate Cox regression analysis. (E) The final gene signature was established using Least Absolute Shrinkage and Selection Operator (LASSO) regression based on the genes identified in (D)

Gene Set Enrichment Analysis (GSEA) was performed to explore the pathways enriched among these differentially expressed genes (DEGs). The top 10 upregulated and downregulated pathways in normal brain tissue (Fig. 1B) and GBM tissues (Fig. 1C) are shown. In normal brain tissue, the most enriched genes were predominantly downregulated (Fig. 1B), whereas in GBM tissues, pathways related to cancer, cell adhesion, and cell cycle were significantly upregulated (Fig. 1C), including the cell cycle (NES = 1.91), p53 signaling (NES = 1.89), TGF-beta signaling (NES = 1.75), DNA replication (NES = 1.75), and cell adhesion molecules (NES = 1.52) (see Supplementary Table S1).

To identify genes associated with GBM prognosis, we conducted univariate Cox regression analysis on the DEGs (Fig. 1D). Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) regression and multivariate Cox regression analysis were used to construct a prognostic model. The optimal penalty parameter (λ) was determined through 10-fold cross-validation (Fig. 1E). Genes with nonzero coefficients at this optimal λ were selected to form a 14-gene prognostic signature, comprising UPP1, ADIRF-AS1, IL27RA, CYP2E1, VPS45, MYLK-AS1, ZNF248, STOX1, FBXO28, TIMP4, STXBP5-AS1, SORBS2, GPR6, and GANC (Fig. 1D and E). Among these, six genes (UPP1, ADIRF-AS1, IL27RA, STXBP5-AS1, SORBS2, GPR6) were upregulated, and eight genes (CYP2E1, VPS45, MYLK-AS1, ZNF248, STOX1, FBXO28, TIMP4, GANC) were downregulated in GBM tissues (Fig. 1D).

Validation of the GBM prognostic model

To validate the established prognostic model, we calculated risk scores for the GBM patients in the TCGA cohort. Based on the median risk score, patients were stratified into high-risk and low-risk groups. Kaplan-Meier analysis with log-rank tests revealed significantly poorer outcomes in the high-risk group compared to the low-risk group (Fig. 2A, Supplementary Figure S1A, S1C, S1E).

Fig. 2.

Validation of the prognostic gene signature. (A, B) Kaplan-Meier survival curves for GBM patients stratified by the gene signature in the TCGA cohort (A) and the GSE108474 cohort (B), with statistical analysis performed using the log-rank test. (C, D) Receiver operating characteristic (ROC) curves for 1-year, 2-year, and 3-year survival predictions based on the gene signature in the TCGA (C) and GSE108474 (D) GBM cohorts. (E, F) Univariate (E) and multivariate (F) Cox regression analyses incorporating the gene signature alongside clinical variables such as gender and age in TCGA GBM cohort

To further assess the robustness of this model, we applied it to an independent GBM cohort from the GSE108474 dataset [14, 15], calculating risk scores and stratifying patients using the same approach. The survival trends in this cohort were consistent with those observed in the TCGA cohort (Fig. 2B, Supplementary Figure S1B, S1D, S1F), supporting the reliability of this model in predicting GBM prognosis.

Receiver Operating Characteristic (ROC) curve analysis demonstrated moderate predictive accuracy of the model for 1-year, 2-year, and 3-year survival in both TCGA and GSE108474 datasets. In the TCGA cohort, the AUC values were 0.744, 0.793, and 0.784 at 1, 2, and 3 years, respectively. In the GSE108474 cohort, the corresponding AUC values were lower, at 0.506, 0.577, and 0.567 (Fig. 2C and D). Furthermore, univariate and multivariate Cox regression analyses incorporating the 14-gene signature alongside clinical variables (e.g., age and sex) identified the gene-based model as an independent prognostic factor for GBM outcomes (Fig. 2E and F). Collectively, these results demonstrate that our 14-gene prognostic model is independently predictive and validated across different GBM cohorts.

Risk score and its association with gene mutation

To explore the biological basis of the prognostic model, we first examined the gene mutation landscape of the TCGA GBM cohort. Missense mutation were the most prevalent variant classification, and single nucleotide polymorphism (SNP) represented the most common variant type (Supplementary Figure S2A). The most frequently mutated genes included PTEN, TP53, EGFR, TTN, MUC16, and NF1 (Supplementary Figure S2A, S2B), consistent with previous findings [16, 17].

