Construction of an actin cytoskeleton-related gene signature for predicting prognosis and therapeutic response in glioblastoma: based on machine learning

Introduction

Based on the latest World Health Organization (WHO) classification of nervous system tumors [2021], glioblastoma (GBM) is delineated as an isocitrate dehydrogenase (IDH) wild-type glioma (1). As the most malignant type of glioma, the efficacy in treating GBM remains extremely poor, with a median survival time of solely about 14–18 months and a 5-year survival rate of <7% (2,3). GBM is also the most prevalent malignant tumor within the central nervous system, accounting for 14.2% of all tumors and 50.9% of all malignant brain tumors registered in the Central Brain Tumor Registry of the United States (4).

The current standard therapy regimen for GBM comprises maximal safe surgical resection and radiotherapy combined with temozolomide chemotherapy (5). Although these therapies may temporarily delay the progression of diseases, most patients eventually experience recurrence. Recently, emerging approaches encompassing tumor-treating fields, molecular targeted therapies, and immunotherapies have been explored in GBM. However, their overall clinical benefits remain limited (6,7).

Therefore, it’s necessary to elucidate the mechanisms underlying the invasion, immune regulation, and metastasis of GBM and to ascertain novel therapeutic targets. This has become a critical scientific issue and clinical challenge for ameliorating the prognosis of patients.

Recent evidence has increasingly demonstrated that aberrant remodeling of the cytoskeletal system is pivotal in the initiation, invasion, metastasis, and other malignant behaviors of tumors (8-10). As a core component of the cytoskeleton, actin regulates the morphology, motility, and adhesion capacity of cells through dynamic polymerization and depolymerization, thereby influencing the migration and invasion of tumor cells (11).

Although existing research has revealed the potential functions of the actin cytoskeleton in the progression of GBM (12,13), systematic identification of actin-related signature genes and evaluation of their prognostic value in GBM remain underexplored. Therefore, this research sought to integrate multi-omics data to identify key genes linked to the regulation of actin cytoskeleton, construct a gene expression feature model with prognostic value, and further elucidate its potential mechanisms in the regulation of malignant phenotypes in GBM. These findings provide new theoretical insights and biomarkers for molecular subtyping, efficacy evaluation, and prediction of prognosis in GBM. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2865/rc).

Methods

Data acquisition and preprocessing

Gene expression data of IDH wild-type gliomas were retrieved from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo), TCGA (https://portal.gdc.cancer.gov/), and Chinese Glioma Genome Atlas (CGGA) (https://www.cgga.org.cn/) databases. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study workflow is illustrated in Figure 1. The GSE145645 dataset included 6 samples (3 tumor tissues and 3 normal controls), while GSE184643 comprised 35 samples (32 tumor tissues and 3 normal controls). CGGA and TCGA provided 179 and 169 tumor samples, respectively. Detailed dataset information is summarized in Table 1.

Figure 1 Analysis workflow of this research. CGGA, Chinese Glioma Genome Atlas; CIBERSORT, cell-type identification by estimating relative subsets of RNA transcripts; GDSC, Genomics of Drug Sensitivity in Cancer; GEO, Gene Expression Omnibus; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; KM, Kaplan-Meier; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas; TIDE, tumor immune dysfunction and exclusion; TMB, tumor mutational burden; WGCNA, Weighted Gene Co-expression Network Analysis.

During preprocessing, raw expression-profile files were downloaded. Probes with weak signals and those that failed to show significant differential expression were removed. The expression values of every gene were then normalized to eliminate systematic differences between samples. Probe annotation was performed utilizing files from official platforms. Probe IDs were converted to corresponding gene symbols. Probes targeting numerous genes were excluded. Finally, an integrated gene expression matrix was identified. GSE145645 and GSE184643 datasets were integrated, and batch effects were corrected and normalized utilizing the ComBat function from the sva package in R (v3.50.0).

Screening of differentially expressed genes (DEGs) and functional enrichment analysis

Differential expression analysis was implemented with the limma package in R (v3.52.4). Genes with adjusted P value (P.adjust) <0.05 and |log₂ fold change (FC)| ≥1.5 were delineated as DEGs. The distribution of DEGs was visualized by volcano and heat maps. Functional enrichment analyses of DEGs for Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways were implemented utilizing clusterProfiler (v4.14.6), with significance set at P<0.05. To further validate the genome-wide significance of enriched pathways, gene set enrichment analysis (GSEA) was implemented utilizing KEGG gene sets from the MSigDB database (1,000 permutations). Significance thresholds were defined as |normalized enrichment score (NES)| >1, P value <0.05, and q-value <0.25.

Screening of key genes and establishment of prognostic gene signature

Based on the 232 genes in the ‘Regulation of actin cytoskeleton’ pathway from the KEGG database with transcriptomic and clinical follow-up data from 175 patients with GBM from the CGGA database, univariate Cox regression analysis was implemented utilizing the survival package in R (v3.8.3). With P<0.05 as the screening threshold, candidate genes significantly linked to overall survival (OS) were identified. To optimize feature selection, ML methods were comprehensively employed. The performance of the model was evaluated through cross-validation and the C-index. Utilizing expression levels of selected core genes, a riskScore was constructed through the coxph() function in the survival package. Patients were stratified into high-risk groups (HRGs) and low-risk groups (LRGs) using the median riskScore as the cutoff for subsequent survival analysis and model evaluation. To minimize platform-related bias, expression data from each cohort were processed independently using their respective normalized matrices, and the calculation of riskScore was performed within each dataset without cross-cohort normalization. The riskScore formula is defined as: riskScore = h0(t) exp (β1 X1 + β2 X2 + … + βn Xn).

Survival analysis and external validation of gene features

Based on the constructed gene signature, Kaplan-Meier (K-M) survival curves were generated leveraging the survival package (v3.8.3) in R. The log-rank test was utilized to contrast the differences of OS between the HRGs and LRGs. Furthermore, time-dependent receiver operating characteristic (ROC) curves for 1, 2, 3, 4, and 5 years were plotted utilizing the timeROC package (v0.4) to estimate the predictive performance of the prognostic signature.

To validate the robustness of the signature, independent validation was implemented in the TCGA cohort (n=169) for OS curves and time-dependent ROC analysis.

Moreover, univariate and multivariate Cox regression analyses incorporating clinical variables were implemented to calculate hazard ratios (HRs) and 95% confidence intervals (CIs), thereby assessing whether the riskScore served as an independent prognostic factor.

Comparative analyses between IDH wild-type and IDH-mutant gliomas were performed in the CGGA cohort. Gene expression differences were assessed using the Wilcoxon rank-sum test. K-M survival analyses stratified by IDH status were conducted to evaluate subgroup-specific prognostic performance.

Weighted Gene Co-expression Network Analysis (WGCNA) for ascertaining risk-related modules and enrichment analysis

To ascertain the characteristics of gene modules linked to risk groups, WGCNA (v1.73) was implemented on the transcriptomic data of 179 GBM samples from the CGGA dataset. First, the expression matrix was normalized, and the top 50% of highly variable genes were selected based on expression variance. Subsequently, a weighted gene co-expression network was developed with a soft threshold of β=12. The dynamic tree-cutting method was leveraged to partition the modules, with a minimum module gene size of 200. The association between module eigengenes and clinical traits (HRGs and LRGs) was ascertained, and heatmaps of relationships between module and trait were generated. Modules significantly linked to the HRG were selected utilizing a threshold of P<0.05, and genes within these modules were extracted for GO and KEGG enrichment analysis. The results were visualized using bubble plots.

Immune infiltration analysis

To assess the relationship between risk groups and the immune microenvironment, the LM22 signature gene set and the cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT) method (https://cibersort.stanford.edu, permutations =1,000) were leveraged to perform deconvolution analysis on the transcriptomic data of patients from CGGA. The infiltration proportions of 22 immune cell types were estimated. Only samples with CIBERSORT results meeting P<0.05 were retained for subsequent analysis. Spearman correlation analysis was further employed to ascertain associations between the 7 key genes and the levels of immune cell infiltration (ICI). The results were visualized utilizing a lollipop plot. Stratified immune analyses based on IDH status were additionally performed in the CGGA cohort. To forecast the potential efficacy of immunotherapy, the tumor immune dysfunction and exclusion (TIDE) tool (http://tide.dfci.harvard.edu) was employed to calculate the TIDE score, T cell exclusion, interferon gamma (IFNG), and related immune suppressive factors [cancer-associated fibroblast (CAF), myeloid-derived suppressor cell (MDSC), and others] for each patient. Differences between the HRGs and LRGs were compared.

Tumor mutational burden (TMB)

Using MAF files from TCGA, mutation data were processed and standardized. The TMB for every patient was ascertained utilizing the maftools package in R (v2.22.0). Information on gene mutation was also collected. The TMB levels and mutation differences of key genes between the HRGs and LRGs were compared. Additionally, maftools was used to generate an oncoplot to visualize the mutation landscape. The most frequently mutated genes and their distribution patterns in the TCGA cohort were also exhibited.

Drug sensitivity prediction analysis

To ascertain potential therapeutic drugs, data on drug response [half maximal inhibitory concentration (IC₅₀)] and gene expression profiles were attained from the GDSC2 database. The oncoPredict package in R (v1.2) was leveraged to develop a drug sensitivity prediction model. The CGGA cohort served as input to predict the sensitivity (IC₅₀) of each patient to the drugs. Patients were then stratified into HRGs and LRGs based on the median riskScore. Differences in predicted IC₅₀ values between HRGs and LRGs were compared utilizing the Wilcoxon rank-sum test (P<0.05). We ultimately identified candidate drugs that showed both lower predicted IC₅₀ values in the HRG and potential prognostic benefit.

