Construction and validation of a model based on necroptosis-related genes for predicting glioma prognosis

Abstract

Background

Gliomas are highly aggressive brain tumors that pose significant challenges in terms of prognosis and treatment. The precise expression patterns and prognostic significance of necroptosis-related genes (NRGs) in gliomas remain poorly understood, despite necroptosis demonstrating promise as a therapeutic target in cancer.

Methods

We downloaded and preprocessed glioma datasets from TCGA (training cohort) and GEO (GSE43378, serving as an independent external validation cohort). An additional independent validation was performed using the Chinese Glioma Genome Atlas (CGGA) cohort. The KEGG, MSigDB, and GeneCards databases were utilized to develop an extensive list of NRGs. The necroptosis score (N-Score) and immune infiltration were assessed using single-sample GSEA. To identify prognostic genes, we performed WGCNA to identify necroptosis-related modules, followed by univariate Cox regression and LASSO-penalized multivariate Cox regression for gene selection.

Results

Glioma cases were assigned to high- and low-risk groups using the N-Score, and a marked difference in survival was observed (p < 0.001). We identified 11 prognostic genes (CASP6, DDIT3, FASN, FERMT1, GLUD1, GNLY, MYD88, NINJ1, PLA2G4A, PPIA, and TNFRSF1A) that were crucial for constructing a robust risk model that accurately predicts survival at one, two, and three years. Important hub genes, such as TNFRSF1A, were identified by the protein–protein interaction and mRNA–miRNA regulatory networks. These hub genes also showed strong correlations with different types of immune cells, such as myeloid-derived suppressor cells and central memory CD4 + T cells, suggesting that necroptosis may affect the tumor immune microenvironment.

Conclusion

NRGs are crucial in glioma prognosis and immune modulation. The risk model and identified hub gene TNFRSF1A provide valuable insights for further research and potential therapeutic targeting.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05414-1.

Introduction

Epidemiological data indicate that glioma, a form of tumor that develops in the central nervous system, accounts for 30% to 40% of all brain tumors [1]. The most aggressive subtype of gliomas is glioblastoma multiforme (GBM), with five-year survival rates below 5% [2]. Current therapy options are frequently hampered by the tumor’s highly infiltrative nature, resistance to traditional medicines, and the presence of the blood-brain barrier, which inhibits drug administration, resulting in considerable adverse effects and high recurrence rates [3, 4]. Therefore, novel therapeutic targets and predictive biomarkers are urgently needed to improve patient outcomes.

Necroptosis, a regulatory form of cell death, has recently received significant attention owing to its implications in cancer biology. Unlike apoptosis—a caspase-dependent, immunologically silent process that maintains membrane integrity and enables phagocytic clearance without inflammation—necroptosis is a caspase-independent pathway characterized by plasma membrane rupture and release of damage-associated molecular patterns (DAMPs), leading to potent inflammatory responses [5]. Although these pathways are molecularly distinct, they share overlapping signaling components, particularly at the receptor level (e.g., TNFRSF1A), through shared adaptor molecules (e.g., MYD88), and via certain downstream effectors (e.g., CASP6) that can participate in both apoptotic and necroptotic cascades depending on cellular context. This molecular crosstalk reflects the complex integration of cell death decisions, where the same initiating signal (e.g., TNFα engagement of TNFRSF1A) can trigger apoptosis, necroptosis, or survival pathways based on the availability of caspases, RIPK kinases, and other regulatory factors [6].

In gliomas, activation of necroptosis has been shown to augment anti-tumor immune responses and sensitize glioma cells to chemotherapeutic agents [7]. Moreover, the expression of necroptosis-related genes (NRGs) has been correlated with glioma grade and patient survival, suggesting their potential as prognostic biomarkers [8]. For instance, Zhou et al. and Yuan et al. independently constructed NRG-based models demonstrating significant prognostic value [9, 10]. However, these existing models show considerable heterogeneity in their gene compositions, limiting their clinical consensus and applicability. Importantly, while some studies primarily focus on predictive modeling, leaving the specific biological functions of the core prognostic NRGs, particularly their function in modulating the immunosuppressive glioma microenvironment, largely unexplored.

The immune microenvironment surrounding gliomas is instrumental in shaping both tumor progression and patient outcomes. Glioma cells interact with a variety of immune cell types, such as regulatory T cells, macrophages, and myeloid-derived suppressor cells (MDSCs), thereby enhancing both tumorigenesis and immune evasion [11]. According to research, certain small-molecule drugs can induce necroptosis in glioma cells, thereby enhancing immune cells’ capacity to identify and eradicate malignancies [12]. Therefore, understanding the mechanism of necroptosis and its influence on immune-cell infiltration may yield innovative therapies for gliomas. However, a systematic investigation into the relationship between NRG modules and immune infiltration is still lacking, which limits the translation of these findings. Furthermore, the regulatory mechanisms governing NRG expression in glioma—particularly at the post-transcriptional level via microRNAs (miRNAs)—remain largely unexplored. miRNAs are key epigenetic regulators that bind to target mRNAs and suppress their translation or promote degradation, thereby fine-tuning gene expression in response to cellular stress and microenvironmental cues [13]. Given that many NRGs are subject to miRNA-mediated regulation, and that miRNAs themselves are frequently dysregulated in glioma [14], constructing an mRNA-miRNA regulatory network can reveal upstream controllers of necroptosis signaling and identify potential therapeutic nodes.

We build a predictive model and use gene correlation analysis to investigate the relationship between glioma prognosis and necroptosis, thereby closing these gaps. We also examine the complex associations between these genes and infiltrating immune cells in gliomas. We hypothesize that the core genes in our signature are not merely prognostic markers but may also functionally modulate the glioma immune microenvironment, thereby providing valuable candidate targets for further mechanistic and therapeutic investigation.

Study methodology

Data acquisition

This study used a training-validation approach to develop and validate a prognostic model. The TCGA-GBMLGG cohort served as the training set for model development, two independent cohorts were used for external validation: the GSE43378 cohort from the GEO database and the CGGA cohort from the Chinese Glioma Genome Atlas (CGGA).