Patients in the TCGA GBM cohort were then stratified into high-risk and low-risk groups based on the median risk score, as previously described. Mutation analysis revealed that the overall distribution of mutation types was not significantly different between the two groups (Fig. 3A and B). However, differences in the mutation frequencies of specific genes were observed. While the top five most frequently mutated genes remained PTEN, TP53, EGFR, TTN, and MUC16, their mutation rates varied between the two groups: PTEN (39% in high-risk group vs. 29% in low-risk group), TTN (31% in high-risk group vs. 27% in low-risk group), TP53 (28% in high-risk group vs. 43% in low-risk group), EGFR (22% in high-risk group vs. 33% in low-risk group), MUC16 (14% in high-risk group vs. 20% in low-risk group) (Fig. 3C and D).

Fig. 3.

Comparison of gene mutations between high-risk and low-risk GBM patients from the TCGA database. (A, B) The types of gene mutations in the high-risk (A) and low-risk (B) patient groups. (C, D) The distribution of the top 30 mutated genes in the high-risk (C) and low-risk (D) patient groups. (E, F) The co-mutation analyses of the top 20 mutated genes in the high-risk (E) and low-risk (F) patient groups

Co-mutation analysis further revealed distinct patterns between the high- and low-risk groups. For example, in the high-risk group, TTN mutations commonly co-occurred with GLYR1 and FLG mutations, while PTEN and TP53 mutations did not show significant co-mutation associations. Conversely, in the low-risk group, TTN mutations frequently co-occurred with mutations in UGGT1, OBSCN, FLG, and RB1, and both PTEN and TP53 mutations were commonly associated with RB1 (Fig. 3E and F), indicating a more prevalent co-mutations in low-risk patients.

These findings suggest that the high-risk and low-risk groups, as defined by our prognostic model, exhibit distinct gene mutation profiles, highlighting potential biological differences underlying GBM prognosis.

Risk score and its association with immune infiltration

To assess immune cell infiltration in the high- and low-risk groups, we applied single-sample Gene Set Enrichment Analysis (ssGSEA). The high-risk group demonstrated significantly elevated immune infiltration scores, characterized by increased levels of various immune cells, including multiple T cell subsets, dendritic cells, neutrophils, macrophages, as well as immune-related molecules (Fig. 4A). In line with these findings, tumor purity scores indicated a more complex immune microenvironment in high-risk patients (Fig. 4B). Additionally, the expression of most human leukocyte antigen (HLA) genes and PD-L1 were upregulated in the high-risk group (Fig. 4C and D).

Fig. 4.

The immune landscape of high-risk and low-risk GBM patients from the TCGA database. (A) Comparison of the immune pathway activities between the high-risk (Risk_H) and low-risk (Risk_L) groups. (B) Comparison of the tumor purity between the high-risk (Risk_H) and low-risk (Risk_L) groups. (C) Comparison of the human leukocyte antigen (HLA) expression between the high-risk (Risk_H) and low-risk (Risk_L) groups. (D) Comparison of the PD-L1 expression between the high-risk (Risk_H) and low-risk (Risk_L) groups. (E) Comparison of the immune cell infiltration profiles between the high-risk (Risk_H) and low-risk (Risk_L) groups. ns, no significance; *, p < 0.05; **, p < 0.01; ***, p < 0.001

To further validate these results, we used the CIBERSORT algorithm to estimate the relative abundance of 28 immune cell types. The analysis confirmed a generally enhanced immune cell infiltration profile in high-risk patients (Fig. 4E).

STOX1 and ZNF248 as risk genes in GBM

Based on the above bioinformatics analyses, we conducted a detailed investigation of the 14 genes included in the prognostic model. Among them, CYP2E1, ZNF248, and STOX1 were significantly downregulated in GBM tissues compared to normal brain tissues (Fig. 5A). To validate these findings, we examined gene expression in GBM cell lines (U251 MG, LN229, and T98G), using the human astrocyte cell line SVG P12 and human immature oligodendrocyte cell line MO3.13 as controls. The expression patterns were consistent with TCGA data, confirming the significant downregulation of these genes in GBM cells, compared to normal brain cells (Fig. 5B–D). While CYP2E1 has been previously reported in the literature [18, 19], STOX1 and ZNF248 remain largely unexplored. Therefore, we focused on STOX1 and ZNF248 for subsequent analyses.

Fig. 5.

