Immune-related hallmark pathway heterogeneity identifies a prognostic risk signature with independent external validation in glioblastoma Authors: Ahmed M Abdelmaguid, Rasha M A Eltanany, Manal F El-Khadragy, Ahmed E Abdel Moneim Published: 2026 ABSTRACT AbstractGlioblastoma (GBM) is the most aggressive primary malignant brain tumor in adults and remains associated with poor clinical outcomes despite advances in surgery, radiotherapy, and chemotherapy. Increasing evidence suggests that the glioblastoma immune microenvironment and coordinated oncogenic signaling pathways play critical roles in tumor progression, therapeutic resistance, and survival heterogeneity. However, the prognostic relevance of pathway-level immune activity in GBM remains incompletely understood. In this study, we performed a comprehensive transcriptomic and pathway-centered analysis of glioblastoma using RNA-sequencing data and clinical information obtained from The Cancer Genome Atlas (TCGA-GBM) cohort. Single-sample gene set enrichment analysis (ssGSEA) was applied to quantify hallmark pathway activity across tumors, followed by immune subtype clustering, survival analysis, Cox regression modeling, machine learning-based pathway prioritization, and prognostic risk signature construction. The reproducibility of the developed prognostic model was subsequently evaluated using an independent Chinese Glioma Genome Atlas (CGGA) validation cohort. Our analysis demonstrated substantial inter-patient variability in immune-related and oncogenic hallmark signaling pathways. Inflammatory and mesenchymal-associated pathways, including TNF-α/NF-κB signaling, IL6/JAK/STAT3 signaling, interferon responses, epithelial–mesenchymal transition, hypoxia, complement activation, and KRAS signaling, showed marked differences across patients and were strongly associated with aggressive disease phenotypes. Two immune-associated molecular subtypes with distinct pathway activation patterns were identified. Furthermore, multiple hallmark pathways demonstrated significant prognostic associations and were integrated into a pathway-derived prognostic risk signature that successfully stratified patients into high-risk and low-risk groups with significantly different overall survival outcomes. Multivariate Cox regression analysis confirmed that the calculated risk score retained independent prognostic significance after adjustment for clinical variables. Importantly, external validation using the CGGA cohort demonstrated stable prognostic performance and reproducibility across independent patient populations. Taken together, these findings demonstrate that immune-related hallmark pathway activity contributes substantially to glioblastoma biological heterogeneity and patient prognosis. The proposed pathway-based prognostic framework provides a biologically interpretable approach for survival stratification and may support future efforts toward precision oncology and pathway-oriented therapeutic investigation in glioblastoma.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05696-5.Keywords: Glioblastoma, Hallmark pathways, Tumor microenvironment, ssGSEA, Prognostic signature, Immune heterogeneity, TCGA, CGGA, Survival analysis, Bioinformatics Introduction Glioblastoma (GBM) is the most aggressive and lethal primary malignant tumor of the central nervous system in adults, accounting for the majority of high-grade gliomas and remaining associated with extremely poor clinical outcomes despite decades of therapeutic advances [1]. Current standard treatment, which includes maximal surgical resection followed by radiotherapy and temozolomide-based chemotherapy, provides only limited survival benefit, with median overall survival typically ranging between 12 and 18 months [2]. The persistent therapeutic resistance and rapid recurrence observed in GBM are largely attributed to extensive intratumoral heterogeneity, diffuse infiltrative growth, metabolic adaptability, and profound remodeling of the tumor microenvironment [3]. Recent advances in transcriptomic profiling have substantially improved the understanding of glioblastoma biology and revealed that GBM represents a biologically heterogeneous disease composed of multiple molecular and cellular states rather than a single uniform entity [4]. Large-scale projects such as The Cancer Genome Atlas (TCGA) have demonstrated that glioblastoma exhibits distinct transcriptional subtypes associated with differences in oncogenic signaling, immune activity, treatment response, and clinical prognosis [5]. However, despite these advances, the clinical translation of molecular stratification strategies in GBM remains limited, partly because many existing approaches rely primarily on individual gene expression patterns that may not fully capture the complex biological interactions driving tumor progression. Increasing evidence suggests that pathway-level analyses may provide a more biologically interpretable framework for understanding tumor behavior than single-gene approaches alone [6]. Hallmark signaling pathways integrate coordinated transcriptional programs associated with inflammation, cell-cycle progression, hypoxia, epithelial–mesenchymal transition (EMT), immune regulation, and metabolic adaptation, all of which are critically involved in glioblastoma progression [7]. In particular, the glioblastoma immune microenvironment has emerged as a major determinant of disease behavior. GBM tumors are characterized by extensive infiltration of tumor-associated macrophages, microglia, and immunosuppressive immune populations that collectively shape inflammatory signaling networks and contribute to immune evasion, therapeutic resistance, and tumor aggressiveness [8]. Several inflammatory pathways, including TNF-α/NF-κB signaling, IL6/JAK/STAT3 signaling, interferon-associated pathways, complement activation, and hypoxia-driven transcriptional programs, have been implicated in glioblastoma progression and poor prognosis [9, 10]. Nevertheless, the prognostic significance of coordinated immune-related pathway activity in GBM remains incompletely understood. Moreover, although machine learning and computational prognostic modeling approaches have become increasingly common in cancer bioinformatics, many previously proposed glioblastoma signatures suffer from limited reproducibility, overfitting, or lack of external validation [11]. In the present study, we performed a comprehensive pathway-centered transcriptomic analysis of glioblastoma using ssGSEA-derived hallmark pathway activity profiles generated from TCGA-GBM transcriptomic data. By integrating immune landscape characterization, immune subtype clustering, survival analysis, Cox regression modeling, machine learning-based pathway prioritization, and prognostic risk signature development, we aimed to identify biologically meaningful pathway patterns associated with patient survival. Importantly, the prognostic framework was externally validated using an independent Chinese Glioma Genome Atlas (CGGA) cohort to evaluate its reproducibility and generalizability across independent patient populations. Collectively, this study sought to establish a biologically interpretable and clinically relevant pathway-based prognostic framework capable of improving understanding of immune-associated signaling heterogeneity in glioblastoma while identifying candidate pathways with potential prognostic and therapeutic relevance. Materials and methods Data collection and preprocessing RNA sequencing (RNA-seq) transcriptomic profiles and corresponding clinical information for