The glioma dataset (TCGA-GBMLGG), which served as a training set, was obtained via the R TCGAbiolinks database [15] to provide clinical data. A total of 699 glioma samples with World Health Organization (WHO) grade, age, and prognosis information were obtained. The UCSC Xena database provided all pertinent clinical data [16]. GEOquery in R (version 2.70.0) [17] was used to obtain the GSE43378 glioma dataset as a validation set from the GEO database [18]. The chip platform of the dataset GSE43378 [19] was GP570. We processed the GSE43378 dataset and standardized and normalized the annotation probes using limma in R (version 3.58.1) [20].

For further validation in an independent cohort of different ethnic background, the CGGA mRNAseq_693 dataset was obtained from the CGGA website (http://www.cgga.org.cn/). This dataset includes 693 glioma patients with RNA-seq data and corresponding clinical information, including survival data, age, gender, grade, and IDH mutation status. After quality control and removal of samples with missing survival data, a total of 657 patients were included in the final CGGA validation analysis.

NRGs were compiled from two databases: (1) KEGG; and (2) GeneCards, using keyword “necroptosis” with relevance score > 2.0. Genes were included if they were: (a) core necroptosis machinery components (RIPK1, RIPK3, MLKL) or their regulators; (b) experimentally validated to initiate or modulate necroptosis; or (c) involved in upstream signaling leading to necroptosis under caspase-inhibited conditions (e.g., TNFRSF1A). For genes with roles in multiple cell death pathways (e.g., TNFRSF1A, CASP6, MYD88), we required evidence of direct necroptosis involvement based on established criteria [5]: inhibition attenuates necroptosis, interaction with core necroptotic effectors, or requirement for necroptosome formation. Genes exclusively involved in other cell death forms were excluded. After deduplication, 271 NRGs were identified (Supplementary Table 1).

Calculation of the necroptosis score (N-Score) and weighted gene co-expression network analysis (WGCNA)

Using the expression matrix of NRGs and the glioma dataset from TCGA-GBMLGG, the single-sample gene set enrichment analysis (ssGSEA) algorithm was used in GSVA in R (version 1.50.0) [21] to calculate the necroptosis score (N-Score) for all samples. Afterward, Kaplan–Meier (KM) (21) analysis was performed using the N-Score data. Subsequently, WGCNA [22, 23] was conducted using WGCNA in R (version 1.72-5). To ensure that the network’s connections followed a scale-free distribution, the correlation coefficient between any two genes was first computed. The weighted correlation coefficient was then utilized. The correlation coefficients between genes were then used to build hierarchical grouping trees. The cluster tree’s many branches correspond to distinct gene modules, each represented by a different color. The significance of the modules was determined. The top 5000 genes with the highest variance were identified by computing variance for each gene in the glioma dataset (TCGA-GBMLGG). The scale-free fitting index was set to 0.87, the minimum distance was set to 0.2, the combined shear height of modules was set to 0.2, the optimal soft threshold was set to 9, and the minimum number of module genes was set to 100. The genes in each module were identified, and the association between the N-Score and the modules was assessed. Each module’s genes were considered module signature genes. Lastly, we looked at the relationship between the discovered modules and the N-Score. We identified genes that overlapped with NRGs in modules with |r| > 0.5.

Enrichment analyses

We conducted GO and KEGG enrichment analyses [24] of the module genes using clusterProfiler in R (version 4.10.0) [25], with the criteria of p < 0.05 and false discovery rate (FDR) value (q value) < 0.25.

Screening for prognostic genes in gliomas

To assess the impact of module genes on patient prognosis, we performed univariate Cox regression analyses. To account for multiple testing and reduce the risk of false-positive findings, we applied a stringent significance threshold (p < 0.01) for gene selection prior to LASSO regression. We next used glmnet in R (version 4.1-8) [26] to perform LASSO regression for all module genes with p-values < 0.01. This was done by designating the family parameter as “Cox,” indicating the Cox proportional hazards model, and setting the number of cycles to 10. The module genes identified through LASSO were determined to be prognostic indicators. In the end, the LASSO risk score (RS) was determined, allowing glioma specimens in the dataset to be classified into high- and low-risk groups.

Prognostic analysis of glioma

We constructed KM curves based on LASSO RS to evaluate variation in overall survival (OS) across glioma specimens categorized as high- or low-risk within the glioblastoma dataset (TCGA-GBMLGG). Furthermore, survivalROC in R (version 1.0.3.1) [27] was utilized to create a time-dependent receiver operating characteristic (ROC) curve. The area under the curve (AUC) was then computed for both LASSO RS and OS.

Assessment of the glioma prognostic risk model

First, RS, age, and WHO grade were used in a univariate regression analysis. Subsequently, factors highly associated with OS (p < 0.010) were identified for further study using multivariate Cox regression analysis (MCRA). A nomogram [28] based on the MCRA results was created using rms in R (version 1.1) to assess the interaction between RS and clinical factors in a multivariate Cox regression framework, particularly for survival outcomes at 1, 2, and 3 years.

Differential expression and correlation analysis of prognostic genes

To assess differences in the levels of the prognostic genes between the high-risk and low-risk glioma cohorts within the glioblastoma dataset (TCGA-GBMLGG), we created comparative maps. In addition, we evaluated the association among these prognostic gene levels in glioma specimens using the Spearman correlation method. This analysis was facilitated by pheatmap (version 1.0.12) in R. We identified the two prognostic genes with the most significant correlation and illustrated their relationship in a scatter plot generated with ggplot2 (version 3.4.4) in R.

Protein–protein interaction (PPI) network and mRNA-miRNA regulatory network

A useful tool for examining interactions between identified and predicted proteins is the STRING database [29]. To create a PPI network associated with prognostic genes, the current study set the confidence threshold to 0.150. Hub genes were then chosen for additional analysis based on their interactions with other genes in the PPI network. We collected miRNAs associated with the hub genes from starBase v3.0 [30] to evaluate associations between the hub genes and miRNAs.