Reduced expression ofCYP2E1, ZNF248, and STOX1 in GBM tissue and cell lines. (A) Comparison of CYP2E1, ZNF248, and STOX1 overall expression levels between control normal brain tissue from the GTEx cohort and GBM samples from the TCGA cohort. (B-D) Comparison of CYP2E1 (B), STOX1 (C), and ZNF248 (D) expression levels in a human astrocyte cell line (SVG P12) and a human immature oligodendrocyte cell line (MO3.13), compared with GBM cell lines (U251MG, LN229, T98G, and U87MG). ns, not significant; *, p < 0.05; **, p < 0.01; ***, p < 0.001; ****, p < 0.0001

Using spatial transcriptomic data from the GBM dataset available at Dryad (https://datadryad.org/stash/dataset/) [20], we found that STOX1 was scarcely expressed in tumor cells, whereas ZNF248 was expressed in tumor cells, infiltrating immune cells, and normal brain cells (Fig. 6A). Notably, both genes showed high expression in normal central nervous system cells, including neurons and glial cells, across multiple patient tissue sections, but were markedly downregulated in GBM tumor cells and infiltrating immune cells (Fig. 6B). Interestingly, STOX1, but not ZNF248, was also highly expressed in endothelial cells (Fig. 6B).

Fig. 6.

Expression profiles of STOX1 and ZNF248 across various cell types in a spatial transcriptomic dataset of GBM patients. (A) Spatial distribution and expression of STOX1 and ZNF248 across different cell types in a GBM tissue section. (B) Comparison of STOX1 expression (from six GBM samples) and ZNF248 expression (from five GBM samples) across various cell types. CNS cells, central nervous system cells; DCs, dendritic cells; NK cells, natural killer cells

To further identify the specific brain cell types expressing STOX1 and ZNF248, we analyzed a publicly available single-cell sequencing dataset of GBM (GSE84465). STOX1 was primarily localized in astrocytes with high expression, and also present in astrocyte-like (AC-like) malignant cells at lower levels (Fig. 7A–C). In contrast, ZNF248 was mainly expressed in oligodendrocyte progenitor cells (OPCs) with high expression, followed by neurons with moderate expression (Fig. 7A–C). ZNF248 was also broadly detected in monocytes/macrophages and AC-like malignant cells, though with lower expression (Fig. 7A–C). These findings are consistent with our in vitro results, which showed that STOX1 is highly expressed in astrocytes, and ZNF248 in both astrocytes and immature oligodendrocytes, whereas both genes exhibit relatively low expression in GBM cells (Fig. 5B–D).

Fig. 7.

Expression patterns of STOX1 and ZNF248 across various cell types in an independent single-cell transcriptomic dataset (GSE84465) of GBM patients. (A) Expression patterns of STOX1 and ZNF248 across multiple cell types, including vascular cells, oligodendrocyte progenitor cells (OPCs), oligodendrocytes, neurons, monocytes, M1-like macrophages (M1), malignant cells, astrocytes, and astrocyte-like (AC-like) malignant cells. (B) Proportions of each cell type among STOX1-negative and STOX1-positive cells (left), and ZNF248-negative and ZNF248-positive cells (right). (C) The relative expression of STOX1 and ZNF248 in each cell type

Pathway enrichment analysis further revealed that GBM tumor cells with high STOX1 or ZNF248 expression exhibited activation of programmed cell death pathways, including apoptosis, necroptosis, cuproptosis, pyroptosis, and ferroptosis (Supplementary Figures S3A, S3B). Specifically, ferroptosis was the most significantly upregulated form of cell death in astrocytes expressing high levels of STOX1 (Supplementary Figures S3A), and similarly, in OPCs with high ZNF248 expression (Supplementary Figures S3B). In contrast, ferroptosis was not prominently activated in malignant tumor cells or AC-like malignant cells (Supplementary Figures S3A, S3B).

Based on these findings, we proceeded to experimentally investigate the roles of STOX1 and ZNF248 in GBM tumor cells.