glioblastoma (GBM) were obtained from The Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM) cohort through the UCSC Xena platform and the Genomic Data Commons (GDC) data portal [12, 13]. Gene expression datasets generated using the STAR alignment pipeline and TPM normalization were downloaded, along with corresponding clinical phenotypes and overall survival information. The TCGA cohort initially included RNA-seq expression data, clinical data, and survival information for glioblastoma patients. To externally validate the proposed prognostic model, an independent glioblastoma cohort was obtained from the Chinese Glioma Genome Atlas (CGGA) database [14]. Transcriptomic expression profiles and corresponding clinical data were downloaded from the official CGGA portal (http://www.cgga.org.cn/download.jsp). All preprocessing and downstream bioinformatics analyses were performed in Python using Google Colaboratory. Patient identifiers were extracted from TCGA sample barcodes using the first 12 characters of each sample identifier to ensure accurate matching between transcriptomic and survival datasets. Duplicate samples were subsequently removed to ensure that each patient was represented only once during downstream analyses. Genes with missing expression values were excluded before dimensionality reduction and pathway enrichment analyses. Ensembl gene identifiers were converted into official HGNC gene symbols using the MyGene Python package to improve compatibility with downstream analyses [15]. Unmapped genes and duplicated identifiers were removed to generate a cleaned transcriptomic matrix suitable for downstream analyses. Exploratory transcriptomic analysis and quality control An exploratory transcriptomic analysis was performed to evaluate global gene expression distributions, assess normalization consistency, and identify transcriptomic heterogeneity among glioblastoma samples. Principal component analysis (PCA) was conducted following z-score normalization of the expression matrix using the scikit-learn Python package [16]. PCA visualization was subsequently used to assess transcriptomic clustering patterns and identify potential outlier samples. Outlier detection was additionally performed using PCA coordinates to identify transcriptomic samples with extreme expression profiles. Hierarchical sample correlation heatmaps were generated using highly variable genes to evaluate transcriptomic similarity among glioblastoma samples. All generated figures were exported in publication-quality PNG and PDF formats. Single-sample gene set enrichment analysis (ssGSEA) Pathway enrichment analysis was performed using single-sample Gene Set Enrichment Analysis (ssGSEA) implemented through the GSEApy Python package [17]. Hallmark gene signatures from the Molecular Signatures Database (MSigDB Hallmark 2020 collection) were used to quantify pathway activity scores across glioblastoma samples [6]. The Hallmark pathway collection was selected because it summarizes biologically coherent signaling pathways while reducing redundancy among gene sets. The generated pathway enrichment matrices were subsequently used for immune landscape characterization, clustering analysis, survival modeling, machine learning analysis, and prognostic signature construction. Immune landscape characterization and clustering analysis To characterize the biological and immune heterogeneity of glioblastoma samples, pathway activity heatmaps were generated using ssGSEA-derived enrichment scores. Variance analysis was performed to identify highly variable pathways across the cohort. Pathways associated with inflammatory signaling, interferon response, epithelial–mesenchymal transition, hypoxia, cytokine signaling, and proliferative programs were prioritized for downstream analyses. Unsupervised clustering analysis was subsequently performed using pathway enrichment profiles to identify immune-associated glioblastoma subtypes. Patients were stratified into immune-related molecular subgroups according to transcriptomic pathway activity patterns. Cluster assignments were integrated with survival datasets for downstream prognostic analyses. Survival and cox regression analyses Kaplan–Meier survival analysis was performed to evaluate survival differences between immune-associated patient subgroups and pathway-derived prognostic groups. Patients were stratified into high- and low-risk groups based on the median calculated pathway-derived risk score. Survival curves and Cox proportional hazards regression analyses were conducted using the Lifelines Python package [18]. Hazard ratios (HRs), confidence intervals (CIs), and corresponding p-values were subsequently calculated for both univariate and multivariate Cox regression analyses. An initial univariate Cox proportional hazards regression analysis was performed to identify prognostic hallmark pathways associated with overall survival. Significant pathways identified during univariate analyses were subsequently integrated into downstream prognostic modeling. Multivariate Cox proportional hazards regression was subsequently conducted to determine whether the proposed pathway-derived risk signature was an independent prognostic factor after adjustment for clinical covariates. Machine learning analysis Machine learning approaches were implemented as complementary exploratory analyses to prioritize biologically relevant hallmark pathways based on ssGSEA-derived pathway enrichment scores. Random Forest and Extreme Gradient Boosting (XGBoost) classifiers were constructed to estimate pathway importance rather than to develop the final prognostic model. The prognostic risk signature was subsequently established using Cox proportional hazards regression analysis based on survival-associated pathways. Model performance was evaluated using receiver operating characteristic (ROC) analysis, classification accuracy, and pathway importance ranking. Given the limited discriminative performance of the machine learning models, these analyses were interpreted primarily as supportive biological evidence for pathway prioritization rather than as independent prognostic classifiers. The machine learning workflow was implemented using the scikit-learn and XGBoost Python frameworks [16]. Construction of the prognostic risk signature To establish a robust and reproducible pathway-based prognostic framework, pathway activity scores were first quantified for all 47 Hallmark gene sets using the single-sample Gene Set Enrichment Analysis (ssGSEA) algorithm implemented in the GSEApy package. The resulting normalized enrichment scores (NES) were calculated for each patient and used as pathway activity scores for subsequent survival analyses. Univariate Cox proportional hazards regression analysis was then performed to evaluate the association between each Hallmark pathway and overall survival. Pathways with P < 0.05 were considered statistically significant prognostic candidates. To construct a parsimonious prognostic signature while minimizing model complexity, the five most statistically significant pathways (ranked according to ascending P-values) were selected and incorporated into a multivariate Cox proportional hazards regression model to estimate their independent regression coefficients (β). The prognostic risk score for each patient was subsequently calculated as a weighted linear combination of the selected pathway activity scores according to the following equation: where n represents the number of pathways included in the final prognostic model (n = 5), βi denotes the regression coefficient estimated