Immune-cell infiltration analysis

We employed ssGSEA [31] to derive immune cell infiltration matrices for glioma specimens procured from the TCGA-GBMLGG dataset. Immune cell types that differed significantly between the two risk groups were chosen for further analysis, and Spearman’s correlation coefficients were used to assess the relationships between them. A heatmap was created using heatmap (version 1.0.12) in R, while ggplot2 (version 3.4.4) was utilized to generate a bubble plot illustrating gene-immune cell correlations.

To investigate the molecular differences between high-risk and low-risk groups beyond immune cell infiltration, we analyzed the expression of key inflammatory factors and immune checkpoint molecules. Gene expression data in FPKM (Fragments Per Kilobase Million) format were obtained from the TCGA-GBMLGG cohort. A comprehensive list of inflammatory factors (n = 18) was compiled based on literature review, including interleukins (IL1B, IL6, IL18, IL33), chemokines (CCL2, CCL5, CXCL8, CXCL10), tumor necrosis factor (TNF), and other inflammation-related genes (NFKB1, RELA, HMGB1, PTGS2, NOS2, ICAM1, VCAM1, SELE, SELP). Immune checkpoint molecules (n = 30) were selected based on established lists from previous studies [32], including the PD-1/PD-L1 pathway (PDCD1, CD274, PDCD1LG2), CTLA4 pathway (CTLA4, CD80, CD86), LAG3, TIM3 (HAVCR2), TIGIT, and other checkpoint regulators.

For each gene, expression levels were compared between high-risk and low-risk groups using the Wilcoxon rank-sum test. Fold changes were calculated as the ratio of mean expression in the high-risk group to that in the low-risk group. Genes with p < 0.05 were considered statistically significant, with significance levels indicated as *p < 0.05, **p < 0.01, and ***p < 0.001. Heatmaps were generated using the pheatmap package in R to visualize the expression patterns of all inflammatory factors and immune checkpoint molecules across risk groups, with rows scaled to Z-scores for better comparability.

Statistical analysis

Data were analyzed using R (version 4.2.2). Unless otherwise noted, independent-samples t-tests were used to compare normally distributed continuous variables, while non-normally distributed data were analyzed using Mann-Whitney U tests. Differences among multiple groups were assessed using Kruskal-Wallis tests. Associations were assessed using Spearman’s rank correlations. p-values < 0.05 were considered statistically significant, with all p-values being two-sided unless otherwise noted.

To evaluate the prognostic performance of our necroptosis-related risk model within the updated WHO classification framework, patients were stratified based on IDH mutation status and 1p/19q codeletion status obtained from the TCGA clinical data. According to the 2021 WHO guidelines, patients were classified into molecular subtypes: IDH-wildtype and IDH-mutant with 1p/19q codeletion (oligodendroglioma). Due to the absence of IDH-mutant without 1p/19q codeletion (astrocytoma) cases in our cohort with complete molecular data, this subtype was not analyzed. KM survival curves with log-rank tests were used to compare overall survival between high-risk and low-risk groups within each molecular subtype. Hazard ratios (HR) and 95% confidence intervals (CI) were calculated using univariate Cox proportional hazards regression models.

Results

Identification of module genes

To assess the clinical relevance of necroptosis in glioma, we first calculated an N-Score for each sample. Necroptosis was shown to be a prognostically meaningful process since patients with a high N-Score had a considerably lower overall survival than those with a low N-Score (p < 0.001; Fig. 1A).

Fig. 1.

Weighted gene co-expression network analysis (WGCNA) for TCGA-GBMLGG. a. Overall survival (OS) of the glioma group in the high and low N-Score survival Kaplan-Meier (KM) curves. b. Scale-free network representation of the optimal soft threshold from the WGCNA; the network connectivity under various soft threshold conditions is displayed in the right panel, while the optimal soft threshold is displayed in the left panel. c. The top 5000 genes with variance are shown in the module clustering results. d. The correlation analysis between N-Score and the top 5000 variance gene cluster modules is displayed. e. Clustering results of the top 5000 genes with variance are presented. Gene modules make up the lower portion, while a hierarchical clustering dendrogram makes up the upper portion. f-h. A total of 271 NRGs with blue (f), turquoise (g), and yellow (h) modules are shown in a Venn diagram

We computed the scale-free fitting index (Fig. 1B) at different soft-threshold levels to identify co-expression modules relevant to the glioma cohort in our dataset. This ensured that the created network more closely followed the concepts of scale-free topology. A fitting index of 0.87 was found to correspond to a minimal soft threshold of 9, considered ideal for building a scale-free network. The top 5000 DEGs were then clustered using a dendrogram (Fig. 1C), annotated with their corresponding groupings, and a co-expression network was built using the identified optimal soft threshold. The results indicated that the top 5000 genes with variance were divided into seven different modules using a screening criterion of 0.2: turquoise, gray, green, yellow, brown, blue, and red. In addition, we used the levels of the module-specific genes to evaluate the relationship between these seven modules and the glioma cohort’s N-Score (Fig. 1D). The top 5000 genes with variable expression were then subjected to clustering analysis, allowing visualization of connections between these genes and the combined modules (Fig. 1E). Lastly, three modules containing genes for further analysis were selected: blue, turquoise, and yellow, using the criterion |r| > 0.5. A Venn diagram was created from 271 NRGs, including all filter modules that intersected (Fig. 1F-H, Supplementary Table 2).

Functional enrichment analysis based on module genes

We performed GO and KEGG analyses to assess the functions of the 42 candidate module genes. The findings demonstrated the roles of these genes in controlling cell death and immune-inflammatory reactions (Fig. 2). These genes were significantly elevated in phrases such as “cytokine-mediated signaling pathway” and “regulation of I-κB kinase/NF-κB signaling” in biological processes (BP), suggesting a possible connection to tumor-associated inflammation. Concurrently, enrichment in ‘necroptotic process’ and ‘extrinsic apoptotic signaling pathway’ confirmed their core relevance to programmed cell death.