Furthermore, given the distinct mutation frequencies and co-mutation patterns observed between high- and low-risk groups (Fig. 3), as well as differences in immune infiltration landscapes (Fig. 4), we further investigated the potential links between STOX1 and ZNF248 expression and both mutational characteristics and immune-related features. Among the mutations in PTEN, EGFR, and TP53, only TP53 mutation was significantly associated with patient survival (Supplementary Figure S4A–C). We therefore focused on its regulatory functions on STOX1 and ZNF248 expression. Using KnockTF 2.0 tool, we found that TP53 knockdown downregulated STOX1 expression in the glioma cell line HSJD-DIPG-007, but upregulated ZNF248 in U87 MG cells (Supplementary Figure S5A, S5B). In terms of immune context, analysis using the Tumor Immune Dysfunction and Exclusion (TIDE) platform suggested that high STOX1 expression correlates with enhanced responsiveness to immune checkpoint blockade therapy, whereas ZNF248 showed no significant association with immune infiltration signatures (Supplementary Figure S6A, S6B).

STOX1 and ZNF248 contribute to ferroptosis sensitivity in GBM cells

Based on the above findings, we overexpressed STOX1 or ZNF248 in the human GBM cell lines U251MG and T98G to further investigate their functional roles, and confirmed successful overexpression compared to the empty vector control (Supplementary Figures S7A, S7B). Cell proliferation assay revealed that overexpression of either gene can significantly inhibited the growth of U251MG and T98G under standard culture conditions (Fig. 8A and B).

Fig. 8.

Functional analysis of STOX1 and ZNF248 in GBM cells. (A, B) Comparative analysis of cell proliferation in U251 MG or T98G GBM cells. Cells were transiently transfected with either an empty vector (EV) as a control, or plasmids overexpressing STOX1 (A) or ZNF248 (B). Cell proliferation was assessed at 24, 48, 72, 96, and 120 h post-transfection using the Cell Counting Kit-8 (CCK-8) assay, which measures cell viability in living cells. (C) Evaluation of ferroptosis sensitivity in T98G cells overexpressing STOX1. Ferroptosis was induced using erastin at the indicated concentrations. Ferrostatin-1 (Fer-1), a ferroptosis inhibitor, was used as at 2 µM a control. Cell viability was measured by CCK-8 assay. Malondialdehyde (MDA) levels were measured as a marker of lipid peroxidation with 10 µM erastin. (D) Evaluation of ferroptosis sensitivity in U251 MG cells overexpressing ZNF248. Ferroptosis was induced using erastin at the indicated concentrations. Fer-1 was used as at 2 µM a control. Cell viability was measured by CCK-8 assay. MDA levels were measured with 2.5 µM erastin. (E) Measurement of intracellular reactive oxygen species (ROS) levels in T98G cells overexpressing STOX1 or ZNF248, with or without treatment with 10 µM erastin. ROS were detected using the DCFH-DA fluorescent probe. (F) Measurement of lipid peroxidation levels in T98G cells overexpressing STOX1 or ZNF248, with or without treatment with 10 µM erastin. The ratio of oxidized (green) to reduced (red) BODIPY-C11 fluorescence was quantified. (G) Measurement of mRNA expression of ferroptosis-related genes (GPX4, SLC7A11, and FTH1) in T98G cells overexpressing STOX1 or ZNF248, with or without treatment with 10 µM erastin. (H) Schematic illustration of the survival experiment following orthotopic implantation. Nude mice were orthotopically implanted with T98G cells and divided into two groups: STOX1 overexpression (OE) with solvent control, and STOX1 overexpression (OE) with dihydroartemisinin (DHA) treatment. DHA or solvent control was administered intraperitoneally (IP) once daily for ten consecutive days starting on day 10 after tumor implantation. (I) Kaplan–Meier survival curves with indicated p value (N = 12). ns, no significance; *, p < 0.05; **, p < 0.01; **, p < 0.001; ****, p < 0.0001

To evaluate their impact on ferroptosis sensitivity, T98G cells overexpressing STOX1, U251MG cells overexpressing ZNF248, and corresponding empty vector controls were treated with increasing concentrations of the ferroptosis inducer erastin for 24 h. Erastin alone significantly reduced cell viability in both cell lines (Fig. 8C and D). Notably, STOX1-overexpressing T98G cells and ZNF248-overexpressing U251MG cells exhibited an even greater reduction in viability compared to their respective controls following erastin exposure, indicating enhanced ferroptosis sensitivity (Fig. 8C and D). As a control, co-treatment with the ferroptosis inhibitor ferrostatin-1 (Fer-1) effectively rescued cell viability, confirming that the observed effects were ferroptosis-specific (Fig. 8C and D).