from the multivariate Cox proportional hazards model for the i-th pathway, and Si represents the corresponding ssGSEA enrichment score of that pathway for an individual patient. The regression coefficients of all pathways included in the final prognostic signature are presented in the Results section. Following risk score calculation, the median risk score of the TCGA discovery cohort was used as the predefined cutoff to stratify patients into high-risk and low-risk groups. The prognostic performance of the constructed signature was subsequently evaluated using Kaplan–Meier survival analysis, risk score distributions, pathway expression heatmaps, and time-dependent receiver operating characteristic (ROC) curves. Finally, the prognostic signature was externally validated using the independent CGGA cohort following the same risk score formula and stratification strategy. Independent prognostic analysis To evaluate the independent prognostic value of the proposed pathway-based risk signature, multivariate Cox regression analysis was performed using both pathway-derived risk scores and clinical variables. Clinical covariates included age and gender information extracted from TCGA clinical datasets. Hazard ratios, confidence intervals, and corresponding p-values were calculated to determine the prognostic contribution of each variable. External validation using the CGGA cohort To validate the robustness and generalizability of the proposed prognostic model, an external validation analysis was performed using the independent CGGA glioblastoma cohort. The same preprocessing workflow and ssGSEA analytical pipeline applied to the TCGA cohort were also applied to the CGGA dataset to ensure analytical consistency. Pathway-derived risk scores were subsequently calculated in the validation cohort using the regression coefficients previously established in the TCGA cohort. Patients in the CGGA cohort were stratified into high- and low-risk groups based on the median risk score. Kaplan–Meier survival analysis, Cox proportional hazards regression analysis, and receiver operating characteristic analyses were subsequently performed to evaluate validation performance. Time-dependent ROC analysis and statistical analysis Time-dependent receiver operating characteristic (ROC) analysis was performed to evaluate prognostic model performance across multiple survival intervals. Predictive performance was assessed at 1-year, 3-year, and 5-year survival endpoints using time-dependent area under the curve (AUC) analysis. All statistical analyses and computational workflows were performed in Python using reproducible computational pipelines. Computational analyses were conducted using the pandas, numpy, scipy, scikit-learn, matplotlib, seaborn, lifelines, gseapy, and xgboost Python libraries. Kaplan–Meier survival curves were compared using the log-rank test, while Cox proportional hazards regression models were used for survival modeling. ROC analyses were used to evaluate predictive performance. A two-sided p-value less than 0.05 was considered statistically significant throughout all analyses. Results Data collection, preprocessing, and quality control Transcriptomic expression profiles, clinical annotations, and survival information for glioblastoma (GBM) patients were obtained from the TCGA-GBM cohort through the UCSC Xena platform. Following sample harmonization and preprocessing, a total of 162 GBM tumor samples with matched transcriptomic and survival information were retained for downstream analyses. Initial quality assessment revealed the presence of missing values primarily within the clinical dataset, whereas the survival dataset contained no missing entries (Fig. 1A; Table 1). The baseline demographic and clinicopathological characteristics of both the TCGA discovery cohort and the independent CGGA validation cohort are summarized in Supplementary Table S1. Quality control and exploratory transcriptomic analysis of the TCGA-GBM cohort. A Distribution of normalized global gene expression values after log2(TPM + 1) transformation. B Principal component analysis (PCA) showing transcriptomic heterogeneity among GBM samples. C PCA-based outlier detection identifies transcriptomic outlier samples within the cohort Table 1Summary of TCGA-GBM preprocessing and quality control proceduresParameterValueInitial expression matrix dimensions4,990 genes × 162 samplesSamples with missing transcriptomic values1Final cleaned expression matrix4,989 genes × 162 samplesMissing values in the expression dataset168Missing values in clinical dataset11,502Missing values in survival dataset0Variance explained by PC119.48%Variance explained by PC28.71%Number of PCA-defined outliers9Final cohort used for downstream analyses162 patients Summary of TCGA-GBM preprocessing and quality control procedures To improve dataset consistency and analytical robustness, preprocessing procedures were subsequently applied to the transcriptomic matrix, including removal of unresolved missing values, harmonization of patient identifiers, and expression matrix cleaning. After preprocessing, one transcriptomic sample containing unresolved missing values was excluded, resulting in a final cleaned expression matrix consisting of 4,989 genes across 162 GBM samples (Table 1). Global transcriptomic distribution analysis demonstrated that the normalized gene expression values followed an approximately log-normal distribution after log2(TPM + 1) transformation, supporting the suitability of the dataset for downstream computational analyses (Fig. 1A). Principal component analysis (PCA) was subsequently performed to investigate transcriptomic variability and global sample structure within the GBM cohort. The first principal component (PC1) explained 19.48% of the total variance, whereas the second principal component (PC2) explained 8.71% of the variance (Fig. 1B). PCA-based outlier analysis identified nine transcriptomic outlier samples exhibiting marked deviation from the major sample distribution (Fig. 1C). Nevertheless, the majority of samples demonstrated relatively stable clustering behavior, indicating acceptable transcriptomic consistency across the cohort. To further evaluate inter-sample transcriptomic similarity and molecular heterogeneity, sample-to-sample correlation analysis was performed using the top 500 most variable genes. Hierarchical clustering revealed distinct transcriptional substructures among GBM patients, highlighting the substantial biological heterogeneity of the TCGA-GBM cohort (Fig. 2). Multiple highly correlated sample clusters were observed, suggesting the presence of diverse molecular states within GBM tumors. The corresponding non-clustered correlation heatmap is additionally provided in Supplementary Figure S1. Clustered sample-to-sample correlation heatmap of the TCGA-GBM cohort generated using the top 500 most variable genes. Hierarchical clustering revealed distinct transcriptional substructures and heterogeneous molecular similarity patterns among GBM samples, indicating substantial biological heterogeneity across the cohort Collectively, these preprocessing and quality control analyses confirmed the robustness and suitability of the TCGA-GBM cohort for downstream immune infiltration analysis, prognostic modeling, machine learning prioritization, and external validation studies. Immune landscape characterization of GBM using ssGSEA To investigate the immune and biological heterogeneity of glioblastoma, single-sample gene set enrichment analysis (ssGSEA) was performed using Hallmark gene sets across the TCGA-GBM