Fig. 2.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses for module genes. a-b. Bar (a) and bubble (b) charts display the GO and KEGG enrichment analyses of module genes: Biological Process (BP), Cellular Component (CC), Molecular Function (MF), and KEGG; GO and KEGG terms are on the ordinate. c-f. GO and KEGG enrichment analysis results of module genes: BP (c), CC (d), MF (e), and KEGG (f). The size of each dot represents the number of genes (Count) from the module gene list that are enriched in the corresponding term. The numerical values in the legend indicate the range of gene counts, with larger dots corresponding to higher gene counts

At the molecular function (MF) level, enrichments included ‘death receptor activity’ and ‘cytokine receptor binding’, aligning with the BP findings. Importantly, KEGG pathway analysis further substantiated these themes, with significant enrichment in the ‘NOD-like receptor signaling pathway’ (a key mediator of inflammation and pyroptosis) and, as expected, the ‘Necroptosis’ pathway itself.

Collectively, these enrichment results validate that our candidate module genes are not random but are functionally coherent, primarily converging on pathways governing immune regulation and necroptosis, which aligns perfectly with our investigative focus on the necroptosis-immune axis in glioma.

Prognostic genes

We used univariate Cox regression and LASSO-penalized multivariate Cox regression to extract the most reliable prognostic indicators from the 42 candidate module genes. Univariate Cox regression analysis of the 42 candidate module genes revealed that 37 genes were significantly associated with OS in glioma patients (P < 0.001). These 37 genes were subsequently included in multivariate Cox regression analysis to construct the prognostic model. This process yielded a compact signature comprising 11 prognostic genes: CASP6, DDIT3, FASN, FERMT1, GLUD1, GNLY, MYD88, NINJ1, PLA2G4A, PPIA, and TNFRSF1A (Fig. 3A-D).

Fig. 3.

Prognostic genes of glioma. a. Univariate Cox regression analysis identifying 37 module genes significantly associated with overall survival (P < 0.001) from an initial pool of 42 candidate genes. b-c. Plots of the LASSO regression model’s variable trajectories (c) and prognostic risk models (b). d. Multivariate Cox regression analysis of prognostic genes using a forest plot

Risk score and internal validation of the model

Each patient in the TCGA cohort had an RS determined using the algorithm based on the 11 prognostic genes:

RS = CASP6*0.0071 + DDIT3*0.0885 + FASN*(-0.0050) + FERMT1*(-0.1150) + GLUD1*(-0.4080) + GNLY*0.0419 + MYD88*0.3290 + NINJ1*0.0825 + PLA2G4A*0.1390 + PPIA*0.4060 + TNFRSF1A*0.0635.

This risk score served as a powerful prognostic classifier. Overall survival was significantly different between patients in the high- and low-risk groups based on the median RS (p < 0.001; Fig. 4A). With AUC values of 0.891, 0.896, and 0.915 for 1-, 2-, and 3-year survival, respectively, the time-dependent ROC analysis confirmed the signature’s robust, consistent predictive power across multiple time points (Fig. 4B).

Fig. 4.

Prognostic analysis and model internal validation. a. Prognostic KM curve between high- and low-risk score groups of LASSO and OS for glioma. b. ROC curves for glioma samples over time from the glioma dataset (TCGA-GBMLGG). c-d. The glioma prognostic risk model’s univariate and multivariate Cox regression analysis forest plots (c and d). e. Glioma prognostic risk model nomograms for the first, second, and third years. f-h. 1-year (f), 2-year (g), and 3-year (h) calibration curves of the glioma prognostic risk model

We next evaluated the clinical utility of the RS alongside established clinical factors. Univariate Cox regression identified the RS, age, and WHO grade as significant predictors (p < 0.001, Fig. 4C). Crucially, in multivariate analysis, the RS remained a strong and independent prognostic factor (p < 0.001, Fig. 4D), demonstrating that these clinical variables do not confound its predictive value.

We constructed a nomogram that combines RS and age. WHO grade to provide personalized estimates of 1-, 2-, and 3-year survival probabilities, thereby converting this model into a therapeutically valuable tool (Fig. 4E). The model’s dependability for clinical forecasting was demonstrated by calibration plots, which showed outstanding agreement between the nomogram’s predictions and observed outcomes, especially at the 2-year mark (Fig. 4F-H).

Given that the TCGA-GBMLGG cohort is an integrated dataset combining both low-grade glioma (LGG) and glioblastoma (GBM) patients, and considering the substantial prognostic differences between these two subtypes, we further evaluated the performance of our necroptosis-related risk model in each subtype separately. As shown in Supplementary Fig. 1, the risk model effectively stratified LGG patients into high-risk (n = 181) and low-risk (n = 349) groups with significantly different overall survival outcomes (log-rank p < 0.001). The median survival time for high-risk LGG patients was 606 days, compared to 1300.5 days for low-risk patients, demonstrating the excellent prognostic performance of our model in low-grade gliomas. Interestingly, in the GBM cohort, we observed that all 169 patients were classified into the high-risk group, with no patients assigned to the low-risk group. Therefore, Kaplan-Meier curve analysis could not be performed due to the absence of a low-risk group in this subtype.

External validation of the model

Glioma samples from the RS collaboration dataset GSE43378 were subjected to a prognostic KM curve analysis utilizing median OS grouping (Fig. 5A). Furthermore, we found that the association was highly significant (p < 0.001). At one, two, and three years, the glioma risk score demonstrated good predictive accuracy in the external validation cohort, with AUC values of 0.778, 0.859, and 0.884, respectively, with the highest accuracy observed at three years (Fig. 5B). The univariate Cox regression analysis revealed that RS and age were significant predictors of OS (p-value < 0.1; Fig. 5C). Subsequently, an MCRA was conducted (Fig. 5D). This showed that both RS and age were independent prognostic factors (p < 0.05). The nomogram also reveals that RS was markedly more effective than other clinical variables in assessing glioma prognosis (Fig. 5E). After calibrating the glioma prognostic risk model for one year (Fig. 5F), 2 years (Fig. 5G), and three years (Fig. 5H), we found that the model produced the most accurate clinical predictions at three years.