To further validate these findings, we assessed malondialdehyde (MDA) levels as a marker of lipid peroxidation. MDA assay revealed that STOX1-overexpressing T98G cells and ZNF248-overexpressing U251MG cells had significantly elevated MDA levels following erastin treatment compared to corresponding empty vector controls (Fig. 8C and D). Consistently, Fer-1 treatment reduced MDA accumulation in control cells, supporting that the effects were ferroptosis-specific (Fig. 8C and D). To rule out the possibility that STOX1 and ZNF248 overexpression affects the general metabolic stress of GBM cells rather than specifically modulating ferroptosis sensitivity, we examined changes in reactive oxygen species (ROS), lipid peroxidation levels, and ferroptosis-related gene expression under conditions with or without ferroptosis induction. In the absence of a ferroptosis inducer, STOX1 overexpression did not affect basal ROS levels, whereas ZNF248 overexpression significantly increased intracellular ROS. Under erastin-induced ferroptosis, ROS levels were elevated in the empty-vector group and further enhanced by either STOX1 or ZNF248 overexpression, both exceeding those of the empty-vector control (Fig. 8E). With respect to lipid peroxidation, neither STOX1 nor ZNF248 overexpression altered the oxidized/reduced BODIPY C11 ratio under basal conditions. Erastin treatment increased lipid peroxidation across all groups, with STOX1- or ZNF248-overexpressing cells showing higher levels than control cells (Fig. 8F). Interestingly, under basal conditions, overexpression of STOX1 and ZNF248 suppressed the ferroptosis-associated genes GPX4 and SLC7A11, whereas only STOX1 reduced FTH1 expression. Following erastin treatment, both STOX1 and ZNF248 overexpression further downregulated GPX4, SLC7A11, and FTH1 (Fig. 8G).

In addition, we assessed glioma patient survival using the Chinese Glioma Genome Atlas (CGGA) database. High expression of either STOX1 or ZNF248 was consistently associated with improved survival in primary gliomas across three independent cohorts (Supplementary Figure S8A, S8B). In Grade IV GBM, higher expression of STOX1 correlated with better survival in two of the three cohorts, and ZNF248 in one of the three cohorts, likely reflecting limitations due to smaller sample sizes (Supplementary Figure S8C, S8D). Taken together, these results indicate that both STOX1 and ZNF248 enhance the ferroptosis sensitivity of GBM cells, which may in part contribute to improved survival in low-risk group of GBM patients.

To validate our in vitro observations, we assessed survival in an intracranial glioblastoma model. To enhance translational potential, we sought a ferroptosis inducer suitable for clinical application. Erastin has unknown blood–brain barrier (BBB) permeability. And RSL3, another commonly used experimental inducer, exhibits poor solubility and limited BBB penetration [21, 22]. In contrast, dihydroartemisinin (DHA), a widely used antimalarial drug with an established clinical safety profile, has been identified as a ferroptosis inducer in multiple cancers [23–25]. Importantly, DHA readily crosses the BBB and does not strongly impair normal astrocytic function [26, 27]. Based on these advantages, we selected DHA as the ferroptosis inducer for our animal survival experiments.

Since STOX1 showed significant correlation on patient survival in two of the three GBM cohorts, while ZNF248 only in one of the three cohorts (Supplementary Figure S8C and S8D), we focused on STOX1 and tested whether it sensitize ferroptosis in GBM in vivo. Nude mice were orthotopically implanted with T98G cells and assigned to two groups: STOX1 overexpression + control vehicle treatment, and STOX1 overexpression + DHA treatment. The results demonstrated that the combination of DHA treatment with STOX1 overexpression significantly prolonged animal survival compared to STOX1 overexpression group with control treatment (Fig. 8H and I). These in vivo findings provide compelling evidence that STOX1 exerts its tumor-suppressive effect, at least partially, through sensitizing GBM cells to ferroptosis.