cohort. The resulting pathway activity matrix demonstrated substantial inter-patient variability in both immune-associated and oncogenic signaling pathways. Hierarchical clustering of pathway enrichment scores revealed distinct transcriptional activity patterns among GBM samples, indicating the presence of biologically heterogeneous immune states within the cohort (Fig. 3). ssGSEA-based immune pathway activity heatmap showing the heterogeneous immune and oncogenic signaling landscape across TCGA-GBM samples. Hierarchical clustering was used solely for heatmap visualization and revealed distinct pathway activation patterns among patients, with marked enrichment of inflammatory and proliferative hallmark pathways Several inflammatory and immune-related pathways showed marked enrichment across subsets of patients, including TNF-α signaling via NF-κB, interferon-α response, interferon-γ response, inflammatory response, allograft rejection, and IL6/JAK/STAT3 signaling. In parallel, multiple proliferation-associated pathways, including E2F targets, G2M checkpoint, mitotic spindle, MYC targets V2, epithelial–mesenchymal transition (EMT), and hypoxia signaling, also demonstrated elevated activity across multiple GBM samples. These findings collectively suggest that GBM exhibits simultaneous activation of immune-regulatory and aggressive tumor-promoting transcriptional programs. Among the evaluated hallmark pathways, the highest variability across patients was observed for Pancreas Beta Cells, TNF-α signaling via NF-κB, E2F targets, interferon-α response, interferon-γ response, inflammatory response, and G2M checkpoint pathways (Table 2). The substantial variance observed within these pathways further supports the existence of distinct immune and molecular phenotypes among GBM tumors. Table 2Top highly variable hallmark pathways identified by ssGSEA analysis in the TCGA-GBM cohortPathwayVariancePancreas Beta Cells0.007680TNF-alpha Signaling via NF-kB0.006567E2F Targets0.006494Interferon Alpha Response0.006103Interferon Gamma Response0.006060Inflammatory Response0.005991G2-M Checkpoint0.005816IL-6/JAK/STAT3 Signaling0.005525Epithelial-Mesenchymal Transition0.004424Allograft Rejection0.004309Mitotic Spindle0.003410Apical Surface0.003142Myc Targets V20.002793Hypoxia0.002564KRAS Signaling Up0.002161 Top highly variable hallmark pathways identified by ssGSEA analysis in the TCGA-GBM cohort Identification of immune subtypes in GBM To further characterize the immune heterogeneity of glioblastoma, unsupervised K-means clustering was performed using standardized ssGSEA-derived hallmark pathway enrichment scores. Patients were classified into two immune-associated molecular subtypes, designated as Immune_High and Immune_Low, based on Euclidean distance to the nearest cluster centroid. The Immune_High subtype included 90 patients, whereas the Immune_Low subtype consisted of 72 patients (Fig. 4). Heatmap showing differential hallmark pathway activation patterns between the Immune_High and Immune_Low glioblastoma subtypes identified using unsupervised K-means clustering of standardized ssGSEA enrichment scores The Immune_High subtype demonstrated markedly elevated enrichment of multiple inflammatory and immune-related signaling pathways, including TNF-α signaling via NF-κB, IL6/JAK/STAT3 signaling, IL2/STAT5 signaling, inflammatory response, allograft rejection, interferon-α response, and interferon-γ response. In contrast, the Immune_Low subtype exhibited comparatively reduced activation of immune-associated pathways and lower inflammatory signaling activity (Fig. 4). To further validate subtype separation, principal component analysis was performed using pathway enrichment scores. PCA demonstrated clear transcriptional segregation between the Immune_High and Immune_Low groups, supporting the robustness of the identified immune subtypes and confirming substantial pathway-level heterogeneity among GBM patients (Fig. 5). Principal component analysis demonstrating the separation of the Immune_High and Immune_Low subgroups identified by unsupervised K-means clustering based on standardized ssGSEA-derived pathway activity profiles Table 3Distribution of immune subtypes identified in the TCGA-GBM cohortImmune SubtypeNumber of PatientsImmune_High90Immune_Low72 Distribution of immune subtypes identified in the TCGA-GBM cohort . Prognostic impact of immune subtypes in GBM To evaluate the prognostic significance of the identified immune-associated subtypes, Kaplan–Meier survival analysis was performed comparing patients classified as Immune_High and Immune_Low. Survival analysis demonstrated distinct overall survival trends between the two groups; however, the observed difference did not reach statistical significance (Fig. 6). Kaplan–Meier survival analysis comparing overall survival between Immune_High and Immune_Low glioblastoma subtypes identified through ssGSEA-based pathway clustering. No statistically significant survival difference was observed between the two immune-associated subgroups (log-rank p = 0.43872) Patients within the Immune_High subgroup exhibited a median overall survival of 442 days, whereas patients classified as Immune_Low demonstrated a median overall survival of 375 days (Table 4). Despite the slightly prolonged survival observed in the Immune_High subgroup, log-rank testing revealed no statistically significant survival difference between the two immune phenotypes (log-rank p = 0.43872). Table 4Survival characteristics of immune-associated GBM subtypesImmune subtypeNumber of patientsMedian overall survival (Days)Log-rank p-valueImmune_high904420.43872Immune_low723750.43872 Survival characteristics of immune-associated GBM subtypes These findings suggest that although substantial immune-related transcriptional heterogeneity exists across GBM tumors, immune subtype classification alone may not be sufficient to fully predict patient prognosis. This observation further supports the need for integrative prognostic models incorporating multiple molecular and pathway-level features. Identification of prognostic hallmark pathways in GBM To identify biological pathways associated with patient prognosis, univariate Cox proportional hazards regression analysis was performed using ssGSEA-derived hallmark pathway enrichment scores. A total of 12 hallmark pathways demonstrated statistically significant associations with overall survival in the TCGA-GBM cohort (Table 5). Table 5Significant hallmark pathways identified by univariate cox regression analysis in the TCGA-GBM cohortPathwayHazard ratio (HR)p-valueEstrogen response late1.350.000945Coagulation1.330.002821UV response up1.330.002961Apical surface1.330.003274Epithelial mesenchymal transition1.310.003850KRAS signaling up1.280.009448Complement1.250.014906Glycolysis1.220.020757Apical junction1.240.023038TNF-alpha signaling via NF-kB1.230.024577IL-2/STAT5 signaling1.210.032632Estrogen response early1.200.040872 Significant hallmark pathways identified by univariate cox regression analysis in the TCGA-GBM cohort Among the identified prognostic pathways, Estrogen Response Late, Coagulation, UV Response Up, Apical Surface, and Epithelial–Mesenchymal Transition (EMT) exhibited the strongest associations with unfavorable prognosis. In addition, several immune-related pathways, including TNF-α signaling via NF-κB, IL2/STAT5 signaling, inflammatory response, complement activation, and allograft rejection, were also significantly associated with patient survival (Fig. 7). Forest plot of significant hallmark pathways identified by univariate Cox regression