Fig. 5.

External validation of the prognostic model in the independent GSE43378 cohort. a. Prognostic KM curve between high- and low-risk score groups of LASSO and OS for glioma. b. Time-dependent ROC curve of glioma samples from the GSE43378 dataset. c-d. The glioma prognostic risk model’s univariate and multivariate Cox regression analysis forest plots (c and d). e. 1-, 2-, and 3-year nomograms of the glioma prognostic risk model. f-h. 1- (f), 2- (g), and 3-year (h) calibration curve of the glioma prognostic risk model

To further validate the robustness of our model in an independent cohort of different ethnic background, we applied the same risk score formula to the CGGA mRNAseq_693 dataset. After quality control, a total of 657 patients with complete survival data were stratified into high-risk (n = 328) and low-risk (n = 329) groups based on the median risk score of the CGGA cohort.

As shown in Supplementary Fig. 2A, patients in the high-risk group had significantly worse overall survival compared to those in the low-risk group (HR = 1.50, 95% CI 1.23–1.83, log-rank p = 6.31 × 10⁻⁵). Time-dependent ROC analysis demonstrated moderate predictive accuracy, with AUC values of 0.565, 0.608, and 0.591 for 1-, 2-, and 3-year survival, respectively (Supplementary Fig. 2B).

Univariate Cox regression analysis revealed that risk score, age, grade, and IDH mutation status were significantly associated with overall survival in the CGGA cohort (Supplementary Table 3). In multivariate Cox regression analysis adjusting for age, gender, grade, and IDH mutation status, the risk score remained an independent prognostic factor (HR = 1.002, 95% CI 1.001–1.003, p = 0.003; Supplementary Table 4). Grade (WHO III: HR = 2.82, 95% CI 2.01–3.94, p < 0.001; WHO IV: HR = 4.83, 95% CI 3.35–6.96, p < 0.001) and IDH mutation status (HR = 0.50, 95% CI 0.39–0.64, p < 0.001) were also independent predictors.

Calibration plots demonstrated acceptable agreement between predicted and observed survival probabilities at 1, 2, and 3 years (Supplementary Fig. 2C-E), indicating good clinical utility of the model across different populations.

Validation of the prognostic model in molecular subtypes defined by the 2021 WHO classification

To evaluate the performance of our necroptosis-related risk model within the updated WHO classification framework, we stratified patients based on IDH mutation status and 1p/19q codeletion status obtained from the TCGA clinical data. After matching samples and removing duplicates, a total of 672 unique patients with complete molecular and survival data were included in this subgroup analysis. According to the 2021 WHO guidelines, patients were classified into two major molecular subtypes: IDH-wildtype (n = 237) and IDH-mutant with 1p/19q codeletion (oligodendroglioma, n = 421). No patients with IDH-mutant without 1p/19q codeletion (astrocytoma) were identified in our cohort with complete molecular data.

In the IDH-mutant with 1p/19q codeletion subtype, the risk model effectively stratified patients into high-risk and low-risk groups with significantly different overall survival outcomes. As shown in Supplementary Fig. 3, patients in the high-risk group had significantly worse prognosis compared to those in the low-risk group (HR = 2.94, 95% CI 1.81–4.77, log-rank p = 4.67 × 10⁻⁶). In the IDH-wildtype subtype, due to the extremely low number of low-risk patients (n = 7), a meaningful Kaplan-Meier comparison could not be performed. This finding further supports the biological relevance of our necroptosis-related signature, which correctly identifies these aggressive tumors as uniformly high-risk.

Correlation of prognostic gene expression

We initially verified that expression of the 11 prognostic genes differed substantially between the high- and low-risk groups identified by our signature, to investigate whether these genes function in concert (Fig. 6A). We then performed a systematic correlation analysis to decipher their interrelationships.

Fig. 6.

Differential expression and correlation analysis of prognostic genes. a. Prognostic gene comparison between high- and low-risk groups in glioma samples from the Glioma dataset (TCGA-GBMLGG). b. Correlation heatmap of prognostic genes in high- and low-risk glioma sample groups. c. Correlation scatter plot of the prognostic genes CASP6 and TNFRSF1A. d. A scatter plot showing the correlation between the prognostic genes CASP6 and MYD88. e. Scatter plot of the association between prognostic genes CASP6 and FASN. f. Scatter plot of the association between prognostic genes (TNFRSF1A) and FASN. ***p < 0.001, highly statistically significant

Importantly, the analysis uncovered a tightly coordinated pro-inflammatory and cell death module. The executioner caspase CASP6 and the hub gene TNFRSF1A, a crucial modulator of TNF signaling, showed a substantial positive association (r = 0.776, p < 0.001; Fig. 6C), suggesting a possible functional linkage between death receptor signalling and caspase activation. This module was further reinforced by the strong positive correlation between MYD88 (a central adaptor in innate immune signaling) and CASP6 (r = 0.719, p < 0.001; Fig. 6D), suggesting cross-talk between immune activation and apoptotic/necroptotic pathways.

In contrast, these pro-death signals were inversely correlated with metabolic reprogramming. Both CASP6 and TNFRSF1A demonstrated significant negative correlations with FASN (Fatty Acid Synthase) (r = -0.684 and r = -0.716, respectively; p < 0.001, Fig. 6E, F), a key enzyme in lipid biosynthesis often upregulated in cancers. This inverse relationship suggests a possible trade-off or regulatory antagonism between lipid anabolic processes and cell death pathways within the glioma microenvironment.

The comprehensive correlation heatmap (Fig. 6B) visualizes distinct relational clusters, supporting the notion that the prognostic signature captures not only individual markers but also a network of biologically interacting genes.