Discussion

The prognosis for GBM remains poor. In this study, by comparing transcriptomic data from GBM samples and normal brain tissues, we identified 1923 DEGs, including well-characterized GBM-associated genes such as PTEN, TP53, EGFR, and RB1. For example, PTEN was significantly downregulated in GBM, and its loss is known to result in sustained activation of the PI3K/Akt pathway, thereby promoting tumor progression [28, 29]. TP53, a well-established tumor suppressor in GBM, plays critical roles in DNA repair and apoptosis induction [30, 31]. These results underscore the relevance of the identified genes and highlight their potential for further investigation. Subsequent Gene Set Enrichment Analysis (GSEA) revealed a strong enrichment of pathways involved in cell cycle regulation, cancer-related signaling, and TGF-β signaling in GBM. In contrast, pathways related to purine and pyrimidine metabolism were predominantly enriched in normal brain tissue, offering important insights into the distinct molecular landscapes of GBM versus healthy brain. Using the 1923 DEGs, we refined a 14-gene prognostic model using univariate Cox regression, followed by LASSO and multivariate Cox regression. While multi-gene prognostic models have been widely adopted in cancer research [32–34], such models remain limited in GBM. To assess its predictive power, we validated our model in an independent cohort (GSE108474, N = 130), where it successfully stratified patients by risk. ROC curve analysis showed moderate predictive accuracy for 1-, 2-, and 3-year survival, further supporting the clinical utility of this 14-gene signature.

Patients were stratified into high- and low-risk groups based on median risk scores. Although the overall mutation types did not differ significantly between the groups, PTEN mutations were more prevalent in high-risk patients, while low-risk patients exhibited fewer PTEN alterations. Co-mutation analysis revealed that high-risk patients had fewer concurrent mutations, whereas low-risk patients displayed more complex mutational landscapes, potentially reflecting compensatory effects from other genetic alterations. Immune infiltration analysis further showed that the tumor microenvironment in high-risk GBM patients was characterized by increased infiltration of immunosuppressive cells, including regulatory T cells (Tregs). Despite the elevated presence of immune cells, the enrichment of Tregs likely promotes the secretion of immunosuppressive cytokines such as IL-10, TGF-β, and IL-35, thereby dampening the tumor-killing activity of cytotoxic T cells and natural killer (NK) cells [35, 36]. Intriguingly, tumor cells in high-risk patients exhibited upregulated expression of HLA molecule alongside elevated levels of PD-L1. Notably, PD-L1, a well-established immune checkpoint molecule, facilitates GBM immune evasion by suppressing T cell-mediated cytotoxicity [37–39]. Collectively, our prognostic model not only stratifies GBM patients into distinct risk subgroups but also reflects established GBM immunobiology, providing strong theoretical validation. However, further experimental and clinical studies are needed to fully elucidate these immune regulatory networks.

One emerging area of interest in cancer biology is ferroptosis—a regulated, iron-dependent form of non-apoptotic cell death characterized by the accumulation of lipid peroxides. Unlike apoptosis or necrosis, ferroptosis is driven by oxidative stress and impaired antioxidant defenses [40]. Mounting evidence suggests that inducing ferroptosis can selectively target cancer cells and suppress tumor growth across a variety of malignancies, including melanoma, ovarian cancer, epithelial ovarian cancer, gastric cancer, and hepatocellular carcinoma [41–49]. A variety of agents have been shown to induce ferroptosis in cancer cells and inhibit tumor growth, including small-molecule inducers (e.g., erastin and RSL3), clinically approved drugs (e.g., sorafenib, sulfasalazine, statins, and artemisinin), radiotherapy, and cytokines such as IFNγ and TGFβ1 [50]. In addition, Immune cells, particularly cytotoxic T cells, can promote ferroptosis, contributing to immune-mediated tumor suppression [51]. In fact, ferroptosis plays a key role in T cell-mediated tumor killing [51, 52], adding another layer of complexity and potential therapeutic opportunity. One promising therapeutic agent, cyst(e)inase, an engineered enzyme that degrades cystine and cysteine, has been shown to effectively induce ferroptosis in prostate cancer and chronic lymphocytic leukemia models. Notably, in a Kras/TP53-driven pancreatic ductal adenocarcinoma mouse model, cyst(e)inase significantly suppressed tumor growth without apparent toxicity [53]. This approach also shows potential in treating EGFR-mutant non-small cell lung cancer and breast cancer [54, 55]. Taken together, these findings suggest that ferroptosis represents a promising diagnostic and therapeutic target across multiple cancer types. The combination of temozolomide and erastin has been shown to activate ferroptosis in GBM cell lines, exhibiting strong anti-tumor effects [56]. Furthermore, DHA selectively induces ferroptosis in GBM cell lines, while exhibiting minimal cytotoxicity in normal human astrocytes [26]. These findings highlight the therapeutic potential of ferroptosis in GBM treatment. However, further investigation is necessary to fully elucidate its clinical applicability.