analysis in the TCGA-GBM cohort. Multiple inflammatory, proliferative, metabolic, and epithelial transition-related pathways demonstrated significant associations with overall survival Multiple proliferation-associated and stress-response pathways, including KRAS signaling, glycolysis, and apical junction signaling, similarly demonstrated elevated hazard ratios, suggesting that coordinated activation of inflammatory, metabolic, and proliferative transcriptional programs contributes substantially to GBM aggressiveness and poor clinical outcomes. These prognostic hallmark pathways were subsequently utilized for downstream machine learning analysis and prognostic risk signature construction. Machine learning-based identification of prognostic hallmark pathways To complement the survival-based statistical analyses, exploratory machine learning approaches were applied to prioritize biologically relevant hallmark pathways according to their relative importance. Random Forest and Extreme Gradient Boosting (XGBoost) models were constructed using the ssGSEA-derived hallmark pathway enrichment scores. These models were implemented as complementary feature-prioritization tools rather than as standalone prognostic classifiers. Feature importance analysis derived from the XGBoost model identified several inflammatory and mesenchymal signaling pathways as the most influential contributors within the machine learning framework. Among the highest-ranked pathways were TNF-alpha signaling via NF-kB, Complement, KRAS signaling Up, Apical Junction, and Estrogen Response Late, suggesting that immune-inflammatory activation and epithelial–mesenchymal transition-associated processes may contribute to the biological heterogeneity of glioblastoma (Fig. 8A). Additional pathways, including Coagulation, Apical Surface, and UV Response Up, also demonstrated notable importance scores. Exploratory machine learning analysis for hallmark pathway prioritization in glioblastoma. A XGBoost feature importance ranking identifying the most biologically informative hallmark pathways based on ssGSEA-derived pathway activity scores. B SHAP (SHapley Additive exPlanations) summary plot illustrating the relative contribution of individual hallmark pathways to model output across patients. C Receiver operating characteristic (ROC) curves comparing the discriminative performance of the Random Forest and XGBoost models, demonstrating their limited predictive performance and supporting their use as exploratory feature-prioritization tools rather than independent prognostic classifiers To further improve the biological interpretability of these findings, SHAP (SHapley Additive exPlanations) analysis was performed to quantify the contribution of individual hallmark pathways to model predictions. The SHAP summary plot demonstrated heterogeneous pathway-specific effects across patients, with pathways such as Complement, Estrogen Response Early, Epithelial–Mesenchymal Transition, and Apical Junction exhibiting both positive and negative contributions depending on pathway activity levels (Fig. 8B). These observations further support the biological heterogeneity of glioblastoma and illustrate the complex interactions among hallmark pathways. The discriminative performance of the machine learning models was subsequently evaluated using receiver operating characteristic (ROC) curve analysis. The Random Forest and XGBoost models achieved AUC values of 0.457 and 0.493, respectively (Fig. 8C), indicating limited discriminative performance for survival prediction within the present cohort. Therefore, these machine learning models were not considered suitable as independent prognostic classifiers. Instead, their primary contribution was to provide complementary biological evidence through pathway importance ranking, whereas the final prognostic signature was established exclusively using Cox proportional hazards regression based on survival-associated hallmark pathways (Table 6). Table 6Exploratory machine learning analysis for hallmark pathway prioritization in glioblastomaModelAUC valueTop predictive hallmark pathwaysRandom forest0.457Complement, KRAS Signaling Up, TNF-alpha Signaling via NF-kBXGBoost0.493TNF-alpha Signaling via NF-kB, Complement, KRAS Signaling Up Exploratory machine learning analysis for hallmark pathway prioritization in glioblastoma Construction and evaluation of the prognostic risk signature To further investigate the prognostic significance of hallmark pathway activity in glioblastoma, a prognostic risk signature was constructed using the survival-associated hallmark pathways identified through Cox proportional hazards regression analysis. Individual risk scores were calculated for each patient using the regression coefficients derived from the multivariable Cox model, and patients were subsequently stratified into high-risk and low-risk groups according to the predefined median risk score. Kaplan–Meier survival analysis demonstrated a significant difference in overall survival between the two risk groups (Fig. 9A). Patients classified into the high-risk group exhibited markedly poorer survival outcomes compared with those in the low-risk group (log-rank p = 0.01568), indicating that the constructed prognostic signature successfully stratified glioblastoma patients according to survival probability. Construction and evaluation of the prognostic risk signature in glioblastoma, including Kaplan–Meier survival analysis comparing overall survival between high-risk and low-risk patient groups stratified according to the prognostic risk signature (A), ROC curve analysis evaluating the predictive performance of the prognostic model (B), and distribution analysis of calculated patient risk scores across the TCGA glioblastoma cohort (C) The predictive performance of the risk signature was further evaluated using receiver operating characteristic (ROC) curve analysis. The prognostic model achieved an AUC value of 0.623, indicating acceptable discriminative performance for survival stratification within the TCGA cohort (Fig. 9B). Distribution analysis of the calculated risk scores demonstrated substantial heterogeneity among glioblastoma patients (Fig. 9C). Most patients exhibited intermediate risk values clustered around the median cutoff, whereas smaller subsets of patients demonstrated markedly elevated or reduced risk scores, reflecting distinct prognostic profiles within the cohort (Table 7). Table 7Prognostic performance of the glioblastoma risk signatureParameterValueTotal patients162High-risk patients81Low-risk patients81Log-rank p-value0.01568ROC AUC0.623Median risk score cutoff0.032 Prognostic performance of the glioblastoma risk signature Independent prognostic value of the risk signature To determine whether the constructed prognostic risk signature could independently predict survival outcomes in glioblastoma patients, multivariate Cox proportional hazards regression analysis was performed by integrating the calculated risk score with available clinical variables, including age and gender. The multivariate Cox regression model demonstrated that the prognostic risk score remained a significant independent predictor of overall survival after adjustment for clinical covariates (Fig. 10). The calculated hazard ratio for the risk score was markedly elevated (HR = 2.54), indicating that patients with increased risk scores exhibited substantially poorer survival outcomes compared with low-risk patients. In contrast, gender did not demonstrate significant prognostic association, whereas age showed only a limited contribution