Construction of DNA methylation and transcription factor regulatory networks

Methylation data for GBM were downloaded from the TCGA database, with β values serving as the basis for subsequent analysis. The missing rate for samples and probes was first calculated, and probes with <10% missing values were retained. Missing data were imputed through the impute.knn method (k=10) from impute package in R (v1.80.0). To explore the interactions between key genes and methylation sites, methylation sites with P<0.05 and their associated genes were imported into Cytoscape (v3.10.3) for visualization and topological feature analysis. The seven key genes were input into the NetworkAnalyst platform. Potential transcription factor binding sites (TFBSs) were predicted based on the JASPAR database. Gene-transcription factor regulatory relationships were attained. Subsequently, a transcriptional regulatory network was constructed and visualized in Cytoscape.

Proteomic validation and enrichment analysis

Data on GBM-related protein expression for 487 proteins were retrieved from the TCGA database. After excluding proteins missing in more than 50% of samples, missing values were imputed through the impute.knn function (k=10) from the impute package in R (v1.80.0). Differential expression analysis was implemented through the limma package (v3.62.2). Multiple testing was corrected utilizing the Benjamini-Hochberg method. A significance threshold of P.adjust <0.05 and |log₂FC| >log₂(1.2) was leveraged to ascertain 52 differentially expressed proteins (DEPs). Functional enrichment analyses of DEPs for GO and KEGG pathways were implemented utilizing the clusterProfiler package (v4.14.6), with a significance threshold of P.adjust <0.05. The results were demonstrated in a bubble plot.

Statistical analysis

All statistical analyses were performed using R software (v4.4.2). A two‑sided P<0.05 was considered statistically significant. Continuous variables were compared using the Wilcoxon rank‑sum test, and categorical variables were analyzed by the chi‑square test. Survival curves were constructed using the Kaplan-Meier method with the log‑rank test. Univariate and multivariate Cox regression were used to identify independent prognostic factors. Model performance was evaluated by time‑dependent ROC and C‑index. Correlations were assessed using Spearman’s correlation. Batch effects were corrected with the ComBat algorithm. Multiple testing was adjusted using the Benjamini-Hochberg method (P.adjust <0.05).

Results

Enrichment of DEGs reveals significance of the actin cytoskeleton pathway

This research integrated the GSE145645 and GSE184643 datasets, and batch effects were removed. The box plots and principal component analysis (PCA) plots before and after batch correction exhibited that the samples became more evenly distributed after correction (Figure 2A-2D). Based on the integrated data, DEGs were identified with a threshold of P.adjust <0.05 and |logFC| ≥ 1.5. In total, 2,450 DEGs were ascertained, of which 1,149 were upregulated and 1,301 were downregulated [supplementary table 1 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. The corresponding volcano plot and heatmap presented the distribution of the DEGs (Figure 2E,2F). GO and KEGG enrichment analyses revealed that the DEGs were significantly enriched in several signaling pathways, including regulation of membrane potential, neuron to neuron synapse, actin binding, tubulin binding, human papillomavirus infection, and neuroactive ligand signaling [Figure 2G,2H, supplementary tables 2,3 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. Among these, entries strongly linked to the regulation of the actin cytoskeleton included regulation of actin filament-based processes, regulation of actin cytoskeleton organization, regulation of actin filament polymerization, cerebral cortex cell migration, cell cortex, cortical actin cytoskeleton, stress fibers, actin binding, structural constituents of the cytoskeleton, regulation of actin cytoskeleton, focal adhesion, and cell adhesion molecules. To further validate the enrichment trend of this pathway at the whole transcriptome level, GSEA was performed on the regulation of actin cytoskeleton pathway. The results exhibited an upregulation trend among individuals with GBM (NES =1.414, P=0.01, Q=0.02165) (Figure 2I), suggesting its potentially important biological function within the context of this study.

Figure 2 Batch effect correction, identification of differentially expressed genes, and functional enrichment analysis. (A) Overall distribution and dispersion trend of each sample before batch correction, showing significant differences between samples; (B) after batch correction, the overall distribution of samples converged significantly, and batch effects were effectively mitigated; (C) PCA scatter plot (before batch correction), colored by dataset (GSE145645 and GSE184643). The first principal component (Dim1) explains 25.2% of the variance, and the second principal component (Dim2) explains 10.2%. The two batches are significantly separated in the PC space, indicating significant batch effects; (D) PCA scatter plot (after batch correction), colored by dataset. Dim1 explains 20.9% of the variance, and Dim2 explains 12.0%. The two batches significantly converge toward overlapping in the PC space, demonstrating effective reduction of batch effects; (E) volcano plot of DEGs. The horizontal axis signifies log₂(fold change), and the vertical axis signifies −log₁₀ (P.adjust). Upregulated and downregulated genes are indicated by different colors; (F) heatmap of DEGs expression, displaying gene expression patterns and clustering relationships among samples based on groupings; (G) bubble plot of GO enrichment analysis; (H) bubble plot of KEGG enrichment analysis; (I) GSEA enrichment analysis of genes related to the regulation of actin cytoskeleton pathway. BP, biological process; cAMP, cyclic adenosine monophosphate; CC, cellular component; cGMP, cyclic guanosine monophosphate; DEGs, differentially expressed genes; GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; NES, normalized enrichment score; P.adjust, adjusted P value; PC, principal component; PCA, principal component analysis; PKG, protein kinase G.

Construction and validation of RiskScore based on ML method

In the CGGA cohort of 179 patients with GBM, uni-Cox was implemented on 232 genes linked to the regulation of the actin cytoskeleton. The results detected that 8 genes were significantly linked to prognosis [P<0.05, Figure 3A, supplementary table 4 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. Subsequently, the predictive performance of 101 ML methods was evaluated across both CGGA and TCGA datasets. The optimal strategy was the combination of CoxBoost and RSF (Figure 3B). Seven key genes were selected: APC2, PPP1R12A, FGFR1, EGF, PIP5K1A, AKT1, and LPAR2. A riskScore was constructed based on the expression levels of these 7 genes. To contextualize the HRs, we summarized the expression distributions of the seven signature genes within the CGGA cohort [supplementary table 5 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. Most genes exhibited substantial variability across samples. For example, APC2 expression ranged from 0.49 to 321.57 [median 47.48; interquartile range (IQR), 49.55], while AKT1 ranged from 5.86 to 232.92 (median 65.65; IQR, 40.79) (Figure S1). Given the wide range of expression, even a modest per-unit HR may lead to substantial risk differences between low- and high-expression individuals.

Figure 3 Establishment and performance evaluation of the prognostic signature. (A) Forest plot of univariate Cox regression analysis. The plot displays the HR with 95% confidence interval and P value for each candidate gene in OS. A HR >1 indicates that high expression is linked to poor prognosis; (B) performance comparison plot of multiple modeling strategies. The figure shows that integrated or two-step combinations (e.g., CoxBoost + RSF) generally exhibit superior performance; (C) distribution plot of samples sorted by riskScore from low to high. The vertical axis signifies individual riskScore, while the horizontal axis signifies patients (ordered by increasing risk). Patients are stratified into high-risk groups (HRGs) and low-risk groups (LRGs) based on the median riskScore in the training set; (D) scatter plot of survival status increasing with riskScore. The horizontal axis signifies patients (ordered by increasing risk), and the vertical axis signifies follow-up time. The color and shape of the points distinguish between survival and death events; (E) K-M survival curves for OS between HRGs and LRGs. The HRG exhibits significantly worse survival relative to the LRG, with the number of risk samples provided at each time point; (F) time-dependent ROC curve estimating the discriminatory ability of the model. AUC, area under the curve; CGGA, Chinese Glioma Genome Atlas; CI, confidence interval; CoxBoost, Cox model with boosting; Enet, elastic net; GBM, gradient boosting machine; HR, hazard ratio; HRG, high-risk group; K-M, Kaplan-Meier; Lasso, least absolute shrinkage and selection operator; LRG, low-risk group; OS, overall survival; plsRcox, partial least squares regression for Cox model; plsRcox, partial least squares regression for Cox model; ROC, receiver operating characteristic; RSF, random survival forest; StepCox, stepwise Cox regression; SuperPC, supervised principal components; survivalSVM, survival support vector machine; TCGA, The Cancer Genome Atlas.

To clarify the relative contribution of each gene to the overall riskScore, we examined the regression coefficients derived from the multivariate Cox model. Among the seven genes, EGF exhibited the largest absolute coefficient, accounting for approximately 66.8% of the total absolute coefficient weight [supplementary table 6 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. However, single-gene survival analyses demonstrated limited predictive performance for individual genes, with C-indices ranging from 0.525 to 0.551 [supplementary table 6 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. In contrast, the combined seven-gene signature achieved a substantially higher C-index of 0.673 (Figure 3B). These findings indicated that although certain genes contributed more strongly to the linear predictor, the overall prognostic performance was primarily attributable to the integrated multi-gene model, not to any single dominant gene.

Patients were stratified into HRGs and LRGs based on the median riskScore (Figure 3C). In the CGGA cohort, the distribution of survival status indicated that patients within the HRG had poorer prognosis (Figure 3D). K-M analysis exhibited a significant difference in OS between the HRGs and LRGs (Figure 3E). The time-dependent ROC curve exhibited that the areas under the curve (AUCs) for the gene signature within 1–5 years of follow-up were greater than 0.6, suggesting favorable predictive performance (Figure 3F). Furthermore, survival analysis for each of the 7 genes individually indicated that high expression was associated with worse prognosis (P<0.05, Figure S2).