PPI and mRNA-miRNA regulatory network

We conducted PPI and mRNA-miRNA network analysis to further confirm the biological coherence of the prognostic gene set and investigate its regulatory mechanisms.

The PPI interaction analysis showed that the following 10 prognostic genes (hub genes) were related: CASP6, DDIT3, FASN, GLUD1, GNLY, MYD88, NINJ1, PLA2G4A, PPIA, and TNFRSF1A (Fig. 7A). Importantly, TNFRSF1A is linked to eight other prognostic genes and is a key target in the PPI network. Subsequently, we constructed the mRNA–miRNA regulatory network, which illuminated the complex interactions governing glioma biology (Fig. 7B). A total of 34 miRNAs and seven hub genes were identified (Supplementary Table 5).

Fig. 7.

PPI network and mRNA-miRNA regulatory network. a. PPI network of prognostic genes. b. mRNA-miRNA regulatory network of hub genes. PPI, Protein–Protein Interaction. In the regulatory network diagram, mRNAs are shown in red and miRNAs in orange

Links between hub genes and immune cells

Given the central role of TNFRSF1A identified in our network analyses and its established function in immune regulation, we sought to dissect its specific relationship with the tumor immune landscape across different risk contexts.

When comparing high- and low-risk cohorts from the glioma dataset (TCGA-GBMLGG), the intergroup analysis shows that infiltration of 28 immune cell types differed markedly (p < 0.05) (Fig. 8A). The high-risk group exhibited significantly higher infiltration of immunosuppressive cells (MDSCs, Tregs, macrophages, all p < 0.001), while anti-tumor immune cells (CD8 + T cells, NK cells, all p < 0.001) were significantly reduced (Supplementary Table 6). T follicular helper (Tfh) cells and MDSCs showed the strongest positive connection (r = 0.888, p < 0.05) among the immune cell types in the low-risk glioma group (Fig. 8B). The majority of immune cells in the high-risk group again showed positive associations, with Tfh cells and MDSCs once more revealing the strongest positive association (r = 0.911, p < 0.05) (Fig. 8C). Furthermore, TNFRSF1A and MDSCs showed the highest positive association in the low-risk glioma samples (r = 0.652, p < 0.05) (Fig. 8D). On the other hand, TNFRSF1A and central memory CD4 + T cells showed the most marked positive association in the high-risk group (r = 0.776, p < 0.05) (Fig. 8E).

Fig. 8.

Risk group immune-cell infiltration analysis using the ssGSEA algorithm for TCGA-GBMLGG. a. Immune cell groupings in the glioblastoma dataset’s low- and high-risk glioma sample groups are compared (TCGA-GBMLGG). b-c. The low-risk (b) and high-risk (c) glioma groups’ immune-cell infiltration abundance correlation analysis results are displayed. d-e. Bubble plot showing the relationship between hub genes and immune-cell infiltration abundance in glioma samples from the low-risk (d) and high-risk (e) groups. **p < 0.01, highly statistically significant; ***p < 0.001, highly statistically significant

As shown in Supplementary Figure S4A and Supplementary Table S7, the high-risk group exhibited significantly higher expression of most pro-inflammatory factors compared to the low-risk group. Notably, CXCL10 showed the most dramatic increase (4.98-fold, p < 0.001), followed by IL6 (3.86-fold, p < 0.001), CXCL8 (2.96-fold, p < 0.001), and CCL5 (2.47-fold, p < 0.001). Other significantly upregulated inflammatory factors included CCL2, ICAM1, IL18, NFKB1, PTGS2, VCAM1, IL1B, SELP, and NOS2 (all p < 0.001). Only TNF showed no significant difference between risk groups (p = 0.383). These results indicate that the high-risk group is associated with a chronic inflammatory microenvironment, which may promote tumor progression and immune evasion.

Analysis of immune checkpoint molecules revealed widespread upregulation in the high-risk group (Supplementary Figure S4B, Supplementary Table S8). Among the 30 immune checkpoint genes analyzed, 27 showed significant differences between risk groups. The most striking differences were observed for CD70 (12.40-fold, p < 0.001), ICOS (7.22-fold, p < 0.001), and CD28 (3.69-fold, p < 0.001). Critically, multiple clinically relevant immune checkpoints were significantly upregulated in the high-risk group. Only three checkpoint genes (VSIR, TNFSF9) showed no significant difference, and two genes (HHLA2, TNFSF18) were downregulated in the high-risk group.

The upregulation of multiple immune checkpoints in the high-risk group suggests that these patients may be more susceptible to immune checkpoint blockade therapy, providing a potential therapeutic implication for our risk model. A heatmap summarizing the expression patterns of all inflammatory factors and immune checkpoints is shown in Supplementary Figure S4C.

Discussion

By developing an N-Score and using WGCNA to identify key gene modules, we systematically investigated the functions of NRGs in glioma prognosis and immune infiltration. Beyond confirming its association with immune-inflammatory pathways, we identify TNFRSF1A as a central hub gene with risk-stratified correlations with the tumor immune microenvironment. Consistent with the established link between necroptosis and inflammation, our functional enrichment analysis confirmed that the identified gene modules are strongly involved in cytokine-mediated signaling and the regulation of cell death pathways. This aligns with recent work by Yuan et al. and Zhou et al., who also developed necroptosis-related prognostic signatures for glioma, underscoring the clinical relevance of this pathway [9, 10]. However, our study extends these findings in several important directions. While previous studies focused on gene selection through direct regression against survival, our network-based, N-Score-guided approach identified a partially overlapping but distinct gene set and, more importantly, revealed the central network position of TNFRSF1A and its context-dependent immune correlations—aspects that were not explored in prior work.