to survival prediction. Independent prognostic analysis of glioblastoma survival demonstrating the hazard ratios and 95% confidence intervals for clinical variables and the constructed prognostic risk signature using multivariate Cox regression analysis These findings suggested that the constructed hallmark pathway-based risk signature retained independent prognostic value beyond conventional clinical characteristics and may therefore represent a robust molecular predictor of glioblastoma patient survival (Table 8). Table 8Multivariate Cox regression analysis of independent prognostic factors in glioblastomaVariableHazard ratio (HR)95% CIP-valueRisk score2.5431.568–4.1240.000154Age1.0211.007–1.0350.003622 Gender 1.084 0.766–1.534 0.648712 Multivariate Cox regression analysis of independent prognostic factors in glioblastoma External validation of the prognostic risk signature using the CGGA cohort To validate the robustness and reproducibility of the prognostic risk signature, an independent external validation analysis was performed using the CGGA glioblastoma cohort comprising 313 patients. Patients were stratified into a high-risk group (n = 159) and a low-risk group (n = 154) according to the median risk score calculated in the validation cohort. Kaplan–Meier survival analysis demonstrated a clear survival disadvantage in the high-risk group compared with the low-risk group (Fig. 11A). The high-risk subgroup exhibited significantly reduced overall survival, whereas patients in the low-risk subgroup showed prolonged survival, confirming the reproducibility and robustness of the proposed prognostic model in an independent external cohort. External validation of the TCGA-derived prognostic risk signature in the CGGA glioblastoma cohort, including Kaplan–Meier survival analysis between high-risk and low-risk patient groups (A), ROC curve analysis evaluating predictive performance of the external validation model (B), time-dependent ROC analysis across multiple survival periods (C), and nomogram-style visualization of patient-specific risk score distribution and clinical characteristics (D) The predictive capability of the prognostic signature was further assessed using receiver operating characteristic (ROC) curve analysis. The external validation ROC curve showed favorable discrimination, with an AUC of 0.712 (Fig. 11B), supporting the reliability of the prognostic model in the independent CGGA cohort. In addition, time-dependent ROC analysis demonstrated stable predictive performance across multiple follow-up periods, with AUC values gradually increasing over time (Fig. 11C). These findings indicate that the prognostic model retained stable long-term predictive performance during external validation (Table 9). Table 9External validation performance of the prognostic risk signature in the CGGA glioblastoma cohortParameterResultValidation cohortCGGA glioblastoma cohort (n = 313)High-risk patients159Low-risk patients154Survival comparisonHigh-risk vs. Low-riskLog-rank p-value< 0.001ROC AUC0.7121-year AUC0.693-year AUC0.725-year AUC0.73Prognostic performanceStable external validationClinical applicabilityConfirmed External validation performance of the prognostic risk signature in the CGGA glioblastoma cohort A nomogram-style distribution analysis was additionally performed to visualize the relationship between patient-specific risk scores and clinical variables within the CGGA cohort (Fig. 11D). Considerable heterogeneity in patient risk distribution was observed, further supporting the potential clinical utility of the proposed prognostic signature for individualized survival risk stratification. Discussion Glioblastoma remains one of the most aggressive and clinically challenging primary brain tumors, characterized by extensive molecular diversity, profound immune dysregulation, and persistently poor patient survival despite multimodal therapy. Rather than focusing exclusively on individual genes, the present study investigated glioblastoma at the pathway level using ssGSEA-derived hallmark signaling activity profiles. By integrating immune landscape characterization, survival modeling, machine learning prioritization, and external cohort validation, we established a pathway-based prognostic framework capable of stratifying GBM patients into biologically and clinically distinct risk groups. Our analysis demonstrated marked variability in immune-related and oncogenic pathway activity across the TCGA-GBM cohort. Pathways associated with inflammatory signaling, interferon responses, cell-cycle progression, epithelial–mesenchymal transition (EMT), and hypoxia showed considerable differences between patients, supporting the concept that GBM is not a uniform disease but rather a collection of transcriptionally diverse tumor ecosystems. Unsupervised K-means clustering identified two immune-associated subtypes based on standardized ssGSEA-derived pathway enrichment scores. The Immune_High subtype showed enrichment of TNF-α/NF-κB signaling, IL6/JAK/STAT3 signaling, inflammatory response pathways, and interferon-associated programs, whereas the Immune_Low subtype exhibited comparatively reduced immune pathway activity. Similar observations have been reported previously, where inflammatory signaling networks and tumor-associated macrophages were shown to contribute substantially to GBM progression and immune remodeling [10, 19]. Among the identified pathways, the TNF-α/NF-κB axis emerged as one of the most biologically relevant components associated with aggressive disease behavior. NF-κB signaling has been repeatedly implicated in glioma proliferation, resistance to apoptosis, inflammatory cytokine production, and treatment adaptation [20, 21]. Previous studies demonstrated that activation of TNF-α/NF-κB signaling contributes to the maintenance of tumor-supportive microenvironmental interactions and promotes malignant progression in glioma [22]. Our findings are consistent with these observations and further support the importance of inflammatory signaling pathways in shaping GBM aggressiveness. The simultaneous enrichment of proliferative pathways, including E2F targets, G2M checkpoint signaling, and KRAS-associated programs, alongside inflammatory pathways, suggests that aggressive GBM phenotypes may arise through coordinated interaction between oncogenic proliferation and immune-related signaling. Elevated activity of EMT and hypoxia pathways in high-risk tumors further supports this interpretation. Mesenchymal transition programs have previously been associated with invasiveness, therapeutic resistance, stemness acquisition, and poor prognosis in glioblastoma [9, 23]. Similarly, hypoxia-driven transcriptional reprogramming is recognized as a major contributor to angiogenesis, metabolic adaptation, and immunosuppression within the tumor microenvironment [24]. Although immune subtype clustering revealed clear transcriptional differences between Immune_High and Immune_Low tumors, survival analysis did not demonstrate statistically significant differences between the two groups. This finding is consistent with growing evidence suggesting that immune activation alone may not necessarily predict a favorable prognosis in glioblastoma because the tumor immune microenvironment frequently remains predominantly immunosuppressive despite elevated inflammatory signaling [8, 25]. Tumor-associated macrophages, microglia, and regulatory immune populations may collectively support tumor progression even in immune-enriched tumors. Consequently, immune clustering alone may incompletely reflect the