External validation was performed on the gene signature within the TCGA cohort with 169 samples. The results similarly demonstrated a significant difference in OS between the HRGs and LRGs (P=0.004, Figure 4A,4B), with individuals in the LRG exhibiting better prognosis relative to the HRG (Figure 4C). The time-dependent ROC curve indicated that the model maintained robust predictive performance within the TCGA cohort (Figure 4D).

Figure 4 Survival characteristics stratified by riskScore and performance validation in the TCGA dataset. (A) Distribution plot of samples ranked by individual riskScore from low to high. The vertical axis signifies the riskScore, and the horizontal axis signifies the order of patients; (B) scatter plot of survival status increasing with riskScore. The horizontal axis signifies the order of patients, and the vertical axis signifies follow-up time. The color of the points distinguishes between survival and death events; (C) K-M survival curves for HRGs and LRGs. The HRG exhibits significantly poorer survival compared to the LRGs. The number of patients surviving at each time point are demonstrated below; (D) time-dependent ROC curve estimating the discriminatory ability of the model. AUC, area under the curve; HRGs, high-risk groups; K-M, Kaplan-Meier; LRGs, low-risk groups; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas.

Uni-Cox revealed that the riskScore was considerably linked to prognosis among individuals with GBM (HR =1.501, 95% CI: 1.238–1.819, P<0.001), indicating that the riskScore was an adverse prognostic factor affecting patient survival (Figure 5A). In contrast, age, gender, and MGMT promoter methylation status exhibited no significant association with OS (P>0.05). Multivariate Cox regression analysis demonstrated consistent results (Figure 5B).

Figure 5 Prognostic value of the riskScore in OS. (A) Forest plot of univariate Cox regression analysis, showing the impact of riskScore, age, gender, and MGMTp methylation status on OS (total n=122); (B) forest plot of multivariate Cox regression analysis incorporating covariates riskScore, age, gender, and MGMTp methylation status (n=122). The riskScore remains an independent adverse prognostic factor. CI, confidence interval; MGMTp, MGMT promoter; OS, overall survival.

To evaluate the incremental prognostic value of the gene signature, we compared its performance with established clinical and molecular indicators.

In the CGGA cohort, for which MGMT promoter status was available, the clinical model (age, gender, and MGMT) achieved a C-index of 0.532. The signature improved the C-index to 0.597, and the combined model further increased it to 0.599 (Figure S3A). At 3 years, the AUC increased from 0.529 (clinical model) to 0.610 (signature) and reached 0.625 (combined model) (Figure S3B).

In the TCGA cohort, the clinical model comprising age and gender achieved a C-index of 0.545, whereas the gene signature enhanced discriminatory power to 0.658. The combined model integrating riskScore with clinical variables further elevated the C-index to 0.662 (Figure S3C). Similarly, the 3-year AUC improved from 0.579 (clinical model) to 0.669 (signature) and reached 0.689 (combined model) (Figure S3D).

These results demonstrated that the actin cytoskeleton-related signature provided complementary prognostic value beyond conventional clinical and molecular markers.

Comparative analysis between IDH wild-type and IDH-mutant gliomas

To further determine whether the actin cytoskeleton-related gene signature was specific to IDH wild-type gliomas or represented a broader glioma-associated molecular pattern, comparative analyses were conducted between IDH wild-type and IDH-mutant tumors in the CGGA cohort.

First, differential expression analysis revealed that six of the seven core genes (APC2, PPP1R12A, FGFR1, EGF, LPAR2, and PIP5K1A) exhibited significant expression differences between IDH wild-type and IDH-mutant tumors (all P<0.05). However, AKT1 did not show a statistically significant difference (Figure S4A). These findings suggested that the majority of signature genes showed IDH-status-associated expression patterns.

Next, comparison of the calculated riskScore demonstrated that IDH wild-type tumors exhibited significantly higher riskScores compared with IDH-mutant tumors (Figure S4B), indicating stronger activation of actin cytoskeleton–related molecular programs in IDH wild-type gliomas.

Furthermore, K-M survival analyses stratified by IDH status showed that the prognostic value of the riskScore remained significant within both molecular subgroups. In the IDH wild-type cohort, patients in the HRG exhibited significantly worse OS than those in the LRG (Figure 3E). Similarly, in the IDH-mutant cohort, risk stratification effectively distinguished survival outcomes, with the HRG showing significantly inferior survival (Figure S4C). Collectively, these results indicated that although the actin cytoskeleton-related signature demonstrated distinct expression patterns associated with IDH mutation status, its prognostic predictive capability was preserved across both IDH wild-type and IDH-mutant gliomas. Notably, the elevated riskScore in IDH wild-type tumors suggested that cytoskeletal remodeling may play a more prominent role in the aggressive phenotype of this subgroup.

Risk-associated modules identified by WGCNA and their functional enrichment analysis

Using WGCNA, a weighted gene co-expression network was developed. A soft threshold of β=12 was selected to ensure the network exhibited scale-free topology (Figure 6A), and a clustering dendrogram of different modules was generated (Figure 6B). Module-trait correlation analysis revealed that the brown and turquoise modules exhibited significant positive associations with risk stratification (r=0.2, P=0.008; r=0.28, P=2e−04; r=0.63, P=1e−21) (Figure 6C), suggesting heightened co-expression activity of these modules in high-risk individuals. This may potentially contribute to poor prognosis. In the scatter plots of module membership and gene significance, genes in the brown module exhibited weak and negative correlations with the HRG (Figure 6D), whereas the core genes in the turquoise module exhibited strong and positive correlations with the HRG (Figure 6E). Subsequent functional enrichment analysis demonstrated that genes in the brown module were considerably enriched in immune-related biological processes (BPs). GO enrichment showed that these genes were primarily involved in immune activation and effector processes such as positive regulation of lymphocyte activation, positive regulation of leukocyte activation, regulation of immune effector process, positive regulation of cell activation, and leukocyte cell-cell adhesion. For cellular components, enrichment was observed in tertiary granule, secretory granule lumen, vesicle lumen, cytoplasmic vesicle lumen, and external side of plasma membrane. Regarding molecular functions, enrichment was observed in immune-related functions like major histocompatibility complex (MHC) protein complex binding, peptide antigen binding, immune receptor activity, cytokine activity, and cytokine receptor binding (Figure 6F).

Figure 6 Workflow of WGCNA and phenotypic association and functional annotation of key modules. (A) Selection of soft-threshold (β) and assessment of scale-free topology. The left panel exhibits the change in the Scale-free topology fit (signed R²) as β increases, while the right panel presents the corresponding change in the average connectivity curve. The β value that achieves a high R² while maintaining reasonable connectivity is selected as the threshold for subsequent analysis; (B) dendrogram of gene clustering and module division through dynamic tree cut. Hierarchical clustering based on TOM is performed, with initial modules marked in different colors; (C) heatmap of module-trait correlation. Every cell demonstrates the Pearson correlation coefficient and P value between the module eigengenes with the HRGs and LRGs; (D) scatter plot of associations between module membership and gene significance for the brown module; (E) scatter plot of associations between module membership and gene significance for the turquoise module; (F,G) bubble plots of GO and KEGG enrichment for the brown module; (H,I) bubble plots of GO and KEGG enrichment for the turquoise module. ATP, adenosine triphosphate; BP, biological process; CC, cellular component; GO, Gene Ontology; HRGs, high-risk groups; KEGG, Kyoto Encyclopedia of Genes and Genomes; LRGs, low-risk groups; ME, module eigengene; MF, molecular function; MHC, major histocompatibility complex; TOM, topological overlap measure; WGCNA, Weighted Gene Co-expression Network Analysis.

KEGG enrichment analysis further revealed that genes in the brown module were considerably enriched in immune- and inflammation-related pathways, including complement and coagulation cascades, antigen processing and presentation, hematopoietic cell lineage, NF-κB signaling pathway, cytokine-cytokine receptor interaction, phagosome, leishmaniasis, and tuberculosis (Figure 6G).

Collectively, the brown module consistently exhibited features of immune response and inflammation regulation in both GO and KEGG analyses, suggesting its potential involvement in the immune microenvironment and immune evasion mechanisms of GBM.

The turquoise module contained numerous genes associated with RNA processing and chromosomal dynamics. GO enrichment analysis revealed that, among BPs, genes in this module were mainly enriched in post-transcriptional processing and chromosome-segregation events, including mRNA splicing, via spliceosome, RNA splicing, via transesterification reactions with bulged adenosine as nucleophile, chromosome segregation, and RNA splicing. For cellular components, genes were significantly enriched in spliceosomal complex, nuclear speck, chromosomal region, and chromosome, centromeric region. Regarding molecular functions, genes were significantly enriched in histone binding, transcription coactivator activity, ubiquitin-like protein transferase activity, aminoacyltransferase activity, and DNA-binding transcription factor binding (Figure 6H).

KEGG enrichment analysis further demonstrated that genes in the turquoise module were considerably enriched in ubiquitin-mediated proteolysis, protein processing in endoplasmic reticulum, cell cycle, endocytosis, mRNA surveillance pathway, autophagy-animal, and several cancer-related pathways like renal cell carcinoma, colorectal cancer, and thyroid hormone signaling pathway (Figure 6I).

Collectively, results of GO and KEGG enrichment analyses suggested that the turquoise module was primarily involved in fundamental BP like RNA splicing, chromosome segregation, protein processing, and protein degradation. Meanwhile, it was significantly linked to multiple tumor-related pathways. This suggested its potential role in transcriptional regulation and proteostasis maintenance in GBM.