Compared with similar studies (7, 8, 32), our prognostic model demonstrates strong performance, as evidenced by its time-dependent ROC analysis (AUC > 0.89). After validating the prognostic model using nomogram charts and GEO datasets, we conducted a correlation analysis of the prognostic genes to deepen our understanding of glioma biology. The association between high N-Score and poor prognosis, while seemingly counterintuitive given necroptosis’s proposed anti-tumor effects [7], aligns with the emerging understanding that necroptosis plays a dual, context-dependent role in cancer [33]. Several mechanisms may explain this paradox. First, chronic necroptotic signaling in the tumor microenvironment can promote immunosuppression through sustained DAMP release, leading to recruitment of MDSCs and regulatory T cells rather than effective anti-tumor immunity [34]. Second, components of the necroptosis machinery, such as TNFRSF1A and RIPK1, can exert non-cell-death functions including NF-κB activation and pro-survival inflammatory signaling, potentially uncoupling NRG expression from actual cell death execution [35]. Third, the balance between necroptosis and apoptosis may be critical; high NRG expression could indicate a state where apoptotic pathways are suppressed while necroptotic machinery is upregulated but incompletely executed [36]. Our observation that TNFRSF1A correlates with different immune populations in low-risk versus high-risk groups (MDSCs vs. central memory CD4 + T cells) supports this context-dependent model, suggesting that the immunological consequences of necroptosis-related gene expression vary with tumor context.

TNFRSF1A is a tumor necrosis factor receptor superfamily member. It regulates cell apoptosis and necrosis by activating several signaling pathways, including NF-κB and MAPK [37]. CASP6 is a caspase that is essential for signal transmission and effector activities during apoptosis [38]. The two genes showed a moderate positive correlation, indicating a potentially shared pathway or BP. Moreover, MYD88 is another gene that is positively correlated with CASP6. It is an important adaptor protein in the Toll-like receptor (TLR) axis, which is necessary for triggering innate and inflammatory reactions [39]. The positive correlation between MYD88 and CASP6 suggests that TLR signaling may modulate apoptotic pathways in glioma, thereby influencing tumor progression and response to therapy. FASN is a key enzyme in adipogenesis, and its elevated levels are often associated with a poor prognosis in various cancers, including glioma, as it provides lipids for membrane synthesis and energy storage in rapidly proliferating tumor cells [40]. The inverse relationship between FASN and the apoptotic genes CASP6 and TNFRSF1A suggests that lipid metabolism may counteract apoptotic signaling, thereby promoting tumor growth and resistance to cell death.

The remaining seven genes in our signature—DDIT3, FERMT1, GLUD1, GNLY, NINJ1, PLA2G4A, and PPIA—each contribute to the prognostic model through distinct but interconnected biological roles. DDIT3 (also known as CHOP) is a key mediator of endoplasmic reticulum stress-induced apoptosis and has been implicated in the crosstalk between ER stress and necroptosis [41]. In glioma, DDIT3 expression correlates with tumor grade and may modulate the unfolded protein response in the hypoxic microenvironment [42]. FERMT1 is involved in cell adhesion and integrin signaling, with emerging roles in regulating immune cell trafficking and cancer cell invasion [43]. GLUD1, a key metabolic enzyme, promotes cell death through reactive oxygen species production and activation of the p38/JNK MAPK signaling pathway, linking metabolic reprogramming to cell fate decisions [44]. GNLY (granulysin) is a cytolytic molecule expressed by cytotoxic T lymphocytes and NK cells, directly contributing to anti-tumor immune responses [45]. NINJ1 has recently been identified as a critical executor of plasma membrane rupture during necroptosis and other forms of lytic cell death, providing a direct mechanistic link between our signature and the necroptosis pathway [45]. PLA2G4A is involved in lipid metabolism and inflammatory signaling through eicosanoid production, and has been implicated in both ferroptosis and necroptosis [46]. PPIA (cyclophilin A) functions as a chemotactic factor for immune cells and may contribute to the recruitment of immunosuppressive populations such as MDSCs in the glioma microenvironment [47]. The diversity of functions represented—from metabolic enzymes to immune effectors to cell death executors—reflects the multifaceted nature of the necroptosis-immune axis in glioma. Understanding these pathways not only enhances our understanding of glioma pathogenesis but also identifies potential therapeutic targets and prognostic biomarkers.

PPI networks and mRNA-miRNA regulatory networks have been extensively studied in various cancers, revealing critical insights into tumor biology [48–50]. Beyond protein-level interactions, our mRNA-miRNA network adds a critical layer of regulatory insight by revealing how hub genes are controlled at the post-transcriptional level. The finding that TNFRSF1A is targeted by multiple miRNAs from the oncogenic miR-17-92 cluster [51] is particularly significant, as it suggests a mechanism by which proliferative signals (via miRNA upregulation) could simultaneously suppress necroptotic potential. This miRNA-mediated suppression may explain why some gliomas with high TNFRSF1A expression fail to execute necroptosis—the mRNA is present but translationally repressed. Similarly, the targeting of CASP6 by miR-26 family members links necroptotic capacity to p53 pathway status [52], providing a molecular bridge between two major tumor suppressor networks.

The biological significance of this regulatory layer extends to therapeutic considerations. miRNAs are attractive drug targets, with anti-miR therapies already in clinical development for cancer [53]. If validated experimentally, the miRNA-hub gene interactions identified here could inform strategies to enhance necroptotic sensitivity: for example, inhibiting miR-17-92 cluster members might derepress TNFRSF1A and restore necroptotic potential in resistant tumors. Conversely, miRNAs that suppress pro-survival genes (e.g., FASN-targeting miRNAs) could be therapeutically mimicked. Thus, our network not only maps regulatory relationships but also generates testable hypotheses for combination therapies that integrate necroptosis induction with miRNA modulation.

Similarly, we built a PPI network for glioma prognostic genes and identified 10 hub genes: CASP6, DDIT3, FASN, GLUD1, GNLY, MYD88, NINJ1, PLA2G4A, PPIA, and TNFRSF1A. These genes demonstrated significant interactions, highlighting their potential as critical nodes in glioma pathogenesis. Interestingly, we also found that TNFRSF1A interacts with 8 other hub genes, suggesting a potential role as a key regulatory factor in glioma development. The mRNA–miRNA regulatory network we constructed reveals the complex regulatory mechanisms underlying glioma biology. During apoptosis, GLUD1 promotes cell death by regulating reactive oxygen species production and activating the p38/JNK MAPK signaling pathway.