Among the 14 genes identified in our prognostic model, CYP2E1, STOX1, and ZNF248, were significantly downregulated in tumors, with higher expression levels correlating with better prognosis, suggesting potential tumor-suppressive roles. Validation using normal brain cells and GBM cell lines confirmed the downregulation of these genes in tumors, with observations from the TCGA dataset. A literature review revealed that CYP2E1 is a well-characterized enzyme involved in ferroptosis and other forms of regulated cell death [18, 57, 58], providing external validation for our screening approach. Expression profiling of CYP2E1, STOX1, and ZNF248 between normal brain tissue and GBM tissue confirmed their significant downregulation in tumors. Given the extensive prior research on CYP2E1, we focused subsequent investigations on the less known tumor suppressors STOX1 and ZNF248. STOX1 has been previously implicated in preeclampsia, pulmonary artery smooth muscle cell proliferation, cerebellar granule neurogenesis, and medulloblastoma [59–62], yet its role in cancer remains largely unexplored. Likewise, ZNF248 has been studied in the context of late-onset asthma, and mentioned in several genomic studies, but has not been previously investigated in cancer [63–66], making this the first study to implicate it in GBM biology. In this study, pathway enrichment analyses linked high expression of STOX1 and ZNF248 to the activation of programmed cell death pathways, particularly ferroptosis. Functional assays demonstrated that overexpression of STOX1 or ZNF248 in GBM cells significantly inhibited proliferation and enhanced sensitivity to ferroptosis. Together, our findings identify STOX1 and ZNF248 as novel tumor suppressors in GBM with potential therapeutic relevance, possibly through enhancing tumor cell sensitivity to ferroptosis.

Conclusions

​This study presents the development and validation of a 14-gene prognostic model for GBM, highlighting STOX1 and ZNF248 as potential tumor suppressors that inhibit GBM progression and enhance sensitivity to ferroptosis. These findings open new avenues for GBM therapy, although further mechanistic studies are necessary to fully elucidate their functional roles.

Abbreviations

AC-like

Astrocyte-like

AUC

Area Under the Curve

BODIPY

Boron dipyrromethene

CCK-8

Cell Counting Kit-8

CI

Confidence interval

CNS

Central nervous system

DCFH-DA

Dichlorodihydrofluorescein diacetate

DEG

Differentially expressed gene

DHA

Dihydroartemisinin

DMSO

Dimethyl sulfoxide

ES

Enrichment score

FBS

Fetal bovine serum

FC

Fold change

FDR

False Discovery Rate

Fer-1

Ferrostatin-1

GBM

Glioblastoma

GSEA

Gene Set Enrichment Analysis

GTEx

Genotype-Tissue Expression

HLA

Human leukocyte antigen

HR

Hazard ratio

LASSO

Least Absolute Shrinkage and Selection Operator

mNSS

modified Neurological Severity Score

NES

Normalized Enrichment Score

NK cell

Natural killer cell

OPC

Oligodendrocyte progenitor cell

OS

Overall survival

PBS

Phosphate-buffered saline

PCA

Principal Component Analysis

PEG300

Polyethylene glycol 300

PFS

Progression-free survival

PMSF

Phenylmethanesulfonyl fluoride

ROC

Receiver Operating Characteristic

ROS

Reactive oxygen species

SEM

Standard deviation mean

SNP

Single nucleotide polymorphism

SNR

Signal-to-Noise Ratio

ssGSEA

Single-sample Gene Set Enrichment Analysis

TCGA

The Cancer Genome Atlas Program

TPM

Transcripts per million

Treg

Regulatory T cell

UMAP

Uniform Manifold Approximation and Projection

UMI

Unique Molecular Identifier

VEGF-A

Vascular endothelial growth factor A

VST

Variance stabilizing transformation

Author contributions

SW, XZ, and WP conceived and designed this study. SW conducted bioinformatics analysis. SW, SX, KL, SL, and YY conducted experiments. SW and XZ wrote the original manuscript and generated the figures. WP and XZ reviewed the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Shenzhen Science and Technology Program (No.JCYJ20220818100803007, KCXFZ20211020163549011), Shenzhen Key Laboratory of Neuroimmunomodulation for Neurological Diseases (No. ZDSYS20220304163558001).

Data availability

All the data in this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Ethics approval is not applicable for this study. Written informed consent for participation is not required for this study.

Consent for publication

All authors consent to the publication of this work.

Competing interests

The authors declare no competing interests.