Gene signature reveals differences in ICI and gene correlation

To further assess the association between risk stratification and the immune microenvironment, the CIBERSORT method was leveraged to estimate the proportion of ICI in the samples (Figure 7A). Box plots were generated to show the differential distribution of immune cells between HRGs and LRGs (Figure 7B). The results indicated that, in the HRG, the levels of activated dendritic cell and resting mast cell infiltration were elevated (P<0.05). The levels of monocyte and activated natural killer (NK) cell infiltration were decreased (P<0.001). To estimate the association between key genes and the tumor immune microenvironment, Spearman correlation analysis was implemented between the expression levels of the 7 candidate genes (AKT1, APC2, EGF, FGFR1, LPAR2, PIP5K1A, PPP1R12A) and the infiltration levels of 22 immune cell types. The results were displayed in a lollipop plot. Overall, most genes exhibited associations with the infiltration levels of monocytes, activated NK cells, and resting memory CD4 T cells. Specifically, AKT1 was primarily linked to monocytes, activated NK cells, and resting memory CD4 T cells, suggesting its close relationship with the monocyte/NK cell axis infiltration (Figure 7C). APC2 was linked to neutrophils and activated mast cells (Figure 7D). EGF was linked to activated NK cells, follicular helper T cells (Tfh), and B cell memory (Figure 7E). FGFR1 was linked to activated NK cells and monocytes (Figure 7F). LPAR2 was linked to resting memory CD4 T cells, activated mast cells, and monocytes (Figure 7G). PIP5K1A was linked to activated NK cells, monocytes, and CD8 T cells (Figure 7H). PPP1R12A was correlated with activated NK cells, CD8 T cells, and plasma cells (Figure 7I). These findings demonstrated that the candidate genes were stably linked to the infiltration of innate immune cells (like monocytes and NK cells) and specific T cell subsets, suggesting that they may participate in remodeling the immune microenvironment and influence the tumor immune phenotype.

Figure 7 Immune infiltration profile based on the gene signature using CIBERSORT and gene-immune correlation. (A) Stacked plot of immune cell composition at the sample level. This plot shows the relative distribution of 22 types of tumor-infiltrating immune cells across the samples, reflecting the overall heterogeneity of the immune ecosystem; (B) box plot of differences in immune cell proportion between HRGs and LRGs. The horizontal axis signifies the cell types, while the vertical axis signifies the estimated proportions; (C-I) correlation plots between the expression of AKT1, EGF, FGFR1, PPP1R2A, APC2, PIP5K1A, LPAR2 and immune cell proportions. The horizontal axis signifies the correlation coefficient. The size of the points indicates the P value and the strength of the correlation. The vertical axis signifies different immune cell types. *, P<0.05; ***, P<0.001. CIBERSORT, cell-type identification by estimating relative subsets of RNA transcripts; Cor, correlation coefficient; HRGs, high-risk groups; LRGs, low-risk groups; NK, natural killer; ns, not significant.

To further explore the immunological relevance of the signature, we evaluated its association with established TCGA immune subtypes (C1–C6). Most TCGA-GBM samples were classified as the C4 (lymphocyte-depleted) subtype (Figure S5A), and the distribution of immune subtypes did not differ significantly between HRGs and LRGs [supplementary table 7 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)].

Importantly, in a multivariate Cox model incorporating immune subtype, the riskScore remained an independent prognostic factor (HR =1.90, 95% CI: 1.44–2.51, P<0.001), whereas immune subtype (C4 vs. C1) was not significant (HR =1.47, 95% CI: 0.36–6.04, P=0.59) (Figure S5B). These results suggested that the prognostic signature facilitated risk stratification beyond existing immune-based subtype classification.

Immune infiltration patterns in relation to IDH status

Given that immune infiltration was significantly influenced by molecular subtypes, this study first compared the immune cell composition of IDH wild-type and IDH mutant gliomas in the CGGA cohort. Significant differences were observed in multiple immune populations, including macrophages M0, monocytes, regulatory T cells (Tregs), Tfh, and γδ T cells. This observation indicated that IDH mutation status had a significant impact on the immune microenvironment (Figure S6A).

Importantly, to determine whether the immune characteristics associated with risk signatures were independent of IDH status, a stratified analysis was conducted in the IDH mutant subgroup. Notably, some immune differences still exist within IDH mutant tumors. Specifically, B cell memory, T cells CD4 memory resting, and T cells CD4 naive showed significant differences between the HRGs and LRGs (Figure S6B). This was consistent with the presence of differences in immune infiltrating cells within IDH wild-type tumors.

These findings suggested that although IDH status led to immune heterogeneity, the immune remodeling captured by actin cytoskeleton-related signatures reflected additional biological mechanisms beyond molecular subtypes.

Differences in immune suppression and response to immunotherapy across HRGs and LRGs

To further assess the association between riskScore and characteristics of immune evasion, this research analyzed different risk groups utilizing the TIDE method. The results indicated significant differences in several immune-related indicators between HRGs and LRGs. Specifically, the TIDE score for the HRG was significantly elevated relative to the LRG (Figure 8A), suggesting that the HRG may exhibit stronger capabilities of immune evasion. Regarding immune suppression-related components, the level of MDSC exhibited no statistical difference between the HRGs and LRGs (Figure 8B), while the Exclusion score exhibited an increasing trend (Figure 8C). Regarding immune activation-related indicators, the levels of CD8⁺ T cell infiltration in the HRG were significantly elevated relative to the LRG (Figure 8D). The expression of the immune checkpoint molecule CD274 (PD-L1) was also higher in the HRG (Figure 8E). Additionally, the HRG exhibited stronger IFNG signaling activity (Figure 8F).

Figure 8 Differences in the expression or scores of key immune-related indicators between HRGs and LRGs. (A) The TIDE score is considerably elevated in the HRG (P=0.01), suggesting that the tumor may have stronger capabilities of immune evasion; (B) the MDSC score increases in the HRG (P=0.2), indicating that immune suppressive cells may be more active; (C) the Exclusion score is considerably elevated in the HRG (P=0.04), reflecting potential limitations in ICI into the tumor; (D) CD8⁺ T cell infiltration is considerably decreased in the HRG (P=0.009), suggesting impaired anti-tumor immune responses; (E) the expression of CD274 (PD-L1) is considerably elevated in the LRG (P=5.2×10⁻¹⁰), indicating that it might be more sensitive to immune checkpoint inhibition therapy; (F) IFNG is significantly elevated in the HRG (P=6.1×10⁻⁷), potentially reflecting its role in a high immune activity state. HRGs, high-risk groups; ICI, immune checkpoint inhibitor; IFNG, interferon gamma; LRGs, low-risk groups; MDSC, myeloid-derived suppressor cell; PD-L1, programmed death-ligand 1; TIDE, tumor immune dysfunction and exclusion.

Asoociation between gene signature and immune checkpoint molecules

To further investigate the immunological relevance of the actin cytoskeleton-related gene signature, the expression levels of multiple immune checkpoint molecules were compared between the HRG and LRG in the CGGA cohort.

The results showed that several classical immune checkpoint genes were significantly upregulated in the HRG, including CD274 (PD-L1), PDCD1 (PD-1), CTLA4, LAG3, TIGIT, HAVCR2 (TIM-3), BTLA, IDO1, and LGALS9. In addition, immunosuppressive cytokine-related genes such as TGFB1 and IL10 also exhibited higher expression levels in the HRG (Figure S7A). These findings suggested that tumors in the HRG exhibited a more immunosuppressive tumor microenvironment characterized by elevated immune checkpoint activity.

Furthermore, correlation analysis between the seven signature genes and immune checkpoint molecules revealed distinct regulatory patterns. Notably, LPAR2, EGF, FGFR1, and AKT1 showed significant positive associations with multiple checkpoint genes, including CD274, CTLA4, PDCD1, and LAG3. In contrast, APC2 exhibited weakly negative or non-significant associations with several immune checkpoint molecules (Figure S7B).

These findings indicated that the actin cytoskeleton-related gene signature was closely associated with immune checkpoint activation and may contribute to immune evasion in GBM.

Differences in mutational profiles and whole-cohort mutation characteristics in risk stratification

In the 159 samples from the TCGA cohort, mutation characteristics were compared between HRGs and LRGs. The overall mutational profile of the TCGA cohort revealed that missense mutations were the most prevalent mutation type. Single nucleotide polymorphisms (SNPs) were the dominant mutation form, with C>T transitions being the most common. The genes with the highest mutation frequencies across the entire cohort were TTN, TP53, EGFR, PTEN, MUC16, NF1, FLG, RYR2, LRP2, and PIK3CA (Figure 9A). Though no significant difference in the overall TMB between the two groups was detected (P=0.95, Figure 9B), there were differences in the types and proportions of gene mutations between the groups. In the HRG, 68 samples (86.08%) had mutations. The most common mutated genes were TP53, EGFR, PTEN, TTN, MUC16, and NF1 (Figure 9A). In the LRG, 76 samples (95%) had mutations. The most frequently mutated genes were TP53, TTN, EGFR, PTEN, MUC16, IDH1, and ATRX (Figures 9C,9D). Notably, K-M survival curves for TMB exhibited statistically significant differences (Figure 9E).