In contrast, FASN influences cell survival and death decisions by modifying lipid metabolism [44, 54]. Our research suggests that GLUD1 and FASN are not independent metabolic pathways; instead, they form an interdependent network regulated by miRNAs that affects the growth and survival of glioma cancer cells. Therefore, further research on the relationship between GLUD1 and FASN, along with their regulated miRNAs, will enhance our understanding of tumor metabolic reprogramming and help develop new targeted therapies.

Immune infiltration in gliomas is a critical factor influencing tumor progression and patient outcomes [55]. For instance, high levels of immune cell infiltration have been associated with poorer survival in glioma patients [56]. Our findings align with those of earlier studies, particularly highlighting the roles of Tfh cells and MDSCs. Tfh cells, a specific subset of CD4 + T cells, are crucial for forming germinal centers and enabling B cells to produce high-affinity antibodies. In gliomas, Tfh cells participate in immune responses against tumor cells; however, their precise impact on glioma progression and patient prognosis remains unclear. Recent research has suggested that Tfh cells can demonstrate both protumor and anti-tumor effects, depending on the TME [57, 58]. MDSCs are a diverse group of cells derived from myeloid progenitors and are known for their potent immunosuppressive properties. MDSCs can also bind to PD-1 on Tfh cells, thereby inhibiting T cell activation and promoting tumor growth and metastasis [59]. Tfh cells and MDSCs showed a substantial association in both risk categories, suggesting a persistent synergistic relationship that may be important for glioma progression. For instance, Tfh cells interact with MDSCs in the lymphoma microenvironment via CXCR5 and ICOS signaling, leading to MDSC accumulation and enhanced function. This process inhibits anti-tumor T cell activity and weakens the body’s anti-tumor immune response [60]. Furthermore, the correlation between TNFRSF1A and MDSCs in the low-risk group suggests that TNFRSF1A may regulate MDSC infiltration, consistent with evidence that TNFRSF1A can modulate immune responses and potentially alter the TME [61]. Central memory CD4 + T cells are vital for long-term immune surveillance, and their association with TNFRSF1A in the high-risk group suggests a potential mechanism by which gliomas may evade immune detection.

Overall, this research provides valuable insights into the role of necroptosis in gliomas. However, some limitations still exist. First, while we performed multivariable Cox regression analyses adjusting for age and WHO grade in both training and external validation cohorts, we were unable to include additional clinical and molecular variables such as gender, IDH mutation status, and 1p/19q codeletion status in a comprehensive multivariable model. This was due to incomplete availability of gender data in our datasets and missing molecular information for some patients, which would have substantially reduced the sample size and statistical power. Although our subgroup analysis according to the 2021 WHO classification demonstrated that the risk model maintains prognostic significance within molecularly defined subgroups, future studies with larger, well-annotated cohorts are needed to definitively confirm the independent prognostic value of our signature after adjusting for all established clinical and molecular factors. Second, while we validated our prognostic model in two independent external cohorts (GSE43378 and CGGA), further validation in larger, prospective, and more diverse multi-center cohorts is necessary to confirm its clinical generalizability and robustness across different populations, platforms, and treatment contexts. The successful validation in the CGGA cohort, comprising Chinese patients, partially addresses the need for cross-ethnic validation, but additional cohorts from other ethnic backgrounds would further strengthen the evidence. Third, the sample size, although sufficient for initial observations, may not be large enough to capture the full heterogeneity of gliomas. Fourth, although we identified TNFRSF1A as a central hub gene with context-dependent immune correlations, wet-lab experimental validation of its biological functions is urgently needed. The current study is primarily bioinformatics-based and lacks validation using glioma cell lines or immunohistochemistry cohorts from independent institutions. Such experimental validation would be essential to: (1) confirm the functional role of TNFRSF1A in glioma cell proliferation, migration, and invasion; (2) validate its involvement in necroptosis signaling under controlled conditions; and (3) assess its protein-level expression patterns in clinical tissue samples. We acknowledge this as a major limitation and are actively planning collaborative studies to perform these experiments in the future. To overcome these constraints, subsequent studies should include adequate experimental validation and conduct thorough clinical assessments.

Conclusion

In conclusion, by integrating multi-omics analyses, we have not only developed and confirmed a robust necroptosis-related prognostic signature for glioma but have also transcended mere prediction by identifying TNFRSF1A as a potential master regulator at the nexus of necroptosis signaling and immune modulation. Our findings propose a novel model in which TNFRSF1A function may dynamically shape the immune landscape in a risk-dependent manner, contributing to the disparate outcomes captured by our signature. By identifying TNFRSF1A as a strong contender for future functional research and a possible target for combinatory immunotherapy approaches in glioma, this work shifts the emphasis from correlation to mechanism.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Kaixiang Wu: Conceptualization; Methodology, Validation; Writing–original draft; Writing–review and editing. Lu Cao: Conceptualization; Methodology; Writing–review and editing. Hongchuan Yang: Conceptualization; Software; Validation; Writing–review and editing. Wenwen Sun: Formal Analysis; Writing–review and editing. Kaifu Wu : Supervision; Writing–review and editing. All authors have read and agreed to the published version of the manuscript.

Funding

Not applicable.

Data availability

The datasets generated and/or analysed during the current study are available from Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/), Gene Expression Omnibus (GEO: ID GSE43378) database (https://www.ncbi.nlm.nih.gov/geo/), GeneCards ( [https://www.genecards.org/](https:/www.genecards.org) ) and Kyoto Encyclopedia of Genes and Genomes (KEGG) ( [https://www.gsea-msigdb.org/gsea/msigdb](https:/www.gsea-msigdb.org/gsea/msigdb) ) databases. Data is provided within the manuscript or supplementary information files. The data used to support the findings of this study are available from the corresponding authors upon request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have read and agreed to the published version of the manuscript.

Competing interests

The authors declare no competing interests.