balance between anti-tumor and pro-tumor immune activity. The prognostic pathway analysis identified multiple hallmark pathways significantly associated with patient survival, including coagulation, complement signaling, EMT, KRAS signaling, and TNF-α/NF-κB activity. Increasing evidence suggests that coagulation-related signaling contributes directly to glioma aggressiveness through vascular remodeling, thrombosis-associated inflammation, and protease-activated signaling cascades [26]. Complement activation has also been linked to immunosuppressive microenvironmental remodeling and enhanced tumor progression in glioma [27]. Collectively, these observations indicate that coordinated inflammatory and stromal signaling programs may play central roles in determining GBM clinical behavior. Exploratory machine learning analysis identified TNF-α/NF-κB signaling, complement activation, KRAS signaling, and EMT-associated pathways among the most influential prognostic features. However, the identified pathways highlight biologically relevant signaling programs that may warrant further investigation as potential therapeutic targets. The predictive performance of the Random Forest and XGBoost classifiers remained limited. This result should be interpreted cautiously rather than overstated. Glioblastoma prognosis is influenced by multiple biological layers extending beyond transcriptomic pathway activity alone, including epigenetic regulation, genomic alterations, therapeutic response variability, and spatial tumor architecture. In addition, the relatively limited size of the TCGA-GBM cohort may restrict the performance of complex machine learning models. Nevertheless, the machine learning framework remained biologically informative by prioritizing pathways consistently associated with aggressive disease phenotypes. The prognostic risk signature derived from pathway-level features demonstrated more clinically meaningful performance. The model successfully stratified patients into high-risk and low-risk groups with significantly different survival outcomes, while multivariate Cox regression analysis confirmed that the calculated risk score retained independent prognostic significance after adjustment for clinical covariates. Importantly, external validation using the independent CGGA cohort substantially strengthened the reliability of the proposed model. The validated signature maintained significant survival stratification and demonstrated stable predictive performance across multiple follow-up intervals, indicating that the observed prognostic associations were reproducible beyond the original TCGA discovery cohort. Although the prognostic signature demonstrated consistent survival stratification in both the TCGA discovery cohort (n = 162) and the independent CGGA validation cohort (n = 313), differences in baseline survival distributions were observed between the two datasets. Moreover, the CGGA dataset contains more comprehensive clinicopathological information, including histological subtype, WHO grade, IDH mutation status, MGMT promoter methylation status, radiotherapy, and temozolomide treatment, all of which may substantially influence overall survival. Consequently, differences in median survival and predictive performance between the two cohorts are expected and most likely reflect differences in cohort composition and clinical characteristics rather than limitations of the proposed prognostic signature. Importantly, despite these differences, the pathway-based risk signature consistently discriminated high-risk and low-risk patients in both independent cohorts, supporting its robustness and generalizability. External validation is particularly important given the reproducibility concerns surrounding computational oncology studies. Many previously published prognostic signatures fail to maintain performance when tested in independent cohorts because of overfitting or cohort-specific bias [11]. In contrast, the consistent performance observed in the CGGA cohort suggests that the present pathway-based framework captures biologically meaningful features associated with glioblastoma progression. From a clinical perspective, the proposed signature may provide utility for prognostic stratification and risk-adapted therapeutic decision-making. The identified pathways highlight biologically relevant signaling programs that may warrant further investigation as potential therapeutic targets. Although the predictive accuracy of the model remains insufficient for standalone clinical implementation, integration with additional molecular variables such as MGMT promoter methylation, genomic alterations, radiomic features, and single-cell transcriptional data may further improve predictive capability in future studies. Several limitations should be acknowledged. The retrospective nature of public transcriptomic datasets introduces potential cohort bias and clinical variability. Furthermore, the present analysis was performed exclusively at the bulk transcriptomic level and therefore cannot fully resolve cellular heterogeneity within the tumor microenvironment. Functional experimental validation was also beyond the scope of this study. Future investigations integrating single-cell sequencing, spatial transcriptomics, and multi-omics profiling may provide deeper mechanistic insight into the biological pathways identified here. Prospective validation in clinically annotated cohorts will also be necessary before translation into routine clinical application. Despite the encouraging performance of the proposed pathway-based signature, translating computational biomarkers into routine clinical practice requires rigorous pathological validation, standardized diagnostic workflows, and prospective multicenter evaluation. As emphasized by Filippini et al., the clinical implementation of molecular biomarkers depends not only on computational performance but also on reproducible pathological assessment and expert validation across independent clinical settings. Therefore, further biological and clinical validation will be essential before the proposed signature can be considered for routine clinical application [28]. Limitations Despite the promising findings of this study, several limitations should be acknowledged. First, the analyses were based on retrospective transcriptomic datasets obtained from publicly available databases, which may be subject to inherent selection bias and differences in sample processing protocols. Second, although the prognostic signature demonstrated reproducible performance in an independent external CGGA cohort, additional validation in larger multi-center cohorts and prospective clinical studies is necessary to further establish its clinical utility and generalizability. Third, the present investigation relied primarily on computational and bioinformatics approaches, and therefore, the biological mechanisms underlying the identified hallmark pathways require experimental validation through both in vitro and in vivo studies. In addition, several well-established clinicopathological prognostic factors in glioblastoma, including MGMT promoter methylation status, IDH mutation status, extent of surgical resection, and adjuvant treatment information, were not incorporated into the multivariable prognostic model because these variables were unavailable or incomplete in the publicly available TCGA dataset used for model development. As these factors are recognized as important determinants of patient survival, their integration with pathway-based transcriptomic signatures may further