Figure 9 The correlation analysis between TMB, mutational profiles of genes, and survival. (A) The mutational profile of the top 30 most frequently mutated genes within the high-TMB group. The most commonly mutated genes include TP53 (46%), EGFR (27%), TTN (25%), and MUC16 (21%). Missense mutations are the predominant type. Mutations are detected in 69 samples (85.19%); (B) the top 30 high-frequency mutated genes within the low TMB group. Relative to the high TMB group, genes such as PIK3CA, ATRX, and IDH1 exhibit higher mutation frequencies in this group. The mutation types are more diverse. Mutations are detected in 75 samples, representing 96.15% of low-TMB samples; (C) the distribution of TMB in the overall population, showing differences in TMB between the HRGs and LRGs; (D) somatic mutational landscape of TCGA patients; (E) K-M survival curve shows that individuals within the high TMB group have considerably better OS relative to the low TMB group (P=0.004), denoting that TMB might serve as a potential prognostic biomarker. DEL, deletion; HRGs, high-risk groups; INS, insertion; K-M, Kaplan-Meier; LRGs, low-risk groups; OS, overall survival; SNP, single nucleotide polymorphism; TCGA, The Cancer Genome Atlas; TMB, tumor mutational burden.

Differences in drug sensitivity between predicted risk groups in the GDSC2 dataset

To further ascertain the potential therapeutic significance of the risk model, drug sensitivity analysis was conducted on 179 patients with GBM from the CGGA cohort utilizing the GDSC2 drug database. The top 5 drugs showing the most significant differences were identified, including Tozasertib_1096, Savolitinib_1936, AZD4547_1786, IWP-2_1576, and GSK591_2110 (Figures 10A-10E). These results implied that high-risk patients might be more sensitive to these drugs, providing support for the potential of the risk model in guiding personalized treatment strategies.

Figure 10 Predicted drug sensitivity for five drugs with significantly differences between HRGs and LRGs. (A-E) The differences in IC₅₀ for Tozasertib_1096, Savolitinib_1936, AZD4547_1786, IWP-2_1576, and GSK591_2110 between the HRGs and LRGs. HRGs, high-risk groups; IC₅₀, half maximal inhibitory concentration; LRGs, low-risk groups.

Methylation modifications of key genes and transcription factor regulatory network analysis

Network analysis based on methylation data revealed significant associations between the seven candidate genes and multiple specific loci. For instance, PPP1R12A was associated with several probes like cg14559388 and cg03165176. APC2 was linked to loci such as cg22816048 and cg15912732. The results indicated that the expression of these genes might be regulated by the status of DNA methylation (Figure 11A) [supplementary table 8 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)]. Transcription factor prediction results indicated that these candidate genes were potentially regulated by several classical transcription factors, including TP53, STAT3, NF-κB1, SP1, CREB1, FOXC1, and others (Figure 11B). These findings suggested that the seven key genes were influenced by epigenetic modifications and jointly regulated by multiple transcriptional regulatory pathways.

Figure 11 The interaction networks (PPI) between the seven core genes, methylation sites, and transcription factors. (A) The relationship network between core genes and their significantly associated DNA methylation sites (CpG sites); (B) the PPI network between the seven core genes and transcription factors. PPI, protein-protein interaction.

Proteomic characteristics and pathway enrichment analysis in risk stratification

Based on the riskScore constructed utilizing the genes EGF, PIP5K1A, LPAR2, APC2, PPP1R12A, FGFR1, and AKT1, 83 GBM samples from the TCGA cohort were stratified into HRGs and LRGs according to the median. Subsequently, differential protein analysis was performed. The volcano plot exhibited the overall distribution of significantly DEPs under the threshold of P.adjust <0.05 and |logFC| >log₂(1.2) (Figure 12A). In total, 52 DEPs were ascertained [supplementary table 9 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)], including key nodes related to receptor tyrosine kinases/PI3K-AKT and adhesion (e.g., EGFR, AKT/MTOR, FAK_pY397, fibronectin, PLC-γ1, GAB2, and PTEN/PHLPP).

Figure 12 Analysis of DEPs between HRGs and LRGs. (A) Volcano plot of DEPs. Screening results are plotted with log₂(fold change) on the x-axis and −log₁₀ (P.adjust) on the y-axis; (B) GO enrichment analysis of DEPs; (C) KEGG enrichment analysis of DEPs. BP, biological process; CC, cellular component; DEPs, differentially expressed proteins; GO, Gene Ontology; HRGs, high-risk groups; KEGG, Kyoto Encyclopedia of Genes and Genomes; LRGs, low-risk groups; MF, molecular function; P.adjust, adjusted P value.

GO enrichment analysis (P.adjust <0.05) revealed that, among BPs genes were mainly enriched in phosphatidylinositol 3-kinase/protein kinase B signal transduction, positive regulation of cell adhesion, morphogenesis of an epithelium, and negative regulation of the cell cycle. For cellular components, genes were concentrated in chromosomal region, pericentric heterochromatin, site of DNA damage/DNA double-strand break (DSB), promyelocytic leukemia (PML) body, and others. Regarding molecular functions, genes were involved in histone reader activity, transmembrane receptor protein tyrosine kinase adaptor activity, helicase/adenosine triphosphate (ATP)-dependent activity acting on DNA, and catalytic activity acting on DNA [Figure 12B, [supplementary tables 10,11 (available at https://cdn.amegroups.cn/static/public/tcr-2025-1-2865-1.zip)].

KEGG enrichment analysis (P.adjust <0.05) demonstrated that significant pathways included glioma, central carbon metabolism in cancer, cellular senescence, FoxO signaling pathway, and microRNAs in cancer (Figure 12C). These results suggested that the DEPs were concentrated in key signaling networks related to glioma development, migration, metabolic reprogramming, cellular senescence, and cellular response.

Discussion

The actin cytoskeleton is crucial in regulating the initiation, differentiation, invasion, and metastasis of tumors. Its dynamic remodeling significantly influences the malignant progression of GBM and prognosis of patients. This research identified seven core genes (APC2, PPP1R12A, FGFR1, EGF, PIP5K1A, AKT1, and LPAR2) from genes associated with the regulation of actin cytoskeleton based on 101 ML methods. A stable and prognostically significant riskScore model was constructed. This model demonstrated robust performance across multiple cohort validations, effectively distinguishing survival differences among patients with different risk profiles. It was further confirmed as an independent prognostic indicator for GBM. Although the magnitudes of several individual HRs were modest, interpretation of effect sizes should account for the observed expression variability across the cohort. Most signature genes demonstrated broad dynamic ranges, spanning several-fold differences between patients. Therefore, cumulative differences in expression between LRGs and HRGs may result in clinically relevant survival disparities. Importantly, the prognostic impact of the model arises from the synergistic effects of multiple genes, not any single component.

Importantly, the prognostic signature demonstrated incremental predictive value beyond established clinical variables, including MGMT status. The consistent improvement observed in both training and external validation cohorts supports its potential utility as a complementary prognostic tool in GBM.

It should be noted that the KEGG ‘regulation of actin cytoskeleton’ pathway, though comprehensive, may not encompass all core regulators of actin dynamics, including certain actin-binding proteins (ABPs) and actin-related proteins (ARPs) (14,15). Moreover, key regulators such as Rho GTPases are predominantly controlled at the post-translational level, and their functional activity may not be fully reflected by mRNA expression levels (16). Therefore, this study focuses specifically on transcriptome-level actin cytoskeleton-associated genes rather than characterizing the complete regulatory landscape of actin remodeling. Despite this limitation, the selected pathway provides a biologically relevant framework for identifying transcriptionally deregulated components associated with cytoskeletal dynamics in GBM.

Furthermore, we systematically ascertained the immune characteristics of different risk groups, revealing that the HRG exhibited stronger features of immune activation, immune exclusion, and immune evasion. This suggests that the model reflects the molecular heterogeneity of tumors and uncovers functional differences within their immune microenvironment. Based on risk stratification, predictions of drug sensitivity were further conducted. Five compounds with potential therapeutic value were identified, providing novel therapeutic insights for individualized treatment of high-risk patients. Overall, this model combines high predictive accuracy with biological interpretability, demonstrating promising clinical utility. It offers novel theoretical basis and decision-making support for risk assessment and precision therapy in GBM. Additional comparative analyses between IDH wild-type and IDH-mutant gliomas further demonstrated that the signature was more strongly activated in IDH wild-type tumors while retaining prognostic significance across both molecular subtypes. This suggests that the signature reflects cytoskeleton-related biological heterogeneity rather than serving merely as a surrogate for IDH mutation.

Firstly, among the seven key genes, APC2 is significantly upregulated by the drug AQB in both GBM and breast cancer, thereby inhibiting the growth and metastasis of tumors (17). In GBM, the complex formed by PPP1R12A binding to protein phosphatase 1 (PP1) can dephosphorylate MLC, thereby antagonizing MRCK-induced effects and inhibiting the invasiveness of tumor cells (18). Numerous studies have detected that FGFR1 facilitates the progression, invasion, and metastasis of GBM through pathways like the transcription factor ZEB1, and the Akt/GSK3β/snail signaling axis (19,20). This study suggests that in IDH wild-type GBM, FGFR1 may improve prognosis by inhibiting the growth and migration of GBM through the regulation of the actin cytoskeleton pathway. Numerous studies in GBM have demonstrated that EGF is strongly linked to the growth and metastasis of GBM cells (21,22). PIP5K1A may influence survival of GBM cells by activating phosphatidylinositol-4,5-bisphosphate (PIP2) or promoting stress-induced actin fiber formation (23-25). The PI3K/AKT signaling pathway is a major oncogenic pathway in GBM. Within this pathway, AKT1 functions as a key kinase. Its activation promotes the proliferation, migration, invasion, and tumorigenesis of tumor cells by mediating signals from upstream regulators (26,27). LPAR2 is an important member of the lysophosphatidic acid receptor family. It regulates various cellular processes, encompassing proliferation, migration, inflammatory response, and immune modulation, through activation of classical signaling pathways like MAPK/NF-κB. Prior evidence has reported that high expression of LPAR2 in head and neck squamous cell carcinoma and clear cell renal carcinoma is strongly linked to unfavorable prognosis (28,29), suggesting its potential pro-tumorigenic role within the progression of various malignant tumors. However, the function of LPAR2 in GBM remains underexplored. Given the highly invasive nature of GBM cells and their capability to remodel the immune microenvironment, it is hypothesized that LPAR2 may facilitate the migration and immune evasion of tumor cells by promoting the activation of the MAPK/NF-κB signaling axis. Moreover, studies have exhibited that NF-κB signaling is strongly linked to various drug resistance mechanisms (30). Sustained activation of the MAPK pathway may promote the maintenance of tumor stem cell properties (31,32). This implies that LPAR2 may contribute to therapeutic resistance in GBM. This offers theoretical support for the potential application of LPAR2-targeted inhibitors in treating GBM. Through the regulatory network analysis of the seven signature genes, this study identified significant correlations between several candidate genes and specific methylation loci, suggesting that the aberrant expression of these genes might be partially regulated by levels of DNA methylation.

Although this study provides comprehensive multi-omics validation supporting the prognostic and biological relevance of the seven-gene signature, several limitations should be acknowledged. First, this study primarily relies on retrospective transcriptomic data from public databases, and functional validation in vitro and in vivo was not performed. Therefore, the causal roles of these genes in regulating malignant phenotypes of GBM remain to be experimentally confirmed. Future studies integrating molecular perturbation experiments, such as gene knockdown/overexpression assays and animal models, are needed to further elucidate the mechanistic contributions of these genes to the progression of GBM and therapeutic response. Nevertheless, the multi-cohort validation and proteomic corroboration performed in this study provide strong supportive evidence for their biological relevance.

Further transcription factor-gene regulatory network analysis revealed that key genes (including EGF, PIP5K1A, LPAR2, APC2, PPP1R12A, FGFR1, and AKT1) were strongly linked to various classical cancer-related transcription factors, such as TP53, NF-κB1, STAT3, SP1, CREB1, and FOXC1. Previous studies have exhibited that oncogenic mutations in TP53 are linked to reduced survival within 1 year among individuals with GBM (33). However, in patients with OS exceeding 36 months, the presence of TP53 mutations was considerably correlated with prolonged survival (34), suggesting that TP53 may have heterogeneous biological effects across varied patient subgroups. Sustained activation of STAT3 and NF-κB signaling is a critical mechanism promoting the invasion and immune evasion of tumor cells. Additionally, SP1 and CREB1 implicate in the regulation of multiple pro-proliferation signaling pathways (35,36). High expression of FOXC1 is strongly linked to elevated invasiveness and poor prognosis in gliomas (37). These results suggest that key genes may promote the malignant progression of GBM under dual regulation by DNA methylation modifications and transcription factor networks. Notably, the analyses of DNA methylation and transcription factor regulation in this study were based on publicly available datasets and computational prediction algorithms. Although these findings suggest potential regulatory mechanisms underlying the identified signature, they remain hypothesis-generating observations.

Experimental validation, such as methylation-specific polymerase chain reaction (PCR), chromatin immunoprecipitation assays, or functional perturbation studies, is required to confirm causal regulatory relationships. Therefore, the proposed regulatory mechanisms should be interpreted cautiously and require further biological investigation. In validation at the proteomic level, DEPs were primarily enriched in pathways linked to cell proliferation, cell invasion, cellular senescence, and cancer-associated metabolic processes. Notably, numerous studies in GBM have demonstrated the AKT/mTOR and PI3K/FoxO pathways within the enriched pathways are correlated with drug resistance of tumors, maintenance of stemness, and poor prognosis (38,39). Conversely, aberrations in repair of DNA damage, progression of cell cycle, and cellular senescence may represent another crucial mechanism implicating in the progression of GBM and treatment resistance. These findings validate the gene signatures identified in this study at the protein level, suggesting that combined therapeutic strategies targeting the PI3K/AKT/mTOR and DNA damage repair pathways may offer new insights for improving prognosis of patients. Simultaneously, the results of the WGCNA enrichment analysis revealed the complex characteristics of immune microenvironment modulation and molecular metabolic reprogramming in GBM. Therefore, this research hypothesizes that the unfavorable prognosis of patients in HRGs might be linked to abnormal expression of protein, disruption of cell cycle, and abnormal remodeling of the immune microenvironment.

Regarding immune infiltration, higher infiltration of activated dendritic cells in the HRG may reflect altered antigen-presenting cell states in GBM. Prior studies have shown that glioma-associated dendritic cells can exhibit dysfunctional or tolerogenic features with impaired T-cell priming capacity, which may contribute to immune evasion (40,41). Increased resting mast cells in the HRG may also indicate microenvironmental remodeling, as mast cells have been implicated in shaping tumor-associated inflammation, angiogenesis, and immune regulation in the tumor microenvironment (42). In contrast, higher proportions of resting NK cells in the LRG may be consistent with relatively preserved innate immune surveillance, given that NK cells are critical for anti-tumor immunity but are often functionally suppressed in glioma (43). Finally, the observed monocyte difference should be interpreted cautiously given the heterogeneity of monocyte/macrophage lineages in GBM and their context-dependent roles. Nevertheless, monocytes can contribute to the development of tumor-associated macrophage programs that modulate immunosuppression and therapy resistance (44). This suggests that these genes are pivotal in regulating the immune evasion, inflammatory states, and immune therapy responses of tumors. Subsequent TIDE analysis demonstrated significant differences in TIDE and Exclusion scores between HRGs and LRGs, indicating enhanced capability of immune escape in the HRG. This aligns with prior studies, such as Du et al., which highlighted the inherent immune-evasion microenvironment in GBM (45). However, no significant differences were observed in MDSC, suggesting that the immunosuppressive effects between the LRGs and HRGs are not significantly different (46,47). Regarding immune activation indicators, the HRG exhibited elevated levels of immune activation. The enhanced synergy between CD8⁺ T cells and CD274 (PD-L1), along with the upregulation of immunosuppressive receptors (PD-1, CTLA-4, TIGIT, LAG3) in HRGs, indicates that the host has the fundamental potential to combat tumors. These patients may benefit from immunotherapy. Notably, IFNG is linked to unfavorable prognosis among individuals with acute myeloid leukemia (48). Ultimately, the results emphasize that immune evasion is more potent and dominates disease progression in GBM. This might be linked to certain tumor-secreted mediators. Reversing the suppressive tumor microenvironment could unleash more effective intrinsic anti-tumor responses.

Moreover, the mutational profiles of HRGs and LRGs revealed distinct patterns. Results revealed that the most common mutation type was missense mutations. SNPs were the dominant mutation form, with C>T transitions being the most common. This is consistent with the mutation characteristics of GBM, thus demonstrating the reliability of the samples. After comparing the mutational profiles between the HRGs and LRGs, both groups exhibited mutations in classic GBM driver genes, such as TP53, EGFR, and PTEN. This suggests that these alterations are common features in the development of disease. No significant differences were observed in TMB between the HRGs and LRGs. This indicates that the biological characteristics reflected by the riskScore are not solely driven by the accumulation of mutations but may be more strongly linked to the activation of signaling pathways, cytoskeletal remodeling, and changes in immune microenvironment. However, in the full cohort analysis, patients with high TMB exhibited significantly reduced OS (P=0.004), suggesting that an elevated mutation burden remains strongly linked to unfavorable prognosis (49). The HRG exhibited enriched mutations in PDGFRA and multiple PI3K family genes, suggesting aberrant activation of PI3K/AKT pathway may enhance the proliferation, invasion, and migration of cells (50-52). This is highly consistent with the results of differential protein enrichment. Simultaneously, mutations in chromatin remodeling genes such as CHD3 (53) reflected epigenetic dysregulation. The LRG exhibited significant enrichment of IDH1 and ATRX mutations. Alterations in genes such as RB1 and PIK3CA were also observed. This mutational profile aligns with the previously reported IDH1/ATRX subtype (54), typically associated with favorable prognosis. Collectively, these findings reveal significant differences in driver-gene mutation profiles between the HRGs and LRGs, indicating that the signature stratifies clinical outcomes and may reflect distinct molecular pathological mechanisms.