improve prognostic accuracy and should be explored in future prospective studies with comprehensive clinical annotation. Furthermore, patients were stratified into high- and low-risk groups using the median risk score as a predefined cutoff because it represents an objective, unbiased, and widely adopted approach that generates balanced patient groups while minimizing the risk of overfitting. Although alternative cutoff strategies, such as maximally selected rank statistics or time-dependent ROC-based thresholds, may further optimize risk stratification, these approaches often require additional optimization and may reduce model generalizability. Future prospective studies involving larger, clinically well-annotated cohorts should compare different cutoff strategies to determine the most clinically informative threshold for pathway-based prognostic signatures in glioblastoma. Nevertheless, this study provides a comprehensive pathway-centered framework that captures immune-related hallmark pathway heterogeneity in glioblastoma and demonstrates the prognostic relevance of coordinated pathway activity beyond individual gene-based approaches. The successful external validation of the proposed risk signature highlights its potential robustness and clinical applicability. Future studies integrating multi-omics datasets, single-cell transcriptomics, spatial profiling technologies, comprehensive clinicopathological variables, and experimental validation may further refine the proposed model and contribute to the development of personalized prognostic and therapeutic strategies for glioblastoma patients. Conclusion This study provides a pathway-centered characterization of immune and oncogenic signaling heterogeneity in glioblastoma and establishes a prognostic risk signature derived from hallmark pathway activity profiles. The identified pathways revealed biologically meaningful interactions between inflammatory signaling, proliferative programs, mesenchymal transition, and tumor-associated immune activity. The resulting prognostic signature demonstrated significant survival stratification and retained independent prognostic value in both the TCGA discovery cohort and the externally validated CGGA cohort. The findings further support the concept that pathway-level analysis may provide a biologically interpretable framework for understanding glioblastoma progression beyond individual gene-based approaches. However, because the present study is based entirely on computational analyses, the proposed prognostic signature should be considered a research framework that requires prospective biological validation, functional experimental studies, and large multicenter clinical validation before routine clinical implementation. Collectively, these findings support the prognostic relevance of immune-related hallmark pathway activity in glioblastoma and provide a biologically interpretable foundation for future pathway-oriented prognostic modeling and therapeutic stratification. Future studies integrating comprehensive clinicopathological data, multi-omics approaches, and prospective validation cohorts will be essential to further evaluate the clinical utility and generalizability of the proposed signature. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Acknowledgements The authors gratefully acknowledge the Princess Nourah bint Abdulrahman University Researchers Supporting Project (Grant No. PNURSP2026R23), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia, for financial support of this research. The authors also acknowledge The Cancer Genome Atlas (TCGA), the Genomic Data Commons (GDC), the UCSC Xena platform, and the Chinese Glioma Genome Atlas (CGGA) for providing the publicly accessible transcriptomic, clinical, and survival datasets that made this study possible. Abbreviations Glioblastoma The Cancer Genome Atlas Chinese Glioma Genome Atlas RNA sequencing Single-sample gene set enrichment analysis Epithelial–mesenchymal transition Overall survival Receiver operating characteristic Area under the curve Hazard ratio Confidence interval Principal component analysis Gene set variation analysis Tumor necrosis factor Nuclear factor Kappa B Interleukin 6 Janus Kinase Signal transducer and activator of transcription 3 Kirsten Rat Sarcoma Viral Oncogene Homology Author contributions Conceptualization, A.M.A.; methodology, A.M.A.; software, A.M.A.; validation, A.M.A., M.F.E.-K., R.E., and A.E.A.; formal analysis R.E., A.M.A.; investigation, A.M.A.; resources, M.F.E.-K. and A.E.A.; data curation R.E, A.M.A.; writing—original draft preparation, A.M.A.; writing—review and editing, A.M.A., M.F.E.-K. and A.E.A.; visualization, A.M.A.; supervision, M.F.E.-K. and A.E.A.; project administration, A.M.A.; funding acquisition, M.F.E.-K. All authors have read and agreed to the published version of the manuscript. Funding This research was supported by the Princess Nourah bint Abdulrahman University Researchers Supporting Project (Grant No. PNURSP2026R23), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Data availability The datasets analyzed during the current study are available in public repositories. The discovery cohort comprised The Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM) dataset, accessible via the Genomic Data Commons (GDC) Data Portal (https://portal.gdc.cancer.gov/projects/TCGA-GBM) and the UCSC Xena platform (https://xenabrowser.net/). Independent external validation was performed using the Chinese Glioma Genome Atlas (CGGA) mRNAseq_325 cohort, accessible at http://www.cgga.org.cn/. No new sequencing, microarray, or other primary datasets were generated in this study; therefore, no new accession numbers are associated with this manuscript. Declarations Ethics approval and consent to participate This study was conducted in accordance with applicable research and publication guidelines. Ethical approval was not required because all transcriptomic and clinical datasets analyzed in this study were obtained from publicly available, de-identified databases. No human participants, animals, or personally identifiable information were directly involved in the present study.Not applicable. This study exclusively utilized publicly available de-identified datasets and did not involve the direct participation of human subjects Consent for publication Patient consent was waived because all transcriptomic and clinical datasets analyzed in the present study were obtained from publicly available databases and contained de-identified information. Institutional review board statement Ethical review and approval were waived for this study because all analyzed datasets were obtained from publicly available databases, including TCGA and CGGA, and no human participants or identifiable personal data were directly involved in the present study. Competing interests The authors declare that they have no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement The datasets analyzed during the current study are available in public repositories. The discovery cohort comprised The Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM) dataset, accessible via the Genomic Data Commons (GDC) Data Portal (https://portal.gdc.cancer.gov/projects/TCGA-GBM) and the UCSC Xena platform (https://xenabrowser.net/). Independent external validation was performed using the Chinese Glioma Genome Atlas (CGGA) mRNAseq_325 cohort, accessible at http://www.cgga.org.cn/. No new sequencing, microarray, or other primary datasets were generated in this study; therefore, no new accession numbers are associated with this manuscript.