Finally, regarding drug sensitivity analysis, tozasertib, savolitinib, AZD4547, IWP-2, and GSK591 demonstrated higher predicted sensitivity in the HRG. This suggests that high-risk stratification indicates unfavorable prognosis and may reveal vulnerabilities that can be targeted by specific pathways. As a broad-spectrum Aurora kinase inhibitor, tozasertib can block mitosis and induce apoptosis. It has demonstrated selective cytotoxicity in multiple glioma models and is linked to radiotherapy or resistance mechanisms, supporting its potential utility for HRGs (55). Although the therapeutic effect of tozasertib in GBM remains unreported, existing studies have confirmed that Aurora kinase inhibitors could serve as potential therapeutic agents for GBM (56). As a selective MET inhibitor, Savolitinib has demonstrated feasibility in pediatric diffuse high-grade glioma (57). Given the enrichment of growth factor/RTK axes (EGF/FGFR1) and PI3K-AKT signaling in the HRG, the higher predicted sensitivity of savolitinib in this subgroup is biologically plausible. This provides statistical robustness but requires further validation for risk-stratified treatment in this population. However, the role of savolitinib in GBM has not been reported, and its specific function and mechanism in this disease remain to be further investigated. AZD4547 targets FGFR1–3 and has been evaluated in tissue-agnostic basket trials and recurrent high-grade gliomas (including those with FGFR-TACC fusions), suggesting the potential for precision treatment in FGFR-dependent subgroups. This aligns with the enrichment of the FGFR pathway in our HRG (58). The efficacy of this drug in GBM is inferior to that of CYY292, which is a small-molecule inhibitor sharing the same mechanism of action. This discrepancy is largely attributed to restricted drug delivery across the blood-brain barrier. Therefore, alternative administration routes may be needed to improve efficacy (19). As an antagonist of the Wnt pathway, IWP-2 inhibits the viability of various cancerous cell lines (59). However, its application in GBM remains unexplored. Therefore, IWP-2 could serve as a candidate drug for future research. GSK591 is a PRMT5 inhibitor. Studies targeting GBM stem-like cells have exhibited that PRMT5 inhibition disrupts splicing and stemness, significantly inhibiting growth. This supports the higher sensitivity signals observed in the HRGs (60). Meanwhile, this study also implies that GSK591 moderately penetrates the blood-brain barrier. Importantly, preclinical studies have confirmed that it confers significant survival benefits in GBM xenograft mouse models (60). Collectively, these five agents target critical nodes in cell cycle/mitosis (Aurora), growth factor receptor axes (MET/FGFR), secretion dependence of Wnt, and epigenetic splicing (PRMT5). This aligns with the high-risk features of the EGFR-FGFR-AKT axis, Wnt/β-catenin, and RNA/epigenetic regulation identified by our risk model. These results support the biological reliability of the gene signature and provide potential directions for integrating riskScore with prediction of drug response, offering insights into precision treatment in GBM. Nevertheless, further pharmacokinetic evaluations, in vitro and in vivo validations, and clinical investigations remain essential to assess the clinical utility of these predicted drugs in GBM, particularly regarding blood-brain barrier penetration and efficacy within the central nervous system.

Although this prognostic signature has been validated in independent TCGA and CGGA cohorts, differences in patient demographics, sequencing platforms, and preprocessing procedures may introduce variability in gene expression profiles. Despite these potential sources of heterogeneity, the gene signature has shown consistent survival stratification and stable predictive performance across different datasets. This suggests that its robustness is not affected by dataset-specific effects. However, it is still necessary to conduct prospective multicenter validation under standardized sequencing conditions to further confirm its clinical utility.

We systematically ascertained the role of actin cytoskeleton-associated genes in GBM at the transcriptomic level and constructed a prognostically significant gene signature utilizing ML method. Nevertheless, this research has certain limitations. First, this research primarily relied on transcriptomic data from public databases like TCGA and CGGA for bioinformatics analysis. The results lack validation through cellular and animal experiments and functional mechanisms of key genes. What’s more, although external data validation was performed, the limited sample size might affect the generalizability of the model. Moreover, the drug sensitivity analysis was based on computational predictions, which still require confirmation through in vitro drug sensitivity experiments and clinical sample validation to further assess their clinical feasibility. Additionally, though preliminary exploration was conducted on ICI and the regulation of transcription factor, the causal relationship between the gene signature and immune modulation remains incompletely elucidated.

Therefore, future studies should integrate multi-omics analyses with experimental validation to clarify the mechanistic roles of this signature in the pathogenesis of GBM at cellular, animal, and clinical levels, and evaluate its potential as a prognostic biomarker and therapeutic target.

Conclusions

Through multi-omics integrative analysis, this research developed and validated a prognostic signature model based on seven key genes. This model serves as an independent indicator of OS among individuals with GBM, demonstrating robust performance and generalizability across multiple cohorts from CGGA and TCGA. Further analysis suggested that the genes linked to the model are co-regulated by DNA methylation and transcription factor networks. They are significantly enriched in pathways linked to actin cytoskeleton remodeling, PI3K/AKT/mTOR signaling, and immune evasion. This suggests their potential role in the invasive growth of tumors, regulation of immune microenvironment, and therapeutic response. This research offers insights into the molecular heterogeneity of GBM, providing novel candidate biomarkers and a theoretical foundation for personalized prognostic assessment and targeted therapeutic strategies. Future studies integrating proteomics, phosphoproteomics, and small GTPase activity assays would be necessary to comprehensively characterize post-translational regulation of actin dynamics in GBM.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2865/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2865/prf

Funding: This study was supported by the Health Commission of Henan Province (No. CKQ20250026).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2865/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Louis DN, Perry A, Wesseling P, et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol 2021;23:1231-51. [Crossref] [PubMed]
2. Gittleman H, Boscia A, Ostrom QT, et al. Survivorship in adults with malignant brain and other central nervous system tumor from 2000-2014. Neuro Oncol 2018;20:vii6-vii16. [Crossref] [PubMed]
3. McKinnon C, Nandhabalan M, Murray SA, et al. Glioblastoma: clinical presentation, diagnosis, and management. BMJ 2021;374:n1560. [Crossref] [PubMed]
4. Weller M, Wen PY, Chang SM, et al. Glioma. Nat Rev Dis Primers 2024;10:33. [Crossref] [PubMed]
5. Dapash M, Hou D, Castro B, et al. The Interplay between Glioblastoma and Its Microenvironment. Cells 2021;10:2257. [Crossref] [PubMed]
6. Colamaria A, Leone A, Fochi NP, et al. Tumor treating fields for the treatment of glioblastoma: Current understanding and future perspectives. Surg Neurol Int 2023;14:394. [Crossref] [PubMed]
7. Kaur R, Bhardwaj A, Gupta S. Cancer treatment therapies: traditional to modern approaches to combat cancers. Mol Biol Rep 2023;50:9663-76. [Crossref] [PubMed]
8. Chen X, Xu Z, Tang K, et al. The Mechanics of Tumor Cells Dictate Malignancy via Cytoskeleton-Mediated APC/Wnt/β-Catenin Signaling. Research (Wash D C) 2023;6:0224.
9. Liang EY, Huang MH, Chen YT, et al. Tanshinone IIA modulates cancer cell morphology and movement via Rho GTPases-mediated actin cytoskeleton remodeling. Toxicol Appl Pharmacol 2024;483:116839. [Crossref] [PubMed]
10. Sun J, He D, Fu Y, et al. A novel lncRNA ARST represses glioma progression by inhibiting ALDOA-mediated actin cytoskeleton integrity. J Exp Clin Cancer Res 2021;40:187. [Crossref] [PubMed]
11. Meng Y, Huang K, Shi M, et al. Research Advances in the Role of the Tropomyosin Family in Cancer. Int J Mol Sci 2023;24:13295. [Crossref] [PubMed]
12. Akdaş EY, Temizci B, Karabay A. miR96- and miR182-driven regulation of cytoskeleton results in inhibition of glioblastoma motility. Cytoskeleton (Hoboken) 2023;80:367-81. [Crossref] [PubMed]
13. Wang J, Yang Q, Li X, et al. Engineering overexpressing SYNGR1 inhibited the progression of GBM cells by suppressing the intracellular FGF1-mediated LDs accumulation and cytoskeleton remodeling. J Neurooncol 2025;174:667-87. [Crossref] [PubMed]
14. Xue Y, Xue C, Song W. Emerging roles of deubiquitinating enzymes in actin cytoskeleton and tumor metastasis. Cell Oncol (Dordr) 2024;47:1071-89. [Crossref] [PubMed]
15. Yadav T, Gau D, Roy P. Mitochondria-actin cytoskeleton crosstalk in cell migration. J Cell Physiol 2022;237:2387-403. [Crossref] [PubMed]
16. Gorla M, Guleria DS. Rho GTPase Signaling: A Molecular Switchboard for Regulating the Actin Cytoskeleton in Axon Guidance. J Cell Physiol 2025;240:e70005. [Crossref] [PubMed]
17. Li Y, Ren Y, Wang Y, et al. A Compound AC1Q3QWB Selectively Disrupts HOTAIR-Mediated Recruitment of PRC2 and Enhances Cancer Therapy of DZNep. Theranostics 2019;9:4608-23. [Crossref] [PubMed]
18. Birch JL, Strathdee K, Gilmour L, et al. A Novel Small-Molecule Inhibitor of MRCK Prevents Radiation-Driven Invasion in Glioblastoma. Cancer Res 2018;78:6509-22. [Crossref] [PubMed]
19. Bi Y, Zheng R, Hu J, et al. A novel FGFR1 inhibitor CYY292 suppresses tumor progression, invasion, and metastasis of glioblastoma by inhibiting the Akt/GSK3β/snail signaling axis. Genes Dis 2024;11:479-94. [Crossref] [PubMed]
20. Jermakowicz AM, Kurimchak AM, Johnson KJ, et al. RAPID resistance to BET inhibitors is mediated by FGFR1 in glioblastoma. Sci Rep 2024;14:9284. [Crossref] [PubMed]
21. Yamaguchi M, Shimokawa N, Murata T. Extracellular regucalcin reveals anti-cancer activity by suppressing cell growth and metastatic activity by blocking EGF signaling pathway in human glioblastoma cells in vitro. Cell Signal 2025;132:111844. [Crossref] [PubMed]
22. Zheng Y, Yang W, Aldape K, et al. Epidermal growth factor (EGF)-enhanced vascular cell adhesion molecule-1 (VCAM-1) expression promotes macrophage and glioblastoma cell interaction and tumor cell invasion. J Biol Chem 2013;288:31488-95. [Crossref] [PubMed]
23. Bassi C, Mello SS, Cardoso RS, et al. Transcriptional changes in U343 MG-a glioblastoma cell line exposed to ionizing radiation. Hum Exp Toxicol 2008;27:919-29. [Crossref] [PubMed]
24. Frisan T, Cortes-Bratti X, Chaves-Olarte E, et al. The Haemophilus ducreyi cytolethal distending toxin induces DNA double-strand breaks and promotes ATM-dependent activation of RhoA. Cell Microbiol 2003;5:695-707. [Crossref] [PubMed]
25. Mejillano M, Yamamoto M, Rozelle AL, et al. Regulation of apoptosis by phosphatidylinositol 4,5-bisphosphate inhibition of caspases, and caspase inactivation of phosphatidylinositol phosphate 5-kinases. J Biol Chem 2001;276:1865-72. [Crossref] [PubMed]
26. Shetty C, Tamatta R, Dhas N, et al. Evolving therapeutic strategies in glioblastoma: traditional approaches and novel interventions. 3 Biotech 2025;15:318.
27. Zheng T, Huang Y, Chu D, et al. COL22A1 Activates the PI3K/AKT Signaling Pathway to Sustain the Malignancy of Glioblastoma. Int J Genomics 2025;2025:6587097. [Crossref] [PubMed]
28. Sun K, Chen RX, Li JZ, et al. LPAR2 correlated with different prognosis and immune cell infiltration in head and neck squamous cell carcinoma and kidney renal clear cell carcinoma. Hereditas 2022;159:16. [Crossref] [PubMed]
29. Wang Y, Qi Z, Li Z, et al. LPAR2-mediated action promotes human renal cell carcinoma via MAPK/NF-κB signaling to regulate cytokine network. J Cancer Res Clin Oncol 2023;149:2041-55. [Crossref] [PubMed]
30. Ryan SL, Beard S, Barr MP, et al. Targeting NF-κB-mediated inflammatory pathways in cisplatin-resistant NSCLC. Lung Cancer 2019;135:217-27. [Crossref] [PubMed]
31. He J, Huang Z, He M, et al. Circular RNA MAPK4 (circ-MAPK4) inhibits cell apoptosis via MAPK signaling pathway by sponging miR-125a-3p in gliomas. Mol Cancer 2020;19:17. [Crossref] [PubMed]
32. Xu P, Zhang G, Hou S, et al. MAPK8 mediates resistance to temozolomide and apoptosis of glioblastoma cells through MAPK signaling pathway. Biomed Pharmacother 2018;106:1419-27. [Crossref] [PubMed]
33. Esperante D, Galicia KD, Rivas-Cuervo KG, et al. TP53 oncogenic variants as prognostic factors in individuals with glioblastoma: a systematic review and meta-analysis. Front Neurol 2024;15:1490246. [Crossref] [PubMed]
34. Di Nunno V, Gatto L, Tosoni A, et al. TP53 mutations and survival in patients with histologically defined Glioblastoma, IDH-wildtype. Pathol Res Pract 2024;262:155516. [Crossref] [PubMed]
35. Zhao X, Shen F, Ma J, et al. CREB1-induced miR-1204 promoted malignant phenotype of glioblastoma through targeting NR3C2. Cancer Cell Int 2020;20:111. [Crossref] [PubMed]
36. Liu X, Sun C, Wang Q, et al. Sp1 Upregulation Bolsters the Radioresistance of Glioblastoma Cells by Promoting Double Strand Breaks Repair. Int J Mol Sci 2023;24:10658. [Crossref] [PubMed]
37. Cao Q, Wang X, Shi Y, et al. FOXC1 silencing inhibits the epithelial to mesenchymal transition of glioma cells: Involvement of β catenin signaling. Mol Med Rep 2019;19:251-61. [Crossref] [PubMed]
38. Sun P, Liu F, Huo K, et al. Adiponectin facilitates the cell cycle, inhibits cell apoptosis and induces temozolomide resistance in glioblastoma via the Akt/mTOR pathway. Oncol Lett 2025;29:127. [Crossref] [PubMed]
39. Cheng Y, Xiao Z, Cai W, et al. Suppression of FOXO1 activity by SIRT1-mediated deacetylation weakening the intratumoral androgen autocrine function in glioblastoma. Cancer Gene Ther 2025;32:343-54. [Crossref] [PubMed]
40. Woroniecka K, Chongsathidkiet P, Rhodin K, et al. T-Cell Exhaustion Signatures Vary with Tumor Type and Are Severe in Glioblastoma. Clin Cancer Res 2018;24:4175-86. [Crossref] [PubMed]
41. Jin X, Kang J, Lu Q, et al. Fc gamma receptor IIb in tumor-associated macrophages and dendritic cells drives poor prognosis of recurrent glioblastoma through immune-associated signaling pathways. Front Genet 2022;13:1046008. [Crossref] [PubMed]
42. Varricchi G, Galdiero MR, Loffredo S, et al. Are Mast Cells MASTers in Cancer? Front Immunol 2017;8:424. [Crossref] [PubMed]
43. Arnesen VS, Gras Navarro A, Chekenya M. Challenges and Prospects for Designer T and NK Cells in Glioblastoma Immunotherapy. Cancers (Basel) 2021;13:4986. [Crossref] [PubMed]
44. Chen Z, Hambardzumyan D. Immune Microenvironment in Glioblastoma Subtypes. Front Immunol 2018;9:1004. [Crossref] [PubMed]
45. Du R, Jing L, Fu D. Immune intrinsic escape signature stratifies prognosis, characterizes the tumor immune microenvironment, and identifies tumorigenic PPP1R8 in glioblastoma multiforme patients. Front Immunol 2025;16:1577920. [Crossref] [PubMed]
46. Yin B, Cai Y, Chen L, et al. Immunosuppressive MDSC and Treg signatures predict prognosis and therapeutic response in glioma. Int Immunopharmacol 2024;141:112922. [Crossref] [PubMed]
47. He H, Yan M, Ye K, et al. Predicting Prognosis and Immunotherapy Response in Glioblastoma (GBM) With a 5-Gene CAF-Risk Signature. Cancer Rep (Hoboken) 2025;8:e70158. [Crossref] [PubMed]
48. Wen XM, Xu ZJ, Ma JC, et al. Identification and validation of necroptosis-related gene signatures to predict clinical outcomes and therapeutic responses in acute myeloid leukemia. Aging (Albany NY) 2023;15:14677-702. [Crossref] [PubMed]
49. Khasraw M, Walsh KM, Heimberger AB, et al. What is the Burden of Proof for Tumor Mutational Burden in gliomas? Neuro Oncol 2020;23:17-22. [Crossref] [PubMed]
50. Ortiz Gómez LD, Contreras Martínez HJ, Galvis Pareja DA, et al. Mutations in the PIK3C2B, ERBB3, KIT, and MLH1 Genes and Their Relationship with Resistance to Temozolomide in Patients with High-Grade Gliomas. Biomedicines 2024;12:2777. [Crossref] [PubMed]
51. Higa N, Akahane T, Yokoyama S, et al. Prognostic impact of PDGFRA gain/amplification and MGMT promoter methylation status in patients with IDH wild-type glioblastoma. Neurooncol Adv 2022;4:vdac097. [Crossref] [PubMed]
52. Carlotto BS, Trevisan P, Provenzi VO, et al. PDGFRA, KIT, and KDR Gene Amplification in Glioblastoma: Heterogeneity and Clinical Significance. Neuromolecular Med 2023;25:441-50. [Crossref] [PubMed]
53. Li Y, Min W, Li M, et al. Identification of hub genes and regulatory factors of glioblastoma multiforme subgroups by RNA-seq data analysis. Int J Mol Med 2016;38:1170-8. [Crossref] [PubMed]
54. Wang M, Yang C, Zhang X, et al. Characterizing genomic differences of human cancer stratified by the TP53 mutation status. Mol Genet Genomics 2018;293:737-46. [Crossref] [PubMed]
55. Jane EP, Premkumar DR, Rajasundaram D, et al. Reversing tozasertib resistance in glioma through inhibition of pyruvate dehydrogenase kinases. Mol Oncol 2022;16:219-49. [Crossref] [PubMed]
56. Lee PY, Chen CL, Lin ZZ, et al. The Aurora kinases inhibitor VE-465 is a novel treatment for glioblastoma multiforme. Oncology 2013;84:326-35. [Crossref] [PubMed]
57. Zuckermann M, He C, Andrews J, et al. Capmatinib is an effective treatment for MET-fusion driven pediatric high-grade glioma and synergizes with radiotherapy. Mol Cancer 2024;23:123. [Crossref] [PubMed]
58. Chae YK, Hong F, Vaklavas C, et al. Phase II Study of AZD4547 in Patients With Tumors Harboring Aberrations in the FGFR Pathway: Results From the NCI-MATCH Trial (EAY131) Subprotocol W. J Clin Oncol 2020;38:2407-17. [Crossref] [PubMed]
59. García-Reyes B, Witt L, Jansen B, et al. Discovery of Inhibitor of Wnt Production 2 (IWP-2) and Related Compounds As Selective ATP-Competitive Inhibitors of Casein Kinase 1 (CK1) δ/ε. J Med Chem 2018;61:4087-102. [Crossref] [PubMed]
60. Sachamitr P, Ho JC, Ciamponi FE, et al. PRMT5 inhibition disrupts splicing and stemness in glioblastoma. Nat Commun 2021;12:979. [Crossref] [PubMed]